A New Intelligent Autoreclosing Scheme Using Artificial Neural ...

306

IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 47, NO. 1, JANUARY/FEBRUARY 2010

A New Intelligent Autoreclosing Scheme Using Artificial Neural Network and Taguchi’s Methodology Fitiwi Desta Zahlay, K. S. Rama Rao, and Taib B. Ibrahim

Abstract—This paper presents a novel intelligent autoreclosure technique to discriminate temporary faults from permanent faults and accurately determine fault extinction time. A variety of fault simulations are carried out on a specified transmission line on the standard IEEE 9-bus electric power system using MATLAB/SimPowerSytems. FFT and Prony analysis methods are employed to extract data features from each simulated fault. The fault identification prior to reclosing is accomplished by an artificial neural network trained by standard Error Back-Propagation, Levenberg Marquardt, and Resilient Back-Propagation algorithms which are developed using MATLAB. Some important parameters which strongly affect the entire training process are fine tuned with Taguchi’s method to their corresponding best values. The robustness of the developed ANN identifier is verified by testing it with the data patterns which consists of high impedance faults obtained from IEEE 14-bus benchmark system. Test results show the efficacy of the proposed AR scheme. Index Terms—Adaptive automatic reclosure, artificial neural networks (ANNs), Error Back Propagation (EBP), Levenberg Marquardt (LM), Resilient Back-Propagation, Taguchi’s method.

I. I NTRODUCTION

A

N AUTORECLOSURE (AR) system is a scheme that is used to automatically re-energize a power line after a fault. Automatic reclosing of the tripped line offers a substantial improvement on the reliability, stability and security of the overall power system, particularly overhead industrial distribution systems. These improvements are based on the assumption that majority of faults on a transmission line are transient in nature i.e., once the fault arc has been extinguished by isolating the line under a fault, it does not restrike. However, the risk of unsuccessful reclosing emanates if the line under a fault is exposed to an elongated arc or permanent fault. Hence, the reliability, stability and security of the power system grid are put in danger.

Manuscript received February 25, 2010; accepted May 18, 2010. Date of publication December 6, 2010; date of current version January 19, 2011. Paper 2010-PSPC-035, presented at the 2010 IEEE/IAS Industrial and Commercial Power Systems Technical Conference, Tallahassee, FL, May 9–13, and approved for publication in the IEEE T RANSACTIONSON I NDUSTRY A PPLI CATIONS by the Power System Protection Committee of the IEEE Industry Applications Society. This work was supported by the Universiti Teknologi PETRONAS, Malaysia. The authors are with the Department of Electrical and Electronics Engineering, Universiti Teknologi PETRONAS, 31750 Tronoh, Malaysia (e-mail: [email protected]; [email protected]; taibib@petronas. com.my). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TIA.2010.2090936

Basically, conventional reclosures operate for permanent and transient faults without any discrimination and practice employing fixed dead time i.e., the time delay required prior to reclosing. Thus, closing decision is issued following circuit breaker opening [1]. For a temporary fault whose extinction time is less than the fixed time delay used by the AR system, subsequent re-energizing of the line will usually be successful. However, if it is a permanent fault or an extended arcing fault, excessive breaking duty is undesirably applied to the breaker unit resulting in not only an unsuccessful reclosure but also possible shock and damage to the system in general. In any case, reclosing before a fault totally cleared would potentially bring about irreparable damage to some system equipment. Therefore, discrimination of temporary fault from the permanent one and accurate determination of fault extinction time are highly demanded before any action is executed. In the very recent years, a number of researchers paid more attention to solve the problem associated with the conventional autoreclosure, and the recent developments on the issue, which explain the need for optimal reclosure in transmission lines, are summarized as follows. The application of artificial neural networks (ANNs) in power engineering is recently gaining interest [2]. As a result, methods for fault classification have been adopted using ANNs [3], [4], and adaptive AR for medium and high voltage transmission lines have been developed using different neural network configurations [5]–[8]. An autoreclosure which utilizes ANN as a pattern classifier and different harmonic components of positive sequence of voltage as inputs to ANN is developed in [9]. Similarly, a research group [10] suggested a single-phase AR technique which uses information extracted from residual voltage of a tripped phase using Discrete Wavelet Transform (DWT) as input to an ANN for classification. On the other hand, different mathematical models based on monitoring the fundamental component of the zero sequence power as in [11], tracking Harmonic Distortion Index (HDI) from the behavior of the low frequency components of voltage and current signals [12], estimating arc voltage minimalmaximal amplitude [13], detecting resonant component [14], and defining and identifying the waveform patterns of the voltage transients following initial breaker opening [15] have been proposed to develop an adaptive AR. A study on an adaptive single phase reclosing scheme on EHV/UHV transmission lines based on calculation of Bergeron model by utilizing harmonics energy ratio to identify arc extinguishment [16], by considering

0093-9994/$26.00 © 2011 IEEE

ZAHLAY et al.: AN AUTORECLOSING SCHEME USING ARTIFICIAL NEURAL NETWORK AND TAGUCHI’S METHOD

a secondary arc model rooted in gray-box model [17], and by distinguishing transient faults from permanent faults based on the carrier channel of carrier protections of EHV transmission lines [18] are reported in recent times. A different approach for adaptive AR with reference to power system stability based on multi-agent system using java agent development framework is proposed [19]. In reference [20], a self-adaptive auto-reclosure criterion using dual-window transient Energy Ratio (ER) for transmission line is proposed, and a novel concept of Close-Opening–Open-Closing (COOCG) morphological gradient is put forward. A numerical algorithm for determining adaptive dead time and blocking automatic reclosing during permanent faults on overhead lines which is based on terminal voltage input data processing has been proposed [21]. In this case, the decision if ever to reclose is determined by the total harmonic distortion factor of the fault voltage signal that is calculated by Discrete Fourier Transform (DFT). This paper focuses on developing an AR that possesses adaptive capability to identify a permanent fault from a temporary or the other way. The approach is mainly based on an optimized ANN. The inputs to the ANN are obtained by thoroughly analyzing the fault voltage waveforms using FFT and Prony Analysis (PA), a method that approximates a function with a sum of damped sinusoids [22], [23]. Accordingly, it has been rigorously found out that the features extracted from the energies contained in DC, fundamental and the first four harmonic components are adequate enough to identify one fault from the other. Besides, for temporary fault case, the paper presents a method, outlined to accurately determine the fault extinction time to avoid problems that may occur due to the predetermined and unchanging extinction time in conventional ARs. The proposed AR is more advantageous over the conventional counterpart. For instance, it increases rate of successful reclosure by avoiding reclosing onto a fault.

307

B. Levenberg Marquardt (LM) Algorithm LM algorithm is an iterative technique that locates a local minimum of a multivariate function that is expressed as the sum of squares of several nonlinear, real-valued functions. This is based on minimizing the objective function F (w) defined as in (1) which can also be written as in (4), [24] F (w) = EE T .

(4)

The increment of weights Δw is obtained as in Δw = [J T J + λI]−1 J T E

(5)

where E = [e11 , . . . , ek1 , e12 , . . . , ek2 , . . . , e1P , . . . , eKP ]T ; ekp = dkp − okp , k = 1, 2, . . . , K, p = 1, 2, . . . , P ; E is the cumulative error vector; J is Jacobian matrix; λ is the learning parameter; and I is identity matrix. The weight updates are expressed as in wt+1 = wt + Δwt , or −1 T Jt Et . wt+1 = wt + JtT Jt + λt I

(6)

In particular, λ is multiplied by a decay rate β (0 < β < 1) whenever F (w) decreases, whereas; λ is divided by β whenever F (w) increases in a new step. The training process can be illustrated in the following pseudo-codes: (i) initialize the weights and λ; (ii) compute the sum squared errors F (w) over all inputs; (iii) obtain the increment of weights Δw; (iv) re-compute F (w); using w + Δw as the trial w, and IF {F (w) < F (w) in step 2} THEN: {w = w + Δw; λ = λ ∗ β (where β = 0.1); Go back to step 2} ELSE: {λ = λ/β; Go back to step 4} END IF. C. Resilient Back Propagation (RPROP) Algorithm

(3)

RPROP is a local adaptive learning scheme, performing supervised batch learning in feed forward ANNs. The basic principle of RPROP is to eliminate the harmful influence of the size of the partial derivative on the weight step which is a common problem in EBP [24]. Consequently, only the sign of the derivative is considered to indicate the direction of the weight update. For each weight wij , its individual update-value Δij (t), which determines the size of the weight-update, and a second learning rule, which determines the evolution of the update-value Δij (t), are introduced. The estimation is based on the observed behavior of the derivative during two successive weight-steps as in ⎧ + η · Δij (t − 1), if ∂E/∂wij (t) ⎪ ⎪ ⎪ ·∂E/∂wij (t − 1) > 0 ⎨ (7) Δij (t) = η − · Δij (t − 1), if ∂E/∂wij (t) ⎪ ⎪ (t − 1) < 0 ·∂E/∂w ⎪ ij ⎩ else Δij (t − 1),

where w = [w1 , w2 , . . . , wN ]T consist of all weights; dij and oij are the desired value and the actual value of the ith output and the jth pattern; N is the number of weights; P is the number of patterns; and K is the number of outputs; η is the learning rate; and α is the momentum term.

where 0 < η − < 1 < η + . Each time the partial derivative of the corresponding weight wij changes its sign, which indicates that the last update was too big and the algorithm has jumped over a local minimum, the update-value Δij (t) is decreased by the factor of η − .

II. ANN A LGORITHMS AND TAGUCHI ’ S M ETHODOLOGY A. Error Back Propagation (EBP) Algorithm In this algorithm, error is propagated backwards in an ANN by apportioning it to each unit according to the amount of this error the unit is responsible for. The algorithm changes current weights of the network iteratively such that system error function, F (w), as in (1), is minimized [24], [25]. The weight update formulas during training are as in (2) and (3) F (w) =

P K (dij − oij )2

(1)

j=1 i=1

Δwji (t) = η ∗ ∂E/∂wij + α ∗ Δwji (t − 1) wji (t + 1) = wji (t) + Δwji (t)

(2)

308


Otherwise, if the derivative retains its sign, the update-value is slightly increased in order to accelerate convergence in shallow regions. The weight-update is performed in such a way that if the derivative is positive, the weight is decreased by its updatevalue, otherwise; the update-value is added as in (8) and (9). However, if the partial derivative changes sign that is the previous step was too large and the minimum was missed, the previous weight-update is reverted as in (10) ⎧ ⎨ −Δij (t), if ∂E/∂wij (t) > 0 (8) Δwij (t) = Δij (t), if ∂E/∂wij (t) < 0 ⎩ 0, else wij (t + 1) = wij (t) + Δwij (t)

(9)

Δwij (t) = − Δwij (t − 1), if ∂E/∂wij (t) · ∂E/∂wij (t − 1) < 0.

(10)

D. Taguchi’s Methodology Taguchi method is a scientifically disciplined mechanism for evaluating and implementing improvements in products, processes, materials, equipment, and facilities. These improvements are aimed at improving the desired characteristics and simultaneously reducing the number of defects by studying the key variables controlling the process and optimizing the procedures or design to yield the best results. This method which is applicable over a wide range of engineering fields is useful for “tuning” a given process for “best” results. Generally, a process to be optimized has several Control Parameters (CPs) which directly or indirectly decide the target or desired value of the output. The optimization then involves determining the best control factor levels so that the output hits the target value. In Taguchi’s Method, the word optimization implies determination of best levels of CPs. In turn, the best levels of control parameters are those that minimize the overall error. The experiments, that are conducted to determine the best levels, are based on orthogonal arrays and are balanced with respect to all control parameters, and yet are minimal in number. This in turn implies that the resources (materials and time) required for the experiments are also minimum [26]. To illustrate this, in an ANN, there is a need of optimizing some of its parameters to effectively hit the target. For instance, the usual practice to set the number of hidden layer neurons is trial and error. This practice, however, is inefficient and may not deliver optimal solution. But, it is possible to get the optimal number of hidden layer neurons by using Taguchi’s method. Taguchi proposed a standard eight-step procedure for optimizing any process: (i) Identify the main function, side effects, and failure mode. (ii) Identify the noise factors, testing conditions, and quality characteristics. (iii) Identify the objective function. (iv) Identify the control parameters and their levels. (v) Select the orthogonal array matrix. (vi) Conduct the matrix experiments.

(vii) Analyze the data; predict the optimum levels and performance. (viii) Perform the verification experiment, plan the future action. III. S YSTEM S IMULATION A. Transmission Line Faults and AR Scheme Various studies have shown that 70%–90% of faults on most overhead lines are transients which can successfully be cleared by momentarily de-energizing to allow the fault to vanish, and reclosing can restore service to the line [1]. The remaining 10–30% of faults is of semi-permanent or permanent in nature. In this case, an immediate de-energizing of the line and subsequent reclosing do not clear the fault but result in damage to system equipment. Despite the widespread use of conventional AR, there are some disadvantages associated with it: (i) fixed dead time; hence, reclosing onto permanent faults or secondary arcs; (ii) re-strike due to insufficient de-ionization; (iii) shock to the system causing beyond-repair damage to equipments. Basically, on EHV systems, an unsuccessful reclosure is more severe to the system than no reclosure at all. Thus, the development of AR system, that can differentiate the type of fault, is mandatory to prevent problems that can occur due to conventional reclosing technique. The technique proposed in this paper transforms the conventional AR system which is based on “restore service” into “reclose only if safe.” B. Simulation of Standard IEEE 9–Bus Power System In our previous work, the power system model used in the study of adaptive autoreclosure [27], was based on a single generator connected via an EHV transmission line to an infinite bus where the voltage is considered to be constant. However, this is an ideal case. The findings of the previous research works needed to be verified with the consideration of real or benchmark networks. Owing to this fact, a standard IEEE 9-bus electric power system with a 400-kV voltage level has been utilized to illustrate the capabilities and effectiveness of the tool developed in earlier works [27]. Fig. 1 shows the benchmark IEEE 9-bus electric power system. The data for the IEEE 9-bus system are taken from reference [28]. The system consists of three power generators G1, G2 and G3, six transmission lines, three transformers and three loads. The transmission line extending from busbar 8 to busbar 7 is the line of interest for simulations during the study. For more generality, it is empirical to consider the transmission line as series and parallel compensated. Moreover, a standard flat transmission line tower configuration is considered during the simulation. The frequency dependent distributed model is used to represent the line. The model is set up in Simulink and is simulated by generating several faults. The simulation of faults is based on a fault model [29], where a time-dependent dynamic resistance representation of a primary arc and improvements to


Fig. 1.

Single-line diagram of an IEEE 9-bus system.

Fig. 2.

Temporary single-phase-to-ground fault voltage.

309

Fig. 3. Normalized energies of harmonics for various faults.

Fig. 4. Amplitude of Prony terms for various faults.

the dynamic conducting characteristics of secondary arc models are adopted. Exhaustive study on various fault waveforms is conducted by taking into consideration factors that have significant influence on the faults–fault inception, fault resistance, fault duration and location. Fault signals are taken from measurements at busbar 8. A typical single-phase-to-ground temporary fault voltage signal generated at 10 km away from the sending end is shown in Fig. 2. C. Feature Extraction Before the ANN implementation, time domain simulations considering several contingencies are carried out for the purpose of gathering the training data sets, which is termed as feature extraction, from simulated fault voltage waveforms. Apparently, the data is rendered into a form which makes computation easier and faster for the ANN to effectively make decisions. This is carried out by taking information from energies of harmonics of the fault voltages. FFT is employed to examine the simulated fault voltage waveforms for feature extraction scheme. The FFT analyses clearly show the differences in the spectra of both types of faults: (i) There is more high frequency energy while a temporary fault exists than when the fault has extinguished. (ii) Permanent faults contain smaller fundamental component and harmonic components than temporary ones.

(iii) Cleared faults contain higher system frequency component and smaller harmonic components. Fig. 3 shows a plot of the energy contained in each harmonic component for 100 different fault cases (including temporary, cleared, and permanent fault samples). The first 30 fault data patterns are obtained by immediately examining the first cycle after the fault inception, and the remaining 30 data are taken from the 10th cycles where the faults fully extinguish. The remaining 40 fault data patterns are obtained from analyzing different permanent faults. Similar to the case of temporary faults, permanent fault data have been taken twice: one immediately after the fault inception and the other one, ten cycles after the fault inception. Unlike to the temporary fault case, the two cases more or less have the same features. This is due to the fact that a permanent fault persists for a longer time as opposed to the transient fault case which usually vanishes shortly following a trip. On the other hand, data patterns have been also extracted using Prony analysis as shown in Fig. 4. D. Neural Network Simulation and Taguchi’s Experiments The target or the output of the neural network is considered to be “1” whenever there is a fault and “0” when the fault is cleared. It is known that temporary fault extinguishes out after short period of time while permanent fault persists for long time. Thus, as demonstrated in Figs. 5 and 6, the ANN output pattern for each sample taken from the fault voltages will have

310


Prony analysis to develop similar ANN-based autoreclosure. Eventually, a comparison of the two feature extraction methods will be reported. Each of the algorithms used in this paper is coded in MATLAB with a GUI from which the parameters are varied. To setup Taguchi’s experiments, certain levels of each control parameter, within recommended range, are considered for each algorithm. During training, the performance of ANN is evaluated in several ways such as tracking root mean square error (RMSE), percent error index (% EI), etc. Expression for % EI is given in P K

Fig. 5. ANN output pattern for transient fault (volts versus time in seconds).

%EI =

p=1 k=1

Tkp − Akp

K P p=1 k=1

Fig. 6. ANN output pattern for a permanent fault (volts versus time in seconds). TABLE I S OME DATA PATTERNS E XTRACTED U SING FFT AND P RONY A NALYSIS FOR T RAINING THE ANN

∗ 100% Tkp

where E is the system error, Tp and Ap are the target and actual values for pth pattern and kth output, respectively, P is the number of patterns, and K is the number of network outputs. None of the above approaches takes the speed of training process into account. Time is an important factor in training. As a matter of fact, fast convergence is required during training. For this reason, it is required that the training process time (or equivalently saying the number of epochs) be small while maintaining minimum overall error. A new performance measuring index obtained by multiplying the percent EI by the number of epochs, as in (12), is used in this paper. This approach considers the combined effect of both % EI and time by effectively assessing the error index over time ε = n × %EI

patterns which resemble 111111000 . . . 0 (for temporary fault) and 11111111 . . . 1 (for permanent fault). Table I shows some of the extracted features for temporary and permanent faults using both FFT and Prony analysis. As mentioned earlier, each of the temporary faults cases, shown in Table I, starts after two cycles from the start of simulation, lasts for about 3–4 cycles, and fully extinguish after the 8th cycle. The development of ANN-based autoreclosure which encompasses training, testing and validation of the ANN is conducted using the data patterns obtained from FFT for this paper. The authors are working with the data patterns extracted using

(11)

(12)

where ε is the modified error index over time, n is the number of epochs (or iterations) a neural network took for training, and EI is the overall error index after training. Equation (12) is adopted in this paper to measure the performance of the ANN for each combination during the Taguchi’s experiments. Combinations of control parameters and levels identical to the number of Taguchi’s orthogonal arrays are set up and the corresponding percent error indices are determined accordingly [26]. After all, the objective of Taguchi’s experiment is to find a combination of selected control parameters with certain levels which gives rise to minimum error index and minimum number of epochs after training the ANN. This is based on the notion of Taguchi–“smaller-thebetter—SB”–which means the smaller the EI, the better the classification performance of the ANN is. For clarity purpose, Taguchi’s experiment for LM algorithm is presented in this paper. The control parameters are initial learning parameter (λ0 ), increment factor (λ+ ), decrement factor (λ− ), and number of hidden units (h). And, the corresponding levels for each of these four control parameters are shown in brackets in Table II. With L number of levels and P control parameters, there are N = LP combinations; in this case, 44 = 256 combinations. It is cumbersome to conduct experiments for all these combinations to come up with the optimal values of the parameters. Taguchi’s experiments,


TABLE II TAGUCHI ’ S E XPERIMENT FOR LM A LGORITHM

311

TABLE V AVERAGE ε FOR E ACH CP AND E ACH L EVEL –EBP

TABLE VI O PTIMIZED VALUES OF CP S

TABLE III AVERAGE ε FOR E ACH CP AND E ACH L EVEL –LM

TABLE IV AVERAGE ε FOR E ACH CP AND E ACH L EVEL –RPROP

however, reduce the number of experiments, which is based on calculating the statistical properties of orthogonal arrays, required to find the best level for each factor. For instance, in this paper, it is required to conduct experiments only on 16 of the combinations which correspond to L16 orthogonal arrays as in Table II. In Table II, the values of the control parameters considered during the Taguchi’s experiments are shown in brackets, and the modified error indices (ε) for each experimental run are shown in the last column. The next step is to obtain the net effect of each level value in each factor which is carried out by averaging the results which contain the same level and factor. This is shown in Table III from which it is inferred that the optimal combination of parameters with smallest effective ε is {λ0 λ+ λ− h} = {1 4 2 4}. In other words, the optimal combination of parameters is achieved when the values are set as: λ0 = 0.001, λ+ = 20, λ− = 0.2 and h = 5. Similarly, parameters of RPROP (Δ0 , η − , η + , α and h) are optimized by conducting Taguchi’s experiments on 16 orthogonal arrays −L16 . When Taguchi’s methodology is applied, the results are obtained as shown in Table IV where {Δ0 η − η + α h} = {4 3 2 1 2} makes the optimal combination vector with minimal ε output. This effectively means parameter

Fig. 7. Performance before (a) and after (b) optimization for LM algorithm.

values set Δ0 = 0.2, η − = 1.8, η + = 0.5, α = 0.001 and h = 3 give the best result compared to other sets of values. An experiment on L16 Taguchi’s orthogonal arrays for EBP algorithm with parameters (η, α and h, each with four levels as in Table V) has given rise to an optimal combination of η1 α2 h4 or simply {1 2 4} as in Table V. This corresponds to η = 1.0, α = 0.2, and h = 5. Table VI summarizes the values of parameters optimized using Taguchi’s experiments for each algorithm implemented to train the neural network. Figs. 7 and 8 illustrate the differences in training performances of ANN with randomly selected parameters and with the optimized values of parameters, as in Table VI, for LM and RPROP algorithms, respectively. It is evident to see the improvements to the training performance of the neural network achieved with the help of Taguchi’s method. A comparison of the three algorithms used in this paper shows that both LM and RPROP yield accuracy greater than 99% and guarantee fast convergence. Whereas, EBP is by far substandard to the rest as far as accuracy and fast

312


Fig. 9. Fig. 8. Performance before (a) and after (b) optimization for RPROP algorithm. TABLE VII T EST R ESULTS OF O PTIMIZED ANN

ANN outputs for temporary and permanent faults.

is 16.7 ms, and secondary arc starts at 66.7 ms and remains for another cycle as in Fig. 9. For the temporary fault case, the output of the developed ANN which has been already optimized to detect the secondary arc extinction time goes from “1” to “0”. Thus, for this case, the fault extinction time becomes 66.7 ms which is less than the preset dead time usually practiced in conventional autoreclosure. However, the output of the ANN remains “1” when the fault is permanent as shown in Fig. 9. VI. C ONCLUSION

convergence are concerned. An accuracy of 94.79% is attained after 10 000 iterations when EBP is used. Whereas, using LM and RPROP algorithms, training process took only one and five iterations to converge, respectively. IV. T ESTING W ITH DATA F ROM IEEE 14-B US E LECTRIC M ODEL Data collected from simulation of IEEE 14-bus have been utilized for testing the ANN-based AR scheme developed in IEEE 9-bus electric system model. The test results show the robustness of the method which showed a sound identification performance even for High Impedance Faults (HIFs), which has not been considered in the previous model. The ANN has not been trained with HIFs but it can still map new data accordingly. Table VII illustrates some of the results obtained by testing the already optimized and trained ANN. The response of the ANN for high impedance fault cases (see highlighted) has been outstanding except for EBP algorithm which gave poor performance compared to LM and RPROP algorithms, yet quite enough to make decisions. While the results show RPROP has a comparable output for this application, LM is exceptionally the most suitable algorithm as far accuracy and robustness are concerned. V. D ETERMINATION OF FAULT E XTINCTION T IME The fault extinction time for temporary fault is accurately determined from the fault voltage waveforms by detecting the secondary arc extinction time. Fig. 9 shows the way how the fault extinction time is obtained by analyzing a series of samples following a fault inception. The time of fault inception

This paper has proposed a novel and new ANN based AR scheme, based on defining the pattern of fault voltage waveforms simulated on IEEE 9-bus system, by adopting Taguchi’s optimization technique. The harmonic components of the simulated voltage waveform are analyzed by FFT to accurately identify the type of fault. EBP, LM and RPROP algorithms are used for training the network and Taguchi’s methodology is employed for optimization of control parameters in each algorithm. It is verified that the optimal parameter combination determined by Taguchi’s method yields the maximum accuracy. Hence, the ANN used is effectively optimized and the technique developed in this study has the ability to distinguish a transient fault from a permanent fault. It has been found out that LM is the best training technique for the ANN employed in this paper. In addition, the robustness of the developed ANN is verified by testing it with the data patterns obtained from IEEE 14-bus benchmark system, which consists of high impedance faults. The developed scheme is also able to accurately determine the fault extinction time and introduce a variable dead time for the autoreclosure scheme. The authors are working on to develop similar ANN-based autoreclosure using the data patterns extracted using Prony analysis, and a comparison of the two feature extraction techniques will be reported. R EFERENCES [1] E. Kuffel, High Voltage Engineering—Fundamentals, 2nd ed. Oxford, U.K.: Butterworth-Heinemann, 2000. [2] R. C. Bansal, “Optimization methods for electric power systems: An overview,” Int. J. Emerging Elect. Power Syst., vol. 2, no. 1, pp. 1–23, Mar. 2005. [3] M. Joorabian, S. M. A. Taleghani, and R. K. Aggarwal, “Accurate fault locator for EHV transmission lines based on radial basis function neural networks,” Elect. Power Syst. Res., vol. 71, no. 3, pp. 195–202, Nov. 2004. [4] P. K. Dash and S. R. Samantray, “An accurate fault classification algorithm using a minimal radial basis function neural network,” Eng. Intell. Syst., vol. 4, pp. 205–210, 2004.


[5] A. I. Megahed, H. M. Jabr, F. M. Abouelenin, and M. A. Elbakry, “Arc characteristics and a single-pole auto-reclosure scheme for Alexandria HV transmission system,” Elect. Power Syst. Res., vol. 76, no. 8, pp. 663–670, May 2006. [6] M. Lukowicz, “A new scheme for single-pole autoreclosure based on recurrent ANNs,” in Proc. Develop. Power Syst., 2004, vol. 2, pp. 591– 594. [7] V. V. Terzija and Z. M. Radojevic, “Numerical algorithm for adaptive autoreclosure and protection of medium-voltage overhead lines,” IEEE Trans. Power Del., vol. 19, no. 2, pp. 554–559, Apr. 2004. [8] I. K. Yu and Y. H. Song, “Wavelet analysis and neural network based adaptive single pole autoreclosure scheme for EHV transmission systems,” Int. J. Elect. Power Energy Syst., vol. 20, no. 7, pp. 465–474, Oct. 1998. [9] H. Khorashadi-Zadeh, “Artificial neural network approach for autoreclosure in transmission lines,” in Proc. IEEE Russia Power Tech, 2005, pp. 1–4. [10] M. A. El-Hadidy, D. H. Moustafa, and A. S. Attia, “Using neuro-wavelet technique for adaptive single phase autoreclosure of transmission lines,” in Proc. 39th Univ. Power Eng. Conf., 2004, vol. 2, pp. 684–688. [11] N. I. Elkalashy, H. A. Darwish, A. M. I. Taalab, and M. A. Izzularab, “An adaptive single pole autoreclosure based on zero sequence power,” Elect. Power Syst. Res., vol. 77, no. 5/6, pp. 438–446, Apr. 2007. [12] M. E. Golshan and N. Golbon, “Detecting secondary arc extinction time by analyzing low frequency components of faulted phase voltage or sound phase current waveforms,” Elect. Eng., vol. 88, no. 2, pp. 141–148, Jan. 2006. [13] K. J. Zoric, M. B. Djuric, and V. V. Terzija, “Arcing faults detection on overhead lines from the voltage signals,” Int. J. Elect. Power Energy Syst., vol. 19, no. 5, pp. 299–303, Jun. 1997. [14] W. Zengping, L. Haofang, X. Yan, M. Jing, and L. Junling, “Prediction method for preventing reclosing on permanent fault of shunt compensated EHV/UHV transmission lines,” in Proc. IEEE Power Eng. Soc. Gen. Meeting, 2006, pp. 1–6. [15] P. K. Aggarwal, Y. H. Song, and A. T. Johns, “Adaptive single-pole autoreclosure scheme based on defining and identifying fault induced voltage waveform patterns,” in Proc. Athens Power Tech, Joint Int. Power Conf., 1993, vol. 1, pp. 411–416. [16] L. Bin, S. Zhang, P. Crossley, and B. Zhiqian, “The scheme of singlephase adaptive reclosing on EHV/UHV transmission lines,” in Proc. IET 9th Int. Conf. Develop. Power Syst. Protection, 2008, pp. 116–120. [17] S. C. Alessandra, M. P. Carlos, and T. M. Cristina, “Single-phase autoreclosure studies considering a robust and reliable secondary arc model based on a gray-box model,” in Proc. Int. Conf. High Voltage Eng. Appl., 2008, pp. 486–489. [18] H. Qiang, L. Yongli, and L. Bin, “A new adaptive autoreclosure scheme to distinguish transient faults from permanent faults,” in Proc. Int. Conf. Power Syst. Technol., 2002, vol. 2, pp. 671–674. [19] Y. Lee, N. Sung, C. Kim, and R. K. Aggarwal, “Development of autoreclosing algorithm using multi agent system,” in Proc. IET 9th Int. Conf. Develop. Power Syst. Protection, 2008, pp. 126–131. [20] X. Lin, B. Wang, X. Deng, Z. Q. Bo, and A. Klimek, “A novel adaptive reclosure criterion for EHV transmission line,” in Proc. IET 9th Int. Conf. Develop. Power Syst. Protection, 2008, pp. 121–125. [21] Z. M. Radojevic and V. V. Terzija, “A new digital algorithm for overhead lines disturbance records analysis,” in Proc. IET 9th Int. Conf. Develop. Power Syst. Protection, 2008, pp. 658–663. [22] S. L. Marple, Digital Spectral Analysis With Applications. Englewood Cliffs, NJ: Prentice-Hall, 1987. [23] L. Scharf, Statistical Signal Processing: Detection, Estimation, and Time Series Analysis. New York: Addison-Wesley, 1991. [24] T. C. Chen, D. J. Han, F. T. K. Au, and L. K. Tham, “Acceleration of Levenberg-Marquardt training of neural networks with variable decay rate,” in Proc. IEEE Int. Joint Conf. Neural Netw., Jul. 2003, vol. 3, pp. 1873–1878. [25] M. Riedmiller, “Advanced supervised learning in multi-layer perceptrons—From back propagation to adaptive learning algorithms,” Comput. Stand. Interfaces, vol. 16, no. 3, pp. 265–278, Jul. 1994.

313

[26] H. R. Lochner, Ed., Designing for Quality. An Introduction to the Best of Taguchi and Western Methods of Statistical Method Design. London, U.K.: Chapman & Hall, 1990. [27] F. D. Zahlay and K. S. Rama Rao, “Taguchi’s method for optimized neural network based autoreclosure in extra high voltage lines,” in Proc. IEEE Power Energy Conf., 2008, pp. 901–906. [28] P. M. Anderson and A. A. Fouad, Eds., Power System Control and Stability. Hoboken, NJ: IEEE Press, 2003. [29] A. T. Johns, P. K. Aggarwal, and Y. H. Song, “Improved technique for modeling fault arcs on faulted EHV transmission systems,” Proc. Inst. Elect. Eng.—Gener., Transm. Distrib., vol. 141, no. 2, pp. 148–154, Mar. 1994.

Fitiwi Desta Zahlay was born in Ethiopia in 1980. He received the B.Sc. degree in electrical and computer engineering from Addis Ababa University, Addis Ababa, Ethiopia, in July 2005. He received a research scholarship grant in January 2007 offered by the Universiti Teknologi PETRONAS, Tronoh, Malaysia, and in July 2009, he received the M.Sc. degree in electrical and electronics engineering. Currently, he is working toward the Ph.D. degree under the same scheme. He worked for two years on the development of the Ethiopian Load Dispatch Center, the control center of the Ethiopian power grid, with French partners Areva T&D and Electricite De France (EDF). His research interests include application of artificial intelligence in power systems, power system transient analysis, and power system quality optimization.

K. S. Rama Rao received the B.E. (Hons.) degree in electrical engineering from Andhra University, Visakhapatnam, India, in 1962, the M.Sc. (Engg.) degree in electrical engineering from the University of Madras, Chennai, India, in 1963, and the Ph.D. degree in electrical engineering from the Indian Institute of Technology, Kanpur, India, in 1979. In 1963, he joined the Department of Electrical and Electronics Engineering, J.N.T. University, Kakinada, India, where he became a Professor in 1990. He was working at the university in various positions until 2002. He then joined the School of Electrical and Electronics Engineering, Universiti Sains Malaysia, Malaysia and was a member of the faculty until 2006. Presently, he is with the Department of Electrical and Electronic Engineering, Universiti Teknologi PETRONAS, Tronoh, Malaysia. His research interests include optimal design of power electronic controllers, electrical machine drives, artificial intelligence, and power systems.

Taib B. Ibrahim was born in Kedah, Malaysia, in 1972. He received the B.Eng. (Hons.) degree in electrical and electronics engineering from Coventry University, Coventry, U.K., in 1996, the M.Sc. degree in electrical power engineering from the University of Strathclyde, Glasgow, U.K., in 2000, and the Ph.D. degree in electrical machine design, from the University of Sheffield, Sheffield, U.K., in 2009. His employment experience includes Airod (M) Sdn Bhd and the Universiti Teknologi PETRONAS (UTP), Tronoh, Malaysia. Currently, he is the leader for power and energy cluster and the co-leader for mission-oriented research (energy) at UTP. His research interests range from electrical machines development to their associated drives.