IJEST Special Issue Complete

MultiCraft
ISSN 2141-2839 (Online); ISSN 2141-2820 (Print)

Available online at www.ijest-ng.com
International Journal
of Engineering,
Science and
Technology
Special Issue:
Application of Computational Intelligence in Emerging Power Systems
Guest Editors:
S.N. Singh
Department of Electrical Engineering, Indian Institute of Technology Kanpur
Kanpur- 208016, India
K.S. Verma
Department of Electrical Engineering, Kamla Nehru Institute of Technology
Sultanpur- 228118, India
Vol. 2, No. 3, 2010
International Journal of Engineering,
Science and Technology (IJEST)
Aims and scope
IJEST is an international refereed journal published by MultiCraft. The aim and scope of the journal is to provide an academic
medium and an important reference for the advancement and dissemination of research results that support high-level learning,
teaching and research in the fields of engineering, science and technology. Original theoretical work and application-based studies,
which contributes to a better understanding of engineering, science and technological challenges, are encouraged.
Journal policy
International Journal of Engineering, Science and Technology (IJEST) publishes articles that emphasize research, development
and application within the fields of engineering, science and technology. All manuscripts are pre-reviewed by the editor, and if
appropriate, sent for blind peer review. Contributions must be original, not previously or simultaneously published elsewhere, and
are critically reviewed before they are published. Papers, which must be written in English, should have sound grammar and
proper terminologies.
Call for papers
We invite you to submit high quality papers for review and possible publication in all areas of engineering, science and
technology. All authors must agree on the content of the manuscript and its submission for publication in this journal before it is
submitted to us. Manuscripts should be submitted by e-mail to the Editor at: sa_oke@yahoo.com
Call for Reviewers
Scholars interested in serving as volunteer reviewers should indicate interest by sending their full curriculum vitae to us.
Reviewers determine submissions that are of quality. Since they are expected to be experts in their areas, they should comment on
the significance of the reviewed manuscript and whether the research contributes to knowledge and advances both theory and
practice in the area.
Editor: S.A. Oke, PhD, Department of Mechanical Engineering, University of Lagos, Nigeria
E-mail : sa_oke@yahoo.com
Associate Editor (Electrical Engineering): S.N. Singh, PhD, Department of Electrical Engineering, Indian Institute of Technology
Kanpur, Kanpur-208016, India
EDITORIAL BOARD MEMBERS
Kyoji Kamemoto (Japan)
M. Abdus Sobhan (Bangladesh)
Sri Niwas Singh (India)
Shashank Thakre (India)
Jun Wu (USA)
Jian Lu (USA)
Raphael Jingura (Zimbabwe)
V. Sivasubramanian (India)
Shaw Voon Wong (Malaysia)
S. Karthikeyan (Sultanate of Oman)
Amir Nassirharand (Malaysia)
N. W. Ingole (India)
Fatih Camci (Turkey)
Milorad Bojic (Serbia)
Asim Kumar Pal (India)
Prasanta Sahoo (India)
Vidosav D. Majstorovich (Serbia)
K. Somasundaram (India)
Shashi Anand (India)
Ian Blenkharn (UK)
Petr Konas (Czech Republic)
Angelo Basile (Italy)
Syed Asif Raza (Qatar)
Atif Iqbal (India)
Kampan Mukherjee (India)
P.K. Kapur (India)
Haitao Huang (Hong Kong)
P. Thangavelu (India)
Yechun Wang (USA)
Abdul Ravoof Shaik (Australia)
Alistair Thompson McIlhagger (UK)
K.I. Ramachandran (India)
Jian Ma (USA)
P.K. Tripathy (India)
Alan Rennie (UK)
J. Paulo Davim (Portugal)
MKS Sastry (West Indies)
Rajneesh Talwar (India)
A. Moreno-Muoz (Spain)
Prabin K Panigrahi (India)
Vinay Gupta (India)
Mohammed Al-Nawafleh (Jordan)
Bhu Dev Sharm (India)
Tien-Fu Liang (Taiwan)
Ranjit Kumar Biswas (Bangladesh)
Siba Sankar Mahapatra (India)
Sangeeta Sahney (India)
Kuang-Yuan Kung (Taiwan)
Eleonora Bottani (Italy)
Evangelos J. Sapountzakis (Greece)
A.M. Rawani (India)
Saurabh Mukherjee (India)
P. Dhavachelvan (India)
A. Bandyopadhyay (India)
Velusamy. Sivasubramanian (India)
S. Vinodh (India)
Vctor Hugo Hinojosa Mateus (Chile)
REVIEWERS
The following reviewers have greatly helped us in reviewing our manuscripts and have brought such submissions to high quality
levels. We are indebted to them.
Marcus Bengtsson (Sweden)
Kit Fai Pun (West Indies)
Peter Koh (Australia)
Erhan Kutanoglu (USA)
Jayant Kumar Singh (India)
Maneesh Singh (Norway)
RRK Sharma (India)
Jamil Abdo (Oman)
Agnes S. Budu (Ghana)
Yuan-Ching Lin (Taiwan)
Withaya Puangsombut (Thailand)
Abd Rahim Abu Bakar (Malaysia)
Fakher Chaari (Tunisia)
Ghosh Surojit (India)
Umut Topal (Turkey)
Maloy Singha (India)
Parviz Malekzadeh (Iran)
G. Possart (Germany)
Masoud Rashidinejad (India)
Vera Meshko (Republic of Macedonia)
Jun Luo (China)
Uday Kumar (Sweden)
Tamer Samir Mahmoud (Egypt)
Arijit Bhattacharya (Ireland)
M.R. Sharma (India)
Hyung Hee Cho (Korea)
Souwalak Phongpaichit (Thailand)
Elsa Rueda (Argentina)
Ming-Kuang Wang (Taiwan)
Ruey-Shin Juang (Taiwan)
Marisa Viera (Argentina)
Shiguo Jia (China)
S. Devasenapati Babu (India)
Rajeeb Dey (India)
Subrata Kumar Ghosh (India)
Timothy Payne (Australia)
Diwakar Tiwari (India)
Mustafa Soylak (Turkey)
Jerzy Merkisz (Poland)
Md Fahim Ansari (India)
Jiun-Hung Lin (Taiwan)
Tzong-Ru Lee (Taiwan)
Subir Kumar Sarkar (India)
Kee-hung Lai (Hong Kong)
Jochen Smuda (Switzerland)
Roland Hischier (China)
Ahmed Abu-Siada (Australia)
Hamzah Abdul Rahman (Jordan)
Chih-Huang Weng (Taiwan)
Yenming Chen (Taiwan)
Dinesh Verma (USA)
Devanandham Henry (USA)
M. Habibnejad Korayem (Iran)
Radu Radescu (Romania)
Hsin-Hung Wu (Taiwan)
Amy Trappey (Taiwan)
A.B. Stevels (Netherlands)
Liang-Hsuan Chen (Taiwan)
Richard Hischier (Switzerland)
Shyi-Chyi Cheng (Taiwan)
Andrea Gerson (Australia)
Ingrid Bouwer Utne (Norway)
Maruf Hossain (Bangladesh)
Enso Ikonen (Finland)
Kwai-Sang Chin (Hong Kong)
Jiunn I Shieh (Taiwan)
Hung-Yan Gu (Taiwan)
Pengwei (David) Du (US)
Min-Shiang Hwang (Taiwan)
Ekata Mehul (India)
Shashidhar Kudari (India)
Khim Ling Sim (USA)
Rong-Jyue Fang (Taiwan)
Chandan Guria (India)
Rafael Prikladnicki (Brazil)
Juraj Kralik (Slovak)
Indika Perera (Sri Lanka)
R K Srivastava (India)
Ramakrishnan Ramanathan (UK)
Suresh Premil Kumar (India)
Fernando Casanova Garca (Colombia)
J. Ashayeri (The Netherlands)
Siddhartha Kumar Khaitan (USA)
Jim Austin (UK)
Rafael Prikladnicki (Brazil)
V. Balakrishnan (India)
P. Dhasarathan (India)
R. Venckatesh (India)
Sarmila Sahoo (India)
Avi Rasooly (USA)
Barbara Bigliardi (Italy)
Huiling Wu (China)
Ahmet N. Eraslan (Turkey)
Ryoichi Chiba (Japan)
P.K.Dutta (India)
Kimon Antonopoulos (Greece)
Josefa Mula (Spain)
Amiya Ku.Rath (India)
Fabio Leao (Brazil)
Francisco Jesus Fernandez Morales (Spain)

MultiCraft

International Journal of Engineering, Science and Technology
Vol. 2, No. 3, 2010, pp. i-ii

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com
2010 MultiCraft Limited. All rights reserved

Editorial

Due to increased interconnection and loading along with deregulation and environmental concerns, electric power systems,
around the world, are changing in terms of structure, operation, management and ownership. This makes the electric power system
complex, heavily stressed and thereby vulnerable to cascade outages. Electric power utilities are trying to provide intelligent and
smart solutions with economical, technical (secure, stable and good power quality) and environmental goals. There are several
challenging issues in the smart grid solutions such as, but not limited to, forecasting of load, price, ancillary services; penetration
of new and renewable energy sources; bidding strategies of participants; power system planning & control; operating decisions
under missing information; increased distributed generations and demand response in the electric market; tuning of controller
parameters in varying operating conditions, etc.
The conventional methods in solving the power system design, planning, operation and control problems have been very
extensively used for different applications but these methods suffer from several difficulties due to necessities of derivative
existence, providing suboptimal solutions, etc. Computational intelligence (CI) is a new and modern tool for solving complex
problems which are difficult to be solved by the conventional techniques. Computation intelligent methods can give better solution
in several conditions and being widely applied in the electrical engineering applications.
In this special issue, sixteen papers, which are focusing on various areas of emerging power systems such as improving power
transfer capability, transformer protection, multi-area economic dispatch, power system security, distribution system planning and
reliability, power quality, shipboard power systems, automatic generation control, flexible ac transmission systems, collaborative
control, etc., are considered. These papers utilize and address the various computational intelligence techniques such as artificial
neural networks, genetic algorithms, particle swarm optimization, evolutionary techniques, ant colony and harmony searches,
fuzzy systems, etc. First paper summarizes the various power system problems and potential CI techniques to solve those
problems. Key issues and challenges in power quality problems are discussed in detail in one of the papers. These papers will be
very useful to the researchers and academicians, but not limited to, in carrying out the research in power systems using
computational intelligence.
Being guest editors of the special issue on Application of Computational Intelligence in Emerging Power Systems for the
International Journal of Engineering, Science and Technology (IJEST), we would like to thank all the authors for publishing their
papers, all the reviewers for reviewing the papers of this special issue and the Editor, IJEST, for encouragement.

Prof. S.N. Singh and Prof. K.S. Verma
Guest Editors,
Application of Computational Intelligence in Emerging Power Systems

ii

Sl.
No.
Title of paper Page
No.
1 Editorial
S.N. Singh, K.S. Verma
i
2 Application of computational intelligence in emerging power systems
D. Saxena, S.N. Singh, K.S. Verma
1-7
3 Application of artificial neural networks to improve power transfer capability through OLTC
A. Abu-Siada, S. Islamand, E.A. Mohamed
8-18
4 Application of radial basis neural network for state estimation of power system networks
J.P. Pandey, D. Singh
19-28
5 Improved transformer protection using probabilistic neural network and power differential method
Manoj Tripathy, R. P. Maheshwari, H. K. Verma
29-44
6 Intelligent agents under collaborative control in emerging power systems
H.F. Wedde, S. Lehnhoff, C. Rehtanz, O. Krause
45-59
7 Requirement analysis for autonomous systems and intelligent agents in future Danish electric power
systems
Arshad Saleem, Morten Lind
60-68
8 Influence of TCSC on social welfare and spot price- A comparative study of PSO with classical
method
S. K. Joshi, K. S. Pandya

69-81
9 Optimal trading strategy for GenCo in LMP-based and bilateral markets using self-organising
hierarchical PSO
C. Boonchuay, W. Ongsakul, J. Zhong, F. F. Wu
82-93
10 Allocation of optimal distributed generation using GA for minimum system losses in radial
distribution networks
T.N. Shulka, S.P. Singh, K.B. Naik
94-106
11 Distribution system reliability evaluation considering fuzzy factors
Xufeng Xu,

Joydeep Mitra

107-118
12 Application of ant colony optimization for reconfiguration of shipboard power system
Sri Hari Krishna Vuppalapati,

Anurag K. Srivastava
119-131
13 Multi-area economic dispatch with tie-line constraints employing evolutionary approach
Manisha Sharma , Manjaree Pandit, Laxmi Srivastava
132-149
14 Overall performance evaluation of evolutionary designed conventional AGC controllers for
interconnected electric power system studies in a deregulated market environment
Y.L. Karnavas, K.S. Dedousis
150-166
15 A modified harmony search based method for rural radial line planning
Yan Yang, Wenxin Guo, Fushuan Wen, Danyue Wu, Yan Lin
167-177
16 Supervised fuzzy C-means clustering technique for security assessment and classification in power
systems
S. Kalyani, K.S. Swarup
175-185
17 Power quality event classification: Overview and key issues
D. Saxena, K.S. Verma, S.N. Singh
186-199

MultiCraft

Vol. 2, No. 3, 2010, pp. 1-7

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Application of computational intelligence in emerging power systems

D. Saxena
*
, S.N. Singh
+1
and K.S. Verma
#

*
Department of Electrical and Electronics Engineering, Invertis Inst. of Engg.& Tech., Bareilly (UP), INDIA.
+
Department of Electrical Engineering, CET, Denmark Technical University, Kgs. Lyngby, DENMARK.
#
Department of Electrical Engineering, K.N.I.T Sultanpur (UP), INDIA.
E-mails: diptisx@gmail.com (D. Saxena), snsingh@iitk.ac.in (S.N. Singh
1
,
1
Corresponding author), ksv02@rediffmail.com (K.S. Verma)

Abstract

Electric power systems, around the world, are changing in terms of structure, operation, management and ownership due to
technical, financial and ideological reasons. Power system keeps on expanding in terms of geographical areas, assets additions,
and penetration of new technologies in generation, transmission and distribution. This makes the electric power system complex,
heavily stressed and thereby vulnerable to cascade outages. The conventional methods in solving the power system design,
planning, operation and control problems have been very extensively used for different applications but these methods suffer
from several difficulties due to necessities of derivative existence, providing suboptimal solutions, etc. Computation intelligent
(CI) methods can give better solution in several conditions and are being widely applied in the electrical engineering
applications. This paper highlights the application of computational intelligence methods in power system problems. Various
types of CI methods, which are widely used in power system, are also discussed in the brief.

Keywords: Power systems, computational intelligence, artificial intelligence.

1. Introduction

Increased interconnection and loading of the power system along with deregulation and environmental concerns has brought new
challenges for electric power system operation, control and automation. In liberalized electricity market, the operation and control
of power system become complex due to complexity in modeling and uncertainties. Power system models used for intelligent
operation and control are highly dependent on the task purpose. In competitive electricity market along with automation,
computational intelligent techniques are very useful. As electric utilities are trying to provide smart solutions with economical,
technical (secure, stable and good power quality) and environmental goals, there are several challenging issues in the smart grid
solutions such as, but not limited to, forecasting of load, price, ancillary services; penetration of new and renewable energy
sources; bidding strategies of participants; power system planning & control; operating decisions under missing information;
increased distributed generations and demand response in the electric market; tuning of controller parameters in varying operating
conditions, etc. Risk management and financial management in electric sector are concerned with finding an ideal trade-off
between maximizing the expected returns and minimizing the risks associated with these investments.
Computational intelligence (CI) is a new and modern tool for solving complex problems which are difficult to be solved by the
conventional techniques. Heuristic optimization techniques are general purpose methods that are very flexible and can be applied
to many types of objective functions and constraints. Recently, these new heuristic tools have been combined among themselves
and new methods have emerged that combine elements of nature-based methods or which have their foundation in stochastic and
simulation methods. Developing solutions with these tools offers two major advantages: development time is much shorter than
when using more traditional approaches, and the systems are very robust, being relatively insensitive to noisy and/or missing
data/information known as uncertainty.
Due to environmental, right-of-way and cost problems, there is an increased interest in better utilization of available power
system capacities in both bundled and unbundled power systems. Patterns of generation that results in heavy flows, tend to incur

1
Corresponding author. Tel: +91-512-2597009
Saxena et al. / International Journal of Engineering, Science and Technology, Vol. 2 No. 3, 2010, pp. 1-7

2

greater losses, and to threaten stability and security, ultimately make certain generation patterns economically undesirable. Hence,
new devices and resrources such as flexible ac transmission systems (FACTS), distributed generations, smart grid technologies,
etc. are being utilized. In the emerging area of power systems, computation intelligence plays a vital role in providing better
solutions of the existing and new problems. This paper lists various potential areas of power systems and provides the roles of
computational intelligence in the emerging power systems. A brief review of computational techniques is also presented.

2. Potential Area of Research in Power System Using Computational Intelligence

There are several problems in the power systems which cannot be solved using the conventional approaches as these methods are
based on several requirements which may not be true all the time. In those situations, computational intelligence techniques are
only choice however these techniques are not limited to these applications. The following areas of power system utilize the
application of computational intelligence.
Power system operation (including unit commitment, economic dispatch, hydro-thermal coordination, maintenance
scheduling, congestion management, load/power flow, state estimation, etc.)
Power system planning (including generation expansion planning, transmission expansion planning, reactive power
planning, power system reliability, etc.)
Power system control (such as voltage control, load frequency control, stability control, power flow control, dynamic
security assessment, etc.)
Power plant control (including thermal power plant control, fuel cell power plant control, etc.)
Network control (location and sizing of facts devices, control of facts devices, etc.)
Electricity markets (including bidding strategies, market analysis and clearing, etc.)
Power system automation (such as restoration and management, fault diagnosis and reliability, network security, etc.)
Distribution system application (such as operation and planning of distribution system, demand side management &
demand response, network reconfiguration, operation and control of smart grid, etc.)
Distributed generation application (such as distributed generation planning, operation with distributed generation, wind
turbine plant control, solar photovoltaic power plant control, renewable energy sources, etc.)
Forecasting application (such as short term load forecasting, electricity market forecasting, long term load forecasting,
wind power forecasting, solar power forecasting, etc.)

Several research papers are published in various journals and conferences. Some conferences in the power system areas are
completely dedicated to intelligent system applications and organized regularly such as intelligent system application to power
systems (ISAP) held alternate years in different locations of the world. Power system computing conference is another very
important conference held once in three years. Similarly, several reputed journals are dedicated to CI applications in the field of
engineering and science. There are several books which address of CI application in power systems (Fogel et al, 1966; Sobajic,
1993; Song, 1996; Warvick, 1997; El-Hawary, 1998; Lai, 1998; Wehnekel, 1998; Momoh, 2000).

3. Various Computational Intelligence Techniques

Computational intelligence (CI) methods, which promise a global optimum or nearly so, such as expert system (ES), artificial
neural network (ANN), genetic algorithm (GA), evolutionary computation (EC), fuzzy logic, etc. have been emerged in recent
years in power systems applications as effective tools. These methods are also known as artificial intelligence (AI) in several
works. In a practical power system, it is very important to have the human knowledge and experiences over a period of time due to
various uncertainties, load variations, topology changes, etc. This section presents the overview of CI/AI methods (ANN, GA,
fuzzy systems, EC, ES, ant colony search, Tabu search, etc.) used in power system applications.

3.1 Artificial Neural Networks

An artificial neural network (ANN) is an information-processing paradigm that is inspired by the biological nervous systems,
such as the brain, process information (Bishop, 1995). The key element of this paradigm is the novel structure of the information
processing system composed of a large number of highly interconnected processing elements (neurons) working in unison to solve
the specific problems. ANNs, like people, learn by example. The starting point of ANN application was the training algorithm
proposed and demonstrated, by Hebb in 1949, how a network of neurons could exhibit learning behaviour. During the training
phase, the neurons are subjected to a set of finite examples called training sets, and the neurons then adjust their weights according
to certain learning rule. ANNs are not programmed in the conventional sense, rather they learn to solve the problem through
interconnections with environment. Very little computation is carried out at the site of individual node (neuron). There is no
explicit memory or processing locations in neural network but are implicit in the connections between nodes. Not all sources of

3

input feeding a node are of equal importance. It all depends on weight which can be negative or positive. Inputs arriving at a node
are transformed according to the activation function of node.

The main advantages of ANNs are as follows:
ANNs with suitable representation can model any degree of non-linearity and thus, are very useful in solving the non-
linear problems.
These do not require any apriori knowledge of system model.
ANNs are capable of handling situations of incomplete information, corrupt data and they are fault tolerant.
Massive parallelism provided by the large number of neurons and the connections amongst them provide good search and
distributed representations.
ANN is fast and robust. It possesses learning ability and adapts to the data.

Though the neural network (NN) training is generally computationally expensive, it takes negligible time to evaluate correct
outcome once the network has been trained. Despite the advantages, some disadvantages of the ANN are: (i) large dimensionality,
(ii) selection of the optimum configuration, (iii) choice of training methodology, (iv) the black-box representation of ANN they
lack explanation capabilities and (v) the fact that results are always generated even if the input data are unreasonable. Another
drawback of neural network systems is that they are not scalable i.e. once an ANN is trained to do certain task, it is difficult to
extend for other tasks without retraining the NN. Artificial neural networks are most promising method for many power system
problems and have been used for several applications.
ANNs are mainly categorized by their architecture (number of layers), topology (connectivity pattern, feed forward or recurrent,
etc.) and learning regime. Based on the architecture ANN model may be single-layer ANN which includes perceptron model
(suggested by Rosenblot, in 1959) and ADALINE (suggested by Widrow & Hoff in 1960). ANN model can be further categorized
as Feed forward NN and Feed Backward NN based on neuron interactions. Learning of ANN may be Supervised Learning,
Unsupervised Learning, and Reinforcement Learning. Based on neuron structure ANN model may be classified as multilayer
perceptron model, Boltzman machine, cauchy machine, Kohonen self-organizing maps, bidirectional associative memories,
adaptive resonance theory-I (ART-1), adaptive resonance theory-2 (ART-2), counter propagation ANN. Some other special ANN
models are parallel self-hierarchical NN, recurrent NN, radial basis function NN, knowledge based NN, hybrid NN, wavelet NN,
cellular NN, quantum NN, dynamic NN, etc.

3.2 Genetic Algorithms

Genetic algorithm (GA) is an optimization method based on the mechanics of natural selection and natural genetics. Its
fundamental principle is the fittest member of population has the highest probability for survival. The most familiar conventional
optimization techniques fall under two categories viz. calculus based method and enumerative schemes. Though well developed,
these techniques possess significant drawbacks. Calculus based optimization generally relies on continuity assumptions and
existence of derivatives. Enumerative techniques rely on special convergence properties and auxiliary function evaluation. The
genetic algorithm, on the other hand, works only with objective function information in a search for an optimal parameter set. The
GA can be distinguished from other optimization methods by following four characteristics.
(i) The GA works on coding of the parameters set rather than the actual parameters.
(ii) The GA searches for optimal points using a population of possible solution points, not a single point. This is an important
characteristic which makes GA more powerful and also results into implicit parallelism.
(iii) The GA uses only objective function information. No other auxiliary information (e.g. derivatives, etc.) is required.
(iv) The GA uses probability transition rules, and not the deterministic rules.

Genetic algorithm is essentially derived from a simple model of population genetics. It has five following components:
Chromosomal representation of the variable characterizing an individual.
An initial population of individuals.
An evaluation function that plays the role of the environment, rating the individuals in terms of their fitness that is
their aptitude to survive.
Genetic operators that determine the composition of a new population generated from the previous one by a mechanism
similar to sexual reproduction.
Values for the parameters that the GA uses.

The advantages of GA over traditional techniques are as follows:
It needs only rough information of the objective function and puts no restriction such as differentiability and convexity on
the objective function.
The method works with a set of solutions from one generation to the next, and not a single solution, thus making it less
likely to converge on local minima.

4

The solutions developed are randomly based on the probability rate of the genetic operators such as mutation and
crossover as the initial solutions would not dictate the search direction of GA.

Major disadvantage of GA method is that it requires tremendously high computational time in case of the large variables and
constraints. The treatment of equality constraints is also not well established in G.A. Alander (1996) has presented a bibliography
of genetic algorithm in the power systems. GA has been widely used in the power system. There are several versions of GA
available for different applications.

3.3 Fuzzy Logic

Fuzzy logic (FL) was developed by Zadeh (Zadeh, 1965) in 1964 to address uncertainty and imprecision, which widely exist in
the engineering problems. FL was first introduced in 1979 for solving power system problems. Fuzzy set theory can be considered
as a generalization of the classical set theory. In classical set theory, an element of the universe either belongs to or does not
belong to the set. Thus, the degree of association of an element is crisp. In a fuzzy set theory, the association of an element can be
continuously varying. Mathematically, a fuzzy set is a mapping (known as membership function) from the universe of discourse to
the closed interval [0, 1]. Membership function is the measure of degree of similarity of any element in the universe of discourse to
a fuzzy subset. Triangular, trapezoidal, piecewise-linear and Gaussian functions are most commonly used membership functions.
The membership function is usually designed by taking into consideration the requirement and constraints of the problem. Fuzzy
logic implements human experiences and preferences via membership functions and fuzzy rules.
Due to the use of fuzzy variables, the system can be made understandable to a non-expert operator. In this way, fuzzy logic can
be used as a general methodology to incorporate knowledge, heuristics or theory into controllers and decision makers. The
advantages of fuzzy theory are as follows:
(i) It more accurately represents the operational constraints of power systems and
(ii) Fuzzified constraints are softer than traditional constraints.

Momoh et al. (2000) have presented the overview and literature survey of fuzzy set theory application in power systems. A
recent survey shows that fuzzy set theory has been applied mainly in voltage and reactive power control, load and price
forecasting, fault diagnosis, power system protection/relaying, stability and power system control, etc.

3.4 Evolutionary Computation: Evolutionary Strategies and Evolutionary Programming

Natural evolution is a hypothetical population-based optimization process. Simulating this process on a computer results in
stochastic optimization techniques that can often perform better than classical methods of optimization for real-world problems.
Evolutionary computation (EC) is based on the Darwins principle of survival of the fittest strategy. An evolutionary algorithm
begins by initializing a population of solutions to a problem. New solutions are then created by randomly varying those of the
initial population. All solutions are measured with respect to how well they address the task. Finally, a selection criterion is applied
to weed out those solutions, which are below standard. The process is iterated using the selected set of solutions until a specific
criterion is met. The advantages of EC are adaptability to change and ability to generate good enough solutions but it needs to be
understood in relation to computing requirements and convergence properties. EC can be subdivided into GA, evolution strategies,
evolutionary programming (EP), genetic programming, classified systems, simulated annealing (SA), etc. The first work in the
field of evolutionary computation was reported by Fraser in 1957 (Fraser, 1957) to study the aspects of genetic system using a
computer. After some time, a number of evolutionary inspired optimization techniques were developed.
Evolution strategies (ES) employ real-coded variables and, in its original form, it relied on mutation as the search operator and a
population size of one. Since then, it has evolved to share many features with GA. The major similarity between these two types of
algorithms is that they both maintain populations of potential solutions and use a selection mechanism for choosing the best
individuals from the population. The main differences are:
ESs operates directly on floating point vectors while classical GAs operate on binary strings,
GAs rely mainly on recombination to explore the search space while ES uses mutation as the dominant operator and
ES is an abstraction of evolution at individual behavior level, stressing the behavioral link between an individual and its
offspring, while GA maintains the genetic link.

Evolutionary programming (EP), which is a stochastic optimization strategy similar to GA, places emphasis on the behavioral
linkage between parents and their offspring, rather than seeking to emulate specific genetic operators as observed in nature. EP is
similar to evolutionary strategies, although the two approaches were developed independently. Like ES and GA, EP is a useful
method of optimization when other techniques such as gradient descent or direct analytical discovery are not possible.
Combinatorial and real-valued function optimizations, in which the optimization surface or fitness landscape is rugged and
possessing many locally optimal solutions, are well suited for evolutionary programming.


5

3.5 Simulated Annealing

This method was independently described by Scott Kirkpatrick, C. Daniel Gelatt and Mario P. Vecchi in 1983 (Kirkpatrick et al.,
1983), and by Vlado Cerny in 1985 (Cerny, 1985). Based on the annealing process in the statistical mechanics, the simulated
annealing (SA) was introduced for solving complicated combinatorial optimization problems. In a large combinatorial
optimization problem, an appropriate perturbation mechanism, cost function, solution space, and cooling schedule are required in
order to find an optimal solution with simulated annealing. SA is effective in network reconfiguration problems for large-scale
distribution systems and its search capability becomes more significant as the system size increases. Moreover, the cost function
with a smoothing strategy enables the SA to escape more easily from local minima and to reach rapidly to the vicinity of an
optimal solution.
The advantages of SA are its general applicability to deal with arbitrary systems and cost functions; its ability to refine optimal
solution; and its simplicity of implementation even for complex problems. The major drawback of SA is repeated annealing. This
method cannot tell whether it has found optimal solution or not. Some other methods (e.g. branch and bound) are required to do
this. SA has been used in various power system applications like transmission expansion planning, unit commitment, maintenance
scheduling, etc.

3.6 Expert Systems

AI programs that achieve expert-level competence in solving the problems by bringing knowledge about specific tasks are called
knowledge-based or expert systems (ES) which was first proposed by Feigenbaum et al. in the early 1970s (Feigenbaum et al,
1971). ES is a knowledge-based or rule based system, which uses the knowledge and interface procedure to solve problems that
are difficult enough to be solved by human expertise. Main advantages of ES are:
(a) It is permanent and consistent
(b) It can be easily transferred or reproduced
(c) It can be easily documented.

Main disadvantage of ES is that it suffers from a knowledge bottleneck by having inability to learn or adapt to new situations.
The knowledge engineering techniques started with simple rule based technique and extended to more advanced techniques such
as object-oriented design, qualitative reasoning, verification and validation methods, natural languages, and multi-agent systems.
For the past several years, a great deal of ES applications has been developed to prepare plan, analyze, manage, control and
operate various aspects of power generation, transmission and distributions systems. Expert system has also been applied in recent
years for load, bid and price forecasting.

3.7 Ant Colony and Tabu Search

Dorigo introduced the ant colony search (ACS) system, first time, in 1992 (Dorigo, 1992). ACS techniques take inspiration from
the behavior of real ant colonies and are used to solve functional or combinational problems. ACS algorithms to some extent
mimic the behavior of real ants. The main characteristics of ACS are positive feedback for recovery of good solutions, distributed
computation, which avoids premature convergence, and the use of a constructive heuristic to find acceptable solutions in the early
stages of the search process. The main drawback of the ACS technique is poor computational features. ACS technique has been
mainly used in finding the shortest route for transmission network.
Tabu search (TS) is basically a gradient-descent search with memory. The memory preserves a number of previously visited
states along with a number of states that might be considered unwanted. This information is stored in a Tabu list. The definition of
a state, the area around it and the length of the Tabu list are critical design parameters. In addition to these Tabu parameters, two
extra parameters are often used such as aspiration and diversification. Aspiration is used when all the neighboring states of the
current state are also included in the Tabu list. In that case, the Tabu obstacle is overridden by selecting a new state.
Diversification adds randomness to this otherwise deterministic search. If the Tabu search is not converging, the search is reset
randomly.
TS is an iterative improvement procedure that starts from some initial solution and attempts to determine a better solution in the
manner of a greatest descent neighborhood search algorithm. Basic components of TS are the moves, Tabu list and aspiration
level. TS is a metahuristic search to solve global optimization problem, based on multi-level memory management and response
exploration. TS has been used in various power system application like transmission planning, optimal capacitor placement, unit
commitment, hydrothermal scheduling , fault diagnosis/alarm processing, reactive power planning, etc.

3.8 Particle Swarm Optimization

The particle swarm optimization (PSO) method introduced by Kennedy and Eberhart (Kennedy et al., 1995) is a self-educating
optimisation algorithm that can be applied to any nonlinear optimisation problem. In PSO, the potential solutions, called particles,

6

fly through the problem space by following the best fitness of the particles. It is easily implemented in most programming
languages and has proven to be both very fast and effective when applied to a diverse set of optimization problems. In PSO, the
particles are flown through the problem space by following the current optimum particles. Each particle keeps the track of its
coordinate in the problem space, which is associated with the best solution (fitness) that it has achieved so far. This implies that
each particle has memory, which allows it to remember the best position on the feasible search space that has ever visited. This
value is commonly called as pbest . Another best value that is tracked by the particle swarm optimizer is the best value obtained so
far by any particle in the neighborhood of the particle. This location is commonly called as . gbest
The position and velocity vectors of the i
th
particle of a d-dimensional search space can be represented as ) ,..... , (
2 1 id i i i
x x x X =
and ) ,..... , (
2 1 id i i i
v v v V = respectively. On the basis of the value of the evaluation function, the best previous position of a particle is
recorded and represented as ) ,...... , (
2 1 id i i i
P P P pbest = . If the
th
g particle is the best among all particles in the group so far, it is
represented as
1 2
( , ,.. )
g g g gd
gbest pbest P P P = . The particle tries to modify its position using the current velocity and the distance
from pbest and gbest . The modified velocity and position of each particle for fitness evaluation in the next iteration are
calculated using the following equations

) ( ) (
2 2 1 1
` ` 1 k
id gd
k
id id
k
id
k
id
x gbest rand c x pbest rand c v w v + + =
+
(1)
` 1 1 + +
+ =
k
id
k
id
k
id
v x x (2)
where, wis the inertia weight parameter, which controls the global and local exploration capabilities of the particle.
2 1
, c c are
cognitive and social coefficients and
1
rand and
2
rand are random numbers between 0 and 1. A large inertia weight factor is used
during initial exploration and its value is gradually reduced as the search proceeds. The concept of time-varying inertial weight
(TVIM) is given by

min
max
max
min max
) ( w
iter
iter iter
w w w +
= (3)
where
max
iter is the maximum number of iterations.
The velocity update expression (1) can be explained as follows. Without the second and third terms, the first term (representing
inertia) will keep a particle flying in the same direction until it hits the boundary. Therefore, the first term tries to explore new
areas and corresponds to the diversification in the search procedure. In contrast, without the first term, the velocity of the flying
particle is only determined by using its current position and its best positions in history. Therefore, the second representing
memory) and third terms (representing cooperation) try to converge the particles to their Pbest and/or Gbest and correspond to the
intensification in the search procedure.

3.9 Support Vector Machines

Support vector machine (SVM) is one of the relatively new and promising methods for learning, separating functions in pattern
recognition (classification) tasks as well as performing function estimation in regression problems. It is originated from supervised
learning systems derived from the statistical learning theory introduced by Vapnik for distribution free learning from data
(Vapnik, 1998). In this method, the data are mapped into a high dimensional space via a nonlinear map, and using this high
dimensional space, an optimal separating hyper-plane or linear regression function is constructed. This process involves a
quadratic programming problem and produces a global optimal solution. The great advantage of SVM approach is that it greatly
reduces the number of operations in the learning mode and minimizes the generalization error on the test set under the structural
risk minimization (SRM) principle. Main objective of the SRM principle is to choose the model complexity optimally for a given
training sample. The input space in a SVM is nonlinearly mapped onto a high dimensional feature space. The idea is to map the
feature space into a much bigger space so that the boundary is linear in the new space. SVMs are able to find non-linear boundaries
if classes are linearly non-separable. Another important feature of the SVM is the use of kernels. Kernel representations offer an
alternative solution by nonlinearly projecting the input space onto a high dimensional feature space. The advantage of using SVMs
for classification is the generalization performance. SVM performs better than neural networks in term of generalization. There is
problem of over fitting or under fitting if so many training data or too few training data are used. The computational complexity is
other factor for the choice of SVMs as classifier. The other advantage of SVM based system is that it is straight forward to extend
the system when new types of cases are added to the classifier.

4. Conclusion

The current interest in using the computation intelligence (CI) for power system applications is increasing amongst the
researchers and academicians. It is noticed that huge amount of research papers and articles are available in all the area of

7

engineering and science. There is no specific guideline for the application of CI techniques in the power systems. In this paper,
various computational techniques widely used in power system applications are briefly described. The potential areas of CI
application are also highlighted. New intelligent system technologies using digital signal processing techniques, expert systems,
artificial intelligent and machine learning provide several unique advantages in solving the power system problems.

Acknowledgement

Authors acknowledge the suggestions and corrections made by Dr Guang Ya Yang, Centre of Electric Technology, Electrical
Engineering Department, Denmark Technical University, Denmark. Dipti Saxena is grateful to the CM and Director, Invertis
Institute of Engineering and Technology, Bareilly for permitting her to carry our PhD work from UP Technical University,
Lucknow, India.

References

Alander J. T., 1996, An indexed bibliography of genetic algorithm in power engineering, Power Report Series 94-1.
Bishop C.M., 1995, Neural networks for pattern recognition, Oxford University Press, Oxford.
Cerny V., 1985, A thermodynamical approach to the travelling salesman problem: an efficient simulation algorithm. Journal of
Optimization Theory and Applications, vol.45, pp.41-51.
Dorigo M., 1992, Optimization, learning and natural algorithms, PhD thesis, Politecnico di Milano, Italy.
El-Hawary, Mohamed E., 1998, Electric power applications of fuzzy systems, John Wiley USA.
Feigenbaum, E.A., Buchanan, B.G., Lederberg J., 1971, On generality and problem solving: A case study using the DENDRAL
program, in Machine Intelligence 6, Edinburgh University Press.
Fogel L.J., Owens A.J., Walsh M.J., 1966, "Artificial intelligence through simulated evolution", Wiley, New York
Fraser A. S. 1957, Simulation of genetic systems by automatic digital computers, I. Introduction.Australian. J. Biol. Sci., vol.10,
pp.484491.
Kennedy J., and Eberhart R., 1995, Particle swarm optimization, Proc. of the IEEE Int. Conf. on Neural Networks, Piscataway,
NJ, pp. 19421948.
Kirkpatrick S., Gelatt C. D., Vecchi M. P., 1983, "Optimization by simulated annealing". Science. New Series 220, pp.671680.
Lai, Loi Lei, 1998, Intelligent system applications in power engineering: evolutionary programming and neural networks, John
Willey & Sons, UK.
Momoh James A., EL-Hawary Mohamed E., 2000, Electric systems, dynamics, and stability with artificial intelligence, Marcel
Dekker, Inc. USA.
Sobajic Dejan J., 1993, Neural network computing for the electric power industry, Routledge Publisher, USA.
Song Yong-Hua Song, Johns Allan, Aggrawal Raj,1996. Computation intelligence applications to power systems, Kluwer
Academic Publishers, USA.
Vapnik N., 1998, Statistical learning theory, John Wiley and Sons, New York.
Warwick K., Ekwue Arthur, Aggarwal Raj, 1997, Artificial intelligence techniques in power systems, IEE UK.
Wehenkel Louis A., 1998, Automatic learning techniques in power systems, Kluwer academic publisher, USA.
Zadeh L.A.,1965, Information and Control, vol. 8, no. 3, pp. 338-353.

Biographical notes:

D. Saxena obtained B. Tech. (Hons) in Electrical Engineering from KNIT Sultanpur (UP), India and M. Tech. in Process Control from Netaji Subhas Institute of
Technology, New Delhi, India in 1999 and 2003, respectively. Presently, she is working as reader in Electrical and Electronics Engineering Department at Invertis
Institute of Engineering And Technology, Bareilly. She is registered as PhD candidate in UP Technical University Lucknow, India. Her research interests are
power quality, power electronics, control systems and DSP application in power.

S. N. Singh received M. Tech. and Ph.D. from Indian Institute of Technology Kanpur, India in 1989 and 1995, respectively. He is a Professor in the Department of
Electrical Engineering, Indian Institute of Technology Kanpur, India. Presently he is on leave from IIT Kanpur and working with the Centre for Electric
Technology (CET), Technical University of Denmark, Denmark. His research interests include power system restructuring, power system optimization & control,
voltage security and stability analysis, power system planning, and ANN application to power system problems. He is a Fellow of IE (India), Fellow of IETE
(India) and senior member of IEEE.

K.S. Verma received his B. Tech. (Hons) in Electrical Engineering and M. Tech. in Electrical Engineering (Power Systems) both from KNIT Sultanpur (India) in
1987 and 1998, respectively. He obtained his Ph.D. degree in Electrical Engineering (Power Systems) from Indian Institute of Technology, Roorkee, Roorkee
(India) in 2003. Presently, Prof Verma is Director, KNIT Sultanpur (India). His research interests include FACTS Technology, distributed generation, power
system optimization & control, power quality and AI application in power system.

Received February 2010
Accepted March 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 8-18

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Application of artificial neural networks to improve power transfer
capability through OLTC

A. Abu-Siada
1
*, S. Islam
1
and E.A. Mohamed
2

1*
Department of Electrical and Computer Engineering, Curtin University of Technology, Perth, AUSTRALIA
2
Department of Electrical Power Engineering, Ain Shams University, Cairo, EGYPT
*
Corresponding Author: e-mail: a.abusiada@curtin.edu.au, Tel +61-8-92667287, Fax.+61-8-92662584

Abstract

On load tap changing (OLTC) transformer has become a vital link in modern power systems. It acts to maintain the load bus
voltage within its permissible limits despite any load changes. This paper discusses the effect of different static loads namely;
constant power (CP), constant current (CI) and constant impedance (CZ) on the maximum power transfer limit from the
generation to the load centre through the OLTC branch and in turn on the static stability margin of power systems. Then the
paper introduces a novel approach for the on-line determination of the OLTC settings using artificial neural network (ANN)
technique in order to improve the power transfer capability of transmission systems. The proposed approach is tested on a six-
bus IEEE system. Numerical results show that the setting of OLTC transformer in terms of the load model has a major effect on
the maximum power transfer in power systems and the proposed ANN technique is very accurate and reliable. The adaptive
settings of OLTC improve the power transfer capability according to the system operating condition.

Keywords: OLTC, Static Load, Voltage Stability, ANN.

1. Introduction

Modern power systems are being stressed with the continuous growth of load requirements. Power system operator should
ensure the quality and reliability of supply to the loads by maintaining the load bus voltages within their permissible limits. The
power transfer from generators to load centers affects load bus voltages. In addition to restoring the load, OLTC also extend the
power transfer ability to the load center (Zhu et al., 2000). It is well known that the operation of OLTC has a significant influence
on voltage stability. Most literature has concentrated on the contribution of OLTC to voltage stability (Tylor 1994, Afzalian et al.
2008). The power transfer capability is measured by the margin from a base case operating point to the operation-limiting
boundary, measured along the load growth trajectory. The static load models have an impact on the power transfer capability from
generating station to the load center and thus on the static stability margin of power systems (Abu-Siada et al. 2008, Feng et al.
2004). The paper considers first the effect of different static load models namely; constant power (CP), constant current (CI) and
constant impedance (CZ) load models on the power transfer limits to the load centre. Then the paper presents a novel approach for
evaluating the on-line optimal settings of OLTC transformer that corresponds to the maximum power transfer to the load and
hence to improve the static stability limits using artificial neural network (ANN).

2. Effect of OLTC Setting on Power Transfer

Assume the simple power system shown in Figure 1. The load power is supplied from an infinite bus via OLTC. The load flow
equation at load bus is given by:

c
L
L T
L
jX
V
V
jQ P
X jn
V nE
* 2
(1)
Abu-Siada et al. / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 8-18

9
where 0 j E E + = is the emf of the equivalent voltage source,
i r L
jV V V + =
is the load bus voltage,
tr eq T
X X X + =
is the
equivalent reactance of power system & transformer, P and Q are the load active and reactive power and X
c
is the reactance of the
shunt capacitor compensated load.

Figure. 1 Simplified power system

2.1 Constant Power (CP) Load: For constant Power load;
K
c
c
P
Q
Let
c Q and c P
q
p
q p
= =
= =

Equation (1) can be re-written as below:

jKP P
X X n
X n X
V V j
nX
EV
nX
EV
j
c T
T c
i r
T
i
T
r
=
+ +
2
2
2 2
) (
(2)
Equating real parts in both sides,

P
E
nX
V
T
i
=
(3)
Equating imaginary parts in both sides and substituting for V
i
from (3),

0 ) (
2
2
2
2
2
2
2 2
=
+
r
T c
c
r
T C
C T T
V
X n X
E nX
V P
X n X
X KX n
P
E
X n
(4)
Solving for P,

) ( ) ( )
) ( 2
(
) ( 2
2
2 2 2
2
2
2
2
r
T c
c
r
T T c T
c
T c T
c
V
X n X
E nX
V
nX
E
X n X X
E KX
X n X X
E KX
P
=
(5)
This solution is subject to:

T
c
r
T c
c
r
T T c T
C
X
X
n
and
V
X n X
E nX
V
nX
E
X n X X
E KX
0 ) ( ) ( )
) ( 2
(
2
2 2 2
2
2
f
(6)
From equation (6), one can conclude that there is a limit to the power transfer to the load depends on the degree of
compensation, OLTC setting and load bus voltage. Figure 2 shows the effect of OLTC tap ratio on the maximum power transfer
for compensated and uncompensated load (X
c
=10 pu. and X
c
= respectively). For both cases the maximum power transfer limit is
increased by increasing the tap ratio. Once the optimal power is reached (0.5 pu. in case of uncompensated load and 0.56 pu. in
case of compensated load) the power slightly decreases with the increase of tap ratio. It should be noted that for the compensated
load, the power is dramatically increased when the resonant case is reached (
472 . 4 = =
T
c
X
X
n
in this case).

Figure 2. Effect of OLTC on power transfer to the CP load (K=0.75, X
T
=0.5 pu., V
r
=0.8 pu.)
Rearranging equation (4) and solving for V
r,

10

) ( )
) ( 2
(
) ( 2
2 2
2 2
2 2
P
E
X
X n X
KX
P X n
X n X
E nX
X n X
E nX
V
T
T c
c
T
T c
c
T c
c
r
+
=
(7)
Figure 3 shows the PV curve for the uncompensated load for different tap ratios. The figure shows that the maximum power
limit is constant and equal 0.5 pu. in all cases.

Figure 3. P-V curve at different tap ratios (X
c
=)
The PV curves in Figures 4 and 5 show that the maximum power transfer limit to a compensated load can be increased by either
increasing the tap ratio of OLTC or by increasing the degree of compensation. The power increase is attributed to the fact that the
OLTC tap settings allow the match between the network impedance and the reflected compensated load impedance.

Figure 4. P-V curve at different tap ratios (X
c
=1 pu)

Figure 5. P-V curve at different load compensation (n=1)

2.2 Constant Current (CI) Load: For constant current load;
K
P
Q
V c Q V c P
L q L p
= = = ,

Then equation (1) can be re-written as:

) 1 ( ) (
2
2
2 2
jK V c
X X n
X n X
V V j
nX
EV
nX
EV
j
L p
c T
T c
i r
T
i
T
r
=
+ +
(8)
Equating real parts in both sides,

r
p T
p T
i
i r p T i
L p
T
i
V
c nX E
c nX
V
V V c nX V E
V c
nX
EV
2 2
2 2 2 2 2
) (
) ( ) (
=
+ =
=
(9)

11
Equating imaginary parts and substituting for V
i

2 ) (
) ( ) (
) (
2
2
2
2
2 2
p T
T c
p c T
c
T c
p T
r
c nX E
X n X E
Kc X X n
nX
X n X E
c nX E
V
=
(10)

2 2
2 2
2
2 2 2
2
) (
)
) (
) (
1 (
p T
r
L
p T
p T
r i r L
c nX E
EV
V
c nX E
c nX
V V V V
+ = + =
(11)

T c
p T c
p T
T c
c p
L p
X n X
K c X X n
c nX E
X n X
X nc
V c P
2
2 2
2 2
2
) (
= =
(12)
This solution is subject to:

T
c
p T
p T
X
X
n
and
c X
E
n
c nX E
p
f 0 ) (
2 2
(13)

Figure 6. Effect of OLTC on power transfer to the CI load
Figure 6 shows the effect of OLTC tap ratio on the power transfer limit to constant current load. The maximum power transfer
limit is increasing with the increase of the tap ratio. When the maximum limit is reached (0.5 pu. in case of uncompensated load
and o.62 pu. in case of compensated load), the power is decreasing with the increase of tap ratio. The shunt capacitor increases the
power transfer limit and shifts the optimal setting of OLTC to a higher value.

2.3 Constant Impedance (CZ) Load: For constant impedance load;

K
P
Q
V c Q V c P
L q L p
= = =
2 2
(14)
Then equation (1) can be re-written as:

) 1 ( ) (
2
2
2 2
jK P
X X n
X n X
V V j
nX
EV
nX
EV
j
c T
T c
i r
T
i
T
r
=
+ +
(15)
Equating the real parts in both sides,

E
P nX
V
T
i
=
(16)
Equating imaginary parts,

P
X X n
X n X
Kc
Ec
nX
V
c T
T c
p
p
T
r
) (
2
2
+ =
(17)
p
T
c T
T c
p
p
T
i r L
c
P
P
E
nX
P
X X n
X n X
Kc
Ec
nX
V V V = +
+ = + =
2 2 2 2
2
2
2 2 2
2
) ( ) ( ) (
2 2
2
) ( ) (
/ 1
E
nX
E
KnX
E X nc
X n X
c
P
T T
c p
T c
p
+ +
=
(18)
Figure 7 shows the effect of OLTC tap ratio on the power transfer limit to constant impedance load. Again, the power is
increasing with the increase of OLTC tap ratio till the optimal power is reached then the power is decreasing with the increase of
tap ratio. The optimal power in case of uncompensated load (X
c
= ) is 0.5 pu. corresponding to 1.72 tap ratio. However, the
maximum power for a compensated load (X
c
= 10 pu.) is 0.62 pu. corresponding to 2.1 tap ratio. It should be noted that in all load
models (CP, CI and CZ) the power transfer limit is increased by increasing the degree of compensation i.e. decreasing the value of
X
c
. When the load power cannot be met with the increasing compensation degree, the power can be increased by increasing the
OLTC tap ratio.

12

Figure 7. Effect of OLTC on power transfer to the CZ load

3. Digital Simulation

Figure 8 shows a one-line diagram of the IEEE 6-bus test system. The system consists of 6-nodes, 7 lines, two OLTC
transformers, two generators and four loads (Tamura et al. 1993). The two OLTCs are varied one at a time, during the calculation
of the maximum power transfer to the load centers.

Figure 8. One line diagram of the system under study
The results obtained using the conventional methodology cover a wide range of operating conditions. A sample of the
simulation results is shown in Figures 9-14. These results are presented to explore the effect of both OLTC settings and the load
model on the maximum power transfer. Figures 9-11 show the relation between the power transfer to the three load buses (1, 2&
4) and the transformer tap ratio n under loading condition P
L
= Q
L
= 0.35 pu.(pf= 0.707 lag), for the three different load models
(CP, CI & CZ). From these figures, it can be seen that the optimum value of n and the corresponding maximum power transfer to
the loads are different for the three load models. Figures 12-14 present similar results at another loading condition P
L
= 0.8 pu. and
Q
L
= 0.13 pu. (pf = 0.987 lag). It can be seen from these figures that the maximum power transfer limit is higher at this operating
condition than the previous case and the OLTC setting n
1
has a strong effect on the load at bus #1 while the OLTC setting n
2
has a
stronger impact on the load at bus #2. On the other hand, the OLTC setting n
2
has a unique correlation and effect on the load at
bus #4. These figures prove that the power transfer limit is affected by the OLTC setting for all load models.
Also, Results show that before the reverse points are reached, the upward of OLTC operation increases the maximum power
transfer to the load centre. Also these results give an indication that the maximum point of power transfer and thus the levels of
these powers are different for different load models. It must be noted here that, in the constant impedance (CZ) load model, the
nonlinear load behavior is very sensitive to any system variation near to the static stability limit. That can be depicted from the
sharp peaks indicated on Figures 10 and 14. In case of constant power (CP) load model, the load power is constant however, the
power transfer to the load bus through the OLTC branch is maximized and the other shares through other branches are varied.

13

Figure 9. Effect of OLTC (n1) on Power transfer to load 1





14



3. Determination of OLTC Setting using ANN

Owing to the fast development in computing systems, application of intelligent systems such as artificial neural network (ANN)
and fuzzy logic have paid a great attention in power system applications (Abu-Siada et al. 2009, Nor et al. 2003). Fuzzy logic
needs a prior experience about the system to enable designing fuzzy logic rules based on the input output behaviors. ANN is
selected in this paper because of its capability of learning a tremendous variety pattern mapping relationships without having prior
knowledge of a mathematical function. ANN is able to handle complex non-linear problems and it overcomes the complex tedious
calculation problems. The enhancement of power transfer and voltage stability can be achieved in real-time using the proposed
method without any additional cost involved for installing capacitor banks and the related switches (El-Keib et al. 1999). The
developed ANN model can be implemented using the currently available neural chips. The main drawback of ANN is the
extensive training data required to ensure the reliability of the network.

Figure 15. Typical ANN architecture model

15
3.1 Topology of the ANN: The proposed ANN scheme uses a multi-layer feed-forward ANN which consists of an input layer, a
hidden layer and an output layer as shown in Figure 15 (El-Sharkawi 1991, Kashem et al. 2001).
3.1.1 Input layer: Appropriate selection of input variables is the key to the success of the ANN application. Usually heuristic
knowledge is required in choosing the appropriate variables. The power transfer from generation to the load centre is mainly
concerned with the transformer OLTC settings. The optimal setting of the OLTC is affected by the load level. Therefore, these
loading conditions are chosen as inputs to the ANN. These inputs are the load active and reactive power in addition to the
generation active and reactive power at each bus in the system.
3.1.2 Hidden layer: The computational power of the ANN increases with the addition of hidden layers. There are no general
guidelines to determine the number of hidden layers and number of neurons per layer. Many applications have proved that ANN
with single hidden layer has sufficient capability of capturing complicated relations between input and output variables. In this
paper, one hidden layer with five neurons was found satisfactory for estimating the optimal transformer OLTC settings.
3.1.3 Output layer: Determination of the output layer is quite forward. Two output neurons were used for setting the OLTC of the
two transformers. The sigmoidal transfer functions have been used for all the neurons as they are suitable for any nonlinear
mapping of input-output combinations because they have noise immunity for low inputs, normal outputs for middle range inputs
and saturation for large inputs. Appropriate scaling of the input and output variables were carried out in the (0, 1) range.

3.2 Training Scenario: The scenario of the training process was started by initiating the ANN topology; i.e. the number of input
nodes, the number of hidden layers, the number of neurons in different hidden layers, the number of neurons in the output layer,
the type of hidden and output activation functions, and the number of presentation cycles (epochs). It must be noted, that, by
adjusting the number of hidden layers and the number of simulated neurons within each layer, the performance of an ANN can be
either enhanced or degraded. There are no general guidelines for a priori knowledge of which ANN architecture would perform the
best for a given application. The researchers have experimented with different architectures to find out the most suitable
configuration. The results obtained from the digital simulation of the system in Figure 8 were used in training and testing the
proposed ANN. Back-propagation algorithm was used to train the proposed ANN using 100 training patterns with a learning rate
of 0.96. The Neural Network toolbox in MATLAB is used to implement the proposed ANN. The network weights and biases
initialized using random values. During training, the weights and biases are adjusted to minimize the error between the network
outputs and the targets. It has been found from the simulation results that the OLTC settings and the corresponding maximum
power transfer to load centers are varying with the load type. Therefore, different ANNNs have been trained for different load
models.

3.3 Testing Scenario: For validating the ANN performance the scenario of the testing process is started once the training process is
completed and a set of new input variables which were not used for training the ANN model are applied to the designed ANN. The
proposed method has been applied to the system shown in Figure 8 and the ANN results are compared with the simulation results.
The obtained results for this scenario are recorded in Tables 1, 2 and 3.

4. Results and Discussions

The three load models under consideration are tested on different operating conditions. The transformers turns ratios are varied
one at a time. The results in Table 1 have been obtained for the constant power (CP) load model. The table shows that the
maximum testing absolute errors are: 5.06% (n
1
) & 5.623% (n
2
). Similarly, Tables 2 & 3 show these errors for the constant current
(CI) and the constant impedance (CZ) load models, respectively. The maximum errors obtained in this case are: 6.0% (n
1
), 3.94%
(n
2
) and 3.66% (n
1
), 3.49% (n
2
), respectively. Also tables present the average and the standard deviations for the three different
load models. These deviations are: 1.092/1.731 (n
1
), 1.762/2.215 (n
2
) for CP load; 1.499/2.375 (n
1
), 1.275/1.636 (n
2
) for CI model;
and 1.66/1.474 (n
1
), 0.825/1.19 (n
2
) for CZ model. From these results, it can be seen that the CP load and the CI load models have
nearly the same error level. While, the CZ load model has the less error level. The ANN is efficient for the prediction of the
optimal value of transformer turns ratio settings corresponding to the maximum power transfer to the load centre.

5. Conclusions

An efficient method for the on line prediction of OLTC transformer settings giving maximum power transfer to the load centre is
analysed. The proposed technique has been tested on the IEEE Six-bus power system. Numerical results show that the operation of
OLTC transformer has a major effect on the maximum power transfer and thus on the static stability margin. Static load models
such as constant power, constant current and constant impedance have a great effect not only on the maximum power transfer but
also on the optimal settings of transformer OLTCs. The Artificial neural network (ANN) proved to be an efficient tool for the
prediction of the optimum value of OLTC transformer settings corresponding to the maximum power transfer. The developed
ANN model can be implemented using the currently available neural chips.


16
Table 1. Results of ANN in case of constant power (CP) load model
Training Scenario Testing Scenario
Predicted
Values
Desired
Values
%Error Predicted
Values
Desired
Values
% Error
n
1
n
2
n
1
n
2
n
1
n
2
n
1
n
2
n
1
n
2
n
1
n
2
.9983
.9985
.997
.9985
.9981
.9975
.9985
.9983
.9975
.9983
.9972
.9972
.9982
.998
.9962
.9943
.9952
.8988
.9933
.9852
.9202
.994
.9909
.9339
.9898
.9324
.8942
.9914
.9839
.8411
1
1
1
1
1
1
1
1
1
.98
1
1
.98
1
1
1
1.02
.9
1
1
.93
1
1
.93
1
.93
.9
1
.95
.85
-.17
-.15
-.30
-.15
-.19
-.25
-.15
-.17
-.25
1.867
-.28
-.28
1.857
-.2
-.38
-.57
-2.48
-.133
-.670
-1.48
-1.05
-.6
-.91
.419
-1.02
.258
-.644
-.86
3.58
1.07
.9983
.9982
.9965
.9984
.9976
.9974
.9984
.9981
.9971
.9981
.997
.9968
.9981
.9978
.9963
.993
.992
.852
.991
.966
.917
.993
.989
.897
.986
.919
.843
.982
.983
.873
.98
1
1
.98
1
1
.99
1
1
.95
1
1
.96
1
1
1
.98
.85
1
.95
.9
1
.98
.88
1
.9
.85
1
.93
.83
1.86
-.18
-.35
1.88
-.24
-.26
.899
-.19
-.29
5.06
-.3
-.32
3.99
-.22
-.37
-.77
1.143
.897
-.9
1.684
1.888
-.670
-.908
2.125
-1.54
1.544
-.318
-1.08
5.623
5.337
Average deviation
Standard deviation
.443
.7368
1.047
1.361
1.092
1.71
1.762
2.215

Table 2. Results of ANN in case of constant Impedance (CI) load model
Predicted
Values
Desired
Values
% Error Predicted
Values
Desired
Values
% Error
n
1
n
2
n
1
n
2
n
1
n
2
n
1
n
2
n
1
n
2
n
1
n
2
.9703
.9843
.9927
.9489
.9910
.9931
.9483
.9878
.9931
,9581
.9952
,9947
.9579
.9849
.9929
.9915
.9406
.8067
.9920
.9440
.8606
.9922
.9458
.8328
.9876
.7850
.8032
.9876
.9171
.7577
.98
1
1
.95
1
1
.95
1
1
.93
.78
1
.93
1
1
1
.95
.80
1
.93
.85
1
.95
.83
1
1
.78
1
.90
.75
-.989
-1.57
-.73
-.116
-.90
-.69
-.179
-1.22
-.69
3.02
-.48
-.53
3.0
-1.51
-.71
-.85
-.989
.838
-.80
1.51
1.24
-.78
-.442
.337
-1.24
.641
.40
-1.24
1.9
1.027
.9456
.9862
.9935
.9453
.9924
.9939
.9466
.9893
.9943
.9531
.9958
.9953
,9540
.9869
.9947
.9929
.9275
.7767
.9929
.9252
.8383
.9926
.9342
.7923
.9897
,7319
,7753
.9899
.8990
.6724
.95
1
1
.93
1
1
.94
1
1
.9
1
1
.9
1
1
1
.93
,78
1
.9
.83
1
.93
.8
1
.75
.78
1
.88
.70
-.463
-1.38
-.65
1.645
-.76
-.61
-.702
-1.07
-.57
5.9
-.42
-.47
6.0
-1.31
-.53
-.71
-.269
-.33
-.71
2.8
1
-.74
.45
-.963
-1.03
-2.41
-.602
-1.01
2.159
-3.94
Average deviation
Standard deviation
1.225
1.563
.9489
1.067
1.499
2.375
1.275
1.636


17
Table 3. Results of ANN in case of constant current (CZ) load model
Predicted
Values
Desired
Values
% Error Predicted
Values
Desired
Values
% Error
n
1
n
2
n
1
n
2
n
1
n
2
n
1
n
2
n
1
n
2
n
1
n
2
.9444
.9955
.9965
.9872
.9980
.9982
.9750
.9975
.9978
.7996
.9979
.9972
.8527
.9963
.9951
.9956
.9494
.8510
.9856
.7494
.785
.9912
.8291
.7706
.9984
.7996
.9343
.9977
.8299
.8023
.95
1
1
.98
1
1
.97
1
1
.8
1
1
.85
1
1
1
.95
.85
1
.75
.78
1
..83
.78
1
.8
.93
1
.83
.8
-.59
-.45
-.35
-.735
-.2
-.18
.515
-.25
-.22
-.05
-.21
-.28
.318
-.37
-.49
-.44
-.06
.118
-1.44
-.08
.641
-.88
-.108
-.94
-.16
-.05
.46
-.23
.012
.288
.9437
.9959
.9965
.964
.9978
.9981
.9695
.9975
.9977
.9844
.9978
.9974
.8226
.9965
.9954
.9948
.9251
.8280
.9919
.7437
.7498
.9910
.8018
.7407
.9651
.7883
.9052
.9978
.7965
.7821
.93
1
1
.93
1
1
.95
1
1
.79
1
1
.8
1
1
1
.93
.83
1
.73
.75
1
.8
.75
1
.78
.90
1
.80
.78
1.47
-.41
-.35
3.66
-.22
-.19
2.053
-.25
-1.56
1.823
-.22
-.26
2.83
-.35
-.35
-.52
-.527
-.24
-.810
1.877
-.27
-.90
.18
-1.24
-3-49
1.064
.578
-.22
-.438
.269
Average deviation
Standard deviation
.349
.3836
.394
.547
1.66
1.474
.825
1.19

References

Abu-Siada A., S. Islam and E. A. Mohamed, 2008. Adaptive Setting of OLTC to Improve Power Transfer Capability of Power
Systems, CMD08, Beijing, China, March.
Abu-Siada A., S. Islam and S. Lai, 2008. A Novel Application of Gene Expression Programming in Transformer Diagnostics,
AUPEC08, Sydney, December.
Abu-Siada A., S. Islam and S. Lai, 2009. Remnant Life Estimation of Power Transformer using Oil UV-Vis Spectral Response,
IEEE PES Power Systems Conference & Exhibition (PSCE), USA, March.
Afzalian A., A. Saadatpoor and W. Wonham, 2008. Systematic supervisory solutions for under-load tap-changing transformers,
Control Engineering Practice, Vol. 16, pp1035-1054.
Dilon T. S., 1991. Artificial Neural Network Applications to Symbolic Methods, Journal of Electrical Power and Energy System,
Vol. 13, No. 2, pp. 66-72.
El-Keib A. and X. Ma, 1999. Application of Artificial Neural Networks in Voltage Stability Assessment, IEEE Transaction on
Power Systems, Vol. 10, Nov.
EL-Sharkawi M. A., R, J. Marks and Siri Weerasooriya, 1991. Neural Networks and Their Application to Power Engineering,
control and Dynamics serial Academic Press.
Feng Dong et al, 2004. Impact of Load Tab Changing Transformers on Power Transfer Capability, Electric Power Components
and Systems, Vol. 32, pp. 1331-1346.
Gwang Won Kim and Kwang Y. Lee, 2005. Coordination Control of ULTC Transformer and STATCOM Based on an Artificial
Neural Network, IEEE Transactions on Power Systems, Vol. 20, No. 2, pp 580-586.
Kashem M. A., V. Ganapathy and G. B. Jasmon, 2001. On-line Network Reconfiguration for Enhancement of Voltage Stability in
Distribution Systems using Artificial Neural Networks, Electrical Power Components and Systems, Vol. 29, pp. 361-373.
Lai S., A. Abu-Siada, S. Islam. 2008. Correlation between UV-Vis Spectral Response and Furan Measurement of Transformer
Oil, CMD08, Beijing, China, March.
Les M.Hajagos, and Behnam Danai, 1998. Laboratory Measurements and Models of Modern Loads and Their Effects on voltage
Stability Studies, IEEE Transaction on Power Systems, vol. 6, No.2 May, pp.584-591.
Nor Haidar et al., 2003. Application of ANN to Determine the OLTC in Minimizing the Real Power Losses in Power System,
National Power and Energy Conference Proceddings, Malaysia.
Qin Zhou et.al., 1994. Application Of Artificial Neural Networks In Power System Security And Vulnerability Assessment, IEEE
Transaction on Power Systems, Vol.9, No.1, February, pp. 525-532.
Qiu J. and S.M.Shahidelhpour; 1987. A New Approach For Minimizing Power Losses and Improving Voltage Profile. IEEE
Transaction on Power Systems, Vol. PWRS-2, No.2, May, pp. 287-295.
Tamura, Y.Mori, H.Iwamoto,S., 1993. Relationship between Voltage Instability and Multiple Load Flow Solutions in Electric
Power Systems, IEEE Transaction on Power Apparatus and Systems, PAS-102, No.5, pp.1115-1125.
Taylor C. W., 1994, Power System Voltage Stability, New York; McGraw-Hill, Inc.
Zhu T. X., S.K. Tso. and K.L.Lo., 2000. An Investigation into the OLTC Effects on Voltage Collapse, IEEE Transaction on
power Systems, vol.15. No.2, May, pp.515-521.

18
Biographical notes:

Ahmed Abu-Siada received the B.Sc. and M.Sc. degree from Ain Shams University, Egypt and the PhD degree from Curtin University of Technology, Perth,
Australia, All in Electrical Engineering. Currently, he is a lecturer in the Department of Electrical and Computer Engineering at Curtin University of Technology.
His research interests include power System Stability and Control, Power Electronics, Power Quality, Condition Monitoring, Energy Technology and System
Simulation. He is a member of IEEE.

Syed Islam (S81, M, 83, SM93) received the B.Sc., MSc, and PhD degree all in electrical power engineering in 1979, 1983, and 1988 respectively. He is
currently the Chair Professor in Electrical Power Engineering and Head of Department of Electrical and Computer Engineering at Curtin University of Technology,
Perth, Australia. He received the IEEE T Burke Hayes Faculty Recognition award in 2000. He has published over 140 technical papers in his area of expertise. His
research interests are in Condition Monitoring of Transformers, Wind Energy Conversion, and Power Systems. He has been a keynote speaker and invited speaker
at many international workshops and conferences. He is the current Vice-Chair of the Australasian Committee for Power Engineering (ACPE) and a member of the
steering committee of the Australian Power Institute. He is a Fellow of the Engineers Australia, a senior member of the IEEE IAS, PES and DEIS, a Fellow of the
IET and a chartered engineer in the United Kingdom. He is regular reviewer for the IEEE Trans. on Energy Conversion, Power Systems and Power Delivery. Prof.
Islam is an editor of the IEEE Transaction on Sustainable Energy.

E.A. Mohamed received the B.Sc. and M.Sc. degree from Ain Shams University, Egypt and the PhD degree from the University of Manitoba, Canada, All in
Electrical Engineering. Currently, he is a Professor in the Department of Electrical Engineering, Ain Shams University, Egypt His research interests include power
System Stability and Control, Power Quality, Application of AI to power systems and System Simulation.

Received August 2009
Accepted March 2010
Final acceptance in revised form March 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 19-28

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Application of radial basis neural network for state estimation of power
system networks

J.P. Pandey
1
*, D. Singh
2

1*
Department of Electrical Engineering, Kamla Nehru Institute of Technology Sultanpur INDIA
2
Department of Electrical Engineering, I.T,B.H.U Varanasi. INDIA
*
Corresponding Author: e-mail: tojppandey@rediffmail.com,

Abstract

An original application of radial basis function (RBF) neural network for power system state estimation is proposed in this
paper. The property of massive parallelism of neural networks is employed for this. The application of RBF neural network for
state estimation is investigated by testing its applicability on a IEEE 14 bus system. The proposed estimator is compared with
conventional Weighted Least Squares (WLS) State Estimator on basis of time, accuracy and robustness. It is observed that the
time taken by the proposed estimator is quite low. The proposed estimator is more accurate and robust in case of gross errors and
topological errors present in the measurement data.

Keywords: Radial Basis Function Neural Networks, State Estimation.

1. Introduction

Electric Power System deregulation has transformed state estimation from an important application to a critical one. The system
operator has to make equitable, security related, congestion management decisions to curtail or deny power transfer rights in real
time. It has to be founded and justified on a precise model of the power system derived from the state estimation process.
Moreover fast and accurate state estimation is foundation of locational marginal pricing methodologies for transmission
management costing.
The state estimation provides the real time representation of the conditions in a power network. A state estimator is a data
processing algorithm, which transforms meter readings and the switch status information into an estimate of the systems state
(voltage magnitudes and phase angles at all the nodes). Real and reactive bus power injections, and real and reactive line flows and
bus voltage magnitudes are the measurements, which are transmitted to computer control system via telemetry system. These
measurements contain random noise due to instrument and phenomenon errors. The state estimation program obtains a best fit for
the power system state variables by minimizing these errors. Ideally state estimation should run at the scanning rate of the
telemetry system (say at every two seconds). Due to computational limitations, most practical state estimators run every few
minutes or when major changes occur.
Most of the state estimation problems are formulated as over determined system of non-linear equations and solved as a
weighted least squares problem. The Weighted Least Squares Estimation (WLSE) is by far the most popular approach in industry.
The least squares technique is slow and computational requirements are prohibitively large since there is large number of
redundant measurement data normally available in form of nodal injection and line flows. To overcome this difficulty, a number of
alternatives algorithms including modification and refinements of the basic WLSE have been presented (Horisberger et al, 1976;
Garcia et al, 1979). The state of the art in state estimation algorithms is presented in (Wu 1990; Monticelli, 2000). Most of the
practical implementation of state estimation in electric power systems is based on the Gauss-Newton methods. The state estimates,
i.e., the voltage magnitudes and the bus voltage angles are solved through an iterative procedure in least squares sense.
Pandey et al. / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 19-28

20

The slowness of the state estimators due to computational requirements has been a major drawback of the present methods of
state estimation employed. The conventional state estimation is based on algorithmic method of solving a large number of non-
linear equations based on network line flows and /or bus injections and network constraints similar to power flow problem.
The present work proposes entirely different paradigm of state estimation problem. In this paper state estimation is addressed as
a pattern recognition problem and solved using learning approach. The learning based methods have found wide applications in
some EMS applications such as load forecasting, Topology Processing, Optimal unit commitment, Load Flows etc (Singh et al,
2001). In this paper an original application of radial basis function neural network for state estimation is proposed. The property of
massive parallelism of neural network is employed for faster state estimation. The proposed estimator is studied for various cases
to show its utility for state estimation in terms of accuracy and time requirements.

2. WLS State Estimator

Most state estimation programs in practical use are formulated as over determined systems of nonlinear equations and solved as
WLSE problem.
Consider the nonlinear measurement model
j j
e x h
j
z + = ) ( (1)
where
j
z is the
th
j measurement, x is the true state vector, ) (x h
j
is a nonlinear scalar function relating the
th
j measurements to
states, and
j
e is the measurement error, which is assumed to have zero mean and variance
2
j
. There are m measurements and n
state variables, n<m.
The WLS state estimation can be formulated mathematically as an optimization problem with a quadratic objective function
and with equality and inequality constraints,
minimize
=
=
m
j
j j j
x h z x J
1
2
/ )) ( (
2
1
) ( (2)
subject to
; 0 ) ( = x g
i

g
n i , 1 = (3)
; 0 ) ( x c
i

c
n i , 1 = (4)
Equation (1) is an objective function, and ) (x g
i
and ) (x c
i
are the functions representing power flow quantities.

3. Radial Basis Function Neural Network

The RBF model consists of two layers namely, a hidden layer consisting of transfer function and a linear output layer. A radial
basis neuron receives the vector distance between its weight vector (cluster center) W and the input vector divided by the spread
constant factor unlike sum of product of the inputs and respective synaptic weights as in case of feed forward network.
The RBF unit or transfer function is similar to Gaussian density function, which is defined by center position and spread. The
output of the RBF unit is given by
=
r
j
ij
W
jp
x
i
o
1
2
) (
exp
(5)
where
pattern input the of variable j
unit RBF the of Spread
j able input vari for unit RBF i the of Center
th
th
=
=
=
x
W
ij

The RBF neural network generalizes on the basis of pattern matching. The different patterns are stored in a network in form of
cluster centers of the neurons of the hidden units. The number of neuron determines the number of cluster centers that are stored in
the network. If the number of neurons are as large as number of training patterns, permissible maximum number of neurons, all the
input patterns will be recognized as separate cluster center, thus it acts like a memory. In case the number of neurons is less than
the training patterns, the network will group the similar inputs patterns a single cluster. Thus it will act like a generalizer. The
response of particular hidden layer node is maximum (i.e. 1) when the incoming pattern matches the cluster center of the neuron
perfectly and the response decays monotonically as the input patterns mismatches the cluster center; the rate of decay can be small
or large depending on the value of the spread. Neurons with large spread will generalize more, as it will be giving same responses

21

(closer to 1) even for the wide variation in the input pattern and the cluster centers whereas a small spread will reduce the
generalization property and work as a memory. Therefore, spread is an important parameter and depends on the nature of input
pattern space.
The output linear layer simply acts as an optimal combiner of the hidden layer neuron responses. The weights w for this layer
are found by multiple linear regression technique. The output of the linear layer is given by
=
+ =
N
i
i i mi mp
b o w y
1
(6)
where,
units) (RBF nodes layer hidden of number = N
pattern incomming i for the
layer output in the node m the of ue output val
th
th
=
mp
y

node output
th
m the of strength biasing =
i
b

layer. linear the input to
th
i =
i
o
node output
th
m and unit RBF
th
i between weight
mi
w =

The values of the different parameters of the RBF networks are determined during training. These parameters are spread, cluster
centers, and weights and biases of the linear layer. The number of neurons for the network and spread is determined through
experimentation with a large number of combinations of spread and number of neuron. The best combination is one which
produces minimum Sum Squared Error (SSE) on the testing data.
The optimal set of parameters is determined through following procedure. Initially start training with two neurons. The number
of cluster centers is the same as number of neurons. The number of neurons is increased gradually till the SSE is reduced to
required goal on the training data for a predefined spread. This process is repeated for different number of neurons and spread
constants, and finally the one giving the lowest error on the testing data is taken as final architecture.

4. Design of Neural Network State Estimator

The radial basis function neural network state estimator is designed and tested for IEEE 14 bus test system (Wallach, 1986). The
line flows and bus injection patterns were generated around the base case using Newton-Raphson power flow. The patterns were
randomly generated taking random perturbations of % 25 around the base case. The patterns included data for intact system and
various single line outage cases. The power flow patterns were first divided into testing and training patterns. The measurement set
containing 28 real and reactive bus injections, 80 real and reactive line flows and 20 line switch status were taken. In all 2000
patterns were generated around the base case out of which 1700 patterns were taken as training patterns and 300 as testing patterns.
The radial basis function network was designed using MATLAB Neural Network tool box (Demuth et al, 1992). The network
parameters were selected in the sum squared error range (for training data) of 3.00 to 0.5 in the steps of 0.5 and the spread range of
5.00 to 0.2 in steps of 0.2. The range was decided by making trail and error guesses. The exhaustive search for the best network in
the above ranges was made. The network was selected on the basis of sum squared error on the testing data set in the said range of
sum squared error and spread. The RBF network which had sum squared error of 0.5 on the training data and spread of 4.6 had the
minimum error on the testing data (0.000233). The network had 19 hidden layer RBF units. This network was selected as state
estimator for further investigations

5. Test Result and Discussions

The performance of the proposed estimator is compared with a conventional WLS state estimator. The performance comparison
is made for following test cases.
Case 1: All the measurements are correct (no gross errors).
Case 2: Gross errors in the measurement data.
Case 3: Topological errors in the measurement data.
Gaussian noise of 2% and 4% was introduced in the actual bus power measurements (real and reactive) and line measurements
(real and reactive) respectively. Tables 1 and 2 shows the estimated voltage magnitudes (V) and bus angle for such a case where
no gross errors were present. The Absolute Error, Mean Average Error (MAE) and the Maximum Error (Max.) is also compared
for conventional WLS (indicated as Con.) and RBF neural network (indicated as ANN) estimator in these tables. It is observed that
the accuracy of the conventional (WLS) method was slightly superior when compared to the proposed method for test case 1.
However, estimated states by the proposed method were quite accurate for practical purposes.

22

Table 1. Voltage estimates for case 1
Bus
No.
V
(Act.)
V
(Con.)
V
(ANN)
Abs Error
(Con.)
Abs Error
(ANN)
1 1.06 1.06 1.06 0 0
2 1.045 1.0454 1.0456 0.0004 0.0006
3 1.01 1.0104 1.0103 0.0004 0.0003
4 1.014 1.0143 1.0152 0.0003 0.0012
5 1.0168 1.0171 1.0178 0.0003 0.001
6 1.07 1.0704 1.0701 0.0004 1E-04
7 1.0589 1.0589 1.0627 00000 0.0038
8 1.09 1.0894 1.0915 0.0006 0.0015
9 1.052 1.0518 1.0549 0.0002 0.0029
10 1.0477 1.0478 1.0501 0.0000 0.0024
11 1.0552 1.0557 1.0573 0.0005 0.0021
12 1.0551 1.0556 1.0569 0.0005 0.0018
13 1.0493 1.0495 1.0511 0.0002 0.0018
14 1.0311 1.029 1.0351 0.0021 0.004
MAE 0.000429 0.001679
Max. 0.0021 0.004

Table 2. Bus angle estimates for case 1
Bus
No.
Angle
(Act.)
Angle
(Con.)
Angle
(ANN)
Abs Error
(Con.)
Abs Error
(ANN)
1 0 0 0 0 0
2 -0.0864 -0.0855 -0.0853 0.0009 0.0011
3 -0.2188 -0.2174 -0.2169 0.0014 0.0019
4 -0.1789 -0.1774 -0.1774 0.0015 0.0015
5 -0.1522 -0.1508 -0.1508 0.0014 0.0014
6 -0.2467 -0.2452 -0.2461 0.0015 0.0006
7 -0.2319 -0.2304 -0.2308 0.0015 0.0011
8 -0.2319 -0.2303 -0.2308 0.0016 0.0011
9 -0.2594 -0.2581 -0.2599 0.0013 0.0005
10 -0.262 -0.2605 -0.2625 0.0015 0.0005
11 -0.2562 -0.2547 -0.2557 0.0015 0.0005
12 -0.2607 -0.2593 -0.2598 0.0014 0.0009
13 -0.2629 -0.2613 -0.2616 0.0016 0.0013
14 -0.2802 -0.2796 -0.278 0.0006 0.0022
MAE 0.001264 0.001043
Max 0.0016 0.0022

Tables 3 and 4 show the estimates for (V) and bus angle for test case 2. The gross errors were introduced randomly in four
measurements, one each in real power injection, reactive power injection, real power line flows and reactive power line flow
measurements. These gross errors are indicated in tables 5-8. The measurements given in the brackets are the actual measurements
which were replaced by gross errors. The gross errors were introduced in real power on bus 3 (Table 5), reactive power on bus 2
(Table 6), real power flow in line 5-1 (Table 7) and reactive power flow in line 7-8 (Table 8). It is observed from tables 3 and 4
that the proposed estimator (ANN) outperformed the conventional WLS estimator both on account of MAE (as 0.0114640.0146
for V, and 0.0209290.036957 for angles) and maximum errors (as 0.05010.1128 for V and 0.0686 0.129 for angles). The
Radial Basis Function neural network state estimator was robust when compared with WLSE state estimator for the test case 2.
This shows that the gross errors present in the measurement data can deteriorate the performance of the conventional state
estimator and have to be removed or reweighed and are to be re-estimated by going through state re-estimation. However, the

23

proposed state estimator is robust in such cases. The bad data can easily be detected in case of proposed method, as there is no data
smearing possible unlike conventional state estimation.

Bus
No.
V
(Act.)
V
(Con.)
V
(ANN)
Abs Error
(Con.)
Abs Error
(ANN)
1 1.06 1.06 1.06 0 0
2 1.0448 1.0427 1.0382 0.0021 0.0066
3 1.01 1.0459 1.0084 0.0359 0.0016
4 1.0138 1.0132 0.9947 0.0006 0.0191
5 1.015 1.0127 1.0028 0.0023 0.0122
6 1.07 1.0687 1.0698 0.0013 0.0002
7 1.0658 1.0677 1.0157 0.0019 0.0501
8 1.09 1.0907 1.0497 0.0007 0.0403
9 1.0236 1.0443 1.0208 0.0207 0.0028
10 1.0231 1.0316 1.014 0.0085 0.0091
11 1.0419 1.0395 1.0518 0.0024 0.0099
12 1.0527 1.0447 1.0497 0.008 0.003
13 1.0446 1.0518 1.0417 0.0072 0.0029
14 1.0127 1.1255 1.0154 0.1128 0.0027
MAE 0.0146 0.011464
Max. 0.1128 0.0501

Bus
No.
Angle
(Act.)
Angle
(Con.)
Angle
(ANN)
Abs Error
(Con.)
Abs Error
(ANN)
2 -0.0888 -0.0741 -0.0828 0.0147 0.006
3 -0.2249 -0.0959 -0.1563 0.129 0.0686
4 -0.1773 -0.1688 -0.1933 0.0085 0.016
5 -0.1557 -0.1554 -0.1636 0.0003 0.0079
6 -0.2845 -0.3194 -0.2561 0.0349 0.0284
7 -0.1773 -0.2082 -0.2058 0.0309 0.0285
8 -0.1773 -0.2209 -0.2058 0.0436 0.0285
9 -0.3316 -0.3727 -0.3336 0.0411 0.002
10 -0.3279 -0.3728 -0.3322 0.0449 0.0043
11 -0.3083 -0.3518 -0.2799 0.0435 0.0284
12 -0.3021 -0.3408 -0.2734 0.0387 0.0287
13 -0.3052 -0.3477 -0.2791 0.0425 0.0261
14 -0.3399 -0.3847 -0.3203 0.0448 0.0196
MAE 0.036957 0.020929
Max 0.129 0.0686


24

Table 5. Real power injection measurements and estimates (the bracketed values are actual values)
Bus
No.
P Meas. P Est.
(Conv)
P Est.
(ANN)
Error
(Con.)
Error
(ANN)
1 2.3494 2.1243 2.3331 0.2251 0.0163
2 0.141 -0.1385 0.0972 0.2795 0.0438
3 0.9747
(-0.9747)
0.3548 -0.1824 0.6199* 1.1571*
4 -0.4583 -0.6902 -1.2235 0.2319 0.7652
5 -0.0709 -0.1856 -0.1492 0.1147 0.0783
6 -0.11 -0.1747 0.0702 0.0647 0.1802
7 0 -0.1251 -0.0621 0.1251 0.0621
8 0 -0.0835 0 0.0835 0
9 -0.293 -0.4225 -0.2854 0.1295 0.0076
10 -0.09 -0.1658 -0.3263 0.0758 0.2363
11 -0.0361 -0.0912 0.1768 0.0551 0.2129
12 -0.0622 -0.115 -0.0584 0.0528 0.0038
13 -0.1265 -0.2326 -0.147 0.1061 0.0205
14 -0.1703 0.0812 -0.0964 0.2515 0.0739

Table 6. Reactive power injection measurements and estimates
Bus
No.
Q Meas. Q Est.
(Con)
Q Est.
(ANN)
Error
(Con.)
Error
(ANN)
1 -0.1504 -0.0446 0.0368 0.1058 0.1872
2 0 (0.36463) 0.0962 0.289 0.0962 0.289*
3 0.0862 0.0354 -0.0829 0.0508 0.1691
4 -0.0408 -0.004 0.0481 0.0368 0.0889
5 -0.0183 0.0397 -0.028 0.058 0.0097
6 0.1294 0.0822 0.1418 0.0472 0.0124
7 0 0.0248 -0.2045 0.0248 0.2045
8 0.147 0.1427 0.2023 0.0043 0.0553
9 -0.1803 -0.3108 -0.0783 0.1305 0.102
10 -0.0666 -0.1276 -0.1341 0.061 0.0675
11 -0.0188 -0.0569 0.0529 0.0381 0.0717
12 -0.0166 -0.0779 -0.0334 0.0613 0.0168
13 -0.0642 -0.1993 -0.0845 0.1351 0.0203
14 -0.0514 0.5411 -0.0461 0.5925* 0.0053

For proposed estimator it is clear from the tables 5-8 that the maximum errors occur exactly at the same data where the gross
error is present, which is not the case with conventional WLS state estimator. This is highlighted (*) in Table 5 to Table 8. For
case 3 a topological error was simulated as inclusion error of line 2-4. The line was actually out but the status (measurement)
showed it to be in the system. WLS program was run with the line flow measurement as zero (both real and reactive), as
acceptable. Results for this case are depicted in Tables 9 and 10. It is observed that the proposed estimator out performed the WLS
estimator both an account of maximum error and MAE for voltage magnitude and bus angles.
Both the programs were coded in MATLAB for fair comparison. For same data set and file operations. Both programs were run
on Pentium IV, Compaq Presario machine for 10 different data sets. The average time for the proposed neural network was 0.0550
seconds, whereas for WLS. State estimator it was found to be 0.8767 seconds. Time is one of the major advantages of the proposed
estimator


25

Table 7. Real power flow measurements and estimates
LINE P Flow Meas P Est.
(Conv)
P Est.
(ANN)
Error
(Con.)
Error
(ANN)
1 2 1.5179 1.3559 1.5233 0.162 0.0054
2 3 0.7517 0.1103 0.4063 0.6414 0.3454
2 4 0.5323 0.5707 0.6693 0.0384 0.137
1 5 0.7992 0.7684 0.8099 0.0308 0.0107
2 5 0.4432 0.5046 0.5048 0.0614 0.0616
3 4 -0.257 0.4644 0.2165 0.7214 0.4735
4 5 -0.51 -0.2918 -0.6908 0.2182 0.1808
5 6 0.5593 0.752 0.4221 0.1927 0.1372
4 7 0 0.2086 0.0621 0.2086 0.0621
7 8 0 0.0835 0 0.0835 0
4 9 0.2821 0.3975 0.2635 0.1154 0.0186
9 10 -0.0356 0.0527 0.0119 0.0883 0.0475
6 11 0.17 0.2098 0.1479 0.0398 0.0221
6 12 0.0883 0.1155 0.0945 0.0272 0.0062
6 13 0.2409 0.2519 0.2499 0.011 0.009
9 14 0.0411 -0.0777 -0.0337 0.1188 0.0748
10 11 -0.129 -0.1136 -0.3146 0.0154 0.1856
12 13 0.0294 -0.0012 0.0349 0.0306 0.0055
13 14 0.1203 0.0142 0.1333 0.1061 0.013
2 1 -1.5554 -1.3241 -1.4832 0.2313 0.0722
3 2 -0.7371 -0.1097 -0.3989 0.6274 0.3382
4 2 -0.5348 -0.5533 -0.6449 0.0185 0.1101
5 1 0.7467 (-0.747) -0.7396 -0.7774 1.4863* 1.5241*
5 2 -0.408 -0.4912 -0.4911 0.0832 0.0831
4 3 0.2561 -0.4512 -0.2134 0.7073 0.4695
5 4 0.4644 0.2931 0.6972 0.1713 0.2328
6 5 -0.6321 -0.752 -0.4221 0.1199 0.21
7 4 0 -0.2086 -0.0621 0.2086 0.0621
8 7 0 -0.0835 0 0.0835 0
9 4 -0.3218 -0.3975 -0.2635 0.0757 0.0583
10 9 0.0388 -0.0521 -0.0117 0.0909 0.0505
11 6 -0.1703 -0.2059 -0.146 0.0356 0.0243
12 6 -0.0938 -0.1138 -0.0934 0.02 0.0004
13 6 -0.2144 -0.2482 -0.2456 0.0338 0.0312
14 9 -0.0423 0.0873 0.034 0.1296 0.0763
11 10 0.1306 0.1146 0.3228 0.016 0.1922
13 12 -0.0277 0.0014 -0.0347 0.0291 0.007
14 13 -0.1169 -0.0061 -0.1304 0.1108 0.0135


26

Table 8. Reactive power flow measurements and estimates
LINE Q Flow
Meas
Q Est.
(Con)
Q Est.
(ANN)
Error
(Con.)
Error
(ANN)
1 2 -0.1922 -0.1134 -0.075 0.0788 0.1172
2 3 0.0345 -0.0653 0.0507 0.0998 0.0162
2 4 0.0092 -0.007 0.0513 0.0162 0.0421
1 5 0.0604 0.0688 0.1118 0.0084 0.0514
2 5 0.0366 0.0162 0.0477 0.0204 0.0111
3 4 0.0697 0.0154 -0.0176 0.0543 0.0873
4 5 0.1273 0.0994 0.0327 0.0279 0.0946
5 6 0.1551 0.1621 0.0685 0.007 0.0866
4 7 -0.139 -0.1506 0.0095 0.0116 0.1485
7 8 -0.0141
(-0.141)
-0.1386 -0.1958 0.1245 0.1817*
4 9 0.0669 0.045 0.031 0.0219 0.0359
9 10 0.0205 0.137 0.0771 0.1165 0.0566
6 11 0.0714 0.0598 0.0277 0.0116 0.0437
6 12 0.03 0.0459 0.0391 0.0159 0.0091
6 13 0.0895 0.0147 0.1057 0.0748 0.0162
9 14 0.0222 -0.2768 0.0365 0.299 0.0143
10 11 -0.0463 0.0077 -0.0575 0.054 0.0112
12 13 0.011 -0.0355 0.0034 0.0465 0.0076
13 14 0.0377 -0.2276 0.016 0.2653 0.0217
2 1 0.3016 0.1522 0.1392 0.1494 0.1624
3 2 0.0181 0.02 -0.0653 0.0019 0.0834
4 2 -0.0016 0.0203 -0.0159 0.0219 0.0143
5 1 0.0065 -0.0026 -0.0301 0.0091 0.0366
5 2 -0.0414 -0.0112 -0.0413 0.0302 1E-04
4 3 -0.0871 -0.0182 -0.0092 0.0689 0.0779
5 4 -0.1305 -0.1086 -0.0251 0.0219 0.1054
6 5 -0.0442 -0.0382 -0.0307 0.006 0.0135
7 4 0.1556 0.1634 -0.0088 0.0078 0.1644
8 7 0.1517 0.1427 0.2023 0.009 0.0506
9 4 -0.0189 0.0363 0.006 0.0552 0.0249
10 9 -0.019 -0.1353 -0.0766 0.1163 0.0576
11 6 -0.0613 -0.0515 -0.0238 0.0098 0.0375
12 6 -0.0282 -0.0424 -0.0368 0.0142 0.0086
13 6 -0.0869 -0.0074 -0.0973 0.0795 0.0104
14 9 -0.0217 0.2972 -0.0359 0.3189* 0.0142
11 10 0.0483 -0.0053 0.0766 0.0536 0.0283
13 12 -0.0107 0.0357 -0.0032 0.0464 0.0075
14 13 -0.0302 0.2439 -0.0103 0.2741 0.0199


27

BUS
No.
V
(Act.)
V
(Con.)
V
(ANN)
Abs
Error
(Con.)
Abs
Error
(ANN)
1 1.06 1.06 1.06 0 0
2 1.045 1.0428 1.0443 0.0022 0.0007
3 1.01 1.0136 1.0092 0.0036 0.0008
4 1.0019 1.0137 1.0103 0.0118 0.0084
5 1.0073 1.0171 1.0142 0.0098 0.0069
6 1.07 1.0776 1.0699 0.0076 1E-04
7 1.0532 1.0613 1.058 0.0081 0.0048
8 1.09 1.0976 1.0928 0.0076 0.0028
9 1.0461 1.0536 1.0517 0.0075 0.0056
10 1.0426 1.0497 1.0459 0.0071 0.0033
11 1.0525 1.0597 1.0576 0.0072 0.0051
12 1.0538 1.0614 1.0557 0.0076 0.0019
13 1.0484 1.0559 1.0493 0.0075 0.0009
14 1.0271 1.0331 1.0328 0.006 0.0057
MAE 0.006686 0.003357
Max 0.0118 0.0084

BUS
No.
Angle
(Act.)
Angle
(Con.)
Angle
(ANN)
Abs
Error
(Con.)
Abs
Error
(ANN)
1 0 0 0 0 0
2 -0.0824 -0.0903 -0.0832 0.0079 0.0008
3 -0.2588 -0.2362 -0.2347 0.0226 0.0241
4 -0.2377 -0.1876 -0.1908 0.0501 0.0469
5 -0.1932 -0.1578 -0.1665 0.0354 0.0267
6 -0.2959 -0.2522 -0.2601 0.0437 0.0358
7 -0.287 -0.2423 -0.2442 0.0447 0.0428
8 -0.287 -0.2438 -0.2442 0.0432 0.0428
9 -0.3124 -0.2681 -0.2769 0.0443 0.0355
10 -0.3144 -0.2701 -0.2806 0.0443 0.0338
11 -0.3075 -0.263 -0.2688 0.0445 0.0387
12 -0.3111 -0.2654 -0.2745 0.0457 0.0366
13 -0.3123 -0.2669 -0.277 0.0454 0.0353
14 -0.331 -0.2863 -0.2939 0.0447 0.0371
MAE 0.036893 0.031207
Max. 0.0501 0.0469

6. Conclusion

Radial Basis Function neural network is designed and tested for IEEE 14-bus system. The performance of the proposed neural
network estimator is compared with conventional WLSE on basis of time and accuracy. The time taken by the RBF estimator was
very less, compared to its conventional counterpart. It is shown that for gross error and topological error present in the
measurement data, the proposed RBF state estimator was more accurate as compared to WLSE. The bad data analysis shows that
the gross errors can easily be detected in case of RBF estimator as there is no data smearing as in case of conventional WLSE.
However, WLSE method was superior in cases where the gross is not present in the measurements. However, state estimates
obtained by the RBF estimator were accurate for practical purposes.

28

References

Demuth Howard, Beale Mark, 1992, Neural Networks Toolbox, The Mathworks Inc.
Garcia A., Monticelli A., and Abreu P., 1979, Fast decoupled static state estimation and data processing, IEEE Transactions,
PAS, vol. 98, No. 5, pp. 1645-1652.
Horisberger H.P., Richard J.C. and Rossier C., 1976, A fast decoupled static state estimator for electric power system, IEEE
Transactions, PAS, vol. 95, issue 1, pp. 208-215
Monticelli A., 2000, Electric power system state estimation, Proc. IEEE, vol. 88, No.2, pp.262-282.
Singh D. and Singh S.P., 2001, A self selecting neural network for short-term load forecasting, Journal of Electric Power
Components and Systems, No. 29, pp.117- 130.
Wallach Y., 1986, Calculations and Programs for Power System Networks, New Jersy, Prentice Hall Inc., pp. 40-41.
Wu F.F., 1990, Power system state estimation: A survey, Elect. Power Eng. Syst. Vol. 12, pp. 80-87.

Biographical notes:

Dr. J.P. Pandey received his BTech and MTech degree in Electrical Engineering from Kamala Nehru Institute of Technology Sultanpur (UP), India. He obtained
his PhD degree from UP Technical University, Lucknow, India. Presently he is professor in the department of electrical engineering at KNIT Sultanpur. His
research interests include applications of artificial techniques to electrical engineering problems and power system.

Dr. D. Singh received his Ph.D. degree from Institute of Technology (IT), BHU, India. He is working as Associate Professor at IT BHU, India. His research
interests include applications of artificial techniques to electrical engineering problems and power system.

Received September 2009
Accepted March 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 29-44

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Improved transformer protection using probabilistic neural network and
power differential method

Manoj Tripathy
1
*, R. P. Maheshwari
2
, H. K. Verma
3

1*
Department of Electrical Engineering, Motilal National Institute of Technology Allahabad, INDIA
2,3
Department of Electrical Engineering, Indian Institute of Technology Roorkee, INDIA
*
Corresponding Author: e-mail: manoj_tripathy1@rediffmail.com, Tel +91-9412015058, Fax. +91-532-2445101

Abstract

This article presents a novel technique to distinguish between magnetizing inrush current and internal fault current of power
transformer. An algorithm has been developed around the theme of the conventional differential protection method in which
parallel combination of Probabilistic Neural Network (PNN) and Power Differential Protection (PDP) methods have been used.
Both PNN and PDP method are independent of harmonic contents of differential current. The proposed algorithm is capable of
detecting fault and its type in the eventuality of fault in the transformer. Moreover, the combination of PDP method with the
PNN makes it capable to detect light internal faults for all ratings of transformers which improve the overall performance of
digital differential protection scheme. For evaluation of presented algorithm, relaying signals of various operating conditions of
power transformer, including internal faults, external faults, over-excitation and inrush conditions were obtained through
modeling of transformer in PSCAD/EMTDC. The performance of proposed amalgamated technique (i.e. combined PNN and
PDP method) is compared with the PNN, Feed Forward Back Propagation (FFBP) neural network and the conventional
harmonic restraint methods. The results amply demonstrate the capability of the proposed algorithm in terms of accuracy and
speed. The algorithm has been implemented in MATLAB.

Keywords: Digital differential protection, Protective relaying, Probabilistic neural network, Active power relays, Power
differential method.

1. Introduction

Protection of large and medium power transformers by means of differential relaying has been a common practice. Differential
relaying technique is based on comparison of the transformers two winding currents. When these currents deviate from a pre-
defined relationship an internal fault is assumed and relay operates. However, during magnetizing inrush condition, very high
current of the order of 10-15 times of full load current of transformer may pass through the primary side of power transformer
(Sidhu et al, 1992). This high current causes mal-operation of the relay. Therefore, main challenge is to precisely distinguish
between magnetizing inrush and fault current to avoid any mal-operation of relay. Literature review suggests that broadly two
approaches are applied to discriminate between magnetizing inrush and fault currents; these are Harmonic Restraint (HR) based
method and Waveform Identification (WI) based method (Tripathy et al, 2005). The HR method is based on the fact that the
second/fifth harmonic component of the magnetizing inrush current is considerably larger than that in a typical fault current. It has
been extensively used in comparison of WI method (Tripathy et al, 2005, Wang et al, 2009). However, this method sometimes fail
to discriminate between magnetizing inrush and internal fault currents because high second harmonic components are generated
during internal faults and low second harmonic component are generated during magnetizing inrush having modern core material
of power transformer and due to the presence of shunt capacitance or distributive capacitance in long Extra High-Voltage (EHV)
transmission line to which power transformers are connected (Shin et al, 2003). Therefore, the detection of second/fifth harmonic
is not a sufficient index to discriminate between the inrush and fault currents of a power transformer.
The second method distinguishes magnetizing inrush current from internal fault current on the basis of WI method. In 1986,
Verma and Basha reported microprocessor based waveform differential relaying scheme (Verma et al, 1986). The magnetizing
Tripathy et al. / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 29-44

30

inrush current exhibits a characteristic peaked wave, which is caused by asymmetric saturation of the transformer core. In 2006, B.
He et al. presented an algorithm to identify inrush current and internal fault current in transformer by using the characteristic of
inrush current i.e. peaked wave and dead angle (He et al, 2006). Another group of researchers have investigated the use of wavelet
analysis for the classification between fault and healthy state of power transformer. Wavelet transform based methods have better
ability of time-frequency location. Their disadvantages are that they need long data window and are also sensitive to noise and
unpredicted disturbances, which limit their application in relaying (Arboleya et al, 2006, Tripathy et al, 2009).
It has been demonstrated that the advanced Digital Signal Processing (DSP) techniques and Artificial Intelligence (AI)
approaches to power system protection can improve discrimination between normal and fault conditions and facilitate faster, more
secure and dependable protection for power transformers. Owing to its superior learning and generalization capabilities Artificial
Neural Network (ANN) can considerably enhance the scope of WI method. ANN approach is faster, robust and easier to
implement than the conventional waveform approach. The use of neural network can provide an intelligent digital differential
protection scheme. Since 1994, many researchers have proposed ANNs based transformer differential protection with various
topologies. Most of them used Multilayer Feed Forward Neural Network (MFFNN) with back-propagation learning technique
(Bastard et al, 1995, Tan et al, 2004). Another ANN model called Radial Basis Function Neural Network (RBFNN) has also been
reported in the literature for power transformer protection (Moravej et al,2003, Borghetti et al, 2008). However, they have so far
left some unsolved problems, including those of local minima and slow convergence in training and the need of empirical
determination of structure and neural network parameters. In (Tripathy et al,2007), Probabilistic Neural Network (PNN) has been
used to discriminate between inrush current and internal fault current of power transformer but it fails to detect turn-turn fault
current (i.e. light internal fault current).
This paper presents an algorithm in which PNN is applied in conjunction with Power Differential Protection (PDP) method
(Yabe, 1997). The PNN and PDP methods together are used to discriminate between magnetizing inrush current and internal fault
current, and also used to determine the type of fault in a power transformer. They are able to detect internal fault current even if the
fault current contains a large second harmonic component, as PNN method and combined PNN and PDP method are independent
of harmonic components present in operating signal to the relay. PNN method is based on pattern recognization while PDP method
is based on the average instantaneous power flowing into the transformer. The proposed amalgamated technique is used to
improve the reliability and accuracy as compare to PNN based method and conventional HR method. In this paper, the results of
the proposed algorithm are compared with the conventional Discrete Fourier Transform (DFT) based method, Feed Forward Back
Propagation (FFBP) based method and optimal PNN based transformer differential protection method.

2. Probabilistic Neural Network (PNN)

PNN is a kind of feed forward neural network. The original PNN structure is a direct neural network implementation of Parzen
nonparametric Probability Density Function (PDF) estimation and Bayes classification rule (Specht, 1990). The standard training
procedure of PNN requires a single pass-over all the patterns of training set (Specht, 1990). This characteristic renders PNN faster
to train as compared to FFBP neural network and RBFNN (Specht et al, 1991). The only drawback of PNN is the requirement of
larger storage for exemplar patterns. As the computer memory has become very cheap and effective, the cost and size of large
storage are no longer of concern these days. PNN is widely used in the area of pattern recognization, nonlinear mapping, fault
detection and classification, estimation of probability of class membership and likelihood ratios (Tian et al, 2001).
x1
xi
xd
out put
Input Layer Pattern Layer
Summation
Layer
Output Layer

Figure 1. Probabilistic neural network structure

The PNN structure is shown in Figure 1. It is a four layer feed forward neural network that is capable of realizing or
approximating the optimal classifier. Generally, Gaussian activation function is used in PNN because if the pattern falls within
certain region then the function output is 1 otherwise function output is 0. It is not related to any assumption about normal

31

distribution. The activation function for PNN is derived from estimates of PDF based on the training patterns as following (Bose et
al, 1996):
Let
d
R X be a d-dimensional pattern vectors and its associated class be
( )
1 2 3
, , , ...,
k
i S S S S
. Where, k is the number of
possible class. If a posteriori probability,
( ) x S P
i r
/
that is from class
i
S , is by Bayes rule,
( )
( ) ( )
( ) x P
S P S x P
x S P
i r i r
i r
/
/ = (1)
where,
( ) / , 1, 2, 3, ...,
r i
P x S i k = is a priori PDF of the pattern in classes to be separated.
( ), 1, 2, 3, . . .,
r i
P S i k =
are the priori probabilities of the classes.
( ) x P is assumed to be constant.
The decision rule is to select class
i
S for which ( ) x S P
i r
/ is maximum. This will happen if for all
i j

( ) ( ) ( ) ( )
j r j i r i
S P S x P S P S x P / / >
(2)
It is assumed that a priori probabilities
( )
i r
S P
of the classes are known and the a priori PDF,
( )
i
S x P /
is Gaussian then the
estimator for a priori PDF is
( )
( )
( ) ( )
^
2
2
1
1
/ ex p
2
2 | |
i
T
i i
n
j j
i d
d
j i
i i
x x x x
P x S
S

=

=

(3)
Where,
i
j
x
is

j
th
exemplar pattern from class
i
S

| |
i
S =
i
n

the cardinality of the set patterns in class
i
S
i

= Smoothing factor
The input layer has d units, to which the d-dimensional input vector
d
X R is applied. The first hidden layer has one pattern
unit for each pattern exemplar. Therefore, each such pattern unit may be associated with a generic term depicted in the summation
of equation (3) for the i
th
class. The second hidden layer contains one summation unit for each class. The output layer is decision
layer used for implementing the decision rule by selecting maximum posteriori probability, ( | )
r
p S x
i
from outputs preceding
summation layer for each i. The network is constructed by setting weight vector to one of the pattern unit equal to each distinct
pattern vector in the training set from a certain class and then connecting the outputs of the pattern units to the appropriate
summation units for that class.
For PNN, many algorithms are available in the literature to achieve optimized exemplar pattern set that means removing
redundant data or duplicate information. It increases the speed of classification with reduced size of exemplar pattern set while still
providing sufficient data to fill the data space (Berthold et al, 1998). The optimal selection of smoothing parameter in PNN
classifier is very important factor. The PNN decision boundary varies from a hyper plane to a very nonlinear boundary when the
smoothing parameter varies from 0 (Zero) to (infinite). In literature many methodology are reported for the selection of
appropriate widths or smoothing factor (Hammond et al, 2004, Musavi et al, 1992). In the present paper a very simple method is
used for the calculation of smoothing factor to avoid complex calculation (Bose et al, 1996):
i
javg
g
i
d =

(4)
Where,
i
j
d = Distance between the j
th
exemplar pattern and nearest exemplar pattern in class i.
g = Constant that has been by found trial and error.

3. Power Differential Protection (PDP) Method

The inflow and outflow of instantaneous power through transformer is according to the magnetic energy stored in transformer
winding. Under normal operating condition, about 1% of the transformer power flows in the magnetizing circuit. In normal
operating condition, total power flowing into transformer is about 1% of the transformer capacity because the copper losses and
core losses are of the same order. The PDP method is based on average instantaneous power flow into transformer winding during

32

one cycle period T (20 ms for 50 Hz system). The PDP algorithm calculates the products of instantaneous current and voltage and
then calculates the average instantaneous power as given by
( ) ( )
2 2
1 1 2 2 1 1 2 2
1
i i - -
t
T t
w t u u r i r i dt
T

= +
(5)
To implement the PDP method by digital relay the following mathematical expressions are given as (Yabe, 1997):
2 2
1 1 2 2 1 1 2 2
( ) i i - - p t u u r i r i = + (6)

1
1
( ) T n 1, 2 , 3 , . . . ,
T
N
n
n
w t p t N
N
=

= =

(7)
Where,
1 2 1 2
, , , u u i i are instantaneous voltage and current at primary and secondary winding terminals respectively.
r
1
, r
2
= Primary and secondary winding resistances respectively
) ( t p
= Instantaneous power
) ( t w
= Average power of one period time
N = Number of samples per cycle

PDP method is not affected by the harmonic components present in either internal fault current or inrush current because it uses
average instantaneous power (Yabe, 1997). In magnetizing inrush condition, average instantaneous power from second period after
energization is almost equal to core losses plus stray losses. On the other hand, under an internal fault condition large amount of
power is consumed proportional to fault degree (as a portion of winding is short circuited). Therefore, by setting a suitable
threshold of average power flowing into the transformer, magnetizing inrush and fault condition can be discriminated.
In the proposed algorithm, the PDP method is utilized to distinguish light internal fault current only as it needs at least one cycle
to discriminate different operating condition of power transformer and this is affordable in case of light internal fault only. This
technique is capable of detecting fault and its type in the eventuality of fault in transformer. It is operated in conjunction with PNN
method. The trip decision of the relay is based on ORing of the two decisions obtained.

4. Simulation and Training Cases

During power transformer operation, it encounters any one of the following conditions:
Normal condition
Over-excitation condition
Magnetizing / sympathetic inrush condition
Internal fault condition
External fault condition

In normal condition, rated or less current flows through the transformer. In this condition normalized differential current is
almost zero (only no load component of current). Whenever, there is large and sudden change in input terminal voltage of
transformer, either due to switching-in or due to recovery from external fault, a large current is drawn by transformer from the
supply. As a result, the core of transformer gets saturated. This phenomenon is known as magnetizing inrush.
Magnetizing inrush can also occur in an already energized transformer when a nearby transformer is energized. A common
situation of sympathetic inrush is encountered when a transformer is energized in parallel with another transformer already in
service. The phenomenon which causes inrush current to flow in a previously energized transformer is known as the sympathetic
inrush. As the paralleled transformer is being energized by closing the breaker, an inrush current is established in the primary of
this transformer and this inrush current has DC component. The DC component of the inrush current can also saturate the already
energized transformer, resulting in an apparent inrush current. This transient current, when added to the current of already
energized transformer, results in an asymmetrical current that is very low in harmonics. This would be the current flowing in the
supply circuit to both transformers. Sympathetic inrush current may not have sufficient amount of the second harmonic in it to
prevent the relay from tripping. Sympathetic inrush current depends on same factors on which switching-in and recovery from
fault magnetizing inrush current depends.

33

PSCAD/EMTDC simulation is used to generate training as well as testing signals under different operating conditions of
transformer as mentioned above. The simulation set up is given in Figures 2-4. While simulating magnetizing inrush condition,
energization angle, remanent flux in the core and load condition are considered because the magnitude and the wave-shape of
magnetizing inrush current depends on these factors.
Figure 2. Simulation diagram of magnetizing / sympathetic inrush condition of transformer
Figure 3. Simulation diagram of phase-to-ground fault under full-load condition of transformer

Figure 4. Simulation diagram of magnetizing inrush under no-load considering remanence flux in transformer

Energization angle is varied from 0 to 360 degrees in steps of 30 degrees, and remanent flux is varied from 0 % to 80 % of the
peak flux linkages generated at rated voltage with no load and full load conditions to generate training signals, whereas, the testing

34

signals are generated by varying energization angle in step of 15 degrees. The desired remanence can be set in un-energized
transformer with controlled DC current sources in PSCAD/EMTDC simulation model (Woodford, 2001).
For internal fault, training and testing is formed by simulating fault from 1% to 99% of power transformer winding turns. Phase-
to-ground fault at different locations as 5%, 15%, 25%, 40% and 50% of the winding as well as terminal fault are simulated. The
detailed information of power transformer PSCAD/EMTDC simulation model to simulate internal fault (Figure 18) is given in the
Appendix-1. Three-phase transformer of 315 MVA at 400/220 kV, 200 MVA at 220/110 kV and 160 MVA at 132/220 kV are
modeled using PSCAD/EMTDC. For the simulation of these transformers through PSCAD/EMTDC, the parameters are used that
are obtained from M. P. State Electricity Board, Jabalpur India. The test signals so acquired by simulating various operating
conditions of a transformer are shown in Figures.5-11. The simulation was done at the rate of 12 samples per cycle of 50 Hz A.C.
supply in view of reported experience on different digital relay designs (Sachdev, 1998). The developed fault detection algorithm
was implemented in MATLAB.

-0.08
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
0.08
0 0.1 0.2 0.3 0.4 0.5
Time (s)
O
p
e
r
a
t
i
n
g

S
i
g
n
a
l

(
p
u
)

Figure 5. Typical differential current waveform under normal operation
-6
-4
-2
0
2
4
6
8
10
0 0.1 0.2 0.3 0.4 0.5
Time (s)
O
p
e
r
a
t
i
n
g

S
i
g
n
a
l

(
p
u
)

Figure 6. Typical differential current waveform for ground fault

-1.0
0.0
1.0
2.0
3.0
4.0
5.0
6.0
7.0
0 0.1 0.2 0.3 0.4 0.5
Time (s)
O
p
e
r
a
t
i
n
g

S
i
g
n
a
l

(
p
u
)

Figure 7. Typical differential current waveform for magnetizing inrush

Switching-in occurred here
Fault occurred here

35

-4
-3
-2
-1
0
1
2
3
4
0 0.1 0.2 0.3 0.4 0.5
Time (s)
O
p
e
r
a
t
i
n
g

S
i
g
n
a
l

(
p
u
)

Figure 8. Typical differential current waveform for over-excitation

-1.5
-1
-0.5
0
0.5
1
1.5
0 0.05 0.1 0.15 0.2
Time (s)
P
o
w
e
r

(
p
u
)

Figure 9. Typical instantaneous power waveform for ground fault

0
0.2
0.4
0.6
0.8
1
1.2
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18
Time (s)
P
o
w
e
r

(
p
u
)

Figure 10. Typical average power waveform for magnetizing inrush

0.8
0.85
0.9
0.95
1
1.05
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18
Time (s)
P
o
w
e
r

(
p
u
)

Figure 11. Typical average power waveform for ground fault

36

5. Implementation and Relay Logic
5.1. Implementation of PNN Based Algorithm and Comparison with FFBP Neural Network
Four layered PNN structure is used as shown in Figure 1. Input layer of PNN model has 12 neurons. The number of neuron of
input layer are decided based on the dimension of feature space i.e. 12 samples per cycle. The first hidden layer consists of 777
neurons and it is decided based on total number of exemplar pattern set used to construct the PNN model. In training exemplar
pattern set, 444 patterns of magnetizing inrush (including sympathetic inrush patterns), and 333 patterns of faults are used. Second
hidden layer has two neurons, as there are only two classes to be classified. In output layer, single neuron is used, as it is the
decision layer and it is used for selecting the maximum posteriori probability, from the outputs of the summation layer.
The PNN requires one node or neuron for each exemplar pattern. Various clustering techniques have been used to reduce this
requirement to one node per cluster center. Clustered data is required to construct the PNN because the output of first hidden layer
is added and this belongs to one specific cluster that is clear from the PNN architecture. To construct an optimized and efficient
PNN, the training data are clustered by using K-means clustering (Specht, 1992), and the smoothing factor for each class is
calculated by equation (4). The optimal value of multiplication factor (g) is obtained by trial and error method and thus optimal
smoothing factor is achieved (see Appendix-2).
Out of 925 sets of data (patterns), 777 patterns sets are used to construct PNN with optimal smoothing factor which is already
obtained by the conventional method and remaining 148 sets are used to test the networks generalization ability that are different
from those used to train the network. These 148 test exemplar pattern sets contain internal fault and magnetizing inrush condition
only as these two conditions are very difficult to discriminate as compared to other operating conditions like external fault
condition, over-excitation and normal operation from protection point of view. Out of these 148 cases 74 are that of magnetizing
inrush and rest are for internal faults. As other conditions have been taken care before this stage, hence there is no chance of
getting exemplar patterns of other operating conditions at this stage.

Start
Input data (Operating Signal )
Over excitation/ Normal Operation ?
Issue Trip Signal
Fault
NO
Check for Inrush or Fault by PNN ?
Yes
Inrush

Figure 12. Flow chart of PNN based algorithm
External fault and normal operating condition are ruled-out by comparing two consecutive peaks of operating signal whereas the
over-excitation condition is determined by comparing actual value of voltage-to-frequency ratio with its rated value (shown in
Figure 12). If these conditions do not exist then the processing is continued to detect the magnetizing inrush and internal fault
conditions by PNN. The posteriori probabilities are calculated by the summation layer of PNN and the decision layer selects
maximum posteriori probability from the preceding summation layer and on the basis of maximum posteriori probability, the
patterns are classified. Accordingly, the PNN gives tripping signal if an internal fault condition is detected.

37

Many experiments have been made to evaluate the performance of the PNN model. The fault and magnetizing inrush conditions
are tested with all possible patterns. The results are given in Table-1 which presents the natural logarithm of probability of the test
pattern to be internal fault or inrush condition. By comparing the magnitude of these probabilities, the test data belongingness is
decided. The values given in the Table-1 are the conditional probability estimated by PNN. On the basis of these conditional
probability PNN based classifier takes decision i.e. the test data represents either internal fault current or inrush current.
To obtain data for the testing of relaying algorithm fault current magnitude, remanent flux, load condition and switching-in angle
are varied to investigate their effects on the performance of the PNN model. Since the profile of magnetizing inrush current wave
changes with variation of switching-in instant of transformer hence, it is varied between 0 to 360 degrees. Similarly, due to
remanence flux, the magnitude of magnetizing inrush current may be as high as 2 to 6 times to that of the magnetizing inrush
without that, although the wave-shape remains same. It is found that the PNN classifier based relay is stable even with such high
magnitude of magnetizing inrush current caused by remanence flux whereas the conventional harmonic based relay may mal-
operate due to such high magnitude of magnetizing inrush current.

Table-1. Conditional probability estimated by PNN
Type of test data Probability of fault Probability of inrush
Inrush data 18.901 26.004
Fault data 31.742 22.991

The PNN is faster than FFBP neural network. The training required for PNN is very different and much faster than that required
for FFBP neural network (Specht et al, 1991). The training process of PNN is one pass without any iteration for the weight
adaptation, as against a large number of iteration (epochs) necessary in case of FFBP neural network. As an example, 1000
iterations are required for convergence in case of FFBP neural network implemented by the authors, thereby giving a ratio of about
1000:1 in terms of the training time. Therefore, the PNN is free from the demerits like local minima and slow convergence in
training and empirical determination of network structure and parameters. Furthermore, in PNN, single parameter namely
smoothing factor is to be tuned whereas in the FFBP neural network at least four parameters like learning rate, momentum
coefficient, weights etc. are to be tuned during training of neural network. This makes PNN easy to design and simple in use than
the classical ANN.

Figure 13. Variation of accuracy versus multiplying factor (g) on 315MVA power transformer using PNN classifier

Table-2 demonstrates that the optimal PNN fails to discriminate some light internal fault i.e. false negative error. This optimal
PNN is designed with optimal value of smoothing factor that is obtained by classical method given in Table-3. Figure 13 shows the
effect of the multiplying factor on the classification accuracy of a PNN classifier in case of 315MVA power transformer. From the
experimental results, it is found that when PNN is trained with one transformer and tested on different rating of transformer, it fails
in case of light internal fault condition only. Figure14 illustrates the typical waveform of light internal fault case. The nature of
waveform is similar to the typical inrush waveform, and because of similarity in wave-shapes of typical inrush waveform and light
internal fault waveform, the boundary of classification becomes narrow and therefore, it is difficult to discriminate the light
internal fault cases. Therefore, for such cases PDP technique is applied, which is also independent of harmonic component present
in fault current and inrush current.

38

Figure 14. Typical waveform of phase-to-ground fault at 2% of the winding from the neutral end

Start
Input data (Operating signals)
Over-excitation /
Normal operation?
Check for inrush or fault by
P D P method?
Check for inrush or fault
by PNN?
OR Logic
Fault
Fault
No
Yes
Inrush
Inrush
Issue trip signal

Figure 15. Flow chart of PNN and PDP algorithm

5.2. Implementation of Hybrid PNN & PDP method and Result Discussion
The PNN and PDP methods are parallely operated. The flow chart (Figure 15) illustrates all the steps of the hybrid PNN and
PDP method to discriminate between inrush and internal fault conditions. In PDP method instantaneous power (shown in Figure 9)
and average power (shown in Figure 11) are calculated by using equation (6) and (7) respectively. From Figure 10 and Figure 11,
it is clear that there is large difference in average power after one cycle of time. Therefore, with proper setting of threshold value,
operating condition of transformer can be monitored. The threshold value depends on transformer rating and its other parameters.
By this method, the type of fault in transformer can also found due to large difference among the threshold values and that for light
internal fault, external fault and phase-to-phase fault conditions. This technique is used for the detection of low level internal faults
only as there is one cycle delay in decision making which is affordable in this case only, whereas in other cases PNN will
-0.5
0
0.5
1
1.5
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
Time (s)
O
p
e
r
a
t
i
n
g

S
i
g
n
a
l

(
p
u
)

39

recognize the different operating conditions of power transformer within half cycle of its occurrence. Therefore, by combining
these two methods, fault can be detected with certainty which improve the reliability and stability of differential relaying
algorithm. The relay utilize OR logic that means, any input signal either OA or OP is high the relay will issue the trip signal.
Where, OA and OP are the output of PNN method and PDP method respectively. O is output of the relay and the high state of
output implies the trip signal to be generated.

No.of False Positive No.of False Negative
Classification Error (in %) 100
Total Number of Test Cases
+
=

Classification Accuracy (in %) = 100 Classification Error (in %) (8)

Table-4 presents the classification in accuracies (in percentage) with FFBP neural network, PNN and combined PNN and PDP
method. The classification in accuracy is calculated by using (8). From the results of Table-4, it is clear that the classification
ability of PNN is better than the FFBP neural network in this application. However, from Table-4, it is observed that the proposed
combined method has ability to discriminate between the light internal fault (turn-to-turn fault) as well as magnetizing inrush
condition and classification in accuracy of the combined (PNN and PDP) method has improved as compare to the FFBP neural
network and PNN method.
Discrete Fourier Transform (DFT) based harmonic restraint method is implemented, to compare performance of the proposed
optimal PNN based algorithm in power transformer differential protection. Figures 16-17 show the ratio of second harmonic to
fundamental of the differential current under typical magnetizing inrush and internal fault condition respectively. During one cycle
under internal fault condition, the ratio of the second harmonic is quite high and in the same range as in case of magnetizing inrush
condition. Therefore, it is difficult to discriminate between internal fault and inrush conditions merely setting a preset threshold.
From Figures 16-17, it is also clear that the ratio values are fluctuating, which create problem to decide a preset threshold.
Moreover, due to the presence of second harmonic during internal fault condition digital relay will take longer time to make trip
decision (one cycle or more than one cycle). In contrast, the optimal PNN based method is able to detect such a fault in 6 ms (half
cycle or with in half cycle) except light internal fault (turn-to-turn fault) cases. However, the harmonic restraint method is capable
to discriminate between these two conditions but does not seems to be intelligent to take decision in case of fluctuating ratio of
second harmonic to fundamental of the differential current due to different loading conditions, severity of internal faults,
switching-in angles etc. and hence mal-operation of relay will occur. While the proposed amalgamated technique (i.e. combined
PNN and PDP method) is intelligent and more reliable than the conventional harmonic restraint method, FFBP based method and
PNN based method.

0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.04 0.06 0.08 0.1 0.12 0.14
Time(s)
r
a
t
i
o

o
f

s
e
c
o
n
d

t
o

f
u
n
d
a
m
e
n
t
a
l

c
o
m
p
o
n
e
n
t

Figure 16. Ratio of second harmonic to fundamental of the differential current under typical inrush condition (inrush occurs at
0.04 sec.)

40

0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
0.04 0.06 0.08 0.1 0.12 0.14
Time(s)
r
a
t
i
o

o
f

s
e
c
o
n
d

t
o

f
u
n
d
a
m
e
n
t
a
l

c
o
m
p
o
n
e
n
t

Figure17. Ratio of second harmonic to fundamental of the differential current under typical internal fault condition (internal fault
occurs at 0.04 sec.)

Table-5 shows the number of post abnormality samples required for decision by FFBP neural network, PNN and PNN
conjunction with PDP method based relays. From Table-5, it is clear that the PNN is faster than the FFBP neural network and PNN
conjunction with PDP method. However, it fails to recognize some typical light internal fault (turn-to-turn fault) cases whereas the
PNN conjunction with PDP method has 100% classification capability with a reasonable speed as compared among these three
methods. In this table actual value indicates the number of post abnormality sample required for detection of a test pattern by
concerned algorithm while logical value provides information regarding the of maximum number of post abnormality samples
available in a pattern.
Tremendous capability of PNN for classification problems shows suitability for differential transformer protection. It is also
immune from the different harmonics contained in operating signals which makes it simpler and robust than the conventional
digital filtering algorithms.

Table 2. Number of false detections in optimal PNN
Tested transformer ratings
315 MVA 200 MVA 160 MVA
False positives False negatives False positives False negatives False positives False negatives
Inrush Faults Inrush Faults Inrush
Switching -in
angles
Phase-to-
ground
fault
Phase-to-
phase
fault
Switching -in
angles
Phase-to-
ground
fault
Phase-to-
phase
fault
Switching -in
angles
Phase-to-
ground
fault
Phase-to-
phase
fault
Training
transform
er ratings
%
**
Degrees
%
**
%
*
%
**
%
*
%
**
Degrees
%
**
%
*
%
**
%
*
%
**
Degrees
%
**
%
*
%
**
%
*
315 MVA 0 - 0 - 0 - 0 - 1 2 0 - 0 - 0 - 0 -
200 MVA 0 - 0 - 0 - 0 - 0 - 0 - 0 - 2 2,5 0 -
160 MVA 0 - 1 2 0 - 0 - 2 2, 5 0 - 0 - 0 - 0 -
% * Represents location of winding from the neutral end at which fault occurs
% ** Represents number of recurrence


41

Table-3. Optimal multiplication factor (g)
Trained transformer ratings
315 MVA 200 MVA 160 MVA
315 MVA 1.1 1.4 1.1
200 MVA 0.5 0.3 0.5
160 MVA 0.3 0.3 1.1

Table -4. Classification accuracies (in %) with FFBP, PNN and combined PNN and PDP method
315 MVA 200 MVA 160 MVA
Trained transformer
ratings
FFBP
(%)
PNN
(%)
PNN
&
PDP
(%)
FFBP
(%)
PNN
(%)
PNN
&
PDP
(%)
FFBP
(%)
PNN
(%)
PNN
&
PDP
(%)
315 MVA 99.32 100 100 94.59 98.65 100 98.64 100 100
200 MVA 96.62 100 100 99.32 100 100 98.64 96.93 100
160 MVA 98.64 98.65 100 96.62 97.3 100 100 100 100

Table 5. Number of post abnormality samples required for decision by FFBP, PNN and
PNN & PDP method based relays
FFBP neural
network
PNN
PNN and PDP
method
Number of samples
required
Number of samples
required
Number of samples
required
Cases
Actual Logical Actual Logical Actual Logical
Magnetizing inrush
(0
0
)
11 12 05 12 05 12
Internal fault (Light phase-
to-ground fault at 2%)
10 12 06 12 18 24

6. Conclusions

This paper presents a novel approach, by amalgamation of PNN and PDP methods, to enhance discrimination between
transformer internal fault and magnetizing inrush and to classify the type of fault in power transformer. The PNN is faster than the
classical ANNs and easy to design. The reported PNN algorithm is based on wave-shape identification technique which is
independent of amount of harmonic contents of operating signal of relay, and suitable for modern power transformers that use
high-permeability low coercion core materials. With these core materials high second harmonic components may be generated
during internal faults while these components may remain low during magnetizing inrush, and hence the conventional harmonic
restraint technique that uses second harmonic component as the indicator of magnetizing inrush may fail. As the PDP method
monitors the power flow into transformer irrespective of the harmonic contents of fault currents, and hence is suitable for
protection of modern power transformers. In the proposed hybrid method, stability of differential relay is ensured during the
magnetizing inrush, sympathetic inrush, over-excitation and external fault conditions. The combination of PDP method with PNN
makes it capable to detect light internal faults (turn-to-turn faults) for all ratings of transformers which improve the overall
performance of digital differential protection scheme. In addition to that the hybrid method is intelligent, reliable and capable to
take decision even in case of fluctuating ratio of second harmonic to fundamental of differential current unlike the conventional
harmonic method. Real time implementation of differential relaying using the proposed algorithm applying PNN as the core
classifier would essentially require a PNN processor and it is matter of further research.


42

Nomenclature
x Input of neural network
y Output of neural network
| |
i
S Cardinality of the set patterns in class
i
S
( ) /
r i
P x S Class conditional probability density function of x
( )
r i
P S Probability that a vector belongs to class
i
S regardless of the identity of that vector
d
i
j
Distance between the j
th
exemplar pattern and nearest exemplar pattern in class i
g
Multiplication factor (Constant)
i
Smoothing factor
i
j
x

j
th
exemplar pattern from class
i
S
1 2
, u u Instantaneous voltage at primary and secondary winding terminals respectively
1 2
, i i Instantaneous current at primary and secondary winding terminals respectively
r
1
, r
2
Primary and secondary winding resistances respectively
) ( t p
Instantaneous power
( ) w t Average power of one period time
N Number of samples per cycle

Appendix-1

Figure 18 shows a typical PSCAD/EMTDC transformer model to simulate internal faults (turn-to-turn, phase-to-ground, and
phase-to-phase) at different location of transformer winding from the neutral end of the windings. In this model MVA rating,
voltage rating, base frequency, leakage reactance, magnetizing current, and fault location (in %) etc. can be defined.

Figure18(a). Typical PSCAD/EMTDC transformer model to simulate internal fault


43

Figure 18(b). Typical PSCAD/EMTDC transformer model to simulate internal faults at different locations

Appendix-2

i. Calculate smoothing factor ( ) by using Euclidean distance between the points representing the training
patterns in feature space for each class i.e.
k
ij
d Euclidean distance between
k
i
x and
k
j
x in class k.
ii. Take average of Euclidean distances
iii. Select a constant g by trial and error method
Find optimal value of g based on the application.
References

Arboleya, P., Diaz, G., Aleixandre, J. G., and Moran, C. G. 2006. A solution to the dilemma inrush/fault in transformer differential
relaying using MRA and wavelets. Electric Power Systems and Research, Vol. 34, No. 3, pp. 285- 301.
Bastard, P., Meunier, M., and Regal, H. 1995. Neural network based algorithm for power transformer differential relays. IEE Proc.
Generation Transmission & Distribution, Vol. 142, No. 4, pp.386 -392.
Berthold, M. R., and Diamond, J. 1998. Constructive training of probabilistic neural networks. Elsevier Science, Neurocomputing,
Vol. 19, No. 1-3, pp.167-183.
Borghetti, A., Bosetti, M., Silvestro, D. M. and Nucci, A. C. 2008. Continuous-wavelet transform for fault location in distribution
power networks: definition of mother wavelets inferred from fault originated transients. IEEE Trans. on Power Delivery, Vol. 2,
No. 23, pp.380-388.
Bose, N. K., and Liang, P. 1996. Neural Network Fundamentals with Graphs, Algorithms, and Applications. McGraw-Hill Book
Co. International Editions.
Hammond, M. H., Riedel, C. J., Rose-Pehrsson, S. L., and Williams, F. W. 2004. Training set optimization methods for a
probabilistic neural network. Elsevier Science, Chemometrics and Intelligent Laboratory System, Vol.71, No.1, pp.73-78.
He, B., Zhang, X., and Bo, Z. Q. 2006. A new method to identify inrush current based on error estimation. IEEE Trans. on Power
Delivery, Vol. 21, No. 3, pp.1163-1168.
Moravej, Z., Vishwakarma, D. N. and Singh, S. P.: 2003. Application of radial basis function neural network for differential
relaying of a power transformer. Computer and Electrical Engineering, Vol. 29, No. 3, pp.421-434.
Musavi, M. T., Kalantri, K., and Ahmed, W. 1992. Improving the performance of probabilistic neural networks. Proc. IEEE Int.
Conf. Neural Networks, Vol. 1, pp.595-600.
Sachdev, M. S. (coordinator). 1988. Microprocessor relays and protection systems. IEEE Tutorial Course Text, (Publi.
No.88EH0269-1-PWR).
Shin, M. C., Park, C. W., and Kim, J. H. 2003. Fuzzy logic based relaying for large power transformer protection. IEEE Trans. on
Power Delivery, Vol. 18, No. 3, pp.718-724.
Sidhu, T. S., and Sachdev, M. S. 1992. On line identification of magnetizing inrush and internal faults in three phase transformers.
IEEE Trans. on Power Delivery, Vol.7, No.4, pp.1885-1891.

44

Specht, D. F. 1990. Probabilistic neural network. Neural Networks, Vol. 3, No.1, pp.190-118.
Specht, D. F., and Shapiro, P. D. 1991. Generalization accuracy of probabilistic neural networks compared with back-propagation
networks. Proc. IEEE Int. Joint Conf. Neural Networks, Seattle, pp. 887-892.
Specht, D. F.1992. Enhancements to probabilistic neural networks. Proc. IEEE Int. Conf. on Neural Networks, Vol. 1, pp.761-
767.
Tan, K. C., and Tang, H. J. 2004. New dynamical optimal learning for linear multilayer FNN. IEEE Trans. on Neural Networks,
Vol. 15, No.6, pp.1562-1568.
Tian, B., Azimi-Sadjadi, M. R. 2001. Comparison of two different PNN training approaches for satellite cloud data classification.
IEEE Trans. on Neural Networks, Vol. 12, No. 1, pp.164-168.
Tripathy, M., and Ala, S. 2009. Optimal radial basis function neural network transformer differential protection. Proc. IEEE Int.
Conf. on Innovative Ideas towards the Electrical Grids of the Futures, 28 June-2 July 2009, Bucharest, Romania, pp.1-8.
Tripathy, M., Maheshwari, R. P., and Verma, H. K. 2005. Advances in transform protection: a review. Electric Power Components
and Systems, Vol. 33, No. 11, pp.1203-1209.
Tripathy, M., Maheshwari, R. P., and Verma, H. K. 2007. Probabilistic neural network based protection of power transformer. IET
Electrical Power Application, Vol. 1, No. 5, pp.793-798.
Verma, H. K., and Basha, A. M. 1986. A microprocessor-based inrush restrained differential relay for transformer protection.
Journal of Microcomputer Application, Vol. 9, No.4, pp.313 318.
Wang, H., and Lin, X. 2009. Studies on the unusual mal-operation of transformer differential protection during the nonlinear load
switch-in. IEEE Trans. on Power Delivery, Vol. 4, No. 24, pp.1824-1831.
Woodford, D. 2001. Introduction to PSCAD V3. Manitoba HVDC Research Centre Inc., 400-1619 Pembina Highway, Winnipeg,
Manitoba, R3T 3Y6, Canada.
Yabe, K. 1997. Power differential method for discrimination between fault and magnetizing inrush current in transformers. IEEE
Trans. on Power Delivery, Vol.12, No.3, pp.1109-1118.

Biographical notes

Manoj Tripathy was born in Gorakhapur, India, in 1976. He received the B.E. degree in electrical engineering from Nagpur University,Nagpur, India, in 1999, the
M. Tech. degree in instrumentation and control from Aligarh Muslim University, Aligarh, India, in 2002, and the Ph.D. degree from the Indian Institute of
Technology Roorkee, Roorkee, India, in 2008. He was an Academic Staff Member with Shobhit University, Meerut, India. He is presently Assistant Professor with
the Department of Electrical Engineering, Motilal Nehru National Institute of Technology Allahabad, Allahabad, India. His research interests include power system
protection, developments in digital protective relay, and power system monitoring. Dr. Tripathy is a Reviewer for various international journals in the area of power
systems. He is a member of IE (India), and member of IEEE.

Rudra Prakash Maheshwari was born in Aligarh, India, in 1960. He received the B.E. and M.Sc. (Engg.) degrees in electrical engineering from Aligarh Muslim
University (AMU), Aligarh, India, in 1982 and 1985, respectively, and the Ph.D. degree from University of Roorkee, Roorkee, India, in 1996. He was an Academic
Staff Member with AMU. He is presently a Professor with the Department of Electrical Engineering, Indian Institute of Technology Roorkee, and a Consultant in
the area of small hydro power plants. He has published more than 75 research papers in various international/national journals and conferences. His research
interests include power system protection, developments in digital protective relay, and protective relay testing. Dr. Maheshwari is a member of editorial boards
and a reviewer for various international journals in the area of power system protection.

H. K. Verma was born in Gojra, India. He received the B.E. degree in electrical engineering from the University of Jodhpur, Jodhpur, India, in 1967, and the M.E.
degree in power systems engineering and the Ph.D. degree in electrical engineering from the University of Roorkee, Roorkee, India, in 1969 and 1977,
respectively. Currently, he is a Professor and Deputy Director of the Indian Institute of Technology, Roorkee. From 1980 to 1982, he was a Manager (R&D) of
Universal Electrics Ltd., Faridabad (a Birla Group Public Ltd. Co.). His research interests are in the areas of Intelligent instrumentation, digital/ numerical relays,
and power system protection, monitoring, and control. He is associated with many Government projects of national importance. He has published a large number of
research papers in various international/national journals and conferences.

Received November 2009
Accepted March 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 45-59

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Intelligent Agents under Collaborative Control in Emerging Power Systems

H.F. Wedde
1
, S. Lehnhoff
1*
, C. Rehtanz
2
, O. Krause
2

1
Department of Computer Science, Technical University of Dortmund, GERMANY
2
Department of Electrical Engineering, Technical University of Dortmund, GERMANY
*
Corresponding Author: e-mail: sebastian.lehnhoff@tu-dortmund.de, Tel +49-231-755-2768, Fax+49-231-755-5323

Abstract

In the world of liberalized power markets traditional power management concepts have come to their limits. Optimal pricing
can no longer be achieved, e.g. for very short-time needs across grids. Power line overload and grid stability, increasingly
resulting in regional or even global black-outs, are at stake. With the highly desirable expansion of renewable energy production
these challenges are experienced in quite an amplified way: We argue that for this emergent technology the traditional top-down
and long-term power management is obsolete, due to the wide dispersion and high unpredictability of wind and solar-based
power facilities. In the DECENT0F
1
R&D initiative we developed a multi-level, bottom-up solution where autonomous
collaborative software agents negotiate available energy quantities and needs on behalf of consumer and producer groups (the
DEZENT algorithm). We operate within very short time intervals of assumedly constant demand and supply, in our case periods
of 0.5sec (switching delay for a light bulb). The solution has proven to be secure against a relevant variety of malicious attacks.
Within this time interval we are also able to manage the coordinated power distribution, and achieve grid stability. In this paper
the main contribution is to make the negotiation strategies themselves adaptive across periods: We derive the dynamic
distributed learning algorithm DECOLEARN from Reinforcement Learning principles for providing the agents with
collaborative intelligence and at the same time proving substantially superior to conventional (static) procedures. We report
briefly on our extensive comparative simulation experiments.

Keywords: Distributed Energy Management, Reinforcement Learning, Multi-agent Systems, Smart Grids

1. Novel Perspectives and Challenges through Regenerative Electric Energy

In the world of liberalized power markets traditional power management concepts have come to their limits. Optimal pricing can
no longer be achieved, e.g. for very short-time needs across grids. Avoiding power line overload and maintaining grid stability
(increasingly resulting in regional or even global black-outs) are at stake. With the both ecologically and economically desirable
oncome of renewable energy production these challenges exhibit quite a fundamental quality, making a strong plea for these
negotiation and management problem areas to be integrated into an adequate novel approach.

1.1 Distributed Agent Negotiation of Regenerative Power: In Southern Germany approx. 180 cities and townships are already able
to cover their needs through wind craft, solar power, or through small and mid-size block heat & power plants (BHPPs) driven by
seed oil. Even entire regions in Germany and Austria are close to completely support themselves with clean energy. There are
similar efforts and projects within the European Community and world-wide. This will cause a paradigm shift: The traditional
power production, distribution, and management are based on long-term negotiations and on a centralized top-down balancing of
needs and supply, entailing network stability problems (rather frequent black-outs) as much as high environmental and financial
costs for reserve facilities (1-min reserve and below) running permanently on a 20% efficiency level.

1
This work is part of the DECENT R&D initiative. It is partially funded by Deutsche Forschungsgemeinschaft (German Research Foundation,
DFG), under contracts WE 2816/4-1 and HA 937/32-1, and by E.ON (contract no. 4500014170).
Wedde et al./ International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 45-59

46

In turn, regenerative energy production is based on widely dispersed small or mid-size facilities. Balancing production and
consumption are organized bottom-up, coordinated between local sources, and stepwise up between ever larger regions. To a large
extent, producers are also consumers, and vice versa. In the best common interest of the users involved, and at the same time
respecting the responsibility of the local and regional customers (producers as much as consumers), these should plan for largely
covering their normal needs themselves yet count for balancing power from higher levels in (unexpected) peak times. The
management itself can only be done under decentralized control, due to the problem complexity which renders a centralized
management obsolete, even under support of very large computer systems. On the other hand, every single source is a potential
reserve facility (yielding a comparably high fault tolerance). As long as the needed power could be negotiated and distributed in
due time no dedicated reserve power stations need to be kept, contrary to the traditional practice. Since the availability of wind or
solar power is rather unpredictable balancing between such sources could be backed up by BHPPs, or by novel batteries featuring
extremely high capacities and very high power gradients while being charged or discharged (Burnett et al., 2006).

1.2 Negotiation Periods and Distributed Intelligence: Under the highly unpredictable fluctuation of solar and wind energy
transformation facilities and generally of local production and consumption , at the same time respecting the authority and
interests of local customers, our key objective is for the distributed negotiations to take place during time intervals of assumedly
constant need and supply. In fact, we have chosen the smallest reasonable size, namely 0.5 sec (the delay between turning a light
switch and the bulb being in full operation). This assumption is acceptable, due to the laxity of electric devices, actors, and
procedures. These intervals are termed periods, designed for negotiations of distributed agents. Even for software agents this
creates rather rigid real-time deadlines.
A human actor will set up, or modify, a negotiation strategy on his level of granularity, thus even a strategy which is meant to be
dynamic on this scale would appear mostly static in the view of the agents! Furthermore, even 1 min (120 periods!) is too short in
the human perspective for readjusting strategies. To take advantage of the speed of the agents and at the same time close the gap
between human actor and agent behavior we provide for the agents to intelligently revise their strategies, for their next period. In
the absence of global information or any safe prediction basis, we derived distributed learning strategies from distributed
Reinforcement Learning.
In each interval, our strategies are exponential bidding functions. While they formalize organic growth or decline they have been
found most suitable for fast convergence between producer offers and consumer bids. A few of the technical features of the
distributed negotiation algorithm (DEZENT) will be described in section 2. More details can e.g. be found in Wedde et al., 2008.

1.3 Previous and Related Work: There has been quite a variety of work on various aspects of Computational Intelligence research
in emergent power systems for many years. Detailed investigations are concerned with (optimal) pricing (Ahamed et al., 2002,
Anthony et al., 2001, Bakirtzis et al., 2006, Hommelberg et al., 2007, Nanduri et al., 2007, Takano et al., 1996, Tellidou et al.,
2006 and Zhou et al., 2003), either in day-ahead or spot market models, blackboard-based auction systems, or other global tools,
such as global data bases. Let alone the restriction to negotiations none of the solutions is really scalable to large systems
integrating widely dispersed regenerative sources. In turn, the customers in DECENT carry out distributed negotiations through
their agents, and neither look-ahead nor auctioning are needed (Wedde et al., 2008). Grid stability is addressed in Ernst et al.,
2002, Ernst et al., 2004, Hadidi et al., 2009, Pipattanasomporn et al., 2009 and Vlachogiannis et al., 2004 in terms of various
learning strategies where a central agency would be enabled to ensure (optimal) stability. This theme has been an open issue in
practice so far, and the proposed solutions are certainly not scalable. In DECENT we pursued a completely distributed approach
where stability checks are done online for each period of 0.5 sec, under distributed control. The authors in Massoud et al., 2005
make plea for distributed safety control/reconfiguration in a military case study, and in Potter et al., 2009 a plea for a very high-
level control facility is made, as a suggestion for later realistic research. Altogether the quoted work from the other authors is
restricted to singular operational problems while the solution quality certainly depends on its compatibility with all other relevant
aspects.
At least for integrating regenerative facilities into existing grids it is clear from 1.1 and 1.2 that a distributed intelligent solution
is to be developed which integrates all relevant operational aspects. The DECENT initiative is, to the best of our knowledge, the
only comprehensive research effort here.

1.4 Organization of the Paper: After a compact formal introduction into the DEZENT algorithm in section 2, our collaborative
(distributed) learning algorithm DECOLEARN is derived from adapted Reinforcement Learning strategies in section 3. In section
4, it will be broadly compared with a static algorithm (no adjustments for the next period, despite changing production and
consumption patterns) as well as with an ideal algorithm. In the concluding section we also hint to future work.


47

2. Distributed Agent Negotiations (DEZENT Algorithm)
2.1. The Model: From the ideas presented in section 1 we will now continue into more detailed model assumptions (see also figure
1A for a typical power grid structure):

1) The consumer needs can be covered within a regional grid (0.4 kV - 10 kV voltage levels), or between grids (110 kV)
through regenerative energy, with very few exceptions where a reserve capacity in the 380 kV network will be accessed
(see figure 1B). The grids are assumed to have no electrical failures, in our subsequent discussion.

Figure 1. Power Grid and Associated Agents

2) Balancing of needs and/or of excess power may take place on different voltage levels as well as across different balancing
groups on the same level (see figure 1). It will occur bottom-up from the 0.4 kV level (level 1), or between groups on the
10 kV (level 2), and 110 kV level (level 3), respectively.

3) Consumers are normally also producers, and vice versa.

4) Negotiations will be carried out through customer agents (see figure 1B) representing the human or technical actors.
While the agents act autonomously their actions are, on each level, coordinated by agents called balance group managers
(BGMs). These operate in parallel on each grid or inter-grid level. As an example, the grid in figure 1B contains 3 such
balancing levels.

5) Negotiations are carried out within periods of 0.5 sec duration. Just during such intervals the demand and supply
situations may be considered constant, a key feature of the approach. Thus such time periods while enabling stable
negotiations constitute a hard deadline for all agents involved.

6) At the beginning of a period each customer agent checks its own demand and supply situation, and thus determines
whether it will act as a producer (excessive power available), a consumer (additional power needed) or take no action
(balanced situation).

7) Price bids and offers are limited by price frames reflecting the amortization of customer investment, maintenance, and of
(low) bio-fuel costs. There are no long-term contracts, thus no discounts for large quantities (to be purchased in the form
of future claims). Energy quantities are always sold and purchased for the duration of the next negotiation period (0.5 sec)
only.

2.2. The Base DEZENT Algorithm: Within this model framework the main idea for setting up the distributed negotiation algorithm
is as follows:

B
110 kV
10 kV
380 kV
Agent
0.4 kV
Negotiation
Producer
Main
reserve
Producer
Producer
BGM BGM
Producer
BGM
BGM
Consumer
Consumer
Consumer
wind turbine
fuel cell
110 kV
10 kV
bus bar 10 kV
disconnect
0.4 kV
block heat and power plant
photovoltaics
transformer
photovoltaics
A

48

1) If [A
k
, B
k
] is the price frame for level k (1 k 3 in figure 1B) each BGM on this level runs a coordination cycle of 10
rounds. Each round takes 1 msec.

2) After each round the BGMs check whether or not there are bids and offers similar enough to be matched (this is done
based on a global similarity parameter), and in these cases it settles contracts between the parties. The basic idea for the
BGMs is to only grant a (globally) fixed amount of X Wh (Watt hours) at a time, in a Round-Robin fashion (in Electrical
Engineering, energy can be partitioned into arbitrary portions).

3) Negotiation strategies as set by a customer agent C are characterized through an opening bid bid
C
(0) [A
k
, (B
k
+ A
k
)],
an opening offer offer
P
(0) [(B
k
+ A
k
), B
k
], a device-specific urgency urg
0
and strategy parameters s
1C
and t
1P
.
Furthermore, after round n; n [0,9] the unsatisfied agents adjust their bids/offers this will be done according to:

0
2
1
1
( )
C
C
C k urg n
s
s
bid n B
e

+
= +

(1)
0
2
1
1
( )
P
P
P k urg n
t
t
offer n A
e

+
= +

(2)
The s
2C
and t
2P
are determined by the opening bid (bid
C
(0) = bid
0
) or offer (offer
P
(0) = offer
0
), respectively:

2 0
log( )
C k
s B bid =

(3)
2 0
log( )
P k
t offer A =

(4)
The exponential behavior is most suited for fast convergence of bids and offers. Figure 2 gives a pictorial impression of
the negotiations between 6 consumers (ascending curves) and 5 producers (descending curves). Encircled bid/offer pairs
(of similar values) and numbers correspond to the order in which contracts are closed (the similarity range in figure 2 is
set to 2 ). On contracting either the consumer curve ends (contract 2), due to the needed quantities being smaller than
offers, or the producer curve ends (contracts 3 and 4), due to offers being smaller than the needed quantities. Both curves
end when needed and offered quantities match exactly (contracts 1, 5 and 6). In this example two consumers remain
unsatisfied by the end of the tenth round.

4) Unsatisfied customers in a cycle on level k are moved to the next-higher level k+1. The prices frames [A
k+1
, B
k+1
] are the
result of shrinking [A
k
, B
k
] by a fixed shrinking rate Sr according to (shrinking rates of 20% or 40% have been
alternatively selected in figure 3):

0
: ( )
k
A A c k = +
(5)
0
: ( )
k
B B c k =
(6)
with
0 0
( ) :
2
B A
c k Sr k
=
(7)
Initial offers and bids will potentially be adjusted to fit into the shrunken frame [A
k+1
, B
k+1
]. The other strategy parameters
remain unchanged. This creates a better chance for bids and offers to match. However, at the same time contracted prices
are likely to be more unfavorable than on the lower level, for all parties.


49

5) When a consumer C
k
and a producer P
k
are contracted the resulting negotiated prices for the consumer (price
Ck
) and for
the producer (price
Pk
) are calculated according to:

( ) ( )
: ( )
2
k k
k
P C
C
offer n bid n
price c k
+
= +
(8)
( ) ( )
: ( )
2
k k
k
P C
P
offer n bid n
price c k
+
=
(9)
Prices are calculated as arithmetic means with additional charges of c(k) (see formula (7)) being added to consumer prices
and subtracted from producer fees, respectively. These surcharges guarantee that the most favorable energy prices (for
consumers as well as producers) may be negotiated only on the lowest level and thus between regionally close customers.

6) Customers who are still unsatisfied after passing all grid levels (cycles) are directed to the main reserve facility. This is
highly unfavorable for their business as can even be seen in the modest pricing scheme in figure 3.

Figure 2. Contracting for Energy Quantities

For each period, and for high numbers of customers, the deadlines previously mentioned for rounds and cycles are all met (this
has been verified in extensive simulation experiments, on the high-performance Linux cluster LiDO at the Technical University of
Dortmund, Germany). Since in our example as much as in the real system (see section 2.1) negotiations will be finalized within
40 msec this leaves more than 450 msec of the period for communication and configuring the electric power for distribution
according to the negotiation results (and respecting network stability at the same time: This latter issue will be addressed in the
conclusion since it is out of the scope of this paper). Also, the prices for electric energy can be kept considerably lower than under
the traditional contracting between consumers and large power companies (the example in figure 3 gives a first idea. Regarding the
negotiation levels please compare also figure 1B).
Also, the algorithm is robust against a large class of security attacks. Due to space limitations we refer the reader to Wedde et
al., 2008 for further details.

3. Periodic Reinforcement Learning in DECENT
As explained in section 2.2 negotiations are organized in cycles, and the strategies within a cycle as well as strategy adjustments
(of the negotiation frames) between cycles are automated, for every period. Typically only 3-4 cycles (of 30-40 msec total
duration) are needed for finalizing the negotiation process, resulting in covering the consumer needs with regenerative power,
without accessing traditional (reserve) power sources (see figure3).
Between periods, i.e. every 0.5 sec (see section 2.2.1), a different form of adaptation has been established. As announced in
section 1, and in face of lacking both global information and predictive insights, we have assumed a Reinforcement Learning
approach Sutton et al., 1998 which has been adapted to the particular negotiation structures discussed in the previous section.
negotiation round
price

50

Figure 3. Exemplary Negotiation Frames and Adjustment

3.1. Reinforcement Learning: Reinforcement Learning is a computational approach for understanding and automating goal-
directed learning and decision making. It focuses on individual learning from direct interaction with the individuals environment.
Please recall from section 1 that Reinforcement Learning is different from supervised learning, the kind of learning studied in most
forms of machine learning, statistical pattern recognition, and artificial neural networks. In contrast, for interactive problems it is
often impossible to obtain examples of desired behavior that are both correct and representative of all the situations in which the
agent has to act. In an unknown environment, where one would expect learning to be most beneficial, an agent must be able to
learn from its own experience only.

Figure 4. The Agent-Environment Interaction in Reinforcement Learning

In a Reinforcement Learning problem an agent (the learner and decision maker) interacts continuously with an environment that
comprises everything outside the agent. Such an agent is faced repeatedly with a choice among different actions and corresponding
responses by the environment which present new situations to the agent. After each action a numerical reward is received
depending directly on the action taken. The agents goal is to maximize the amount of accumulated reward in the long run (see
figure 4, taken from Sutton et al., 1998). More specifically, agents interact with the environment at discrete time steps. At every
time step an agent evaluates some representation of the environments state s
t
. Based on that state it selects an action a
t
. One time
step later the agent receives a numerical reward r
t
(probably a direct consequence of its action) and finds itself in a new state s
t+1
.
Based on this new situation the agent selects a new action a
t+1
and so on and so forth.
One of the challenges that arise in Reinforcement Learning and not in other kinds of learning is the trade-off between
exploration and exploitation when selecting new actions. To obtain a high reward, an agent will prefer actions that it has tried in
the past, and found to be effective in producing reward. But for discovering favorable actions, it will equally explore actions that it
has not selected before (in order to be prepared, or educated, for changing situations to come). So the agent will exploit the past
in order to obtain a reward, but it also will explore potential strategies in order to receive a high award in the future. The dilemma
Level 1
Level 2
Level 3
Level 4
frame
adjustment by
20 % 40 %
negotiation
frame
8.8-17.2 10.2-15
Negotiation
Main reserve
capacity
Balancing Group
Manager
Balancing Group
Manager

Balancing Group
Manager

Balancing Group
Manager
Agent
producers refund: 5
consumer price: 21
negotiation frame: 6-20
frame
adjustment by
20 % 40 %
negotiation
frame
7.4-18.6 8.8-17.2

51

is that neither exploitation nor exploration can be pursued exclusively without failing at the task at hand (Sutton et al., 1998). It is
in a combination of both attitudes that an agent aims at performing best actions. In a stochastic way, each action will have to be
tried many times to gain a reliable estimate of its expected reward. Exploitation has been found to be the right thing for
maximizing the expected reward within the next period, but exploration may produce the greater total reward in the long run
(Sutton et al., 1998).

3.2. The Collaborative Learning Model in DECENT: Reinforcement Learning in DECENT has to be applied by each consumer
and producer based on local information only, since no agent has global information. In DECENT the only information an agent is
continuously faced with is its demand (or supply) and the negotiated energy price that is valid for the duration of next period (see
section 2.1.7). The negotiation scheme developed in section 2.2 guarantees that best energy prices may only be negotiated within
balancing groups on the lowest level, thus between regionally close customers (see formulas (7) through (9)). Hence, high or low
energy prices, respectively, are a direct indication of contracts that have been closed on higher levels, due to inappropriate
negotiation strategies (opening offers or bids and gradients of the negotiation curves, see formulas (1) through (4)). Thus, price
feedbacks can be utilized by the agents to learn from and adequately adjust their strategies to changing supply situations in
DECENT.
The negotiation functions defined in section 2.2 through equations (1) through (4), are characterized each by a tuple of the form
(s
1Ci
, bid
Ci
(0)) or (t
1Pj
,offer
Pj
(0)), for consumers C
i
or producers P
j
, respectively. The parameters urg
0
, A
k
, B
k
as well as the shrinking
factor are constants (see formulas (3) and (4) in section 2.2). The strategy parameters s
1Ci
and t
1Pj
are to be selected each from a
finite sequence of equidistant values, so are bid
Ci
(0) and offer
Pj
(0). The intervals are determined so as to avoid excessive
bargaining strategies under which contracts very likely will be settled on the highest level (reserve capacity) which is unfortunate
for all parties involved. (We omit the technical details here. They are part of a Ph.D. thesis and will be subject to a separate ICT
publication.) These 2-tuples uniquely define a negotiation strategy for C
i
and

P
j
, respectively. Each strategy is meant to be operated
for a period, and it will be revised it after each period.
Let the set of all feasible s
1Ci
be S
Ci
, the set of all feasible t
1Pj
be T
Pj
. Having in mind that A
1
bid
Ci
(0) A
1
+ (A
1
+ B
1
) and A
1

+ (A
1
+ B
1
) offer
Pj
(0) B
1
let us denote the set of feasible bids and offers for C
i
and P
j
by O
Ci
and O
Pj
, respectively. Then S
Ci
O
Ci
and T
Pj
O
Pj
are called the strategy space for C
i
and P
j
, respectively. Hence, the strategy space of an agent is the finite set of
strategies, the agent chooses from at the beginning of each negotiation period.

Figure 5. Varying Consumer Energy Prices (in ) throughout the Strategy Space

Figure 5 shows the consumer energy prices for varying strategies within a consumers strategy space. The price relief depicted
here was constructed during a balanced supply situation. In this scenario 25 producers and 24 consumers are negotiating with
constant energy demands and static negotiation strategies (normally distributed opening bids and offers and strategy parameters s
1

and t
1
, respectively). In this scenario a test consumer sequentially probes every strategy within its strategy space. Figure 5 depicts
the resulting energy prices plotted against opening bids and values of parameter s
1
.
However, in extensive experiments we have found that it is not reasonable for agents to regard the entire strategy space when
exploring new and possibly better strategies. On the one hand the strategy space contains extreme strategies at its borders (extreme
opening bids or offers with very steep or flat curves, respectively) that may only be successful if the system itself exhibits extreme
demand or supply situations (e.g. critical undersupply or extreme surpluses of generated renewable energy). On the other hand it
turned out that a multi-agent system of individual consumers and producers under distributed control and acting on local
information alone does not arbitrarily changes its states. DECENT exhibits a certain inertia in its reaction to changes in supply
situations: Successful strategies for unknown supply situations are likely to be close to the last known best strategy.

52

Taking the above phenomena into account we will introduce a neighborhood relation for the strategy spaces for C
i
and P
j
,
respectively, as follows:

Two consumer strategies strategy
Ci
and strategy
Ci
are k-neighbors iff (for a given parameter k):
1 1
i i
C C
s s k
and
(0) (0)
i i
C C
bid bid k
. (10)
In the same way two producer strategies strategy
Pj
and strategy
Pj
are k-neighbors iff:
1 1
j j
P P
t t k
and
(0) (0)
j j
P P
offer offer k
. (11)
Hence, k-neighborhoods are designed as finite subsets of the strategy spaces S
Ci
O
Ci
and T
Pj
O
Pj
, respectively. An energy
price r achieved in a period is computed as the sum of the prices achieved for every quantum, including the price possibly paid, or
received, at the reserve facility. In this way, after each period an agent is faced with a total energy price under the chosen strategy.
Consumer agents try to keep their energy costs low while producer agents try to get reimbursed their investment and maintenance
costs.
Strategic adaptations on behalf of C
i
or P
j
will be based on normalized energy prices. Formally speaking prices or
negotiated within period t will be normalized according to:

1
1 1
( )
( )
i
i
C
C
r t A
r t
B A
for consumers, and (12)

1
1 1
( )
( ) 1
j
j
P
P
r t A
r t
B A
for producers. (13)

3.2.1. Strategic Preference: In the DEZENT algorithm the individual consumption and production vary unpredictably (please
remember from section 2.1 that it does not only depend on a customer agent whether or not he will act as a consumer, producer, or
not act at all, during a period). Still customer agents want to keep their prices on a favorable level through the upcoming periods.
This is pursued based on a judgment about the quality of strategy a selected for the current period t, in relation to the negotiated
price r. For this purpose the strategic preference for the next period will be determined:

Definition: Let, for period t, and be the normalized prices negotiated under strategy a. The strategic preference p(t,
a)is then defined for 2 different cases:

1) (regular strategic preference)
(0, ) : 0
( 1, ) : ( , ) ( '( ) ( , )); 0 1
p a
p t a p t a r t p t a
=
+ = + <
(14)

The step-size parameter determines how recent rewards are weighted against long-past rewards. In highly dynamic
systems it makes sense to weight recent rewards more heavily than long-past ones, thus a higher value is deemed appropriate. In
any case the step-size parameter will need to be adjusted specifically to the application at hand. Under certain circumstances is
inappropriate to calculate the strategic preference as a weighted average over recent rewards. If the last known reward is far
outdated it is not reasonable to calculate the strategic preference as the weighted average from those values and a most-recent
reward (they may reflect completely opposing supply situations). Under such circumstances the strategic preference is
set to the most recent reward:

2) (exceptional strategic preference)
(0, ) : 0
( 1, ) : '( )
p a
p t a r t
=
+ =
(15)


53

Over the periods, the strategies will be dynamically sorted by each agent according to their strategic preferences. During period
t, the currently best strategy (to be selected for period t +1) is the strategy with the highest strategic preference value in the
(dynamic) order.

There are 3 different modes for selecting the strategy for period t +1: Exploitation, Explore
1
and Explore
2
:

1) Exploitation selects the currently best strategy.
2) Explore
1
randomly picks a strategy which is in the k-neighborhood of the currently best strategy.
3) Explore
2
randomly picks any strategy.

3.3. The Collaborative Learning Algorithm in DECENT (DECOLEARN): Every customer agent executes the following algorithm:

1) Initially the strategies will be chosen randomly according to Explore
2
(or may be prescribed by the customers).
2) For selecting the strategy for period t +1; t 1 the selection modes are given the following probabilities:
P(Explore
1
) =
1
, P(Explore
2
) =
2
, P(Exploitation) = 1-(
1
+
2
) with 0
1
,
2
1 and 0 1-(
1
+
2
) 1. According to
these probabilities exactly one of the modes is randomly determined.
3) Through the determined mode, the strategy for period t +1 is selected, and will be executed.
4) If the mode was Exploitation or Explore
1
then the regular strategic preference for the next period will be computed (see
formula (14)). In case of Explore
2
the exceptional strategic preference will be used (see formula (15)).

For an example we refer the reader to our simulation studies (see table 1).

Figure 6. The Agent-Environment (DECENT) Interaction in DECOLEARN

Figure 6 depicts the adjusted Reinforcement Learning model in DECOLEARN (please compare it to figure 4). The only way for
an agent in DECENT to interact with its environment and to generate a reward feedback is by bidding on energy quantities and
evaluating the negotiated energy price. The reward is used to adjust the strategic preference of the appropriate action. Based on the
adjusted preferences a new strategy for the following period is chosen.
Figures 7.a through 7.d depict the dynamic adjustment of preferences for best strategies in a producers strategy space. In this
example the situation shifts from a balanced system-wide demand and supply to a shortage of energy generation on the lower
negotiation levels (see figures 1 and 3). The producer starts from a normally distributed sampling of strategies within his strategy
space (in a balanced demand and supply situation, see figure 7.a). During the undersupply producers learn to favor negotiation
strategies with high opening offers and high strategy parameters t
1
(flat negotiation curves, see figures 7.b through 7.d).
Although the producer depicted in figure 7 favors strategies with highest opening offers, he does not at the same time choose
extreme values for his strategy parameter t
1
. The producer learns to omit this combination of highest opening offers in combination
with extremely flat negotiation curves. Even in undersupply situations, with these negotiation strategies he risks getting propagated
onto higher negotiation levels and ultimately satisfies his needs at the backup reserve facility, thus at lowest possible rates (see
figure 3). However, a producer may still choose these extreme negotiation strategies randomly in the course of an Exploration
2

action choice.


54

4. Simulation Studies
In order to evaluate the impact of distributed learning on negotiations in DECENT we compared the DECOLEARN algorithm
with the previously introduced algorithm for which the strategies remained unchanged across the periods. We term it here the
STATIC algorithm.
In previous publications (Wedde et al., 2008) it has been demonstrated that a very large majority of customers got settled on
level 1 (see figures 1 and 3). Into the same vein it has been argued that early satisfaction within a period was beneficial for all
parties involved. In particular, the flexibility for selecting negotiation parameters was better: The intervals (A
k
, B
k
) are larger for
smaller k (see section 2.2) thus the exponential curves for producers and consumers would converge faster yet at less favorable
rates, for both of them. Our comparative studies were consequently based on 2 major criteria:

1) Minimize, if not avoid, the access to the main reserve capacity (see figure 1);
2) Allow the customers for contracting as early within a period as possible.

We conducted a series of experiments with 3000 agents and a layered agent negotiation structure exhibiting 3 levels (for the
experimental setup see table 1). The experiments were conducted on the high-performance LiDO cluster at the Technical
University of Dortmund, over 7200 periods (=1 hour) with realistic profiles for individual producers and consumers of renewable
energy. The resulting demand and supply pattern utilized in our experiments is depicted in figure 8. Every energy quantum not
successfully negotiated on these levels is sold to, or purchased at, a system-wide reserve facility.

Figure 7. Development of Strategic Preferences Values

Producers are modeled based on real world profiles derived from wind turbines and photovoltaic installations of various
capacities (50-300 W/h). Consumers are modeled based on empirical studies on average power requirements of households. (A
detailed description of the tools utilized in modeling realistic behavior of producers and consumers of renewable energy is beyond
die scope of this presentation yet this is subject of a separate publication).

a
d
c
b

55

Table 1. Experimental Setup
Negotiation Levels 3
Number of BGMs on Level 1 30
Number of Clients per BGM on Level 0 100
Number of Consumers (50 W/h average capacity) 2160
Number of Producers (50-300 W/h average capacity) 840
Simulation Time (negotiation periods) 7200
A
1
= 5 , B
1
= 20 , similarity = 1.2 ,
allowance = 50 W, shrinking rate = 20 %

energy refund = 4 /W, energy price = 21 /W (reserve facility)
1 1
, 0.5
i i
C P
s t =
, 1 1
, 50
i i
C P
s t =
= 0.3, k = 2,
1
= 0.45,
2
= 0.05

The demand and supply pattern shows an overall energy demand that fluctuates around 120 kW/h with an amplitude of
approximately 10 kW/h (see figure 8). The overall power production oscillates smoothly with a period of approximately 2000
seconds. In the course of the simulation the overall production rises from 70 kW/h to 140 kW/h. Thus, for the first 4300 periods
the overall energy demand slightly exceeds the overall energy production. After 4300 periods the overall energy demand and
supply situation improves as it features a production surplus throughout the remainder simulation time. A third curve at the bottom
of figure 8 resembles this imbalance in overall demand and supply.

Figure 8. Overall Demand and Supply Pattern

In order to evaluate the performance of distributed learning in DECENT, the overall renewable energy contracted on the 3
negotiation layers before reaching the system-wide reserve facility is plotted against the simulation time (see figure 9). The results
are compared against the maximum amount of energy that can be purchased under ideal negotiations (reflecting the global power
balance before imbalances are left with the reserve facility as indicated in figure 3). This ideal curve coincides with the
minimum of both the overall production and the overall demand (see figure 8). Imbalances in demand and supply beyond level 3
have to be regulated at the system-wide backup capacity (level 4, at 380 kV in figure 3).


56

Figure 9. Energy Purchased throughout the Multi-Level Negotiations in DECENT

In the course of these experiments the DECOLEARN algorithm has been evaluated against the distributed STATIC algorithm.
For this purpose additional experiments have been conducted on the basis of the setup in Table 1. Instead of dynamic strategy
adjustments after each period, strategies have been chosen at random once at the beginning of the experiment, and then kept
constant throughout the 7200 periods. Simulations have been repeated 50 times with different initial strategy settings. The
resulting averaged curve is depicted in figure 9.

Figure 10. Experiments with Various Configurations of k

With the chosen setup the negotiations based on static strategies are unable to adjust to varying supply situations and thus
perform significantly worse than ideally possible throughout all 7200 periods. This is illustrated in the gap between the STATIC
curve and the ideal curve that the algorithm is unable to close. This underperformance increases after approx. 4000 periods when
overall system dynamics increase. Distributed learning performs nearly ideally, and far better than STATIC, throughout the entire
simulation. After approximately 4000 periods the performance of DECOLEARN changes. The maximum amount of energy that
can be sold before reaching the backup facility fluctuates more rapidly as it has to adapt to the high variation of the demand curve
(see figure 8). So the performance of DECOLEARN decreases somewhat (although it remains still far superior to STATIC). A

57

reasonable interpretation may come from problems in finding correct or good adjustments of strategic preferences within the
agents strategy neighborhoods.
In order to indicate the influence of the size of parameter k (neighborhood width, see section 3.2) on the DECOLEARN
performance we repeated experiments with different values for k. The results between the periods 3000 and 7200 are depicted in
figure 10.
With growing k DECOLEARN exhibits problems while coping with the increasing fluctuation in the maximum amount of
energy that can be sold before reaching the backup facility. This supports the conjecture made in section 3.2 that distributed
learning in large neighborhoods is suffering from an increased uncertainty about good adjustments under more rapidly varying
supply and demand situations. It is not reasonable for agents to regard the entire strategy space when searching for better
strategies. Successful strategies for unknown supply situations are likely to be within a narrow k-neighborhood of the last known
best strategy.
Regarding the criterion of early satisfaction on the lowest negotiation level we assessed the impact of distributed learning on the
amount of energy negotiated on the lowest level. Figure 11 depicts the amount of energy sold and purchased on the first level for
both DECOLEARN and STATIC strategies. The positive DECOLEARN performance trend that was already observed in figure 9
results to a large extent from its stronger performance on the first negotiation level, throughout the whole simulation. We found
that negotiations on the second and third level performed alike for DECOLEARN and STATIC yet with a relevant advantage for
DECOLEARN, after 4000 periods.

Figure 11. Negotiated Energy on the Levels 1, 2 and 3

5. Conclusion
The general idea behind DECOLEARN as derived from Reinforcement Learning was that individually favorable prices should
be adaptively found, and at the same time such rates should be dynamically achieved during later periods as well.
Due to page limitations, we restricted ourselves to a representative report on our experimental evaluation. However, due to the
distributed control concept, the results do not depend on the number of agents involved: Even with 300.000 agents (simulated on
the HPC Cluster at the Technical University of Dortmund) we found the picture to be equally favorable.
In the model scenario there is a large amount of uncertainty on different levels, last but not least due the fact that for every agent,
be it a producer or consumer, it is unpredictable whether it will act as a producer or consumer agent, or remain inactive during the
next period. So at first glance the STATIC algorithm may well have its own merits as a very simple framework for handling
unpredictable negotiations dynamically. Compared to Explore
1
or Explore
2
as degenerative forms of DECOLEARN, i.e. for
P(Explore
1
)=
1
=1, P(Explore
2
)=
2
=0=P(Exploitation), or P(Explore
2
)=
2
=1, P(Explore
1
)=
1
=0=P(Exploitation), STATIC may be
expected to exhibit a similar behavior, depending on the amount of variation or customer unpredictability. In this way STATIC
served as a natural and simple candidate for the (comparative) evaluations.

58

We have argued above that a bottom-up management is most adequate for the widely dispersed sources of electric power. It
should also be very clear from the previous discussion that a distributed algorithmic solution and the corresponding computer
network support are indispensable for the purpose.
The results show very clearly the strong improvement that comes into the picture through the flexible and adaptive distributed
learning methods in DECOLEARN. With DECOLEARN we successfully introduced a highly dynamic concept for negotiations
across periods and, indeed, for providing the macro-customers (human or technical actors) with decent solutions at reasonable
costs (see figure 3 and Wedde et al., 2008), despite the unpredictability of all actions and influences.
In the DECENT initiative we have pursued a novel distributed approach for negotiating, distributing, and monitoring electric
energy in large power grids. While in the world of liberalized power markets traditional power management concepts have come to
their limits, considering both in achieving optimal prices, in avoiding power line overloads, and in maintaining grid stability
(absence of regional or global black-outs), the rapidly emerging regenerative power production while offering an ecologically
highly desirable future poses even stronger challenges on a traditional grid management, in all aspects discussed above. In this
situation we have defined a distributed control approach which combines, and completely integrates, measures and solutions in the
economic, ecologic, electric, and ICT aspects. Distributed agents take care of producer and consumer needs in a bottom-up
fashion, balancing at the same time the needs within balancing groups and beyond. Stability is taken care of through novel
distributed algorithms Krause et al., 2009, allowing for testing and achieving a stable distribution pattern for each period. Peak
demand and supply management have been successfully handled in a distributed procedure, an extension of the DEZENT
negotiation algorithm Wedde et al., 2008. Please recall that the DECENT bottom-up approach allows for a high amount of
parallelism which is the basis for the timely micro-operations of the agents while allowing getting along without look-ahead
strategies, spot market procedures etc.
Our experimental results so far reach to the 110 kV grid level, however, conceptually there are no limits to expand into larger
dimensions and grids interconnected by even higher voltage levels.

References

Ahamed T.P.I., Rao P.S.N. and Sastry P.S. 2002. A Reinforcement Learning Approach to Automatic Generation Control. Electric
Power Systems Research, Vol.63, No.1, pp.9-26.
Amin S.M. and Wollenberg B.F. 2005. Toward a Smart Grid: Power Delivery for the 21
st
Century. IEEE Power and Energy
Magazine, Vol.3, No.5, pp.34-41.
Anthony P., Hall W., Dang V. and Jennings N. 2001. Autonomous Agents for Participating in Multiple Online Auctions.
Proceedings of the 17
th
International Joint Conference on Artificial Intelligence (IJCAI01) Workshop on E-Business and the
Intelligent Web, Vol.1, No.1, pp.54-64.
Bakirtzis A.G. and Tellidou A.C. 2006. Agent-Based Simulation of Power Markets under Uniform and Pay-as-Bid Pricing Rules
using Reinforcement Learning. Proceedings of the Power Systems Conference and Exposition (PSCE '06), Vol.1, No.1, pp.1168-
1173.
Burnett M.B. and Borle L.J. 2006. A Power System Combining Batteries and Super-Capacitors in a Solar/Hydrogen Hybrid
Electric Vehicle. Proceedings of the 2006 IEEE Vehicle Power and Propulsion Conference, Vol.1, No.1, pp.709-715.
Ernst D. and Wehenkel L. 2002. FACTS Devices Controlled by Means of Reinforcement Learning Algorithms. Proceedings of the
14
th
Power Systems Computation Conference (PSCC 2002), Vol.1, No.1, pp.1-7.
Ernst D., Glavic M. and Wehenkel L. 2004. Power Systems Stability Control: Reinforcement Learning Framework. IEEE
Transactions on Power Systems, Vol.19, No.1, pp.427-435.
Greenwald A.and Kephart J.O. 1999. Shopbots and Pricebots. Proceedings of the 16
th
International Joint Conference on Artificial
Intelligence (IJCAI '99), Vol.1, No.1, pp.506-511.
Hadidi R. and Jeyasurya B. 2009. Near Optimal Control Policy for Controlling Power System Stabilizers Using Reinforcement
Learning. Proceedings of the 2009 IEEE Power & Energy Society General Meeting, Vol.1, No.1, pp.1-7.
Hommelberg M.P.F., Warmer C.J., Kamphuis I.G., Kok J.K. and Schaeffer G.J. 2007. Distributed Control Concepts Using Multi-
Agent Technology and Automatic Markets: An indispensable feature of smart power grids. Proceedings of the 2007 IEEE Power
Engineering Society General Meeting, Vol.1, No.1, pp.1-7.
Krause O., Lehnhoff S., Rehtanz C., Handschin E. and Wedde H.F. 2009. On Feasibility Boundaries of Electrical Power Grids in
Steady State. International Journal of Electric Power & Energy Systems, Vol.31, No.9, pp.437-444.
Nanduri V. and Das T.K. 2007. A Reinforcement Learning Model to Assess Market Power under Auction-based Energy Pricing.
IEEE Transactions on Power Systems, Vol.22, No.1, pp.85-95.
Potter C.W., Archambault A. and Westrick K. 2009. Building a Smarter Smart Grid through Better Renewable Energy
Information. Proceedings of the IEEE Power Systems Conference and Exposition, Vol.1, No.1, pp.1-5.
Pipattanasomporn M., Feroze H. and Rahman S. 2009. Multi-Agent Systems in a Distributed Smart Grid: Design and
Implementation. Proceedings of the IEEE Power Systems Conference and Exposition, Vol.1, No.1, pp.1-8.
Sutton R.S. and Barto A.G. 1998. Reinforcement Learning. MIT Press.
Takano T., Matsumoto K., Oki, I. and Ohashi T. 1996. A Case Learning Tool for Operations of Power Systems. Proceedings of the

59

International Conference on Intelligent Systems Applications to Power Systems, Vol.1, No.1, pp.91-96.
Tellidou A.C. and Bakirtzis A.G. 2006. Multi-Agent Reinforcement Learning for Strategic Bidding in Power Markets.
Proceedings of the 3
rd
International IEEE Conference on Intelligent Systems, Vol.1, No.1, pp.408-413.
Vlachogiannis G. and Hatziargyriou N.D. 2004. Reinforcement Learning for Reactive Power Control. IEEE Transactions
on Power Systems, Vol.19, No.3, pp.1317-1325.
Wedde H.F., Lehnhoff S., Rehtanz C. and Krause O. 2008. Bottom-Up Self-Organization of Unpredictable Demand and Supply
under Decentralized Power Management. Proceedings of the 2
nd
IEEE International Conference on Self-Adaptation and Self-
Organization, Vol.X, No.X, pp.74-83.
Wedde H.F., Lehnhoff S., Moritz K.M, Handschin E. and Krause O. 2008. Distributed Learning Strategies for Collaborative
Agents in Adaptive Decentralized Power Systems. Proceedings of the 15
th
IEEE International Conference on Engineering of
Computer-Based Systems, Vol.1, No.1, pp.26-35.
Ye H., Brocco A., Kuonen P., Courant M. and Hirsbrunner B. 2008. SmartGRID: A Fully Decentralized Grid Scheduling
Framework Supported by Swarm Intelligence. Proceedings of the 7
th
International IEEE Conference on Grid and
Cooperative Computing, Vol.1, No.1, pp.160-168.
Zhou M., Ren J., Li G. and Xu X. 2003. A Multi-Agent based Dispatching Ooperation Instructing System in Electric Power
Systems. Proceedings of the 2003 IEEE Power Engineering Society General Meeting, Vol.1, No.1, pp.436-440.

Biographical notes

Prof. H. F. Wedde received his diploma degree in Pure Mathematics, the Ph.D. degree in Computer Science (1974), all from the University of Bonn, Germany.
His appointments included a Senior Research Staff position at the Gesellschaft fr Mathematik und Datenverarbeitung (GMD) in Bonn, Germany (1969-83),
visiting professor positions at the universities of Pisa (1980) and Turin (1982), Italy, and at the Academy of Sciences in Warsaw, Poland (1977, 1980). Prof. Wedde
was on faculty of Wayne State University/ Detroit (1984 - 1993). From 1994 on, he has been with the School of Computer Science at the University of Dortmund,
Germany. His major research interests are in various areas of Distributed Computing Systems including Theory, Distributed Operating/ File Systems, Distributed
Real-Time Systems, Distributed Security, and Nature-Inspired Routing algorithms, the latter especially in the Bee Hive project. Recent efforts have been devoted to
such application areas as distributed (secure and real-time) power management of renewable energy, integrated production and transportation scheduling in
Distributed Manufacturing, or traffic jam avoidance.

Dr. S. Lehnhoff received his diploma degree in Computer Science in 2005 and his Ph.D. in 2009 from the Technical University of Dortmund, Germany. He is a
research associate with the Institute of Operating Systems and Computer Architecture at Technical University of Dortmund, Germany. His research interests
include but are not limited to smart grids, intelligent power management systems and distributed control algorithms for applications in collaborative self-organizing
systems.

Prof. C. Rehtanz received his diploma degree in Electrical Engineering in 1994 and his Ph.D. in 1997 at the University of Dortmund, Germany. From 2000 he was
with ABB Corporate Research, Switzerland and from 2003 Head of Technology for the global ABB business area Power Systems. From 2005 he was Director of
ABB Corporate Research in China. From 2007 he is professor and head of chair for power systems and power economics at the Technical University of Dortmund.
His research activities include technologies for network enhancement and congestion relief like stability assessment, wide-area monitoring, protection, and
coordinated FACTS- and HVDC-control.

Dr. O. Krause received his diploma degree in Electrical Engineering in 2005 and his Ph.D. in 2009 at the Technical University of Dortmund, Germany. He is a
research associate at the Institute of Electrical Power Systems and Power Economics, Technical University of Dortmund, Germany. His research interests are
strategies for the coordinated use of power systems and issues related to network stability.

Received December 2009
Accepted March 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 60-68

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Requirement analysis for autonomous systems and intelligent agents in
future Danish electric power systems

Arshad Saleem
*
and Morten Lind

Center for Electric Technology CET, Department of Electrical Engineering
Technical University of DENMARK
*
Corresponding Author: e-mail: asa@elektro.dtu.dk, Tel. +45-45253532, Fax.+45-45886111

Abstract
Denmark has already achieved a record of 20% penetration of wind power and now moving towards even higher targets with
an increasing part of the electricity produced by distributed generators (DGs). In this paper we report work from a sub activity
"subgrid design" of the EcoGrid.dk
1
project. First we review innovative control architectures in electric power systems such as
Microgrids, Virtual power plants and Cell based systems. We evaluate application of autonomous systems and intelligent agents
in each of these control architectures particularly in the context of Denmark's strategic energy plans. The second part formulates
a flexible control architecture for electric power systems with very high penetration of distributed generation. This control
architecture is based upon the requirements identified in the first part. We also present development of a software framework to
test such flexible control architectures.
Keywords: Electric power system, distributed control, autonomous systems, intelligent agents

1. Introduction

Electric power systems is one of the most critical and strategic infrastructures of industrial societies and is currently going
through a revolutionary change. Deregulation, security of supply, environmental concerns and rapid growth in Information and
Communication Technologies (ICT) are basic driving factors behind trends like renewables and distributed generation, free
market, decentralized control, self healing and automatic (dynamic) reconfiguration in today's' power systems. Deregulation aims
at ultimately providing a competitive market based environment which can offer energy to customers at a favorable price.
Environmental Concerns primarily deals with CO2 emissions and global warming. Security of supply ensures the smooth delivery
of energy to customers. Electric power industry has traditionally been slower to adopt modern ICT compared with other industries
e.g. telecommunication. But recently, a rapid growth and significant reduction in price of ICT tools and technologies has made it
inevitable for electric power industry to take initiatives to adopt ICT more effectively. This situation has created an incentive
among both distribution utilities as well as network operators to develop long term strategies and plans that address above
mentioned challenges.
Denmark has already achieved a record of 20% penetration of wind power (2007 data), and innovative control architectures like
microgrids, virtual power plants (VPP), cell based systems, vehicle to grid, are under consideration and development in electric
power system. Furthermore, Denmark is now moving towards a target set out by a new national energy strategy which implies
50% wind power penetration for the electric power system by 2025.
With the valuable experiences accumulated so far, it has been anticipated that the technical challenges for a 2025 scenario of
50% wind penetration will be in balancing the power system, development of new market services and the need for new strategies
for operational security and control. Accomplishing these challenges require that future power systems are of a distributed nature,
consisting of autonomous components or subsystems, are able to coordinate, communicate, compete and adapt to emerging

1
EcoGrid project: http://www.ecogrid.dk/

Saleem and Lind / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 60-68

61

situations and self heal themselves. Efforts are underway to meet these challenges. EcoGrid.dk is a principal research program on
the national level initiated by the Danish TSO (Transmission System Operator) Energinet.dk, and comprised a consortium of stake
holders including power distribution and generation companies, manufacturers, consultants and research institutes both in and
outside Denmark, to propose a strategy for transforming the Danish power system into the world best renewable based electricity
network. In this paper we report work from a sub activity "subgrid design" of the Ecogrid.dk project (phase 1). This activity is part
of the work package II dealing with System Architecture.
The paper is composed of three main parts. In the first part we make a review of innovative control architectures in electric
power systems such as microgrids, virtual power plants and cell based systems. In the second part we evaluate the approach of
using autonomous systems and intelligent agents in each of these control architectures particularly in the context of Denmark's
strategic energy plans. We identify specific requirements and respective capabilities of intelligent agents for each of these control
architectures. In the third part we formulate a generic and flexible subgrid control architecture for the future Danish electric power
system. We base this architecture on the requirements identified in the second part and suggest that power systems with high
penetration of distributed generation may be controlled as a loose aggregation of energy units -- the subgrid control. Finally we
present a multiagent software platform for design and test of the subgrid control concept which is currently under development.
The structure of the paper is as follows: section II presents the review of innovative control architectures in the context of the
Danish Power System and a requirement analysis for the use of intelligent agents and autonomous systems. Section III presents the
generic sub-grid based control architecture and section IV presents our work on developing a software framework for designing
and testing the flexible control strategies mentioned in section III. Section V contains conclusions of the work.

2. Review of Innovative Control Architectures

The objective of this section is to review existing Danish and international proposals for future systems structures e.g.
microgrids, cell based systems and virtual power plants and perform a requirement analysis for the use of intelligent agents and
autonomous systems in these architectures. One purpose of this task within the EcoGrid.dk project was to provide input to other
tasks in the project and to provide inputs to the electric power industry for their future planning and application of intelligent
control technologies.

2.1 Intelligent Agents and Multiagent Systems:
An agent is an encapsulated computer system that is situated in some environment and can act flexibly and autonomously in that
environment to meet its design objectives (Wooldridge 2002).

To further explain:

Agents are encapsulated in a belief, desire, intention BDI metaphor
Agents are situated in physical or software environments
Agents are autonomous and can exercise control over their state and behavior
Agents are proactive and can take initiatives by themselves rather than passively responding to changes in their
environment
Agents are social and can communicate in high level dialogues

This definition of agents and the metaphore presented in Figure 1 is analogous to the notion of engineering autonomous systems
which can react intelligently and flexibly on changing operating conditions and demands from the surrounding processes (Rehtanz
2003). Such intelligent and autonomous systems provide capabilities like decomposition, reasoning, dynamic, flexibility (dynamic
reconfiguration) and cooperation modeling. The remaining part of this section indicates that such capabilities are critical for
realization of innovative control architectures in electric power systems.
It has been suggested to use intelligent agents and multiagent systems from a requirements rather than a technological
perspective. Intelligent agents are considered appropriate for applications that are modular, decentralized, changeable, ill-
structured, and complex (Wernstedt & Davidsson 2002) because multiagent architecture directly supports design and development
of systems with such features.
We investigate innovative control architectures in electric power systems and argue that they have specific characteristics which
makes them candidates for application of intelligent agents. We base our analysis primarily on the three most noteworthy and
emerging concepts: microgrids, virtual power plants and cell based systems. It is important to note that the terminology cell based
systems has been used in very general meaning in electric power system literature. For our study we refer to a concept applied by
the cell project
2
of Danish TSO Energinet.dk, which falls within the general notion of cell based energy systems.

2
The Danish Cell Project: www.energinet.dk

62

Figure 1. Belief, Desire, Intention BDI based abstract intelligent agent metaphor

2.2 Microgrids:
Microgrids are small electrical distribution systems that connect multiple customers to multiple distributed sources of generation
and storage. A microgrid can be connected to the main power network or be operated autonomously, similar to an island operation.
Sources in a microgrid are usually small (< 100 kw) units with power electronic interfaces (Hatziargyriou 2008). Microgrids can
be characterized by following specifications based upon current implementation and test efforts:

Different resources in a microgrid may be owned by different owners, thus making environments where the components
have competing goals, while still serving the common goal of ensuring security of supply
Resources in a microgrid are of a heterogeneous nature
In a grid connected mode, all the resources in microgrid participate in the market, whereas in islanded mode microgrid
have its own market managed by local microgrid controller
A microgrid has to be able to shift to (and from) islanded mode to grid-connected mode

Considering this characterization, intelligent software agents have a very obvious case for application in microgrid control and
operation. Intelligent agents have sophisticated cooperation mechanism (Jennings 1995; Liu et al. 2006) which could help
implement the required cooperation between different DGs in a microgrid. The adaptive nature of intelligent agents (Bernon et al.
2003) can help to implement dynamic adaptability in microgrids e.g., from grid-connected to islanding mode. The autonomous
nature of agents is vital for the local and distributed control of microgrids and to facilitate the economic interest of individual DG
owners. The ability of agents to coordinate, communicate and resolve issues can help to implement market mechanism. Also
agents have the capability to strike a balance between individual and collective system goals which is relevant for implementing
the above scenario of maintaining a balance between individual economic goal of DG owners and overall system goal of security
of supply and balancing.
There are several efforts and demonstrations for utilizing intelligent agents and autonomous systems for design and control of
microgrids (Dimeas 2004; Farred 2008).

2.3 Virtual Power Plants:
Unlike a microgrid which is a geographical aggregation, a virtual power plant is a commercial aggregation of distributed energy
resources. VPP is a group of DG units which are controlled to function together like a single power plant, thereby improving the
market function and providing valuable flexibility to the power system. Through the VPP concept, individual DGs will be able to
gain access and visibility across the energy markets, and benefit from VPP market intelligence in order to get their revenue

63

maximized, which otherwise would have not been able to participate in the market due to their small size or stochastic energy
supply. Virtual power plants can be characterized further as following:

Same as microgrids, different components in a VPP may be owned by different owners with competing goals/interests
The resources in a VPP may be geographically dispersed at long distances, making the communication requirements more
significant
All nodes participate in the market, individually in the local market and in the main grid through a central facilitator
(controller) of VPP
Nodes of a VPP should have enough level of intelligence to take local decisions and perform well in the market

Besides the requirement of autonomy, local control capability, high level communication and decision making as in microgrids,
it is apparent that VPP has a higher requirement for (real time) market mechanisms. Current research has shown that economic
feasibility of VPP is very much dependent on good local market mechanisms, ability of the DGs inside a VPP to respond
intelligently to price signals, and negotiation patterns of DGs for participating in the market (Shi 2009). Intelligent agents and
other reasoning mechanism has been used widely in different market processes e.g., stock exchanges and has great potential to be
used in VPP particularly enabling for individuals DGs to participate effectively in the market.

2.4 Cell based systems:
A cell is defined as the portion of a distribution system below 150/60 KV sub station typically consisting of 20-100 MW of
conventional loads and a mix of CHP and wind turbine generators (Lund 2007; Cherian & Knazkins 2007).
A cell has two main operational modes:

In the case of a cell-area operating in parallel with the HV-grid, the main focus of control is a fully automated VPP
operation on existing and the future market conditions. It should also be possible for TSO to interact with cells from a
control center just as they would interact with a conventional power plant

In emergency situations in a 400 KV grid, it performs intentional islanding i.e., on receiving a signal from the control
center it quickly manages a balance between generation and load within the cell and restarts itself using local resources (<
60 KV)
In the case of a cell-area operating in parallel with the HV-grid, the focus of control can be fully automated VPP operation on
existing and the future market conditions. It should also be possible for TSO to interact with cells from a control center just as they
would interact with a conventional power plant. In emergency situations in a 400 KV grid, it performs intentional islanding i.e., on
receiving a signal from the control center it quickly manages a balance between generation and load within the cell and restarts
itself using local resources (< 60 KV)
In the first case of VPP operation the control is commercially motivated and can incorporate both active and reactive power at
the same time. Furthermore the controller can be asked to do voltage control in a certain grid point utilizing all available generators
under its control. During the second case of emergency islanding situations, islanding occurs and control is transferred to local
cells. In such cases each cell behaves like a microgrid (or island) which is not connected to the main network. When cells are
islanded (in emergency situations) they are disconnected from the main grid and are no longer able to participate in the normal
market.
It is clear that cell operation can either be commercially motivated, the first case of VPP operation where more centralized
approach and competitive market mechanisms are required, or it can be technically motivated, the second case of emergency
islanding where a decentralized (or local) control with cooperative behavior of individual resources is anticipated.
Cell based systems have a high requirement for situation awareness and being able to reconfigure in a changed situation e.g.,
when going to cell mode. The behavior of cell components has to be different during islanding and VPP operation modes. Thus it
is anticipated that characteristics of intelligent agent technology like knowledge representation, situation reasoning and cooperative
decision models can facilitate realization of this concept. Aspects of agent technology are already under testing in the cell project
demonstrations (Lund 2007; Cherian & Knazkins 2007). Because of the requirement for more centralized nature of control during
VPP operation and more decentralized one during islanding, it should be interesting to consider heterarchical control strategies
where different modes of control are implemented at different control levels.
Table 1 summarizes a mapping of intelligent capabilities into innovative power systems control requirements. It shows the
important characteristics of intelligent based systems which have a potential for application in innovative control architectures.
The application cases have been analyzed with two levels of significance: high requirement and low requirement.


64

Table 1. Mapping of intelligent system capabilities into innovative power systems control requirements

3. Subgrid Architecture

As the second part of this work, a generic flexible control architecture was envisioned for future scenarios where the electric
power system is a loose aggregation of units that could be a microgrid, virtual power plant or cell like structure. These subgrids not
only have to optimally perform local control within the subgrid, but also must comply with responsibilities towards the main grid.
This two way responsibility is particularly interesting for scenarios of electric power systems with very high penetration of
distributed generation and where a large part of the grid are sub-aggregated units -- the subgrids. The subgrid control architecture
tries to organize the grid in a flexible way which allows dynamic aggregation and de-aggregation of resources at different control
levels. The process of (de)aggregation is supposed to be flexible enough to incorporate both technical and commercial motivations
for aggregation and should have a mechanism for capturing semantics for differentiating these modes. Figure 2 shows a symbolic
presentation of such a scenario. The concept of subgrid based control has been motivated and based upon the capabilities of
intelligent systems e.g., modularity, decomposition, local/distributed control and its implementation is anticipated to be done using
sophisticated mechanisms of coordination, cooperation and competition provided by multiagent technology.

Figure 2. Decomposition of the power grid into a main grid interconnected with a number (N) of sub-grids (Lind et al. 2008; Xu et
al. 2008)
Figure 3 presents different the modes of operation, and transition among these modes for a subgrid. In the following we briefly
describe these modes and their operation. It should be noted that "connected", "islanding" and "commercially aggregated" are
stable modes, whereas "blackout", "connected alert", "synchronization" and "aggregated alert" are modes of transition.

3.1 Connected:
This is a mode of normal operation when the subgrid is part of the main grid and taking part in normal operations. From this state
the subgrid can either go to "connected alert ", in the case where early warning systems work (emergency), directly go to

65

"blackouted" state when early warning systems does not work, or to "aggregation alert" state where it starts
planning/synchronizing for commercial aggregation.

3.2 Connected Alert:
An alert message may come from a PMU based early warning system informing about some disturbance or fault and the subgrid
goes to the connected alert state. The subgrid controller suggest an optimized plan for islanding operation at this state which may
include reconfiguration and load shedding schemes based on the current situation. From "connected alert" mode the subgrid can go
to one of following states:

Connected mode: In the situation when a disturbance is not very severe, e.g., under-voltage signal and the grid was able to
overcome it without any need to go in islanded mode (restoring)

Islanded mode: In the situation when the system has prepared a (partial) optimized plan for islanding operation and goes
safely to islanding (optimized islanding)

Blackouted mode: in the case of a very severe disturbance and there was not enough time to prepare a plan for islanding.

3.3 Islanded:
The sub-grid may enter this mode from one of following:

Connected mode: when early warning systems don't work and subgrid directly goes to islanded (non optimized islanding)

Connected alert mode: when an alert message from early warning system takes subgrid to connected and it comes to
islanded state with already prepared (semi) optimized plan for islanding operation

Blackouted mode: when system was blackouted initially but then was able to prepare an optimized plan for islanding
operation and blackstarts later

Figure 3. Control modes and transition for subgrid operation

3.4 BlackOuted:
In this subgrid operating mode there is no operation going on in the subgrid and the power is lost by all the loads. The system
tries to make an optimal islanding plan at this state and tries to blackstart.


66

3.5 Aggregated alert:
It is a transitional mode and is reached when some agents controlling the grid resources decide to create a dynamic aggregation
based upon some commercial motives. During this mode synchronization and planning is performed to prepare for commercial
aggregated operation.

3.6 Commercially Aggregated:
This is a stable mode where aggregated resources perform a commercial operation e.g. VPP operation or market based
intentional islanding of microgrid.
A more detailed description of the subgrid based control architecture and discussion on its usability in electric power systems with
very high penetration of distributed generation can be found in (Lind et al. 2008).

4. Development of Software platform for design and testing of Flexible Control

Section II performed a requirement analysis for applying intelligent agents and autonomous systems in electric power systems
with high penetration of distributed generation, a flexible control architecture was presented for such systems in section III. This
section describes development of a software platform to design and test such multiagent based flexible control strategies. A
dynamic multiagent platform has been implemented in the Java agent development framework (JADE)
3
. The platform consists of
one main container, and several sub containers. Each sub container represents a sub-aggregation unit in a electric power network
consisting of one load shedding agent LS and several DG and load agents. Both DG and load agents can join or leave the network
dynamically according to the changes in the network. New such sub-aggregation units can also be created following any situation
in the network. The software platform also includes JADE utility agents and services. Some of the important utilities and agents
are: i) DF (directory facilitator) agent which provides yellow page services. DG and load agents interact with this agent to register
and discover agent services, ii) AMS (agent management services) agent which provide white page services. This agent is
responsible for creating, destroying and managing agents and containers in a JADE platform, iii) MTS (message transport service)
is a service responsible for message transportation between agents within a container and across containers. This service also
enables synchronization of messages when several messages are sent and received from different agents in parallel. In order to take
full advantage of agent capabilities such as autonomy, local control, scalability and high level communication, the software
platform is implemented as fully compliant with FIPA (foundation for physical intelligent agents)
4
standards. Figure 4 shows the
structure of the software platform. It presents a symbolic representation of containers which represent electrical islands in a
distribution system, the agents inside these containers and the JADE utility services described above. Detailed discussion on
development and effectiveness of this platform has been described in (Saleem et al. 2009), where results have been presented from
several experiments of using this platform in distributed control scenarios.

Figure 4. Software platform for design and testing of flexible control strategies

3
Java Agent Development Framework (JADE): http://jade.tilab.com
4
Foundation for Physical Intelligent Agents (FIPA): http://www.fipa.org

67

5. Conclusions

A requirement analysis has been performed for utilization of intelligent agents and autonomous systems in innovative control
concepts of electric power system. A flexible subgrid based control architecture has been proposed for control of electric power
systems with very high penetration of distributed generation. A multiagent software platform has been described. The purpose of
this software platform is to support design and test of multiagent based flexible control strategies. This platform has been based on
the requirement identified in section II and the control architecture envisioned in section III.
The approach of the paper is to identify requirement and to map them into the capabilities of intelligent systems. Subsequently,
to suggest a control architecture based upon these requirements and present the development of a software platform for design and
test of such control architecture.

Nomenclature
DG Distributed Generation
VPP Virtual Power Plant
JADE Java Agent Development Framework
FIPA Foundation of Physical Intelligent Agents
PMU Phasor Measurement Unint

Acknowledgement
The authors acknowledge the support obtained from Energinet.dk and the inspiration and insights gained from discussions with
members of the EcoGrid.dk work package II team. It is emphasized that the paper represents the personal views and conclusions of
the authors and not that of Energinet.dk.

References

Bernon, C. Gleizes, M. & Picard, G. 2003. ADELFE: A Methodology for Adaptive Multi-agent Systems Engineering Lecture
Notes in Artificial Intelligence (Subseries of Lecture Notes in Computer Science
Cherian, S. & Knazkins, V. 2007 The Danish Cell Project - Part 2: Verification of Control Approach via Modeling and Laboratory
Tests. In Proceedings of the IEEE Power Engineering Society General Meeting
Dimeas, A. & Hatziargyriou, N. 2004. A multi-agent system for Microgrids Lecture Notes in Artificial Intelligence (Subseries of
Lecture Notes in Computer Science), Springer Verlag
Farred, K. Reza, I. Hatziargyriou, N. & Dimeas, A. 2008. Microgrid Management: control and operation aspects of microgrids
IEEE Power and Energy Magazine, 6, 54-63
Hatziargyriou, N.,2008. Microgrids (guest editorial), IEEE Power and Energy Magazine, PP, 26-29
Jennings, N., 1995. Controlling cooperative problem solving in industrial multi-agent systems using joint intentions Elsveir
Artificial Intelligence, 75, 195-240
Lind, M. et al. 2008. EcoGrid: 50% Wind power in the Danish Electric Power System: System Architecture WP-2 (Phase 1) In
Series of Energinet.dk
Liu, Y., Dong, L. & Sun, Y., 2006. Cooperation Model of Multi-agent System Based on the Situation Calculus IAT '06:
Proceedings of the IEEE/WIC/ACM international conference on Intelligent Agent Technology, IEEE Computer Society
Lund, P. 2007. The Danish Cell Project - Part 1: Background and General Approach. In Proceedings of the IEEE Power
Engineering Society General Meeting
Rehtanz, C., 2003. Autonomous Systems and Intelligent Systems in Power System Control and Operation, Springer Verlag
Saleem, A., Heussen, K. & Lind, M. 2009. Agent Services for Situation Aware Control of Power Systems With Distributed
Generation. In Proceedings of the IEEE PES General Meeting
Shi, Y. 2009. Genetic Virtual Power Plant for Combined Heat and Power Units: PhD project Technical Universty of Denmark
DTU
Wernstedt, F. & Davidsson, P., 2002. Adaptive Control - An Agent-Based Approach to Monitoring and Control of District Heating
Systems Lecture Notes in Computer Science, Springer Verlag
Wooldridge, M., 2002. An Introduction to Multiagent Systems, John Wiley & Sons
Xu, Z., Gordon, M., Lind, M. & Ostergaard, J. 2009. Towards a Danish Power System with 50% Wind - Smart Grids Activities
in Denmark. In Proceedings of the IEEE PES General Meeting


68

Biographical notes:

Arshad Saleem is currently pursuing a Ph.D.degree from the Department of Electrical Engineering of Technical University of Denmark. He received his bachelors
degree in computer science and and an M.Sc degree in Artificial Intelligence from Blekinge Institute of Technology Karlskrona, Sweden in 2003 and 2007
respectively. His research interests are intelligent systems applications to power system, power system analysis and distributed control.

Morten Lind is Professor of Automation at Department of Electrical Engineering at Technical University of Denmark. His research interests include automation
design, supervisory control of complex industrial systems and infrastructures, and application of agent technology and knowledge based system in automation.

Accepted March 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 69-81

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Influence of TCSC on social welfare and spot price - A comparative study of
PSO with classical method

S. K. Joshi
1
*, K. S. Pandya
2

1*
Electrical Engineering Department,,The M.S.University of Baroda, Vadodara,INDIA
2
Electrical Engineering Department, Charotar Institute of Technology, Changa, INDIA
*
Corresponding Author: e-mail: skjoshi@ieee.org, Tel +91-265-2352958, Fax.+91-265-2423898

Abstract

In this paper, Particle Swarm Optimization (PSO) based algorithm has been suggested to find optimal location and setting of
Thyristor Controlled Series Compensator (TCSC) to maximize Social Benefit (SB), considering its installation cost in
competitive electricity market. PSO has simultaneously optimized generators active powers output, generators bus voltages,
TCSC reactance and its location. In addition, the influence of optimally located TCSC on the magnitude of real power spot
price, reactive power spot price, wheeling charges and bilateral transaction matrix has been investigated. The effectiveness of the
proposed approach has been tested on IEEE 6 bus system and results obtained are compared with those obtained from various
classical methods.

Keywords: Bilateral transaction, Particle Swarm Optimization, Social Benefit, Spot price

1. Introduction

In a last decade, many monopolistic electricity companies and electricity boards have been transformed into competitive
structure to obtain higher efficiency, revenue, better service to consumers, to increase stability and much more. But, rapid rise in
power demand, competition, outage and scare natural resources are some factors due to which transmission systems operate very
near to their thermal limits. But because of economic, environmental and political reasons it is not preferable to build new
transmission lines. So there is an interest in better utilization of existing capacities of power system by installing Flexible A.C.
Transmission System (FACTS) device such as Thyristor Controlled Series Compensator (TCSC)(Hingorani et al, 2000). It is the
power electronics based device which can enhance Total Transfer Capability, voltage stability, loadability, security etc. and can
reduce losses, cost of generation, can remove congestion and fulfill transaction requirement rapidly, dynamically and efficiently.
Because of following reasons it is important to optimally locate TCSC in order to obtain its full benefits. (1) It is a costly
device; (2) It has been proved that TCSC may have negative effect on system stability unless it is optimally placed (Yu et al,
2003); (3) It may adversely affect some market participants in deregulated power system.

1.1 Literature Review:
Various classical and artificial intelligence methods have been suggested to optimally locate FACTS devices with different kinds
of objective functions.
Interior point method was used for system expansion with UPFC to maximize social welfare and to manage congestion (Lin et
al, 2000). Genetic algorithm was used for multi-type FACTS placement to enhance system loadability. However, number of
devices to be installed was assigned before optimization and installation cost was ignored (Gerbex et al, 2001). Mixed Integer
Linear Programming was used to find optimum number, locations and setting of Thyristor Control Phase Shifting Transformer
(TCPST) to optimize loadability and investment with respect to dc load flow model (Lima et al, 2003). Mixed Integer Non-linear
programming was used for optimal location of FACTS to maximize social welfare based on multiple time periods. But it ignored
the impact of FACTS on reactive power flow (Yu et al, 2004). Predictor corrector primal duel interior point linear programming
was used to enhance Available Transfer Capability (ATC) using various FACTS devices. However, it did not optimize their
Joshi et al. / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 69-81

70

ratings and locations (Xiao et al, 2003). A hybrid evolutionary algorithm was proposed to optimally place multi-type FACTS
devices to maximize TTC and minimize losses (Jirapong et al, 2007). A sensitivity based approach was used to study the impact
of Unified Power Flow Controller (UPFC) on real and reactive power spot prices (Verma et al, 2006). Sequential quadratic
programming was used to study impact of TCSC on congestion and spot price. But it did not optimize TCSC setting and location
(Archarya et al, 2007). A Mixed-Integer Quadratic programming was used to find optimal location of TCSC for enhancing system
loadability, voltage profile considering its investment cost (Yang et al, 2007). A real genetic algorithm was proposed to optimally
locate two TCSC for enhancing ATC and to get best voltage profile But generators active output powers and bus voltages were
not optimized (Rashidinejad et al, 2008). A mixed integer nonlinear programming was used to locate UPFC to maximize system
loadability for pool and bilateral electricity markets (Kumar et al, 2009). A reactive power spot price index was proposed to place
SVC to minimize total generation cost and to maximize loading margin (Singh et al, 2007).
So, it is revealed that most of the OPF problems are non-linear and non-convex. With the inclusion of FACTS control variables,
they become even more nonlinear because they change the size of bus admittance matrix and dimension of the problem.
Conventional classical optimization methods like gradient method, lamda iteration, linear programming etc. rely on the convexity
assumption of objective function. They fail to capture discontinuities of objective function and may get trapped into local minima
or diverge at all. Choice of initial starting point also affects the quality of solution. Also, they could find only single optimized
solution in a single simulation run.
Thus, to find global optimum solution is a challenging task in optimization problem incorporating FACTS devices. To solve
such problem, an artificial intelligent method called Particle Swarm Optimization may be used as it is a fast method and it provides
global or near global solution (Kennedy et al, 1995). PSO has shown its superiority over other classical and AI methods with
respect to execution time and global solution in solving economic dispatch problem (Park et al, 2005), optimal reactive power
dispatch problem (Zhao et al, 2005)and congestion management (Dutta et al, 2008)
So in this paper, PSO based algorithm has been suggested to find the best location and setting of TCSC to maximize social
benefit and to minimize total generation cost while satisfying various constraints of IEEE 6 bus test system. The same algorithm
can be modified to maximize social benefit without considering TCSC. Comparisons are carried out with solutions obtained from
Non-linear programming method (MINOS) (Murtagh et al) used in MATPOWER (Zimmerman et al, 2007). Transmission service
pricing is also an important issue of deregulated market. Out of many suggested pricing methods, Locational Marginal Pricing
(LMP) method is popular because it considers all system constraints and losses. As PSO cannot provide Lagrange multipliers
which are required for finding LMP, an interior point method is used to calculate LMP. But choice of initial starting points greatly
affects the quality of solution of interior point method (Xie et al, 2001). So in this paper, optimized output obtained from PSO are
used as starting points and results are compared with default setting of Primal Dual Interior Point Method (Wang et al, 2007)used
in MATPOWER. Lastly, influence of optimally placed TCSC by PSO on wheeling charges and secure bilateral transaction matrix
has been studied and results are compared with those obtained by Primal Dual Interior Point Method.
This paper can be broadly divided into two parts. Firstly, PSOs superiority over various classical methods like non-linear
programming method and interior point method in obtaining global solution without using TCSC has been searched out. Secondly,
impact of optimally placed TCSC by PSO on social benefit, total generation cost, losses, real power and reactive power spot
prices, wheeling charges and secure bilateral transaction has been investigated.
This paper has been organized as follows: Section 2 describes static modeling of TCSC. Section 3 explains overview of PSO
method. Section 4 explains the optimal power flow formulation and major steps involved in the algorithm. Section 5 explains the
PSO based algorithm in detail. Section 6 discusses simulation results and in Section 7, conclusive remarks have been given.

2. Static modeling of TCSC

As shown in Figure 1., the TCSC has been represented by a variable capacitive/inductive reactance inserted in series with the
transmission line (Sharma, 2006). So the reactance of the transmission line is adjusted by TCSC directly. Let,
mn
X is the reactance
of the transmission line, Xc is the reactance of TCSC and
new
X is the new reactance of the line after placing TCSC between bus m
and n. Mathematically, equation is written as:
new mn
X X Xc = (1)
The modified power flow equations of the transmission line in the presence of TCSC are given as below:
2
( cos sin )
mn m mn m n mn mn mn mn
P V G V V G B = + (2)
2
( sin cos )
2
mn m mn m n mn mn mn mn
B
Q V B V V G B

= +

(3)
2
( cos sin )
nm n mn m n mn mn mn mn
P V G V V G B = (4)
2
( sin cos )
2
nm n mn m n mn mn mn mn
B
Q V B V V G B

= + + +

(5)


71

where,
2 2 2 2
( )
,
( ) ( )
mn mn c
mn mn
mn mn c mn mn c
R X X
G B
R X X R X X

= =
+ +

,
mn mn
P Q : Active and reactive power flow from bus m to n
,
nm nm
P Q : Active and reactive power flow from bus n to m
mn
G : New line conductance between bus m and n
mn
B : New line susceptance between bus m and n
mn
R : Line resistance between bus m and n

3. Basics of the Particle Swarm Optimization

PSO is a fast, simple and efficient population-based optimization method which was proposed by Eberhart and Kennedy. It has
been motivated by the behavior of organisms such as fish schooling and bird flocking. In PSO, a Swarm consists of number of
particles which represent the possible solutions. The coordinates of each particle is associated with two vectors, namely the
position ( )
i
x and velocity ( )
i
v vectors. The size of both vectors is same as that of the problem space dimension. All particles in a
swarm fly in the search space to explore optimal solutions. Each particle updates its position based upon its own best position,
global best position among particles and its previous velocity vector according to the following equations:
1
1 1 2 2
( ) ( )
i
k k k k
i i best i best i
v w v c r p x c r g x
+
= + + (6)
1 1 k k k
i i i
x x v
+ +
= + (7)
where,

1 k
i
v
+
: The velocity of
th
i particle at ( 1)
th
k + iteration
w : Inertia weight of the particle

k
i
v : The velocity of
th
i particle at
th
k iteration
1, 2
c c : Positive constants having values between [0, 2.5]
1 2
, r r : Randomly generated numbers between [0, 1]
i
best
p : The best position of the
th
i particle obtained based upon its own experience
best
g : Global best position of the particle in the population
1 k
i
x
+
: The position of
th
i particle at ( 1)
th
k +

iteration
k
i
x : The position of
th
i particle at
th
k iteration
: Constriction factor. It may help insure convergence. Its low value facilitates fast convergence and little exploration while high
value results in slow convergence and much exploration.
If no restriction is imposed on the maximum velocity
max
( ) v of the particles then there is likelihood that particles may leave the
search space. So velocity of each particle is controlled between
max
( ) v to
max
( ) v .
Suitable selection of inertia weight w provides good balance between global and local explorations. It is set according the
following equation.
max min
max
max
w w
w w iter
iter
= (8)

72

Where,
max
w is the value of inertia weight at the beginning of iterations,
min
w is the value of inertia weight at the end of iterations,
iter is the current iteration number and
max
iter is the maximum number of iterations.

4. Proposed solution

The proposed algorithm is given in section 5. The major steps involved in the algorithm are summarized in the following.

4.1 Initial population

In this paper, a particle is made of continuous and integer control variables. The continuous variables include the generators
active powers output (PG
2, ,
PG
NG
), generators bus voltages (VG
1, ,
VG
NG
)and reactance value of TCSC (X
C
). Generators
active powers and generators bus voltages are generated randomly within their permissible minimum and maximum limits.
Reactance of TCSC is generated randomly between 0 and 1(normalized form, X
(N)
) and its actual value (X
(D)
) is found using
eqn. (9).
( ) ( )
( )
D (min) max min N
X X X X X = + . (9)
Where,
min
X and
max
X are minimum and maximum values of the variable.
The integer variable consists of possible location (Loc) of TCSC. The particles can be generated in matrix form as shown in
Figure 2.

Particle No. PG
2
PG
NG
VG
1
VG
NG
X
C
Loc
1 PG
2,1
PG
NG,1
VG
1,1
VG
NG,1
X
C1
Loc
1
2 PG
2,2
PG
NG,2
VG
1,2
VG
NG,2
X
C2
Loc
2

i PG
2,i
PG
NG,i
VG
1,i
VG
NG,i
X
Ci
Loc
i

Figure 2. Representation of particle

Where, PG
2,i
, PG
NG,i
: From 2
nd
to NG
th
generators active output powers corresponding to i
th
particle excluding slack bus generator
power.
VG
1,i,
VG
NG,i
: From 1
st
to NG
th
generators voltage magnitudes corresponding to i
th
particle including slack bus generator voltage.
X
Ci
: Reactance of TCSC corresponding to i
th
particle
Loc
i
: Location (line number) of TCSC corresponding to i
th
particle.
If TCSC is not included in the transmission system, then variables X
c
and Loc are not considered.
If there are total i number of particles and if each particle consists of j number of control variables, then the dimension of a
population becomes i * j.

4.2 Installation of TCSC.

After generating initial population of particles, for each particle TCSC is randomly installed in the transmission line with
randomly generated reactance. Then after, new value of bus admittance matrix is found out.

4.3 Power flow

Run full a.c. Newton-Raphson power flow for all particles to obtain generators active and reactive output powers, bus voltages,
load angles, line flows, active and reactive power losses of transmission lines.

4.4 Optimal power flow problem formulation

An OPF (fitness function) to maximize Social Benefit (to minimize total generation cost) considering installation cost of TCSC
subject to various equality and inequality constraints using PSO can be formulated as (10).
( ) ( )
1 1
min{ ( ) ( )}
G D
N N
Gm Gm Dn Dn TCSC
m n
C P B P IC PF
= =
+ +

(10)


73

An OPF for maximizing social benefit without considering TCSC can be formulated as (11)
( ) ( )
1 1
min{ ( )}
G D
N N
Gm Gm Dn Dn
m n
C P B P PF
= =
+

(11)
Subject to the power balance equations (equality constraints)
1
1
- - cos( - - ) 0
- - sin( - - ) 0
b
b
N
Gm Dm m n mn m n mn
n
N
Gm Dm m n mn m n mn
n
P P V V Y
Q Q V V Y

=
=

=

=

(12)
Various operating constraints (inequality constraints)
min max
,
Gm Gm Gm G
P P P m N (13)
min max
,
Gm Gm Gm G
Q Q Q m N (14)
max
,
l l L
S S l N (15)

min max
,
m m m b
V V V m N (16)
min max
C C C
X X X p.u. (17)
Where,
( )
Gm Gm
C P : Bid function of m
th
generator bus (seller bus)
( )
Dn Dn
B P : Bid function of n
th
consumer bus (buyer bus)
TCSC
IC : Installation cost of TCSC ($)
PF : Penalty Function
,
Gm Gm
P Q : Active and reactive power generation at bus m
,
Dm Dm
P Q : Active and reactive power demand at bus m
m m
V : Complex voltage at bus m
mn mn
Y :
th
mn element of bus admittance matrix
min max
,
Gm Gm
P P : Active power generation limits at bus m
min max
,
Gm Gm
Q Q : Reactive power generation limits at bus m
max
l
S : Thermal limit of
th
l transmission line
min max
,
m m
V V : Voltage magnitude limits at bus m
min
0.85
C mn
X X = : Lower limit of reactance of TCSC
max
0.2
C mn
X X = : Upper limit of reactance of TCSC
L
N : Total number of transmission lines
b
N : Total number of buses
G
N : Total number of generator buses
D
N : Total number of load buses
The cost function of TCSC is given in Siemens database and used in (Cai et al, 2004). Mathematically it is written as equ (18)
2
0.0015( ) 0.7130( ) 153.75
TCSC
C S S = + (18)
Where,
TCSC
C is the cost of TCSC in US$/KVAR and S is the operating range of the TCSC in MVAR.
1 2
S Q Q = (19)
Where
1
Q is the reactive power flow in the line before placing TCSC in MVAR and
2
Q is the reactive power flow in the line after
placing TCSC in MVAR.
The installation cost ($) of TCSC is given by (20)
1000
TCSC TCSC
IC C S = (20)
Square penalty function is used to handle inequality constraints such as reactive power output of generator buses, voltage
magnitude of all buses and transmission lines thermal limits respectively, as shown in (21) and (22).
1 2 3
1 1 1
( ) ( ) ( )
G b L
N N N
Gm m lm
m m m
PF k f Q k f V k f S
= = =
= + +

(21)

74

min max
max 2 max
min 2 min
0
( ) ( )
( )
if x x x
f x x x if x x
x x if x x

= >

<

(22)

where,
1 2 3
, , k k k : Penalty coefficients for reactive power output of generator buses ( )
Gm
Q , voltage magnitude ( )
m
V of all buses and
transmission line loading ( )
lm
S , respectively. The value of each coefficient is equal to 1000.
min max
, x x : Minimum and maximum limits of variable x.

4.5 Spot Price

In the second phase of optimization, Real power spot price and Reactive power spot price (Locational Marginal Price) at each
generator bus and load bus is found. But, PSO does not provide Lagrange multipliers (required for determination of spot price)
during optimization. So an interior point method is used to calculate LMP. But choice of initial starting points greatly affects the
quality of solution and convergence of interior point method. So in this work, optimized output obtained from PSO are used as
starting points for interior point method and results are compared with those obtained by using default setting of Primal Dual
Interior Point Method used in MATPOWER.

4.6 Wheeling Charge

Wheeling is the transmission of electrical energy from a seller bus to a buyer bus through a transmission network owned by a
third party. Wheeling rates are the charges given to the transmission owner.
Wheeling rate ( )
p
W for the real power can be expressed as (23)
p pB pS
W = (23)
Where,
pB
is the spot price of real power at the buyer bus-B,
pS
is the spot price of real power at the seller bus-S
Wheeling charge for purchase of real power ( )
B
P is expressed as (24)
p B p
WC P W = (24)
Wheeling rate ( )
q
W for the reactive power can be expressed as (25)
q qB qS
W = (25)
Where,
qB
is the spot price of reactive power at the buyer bus-B,
qS
is the spot price of reactive power at the seller bus-S
Wheeling charge for purchase of reactive power ( )
B
Q is expressed as (26)
q B q
WC Q W = (26)

4.7 Social Benefit of various market participants.

Social Benefit was originally defined by the Bergson-Samuelson social welfare function and latter on modified by the
economists to refer to the benefit of various participants (Luenberger, 1995). The benefit to the generator is the revenue minus the
generation cost, referred to as the generator surplus (Producer surplus). It can be also expressed as (27), if the spot price at the
generator bus is known. Transmission network is managed by a non-profit organization known as the Wholesaler( Merchandize).
Different spot prices exist at the generator buses and load buses due to the losses and congestion. So merchandize surplus is the
revenue obtained to wholesaler to compensate losses and congestion as given in (29). Consumer surplus is the benefit obtained to
consumer from the consumption of electrical energy as given in (28). Social Benefit is the addition of Producer surplus,
Merchandize surplus and Consumer surplus.
( )( )
min
1
1
2
G
N
m gm Gm Gm
m
producersurplus b P P
=
=
(27)
Where,
m
is the spot price at generator bus m
gm
b is the linear coefficient in the quadratic generator bid function
Gm
P is the real power output of generator m

75

min
Gm
P is minimum real power generation limit of generator m
( )( )
min
1
1
2
D
N
dn n Dn Dn
n
consumersurplus b P P
=
=
(28)
Where,
dn
b is the linear coefficient in the quadratic demand function
n
is the spot price at load bus n
Dn
P is the real power demand at load bus n
min
Dn
P is minimum real power demand limit at load bus n
1 1
G D
N N
n Dn m Gm
n m
merchandize surplus P P
= =
=

(29)

4.8 Effect of TCSC on the magnitude of bilateral transactions

A bilateral transaction is made directly between a seller and a buyer without any third party intervention. Mathematically, each
bilateral transaction between a seller at bus m and buyer at bus n satisfies the following power balance relationship:
0
Gm Dn
P P = (30)
The secure bilateral transaction matrix is obtained from the proposed bilateral transaction matrix after obtaining social benefit.
2
min ( )
o
mn mn
m n
T T

(31)
Where
mn
T is the secure bilateral transaction between seller and buyer buses.
o
mn
T is the proposed bilateral transaction between seller and buyer buses.
The value of secure bilateral transaction matrix has been obtained without and with optimally placed TCSC using PSO. Its value is
also obtained by interior point method. Results obtained using PSO method is compared with those obtained from interior point
method.

5. Algorithm to optimally locate TCSC to maximize social benefit using PSO

Step 1: Input the data of lines, generators, buses and loads. Choose population size of particles and convergence criterion. Define
type of power transaction.
Step 2: Select generators active power output excluding slack bus generator power, voltage magnitude of generator buses
including slack bus voltage, reactance setting of TCSC and location (line number) of TCSC as control variables. If TCSC
is not included in the OPF problem, then variables- reactance setting of TCSC and location are not considered.
Step 3: Randomly generate population of particles in such a way that their variables fall within their feasible limits.
Step 4: Modify the bus admittance matrix.
Step 5: Run full a.c. Newton-Raphson load flow to get line flows, active power generations, reactive power generations, line losses
and voltage magnitude of all buses.
Step 6: Calculate the penalty function of each particle using eqn. (21).
Step 7: Calculate the fitness function of each particle using eqn. (10) (if TCSC is included in optimization)
Or eqn. (11)(if TCSC is not included in optimization).
Step 8: Find out the global best ( )
best
g particle having minimum value of fitness function in the whole population and personal
best ( )
i
best
p of all particles.
Step 9: Generate new population using eqns. (6) and (7).
Step 10: Go to step 4 until maximum number of iterations are completed.
Step 11: After the optimization, the fitness value of
best
g particle gives minimized value of total generation cost (if the loads are
inelastic) including installation cost of TCSC or gives maximized value of Social Benefit (if the loads are elastic).
Coordinates of
best
g particle give the optimized values of generators active power output, generators bus voltages,
optimal reactance setting of TCSC and location of TCSC respectively.
Step 12: Real power and reactive power spot price at each generator and load bus is found using interior point method with and
without using TCSC. Optimized outputs obtained by PSO in step 11 are used as the starting point in interior point method.
Step 13: Wheeling charges for real power and reactive power can be calculated using eqn (24) and (26).
Step 14: Surplus of various market participants could be obtained using eqn (27), (28) and (29).


76

6. Simulation results and discussions

The simulation studies were carried out on intel CORE 2 Duo, 2GB of RAM, 2.20 GH
Z
system in MATLAB 7.6 platform.
The IEEE 6 bus system has been used to test the proposed algorithm. The bus, line and generator data was taken from
(Zimmerman et al, 2007). It consists of 3 generators and 11 transmission lines.
As shown in Table 1, generation cost minimization OPF is solved using PSO (Case 1A) and non-linear programming (MINOS
solver, Case 1C) without using TCSC. The obtained total generation cost is 3,128.8 $/h by PSO, whereas it is 3,143.97 $/h by
MINOS. It is due to that fact that PSO can simultaneously optimize generators active powers output and generator bus voltages.
So it can find a global solution. But MINOS trapped into local minima. Active power loss and reactive power loss obtained from
PSO are lesser than those obtained from MINOS. Reactive power output of generators no. 2 and 3 are reduced by PSO, so they can
be even more efficiently utilized. In case 1B, optimal reactance setting and location of TCSC are found by PSO to minimize the
composite objective function which consists of total generation cost and installation cost of TCSC. Total generation cost is 3,125.3
$/h, which is lesser than those obtained in cases 1A and 1C. Optimal installation cost of TCSC obtained was 0.3977*10
6
($).
Optimal reactance of TCSC was negative, means it operated in capacitive mode. Optimal location of TCSC was the line that was
connected between bus 1 and bus 4. Also, Optimally placed TCSC significantly reduced active power losses and reactive power
losses.

Table 1. Comparison of results of Cost minimization by PSO and MINOS

PSO
Non-Linear Programming (MINOS)
in MATPOWER

Without TCSC
Case 1A
With TCSC
Case 1B
Without TCSC
Case 1C
Total generation cost ($/h) 3,128.8 3,125.3 3,143.97
Real power o/p of gen.G
1
(P
1
MW) 52.96 49.98 77.22
P
2
86.28 89.02 69.27
P
3
77.58 77.62 70.42
Total generation (MW) 216.83 216.61 216.91
Generator bus voltage, V
1
(p.u.) 1.0546 1.042 1.050
V
2
1.0351 1.048 1.050
V
3
1.0508 1.048 1.070
TCSC setting (p.u) ---- -0.0980 ---
Location ---- Bus 1-4 ---
Installation cost of TCSC ($) ---- 0.3977*10
6
---
Active power loss (MW) 6.826 6.611 6.908
Reactive power loss (MVAR) 21.15 18.53 21.21
Reactive power of G
1,
(Q
1
MVAR) 58.06 43.36 25.72
Q
2
43.05 68.84 64.65
Q
3
77.11 63.02 86.64

As shown in Figures 3(a) and 3(b), Real power spot price and Reactive power spot price (Locational Marginal Price) at each
generator bus and load bus were found. The generators are connected at bus 1, 2 and 3. The loads are connected at bus 4, 5 and 6.
The spot prices of real power are very high as compared to that of reactive power. Most of the spot prices obtained at load buses in
Case 3A were lower than that of Case 3C. So it is cleared that consumers surplus increase when PSOs optimized outputs are used
as starting point for interior point method. Similarly, most of the spot prices obtained at generator buses in Case 3A were higher
than that of Case 3C. So producers surpluses increase when PSOs optimized results are used as starting point for IP method.
Optimally placed TCSC (Case 3B) reduce real power spot price and reactive power spot price of all load buses and of some
generator buses. It is because TCSC redistributes power flow in the transmission lines in such a way to decrease the losses and
thus it tries to equalize spot prices at various load buses. Thus, TCSC increases consumers surpluses and producers surpluses.


77

Case 3A- Optimized solutions of PSO used as starting points in PDIPM (Without TCSC)
Case 3B- Optimized solutions of PSO used as starting points in PDIPM. (With TCSC)
Case 3C- Default setting of Primal/dual Interior point method (PDIPM) in MATPOWER (Without TCSC)

0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
R
e
a
c
t
i
v
e

p
o
w
e
r

s
p
o
t

p
r
i
c
e

(
$
/
M
V
A
R
H
)
1 2 3 4 5 6
Bus
Case 3A
Case 3B
Case 3C

Figure 3(a). Comparison of real power spot price obtained Figure 3(b). Comparison of reactive power spot price obtained
from PSO based starting points and default from PSO based starting points and default starting
starting points in interior point method. points in interior point method.

Figure 4 shows the wheeling charges for real power of various bilateral transactions. Comparing the results of Case 4A with
Case 4C reveals that most of the wheeling charges obtained by PSO have lesser value than those obtained by IP method. For
transaction between buses 3-6 wheeling charge of Case 4A is greater than that of Case 4C, because load at bus 6 is responsible for
large power flow in some transmission lines which increases its spot price. When TCSC is optimally placed (Case 4B) by PSO, it
significantly reduces wheeling charges for real power. For transaction between bus 2-4 , highest reduction in wheeling charge is
obtained because the value of load at bus 4 is such that it requires less power flow in some transmission lines and TCSC reduces
losses to a great extend and thus wheeling charge is significantly reduced.

Case 4A-Wheeling charges obtained by PSO (Without TCSC)
Case 4B- Wheeling charges obtained by PSO (With TCSC)
Case 4C-Wheeling charges obtained by Interior point method (Without TCSC)

0
50
100
150
200
W
h
e
e
l
i
n
g

c
h
a
r
g
e
s

(
$
/
M
W
H
)
(1-4) (1-5) (2-4) (2-5) (2-6) (3-5) (3-6)
Seller bus-Buyer bus
Case 4A
Case 4B
Case 4C

Figure 4. Wheeling charges ($/MWH) of real power of selected bilateral transactions

Figure 5 shows the wheeling charges for reactive power of various bilateral transactions. It can be seen that even though the spot
price of reactive power are very less as compared to that of real power spot price, the wheeling charge of reactive power is quite
significant and should be considered while formatting the bilateral contracts. Wheeling charge of reactive power for transaction
between buses 2-6 becomes negative after placing TCSC (Case 5B). So buyer bus 6 will get discount for purchasing reactive
power from seller bus 2. So TCSC encourages buyer bus 6 to reschedule its load in such a way that it may even get more discounts
from the transaction.

0
2
4
6
8
10
12
14
16
R
e
a
l

p
o
w
e
r

s
p
o
t

p
r
i
c
e

(
$
/
M
W
H
)
1 2 3 4 5 6
Bus
Case 3A
Case 3B
Case 3C

78

Case 5A:Wheeling charges obtained by PSO (Without TCSC)
Case 5B:Wheeling charges obtained by PSO (With TCSC)
Case 5C:Wheeling charges obtained by Interior point method (Without TCSC)

-50
0
50
100
150
200
W
h
e
e
l
i
n
g

c
h
a
r
g
e
s

(
$
/
M
V
A
R
H
)
(1-4) (1-5) (2-4) (2-5) (2-6) (3-5) (3-6)
Case 5A
Case 5B
Case 5C

Figure 5. Wheeling charges ($/MVARH) of reactive power of selected bilateral transactions

As shown in Figure 6, Consumers surplus, Merchandize surplus and Producers surplus can be obtained using eqn (28), (29) and
(27) respectively. Optimally placed TCSC (Case 6B) can significantly increase Consumers surplus and Producers surplus, which
are the main aims of deregulation. Merchandize surplus includes cost of losses and cost of congestion. As TCSC reduces losses of
the transmission system and thus removes congestion, it drastically decreases Merchandize surplus.
Case 6A: Surplus obtained by PSO (Without TCSC)
Case 6B: Surplus obtained by PSO (With TCSC)
Case 6C: Surplus obtained by Interior point method (Without TCSC)

0
500
1000
1500
2000
2500
3000
3500
S
u
r
p
l
u
s

(
$
/
h
)
Consumer Merchandize Producer
Variousparticipants
Case 6A
Case 6B
Case 6C

Figure 6. Comparison of surplus obtained by PSO and Interior point method

Figure 7 shows the proposed bilateral transaction matrix. The elements of the figure represent the bilateral contract between the
i
th
seller bus (generator bus) and j
th
buyer bus (load bus). The elements in the table have positive real values. Some contracts
between the generator bus and load bus have zero values. The secure bilateral transaction matrix is obtained from proposed
bilateral transaction matrix after maximizing social benefit function.
Figure 8 shows the secure bilateral transaction matrix obtained by PSO and interior point method, without using TCSC. By
comparing them, it was seen that PSO could fulfill more contractual demand than interior point method in most of the transactions.
It was because PSO could simultaneously optimized generators active power output and generators bus voltages, whereas interior
point method could not find the same.


79

0
5
10
15
20
25
30
35
40
45
50
T
r
a
n
s
a
c
t
i
o
n

v
a
l
u
e

(
M
W
)
(1-
4)
(1-
5)
(1-
6)
(2-
4)
(2-
5)
(2-
6)
(3-
4)
(3-
5)
(3-
6)
0
10
20
30
40
50
T
r
a
n
s
a
c
t
i
o
n

v
a
l
u
e

(
M
W
)
(1-
4)
(1-
5)
(1-
6)
(2-
4)
(2-
5)
(2-
6)
(3-
4)
(3-
5)
(3-
6)
PSO Interior point method

Figure 7. Proposed bilateral transaction matrix Figure 8. Comparison of secure bilateral transaction matrix obtained by
(without TCSC) PSO and Interior point method (without TCSC)

Figure 9 shows the impact of optimally placed TCSC by PSO on the magnitude of secure bilateral transaction matrix. Comparing
the figures 7 and 9, it can be seen that the pattern of secure bilateral transactions is different and it is more uniform in the presence
of TCSC. It is because of power flow control capability of TCSC.
0
10
20
30
40
50
T
r
a
n
s
a
c
t
i
o
n

v
a
l
u
e

(
M
W
)
(1-4) (1-5) (1-6) (2-4) (2-5) (2-6) (3-4) (3-5) (3-6)

Figure 9. PSO based secure bilateral transaction matrix with TCSC

6.1 Effects of PSO parameters on convergence

PSO parameters such as cognitive parameter (C
1
), social parameter (C
2
), Constriction factor ( ) , maximum inertia weight
(W
max
), minimum inertia weight (W
min
), upper limit of velocity (V
max
), and lower limit of velocity (-V
max
) were selected through
experiments and their effects on the value of optimal generation cost have been studied. The results are shown in Table 3. The
population of 35 particles was taken for all cases 3TA-3TE and 25 independent trials were carried out for each case. It was seen
that case 3TC (C
1
=1.2, C
2
=1, =0.3, W
max
=0.7, W
min
=0.3, V
max
= 0.25, and -V
max
=-0.25) gave the minimum generation cost of
3,125.3 $/h. To check the convergence characteristics of PSO with the selected parameters of case 3TC, number of simulations
was carried out. Convergence criterion taken was 25 iterations. The variation in generation cost with the iteration number is shown
in Figure 10 for PSO method and in Figure 11 for NLP (MINOS) method. Comparing both the figures, it is observed that PSO
provides global solution in less than 15 iterations. Whereas, NLP was trapped into local minima.

Table 3. Effect of PSO parameters on convergence

Case C
1
, C
2
Constriction factor ( ) W
max
, W
min
V
max
, -V
max
Best generation cost ($/h)
3TA 1, 1.3 0.6 1, 0.5 0.35, -0.35 3,125.5
3TB 1.1, 2 0.4 0.8, 0.4 0.30, -0.30 3,127.7
3TC 1.2,1 0.3 0.7, 0.3 0.25, -0.25 3,125.3
3TD 2, 1.6 0.2 0.75, 0.35 0.20, -0.20 3,128.2
3TE 2, 2 0.1 0.9, 0.4 0.10, -0.10 3,126.1


80

Figure 10. Convergence characteristic of PSO Figure 11. Convergence characteristic of NLP

7. Conclusion

This paper has proposed PSO based algorithm to maximize social benefit with/without using TCSC. The major outcome of the
paper can be summarized as follows (1) Solutions obtained from PSO method are superior to non-linear programming method in
solving cost minimization OPF problem. (2) Optimally placed TCSC by PSO could significantly reduce total generation cost, real
and reactive power losses, total wheeling charges, real and reactive power spot prices and could increase social benefit of market
participants and secure bilateral transaction. (3) When optimized output parameters obtained from PSO are used as initial staring
points in interior point method to obtain spot prices, then it gives better results than those obtained from their default settings used
in MATPOWER. (4)PSO outperforms interior point method in obtaining secure bilateral transaction matrix. So the proposed PSO
based algorithm can be applied at planning stage while installing TCSC and it is also suitable for on line application at the energy
management centre due to its robust convergence characteristic.

References

Acharya Naresh and Mithulananthan N., 2007. Influence of TCSC on congestion and spot price in electricity market with bilateral
contract, Electric power system research, Vol. 77, pp. 1010-1018.
Cai L.J., Erlich I. and Stamtsis G., 2004. Optimal choice and allocation of FACTS devices in deregulated electricity market using
genetic algorithm, in:proc.of IEEE PES power system Conf. and exposition, USA, pp. 10-13.
Dutta Sudipta and Singh S.P., 2008. Optimal rescheduling of generators for congestion management based on particle swarm
optimization, IEEE Trans. Power Syst., Vol. 23, No. 4, pp. 1560-1568.
Gerbex S, Cherkaoui R and Germond A.J., 2001. Optimal location of multi-type FACTS devices in a power system by means of
Genetic algorithms, IEEE Trans. Power Syst. 16 (3) (2001) pp. 537-544.
Hingorani N.G. and Gyugyi L., 2000. Understanding FACTS Concepts and Technology of Flexible AC Transmission Systems,
IEEE press, ISBN-0-7803-3455-8.
Jirapong Peerapol and Ongsakul Weerakorn, 2007. Optimal placement of multi-type FACTS devices for total transfer capability
enhancement using hybrid evolutionary algorithm, Electr. Power Components and Syst., Vol.35, pp. 981-1005.
Kennedy J. and Eberhart R, 1995. Particle swarm optimization, in: Proceedings of the IEEE International Conf. on Neural
Networks, pp. 1942-1948.
Lima F.G.M., Galiana F.D., Kockar Ivana and Munoz J., 2003. Phase shifter placement in large-scale systems via mixed integer
linear programming, IEEE Trans. Power Syst., Vol. 18, No. 3, pp. 1029-1034.
Lin Whei-Min, Chen shi-Jaw and Su Yuh-Sheng., 2000. An application of interior-point based OPF for system expansion with
FACTS devices in a deregulated environment, in Proc. IEEE conference (0-7803-6338-8/00/$10).
Luenberger D.G., 1995. Microeconomics theory. New York: McGraw-Hill, p. 187.
Murtagh B.A. and Saunders M.A., MINOS 5.5 users guide. Stanford Univ. Systems, optimization laboratory tech. Rep. SOL 83-
20R.
Park J.B., Lee K.S., Shin J.R. and Lee K.Y., 2005. A particle swarm optimization for economic dispatch with nonsmooth cost
functions, IEEE Trans. Power Syst., Vol. 20, No. 1, pp. 34-42.
Rashidinejad M. et al, 2008. ATC enhancement using TCSC via artificial intelligent techniques, Electric Power System Research,
Vol. 78, 2008, pp. 11-20.
Sharma Ashwani kumar, 2006. Optimal number and location of TCSC and loadability enhancement in deregulated electricity
markets using MINLP, International Journal of Emerging Electric Power Systems, Vol. 5, No. 1.


81

Sharma Ashwani kumar and Chanana S., 2009. New secure bilateral transaction determination and study of pattern under
contingencies and UPFC in competitive hybrid electricity markets, Electrical power and energy systems, Vol. 31, pp. 23-33.
Singh J.G., Singh S.N.and Srivastava S.C., 2007. An approach for optimal placement of static VAR compensators based on
reactive power spot price, IEEE Trans. Power Syst., Vol. 22, No.4, pp. 2021-2029.
Verma K. S.and Gupta H.O., 2006. Impact on real and reactive power pricing in open power market using unified power flow
controller, IEEE Trans. Power Syst., Vol. 21. No 1, pp. 365-371.
Wang H., Murillo-sanchez C.E., Zimmerman R.D. and Thomas R.J., 2007. On computational issues of market-based optimal
power flow, IEEE Trans. Power Syst., Vol. 22, No.3, pp. 1185-1193.
Xiao Ying, Song Y.H., Liu Chen-Ching and Sun Y.Z., 2003. Available transfer capability enhancement using FACTS devices,
IEEE Trans. Power Syst., Vol. 18. No.1, pp. 305-312.
Xie K. and Song Y.H., 2001. Power market oriented optimal power flow via an interior point method, IEE Proc.-Gener. Transm.
Distrib., Vol. 148, No. 6, pp. 549-556.
Yang Guang Ya et al, 2007. TCSC allocation based on line flow based equations via mixed integer programming, IEEE Trans.
Power Syst. Vol. 22, No 4, pp.2262-2269.
Yu X., Singh C, Jakovljevic S, Ristanovic D and Huang G., 2003. Total transfer capability considering FACTS and security
constraints, in: Proc. IEEE Conference (0-7803-8110-6/03/$17.00 ).
Yu Zuwei and Lusan D., 2004. Optimal placement of FACTS devices in deregulated systems considering line losses, Electrical
power and energy systems, vol. 26, pp. 813-819.
Zhao B., Guo C.X. and Cao Y.J., 2005. A multiagent-based particle swarm optimization approach for optimal reactive power
dispatch, IEEE Trans. Power Syst., Vol. 20, No. 2, pp. 1070-1078.
Zimmerman R.D. and Sanchez C.E.Murillo, MATPOWER, a Matlab power system simulation package, Version 3.2, 2007.
[Online] . Available http:// www. pserc.cornell.edu/matpower/.

Biographical notes:

S. K. Joshi received M. E. and Ph.D. from the M. S. University of Baroda and Indian Institute of Technology Kanpur, India, respectively. He is a Professor in the
Department of Electrical Engineering, M. S. University of Baroda, India. His research interests include power system restructuring, power system optimization &
control, voltage stability analysis and Artificial intelligence applications to power system. He is a member of IEEE.

K. S. Pandya received M.E. from L.D.College of Engineering, Gujarat, INDIA. Currently, he is pursuing Ph.D. from M.S.university of Baroda, INDIA. He is an
Associate Professor in the Department of Electrical Engineering, Charotar Institute of Technology, CHANGA, INDIA. His current area of research includes power
system restructuring, power system optimization, FACTS devices and artificial intelligence applications to power system. He is a member of IEEE.

Accepted March 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 82-93

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Optimal trading strategy for GenCo in LMP-based and bilateral markets
using self-organising hierarchical PSO

C. Boonchuay
1
, W. Ongsakul
1
*, J. Zhong
2
, F.F. Wu
2

1
Energy Field of Study, Asian Institute of Technology, THAILAND
2
Department of Electrical and Electronic Engineering, The University of Hong Kong, CHINA
*
Corresponding Author: e-mail: ongsakul@ait.asia, Tel +66-2-524-5398, Fax. +66-2-524-5439

Abstract

This paper proposes an optimal trading strategy for a generation company (GenCo) in multi-market environment including
day-ahead spot and long term bilateral contract markets using self-organising hierarchical particle swarm optimisation with
time-varying acceleration coefficients (SPSO-TVAC). The proposed trading strategy is formulated as a two-stage optimisation
problem. Firstly, the GenCos objective model which is to maximise expected profit and to minimise risk of profit variation is
solved by SPSO-TVAC. Secondly, the market clearing model which is to minimise system cost of locational marginal price
(LMP) based market is solved by DC optimal power flow (DCOPF). The Monte Carlo method is employed to simulate other
bidders behaviour in competitive environment. Test results on the PJM 5-bus system indicate that SPSO-TVAC is superior to
inertia weight approach particle swarm optimisation (IWAPSO) and genetic algorithm (GA) in searching the optimal trading
solution. In addition, different bilateral contract prices and spot demand significantly impact GenCos trading behaviour.
Accordingly, the proposed approach could be a beneficial decision-making tool for a GenCo in energy trading.

Keywords: Optimal bidding strategy, energy trading, locational marginal price, bilateral contract market, particle swarm
optimisation, Monte Carlo simulation.

1. Introduction

In energy market, generation companies (GenCos) should have strategies to maximise their profits and to avoid price risk
(Shahidehpour et al., 2002). For example, GenCos may allocate their power into several markets including spot, bilateral contract
or reserve markets. For trading in spot market, a GenCo could develop a bidding strategy to maximise its profit especially in
imperfect competitive environment. In fact, spot price could be influenced by market participants behaviour and demand
fluctuation. Therefore, hedging price risk with low risk transaction such as bilateral contract is necessary for GenCos.
Recently, there are a number of researchers investigating optimal bidding strategies in locational marginal price (LMP) based
markets. In Li and Shahidehpour (2005), a game-based bidding strategy for GenCos with incomplete information is proposed to
find the supply function equilibrium. However, with the complexity of the problem, the global optimal solution is difficult to be
found by this approach. In Gountis and Bakirtzis (2004), genetic algorithm (GA) and Monte Carlo (MC) simulation are applied to
provide the optimal bidding strategy. Different risk profiles of market participants are also considered. However, only single
period trading without technical constraints is considered. Finally, in Badri et al. (2008), a network-constrained bidding strategy
considering bilateral contracts based on the primal dual interior point method (PDIPM) is proposed. It is assumed that all players
cost functions are known and the equilibrium point is provided. However, it is virtually impossible to know such confidential
information.
With the complexity and large scale of the network-constraint bidding strategy problem, an efficient heuristic approach is
needed to provide the optimal solution. In Bajpai and Singh (2008), fuzzy adaptive particle swarm optimisation (FAPSO) is
proposed to solve the strategic bidding problem. The simulation results indicate that FAPSO is superior to GA in both searching
performance and execution time. However, a proper design of the fuzzy set is needed to adapt the velocity weight for the particular
problem.
Boonchuay et al./ International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 82-93

83

In Ratnaweera et al. (2004), self-organising hierarchical particle swarm optimisation with time-varying acceleration coefficients
(SPSO-TVAC) is proposed. It is applied to a non-convex economic dispatch (Chaturvedi et al., 2008). The results show that
SPSO-TVAC is efficient to solve a non-convex and non-differentiable function and its performance is superior to several other
particle swarm optimisation (PSO). In addition, the velocity weight coefficient which is a key parameter of PSO is eliminated.
Thus, it is applicable to solve various optimisation problems.
This paper proposes an optimal trading strategy for a GenCo in multi-market environment including day-ahead LMP-based and
long term bilateral contract (Hou and Wu, 2007) markets using a novel stochastic search technique, SPSO-TVAC. The trading
strategy is formulated as a two-stage optimisation problem. The first subproblem is the GenCos objective model, maximising
expected profit from trading in the both markets and minimising risk of profit variation. The second is the market clearing process
based on DCOPF to minimise system cost. MC simulation is employed to deal with the uncertainties from other bidders
behaviour in spot market. The proposed strategy is illustrated on the PJM 5-bus system and compared to the well-known stochastic
search approaches, inertia weight approach particle swarm optimisation (IWAPSO) and GA.
The rest of the paper is organised as follows: Section 2 expresses problem formulation including the GenCos objective model
and the market clearing model. Section 3 describes the solution methodologies including SPSO-TVAC and MC simulation, and
the solution algorithm. Section 4 illustrates the proposed strategy through numerical examples and discussions. Finally, Section 5
concludes the paper.

2. Problem Formulation

2.1 GenCos objective model: To maximise the expected profit for a GenCo from spot and bilateral contract markets and to
minimise risk of profit variation, the GenCos objective model could be expressed as

( )
1 1
max E
T N
d b
nt nt nt
t n

= =
( +

(1)

min max
s.t. ,
d b
n nt n nt nt nt
q q q q q q = + (2)
, if 0
nt n nt
q UR q > (3)
, if 0
nt n nt
q DR q <
(4)
In (1), the multiple objectives of a GenCo are considered as the single objective function. Generally, for conflicting objectives,
the tradeoff technique could also be implemented as
( ) 1 2
E
d b
nt nt nt
w w ( +

(Boonchuay and Ongsakul, 2009). Here, it is
assumed that all GenCos objectives are equally weighted so that all weight coefficients could be set to one. The technical
constraints, which are the maximum and minimum generation limits in (2), and ramping up/down limits in (3) and (4), are
considered. The minimum up/down time constraints could also be included if some units need to be shut off.
The detail of the GenCos objective model in (1) is written as

1
1
E[ ] ( , )
d M
d d d d b nmt
nt mt nmt nt nmt nt d b
m nmt nt
q
q c q q
M q q

=
| |
=
|
+
\
(5)
( , )
d
d d d d b nmt
nmt mt nmt nt nmt nt d b
nmt nt
q
q c q q
q q
=
+
(6)
( , )
b
b b b d b nt
nt t nt nt nt nt d b
nt nt
q
q c q q
q q
=
+
(7)
( )
2
1
1
E[ ]
M
d d
nt nmt nt
m
M

=
=
(8)

In (9) to (10), the cost function of unit i at period t is described as (Bajpai and Singh, 2007)

( ) ( )
2
0 1 2 3 4 min
( ) sin
nt nt n n nt n nt n n n nt
c q c c q c q c c q q = + + + (9)
d b
nt nt nt
q q q = + . (10)


84

2.2 Market clearing model using DCOPF: In LMP-based market, the system operator generally provides electricity price with
minimum system cost considering transmission limits. ACOPF is an efficient tool to determine electricity price and quantity for
each bus. However, ACOPF requires a long execution time to obtain the optimal solutions. Therefore, DCOPF is used to provide
market clearing price and quantity, which can be expressed as
1
min
I
i i
i
p q
=

(11)
s.t.
1 1
0
I I
i i Loss
i i
q L P
= =
=

(12)

k k
F Limit , for k = 1, 2,, K (13)
min max i i i
q q q , for i = 1, 2,, I (14)

where the system loss
Loss
P is calculated by
2
1
K
k k
k
F R
=

, and the line flow

k
F is provided by
1
( )
I
k i i i
i
GSF q L
(Wood and
Wollenberg, 1996).
In the DCOPF model, the entire system loss will be compensated by the generator at the reference bus. Therefore, for more
accurate LMP calculation, the fictitious nodal demand (FND) is included in the line flow equation as (Li and Bo, 2007)
1
( )
I
k k i i i i
i
F GSF q L E
=
=
. (15)
In (15),
i
E is the FND at bus i, which can be calculated by
2
1
1
2
i
K
i k k
k
E F R
=
=
. (16)
LMP at the reference bus is equal to the Lagrangian multiplier of the equality constraint in (12). For providing LMP at the other
buses, the inequality constraint Lagrangian multiplier and the marginal loss factor have to be determined (Li and Bo, 2007).

3. Solution Methodologies

3.1 Self-organising hierarchical PSO with time-varying acceleration coefficients: SPSO-TVAC is an efficient population-based
optimisation technique, which is appropriate for non-convex optimisation problems (Chaturvedi et al., 2008). Mathematically, it is
defined as

( ) ( )
1
1 1 2 2
k k k k k k k
id id id gd id
v a rand p x a rand p x
+
= +
(17)
1 1 k k k
id id id
x x v
+ +
= +
. (18)
In (17), the first component is the cognitive component that represents the individual experience of each particle. The last
component is the social collaboration of the particles in finding the global optimal solution. The position of the ith particle on the
dth dimension is updated by (18). With the TVAC concept, the acceleration coefficients are given as
1 1
1 1
max
i f k
i
a a
a a k
k
=
(19)
2 2
2 2
max
.
i f k
i
a a
a a k
k
=
(20)
In addition, SPSO-TVAC has a reinitialised process to enhance its performance when particles stagnate during the search. The
pseudocode is shown as follows.


85

Main procedure
Velocity update equation in (17)
if vid=0
if rand3<0.5
vid=rand4*vdmax
else vid=-rand5*vdmax
end if
end if
vid=sign(vid)*min(abs(vid,vdmax))
Position update equation in (18)
v
id
is the velocity of the ith particle on the dth dimension, and v
dmax
is the maximum velocity limit on the dth dimension, which
could be set as 10% of the dynamic range of the variable on each dimension (Chaturvedi et al., 2008).

3.2 Monte Carlo simulation for bidding strategy: MC simulation is a stochastic computational technique which is performed by
statistical sampling experiments on a computer. It needs a probability density function (PDF) of uncertainty source to generate the
sampling. For strategic bidding problems, the uncertainties come from bidding behaviour of market participants. A PDF that
appropriately represents the distribution of historical bidding behaviour could be used. However, in this paper, a normal PDF is
employed to estimate rivals bidding behaviour, which can be expressed as (Bajpai and Singh, 2007)
2
2
( ) 1
( ) exp
2( ) 2
r r
r n n
n r r
n n
p
pdf p

| |
=
|
\
. (21)
3.3 Solution algorithm: The proposed solution algorithm based on PSO and MC simulation for the strategic trading problem is
described in 10 steps as shown in Figure 1.

Import input data, i.e., gen.,
rivals, lines, load, contract price
Start
Initialise decision variables
Specify iter_max and mc_max
Set mc = 1
Generate rivals bid prices
Provide market settlement Calculate GenCos profit
7) mc < mc_max
Evaluate fitness function
Define best particles
Update particles
10) iter < iter_max
End
mc = mc + 1
No
No
iter = iter + 1
1
2
3
4
5 6
8
9

Figure 1. Solution procedure

Step 1) Input data: The concerned GenCos cost data, historical rivals bidding data, transmission lines, load, and bilateral contract
price are given.
Step 2) Initial value: Decision variables including bid price and quantity, and amount of bilateral transaction are randomly
initialised. The maximum number of PSO iterations and MC trials are specified.
Step 3) Reset MC counter: The MC counter is set/reset to be 1.
Step 4) Rivals strategy estimation: The rivals bid prices are generated based on their PDF.
Step 5) Market settlement process: The DCOPF model in (11) to (16) is solved by a linear programming approach to provide LMP
and dispatched quantity for each bus.

86

Step 6) Profit evaluation: GenCos profit from trading in spot market is calculated using (6). This finishes one trial of MC
simulation.
Step 7) MC loop stopping criteria: If the maximum number of the simulation is not reached, go to Step 4 and the MC counter is
increased by 1. Otherwise go to the next step.
Step 8) Evaluating fitness value: The fitness function including the objective function in (1) and the penalty function (PF) is
evaluated by
( )
1 1
E
T N
d b
nt nt nt
t n
Fitness PF
= =
( = +

(22)
where PF is a constant that is larger than zero if the constraints in (2) to (4) are not satisfied. Otherwise PF equals to 0.
Step 9) Updating particles: The best fitness particles, global and local best particles, are defined and all particles are updated by
(17) and (18).
Step 10) PSO loop stopping criteria: If the PSO counter is less than maximum number of the PSO iterations, go to Step 3 and
increase the PSO counter by 1. Otherwise stop.

4. Numerical Example and Discussion

The proposed optimal trading strategy is tested on the PJM 5-bus system. The optimisation problem in (1) to (4) is solved by
different optimisation techniques including SPSO-TVAC, IWAPSO, and GA. In addition, the DCOPF model in (11) to (16) is a
suboptimisation problem. As the simulation is performed on the MATLAB environment, the DCOPF model is conveniently solved
by fmincon function included in the software package. To reduce the computing effort, an artificial neural network (ANN) is
used to approximate the LMP. The simulation is performed on a 2.8 GHz Pentium IV personal computer with 1GB RAM.
The configuration and parameters of the PJM 5-bus system are shown in Figure 2 and Table 1, respectively. There are four
GenCos participating in the market. In this example, the concerned GenCo, GenCo-A, needs to provide the optimal bid price, bid
quantity, and amount of bilateral transaction to maximise expected profit and to minimise risk of profit variation. Therefore, the
rest GenCos are the opponent bidders and the bidding data is given in Table 2.
The production cost function of GenCo-A is expressed in (9). The cost coefficients and the technical parameters are given in
Table 3.

Figure 2. PJM 5-bus system

Table 1. Transmission line data of PJM 5-bus system
A-B A-D A-E B-C C-D D-E
R(%) 0.281 0.304 0.064 0.108 0.297 0.297
X(%) 2.810 3.040 0.640 1.080 2.970 2.970
Limit 999 999 999 999 999 240


87

Table 2. Rivals bidding data
GenCo
Mean of bid price,
$/MWh
SD. of bid price,
$/MWh
Bid quantity, MW
C 30 4 520
D 35 5 200
E 10 3 600

Table 3. GenCo-As data
c
0
c
1
c
2
c
3
c
4

q
max
,

MW
q
min
,
MW
UR,
MW
DR,
MW
0.00482 7.97 78 150 0.063 210 20 30 50

In the simulation, it is assumed that there is no demand elasticity and each load bus equally shares the spot demand. In addition,
there is a large customer at Bus-D who prefers to purchase power from GenCo-A through bilateral transaction at specific prices
and the bilateral congestion charge is ignored.

4.1 Solution convergence comparison: Three different global search approaches including SPSO-TVAC, IWAPSO, and GA are
employed to provide the optimal trading strategy. For the PSO parameters, the maximum number of iterations is taken as 300 with
the swarm size of 20 particles. For SPSO-TVAC, the coefficient of cognitive component (a
1
) is decreased in the range of 2.5 to 0.5
and the social learning factor (a
2
) is increased from 0.5 to 2.5. For IWAPSO, the inertia weight is varied from 0.9 to 0.4 and the
acceleration coefficients are equal to 2. For the GA parameters, the maximum number of generation, population size, and mutation
and crossover probabilities are used as 300, 20, 0.2, and 0.6, respectively.
The single period trading with spot demand of 900 MW and bilateral contract price of $10/MWh is used to compare the
performances of the optimisers in the strategic trading problem. The 100-trial MC simulation is performed to simulate the rivals
behaviour. By 20 runs of the simulation, the convergence characteristics of the best solutions from SPSO-TVAC, IWAPSO, and
GA are compared in Figure 3.
0 50 100 150 200 250 300
1300
1350
1400
1450
1500
Iterations
F
i
t
n
e
s
s

v
a
l
u
e

(
$
)

SPSO-TVAC
IWAPSO
GA

Figure 3. Convergence characteristics of different optimisers

In the beginning of the search, the SPSO-TVAC and IWAPSO fitness values dramatically increase, while the GA fitness value
slightly increases. After the 50th iteration, SPSO-TVAC provides higher fitness solutions than IWAPSO and GA. An advantage of
SPSO-TVAC is that the reinitialised process will be active when a particle stagnates during the search. Therefore, all particles are
always encouraged to find a better solution. At the last iteration, the SPSO-TVAC fitness value is $1,478.8, whereas the IWAPSO
and GA fitness values are $1,461.2 and $1,410.5, respectively. For the execution time, the PSO-based approaches require
approximately 60 minutes, but the GA approach needs around 80 minutes to provide the optimal solution.

4.2 LMP approximation using ANN: A feed-forward ANN with 15 hidden nodes and 3 hidden layers is used to find LMP. The
ANN has been trained by 30,000 input/output patterns from the DCOPF solutions. The ANN diagram for the LMP approximation
is shown in Figure 4. In this case, a single period trading with spot demand of 900 MW and bilateral contract price of $10/MWh is
considered. SPSO-TVAC is employed to maximise the GenCos objective model. The solutions from the DCOPF and the ANN-
based LMP approximations are shown in Table 4.


88

Figure 4. Diagram of the ANN-based LMP approximation

Table 4. Comparative results from different LMP approximation approaches
100 trials of MC 500 trials of MC
DCOPF ANN DCOPF ANN
Bid price, $/MWh 15.26 16.54 15.48 16.65
Bid quantity,MW 84.38 83.84 84.46 83.83
Bilateral transaction, MW 115.43 115.64 114.95 115.67
Expected LMP at Bus-A, $/MWh 29.68 29.67 29.60 29.65
Expected profit from spot,$ 1716.52 1705.54 1713.16 1703.51
Profit from bilateral,$ 76.18 78.35 77.65 78.37
Risk of profit variation,$ 313.81 324.64 332.18 327.49
Execution time, min 53.12 2.51 255.48 14.35
Computer Configuration: 2.8 GHz, PIV Processor, 1GB RAM

In Table 4, the decision variables of GenCo-A are bid price, bid quantity and amount of bilateral transaction. The other
parameters are state variables. For the ANN approach with 100-trial MC simulation, on average, the decision variables differ
3.06% from the DCOPF solutions while the other parameters differ 1.74%. Using 100-trial MC simulation, the ANN approach
requires 2.51 minutes, whereas the DCOPF approach needs 53.12 minutes to provide the optimal solution. For the ANN approach
with 500-trial MC simulation, the decision and the state variables differ from the DCOPF solutions 2.98% and 0.76%,
respectively. Using 500-trial MC simulation, the execution times of the ANN and the DCOPF approaches are 14.35 minutes and
255.48 minutes, respectively. The results indicate that the ANN approach could significantly reduce the computing effort in
providing the optimal trading strategy.
Note a higher number of MC simulation could provide more accurate solutions. However, it is necessary to consider the
execution time. When MC simulation is taken as 100 trials, the execution time is reduced by 5 times, but the solutions are not as
diverse as the 500-trial solutions.

4.3 Different bilateral contract prices: In this case, the optimal trading strategies for GenCo-A with different bilateral contract
prices are illustrated. The bilateral contract price at Bus-D varies from $15 to $28/MWh with the spot demand of 900 MW. The
optimal solutions are shown in Table 5.

Table 5. Optimal solutions with different bilateral contract prices
Bilateral contract price $/MWh
15 20 25 28
Bid price, $/MWh 16.94 16.44 15.50 15.28
Bid quantity, MW 83.83 83.69 80.42 2.26
Bilateral transaction, MW 115.66 126.22 129.56 207.24
Expected LMP at Bus-A, $/MWh 29.67 29.62 29.62 29.49
Expected profit from spot, $ 1705.31 1660.93 1594.36 54.54
Profit from bilateral, $ 656.69 1288.54 1970.42 3773.51
Risk of profit variation, $ 324.47 324.44 311.20 18.48

In Table 5, the optimal solutions for GenCo-A with different bilateral contract prices are compared. The bid price and the
expected LMP at Bus-A are quite stable even the bilateral contract price varies. Furthermore, the relation between the allocated
power in each market and the bilateral contract price is non-linear as shown in Figure 5. When the bilateral contract price is less
than $24/MWh, GenCo-A keeps the same amount of allocated power in each market. However, when the bilateral contract price
increases to $27/MWh, GenCo-A might allocate the entire power to the bilateral contract market.
Bid prices and
quantities from
GenCos

ANN
Load
LMP
Dispatched
quantities

89

16 18 20 22 24 26 28
0
50
100
150
200
Bilateral contract prices ($/MWh)
D
i
s
p
a
t
c
h
e
d

p
o
w
e
r

(
M
W
)

Total dispatched power
Allocated power in spot market
Alloacted power in bilateral market

Figure 5. GenCo-As trading behaviour versus different bilateral contract prices
In Figure 6, GenCo-As expected profit and risk of profit variation from spot market, and profit from bilateral transaction with
different bilateral contract prices are shown. Varying the bilateral contract prices from $15 to $24/MWh, the expected profit is not
diverse because the expected LMP and dispatched quantity at Bus-A are rather flat as shown in Table 5. In bilateral contract
market, as the price increases, the profit is linearly growing. Nevertheless, when the bilateral contract price is higher than
$25/MWh, profit from the both markets will sharply changes. The risk correlates with the expected profit from spot market
trading, which is close to zero when the bilateral contract price is equal to $28/MWh.

16 18 20 22 24 26 28
0
500
1000
1500
2000
2500
3000
3500
4000
Bilateral contract prices ($/MWh)
P
r
o
f
i
t

(
$
)

Spot market trading
Bilateral trading
Risk

Figure 6. Profit and risk of GenCo-A versus different bilateral contract prices

4.4 Multi-hourly trading strategy: Day-ahead spot and bilateral contract markets are considered to provide the optimal bid price
and power allocation for GenCo-A. The day-ahead spot demand is shown in Figure 7. It is assumed that the bilateral contract price
at Bus-D is $15/MWh for the entire trading periods.

3 6 9 12 15 18 21 24
0
200
400
600
800
1000
1200
1400
Hours
D
e
m
a
n
d

(
M
W
)

Figure 7. Forecast demand for day-ahead spot market

90

In Figure 8, the allocated power in spot and bilateral contract markets and total dispatched power of GenCo-A for 24-hour
trading period are shown. The results indicate that different spot demands cause strategy changes of GenCo-A. However, in entire
trading periods, GenCo-A allocates a higher energy proportion in bilateral contract market to hedge price risk. The average
dispatched output of GenCo-A is 199.58 MW. With the different demands, dispatched levels of GenCos vary as shown in Table 6.
2 4 6 8 10 12 14 16 18 20 22 24
0
50
100
150
200
Hours
P
o
w
e
r

(
M
W
)

Total dispatched power
Allocated power in spot market
Allocated power in bilateral market

Figure 8. Total dispatch and allocated power of GenCo-A in day-ahead spot and bilateral contract markets

Table 6. Expected power dispatch of each GenCo with different spot demands
GenCo
dispatch, MW
Spot demand, MW
900 1000 1100 1200 1300
A 84.39 68.13 51.84 57.23 91.78
C 186.35 300.88 415.40 486.86 499.71
D 35.41 36.72 38.20 61.44 114.60
E 600.00 600.00 600.00 600.00 600.00

In Figure 9, the expected LMP at Bus-A and the optimal bid price of GenCo-A for 24-hour trading period are shown. During
peak demand periods of 1300 MW, the expected LMP at Bus-A increases to $34.94/MWh, whereas the lowest LMP, with spot
demand of 900 MW, is $29.68/MWh. The optimal bid prices of GenCo-A are, on average, 58.11% lower than the LMPs.

2 4 6 8 10 12 14 16 18 20 22 24
10
15
20
25
30
35
Hours
P
r
i
c
e

(
$
/
M
W
h
)

Expected LMP
Optimal bid price

Figure 9. Optimal bid price and expected LMP at Bus-A

In Figure 10, the expected profit and risk of profit variation from spot market, and profit from bilateral transaction of GenCo-A
for 24-hour trading period are shown. For all trading periods, the cumulative expected profit and risk of profit variation from spot
market are $40,722 and $7,060, respectively. Profit from bilateral trading depends on amount of sold energy, which the cumulative
bilateral profit is equal to $16,969.


91

2 4 6 8 10 12 14 16 18 20 22 24
0
500
1000
1500
2000
2500
3000
Hours
P
r
o
f
i
t

(
$
)

Spot market trading
Bilateral trading
Risk

Figure 10. Profit and risk of GenCo-A in day-ahead spot and bilateral contract markets

5. Conclusion

In this paper, SPSO-TVAC and MC simulation are used to provide the optimal trading strategies for a GenCo in LMP-based and
long term bilateral contract markets. The proposed trading strategy includes two optimisation stages. The first stage is the GenCos
objective model, which is to maximise expected profit in spot and bilateral contract markets and to minimise risk of profit
variation, solved by SPSO-TVAC. The second stage is the market clearing model based on DCOPF to provide the LMP. The
uncertainties from other bidders behaviour are estimated by the MC approach. Test results indicate that SPSO-TVAC could
provide a better solution than IWAPSO and GA in the strategic trading problem. The ANN-based LMP approximation approach is
effective to reduce the computing effort in the optimisation process. With different bilateral contract prices and spot demands, the
proposed approach could be used as a decision-making tool for a GenCo in energy trading.

Nomenclature

1
k
a ,
2
k
a Acceleration coefficients of cognitive and social components at iteration k, respectively.
1i
a
,
2i
a Initial values of cognitive and social acceleration coefficients, respectively.
1 f
a ,
2 f
a Final values of cognitive and social acceleration coefficients, respectively.
0n
c ,
1n
c ,
2n
c
Production cost coefficients of unit n.
3n
c ,
4n
c
Constants of the valve point loading effect of unit n.
( )
nt
c
Production cost function of unit n at period t ($).
n
DR
Ramping down limit of unit n (MW).
E[ ]
d
nt
Expected profit from trading in spot market of unit n at period t ($).
k
F
Line flow at line k (MW).
k
F
Line flow at line k from the previous iteration of the DCOPF calculation.
k i
GSF
Generator shift factor of line k and bus i.

I
Number of buses.
k
Iteration index.
max
k Maximum number of iterations.
K
Number of transmission lines.
i
L
Spot market load at bus i (MW).
k
Limit Line flow limit at line k (MW).
M
Maximum number of rivals strategies.
N
Number of generation units of the concerned GenCo.
i
p Bid price at bus i ($/MWh).
r
n
p
Bid price of unit n of rival r.
k
id
p Best position of particle i on dimension d reached at iteration k.
k
gd
p Global best particle on dimension d reached at iteration k.
Loss
P
System loss (MW).
i
q
Generation dispatch at bus i (MW).
nt
q Generation output of unit n at period t (MW).
maxi
q Maximum generation dispatch at bus i (MW).

92

mini
q
Minimum generation dispatch at bus i (MW).
max n
q
Maximum generation limit of unit n (MW).
minn
q Minimum generation limit of unit n (MW).
d
nt
q Bid quantity in spot market of unit n at period t (MW).
b
nt
q Sold power through bilateral contract of unit n at period t (MW).
d
nt
q Expected dispatched quantity in spot market of unit n at period t (MW).
d
nmt
q
Dispatched quantity in spot market of unit n for the mth rivals strategy at period t (MW).
nt
q
Generation change from the previous period of unit n at period t (MW).
i
rand
Random number between 0 and 1, i = 1, 2,, 5.
k
R Resistance at line k (pu).
T
Maximum number of trading periods.
n
UR Ramping up limit of unit n (MW).
k
id
v
Velocity of particle i on dimension d at iteration k.
k
id
x Position of particle i on dimension d at iteration k.
b
nt
Profit from bilateral transaction of unit n at period t ($).
d
nmt
Spot market profit of unit n for the mth rivals strategy at period t ($).
nt
Standard deviation of profit of unit n at period t ($).
r
n
,
r
n
Mean value and standard deviation of
r
n
p
, respectively.
b
t
Bilateral contract price at period t ($/MWh).

d
mt
Spot price at the concerned bus for the mth rivals strategy at period t ($/MWh).

Acknowledgement

This research is supported by the Office of the Higher Education Commission, under the Ministry of Education, Thailand, and
the Energy Policy and Planning Office (EPPO), under the Ministry of Energy, Thailand.

References

Badri A., Jadid S., Rashidinejad M. and Moghaddam M.P., 2008. Optimal bidding strategies in oligopoly markets considering
bilateral contracts and transmission constraints. Electric Power Systems Research, Vol. 78, No. 6, pp. 1089-1098.
Bajpai P. and Singh S.N., 2007. Fuzzy adaptive particle swarm optimization for bidding strategy in uniform price spot market.
IEEE Transactions on Power Systems, Vol. 22, No. 4, pp. 2152-2160.
Bajpai P. and Singh S.N., 2008. Strategic bidding in network constrained electricity markets using FAPSO. International Journal
of Energy Sector Management, Vol. 2, No. 2, pp. 274-296.
Boonchuay C. and Ongsakul W., 2009. A Risk-Constrained Optimal Bidding Strategy for a Generation Company by IWAPSO.
Proceedings of IEEE Bucharest PowerTech, Romania.
Chaturvedi K.T., Pandit M. and Srivastava L., 2008. Self-organizing hierarchical particle swarm optimization for nonconvex
economic dispatch. IEEE Transactions on Power Systems, Vol. 23, No. 3, pp. 1079-1087.
Gountis V.P. and Bakirtzis A.G., 2004. Bidding strategies for electricity producers in a competitive electricity marketplace. IEEE
Transactions on Power Systems, Vol. 19, No. 1, pp. 356-365.
Hou Y. and Wu F.F., 2007. Long-term bilateral contract pricing with risks of congestion charge. Proceedings of IEEE PES
General Meeting, pp 1-6.
Li F. and Bo R., 2007. DCOPF-based LMP simulation: Algorithm, comparison with ACOPF, and sensitivity. IEEE Transactions
on Power Systems, Vol. 22, No. 4, pp. 1475-1485.
Li T. and Shahidehpour M., 2005. Strategic bidding of transmission-constrained GENCOs with incomplete information. IEEE
Optimization toolbox, available at: www.mathworks.com.
PJM Training MaterialsLMP 101, available at: www.pjm.com.
Ratnaweera A., Halgamuge S.K. and Watson H.C., 2004. Self-organizing hierarchical particle swarm optimizer with time-varying
acceleration coefficients. IEEE Transactions on Evolutionary Computation, Vol. 8, No. 3, pp. 240-255.
Shahidehpour M., Yamin H. and Li Z., 2002. Market Operations in Electric Power Systems. Wiley Interscience, New York.
Wood A.J. and Wollenberg B.F., 1996. Power Generation, Operation, and Control. 2nd edition, Wiley Interscience, New York.


93

Biographical notes

Chanwit Boonchuay received the M.Eng. degree in electrical engineering from King Mongkut's Institute of Technology North Bangkok, Thailand, in 2005. He is
currently pursuing the D.Eng. degree in electric power system management at Asian Institute of Technology, Thailand. He spent two months as a visiting scholar at
the University of Hong Kong in 2009. At present, he is a visiting scholar at the University of Tennessee, Knoxville. His current interests include power system
optimisation, AI applications to power systems, and power system restructuring and deregulation.

Weerakorn Ongsakul received the M.S. and Ph.D. degrees in electrical engineering from Texas A&M University, College Station, in 1991 and 1994, respectively.
He is currently an Associate Professor of Energy and Dean of School of Environment, Resources, and Development, Asian Institute of Technology, Thailand. His
current interests are in parallel processing applications, power system optimisation, AI applications to power systems, and power system restructuring and
deregulation.

Jin Zhong received the B.Sc. degree from Tsinghua University, Beijing, China, in 1995, the M.Sc. degree from China Electric Power Research Institute, Beijing,
in 1998, and the Ph.D degree from Chalmers University of Technology, Gothenburg, Sweden, in 2003. At present, she is an Assistant Professor in the Department
of Electrical and Electronic Engineering at the University of Hong Kong. Her areas of interest are electricity sector deregulation, ancillary service pricing, and
power system planning.

Felix F. Wu received the Ph.D. degree from the University of California at Berkeley (UCB). He is currently the holder of Philip K H Wong Wilson K L Wong
Professor in Electrical Engineering and Chair Professor of Electrical Engineering at the University of Hong Kong (HKU). He is emeritus Professor of UCB and
Director of Center for Electrical Energy System at HKU and Director of HKU-Tsinghua Shenzhen Power System Research Institute in Shenzhen, China. His
research interests are electric energy industry restructuring, power system investment planning, design of modern control centers, distribution automation,
distributed processing.

Received October 2009
Accepted February 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 94-106

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Allocation of optimal distributed generation using GA for minimum system
losses in radial distribution networks

T. N. Shukla
1*
, S.P. Singh
2
, K. B. Naik
3

*1
Department of Electrical Engineering, Kamla Nehru Institute of Technology, Sultanpur, (U. P.), INDIA
2
Department of Electrical Engineering, Institute of Technology, Banaras Hindu University, Varanasi (U. P.), INDIA
3
Director, College of Engineering Science and Technology, Lucknow (U. P.), INDIA
*
Corresponding Author: e-mail: tns_shukla55@yahoo.com, tns.shukla@gmail.com,Tel. 05362-226278, Mob. +919415962452

Abstract
The distributed generation (DG) is one of the viable options for mitigation of problems of load growth, overloading of lines,
quality of supply and reliability in tern extending equipment maintenance intervals and to reduce line losses. However, the line
loss reduction is the obvious parameter easily expressible in terms of system parameters. Therefore, this paper aims to minimize
active power loss by placing DG strategically in a radial distribution system. The problem is formulated as an optimization
problem and solution is obtained using genetic algorithm (GA). The strategic locations are decided on the basis of loss
sensitivity to active power injection at various nodes. This approach helps in reducing the computational efforts of selecting
appropriate location(s). The performance of the method is tested on 33-bus test system and comparison of the results with a
reported method reveals that the proposed method yields superior results. In addition, long term economic benefit of optimal DG
placement is also demonstrated.

Keywords: distributed generation, line loss reduction, optimal location, radial distribution networks.

1. Introduction

Earlier the distribution networks had been designed to convey electrical energy from high voltage transmission networks,
whereby the majority of electrical generation plants were connected, to the customers (Jenkins, 1996). By the addition of
embedded generation plants to distribution networks, voltage profile of the network is improved and energy demand decreases
from the utility network. Also, passive distribution networks are transformed into active networks after embedded units have been
added to them (Ijumba et al., 1999; Milanovic and David, 2002).
Traditionally load growth is forecasted by distribution companies until a predetermined amount is reached, whereby a new
capacity must be added to the network. This new capacity is usually the addition of new substations or expanding existing
substations capacities and their associated new feeders or both. However, the flexibility, technologies, technical & monetary
benefits and concepts of DG planning is challenging this state of matter and gaining credibility as a solution to the distribution
planning problems with the prohibitively high cost of power curtailment in the changing regulatory and economic scenarios and
enhancing DG as an attractive distribution planning option that avoids causing degradation of power quality, reliability and control
of the utility systems (Dugan et al.,2001 and Rackliffe, 2000).
Quinta et al. (1993) and Khator and Leung (1997) reported that the distribution system planning problem is to identify a
combination of expansion projects for the least cost network investment that satisfies load growth requirements without violating
any system and operational constraints. The DG benefits (Daly and Morisson, 2001 and Chiradega and Ramkumar ,2004) are
numerous and the reasons (Jenkisin,2000) for implementing DGs are an energy efficiency or rational use of energy, deregulation
or competition policy, diversification of energy sources, availability of modular generating plant, ease of finding sites for smaller
generators, shorter construction time and lower capital costs of smaller plants and proximity of the generation plant to heavy loads,
which reduces transmission costs. Distributed generation (DG) can offer an alternative planning approach to utilities to satisfy
demand growth and distribution network security, planning and management issues. As DG may provide many benefits for
distribution network operators that can choose where to place it, as well as controlling its operating pattern through peak load
Shukla et al. / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 94-106

95

operation, the recognition of DG deferment benefits [by Piccolo and Siano (2009)] may influence the optimal connection of new
generation within existing networks. Brown et al. (1997) suggested that proper siting of DG may defer T&D expansion. Also it is
accepted by many countries in Kyoto Protocol (1997) that the reduction in gaseous emissions (mainly CO
2
) offered by DGs is
major legal driver for DG implementation.
DG are increasingly becoming an attractive alternative to network reinforcement and expansion for their small size, low
investment cost, modularity and ability to exploit renewable energy sources and due to their high efficiency. Numerous studies
used different approaches to evaluate the benefits from DGs to a network in the form of loss reduction and loading level reduction.
In the optimization process, Daz-Dorado et al. (2002) proposed to solve MV networks with ring, clasp, and inter-connective
configurations, considering investment and losses costs taking into account the constraints of conductor capacities and voltage
drop. The paper of Mardaneh and Gharehpetian (2004) deals with the task of finding the optimal siting and sizing of distributed
generation (DG) units for a given distribution network to minimize the cost of active and reactive power generation. Celli et al
(2005) proposed a multi-objective formulation based on the application of GA for the siting and sizing of DG resources into
existing distribution networks. They considered the power loss cost and the cost of service interruptions (i.e. number and duration)
as the master objective function. GA based optimization technique (which can give near optimal results), suitable for multi-
objective problems like DG allocation with optimal power flow (OPF) calculations has been used by Silvestri et al. (1999). A
hybrid GA-OPF approach was proposed by Harrison et al. (2008) for finding optimal location for connecting a predefined numbers
of DGs in a distribution network. Jabr and Pal (2009) presented an ordinal optimization (OO) method for specifying the locations
and capacities of distributed generation (DG) such that a trade-off between loss minimization and DG capacity maximization is
achieved. Acharya et al (2006) suggested a heuristic method to select appropriate location and optimal value of DG capacity for
minimum real power losses of the system by calculating DG size at different buses. Though the method is effective in selecting
location, it requires more computational effort. The heuristic method used to calculate DG size is based on approximate loss
formula and it may lead to an inappropriate solution.
Size and location of DG are crucial factors in the application of DG for its maximum benefits. In the present work, GA based
technique has been proposed and an attempt has been made to determine optimal DG size for minimization of real power losses in
a radial distribution network and its monetary benefits. The number of candidate locations is decided from the results of loss
sensitivity obtained by injecting real power into the network.


The installation of DG units at non-optimal places may not result as beneficial as it could have been otherwise. Since the impacts
of distributed generation on system performance depend on system operating conditions and the characteristics of the distributed
generation, it is necessary to use some solutions in planning and operation to attain the best performance. In large distribution
systems to select best place(s) for installation of optimal size DG units is a complex combinatorial optimization problem.

2.1. Selection of Location
In order to reduce the efforts to select appropriate nodes for placement of distributed generation devices a priori, sensitiveness of
nodes towards the change in active power loss with respect to change in active power injection at various nodes are identified. The
loss sensitivity factors at different buses have been evaluated to select appropriate nodes for DG planning by using load flow
program suitable to radial networks suggested by Afsari et al. (2002). These sensitivity factors reflect how the feeder power losses
change if more real power is injected at a particular node and it also allows obtaining the candidate nodes to locate DG.
Loss sensitivity factors are evaluated for the base case first to decide the first appropriate location. In order to select next
candidate location(s), successive sensitivity analysis (computation of loss sensitivity of various buses taking into account the
previously selected location(s) at which DG could be placed) has been used. This process is repeated till optimum number of
locations is ascertained. The optimal number of locations is that number of buses which gives maximum benefits (saving) for
optimal size DGs placed at selected locations. The expression for line losses describe by Elgard (1971) has been used for the
purpose. The change in active power loss of the system due to change in active power injection at a node is expressed as
( )
=
=
n
j
i ij i ij
i
Loss
Q P
P
P
1
2 (1)
2.2. Loss Minimization
Among the many benefits of distributed generation, reduction in system line losses is one of them. Nevertheless, reactive power
loss is obviously not less important. This is due to the fact that reactive power flow in the system needs to be maintained at a
certain amount for sufficient voltage level. Consequently, reactive power makes it possible to transfer real power through
transmission and distribution lines to customers. Normally, the real power loss draws more attention for the utilities, as it reduces
the efficiency of transmitting energy to customers. System loss reduction by strategically placed DG along the network feeder, can
be very useful if the decision maker is committed to reduce losses and to improve network performance (e.g. on the level of losses
and/or reliability) maintaining investments to a reasonable low level. This feature may be very useful in case of revenue recovered

96

by distribution company (DISCO) which is not only based on the asset value but also on network performance. So in the present
work, real power injection at selected location(s) through DGs is considered to achieve the objective.
An established loss formula; Elgard (1971) based on loss coefficients and is represented in Eq. (2) and has been used for the
calculation of real power loss in the present work.
( ) ( )
j i j i ij j i j i
n
i
n
j
ij Loss
Q P P Q Q Q P P P + + =
= =

1 1
(2)
The objective function is to minimize the active power loss using Equation (10) and formulated as to

Minimize

=
=
LN
k
k Loss
Loss P
1
(3)
Subject to

( )
Loss i
n
i
n
i
Gi
P P P +

= = 1 1
(4)

max min
i i i
V V V

(5)

scheduled j i j i
I I
(6)

2.3. Demand Curve
Energy losses are usually accounted per year. Since constant loading condition of a distribution system is not realistic and
variation of load demand can also not be predicted, load duration curve can be constructed using the demand curve data and can be
approximated in discrete levels. Day to day demand curves vary as per the demand of loads. Seasonal variations, social
commotion, economic and environmental aspects also dictate the changes in demand curve. Since the final solution depends on
proper choice of demand curve, a careful analysis is required. In order to achieve the objective, a steady demand curve is
considered in this work wherein annual demand curve is approximated by 360 daily demand variation curves and is used to
compute energy loss. The piecewise linear load duration curve is assumed to include the effect of varying loads () and is divided
into three load levels as average (L1), base (L2) and peak load (L3) conditions of 62.5%, 100% and 125% of scheduled system
loads of duration 1000 hrs, 6760 hrs and 1000 hrs respectively as shown in Figure 1and the these description with the base case
loss and the lowest voltage are also shown in Table 1.
All the individual buses demand curves are considered identical to the demand curve taken at substation. This assumption is
really necessary because individual demand curves for each bus are rarely available. In addition, this assumption does not interfere
in the result significantly, especially when the feeder supplies many loads and the influence area is homogenous like a residential,
commercial, rural or industrial area.

Table1: Duration of various load levels and other details for the considered systems

Load Levels L1 L2 L3
Load level 0.625 1.00 1.25
Duration time(Hrs) 1000 6760 1000
Loss without DG (kW) 79.8824 216.00 351.506
Lowest voltage V 0.9443 0.9078 0.8820


97

T 1 T 2 T3
L1
L2
L3
Time
Load
Level

Figure 1: Approximated Load Duration Curve
3. Genetic Algorithm

Genetic Algorithm simulates the biological processes that allows the consecutive generations in a population to adapt to their
environment. Genetic Algorithms are unconstrained optimization methods, which model the evolutionary adaptation in nature.
They work with a population of solutions and create new generations of solutions by appropriate genetic operators. In depth
description of the method is not provided here as GA has been applied in several problems and excellent texts; Goldenberg (1989)
and Kalyanmoy Deb (1995), are available. However, its implementation for DG allocation problem is delineated in following
sections.
GA is effective parameter search techniques. They are considered when conventional techniques have not achieved the desired
speed, accuracy or efficiency. GA is different from conventional optimization and search procedures in the following ways.
GA work with coding of parameters rather than the parameters themselves.
GA search from a population of points rather than a single point.
GA use only objective functions rather than additional information such as their derivatives.
GA use probabilistic transition rules, and not deterministic rules.
Following are the advantages of GA.
They require no knowledge of gradient information about the response surface.
They are resistant to becoming trapped in local optima therefore can be employed for a wide variety of optimization
problems.
It can quickly scan a vast solution set.
Bad proposals do not affect the end solution negatively as they are simply discarded.
It doesn't have to know any rules of the problem - it works by its own internal rules.
For the advantages of parallel searching, robust searching, and searching mechanism based on the principle of natural evolution,
genetic algorithm has found applications in many areas and has become one of the most successful optimization algorithms. The
GA become particularly suitable for the problem posed here due to above mentioned features.
3.1. Coding Strategy
It is expected that the GA should produce optimal values of active power to be injected from the DG installed at appropriate
location(s) in the network. Since the solution variables are coded in binary form in GA implementation, it becomes imperative to
represent solution variables (i.e. DG values in present study) in binary coded form. The coding of active power to be injected in the
system at the candidate buses through DGs is done in binary form using 12 binary bits for single location to take care of real power
capacity of 4096 kW available at the source node (substation) of example test system discussed in section 5. These 12 bits would
get repeated as many times as the number of locations in case of more than one location.

3.2. Fitness Function
The fitness function in GA implementation plays a major role and generally expressed in terms of objective function to be
optimized. Since in the present work, objective function is to minimize active power loss, the fitness function is defined as the
inverse of objective function.


98

3.3. Selection of GA parameters
Selection of appropriate GA parameters is crucial for its faster convergence. In absence of any guide line to choose these
parameters, some mechanism has to be devised. However, the types of crossover and mutation are based on user choice. In present
implementation, one point crossover and constant mutation types have been chosen. In order to quantify the values of crossover
and mutation probabilities, effect of variations in mutation probability at fixed value of crossover probability and crossover
probability at fixed value of mutation probability on convergence properties have been studied and demonstrated in Figure 2 and
Figure 3. The population size and number of generations for these studies have been chosen to be 100 and 50 respectively. It can
be observed from Figure 2 that the mutation probability of 0.001 appears to be the most suitable as objective function value
declines fast and attains lowest value at the earliest compared to other values of mutation. With this value of mutation probability,
convergence properties at various crossover probabilities were plotted as shown in Figure 3. The best performance was found at
0.85 crossover probability. Therefore the final value of crossover and mutation probabilities was chosen 0.85 and 0.001
respectively for the present studies. In order to confirm the suitability of chosen value of population size, the performance was
further studied at various values of population size at cross over and mutation probabilities 0.85 and 0.001 respectively and best
performance was observed at population size of 100. Therefore, this combination of 100, 50, 0.85 and 0.001 respectively for
population size, number of generations, crossover probability and mutation probability was taken as the final value for further
studies.


99

3.4. GA Implementation
The GA technology is particularly suitable for the solution of combinatorial optimization problems. The advantages of using GA
are that they require no knowledge of gradient information about the response surface; they are resistant to becoming trapped in
local optima and can be employed for a wide variety of optimization problems. The GA methodology discussed above is
implemented using following steps.

Step 0: Initialized number of locations to unity.
Step 1: Determination of candidate locations:
Input the distribution system branch impedances and complex bus powers.
Determine the sensitivity factors.
Arrange the buses in descending order of their sensitivities.
Bus at the top is selected as the present location for DG placement.

Step 2: Input genetic algorithm control data.
Step 3: Initialize population with random strings and copy into mating pool.
Step 4: Do while generation number is less than maximum number of generation taken
Do while population number is less than population size
Pick up the string corresponding to population number from mating pool and decode it into test configuration
Apply load demand
Call distribution load flow solver
Check voltage constraints
Compute fitness function
Increment population number by one
Use mating pool to create new population for next generation
Carry out reproduction, cross over and mutation in mating pool
Increment generation number by one
Step5: Obtain desired solution i.e. optimal DG size, minimum system loss and savings
Step 6: If the saving is more than the previous value, increment number of locations by one. Otherwise go to step 7.
Step 7: Stop

4. Results and Discussion

4.1. System and other details
Effectiveness of the proposed methodology is tested on the widely used 33 bus-32 branch test system at different loading
conditions. The system data for this system has been taken from Kashem et al (2000). The active power (sum of total connected
load and the base system losses) at the source node in the beginning of the period is about 3936 kW. The entire algorithm has been
written in Visual C++ language and implemented on dual core 2.56 MHz, 1 GB PC. In order to demonstrate the long term impact
of DG, planning period of 10 years have been considered with a uniform load growth of 2.5% per year and maintaining same
configuration of the network. DG sources are assumed to be always available throughout the planning period.
The cost benefits of optimally allocated DG for a utility are explained by a reduction in the energy bought by the amount
produced by the DG units. This also implies a reduction in losses and saving in energy that the utility avoids buying from large
generators because of self production. The cost of DG generated power of US $30.00 per kW has been arbitrarily taken for this
study which includes capital cost of DG with installation, operation and maintenance cost. The generation cost may be low
however, it can be considered appropriate for demonstration of the effectiveness of the methodology in terms of energy savings.
The rate of energy cost has been considered equal to US $0.05 per kWh for the cost benefit analysis. It should be highlighted that
in presence of liberalized electricity market, different retail sales rate of the energy produced by a DG unit should be considered.
These retail sales depend on the technology adopted (mini gas turbine, CHP, wind turbine, etc.), the regulatory actions and the
willingness to harness renewable energy.
Energy saving cost (ESC) has been calculated as the difference of energy loss cost without DG and the energy loss cost with DG
for variable loads (ELC) and the monetary benefits are calculated as a difference of ESC and the DG cost which includes cost of
DG and its installation, and the expenditure incurred towards their operation & maintenance.

100

The base case loss sensitivity results of six most sensitive buses arranged in descending order are shown in Table 2 for the
system studied. Although this paper aims to illustrate technical and economic benefits of DG allocation at multiple locations, the
demonstration deals with single location case also in order to compare it with results reported in literature by Acharya (2006).

4.2. Discussion
Bus no. 6 being the most sensitive mode as seen in Table 2, is selected as first candidate location for DG placement in this
system. The same node has also been selected as the most suited location by Acharya (2006) using heuristic method. Results of
DG capacity, loss reduction and energy savings are reported in Table 3 for proposed and heuristic methods. Acharya et al. (2006)
reported the base case system loss of 211.20 kW is based on an approximate load flow method Distflow used by the authors.
However, this value was found to be 216. 00 kW using backward sweep power flow method and the same is shown in Table 3 for
fair comparison. The optimal DG size of 2380 kW was obtained by proposed method whereas 2490 kW has been reported by
Acharya (2006). The system loss dropped from 216 kW to 132.83 kW in heuristic method and 132.64 kW in proposed method. It is
to be noted that the evaluated DG capacity is much less than the reported in heuristic method with higher loss reduction though
marginal. It can also be seen that the energy savings obtained for one year by proposed method is $36511.68 while it is $36428.46
by heuristic method.

Table 2: Sensitivity Analysis Results

Bus No. 6 2 5 20 8 7
Sensitivity values 0.06779 0.0670792 0.0664193 0.0573545 0.0537697 0.0476795

Table 3: Summary for 33 bus system with rated loads for Single location

Power Loss(in KW)
Methodology

Optimal
location

DG size
(kW)
Without
DG
With
DG
Loss
reduction
(kW)
Energy
Savings ($)
Heuristic 13] 6 2490 216.00 132.83 83.17 36428.46
Proposed GA 6 2380 216.00 132.64 83.36 36511.68

In order to study the effect of DG installation in radial network, an attempt has been made to place DG at more than one location
(multiple locations). The next location has been decided on the basis of successive sensitivity explained in section 2.1. Following
this procedure, bus no. 8 was found to be the next location for allocating DG at two locations. The GA produced DG value of 1718
kW and 840 kW at buses 6 and 8 respectively. Proceeding in this manner, analysis was carried out up-to four locations considering
the variable loads and results are shown in Table 4. It can be seen from this table that the line loss decreases as the number of
locations increases. However, amount of decrease in line loss reduces in successive locations. The saving in energy cost (ESC)
follows the reverse trend. It increases with increase in number of locations but at higher number of locations this increase is
marginal. Since marginal reduction in loss and marginal increase in energy saving cost was observed when the number of locations
was increased to four, further increase in number of locations were not considered.
Optimal values of DG to be placed for various loading conditions were obtained using proposed technique. Effect of variations
in number of locations on DG values for three considered loading conditions are also tabulated in Table 5. It is observed that the
losses reduce at all load levels with the improved voltage profile as the number of locations increase. The value of loss reduction
becomes smaller and smaller as the number of locations increase which can be seen from Table 5.

Table 4: Effect of no. of locations on DG size, losses, voltage profile and one year ELC, ESC

Another advantage of improvement in voltage profiles due to installation of DG (at the end of planning period of 10 years) has
also been demonstrated in Table 5. One can see the variation in voltage profile at various loading conditions with the variation in
No. of
location
1
(6)
2
(6,8)
3
(6,8,20)
4
(6,8,20,24)
L1 41.9792 36.701 28.4856 27.8815
L2 132.64 96.58 76.301 67.7772
Losses (kW) at various
load levels
L3 177.863 153.986 123.63 113.667
Lowest Voltage (p. u.) 0.934575 0.94195 0.943189 0.945094
Cost of Energy Loss (ELC) in $ 48500.00 42178.34 33395.52 29986.119
Cost of Energy Saving (ESC) in $ 46108.00 52429.66 61212.48 64621.881

101

number of locations. It can be very well understood that the voltage profiles get improved as the number of locations increases.
This increase has been observed from 0.8820 p. u. (without DG) to 0.9451 p. u. (with DG placed at four locations) at peak load
level (L3).
It can further be observed that the values of loss reduction reduced by 0.381kW, 0.921kW and 1.550kW at load levels L1, L2 &
L3 respectively when number of locations was increased from three to four while the corresponding values of loss reduction when
number of locations was increased from one to three can be observed equal to 9.3952 kW, 46.889 kW and 41.294 kW respectively.
It suggests that the loss reduction reduces as the number of locations increases and beyond certain number of locations, the loss
reduction may cease. However, the limiting number is decided on the basis of economic benefits as discussed in the next section.
DGs installed ones will remain in system as long as their performance is satisfactory. Thus the money invested once on DG will
help investor to harness its benefit throughout its life. Although the systems losses reduce as the number of DG at various locations
increase, the investment on DG also increase. Thus the reduction in cost of energy loss is achieved at additional investment on DG
and investor can be only benefited as long as saving is more than investment. Therefore, there must be some limit on number of
DG beyond which it may be uneconomical.
In order to investigate this limit, economic analysis has been performed. The benefit achieved was obtained with varying loads
and varying number of DG locations. Considering the load growth of 2.5% per year, the loads at peak loading condition have risen
to about 1.5 times the base case loading of the considered system at the end of 10 year planning period. The results of energy loss
cost (ELC) and the benefits in US$ are shown in Figure 4. It can be seen that the benefit increases if the number of locations is
increased from one to three. However, the value of the benefit (compensation) gets reduced beyond three no. of locations.
To assess the speed of GA for the solution of present problem for the DG placed at multiple locations, computational time was
also recorded and has been plotted in Figure 5 for one year and ten year planning periods. It can be observed that the
computational time increases with increase in locations but there is marginal increase in time for ten years planning period
compared to one year period.

Table 5: Effect of optimally located DG on DG size, losses and voltage profile for different locations

DG Size (kW) at bus no. No. of
locations
Load
Levels 6 8 20 28
Losses
(kW)
Lowest
voltage (p .u)
L
1
1581 - - - 41.979 0.964421
L
2
2380 - - - 132.64 0.934575

one
L
3
3293 - - - 177.86 0.908954
L
1
1060 511 - - 36.695 0.97167
L
2
1718 840 - - 96.58 0.96639

Two
L
3
2047 1111 - - 153.99 0.94195
L
1
897 511 599 - 32.584 0.972824
L
2
1536 846 946 - 85.751 0.955147
Three
L
3
1882 1023 1280 - 136.57 0.943189
L
1
889 516 189 593 32.203 0.96981
L
2
1391 896 298 956 84.830 0.95974 four
L
3
1839 1057 373 1216 135.02 0.945094


102

4. Conclusions

Placement of DG in radial distribution network has been demonstrated in this paper. The appropriate location of DG had been
decided on the basis of sensitivity of active power loss with respect to real power injection through DG. In case of multiple
locations, successive sensitivity had been made use of using same algorithm. The problem is formulated as an optimization
problem with minimization of real power loss (hence evaluation of monetary benefits) subject to equality and inequality
constraints and solved for optimal value of DG using GA. It is demonstrated that the benefit increases with increase number of
locations within certain locations beyond which it is uneconomical. Significant improvement in voltage profile as an additional
advantage of optimal DG placement has also been observed and demonstrated.

Nomenclature

r System resistance per unit length,
( ) rL R = Total resistance of the line,
L Total length of the line,
LN
Number of line sections,

k
Loss
Real power loss in section k,
SG
Loss
Line loss from source to DG location,

103

GL
Loss
Line loss from the location of DG to load
P
i
and Q
i
Real and reactive power demands at
th
i bus respectively,
P
Gi
and Q
Gi
Real and reactive power generation by DG placed at bus i respectively,
P
Loss
System real power loss,
N Number of buses,
scheduled
ij
I Current carrying capacity of line section connected between nodes j and i .
i
V And
i
voltage magnitude and angle at i
th
bus respectively.
min
i
V ,
max
i
V Acceptable voltage limit at bus i
V
P
System phase voltage,

ij ij ij
jx r Z + = ;
th
j i

Elements of [ ]
BUS
Z matrix
ij
,
ij
are the loss coefficients represented as
j i j i j i ij
V V r / ) cos( =

Appendix

A.1 Loss Saving Analysis
Loss savings are classified as either capacity or energy loss savings. Capacity loss savings reduce load on T&D and generation
system equipment. This lessens the need for capital upgrades. They are calculated by developing feeder and transformer loss
savings equations and evaluating them during peak load conditions. Energy loss savings reduce electricity generation
requirements. Their value is the cost savings realized by reducing operation and maintenance expenses of existing plants.
A.2 System Modeling
This section focuses on line loss reduction analysis. In this study, one-concentrated load is assumed at the end of the line. With
the introduction of DG, line loss reduction can be expected. This factor is analyzed, quantified and presented in this paper for
multiple locations of the DG along the feeder.
Two simple radial systems are considered:
i) System without DG
ii) System with the inclusion of DG.
Both systems have a concentrated load at the line end. The total length of the line is assumed to be L km. Schematics of the two
cases are shown in Figure A.1 and Figure A.2.

Figure A.1: A simple radial distribution system without WTG.

For the system with DG, the location of DG is assumed to be G km from source.
j i j i j i ij
V V Sin r / ) ( =

104

Figure A.2: Schematic of a radial system with the inclusion of DG.
The following assumptions are made in the study:
1) Load is Y-connected; line current is the same as phase current; I
L
= I
P

2) Load absorbs real power at some specified power factor
3) DG produces real power at a lagging or leading or unity power factor
4) V
P
is the RMS load phase voltage. V
P
is the reference phasor.
The load complex power is S
L
= P
L
+ jQ
L
, therefore, the current absorbed by the load is

( )
P
L L
L
V
Q j P
I
3
=
(A.1)
A.3 Line Loss Reduction Analysis
Electrical line loss occurs when current flows through transmission and distribution systems. The magnitude of the loss depends
on amount current flow and the line resistance. Therefore, line loss can be decreased by reducing either line current or resistance or
both. If DG is used to provide energy locally to the load, line loss can be reduced because of the decrease in current flow in some
part of the network.
The real power injection at a bus i,
i inj
P
,
considered here as the difference between real power generation and the real power
demand at that bus, and can be expressed as

i Gi i inj
P P P =
,
(A.2)
From the schematic of the system shown in Figure 1, instantaneous base case losses for a three-phase system can be expressed as

( )
( )
2
2 2
3
P
i i
B
V
Q P L r
Loss
+
=
(A.3)
Now in case of line loss with DG, assuming that the line is short, voltage drop along the line is neglected. Schematic of this
system is shown in Figure 2.
The output current of DG supplying complex power {S
Gi
= P
Gi
+ jQ
Gi
}

is given by

=
P
Gi Gi
G
V
jQ P
I
3
(A.4)
The line loss with the integration of DG is a combination (sum) of two parts:
i) Line loss from source to the location of DG.
ii) Line loss from DG location to the location of load.
In presence of DG (Figure A.2), the feeder current I
S
will be the difference of load current I
L
and DG output current I
G
. The line
loss from source to DG location can be expressed as

2
3
) 2 2
2
Q
2
P
2
Q
2
P (
P
V
G
Q
L
Q
G
P
L
P
G G L L
rG
rG Loss
SG
+ + +
=
(A.5)
Again, the line loss from the location of DG to load (
GL
Loss ) will be

2
2 2
3
) ( ) (
P
L L
GL
V
Q P G L r
Loss
+
=
(A.6)
The total line loss (
) ( AT
Loss ) in presence of DG placed at a X distance from the source can be calculated by combining
Equations (A.5) and (A.6) and expressed as

105

( )

+ + + =
L
X
Gi
Q
i
Q
Gi
P
i
P
Gi
Q
Gi
P
i
Q
i
P
P
V
R
AT
Loss 2 2
2 2 2 2
2
3
) (
(A.7)
Instantaneous loss savings (LS) at any point on a feeder can be represented as the difference between losses without DG and
losses with DG.
LS = Loss
(B)
- Loss
(AT).

Hence,

( )
2 2
2
2 2
3
Gi Gi Gi i Gi i
P
Q P Q Q P P
L V
RX
LS + =
(A.8)
The positive sign of LS indicates that system loss reduces with the integration of DG but the negative sign implies that DG
causes higher loss in the system.

References
Acharya N., Mahat P., Mithulanathan N., 2006. An analytical approach for DG allocation in primary distribution network, Electric
Power and Energy Systems, Vol. 28, pp. 669-678.
Afsari, M., Singh, S.P. , Raju, G.S. and Rao, G.K. 2002. A fast power flow solution of radial distribution networks, Electric
Power Components and Systems (USA), Vol. 30, No.10, 1065-1074.
Brown R.E., Pan J., Feng X., and Koutlev K., 1997. Siting distributed generation to defer T&D expansion, Proc. IEE. Generation,
Transmission and Distribution, Vol. 12, pp. 1151- 1159.
Chiradeja P., Ramakumar R. 2004. An approach to quantify the technical benefits of distributed generation, IEEE Transactions on
Energy Conversion, Vol. 19, No. 4, pp. 764-773.
Celli G., Ghiani E., Mocci S., and Pilo F., 2005. A multiobjective evolutionary algorithm for the sizing and siting of distributed
generation, IEEE Transactions on Power Systems, Vol. 20, No. 2, pp. 750-757.
Dugan R.C., McDermott T.E. and G.J. Ball 2001. Planning for distributed generation, IEEE Industrial Application Magazine, Vol.
7, pp. 80-88.
Daly P.A., Morrison J. 29 April 1May 2001. Understanding the potential benefits of distributed generation on power delivery
systems, Rural Electric Power Conference, pp. A211 A213.
Diaz-Dorado E., J. Cidras, E. Miguez 2002. Application of evolutionary algorithms for the planning of urban distribution networks
of medium voltage, IEEE Transaction on Power Systems, Vol. 17, No. 3, pp. 879-884.
Elgerd O.I., 1971. Electric energy system theory: an introduction, McGraw- Hill Inc.
Goldenberg D. E. 1989. Genetic Algorithm in Search, Optimization and Machine Learning: Addition-Wesley Publishing Co. Inc
Harrison G.P., Piccolo A., Siano P., Wallace A.R. 2008. Hybrid GA and OPF evaluation of network capacity for distribution
generation connections, Electrical Power Energy System, Vol. 78, pp. 392398.
Ijumba NM, Jimoh AA, Nkabinde M (1999). Influence of Distribution Generation on Distribution Networks Performance,
Proceeding of AFRICON 99, 2: 961-964.
Jenkins N (1996). Embedded Generation-Part 2, IEE Power Engine. J., pp. 233-239.
Jabr, R. A., Pal, B. C. 2009. Ordinal optimization approach for locating and sizing of distributed generation, IET proceedings
Generation, Transmission & Distribution, Vol. 3, No. 8, pp 713 723.
Khator K. and Leung, L.C. 1997. Power distribution planning: A review of models and issues, IEEE Transaction on Power
Systems, Vol. 12, No. 3, pp.1151-1159.
Kyoto Protocol to the United Nations Framework Convention on climate change, Dec.1997.
Kalyanmoy Deb, 1995. Optimization for engineering design: Algorithm and Examples: Prentice Hall of India Ltd., New Delhi.
Kashem M. A., Ganpathy, V., Jamson, G.B., Buhari, M. I. 2000. A novel method for loss minimization in distribution networks,
Proc. International Conference on Electric Utility Deregulation and Restructuring and Power Technologies, pp. 251-255.
Mardaneh, M., Gharehpetian G. B..2004, Siting and sizing of DG units using GA and OPF based technique, TENCON. IEEE
Region 10 Conference, Vol. 3, pp. 331-334, 21-24.
Milanovic JV, David TM (2002). Stability of Distribution Networks with Embedded Generators and Induction Motors,
Proceeding of the IEEE PES Winter Meeting, 2: 1023-1028.
Piccolo, A. Siano, P. 2009. Evaluating the impact of network investment deferral on distributed generation expansion, IEEE
Transactions on Power System, Vol. 24 No. 3, pp 1559 1567.
Quintana V.H., Temraz H.K., Hipel K.W. 1993. Two stage power system distribution planning algorithm, Proc. IEE Generation,
Transmission and Distribution, Vol. 140, No. 1, pp. 17-29.
Rackliffe G. 2000. Guidelines for planning distributed generation systems, Proc. of IEEE Power Engineering Society Summer
Meeting, vol. 3. Seattle (WA, USA): pp1666-1667.
Silvestri A., Berizzi, S., Buonanno, 1999. Distributed generation planning using genetic algorithms, Electric Power Engineering,
Power Tech Budapest 99, Inter. Conference, pp.257.

106

Biographical notes:

T. N. Shukla was born in India in 1955. He received his B.E. and M.E. in Electrical Engineering from Allahabad University in 1976 and 1981 respectively.
Presently he is pursuing his Ph. D. from UP Technical University, Lucknow. He joined the Electrical Engineering Department at Kamla Nehru Institute of
Technology, Sultanpur, as lecturer in 1982 where he is presently serving as Professor. He is currently pursuing his Ph.D. from U.P. Technical University, Lucknow.
His main research interests are in optimal operation and AI applications to power distribution systems.

S. P. Singh was born in India in 1957. He received his B.E. and M.E. in Electrical Engineering from Allahabad University in 1978 and 1981 respectively. He
obtained his Ph.D. from Banaras Hindu University in 1990.He was Post Doctoral scholar at University of Calgary, Canada during 1993-94. He joined the Electrical
Engineering Department of Banaras Hindu University as faculty in 1981 where he is presently serving as Professor. His main research interests are in secure and
optimal operations of power systems, deregulation, distribution automation and AI applications to power systems. Dr. Singh is a Fellow of Institution of Engineers
(India), Senior Member of IEEE and life member of Systems Society of India.

K.B.Naik was born in India in 1947. He received his B.E. from Kanpur University and M.E. from Nagpur University in Electrical Engineering. He obtained his
Ph.D. from Kanpur University in 1984. He joined the Electrical Engineering Department at Kamla Nehru Institute of Technology, Sultanpur, as professor in 1986
and served as professor till Jan. 2007. He has also been the Director of the institute for about six years. After retirement of his service, presently he is serving as the
Director of College of Engineering, Sciences and Technology, Lucknow (UP), India. His main research interests are in optimal operation, power electronics and AI
applications to power distribution systems.

Accepted March 2010

MultiCraft
Vol. 2, No. 3, 2010, pp. 107118
INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY
www.ijest-ng.com
2010 MultiCraft Limited. All rights reserved.
Distribution system reliability evaluation using credibility theory
Xufeng Xu, Joydeep Mitra*
Department of Electrical & Computer Engineering, Michigan State University, USA, MI 48824
*
Corresponding Author: e-mail: mitra@ieee.org
Abstract
This paper describes a new method of using credibility theory to evaluate distribution system reliability. The advantage of this
method lies in its ability to account for both objective and subjective uncertainty by integrating stochastic and fuzzy approaches.
Equipment failures are modeled as random events, while the uncertainty in failure indices is modeled using credibility theory. In
this manner, distribution system reliability is evaluated by simultaneously considering objective and subjective uncertainty. In this
paper, a hybrid algorithm based on fuzzy simulation and Failure Mode and Effect Analysis (FMEA) is applied to determine fuzzy
reliability indices of distribution system. This approach can obtain fuzzy expected values and their variances of reliability indices,
and the credibilities of reliability indices meeting specified targets. The method is demonstrated using an IEEE test system
comprising sixty seven nodes and thirty eight load points.
Keywords: Distribution system, reliability evaluation, credibility theory, fuzzy reliability indices, fuzzy simulation
1. Introduction
Distribution system reliability indices quantify, in various ways, the ability of a distribution system to meet customer demand at
all its load points. Due to the larger impact of failures in the bulk power system, the development of predictive reliability
approaches has received more attention from the research community. However, distribution system components are relatively
more numerous and they are exposed to a more complex set of uncertain events. As distribution systems grow in complexity, with
the deployment of distributed generation and storage, and of distribution automation equipment, it is important to develop more
complex methods for predictive reliability analysis of distribution systems so as to suitably address these uncertainties.
Recent research efforts in distribution system reliability evaluation have focused on the following: (1) consideration of the
impact of distributed generation (Atwa and Saadany, 2009; Bae and Kim, 2007) and microgrids (Bae and Kim, 2008; Kennedy,
2009); (2) incorporating weather effects (Billinton and Acharya, 2005; Billinton and Acharya, 2006); (3) time-varying failure rate
(Retterath et al, 2004; Moon et al, 2004). Since all of these factors are indeterminate, some research has considered them
collectively (Dong et al, 2009; Wang et al, 2002).
Uncertainty factors can be classified into two categories: stochastic factors and fuzzy factors. Usually, randomness is related to
objective uncertainty factors that have statistical characteristic, while fuzziness is commonly used to define factors that have
subjective uncertainty (Wang et al, 2007). Reliability evaluation methods in which only randomness is taken into account often
fails to include the effect of fuzzy factors which can be significant. Models with both stochastic and fuzzy factors are often more
appropriate; however, modeling this dual uncertainty can be challenging. Some recent efforts have modeled these uncertainties:
trapezoidal fuzzy numbers have been used to express uncertainties in Lei et al, 2005; Yuan et al, 2007 have used interval
algorithm to deal with the uncertainty of component data to calculate the interval reliability indices.
Most of fuzzy methods for reliability evaluation of distribution system are based on fuzzy set theory. In recent years researchers
have found that credibility theory, which broadens the scope of fuzzy set theory, is an effective tool for representing fuzzy events,
and have developed a theoretical framework similar to that of probability theory (Liu, 2004). Credibility theory has been
successfully applied in power system analysis fields, such as fuzzy power flow (Wang et al, 2007) and fuzzy risk assessment (Feng
et al, 2008).
Xu and Mitra / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 107118 108
This paper focuses on the reliability evaluation of distribution system under uncertain environment and proposes an approach
based on credibility theory. The approach intensively analyzes the fuzzy factors in evaluation process and defines fuzzy reliability
counterparts of traditional indices. Using this approach, a hybrid algorithm based on FMEA and fuzzy simulation is developed and
implemented. The algorithm is demonstrated using the distribution system connected to bus 4 of the IEEE test system commonly
referred to the RBTS (Allan et al, 1991). The case study includes the computation of fuzzy expected values and variances of
reliability indices, as well as ranges of these indices, which could be useful in analysis and decision making processes.
2. Credibility Theory
Definition 1: Let
~
be a fuzzy variable; then the possibility, the necessity, and the credibility of the fuzzy event A are expressed as
follows:
{ } ) ( sup Pos ~ x
A x

e
= A (1)
{ } ) ( sup 1 Nec ~ x
c
A x

e
= A (2)
{ } { } { } ( ) A A A Nec Pos
2
1
Cr + = (3)
where ( ) x
~ is membership function of
~
; A
c
is the complementary event of A.
In traditional fuzzy set theory, possibility measure is considered to be a parallel concept of probability measure. But it does not
possess self-duality: when Pos{A} = 1, fuzzy event A is not always true; when Nec{A} = 0, A may also be true. In order to solve
this problem, reference (Liu and Liu, 2002) put forward the credibility measure and established credibility theory in 2004. In this
theory, credibility measure which corresponds to probability measure in probability theory possesses self-duality and monotonicity
when Cr{A} = 1, fuzzy event A is inevitably true, otherwise it is false.
Definition 2 (Liu and Peng, 2005): Let
~
be a fuzzy variable. Then the expected value of is defined by
{ } { }dr r Cr dr r Cr E
} }

+
< > =
0
0
~ ~
]
~
[ (4)
provided that at least one of the two integrals is finite.
Definition 3 (Liu and Peng, 2005): Let
~
be a fuzzy variable with finite expected value ]
~
[ E . Then the variance of is defined
as
] ])
~
[
~
[( ]
~
[
2
E E V = (5)
3. The Definition of Interrelated Fuzzy Indices
Equipment failure rate and its repair time are two basic parameters used in the reliability evaluation of distribution systems.
They are statistical values with a lot of subjectivity and are also used to evaluate equipment of similar types. For this reason, this
approach adopts trapezoidal fuzzy number which affords comparatively broad adaptability in engineering area.
3.1 Modeling of fuzzy factors
1) Failure rate
i
~
The equipment failure rate is generally used to quantify the probability of failure of a specific piece of equipment. Although
equipment status in the future is not predictable, the failure rate of this equipment can be obtained from tests or experience. Also,
i
~
can be modeled by trapezoidal fuzzy number ( )
4 3 2 1
, , ,
~
i i i i i
= , in which | |
4 1
,
i i
is the failure rate interval of specific
equipment i and | |
3 2
,
i i
is the most possible limits. The membership function of
i
~
is defined as below:
> <
s s
< <
s s
=
4 1
4 3
3 4
4
3 2
2 1
1 2
1
~
or , 0
,
, 1
,
) (
i i
i i
i i
i
i i
i i
i i
i
i
x x
x
x
x
x
x
x

(6)
2) Repair time
i
r
~
Since the insufficiency of information will result the inaccuracies of forecasting failure time, for the equipment i, trapezoidal
fuzzy number ( )
4 3 2 1
, , ,
~
i i i i i
r r r r r = is used to describe repair time because this number is more reasonable than point value, and its
membership function
i
r
~ is shown in Figure 1.
i
r
~
1
i
r
i
r
~
2
i
r
3
i
r
4
i
r
Figure 1. Membership function of repair time
3.2 Fuzzy reliability indices
Distribution networks are often radial and consist of series connections of lines, transformers, isolating switches, fuses, etc.
Service continuity requires each equipment between power resource and load to be online. Therefore, the effect of the series
system failure rate
s
~
, repair time
s
r
~
and time to failure (up time)
s
T
~
are taken into account as follows:
=
i s

~ ~
(7)
=
i
i i
s
r
r
~
~
~
~
(8)
~
+
=
i i
s s
s
s
r
r
T
~
~
~
/ 1
~
~
~

(9)
where
i
~
,
i
r
~
are failure rate and repair time of equipment i respectively. Subscript s expresses equivalent series network from
power sending end to the receiving end.
The definition of fuzzy reliability indices of distribution system is given as follows, based on Billinton and Allan, 1996.
Comparing with regular standard, those fuzzy indices indicate the impact of fuzzy parameters on the reliability indices.
1 Fuzzy Energy Not Supplied (ENS)
s s
T d
~
ENS

= (MWh/yr) (10)
where
s
d is the load of the equivalent series network. As
i
~
and
i
r
~
are fuzzy variables, ENS is also a fuzzy index reflecting
the influence of parameter ambiguity on reliability index.
2 System Average Interruption Frequency Index (SAIFI)
=
s
s s
N
N
~
SAIFI (int/cus-yr) (11)
where
s
N is the number of users on load node s.
3 System Average Interruption Duration Index (SAIDI)
=
s
s s
N
N T
~
SAIDI (hr/cus-yr) (12)
4 Customer Average Interruption Duration Index (CAIDI)
=
s s
s s
N
N T
~
~
CAIDI (hr/int) (13)
5 Average Service Availability Index (ASAI)

=
s
s s s
N
N T N
8760
~
8760
ASAI (14)
6 Average Energy Not Supplied (AENS)
=
s
s s
N
d T
~
AENS (MWh/cus-yr) (15)
4. Solution Procedure
This approach adopts Failure mode and effect analysis (FMEA) (Billinton and Allan, 1996) to evaluate reliability. The entire
process based on fuzzy sampling can be divided into three main parts and it is described as follows.
4.1 Distribution system analysis
First, the branch set connected by closed switches is defined as Sub-Power Supply System (SPSS). Based on the status of
switches, the distribution system can be divided into multiple SPSS, which are connected with tie switches. For example, SPSS
corresponding to feeder F1 in Figure 2 is composed of {1, 2,, 12}.
Secondly, based on the definition of SPSS, the power supply area of each feeder is analyzed by depth-first search strategy. The
searching process begins from power source and ends in load point or tie-line.
Finally, two matrices are defined and generated: one is network layer matrix which is used to reflect the level of branches and
their relationships; and the other is loading point information matrix which contains information of each load point, such as
number, active load values, number of customers, branch numbers and feeder numbers.
4.2 Reliability evaluation process
The procedure for evaluating fuzzy reliability indices consists of the following steps.
Step 1: Data input:
Set 0 =
i
u (i = 1, 2,, 6) and fuzzy simulation time N.
Step 2: Fuzzy sampling:
Randomly generating ( )
n
k
' ' ' = , ,
2 1
from c level cut sets of fuzzy vector
~
. The membership degree of
k
can be
obtained by following equation:
( ) ( ) ( )
n
n
k
v

' ' ' = ~
2
2
~
1
1
~
(16)
where c is a positive number small enough; n is equipment number; k = 1, 2,,N.
Similarly, fuzzy random value of r
~
is
k
r , and its membership degree is
k
v
r
.
Step 3: Calculate reliability indices of load point using (7), (8) and (9). Input data required are network layer matrix, load point
information matrix and fuzzy sampling values.
The network reliability is assessed as well as reliability index ( )
k k k
i
f r , using (10)(15).
Step 4: Calculate ( )
k k k
i
f r , , and its membership degree
k k k
v v v
d
= .
Step 5: Repeat steps 24 for a total of N times;
Step 6: Calculate
i
a and
i
b by the following equations, where
i
a and
i
b are both non-negative real numbers.
( ) ( ) ( )
N N N
i i i i
f f f a r r r , , ,
2 2 2 1 1 1
= (17)
( ) ( ) ( )
N N N
i i i i
f f f b r r r , , ,
2 2 2 1 1 1
= (18)
Step 7: Generate
i
t from ] , [
i i
b a randomly.
Step 8: Set ( ) { }
i i i i
f t u u > + r
~
,
~
Cr .
( ) { }
i i
f t > r
~
,
~
Cr are approximately calculated using (19).
( ) { } ( ) { } ( ) { } ( )
i
k k k
i
k
N k
i
k k k
i
k
N k
i i
f v f v f t t t < + > = >
s s s s
r r r , 1 min , max
2
1
~
,
~
Cr
1 1
(19)
Step 9: Repeat steps 7,8 for a total of N' times; N' is a given positive integer.
Step 10: ( ) | | N a b a f E
i i i i i
' + = / ) (
~
,
~
u r , (i = 1, 2,,6).
Fuzzy reliability indices ( ) | | r
~
,
~
i
f V can used to reflect the rate of deviation between fuzzy sampling values and fuzzy expected
value, just as the definition of variance in probability theory. The deriving of variance of fuzzy variable is basically the same as the
above process; the difference lies in the calculation of fuzzy sampling value of variance ( )
k k k
i
g r , which can be obtained from
(20).
( ) ( ) ( ) | | ( )
2
~
,
~
, , r r r
i
k k k
i
k k k
i
f E f g = (20)
4.3 Credibility calculation
Sometimes it may be necessary, during planning or operation, to use credibility of reliability indices meeting some stipulated
levels or guarantees; this credibility can be obtained by following equation:
( ) { } ( ) { } ( ) { } ( )
i
k k k
i
k
N k
i
k k k
i
k
N k
i i
F f v F f v F f > + s = s
s s s s
r r r , 1 min , max
2
1
~
,
~
Cr
1 1
(21)
where
i
F (i = 1, 2,,6) is default value which is also the value expected.
5. Simulation Results
The IEEE-RBTS Bus 4 system (Allan et al, 1991) is used to demonstrate the feasibility of the proposed model and algorithm. A
one-line diagram of the system is shown in Figure 2. This system contains 67 nodes (including 38 load nodes), 71 lines and 26
switches. The main feeders are LGJ-240, trunk lines are LGJ-120, and branch lines are LGJ-50. Parameters used in the algorithm
are as follows: N = 5000; N' = 2000. The equipment failure rate and repair time are fuzzified by using the multiplying factors
[0.80, 0.95, 1.05, 1.2].
Figure 2. Distribution system for RBTS Bus 4
This work analyzes the scenario that includes disconnects, fuses and alternative supply. The fuzzy expected value of reliability
indices are shown in the column three of Table 1. The result derived from the approach in this work compare well with results
reported in Allan and Billinton, 1991. Fuzzy reliability indices of feeder are shown in Table 2, and the fuzzy expected values and
variances of three main reliability indices of each load points are shown in Figure 3, Figure 4 and Figure 5. The stars denote the
fuzzy expected values while the circles denote the traditional reliability indices.
Table 1. Reliability indices of the system
Reliability indices
Result reported in Allan
and Billinton, 1991
Result obtained using
proposed method
Deviation
(%)
SAIFI (int/cus-yr) 0.300 0.2993 0.2331
SAIDI (hr/cus-yr) 3.47 3.4694 0.0184
CAIDI (hr/int) 11.56 11.6412 0.7025
ASAI 0.9996 0.9996 0.0000
ENS (MWh/yr) 54.293 54.3161 0.0425
AENS (MWh/cus-yr) 11.36 11.3655 0.0481
Table 2. Feeder fuzzy reliability indices
Feeders
SAIFI
(int./cus.yr)
SAIDI
(hr/cus.yr)
CAIDI
(hr/int.)
ASAI
ENS
(MWh/yr)
AENS
(MWh/cus.yr)
F1 0.3027 3.4707 11.6110 0.9996 12.1958 11.0857
F2 0.1888 0.3805 2.0120 1.0000 1.3277 443.8710
F3 0.2933 3.4698 12.0096 0.9996 12.0032 11.1142
F4 0.3070 3.4766 11.4866 0.9996 13.9398 10.7227
F5 0.1860 0.3763 2.0332 1.0000 1.1309 376.6532
F6 0.1940 0.3657 1.8882 1.0000 1.2704 423.0530
F7 0.2950 3.4722 11.8993 0.9996 12.4705 9.6672
1 5 10 15 20 25 30 35 38
0.1
0.2
0.3
0.4
F
u
z
z
y
E
x
p
e
c
t
e
d
V
a
l
u
e
Number of Load Point
5 10 15 20 25 30 35
0
0.5
1
1.5
x 10
-3
V
a
r
i
a
b
l
e
Figure 3. Failure rates of load points
1 5 10 15 20 25 30 35 38
0
2
4
6
8
10
12
14
F
u
z
z
y
E
x
p
e
c
t
e
d
V
a
l
u
e
5 10 15 20 25 30 35
0
0.2
0.4
0.6
0.8
1
1.2
1.4
V
a
r
i
a
b
l
e
Figure 4. The average interruption durations of load points
5 10 15 20 25 30 35
0
2
4
F
u
z
z
y
E
x
p
e
c
t
e
d
V
a
l
u
e
5 10 15 20 25 30 35
0
0.005
0.01
V
a
r
i
a
b
l
e
Figure 5. The failure time of load points
The fuzzy expected value ( ) | | r
~
,
~
ENS
f E and its variable ( ) | | r
~
,
~
ENS
f V of ENS is 54.3161 MWh/yr and 0.2676 respectively. If we
set 0.005 as unit interval, the range of fuzzy sampling value of ENS can be divided into several intervals. Defining
j
m as the
frequency that sampling value will be inside band j and
N
m
j
as the subjection frequency function, the relationship between
N
m
j
and sampling value ENS is shown in Figure 6. It can be seen that most of sampling values assemble around ( ) | | r
~
,
~
ENS
f E . In order
to reflect the degree of dispersion, unbiased variance o is defined by square root of variance ( ) | | r
~
,
~
ENS
f V . Within the interval
[53.7988, 54.8334] ( | | o o + E E , ), the total percent of frequency is 0.9936.
53.4 53.6 53.8 54 54.2 54.4 54.6 54.8 55 55.2
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
ENS (MWh/yr)
S
u
b
j
e
c
t
i
o
n
F
r
e
q
u
e
n
c
y
Figure 6. Frequency distribution of simulation results of ENS
In order to evaluate the influence of uncertain parameters on final results, three cases are designed for comparison: Case 1, using
parameters given; Case 2, fuzzifying multiplying factor is [0.70, 0.95, 1.05, 1.3]; Case 3, uses a triangular membership function,
with fuzzifying multiplying factor is [0.8, 1.0, 1.2]. The results are shown in Table 3. It can be found that the influence of
uncertainty on fuzzy variance will be much more than fuzzy expected value, through comparing Case 1 and Case 2. Triangle fuzzy
number which is also widely used in engineering is adopted in Case 3.
Table 3. Fuzzy expected value and variance of system reliability indices
Indices
SAIFI
(int./cus.yr)
SAIDI
(hr/cus.yr)
CAIDI
(hr/int.)
ASAI
ENS
(MWh/yr)
AENS
(MWh/cus.yr)
| |
i
f E 0.2993 3.4694 11.6412 0.9996 54.3161 11.3655
Case1
| |
i
f V 0.0002 0.0007 0.3224 0.0000 0.2676 0.0117
| |
i
f E 0.3008 3.4706 11.6695 0.9996 54.3579 11.3744
Case2
| |
i
f V 0.0006 0.0018 0.8139 0.0000 0.6190 0.0271
| |
i
f E 0.3007 3.4706 11.5725 0.9996 54.3344 11.3693
Case3
| |
i
f V 0.0003 0.0008 0.3408 0.0000 0.3311 0.0145
In our approach, equipment failure rate and repair time are uncertain, so reliability indices will also not assured. Thus,
credibility is also a good option in addition to fuzzy expected value of reliability indices. We design another two cases: Case
4, 32 . 0
1
= F , 5 . 3
2
= F , 3 . 12
3
= F , 9996 . 0
4
= F , 58
5
= F , 15 . 12
6
= F ; Case 5, assigning the reference result in column two of
Table 2 to
i
F (i=1,2,6) respectively. The simulation results are shown in Table 4. It can be seen that the higher the demand on
the reliability indices, the lower credibility will be. The relationship between { }
SAIFI SAIFI
F f Cr < and
SAIFI
F is shown in Figure 7
and the curve shape is dependent on the fuzzy expression of parameters. The relationships of another five reliability indices are
shown in Figure 8, Figure 9, Figure 10, Figure 11 and Figure 12 respectively.
Table 4. Credibility of reliability indices
Indices Case4 Case5
SAIFI { } 32 . 0
1
< f Cr = 0.9773 { } 3 . 0
1
< f Cr = 0.0844
SAIDI { } 5 . 3
2
< f Cr = 0.9904 { } 47 . 3
2
< f Cr = 0.0333
CAIDI { } 3 . 12
3
< f Cr = 0.9900 { } 56 . 11
3
< f Cr = 0.9667
ASAI { } 9996 . 0
4
> f Cr = 0.9990 { } 9996 . 0
4
> f Cr = 0.9990
ENS { } 58
5
< f Cr = 1 { } 293 . 54
5
< f Cr = 0.9684
AENS { } 12.15
6
< f Cr = 1 { } 11.36
6
< f Cr = 0.9684
0.25 0.26 0.27 0.28 0.29 0.3 0.31 0.32 0.33 0.34 0.35
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
F
SAIFI
C
r
e
d
i
b
i
l
i
t
y
Figure 7. The relationship between { }
SAIFI SAIFI
F f Cr < and
SAIFI
F
3.4 3.42 3.44 3.46 3.48 3.5 3.52 3.54 3.56 3.58 3.6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
F
SAIDI
(hr/cus.yr)
C
r
e
d
i
b
i
l
i
t
y
SAIDI SAIDI
F f Cr < and
SAIDI
F
10 10.5 11 11.5 12 12.5 13
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
F
CAIDI
(hr/int.)
C
r
e
d
i
b
i
l
i
t
y
CAIDI CAIDI
F f Cr < and
CAIDI
F
0.995 0.996 0.997 0.998 0.999 1 1.001 1.002 1.003 1.004 1.005
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
F
ASAI
C
r
e
d
i
b
i
l
i
t
y
ASAI ASAI
F f Cr > and
ASAI
F
53 53.5 54 54.5 55 55.5 56
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
F
ENS
(MWh/yr)
C
r
e
d
i
b
i
l
i
t
y
ENS ENS
F f Cr < and
ENS
F
11 11.1 11.2 11.3 11.4 11.5 11.6 11.7 11.8 11.9 12
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
F
AENS
(MWh/cus.yr)
C
r
e
d
i
b
i
l
i
t
y
AENS AENS
F f Cr < and
AENS
F
6. Conclusion
This paper employed credibility theory to evaluate the reliability of a distribution network. This method is able to model both
objective and subjective uncertainties. Compared with traditional methods, the proposed method is provides more information
about the indices by showing how the indices are dispersed, rather than merely providing their expected values. This affords a
better sense of realizable reliability targets, during both planning and operation. The illustrative cases have also demonstrated the
feasibility and value of the approach. It is surmised that this approach based on credibility theory will afford broad application in
the area of distribution system reliability evaluation.
Acknowledgment
This work was supported in part by the U.S. National Science Foundation under Grant ECCS 0702208.
References
Allan R. N., Billinton R., Sjarief I., Goel L. and So K. S., 1991. A reliability test system for education purposes basic
distribution system data and results. IEEE Transaction on Power Systems, Vol. 6, No. 2, pp. 813820.
Atwa Y. M., Saadany E. F. E., 2009. Reliability evaluation for distribution system with renewable distributed generation during
islanded mode of operation. IEEE Transactions on Power Systems, Vol. 24, No. 4, pp. 572581.
Bae I. S., Kim J. O., 2007. Reliability evaluation of distributed generation based on operation mode. IEEE Transactions on Power
Systems, Vol. 22, No. 2, pp. 785790.
Bae I. S., Kim J. O., 2008. Reliability evaluation of customers in a microgrid. IEEE Transactions on Power Systems, Vol. 23, No.
3, pp. 14161422.
Billinton R., Acharya J. R., 2005. Consideration of multi-state weather models in reliability evaluation of transmission and
distribution systems. Electrical and Computer Engineering, pp. 916922.
Billinton R., Acharya J. R., 2006. Weather-based distribution system reliability evaluation. Generation, Transmission and
Distribution, IEE Proceedings, Vol. 153, No. 5, pp. 499506.
Billinton R., Allan R. N., 1996. Reliability evaluation of power systems (Second edition). Plenum press, New York, NY.
Dong L., Li J., and Bao H., 2009. Reliability evaluation of distribution system considering composite uncertainty factors. Power
and Energy Engineering Conference, Asia-Pacific (APPEEC), pp. 15.
Feng Y. Q., Wu W. C., Zhang B. M. and Li W. Y.,2008. Power system operation risk assessment using credibility theory. IEEE
Transaction on Power Systems. Vol. 23, No. 3, pp. 13091318.
Kennedy S., 2009. Reliability evaluation of islanded microgrids with stochastic distributed generation. Power & Energy Society
General Meeting, pp. 18.
Lei X. R., Ren Z., Huang W. Y. and Chen B. Y., 2005. Fuzzy reliability analysis of distribution systems accounting for parameters
uncertainty. The Fourth International Conference on Machine Learning and Cybernetics, Guangzhou, China, pp. 40174022.
Liu B. D., 2004. Uncertainty theory: an introduction to its axiomatic foundations. Springer-Verlag, Berlin.
Liu B. D., Liu Y. K., 2002. Expected value of fuzzy variable and fuzzy expected value models. IEEE Trans on Fuzzy Systems,
Vol.10, No.4, pp.445450.
Liu B. D., Peng J., 2005. A Course in Uncertainty Theory. Tsinghua University Press, Beijing.
Moon J. F., Kim J. C., Lee H. T., Park C. H., Yun S. Y. and Lee S. S., 2004. Reliability evaluation of distribution system through
the analysis of time-varying failure rate. Power Engineering Society General Meeting, pp. 668673.
Retterath B., Venkata S. S. and Chowdhury A. A., 2004. Impact of time-varying failure rates on distribution reliability. 8th
International Conference on Probabilistic Methods Applied to Power Systems, Iowa State University, Ames, IA, pp. 953958.
Wang H., Wu W. C., Zhang B. M., and Pan J. Y., 2007. Fuzzy power flow based on credibility theory. Automation of Electric
Power Systems, Vol. 31, No. 17, pp. 2125.
Wang P., Billinton R., 2002. Reliability cost/worth assessment of distribution systems incorporating time-varying weather
conditions and restoration resources. IEEE Transactions on Power Delivery, Vol. 17, No. 1, pp. 260265.
Yuan Y., Qian K., Zhou C., 2007. The effect of distributed generation on distribution system reliability. Universities Power
Engineering Conference (UPEC), 42nd International, pp. 911916.
Biographical Notes
Xufeng Xu was born in 1981 in Zhejiang, China. He received the B.E. degree and the Ph.D. degree in electrical engineering from Zhejiang University, Hangzhou,
China. He is currently a Post-doctoral research associate in electrical engineering at Michigan State University, East Lansing, USA. His research interests include
power system reliability, power system planning and microgrid architecture. E-mail: xufeng@egr.msu.edu.
Joydeep Mitra received the B.Tech. (Hons.) degree in electrical engineering from Indian Institute of Technology, Kharagpur, India, and the Ph.D. degree in
electrical engineering from Texas A&M University, College Station. He is Associate Professor of Electrical Engineering and Senior Faculty Associate of the
Institute of Public Utilities at Michigan State University, East Lansing. Prior to this, he was Associate Professor at New Mexico State University, Las Cruces,
Assistant Professor at North Dakota State University, Fargo, and Senior Consulting Engineer at LCG Consulting, Los Altos, CA. His research interests include
power system reliability, distributed energy resources, and power system planning. E-mail: mitra@ieee.org.
Received December, 2009
Accepted March, 2010
Final acceptance in revised form March, 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 119-131

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Application of ant colony optimization for reconfiguration of shipboard
power system

Sri Hari Krishna Vuppalapati
1
,

Anurag K. Srivastava
2*

1
Department of Electrical Engineering, Indian Institute of Technology Kharagpur, INDIA 721302

2*
Department of Electrical and Computer Engineering, Mississippi State University, Mississippi State, MS, USA 39762

*
Corresponding Author: e-mail:srivastava@ece.msstate.edu, anurags@ieee.org

Abstract

The shipboard power system (SPS) supplies energy to sophisticated systems of weapons, navigation, services and
communication. Due to battle damage or faults, electric energy delivery may be interrupted to critical loads. Reconfiguration of
electrical network in the SPS is necessary, to either restore the service to all the possible loads or to meet some of the operational
requirements of the naval ship. In this research work, ant colony optimization (ACO) has been utilized to reconfigure given SPS
network, satisfying the operational requirements, priorities of the loads and considering islanding operation and distributed
generation. Graph theory has been applied to represent SPS electrical network to simplify the mathematical formulation to be
used by ACO. Developed technique has been applied to an eight bus and thirteen bus representative shipboard power system
model to reconfigure, while maximizing the load magnitude, or load priority, or load magnitude with priority. Satisfactory
results have been obtained for both the test cases.

Keywords: Shipboard power systems, islanding, reconfiguration, graph theory, ant colony algorithm

1. Introduction

Reconfiguration of an electrical network in shipboard power systems (SPS) is an important task to maximize the survivability,
security and reliability of naval ships (Jayabalan et al, 2005). The protection schemes in SPS only provide fault detection and
isolation functions, but do not consider system stability or power balance after the fault isolation. When a fault occurs, the fault
should be isolated immediately, which may result in un-faulted sections to be left without supply. It is required to quickly
reconfigure to restore supply to these un-faulted sections to increase the system survivability. Reconfiguration refers to the ability
of the system to redirect power by closing or opening of breakers to the loads in the event of a component failure, fault or
generator loss (Davey et al, 2005). Reconfiguration also allows power to be quickly diverted to high power applications/ devices
for very small time frames, which is vital for next generation, high-power weapons.
In SPS, generators are connected in ring configuration through generator switchboards. Bus tie circuit breakers interconnect
generator switchboards to allow the transfer of power from one to another. Load centers and some individual loads are supplied
from generator switchboards. Reconfiguration approach for SPS has been address in several literatures in the past. The
reconfiguration problem for shipboard power systems has been addressed using the heuristic-search approach in (Butler et al
2000). This method is very simple but has problems in restoration with an increasing number of loads. In the method of network
flow approach (Butler et al, 1999; 2001), the load priorities have not been considered. Authors have provided comparison of
several techniques for reconfigurations of distribution system in (Yu et al, 2008; Meek et al, 2004; Xiuxia et al, 2006; Kleppinger
et al, 2010). A hybrid approach for reconfiguration has been discussed in (Ahuja et al, 2007). Particle swarm based reconfiguration
for shipboard power system has been presented in (Kumar et al, 2007). Reconfiguraion using genetic algorithm for SPS have been
discussed in (Padamati et al, 2007; Luan et al, 2002; Pendurthi et al, 2009). Ant colony based reconfiguration for distribution
system has been presented in (Daniel et al, 2005) and multi agent based in (Huang et al, 2009). While several methods exist for
terrestrial distribution system, not many techniques exist in literature for reconfiguration of shipboard power system with all the
practical constraints including consideration of islanding or consideration of load priority.
Vuppalapati and Srivastava / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 119-131

120

In this paper, a reconfiguration methodology, using an ant colony optimization algorithm is presented that will find post-fault
optimal configuration meeting designated objectives and system constraints. Graph theory is used to represent the power system
model. As long as a power network is represented with a graph, the corresponding graph matrices based on the graph adjacency
and incidence structures can describe the system topology (Huang, 2006). The matrix specification provides far more efficient
ways of representing a larger or more complicated graph than a pictorial representation. The optimization algorithms are then
applied for power supply restoration and optimal load shedding. The remaining part of the paper is organized as follows. Section 2
describes the graph representation of SPS. Section 3 presents the mathematical problem formulation for reconfiguration. The ant
colony algorithms and developed reconfiguration techniques are described in section 4. Section 5 presents the results for different
fault scenarios applied to eight and 13 bus SPS. Finally, section 6 concludes the paper.

2. Graph Representation of Shipboard Power System

Consider a simple SPS model as shown in figure 1. This system consists of four generators and six loads. Graphical
representation of system is shown in figure 2. In a bus arranged power system, the main components include service switchboards,
buses, cables, generators, and load feeders. These components are interconnected by breakers. Table 1 shows the relationship
between the major components of Shipboard Power System (SPS) and the corresponding graphical elements.

Table 1. Graph Representation of Shipboard Power System Model

Figure 1. A shipboard power system model

Figure 1 shows a SPS model, which consists of six bus switchboards (bus 1, bus2, bus 3, bus 5, bus 6 and bus 7), two cables (bus
4 and bus 8), four generators (G1, G2, G3, and G4), and eighteen breakers (with current transformer). Figure 2 shows the graph
representation of the power system model in figure 1. The direction of the edge is related to the direction of flow of power through
the breaker. The power system is divided into different zones. Each bus with directly connected breakers is defined as a zone in the
protection scheme. In the corresponding system of graph representation, each vertex with the directly connected edges represents a
protection zone (Huang, 2006).
After simplifying a power system topology into a graph, the corresponding matrices can represent the graph mathematically for
computer implementation. This matrix S(ZoneI) is called breaker-to-zone matrix or edge-to-vertex (EtoV) matrix. For the power
system shown in figure 1 the incidence vectors of the matrix EtoV are shown in table 2. The EtoV matrix is an 8x18 matrix
Components In Power System Elements In Graph

Generator, Bus bar, Cable, Load Vertex
Circuit Breaker Edge

121

corresponding to eight zones and eighteen breakers. The positive sign represents the flow of power towards a zone and negative
sign represents flow of power away from the zone.

Figure 2. Graph representation of shipboard power system model

Table 2. Input data for Reconfiguration Algorithm

Additional data required for the system include the breaker type (BRK_TYPE), breaker status (BRK_STATUS), breaker flow
(BRK_FLOW), generator capacity (GEN_CAP), loads and load priority as shown in table 2. There are three types of breaker: 1
refers to the generator breaker, which connects a generator with a zone, 2 refers to the load breaker, which connects a load feeder
with a zone, and 3 refers to the tie breaker, which connects two zones.
The breaker status matrix represents the system breaker open/close status. A 1 means that the breaker is closed (ON), and a 0
means that the breaker is open (OFF). The instantaneous power flow through each breaker can be obtained from the power flow.
The generator capacity (GEN_CAP) matrix is a two-column matrix. The first column contains the index of the generator breaker
and the second column contains the corresponding generator capacity. The instantaneous power flow (BRK_FLOW) matrix and
generator capacity matrix are also shown in table 2 with the unit of MW. The load priority (LOAD_PRIORITY) matrix is another
important piece of information for reconfiguration with possible load shedding. In the load priority matrix shown in table 2, the
first column contains the index of the breaker; the second column contains the load priority, with larger numbers indicating the
higher priority. In the eight bus power system model shown in figure 1, loads 2, 3, and 6 have higher priority than loads 1, 4 and 5.

S(Zone1) = [1, -1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1]
S(Zone2) = [0, 0, 1, -1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
S(Zone3) = [0, 0, 0, 0, -1, -1, 1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
S(Zone4) = [0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0]
S(Zone5) = [0, 0, 0, 0, 0, 0, 0, 0, -1, 1, -1, -1, 0, 0, 0, 0, 0, 0]
S(Zone6) = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, 1, 0, 0, 0, 0]
S(Zone7) = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, -1, 1, 1, 0]
S(Zone8) = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, 1]
BRK_TYPE = [1, 3, 2, 3, 2, 3, 1, 2, 2, 1, 3, 2, 3, 2, 3, 1, 2, 2]
BRK_STATUS = [1, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 1, 1, 0, 1, 1, 1, 1]
BRK_FLOW = [24, 3, 20, 20, 0, 1, 1, 0, 0, 23, 3, 20, 20, 0, 1, 1, -1, 1]
GEN_CAP = [1, 36; 7, 4; 10, 36; 16, 4]
LOADS = [2, 3; 4, 20; 6, 1; 11, 3; 13, 20; 15, 1]
LOAD_PRIORITY = [2, 1; 4, 150; 6, 30; 11, 1; 13, 1; 15, 150]

122


The problem statement can include objectives such as minimization of power loss, maximum power delivery to the loads, and
maximum power delivery to a single load. The objective function considered here is to maximize the power supply to unaffected
loads and supply power to high priority loads through load shedding, if necessary. The loads are classified as non-vital, semi-vital
and vital loads. These loads are multiplied by a given weighting factor
i
P . The weighting factor is selected so that the vital and
semi-vital load contributions are greater than the largest non-vital load contribution in objective function. The objective function to
maximize the load magnitude is defined as follows,

Maximize
n
L L L L + + + + ...
3 2 1

Subject to
load gen
P P
The objective function for maximization of load with given priority includes two objectives with different weighting factors.
These weighting factors are selected depending on the situation whether to restore the loads by considering load priority or without
considering load priority. The magnitude of the objective function depends on switch configuration
x
. The objective function as
given in (Kumar et al, 2007; Padamati et al, 2007) and defined as equation (1),

] ) ( ) 2 ( ) 1 ( [ ] ) ( ) 2 ( ) 1 ( [
2 2 1 1 2 1 n n P n M
L n x P L x P L x P W L n x L x L x W F + + + + + + + = L L (1)

Where,
x is the switch configuration.
x (i) =1 indicates breaker closed for i = 1,2,,n.
x (i) = 0 indicates breaker open for i = 1,2,,n.

n
L L L L ,...,
3
,
2
,
1
are load values.
Pn P P P ,..., 3 , 2 , 1 are priorities of loads.
M
W is weighting factor for load selection by considering load magnitude.

P
W is weighting factor for load selection with considering load priority.

4. Ant Colony Algorithms for Reconfiguration

4.1 Ant colony discrete algorithm: The ant colony algorithm is a probabilistic technique for solving computational problems and
was first proposed by Marco Dorigo in 1992. The procedure can be applied irrespective of the objective function complexity in
most of the cases, making it useful for functions that are highly nonlinear. The first algorithm was aiming to search for an optimal
path in a graph, based on the behavior of ants seeking a path between colony and food.
Initially ants wander randomly and upon finding food they return to their colony while laying down pheromone trails. The
pheromone level evaporates with time, so that other ants can reach the food by following the shortest paths marked with
strongest pheromone quantities. A shortest path by comparison, gets marched over faster, and thus the pheromone density remains
high as it is laid on the path as fast as it can evaporate. An Ant tends to choose a path positively correlated to the pheromone
intensity of founded trails. If there were no evaporation at all, the paths chosen by first ants would be excessively attractive to the
following ones.
The discrete algorithm is a constructive meta-heuristic for the travelling salesman problem (Dorigo et al, 1997) based on this
biological metaphor. It associates an amount of pheromone (i, j) with the connection between two cities. Each ant is placed on a
random start city, and builds a solution going from city to city, until it has visited all cities. The probability that an ant k in a city i
choose to go to a city j is given by equation (2):
[ ] [ ]
[ ] [ ]

Otherwise 0
k(t) allowed j if (t) (t)
(t) (t)
P
k(t) allowed p
ip
ip
ij
ij
k
ij
(2)

123

In this equation, (i, j) is the pheromone between i and j and (i, j) is a simple heuristic function and this is the inverse of the
cost of the connection between i and j. It is the parameter which defines the relative importance of the heuristic information as
opposed to the pheromone evaporation. Once all ants have built a tour, the pheromone trail intensity will be updated. This is done
according to following equations.
) , ( ) ( ) 1 ( ) ( n t t t n t
ij ij ij
+ + = + (3)

The left part of the equation makes the pheromone on all edges decay where as the right part of the equation increases the
pheromone level on all edges visited by all ants. The amount of pheromone an ant k deposits on an edge is defined by:
= +
Otherwise 0
tour its in j) edge(i, uses k ant if
) (
k
k
ij
L
Q
n t (4)

L
k
is the length of a tour created by ant k. In this way, the increase of pheromone of an edge depends on the number of ants that
use this edge. More details on ant colony algorithm have been provided in (Dorigo et al, 2004). The power system optimization
problems are dealt with discrete ant colony optimization in (Daniel et al, 2005 and Hamdaoui et al, 2008). Problem addressed in
this paper is unique due to possible load shedding requirement and the output being a binary value.

4.2 Modified ant colony optimization-I: The discrete algorithm is ordering problem such as traveling salesman problem or
quadratic assignment problem as discussed before. But subset problems are quite different from ordering problems. We need to
select the best subset out of the whole set, possibly satisfying some additional constraints. There is no concept of path here, so it is
difficult to apply the concepts developed in discrete ant colony algorithm in subsection 4.1. For the ordering problems solution is
fixed length, as we have to search for a permutation of a known number of elements. Solutions for subset problems do not have a
fixed length. Thus it is necessary to establish a number, N
max,
which will be used to determine the end of construction cycle for all
the ants (Leiguizamon et al, 1999). Here are the procedural steps:
The intensity of a pheromone trail on item i at time t+N
max
is given by following equation, where N
max
<n is the maximum
of items allowed to be added to some solution by some ant.

) , ( ) ( ) 1 ( ) (
max max
N t t t N t
k
i i i
+ + = + (5)

The increment in the pheromone level in the time t+Nmax is given by summing the contribution of each ant. In the
function G depends on the problem and gives the amount of trail being added to item i. Usually G (L
k
) = Q/L
k
or G
(L
k
) = QL
k
for minimization or maximization problems and Q is a parameter of the method. L
k
is the objective or
fitness function found by k
th
ant .

=
+ = +
n
i
k
i i
N t t N t t
1
max max
) , ( ) , ( (6)

= +
Othewise 0
i item es incorporat k ant if ) G(
) , (
max
k
k
i
L
N t t (7)

The probability of selecting a particular item ip as the next item is given by following equation, where allowed k(t) is
the set of remaining feasible items. Thus, the higher value of
ip
and/or
ip
the more profitable to include item ip in the
partial solution.


124

[ ] [ ]
[ ] [ ]

Otherwise 0
k(t) allowed ip if (t) (t)
(t) (t)
) ( P
k(t) allowed ip
ip
ip
k
ip
t (8)
We consider two types of heuristics, static and dynamic.
Static: It is set at the beginning of run to a fixed value.
Dynamic: It depends on the partial solution.

The subset-based and ordering-based algorithms have many features in common. However, in the ordering based, the pheromone
is laid on the paths while for subset problems no path exists connecting the items. In the ordering problem the more amount of
trail on a particular path, the more profitable is that path. This idea was adapted here in the following way: the more pheromone
trail on a particular item, the more profitable that item is. In other words we move the pheromone from paths to items. At the
same time, a local heuristic is also used in the new version, considering items only instead of connections between them.

4.3 Modified ant colony optimization-2: In this algorithm, the variable selection is the main criteria. There is no concept of path
here, so it is difficult to apply the discrete ACO algorithm over the variable selection. According to information positive feedback
and indirect communication mechanism of ACO, a modified ACO for variable selection is proposed.
For Variable selection problem expressed in binary notation, an ant moves in a binary notation, an ant moves in a N
dimensional search space of N variables, its motion is restricted to either 0 or 1 on each dimension. Every ant selects the variables
depending upon the moving probability of 0 or 1. If the variable is selected then it can be shown by 1 else 0.
In the Variable selection problems the concept of pheromone trail intensity is different from ordering problems; here the
pheromone levels on each dimension are divided into two kinds,
k
i1
and
k
i0
which represent the pheromone of dimension i taking
the value 1 and 0, respectively. The pheromone levels are updated according to the updating rule (Shen et al, 2005).

=
=
=
+ =
=
+ =
m
k
k
i i
i i i
m
k
k
i i
i i i
old new
old new
1
1 1
1 1 1
1
0 0
0 0 0
) ( ) (
) ( ) (

(9)

For each dimension, the intensity of the pheromone at time 0 is set to 0.
i1
= F+F
H
if k
th
ant selected variable i both in the current iteration and in its global best solution
i1
=F if k
th
ant selected variable i only in the current iteration
i1
=F
H
if k
th
ant selected variable i only in its historical global best solution.
i0
=F+ F
H
if variable i was not selected by ant k in the current iteration either in its historical global best solution
i0
= F if variable i was not selected by ant k in current iteration
i0
= F
H
if variable i was not selected by ant k in its historical global best solution

Where F and F
H
are defined by fitness function. For improving the convergence velocity, the information F
H
which corresponds
to the historical global best result of the i
th
ant was introduced to the increment of pheromone. An ant K makes a decision
concerning the variable selection according to the pheromone amount. The moving probability is given by equation 10:

k
i
k
i
k
i k
i
P
0 1
1

+
= (10)


125

Figure 3. Flow chart of reconfiguration procedure using ACO

In the modified ACO, m ants select variables from all N variables, according to the probability. After selection, the amount of
pheromone is updated. This process is iterated until the minimum error criterion is attained or the number of iteration reaches a
user defined limit.
To apply developed ant colony technique for reconfiguration, power system need to be simplified using graph theory. After
simplifying a power system topology into a graph, the corresponding matrices can represent the graph mathematically for
computer implementation. These matrices include breaker-to-zone matrix, breaker status matrix, generator capacity matrix, load
priority matrix and power flow matrix. When the power system experiences fault(s), the fault detection functions will quickly
locate the fault and issue a trip signal to the breakers surrounding the faulted zone to isolate the fault. Then the breaker-to-zone
matrix will be updated by removing the isolated zones and tripped breakers. The zone balance matrix will be calculated to find out
any negative zones. The path searching function is then activated to find out possible power supply paths for these negative zones.
All the loads and generators are extracted along the path and then ACO is used for restoration and load shedding if necessary. The
flow chart of the reconfiguration process is shown in figure 3.

6. Results
The reconfiguration process using ACO has been developed and implemented in MATLAB. The algorithms have been tested on
an 8-bus power system model and on 13-bus power system model with distributed generators. Loads are served on the basis of
load magnitude only, load priority only, as well as both load magnitude and load priority. For 8 bus power system shown in figure
1, loads 2, 3, and 6 have higher priority than loads 1, 4 and 5. Several test case scenarios have been created by simulating fault at
different buses or combinations of buses to make it total of 8 test case scenarios.


126

Table 3. Based on Load Magnitude for 8-bus SPS

In the table 3, for a particular fault bus, the negative bus formed is shown with the possible power paths and corresponding load
shedding. Amount of load served in MW and the breaker reconfiguration showing open and close status are also shown.
Table 4 shows the test results for eight bus power system with consideration of load priority. In case 1, when the fault happened
on bus 1 and protection system isolate bus 1 from rest of the system, bus 2 had negative power balance of -20MW, which means
that bus 2 imported 20MW power from faulted bus 1 before fault occurred. Then, the reconfiguration algorithm started to search
possible paths with non-negative power balance for the negative power bus 2. The only possible path found was B2-B3-B4-B5-
B6-B7-B8 using graph theory. All the loads and generators are extracted along this path using the ant colony algorithm to restore
the loads based on the priority and the available generation. In this case the generation was found to be 44 MW and the total load
was 46 MW. So load 4 (L4) was shed because it has lowest priority compared to other loads and the remaining loads are restored
along the path. In case 6, two faults happened on bus 1 and bus 5. After fault isolation, bus 2 and bus 6 both have a power deficit.
In this case two islands were formed corresponding to two negative buses. Possible power supply paths are B2-B3-B4 and B6-B7-
B8. Ant Colony algorithm was applied separately for each island and loads L2 and L5 were shed even though they have the
highest priority because there was not enough power generation. It can be seen from table 3 and 4, that there are not much
differences in results obtained for load magnitude and load priority. The reason for similarity between these results are less
available options for optimization given small test system with only two generators and two distributed generations.

Table 4. Based on Load Priority for 8-bus SPS
Test
Case
Faulted
Bus
Number
Negative
power bus
Possible
power
supply path
Load
Shedding
Breaker
Reconfiguration
MW
Served
Case1 B1 B2 B2-B3-B4-B5-B6-B7-B8 L4
BK 11, 1, 2, 3, 18 (O)
BK 5, 8, 9, 14 (C)
44
Case2 B3 No No No BK 6,7 (O) 46
BK 2 , 10, 11, 12 (O)
BK 14, 5,8 (C)
44
Case4 B7 No No No BK 15, 16, 17 (O) 46
Case5 B1, B3 B2
No possible
generation
L2 BK 4, 1, 2, 3, 18, 6, 7 (O) 24
Case6 B1, B5
B6
B2
B6-B7-B8
B2-B3-B4
L2
L5
BK 4, 13 , 1, 2 ,3 ,18, 10, 11, 12(O)
BK 5, 8, 14 (C)
1
1
Case7 B3, B7 No No No BK 6,7 15, 16, 17 (O) 44
Case8 B5, B7 B6
No possible
generation
L5 BK 13, 10, 11, 12, 15, 16, 17 (O) 24

Test
case
Faulted
Bus
Number
Negative
power bus
Possible power supply path
Load
Shedding
Breaker
reconfiguration
MW
Served
BK 11, 1, 2, 3, 18 (O)
BK 5, 8, 9, 14 (C)
44
BK 2, 10, 11, 12 (O)
BK 14, 5,8 (C)
44
Case4 B7 No No No BK 15, 16, 17 (O) 47
Case5 B1, B3 B2 No possible generation L2 BK 4, 1, 2, 3, 18, 6, 7(O) 24
Case6 B1, B5
B6
B2
B6-B7-B8
B2-B3-B4
L2
L5
BK 4, 13, 1, 2 ,3 ,18, 10, 11, 12(O)
BK 5, 8, 14 (C)
2
2
Case7 B3, B7 No No No BK 6,7, 15, 16, 17 (O) 44
Case8 B5, B7 B6 No possible generation L5 BK 13, 10, 11, 12, 15, 16, 17 (O) 24

127

Table 5. Based on Load Magnitude and Priority for 8-bus SPS

Test results for simulations objective of maximizing the load magnitude and priority for eight bus SPS. As shown in table 3, 4
and 5, those results for this particular scenario is similar to reconfiguration with load priority.

Figure 4. 13-bus Shipboard Power System with Distributed Generators
Test
Case
Faulted
Bus
Number
Negative
power
bus
Possible power
supply path
Load
Shedding
Breaker
reconfiguration
MW
Served
BK 15, 1, 2, 3, 18 (O)
BK 5, 8, 9, 14 (C)
44
BK 15, 10, 11, 12 (O)
BK 14, 5,8 (C)
44
Case4 B7 No No No BK 15, 16, 17 (O) 46
Case5 B1, B3 B2
No possible
generation
L2 BK 4, 1, 2, 3, 18, 6, 7 (O) 24
Case6 B1, B5
B6
B2
B6-B7-B8
B2-B3-B4
L2
L5
BK 4, 13 , 1, 2 ,3 ,18, 10, 11, 12(O)
BK 5, 8, 14 (C)
1
1
Case7 B3, B7 No No No BK 6,7 15, 16, 17 (O) 44
Case8 B5, B7 B6
No possible
generation
L5 BK 13, 10, 11, 12, 15, 16, 17 (O) 24

128

Developed reconfiguration algorithm was also tested on bigger shipboard power system with thirteen buses. This system has 5
generators and 3 distributed generators as shown in figure 4. There are 13 loads with different MW value in the system. There are
more number of possible options to reconfigure the network in this system compared to eight bus power system shown in figure 1.
Table 6 shows the priority preferences for all the loads for thirteen bus SPS.

Table 6. Load Priority for 13-bus SPS
Load Priority
Load 1 Semi-vital
Load 2 Non-vital
Load 3 Semi-vital
Load 4 Non-vital
Load 5 Semi-vital
Load 6 Non-vital
Load 7 Vital
Load 8 Non-vital
Load 9 Vital
Load 10 Semi-vital
Load 11 Non-vital
Load 12 Semi-vital
Load 13 Non-vital

Test case 2, 13 bus SPS was represented by graph theory in matrices form similar to eight bus power system to apply ACO
reconfiguration algorithm. Table 7, 8 and 9 shows the simulation results for optimization based on load priority, load magnitude
and simultaneous optimization of load magnitude and load priority.

Table 7. Based on Load Priority for 13-bus SPS
Test
case
Faulted
Bus
Number
Negative
power
bus
Possible power
supply path
Load
Shedding
Breaker
Reconfiguration
MW
Served
Case 1 B1 B2, B13
B2-B3-B4-B5-B6-B7-B8-
B9-B10-B11-B12-B13
L3
BK 1, 2, 3, 6, 32 (O)
BK 15, 17, 23, 28, 7, 16, 30 (C)
36
Case 2 B4 B3, B5
B3-B2-B1-B13-B12-B11-
B10-B9-B8-B7-B6-B5
L2, L13
BK 4, 8, 9, 10, 11,29 (O)
BK 28, 33, 17, 15, 7, 16, 30 (C)
71
Case 3 B8 B9
B9-B10-B11-B12-B13-
B1-B2-B3-B4-B5-B6-B7
L2, L4, L6, L10,
L11,L13
BK 4, 9, 34, 24 27, 29, 18, 19, 20
(O)
BK 23, 28, 15, 7, 16, 30 (C)
65.9
Case 4 B9 No No No BK 20, 21, 22 (O) 49
Case 5 B10 B11 B11-B12-B13-B1 No BK 24, 25, 26 (O) 72
Case 6 B1, B4
B2
B3
B5
B13
B2-B3

B5-B6-B7-B8-B9
B13-B12-B11-B10
L2,L3,L6,L7,L8,L9,L10,
L11,L12,L13
BK 4, 6, 1, 2, 3, 32, 8, 9, 10,
11,34,14,19,25,24,27,33,29 (O)

0.5
Case 7 B1, B8
B2
B9
B13
B2-B3-B4-B5-B6-B7
B9-B10-B11-B12-B13
L3
L9
BK 6, 21, 1, 2, 3, 32, 18, 19, 20 (O)
BK 7, 15, 16, 23, 28, 30 (C)

9
Case 8 B4, B10
B3
B5
B11
B3-B2-B1-B13-B12-B11
B5-B6-B7-B8-B9
L3,L8,L13
BK 6, 8, 9, 10, 11 , 24, 25, 26,19,29
(O)
BK 7, 15, 16, 28, 30 (C)
32


129

Table 8. Based on Load Magnitude for 13-bus SPS

As shown in the table, optimization with load magnitude has more MW served compared to MW served with load priority. It is
also evident from these results that high priority loads do not get shed in reconfiguration algorithm as shown in table 7. Similar to
eight bus test cases, several test case scenarios have been created by simulating fault at different buses or combinations of buses to
make it total of 8 test case scenarios for 13-bus test case. Numbers of negative power buses are generally more for multiple bus
faults, as expected.

Table 9. Based on Load Magnitude and priority for 13-bus SPS
Test
case
Faulted
Bus
Number
Negative
power
bus
Possible power
supply path
Load
Shedding
Breaker
reconfiguration
MW
Served
Case1 B1 B2, B13
B2-B3-B4-B5-B6-B7-B8-
B9-B10-B11-B12-B13
L2,L4,L5,L6,L7,L8,L10,
L11,L12,L13
BK 1, 2, 3,
32,4,9,12,34,14,19,24,27,33 (O)
BK 15, 17, 23, 28, 7, 16, 30 (C)
62
Case2 B4 B3, B5
B3-B2-B1-B13-B12-B11-
B10-B9-B8-B7-B6-B5
L12, L6, L13
BK 33, 34, 29, 9 10, 11 (O)
BK 28, 23, 17, 15, 7, 16, 30 (C)
71.1
Case3 B8 B9
B9-B10-B11-B12-B13-B1-
B2-B3-B4-B5-B6-B7
L1, L5, L7, L12, L13
BK 2,14,33,17, 29, 18, 19, 20 (O)
BK 23, 28, 15, 7, 16, 30 (C)
66.1
Case4 B9 No No No BK 20, 21, 22 (O) 49
Case5 B10 B11 B11-B12-B13-B1 No BK 24, 25, 26 (O) 72
Case6 B1, B4
B2
B3
B5
B13
B2-B3
B5-B6-B7-B8-B9
B13-B12-B11-B10
L2,L3,L5,L7,L8,L9,L10,
L11,L12,L13
BK 4, 6, 1, 2, 3, 32, 8, 9, 10,
11,12,14,19,25,24,27,33,29 (O)

0.5

Case7 B1, B8
B2
B9
B13
B2-B3-B4-B5-B6-B7
B9-B10-B11-B12-B13
L3
L9
BK 6, 21, 1, 2, 3, 32, 18, 19, 20
BK 7, 15, 23, 28, 30 (C)

5
4
Case8 B4, B10
B3
B5
B11
B3-B2-B1-B13-B12-B11
B5-B6-B7-B8-B9
L1,L2,L3,L6,L12
BK 6, 8, 9, 10, 11 , 24, 25,
26,2,4,34,33 (O)
BK 7, 15, 16, 28, 30 (C)
31.1
Test
case
Faulted
Bus
Number
Negative
power
bus
Possible power
supply path
Load
Shedding
Breaker
reconfiguration
MW
Served
Case1 B1 B2, B13
B2-B3-B4-B5-B6-B7-B8-B9-
B10-B11-B12-B13
L3
BK 6, 1, 2, 3, 32 (O)
BK 15, 17, 23, 28, 7, 16, 30 (C)
36
Case2 B4 B3, B5
B3-B2-B1-B13-B12-B11-
B10-B9-B8-B7-B6-B5
L2,L13
BK 4,29, 8, 9, 10, 11 (O)
BK 28, 33, 17, 15, 7, 16, 30 (C)
71
Case3 B8 B9
B9-B10-B11-B12-B13-B1-
B2-B3-B4-B5-B6-B7
L2, L4, L6,L10,
L11,L13
BK 4, 9, 34,24,27,29, 18, 19, 20 (O)
BK 23, 28, 15, 7, 16, 30 (C)
65.9
Case4 B9 No No No BK 20, 21, 22 (O) 49
Case5 B10 B11 B11-B12-B13-B1 No BK 24, 25, 26 (O) 72
Case6 B1, B4
B2
B3
B5
B13
B2-B3

B5-B6-B7-B8-B9
B13-B12-B11-B10
L2,L3,L6,L7,L8,L9,
L10,L11,L12,L13
BK 4, 6, 1, 2, 3, 32, 8, 9, 10,
11,34,14,19,25,24,27,33,29 (O)

0.5
Case7 B1, B8
B2
B9
B13
B2-B3-B4-B5-B6-B7
B9-B10-B11-B12-B13
L3
L9
BK 6, 21, 1, 2, 3, 32, 18, 19, 20 (O)
BK 7, 15, 23, 28 (C)

9
Case8 B4, B10
B3
B5
B11
B3-B2-B1-B13-B12-B11
B5-B6-B7-B8-B9
L3,L2,L8,L13
BK 6, 8, 9, 10, 11 , 24, 25, 26,4,19,29
(O)
BK 7, 15, 16, 28, 30 (C)
31

130

Similar to eight bus test cases, test results for simulations objective of maximizing the load magnitude and priority is similar to
reconfiguration with load priority. There are several other test cases scenarios for different fault scenarios are also simulated and
satisfactorily results were obtained. Those results are not presented here and only 8 test scenarios have been shown in these tables.

7. Conclusions
The reconfiguration problem for shipboard power system was investigated using novel application of ant colony and graph
theory. Graph theory is used to represent the shipboard power system and ant colony was used to maximize the load magnitude or
load priority. The mathematical problems were formulated to maximize the objective function considering (i) load magnitude only
(ii) load priority only (iii) both load magnitude and load priority. Developed ant colony optimization technique has been tested for
two test cases under several fault scenarios. Simulation results obtained are satisfactory and proposed method can be easily
extended for application to bigger distributed power system.

Acknowledgement
Authors would like to thank financial support provided by the Office of Naval Research (ONR) under Grant N00014-02-1-0623.

References

Ahuja, A. Das, S., and Pahwa A. 2007. An AIS-ACO hybrid approach for multi-objective distribution system reconfiguration.
IEEE Transactions on Power Systems. Vol. 22, No. 3, pp. 1101-1111.
Bulter K. L., Sarma N. D. R. 2000. General reconfiguration methodology for AC radial shipboard power systems. IEEE 2000
Power Engineering Society Winter Meeting. pp. 1226-1230.
Bulter K. L., Sarma N. D. R. and Prasad V. R. 1999. A new method of network reconfiguration for service restoration in shipboard
power systems. Proc. of IEEE Power Engineering Society Transmission and Distribution Conf, pp. 658-662.
Bulter K. L., Sarma N. D. R., and Prasad V. R. 2001. Network reconfiguration for service restoration in shipboard power
distribution system. IEEE Transactions on Power System. Vol. 16, No. 4, pp. 653-661.
Daniel C. L., Khan, H. I., and Ravichandran S. 2005. Distribution network reconfiguration for loss reduction using ant colony
algorithm. Annual IEEE India Conference (Indicon), pp. 619-622.
Davey K. R., and Hebner, R. E. 2005. Reconfiguration: A tool for designing new ships. IEEE Electric Ship Technologies
Symposium (ESTS). Philadelphia, Pennsylvania, U.S.A.
Dorigo M. and Stutzle, T. 2004. Ant Colony Optimization. A Bradford book. MIT press, Cambridge, MA, USA.
Dorigo, M., and Gambardella, L. M. 1997. Ant colony system: a cooperative learning approach to the traveling salesman problem.
IEEE transcations on Evolutionary Computation. Vol. 1, No. 1, pp. 53- 66.
Hamdaoui H., Hadjeri, S. and Zeblah, A. 2008. A new constructive method for electrical power system using ant colony. Leonardo
Electronic Journal of practices and Technologies. Vol. 7, No. 12, pp. 49-60.
Haung Y. 2006. Fast Reconfiguration Algorithm Development for Shipboard Power Systems. M.S. thesis, Dept. of Electrical Eng.
Mississippi State University, MS, USA.
Huang K, Srivastava S.K., Cartes D.A., and Sun L.-H. 2009. Market-based multiagent system for reconfiguration of shipboard
power systems. Electric Power Systems Research, Vol. 79, No. 4, pp. 550-556.
Jayabalan, R., and Fahimi, B. 2005. Naval shipboard power system. IEEE Vehicle Power and Propulsion Conference. Chicago, IL.
Kleppinger D., Broadwater R., and Scirbona C. 2010. Generic reconfiguration for restoration. Electric Power Systems Research.
Vol. 80, issue 3, pp. 287-295.
Kumar N., Srivastava, A. K., and Schulz, N. N. 2007. Shipboard power system restoration using binary particle swarm
optimization. 39
th
North American Power Symposium, (NAPS). pp. 164-169.
Leiguizamon G. and Michalewicz Z. 1999. A new Version of ant system for subset problems. Proceedings of the congress on
Evolutionary Computation. Vol. 2.
Luan W. P., Irving, M. R., and Daniel, J. S. 2002. Genetic algorithm for supply restoration and optimal load shedding in power
system distribution networks. IEE Proceedings of Generation, Transmission and Distribution. Vol. 149, No.2, pp. 145-151.
Meek, C. D. 2004. Investigation of optimization methods used for reconfiguration of the naval electric ship power system. M. S.
thesis, University of Texas, Austin.
Padamati, K. R., Schulz N. N., and Srivastava A. K. 2007. Application of genetic algorithm for reconfiguration of shipboard power
system. 39
th
North American Power Symposium, (NAPS). pp. 159-163.
Pendurthi, V.K., Schulz, N.N., Doane, S., and Srivastava, A.K. 2009. Cognitive engineering studies of DSS and dealing with
uncertainty in load for real-time adaptive power system reconfiguration. IEEE Electric Ship Technologies Symposium. pp. 79
85.

131

Shen Q., Jiang, J. H., Tao, J. C., Shen, G. L., and Yu R. Q. 2005. Modified ant colony optimization algorithm for variable selection
in QSAR modeling: QSAR studies of cyclooxygenase inhibitors. Journal of Chem. Inf. Model. Vol. 45, pp. 1024-1029.
Xiuxia Y., and Yi, Z. 2006. Study of shipboard power system intelligent network reconfiguration. Proceedings of the Sixth
International Conference on Intelligent Systems Design and Applications.
Yu Q., Solanki, J., Padmati, K., Kumar, N., Srivastava, A., Bastos, J.L., and Schulz, N. N. 2008. Intelligent methods for
reconfiguration of terrestrial and shipboard power systems. IEEE Power and Energy Society General Meeting, Tampa, FL,
USA.

Biographical notes:

Sri Hari Krishna Vuppalapati is currently pursuing his B.Tech final year in Electrical Engineering from Indian Institute of Technology, Kharagpur, West Bengal,
India. He has done research projects as summer internship at Mississippi State University, Mississippi State, USA in 2009. His research interests include
applications of optimization techniques to electrical engineering problems and power system reconfiguration.

Dr. Anurag K. Srivastava received his Ph.D. degree from Illinois Institute of Technology (IIT), Chicago, in 2005, M. Tech. from Institute of Technology, India in
1999 and B. Tech. in Electrical Engineering from Harcourt Butler Technological Institute, India in 1997. He is working as Assistant Research Professor at
Mississippi State University since September 2005. Before that, he worked as research assistant and teaching assistant at IIT, Chicago, USA and as Senior Research
Associate at Electrical Engineering Department at the Indian Institute of Technology, Kanpur, India as well as Research Fellow at Asian Institute of Technology,
Bangkok, Thailand. His research interest includes power system security, power system modeling and real time simulation, electricity market and artificial
intelligent application in power system. Dr. Srivastava is member of IEEE, IET, IEEE power and energy society, Sigma Xi, ASEE and Eta Kappa Nu. He serves as
vice-chair of power engineering career promotion subcommittee and secretary of student activities subcommittee within IEEE power and energy society. He is
recipient of several awards and serves as reviewer for several international journals and conferences.

Accepted March 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 132-149

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Multi-area economic dispatch with tie-line constraints employing
evolutionary approach

Manisha Sharma, Manjaree Pandit*, Laxmi Srivastava

1*
Department Department of Electrical Engineering
M.I.T.S., Gwalior, INDIA
*
Corresponding Author: e-mail: manjaree_p@hotmail.com Tel.+91- 0751-2665962, +91-0751-2409380

Abstract

Economic dispatch (ED) has the objective of generation allocation to the power generators in such a manner that the total fuel
cost is minimized while all operating constraints are satisfied. For the sake of simplicity, ED is generally solved without
accounting for transmission constraints. However, in deregulated power system environment it is essential to model the ED
problem for practical multi-area cases with tie line constraints. Most of the conventional gradient based methods are time
consuming, suffer from dimensionality problem and assume the fuel cost curves of generating units to be piecewise linear,
monotonically increasing in nature. The resulting dispatch solutions are therefore inaccurate; sometimes producing infeasible
solutions for modern generating units having non convex cost curves. On the other hand evolutionary methods do not suffer
from convexity assumptions and achieve fast solutions even for complex non-linear, non-convex, multi-modal optimization
problems. This paper reviews and compares some evolutionary techniques for multi-area economic dispatch (MAED).
The paper presents an extensive comparison of the search capability and convergence behavior of i) Classical differential
evolution (DE) and its various strategies ii) Classical particle swarm optimization (PSO) and iii) An improved PSO with a
parameter automation strategy having time varying acceleration coefficients (PSO_TVAC) for solving MAED problems for two
area and three area test power systems with 4, 10 and 40- generating units. The results are found to be superior compared to
some recently published results.

Keywords: Differential evolution, Multi-area economic dispatch, multiple fuel options, particle swarm optimization,
transmission capacity constraints. Time varying acceleration coefficients (TVAC), Valve point loading effects.

1. Introduction

The objective of Economic dispatch (ED) is to allocate power generation among available generators in the most economical
manner, while satisfying the physical and operational constraints. The cost of power generation, particularly in fossil fuel plants, is
very high and economic dispatch helps in saving a significant amount of revenue. Conventional methods like lamda iteration, base
point participation factor, gradient methods etc. rely heavily on the convexity assumption of generator cost curves and hence
approximate these curves using quadratic or piecewise quadratic, monotonically increasing cost functions (Wood et al., 1984). In
actual practice however, this assumption is not valid because the cost functions exhibit higher order non-linearities and
discontinuities due to prohibited operating zones (POZ), ramp rate limits and valve point loading effects (Walter et al., 1993,
Orero et al., 1996). In the practical ED, the cost function must be expressed as a piecewise non-linear function in place of a single
quadratic function. Therefore, ED problem with valve point effects gives rise to a non smooth optimization problem with heavy
equality and inequality constraints, having complex and nonconvex characteristics with multiple minima, which make the
challenge of obtaining the global minima very difficult. Most traditional methods fail for this NCED problem except dynamic
programming (Shoults et al., 1986) in which no restriction is imposed on the shape of cost curves, but this method suffers from the
problem of dimensionality and excessive evaluation at each stage.
Power utilities try to achieve high operating efficiency to produce cheap electricity. Competition exists in the electricity supply
industry in generation and in the marketing of electricity. The operating cost of a power pool can be reduced if the areas with more
economic units generate larger power than their load, and export the surplus power to other areas with more expensive units. The
Sharma et al. / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 132-149

133

benefits thus gained will depend on several factors like the characteristics of a pool, the policies adopted by utilities, types of
interconnections, tie-line limits and load distribution in different areas. Therefore, transmission capacity constraints in production
cost analysis are important issues in the operation and planning of electric power systems.
The economic dispatch problem is frequently solved without considering transmission constraints. Compared to the classical
ED problem the MAED problem is more complex due to the additional tie-line constraints and area power balance requirements.
However, some researchers have taken transmission capacity constraints into consideration. A complete formulation of multi-area
generation scheduling with import/export constraints was presented in a reference (Shoults et al., 1980). Desell et al. (1984)
proposed an application of linear programming to transmission constrained production cost analysis. Farmer et al. (1990)
presented a probabilistic method which was applied to the production costing of transmission constrained multi-area power
systems. Hopfield neural network based approach was proposed to solve the MAED problem (Yalcinoz et al., 1998). Doty and
McIntyre et al. (1982) solved multi-area economic dispatch problem by using spatial dynamic programming and optimal results
were reported considering transmission constraint with linear losses. Linear programming application is proposed in Desell et al.
(1984) to production cost analysis with transmission constraint. Area control error is solved in multi-area economic dispatch
(Hemick et al., 1985). Wang and Shahidehpour et al. (1992) proposed a decomposition approach for solving multi-area generation
scheduling with tie-line constraints using expert systems. They presented efficiency of decomposition approach by testing it on a
four area system with each area consisting of 26 units. The Newton-Rapshon's method is applied to solve multi-area economic
dispatch problem (Wernerus et al., 1995) by calculating short range margin cost based prices. An incremental network flow
programming algorithm was proposed for the MAED solution with tie-line constraints (Streifferet et al., 1995). The MAED is
solved by the direct search method with considering transmission constraint (Chen et al., 2001). Evolutionary programming is
proposed in Jayabarathi et al. (2000) for multi-area economic dispatch problem. Recently covariance matrix adapted evolutionary
strategy has been proposed for MAED problems where a Karush Kuhun Tucker (KKT) optimality based stopping criterion is
applied to guarantee optimal convergence (Manoharan et al., 2009). Determining the most economical fuel to burn poses another
optimization challenge for generators with multiple fuel options. In such cases the fuel cost curve is represented as a segmented
piecewise quadratic function similar to the valve point loading effects (Park et al., 1993). The economic dispatch problem with
multiple areas and multiple fuel options translates into a nonconvex optimization problem with complex constraints. Such
optimization problems require algorithms which avoid approximation of cost function and still do not require large computational
time. The methods found suitable include tabu search, simulating annealing, neural networks (Yalcinoz et al., 1998; Park et al.,
1993), genetic algorithm (Orero et al., 1996) , particle swarm optimization (Chaturvedi et al., 2009), harmony search (Vasebi et
al., 2007), ant colony optimization (Song et al., 1999), bacterial foraging, (Panigrahi et al., 2009), artificial immune system
(Vanaja et al., 2008) and differential evolution (DE) (Coelho et al., 2006). Among these techniques, PSO, DE and their variants
have been extensively popular due to their superior convergence characteristics, consistency and ease of implementation.
Although these methods do not always guarantee global best solutions, they often achieve a fast and near global optimal
solution. Researches have constantly observed that all these methods very quickly find a good local solution but get stuck there for
a number of iterations without further improvement sometimes causing premature convergence. Time varying acceleration
coefficients (TVAC), (Chaturvedi et al., 2009) are employed countering the effect of premature convergence in PSO. The TVAC
strategy strikes a proper balance between the cognitive and social component during the initial and latter part of the search and
hence is found to avoid premature convergence of the swarm. The paper aims to test the potential of all the basic DE variants in
producing feasible solutions for the MAED problem formulated with many different constraints. The paper also compares the
solution quality of DE variants with the PSO_TVAC strategy with classical PSO. The results of all three evolutionary strategies
are found to be feasible and superior to reported results (Yalcinoz et al., 1998; Chen et al., 2001; Manoharan et al., 2009).

2. Multi-area Economic Dispatch with Multiple fuel Options

The objective of MAED is to determine the generation levels and the power interchange between areas which would minimize
total fuel costs in all areas while satisfying power balance, generating limit and transmission capacity constraints. If an area with
excess power is not adjacent to a power deficient area, or the tie-line between the two areas is at the transmission limit, it is
necessary to find an alternative path between these two areas in order to transmit additional power.
The generator cost function is obtained from the data taken during the heat-run tests, in which the input-output data is measured
to cover the operating region. Large turbine generators usually have a number of fuel admission valves which are opened one by
one when the unit is called upon to increase production. When a valve is opened, the throttling losses increase rapidly as a result of
which, the incremental heat rate rises suddenly. The valve-point effects introduce ripples in the heat-rate curves and make the
objective function discontinuous, nonconvex and with multiple minima. The fuel cost of the i
th
unit can be calculated as.
( ) ( ) ( )
i i i i i i i i i i i
P P f e c P b P a P F + + + =
min 2
sin (1)
When the generating units are supplied with multiple fuel sources, the cost of each unit is represented with several piecewise
quadratic functions reflecting the effects of fuel changes; the generator has to identify the most economic fuel to burn from the
options available. The fuel cost function for such a case is represented as in Lin et al. (1984).


134

(2)

where ,
mik
a ,
mik
b and
mik
c are the fuel-cost coefficients of the i
th
unit, of m
th
area and k = 1, 2,. . ., K are the available fuels. The
ED problem is to determine the generated powers Pi of units for a total load of PD so that the total fuel cost for the N number of
generating units is minimized subject to the power balance constraint and unit upper and lower operating limits. Taking into
consideration the cost of transmission though each tie-line, the objective function of multi-area economic dispatch is given in Eq.
(3) as

=
+ =
N
i
M
j
M j j i i
T f P F MinF
T
1
1
) 1 (
) (
(3)
f
i
is the cost function associated with tie line power flow from area j to area (M-1).
i)Area Power Balance Constraints
The power balance constraints for area m neglecting losses can be given as
0
1
1
) 1 (
=
+ =

=
N
i
M
j
M j Dm im
T P P
(4)
for m=1,2..M (areas). P
Dm
is the load demand in m
th
area and T
j
represents the tie line flows to the j
th
area from other areas.
ii)Generating Limit Constraints
The power output of a unit must be allocated within the range bounded by its lower and upper limits of real power generation as
given by
N i P P P
i i i
,..., 2 , 1
max min
=
(5)
iii)Tie-line Limit Constraints
The tie line power flows to area j should be between the maximum and minimum
M j T T T
M j M j M j
,..., 2 , 1
max
) 1 ( ) 1 (
min
) 1 (
=

(6)
where T
j
is the power flow through the tie line.

3. Review of Evolutionary Strategies

During the last decade different Evolutionary strategies have increasingly been applied by researchers for solving economic
dispatch problems (Walter et al., 1993; Orero et al., 1996; Jayabarathi et al., 2000; Coelho et al., 2006; Sinha et al., 2003) due to
their powerful search capability and ability to handle different types of cost functions. Out of the different evolutionary techniques
proposed PSO and DE have emerged as the most popular, looking at the number of papers published during the past few years.
The present paper aims to present a brief review and comparison of both the techniques and their different variants using
performance metrics such as convergence behavior, consistency and solution quality for solving the MAED problem with multiple
fuel options.

3.1 Classical PSO

A PSO is a population based modern heuristic search method that traces its evolution to the emergent motion of a flock of birds
searching for food. It scatters random particles i.e. solutions into the problem space. These particles, called swarms, collect
information from each other through their respective positions. The particles update their positions using their own experience and
the experience of their neighbors. The update mode is termed as the velocity of particles. The position and velocity vectors of the
i
th
particle of a d-dimensional search space can be represented as ) ... ,......... , (
2 1 id i i i
x x x X = and ) ,........ , (
2 1 id i i i
v v v V =
respectively. On the basis of the value of the evaluation function, the previous best position of a particle is recorded and
represented as ) ........ , (
2 1 id i i i
p p p pbest = If the g
th
particle is the best among all particles in the group so far, it is represented
as ) . ,......... , (
2 1 gd g g g
p p p gbest pbest = = The particle tries to modify its position using the current velocity and the
+ +
+ +
+ +
=
k fuel P P P if C P b P a
fuel P P P if C P b P a
fuel P P P if C P b P a
P F
mi mi mik mik mi mik mi mik
l mi l mi mi mi mi mi
l mi mi mi mi mi mi mi
mi
... , ..... .....
.....
.....
2 .. , ..... .....
1 .. , ...... .......
) (
(max) 1 ,
2
2 1 , 2 2
2
2
1 (min) , 1 1
2
1

135

distance from pbest and gbest and .The modified velocity and position of each particle for fitness evaluation in the next iteration
are calculated using the following equations:
( )
id id
k
id
k
id
x pbest rand c v w C v + =
+
1 1
1
[ ( )]
2 2 id gd
x gbest rand c + (7)
1 1 + +
+ =
k
id id
k
id
v x x (8)
Here w is the inertia weight parameter, C is constriction factor, c
1
,c
2
are cognitive and social coefficients,rand
1
and rand
2
are
random numbers between 0 and 1. A large inertia weight helps in good global search while a smaller value facilitates local
exploration. Therefore, the practice is to use larger inertia weight factor during initial exploration and gradual reduction of its value
as the search proceeds in further iterations. The time varying inertial weight is given by
( )
( )
min
max
max
min max
w
iter
iter iter
w w w +
=
(9)
where iter is the current iteration number while iter
max
is the maximum number of iterations. Usually, the value of w is varied
between 0.9 and 0.4.

3.2 PSO with Time-Varying Acceleration Coefficients (PSO_TVAC)

Though the PSO technique with time varying inertia weight can locate good solution at a significantly fast rate, its ability to fine
tune the optimum solution is weak, mainly due to the lack of diversity at the end of the search. It has been observed by most
researchers that in PSO, problem-based tuning of parameters is a key factor to find the optimum solution accurately and
efficiently. (Kennedy and Eberhart et al., 1995) stated that a relatively higher value of the cognitive component, compared with the
social component, results in roaming of individuals through a wide search space. On the other hand, a relatively high value of the
social component leads particles to a local optimum prematurely. In population-based optimization methods, the policy is to
encourage individuals to roam through the entire search space during the initial part of the search, without clustering around local
optima. During the latter stages, convergence towards the global optima is encouraged, to find the optimal solution efficiently. The
idea behind TVAC is to enhance the global search in the early part of the optimization and to encourage the particles to converge
towards the global optima at the end of the search. This is achieved by changing the acceleration coefficients c
1
and c
2
with time in
such a manner that the cognitive component is reduced while the social component is increased as the search proceeds. With a
large cognitive component and small social component at the beginning, particles are allowed to move around the search space
instead of moving toward the population best during early stages. On the other hand, a small cognitive component and a large
social component allow the particles to converge to the global optima in the latter part of the optimization process. The
acceleration coefficients are expressed as (Ratnaweera et al., 2004):
( )
i i f
c
iter
iter
c c c
1
max
1 1 1
+ =
(10)
( )
i i f
c
iter
iter
c c c
2
max
2 2 2
+ = (11)
where c
1i
, c
1f
, c
2i
and c
2f
are initial and final values of cognitive and social acceleration factors respectively.

3.3 Differential evolution

DE is a population-based stochastic function minimizer (or maximizer) based on evolutionary computation, whose simple yet
powerful and straightforward features make it very attractive for numerical optimization. DE differs from conventional genetic
algorithms in its use of perturbing vectors, which are the difference between two randomly chosen parameter vectors, a concept
borrowed from the operators of Nelder and Meads simplex optimization technique. The DE algorithm was first introduced by
Storn and Price et al. (1995) and was successfully applied in the optimization of some well-known nonlinear, non-differentiable,
and non-convex functions. DE works on three basic operations, namely mutation, crossover and selection.
Mutation is an operation that adds a vector differential to a population vector of individuals according to the chosen variant. The
different variants of DE are classified using the notation DE /// where indicates the method for selecting the parent
chromosome that will form the base of the mutated vector, indicates the number of difference vectors used to perturb the base
chromosome, and indicates the recombination mechanism used to create the offspring population. Most papers have explored the
variant DE / rand / 1 / bin (Coelho et al., 2006). The best performing variant is found to be problem specific and needs detailed
investigation. The donor or mutant vector for each population member is generated for different variants in classic DE as given
below
1) DE/rand/1
)] ( ) ( [ ) ( ) 1 (
3 , 2 , 1 ,
t x t x f t x t Z
r i r i m r i i
+ = +
2) DE/best/1
)] ( ) ( [ ) ( ) 1 (
3 , 2 , ,
t x t x f t x t Z
r i r i m best i i
+ = +


136

3) DE/rand-to-best/1
)] ( ) ( [ )] ( ) ( [ ) ( ) 1 (
2 1 ,
t r x t r x fm t x t x f t x t Z
i i i best i m i i
+ + = +
4) DE/best/2
)] ( ) ( [ )] ( ) ( [ ) ( ) 1 (
4 3 2 1 ,
t r x t r x fm t r x t x f t x t Z
i i i r i m
best
i i
+ + = +
5) DE/rand/2
)] ( ) ( [ )] ( ) ( [ ) ( ) 1 (
4 3 2 1 ,
5
t r x t r x fm t r x t x f t x t Z
i i i r i m
r
i i
+ + = + (12)
where i = 1, 2 , R is the individuals index of population and j = 1, 2,,N is the position in n-dimensional individual; t is the
time (generation); r
1
, r
2
, r
3
, r
4
and r
5
are mutually different integers and also different from the running index, i, randomly selected
with uniform distribution from the population set and f
m
> 0 is a real parameter called mutation factor, which controls the
amplification of the difference between two individuals so as to avoid search stagnation and is usually taken from the range [0,
2].Following the mutation operation, recombination is applied to the population. Recombination is employed to generate a trial
vector by replacing certain parameters of the target vector with the corresponding parameters of the randomly generated donor
vector.
>
= +
= +
)) int( ( ) ) ( ( . ),........ (
)) int( ( ) ) ( ( ),..... 1 (
) 1 (
i rand j or CR j rand if t x
i rand j or CR j rand if t Z
t U
ij
ij
ij
(13)
In the above rand(j) is the j
th
evaluation of a uniform random number generation within range [0, 1], and CR is a crossover or
recombination rate in the range [0, 1]. The performance of a DE algorithm usually depends on three variables; the population size
N, the mutation factor f
m
and the recombination rate CR. Selection is the procedure of producing better offspring. To decide
whether or not the vector should be a member of the population comprising the next generation, it is compared with the
corresponding vector. Thus, it denotes the objective function under minimization, and
< + +
= +
otherwise t x
t x f t u if t U
t x
i
i i
i
),.... (
)) ( ( ) 1 ( ( )...... 1 (
) 1 (
(14)
In this case, the cost of each trial vector U
i
(t+1) is compared with that of its parent target vector x
i
(t). Here, a detailed study of
all the basic DE variants is carried out to find the best strategy for a given MAED problem with multiple fuel options. The
performance is then compared with classical PSO and PSO_TVAC.

4. Implementation of MAED Problem with Multiple Fuel Option

The paper presents a solution of the MAED problem with multiple fuel options and valve point loading employing PSO and
DE strategies and critically compares their features for practical power system operation.

Step 1) Parameter setup
The PSO and DE parameters such as population size, the boundary constraints of optimization variables, cognitive and social
acceleration coefficients, the mutation factor (f
m
) ,the crossover rate (CR), and the stopping criterion of maximum number of
iterations (G
max
), are selected.

Step 2) Initialization of an individual population
For a population size R, the particles are randomly generated and normalized between the maximum and the minimum operating
limits of the generators. If there are N units, the i
th
particle is represented as
( )
M M M M
n
iN
n
i
n
i
n
i i
T T T T T T T P P P P P
) 1 ( 2 ,... 24 , 23 1 ........ ,......... 13 , 12 3 2 1
)...... ......... ( ), ( ........ , ,

= (15)
The j
th
Dimension of The i
th
Particle is normalized as given below to satisfy the generation limit constraint given by (5). Here, r
[0,1].
) (
min max min ij ij ij
n
ij
P P r P P + = (16)

Step 3) Evaluation of the individual population
The strength of each individual particle in the swarm is evaluated to judge its merit using a fitness function called evaluation
function. The evaluation function should be such that cost is minimized while constraints are satisfied. One of the methods for this
is the popular penalty function method. In this method, the penalty functions composed of squared or absolute violations are
incorporated in the fitness function, and are set to reduce the fitness of the particle according to the magnitude of the violation. The
penalty parameters are chosen such that an infeasible solution is awarded lesser fitness than the weakest feasible particle string.
Since two infeasible particles are not treated equally, the string further away from the feasibility boundary is more heavily
penalized. The penalty function approach, thus, converts a constrained optimization problem into an unconstrained optimization
problem. The fitness function values need to be calculated for each particle in order to find its merit. The evaluation function used
here is given by

137

( )

= =

+ +
N
i
N
i
M j
M
j
Dm im
M
m
i i
T P P P F
1
2
1
) 1 (
) 1 (
1
min
(17)
Here, is the penalty parameter. The second term imposes a penalty on the particle in terms of increased cost, if power balance
constraints of all the areas are not satisfied. The first term is calculated using Fi(Pi) from eq. (1) for solution considering the valve
point effects and eq. (2) for units with multiple fuel options. Transmission losses are neglected here for the sake of simplicity.

Step 4) Iterative Parameter Updation
In each iteration the parameters are updated to improve the fitness. In PSO the parameters are updated using eq. (7)-Eq. (11) while
in DE mutation adds a vector differential to a population vector of individuals; the donor or mutant vector is generated by using eq.
(12) corresponding to the chosen DE variant.

Step 5) Recombination operation
Recombination is applied in DE using eq. (13) to generate a trial vector by replacing certain parameters of the target vector with
the corresponding parameters of the randomly generated donor in step 4.

Step 6) Selection operation
Finally the selection operation produces better offspring. The values of the evaluation function are calculated for the updated
positions of the particles. In PSO if the new value is better than the previous pbest, the new value is set to pbest. Similarly, value
of gbest is also updated as the best of pbest. In DE the trial vector U
i
(t+1) replaces its parent target vector x
i(
t) if its cost is found
to be better otherwise the target vector is allowed to advance to the next generation.

Step 7) Stopping criterion: A stochastic optimization algorithm is stopped either based on the tolerance limit or maximum number
of iterations. For comparison with other strategies, the number of iterations is adopted as the stopping criterion in this paper.


The additional tie-line constraints and area power balance constraints make the MAED problem much more complex and
difficult to solve as compared to the classical ED problem. The PSO and DE based evolutionary strategies are tested for the
proposed practical MAED problem on three test systems having different sizes and nonlinearities. The performance was compared
with previously published results (Yalcinoz et al., 1998; Chen et al., 2001 and Manoharan et al., 2009) and was found to be better.

5.1 Description of the test systems

i) The first test system consists of a two-area system with four generating units (Yalcinoz et al., 1998; Chen et al., 2001) as shown
in Fig. 1. This system is considered here for the purpose of comparison with previous results. The percentage of the total load
demand in area 1 is 70% and 30% in area 2. The cost coefficients and limits are taken from (Chen et al., 2001). The load demand
(PD) and tie-line flow limit are set at 1120 MW and 200MW respectively. The global best for this system has been reported at
$10,605 (Yalcinoz et al.1998,Chen et al.2001) . Reference (Manoharan et al., 2009) has reported $10,574 but the reported results
are infeasible because though their solutions satisfy the power balance and generating limit constraints, it does not satisfy the area
power balance constraints.

ii) The second system (Manoharan et al., 2009, Lin et al., 1984) comprises of three areas, 10 generating units, with three fuel
options. The total system demand is 2700 MW. The 10 generating units are divided into three areas, as shown in Fig. 2. Area 1
comprises the first four units (P1, P2, P3, P4); area 2 includes three units (P5, P6, P7); and area 3 has the remaining three units (P8,
P9, P10). Each area has both generation and load and each area is represented as having tie-line connections to each of the other
areas. The load demand in area 1 is assumed as 50% of the total demand. The load demand in area 2 is assumed as 25% and in area
3 the load is 25% of the total demand. The tie line flow limit is set at 100 MW for each tie line. The global best for this system
has not yet been reported. The minimum cost for this system using covariance matrix adapted evolutionary strategy (CMAES)
technique (Manoharan et al., 2009 ) is $686.9850 which is infeasible as area power balance constraints are not satisfied in this case
too. DE and PSO_TVAC algorithms used in this paper have achieved lower and feasible results for this system.

iii) Test system three was selected for testing the performance of the evolutionary methods for MAED problem of a large system.
This system has 40-generating units with valve point loading effects taken from (Sinha et al., 2003), randomly distributed into two
areas such that both areas include half the units. The total system load is 10,500MW (Sinha et al., 2003). The classical PSO is not
able to find the best solution for this system but DE and PSO_TVAC converge to global best solutions.

138

For the purpose of comparison the cost of tie-line power flow is neglected in all three test cases. In classical PSO both the
acceleration coefficients are taken equal to 1.5 for all systems. Simulations were carried out using MATLAB 7.0.1 on a Pentium
IV processor, 2.8 GHz. with 512 MB RAM.

Figure 1.Two-area, four unit system.

Figure 2. Three area ten generating unit system.

5.2 Testing Strategies

The MAED problem was solved using the classical PSO, PSO_TVAC and DE and their performance was compared with some
already reported results (Yalcinoz et al., 1998; Chen et al., 2001; Manoharan et al., 2009). During recent years DE and its hybrid
variants have been proposed for many power system applications but a detailed comparative study of all its basic variants for
MAED problem with complex constraints has not been reported yet. The paper i) Compares different DE variants for the MAED
problem ii) Compares the best DE variant with its close competitor PSO and its effective variant PSO_TVAC iii) Investigates the
influence of tie-line power limits on the total fuel cost. Through out this paper DE/rand to best/1 variant of DE is applied for
comparison with PSO and PSO_TVAC as it was found to be the best among various DE variants defined by eq. (12).

5.3 Effect of tuning parameters

An attempt has been made here to compare the PSO and DE strategies extensively, to find similarities and differences; merits
and shortcomings; solution quality and consistency and dependence on tuning parameters. A deeper understanding of these
strategies can save a lot of time and effort which is otherwise wasted in tuning the different parameters of these algorithms. The
mean and standard deviation (S.D.) out of 50 trials for DE algorithm are tabulated in Table 1 for 4-unit system and in Table 2 for
10-unit system. From both the tables it is evident that better results are available for fm and CR pairs lying in a diagonal. Higher
fm supports higher CR and best results are obtained for f
m
=CR=0.9 (and/or 0.8) for both the systems. This behavior was found to
be consistent for other mathematical benchmark problems too. It can be seen from the two tables that for some combinations of
fm and CR, the DE algorithm did not converge.
Similar to DE and other evolutionary techniques, the performance of PSO algorithm is quite sensitive to the various parameter
settings. The initial and final values of the acceleration coefficients have a significant effect on the solution. Based on empirical
studies on a number of mathematical benchmarks, reference (Ratnaweera et al., 2004) has reported the best range of variation as
2.5-0.5 for c
1
and 0.5-2.5 for c
2
. In reference (Chaturvedi et al., 2008). It is observed that initial value of the cognitive coefficient

139

c1i and final value of social coefficient c
2f
control the range of the search space. Therefore in PSO_TVAC the values of c
1i
and c
2f

are varied between 2.5-1.8 keeping c
1f
and c
2i
fixed at 0.2 (Chaturvedi et al., 2008). It is seen that a high initial value of the
cognitive coefficient c
1
makes use of full range of the search space while a low social coefficient c
2
helps in avoiding premature
convergence. As the search progresses c
1
is reduced to reduce search space and c
2
is increased to accelerate the solution towards
global convergence. The best parameter combination is found to depend on the nature of the objective function. However all the
tested functions were found to achieve optimal results in the above variation range of c
1
and c
2
.
Table 3 and Table 4 present the effect of acceleration coefficients on the MAED solution of 4-unit and 10-unit system by
PSO_TVAC strategy. It can be seen that for all combinations of c
1i
and c
2f
around 2.0, the PSO_TVAC strategy converges to near
global results indicated by a small value of S.D. Best results in this case are found for c1i = c2f =1.8. On comparison of DE with
PSO_TVAC it can be seen that DE converges to the global solution for f
m
=CR=0.9 for both the test systems with zero S.D. but for
other values of fm and CR the S.D. is quite high, sometimes indicating non convergence. To summarize, it can be said that DE is
capable of producing guaranteed global best results accurately for fm =CR=0.9 while PSO_TVAC converges to near global values
for all combinations of tuning parameters; the global best is also achieved in PSO_TVAC for some of the runs but the optimal
parameter combination for achieving global best is problem specific, based on trial and error. The S.D. of PSO_TVAC is also
higher than that of DE.

Table 1. Effect of mutation factor and cross over rate on mean and S.D.in DE (4-unit system; 50 Trials)

*the bracketed value indicates the standard deviation

Table 2. Effect of mutation factor and cross over rate on mean and S.D.in DE (10-unit system; 50 Trials)

*the bracketed value indicates the standard deviation

140

Table 3. Effect of acceleration coefficients on performance of PSO_TVAC (4 unit system ; 50 trials)

Table 4. Effect of acceleration coefficients on performance of PSO_TVAC (10 unit system ; 50 trials)


141

5.4 Effect of population size

In addition to tuning parameters, the population size is another important issue in stochastic search based optimization
methods. Studies have been carried out on the role of population size and it has been reported that increasing population improved
the performance of PSO algorithm (Chaturvedi et al., 2008, Alrashidi et al., 2007). The optimal population size depends on the
problem dimension. Reference (Storn and Price et al., 1997) suggested a population size of 5-10 times that of the problem
dimension. It was shown that for various ED problems, larger the dimension, larger is the population size required to achieve good
results, (Noman et al., 2008).
In this paper, tests were carried out for different population sizes and results are tabulated in Table 5 and Table 6 for 4-unit and
10-unit systems for the DE algorithm. In Table 7 and Table 8 the results of the two test systems for PSO_TVAC algorithm are
presented. The DE algorithm converges to the global best solution for the 4-unit system for a population size of 20 with a S.D. of
7.31. However when the population is increased to 25, the S.D. becomes zero. Similar results are found for the 3-area, 10-unit, test
system with multiple fuels. The system was found to converge with zero S.D. for a population size of 120. For the 40-unit system a
population size of 1000 was found to be optimum for PSO_TVAC. The DE algorithm converged for a population of 200.
However, the S.D. was found to increase for DE as well as PSO_TVAC for the larger system.

Table 5. Effect of population size on performance of DE (4 unit system; 50 trials) (F=CR=0.9)

Table 6. Effect of population size on performance of DE (10- unit system ; 50 trials) (F=CR=0.9)

Table 7. Effect of population size on performance of PSO_TVAC (4 unit system ; 50 trials)

Table 8. Effect of population size on performance of PSO_TVAC (10 unit system; 50 trials)


142

5.5 Convergence characteristics

To test the convergence behavior of the evolutionary strategies under study, i.e. DE, PSO and PSO_TVAC the convergence
test was carried out employing the same evaluation function and same initial population for the same number of iterations. The
results for all three strategies for one trial of 100 iterations are shown in Fig. 3 and Fig. 4 for the 4-unit and 10-unit systems
respectively. It can be seen that the TVAC strategy provides the PSO algorithm with optimal search capability due to the proper
tuning of social and cognitive coefficients during the search. When search advances and reaches a certain iteration count, the
classical PSO characteristic saturates but the PSO_TVAC still continues to improve and thus shows the best convergence
characteristics. It is seen that DE takes longer to converge than PSO_TVAC.

Figure 3. Convergence characteristics of the three evolutionary strategies. (4-unit system)

Figure 4. Convergence characteristics of the three evolutionary strategies. (10-unit system)


143

5.6 Robustness

The performance of heuristic search based optimization algorithms is not judged by the results of a single trial due to the
randomness involved in its functioning. Many trials with different initial populations were carried out to test the
robustness/consistency of the different evolutionary algorithms. The lowest cost for each of the 50 different trials has been plotted
in Fig. 5 and Fig. 6 from which it can be seen that DE produces lowest cost with zero standard deviation indicating its highest
consistency. The PSO_TVAC performs much better than classical PSO as it achieved near best results in many trials.

Figure 5. Best results of DE, PSO and PSO_TVAC variants (4-unit system); 50 trials


144

5.7 Effect of variation of tie-line power flow

By increasing the power transfer limit between the three areas of the 10-unit system, the operating cost for the same load can
be reduced as cheaper generators generate more and transfer the surplus power to the deficient area. Table 9 shows using DE that
the operating cost can be reduced from $637.4111/hr to $630.9390/hr if the inter area tie line limit is increased uniformly from
100 MW to 150 MW. The cost drops further to $623.6298 if the limit is raised to 200 MW.

Table 9. Effect of variation of Tie line limits

Table 10. Comparison of Classic DE variants with PSO and PSO_TVAC (4-unit system;50 trials)

Table 11. Comparison of Classic DE variants with PSO and PSO_TVAC (10-unit system;50 trials)


145

Table 12. Time (cpu seconds) comparison of DE, PSO and PSO_TVAC

Table 13. Comparison of best results of evolutionary strategies for 4-unit system

* V1 and V2 are area power balance violations for the two areas
Table 14. Comparison of best results of evolutionary strategies for 10-unit system with multiple fuel options

5.8 Comparative analysis

A detailed study of classic DE variants is carried out to judge their merit for the complex MAED problems. Table 10 and table
11 show the best results out of 50 trials for these strategies. Observing the minimum cost and standard deviation it can be
concluded that out of the DE variants, the DE/rand to best /1 strategy performs the best for all the systems followed by
DE/rand/1 and DE/rand/2 strategies. The DE/best/1 and DE/best/2 strategies do not converge in most trials. PSO_TVAC strategy
gives better performance than the classical PSO and unlike DE convergence is achieved for all values of acceleration coefficients.

146

Another significant difference is that DE requires less population but takes larger number of iterations to converge while PSO
works with larger populations and lesser iterations as reported in Table 12. Table 13 and Table 14 give the comparison of reported
results with already published results. The superiority of the reported results is evident from their ability to satisfy all constraints
and produce feasible results. The solution reported in Manoharan et al. (2009) by the CMAES method satisfies the power balance
constraint but the individual area balance constraints are violated.

5.9 Performance on a large system

The performance of DE and PSO strategies is also tested on a large system for ED with tie-line constraints and area power
balance constraints. Fig. 7 shows the cost/MW of generators in the two areas. The overall cost will be heavily influenced by tie-
line limits, area loads and cost curves. The problem is complex as the unit valve point effects are also considered. Cost curves for
some generators are plotted in Fig. 8 to show the effect of valve point loading effects. The DE and PSO_TVAC strategies could
handle this complex problem effectively, with full constraint satisfaction, but the S.D. increases for both. The optimal generation
schedules, the tie line power flow and operating cost for this system are tabulated in Table 15 for different loading conditions. Fig.
9 shows the best results of 50 different trials for DE, PSO and PSO_TVAC. The DE strategy is most consistent followed by
PSO_TVAC.

Figure 7. Per MW generation cost of 2-area, 40 unit system

Figure 8. Cost characteristics of some generators of the 2-area 40-unit system with valve point loading

147


Table 15. DE Results for 2-area, 40-unit system with valve point loading effects


148

6. Conclusions

Evolutionary strategies are increasingly being proposed for solving non-convex, discontinuous, multi-modal ED problems
with complex constraints. In multi-area ED problem additional tie-line and area balance constraints are introduced which present
difficulty in obtaining feasible solutions. The paper explores the ability of DE and PSO strategies to produce global best solutions
for the complex MAED problems. A comparative study of classic DE variants with PSO and its variants is made to find their
abilities and limitations. Simulations on three systems of different sizes and areas, having different complexity levels clearly reveal
that
The DE/rand to best/1 strategy performs best for all tested systems closely followed by DE/rand/1 and DE/rand/2
strategies. The other DE variants do not converge for MAED problems.
PSO_TVAC and classic PSO converge to near global solutions in all trial runs, for all tested values of tuning parameters
though global best performance is not guaranteed; but the S.D. is higher than DE.
DE converges to the global best solution with zero S.D. in a very narrow range of tuning parameters. For other
combinations it either diverges or produces low quality solutions; hence very little parameter tuning is required in DE as
compared to PSO and its variants which work for a very broad range of acceleration coefficients.
DE works with lesser population size but require more number of iterations to converge while the PSO and PSO_TVAC
require larger populations but converge earlier.
Due to their speed and efficiency, the evolutionary optimization algorithms prove to be very effective in analyzing the
effect of tie-line power limits on the cost of operation.
It is shown through different trials that the DE outperforms other methods, particularly for MAED problems, in terms of
solution quality, computational efficiency, dynamic convergence, robustness and stability.

Acknowledgement

The authors sincerely acknowledge the financial support provided by University Grants Commission (UGC), New Delhi, India
under major research project entitled Power System Optimization and Security Assessment Using Soft Computing Techniques,
vide F No.34-399/2008 (SR) dated, 24th December 2008 and AICTE New Delhi, India for financial assistance under RPS project
F No 8023/RID/BOR/RPS-45/2005-06 dated 10/03/2006. The authors also thank the Director, M.I.T.S. Gwalior for providing
facilities for carrying out this work.

References

Alrashidi M.R. and El-Hawary M.E.2007. Hybrid particle swarm optimization approach for solving the discrete OPF problem
considering the valve loading effects, IEEE Trans. on Power Systems, Vol. 22, No. 4, pp. 2030-2038.
Chaturvedi Krishna Teerth, Pandit Manjaree and Srivastava Laxmi. 2009. Particle swarm optimization with time varying
acceleration coefficients for nonconvex economic power dispatch, Electrical Power and Energy Systems, Vol. 31, No.6, pp.
249-257.
Chaturvedi Krishna Teerth, Pandit Manjaree and Srivastava Laxmi 2008. Self-organizing hierarchical particle swarm optimization
for nonconvex economic dispatch, IEEE Transactions on Power System, Vol. 23, No. 3, pp. 1079-1087.
Chen C.L. and Chen N. 2001.Direct search method for solving economic dispatch problem considering transmission capacity
constraints, IEEE Transactions on Power System, Vol. 16, No. 4, pp. 764-769.
Cheng S.-L. and Hwang C.2001. Optimal approximation of linear systems by a differential evolution algorithm, IEEE Trans.
Syst., Man, Cybern.A, Syst., Humans, Vol. 31, No. 6, pp. 698707.
Coelho Leandro dos Santos and Mariani Viviana Cocco 2006. Combining of chaotic differential evolution and quadratic
programming for economic dispatch optimization with valve-point effect, IEEE Trans. Power System, Vol. 21, No. 2, pp. 989-
996.
Desell A.L., Tammar K., McClelland E.C. and Van Home P.R. Transmission constrained production cost analysis in power
system planning, 1984. IEEE Trans. on Power Apparatus and Systems, Vol. PAS-103, No. 8, pp. 2192-2198.
Doty K.W. and McEntire P.L. 1982 .An analysis of electrical power brokerage systems, IEEE Trans Power Apparatus System,
Vol. 101, No.2 , pp. 389-396.
Farmer E.D., Grubb M.J. and Vlahos K. 1990. Probabilistic production costing of transmission-constrained power systems, 10th
PSCC Power System Computation Conference, pp. 663-669.
Hemick SD,Shoults RR. 1985. A practical approach to an interim multi-area economic dispatch using limited computer resources.
IEEE Trans Power Apparatus System, Vol. 104, No. 6, pp. 1400-1404
Jayabarathi T., Sadasivam G. and Ramachandran V. 2000. Evolutionary programming based multi-area economic dispatch with
tie-line constraints, Electrical Machines and Power Systems, Vol. 28, pp.1165-1176.
Kennedy J and Eberhart R.1995. Particle swarm optimization. Proc. IEEE Conf. on Neural Networks (ICNN95), Vol. IV, Perth,
Australia, pp. 1942-48.

149

Lin C.E.and Viviani G.L.1984. Hierarchical economic load dispatch for piecewise quadratic cost functions, IEEE Transactions
Power Apparatus and systems, Vol. PAS-103, pp.1170-1175.
Manoharan P.S., Kannan P.S., Baskar S., Iruthayarajan M. Willjuice. 2009. Evolutionary algorithm solution and KKT based
optimality verification to multi-area economic dispatch, Electrical Power and Energy Systems, Vol. 31, pp. 365373.
Noman Nasimul and Iba Hitoshi.2008. Differential evolution for economic load dispatch problems, Electric Power Systems
Research, Vol. 78, pp. 1322-1331.
Orero SO and Irving MR. 1996. Economic dispatch of generators with prohibited operating zones: a genetic algorithm approach.
IEE proceedings Generation, Transmission and Distribution, Vol. 143, No.6, pp. 529-33.
Panigrahi B.K. and Pandi V. Ravikumar 2009. Bacterial foraging optimisation: NelderMead hybrid algorithm for economic load
dispatch, IET Generation, Transmission and Distribution, Vol. 2, No. 4, pp. 556565.
Park J.H., Kim Y.S., Eom I.K., Lee K.Y. 1993.Economic load dispatch for piecewise quadratic cost function using Hopfield neural
network, IEEE Transactions on Power Systems, Vol. 8, pp. 1030-1038.
Ratnaweera A., Halgamuge S.K. and Watson H.C.2004.Self-organizing hierarchical Particle swarm optimizer with time varying
acceleration coefficients, IEEE Trans Evol Computation, Vol. 8, No. 3, pp. 240-255.
Shoults R.R., Venkatesh S.V., Helmick S.D., Ward G.L., Lollar M.J. 1986. A dynamic programming based method for developing
dispatch curves when incremental heat rate curves are non-monotonically increasing. IEEE Trans Power Syst., Vol. 1, No. 1,
pp. 10-16.
Shoults R.R., Chang S.K., helmick S. and Grady W.M. 1980.A practical approach to unit commitment, economic dispatch and
savings allocation for multiple area pool operation with import/export constraint, IEEE Transactions on Power Apparatus and
Systems, Vol. 99, No.2, pp. 625-35.
Sinha N., Chakraborty R. and Chattopadhyay P.K.2003.Evolutionary programming techniques for economic load dispatch, IEEE
Transactions on. Evolutionary Computation, Vol. 7, No.1, pp. 83-93.
Song YS, Chou CS and Stoham TJ. 1999.Combined heat and power economic dispatch by improved ant colony search algorithm.
Electric Power Systems Research, Vol. 52, No.2, pp. 115-121
Storn R. and Price K.1995.Differential evolution, A simple and efficient adaptive scheme for global optimization over continuous
spaces, International Computer Science Institute, Berkeley, CA, Tech. Rep. TR-95-012.
Storn R.1997.Differential evolutiona simple and efficient heuristic for global optimization over continuous spaces, J. Global
Optimization., Vol. 11, No. 4, pp. 341359.
Streifferet Dan. 1995. Multi-Area economic dispatch with tie-line constraints. IEEE Transactions on Power Systems, Vol. 10, No.
4, pp. 1946-1951.
Vanaja B., Hemamalini S. and Simon S.P. 2008. Artificial immune based economic load dispatch with valve-point effect, IEEE
Region 10 Conference, TENCON 2008, pp. 1-5.
Vasebi A., Fesanghary M., Bathaee S.M.T. 2007.Combined heat and power economic dispatch by harmony search algorithm,
Electrical Power and Energy systems, Vol. 29, pp. 713-719.
Walter D.C. and Sheble G.B.1993 Genetic algorithm solution of economic load dispatch with valve point loading, IEEE
Wang C, Shahidehpour SM. 1992. A decomposition approach to non -linear multi-area generation scheduling with tie-line
constraints using expert system. IEEE Trans Power Apparatus System, Vol. 7, No. 4, pp. 1409-1418
Wernerus J. and Soder L. 1995.Area price based multi-area economic dispatch with tie-line losses and constraints, IEEE/KTH
Stockholm Power Tech Conference. Sweden , pp.710-15.
Wood A. J. and Wollenberg B.F. 1984. Power generation, operation and control, John Wiley & Sons.
Yalcinoz T. and Short M. J. 1998. Neural networks approach for solving economic dispatch problem with transmission capacity
constraints, IEEE Transactions on Power Systems, Vol. 13, No. 2, pp. 307-313.

Biographical notes:

Manisha Sharma obtained her M.E. degree in Electrical Engineering from Madhav Institute of Technology & Science Gwalior (India) in 2007. She is currently
working as Project Fellow in the Department of Electrical Engineering, M.I.T.S., Gwalior (India) in the area of Evolutionary Computation applications.

M. Pandit obtained her M.Tech degree in Electrical Engineering from Maulana Azad College of Technology, Bhopal, (India) in 1989 and Ph.D. degree from Jiwaji
University Gwalior (India) in 2001. She is currently working as Professor in Department of Electrical Engineering, M.I.T.S., Gwalior, (India). Her areas of interest
are Power System Security Analysis, Optimization and soft computing/ evolutionary methods, ANN and Fuzzy neural applications to Power System.

L. Srivastava obtained her M. Tech. degree in Electrical Engineering from the Indian Institute of Technology, Kanpur, India in 1990 and her Ph.D. degree from
University of Roorkee (Presently IIT Roorkee), Roorkee, India in 1998. She is working as a professor in the Department of Electrical Engineering, M.I.T.S.
Gwalior, India. She is currently involved in research in power system optimization and control, security analysis, operation and control of deregulated power
systems, and ANN and fuzzy logic applications to power system.

Accepted March 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 150-166

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Overall performance evaluation of evolutionary designed conventional AGC
controllers for interconnected electric power system studies
in a deregulated market environment

Y.L. Karnavas*, K.S. Dedousis

Lab. of Electric Machines & Installations, Dept. of Electrical Engineering
School of Applied Technology, Technological Educational Institution of Crete (TEIC),
PO Box 1939, Zip 710 04, Heraklion, Crete, Hellas, GREECE

*
Corresponding Author: e-mail: karnab@staff.teicrete.gr

Abstract

Electric power industry is currently in transition from vertically integrated utilities to an industry that will incorporate
competitive companies. This increases the complexity of the load frequency issue and calls for more insight and research. In this
context, the tuning of a multi-area automatic generation control (AGC) system after deregulation and furthermore, the effect of
reheat turbines dynamics in the power system performance, are not yet discussed in depth and are studied in this work. The
effect of bilateral contracts on the dynamics of the system is taken into account and the concept of DISCO participation matrix
for these bilateral contracts is simulated. Genetic algorithms are adopted in order to obtain the optimal parameters of the load-
frequency controllers as well as of the frequency biases of thermal systems with reheat turbines. Also, since the optimum
parameter values of the classical AGC have been obtained in the literature by minimizing the popular integral of the squared
errors criterion (ISE) only, an effort is made in this study to show that this criterion does not give always the best system
performance especially in a deregulated environment. In this work, we investigate the optimum adjustment of the load frequency
controllers using a set of performance indices which are various functions of error and time. In this way, someone can observe
the various performances that such a kind of power system might have when a different performance index is used. It should be
noted that to the extent of the authors' knowledge, this kind of optimization has not been done yet in the literature. The
performances of the tuned twoarea AGC system are obtained using appropriate Matlab/Simulink models. Finally, it is
envisaged that the synthesis procedure highlighted in this paper could be of practical significance for tuning conventional AGC
controllers for an interconnected thermal-electric power system in a deregulated environment.

Keywords: AGC, load frequency control, power systems, deregulation, bilateral contracts, genetic algorithms

1. Introduction

Many investigations have been reported in the past pertaining to AGC of a large interconnected power system (i.e Abdel-Magid
and Dawoud, 1997, Aditya and Das, 2003, Cohn, 1986). A net interchange tie-line bias control strategy has also been widely
accepted by utilities. The frequency and the interchanged power are kept at their desired values by means of feedback of the area
control error (ACE) integral, containing the frequency deviation and the error of the tie-line power, and controlling the prime
movers of the generators. The controllers so designed regulate the ACE to zero. For each area, a bias constant determines the
relative importance attached to the frequency error feedback with respect to the tie-line power error feedback; the bias is very often
equal to the natural area frequency response characteristic. Classical AGC corresponds basically to industry practice for the past
years or so. The key assumptions are: (a) the steady-state frequency error following a step-load change should vanish and also the
transient frequency and time errors should be small, (b) the static change in the tie power following a step-load in any area should
be zero, provided each area can accommodate its own load change and (c) any area in need of power during emergency should be
assisted from other areas. The key advantage of the classical AGC is that the control strategy is a totally decentralized one, in the
sense that each control area carries out its own frequency and power regulation using locally gathered real-time information.
Karnavas and Dedousis / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 150-166

151

On the other hand, the traditional power system industry has a vertically integrated utility structure. In the restructured or
deregulated environment, vertically integrated utilities no longer exist. The utilities no longer own generation, transmission, and
distribution; instead, there are three different entities, viz., GenCos (generation companies), TransCos (transmission companies)
and DisCos (distribution companies). As there are several GenCos and DisCos in the deregulated structure, a DisCo has the
freedom to have a contract with any GenCo for transaction of power. A DisCo may have a contract with a GenCo in another
control area. Such transactions are called bilateral transactions. All the transactions have to be cleared through an impartial entity
called an independent system operator (ISO). The ISO has to control a number of so-called ancillary services, one of which is
AGC. For an in-depth discussion of implications of restructuring the power industry, the reader can refer to (Ilic et al, 1998,
Sheble, 1999, Christie et al, 2000). So, it is obvious nowadays that in a restructured electric power system environment, the
engineering aspects of planning and operation have to be reformulated. In most of the recent reported strategies, attempts have
been made to adapt well-tested classical AGC schemes to the changing environment of power system operation under deregulation
(Kumar et al, 1997, Delfino et al, 2002, Ibraheem et al, 2009, Sinha et al, 2010). Comprehensive studies on simulation and
optimization in an AGC system after deregulation have been carried out by Donde et al as well as one of the authors (Donde,
2001, Karnavas, 2005). In Dondes work the concept of DisCo participation matrix (DPM) is proposed that helps the visualization
and implementation of the contracts. The critical parameters in order to tune such a system are found to be the feedback integral
gains of the conventional integral controllers as well as the frequency biases of the areas. In this work, all the three components of
proportional plus integral plus derivative (PID) load frequency controllers are being examined, and a two-area AGC system block
diagram in restructured environment is used to demonstrate the tuning procedure. Modifications are made in the area controllers
which are extended to be I, PI and PID ones as well. A set of different performance indices which are various functions of error
and time is also used here. In this way, someone can observe the various performances that such a kind of power system might
have when a different performance index is used. It should be noted that to the extent of the authors' knowledge, this kind of
optimization has not been done yet in the literature. The paper is organized as follows. In the next section the block diagram of the
two-area four-GenCos AGC system is modeled, described and its aspects are discussed. The reheat type steam turbine model is
also implemented there. GAs details and the tuning procedure are given next. The GA tuning is applied to three case studies
described in the fourth Section concerning normal and contract violation situations, and the simulation results which are
summarized in fifth section show the effectiveness of the proposed tuning procedure. Finally, in the conclusion Section, some
comments are given as well as some thoughts for future work.

2. The Deregulated Market Environment for AGC

The traditional AGC is well discussed in numerous works (i.e. Elgerd, 1970, Karnavas and Papadopoulos, 2002, Pan and Lian,
2005) while research work in deregulated AGC can be found in (Kumar et al, 1997, Delfino et al, 2002, Donde et al, 2001,
Christie and Bose, 1996, Nobile et al, 2000, Bevrani et al, 2004, Karnavas, 2005, Shayeghi et al, 2006, Demiroren and Zeynelgil,
2007, Srinivasa et al, 2008). In the new restructured environment (see Figure 1), GenCos sell power to various DisCos at
competitive prices. DisCos have the liberty to choose the GenCos for their contracts. They may or may not have contracts with the
GenCos in their own area. This makes various combinations of GenCo-DisCo contracts possible in practice (Figure 2). The block
diagram in Figure 3 shows how the bilateral contracts are incorporated in the traditional AGC system. The system is modeled in
Matlab/Simulink and the area controllers are replaced with I or PI or PID ones. Each area includes two identical GenCos and two
DisCos. With respect to Figure 3, the four loads of the DisCos are stored in the DISCO block. It should be noted also that the two
power systems have equal nominal installed capacities. Furthermore, the developed model for a GenCo is shown in Figure 4, and
the reheater transfer function block is taken from Kumar et al (1985). Finally, the used model for the conventional controllers is
shown in Figure 5. The controller's demand signal is distributed according to the area participation factors (apf) block. Each
GenCo is represented by a governor, a reheater and a turbine. All the system data are given for clarity in the Appendix A.

I
S
O
A
r
e
a
I
C #1 C #2 C #
DC #1 GC #2 GC #
GC #1
A
r
e
a
II
GC #2 GC #
C #1 C #2 C #
DC #1 DC #2 DC #
GC #
A
r
e
a
X
GC #2 GC #1
C # C #2 C #1
DC # DC #2 DC #1
GC #1 GC #2 GC #

Figure 1. Interconnection between generation (GC), transmission (TC) and distribution (DC) companies in a deregulated
environment operated by independent system operator (ISO).

152

Power
System
A
r
e
a
I
A
r
e
a
II
GenCo
Contracts
DisCo
Contracts
Area
Controller
GenCo
Contracts
DisCo
Contracts
Area
Controller

Figure 2. Schematic of a two-area AGC model in a restructured environment.

f1
f2
PL2
PL1
Ptie
f1
f2
1/R4
1/R3
1/R2
1/R1
apf1
apf2 Controller 2
Controller 1
I, PI, PID
I, PI, PID
a121
a122

Figure 3. Developed model of a two-area double-GenCo AGC system with controllers.

Figure 4. Developed model of a GenCo with a reheat turbine.

Figure 5. Discrete PID-type controller used in this study.


153

Depending on the contracts made between GenCos and DisCos, the DPM is set. DPM is a matrix with the number of rows equal
to the number of GenCos and the number of columns equal to the number of DisCos in the system (see Eq. 1). Each entry in this
matrix can be thought of as a fraction of a total load contracted by j
th
DisCo towards the i
th
GenCo and is called contract
participation factor (cpf) as seen in Eq. 2.

11 12 13 14
21 22 23 24
31 32 33 34
41 42 43 44
1 2 3 4
1
2
3
4
+ + + +
(
=
(

(
(
(

DC DC DC DC
cpf cpf cpf cpf GC
cpf cpf cpf cpf GC
cpf cpf cpf cpf GC
cpf cpf cpf cpf GC
DPM (1)
where

DisCo's power demand out of GenCo [pu.MW]
DisCo's total power demand [pu.MW]

=
th th
ij th
j i
cpf
j
(2)

Whenever a load demanded by a DisCo changes, it is reflected as a local load in the area to which this DisCo belongs. This
corresponds to the local loads P
L1
and P
L2
and should be reflected in the deregulated AGC system block diagram at the point of
input to the power system block. As there are many GenCos in each area, ACE signal has to be distributed among them in
proportion to their participation in the AGC. Coefficients that distribute ACE to several GenCos are termed (as described above)
ACE participation factors (apf). Note that
1
( ) 1
=
=
j
m
j
apf where m is the number of GenCos. Thus, as a particular set of GenCos
are supposed to follow the load demanded by a DisCo, information signals must flow from a DisCo to a particular GenCo
specifying corresponding demands. These signals (which were absent from the traditional AGC scenario) describing the partial
demands, are specified by the cpfs and the puMW load of a DisCo. These signals carry information as to which GenCo has to
follow a load demanded by which DisCo. In our case of two areas (I, II), the scheduled steady state power flow on the tie-line is
given as

( )
( )
Demand of DisCos in area II from GenCos in area I
Demand of DisCos in area I from GenCos in area II
A =
scheduled
tie I II
P
(3)

At any given time, the tie line power error is defined:

, ,
- A = A A
i actual i scheduled i
tie tie tie
P P P (4)

This error vanishes in the steady state as the actual tie line power flow reaches the scheduled power flow. This error signal is used
to generate the respective ACE signals as in the traditional scenario. Thus, the area control error (ACE) for the i
th
area at any time
instant t is defined as,
( ) ( ) ( ) ( ) = = A + A
i i i
tie i i
ACE t e t P t B f t (5)

where ( )
2 1 1 2 1 2
A = A
tie r r tie
P P P P and P
r1
, P
r2
are the rated installed capacities of areas I and II, respectively. Consequently,
1 1 1 12 1 2
= A + A
tie
ACE B f a P and
2 2 2 12 1 2
= A + A
tie
ACE B f a P where
12 r1 r 2
P P = a . Therefore, in this work the required GenCos
production is given by:

GENCO = DPMDISCO (6)

Finally, for this work, the controllers gains as well as the frequency biases are set to be equal for both areas.

2.1 Control Strategies under Study
The conventional automatic generation controller found mostly in literature has a linear integral only control strategy of the
following form and also sometimes researchers use continuous time models which are completely inappropriate, i.e.

( ) =
}
i i
i I
u K e t dt (7)

154

In this work, for achieving the basic objectives of load-frequency control problem, i.e., zero steady-state error in frequency and tie-
line power, and also for completeness purposes since our goal is the overall conventional controllers performance evaluation, the
discrete type of integral (I), proportional plus integral (PI) and proportional plus integral plus derivative (PID) controller is used
(taken from Simulink/Matlab library) as already shown in Figure. 5. Parameter tuning and in depth discussion of such type of
controllers can be found in Wu and Huang (1997). In our case, the control law for the i
th
area (i=1,2) is given by

( ) ( )
( )
| |
=
|
|
\ .
}
i
i i i i i
i P I D
de t
u K e t K e t dt K
dt
(8)
which when implemented in the Simulink/Matlab environment takes the form
( ) ( )
( )
0
A
= A
A
| |
|
\ .
i
Tsim
i i i i i
i P I D
e kTs
u K e kTs K e kTs t K
t
(9)

where K
I
i
, K
P
i
and K
D
i
are the gains of the PID controllers (K
P
i
, K
D
i
or K
D
i
are set to zero when a I or a PI is needed respectively),
T
s
is the sampling time, T
sim
is the total simulation time and k=1,2,..,T
sim
/T
s
. In this study, the optimum values of these gains along
with the bias factor B
i
which minimize a whole set of different performance indices are easily and accurately computed using a
genetic algorithm. In a typical run of the GA, an initial population is referred to as the 0
th
generation. Each individual in the initial
population has an associated performance index value. Using the performance index information, the GA then produces a new
population. The application of a GA involves repetitively performing two steps: (a) the calculation of the performance index for
each of the individuals in the current population. To do this, the system must be simulated to obtain the value of the performance
index; (b) the GA then produces the next generation of individuals using the reproduction, crossover and mutation operators. These
two steps are repeated from generation to generation until the population has converged, producing the optimum parameters. A
flowchart of the GA optimization procedure is given in Figure 6.

3. Genetic Algorithm Overview and Implementation Aspects

Genetic algorithms (GA), a way to search randomly for the best answers to tough problems were first introduced by Holland
(1975). Over the past years, it is becoming important to solve a wide range of search, optimization and machine learning problems.
A GA is an iterative procedure which maintains a constant size population of candidate solutions. The algorithm begins with a
randomly selected population of function inputs represented by string of bits. During each iteration step, called a generation, the
structures in the current population are evaluated and on the basis of this evaluation, a new population of candidate solution is
formed. That is GA uses the current population of string to create a new population such that the strings in the new population are
on average better than those in the current population. The idea is to use the best elements from the current population to help form
the new population. If this is done correctly, then the new population will on average be better than the old population. Three basic
processes, namely selection, mating (or crossover) and mutation are used to make the transition from one population generation to
the next. In addition, GA works with a coding of parameters rather than the parameters themselves, thereby freeing itself of the
limitations (e.g. continuity and derivative existence) of conventional techniques such as gradient methods (Wu and Huang, 1997).
The simplified genetic algorithm cycle based on the above is shown in Figure 6, while the main pseudo-code implemented in
MATLAB environment for this work is depicted in Figure 7.

Initial
Population
Evaluation
Selection
Crossover
Mutation
Convergence?
Start
End

Figure 6. Simplified flowchart of a typical GA.


155

3.1 GA's Processes
The above three steps are repeated to create each new generation. And it continues in this fashion until some stopping condition
is reached (e.g. maximum number of generations or resulting new population not improving fast enough).
Selection: This is the first step of the three genetic operations. It determines which strings in the current population will be used to
create the next generation. This is done by using a biased random selection methodology. That is, parents are randomly selected
from the current population in such a way that the "best" strings in the population have the greatest chance of being selected. There
are many ways to do this. One wide known technique is roulette wheel parent selection (Grefenstette, 1986) which is also used
here.
Crossover: This is a randomized yet structured recombination operation. Simple crossover may proceed in two steps. First, the
newly reproduced strings in the mating pool are mated at random. Second, crossover of each pair of strings is done as follows: (i)
an integer position p along a string is selected at random in the intervals [1, L-1], where L is the string length and (ii) two new
strings are created by swapping all characters between position 1 and p inclusively.

1) Initialize Matlab/Simulink Workspace
2) Set Power System Model Parameters
(Governors, Reheaters, Turbines, Droops,
PI Controllers, Tie Line, Gains)
3) Set Deregulated Environment Parameters
(Tie Lines, Demands, Area Participation Factors,
Disco Participation Matrix, Contracts)
4) Set Genetic Algorithm Parameters
(pop size, no generations, crossover and mutation values,
other variables)
5) gen=1
6) Initialize population
(Consisting of chromosomes pertaining controller
values and frequency biases)

7) Set controller gains and freq. biases to system model
8) Simulation of the system
9) Assign fitness-value to entire population
10) Keep best values
11) gen=gen+1
while (gen < Max.Generations) or (Converge)
12) Select individuals for breeding
13) Recombine selected individuals (crossover)
14) Perform mutation on offspring
15) Form new population
16) Set controller gains and freq. biases to system model
17) Simulation of the system
18) Evaluate new population
19) Keep best values
End While

Figure 7. Genetic algorithm pseudo-code as applied to this work.

Mutation: Reproduction and crossover effectively search and recombine the existing chromosomes. However, they do not create
any new genetic material in the population. Mutation is capable of overcoming this shortcoming. It is an occasional random
alteration of a string position. In the binary string representation, this simply means changing a 1 to 0 or vice versa. This random
mutation provides background variation and occasionally introduces beneficial materials into the population.

3.2 GA's Parameters Selection
Genetic parameters, namely population size, crossover rate (P
c
) and mutation rate (P
m
), are the entities that help to tune the
performance of the GAs. The selection of values for these parameters plays an important role in obtaining an optimal solution.
There are no deterministic rules to decide these values, but there are some general guidelines which can be followed to arrive at
optimal values for these parameters and which can be found in Holland (1975).
Control Parameters Selected: First of all, the effect of population sizes for different test cases was observed. Different populations
(20, 50, 60, 80, and 100) were considered and it has been observed that the population size 50 or even 40 was satisfactory. After
selecting the population size, the effect of mutation and crossover probabilities was examined. It has been found that suitable
combination of mutation and crossover probabilities giving the best performances varies with test cases. Different combinations of
mutation probabilities (0.0001, 0.001, 0.005, 0.01) and crossover probabilities (0.6, 0.8, 0.9, and 1.0) were tested and it was found
that P
c
=1.0 and P
m
=0.005 give the best performance for all the test cases.
Encoding: The design variables are mapped onto a fixed-length binary digit string which is constructed over the binary alphabet
{0, 1}, and is concatenated head-to-tail to form one long string called a chromosome. That is, every string contains all design
variables. Each design variable is represented by a -bit string. We have to determine the value of . It is shown by Lin and Hajela
(1992) that:

2
log
c
>
u l
x x
(10)

where x
u
=upper bound on x; x
l
=lower bounds on x; =the resolution. For example, if =0.01, x
u
=60.0, x
l
=20.0, then > 12.
Decoding: The physical value of design variable x is computed from the following equation

2 1
= +
u l
l
x x
x x I (11)

156

For example, if =0.01, x
u
=60.0, x
l
=20.0 and =12, then the bit string 100000000001 is decoded to I=2049 and thus x=40.014652.
During the optimization process upper and lower bounds of all the controllers gain settings and bias factors were selected as [-50,
50] respectively, and the bit size (gene length) of each variable as 20 (i.e. =20).

3.3 Fitness-Objective Function and Performance Indices under Consideration
In GAs, the value of fitness represents the performance which is used to rank the string 0, and the ranking is then used to
determine how to allocate reproductive opportunities. This means that individuals with a higher fitness value will have a higher
opportunity of being selected as a parent. Thus, fitness is some measure of goodness to be maximized. The fitness function is
essentially the objective function for the problem. In unconstrained maximization problem, the objective function can be directly
adopted as the fitness function F=J where F is the fitness function and J is the objective function. The unconstrained minimization
problem according to the equation F=K/J where K is a positive constant multiplier. To maximize the fitness function is the same as
minimizing the objective function.
The transient performance of the two area interconnected power system in the deregulated environment with respect to the
control of the frequency and tie line powers obviously depends on the value of the controllers' gains and the frequency bias. The
optimum parameter values of the classical AGC have been obtained in the literature (using integral or proportional-plus-integral)
by minimizing the popular integral of the squared error criterion (ISE) (Abdel-Magid and Dawoud, 1997). This criterion has been
used because of the ease of computing the integral both analytically and experimentally. A characteristic of the ISE criterion is that
it weights large errors heavily and small errors lightly and it is not very selective. A system designed by this criterion tends to
show a rapid decrease in a large initial error. Hence the response is fast, oscillatory and the system has poor relative stability.
In this work, we investigate the optimum adjustment of the load frequency controllers by extending this issue, by incorporating a
set of performance indices which are various functions of error and time. In this way, someone can observe the various
performances that such a kind of power system might have when a different performance index is used. It should be noted that to
the extent of the authors' knowledge, this kind of optimization has not been done in the literature before. Finally, it is envisaged
that the synthesis procedure highlighted in this paper could be of practical significance for tuning those AGC parameters for
similar power systems. The indices adopted here include:
(a) The integral of the square of the error criterion (ISE) which is given by:
( )
2
0

=
}
ISE e t dt (12)
(b) The integral of time-multiplied absolute value of the error criterion (ITAE) which is given by:
( )
0

=
}
ITAE t e t dt (13)
This criterion penalizes long duration transients and is much more selective than the ISE. A system designed by use of this
criterion exhibits small overshoot and well damped oscillations.
(c) The integral of time-multiplied square of the error criterion (ITSE) which is given by:
( )
2
0

=
}
ITSE te t dt (14)
This criterion weights large initial error lightly, while errors occurring late in the transient response are penalized heavily. This
criterion has a better selectivity than the ISE.
(d) The integral of squared time-multiplied absolute value of the error criterion (ISTAE) which is given by:
( )
2
0

=
}
ISTAE t e t dt (15)
(e) The integral of squared time-multiplied square of the error criterion (ISTSE) which is given by:
( )
2 2
0

=
}
ISTSE t e t dt (16)
Even though performance indices (c) through (e) have not been applied to any great extend in practice due to the increased
difficulty in handling them, they are considered here. If the power system model of Figure 1 is taken into account and also, a
discrete time control is performed, Eqs. (12)-(16), are then translated into the following forms respectively

( )
2 2 2
1 1 2
0
o | = = A + A + A A
Tsim
tie
J ISE P f f t (17)

( )
2 1 2
0
o | = = A + A + A A
Tsim
tie
J ITAE t P f f t (18)

157

( )
2 2 2
3 1 2
0
o | = = A + A + A A
Tsim
tie
J ITSE t P f f t (19)

( )
2
4 1 2
0
o | = = A + A + A A
Tsim
tie
J ISTAE t P f f t (20)

( )
2 2 2 2
5 1 2
0
o | = = A + A + A A
Tsim
tie
J ISTSE t P f f t (21)

where J
m
(m=1..5) is the objective function as described above and , are penalty coefficients which in our case are set to unity.
To compute the optimum parameter values, unit step load changes are assumed in one or both areas and the performance index is
minimized using the GA. In the next Section, the optimum values of the parameters K
P
i
, K
I
i
, K
D
i
and B
i
resulting from minimizing
the five different performance indices are presented.
Since each performance index gives values in different orders of magnitude, a common performance measure must be also adopted
for comparison purposes between them. Thus, the integral or absolute error (IAE) is introduced here which gives the area under
each curve through time and can be expressed as

( )
1 2
0
o | = = A + A + A A
Tsim
c tie
J IAE P f f t (22)

4. Case Studies

The proposed controllers are applied for each control area of the restructured power system as shown in Figure 3. To illustrate
robustness of the proposed control strategies against parametric uncertainties and contract variations, simulations are carried out
for three cases of possible contracts under various operating conditions and load demands (Karnavas, 2005).

4.1 Case 1: Exclusive Case:
The GenCos in each area participate equally in AGC; i.e., all four apf values are equal to 0.5. Contracts are made only between
DisCos in area I and GenCos in area I, to purchase 0.1 puMW for each of them. In other words the load change occurs only in area
1. The following data are applicable for this case:

0.5 0.5 0 0 0.1
0.5 0.5 0 0 0.1
,
0 0 0 0 0
0 0 0 0 0
( (
( (
( (
= =
( (
( (

puMW DPM DISCO (23)

Each areas load is the sum of the local DisCos demand, i.e. PL
1
=0.2 and therefore PL
2
=0.

4.2 Case 2: Normal Case:
All the DisCos contract with the GenCos for power as per the following DPM:

0.5 0.25 0 0.3 0.105
0.2 0.25 0 0 0.045
,
0 0.25 1 0.7 0.195
0.3 0.25 0 0 0.055
( (
( (
( (
= =
( (
( (

puMW DPM DISCO (24)

It is assumed that each DisCo demands 0.1 puMW power from GenCos as defined by cpfs in DPM matrix and each GenCo
participates in AGC by following apfs: apf
1
=0.75, apf
2
=0.25, apf
3
=0.5, apf
4
=0.5. Also, PL
1
=0.15 and therefore PL
2
=0.25.

4.3 Case 3: Contract Violation Case:
It may happen that a DisCo violates a contract by demanding more power than that specified in the contract. This excess power
is not contracted out to any GenCo. This un-contracted power must be supplied by the GenCos in the same area as the DisCo. It
must be reflected as a local load of the area but not as the contract demand. So, consider Case 2 again with a modification that
DisCo1 demands 0.1 puMW of excess power, which is reflected now in PL
1
which now is set to PL
1
=0.25.


158

5. Simulation Results and Discussion

The time responses of the system are simulated and the signals are sampled at 100Hz. These responses are presented in Figures
8-10 for the I-type controller, in Figures 11-13 for the PI-type controller and in Figures 14-16 for the three cases respectively. All
these results show the effects of the load change in: (a) the areas frequency deviations, (b) the generated powers of the various
GenCos following the step change in the load demands of the DisCos and (c) the actual power flow on the tie line (e.g. in a
direction from area I to area II for Case 1). Visually, the GA-tuned system has quite fast frequency response. The frequency
deviation in each area goes to zero in the steady state. Also, in the steady state, generation of a GenCo matches the demand of the
DisCos in contract with it. Generally speaking, it can be said that no matter the type of controller adopted, each one seem to
succeed to come to steady state quite easily and also satisfactorily (but of course they pertain small oscillations around the nominal
value i.e. f=0). By inspection of the aforementioned figures it is also clear that, when PID-type controllers are used, they act too
fast to the generator inputs (this is not desirable due to the wear and tear of the machines) and also exhibit very fast oscillations.
Thus, I-type or PI-type controllers seem to be the better choice for the system under study especially when they are to be tuned
properly. Apart from that, in all cases, an acceptable overshoot and settling time on frequency deviation signal in each control area
is maintained and also, a reasonable limit on the control action signal, in the viewpoint of change speed and amplitude, is assured.
Tables 1-3 give the optimum controller gain (K
I
i
) for the I-type controller, Tables 4-6 give the optimum controller gains (K
P
i
, K
I
i
)
for the PI-type controller and Tables 7-9 give the optimum controller gains (K
P
i
, K
I
i
, K
D
i
) for the PID-type controller. Also in these
Tables, the values of the frequency biases B
i
for the two areas obtained by the application of the GA are given along with the
corresponding objective function value (J
m
) and the values of the common (comparative) performance measure (J
c
). It is clearly
shown that for different cases, different performance indices give the best optimization parameters values. The latter proves that
when research work is done in such kind of power system models, several performance indices must be examined thoroughly
before extracting valuable information about the correct system controllers tuning.
Another interesting thing to observe from Figures 8-16 (in most cases and excluding the PID-type control) is that the quantities
that are involved to the objective function (i.e. f and P
tie
), have the fastest response when the PI-type controllers are applied, but
the rest quantities (i.e. the generated powers of the GenCos) have the fastest response and with less oscillations when the I-type
controllers are applied. This also leads to the conclusion that the choice of the objective function is crucial and the behavior of the
tuned system is directly depended on it.

Table 1. Optimum values for I-type controller gains, frequency biases and fitness function obtained by GA for Case 1.
ISE ITAE ITSE ISTAE ISTSE
K
I
i
0.0692 1.4867 0.7110 1.2059 0.0569
B
i
3.5383 0.2180 0.3395 0.3056 0.2342
J
m
133.8353 548497.4906 21953.0808 420239883.6861 8891752.1239
J
c
1401.0005 1396.5185 1371.8653 1395.9370 1429.2178

K
I
i
0.2871 0.8317 0.6628 0.9254 0.4085
B
i
1.4112 0.6322 0.6617 0.5669 1.2710
J
m
285.1501 550555.8858 33268.4512 431862160.6771 10803805.4679
J
c
1728.1945 1636.2764 1650.6018 1637.7884 1665.6872

K
I
i
0.9218 0.9996 0.8219 0.7247 0.7439
B
i
0.5215 0.6116 0.5884 0.8643 0.7589
J
m
496.0482 594636.4936 50498.0860 438934198.7124 12625614.6107
J
c
2026.8078 2007.5449 2034.7790 2030.2019 2025.1652

Table 4. Optimum values for PI-type controller gains, frequency biases and fitness function obtained by GA for Case 1.
K
P
i
-0.1964 0.3311 -0.1179 -0.6538 -0.7664
K
I
i
-0.0055 0.0105 -0.0038 1.4550 1.8849
B
i
-6.7169 1.6175 -7.7037 0.3761 0.2210
J
m
75.8010 513978.2119 17054.2253 242263510.1152 4594143.2581
J
c
1243.5717 1205.5894 1188.7454 1401.9248 1227.6306


159

K
P
i
-1.2968 -0.4640 0.3264 -0.7073 -0.2618
K
I
i
-0.3681 2.2838 0.2304 -0.7557 -0.7422
B
i
-0.9744 0.2516 2.1264 -1.6999 -1.3028
J
m
115.8009 519320.0472 20667.1899 343451399.8030 7347239.0185
J
c
1168.4162 1783.2774 1341.6084 1328.8407 1404.0484

K
P
i
0.5846 0.1329 0.4658 -0.1014 -0.1699
K
I
i
0.1769 0.5836 0.3323 0.8977 2.1173
B
i
2.5360 1.1443 1.8436 0.7016 0.2515
J
m
199.2844 588128.6877 25992.1344 426312133.5791 13535587.3574
J
c
1766.0177 1853.4039 1587.7634 2166.3469 2067.8091

Table 7. Optimum values for PID-type controller gains, frequency biases and fitness function obtained by GA for Case 1.
K
P
i
2.4297 3.4432 7.0889 6.8616 9.0484
K
I
i
11.8674 47.6944 26.8778 94.9188 47.9638
K
D
i
1.1285 0.7832 2.2484 1.4102 2.3543
B
i
4.4101 5.2857 1.7220 2.6488 1.4293
J
m
0.6660 2996.6590 19.1299 1027809.5262 1101.6672
J
c
35.5077 34.5102 32.8473 40.0194 31.9470

K
P
i
2.1568 7.1090 7.3798 8.3011 9.4361
K
I
i
8.7097 77.2074 22.3201 137.7736 46.2315
K
D
i
0.9369 1.6112 2.2906 1.6732 2.3504
B
i
5.3036 2.4068 1.6806 2.5333 1.4543
J
m
1.3700 3801.0699 35.4209 1027678.5156 1953.7130
J
c
58.3010 56.2698 53.3756 74.7886 51.1510

K
P
i
1.5892 9.4166 9.7988 6.1887 5.0431
K
I
i
8.0690 83.5681 31.6823 101.3438 26.1168
K
D
i
0.6833 2.0075 2.9203 1.2508 1.2350
B
i
7.3000 1.6746 1.3093 3.3909 2.7601
J
m
2.0295 4351.3798 50.4309 1142001.3002 2767.6081
J
c
74.6823 69.3386 63.7078 90.3766 64.0060

6. Conclusions
The important role of AGC will continue in restructured electricity markets, but with modifications. Bilateral contracts can exist
between DisCos in one control area and GenCos in other control areas. The use of a DPM facilitates the simulation of these
bilateral contracts. So, the modified AGC scheme in a deregulated environment includes contract data and measurements along
with the various possible types of contracts combinations. In this new restructured environment, GenCos sell power to various
DisCos at competitive prices, and the minimization of the total cost in this open market, is one of the most important aspects. In
this context, the tuning of area controllers in an AGC deregulated system is discussed and applied. It is evident that in this tuning
process, GAs are a valuable tool and provide quite easily the best answers for such a kind of problems. Controller gains and
frequency biases are obtained for I-type, PI-type and PID-type controllers for a two-area interconnected restructured thermal
power system with reheat turbines. Several performance indices are considered. These include in addition to the popular integral
square of the error (ISE), the integral of time-multiplied absolute value of the error (ITAE), the integral of time-multiplied square
of the error (ITSE), the integral of squared time-multiplied absolute value of the error (ISTAE), and the integral of squared time-
multiplied square of the error (ISTSE). For each performance index, digital simulations of the system are carried out and
0 2 4 6 8 10 12 14 16 18 20
-0.4
-0.2
0
0.2
0.4
Time (sec)
A
f
1

(
H
z
)
0 2 4 6 8 10 12 14 16 18 20
-0.4
-0.2
0
0.2
0.4
Time (sec)
A
f
2

(
H
z
)

(a)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
1
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
2
P
o
w
e
r

(
p
u
M
W
)

0 2 4 6 8 10 12 14 16 18 20
-0.02
0
0.02
0.04
0.06
Time (sec)
G
e
n
C
o
3
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
-0.02
0
0.02
0.04
0.06
Time (sec)
G
e
n
C
o
4
P
o
w
e
r

(
p
u
M
W
)

(b)
0 2 4 6 8 10 12 14 16 18 20
-0.1
-0.05
0
0.05
0.1
Time (sec)
A
P
t
i
e

(
p
u
M
W
)

(c)
ISE ITAE ITSE
ISTAE ISTSE
(grey)
(black)

Figure 8. Transient systems responses for Case 1 with I-type
control w.r.t. the five performance indices: (a) frequency
deviations, (b) Tie line power, (c) GenCos generated power.
0 2 4 6 8 10 12 14 16 18 20
-0.4
-0.2
0
0.2
0.4
Time (sec)
A
f
1

(
H
z
)
0 2 4 6 8 10 12 14 16 18 20
-0.4
-0.2
0
0.2
0.4
Time (sec)
A
f
2

(
H
z
)

(a)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
1
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
2
P
o
w
e
r

(
p
u
M
W
)

0 2 4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
Time (sec)
G
e
n
C
o
3
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
4
P
o
w
e
r

(
p
u
M
W
)

(b)
0 2 4 6 8 10 12 14 16 18 20
-0.04
-0.02
0
0.02
0.04
Time (sec)
A
P
t
i
e

(
p
u
M
W
)

(c)
ISE ITAE ITSE
ISTAE ISTSE
(grey)
(black)


161

0 2 4 6 8 10 12 14 16 18 20
-1
-0.5
0
0.5
Time (sec)
A
f
1

(
H
z
)
0 2 4 6 8 10 12 14 16 18 20
-1
-0.5
0
0.5
Time (sec)
A
f
2

(
H
z
)

(a)
0 2 4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
Time (sec)
G
e
n
C
o
1
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
2
P
o
w
e
r

(
p
u
M
W
)

0 2 4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
Time (sec)
G
e
n
C
o
3
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
4
P
o
w
e
r

(
p
u
M
W
)

(b)
0 2 4 6 8 10 12 14 16 18 20
-0.06
-0.04
-0.02
0
0.02
Time (sec)
A
P
t
i
e

(
p
u
M
W
)

(c)
ISE ITAE ITSE
ISTAE ISTSE
(grey)
(black)

0 2 4 6 8 10 12 14 16 18 20
-0.4
-0.2
0
0.2
0.4
Time (sec)
A
f
1

(
H
z
)
0 2 4 6 8 10 12 14 16 18 20
-0.4
-0.2
0
0.2
0.4
Time (sec)
A
f
2

(
H
z
)

(a)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
1
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
2
P
o
w
e
r

(
p
u
M
W
)

0 2 4 6 8 10 12 14 16 18 20
-0.05
0
0.05
0.1
0.15
Time (sec)
G
e
n
C
o
3
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
-0.05
0
0.05
0.1
0.15
Time (sec)
G
e
n
C
o
4
P
o
w
e
r

(
p
u
M
W
)

(b)
0 2 4 6 8 10 12 14 16 18 20
-0.1
-0.05
0
0.05
0.1
Time (sec)
A
P
t
i
e

(
p
u
M
W
)

(c)
ISE ITAE ITSE
ISTAE ISTSE
(grey)
(black)

Figure 11. Transient systems responses for Case 1 with PI-type

162

0 2 4 6 8 10 12 14 16 18 20
-1
-0.5
0
0.5
Time (sec)
A
f
1

(
H
z
)
0 2 4 6 8 10 12 14 16 18 20
-1
-0.5
0
0.5
Time (sec)
A
f
2

(
H
z
)

(a)
0 2 4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
Time (sec)
G
e
n
C
o
1
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
2
P
o
w
e
r

(
p
u
M
W
)

0 2 4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
Time (sec)
G
e
n
C
o
3
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
4
P
o
w
e
r

(
p
u
M
W
)

(b)
0 2 4 6 8 10 12 14 16 18 20
-0.1
-0.05
0
0.05
0.1
Time (sec)
A
P
t
i
e

(
p
u
M
W
)

(c)
ISE ITAE ITSE
ISTAE ISTSE
(grey)
(black)

0 2 4 6 8 10 12 14 16 18 20
-1
-0.5
0
0.5
Time (sec)
A
f
1

(
H
z
)
0 2 4 6 8 10 12 14 16 18 20
-1
-0.5
0
0.5
Time (sec)
A
f
2

(
H
z
)

(a)
0 2 4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
Time (sec)
G
e
n
C
o
1
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
2
P
o
w
e
r

(
p
u
M
W
)

0 2 4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
Time (sec)
G
e
n
C
o
3
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
0
0.05
0.1
0.15
0.2
Time (sec)
G
e
n
C
o
4
P
o
w
e
r

(
p
u
M
W
)

(b)
0 2 4 6 8 10 12 14 16 18 20
-0.1
-0.05
0
0.05
0.1
Time (sec)
A
P
t
i
e

(
p
u
M
W
)

(c)
ISE ITAE ITSE
ISTAE ISTSE
(grey)
(black)


163

0 2 4 6 8 10 12 14 16 18 20
-0.1
-0.05
0
0.05
0.1
Time (sec)
A
f
1

(
H
z
)
0 2 4 6 8 10 12 14 16 18 20
-0.1
-0.05
0
0.05
0.1
Time (sec)
A
f
2

(
H
z
)

(a)
0 2 4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
Time (sec)
G
e
n
C
o
1
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
Time (sec)
G
e
n
C
o
2
P
o
w
e
r

(
p
u
M
W
)

0 2 4 6 8 10 12 14 16 18 20
-0.01
-0.005
0
0.005
0.01
Time (sec)
G
e
n
C
o
3
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
-0.01
-0.005
0
0.005
0.01
Time (sec)
G
e
n
C
o
4
P
o
w
e
r

(
p
u
M
W
)

(b)
0 2 4 6 8 10 12 14 16 18 20
-10
-5
0
5
x 10
-3
Time (sec)
A
P
t
i
e

(
p
u
M
W
)

(c)
ISE ITAE ITSE
ISTAE ISTSE
(grey)
(black)

Figure 14. Transient systems responses for Case 1 with PID-type
0 2 4 6 8 10 12 14 16 18 20
-0.1
-0.05
0
0.05
0.1
Time (sec)
A
f
1

(
H
z
)
0 2 4 6 8 10 12 14 16 18 20
-0.1
-0.05
0
0.05
0.1
Time (sec)
A
f
2

(
H
z
)

(a)
0 2 4 6 8 10 12 14 16 18 20
-0.1
0
0.1
0.2
0.3
Time (sec)
G
e
n
C
o
1
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
-0.05
0
0.05
0.1
0.15
Time (sec)
G
e
n
C
o
2
P
o
w
e
r

(
p
u
M
W
)

0 2 4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
Time (sec)
G
e
n
C
o
3
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
-0.1
0
0.1
0.2
0.3
Time (sec)
G
e
n
C
o
4
P
o
w
e
r

(
p
u
M
W
)

(b)
0 2 4 6 8 10 12 14 16 18 20
-2
0
2
4
x 10
-3
Time (sec)
A
P
t
i
e

(
p
u
M
W
)

(c)
ISE ITAE ITSE
ISTAE ISTSE
(grey)
(black)


164

0 2 4 6 8 10 12 14 16 18 20
-0.2
-0.1
0
0.1
0.2
Time (sec)
A
f
1

(
H
z
)
0 2 4 6 8 10 12 14 16 18 20
-0.2
-0.1
0
0.1
0.2
Time (sec)
A
f
2

(
H
z
)

(a)
0 2 4 6 8 10 12 14 16 18 20
-0.2
0
0.2
0.4
0.6
Time (sec)
G
e
n
C
o
1
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
-0.05
0
0.05
0.1
0.15
Time (sec)
G
e
n
C
o
2
P
o
w
e
r

(
p
u
M
W
)

0 2 4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
Time (sec)
G
e
n
C
o
3
P
o
w
e
r

(
p
u
M
W
)
0 2 4 6 8 10 12 14 16 18 20
-0.1
0
0.1
0.2
0.3
Time (sec)
G
e
n
C
o
4
P
o
w
e
r

(
p
u
M
W
)

(b)
0 2 4 6 8 10 12 14 16 18 20
-2
-1
0
1
x 10
-3
Time (sec)
A
P
t
i
e

(
p
u
M
W
)

(c)
ISE ITAE ITSE
ISTAE ISTSE
(grey)
(black)

optimization of the parameters of the AGC systems is
achieved in a simple and elegant manner through the
effective application of genetic algorithms. It is clear that the
dynamic performance of the system, using the optimal
parameters, is resulting from the minimization of a different
performance index (and not only the ISE usually used in
literature). It is also seen that the choice of the objective
function is important and affects the behavior of the system.
This way, the lack of poor damping and settling time
(relative to the other indices) as well as the improvement of
the transient error in both the frequency and tie-line power,
can be assured. Also, the results obtained indicate the
appropriateness of I or PI over the PID strategy. Further
work should include multi area systems, new controller
structures as well as different power system characteristics
i.e. hydro and diesel units for the GenCos.

Nomenclature
f
i
, f
i
power system frequency, deviation of
P
ri
areas i nominal installed capacity
P
Li
areas i load disturbance
P
tie
interchange areas tie line error
P
tj
turbines block output signal of generation
company j
P
gj
governors block output signal of
generation company j
K
psi
, T
psi
power systems block gain and time
constant of area i
K
tj
, T
tj
turbines block gain and time constant of
K
gj
, T
gj
governors block gain and time constant of
K
rj
, T
rj
reheaters block gain and time constant of
R
j
governor droop parameter of generation
company j
B
i
areas i frequency bias parameter
cpf contract participation factor
apf area participation factor
u
i
areas i controllers output signal
K
P
, K
I
, K
D
proportional, integral and derivative gains
of a PID structure controller
J performance index
F fitness function of the genetic algorithm
T
s
, t simulation and sampling time
, weight values used by the fitness function
P
c
genetic algorithms crossover rate
P
m
genetic algorithms mutation rate
chromosomes bit length
x
u
, x
l
genetic algorithm variable upper and lower
bounds
z
-1
discrete time system delay


165

Appendix A: Power System Data
P
r1
=P
r2
=2000MW, f
o
=60Hz,
K
ps1
=K
ps2
=120Hz/puMW, T
ps1
=T
ps2
=20sec,
K
t1
=K
t2
=K
t3
=K
t4
=0.5, T
t1
=T
t2
=T
t3
=T
t4
=0.3sec,
K
g1
=K
g2
=K
g3
=K
g4
=1, T
g1
=T
g2
=T
g3
=T
g4
= 0.08sec,
K
r1
=K
r2
=K
r3
=K
r4
=0.5, T
r1
=T
r2
=T
r3
=T
r4
=10.0sec,
R
1
=R
2
=R
3
=R
4
=2.4Hz/puMW, T
12
=0.545puMW,
a
121
=a
122
=-1, B
o
=0.425puMW/Hz.

Appendix B: Genetic Algorithm Parameters
Parameter Value
Maximum no of generations * 30
No of population size 50
Uniform crossover Yes
Crossover probability 1.0
Elitism Yes
Mutation probability 0.005
Creep mutations Yes
Creep mutation probability 0.02

References

Abdel-Magid Y.L., Dawoud M.M., 1997. Optimal AGC tuning with genetic algorithms, Electric Power Systems Research, Vol.
38, pp. 231-238.
Aditya S.K., Das D., 2003. Design of load frequency controllers using genetic algorithms for two area interconnected hydro power
system, Electric Power Components and Systems, Vol. 31, No. 1, pp. 81-94.
Ai-Hamouz Z.M., Abdel-Magid Y.L., 1993. Variable structure load frequency controller for multiarea power systems, Int. Journal
of Electric Power and Energy Systems, Vol. 15, No. 5, pp. 90-101.
Bevrani H., Mitani Y., and Tsuji K., 2004. Robust decentralized AGC in a restructured power system. Energy Conversion &
Management, Vol. 45, pp. 2297-2312.
Christie R., Bose A., 1996. Load-frequency control issues in power systems operation after deregulation. IEEE Trans. on Power
Systems, Vol. 11, pp. 1191-1200.
Christie R.D., Wollenberg B.F., and Wangensteen I., 2000. Transmission management in the deregulated environment. Proc. IEEE
Special Issue on The Technology of Power System Competition, Vol. 88, No. 2, pp. 170195.
Cohn N., 1986. Control of generation and power flow on interconnected systems. Wiley, New York.
Delfino B., Fornari F., and Massucco S., 2002. Load-frequency control and inadvertent interchange evaluation in restructured
power systems. Proc. IEE Gen. Trans. Distr., Vol. 149, No. 5, pp. 607614.
Demiroren A., Zeynelgil H.L., 2007. GA application to optimization of AGC in three-area power system after deregulation.
International Journal of Electrical Power & Energy Systems, Vol. 29, No. 3, pp. 230-240.
Donde V., Pai M.A., and Hiskens I.A., 2001. Simulation and optimization in an AGC system after deregulation. IEEE Trans. on
Power Systems, Vol. 16, No. 3, pp. 481-489.
Elgerd O.I., 1970. Optimum megawatt-frequency control of multi-area electric energy systems. IEEE Trans. on PAS, Vol. 89, No.
4, pp. 556-563.
Fleming P.J., Fonseca C.M., 1993. Genetic algorithms in control systems engineering, Research Report No. 470, Dept. of
Automatic Control and Systems Engineering, University of Sheffield, Sheffield, UK.
Grefenstette J.J., 1986. Optimization of control parameters for genetic algorithms. IEEE Trans. on System, Man and Cybernetics,
Vol. SMC-16, pp. 122-128.
Holland J.H., 1975. Adaptation in Nature and Artificial Systems. University of Michigan Press.
Ibraheem, Singh O., Hasan N., 2009. Genetic Algorithm Based Scheme for Optimization of AGC Gains of Interconnected Power
System. Journal of Theoretical and Applied Information Technology, Vol. 12, No. 1, pp. 33-39.
Ilic M., Galiana F., and Fink L., Eds., 1998. Power Systems Restructuring: Engineering & Economics. Boston: Kluwer Academic
Publishers.
Karnavas Y.L., 2005. On the Optimal Control of Interconnected Electric Power Systems in a Re-Structured Environment Using
Genetic Algorithms, WSEAS Transactions on Systems Journal, Vol. 4, No. 8, pp. 1248-1258.
Karnavas Y.L., Papadopoulos D.P., 2002. AGC for autonomous power station using combined intelligent techniques. Int. J.
Electric Power Systems Research, Vol. 62, pp. 225-239.
Kumar A., Malik O.P., and Hope G.S., 1985. Variable-structure-system control applied to AGC of an interconnected power
system. IEE Proc. Part C, Vol. 132, No. 1, pp. 23-29.
Kumar J., Ng K.H., Sheble G., 1997. AGC simulator for price-based operation-part I: A model. IEEE Trans. on Power Systems,
Vol. 12, No. 2, pp. 527532.
Lin C. Hajela Y.P., 1992. Genetic algorithms in optimization problems with discrete and integer design variables. Engineering
Optimization, Vol. 19, No. 4, pp. 309-327.

166

Nobile E., Bose A., and Tomsovic K., 2000. Bilateral market for load following ancillary services. Proc. PES Summer Power
Meeting, July 15-21, Seattle, WA.
Pan C.T., Lian C.M., 2005. An adaptive controller for power system load frequency control. IEEE Transactions on Power
Systems, Vol. 4, No. 1, pp. 122-128.
Shayeghi H., Shayanfar H.A., Jalili A. 2006. Multi stage fuzzy PID power system automatic generation controller in the
deregulated environment. Energy Conversion and Management, Vol. 47, No. 18, pp. 2829-2845.
Sheble G.B., 1999. Computational Auction Mechanisms for Restructured Power Industry Operation. Boston: Kluwer Academic
Publishers.
Sinha S.K., Prasad R., Patel R.N., 2010. Automatic generation control of restructured power systems with combined intelligent
techniques. International Journal of Bio-Inspired Computation, Vol. 2, No.2, pp. 124-131.
Srinivasa R.C., Siva N.S, Sangameswara R.P., 2008. AGC Tuning of TCPS Based Hydrothermal System under Open Market
Scenario with Particle Swarm Optimization, J. Electrical Systems, Vol. 4, No. 2, pp. 1-13.
Wu C. J., Huang C.H., 1997. A Hybrid Method for Parameter Tuning of PID Controllers. Journal of Franklin Institute, JFI-334B,
pp. 547-562.

Biographical notes:

Dr. Eng. Yannis L. Karnavas is Assistant Professor at the Department of Electrical Engineering, Technological Educational Institution of Crete, Heraklion, Crete,
Hellas. He has engaged in teaching and research activities the last 15 years. His fields of expertise are the analysis and modeling of power systems and electric
machines, the design and application of various types of controllers and artificial intelligence methods application to them. Dr.Eng. Karnavas has published several
papers in various national, international journals and conferences as well as book chapters in international engineering books. He has also participated in research
projects as research leader or scientific associate.

Konstantinos S. Dedousis is a graduate Electrical Engineer from the Department of Electrical Engineering, Technological Educational Institution of Crete,
Heraklion, Crete, Hellas. He has done extensive work on the field of modeling of power systems with various types of generating sources and also the field of
designing load frequency controllers for them. Presently, he is focusing on analysis of permanent magnet synchronous generators and their implementation in
power system studies.

Accepted March 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 167-174

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

A modified harmony search based method for optimal rural radial line
planning

Yan Yang
1
, Wenxin Guo
1
, Fushuan Wen
2
, Danyue Wu
3
, Yan Lin
3

(1. School of Electrical Engineering, South China University of Technology, Guangzhou 510640, CHINA;
2. School of Electrical Engineering, Zhejiang University, Hangzhou 310027, CHINA;
3. Fujian Electric Power Test and Research Institute, Fuzhou 350007, CHINA)

Abstract

Long medium voltage radial lines are usually used to supply power to large areas with a very low population/ load density.
Generally, the planning of this kind of lines involves both continuous and discrete variables, or in other word, this problem is
a mixed integer programming one. Many available algorithms may experience difficulties in solving this kind of problems. In
this work, a Harmony Search (HS) based optimization approach is developed to solve the radial line planning problem.
Furthermore, some modifications to the HS are presented for improving the computational efficiency of optimization
problems with strongly interrelated mixed variables. A sample system is served for demonstrating the feasibility and
effectiveness of the proposed approach.

Key words: radial line planning, harmony search, mixed integer programming; radial line planning

1. Introduction

Power system planning is generally a complicated nonlinear multi-objective optimization problem with different kinds of
constraints. The difficulty of solving this problem stems from the large number of optimization variables involving both
continuous and discrete types. The planning objective is to determine the most economical scheme with security constraints
respected. As it is difficult for traditional mathematical optimization methods to solve the planning problem efficiently, many
different approaches based on Artificial Intelligence (AI) such as Tabu Search (TS), Genetic Algorithm (GA) and Differential
Evolution (DE), have been proposed for this purpose in the past decade.
TS is a restricted neighbourhood search technique with the guidance of the flexible memory of the search history. Due to its
good performance, many TS based approaches have been developed for a wide variety of optimization problems in power
systems such as fault section estimation (Wen et al, 1997), transmission network optimal planning (Wen et al, 1997) and
optimal placement of capacitors in radial distribution systems (Yang et al, 1995). These applications have shown the high
efficiency of TS in handling combinatorial optimization problems. In principle, TS based methods are more suitable for 0-1
integer programming problems. However, many practical optimization problems cannot be formulated as such. GA represents
a computational technique mimicking the concept of the biological evolution process (Goldberg, 1989). Up to now, there have
been an enormous number of GA applications for solving difficult optimization problems (Carpimelli et al, 2001; Huang and
Negnevitsky, 2008; Samaan and Singh, 2004). The major advantages of GA lie in its adaptability to many kinds of problems
and its speed in obtaining optimal solutions, but its encoding requirement limits its efficient applications in some problems
(Dong et al, 2006). DE is reported to be superior to GA and Evolutionary Strategy (ES) in solving the power system planning
problem. DE is an evolutionary algorithm similar to GA, while its ability to operate on floating-point variables makes it more
flexible. Furthermore, some modifications for DE are made to better cater for the need of the planning problems (Dong et al,
2006; Yang et al, 2008; Lampinen and Zelinka, 1999). However, up to now, there is not a definite conclusion on how to well
handle the planning problem involving different kinds of variables.
In this work, Harmony Search (HS) is employed to solve the optimal rural radial line planning problem. HS was proposed
several years ago, and has been successfully applied in a wide variety of optimization problems with good convergence
performances. The rural radial line planning problem, an instance of the more general distribution system planning problem, is
illustrated to verify the efficiency of the developed approach. Line selections and the placements of voltage regulators
(including the decisions regarding whether and where voltage regulators should be installed) are taken into account. Moreover,
the placement of a voltage regulator is regarded as a whole, called a mixed variable, in the process of searching for the optimal
solution. To make HS suitable for handling mixed variables defined in this work, some modifications to HS are also made.
Yang et al. / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 167-174

168
In Section 2, the traditional HS is briefly described. The formulation of the rural line planning problem is given in Section 3.
In Section 4, the specific modifications of HS for the mixed variables involved in the rural line planning is presented. A
sample system is employed for demonstrating the feasibility and efficiency of the developed HS based approach in Section 5.
Finally, the conclusion is presented in Section 6.

2. The basic principles of harmony search

The Harmony Search (HS) algorithm was originally proposed by GEEM Z W in 2001 (Geem et al, 2001). HS is a novel
meta-heuristic search method, mimicking the improvisation process of music players, and utilizing just one solution vector for
each search (Geem et al, 2001) (Lee and Geem, 2005). Up to now, it has been successfully employed in a wide variety of areas
including power systems such as the combined heat and power economic dispatch (Vasebi et al, 2007), power control strategy
of parallel inverters (Baghaee et al, 2008) and fault current reduction in distribution systems with distributed generation units
(Baghaee et al, 2008).
To describe the working procedure of HS, an optimization problem is specified as follows:
min F( ) X
1 2
X=[X , X ]=[x ,x , ,x x ]
con dis i n
L L
Subject to:
x x x
L U
i i i

where X is a n-dimension vector composed of optimization variables to be determined, and its any element x
i
may be
continuous (X
con
) or discrete (X
dis
). F(X) is the objective function;
U
i
x and
L
i
x respectively denotes the upper and lower bounds
of x
i
. The nomenclatures used in the HS algorithm are as follows.
HM: harmony memory. A list in which some solution vectors are included, and sorted by their corresponding values of
the objective function.
HMS: harmony memory size. It represents the number of solution vectors (i.e., optimization variables) in HM.
Memory Consideration: It is used to choose the value for a certain element of the new vector from the specified HM
range.
Pitch Adjustment: It is used to adjust the value of the component obtained by the memory consideration according to a
specified strategy.
Random Selection: It is used to select one value randomly for a certain element of the new vector from the possible
range of values.
HMCR: harmony memory considering rate.
PAR: pitch adjusting rate.
bw: an arbitrary distance bandwidth.

As shown in Fig. 1, HS is mainly composed of three steps: initialization, new solution generation and updating operation.

Start
Step1 Initialize the HM
Step3 Update the HM
1 k =
max
k K
Step2 Improvise a new harmony
Is the new harmony
better than the worst
harmony in the HM?
End
No
Yes
Yes
No

Fig.1 The optimization procedure of the HS algorithm

The HS algorithm is carried out by the following steps:

Step 1 Initialize the Harmony Memory (HM).

169
Similar to most existing meta-heuristic algorithms, in HS a random rule is also employed to generate the initial solution
vector.
( ) is continuous
( ) is discrete
L U L
i and i i i
i
L U L
i and i i i
x r x x x
x
x r x x x
+
=

+

(1)
Where

denotes rounding down to the nearest integer.
and
r is a random number uniformly distributed between 0 and 1. At
the end of this step, the HM list is filled with randomly generated solution vectors.

Step 2 Improvise a new harmony
A new harmony vector,
' ' ' '
1 2
[ , ,..., ]
N
X x x x = , is generated according to the following three rules:
Memory Consideration
Pitch Adjustment
Random Selection
For instance,
'
i
x , the value of the ith optimization variable in the new vector, is generated in accordance with the memory
consideration with a probability of HMCR, while (1-HMCR) is the rate of selecting one value by the random selection. The
component obtained by the memory consideration is adjusted with the rate of PAR. If the pitch adjustment is carried out,
'
i
x is
replaced as follows:
'
'
'
i and i i
i
i and i i
x r bw x
x
x r bw x

=

i s cont i nous
i s di scr et e
(2)
where
i
bw is a distance bandwidth of the ith variable in the new vector.
As described above, the pitch adjustment and random selection are primarily responsible for maintaining the diversity of
solutions and for searching the new territory, while the memory consideration serves to improve the convergence performance.
It is a disadvantage of the traditional HS algorithm to use fixed values for both PAR and bw. In early iterations (generations),
small PAR and large bw could lead the HS algorithm to increase the diversity of solution vectors, while large PAR and small
bw are usually helpful for improving the best solution in the final iteration stage. Given this background, an improved HM
algorithm (Mahdavi et al, 2007) is presented with adaptable PAR and bw as follows:

max
( )
max
ln( )
( )
( )
( )
L
U
U c k
U L
L
PA
bw
bw
c
K
bw k bw e
PAR PAR
PAR k PAR k
K
= +
(3)

where
max
K is the permitted maximum number of generations, i.e., stopping criterion;
L U
bw bw bw and
L U
PAR PAR PAR .
As shown in Fig.2, PAR and bw change dynamically with the increase of the generation (iteration) number.

( ) PAR k
U
PAR
L
PAR
( ) bw k
k
U
bw
L
bw
a
b
k

Fig. 2 The variation of PAR and bw with the increase of the generation number

Step 3 Update the HM
If the new vector generated in Step 2 is better than the worst harmony in the HM in terms of the corresponding objective
function values, the new vector is then added in the HM and the existing worst harmony is excluded from the HM.
HS has the advantage of ease to be implemented, powerful capability at finding areas of the global optimum and at using
mathematical techniques for fine-tuning within those areas. Its effectiveness in solving nonlinear constrained optimization
problems has been proven in many applications (Kim et al, 2001) (Geem et al, 2002) (Kang and Geem, 2004).

170
3. The mathematic model for rural radial line planning

In some developed countries, long medium voltage radial lines are widely employed to supply power to large areas with very
low population and load density. As the lines are normally very long, voltage regulators are usually used to mitigate the
voltage dropping problem and hence the placements of voltage regulators should be taken into account in the planning
procedure. In this work, the line selections and placements of voltage regulators are of primary concern. Fig. 3 displays the
typical topology of a rural radial line with K loads evenly distributed along the feeder.
SLD1=P1+jQ1
SLD2=P2+jQ2
SLD 3=P3+jQ3
SLD4=P4+jQ4
SLD5=P5+jQ5
L1
L2
S0,1 S1, 2 S2,3
S3,4 S4, 5

Fig.3 The typical topology of a rural radial line

Based on the given topology structure, the planning problem can be regarded as an issue of selecting appropriate values for
the design parameters as stated below:
{ }
{ }
{ }
1 2
1 2 ,
0, 1 1, 2 1, , 1,
, , , ,
, ,
, ,
i N
i N
n n K K
b b b b
L L L L
S S S S
=
B
L
S
L
L
L L
L
L
(4)
where
B and L are both N-dimension vectors with N being the maximum permitted number of voltage regulators;
b
i
, the ith element of B, is a 0-1 variable, and represents whether the ith voltage regulator should be installed or not (1
or 0);
i L , the ith element of L, is a continuous variable between zero and the total length of the line, and represents the
location of the ith voltage regulator;
S is a K-dimension vector with K being the total number of loads;
1, n n
S
, the nth element of S , is a discrete variable, and identifies the selected type of cables for the section between
the (n-1)th and nth loads.

In fact, if the ith voltage regulator is not selected for installation (b
i
=0), then it will not be necessary to determine i L , i.e.,
the designed parameters b
i
and i L are interrelated. In most AI based methods, the design parameters are considered separately.
In this work,
i
M , a new type variable called the mixed variable, is defined to combine two related design parameters b
i
and
i L as follows:
i
M [ | ]
i i
b L =
i
M is treated as a single variable like a continuous one. The corresponding modification of HS will be introduced in the next
section to deal with this mixed variable. Then, the solution vector X can be expressed as follow:

2
1 0,1 1,2 1,
{ , } { , , , , }
N K K
M M M S S S
= = X M S L L (5)

The objective of the radial line planning is to reduce the total cost under given security constraints. Issues of installation cost,
line loss and physical constraints all affect the design-making process. In this paper, the mathematical formulation of the rural
line planning problem is as follows:
( ) ( ) ( ) ( )
RV C PL
Min F C C C = + + X X X X (6)
Subject to
min max
1, max
I
n
n n
V V V
I

where ( ) F X represents the total cost consisting of the installation cost, i.e., the cost of voltage regulators (
RV
C ( ) X ) and that
of cables ( ( )
C
C X )), and the cost line losses ( ( )
PL
C X ). The detailed formulas for calculating these costs are as follows:

171
( )
1,
1
0
( )
( ) ( )
( ) ( ) / (1 )
RV
K
C n n
n
Y
i
PL L L
i
C A T
l
C W S
K
C P C H f p
=
=
= +
X
X
X
(7)
where
A is the cost of each voltage regulator;
T is the maximum possible number of voltage regulators to be installed;
l is the total length of the line;
-1,
( )
n n
W S is the cost of per unit length of the selected line between the (n-1)th and nth loads;
L
P is the power loss at the peak load point (kW), and can be approximated by calculating the power loss per section
of the line between two loads;
L
C is the cost of per kW losses;
H is the hours per year (=8760h);
Y denotes the number of years to accumulate the losses;
f is the load factor.

Taking the time value of funding into account, the cost of power loss in the ith year should be transferred to its present value
with annual rate p.

4. Modifications of HS for rural radial line planning

In this section, the modifications to the traditional HS will be detailed so as to deal with the defined mixed variable in the last
section. The modifications are mainly concerned with the processes of initialization and improvisation.
First, the formulas used for the HM initialization are modified as follows:

'
1, if r 0.5
( )
0, else
( ) r ( ( ) ( ) ), if ( ) 1
( )
null, else
and
i
L U L
i and i i i
i
M b
M L M L M L M b
M L
<
=

+ =
(8)
It is not clear that why using harmony search instead of other algorithms.
As shown in Eqn. (8), if the ith voltage regulator is not selected for installation ( ( ) 0
i
M b = ), then the location of this
regulator is no longer an issue to consider ( ( ) null
i
M L = ). These initialization strategies are consistent with the practical
situations.
Secondly, in order to utilize the available information for the placements of voltage regulators in the current HM more
reasonably, b
i
and i L are regarded as a whole in this work. Given this background, the following modifications for
improvisation are presented. If r
and
HMCR < ,
'
i
M , the parameter corresponding to the ith regulator in the new solution vector,
is generated by the memory consideration in Eqn. (9) below, otherwise randomly according to Eqn. (8). For the pitch
adjustment in handling mixed variables, the modifications are defined in Eqn. (10).
'
, i i n
M M = (9)
'
, step
'
, step
( ) ( )
( ) ( ) r bw
i i n
i i n and
M b M b
M L M L
+
+
=
= +
(10)
where
, i n
M denotes the ith element of the nth solution vector saved in the current HM;
n is generated randomly between 0 and (HMS-1) (memory consideration).

To apply the pitch adjustment on the mixed variable, step, a randomly generated integer, is added to n with the
constraint 1 0 + HMS step n .

5. Test results

In this section, a long medium voltage radial line with 50 loads evenly distributed along the feeder will be employed as a test
system to verify the feasibility and effectiveness of the proposed planning method. The details of the test system are detailed
below:
A radial system will be designed to supply power for 50 loads.
The loads are evenly distributed along the feeder, each with its own MV/LV supply transformer.

172
Each load is 10 kVA with the power factor 0.85 lagging (under the peak load condition).
7 different candidate overhead conductors are specified, each with a different cross sectional area (and hence R & X)
and cost (as shown in Table 2).
The maximum permitted number of installed voltage regulators is 3; each can boost the voltage by a maximum of
10%. The cost of each voltage regulator is known.
Different conductors could be employed for different sections along the line (between loads).
The constraint on the voltage at each customer connection point is within 6% of the nominal voltage, i.e.
0.94V1.06.

The available data are respectively listed in Table 1 and Table 2.

Table 1 Parameters for the planning

Table 2 Parameters of candidate lines
Cable Name Cost [$/km] Maximum Current [A] R [Ohm/km] X [Ohm/km]
1 GZ 3/2.75 500 49 11 0.45
2 AC 2/2.75 720 76 5 0.45
3 Quince 800 85 4.37 0.346
4 Raisin 1200 131 2.14 0.324
5 Sultana 1500 181 1.21 0.302
6 Libra 1900 237 0.6 0.284
7 Mercury 3100 388 0.26 0.259

22kV 3-phase system planning

The planning for a 22kV 3-phase system is carried out. The simulation results are listed in Table 3 and Table 4, and the total
cost is $1807679.24.

19.1kV Single Wire Earth Return planning

The single wire earth return (SWER) or single wire ground return is usually used to supply single-phase electrical power
from an electrical grid to remote areas at a low cost. The planning for a 19.1kV SWER is tested here. The planning results are
listed in Table 3 and Table 4, and the total cost is $999351.98.

6. Conclusions

In this work, a modified HS based method is developed for the planning problem. The capability of HS in handling
constrained optimization problems with a large number of variables of different types has been demonstrated with many
application examples. For the given rural line planning problem, a new type mixed variable is defined to deal with the voltage
regulator placement problems more reasonably. Furthermore, some modifications for HS are made so as to make it more
efficient for the rural radial line planning problem. The obtained optimal solution can satisfy the power flow and other physical
constraints with low costs. In fact, the modified HS is applicable to many practical optimization problems in different areas.

Acknowledgement

This work is supported by the Doctoral Fund of the Ministry of Education of China with project number 200805610020.

References

Baghaee H.R., Mirsalim M. and Saniari M.J., 2008. Fault current reduction in distribution systems with distributed generation
units by a new dual functional series compensator. Proceedings of the 13th Power Electronics and Motion Control
Conference, Poland, pp. 750-757.
Baghaee H.R., Mirsalim M. and Saniari M.J., 2008. Power control strategy of parallel inverter interfaced DG units.
Proceedings of the 13th Power Electronics and Motion Control Conference, Poland, pp. 629-636.
Carpimelli G., Celli G., Pilo F. and Russo A., 2001. Distributed generation sitting and sizing under uncertainty. In:
Proceedings of 2001 IEEE Porto Power Tech Conference, Porto.
Parameter
L
C

Y

f

Value $0.076/kWh 20 y 0.333

173
Dong Z.Y., Lu M., Lu Z. and Wong K.P., 2006. A differential evolution based method for power system planning,
Proceedings of 2006 IEEE Congress on Evolutionary Computation, Canada, pp. 2699-2706.
Geem Z.W., Kim J.H. and Loganathan G.V., 2001. A new heuristic optimization algorithm: harmony search. Simulation, 76(2):
60-68.
Geem Z.W., Kim J.H. and Loganathan G.V., 2002. Harmony search optimization: application to pipe network design.
International Journal of Modelling and Simulation, Vol.22, No.2, pp. 125-133.
Goldberg D.E., 1989. Genetic algorithms in search, optimization and machine learning. Addison Wesley.
Huang J.S. and Negnevitsky M., 2008. A messy genetic algorithm based optimization scheme for SVC placement of power
systems under critical operation contingence. Proceedings of 2008 International Conference on Computer and Software
Engineering, China, pp. 467-472.
Kang S.L. and Geem Z.W., 2004. A new structural optimization method based on the harmony search algorithm. Computer
Structure, Vol.82, No.9-10, pp. 781-798.
Kim J.H. and Geem Z.W., Kim E.S., 2001. Parameter estimation of the nonlinear Muskingum model using harmony search,
Journal of American Water Resource Association, Vol.37, No.5, pp. 1131-1138.
Lampinen J. and Zelinka I., 1999. Mechanical engineering design optimization by differential evolution. New Ideas in
Optimization (edited by D. Corne, M. Dorigo, et al). London, McGraw-Hill International (UK) Limited, pp. 129-146.
Lee K.S. and Geem Z.W., 2005. A new meta-heuristic algorithm for continuous engineering optimization: harmony search
theory and practice. Computer Methods in Applied Mechanics and Engineering, Vol.194, No.36-38, pp. 3902-3933.
Mahdavi M., Fesanghary M. and Damangir E., 2007. An improved harmony search algorithm for solving optimization
problems. Applied Mathematics and Computation, Vol.188, No.2, pp. 1567-1579.
Samaan N. and Singh C., 2004. Genetic algorithms approach for the assessment of composite power system reliability
considering multi-state components. Proceedings of 8
th
International Conference on Probabilistic Methods Applied to Power
Systems, USA, pp. 64-69.
Vasebi A., Fesanghary M. and Bathaee S.M.T., 2007. Combined heat and power economic dispatch by harmony search
algorithm. International Journal of Electrical Power and Energy Systems, Vol.29, No.10, pp. 713-719.
Wen F.S. and Chang C.S., 1997. A Tabu search approach to fault section estimation in power systems. Electric Power Systems
Research, Vol.40, No.1, pp. 63-73.
Wen F.S. and Chang C.S., 1997. Transmission network optimal planning using Tabu search method. Electric Power Systems
Research, Vol.42, No.2, pp. 153-163.
Yang H.T. and Huang Y.C., 1995. Solution to capacitor placement problem in a radial distribution system using tabu search
method. Proceedings of 1995 International Conference on Energy Management and Power Delivery (EMPD 95), Singapore,
pp. 388-393.
Yang Y., Dong Z.Y. and Wang K.P., 2008. A modified differential evolution algorithm with fitness sharing for power system
planning. IEEE Transactions on Power Systems, Vol.23, No.2, pp. 514-522.

Table 3 The planning results on line selections
22 kV 3-phase system 19.1 kV SWER Line
Number
Connected Nodes (Unit:
kV)
Cable
Name
Connected Nodes
(Unit: kV)
Cable
Name
1 0 1/21.926 Mercury 0 1/19.012 Libra
2 1 2/21.853 Libra 1 2/18.967 Mercury
3 2 3/21.781 Libra 2 3/18.922 Mercury
4 3 4/21.653 Sultana 3 4/18.768 Raisin
5 4 5/21.525 Raisin 4 5/18.686 Libra
6 5 6/21.490 Mercury 5 6/18.644 Mercury
7 6 7/21.368 Raisin 6 7/18.500 Sultana
8 7 8/21.334 Mercury 7 8/20.212 Raisin
9 8 9/21.217 Raisin 8 9/20.143 Libra
10 9 10/21.155 Libra 9 10/20.108 Mercury
11 10 11/21.044 Raisin 10 11/19.988 Sultana
12 11 12/20.935 Raisin 11 12/19.871 Sultana
13 12 13/20.829 Raisin 12 13/19.808 Libra
14 13 14/22.884 Mercury 13 14/19.696 Raisin
15 14 15/22.834 Libra 14 15/19.332 Quince
16 15 16/22.744 Raisin 15 16/19.225 Raisin
17 16 17/22.659 Sultana 16 17/19.121 Raisin
18 17 18/22.613 Libra 17 18/19.067 Libra
19 18 19/22.532 Sultana 18 19/18.969 Sultana


174
Table 3 (contd) The planning results on line selections
22 kV 3-phase system 19.1 kV SWER Line
Number
Connected Nodes
(Unit: kV)
Cable Name Connected Nodes
(Unit: kV)
Cable
Name
20 19 20/22.489 Libra 19 20/18.875 Sultana
21 20 21/22.467 Mercury 20 21/18.783 Sultana
22 21 22/22.445 Mercury 21 22/18.735 Libra
23 22 23/22.374 Sultana 22 23/18.649 Raisin
24 23 24/22.304 Raisin 23 24/18.603 Libra
25 24 25/22.268 Libra 24 25/18.523 Raisin
26 25 26/22.232 Libra 25 26/18.446 Sultana
27 26 27/22.027 Quince 26 27/18.372 Raisin
28 27 28/21.967 Sultana 27 28/18.333 Libra
29 28 29/21.910 Raisin 28 29/18.296 Libra
30 29 30/21.856 Raisin 29 30/20.094 Libra
31 30 31/21.804 Raisin 30 31/20.037 Raisin
32 31 32/21.754 Raisin 31 32/19.893 Raisin
33 32 33/21.707 Raisin 32 33/19.956 Libra
34 33 34/21.694 Mercury 33 34/19.907 Raisin
35 34 35/21.534 AC2/2.75 34 35/19.862 Raisin
36 35 36/21.495 Sultana 35 36/19.819 Sultana
37 36 37/21.355 AC2/2.75 36 37/19.780 Sultana
38 37 38/21.321 Raisin 37 38/19.638 AC2/2.75
39 38 39/21.312 Mercury 38 39/19.604 Sultana
40 39 40/21.216 Quince 39 40/19.484 AC2/2.75
41 40 41/21.116 AC2/2.75 40 41/19.375 AC2/2.75
42 41 42/21.037 Quince 41 42/19.349 Raisin
43 42 43/20.956 AC2/2.75 42 43/19.273 Quince
44 43 44/20.886 AC2/2.75 43 44/19.200 Quince
45 44 45/20.870 Raisin 44 45/19.196 Libra
46 45 46/20.826 Quince 45 46/19.182 Raisin
47 46 47/20.816 Raisin 46 47/19.171 Sultana
48 47 48/20.813 Quince 47 48/19.138 AC2/2.75
49 48 49/20.793 AC2/2.75 48 49/19.119 Quince
50 49 50/20.776 Quince 49 50/19.099 AC2/2.75

Table 4 The planned results on voltage regulators
22 kV 3-phase system 19.1 kV SWER
Installed or not ? Site Installed or not ? Site
1 Yes 44.1km Yes 29.7km
2 No / Yes 120.9km
3 No / No /

Biographical notes:

Yan Yang is a postgraduate student in South China University of Technology. Her main research interest is power system reliability and planning.

Wenxin Guo is a postgraduate student in South China University of Technology. His main research interest is power system fault diagnosis and restoration.

Fushuan Wen received his BE and ME degrees from Tianjin University, China, in 1985 and 1988, respectively, and PhD from Zhejiang University, China, in
1991, all in electrical engineering. He joined the faculty of Zhejiang University in 1991, and has been a full professor and the director of the Institute of Power
Economics and Information since 1997. He had been a university distinguished professor, the deputy dean of the School of Electrical Engineering and the
director of the Institute of Power Economics and Electricity Markets in South China University of Technology (SCUT), China. His current research interests
lie in power industry restructuring, power system alarm processing, fault diagnosis and restoration strategies, as well as smart grids.

Accepted March 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 175-185

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Supervised fuzzy C-means clustering technique for security assessment and
classification in power systems

S. Kalyani
1
*, K.S. Swarup
1

1
Department of Electrical Engineering, Indian Institute of Technology Madras, Chennai - 600036, INDIA
*
Corresponding Author: e-mail:kal_yani_79@yahoo.co.in, Tel +91-044 -22575461, Fax.+91-044 -22574402

Abstract

Security assessment is an important concern in planning and operation studies of an electric power system. Conventional
method of security evaluation is performed by simulation consisting of load flow program and transient stability analysis,
consuming long computer time and generating voluminous results. This paper presents a practical Pattern Recognition (PR)
approach for security assessment in power systems. The problem of security assessment is focused in two modes, viz., static and
transient. Static security pertains to the study of violation of system components when subjected to contingencies like
line/generator outages. Transient Security study deals with system dynamic behavior when subjected to severe perturbations like
three phase faults. A Supervised Fuzzy C-Means (SFCM) algorithm is proposed in the classification phase of PR system for
security assessment. The proposed algorithm is tested on 39 Bus New England and IEEE 57 Bus systems. The classification
results of the proposed SFCM classifier is compared with the Method of Least Squares (MLS) and Multilayer Perceptron (MLP)
classifiers. The results prove that the former gives high classification accuracy and less misclassification rate compared to the
latter, enhancing the feasibility and applicability of SFCM algorithm for on-line security evaluation.

Keywords: Security Assessment, Static Security, Transient Security, Pattern Recognition, Fuzzy C-Means.

1. Introduction

Nowadays, power systems are forced to operate under stressed operating conditions closer to their security limits. Under such
fragile conditions, any small disturbance could endanger system security and may lead to system collapse (Niazi et al, 2004). Fast
and accurate security monitoring method, therefore, has become a key issue to ensure secure operation of the system and forewarn
the system operators to take necessary preventive actions in case need arises. Power system security is defined as the ability of the
system to withstand unforeseen contingencies without violating normal operating limits.
Security assessment (SA) refers to the analysis performed to determine whether or not a power system can meet specified
reliability and security criteria in both steady-state and transient time frames for all credible contingencies. Security analysis may
be broadly classified as (i) Static Security and (ii) Transient Security (Shahidehpour et al, 2003). Static security evaluation detects
any potential overload of a system branch or an out-of limit voltage following a given list of contingencies (Matos et al, 2000).
Transient security evaluation pertains to system dynamic behavior in terms of rotor angle stability, when subjected to
perturbations. Traditional security assessment involves numerical solution of non-linear load flow equations and transient stability
analysis for all credible contingencies (Luan et al, 2000). Because of the combinatorial nature of problem, this approach requires a
huge amount of computation time and hence found infeasible for real time security analysis of practical power system networks.
Pattern Recognition (PR) techniques have shown great importance as a means of predicting the security of large electric power
systems, overcoming the drawbacks of traditional approaches (Luan et al, 2000; Pang et al, 1973). The first step in applying PR
technique to the security assessment problem is the creation of an appropriate data set for training and testing purposes. The
required data samples called patterns are generated by off-line simulations or obtained from real time occurrences (Pang et al,
1973). The next important aspect in achieving good performance is the selection of input features, a subset of pattern variables.
Many feature selection algorithms are reported in the literature such as Principal Component Analysis, Entropy Maximization, and
Kalyani and Swarup / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 175-185

176

Fisher Discrimination (Jensen et al, 2001). In this paper, a forward sequential method is adopted for feature selection process.
Using the input features selected, a classification function is designed for accessing system security status.
The design of a suitable classifier in the pattern recognition system is an important concern for on-line security evaluation.
Literatures have reported the use of conventional algorithms like linear programming, least squares (Pang et al, 1974) and use of
expert systems like neural networks, decision trees (Luan et al, 2000) for designing the classifier. These existing algorithms seem
to work well with linearly separable classes, but not well established on non-linearly separable classes. This led to the idea of
applying clustering algorithm for power system security evaluation to handle the problem of non-linear separability between
classes. The fuzzy clustering technique of steady state security evaluation proposed by Matos (Matos et al, 2000) uses an
unsupervised learning, making it infeasible for on-line implementation. This paper presents the application of an active Supervised
Fuzzy C-Means (SFCM) clustering algorithm for real time security assessment. A given system operating state is grouped in one
of two clusters - Secure/Insecure, according to its membership value. The proposed SFCM algorithm is implemented in New
England 39 bus and IEEE 57 bus standard test systems and its performance is compared with Method of Least Squares (MLS) and
Multilayer Perceptron (MLP) classifiers. The simulation results prove that the SFCM trained classifier gives high classification
accuracy and less false dismissals, enhancing its feasibility for on-line implementation of security evaluation.

2. Security Assessment (SA)

Security Assessment is the process of determining whether and to what extent, a system is reasonably safe from serious
interference to its operation (Luan et al, 2000). It evaluates the robustness of the system (security level) to a set of contingencies in
its present state or future state. This section describes in brief the process of static security assessment and transient security
assessment carried out in power system networks.

2.1 Static Security Assessment (SSA):
Static security of a power system addresses whether, after a disturbance, the system reaches a steady state operating point
without violating system operating constraints called Security Constraints (Shahidehpour et al, 2003; Pang et al, 1973; Pang et
al, 1974). These constraints ensure the power in the network is properly balanced as given by Eq. (1), magnitude of all bus voltage
and the MVA flow in the transmission line is within acceptable limits as given by Eq. (2). If any one constraint violates, the system
may experience disruption that could result in a black-out.
1
min max
;
N
g
i
P P P P P P
Gi D loss Gi Gi Gi
=
= +
(1)
m i n m a x m a x
; V V V S S
k k k k m k m

(2)
In static security assessment process, the status of the power system is evaluated for various probable contingencies by solving
non-linear load flow equations. The contingencies may include outage of a transmission line or a transformer or a generating unit.
The load flow is solved for various disturbances and the results are compared with system constraints. The system operating state
is labeled as Static Secure (SS: Binary 1) if all the constraints (1)-(2) are satisfied for a specified contingency. If any one
constraint violation is identified, the system state is labeled as Static Insecure (SI: Binary 0).

2.2 Transient Security Assessment (TSA):
Transient security of a power system addresses whether, after a perturbation, the system proceeds to operate consistently within
the limits imposed by the system stability phenomena (Shahidehpour et al, 2003; Pang et al, 1973; Pang et al, 1974). Transient
security assessment consists of determining, whether the system oscillations, following the occurrence of a fault or a large
disturbance, will cause loss of synchronism among system generators (Hakim, 1992). TSA is a subset of transient stability
analysis. Transient stability pertains to rotor angle stability, wherein the stability phenomena are characterized by rotor dynamics
under a severe perturbation. The system state is classified as Transient Secure (TS: Binary 1) if the rotor angle of any generator
does not exceed 180
0
-
0
,
0
being the rotor angle of slack (reference) generator, after fault clearing instant, under specified
transient disturbance. On the contrary, if the rotor angle exceeds 180
0
-
0
, the system state is classified as Transient Insecure (TI:
Binary 0).

3. Application of Pattern Recognition Technique to Security Assessment

Pattern Recognition is an integral part in machine intelligence systems built for decision making. It deals with classification of
data objects, referred as Patterns, into a number of categories or classes (Theodoridis et al, 1992). The main objective of
applying pattern recognition approach to security assessment problem is to reduce on-line computational requirements. This is
done at the expense of an extensive off-line simulation, generating sufficient data points. If the separating surface between the
distinguishing classes is evaluated as a security function, the system security can be accessed at any point of time. This is the basic

177

idea of PR approach. The sequence of steps carried out in off-line in applying PR approach to security assessment is shown in the
form of a flowchart in Figure 1.

Figure 1. Stages in Design of Pattern Recognition (PR) System for Security Evaluation

As shown in Figure 1, the design of pattern recognition system using the proposed Supervised Fuzzy C-Means (SFCM)
algorithm undergoes a series of sequential steps. In this section, we briefly outline the main steps and algorithm of the fuzzy
classification scheme. The main stages are as follows:

Step 1: Data Generation (consists of generating pattern vector)
Step 2: Feature Selection (selecting a subset of pattern attributes called features)
Step 3: Classifier Design (devising a decision function called Security Function)
Step 4: Performance Evaluation (Validating and testing the designed classifier)

This section gives a brief description of each of the above stages involved in the design of pattern recognition system for the
problem of security assessment.

3.1 Step 1: Pattern (Data) Generation:
The success of pattern recognition relies on a good training set. This set must adequately represent the entire range of system
operating states. A large number of characteristic operating points are generated through off-line simulation and the security status
is evaluated for each contingency under study. Each operating point is termed as a pattern (Se-Young oh, 1986). Each pattern is
characterized by a number of attributes such as load level, voltages, power generation, etc. These attributes form the components
of a vector called pattern vector. Evaluating the security status, each pattern is labeled as belonging to secure/insecure class. The
data samples generated in this phase are randomly split into train set and test set.


178

3.2 Step 2: Feature Selection:
Feature selection reduces the dimensionality of data by selecting only a subset of measured features (predictor variables) to
create a model (Weerasooriya et al, 1986). The selected features must be capable of giving more useful information to build the
classification function. The features form the components of a vector called feature vector Z. Features may be selected by
engineering judgment. But such selections will be subjective with the possibility of important variables getting rejected. A
common method of feature selection is sequential feature selection, consisting of two components - an objective function called
criterion and a sequential search algorithm. In this paper, we use a Sequential Forward Selection (SFS) method (Theodoridis et
al, 1992). The criterion which this method seeks to minimize is misclassification for classification models. The SFS method starts
with an empty candidate set and adds feature variables sequentially until addition of further variables does not decrease the
criterion (minimization of misclassification).

3.3 Step 3: Classifier Design:
Having selected the desired feature subset, the next step in the PR system is to design an efficient classifier for the security
assessment problem. There are many training algorithms reported in literature for classifier design. Few of them include least
squares, linear programming, back propagation, etc. These algorithms, although less time consuming, were found to give poor
classification accuracy. The main requirements of a good classifier are high accuracy and less misclassification. This led to the
thought of finding a more efficient learning algorithm for classifier design. In this paper, the Supervised Fuzzy C-Means (SFCM)
clustering algorithm is identified as a suitable tool for design of classifier in the PR system. This section gives a brief outline of
fuzzy clustering and algorithm of SFCM for classification task.

3.3.1 Fuzzy Clustering:
Clustering is a process of partitioning or grouping a set of data objects into a number of clusters such that similar patterns are
assigned to one cluster. The measure of similarity or distance between respective patterns is fundamental for any clustering
technique (Ross, 2004). Depending on the data and application, different types of similarity measures like distance, connectivity
and intensity can be used. These measures control the method of cluster formation. In this paper, Euclidean distance based
similarity measure is used for class identification. Fuzzy clustering is a class of algorithm in cluster analysis wherein the allocation
of data points to clusters is not hard but fuzzy in the same sense as fuzzy logic. Fuzzy logic is a multi-valued logic derived from
fuzzy set theory, proposed by Lofti Zadeh to deal with reasoning that is approximate rather than precise (Ross, 2004).
Fuzzy C Means (FCM) is one of the widely used fuzzy partitioning scheme, originally introduced by Bezdec, as an improvement
on earlier clustering methods (Wei-Che Chen et al, 2009). FCM is an overlapping data clustering technique wherein each data
point, X
k
, belongs to a cluster i to some degree specified by a membership grade, u
ik
. The detailed algorithm of fuzzy c-means
clustering is available in Ref (Wei-Che Chen et al, 2009). The FCM algorithm outputs a list of cluster centers and membership
grades for each point. This information can be used to build a fuzzy inference system by creating appropriate membership
functions to represent the fuzzy qualities of each cluster. FCM is an unsupervised learning algorithm, which performs the
classification of data samples without utilizing the class label information. A common problem with FCM is that the cluster
structure does not necessarily correspond to the classes in the dataset, reducing its classification accuracy and efficiency. Hence,
FCM is preferred only when there is no prior information regarding class labels.

3.3.2 Supervised Fuzzy C-Means (SFCM) Clustering:
Class labels always provide a useful guidance during training process, as being done in all the learning methods. Hence, it
becomes necessary to use the labeled samples in training phase and unlabeled samples in testing phase to improve the performance
of FCM. This idea led to the development of a new algorithm called Supervised Fuzzy C-Means (SFCM) algorithm, a slight
modification of FCM (Hong-Bin et al, 2005). The SFCM clustering technique aims to develop classifiers that can utilize both
labeled and unlabeled samples. In this method of classification, a known fixed set of categories and category-labeled training data
are used to induce a classification function (Li et al, 2008). The supervised clustering can group data using the categories in the
initial labeled data, as well as extend and modify the existing set of categories to reflect other irregularities in the dataset.
The determination of fuzzy partition matrix U (dividing n data sets into c classes) using Supervised Fuzzy C Means clustering is
an iterative optimization procedure. The core of SFCM is to use the labeled data samples to guide the iterative optimization
procedure. The objective function of the SFCM optimization problem is defined as:
( ) ( )
2 2
1 1 1 1
( , )
m
c n c n
m
m ik ik ik ik ik
i k i k
J U v u d a u f d
= = = =
= +

(3)
where
U Fuzzy Partition Matrix
v
Cluster Center
ik
u Membership degree of k
th
data point belonging to the i
th
cluster (value between 0 and 1)

179

ik
d Distance measure of k
th
data point from i
th
cluster center
ik
f Membership degree of k
th
labeled sample belonging to the i
th
cluster (value is either 0 or 1)

The coefficient a denotes scaling factor and m denotes the fuzzy coefficient. The role of a is to maintain a balance between
supervised and unsupervised component within the optimization mechanism and parameter m controls the amount of fuzziness in
the classification. The typical value of m is 2 and a=L/n, L denoting the size of labeled samples (Li et al, 2008). The function J
m

can take a large number of values, the smallest one being associated with best clustering.

Algorithm for SFCM
An effective algorithm for fuzzy classification called iterative optimization procedure is discussed herein. The steps in this
algorithm are as follows:
1. Fix the number of clusters c. Initialize membership values of matrix F of size c x n with 0 or 1 in accordance with class
labels. Initialize fuzzy partition matrix U
(0)
with random values between 0 and 1, where U M
fc
.
[ ]
1 1
0,1 ; 1; 0
ik
c n
fc ik ik
u
i k
M U u u n
= =

= = < <

(4)
2. Start the iterative procedure and set the iteration count, t = 1.
3. Calculate the centers (prototypes) of the clusters using the equation (5) given below

( )
( )
( 1) ( )
( ) 1
( 1)
1
; 1, 2.....
n
m
t train
ik kj
t k
ij
n
m
t
ik
k
U Z
v j m
U
=
= =
(5)
( ) t
i
v is the i
th
cluster center described by m features (m coordinates) and arranged in the vector form represented as
{ }
( ) ( ) ( ) ( )
1 2
, , ,
t t t t
i i i im
v v v v = K
.
( ) train
kj
Z
represents the k
th
data instance corresponding to the m
th
selected feature variable.
The data matrix, Z, is the input train feature vector set obtained for SSA / TSA classification process.
4. Calculate the distance,
( ) t
ik
d , between the i
th
cluster center and k
th
data set (data point in m-space). The distance measure
used is Euclidean Distance as given by equation (6).
( )
2
( ) ( ) ( ) ( )
1
m
t t train t
ik k i kj ij
j
d Z v Z v
=
= =
(6)
5. Update the fuzzy partition matrix,
( 1) t
U
+
, for the next iteration as follows:
1
2
( ) 1
( 1)
( )
1
~
( 1)
(1 )
( ) 0
t m c
t ik
ik ik k
t
j
jk
t
ik
d
u a a f I
d
or u classes i where i
k
I
+
=
+

= + =

=
(7)
where
{ } { }
~
( ) ( 1)
2
; 0 ; 1, 2, , ; 1
k
t t
k ik k i k
c n
i I
I i d c I u
k
I
+
< <
= = = =
K (8)
6. If
( 1) ( ) t t
U U
+
( being iterative accuracy), stop the iteration and output
v
(cluster center), U (fuzzy matrix);
else increment the iteration count, t = t+1 and return to Step 3.

Note: For test set samples, whose class labels are unknown, the fuzzy matrix is updated as follows:

180

1
2
( )
1
( )
1
test
m
c
k i
ik
test
j
k j
Z v
u
Z v
(9)
where
i
v and
j
v
are values obtained from final cluster center v of the SFCM iterative training algorithm described above.
3.4 Step 4: Performance Evaluation:
The performance of the classifier designed using SFCM training algorithm is rated by evaluating the following measures for
train set, test set and combined data set.
(a) Mean Squared Error (MSE)
( )
2
1
1
;
n
k k k k
k
MSE E E DO AO
n
=
= =
(10)
where
k
DO Desired output obtained from off-line simulation (data generation)
k
AO Actual output obtained from the classifier algorithm designed in PR system

(b) Classification Accuracy (CA)
.
(%) 100
.
No of samples classified correctly
CA
Total No of samples in data set
=
(11)
(c) Misclassification (MC) Rate
(i) Secure Misclassification (SMC) or False Dismissal
. 0 ' 1
( % ) 1 0 0
. s e c ( 0 )
N o o f s c l a s s i f i e d a s
S M C
T o t a l N o o f I n u r e S t a t e s
= (12)
(ii) Insecure Misclassification (ISMC) or False Alarm
. 1 ' 0
(%) 100
. ec (1)
No of s classified as
ISMC
Total No of S ure States
=
(13)
In power system security evaluation, the false alarms are not much harmful. In case of false dismissals, failure of control actions
may lead to a severe blackout. It is, therefore, important to ensure that false dismissals are kept at minimal. The classification
system must be efficiently designed to meet this requirement.


The proposed SFCM based design of classifier for security assessment is implemented in 39 Bus New England (Pai, 1989) and
IEEE 57 Bus systems (PSTCA). The bus voltage magnitude is limited to the range of 0.90pu - 1.10pu. The generator data and their
limits are given in Appendix. The data set required for training and testing are generated by off-line simulation with programs
developed in Matlab 7.0 package. We have considered different operating scenarios by varying generation and load from 50% to
200% of their base case value. The variation in generation is limited to their minimum and maximum limits.

4.1. Results of Static Security Assessment:
In Static Security Assessment, single line outages are simulated for each operating condition. For a given operating condition
and specified contingency, load flow solution by Fast Decoupled Load Flow method is obtained. The static security status
(secure/insecure) is determined for feasible solutions by evaluating the security constraints given by equation (1)-(2). The steady
state variables of the load flow solution are recorded as pattern variables X, which includes bus voltage magnitude, bus voltage
angle, complex power generation at generator buses, complex power load at load buses and MVA flow in all branches. An optimal
subset of pattern vector called feature vector (Z) is identified by SFS feature selection method. The results of data generation and
feature selection phases for static security assessment are shown in Table 1.
The data samples of m features are randomly split into train and test set. The classifier is designed by SFCM algorithm based on
train set. Using the resultant cluster centers, the fuzzy partition matrix of test set is determined. The class labels of the test samples
are predicted by the maximal membership function value in the fuzzy matrix, U. The fuzzy coefficient m in SFCM algorithm is
assumed as 2. The performance of SFCM classifier is shown in Table 2. The results of SFCM classifier is compared with the
Method of Least Squares (MLS) and Multilayer Perceptron (MLP) classifiers. The MLS classifier is designed by Multiple Linear
Regression technique using the selected input feature set. The MLP classifier, designed using Neural Network toolbox in Matlab

181

7.0, consists of a hidden layer with 30 neurons of tansig function. The MLP network is trained with Levenberg Marquardt
algorithm (Learning rate=0.05, Goal=0.001, Epochs=600).
Table 1. Data Generation and Features of Static Security Assessment
Test Case Studied New England 39 Bus System IEEE 57 Bus System
Operating Scenarios 531 1378
Static Secure (SS) Classes 330 719
Static Insecure (SI) Classes 201 659
No. of Pattern Variables 153 243
No. of Features Selected 20 22
Dimensionality Reduction 13.072 % 9.054 %

The classification results of various classifiers show that the SFCM trained classifier gives a fairly high classification accuracy
and less secure misclassification rate, compared to MLP or MLS classifiers, as evident from Table 2. The SFCM algorithm is
capable of classifying unlabeled class samples (test set) with high accuracy and less SMC rate, as shown highlighted (boldface) in
Table 2. This feature is highly important for power system operation, where it is unrealistic to expect that all possible cases will be
encountered through off-line simulation. Figure 2 shows the comparison of the performance of classifiers for entire data set, as a
bar plot for the test cases studied. As seen from Figure 2, the results of SFCM prove to more encouraging than other classifier
algorithms, making it suitable for on-line implementation.
Table 2. Comparative Performance of Classifiers for Static Security Assessment
New England 39 Bus System IEEE 57 Bus System Test Case Studied
Classifier Type SFCM MLS MLP SFCM MLS MLP
No. of Samples 434 434 434 1213 1213 1213
Classification Accuracy (%) 100.0 70.97 70.27 100.0 63.48 75.43
Mean Squared Error 0.000 0.291 0.297 0.000 0.365 0.246
Secure Misclassification (%) 0.000 79.75 81.65 0.000 76.25 51.29
Insecure Misclassification (%) 0.000 0.000 0.000 0.000 0.000 0.000
T
R
A
I
N

S
E
T

CPU Time (s) 0.340 1.274 3.719 0.420 0.883 17.96
No. of Samples 97 97 97 165 165 165
T
E
S
T

S
E
T

CPU Time (s) 0.042 0.013 0.023 0.124 0.002 0.024

Figure 2. Overall Performance of Classifiers for Static Security Assessment

182

4.2. Results of Transient Security Assessment:
In transient security assessment process, the static security status of all the operating scenarios is first identified by running load
flow program. The operating scenarios which are static insecure (violating one or more of the constraints (1)-(2)) are ignored. Each
static secure case is subjected to transient security analysis by simulating transient disturbances (three phase faults) on all lines,
one at a time, both near sending and ending buses. The faults are applied at 0 sec and cleared at 0.25 sec (freq. being 60Hz) by
tripping the faulted line. The system dynamic equations are solved by numerical integration technique, viz., fourth-order Runge
Kutta method and the transient security status is evaluated for each disturbance. If the relative rotor angle of any generator with
respect to reference generator exceeds 180
0
-
0
after fault clearing instant, the corresponding data pattern is labeled as Transient
Insecure (0), else classified as Transient Secure (1). A simple classical model with each generator represented by constant voltage
behind transient reactance is used in the transient stability simulation.
The steady state variables from static security assessment and the variables pertaining to system dynamic behavior obtained
from transient security assessment form the components of pattern vector. The pattern variables are bus voltage magnitude and
angle, power generation and load, mechanical input power, electrical output power and relative rotor angle of generators at fault
application time and fault clearing time. The size of the pattern vector being large, we identify those variables having higher
information content by SFS feature selection method. This, in turn, yields the feature vector (Z) for classifier design. The result of
data generation and feature selection phases for transient security assessment is shown in Table 3.
Table 3. Data Generation and Features of Transient Security Assessment
Test Case Studied New England 39 Bus System IEEE 57 Bus System
Static Secure (SS) Classes 15 14
S
S
A

Static Insecure (SI) Classes 16 11
Transient Secure Classes 614 1072
T
S
A

Insecure Classes 406 692
No. of Pattern Variables 153 198
No. of Features Selected 23 7
Dimensionality Reduction 15.032 % 3.535 %

Table 4. Comparative Performance of Classifiers for Transient Security Assessment
New England 39 Bus System IEEE 57 Bus System Test Case Studied
Classifier Type SFCM MLS MLP SFCM MLS MLP
No. of Samples 910 910 910 1589 1589 1589
T
R
A
I
N

S
E
T

CPU Time (s) 0.348 0.609 10.93 0.359 0.369 8.992
No. of Samples 110 110 110 175 175 175
T
E
S
T

S
E
T

CPU Time (s) 0.098 0.002 0.015 0.044 0.002 0.015

183

Figure 3. Overall Performance of Classifiers for Transient Security Assessment

The classification function is designed based on the train set of the feature vector. The results of classifiers obtained during
training and testing phases for transient security assessment is shown in Table 4. The SFCM results are compared with MLP and
MLS classifiers. The MLP and MLS classifier are designed and trained as discussed in previous section. Figure 3 shows the
performance comparison of the classifiers as a bar plot for the test cases studied. It can be seen from Table 4 and Figure 3 that the
SFCM classifier gives a better result in terms of high classification accuracy and less misclassification rate than the other
classifiers. Furthermore, the time taken by SFCM classifier is quite acceptable, making it feasible for on-line security monitoring
system. Data robustness, overload detection, voltage monitoring and contingency analysis are widely studied in security
assessment. The proposed classifier is useful for all these analysis.

5. On-line Implementation

The security system developed based on PR approach using SFCM algorithm is feasible for on-line implementation. In on-line
mode, real time system data of the selected feature variables are measured and system static / transient security status is accessed
as shown in Figure 4. The implementation procedure shown in Figure 4 is a general idea and applicable for both SSA and TSA.
When applied for SSA, real time power flow data measurements are used as input to SFCM classifier. In case of TSA application,
real time dynamic variables like machine angle, electrical power are fed as input to classifier algorithm. For any new operating
point, use of equation (9) gives the degrees of membership to each class. The new point is assigned to the class corresponding to
maximal membership function value. This process involves very little computation effort, i.e., only manipulation of equation (9)
and hence suitable for on-line security evaluation process. Moreover, the classification of operating points among all classes with
degrees of membership helps the operator to take appropriate control decisions, especially in case of critical operating points.

Figure 4. On-line Implementation of Security Assessment Process

6. Conclusions

The application of pattern recognition approach for classifying the input feature vector representing the power system states is
presented. The classifier in the PR system is developed by a fuzzy clustering algorithm called Supervised Fuzzy C-Means
(SFCM). Training set vectors generated from off-line simulations are presented as inputs to SFCM algorithm, which uses active
supervised learning to adapt its weight vectors (cluster centers). The proposed SFCM-PR model was tested on New England 39
Bus and IEEE 57 Bus test systems for both static and transient security assessment. Simulation results show that high accuracy
classifiers with less false dismissal rate are realizable with the SFCM algorithm. Further, the SFCM algorithm involves less
computation effort, making it suitable for real time security evaluation. Future work will focus on further improving the successful
classification level by combining SFCM with machine learning algorithms like Support Vector Machines.


184

Appendix

Generator Data:
New England 39 Bus System IEEE 57 Bus System
Gen
No.
Bus
No.
P
min

(MW)
P
max

(MW)
R
a

(pu)
X
d

(pu)
H
(s)
Gen
No.
Bus
No.
P
min

(MW)
P
max

(MW)
R
a

(pu)
X
d

(pu)
H
(s)
1 30 0 350.00 0.00 0.0310 42.00 1 1 0 575.88 0.00 0.2500 4.000
2 31 0 1150.00 0.00 0.0697 30.30 2 2 0 100.00 0.00 0.2000 3.000
3 32 0 750.00 0.00 0.0531 35.80 3 3 0 140.00 0.00 0.2000 3.000
4 33 0 732.00 0.00 0.0436 28.60 4 6 0 100.00 0.00 0.2500 5.000
5 34 0 608.00 0.00 0.1320 26.00 5 8 0 550.00 0.00 0.2000 2.500
6 35 0 750.00 0.00 0.0500 34.80 6 9 0 100.00 0.00 0.2000 3.000
7 36 0 660.00 0.00 0.0490 26.40 7 12 0 410.00 0.00 0.2500 5.000
8 37 0 640.00 0.00 0.0570 24.30
9 38 0 930.00 0.00 0.0570 34.50
10 39 0 1100.00 0.00 0.0060 500.00

Acknowledgement

The first author likes to thank the Management and Principal of KLN College of Engineering, Madurai for giving an opportunity
to undergo Doctoral programme at IIT Madras under Quality Improvement Programme (QIP) scheme. The authors also thank
Indian Institute of Technology Madras for providing necessary facilities and resources to carry out this research work.

References
Hakim, H., 1992. Application of Pattern Recognition in Transient Security Assessment, Electric Power Components and Systems,
Vol.20, Issue 1, pp.1-15.
Hong-Bin, Jie Yang, Xiao-Jun Liu and Kuo-Chen Chou, 2005. Using Supervised Fuzzy Clustering to Predict Protein Structural
Classes, Biochemical and Biophysical Research Communication, Vol. 334, pp. 577-581.
Jensen, Craig A., Sharkawi, El-Mohamed A., and Marks, Robert J., 2001. Power System Security Assessment using Neural
Networks: Feature Selection using Fisher Discrimination, IEEE Transactions on Power Systems, Vol.16, No.4, pp.757-763.
Li, C., Liu, L. and Jiang, W., 2008. Objective Function of Semi-Supervised Fuzzy C-Means Clustering Algorithm, IEEE
International Conference on Industrial Informatics, Daejeon, Korea, July 13-16, pp. 737-742.
Luan, W.P., Lo, K.L., and Yu, Y.X., 2000. ANN Based Pattern Recognition Technique for Power System Security Assessment,
IEEE International Conference on Electric Utility Deregulation and Restructuring and Power Technologies 2000, City
University, London, pp. 197-202.
Matos, M.A., Hatziargyriou, N.D. and Pecas Lopes, J.A., 2000. Multi-contingency Steady State Security Evaluation using Fuzzy
Clustering Techniques, IEEE Transactions on Power Systems, Vol.15, No.1, pp.177-182.
Niazi, K.R., Arora, C.M., and Surana, S.L., 2004. Power System Security Evaluation using ANN: Feature Selection using
Divergence, Electric Power Systems Research, Vol.69, No. 2-3, pp.161-167.
Pai, M.A., 1989. Energy Function Analysis of Power System Stability, Kluwer Academic Publishers.
Pang, C.K., Kovio, A.J., and El-Abiad, A.H., 1973. Application of Pattern Recognition to Steady State Security Evaluation in a
Power System, IEEE Trans. on Systems, Mans & Cybernetics, Vol. SMC-3, No.6, pp.622-631.
Pang, C.K., Prabhakara, F.S., El-Abiad, A.H., and Kovio, A.J., 1974. Security Evaluation in Power Systems using Pattern
Recognition, IEEE Transactions on Power Apparatus & Systems, Vol. PAS-93, pp.969-976.
Power System Test Case Archive. Available at http://www.ee.washington.edu/research/pstca/
Ross, Timothy J., 2004. Fuzzy Logic with Engineering Applications, John Wiley & Sons, Second Edition.
Se-Young Oh, 1986. Pattern Recognition and Associative Memory Approach to Power System Security Assessment, IEEE
Transactions on Systems, Mans & Cybernetics, Vol. SMC-16, No. 1, pp. 62-72.
Shahidehpour, S.M. and Wang, Y., 2003. Communication and Control in Electric Systems: Applications of Parallel and
Distributed Processing, Wiley-IEEE, John Wiley & Sons.
Theodoridis, S. and Koutroumbas, Konstantinos., 2003. Pattern Recognition, John Wiley & Sons, Prentice Hall, Third Edition.
Weerasooriya, S., and El-Sharkawi, M.A., 1992. Feature Selection for Static Security Assessment using Neural Networks, IEEE
International Symposium on Circuits & Systems, California, May 10-13, pp. 1693-1696.
Wei-Che Chen, and Ming-Shi Wang, 2009. A Fuzzy C-Means Clustering-Based Fragile Watermarking Scheme for Image
Authentication, Expert Systems with Applications, Vol. 36, pp. 1300-1307.


185

Biographical notes:

S. Kalyani received her Bachelors Degree in Electrical and Electronics Engineering from Alagappa Chettiar College of Engineering Karaikudi, in the year 2000
and Masters in Power Systems Engineering from Thiagarajar College of Engineering, Madurai in December 2002. From 2003 to 2007, she was a faculty member
with the Department of Electrical and Electronics Engineering, KLN College of Engineering, Madurai, India. She is currently a Research Scholar in Dept. of
Electrical Engineering, Indian Institute of Technology Madras. Her research interests are power system stability, Pattern Recognition, Neural Networks and Fuzzy
Logic applications to Power System studies. She is a Member of IEEE since June 2009.

K. Shanti Swarup (S'87-M'92-SM'03) is currently a Professor in the Department of Electrical Engineering, Indian Institute of Technology Madras, Chennai, India.
Prior to his current position, he held positions at Mitsubishi Electric Corporation, Osaka, Japan, and Kitami Institute of Technology, Hokkaido, Japan, as a Visiting
Research Scientist and a Visiting Professor, respectively, during 1992 to 1999. He received his Bachelors Degree in Electrical Engineering from Jawaharlal Nehru
Technological University, Kakinada Andhra Pradesh and Masters Degree in Power Systems Engineering from Regional Engineering College, Warangal. He
received his PhD degree from Indian Institute of Science, Bangalore. His areas of research are artificial intelligence, knowledge-based systems, computational
intelligence, soft computing, and object modeling and design of power systems. He is a Senior Member of IEEE since 2003.

Received January, 2010
Accepted March 2010

MultiCraft

Vol. 2, No. 3, 2010, pp. 186-199

INTERNATIONAL
JOURNAL OF
ENGINEERING,
SCIENCE AND
TECHNOLOGY

www.ijest-ng.com

Power quality event classification: an overview and key issues

D. Saxena
*1
, K.S. Verma
#
and S.N. Singh
+

*
Department of Electrical and Electronics Engineering, Invertis Inst. of Engg.& Tech., Bareilly (UP), INDIA.
+
Department of Electrical Engineering, CET, Denmark Technical University, Kgs. Lyngby, DENMARK.
#
Department of Electrical Engineering, K.N.I.T Sultanpur (UP), INDIA.
E-mails: diptisx@gmail.com (D. Saxena
1
,
1
Corresponding author), snsingh@iitk.ac.in (S.N. Singh), ksv02@rediffmail.com (K.S. Verma)

Abstract

Power quality (PQ) issue has attained considerable attention in the last decade due to large penetration of power electronics
based loads and/or microprocessor based controlled loads. On one hand these devices introduce power quality problem and on
other hand these mal-operate due to the induced power quality problems. PQ disturbances/events cover a broad frequency range
with significantly different magnitude variations and can be non-stationary, thus, accurate techniques are required to identify and
classify these events/disturbances. This paper presents a comprehensive overview of different techniques used for PQ events
classifications. Various artificial intelligent techniques which are used in PQ event classification are also discussed. Major Key
issues and challenges in classifying PQ events are critically examined and outlined.

Keywords: Power quality, PQ event classifiers, artificial intelligence techniques, PQ noise, PQ key issues.

1. Introduction

An electrical power system is expected to deliver undistorted sinusoidal rated voltage and current continuously at rated
frequency to the end users. However, large penetration of power electronics based controllers and devices along with restructuring
of the electric power industry and small-scale distributed generation have put more stringent demand on the quality of electric
power supplied to the customers (Arrillaga et al, 2000a; Arrillaga et al, 2000b; Dugan et al, 2003). To define power quality (PQ),
the views of utilities, equipment manufacturers, and customers are completely different. Utilities treat PQ from the system
reliability point of view. Equipment manufacturers, on the other hand, consider PQ as being that level of power supply allowing
for proper operation of their equipment, whereas customers consider good PQ that ensures the continuous running of processes,
operations, and business. A PQ problem can be defined as any power problem manifested in voltage, current and/or frequency
deviations that result in failure or mal-operation of customers equipment. In early days, power quality (Martzloff et al, 1998)
issues were concerned with the power system transient arising due to switching and lightning surges, induction furnace and other
cyclic loads. Increased interconnection, widespread use of power electronics devices with sensitive and fast control schemes in
electrical power networks have brought many technical and economic advantages, but these have also introduced new challenges
for the power engineers (Burke et al, 1990; Domijan et al, 1993).
In new electricity market scenarios, now electricity consumers can shift to the new service providers, if power quality is not
good. Moreover, these customers can demand a higher quality of service. The utilities or other electric power providers have to
ensure a high quality of their service to remain competitive and to retain/attract the customers. The power quality term was earliest
mentioned in a study published in 1968 (Kajihara et al, 1968) following many publications (Martzloff et al, 1998) mentioned PQ
as voltage quality with reference to slow variation in voltage magnitude. The power quality analysis was, first, started at the end of
19
th
century, when rotating machinery and transformers were found to be main sources of the waveform distortion (Acha et al,
2001). PQ problems fall into two basic categories (Dwivedi et al, 2008):

1
Corresponding author. Tel: +91- 9410261368
Saxena et al. / International Journal of Engineering, Science and Technology, Vol. 2, No. 3, 2010, pp. 186-199

187

1. Events or Disturbances: Disturbances are measured by triggering on an abnormality in the voltage or the current.
Transient voltages may be detected when the peak magnitude exceeds a specified threshold. RMS voltage variations (e.g.,
sags or interruptions) may be detected when it exceeds a specified level.
2. Steady-State Variations: Steady state variation is basically a measure of the magnitude by which the voltage or current
may vary from the nominal value, plus distortion and the degree of unbalance between the three phases. These include
normal rms voltage variations, and harmonic and distortion.

The power quality events/disturbances can further be classified according to the nature of the waveform distortion. The
information regarding typical spectral content, duration and magnitude for each category of electromagnetic disturbances are
shown in Table 1. The phenomena listed in the Table 1 can be described further by various appropriate attributes. For steady-state
disturbances, the amplitude, frequency, spectrum, modulation, source impedance, notch depth and notch area attributes can be
utilized. However, for non-steady state disturbances, other attributes such as rate of rise, rate of occurrence and energy potential
are useful.
The main reasons for the increased interest in PQ can be summarized as follows:
Modern electric appliances are equipped with power electronics devices utilizing microprocessor/microcontroller. These
appliances introduce various types of PQ problems and moreover, these are very sensitive to the PQ problems.
Industrial equipments such as high-efficiency, adjustable speed motor drives and shunt capacitors are now extensively
used. The complexity of industrial processes, which results in huge economic losses if equipment fails or malfunctions.
The complex interconnection of systems, resulting in more severe consequences if any one component fails. Moreover,
various sophisticated power electronics equipments, which are very sensitive to the PQ problems, are used for improving
system stability, operation and efficiency.
There has been a significant increase in embedded generation and renewable energy sources which create new power
quality problems, such as voltage variations, flicker and waveform distortions.
Introduction of competitive electricity market gives right to the customers to demand high quality of supply.
Table 1: Classification of various power quality events
S. No. Categories Duration Voltage Magnitude
I
(a)

(b)

(c)
Short Duration Variation
Sag Instantaneous
Momentary
Temporary
Swell Instantaneous
Momentary
Temporary
Interruption Momentary
Temporary

0.5-30 cycle.
30 cycles-3 sec.
3sec-1min.
0.5-30 cycle.
30 cycles-3 sec.
3sec-1min.
0.5cycles-3sec.
3sec-1min.

0.1-0.9 pu.
0.1-0.9 pu.
0.1-0.9 pu.
1.1-1.8 pu.
1.1-1.4 pu.
1.1-1.2 pu.
<0.1 pu.
<0.1 pu.
II
(a)
(b)
(c)
Long Duration Variation
Interruption, Sustained
Under-voltage
Overvoltage

>1min
>1min
>1min

0.0 pu.
0.8-0.9 pu.
1.1-1.2 pu.
III
(a)

(b)
Transients
Impulsive Nanosecond
Microsecond
Millisecond
Oscillatory Low frequency
Medium freq.
High freq.

<50nsec.
50-1msec.
>1msec.
0.3-50msec.
20sec.
5sec.

0-4 pu.
0-8 pu.
0-4 pu.
IV Voltage Imbalance Steady state 0.5-2%
V
(a)
(b)
(c)
Waveform Distortion
Harmonics
Notching
Noise

Steady state
Steady state
Steady state

To minimize power quality disturbances and to devise suitable corrective and preventive measures, efficient detection and
classification techniques are required in the emerging power systems. Classification of power quality disturbances based on the

188

visual inspection of waveforms by human operators is laborious and time consuming. Moreover, it is not always possible to extract
important information from simple visual inspection. The classification of PQ disturbances in electric power systems has become
an important task for proper developing and designing the preventive and corrective measures.
Extensive research work and books are now available in the area of power quality. In this paper, a comprehensive survey of
different classifications techniques used for PQ events has been presented. Various artificial intelligent techniques which are used
in PQ event classification are also discussed. Major key issues and challenges in classifying PQ events are critically analyzed and
presented.

2. Automatic Power Quality Classifiers

A basic block diagram of automatic classifier is shown in Fig. 1. The disturbance signal is passed to the pre-processing unit
having two function blocks: segmentation and feature extraction. Extracted features are used to classify the PQ events. The
classifiers information is used to make the final decision through the post-processing unit. The various blocks of automatic
classification system is described below.

Figure 1. Block diagram of automatic power quality classifier

2.1 Segmentation

It is a preprocessing technique which divides the data sequence into stationary and non-stationary parts i.e. data is divided into
a large number of transition segments (corresponding to large and sudden change in signal, and event segments). Events are the
segments in between transition segments. The features are extracted from the event segments because the signal is stationary in
this segment and normally contains the information that is unique enough to distinguish between different types of disturbances.
To capture disturbance waveform period, a triggering method is required to get start and end time instant of PQ event (Bollen et al,
2006).
The current methods used for detecting PQ disturbances are based on a point-to-point comparison of adjacent cycle or a point-
to point comparison of the RMS values of the distorted signal with its corresponding pure signal and/or frequency domain
transformed data. The recent methods proposed for this purpose are parametric (or model based) such as Kalman filter (KF) and
auto-regressive (AR) models, and nonparametric (or transform based) such as short-term Fourier transform (STFT) and wavelet
transform (WT). Parametric methods exploit the prominent residuals obtained by fitting the captured waveform into the chosen
model and nonparametric methods by finding singular points from multi-state decomposition of PQ signal. Residual signal from a
model can be used for detecting transition points and for analyzing and characterizing the disturbances. The basic idea in KF and
AR residuals based detection and triggering method is to fit the PQ data into a chosen KF or AR model.

2.2 Features Extraction

The selection of suitable features is extremely important for classification of any problem. The features extracted by signal
processing techniques are used as input to the PQ classification system. Feature extraction can be defined as transformation of the
raw signal from its original form to a new form, from which suitable information can be extracted. An appropriately chosen
feature set reduces the burden over the classifiers. Feature may directly be extracted from the original measurements (e.g., RMS
values), or from some transformed domain (e.g., Fourier, wavelet, STFT, and S-transform) or from the parameters of signal
models (e.g., sinusoid KF and AR models).
Fourier transform (FT) has been used for extracting the frequency contents of the recorded signal. According to the frequency
contents of the signal, some of the PQ problems can be detected (Allen et al, 1977; Altes et al, 1980). But, FT is not suitable for
non-stationary signals. This is because FT provides information only about the existence of a certain frequency component, but

189

does not give information about component appearance time. A suitable way to obtain such information is to apply time-frequency
(or time scale) signal decomposition where time-evolved signal components in different frequency bands can be obtained.
Although, STFT can partly alleviate this problem, but STFT still has the limitation of a fixed window width i.e. the trade-off
between the frequency resolution and the time resolution should be determined a priori to observe a particular characteristic of the
signals (Gu et al, 2000). Due to a fixed window width, STFT is inadequate for the analysis of the transient non-stationary signals.
Therefore, more powerful and efficient methods and techniques are required to detect and analyze non-stationary disturbances.
To resolve the fixed resolution problem of STFT, many researchers have proposed the use of the WT approach to analyzing
the power system disturbances (Santoso et al, 1996; Perunicic et al,1998; Santoso et al, 2000a; Ibrahim et al, 2002; Zhu et al,
2004; Abdel-Galil et al, 2004). The WT approach prepares a window that automatically adjusts to give proper resolutions of both
the time and the frequency. In this approach, a larger resolution of time is provided to high-frequency components of a signal, and
a larger resolution of frequency to low-frequency components. These features make the WT well suited for the analysis of the
power system transients caused by various PQ disturbances.
S-transform has been also introduced recently in (Dash et al, 2003) as a new PQ signal analysis and feature extraction tool. In
(Chilukuri et al, 2004), the features of seven simulated signals have been extracted by calculating the minima and maxima of the
S-transform absolute matrix. Similar to the STFT, S-transform also requires significant amount of computational resources. This is
due to the fact that the S-transform matrix is calculated by performing the inverse.

2.3 PQ Classifier

Automatic classifiers used for classification of different power quality disturbances can be broadly classified into deterministic
and statistical classifiers as described below:
Deterministic Classifiers: These classifiers can be designed with limited amount of data with sufficient power system
expert knowledge. Rule based expert system and fuzzy expert system come under deterministic classification method.
Statistical Classifiers: Statistical methods are suitable when a large amount of data from training of the classifiers is
available. Artificial neural network (ANN) and support vector machines (SVM) are the statistical based classifiers. The
artificial neural network is more popular in literature and often combined with set of wavelet filters for feature extraction
and fuzzy logic for decision making.

2.4 Decision-Making

In most of the classifiers, this stage is merged with the classification stage. It assigns the type event to the current event. There
should be a proper decision tool to increase the accuracy of classification. Normally, expert system and fuzzy logic are used as
decision making tools.

3. Classification Approaches

Both conventional and artificial intelligence (AI) based classification methods are reported in the literature. The limitations of
conventional methods are overcome by the AI based methods. Some frequently used AI based classifiers are rule-based expert
systems, fuzzy classification systems, artificial neural networks, kernel machines, and support vector machines. The artificial
neural network is the most popular method in literature often combined with other AI techniques. Reference (Bollen et al, 2007)
has shown the advantage of statistical classifiers over the expert system based approach. Various classifiers are discussed in the
subsections given below.

3.1 Artificial Neural Network Based Classifiers

Neural network is a non-linear, data driven self adaptive method and is a promising tool for classification. These can adjust
themselves to the data without any explicit specification of functional or distributional form for the underlying model (Santoso et
al, 2000c; Bishop, 1995). The neural network recognizes a given pattern by experience which is acquired during the learning or
training phase when a set of finite examples is presented to the network. This set of finite examples is called the training set, and it
consists of input patterns (i.e., input vector) along with their label of classes (i.e., output). In this phase, neurons in the network
adjust their weight vectors according to certain learning rules. After the training process is completed, the knowledge needed to
recognize patterns is stored in the neurons weight vectors. The network is, then, presented to another set of finite examples, i.e.,
the testing data set, to assess how well the network performs the recognition tasks. This process is known as testing or
generalization. ANN is a universal function approximator i.e. this can approximate any function with arbitrary accuracy. All the
above mentioned attributes make ANN flexible in modeling real world complex problems.
Santoso et al (2000b) have classified six types of PQ events using wavelets and multiple neural networks. The classifier uses
wavelet transform coefficient at five-scale signal decomposition level as input to multiple neural. The squared wavelet transform
coefficients (SWTC) at each scale are used as inputs to the multiple neural networks for classifying the disturbances type. The

190

architecture of the network is learning vector quantization (LVQ). The final decision for the disturbances type is made by
combining the outcomes of multiple neural networks by using two- decision making schemes. One is simple voting scheme and
the other is Dempster-Shafer theory of evidence. The proposed classifier is able to provide a degree of belief for the identified
waveform. The classifier is able to achieve the accuracy rate of 90% by rejecting less than 10% of the waveform as ambiguous.
The scheme is not good for sag and monitoring interruptions because of the fact that the irregularities caused by these disturbances
are not significantly irregular compared to 60Hz background sinusoidal signal and the time scale over which they occur is large.
Discrete wavelet transform (DWT) has been widely used to capture the transient occurrence and exact frequency features of
PQ disturbances. DWT technology has been integrated with AI technologies to recognize and classify the PQ disturbances
accurately. It was demonstrated by (Mo et al, 1997) how to extract features from wavelet transform coefficients at different scales
that can be applied as input to neural networks for classification of non stationary signal type. Angrisani et al (1998) proposed to
employ disturbances time duration estimated by continuous wavelet transform (CWT) and disturbances amplitude estimated by
DWT to classify transient disturbance type. To extract the squared wavelet transform coefficients at each scale as inputs to neural
networks for classification of the disturbances have been proposed in (Santoso et al, 1996; Santoso et al, 2000c; Gaouda et al,
1999). (Perunicic et al, 1998) used DWT coefficient as inputs to a single layer self organizing map neural network to train and
classify the transient disturbance type. An effective wavelet multi-resolution single decomposition method was proposed for
analyzing the power quality transient events based on standard derivation (Gaouda et al, 1999) and root mean square value
(Gaouda et al, 2002a).
Although wavelet transform emerges as a powerful tool for PQ analysis, employing DWT coefficients directly as inputs to the
neural networks require large memory space and much learning time. The computing efficiency can be enhanced if decomposition
levels with the number of extraction features can be reduced. (Giang, 2004) proposed a classifier integrating discrete wavelet
transform (DWT) with probabilistic neural network (PNN). The energy distribution features of distorted signal at 13 resolution
levels are extracted using the wavelet multi-resolution analysis (MRA) technique and the Parsevals theorem. It has been found
that energy values of decomposed signal at different level are sensitive to the type of disturbances. With the help of Parsevals
Theorem, the quantity of extracted features of distorted signal is reduced without losing its property and hence requirement of
memory space and computing time for proper classification of disturbance types is less. PNN is one of the supervised learning
networks and it differs from other network in learning process (Specht, 1990). There is no requirement to set the initial weights of
the network and no learning is required.
He et al (2006) presented the classification of seven types of PQ events based on wavelet transform and self organizing
learning array (SOLAR) system. Wavelet MRA is used to calculate energy at each decomposition level which serves as input
variable for SOLAR classification which has three advantages over a typical neural network: data driven learning, local
interconnection and entropy based self organization. The proposed classifier has a robust anti-noise performance and it can achieve
high overall classification accuracy even when the signal-to-noise ratio (SNR) is very low. Comparison with support vector
machine (SVM) and inductive inference approach of classification (Abdel-Galil et al, 2004), SOLAR based on wavelet feature
extraction can effectively classify PQ disturbances.
Though WT exhibits notable capabilities for detection and localization of disturbances, however, its compatibilities are often
degraded due to the presence of noise mixed with measured waveform (Dwivedi et al, 2008; Santoso et al, 2000a; Santoso et al,
1996; Gaouda et al, 1999; Lee et al, 1997; Lin et al, 1998; Yang et al, 2001; McGranaghan et al, 2007). In particular when the
spectrum of noise coincides with that of transient signals, the effects of the noise cannot be excluded by the filters without
affecting the performance of these advanced digital signal processing (DSP) based method.
It has been shown that the use of S-transform for features extraction with neural network classifier led to a very high degree of
classification for signals contaminated with 40dB noise (almost pure signals). However, the accuracy has been dropped to 75% if
contaminated noise strength has increased. Mishra et al (2008) proposed an S-transform based probabilistic neural network (PNN)
classifier for classification of 11 types PQ disturbances with only four extracted features. Integrating S-transform with PNN can
effetely detect and classify PQ disturbances even under noisy condition. Comparison of PNN with other two well known neural
networks i.e., feed forward multilayer back propagation (FFML-BP) and learning vector quantization (LVQ) shows that PNN
classifies events more effectively than FFML and LVQ.
The main drawbacks of ANN for PQ classifications are
There is no rule to determine the data required for the training of each disturbance types. More data available for training
will give better accuracy but will increase time to train the network. A classifier trained with a small data set is very
accurate in identifying the training data set, but it is potentially unable to identify other data sets.
The generalization performance of ANN is not guaranteed and could be poor on selection of training data. The
performance is heavily dependent on the selection of training data set and the structure (or topologies) of neural networks
(e.g. the number of hidden layers, neurons, the interconnection of sub neural networks if employed).

K-nearest neighbor (kNN) pattern recognition technique has been applied to classify the PQ disturbances in (Gaouda et al,
2001; Gargoom et al, 2007). The kNN technique requires a large capacity of memory to store the training data. In (Gaouda et al,
2002b), three pattern recognition techniques (minimum Euclidean distance, kNN, and neural network) have been proposed to
automatically classify PQ disturbances. The difference in energy distribution between the distorted power signal and the pure one,

191

and the measured duration of the distortion event have been used to form feature vector for the proposed pattern recognition
techniques. It was found that the error rate for neural network classifier is generally better than that for the other two pattern
recognition techniques. The NN classifier shows higher overall performance even with high noise levels. The NN classifier is the
least sensitive to noise level variation in the input data.

3.2 Expert System Based Classifier

The main drawback of ANN technique is the need of training cycle and requirement of retraining the entire ANN for every
new PQ event (Chung et al, 2002; Youssef et al, 2004). In (Chung et al, 2002), first, a rule-based method has been used to classify
time-characterized disturbances, and then, a wavelet method has been utilized to obtain a more flexible time frequency
information. A hidden Markov model has also been adopted to determine the disturbance existence. A scalable event
identification system (a rule based expert system) has been introduced in (Santoso et al, 2000b) to present an alternative approach
to ANN in identifying power quality events. However, its poor learning ability and adaptability limit its application in real
practice. Knowledge of power quality experts plays an important role in identification of power quality disturbances. Knowledge
from power quality experts can be implemented in automatic classification system by developing a set of classification rules and
implementing rules in a expert system (Grenier 1983; Agostino et al, 1986; Ashton et al, 1990; Domijan et al, 1993; Shapiro et al,
1993). These expert system consists of a set of rules where the real intelligence by human experts is translated into artificial
intelligence for computers. The expert system consists of three basic elements (Hayes-Roth et al, 1983; Gevarter, 1985) as shown
in Fig. 2.

Figure 2. Basic structure of an expert system

The first element is the inference engine or the control procedure mechanism. This element draws inferences based on
previously available knowledge and controls the flow of analysis. The second element is the knowledge reservoir, which is a
collection of static knowledge represented by production rules (or if then rules). The third element is a user interface which
facilitates the communication between users and the expert system. The structure of the expert system described above is indeed a
common structure found in many expert system literatures.
The expert system identifies the PQ event by comparing the extracted feature from a given waveform from the knowledge
stored in the system knowledge base by rule base of technique which consists of set of production rules. The expert systems are
scalable i.e. they can be easily expanded for any new PQ event identification merely by just adding a new knowledge base module
for the new PQ event, whereas in trained ANN it is difficult. If the expert system misclassifies, the associated production rules can
always be debugged. However, it is not possible in ANN based classifiers without modifying the structure, the number of neuron
or training strategy. In (Styvaktakis et al, 2001), expert systems have been used for classification of voltage dips. Reference
(Styvaktakis et al, 2002) has also proposed an experts system for the classification and analysis of a number of power system
events. Beside, few more papers (Kazibwe et al, 1992; Schlabbach 1994; Kezunovic et al, 1999; Reaz et al, 2007) have used the
expert system as the PQ classifier. As the number of events or features increases, the complexity of the expert systems also
increases. When data are abundant and it is difficult for human experts to characterize the given PQ event, ANN classifiers are
preferred.

3.3 Fuzzy Expert System Based Classifiers

Fuzzy logic with rule based expert system has emerged as a powerful categorization tool for PQ events that is computationally
simple and fairly accurate (Chilukuri et al, 2004; Ibrahim et al, 2000; Ibrahim et al, 2001; Thapar et al, 2003). Fuzzy logic system

192

(FLS) has strong inference capabilities of expert system as well as power of natural knowledge representation. The rules of this AI
technique are based on modeling human experience and expertise. A membership function provides a measure of the degree of
similarity of an element in the fuzzy subset (Mendel, 1995). The basic block diagram of fuzzy logic system is shown in Fig. 3.

Figure 3. Fuzzy logic system

The fuzzifier maps crisp numbers into fuzzy sets. It is needed in order to activate rules which are in terms of linguistic
variables having fuzzy sets associated with them. This step is also called fuzzy matching, which calculates the degree that the input
data match the conditions of the fuzzy rules. The inference engine of the FLS maps fuzzy sets into fuzzy sets. It handles the way in
which rules are combined. There are two common approaches for the inferences, namely, the clipping method and the scaling
method. Both methods generate the conclusion by suppressing the membership function of the consequent. The extent to which
they suppress the membership function depends on the degree to which the rule is matched. The clipping method cuts off the top
of the membership function, whose value is higher than the matching degree. The scaling method scales down the membership
function in proportion to the matching degree.
Knowledge base, which is a set of fuzzy rules, is expressed as a collection of if -then statements which may be provided by the
experts. The antecedent (the rules premise) describes the degree to which input variable matches fuzzy rules and the membership
function of the output variable is assigned by (the rules consequent) the conclusion. In many applications, crisp numbers must be
obtained at the output of a FLS. The defuzzifier maps output sets into crisp numbers. Among the various types of defuzzification
methods, the center of area (COA, or centroid) and maximum are the two most widely used techniques. The COA derives the crisp
number by calculating the weighted average of the output fuzzy set while the maximum membership degree as the crisp number.
Ortiz et al (2006) have proposed a fuzzy expert system for detection and classification of voltage sags. Bizjak et al (2006)
proposed an extended fuzzy reasoning scheme for identifying eight kinds of power quality disturbances. The extracted features in
wavelet domain and fast Fourier transform (FFT) are used to form the linguistic rules. Simple disturbances such as outage can be
detected without the use of DWT and FFT. In (Dash et al, 2000), a hybrid scheme using a Fourier linear combiner and fuzzy
expert system has been presented. The captured waveforms have been passed through a Fourier linear combiner block to extract
amplitude and phase of the fundamental signal. The proposed method was found to be accurate and robust in presence of noise. It
is computationally simple and gives classification result in less than a cycle as compared to the proposed ANN based classifier in
(Dash et al, 1996) which produces large error in presence of noise and is computationally intensive.
The main disadvantage of fuzzy classifier is that system time response slows down with the increase in number of rules. If the
system does not perform satisfactorily, then the rules are reset again to obtain efficient results i.e. it is not adaptable according to
the variation in data. The accuracy of the system is dependent on the knowledge and experience of human experts. The rules
should be updated and weighting factors in the fuzzy sets should be refined with time. Neural networks, genetic algorithms, swarm
optimization techniques, etc. can be used to for fine tuning of fuzzy logic control systems.

3.4 Adaptive Neuro-Fuzzy System Based Classifier

Adaptive neuro-fuzzy system (ANFS) is a hybrid system incorporating the learning abilities of ANN and excellent knowledge
representation and inference capabilities of fuzzy logic (Jang, 1993; Jang et al, 1995; Lin et al, 1991) that have the ability to self
modify their membership function to achieve a desired performance. An adaptive network, which subsumes almost all kinds of
neural network paradigms, can be adopted to interpret the fuzzy inference system. ANFS utilizes the hybrid-learning rule and
manage complex decision-making or diagnosis systems. ANFS has proven to be an effective tool for tuning the membership
functions of fuzzy inference systems. Reference (Ibrahim et al, 2001) proposed an adaptive neuro fuzzy system to learn power

193

quality signature waveform. It is showed that adaptive fuzzy systems are very successful in learning power quality waveform. The
new adaptive neuro-fuzzy tool will enhance the performance of the existing power quality service in several ways:
Maximizing the knowledge gain from the huge amount of power quality data available.
By setting the proper boundaries of abnormal behavior, unnecessary invocation of more complex PQ analysis software
will be prevented, hence the efficiency (including cost effectiveness) of the service currently offered will be increased
drastically.
The adaptive technique used eliminates the need for any prior PQ knowledge on behalf of the customer or system
engineers.

3.5 Support Vector Machines

Support vector machines (SVMs), which are relatively recent development (Vapnik, 1998; Cristianini et al, 2000; Bishop,
2008), are a set of related supervised learning methods, introduced in the last decade, for pattern recognition and regression, and
belong to a family of generalized linear classifiers. In another terms, SVM is a classification and regression prediction tool that
uses machine learning theory to maximize predictive accuracy while automatically avoiding over-fit to the data. SVMs are based
on minimization of the misclassification probability of unseen patterns with an unknown probability distribution of data and have
solid theoretical foundation rooted in statistical learning theory. Real world problems often require hypothesis spaces that are more
complex than those using linear discriminants. SVMs are able to find non-linear boundaries if classes are linearly non-separable.
The main issue of interest in using SVM for classification is its generalization performance. SVM performs better than neural
networks in term of generalization (Cristianini et al, 2000; Bollen et al, 2007; Dwivedi et al, 2008; Bishop, 2008).
Applications within power systems using SVM have been reported in (Moulin et al, 2004; Thukaram et al, 2005; Janik et al,
2006). In (Bishop, 2008), a classifier based on radial basis function (RBF) network and SVM has been proposed and compared for
classification of four classes of PQ disturbances. It is claimed that the SVM classifier is particularly effective in the automatic
classification of voltage disturbances. The investigation revealed that the SVM network has satisfactory generalization ability and
is able to recognize sags and other disturbances correctly, for the wide range of variable parameters. In another work (Axelberg et
al, 2007), SVM based algorithm has been proposed for classification of common types of voltage sag disturbances. The results
have shown high classification accuracy which implies that, the SVM classification technique is an attractive choice for
classification of voltage sag and other PQ disturbances. It has also been found that the accuracy of the proposed method is also
dependent on the features given to the classifier. The other advantage of SVM based system is that it is straight forward to extend
the system when new types of disturbances are added to the classifier.
Bollen et al (2007) proposed a method based on statistical learning and SVMs for classification of five common types of
voltage disturbances. The proposed SVM classifier demonstrated high performance even when training data and test data originate
from different networks. Using features from both time domain (RMS signatures) and frequency domain (harmonic magnitudes
and total harmonic distortions) effectively characterize the disturbance classes. SVM has originally designed for binary
classification and effectively extending it for multiclass classification is still an ongoing research issue (Bunke et al, 2008;
Cristianini et al, 2000; Bishop et al, 2008).

3.6 Summary of PQ Classification Method

In literature, several methods have been tested for the PQ event classifications. Papers related to various approaches are
summarized in Table 2. Reference (Bollen et al, 2007) used the both expert system and support vector approaches for PQ event
classification. The strength and weakness of different artificial techniques such as expert system (ES), fuzzy system (FS), neural
network (NN), genetic algorithm (GA) and support vector machine (SVM) has been presented in Table 3 (Negnevitsky, 2004).

Table 2: List of references using different approaches
Classification Approaches References
Artificial Neural Network Based Classifiers Perunicic et al, 1998; Santoso et al, 2000c; Gaouda et al, 2002a;
Giang, 2004; He et al, 2006; Mishra et al, 2008.
Expert System Based Classifier Santoso et al, 2000b; Styvaktakis et al, 2001; Styvaktakis et al,
2002; Chung et al, 2002; Reaz et al, 2007.
Fuzzy Expert System Based Classifiers Dash et al, 2000; Thapar et al, 2003; Zhu et al, 2004; Chilukuri et
al, 2004; Ortiz et al, 2006; Bizjak et al, 2006.
Adaptive Neuro-Fuzzy System Based Classifier Ibrahim et al, 2001.
Support Vector Machines Moulin et al, 2004; Lin et al, 2006; Axelberg et al, 2007.
Artificial Neural Network & Support Vector Machines Thukaram et al, 2005; Janik et al, 2006.


194

Table 3: Strength and weakness of different AI techniques
AI Techniques
Attributes Expert
System
Fuzzy
System
Neural
network
Genetic
Algorithm
Support Vector
Machine
Knowledge representation Rather good Good Bad Rather bad Bad
Uncertainty tolerance Rather good Good Good Good Very good
Imprecision tolerance Bad Good Good Good Very good
Adaptability Bad Rather bad Good Good Good
Learning ability Bad Bad Good Good Very good
Explanation ability Good Good Bad Rather bad Bad
Knowledge discovery & data mining Bad Rather bad Good Rather good Good
Maintainability Bad Rather good Good Rather good Good
Generalization Performance Bad Bad Good Bad Excellent

4. Classification of Multiple Disturbances

The majority of classification techniques discussed above is for single disturbances. Incidence of more than one disturbance in
an electric system at same time interval is an ordinary situation owing to the presence of several sources of disturbances in power
systems. During an abnormal behavior of power system, power line signal may be corrupted by the various disturbances occurring
simultaneously i.e. multiple disturbances, and the performance of the above methods may have limited application since they have
to deal with multiple disturbances, even though they have not been designed to do so.
In reference (Abdel-Galil et al, 2004; He et al, 2006), classifications of single and two kinds of multiple disturbances have
been proposed but further studies are still required. Ribeiro et al (2007) proposed the principle of divide and conquer to decompose
electric signal into a set of primitive components for classification of single and multiple disturbances in electric network. The
proposed technique is computationally complex and cannot classify flicker, inter harmonics and unbalances. In (Lin et al, 2006),
one-versus-one approach based SVM (OSVM) is proposed to classify multiple power quality disturbances. OSVM could detect
multiple harmonics and voltage disturbance simultaneously. The above proposal method requires less memory space and
processing time by reducing a great quantity of training data. Ferreira et al (2009) proposed (ICA) independent component
analysis (Hyvarinen et al, 2001) for analyzing PQ events with multiple disturbances. The method was able to decouple the
information between independent disturbance sources for simulated and experimental data.

5. Effect of Noise on PQ Event Classifiers

In practice, signals captured by monitoring devices are often accompanied with noise thereby affecting the extraction of
important features from the signal. Noise has an adverse effect on the performances of wavelet based event detection, time
localization and classification schemes due to the difficulty of separating noise and the disturbances. Very few research works
have addressed the effect of electrical noise on performance of wavelet based detection and feature extraction techniques. The
disturbance component of the waveform carries the most important information for detection and classification of the disturbances.
When a PQ data is decomposed using wavelet transform, most of the disturbance components are reflected at higher frequency
bands which are also occupied by noise. Therefore, even if the magnitude of noise level present is not very high compared to
fundamental component, for many PQ disturbances, it is comparable to disturbance energy at these bands. Hence, presence of
noise degrades the detection capability of wavelet based PQ monitoring system. Electrical noise riding on the PQ waveform data
also modulates wavelet domain energy distribution patterns of disturbances which, in turn, degrade performance of the existing
classification schemes utilizing wavelets for feature extraction, under practical conditions.
Wavelet based denoising techniques to remove the effect of noise on PQ waveform data (Yang et al, 2000; Elmitwally et al,
2001; Yang et al, 2001; Gaouda et al, 2002b) have been proposed in the literature but, their performance degrades with decrease in
the signal to noise ratio (SNR). Among these, (Yang et al, 2001) proposed a promising method for denoising of PQ waveform data
to improve the performance of wavelet based PQ monitoring systems. But, there are fundamental problem in the formulation of the
proposed technique. Most of the above works generally deal with the denoising of PQ signals from detection and localization point
of view (Yang et al, 2000; Gaouda et al, 2002b; Elmitwally et al, 2001). However, the robustness of the energy features, extracted
for classification using DWT in the presence of noise and its effect on classification accuracy is rarely addressed.
Dwivedi et al (2006) discussed the effect of noise on the energy features in wavelet domain, extracted for detection and
classification of PQ disturbances, and proposed simple and very low complex technique for the extraction of energy features, in
noisy environment. The proposed feature extraction method does not require the tedious job of denoising at each and every

195

wavelet coefficients of the signal and hence, it is computationally very efficient and is potentially capable in extracting accurately
the required energy features even at very low SNR. Two standard statistical hypothesis test based (likelihood ratio (LRT) based
and KolmogorovSmirnov (KS) goodness-of-fit based) procedure to denoise the PQ waveform have been proposed in (Dwivedi et
al, 2009) to enhance the performance of DWT based monitoring system and to avoid miss-alarming. LTR based denoising scheme
performs superior to KS test based scheme. Reference (Dwivedi et al, 2010) proposed a low complex technique which exploits the
local structure of wavelet coefficients obtained from wavelet multi-resolution decomposition of PQ waveform data as well as high
correlation of adjacent wavelet scales for enhancing detection, segmentation and time localization performance of DWT based PQ
monitoring systems under practical conditions of harmonics and noise.

6. Key Issues and Challenges in Classification of PQ Disturbances

Accurate PQ disturbance classification, which depends on the several factors, is a difficult task. The following are the some of
the major issues and challenges in automatic classification of PQ disturbances.
There exist a few literatures on automatic classification of PQ event based on underlying cause of the disturbance. Most
of the methods deal with the type of the PQ event without specifying the underlying cause. For example, for a voltage sag
event, it is desirable to know not only voltage sag, but also whether it is caused by the switching of a large load, a line-to-
ground fault, or any other reason. The work should be extended towards cause based classification instead of phenomenon
based classification for better understanding of PQ events.
Classification of swell, dips and interruption i.e. the classification of voltage magnitude events are addressed in many
classification algorithms. New algorithms for classification of transients and harmonics distortion have to be explored.
Transient identification requiring high frequency waveform recording devices and more robust classification methods is
rarely addressed in literature with the exception of (Mamishev et al, 1996).
Effort should be aimed at incorporating knowledge and expertise of power system engineers in statistical classifiers also.
The majority of classification techniques proposed is for single disturbances. Therefore, efforts need to be done for
multiple disturbance classification. Currently, there are different approaches available for multi-class SVM. These
approaches can be tested for their suitability in classification of PQ disturbances.
Performance of a classifier is highly dependent on the input extracted features. FT, WT and/or model based methods are
the optimal starting point for generating features to be proved. The wavelet transform has limited utility in detecting,
extracting sag disturbances features because the gradient of the disturbance events are comparable to that of the
background signal.
Choice of a suitable mother wavelet function is another issue of concern in the classifier with the wavelet based extracted
features. Because of its compact and localized properties in time frequency plane, Daubechies db4 has been the most
frequently used. Another issue of concern is the number of decomposition level required to avoid possible loss of some
important information and to have accurate classifier since PQ disturbances cover a wide range of frequency.
To properly monitor the PQ events, the power quality monitors are installed in the system. Since, it is not possible to
install the PQ monitor at all the nodes in the system due to technical and economical reasons, the optimal number of
monitoring devices are to be placed in the system to get complete information about the PQ events.
Most of the studies have done training and testing on synthetic data. Therefore, accumulation of a comprehensive
standard PQ database similar to that of many other signal processing fields, for testing and comparisons of the state of the
art techniques are also needed.
Both the economics and the technical limitations must be considered before reaching at solutions. Also, the solutions need
to be evaluated using system perspective i.e. possible solutions should be identified at all levels of system from utility
supply to the end users equipment being affected.
Noise present in the signal has been a major hurdle in the accurate feature extraction and classification of PQ events.
Some methods have been proposed to denoise the signal but denoising is still a challenge to power engineers working in
the area of PQ analysis and classification.

7. Conclusion

Electric power quality, which is a current interest to several power utilities all over the world, is often severely affected by
harmonics and transient disturbances. Due to increased use of various power electronic devices in modern power systems, power
quality is becoming an important and challenging issue for the power engineers. There is no unique model which can assess the
power quality problem and to identify and classify them properly. Existing automatic recognition methods need improvement in
terms of their versatility, reliability, and accuracy. The AI tools have been applied for the PQ classification.
This paper addresses different PQ classification approaches in general and AI tools are reviewed in particular. Major key
issues and challenges related to these advanced techniques in automatic classification of PQ problems are highlighted. New

196

intelligent system technologies using DSP, expert systems, AI and machine learning provide some unique advantages in intelligent
classification of PQ distortions.

To build an ideal PQ classification system, various issues have to be addressed.
The new algorithms should be able to classify voltage or current transients due to different power system events, such as
motor starting, transformer inrush currents, capacitor switching, and high impedance faults.
The new methods need to be capable of implementing cause-based classifications, in addition to phenomenon-based
classifications.
The new methods should classify the multi PQ events occurring simultaneously.
The new methods should be noise tolerant as PQ signals are normally accompanied with noise.

Acknowledgement

Dipti Saxena is grateful to the CM and Director, Invertis Institute of Engineering and Technology, Bareilly for permitting her
to carry our PhD work from UP Technical University, Lucknow, India.

References

Abdel-Galil T.K., Kamel M., Youssef A.M., El-Saadany E.F., and Salama M.M.A., 2004. Power quality disturbance classification
using the inductive inference approach, IEEE Trans. on Power Delivery, vol. 19, pp. 1812 1818.
Acha E. and Madrial M., 2001. Power system harmonics: computer modeling and analysis, London, John Wiley and Sons Ltd.
Agostino Ralph B D', and Stephens Michael A., 1986. Goodness-of-fit techniques, Marcel Dekker Inc., New York, NY.
Allen B. and Rabiner L.R., 1977. A unified approach to short time Fourier analysis and synthesis, IEEE Proc., vol. 65, pp. 1558-
1564.
Altes R., 1980. Detection estimation and classification with spectrograms, Journal of the Acoustical Society of America, vol. 67.
Angrisani L., Daponte P., Apuzzo M.D and Testa A., 1998. A measurement method based on the wavelet transform for power
quality Analysis, IEEE Trans. on Power Delivery, vol. 13, pp. 990-998.
Arrillaga J., Bollen M.H.J. and Waston N.R., 2000a. Power quality following deregulation, Proc. IEEE., vol. 88, no. 2, pp. 246-
261.
Arrillaga J., Watson N.R. and Chen.S., 2000b. Power system quality assessment, John Wiley and Sons Ltd., London.
Ashton P.T. and Swift G.W., 1990. Measurements of transient electrical noise on low voltage distribution systems, IEEE Industry
Applications Society Annual Meeting, vol. 2, pp. 1740-1742.
Axelberg P.G.V., Gu I.Y.H. and Bollen M.H.J., 2007. Support vector machine for classification of voltage disturbances, IEEE
Trans. on Power Delivery, vol. 22, no.3, pp. 1297-1303.
Bishop C.M., 1995. Neural networks for pattern recognition, Oxford University Press, Oxford.
Bishop C.M., 2008. Pattern recognition and machine learning, Springer, New York, USA.
Bizjak Boris. and Planinsic Peter.,2006. Classification of power disturbances using fuzzy logic, Proc. of IEEE Power Electronics
and Motion Control Conference, EPE-PEMC 2006, Portoroz, Slovenia.
Bollen M.H.J. and Gu I.Y.H., 2006. Signal processing of power quality disturbances, IEEE Press, New York, USA.
Bollen M.H.J., Gu I.Y.H., Axelberg P.G.V. and Styvaktakis E., 2007. Classification of underlying causes of power quality
disturbances: deterministic versus statistical methods, EURASIP Journal on Advances in Signal Processing, pp.1-17.
Bunke H., and Kandel A., Las M., 2008. Applied pattern recognition, Springer, New York, USA.
Burke J.J., Grifith D.C., and Ward J., 1990. Power qualityTwo different perspectives, IEEE Trans. on Power Delivery, vol. 5,
pp. 15011513.
Chilukuri M.V. and Dash P.K., 2004. Multi-resolution S-transform-based fuzzy recognition system for power quality events, IEEE
Trans. on Power Delivery, vol. 19, no. 1, pp. 323 330.
Chung J., Powers E.J., Grady W.M., and Bhatt S.C., 2002. Power disturbance classifier using a rule-based method and wavelet
packet-based hidden markov model, IEEE Trans. on Power Delivery, vol. 17, pp. 233241.
Cristianini N. and Shawe-Taylor J., 2000. An introduction to support vector machines, MA: Cambridge Univ. Press, Cambridge.
Dash P.K., Mishra S., Salama M.M.A., and Liew A.C., 2000. Classification of power system disturbances using a fuzzy expert
system and a Fourier linear combiner, IEEE Trans. on Power Delivery, vol. 15, pp. 472477.
Dash P.K., Panigrahi B.K., and Panda G., 2003. Power quality analysis using S-transform, IEEE Trans. on Power Delivery, vol.
18, no. 2, pp. 406411.
Dash P.K., Swain D.P., Liew A.C. and Raman S., 1996. An adaptive linear combiner for on-line tracking of power system
harmonics, IEEE Trans. on Power Systems, vol. 11, no.4, pp.1730-1736.
Domijan A., Heydt G.T., Meliopoulos A.P.S., Venkata S.S., and West S., 1993. Directions of research in electric power quality,
IEEE Trans. on Power Delivery, vol. 8, pp. 429436.

197

Dugan Roger.C., McGranaghan M.F., Santoso S. and Beaty H.W., 2003. Electrical power system quality, 2
nd
Edn., McGraw Hill
Book Company, New York.
Dwivedi U.D., Shakya Deepti and Singh S.N., 2008. Power quality monitoring and analysis: an overview and key Issues,
International Journal of Systems Signal Control and Engineering Application, Vol., 1, No.1, pp. 74-88.
Dwivedi U. D. and Singh S.N., 2006. A robust energy features estimation for detection and classification of power quality
disturbances, Proc. 2006 IEEE PES Power India Conference, New Delhi, pp. 384390.
Dwivedi U. D., and Singh S.N., 2009. De-noising techniques with change-point approach for wavelet-based power quality
monitoring IEEE Trans. on Power Delivery vol. 24, no. 3, pp. 1719-1727.
Dwivedi U.D., and Singh S.N., 2010. Enhanced detection of power quality events using intra and inter-scale dependencies of
wavelet coefficients, IEEE Trans. on Power Delivery, vol. 25, no. 1, pp. 358-366.
Elmitwally A., Farghal S., Kandil M., Abdelkader S. and Elkateb M., 2001. Proposed wavelet-neurofuzzy combined system for
power quality violations detection and diagnosis, IEE Proc., Generation, Transmission and Distribution, vol. 148, pp. 15 20.
Ferreira Danton D., Seixas Jose M. de and Cerqueira Augusto S.,2009. ICA-based method for power quality disturbance analysis,
Proc of 15th Intelligent System Application to Power, Curitaba, Brazil.
Gaouda A.M., Kanoun S.H., and Salama M.M.A., 2001. On-line disturbance classification using nearest neighbor rule, Electrical
Power System Research, vol. 57, no. 1, pp. 18.
Gaouda A.M., Kanoun S.H., Salama M.M.A., and Chikhani A.Y., 2002a. Pattern recognition application for power system
disturbance classification, IEEE Trans. on Power Delivery, vol. 17, pp. 677683.
Gaouda A.M., Kanoun S.H., Salama M.M.A., and Chakhani A.Y., 2002b. Wavelet-based signal processing for disturbance
classification and measurement, Proc. IEE, Generation, Transmission and Distribution, vol. 149, pp. 310-318.
Gaouda A.M., Salama M.M.A., Sultan M.R., and Chikhani A.Y., 1999. Power quality detection and classification using wavelet-
multiresolution signal decomposition, IEEE Trans. on Power Delivery, vol.14, pp. 1469-1476.
Gargoom A.M., Ertugrul N., and Soong W.L., 2007. Investigation of effective automatic recognition systems of power-quality
events, IEEE Trans. on Power Delivery, vol. 22, no. 4, pp. 319 - 2326.
Gevarter W.B., 1985. Intelligent machines, an introductory perspective of artificial intelligence and robotics, Prentice-Hall.
Giang Z.L., 2004. Wavelet based network for power disturbance recognition and classification, IEEE Trans. on Power Delivery,
vol. 19, no. 4, pp. 1560-1568.
Grenier Y., 1983 Time-dependent ARM modeling of non-stationary signals, IEEE Trans. on Acoustics, Speech, and Signal
Processing, vol. 31, no.4, pp. 899-911.
Gu Y.H., Bollen M.H.J., 2000. Time-frequency and time-scale domain analysis of voltage disturbances survey, IEEE Trans on
Power Delivery, vol. 15, pp. 1279-1284.
Hayes-Roth F., Waterman D.A., and Lenat D.B., 1983. Eds., Building expert systems, Addison-Wesley.
He H. and Starzyk J.A., 2006. A self- organizing learning array system for power quality classification based on wavelet
transform, IEEE Trans. on Power Delivery, vol. 21, no. 1, pp. 286-295.
Hyvarinen, Karhunen J. and Oja E., 2001. Independent component analysis, John Wiley & Sons.
Ibrahim W.R. Anis and Morcos M.M., 2000. Preliminary application of an adaptive fuzzy system for power quality diagnostics,
IEEE Power Engineering Review, vol. 20, pp. 55-58.
Ibrahim W.R.A. and Morcos M.M., 2001. An adaptive fuzzy technique for learning power-quality signature waveforms, IEEE
Power Engineering Review, vol. 21, no.1, pp. 56-58.
Ibrahim W.R.Anis and Morcos M.M., 2002. Artificial intelligence and advanced mathematical tools for power quality
applications: a survey, IEEE Trans. on Power Delivery, vol. 17, pp. 668-673.
Janik P. and Lobos T., 2006. Automated classification of power-quality disturbances using SVM and RBF networks, IEEE Trans.
on Power Delivery, vol. 21, no.3, pp. 1663-1669.
Jang J.S.R.,1 993. ANFIS: Adaptive network based fuzzy inference system, IEEE Trans. on Systems, Man and Cybernetics, vol.
23, no. 3, pp. 665-685.
Jang J.S.R. and Sun C.T., 1995. Neuro-fuzzy modeling and control", IEEE Proc., vol. 83, no.3, pp. 378-406.
Kajihara H.H., 1968. Quality power for electronics, Electro-Technology, vol.82, no.5.
Kazibwe W.E. and Sendaula H.M., 1992. Expert system targets power quality issues, IEEE Computer Application in Power, vol.
5, pp. 2933.
Kezunovic M. and Rikalo I., 1999. Automating the analysis of faults and power quality, IEEE Computer Applications in Power,
vol. 12, pp. 4650.
Lee J.S., Lee C.H., Kim J.O., and Nam S.W., 1997. Classification of power quality disturbances using orthogonal polynomial
approximation and bi-spectra, Electronics Letters, vol. 33, no. 18, pp. 15221524.
Lin C.T. and Lee C.S.G., 1991. Neural-network-based fuzzy logic control and decision system, IEEE Trans. on Computers, vol.
40, no. 12, pp. 1320-1336.
Lin T., Ying X., and Chen D., 1998. Study on wavelet analysis and its application to numerical protection for large synchronous
generator, Proceedings of International Conference on Power System Technology, vol. 2, pp. 1121 1129.

198

Lin Whei-Min, Wu Chien-Hsien, Lin Chia-Hung and Cheng Fu-Sheng., 2006. Classification of multiple power quality
disturbances using support vector machine and one-versus-one approach, International Conference on Power System
Technology, Chongqing, pp. 1-8.
Mamishev A.V., Short S.X., Kao T.W. and Russel B.D., 1996. Analysis of high impedance faults using fractal techniques, IEEE
Trans. on Power Systems, vol. 11, no.1, pp. 435-440.
Martzloff F.D. and Gruzs T.M., 1998. Power quality site surveys: Facts, fiction, and fallacies, IEEE Trans. on Industry
Application, vol. 24, pp. 10051018.
McGranaghan Mark F. and Santoso Surya, 2007. Challenges and trends in analyses of electric power quality measurement data,
EURASIP Journal on Advances in Signal Processing, no. 1, pp. 171-177.
Mendel Jerry M., 1995. Fuzzy logic systems for engineering: a tutorial, IEEE Proc., vol. 83, no. 3,pp. 345-377.
Mishra S., Bhende C.N. and Panigrahi B.K., 2008. Detection and classification of power qulality disturbances using S-transform
and probabilistic neural network, IEEE Trans. on Power Delivery, vol. 23, No. 1.
Mo F. and Kinsner W., 1997. Wavelet modeling of transients in power systems, Proc. Of Conference on Communications, Power
and Computing, Winnipeg, MB, Canada, pp. 132-137.
Moulin L.S., Silva A.P.A. da., El-Sharkawi M.A., and Marks R.J., 2004. Support vector machines for transient stability analysis of
large-scale power systems, IEEE Trans. on Power System, vol. 19, no. 2, pp. 818825.
Negnevitsky M., 2004. Hybrid intelligent systems: neural expert systems and neuro-fuzzy
systems,www.ecst.csuchico.edu/ renner/Teaching/222/Lectures/negnevitsky/lec11_hybrids.pdf, [online].
Ortiz A., Manana M., Ortiz F., C. J., Delgado F.and Eguiluz L. I., 2006. Implementation of a voltage sag characterization based on
DSP platform, Proceedings of ICREPQ 06, Palma de Mallorca.
Perunicic B., Mallini M., Wang Z., Liu Y., and Heydt G.T., 1998. Power quality disturbance detection and classification using
wavelets and artificial neural networks, Proc. 8th Int. Conf. Harmonics and Quality of Power, pp. 7782.
Reaz M.B.I., Choong F., Sulaiman M. S., Yasin F.M. and Kamada M., 2007. Expert system for power quality disturbance
classifier, IEEE Trans. on Power Delivery, vol. 22, no. 3, pp. 1979-1988.
Ribeiro M.V. and Pereira J.L.R., 2007. Classification of single and multiple disturbances in electric signals, EURASIP Journal on
Advances in Signal Processing.
Santoso S., Powers E.J., Grady W.M. and Hofmann P., 1996. Power quality assessment via wavelet transform analysis, IEEE
Trans. on Power Delivery, vol. 11, pp. 924-930.
Santoso S., Grady W.M., Powers E.J., Lamoree J. and Bhatt S.C., 2000a. Characterization of disturbance power quality event with
Fourier and wavelet transforms, IEEE Trans. on Power Delivery, vol. 15, pp.247-254.
Santoso S., Lamoree J., Grady W.M., Powers E.J., and Bhatt S.C., 2000b. A scalable PQ event identification system, IEEE Trans.
on Power Delivery, vol. 15, pp. 738743.
Santoso S., Powers E.J., Grady W.M., and Parsons A.C., 2000c. Power quality disturbance waveform recognition using wavelet-
based neural classifier- Part1: Theoretical foundation, IEEE Trans. on Power Delivery, vol. 15, pp. 222228.
Schlabbach, 1994. Expert system measures harmonics and EMC, IEEE Computer Application in Power, vol. 7, pp. 2629.
Shapiro J.M., 1993. Embedded image coding using zero trees of wavelet coefficients, IEEE Trans. Signal Processing, vol. 41, no.
12, pp. 3445-3462.
Specht D.F., 1990. Probabilistic neural networks, Neural Networks, vol. 3, no. 1, pp. 109-118.
Styvaktakis E., Bollen M.H.J., and Gu I.Y.H., 2001. Expert system for voltage dip classification and analysis, IEEE PES Summer
Meeting, vol. 1, pp. 671676.
Styvaktakis E., Bollen M.H.J., and Gu I.Y.H., 2002. Expert system for classification and analysis of power system events, IEEE
Trans. on Power Delivery, vol. 17, pp. 423428.
Thapar A., Saha T.K., and Dong Z.Y., 2003. A fuzzy logic based recognition technique for power quality categorization,
Proceedings of the Australasian Universities Power Engineering Conference, Christchurch, New Zealand, 28 September-1.
Thukaram D., Khincha H.P. and Vijaynarasimha H.P., 2005. Artificial neural network and support vector machine approach or
locating faults in radial distribution systems, IEEE Trans. on Power Delivery, vol. 20, no. 2, pp. 710720.
Vapnik N., 1998 Statistical learning theory, John Wiley and Sons, New York.
Yang Hong-Tzer. and Liao Chiung-Chou., 2000. A correlation-based noise suppression algorithm for power quality monitoring
through wavelet transform, International Conference on Power System Technology, 2000. Proceedings Power Conf. 2000, vol.
3, pp. 1311 1316
Yang Hong-Tzer, and Liao Chiung-Chou, 2001. A de-noising scheme for enhancing wavelet-based power quality monitoring
system, IEEE Trans on Power Delivery, vol. 16, no. 3, pp. 353 - 360.
Yen J., and Langari R., 1999. Fuzzy logic: intelligence, control, and information, Englewood Clipps, NJ: Prentice-Hall.
Youssef A.M., Abdel-Galil T.K., El-Saadany E.F., and Salama M.M.A., 2004. Disturbance classification utilizing dynamic time
warping classifier, IEEE Trans. on Power Delivery, vol. 19, no. 1, pp. 272278.
Zhu T.X., Tso S.K., Lo K.L., 2004. Wavelet-based fuzzy reasoning approach to power-quality disturbance recognition IEEE
Trans. on Power Delivery, vol. 19, pp. 1928-1935.


199

Biographical notes:
D. Saxena obtained B. Tech. in Electrical Engineering from KNIT Sultanpur (UP), India and M. Tech. in Process Control from Netaji Subhas Institute of
Technology, New Delhi, India in 1999 and 2003, respectively. Presently, she is working as reader in Electrical and Electronics Engineering Department at Invertis
Institute of Engg. & Technology, Bareilly. She is registered as PhD candidate in UP Technical University Lucknow, India. Her research interests are power
quality, power electronics, control systems and DSP application in power.

K.S. Verma received his B. Tech. (Hons) in Electrical Engineering and M. Tech. in Electrical Engineering (Power Systems) both from KNIT Sultanpur (India) in
1987 and 1998, respectively. He obtained his Ph.D. degree in Electrical Engineering (Power Systems) from Indian Institute of Technology, Roorkee, Roorkee
(India) in 2003. Presently, Prof Verma is Director, KNIT Sultanpur (India). His research interests include FACTS Technology, distributed generation, power
system optimization & control, power quality and AI application in power system.

S. N. Singh received M. Tech. and Ph.D. from Indian Institute of Technology Kanpur, India in 1989 and 1995, respectively. He is a Professor in the Department of
Electrical Engineering, Indian Institute of Technology Kanpur, India. Presently he is on leave from IIT Kanpur and working with the Centre for Electric
Technology (CET), Technical University of Denmark, Denmark. His research interests include power system restructuring, power system optimization & control,
voltage security and stability analysis, power system planning, and ANN application to power system problems. He is a Fellow of IE (India), Fellow of IETE
(India) and senior member of IEEE.

Accepted March 2010

No part of this journal may be reproduced by any means, nor transmitted, nor translated into any language without permission of
the publisher
Published in Nigeria by
MultiCraft Limited
14, Olajide Street, Iyana Ipaja, Lagos.

IJEST Special Issue Complete

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

IJEST Special Issue Complete

Încărcat de

Drepturi de autor:

Formate disponibile

MultiCraft

ISSN 2141-2839 (Online); ISSN 2141-2820 (Print)

for consumers, and (12)

for producers. (13)

, and the line flow

Generator shift factor of line k and bus i.

Bilateral contract price at period t ($/MWh).

International Journal of Engineering,

S-ar putea să vă placă și