Sunteți pe pagina 1din 7

ADEQUACY ASSESSMENT OF COMPOSITE POWER SYSTEMS THROUGH

HYBRIDIZATION OF MONTE CARLO SIMULATION AND ARTIFICIAL


IMMUNE RECOGNITION SYSTEM
Lingfeng Wang, Chanan Singh
Department of Electrical and Computer Engineering
Texas A&M University
College Station, TX 77843
l.f.wang@ieee.org, singh@ece.tamu.edu

Abstract - Adequacy assessment of power systems provides an effective mechanism to ensure proper or acceptable
system performance in the presence of different uncertainties. Monte Carlo simulation (MCS) has been widely used to
evaluate system adequacy of complex power systems. However, one major drawback of MCS is its high computational
cost when power flow calculation is involved in determining system state. Also, the target problem is concerned with
highly imbalanced data sets which may hinder the performance of general classifiers. In this investigation, we propose
a novel hybrid algorithm by combining Monte Carlo simulation with artificial immune recognition system (AIRS). AIRS
is inspired by the biological immune system and has shown
to be an effective classifier. The hyridization significantly decreases the state evaluation time by first deriving a set of artificial recognition balls (ARBs) in the training process, which
can then be used to classify system states without power flow
analyses. A comparison is made with respect to artificial neural network based classifiers including standard backpropagation neural network (BPNN) and Self-Organizing Map
(SOM), and it demonstrates a better performance of the proposed algorithm in terms of both classification accuracy and
computational time.

Keywords - Adequacy assessment, composite power


systems, Monte Carlo simulation, artificial immune
recognition system, pattern classification.
1

Introduction

Probabilistic methods have been widely used in power


system operations and planning due to a variety of uncertainties involved. For instance, adequacy assessment is
an important component to ensure the proper operations
of power systems and some methods have been proposed
for this purpose [1-3]. Power generation system adequacy
relates to the facilities necessary to generate sufficient energy in the presence of different uncertainties. Different
adequacy indices are defined to evaluate the existence of
sufficient facilities within the system to satisfy load demand as well as system operational constraints.
Monte Carlo simulation has been used in adequacy assessment of complex electric power systems. Its primary
disadvantage is the lengthy calculation time to achieve satisfactory statistical convergence of adequacy index values. One of the time-consuming procedures in MCS is
the characterization of the sampled states. For each sam-

16th PSCC, Glasgow, Scotland, July 14-18, 2008

pled state, determination needs to be made whether it is


a loss of load state or not. This typically needs a power
flow calculation to be made. Depending on the method of
flow calculation, this step could impose a significant computational burden. Additionally, because of the random
sampling, many similar states may be sampled and their
characteristics determined repeatedly. These two factors
make Monte Carlo simulation time-consuming. To render the MCS more computationally efficient, some pattern classifiers can be combined with the straight MCS. In
such hybrid algorithms, first the classifier is trained, and
then in adequacy evaluation the trained classifier is used
to characterize system states instead of using computationally expensive power flow analyses. In this way, the
computational efficiency of MCS can be significantly improved and it has shown some prospects [1, 2, 4]. Since
the dominant failure states only occupy a very small proportion of the entire state space, the whole data set for
adequacy evaluation is highly imbalanced. This imbalance adds much difficulty to the pattern recognition and
classification task. Many classifiers have been shown to
be sensitive to the class imbalances in the data set, including decision tree learning, support vector machines, neural
networks, and so forth [5]. Very recently, a new metaheuristic algorithm termed artificial immune recognition
system (AIRS) has demonstrated some superiority in serving as a pattern classifier for highly imbalanced data sets
[6]. This paper proposes a novel method for power system
adequacy evaluation by combining MCS and AIRS, which
is promising in reducing the computing burden of the adequacy indices such as loss of load probability (LOLP) as
compared to the straight MCS. A case study based on the
IEEE Reliability Test System (RTS) is presented demonstrating the efficiency of the proposed method. Furthermore, comparative studies are also carried out with respect
to other hybrid MCS algorithms using Artificial Neural
Networks (ANN) based classification algorithms including the standard backpropagation neural network (BPNN)
and Self-Organizing Map (SOM).
The remainder of the paper is organized as follows.
Section 2 presents the LOLP identification and adequacy
evaluation based on Monte Carlo simulation. Section 3 introduces the inner working of AIRS. The proposed MCSAIRS method is detailed in Section 4. Section 5 discusses
the numerical study of adequacy evaluation problem for an
IEEE RTS using the proposed method, where simulation
results and analysis are presented. Finally, conclusions are

Page 1

drawn and future research is suggested.


2

Loss of Load Identification and MCS-based


Adequacy Assessment

Various network flow methods have been used to determine whether a system state constitutes loss-of-load or
not. Typically for multi-area reliability studies, linearized
network flow model or DC load flow are used. In the
composite system reliability studies, either DC or AC load
flow can be used for flow calculations.
2.1

Loss of Load Identification

In this investigation, optimal power flow (OPF) is used


to minimize load shedding at each bus subject to a set of
constraints [7]. The result of the objective function minimization can then be used to determine if a system state is
loss-of-load or not. The problem can be stated as follows.
Minimize:
N
X
LCU Ti
(2.1)

If the objective function in the above formulation is


zero, the power system state does not result in loss-ofload. If the objective function is greater than zero, then the
system state does constitute loss-of-load. Furthermore, the
formulation above not only characterizes a system state as
loss-of-load or not, but also gives the minimum amount of
load which must be shed at each bus if the system state
results in loss-of-load. This information can also be used
to determine EUE (Expected Unserved Energy) index.
2.2

Several reliability indices have been proposed in the


literature. Perhaps loss of load probability (LOLP) is the
most widely known power system reliability index. This
section gives various steps of Monte Carlo simulation for
the LOLP index but the same process can be used for other
indices. Typical steps for LOLP calculation using simple
Monte Carlo simulation can be described as follows.
Step 1: Select the seed for the random number generator. Set the maximum iteration number and let
the initial iteration number K = 1.

i=1

Subject to:
Pgi (Pli LCU Ti )

jN

Step 2: Sample the system state randomly (load


level, generation status and line status) and perform
a flow calculation to classify it as loss-of-load or
otherwise.

1, loss-of-load state;
Xi =
(2.9)
0, otherwise.

Vi Vj Yij cos(ij + j i ) = 0
(2.2)

Qgi (Qli LCU Ti

jN

Qli
)
Pli

Vi Vj Yij sin(ij + j i ) = 0

(2.3)

min
max
Pgi
Pgi Pgi

(2.4)

Qmin
Qgi Qmax
gi
gi

(2.5)

||Sij ||

max
||Sij
||

Step 3: After a certain number of samples, calculate LOLP, variance of the estimated LOLP and the
coefficient of variation.
K
X
\ = 1
Xi
LOLP
K i=1

(2.6)

Vimin Vi Vimax

(2.7)

0 LCU Ti Pli

(2.8)

where
LCU Ti : The amount of real power load curtailed at bus i;
Yij , ij : Magnitude and angle of ij-th element of the admittance matrix Y ;
Vi , i : Magnitude and angle of voltage at bus i;
Pgi , Qgi : Real power and reactive power output of generation at bus i;
Pli , Qli : Real power and reactive power load at bus i;
||Sij ||: Apparent power flow in transmission line between
bus i and j;
min
max
Pgi
, Pgi
: Minimum and maximum real power output
of generator i;
max
Qmin
gi , Qgi : Minimum and maximum reactive power
output of generator i;
Vimin , Vimax : Minimum and maximum of voltage magnitude at bus i;
max
||Sij
||: Maximum allowed apparent power flow in
transmission line between bus i and j;
N : Number of buses.

16th PSCC, Glasgow, Scotland, July 14-18, 2008

Monte Carlo Simulation

(2.10)

2
\ ) = 1 (LOLP
\ LOLP
\ )
V (LOLP
K
q
\)
V (LOLP
=
\
LOLP
where K is the total number of samples.

(2.11)

(2.12)

Step 4: Check whether the coefficient of variation


is less than a specified threshold . If < or
K > Kmax , stop; otherwise, K = K + 1, go to
step 2.
3

Artificial Immune Recognition System (AIRS)

Analogous to the relationship between artificial neural networks (ANN) and biological neural systems (BNS),
artificial immune system (AIS) is the counterpart of biological immune systems (BIS) in the artificial intelligence
domain through mathematical abstraction and modeling.
The immune system in a mammalian body comprises a
set of cells and molecules that guard body against the attacks of various pathogens via complex immunological response mechanisms. AIRS is a BIS-inspired algorithm for

Page 2

supervised pattern classification [8-10]. The objective of


AIRS training in this problem is to evolve a set of Artificial Recognition Balls (ARBs) capable of recognizing
different transformer states. The AIRS learning algorithm
primarily includes the following steps [8, 9]:
Initialization: This can be deemed a data preprocessing phase coupled with a parameter discovery phase. During initialization, first all the input
data are normalized in order to match the affinity
metric defined for any antigen-antibody pair. If the
Euclidean distance is used to measure the affinity
between any two data instances, the data samples
are normalized to guarantee that the reactions between any antigen-antibody pair falls in the range of
[0, 1]. The affinity threshold is the average affinity
value over all training data, which can be calculated
by
Pn Pn
i=1
j=i+1 affinity(agi , agj )
Affinity threshold =
n(n1)
2

(3.13)
where n is the number of training antigens, agi
and agj are the ith and jth training antigens, and
affinity(x, y) returns the Euclidean distance between the two antigens feature vectors. The seeding of initial memory cells and initial ARB population is also accomplished in this step.
Memory cell identification and ARB generation:
Artificial Recognition Ball (ARB) corresponds to
the B-Cell in natural biological systems. An initial set of ARBs and memory cells is randomly generated from the training data. After initialization,
training process is started as a one-shot incremental
learning procedure. Each item in the training data
set is fed to the AIRS for once. The first step is
to locate the memory cell for a given training sample. Each time a new training sample is presented
to the AIRS, the best matching cell from the whole
memory cell population of antigens class is identified, which produces offspring via clonal expansion
operations. Clonal expansion produces the clones
of an antibody proportional to its affinity with respect to the presented antigen. Certain clones undertake mutation by altering some gene information
producing antibody variants. Following this, each
ARB from the refreshed set of ARBs is presented
to the surviving antigen so as to examine its affinity
with the antigen. Then certain amount of system resources proportional to its affinity are assigned to
the ARB appropriately. Given a specific training
antigen, ag, find the memory cell, mcmatch , that
has the following property:
mcmatch = argmaxstimulation(ag, mc),
mc M Cag.c (3.14)
where M C represents the set of memory cells
and mc means an individual member of this set.

16th PSCC, Glasgow, Scotland, July 14-18, 2008

ag.c represents the class of a given antigen, ag,


where ag.c C = {1, 2, . . . , nc} and nc is the
number of classes in the data set. The function
stimulation(x, y) is defined as follows:
stimulation(x, y) = 1 affinity(x, y)

(3.15)

Once mcmatch is identified, this memory cell is


used to create new ARBs, which are injected to the
population of (existing) ARBs.
Competition for resources and development of a
candidate memory cell: Thus far, a set of ARBs is
developed including the matching memory cells and
their mutants, also coupled with surviving ARBs
from other previously presented antigens. When the
total resources assigned to the entire ARB population exceed the permissible threshold, the weakest
ARBs are dumped until the total assigned resources
return below the allowable limit. The survived
ARBs further produce offspring through clonal expansion and mutation until the average affinity value
for all the existing ARBs with respect to the antigen
reaches the pre-specified threshold or satisfy any
other termination criteria for one-antigen training.
The ARB population AB comprises the ARBs able
to compete for resources. The algorithm starts by
determining if the ARBs in AB were sufficiently
stimulated to terminate the training. For this purpose, a vector ~s in length nc is defined containing
the average stimulation value for each class subset
of AB
P|ABi |
j=1 abj .stim
, abj ABi
(3.16)
si =
|ABi |
where ab.stim represents the stimulation level of
the ARB ab. The stopping criterion is satisfied provided that for all elements in ~s = {s1 , s2 , . . . , snc },
the relationship si stimulationthreshold holds.
Memory cell introduction: The final step in the
training phase is to add the just-developed candidate memory cells into the pool of existing memory
cell. After the training process for current antigen
is completed, the best matching ARB with the same
class as the antigen is chosen as the candidate memory cell. It is turned into a long-lived memory cell if
it matches the antigen better than the original memory cell. From a machine learning perspective, this
process offers the data reduction capabilities of the
AIRS. The algorithm also generalizes because the
developed memory cells in the pool are not necessarily the same as training instances. The procedure does not stop until all the training antigens
have been presented.
Classification: After training is completed, the
evolved memory cells can be used for classification,

Page 3

Step 4: Assign a certain amount of resources to each


ARB according to its affinity.

where the class of each presented test antigen is decided by this set of class prototypes through the knearest neighbor (KNN) scheme. Each memory cell
is iteratively presented with each test instance for
stimulation. The classification of a test instance is
decided based on a majority vote of outputs from
the k most stimulated memory cells.
4

Step 5: Dump the weakest ARBs until the level of


resources returns to a pre-specified threshold.
Step 6: If the average affinity of the surviving ARBs
exceeds a certain limit, proceed to step 7. Otherwise, clone and mutate these surviving ARBs depending on their affinity and return to step 3.

Adequacy Evaluation Through Hybridization of


MCS and AIRS

Step 7: Select the best ARB as a candidate memory


cell (mccand ).

The objective of this paper is to show how Monte


Carlo Simulation (MCS) can be made more efficient using AIRS. The input training features for AIRS, the major
steps in AIRS training, and the approach to marriage between AIRS and MCS will be discussed.
4.1

Step 8: If the affinity of mccand for the training instance is better than the affinity of mcmatch , then
inject mccand to the pool of memory cells. Furthermore, if the affinity between mccand and mcmatch
falls within a certain range, then delete mcmatch
from the memory cell pool.

Input Training Features

For the problem of loss-of-load state identification, a


power system state can be characterized by load conditions, network topology and availability status of generators. The input vector corresponding to a system state is:
Xi = [Pi1 , Qi1 , . . . , Pin , Qin , P Gi1 , . . . , P Gim ]
(4.17)
where Pik is the real power load of bus k for state i, Qik
is the reactive power load of bus k for state i, P Gim is
the available real power generation of bus m for state i,
n is the number of load buses, and m is the number of
generation buses.
4.2

Preprocessing of Training Data

When the Euclidean distance between two training


system states is very small, one should be eliminated or replaced since they exhibit a high degree of similarity. This
kind of redundancy may decrease the training efficiency
and slow down the algorithm convergence since the diversity of training antigens is reduced. In this study, the
redundant antigens (i.e., system states) are eliminated
directly and replaced by more distinct ones.
4.3

Training Procedure

A set of ARBs is intended to be evolved during AIRS


training, which will be used to identify various system
states either encountered in the training process or not.
The major computational procedure of AIRS (i.e., the evolution of a set of memory cells) [8] for the classification
problem is laid out in the following.
Step 1: Compare a training system state with all
memory cells of the identical class and seek out the
memory cell with the best affinity for the training
system state. This memory cell is referred to as
mcmatch .
Step 2: Clone and mutate mcmatch proportional to
its affinity to produce a pool of primitive ARBs.
Step 3: Compute the affinity of each ARB with respect to the training system state.

16th PSCC, Glasgow, Scotland, July 14-18, 2008

Step 9: Repeat from step 1 until all training instances have been presented.
After this training procedure is accomplished, AIRS classifies system states based on the k-nearest neighbor voting
mechanism with the evolved set of memory cells.
4.4

Hybridization of MCS and AIRS

The proposed method can efficiently determine the


load loss characteristic of the sampled state. Here the
AIRS is trained to recognize the loss-of-load states. Once
this training is complete, the AIRS is used along with the
Monte Carlo simulation to estimate the reliability of the
system. In this version of MCS-AIRS the state evaluation
is accomplished by the trained AIRS rather than OPF calculation. Overall procedure of the proposed method consists of the following steps.
Step 1: Prepare the training patterns (i.e., input vectors) for AIRS. Training patterns are obtained by
OPF calculations which characterize each training
pattern as loss of load or otherwise. The purification processing of training data is conducted at this
stage.
Step 2: Carry out AIRS training with the prepared
training patterns. A set of memory cells representing loss-of-load or no-loss-of-load states is evolved
at the end of training.
Step 3: After the AIRS is trained, Monte Carlo simulation follows the same procedure as in section 2.2
except that state classification is performed by AIRS
instead of OPF. The class (loss-of-load or no-lossof-load) of each sampled state (i.e., antigen) is determined by the voting of the k most stimulated antibodies (i.e., ARBs).
We can see from the above procedure that the computational efficiency can be improved by carrying out much
fewer power flow calculations. Especially for very complex systems requiring computationally expensive power

Page 4

flow analyses, the advantage of this method will become


more evident in relation to the straight MCS.
5

Simulations and Evaluation

The IEEE Reliability Test System (RTS) [11] was chosen to test the proposed method. It has 24 buses (10 generation buses and 17 load buses), 38 lines and 32 units.
The system annual peak load is 2850 MW and the total installed generating capacity is 3405 MW. Two sets of studies are to be performed, one considering only the peak load
and the other with multiple load levels. The coefficient of
variation in all MCS simulations is set 0.02.
5.1

Peak Load Level

First, reliability analysis is carried out at the peak load


level.
5.1.1

Input selection

The total load is fixed at the peak load of 2850 MW.


The input features for the AIRS consist of the generating
unit statuses. There are 32 units distributed at 10 buses, so
the input vector is:
X = [P G1 , P G2 , . . . , P G9 , P G10 ]
5.1.2

(5.18)

AIRS training

A total of 600 different training patterns were used to


train the AIRS classifier. These training patterns were selected in such a fashion that they were non-repetitive but
most samples were from the high probability region of the
state space. This was achieved by varying the availability
status of units through a preliminary Monte Carlo experiment and evaluation by OPF.
5.1.3

Monte Carlo simulation

After the AIRS classifier was trained, Monte Carlo


simulation was performed to estimate the loss of load
probability (LOLP). For each sampled system state, AIRS,
instead of OPF, was used to characterize it as loss-of-load
or not. The most stimulated ARBs to each sampled system
state determine its load-loss status. Fifteen thousand states
were sampled in the simulation. There were 13,153 noloss-of-load states and 1847 loss-of-load states characterized by AIRS. Thus the estimated LOLP is 0.1231. Monte
Carlo simulation combined with OPF was performed to
obtain the benchmark value of LOLP at peak load level
for use in assessing the accuracy of AIRS. For the 15,000
system states sampled above, there were 13,168 no-lossof-load states and 1832 loss-of-load states characterized
by OPF. The computed benchmark value of LOLP is thus
0.1221. Among the 13,168 no-loss-of-load states classified by OPF, 13,118 states were classified as no-loss-ofload correctly by AIRS, resulting in a classification accuracy of 99.62%. Among the 1832 loss-of-load states, 1797
states were classified as loss-of-load correctly by AIRS,
giving a classification accuracy of 98.11%.

16th PSCC, Glasgow, Scotland, July 14-18, 2008

5.1.4

Computing time

It required 4.4 seconds for the AIRS training. For the


characterization of all the 15,000 sampled states, the computing time was 2.4 seconds. Compared to the straight
Monte Carlo simulation, which needs to perform 15,000
OPFs, the computing time is considerably reduced. The
program was implemented in C language.
5.2

Multiple Load Levels

In the following, reliability analysis is carried out with


multiple load levels.
5.2.1

Input selection

The load profile for IEEE RTS is represented by 15


load levels [12]. In this study system, loads are distributed
at 17 buses and units are distributed at 10 buses, so the
input features for the AIRS consist of load levels and generating unit statuses as follows.
X = [P1 , Q1 , . . . , P17 , Q17 , P G1 , . . . , P G10 ]
5.2.2

(5.19)

AIRS training

Similar to the procedure in selecting training patterns


for the peak-load-level case, a total of 1400 different training patterns were generated by varying the load levels and
the availability status of generators. They were evaluated
by OPF and used to train the AIRS. The AIRS training parameters are set as in [13]. There were 128 ARBs evolved
as loss-of-load or no-loss-of-load states.
5.2.3

Monte Carlo simulation

Monte Carlo simulation was performed after the memory cells were established. AIRS was used to characterize each sampled state as loss-of-load or not. Twenty-five
thousand states were sampled in the simulation. There
were 24,840 no-loss-of-load states and 160 loss-of-load
states as characterized by AIRS. Thus the estimated LOLP
is 0.0064. Monte Carlo simulation combined with OPF
was performed to obtain the benchmark value of LOLP for
multiple load levels. For the 25,000 states sampled above,
there were 24,912 no-loss-of-load states and 88 loss-ofload states as characterized by OPF. Therefore the computed benchmark value of LOLP is 0.0035. Among the
24,912 no-loss-of-load states classified by OPF, 24,837
states were classified correctly by AIRS, resulting in a
classification accuracy of 99.70%. Among the 88 lossof-load states, 85 states were classified correctly by AIRS,
giving a classification accuracy of 97.12%.
5.2.4

Computing time

It required 6.6 seconds for the AIRS training. For the


classification of the total 25,000 sampled states, the computing time was 11.7 seconds.
5.3

Comparative Study

For comparison, two representative neural-network


classifiers including standard backpropagation neural network (BPNN) [14] and self-organizing map (SOM) [2, 15]

Page 5

are used to hybridize the MCS resulting in MCS-BPNN


and MCS-SOM algorithms. The basic procedure is not
changed except that the AIRS is replaced by the corresponding ANN classifier at both training and simulation
stages. For instance, the overall procedure of MCS-SOM
algorithm consists of the following steps.
Step 1: Prepare the training patterns for SOM.
Training patterns are obtained by OPF calculations
which characterize each training pattern as loss of
load or otherwise.
Step 2: Carry out SOM training with the prepared
training patterns.
Step 3: Label a neuron in the map as loss-of-load or
no-loss-of-load according to the majority label voting of the training patterns mapped to that neuron.
Step 4: After the SOM network is trained, Monte
Carlo simulation follows the same procedure as in
section 2.2 except that state classification is performed by SOM instead of OPF. The class (lossof-load or no-loss-of-load) of each sampled state is
determined by the label of the nearest neuron in the
map.
The overall classification accuracy is oftentimes used
to measure the performance of a classification task. However, for the classification task of imbalanced data, this criterion may not be sufficiently informative and sometimes
even be misleading. For instance, in a two-class classification problem with highly skewed class distribution, most
classifiers usually exhibit a low error rate for the majority
class and a high error rate for the minority class. Assume
for example that the majority class occupies 95% of the
data and the minority class contains the remaining 5%.
An overall classification accuracy of 95% can be achieved
even if all the data are classified into the majority category. The error rates for the majority and minority classes
are 5% and 100%, respectively. Thus, a more objective
and effective index should be defined to measure the classification performance of different classifiers in handling
imbalanced data sets.
The concept of g-mean is proposed in [15] for this purpose, where the classification performance on imbalanced
data sets is measured based on the confusion matrix. Table 1 shows the confusion matrix where the failed system
states are considered as positive class and success system
states as negative class.
Actual negative
Actual positive
class
class
Predicted True Negative (TN) False Negative (FN)
negative
class
Predicted False Positive (FP)
True Positive (TP)
positive
class
Table 1: Confusion Matrix.

True positive rate (TPR) indicates the classification


accuracy of failed system states, and the true negative

16th PSCC, Glasgow, Scotland, July 14-18, 2008

rate (TNR) presents the classification accuracy of success system states. They are used to define the new index
g mean:
TP
TPR =
(5.20)
TP + FN
TN
TN + FP

g mean = T P R T N R.
TNR =

(5.21)
(5.22)

The index g-mean examines the classification accuracies on both positive and negative classes. The g-mean
is high only if both TPR and TNR are large and the difference between them is small. For instance, the g-mean
value of the classifier in the previous example is zero, although the overall accuracy is as high as 95%. The results
for comparison in different load levels using straight MCS,
MCS-BPNN, MCS-SOM, and MCS-AIRS are given in
Table 2 and Table 3, which show the mean values over
40 runs. We can find that in all evaluation scenarios, the
straight MCS exhibits the much poorer computational efficiency than the hybrid ones, and between them MCSAIRS is the most computationally efficient. MCS-BPNN
derives solutions of the relatively lowest quality among
these three hybrid algorithms. The solution quality of
MCS-AIRS is also somewhat higher than the MCS-SOM
in terms of different measures including TPR, TNR, and
g-mean.
Methods TPR (%) TNR (%) g-mean Time (s)
MCS
N/A
N/A
N/A
128.8
MCS
89.65
99.14
0.9428
9.2
-BPNN
MCS
96.77
99.39
0.9807
4.3
-SOM
MCS
98.11
99.62
0.9886
2.4
-AIRS
Table 2: Comparative Results for State Classification (Peak Load Level).

Methods
MCS
MCS-BPNN
MCS-SOM
MCS-AIRS

TPR
N/A
86.98
93.39
97.12

TNR
N/A
99.16
99.60
99.70

g-mean
N/A
0.9287
0.9645
0.9683

Time
189.3
28.5
16.9
11.7

Table 3: Comparative Results for State Classification (Multiple Load


Levels).

Two-sample tests of hypothesis are performed using


the 40-run experimental data to compare the g-means of
any two approaches. In this study, the null hypothesis
H0gmean : gmeanM CSAIRS = gmeanM CSAN N
(5.23)
is tested against the alternative hypothesis
H1gmean : gmeanM CSAIRS > gmeanM CSAN N .
(5.24)
where MCS-ANN refers to either MCS-BPNN or MCSSOM. The P-values (Probability values) of the tests can
be used to determine if MCS-AIRS achieves better performance than MCS-ANN. The P-value measures the
strength of evidence against the null hypothesis. Since it
indicates the probability of achieving the existing sample

Page 6

data under the null hypothesis [6], a small P-value results


in the rejection of the null hypothesis. Here the significance level is set 0.05, which means that a P-value under
0.05 will reject the null hypothesis in favor of the alternative hypothesis. The P-value of two-sample tests on gmeans achieved by MCS-AIRS and MCS-BPNN is less
than 0.0001, and that of MCS-AIRS and MCS-SOM is
about 0.01. This indicates that MCS-AIRS increases the
average g-mean values for classification of failed states
in relation to both MCS-BPNN and MCS-SOM. At this
point, we can conclude that the proposed method exhibits
a comparable or better performance with respect to the
MCS-ANN method when measured by multiple criteria.
Since BPNN aims to minimize the overall error rate, it
tends to prioritize the class with more data samples in the
imbalanced data set. This bias will sacrifice the accuracy
of minority-class classification. On the contrary, in AIRS
the matching antibodies are evolved for all training data,
so the biased classification performance does not happen.
6

Concluding Remarks

In electric power systems, the number of generating


units and other components is very large. Inevitably, adequacy assessment of power systems becomes more challenging due to their larger scale and increasing complexity. Thus, in adequacy assessment, exhaustive enumeration is usually impractical due to innumerable system
states incurred. The overall objective of classifier learning is to derive a model based on certain training data,
which is intended to make as few errors as possible when
classifying test data not encountered previously. Artificial
Immune Recognition System (AIRS) has been shown to
be an outstanding machine learner in classifying imbalanced data sets. A hybrid algorithm termed MCS-AIRS
is proposed in this paper for adequacy evaluation, whose
objective is to render Monte Carlo simulation more efficient by utilizing the characteristics of AIRS. MCS-AIRS
here serves as a binary classifier, which maps various system states into two classes. Cases studies are presented
for adequacy evaluation based on an IEEE RTS. we found
that the training time for the AIRS is short, while at the
simulation stage OPFs required in straight Monte Carlo
simulation are not needed anymore.
REFERENCES

[3] N. Samaan and C. Singh, Adequacy assessment of


power system generation using a modified simple genetic algorithm, IEEE Transactions on Power Systems, Vol. 17, No. 4, pp 974981, 2002.
[4] H. Kim and C. Singh, Power system probabilistic
security assessment using Bayes classifier, Electric
Power System Research, Vol. 74, pp 157165, 2005.
[5] V. Hoste, Optimization Issues in Machine Learning
of Coreference Resolution, Ph.D Thesis, University
of Antwerp, Belgium, 2005.
[6] L. Xu, M.-Y. Chow, J. Timmis, and L. S. Taylor,
Power distribution outage cause identification with
imbalanced data using artificial immune recognition
system (AIRS) algorithm, IEEE Transactions on
Power Systems, Vol. 22, No. 1, pp 198204, 2007.
[7] X. Luo, C. Singh, and A.D. Patton, Loss-of-load
state identification using self-organizing map, Proceedings of IEEE PES 1999 Summer Meeting, Vol.2,
pp 670-675, 1999.
[8] A. Watkins, J. Timmis, and L. Boggess, Artificial
immune recognition system (AIRS): An immune inspired supervised machine learning algorithm, Genetic Programming and Evolvable Machines, 5, pp
291317, 2004.
[9] A. Watkins and J. Timmis, Exploiting parallelism
inherent in AIRS, an artificial immune classifier, G.
Nicosia et al. (Eds.): Proceedings of ICARIS 2004,
LNCS 3239, pp 427438, 2004.
[10] A. Watkins and J. Timmis, Artificial immune recognition system (AIRS): revisions and refinements, In
Proceedings of 1st International Conference on Artificial Immune Systems (ICARIS), University of Kent
at Canterbury, September 9-11, pp 173181, 2002.
[11] IEEE Committee Report, IEEE reliability test system, IEEE Transactions on Power Apparatus and
Systems, Vol. PAS-98, No. 6, pp 20472054, 1979.
[12] G. Hamound, Probabilistic assessment of interconnection assistance between power systems, IEEE
Transaction on Power System, Vol.13, No.2, pp 535
542, 1998.

[1] X. Luo, C. Singh, and A. D. Patton, Power system


reliability evaluation using learning vector quantization and Monte Carlo simulation, Electric Power
Systems Research, Vol. 66, pp 163169, 2003.

[13] A. Watkins, Artificial Immune Recognition System


(AIRS) C++ Source Code, 2002.

[2] C. Singh, X. Luo, and H. Kim, Power system adequacy and security calculations using Monte Carlo
Simulation incorporating intelligent system methodology, Proceedings of the 9th International Conference on Probabilistic Methods Applied to Power
Systems, Stockholm, Sweden, 2006.

[15] T. Kohonen, Self-Organizing Maps, SpringerVerlag, Heidelberg, 1995.

16th PSCC, Glasgow, Scotland, July 14-18, 2008

[14] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern


Classification, Wiley, 2001.

[16] M. Kubat, R. Holte, and S. Matwin, Machine learning for the detection of oil spills in radar images,
Machine Learning, Vol. 30, pp 195-215, 1998.

Page 7

S-ar putea să vă placă și