Energy Conversion and Management 47 (2006) 32043215 www.elsevier.

com/locate/enconman

Power factor correction technique based on articial neural networks

S. Sagiroglu a, I. Colak b, R. Bayindir
a b

b,*

Department of Computer Engineering, Faculty of Engineering and Architecture, Celal Bayar Bulvari, Gazi University, Maltepe, Ankara 06570, Turkey Department of Electrical Education, Faculty of Technical Education, Gazi University, Besevler, Ankara 06500, Turkey Received 3 August 2005; accepted 15 February 2006 Available online 29 March 2006

Abstract This paper presents a novel technique based on articial neural networks (ANNs) to correct the line power factor with variable loads. A synchronous motor controlled by the neural compensator was used to handle the reactive power of the system. The ANN compensator was trained with the extended delta-bar-delta learning algorithm. The parameters of the ANN were then inserted into a PIC 16F877 controller to get a better and faster compensation. The results have shown that the proposed novel technique developed in this work overcomes the problems occurring in conventional compensators including over or under compensation, time delay and step changes of reactive power and provides accurate, low cost and fast compensation compared to the technique with capacitor groups. 2006 Elsevier Ltd. All rights reserved.
Keywords: Articial neural network; Power factor correction; Synchronous motor

1. Introduction In electrical systems, all inductive loads fed by alternating current draw active and reactive powers from the line. While the active power is converted into heat, light and mechanical energy or other types of energy, the reactive power cannot be converted. It causes the transformer, alternator, cable, protection relay and other equipment to be larger than their rated values. Therefore, reducing the capacities of production, transmission and distribution of the line is the result of the eects of lower power factor [1,2]. In order to get rid of this eect, the power factor needs to be corrected [3]. In practical applications, reactive power compensations have generally been achieved by employing constant capacitor groups using some relays, timers and contactors. These types of systems are known as classical methods and have some mechanical problems, slow responses, over or under compensation and harmonics
*

Corresponding author. Tel.: +90 312 212 6820/1221. E-mail address: bayindir@gazi.edu.tr (R. Bayindir).

0196-8904/$ - see front matter 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.enconman.2006.02.018

S. Sagiroglu et al. / Energy Conversion and Management 47 (2006) 32043215

3205

in the line voltage due to step changes of the capacitor groups that occur during operation. The changes of reactive power produced by variation of the load or load switching on the line can cause adverse eects on voltage stability and system security [46]. Nevertheless, static volt-ampere reactive (var) compensators (SVCs) can give smooth reactive power compensation without step changes. SVCs can be classied in four categories [712]: thyristor controlled reactor (TCR), thyristor controlled capacitor (TSC), combinations of TCR and TSC with switched or xed capacitor and advanced static var compensation (STATCOM). These are utilized to enhance the integrated voltage stability. TCRs (or TSCs) for reactive power compensation systems are faster and do not have any mechanical problems. However, they generate harmonics in the voltage and current and cause a stability problem in the system [1318]. These problems can be overcome by using a synchronous motor (SM), if it is already available in the system [19]. Synchronous motors can operate at unity, lagging or leading power factor condition [20,21]. When the motor is operating at unity condition, its power factor is equal to 1.0, and it only draws active power from the line to compensate its mechanical losses. When it is operating at lagging condition, its power factor is less than 1.0, and it draws lagging reactive power from the line. When it is operating at leading condition, its power factor is again less than 1.0, but this time, it produces leading reactive power for the line. Since the synchronous motor is operating at an over or under compensation condition, the transmission line can be over loaded by the reactive powers produced or drawn by the synchronous motor. The use of a synchronous motor as a reactive power compensator is a well known method, but if it is only used for reactive power compensation, the system will be very inecient and expensive in comparison with a group of capacitors [1,5]. Synchronous motor operation is important to reduce the cost and var penalties, to improve the voltage stability of the system, it can be operated by exciting its eld circuit either using a xed DC supply for constant loads or a variable direct current (DC) supply for variable loads. The variable DC supply can be achieved either manually or automatically. Manual control requires a serial rheostat in the eld circuit, which causes a step change of the eld current, electrical arcs on the rheostat, over or under compensation and time delay for the compensation. To get rid of these problems, the eld voltage of the SM has to be adjusted using different automatic control techniques. Proportional plus integral (PI), proportional plus integral plus derivative (PID), pulse width modulation (PWM) and fuzzy logic (FL) techniques have been used for improving the compensation [4,22,23]. If one of these automatic controllers is not used in operation, the system might be operating under the eects of potential pole slip, increasing the kVA loading on the plant transformer and reducing the system voltage. So, using an appropriate controller, the system voltage stability of the bus might be achieved or the requirement of the kVA loading for the plant transformer is decreased [5]. Articial neural networks (ANNs) have been very popular for applying in many engineering elds because of their fascinating features, such as learning, generalization, faster computation and ease of implementation. ANNs have been recently applied for power system security, power system stability estimation and optimal SVC and controlling induction, direct current and synchronous motors [2431]. This paper introduces a novel technique based on ANNs to correct the power factor using a dynamic reactive power compensator. The experimental data used to design, train and test the neural controller were achieved from the test rig. Using an ANN compensator totally removes the problems mentioned earlier. Moreover, the exact eld current required by the load can be produced without any delay compared to the other methods presented in the literature. 2. Power factor correction The electrical power produced by an alternator is transmitted, distributed and then used by loads. On a power line, besides the active power, reactive power must also be available for inductive loads. An alternator in the power station can produce the reactive power for the line, but the reactive power also can be supplied from any source, which can either be a synchronous motor or capacitor groups connected near the load. The source of the reactive power must be very close to the load for ecient operation of the system. If the reactive power of any load is supplied from a synchronous motor or a group of capacitors rather than the power line, this system is called a reactive power compensator [32]. So, the power factor of the system can be kept at a required value.

3206

S. Sagiroglu et al. / Energy Conversion and Management 47 (2006) 32043215

The power factor is denoted as pf or cos u. It is the ratio of the active power (P) to the apparent power (S), which can also be calculated from Fig. 1: pf Active Power P . Apparent Power S 1

The typical pf for alternating current (AC) electrical machines and any other inductive loads are between 0.35 and 0.85. Reactive power for these machines must be available to produce the magnetic eld inside the machines for their operation. If the reactive power is compensated somehow, the capacity of the power line is increased. A commonly used reactive power supply is a group of capacitors. On the other hand, synchronous motors can operate at leading power factor conditions using a dynamic compensator, and they have been very reliable for reactive power compensation and do not produce any harmonics in the system [4,33]. The reactive power drawn from the line by an inductive load lags the active power by 90. If a capacitor is connected to the system, it also draws reactive power, but it leads the active power. The direction of the capacitive reactive power (Qcap) is opposite the direction of the inductive reactive power (Qind) as shown in the graphical representation of the power factor correction in Fig. 1. The eective inductive reactive power drawn by the circuit will be reduced by the capacitive reactive power, resulting in a reduction of the apparent power from S to S1. The phase angle between the active power, P, and the new apparent power, S1, also is reduced from u to u1. Thus, the power factor increases from cos u to cos u1. So, the new pf is New pf cos u1 P . S1 2

By selecting a suitable capacitor value, the power factor can be compensated nearly to 1.0. However, in practice, the power factor is improved to fall between 0.90 and 0.95. Since the power factors of consumer type loads are very low, the average power factor of the line becomes lower. The power factor of a running loaded or unloaded synchronous motor can be at unity, lagging or leading conditions by changing its excitation current under constant load and terminal voltage. If the excitation current is reduced while the motor is running at unity condition, the power factor shifts to the lagging condition. If the excitation current is increased, and the motor is running at unity condition, the power factor shifts to the leading condition. Furthermore, the excitation current can also change the reactive component of the apparent power, but the active component stays constant. For an ideal compensation, the features of low cost, fast computation and more accuracy are desired for building a new compensation system.

Q cap

NewVAr (after pf correction)

90

P
90
1

Old VAr (before pf correction)

S1 Q cap

Q ind
Fig. 1. Reactive power compensation with capacitors.

S. Sagiroglu et al. / Energy Conversion and Management 47 (2006) 32043215

3207

3. Articial neural networks Articial neural networks (ANNs) are biologically inspired intelligent techniques. ANNs are generally made of a number of simple and highly interconnected processing elements organized in layers as shown in Fig. 2. These processing elements or neurons process information by its dynamic state response to external inputs. ANNs are capable of learning patterns by being trained with a number of known patterns. The learning process automatically adjusts the weights and thresholds of the processing elements. Once adjusted with minimal dierences between the ANN output and the targeted output, the neural network is said to be trained. Articial neural networks have many structures and architectures [34,35]. Multi-layered perceptrons (MLPs) are the simplest and, therefore, most commonly used neural network architectures [35]. Fig. 2 shows an MLP with three layers: an input layer, an output layer and an intermediate or hidden layer. Neurons are represented with circles. Neurons in the input layer only act as buers for distributing the input signals xi to neurons in the hidden layer. Each neuron j in the hidden layer sums up its input signals xi after weighting them with the strengths of the respective connections wji from the input layer and computes its output yj as a function f of the sum: X  yj f wji xi ; 3 f can be a sigmoidal or a hyperbolic tangent function. The output of the neurons in the output layer is computed similarly. A number of learning algorithms were used to adjust the weights of the ANNs. The extended delta-bar-delta (EDBD) learning algorithm used to train the neural architecture is introduced below. This algorithm is an extension of the delta-bar-delta (DBD) algorithm and is based on decreasing the training time for multi-layered perceptrons. The use of momentum heuristics and avoiding the cause of wild jumps
y1 y2

outputs Output Layer

yn

Weights

Hidden Layer

Weights

Input Layer

x1

x2
Fig. 2. A multi-layered perceptron.

inputs

xm

3208

S. Sagiroglu et al. / Energy Conversion and Management 47 (2006) 32043215

in the weights are the features of the algorithm developed by Minai and Williams [36]. The EDBD algorithm includes a little used error recovery feature that calculates the global error of the current epoch during training [36]. If the error measured during the current epoch is greater than the error of the previous epoch, then the networks weights revert back to the last set of weights that produced the lower error. In this algorithm, the changes in weights are calculated as Dwk 1 ak dk lk Dwk 4

where a(k) and l(k) are the learning and momentum coecients, respectively. Dw(k) is the previous change of the weights and d(k) is the gradient component. The weights are then found as wk 1 wk Dwk 1. The learning coecient change is given as 8 > < ja expca jdk j if dk 1dk > 0; Dak ua ak if dk 1dk < 0; > : 0 otherwise, 5

where ja is the constant learning coecient scale factor, exp is the exponential function, ua is the constant learning coecient decrement factor and ca is the constant learning coecient exponential factor. The change in momentum coecient is also written as 8 > < jl expcl jdk j if dk 1dk > 0; 7 Dlk ul lk if dk 1dk < 0; > : 0 otherwise, where jl is the constant momentum coecient scale factor, ul is the constant momentum coecient decrement factor, cl is the constant momentum coecient exponential factor and  dk is the magnitude of the weight gradient component. As can be seen from Eqs. (6) and (7), the changes of the learning and momentum coecients have separate constants controlling their increase and decrease. d(k) is used whether an increase or decrease is appropriate. Therefore, the increases in both of the coecients were modied to be exponentially decreasing functions of the magnitude of the weighted gradient components j dk j. Thus, greater increases will be applied in areas of small slope or curvature than in areas of high curvature. This is a partial solution to the jump problem. In order to take a step further to prevent wild jumps and oscillations in the weight space, ceilings are placed on the individual connection learning and momentum coecients. For this, a(k) 6 amax and l(k) 6 lmax must be satised for all connections, where amax is the upper bound on the learning coecient, and lmax is the upper bound on the momentum coecient [37]. If the error, E(k), is less than the previous minimum error, the weights are saved as the best current values. A recovery tolerance parameter k controls this phase. Specically, if the current error exceeds the minimum previous error such that E(k) > Emink, all connection weights revert to the best set of weights stored in memory. Further, both coecients are decreased to begin the recovery [37]. 4. Design and implementation of ANN controller for compensation The block diagram of the proposed compensation system used in the design and implementation is given in Fig. 3. The block diagram consists of an ANN controller (ANNC), various inductive loads, cos u meter, a synchronous motor and a eld circuit of a synchronous motor. The frame in the diagram shown with the dashed line demonstrates the novel controller presented in this work. The input variables of the proposed ANNC are the load current (IL), the power factor error (e), the change of excitation current (DIf) and the power factor of the system (cos usystem). The only output of the ANNC is the excitation current (If). The power factor error is calculated as e cos uref cos usystem . 8

S. Sagiroglu et al. / Energy Conversion and Management 47 (2006) 32043215

3209

PIC 16F877
IL

cos ref

Ifk
cos system

If
cos system

Artificial Neural Network Controller

If

(ANNC)

Field circuit of synchronous motor

I f ( k1 )

cos meter
A

If

C 0

Inductive load

Fig. 3. Block diagram of the compensation system based on ANNs.

The excitation current, If(k), of the synchronous motor can be calculated by adding the change of current, DIf, to its previous value, If(k1), as I f k I f k1 DI f . 9 The design process starts with achieving the input and output tuples of the ANNC given in Fig. 3. In order to obtain these tuples, IL, e, cos usystem, If and DIf, a test rig shown in Fig. 4 was setup. The input and output parameters used to establish the ANNC were achieved from this test rig. The test procedures followed are given below: (1) The synchronous motor was driven by an auxiliary machine. (2) AC voltage was applied to the stator windings of the synchronous motor. a. When the speed of the motor was very close to the synchronous speed, DC voltage was applied to the eld winding of the motor and synchronous operation was started. b. After starting synchronous operation, the auxiliary machine was disconnected from the synchronous motor. (3) As the synchronous motor was running, its eld current was adjusted to its minimum value by a rheostat connected in series to the eld circuit. a. At this point of the operation, the motor was drawing minimum current from the supply, and its power factor was at unity. b. Choosing this point as the reference and keeping the load and applied voltage constant, the eld current was adjusted by a serial rheostat. c. When the eld current If was increased, the motor shifted from unity power factor to leading power factor operation.

3210

S. Sagiroglu et al. / Energy Conversion and Management 47 (2006) 32043215

A B C 0

IL

cos

Inductive Load

VL

+ M 3 ~ If
Fig. 4. Test rig used in achieving the input and output tuples.

V A -

(4) This test was repeated for several times at dierent loads. (5) The parameters were then measured and recorded from the test rig. The levels of the inputs and the output were between 3.0 6 IL 6 6.0, 0.65 6 cos u 6 0.95, 0.05 6 e 6 0.35, 1.318 6 If 6 2.186 and 0.138 6 DIf 6 0.946. The ANNC system only requires appropriate parameter settings of the inputs, the weights, the biases and the outputs. The input and output parameters achieved from the test rig were used to establish the compensation model based on the ANNs. According to the number of inputs and the output, the number of input neurons and output neurons were assigned. The numbers of neurons for the hidden layer were also selected after a couple of trainings. A total of 100 patterns were used in training and test. Fifty of them were used in training. The remaining 50 patterns were used in test. The training process starts with applying all the tuples (patterns) in the training set to the network. Training is stopped when the training accuracy of the network is deemed satisfactory according to some criterion (for example, when the t root mean square error between I f and If for all the training set falls below a given threshold) or the maximum allowable number of epochs is reached. Once the model parameters were obtained from a proper ANNC trained with the specied accuracy, the neural parameters achieved were then inserted into a PIC 16F877 controller for a real time implementation as shown in Fig. 3. The microcontroller has only 35 single word instructions. All are single cycle instructions except for the program branches, which are two cycles. This integrated circuit operates at 20 MHz clock frequency and runs each instruction as fast as 200 ns. Flash program memory is up to 8K 14 words. Data memory is partitioned into four banks, which contain the general-purpose registers and the special function registers. Bits RP1 and RP0 are the bank select bits. Each bank extends up to 7Fh (128 bytes). The prepared program in C++ is 1096 lines total. The program is compiled by a HITECH C compiler and then transferred into the PIC. The program was reserved for about two banks or 4 Kbytes. The required excitation current by a specic load was calculated and applied to the eld circuit of the SM very rapidly with the help of the ANNC inserted into the PIC. With this neural controller, the voltage overshoots, and over or lower compensations were eectively and eciently reduced to minimum levels. The ANNC system presented in this work only requires appropriate settings of the parameters: the numbers of inputs and outputs, the number of neurons in the hidden units, the type of activation function, the learning algorithm and the parameters of the learning algorithms. Designing a proper controller is crucial for a successful application. The conguration of the ANN utilized in this study is illustrated in Fig. 5. The neural network is a four layered one. The numbers of nodes in the input (X), in the rst hidden (H1i), in the second hidden (H2j) and the output layers (Y) are 4, 10, 5 and 1,

S. Sagiroglu et al. / Energy Conversion and Management 47 (2006) 32043215

3211

+1

Bias

H11

H12

+1

Bias

H13

H21

IL cos e
If

X1

H14
H 22

+1
X2
H15 H 23

Bias

If
Y1

X3

H16 H24

X4

H17

H18

H25

H19

H 110

Fig. 5. ANNC conguration.

respectively. The neurons in both the hidden and output layers have biases. Sigmoid functions were selected in the hidden and output layers. The inputs were scaled between 1.0 and +1.0. The output was scaled between 0.2 and 0.8. The articial neural network controller (ANNC) was trained with the EDBD learning algorithm as explained earlier. After a proper training, the ANNC parameters, weights, biases activation functions, number of neurons and the levels of the input and output values were rst achieved in an o line process with software developed in C++. The ANNC parameters achieved were nally inserted in a PIC 16F877 controller for the real time application. After this insertion, the test values obtained from the test rig in Fig. 4 were also used to test the ANNC compensator. The test results have shown that the ANNC provides satisfactory results for the correction. Using these parameters, it is possible to calculate the value of the excitation current required by the power factor of the system under a specic load. The owchart of the developed program is given in Fig. 6. The program helps us to achieve IL, DIL, cos uold, cos unew, DIf and If automatically. When a user enters the current and power factor of the system between the dened limits to increase the power factor to 0.98, the changes of the system power factor and the excitation current can be achieved and displayed on the menu. In addition, the old and new values of the power factors are also displayed on the menu as shown in Fig. 7. When the program is run, the calculated values of load current, excitation current and power factor are saved in a le. The values can be used to produce the graphics of controlled power factor as given in Fig. 8. If the test is repeated with dierent loads, the value of the power factor stays constant with

3212

S. Sagiroglu et al. / Energy Conversion and Management 47 (2006) 32043215

Fig. 6. Flow chart of present compensator.

respect to the variable load currents. So, the novel technique presented in this work provides acceptable and reliable results. The experimental and ANNC results in training are given in Fig. 8. It is clear to say that the neural model compensator follows the experimental results very closely. Fifty test samples, which were never applied to the ANNC before, were used to test the neural compensator. Fig. 9 demonstrates the performance of the ANNC in correcting the power factors for the samples applied. The corrected power factors have shown that the

S. Sagiroglu et al. / Energy Conversion and Management 47 (2006) 32043215

3213

Fig. 7. Screen shot of the program developed.

2.2

Field Current (A)

1.8

1.6

1.4

Experiments ANNC

1.2 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49

No of ANN Test Examples

Fig. 8. Verication results achieved from ANNC.

present ANNC provides satisfactorily good results. Even if dierent loads were applied to the ANNC, the power factor was kept nearly constant.

3214

S. Sagiroglu et al. / Energy Conversion and Management 47 (2006) 32043215

Fig. 9. ANNC results in test for power factor compensation.

5. Results and conclusions A novel power compensation system based on an ANNC was successfully presented in this work. The results achieved from this work have shown that the compensation system based on an ANNC performed well. As mentioned earlier, the features such as low cost, accuracy and fast computation of a compensator system are always required. The proposed ANNC achieves the compensation with these features. Using an ANNC for the power factor correction problem provides eective, ecient, reliable and robust results as seen in Fig. 9. Inserting the ANNC parameters into the PIC enables low cost and faster computation for the compensation. The approximate computation time achieved was about one microsecond. Line stability and maximum power transfer can be achieved eectively with the proposed compensation scheme. When the synchronous motor is controlled with an ANN controller, the system becomes very robust and eective. As a result, there have been no over or lower compensations as demonstrated in Fig. 9. In addition, the problems of classical methods, such as the mechanical problems, harmonics in the voltage waveform, time delays, step change of excitation voltage, pole slip, kVA loading on the plant transformer, var penalties and voltage drop on the line, have all been removed, improving the voltage stability, accuracy and eectiveness of the system. So, the novel technique presented in this work can be very attractive for industrial applications, since the synchronous motor is available in the system for any other purposes. The proposed system can also be used as a voltage regulator for variable loads. The disadvantages that come across during this work are the requirement of a synchronous motor for reactive power compensation, selecting appropriate ANNC parameters, training the ANNC in an o line process and then inserting them into the microcontroller system, which indicates the need of much more work. References
[1] Richard E, Frederick PCH, Jayanta KP. Optimal reactive power control for industrial power networks. IEEE Trans Ind Appl 1999;35(3):50614. [2] Jin K, Ortmeyer TH. Application of static compensators in small AC systems. Electr Pow Compo Syst 2002;30(9):96780. [3] Mekhamer SF, El-Hawary ME, Mansour MM, Moustafa MA, Soliman SA. State of the art in optimal capacitor allocation for reactive power compensation in distribution feeders. In: IEEE large engineering systems conference on power engineering, LESCOPE 02; 2002. p. 6175. [4] Colak I, Bayindir R, Bay OF. Reactive power compensation using a fuzzy logic controlled synchronous motor. Energy Convers Manage 2003;44(13):2059215. [5] Schaefer RC. Excitation control of the synchronous motor. IEEE Trans Ind Appl 1999;35(3):694702. [6] Ortega JMM, Payan MB, Mitchell CI. Power factor correction and harmonic mitigation in industry. In: IEEE industry applications conference, vol. 5; 2000. p. 312734. [7] El-Sadek MZ, Fetih NH, Abdelbar FN. Starting of induction motors by static VAR compensators. In: Third international conference on power electronics and variable-speed drives; 1988. p. 4447. [8] Hsu SM, Czarnecki LS. Adaptive harmonic blocking compensator. IEEE Trans Pow Delivery 2003;18(3):895902. [9] Song YH, Johns AT, Aggarwal RK. Nonlinear thyristor-controlled static VAR compensation. In: Fifth European conference on power electronics and applications, vol. 8; 1993. p. 5660.

S. Sagiroglu et al. / Energy Conversion and Management 47 (2006) 32043215

3215

[10] Mohan N, Kamath GR. A novel, high-frequency, per-phase static VAR compensator. In: Industry applications conference, 30th IAS annual meeting, IEEE, IAS95, vol. 3; 1995. p. 25816. [11] Hua J, Goos G, Lopes L. An ecient switched-reactor-based static VAR compensator. IEEE Transactions on Industry Applications 1994;30(4):9981005. [12] Sharaf AM, Snider LA. An intelligent voltage controller for static VAR compensators. In: Proceedings of the 26th southeastern symposium on system theory; 1994. p. 23943. [13] Xu W, Marti JR, Dommel HW. Harmonic analysis of systems with static compensators. IEEE Trans Pow Syst 1991;6(1):18390. [14] Bohmann LJ, Lasseter RH. Stability and harmonics in thyristor controlled reactors. IEEE Trans Pow Delivery 1990;5(2):117581. [15] Gutierrez J, Montano JC, Castilla M, Lopez A. Power-quality improvement in reactive power control using FC-TCR circuits. In: IECON 02 industrial electronics society, IEEE 2002 28th annual conference, vol. 2; 2002. p. 8805. [16] Zhang L, Li QZ. Application of TSC in reactive compensation in traction substations. In: Proceedings of the international workshop on autonomous decentralized systems; 2000. p. 14953. [17] De Oliveira A, Martins RM, Goncalves WKA. Expert system for power quality improvement. In: Proceedings of the ninth international conference on harmonics and quality of power, vol. 2; 2000. p. 5049. [18] Nunes JVR, Braga RAM, Libano FB, Muller SL. A linear displacement power factor compensator. In: IEEE 35th annual power electronics specialists conference, PESC 04, vol. 2; 2004. p. 951957. [19] Al-Hamrani MM, Von Jouanne A, Wallace A. Power factor correction in industrial facilities using adaptive excitation control of synchronous machines. In: Pulp and paper industry technical conference, conference record of the 2002 annual. p. 14854. [20] Kentli F, Birbir Y, Onat N. Examination of the stability limit on the synchronous machine depending on the excitation current wave shape. In: IEEE international electric machines and drives conference, IEMDC 2001; 2001. p. 52832. [21] Chin T-H, Nakano M. Novel control scheme for adjustable speed synchronous machine drives. In: Proceedings of the power conversion conference, PCC Osaka, vol. 3; 2002. p. 129599. [22] Abido MA, Abdel-Magid YL. A fuzzy basis function network for generator excitation control. In: IEEE proceedings of the sixth international conference on fuzzy systems, vol. 3; 1997. p. 144550. [23] Handschin E, Homann W, Reyer F, Stephanblome T, Schlucking U, Westermann D, Ahmed SS. A new method of excitation control based on fuzzy set theory. IEEE Trans Pow Syst 1994;9(1):5339. [24] El-Sharkawi MA, Atteri R. Static security assessment of power system using Kohonen neural network. In: ANNPS 93, Proceedings of the second international forum on applications of neural networks to power systems; 1993. p. 37377. [25] Moghavvemi M, Yang SS. ANN application techniques for power system stability estimation. Electr Mach Pow Syst 2000;28(2):16778. [26] Gubina A, Golob R. Power systems security improvement through ANN-based decentralized secondary voltage control. in: The 5th seminar on neural network applications in electrical engineering, NEUREL 2000. p. 1948. [27] Das B, Velpula SP. Optimal capacitor switching in a distribution system using functional link network. Electr Pow Compo Syst 2002; 30(8):83347. [28] Neily G, Barone R, Josin G. Charney D. Joint VAR controller implemented in an articial neural network environment. In: International joint conference on neural networks, vol. 2; 1991. p. 133642. [29] Senjyu T, Urasaki N, Uezato K. Vector control of brushless DC motors using neural network. In: International conference on power electronics and drive systems, vol. 1; 1997. p. 29196. [30] Senjyu T, Shingaki T, Uezato K. A novel high eciency drive strategy for synchronous reluctance motors considering stator iron loss in transient conditions. In: Power electronics specialists conference, PESC, IEEE 32nd annual, vol. 3; 2001. p. 168994. [31] Weerasooriya S, El-Sharkawi MA. Laboratory implementation of a neural network trajectory controller for a DC motor. IEEE Trans Energy Convers 1993;8(1):10713. [32] Miller TJE. Reactive power control in electric systems. New York: Wiley-Interscience Publication; 1982. p. 182222. [33] Schaefer RC. Excitation control of the synchronous motor. In: Pulp and paper industry technical conference; 1998. p. 108115. [34] Maren A, Harston C, Pap R. Handbook of neural computing applications. London: Academic Press; 1990, ISBN 0-12-471260-6. [35] Haykin S. Neural networks: a comprehensive foundation. New York: Macmillan College Publishing Company; 1994, ISBN 0-02352761-7. [36] Minai AA, Williams RD. Acceleration of back propagation through learning rate and momentum adaptation. In: International joint conference on neural networks, vol. 1; 1990. p. 6769. [37] Neural computing, A technology handbook for professional II/PLUS and NeuralWorks Explorer. Pittsburgh: NeuralWare, Inc., Technical Publications Group; 1996.