Paper. Fault Location in an Unbalanced Distribution System using Support Vector Classification and Regression Analysis

Transcription

1 IEEJ TRANSACTIONS ON ELECTRICAL AND ELECTRONIC ENGINEERING IEEJ Trans ; 3: 37 5 Published online in Wiley Online Library (wileyonlinelibrary.com). DOI:./tee.59 Paper Location in an Unbalanced Distribution System using Support Vector Classification and Regression Analysis Sophi Shilpa Gururajapathy *a, Non-member Hazlie Mokhlis *, Non-member Hazlee Azil Bin Illias *, Non-member Ab Halim Abu Bakar **, Non-member Lilik Jamilatul Awalin ***, Non-member Support vector machine (SVM) is a novel machine for data analysis and has advantageous characteristic of good generalization. Because of this characteristic, SVM is used in this work for fault classification and diagnosis in distribution systems. This work proposes an effective fault location method using SVM to identify the fault type, faulty section, and fault distance. The classification and regression analysis of the SVM are performed to locate a fault. The proposed method utilizes the voltage sag magnitude and measured at the primary substation of a distribution system. First, the fault type is identified using oneversus-one concept of support vector classification. The next step identifies the faulty section by calculating fault resistance, finding possible faulty sections and ranking the possible sections. Finally, the fault distance is identified using support vector regression analysis. The performance of the proposed method is tested using SaskPower distribution system from Canada having line sections. Test cases are carried out under various fault scenarios considering the fault type and fault resistance. The results of fault distance are compared for different kernel functions, and the most accurate kernel is chosen. Test results show that the proposed method obtains reliable fault location. 7 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc. Keywords: support vector machine; fault type; faulty section; fault distance; distribution system Received 3 December ; Revised February 7. Introduction The distribution system plays a vital role in delivering power to customers. It may face fault occurrences that are permanent, transient, or intermittent due to insulation breakdown, change in weather conditions, physical damage, or human error. The fault occurrence causes the current to pass through an improper path, which will damage the equipment and lead to power interruption []. In order to maintain continuous power supply to customers and to ensure high reliability indices for the system, the fault has to be identified and isolated from the system. s in the distribution system are classified into four types, namely single line to ground fault (SLGF) or single line to earth fault, line to line fault (), double line to ground fault () or double line to earth fault, and three-phase to ground fault () or three-phase to earth fault. Earlier, fault in a distribution system used to be identified using conventional methods, which include visual inspection, cut and try method, and sectionalizing by re-fusing. However, these methods typically results in damage to customer and utility equipment due a Correspondence to: Sophi Shilpa Gururajapathy. shilpa.sophi@gmail.com; sophishilpa@siswa.um.edu.my *Department of Electrical Engineering, Faculty of Engineering, University of Malaya, 53 Kuala Lumpur, Malaysia **University Malaya Power Energy Dedicated Advanced Centre (UMPEDAC), Level, Wisma R&D, University of Malaya, 5999 Kuala Lumpur, Malaysia ***University Kuala Lumpur, Electrical Engineering Section, International College, British Malaysian Institute, Batu, Jalan Sungai Pusu, 53 Gombak Selangor, Malaysia to switching surges and fault currents. Because of the importance of locating faults, research on various automated fault location methods is carried out to identify the fault and to expedite the restoration process. location methods are generally classified into impedancebased methods and knowledge-based methods. The impedancebased method utilizes voltage and current measurements from primary substation to locate the fault. A well-known impedancebased method was proposed by Girgis []. The method calculates the fault distance in a balanced network. Another impedancebased method suitable for both balanced and unbalanced systems is presented in [3,]. The method finds the fault distance using direct circuit analysis by utilizing the matrix inverse lemma. The limitation is that the method is implemented only for line to earth faults and line to line faults. Considering this limitation, an approach suitable for all fault types was proposed in [5,]. These methods proposed a fault analysis algorithm based on the phase coordinates and fault resistance. Also, the methods use an iterative approach for fault current calculation, which results in computational error due to a truncation or round-off. Because of the limitations of impedance-based methods, knowledge-based methods such as fuzzy logic [7], artificial neural network (ANN) [ ], and support vector machine (SVM) [ ] are focused on. Fuzzy logic-based fault type classification was proposed in [7], which uses higher order statistics to extract the features of a fault signal. But the limitation of fuzzy logic is in determining the global minimum using fuzzy membership functions. ANN-based fault classification was proposed in [], which identifies the fault type and the faulty section. The method was tested only for and not on all 7 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.

2 S. S. GURURAJAPATHY ET AL. fault types. A combination of ANN and SVM for locating fault is presented in [9,] to identify the fault type and distance. A hybrid algorithm using a two-stage approach using SVM was proposed in []. The method utilizes the fundamental voltage and current components for fault location. The method in [] was proposed for fault type classification using SVM. Two approaches using SVM and extreme learning machine (ELM) are given in [3], which identify the faulty phase, fault type, and the fault distance. The methods are compared, and it is concluded that SVM is more accurate than ELM for faulty phase and fault type whereas their performance is almost similar for fault distance calculation. A combination of wavelet transform and SVM is used in [] to locate the fault. The wavelet transform is used to extract high-frequency components of voltage and current signals. The parameters are optimized, and SVM is established for classifying fault type and locating earth fault. A protective scheme using the S-transform (ST) and SVM was proposed in [5]. The method uses three-line voltage and current signals and zero-sequence current for fault type and faulty section identification. From the past research, it is noticed that most of the work finds the fault type or the fault distance separately. Limited research has been reported on finding the fault type, the faulty section, and the distance []. However, the method in [] was implemented for a transmission system. Different from a transmission system, the distribution system has more complex topological structures with multiple laterals. Considering this limitation and opportunity of improvement, this work identifies the fault type, faulty section, and fault distance for distribution systems. The proposed method uses SVM classification and regression analysis for locating the fault. It follows three steps to locate a fault. First, the fault type is identified utilizing the multiclass support vector classification (SVC). Second, the faulty section is identified by calculating the fault resistance, finding possible faulty sections and ranking the possible sections. Finally, the fault distance is identified using support vector regression (SVR) analysis. The rest of the paper is organized as follows. The proposed method is described in Section. Section 3 presents the test distribution system, Section presents the test results of the proposed method, and Section 5 draws conclusions of the proposed work.. Proposed Method The illustration of the proposed method is shown in Fig.. s are simulated at all nodes of distribution system for various fault resistances, and the voltage sag magnitude and are recorded in a database. The actual voltage sags during the fault are the input to the SVM. The proposed method identifies the fault type using the oneversus-one concept of multiclass SVC, which finds an optimal hyperplane to separate the voltage samples based on the fault type. The output is a binary outcome, which yields the type of fault. Once the fault type is identified, the faulty section needs to be identified. The identification of the faulty section involves fault resistance estimation using SVR. The possible faulty sections are identified by comparing the actual voltage sag due to the fault with the simulated voltage sag in the database, and the possible faulty sections are ranked. The fault distance for possible faulty sections is identified using SVR analysis. SVR maps the input to real numbers based on the training data. The output is real point floating numbers, which yields the fault distance... Database establishment The process of establishing database is summarized by the following steps:. SLGF is simulated at each nodes of the system with resistance. The voltage sag magnitude and are recorded from the measurement node.. Step is repeated for fault at all n nodes of the system. 3. Steps and are repeated for fault resistance of,, and.. Steps 3 are repeated for other fault types of,, and... type identification The proposed method identifies the fault type among SLGF (class ), (class ), (class 3), and (class ) using multiclass support vector classification. Each fault type represents classes from to. The classification utilizes the one-versus-one concept, which constructs SVM classifiers for all pairs of classes. In general, for a total of k classes, SVC constructs a total of k(k )/ pairs, and a binary SVM problem is solved for all pairs. The voltage sag database is trained using a radial basis function (RBF) kernel and SVC finds an optimal hyperplane to separate the voltage samples according to their classes, such that the separation between the classes is a maximum. Figure illustrates fault type classification using SVC. The voltage sag phase and during fault are given as input to SVC. Since there are four classes, k =, a total of six pairs of binary SVC (SLGF/, SLGF/, SLGF/, /, /, and /) are solved. The fault type is identified based on the Max wins strategy. If the maximum output is obtained with SLGF among six pairs of binary SVC, the fault type is finalized as SLGF..3. y section identification Each process of faulty section identification is explained in detail as follows..3.. resistance estimation resistance is estimated using SVR analysis. The voltage sag (V f, φ f ) during a fault is assigned as the input to SVM process. If the fault type is SLGF, then SVR is chosen, which has the training voltage sag data for SLGF at various fault resistance. Similarly, SVR is used for, SVR3 for, and SVR for. The corresponding output is the fault resistance (R f ). The illustration of the fault resistance estimation is given in Fig Selection of possible faulty section The estimated fault resistance R f is compared with the fault resistance in database (for which the voltage sag is simulated). If the estimated fault resistance lies between two adjacent fault resistances in database R f (x) andr f (x + ), then its voltage sag data is used for faulty section identification, as shown in : R f (x) <R f < R f (x + ) () A faulty section is identified by comparing the voltage sag phase and of adjacent fault resistance with the actual voltage sag phase and [7]. For analysis, a faulty section between nodes i and j, with a fault resistance between R f (x) andr f (x + ) is taken as an example. Its voltage sag variation is shown in Fig.. The voltage sag magnitude for R f (x) at section i is (V R f (x) i,min, φr f (x) i,min ) and at section j is (V R f (x) j,min, φr f (x) j,min ). Also, the voltage sag magnitude for R f (x + ) at section i is, φ R f (x+) i,max ) and at section j is (V R f (x+), φ R f (x+) ). (V R f (x+) i,max The shaded area represents the search boundary at resistance values R f (x) andr f (x + ). It is noticed that the actual voltage sag (V f, φ f ) during the fault is not within the shaded area. This problem can be addressed by identifying the minimum and maximum voltage sag profiles of adjacent fault resistances R f (x) and R f (x + ) in a section. The corresponding section is chosen as the faulty section as mentioned in and 3: V R f (x) i,min φ R f (x) i,min V f V R f (x+) () φ f φ R f (x+) (3) 3 IEEJ Trans 3: 37 5 ()

3 FAULT LOCATION IN AN UNBALANCED DISTRIBUTION SYSTEM Data base establishment Input Real time measurement (V f,ϕ f ) SLGF type identification Process ysection estimation distance calculation Output type y section distance Fig.. Illustration of the proposed method Start Input data(v f, ϕ f ) Support vector classification SLGF/ SLGF/ SLGF/ / / / Max wins statergy type Fig.. type identification Fig.. Voltage sag variation for a section at two different resistances V f ϕ f S V M SVR/SLGF SVR/ SVR3/ Calculated fault resistance R f line joining (V R f (x) i,min,φr f (x) distance d s is given in : i,min )and(v R f (x+),φ R f (x+) ). The shortest d s = sin θ BC C () SVR/ where Fig. 3. resistance estimation using SVR θ BC = cos [(B + C A )/( B C )] (5).3.3. Ranking analysis The most probable faulty section can be obtained using ranking analysis, which identifies the shortest distance from the fault point (V f, φ f ) to the line joining the voltage sag data of a line section (selected from database). This process is important since faulty section identification generates multiple sections. This happens because of parallel branches and sub-branches in the distribution system, which makes many sections to overlap and cause equivalent electrical distance from the measurement location. Figure 5 illustrates the ranking analysis with the assumption of two possible faulty sections: section (between nodes i and j ) and section (between nodes p and q). The voltage sag data for section is (V R f (x) i,min, φr f (x) i,min ) and, φ R f (x+) sag for section is (V R f (x) (V R f (x+) ) as mentioned in and 3. Similarly, the voltage p,min, φr f (x) p,min ) and (V R f (x+) q,max, φ R f (x+) q,max ). d s and d s represent the shortest distance between the fault point and the line joining sections and. The shortest distance among possible faulty sections and is chosen as the most possible faulty section as in []. The shortest distance d s at section can be calculated by identifying the perpendicular distance from point (V f, φ f )tothe B = A = C = (φ f φ R f (x) i,min ) + (V f V R f (x) i,min ) () (φ R f (x+) φ R f (x) i,min ) + (V R f (x+) V R f (x) i,min ) (7) (φ R f (x+) φ f ) + (V R f (x+) V f ) () The calculation is repeated for all possible faulty sections. The line section with the shortest distance will be determined as the first rank of possible faulty section, followed by the second rank, and so on. Based on the ranking of the section, the first rank faulty section will be inspected first. In case the first section is incorrect upon inspection, the second possible section will then be inspected. This process is repeated for the next section until the actual faulty section is found. 39 IEEJ Trans 3: 37 5 ()

4 S. S. GURURAJAPATHY ET AL. Fig. 5. Rank analysis Training input Table I. Training data for fault distance estimation Training output V R f (x) i,min, φr f (x) i,min, R f (x) V R f (x) j,min, φr f (x) j,min, R f (x) L V R f (x+) i,max, φ R f (x+) i,max, R f (x + ) V R f (x+), φ R f (x+), R f (x + ) L using SVR for fault distance estimation. The training data is represented in Table I. For each possible faulty section, the voltage sag data from database is analyzed individually and trained using SVR. Here, L represents the length of the line section. The illustration for fault distance estimation is shown in Fig.. Voltage sag data during fault and the estimated fault resistance are given as input to SVR. If the fault type is SLGF, SVR is utilized for fault distance estimation, which has the training data for SLGF at resistance R f (x) andr f (x + ). Similarly, SVR is used for, SVR3 for, and SVR for. SVR maps the input data with the training data (as shown in Table I) for finding the fault distancef d. V f ϕ f R f S V M SVR/SLGF SVR/ SVR3/ SVR/ Fig.. distance estimation Calculated fault distance f d.. distance estimation distance is estimated using SVR analysis. The voltage sag data of the line section from adjacent fault resistances R f (x) andr f (x + ) are trained 3. Case Study A 5-kV power distribution system from SaskPower, Canada, is considered for testing the proposed method, as shown in Fig.7. The system consists of one source of 5 kv representing the grid, single-phase laterals, and three-phase laterals [7,9]. The distribution system has a total of line sections and nodes. A node number is shown on the line, and the line section length is shown in brackets (in km). The loads are shown in rectangular boxes (kilovolt amperes). The line parameters are given in Table II. The voltage sag data is recorded at the measurement node nearer to node. The lowest voltage sag magnitude and were used to locate the fault. The distribution system is modeled using the PSCAD power system simulation software. The cables are modeled as constant 5 kv Measurement node (.) (.3) (.3) (.3) (.3) (.3) (5.5) (.)(.5) (.) [I] [I] [I] [I] [I] [I] [II] [II] [II] [I] a a b b c (.) [I] 5 5 [III] 5 [III] Motor 5 [III] 5 [III] 3 [III] [III] 7.5 (.) 7 5 [III] (.) (.) [III] (.) (.) (.) 5 (3.9) (3.9) (.) [III] Fig. 7. Schematic diagram of SaskPower Distribution system IEEJ Trans 3: 37 5 ()

5 FAULT LOCATION IN AN UNBALANCED DISTRIBUTION SYSTEM Table II. Line parameters /km Line type Z Z,Z [I].55 + j j.5 [II].79 + j j.539 [III] j j Table III. Training and testing data for the proposed method Training details Testing details Voltage sag data 33 point At node At middle of line section resistance,,,, 3, Fig.. Hyperplane for SLGF/ impedance load using the PI model. The voltage sag database is created by simulating faults at all nodes of the distribution system. The performance of algorithm is tested for various fault types such as SLGF,,, and and for various resistances of, 3, and 5. The measured voltage sag data at the fault is analyzed using the MATLAB programming code Results The training and test data for locating the fault are shown in Table III. For training purpose, simulations were performed for fault at the nodes of the distribution system at,,, and resistance. The test location for the fault is chosen at the mid-point of a line section. The proposed method was also tested for three different fault resistances of, 3, and 5... Identification of fault type SVC is trained with 33 voltage samples using the RBF kernel for classification of four classes. After training, SVC identifies support vectors and hyperplane for fault type classification. For analysis, faults at section 7 for various fault resistances of, 3, and 5 are considered. The results are analyzed for all six pairs of combinations and are tabulated in Table IV. The support vectors and the hyperplane for fault type identification are shown from Figs to 3. The x-axis represents the phase voltage in p.u., and the y-axis represents the phase of voltage sag data. Figure shows the classification of voltage sag data Fig. 9. Hyperplane for SLGF/ between SLGF and. It shows the voltage sag data of SLGF (class ) in red and (class ) in green, and the separating hyperplane. The support vectors are marked as circles, and a total of 55 support vectors are identified by mapping using the kernel function for SLGF/ identification. Figure 9 shows the classification of voltage sag data between SLGF and. It shows the voltage sag data of SLGF (class ) Section 7 resistance ( ) SLGF vs. Table IV. type identification for line section 7 SLGF vs. SLGF vs. vs. vs. vs. SLGF SLGF SLGF SLGF SLGF 3 SLGF SLGF SLGF SLGF 5 SLGF SLGF SLGF SLGF SLGF SLGF 3 SLGF 5 SLGF Output IEEJ Trans 3: 37 5 ()

6 S. S. GURURAJAPATHY ET AL Fig.. Hyperplane for SLGF/ Fig.. Hyperplane for / Fig.. Hyperplane for / Fig. 3. Hyperplane for / in red and (class 3) in green, and the separating hyperplane using 3 support vectors. Figure shows the classification of voltage sag data between SLGF and. The voltage sag data of SLGF is class and is class. A total of 3 support vectors are identified using the kernel function. Figures and show the classification of voltage sag data between / and /. The voltage sag of is represented as class, as class 3, and as class. A total of 7 support vectors are identified for / and 7 for / classification. Figure 3 shows the classification of voltage sag data between and with 3 support vectors identified by mapping using kernel function. Test cases were carried out for fault at mid-point of all line sections for fault resistances of, 3, and 5. Test results of SVC gives % accuracy in classifying the fault type for all test cases... y section identification... resistance estimation For analysis, a fault at the mid-point of line section 7 is selected. The actual and calculated fault resistances are shown in Fig.. The analysis was carried out for all fault types (SLGF,,, and ). It can be seen that the calculated fault resistance is close to the actual resistance 5 3 Actual fault resistance Calculated fault resistance SLGF Fig.. resistance using SVR fault resistance in all cases. The performance of fault resistance estimation is analyzed using test cases for fault at the midpoint of all line sections. SVR finds the fault resistance with a maximum deviation of ±3 in all the test cases.... Ranking The possible faulty sections are selected and ranked using shortest distance approach. For analysis, SLGF at the mid-point of nodes 7 and is considered. The possible faulty sections, calculated shortest distance, and its corresponding ranking are given in Table V. The test case is repeated for various IEEJ Trans 3: 37 5 ()

7 FAULT LOCATION IN AN UNBALANCED DISTRIBUTION SYSTEM resistance ( ) Table V. Ranking of possible faulty sections Possible faulty sections Shortest distance Rank Node 5.59 Node Node Node 7. 5 Node 5. Node Node. 3 Node Node No. of possible sections SLGF Rank Rank Rank 3 Rank Rank 5 > Rank 5 Fig. 7. Ranking analysis at 5 resistance sections possible of No. No. of possible sections SLGF Rank Rank Rank 3 Rank Rank 5 > Rank 5 Fig. 5. Ranking analysis at resistance SLGF Rank Rank Rank 3 Rank Rank 5 > Rank 5 Fig.. Ranking analysis at 3 resistance resistances of, 3, and 5. The higher ranking of faulty section 7 is due to the presence of multiple laterals, which makes the voltage sag pattern overlap with other sections. Also, the ranking of sections is due to the effect of fault resistance, which causes the selection area of the possible sections to overlap with the selection area of other sections. The overall ranking performance for fault at mid-point of all line section is given in Figs 5 7. The x-axis represents the rank and the y-axis represents the number of sections identified. Figure 5 shows the ranking performance at resistance for SLGF,,, and. The first rank is achieved in sections for SLGF and, sections for, and 5 sections in. Most of the other sections are identified within to 5 ranking. Figure shows the ranking performance at 3 resistance for SLGF,,, and. It can be noticed that most of the faulty sections are identified within the first five rankings. The first rank is achieved in four sections (SLGF), six sections ( and ), and eight sections (). Figure 7 shows the ranking performance of all fault types at 5 resistance. A maximum of eight sections are identified in first ranking for SLGF. A maximum of seven sections are identified in second ranking for, eight sections in second ranking (), and ten sections in second ranking for. type Table VI. distance calculation at line section 7 resistance ( ) Polyhomog kernel distance (km) Multiquadric kernel RBF kernel SLGF Percentage error (%) 3 SLGF y section Fig.. distance at resistance It can be seen that most of the sections are ranked within the first five ranking, and limited sections of or go beyond ranking 5. These sections are identified with a maximum of eight ranking. In actual practice, the fault location has to be pin-pointed, and the proposed method yields the prior possible faulty sections through ranking analysis with five possible faulty sections in most sections..3. distance calculation The test results are analyzed for kernels functions such as Polyhomog, multiquadric, and RBF functions. Table VI shows the results of fault at the midpoint of the line section 7 at a distance of.575 km for various fault types and resistances. It can be noted that the RBF kernel yields more accurate results compared to other kernel functions. To evaluate the performance of fault distance using SVR, fault at the mid-point of all line sections are considered. The test cases are repeated for various resistances of, 3, and 5. Figure shows the calculated fault distance of SLGF,,, and at a resistance of. The percentage error is calculated with respect to the corresponding line section. A maximum error of % is obtained for SLGF in the line section. In this, the calculated fault distance is.59 km while the actual distance is 3 IEEJ Trans 3: 37 5 ()

8 S. S. GURURAJAPATHY ET AL. (% ) Percentage error SLGF y section in Canada, which shows that the proposed method is able to identify all types of fault and also locate the faulty section and the fault distance. The fault type has been identified using SVC. The possible faulty sections are identified and ranked within first five ranks in most cases. The RBF kernel function is chosen for fault distance calculation after comparison with Polyhomog and multiquadric kernel functions. A maximum error of 5% is obtained in the analyzed test cases of fault distance calculation. Thus the proposed method obtains reliable output in identifying the fault type, faulty section, and the fault distance for various values of fault resistance. Percentage error (%) Fig. 9. distance at 3 resistance SLGF y section Fig.. distance at 5 resistance.7 km. The absolute error is. km, which is small compared to that of the whole distribution system. The maximum error of is.37% (at section 5 ), of is.39% (section 9 ), and of is.93% (section 7). Most of the other fault distances are calculated at a lower error percentage of.%. Therefore, the proposed method has managed to identify the fault distance with greater accuracy. The percentage error of the calculated fault distance at 3 resistance is shown in Fig. 9. A maximum error of 5% is obtained in at line section. The actual fault distance is.95 km, while the calculated fault distance is.5 km. The absolute error is. km, which is small compared to that of the whole distribution system. The maximum percentage error using SLGF is 3% (section ), is % (section 3), and is 9% (section ). Figure shows the percentage error at a fault resistance of 5. The maximum percentage error at SLGF is % (section 5), is % (section ), is % (section 3 ), and is 3.7% (section 9). The absolute error of at section 9 is. km, which is small compared to that of the whole distribution system. Overall, it can be noticed that the proposed method gives a high accuracy at resistance for all types of fault. The error percentage is calculated within a section, and a maximum percentage error of 5% is obtained at (line section at 3 resistance) comparing all test cases, in which the absolute error between the actual and calculated fault distance is. km. Hence the fault distance is identified with a maximum error range of ±. km. Meanwhile, most of the other calculated fault distances have lower error percentage in the range.%. Therefore, the proposed method has managed to identify the fault distance with greater accuracy. 5. Conclusion In this work, an approach using SVC and regression analysis for locating fault has been successfully proposed. The performance of the proposed method is tested on SaskPower distribution system Acknowledgments The authors thank the Ministry of Education and the University of Malaya for supporting this work through research grants HIR (H-- D), UMRG (RG35/AET), and FRGS (FP-A). References () Anderson PM. Analysis of ed Power Systems. The Institute of Electrical and Electronics Engineers Inc.: New York; 995. () Girgis AA, Lubkeman DL. A fault location technique for rural distribution feeders. IEEE Transactions on Industry Applications 993; 9:7 75. (3) Choi MS, Lee SJ, Lee DS, Jin BG. A new fault location algorithm using direct circuit analysis for distribution systems. IEEE Transactions on Power Delivery ; 9:35. () Choi MS, Lee SJ, Lim S, Lee DS, Yang X. A direct threephase circuit analysis-based fault location for line-to-line fault. IEEE Transactions on Power Delivery 7; :5 57. (5) Filomena AD, Alegre P, Salim RH, Resener M, Bretas AS. Ground distance relaying with fault-resistance compensation for unbalanced systems. IEEE Transactions on Power Delivery ; 3:39 3. () Filomena AD, Resener M, Salim RH, Bretas AS. Distribution systems fault analysis considering fault resistance estimation. International Journal of Electrical Power & Energy Systems ; 33: (7) Pradhan AK, Routray A, Biswal B. Higher order statistics-fuzzy integrated scheme for fault classification of a series-compensated transmission line. IEEE Transactions on Power Delivery ; 9:9 93. () Warlyani AJP, Thoke AS, Patel RN. classification and faulty section identification in teed transmission circuits using ANN. International Journal of Computer and Electrical Engineering ; 3:7. (9) Thukaram D, Khincha HP, Vijaynarasimha HP. Artificial neural network and support vector machine approach for locating faults in radial distribution systems. IEEE Transactions on Power Delivery 5; :7 7. () Jiang JA, Chuang CL, Wang YC, Hung CH, Wang JY, Lee CH, Hsiao YT. A hybrid framework for fault detection, classification, and location part I: concept, structure, and methodology. IEEE Transactions on Power Delivery ; :9 99. () Salat R, Osowski S. Accurate fault location in the power transmission line using support vector machine approach. IEEE Transactions on Power Systems ; 9: () Parikh UB, Das B, Maheshwari R. classification technique for series compensated transmission line using support vector machine. International Journal of Electrical Power & Energy Systems ; 3:9 3. (3) Malathi V, Marimuthu NS, Baskar S. Intelligent approaches using support vector machine and extreme learning machine for transmission line protection. Neurocomputing ; 73: 7. () Ekici S. Support vector machines for classification and locating faults on transmission lines. Applied Soft Computing ;:5 5. (5) Coteli R. A combined protective scheme for fault classification and identification of faulty section in series compensated transmission lines. Turkish Journal of Electrical Engineering & Computer Sciences 3; : 5. IEEJ Trans 3: 37 5 ()

9 FAULT LOCATION IN AN UNBALANCED DISTRIBUTION SYSTEM () Ravikumar B, Thukaram D, Khincha HP. Application of support vector machines for fault diagnosis in power transmission system. IET Generation, Transmission and Distribution ; : 9 3. (7) Mokhlis H, Li H. Non-linear representation of voltage sag profiles for fault location in distribution networks. Electrical Power and Energy Systems ; 33: 3. () Awalin LJ, Mokhlis H, Abu Bakar A, Mohamad H, Illias HA. A generalized fault location method based on voltage sags for distribution network. IEEJ Transactions on Electrical and Electronic Engineering 3; :S3 S. (9) Mora-Florez J, Melendez J, Carrillo-Caicedo G. Comparison of impedance based fault location methods for power distribution systems. Electric Power Systems Research ; 7:57. Sophi Shilpa Gururajapathy (Non-member) was born in Tamil Nadu, India, in 9. She received the B.Eng. degree (electrical engineering) in 7 from Anna University, Chennai, India, and the M.Eng. degree in from the University of Malaya (UM), Malaysia. She is currently pursuing the Ph.D. degree at UM. Her research interests include fault location in distribution networks. Hazlie Mokhlis (Non-member) received the B.Eng. degree in electrical engineering in 999, the M.Eng.Sc. degree in from the University of Malaya (UM), Malaysia, and the Ph.D. degree in 9 from the University of Manchester, UK. Currently he is a Senior Lecturer with the Department of Electrical Engineering, UM. He is also an associate member of UM Power Energy Dedicated Advanced Research Centre (UMPEDAC), UM. His research interests include distribution automation and power system protection. Dr. Mokhlis is a member of the IEEE. Hazlee Azil Bin Illias (Non-member) was born in Kuala Lumpur, Malaysia, in 93. He received the B.Eng. degree in electrical and electronic engineering from the University of Malaya (UM), Malaysia, in, and the Ph.D. degree from the University of Southampton, UK, in. Currently, he is a Lecturer with the Department of Electrical Engineering, UM. His research interests include partial discharge modeling and measurements in cavity voids within solid dielectric materials, partial discharge simulation using finite element analysis, and load flow analysis. Ab Halim Abu Bakar (Non-member) received the B.Sc. degree in electrical engineering in 97 from Southampton University, UK, and the M.Eng. and Ph.D. degrees from the University Technology, Malaysia, in 99 and 3, respectively. He has over 3 years of utility experience in Malaysia before joining academia. Currently, he is a Lecturer with the Department of Electrical Engineering, University Malaya, Malaysia. His research interests include power system protection and power system transients. Dr. Halim is a member of the IEEE, CIGRE, and IET. He is a Chartered Engineer. Lilik Jamilatul Awalin (Non-member) was born in East Java, Indonesia, in 977. She received the B.Eng. degree in electrical engineering in 999 from the University of Widya Gama, Indonesia, the M.Eng. degree in from the Institut Teknologi Sepuluh Nopember, Indonesia, and the Ph.D. degree in from University of Malaya. Currently, she is a Senior Lecturer with the International College of University Kuala Lumpur. 5 IEEJ Trans 3: 37 5 ()