IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 19, NO.

2, MAY 2004

979

Accurate Fault Location in the Power Transmission Line Using Support Vector Machine Approach
Robert Salat and Stanislaw Osowski, Member, IEEE
AbstractThe paper presents a new approach to the location of fault in the high-voltage power transmission line, relying on the application of the support vector machine and frequency characteristics of the measured one-terminal voltage and current transient signals of the system. The extensive numerical experiments performed for location of different kinds of faults of the transmission line have proved very good accuracy of fault location algorithm. The average error of fault location in a 200-km transmission line is below 100 m and the maximum error did not exceed 2 km. Index TermsFault location, support vector machine.

I. INTRODUCTION

fault on the basis of fundamental voltage and current components of the measured data. In the second stage, the high-frequency characteristics of the system are analyzed in the vicinity of the predicted fault location and compared to the measured frequency peak in the high-frequency range. In this way, the initial location of fault is corrected. This two-stage approach is called here the hybrid algorithm. It can deliver very accurate results of the fault location for all types of systems and different kinds of faults. The numerical experiments have confirmed that the average error of fault location in 200-km high-voltage transmission line can be reduced in this way to approximately 45 m only.

HE accurate fault location in the power transmission line based on the measurement of the currents and voltages on one side of the system is still an important problem, since a more accurate location results in the minimization of the amount of time spent by the line repair crews in searching for the fault. The problem has attracted widespread attention among researchers in power system technology [1][14]. Different techniques of location have been applied. Most algorithms are based on the computation of the postfault fundamental frequency voltage and current phasors [1][5], used either directly in the appropriate mathematical relations or as the input variables to the artificial neural networks. Another class of algorithms uses global positioning system (GPS) and phasor measurement unit (PMU) techniques or a traveling wave method [6], [7]. There are also many different algorithms that use raw samples of voltage and current data fit to the appropriate model of the line [8][10]. The unknown distance to the fault is then usually solved in the time domain. Also, methods related to two-terminal voltage and current measurements have been developed [11], [12]. Especially difficult is the problem of fault location on the basis of only one end data of the transmission line. In such a case, the accuracy of single-ended fault locators is strongly affected by the assumption about the fault impedance, generally unknown. In this paper, we present a novel method based on the application of artificial neural support vector machine (SVM) network [15][18] and frequency characteristics of the measured transient data. The proposed method works in two stages. In the first one, the SVM network predicts the rough position of the

II. CHARACTERISTICS OF THE POWER TRANSMISSION LINE IN THE FAULTY STATE We consider the power transmission line connecting two systems A and B as shown in Fig. 1. Assume the fault of the line, as the shorting resistance of the unknown value, occurring somewhere at the unknown distance. Different types of faults are considered in the paper, including phase to ground (R-g, S-g, T-g), phase to phase (R-S, S-T, R-T), two phases to ground (R-S-g, S-T-g, R-T-g), and three phase (R-S-T). The power system has been modeled using ATP-EMTP program [19]. The faults have been simulated by applying the universal circuit structure (resistances and switches) as presented in Fig. 1. All shorting resistances are of equal values, changing in the simulations according to the actual assumptions. The location of the particular type of fault should be done on the basis of the measured transient voltages and currents at the origin of the system A. At this stage, we assume that the fault type has already been detected [14]. Any fault occurring in the line causes the transient state, distorting the sinusoidal waveforms of the currents and voltages, and increasing the magnitude of currents in the ill phases. The distortions change with the type of fault, type of the interconnected systems, distance of the fault to the system, etc. Fig. 2 presents the exemplary three-phase transient voltages and currents of the 200-km 400-kV transposed transmission line (Clark model) under R-g phase-to-ground fault occurring at two different places. These transients have been obtained using ATP-EMTP program. The shorting resistance assumed in the . Both systems used in the experiments was equal simulations were weak. Fig. 2(a) and (b) correspond to the fault occurring at the distance of 1 km from the system A and Fig. 2(c) and (d)199 km from the system A. It is seen that the level of current of ill phase as well as the harmonic components of both current and voltage change greatly with the distance to the fault.

Manuscript received April 17, 2003. The paper has been supported by KBN, Poland. R. Salat is with the Warsaw University of Agriculture, Warsaw 02-776, Poland (e-mail: salat@iem.pw.edu.pl). S. Osowski is with the Warsaw University of Technology, Warsaw 00-661, Poland, and also with Military University of Technology, Warsaw 00-908, Poland (e-mail: sto@iem.pw.edu.pl). Digital Object Identifier 10.1109/TPWRS.2004.825883

0885-8950/04$20.00 2004 IEEE

980

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 19, NO. 2, MAY 2004

Fig. 1. EMTP model of the fault of the line connecting two systems A and B.

Fig. 2.

Transients voltages and currents measured at the input side of the line at the fault occurring at the origin (a and b) and end (c and d) of the line.

III. GENERATION OF FEATURES In our approach to fault location, we will rely on the values of the magnitudes of the fundamental harmonics of the voltages and currents of the faulty phases and also on the frequency characteristics of the measured voltage in the high-frequency range. Fig. 3 presents the relationship of the magnitude of the first harmonic of the voltage [Fig. 3(a)] and current [Fig. 3(b)] versus distance to the phase-to-ground fault, obtained using ATP-EMTP program for five different values of the shorting , changing in experiments from 2 to 25 , at resistance different angles of the supplying voltage. The characteristics depend significantly on the value of shorting resistance, especially if it is above 2 . On the other hand, the varying angle of the supplying voltage causes only slight fuzziness of these relationships. Similar relations exist for all different kinds of faults of the line.

It is evident from these curves that relying on the location of fault on the fundamental harmonics of voltages and currents only, will not allow getting very accurate results, since, in practice, the value of the shorting resistance is generally unknown. To correct the estimation of the distance to the fault, we propose to consider also the frequency characteristics of the faulty system in the high-frequency range, usually neglected by the researchers. The first step in the analysis of frequency characteristics is to uncover higher harmonics by high-pass filtering of the measured voltage or current transients. After some numerical experiments performed for different shorting resistances and at different moments of fault occurrence, we have found that the best results are obtained at the application of the elliptic filters of the 8th order, of the ripple in the passband equal to 0.01 dB, and the attenuation in the stopband equal to 120 dB. The cut frequency and the length of the measuring window depend on the type of the

SALAT AND OSOWSKI: ACCURATE FAULT LOCATION IN THE POWER TRANSMISSION LINE

981

Fig. 3. Dependence of the magnitude of the fundamental harmonic of the voltage (a) and current (b) measured at the origin of the line, as the function of the distance to the fault at the phase-to-ground fault.

Fig. 4. Frequency characteristics of the filtered signal in the high-frequency range.

systems under investigation and are in the range above 4000 Hz. Filtering the measured voltage and current signals uncovers the high-pass range. The typical frequency characteristics of the filtered signal at the faulty conditions are shown in Fig. 4. . There is a visible peak corresponding to the frequency This frequency has been found to be the function of the fault distance. Moreover, it is highly insensitive to the value of shorting resistance and the angle of the source voltage signal at the moment of fault occurrence. Many experiments performed at different faulty conditions (different values of shorting resistance, different moments of fault occurrence, different loading) have confirmed this shape of frequency characteristics in the highfrequency range. Fig. 5 shows the typical dependence of on the fault distance to the origin of the line for five different , 5, 10, 14, 25 ) and at values of the shorting resistance ( varying angle of the supply voltage, at the two-phase-to-ground fault. It is seen that this relationship is very weakly dependent on the value of the shorting resistance, in full range of the fault locations in the line. Small disturbances, visible in some chosen locations of characteristics, are the only visible results of different shorting resistance values. Each fault type (one phase to ground, two phases, or three phases faults) is associated with specific relationship between

and the placement of fault in the line. The presented curve is not monotonic and thus cannot be used directly in the learning process of the neural network. However, if we know roughly the range of the fault, we may use this relationship to correct the fault location. This will be exploited later and used in the so-called hybrid SVM approach to the fault location.

IV. SVM NEURAL NETWORK The SVM, the solution of the universal feedforward networks, pioneered by Vapnik [15], [16], is known as the excellent tool for classification and regression problems of good generalization performance. In distinction to the classical neural networks, the formulation of learning problem of SVM leads to the quadratic programming with linear constraints. , Basically, the SVM is a linear machine of one output working in the high dimensional feature space formed by the nonlinear mapping of the -dimensional input vector into through the use of a -dimensional feature space . The number of hidden units is equal to the function number of so-called support vectors that are the learning data points, closest to the separating hyperplane. The learning task is transformed to the minimization of the error function, defined

982

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 19, NO. 2, MAY 2004

Fig. 5.

Relationship in high-frequency range between the peak frequency f

and the distance for two-phases-to-ground faults.

through the so called -insensitive loss function [15] for for (1)

where is the assumed accuracy, is the destination, is the is the actual output of the network under input vector, and excitation of . The actual output signal of the SVM network is defined by (2)

mension of the approximating function [15], [17]. Both must be selected by the user. The solution of so-defined constrained optimization problem is solved by the introduction of the Lagrangian function and Laresponsible for funcgrange multipliers , tional constraints. The minimization of the Lagrangian function has been transformed to the so-called dual problem [15], [16], [18] max (6)

is the weight vector and the basis function vector. The learning task is defined as the minimization problem of the error function (3)

where

at the constraints

(7) at the upper bound on the weight vector , , where is a user-specified constant and is the number of learning . Introducing the slack variables and , the data pairs learning problem can be redefined as the minimization of the cost function (4) at the following functional and boundary constraints: where is a inner-product kernel defined in accordance with Mercers theorem [15] on the basis of the learning data set . Note that the dual problem belongs to the quadratic programming optimization tasks with respect to the Lagrange multipliers. The solution of it is relatively easy and leads to the global minimum. After solving the dual problem, the optimum solution for the vector is given by (8) where is the number of so-called support vectors (equal to the number of nonzero Lagrange multipliers). The network can be expressed through the Lagrange muloutput signal tipliers and kernel function (9)

(5) in (4) is a user-specified parameter. The two The constant variables and are free parameters that control the VC di-

SALAT AND OSOWSKI: ACCURATE FAULT LOCATION IN THE POWER TRANSMISSION LINE

983

and does not need to know the explicit form of the nonlinear . The most known kernel functions used in basis function practice are radial (Gaussian), polynomial, spline, or sigmoidal functions [16], [18]. The most important is the choice of coefficients and . Constant determines the margin within which the error is neglected. The smaller its value, the more support vectors will be found by the algorithm. The constant is the weight, determining the balance between the complexity of the network, characterized by the weight vector and the error of approx. imation, measured by the slack variables For the normalized input signals, the value of is usually , and is much bigger adjusted in the range than 1.

V. HYBRID METHOD OF FAULT LOCATION IN THE POWER TRANSMISSION LINE A. Principle of the Method To solve the problem of fault location in the power transmission line, we will apply the SVM neural network approach. The most important problem in solving any task using neural network is generation of the features, characterizing the process in the best unique way. In the solution presented here, we will rely our approach on two kinds of information: the magnitude of fundamental harmonics of the voltage and current at one terminal of the system; the frequency characteristics in high-frequency range of the measured transient voltage at the same terminal of the transmission line. This two-stage fault location algorithm will be called here the hybrid approach. It has been shown in Section III that the faulty current and voltage characteristics are not unique and greatly depend on the shorting resistance value as well as on the angle of the supply voltage at the fault moment. It means that relying placement of the fault on this information only, will not deliver the accurate results, since the value of shorting resistance is generally unknown. To improve the accuracy of estimation, we propose here to use the information contained in the frequency characteristics of the system in the high-frequency range. The numerical experiments described in Section II have confirmed that in the high-frequency range, there is the frequency corresponding to the maximum (peak) of the magnipoint tude of the voltage that is directly changing with the placement of fault and is practically independent on the value of shorting resistance and on the angle of the supply voltage. It corresponds to the resonance phenomena occurring in the power system. We will use this information to correct the placement of the fault. B. Hybrid Approach to the Fault Location This general idea of the hybrid algorithm of the fault location proposed in the paper is illustrated in Fig. 6. The SVM neural network of inputs related to the fundamental harmonics of voltage (vector ) and current (vector ) of all ill phases and of the supplying voltage determines to the measured angle the initial placement of the fault.

Fig. 6. General idea of the hybrid approach to the fault location.

The second stage uses the high-frequency range characteristics of the power system in the vicinity of the initially located fault and measured peak frequency of the actual system for precise estimation of the distance to the fault. The hybrid approach is then composed of two succeeding stages. Apply the SVM neural network working in the regression mode. The inputs to the SVM neural network are formed by the amplitudes of the first harmonics of voltages and currents of all ill phases of the transmission line, as well as the information of the measured angle of the supply voltage at the fault moment. The number of hidden radial units is adjusted in the learning procedure and is equal to the number of support vectors. The SVM network has only one output, whose signal indicates the initial placement of the fault. In the second stage, the measured peak frequency value is compared with the high-frequency characteristics of the power system in the neighborhood of the place indicated by SVM. This comparison corrects the initial placement to the final estimation pointed by the actually measured . To get the best results of location, we use the high-frequency range characteristics averaged over all cases corresponding to all applied shorting resistance values used in learning the SVM network. Hybrid solution should be applied independently for each type of the fault: phase-to-ground, phase-to-phase (with and without ground), and three-phase fault. It means that each type of the fault needs learning one SVM network and also generation of the appropriate frequency characteristics in the high-frequency range. Once the networks have been trained, they can be used in the retrieval mode, in which the input variables to SVM . The trained hyare determined, as well as the value of brid system, corresponding to the appropriate type of fault (determined earlier) calculates then the exact location of the fault in the line. The numerical experiments performed for different kinds of systems at the faults placed at different locations of the line have proven good performance of this approach to the fault location.

984

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 19, NO. 2, MAY 2004

VI. RESULTS OF NUMERICAL EXPERIMENTS The numerical experiments have been performed for four types of 400-kV noncompensated system interconnections: strong-strong, strong-weak, weak-strong, and weak-weak. Different types of faults have been considered, including: phase-to-ground (R-g, S-g, T-g), phase-to-phase (R-S, S-T, R-T), two-phases-to-ground (R-S-g, S-T-g, R-T-g), and three-phase (R-S-T). The transients of one-terminal phase voltages and currents of the unloaded line, switched on very shortly (for two periods only) at the fault presence, have been registered and processed according to the procedure described above. The data used in learning have been generated by using ATP-EMTP program and then preprocessed according to the described procedure. In learning, we have used the chosen data corresponding to the faults occurring every 100 m of the 200-km transmission line (approximately 2000 data points for each case). This short distance discretization has been assumed to increase the information on the line operating conditions, very essential for obtaining good generalization properties of neural networks. Observe that the neural network fulfills the role of universal approximation and more learning data means its better approximation ability. Different values of shorting resistances have been used, changing from 2 to 25 ohm. Also, different angles of the supplying voltage have been tried to show the neural network different operating conditions at future faults. More than 200 000 learning data have been generated in this way. The learning process of the SVM networks has been performed by using Mangasarian algorithm [17]. We have used the Gaussian radial basis function as the kernel function of the value of . The optimal value of the parameter has been determined after a series of . The parameter used experiments and assumed in experiments was equal . Three different SVM networks have been trained: one for phase-to-ground fault, second for phase-to-phase fault, and the third for three-phase fault. The optimal neural network structures found by the learning program are given in Table I. The first digit in notation of the structure indicates the number of inputs (the voltage and current of all ill phases and the angle of the supply voltage at the moment of fault). The second digit points to the number of hidden radial neurons (support vectors) and the last, equal in all cases one, corresponds to one output linear neuron responsible for initial estimation of the fault distance. For each type of fault, the appropriate high-frequency range characteristics have been determined, and then used in the second step of algorithm to get the corrected location of fault. To get the best results of location, we have used the high-frequency range characteristics averaged over all appropriate cases corresponding to all applied shorting resistance values, used in learning of the SVM network. All experiments have been repeated for different system interconnections. After learning, all parameters of the trained networks have been frozen and then used in the retrieval mode for testing the capabilities of the system on the data not used in learning. The testing data samples have been generated through

TABLE I OPTIMAL STRUCTURES OF SVM NETWORKS

the ATP-EMTP program by placing faults at the distances not used in learning, by applying different resistance values, and different moments of fault occurrence. The results in the form of errors (measured in meters) for different types of faults and different system interconnection are shown in Table II. Two types of errors have been calculated: the maximum and the mean, calculated as the average of all errors at 2000 fault positions, distributed every 100 m along 200-km transmission line. A large number of testing data have been used to check the proposed solution in the most objective way at practically all possible fault locations. As it is seen, the maximum error of fault location is limited to 2 km in the worst case. The maximum errors may happen from time to time for the testing data at conditions highly different from the learning samples and are the results of the error of the first stage of location (SVM alone), magnified by the nonmonotonic character of the frequency characteristics in the high-frequency range (second stage of location). The main sources of these errors are peak-type regions of frequency characteristics, like that of Fig. 5 (the range of 135 km, 160 km, 175 km, and 190 km). In the second (correcting) stage of algorithm, they may magnify the errors, made by SVM alone. Unfortunately, the compensation of these errors is not possible by including them into learning samples, since real testing errors correspond to the unknown fault conditions of the line (shorting resistance, moment of fault occurrence, etc.). Fig. 7 presents the typical histogram of location errors for phase-to-ground fault (R-g) for strong-strong system interconnection. Only seven errors have the magnitude over 1000 m (at 2000 testing samples). Table III depicts the number of location errors exceeding 1 km for different faults and system interconnections. They are very limited and most of them happen at the fault occurring in the ending section of the line. Fig. 8 presents the typical distribution of location errors for 200-km transmission line at the measurement resolution of 100 m at phase-to-ground (T-g) faults. For this particular case, the mean percentage error of fault location, calculated as the ratio of the absolute error to the real distance of the fault, was equal to 0.08%. Similar results of fault locations regarding accuracy have been obtained for other types of faults and system interconnections. Table IV presents the exemplary relative errors of fault location in the form of mean value over 2000 cases for all investigated types of faults for the most demanding strong-strong system interconnections. All mean errors are of very limited magnitude. They depend on the type of the system interconnection and the type of fault. The most difficult seems to be the location of phase-to-ground and two-phases-to-ground faults. On the other hand, the best results have been obtained for weak-weak system interconnection.

SALAT AND OSOWSKI: ACCURATE FAULT LOCATION IN THE POWER TRANSMISSION LINE

985

TABLE II TEST ERRORS OF FAULT LOCATION IN TRANSMISSION LINE SYSTEM AT DIFFERENT CONFIGURATIONS OF SYSTEM INTERCONNECTION AND DIFFERENT TYPES OF FAULTS

Fig. 8. Distribution of location errors for T-g faults at weak-weak system interconnection.

Fig. 7. Histogram of the location error for phase-to-ground faults for strong-strong systems. TABLE III LOCATION ERRORS EXCEEDING 1 km TRANSMISSION LINE

NUMBER

OF

OF

200-km

conditions, shorting resistances, systems interconnections, etc.). The most often cited paper [1] from 1982 investigated 71.2-km line, achieving errors from 200 to 900 m (0.6% to 2.6%). We have applied the method of this paper for our transmission line configuration at strong-strong system interconnection for three values of fault resistance. The obtained results of localization, compared to our methods (SVM alone and hybrid method at random fault resistance values in the range 525 ) are presented in Table V. The maximum error of Takagi algorithm exceeded 3.5 km and is much higher than that obtained by our method. Similar results have been obtained for other types of system interconnections. An interesting neural solution of fault location based on oneterminal measurements, by applying the fuzzy neural network, has been presented in [5]. Their reported accuracy changes significantly from 0.03% to 7.91%, depending on the placement of fault, shorting resistance value, and operating conditions. The average error for all cases reported in the paper is about 4%. Much better results have been presented in the papers [12] at two-terminal measurements. The accuracy of fault location depends on the placement of the fault. For example, of faults occurring at 80% of the line length, the estimated locations of different type faults have ranged from 79.31% to 80.79%. Thus, the individual errors are within the range of approximately . Even better results have been reported in the paper [11]. The fault location errors in 100-km line reported for selected places of faults (1 km, 5 km, 10 km, 15 km, 19 km) are within the range from 0.029% to 0.471%, depending on and 1 ohm) the values of the shorting resistances ( and different fault types. VII. CONCLUSION The paper has presented an effective method of fault location in the power transmission line. The method uses the so-called hybrid approach, composed of two steps. In the first one, the SVM neural network estimates the initial distance to the place of fault using the information contained in the fundamental harmonics of the voltages and currents of ill phases. The second step corrects the final estimation of this distance by using the information contained in the high-frequency range characteristics.

It is interesting to present our results on the background of other results reported by different authors in scientific journals within the last years. However, the main problem of this comparison is that different system interconnections have been used in different papers (different lengths of the line, loading

986

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 19, NO. 2, MAY 2004

TABLE IV MEAN OF THE RELATIVE ERRORS OF FAULT LOCATION FOR ALL INVESTIGATED TYPES OF FAULTS AT STRONG-STRONG SYSTEMS INTERCONNECTION

TABLE V COMPARISON OF LOCALIZATION ERROR (IN KILOMETERS) OF THE PHASE-TO-GROUND FAULT IN 200-km INE AT STRONG-STRONG SYSTEM INTERCONNECTION

The results of numerical experiments performed for the location of faults in different type of power system interconnections and different kinds of fault have proved very good performance of the proposed solution. The mean of all average errors of fault location is below 100 m in the 200-km line, irrespective of the type of system and kind of fault. REFERENCES
[1] T. Takagi, Y. Yamakoshi, M. Yamamura, R. Kondow, and T. Matsushima, Development of a new type fault locator using one-terminal voltage and current data, IEEE Trans. Power App. Syst., vol. PAS-101, pp. 28922898, Aug. 1982. [2] M. S. Sachdev and R. Agarwal, A technique for estimating the fault locations from digital impedance relay measurements, IEEE Trans. Power Delivery, vol. 3, pp. 121129, Jan. 1988. [3] W. Caminhas, H. Tavares, and F. Gomide, A neuro-fuzzy approach for fault diagnosis in dynamic systems, in Proc. IEEE Int. Conf. Fuzzy Syst., vol. 3, New Orleans, LA, 1996, pp. 20322037. [4] B. Cannas, G. Celli, and M. Marchesi, Neural networks for power system condition monitoring and protection, Neurocomputing, no. 23, pp. 111123, 1998. [5] K. Dash, A. K. Pradham, and G. Panda, A novel fuzzy network based distance relaying scheme, IEEE Trans. Power Delivery, vol. 15, pp. 902907, July 2000. [6] J. A. Jiang, J. Z. Yang, Y. H. Lin, C. W. Liu, and J. C. Ma, An adaptive PMU based fault detection technique for transmission line, IEEE Trans. Power Delivery, vol. 15, pp. 486493, Apr. 2000. [7] H. Lee, Development of accurate traveling wave fault locator using global positioning satellites, in Proc. Spring Meeting Can. Elect. Assoc., Montreal, QC, Canada, Mar. 1993. [8] A. Gopalkrishnan, M. Kezunovic, S. M. Kenna, and D. M. Hamai, Fault location using the distributed parameter transmission line model, IEEE Trans. Power Delivery, vol. 15, pp. 11691174, Oct. 2000. , V. Terzija, and M. Djuric , Numerical algorithm for [9] Z. Radojevic overhead lines arcing faults detection and distance and directional protection, IEEE Trans. Power Delivery, vol. 15, pp. 3137, Jan. 2000. [10] C. Booth and J. R. McDonald, The use of artificial neural networks for condition monitoring of electrical power transformers, Neurocomputing, no. 23, pp. 97109, 1998. [11] Y. H. Lin, C. W. Liu, and C. S. Yu, A new fault locator for three terminal transmission lines using two-terminal synchronized voltage and current phasors, IEEE Trans. Power Delivery, vol. 17, pp. 452459, Apr. 2002.

[12] D. Novosel, D. Hart, E. Udren, and J. Garitty, Unsynchronized twoterminal fault location estimation, IEEE Trans. Power Delivery, vol. 11, pp. 130138, Jan. 1996. [13] E. E. Vazquez, H. Altuve, and O. Chacon, Neural network approach to fault detection in electric power systems, IEEE Trans. Power Delivery, vol. 11, pp. 20902095, Oct. 1996. [14] A. Poeltl and K. Frohlich, Two new methods for very fast fault type detection by means of parameter fitting and ANN, IEEE Trans. Power Delivery, vol. 14, pp. 12691275, Oct. 1999. [15] V. Vapnik, Statistical Learning Theory, New York: Wiley, 1998. [16] A. Smola and B. Scholkopf. (1998) A Tutorial on Support Vector Regression, Neurocolt Tech. Rep. NV2-TR-1998-030. [Online]. Available: http://www.neurocolt.com [17] O. L. Mangasarian and P. Lagrangian, Support vector machines, J. Mach. Learning Res., pp. 161177, 2001. [18] S. Haykin, Neural Networks, Comprehensive Foundation. Englewood Cliffs, NJ: Prentice-Hall, 2000. [19] ATP-EMTP, Rule Book, Canadian-American EMTP Users Group, 1997.

Robert Salat was born in Poland in 1973. He graduated from the Cracow University of Technology, Cracow, Poland, in 1997. He received the Ph.D. degree from the Warsaw University of Technology, Warsaw, Poland, in 2002. His areas of interest are neural networks, diagnostics, and fault location in electrical circuits and power systems.

Stanislaw Osowski (M99) was born in Poland in 1948. He received the M.Sc., Ph.D., and Dr.Sc. degrees in electrical engineering from the Warsaw University of Technology, Warsaw, Poland, in 1972, 1975, and 1981, respectively. Currently, he is a Professor of electrical engineering at the Institute of the Theory of Electrical Engineering, Measurement and Information Systems at the Warsaw University of Technology. His research and teaching interests include neural networks, optimization techniques, and computer-aided circuit analysis and design. He is an author or co-author of many scientific papers and ten books.