Identification of Fault Type and Location in Distribution Feeder Using Support Vector Machines

Transcription

1 Identification of Type and in Distribution Feeder Using Support Vector Machines D Thukaram, and Rimjhim Agrawal Department of Electrical Engineering Indian Institute of Science Bangalore INDIA dtram@ee.iisc.ernet.in, rimjhim@ee.iisc.ernet.in Abstract This paper presents a support vector machine (SVMs) approach for locating and diagnosing faults in electric power distribution feeders. The proposed approach is based on the measurements available at the substation and remote terminal units (RTUs). To illustrate the proposed methodology, a practical distribution feeder emanating from 132/11KV-grid substation in India with loads is considered. To show the effectiveness of the proposed methodology, practical situations in distribution systems (DS) such as all types of faults with a wide range of varying fault locations, source short circuit (SSC) levels and fault impedances are considered for studies. The proposed fault location scheme is capable of accurately identify the fault type, location of faulted feeder section and the fault impedance irrespective of SSC level, fault impedance and fault locations. The results demonstrate the feasibility of applying the proposed method in practical distribution automation (DA) system for fault diagnosis. Keywords- distribution systems; fault location; support vector machines E I. INTRODUCTION LECTRIC power distribution system is a complicated network of power systems which comprises of number of radial feeders. These distribution feeders constitute the major link between the power supply utility and consumers, which has to be highly reliable and efficient under normal and contingency condition. The reliability of a distribution system can be improved by proper planning and good maintenance. However, at times, supply interruption is still unavoidable mainly because of fault or overloading in the distribution system consisting of a number of radial feeders. These feeders are frequently subjected to the various types of faults like temporary short circuits, permanent short circuits and open conductor high impedance faults. Detection of high impedance faults (HIFs) on electric distribution system has been one of the most persistent and difficult problem facing the electric utility industry. Consequently, HIFs cannot be detected by conventional over current protection [1, 2]. In recent years, there has been increasing interest in the use of computer aided diagnosis and restoration systems to assist the operations engineer [3]. Various short circuit fault location techniques have been proposed in the literature. However, a survey of previous work has revealed that most of the fault location algorithms were developed for transmission systems and are not suitable for radial distribution network [4]. As electric utilities install monitoring devices on distribution substations and collect data from these devices, new methods are needed to extract useful information from monitored data. Data obtained at the substations can provide valuable information about what is happening on the system, and this will enable the utility to respond to the problems proactively. This objective depends on the success of the distribution automation system and in turn on the availability of reliable measurement database from SCADA systems [5]. An automated distribution system fault location and diagnosis system is described in [6]. The proposed fault location algorithm is based on the integration of information available from distribution recording devices with knowledge contained in a distribution feeder database. Methodologies used to solve the fault location problem in power distribution systems could be divided in two groups: methodologies based on circuital models of the distribution network also known as circuit based models (CBM), and methodologies that use other type of information like weather conditions, fault databases, customer calls, among others, also known as knowledge based methodologies (KBM) [7]. The application of expert systems, fuzzy logic, wavelet, neural networks (NNs) has successfully been applied for fault diagnosis in literature [8-10]. To overcome the line model dependency problem, in [11] artificial neural networks (ANN) based estimator were built to estimate fault locations. This paper describes an automated distribution fault diagnosis system using multi-class support vector machines (SVMs). SVM is a relatively new machine learning method developed by Vapnik [12], which is based on the statistical learning theory. In this paper, SVMs are used as an intelligent tool for identifying the faulted line section that is emanating from a node in conjunction with the fault type and subsequently to find the fault impedance and fault distance from the node. Results on a practical distribution feeder emanating from 132/11KV-grid substation in India are presented to illustrate the proposed method. The paper is organized as follows. Section II introduces the proposed methodology using SVMs. Section III brings in the detail of the Support Vector Machines, multi class classification, and describes the procedure to implement SVM. The test system and the results of the proposed approach are presented in section IV. It gives the brief description of the 1

2 input patterns for different SVM blocks and mainly describes the parameter variation results for both classification and regression. Finally, Section V presents the conclusions. II. PROPOSED APPROACH Distribution automation is a tool which refers to automation of repetitive tasks on the electric utility distribution system and is used to benefit the consumer as well as utility. The status of various switches, RMS values and phase angle of section current and voltage are made available from pole top remote terminal unit to substation remote terminal unit (RTU). Availability of this information at substation RTU greatly helps in location of faults in distribution system. Proposed method is developed using above information to monitor the feeders during fault condition. Feeder Automation SCADA State Estimation Data Base Software Power Flow Figure 1. Automated System Configuration Fig. 1 shows the configuration of the distributed feeder automation with all the facilities as mentioned above. Here it is assumed that all nodes data is available at SCADA through RTUs. Then through SCADA the system data is gathered and stored at central located data base system. Data is then processed via software program (three phase power flow) and short circuit analysis to obtain the required information for distribution dispatch center (DDC). This information is then conveyed to the SVMs which perform the classification and regression task. Suitable control actions are then taken once Three phase Voltages at the substations (V a, V b, V c ) Distribution Dispatch Center Software Short Circuit Analysis SVM (Classification and Regression) Control the faulted line, fault type, and fault impedance value is known. When fault occurs on the system, voltage and current magnitude changes at many places in the network. This paper presents an approach to assist the post-fault diagnosis process immediately after detection of fault. The first goal of our approach is identification of the faulted line section in conjunction with the fault type and consequently the identification of fault location and fault impedance forms the second. Block description of the proposed method using the SVMs is illustrated in Fig 2. The RMS values of three phase voltages and currents during fault, measured with RTUs at the nodes are considered as an input feature vector to the SVMs and the expected output i.e. identifying the fault type, faulted line estimation, fault location on the line emanating from the node and fault impedance value are considered as targets. The relation between the measurements and targets are buildup by SVM, which relates them only in a condition when fault occurs in one of the lines emanating from the node. In the proposed approach identification of fault type and faulted line among the lines emanated from the node is a multi-class classification problem. This is solved as a classification problem using techniques such as support vector classifiers (SVCs), which have well-established advantages over other methods. In Fig. 2, SVM-C corresponds to the faulted line classifier with inherited fault type. After the faulted line is identified, the accurate fault location and fault impedance value is a function approximation problem, which is solved by ε-support vector regression (epsilon-svr) approach. In this regression problem of fault location identification, the block ε-svr uses the fault type information obtained from SVM-C and the same local information available at the substation during the fault. V s 1 V sk Set of three phase Currents in the substation lines (I a, I b, I c ) I L 1 I Lm FEATURE VECTOR ε-svr-1 SVM -C ed Line Classifier ε-svr-2 ε-svr-3 ed Line Number with inherited fault type Triggeer the fault type SVR S-L-G L-L 3-Ph L-L-G ε-svr-4 ε-svr-5 ε-svr-6 ε-svr-7 ε-svr-8 Figure 2. Block description of the proposed method. 2

3 As shown in Fig. 2, the design of the SVR consists of four Support Vector Regression blocks. One of these blocks is triggered based on the fault type identified by the SVM-C. To train each one of these SVR blocks, patterns are generated for each fault type at different locations on the lines with all the parameters varied. The details of training and testing patterns are presented in section 4. In Fig. 2, ε-svr-1, ε-svr-3,ε-svr- 5 andε-svr-7 correspond to the fault locator; ε-svr-2, ε- SVR-4,ε-SVR-6 andε-svr-8 correspond to the fault impedance identifier for four different types of fault. III. SUPPORT VECTOR MACHINES SVM is a relatively new computational learning method based on the statistical learning theory. In SVMs, the original input space is mapped into a high-dimensional dot product space called a feature space, and in the feature space, the optimal hyperplane is determined to maximize the generalization ability of the classifier. The optimal hyper plane is found by exploiting the optimization theory, and respecting insights provided by the statistical learning theory. SVMs have the potential to handle very large feature spaces, because the training of SVMs is carried out in such a way that the dimension of classified vectors does not have a distinct influence on the performance of SVMs, as it has on the performance of conventional classifiers. That is why it is noticed to be especially efficient in large classification problems. This will also benefit in fault classification, because the number of features for fault diagnosis may not have to be limited. The mathematical formulation for support vector classification and support vector regression is given in [13, 14]. A. Multi class classification Basically, there are two types of approaches which can be used for the extension of binary two-class problem to n class problem (first is to modify the design of the SVMs to incorporate the multi-class learning in the quadratic solving algorithm and second is to combine several binary classifiers). In the second case, methods like one-against-all and oneagainst-one have been proposed where typically a multi-class classifier is constructed by combining binary classifiers[15, 16]. In this paper one-against-one method is used for multiclass classification, because of its less training time over oneagainst-all. B. Procedure to use SVM In this section, we will briefly describe the procedures to apply the SVMs in practice for classification and regression task. Scaling: The input and the output data need to be scaled properly for better performance. The main advantage of scaling is to avoid attributes in greater numeric ranges dominating those in smaller numeric ranges [17]. Another advantage is to avoid numerical difficulties during the calculation. Because kernel values usually depend on the inner products of feature vectors, therefore each attribute is linearly scaled to the range [-1, +1] or [0, 1]. In this paper [-1, +1] normalization is used. For the linear scaling method, assuming the maximal and minimal values of the i th attribute as M i and m i respectively, Scaling to [0, 1] means x'=(x-m i )/(M i -m i ) and scaling to [-1, +1] means x''=2(x-m i )/(M i -m i )-1. Choosing the kernel: The next step is to choose an appropriate kernel function. The use of kernel methods provides a powerful way of obtaining nonlinear algorithms capable of handling non-separable datasets in the original input space [18, 19]. Different types of kernels which can be used to train the SVMs are linear kernel, polynomial kernel, radial basis function (RBF) kernel and sigmoid kernel. The RBF kernel nonlinearly maps samples into a higher dimensional space. Hence, unlike the linear kernel it can handle the case when the relation between class labels and attributes is nonlinear. Another reason of choosing RBF kernel is the number of hyper-parameters which influences the complexity of model selection. Hence in this paper, we have used RBF kernel for SVMs training. SVM Model Selection: In any predictive learning task, such as classification/ regression, an appropriate representation of examples as well as the model and parameter estimation method should be selected to obtain a high level of performance of the learning machine. To obtain a good performance, some parameters in SVMs have to be chosen carefully. These parameters include: The penalty term C, which determines the tradeoff between minimizing the training error and minimizing model complexity; Kernel parameters such as gamma (g = 1/2σ 2 ) that implicitly defines the nonlinear mapping from input space to some high-dimensional feature space. An optimum value for the kernel parameter of any specific kernel might be achieved using cross-validation. Training: The next step is to train the SVM using the RBF kernel with optimal parameter, the input and the output (for regression) data. The choice of the appropriate kernel could be influenced by the number of training data. However, in practice, usually crossvalidation technique is the best way to judge the appropriateness of the kernel. Testing: After training, the SVM data can be tested using the validation parameters. IV. TEST SYSTEM STUDIES AND RESULTS The developed algorithm is tested on a practical distribution feeder emanating from 132/11KV-grid substation, which is a part of the three phase 52 bus distribution network in India consisting of three feeders. Each feeder of the system consists of more than one main branch. Figure 3 shows the 19 bus distribution feeder used for analysis. The loads in kva are represented by the distribution transformers at various nodes. Bus number-1 is 11 kv substation bus. The complete analysis is carried out on the feeder where each node is assumed to have a RTU. The data corresponding to the feeder 1 of the practical 52 bus DS is tabulated in Table V of Appendix. 3

4 Substation Bus 1 Bus 4 40 Bus 2 Bus Figure 3. A practical 19 bus distribution feeder in India. A. Input Patterns Bus 3 Bus Bus Bus 8 Bus 9 Bus 10 Bus 11 Bus 13 The data used for training and testing purpose are obtained from RTUs located at different nodes and it is assumed that all nodes data is available at DDC. Considering all 18 nodes of feeder 1 as monitoring nodes, the input patterns to the support vector machine contains these node s three phase voltages (V a k, V b k, V c k ), where k is the number of nodes and set of three phase currents (I a m, I b m, I c m ), where m is the number of all lines emanating from the nodes. Hence the total number of features in the input pattern are 3*(total number of buses) +3*(total number of lines). So input feature space is 111 (three phase-voltages at all 19 buses and 18 currents in three phases of the eighteen lines). While training the SVMs, the target values for classification problem will be the class label and for regression problem, the distance values (finding the fault distance from the node) and fault impedance are considered as the target values. The faulty measurements are simulated considering different types of faults: single line to ground (SLG), line to line (LL) fault, 3-phase symmetrical faults and double line to ground faults on each line of the distribution network. The details of the simulation study for generating the patterns for training and testing of faulted line SVCs during faulted condition are given in Table I. As the proposed method uses only one support vector machine classifier as shown in Fig. 2, the output from the SVM-C is one of the class labels from the set (1, 2,, L; L+1,, 2L; 2L+1,, 3L; 3L+1,, 4L). In the above set, the fault type is inherited with the line number. Lines are labeled from 1 to L for SLG fault, from L + 1 to 2L are for LL fault, from 2L + 1 to 3L are for 3-Ph, from 3L + 1 to 4L are for LLG type of fault. Measurements are taken at each of the above DS conditions. The training parameters during fault simulation are modified as follows: The SSC level is varied over the values of 20, 40, 60, 80, and 100 MVA. 100 Bus Bus 15 Bus 14 Bus Bus 19 Bus 18 Bus 17 The fault impedance values are varied over the values of 0, 25, 50, 75 and 100 Ω. The faults are created at distances of 25% and 75% of their total line length. Overall, the training patterns are generated for 4 types of fault on all the distribution lines considered over 5 SSC levels, 5 fault impedance values and 2 locations. For feeder-1, the numbers of training patterns are 3400(17x5x5x2x4). The testing parameters during fault simulation are modified as follows: The SSC level is varied over the values of 30, 50 and 70 MVA. The fault impedance values are varied over the values of 30, 60 and 90 Ω. The fault location is taken as 50% of their total line length The Test patterns are also generated for all types of fault. The number of test patterns generated for each type of fault is 612 (17x3x3x4). In this paper, LIBSVM is used for training the SVMs in classification and regression modules [20]. TABLE I. DETAILS OF TRAINING AND TEST PATTERNS GENERATION OF THE FAULTED LINE SUPPORT VACTOR CLASSIFIER (SVCS) DURING FAULTED CONDITION Line Variation of parameters Training SSC level :- 20,40,60,80 and 100 MVA :- 0,25,50,75 and 100Ω location :- 25% and 75% of their total line length Testing SSC level :- 30, 50 and 70 MVA :- 30,60 and 90Ω location :- 50% of their total line length The number of classes to be classified by the classifier block (SVM-C) is 17, number of hyperplanes (decision functions) formed are 136, the number of the input patterns for training are 3400 and testing are 612. In case of regression blocks, each block functions as a fault distance locator and fault impedance value identifier for different type of faults. B. Support Vector Classifiers Our primary goal is identification of faulted line section in conjunction with fault type. SVMs training require selection of the cost function (C) and kernel function (φ) parameters, which influence the ensuring model performance. In our simulations, we have considered radial basis function (RBF) as kernel function. Choosing the best parameters for faulted line classification can be time consuming if a systematic approach is not used 4

5 and/or the problem knowledge does not aid for proper selection. Hence, an interactive grid search model selection has been accomplished for SVM-C and the generalized accuracy is evaluated on the training data. The grid search is on for C = 2 0, and for g = 2-7, 2-6, 2-5, 2 6, 2 7. The efficient heuristic way of searching points in that space with small generalization errors will lead to a good understanding of the hyper-parameter space. We can then do a refined search of the (C, g) pairs for proper model selection. The obtained model parameters during grid search are merit listed for selecting the best parameters with highest testing accuracy. TABLE II. THE BEST FIVE RESULTS FROM INTERACTIVE GRID SEARCH SELECTION FOR CHOOSING PARAMETERS FOR FAULTED LINE CLASSIFIER (SVM-C) C g No. of Iters. No. of SVs % Training Accuracy % Testing Accuracy Table II shows the extracted model parameters with their training and testing accuracies, number of iterations for completion of the optimization process, and number of support vectors for the fault type classifier. The highest testing accuracy resulted for SVM-1 is % with extracted model parameters of C =8192.0, g =0.125 and cross validation accuracy of % on the training data. Once the SVM is learned with these parameters, all parameters of the trained SVM are frozen and then used in retrieval mode for testing the capabilities of the system on the data not used in learning. The % testing accuracy is defined by (No. of samples correctly classified*100/total number of samples presented). C. Support Vector Regression Once the faulted line section with fault type is identified from the approach mentioned above, our second goal is to identify the location of the fault from the substation and the fault impedance value. To model the ε -SVRs, RBF kernel is chosen. A series of experiments are run on the ε -SVR module with several values of C (cost coefficient or penalty parameter), g and p (epsilon in loss function) to find which combination of parameters is the best. The value of the penalty parameter range is trailed between 2 0 and Aforementioned, gamma (g) is an important parameter for all the kernel functions. The range of g is varied as [.0001, 3]. TABLE III. TESTING ACCURACY IN IDENTIFYING THE FAULT LOCATION AND FAULT IMPEDANCE IN TERMS OF MSE FROM ε SVR WITH DIFFERENT C, G, AND P VALUES OF FEEDER-1 WITH FAULT AT LINE 8-9 Type C g p MSE SLG e-6 LL e-6 3-Ph e-6 LLG e-6 SLG e-4 LL e-5 3-Ph e-4 LLG e-3 After a series of experiment the best combination of parameters obtained for all the four types of ε -SVR corresponding to different faults for fault location and fault impedance are given in Table III. It gives the information related to the variation of testing accuracy in terms of mean squared error (MSE) for different combinations of the ε -SVR parameters. For illustrative purpose line 8-9 of the feeder is considered. Table IV shows the obtained fault location (%) and obtained fault impedance (in ohms) for fault at line 8-9 of feeder, where actual distance is taken as 50% of the total line length. The fault locations and fault impedance are obtained from the corresponding ε SVR for all types of faults, with the parameter combinations tabulated in Table III. Hence, the parameters considered for ε SVR-1 for SLG fault are C=128, g =0.5 and p = Similarly for ε -SVR-3 (C=256, g =1 and p =0.1), ε -SVR-5 (C=64, g =0.1 and p =0.2) and ε -SVR-7 (C=128, g =0.5 and p =0.001) for LL, 3-Ph and LLG faults respectively. TABLE IV. OBTAINED FAULT LOCATION (%) AND FAULT IMPEDANCE (OHM) IN LINE 8-9 OF FEEDER-1 FOR DIFFERENT FAULT TYPES WITH ε -SVRS SSC (MVA) Actual Obtained Distance (%) Actual Obtained (Ohm) (%) SLG LL 3Ph LLG (Ohm) SLG LL 3Ph LLG

6 V. CONCLUSION The use of SVMs as powerful tool for applications in fault location problems, specific to distribution systems is presented. The proposed approach is illustrated on a practical distribution feeder emanating from 132/11KV-grid substation. Its operation is evaluated by simulating different types of faults under various practical conditions such as different SSC levels, fault impedances and locations. The classification accuracy for fault type with inherited faulted line is found to be 99.39% for feeder-1. The fault location accuracy of the results obtained is found to be in the range of few meters. However, for a change in network configuration, following a contingency, either the SVMs has to be retrained or the SVM trained beforehand for the contingency has to be put into service. If retraining has to be adopted, the time required is not significant with the proposed approach. The test results obtained prove to be encouraging and demonstrate the promising potential of the proposed algorithm for practical use. TABLE V. End Buses of the Line Segment From- Bus To-Bus APPENDIX LINE DATA, LOAD DATA AND INITIAL BUS VOLTAGES OF FEEDER 1 OF THE 52-BUS TEST SYSTEM Line Length (kms) Load at To-Bus Active Power (kw) Reactive Power (kvar) Initial Voltage at To-Bus (Base Case) V p.u. V Feeder o REFERENCES [1] V. N. Gohokar, M. K. Khedkar, s locations in automated distribution system, Electric Power Systems Research, vol. 75, pp , [2] M. Sarlak and S. M. Shahrtash, High Detection in Distribution Networks Using Support Vector Machines Based on Wavelet Transform, IEEE Electrical Power & Energy Conference, pp. 1-6, 6-7 Oct., [3] C. Y. Teo, A computer aided system to automate the restoration of electrical power supply, Electric Power Systems Research, vol. 20, pp , [4] T. Adu, "A new transmission line fault locating system," IEEE Trans. Power Delivery, vol. 16, no. 4, pp , [5] D. Shirmohammadi,W. H. E. Liu, K. C. Lau, and H.W. Hong, Distribution automation system with real-time analysis tools, IEEE Computer Applictions Power, vol. 9, no. 2, pp , [6] Jun Zhu, Member, David L. Lubkeman and Adly A. Girgis, Automated And Diagnosis On Electric Power Distribution Feeders,IEEE Transactions on Power Delivery, 1997, 12, (2), pp [7] J. Mora-Florez, J. Bedoya-Ceballos, L. Perez-Hernandez, Selection of currents patterns using SVMs for locating faults in radial power systems, Electric Power Systems Research, vol. 75, pp , [8] C. L. Yang, H. Okamoto, A. Yokoyama, and Y. Sekine, Expert system for fault section estimation of power systems using time sequence information, Elect. Power Energy Syst., vol. 14, no. 2, pp , [9] T. Dillon and D. Niebur, Tutorial on artificial neural networks for power systems, Eng. Intell. Syst., vol. 7, no. 1, pp. 3 17, [10] H. J. Lee, D. Y. Park, B. S. Ahn, Y. M. Park, and S. S. Venkata, A fuzzy expert system for the integrated fault diagnosis, IEEE Trans. Power Delivery, vol. 15, no. 2, pp , [11] D. Thukaram, H. P. Khincha, and H. P. Vijaynarasimha,"Artificial neural network and support vector Machine approach for locating faults in radial distribution systems," IEEE Transactions on Power Delivery, vol. 20, no. 2, pp , [12] [19] Vapnik, V., Statistical Learning Theory. Wiley, New York, NY, [13] Sastry, P.S., An introduction to Support Vector Machines, Chapter in J.C. Misra (Ed), computing and information sciences: Recent Trends,Narosa Publishing House, New Delhi [14] Burges, C. J. C, A tutorial on support vector machines for pattern recognition, Data Mining and Knowledge Discovery, vol. 2, no. 2, [15] Weston, J., and Watkins, C., Multi-class support vector machines, Technical Report CSD-TR-98-04, Department of Computer Science, Royal Holloway, University of London, Egham, TW20 0EX, UK, [16] C. W. Hsu and C. J. Lin, A comparison of methods for multi-class support vector machines, [17] Chih-Wei Hsu, Chih-Chung Chang, and Chih-Jen Lin, A Practical Guide to Support Vector Classification, [18] Platt, J., Fast Training of Support Vector Machines using Sequential Minimal Optimization, Advance in Kernel Methods: Support Vector Learning, pp , MIT Press, Cambridge, MA, [19] B. Scholkopf, C. Burges, and A. Smola, "Advances in Kernel Methods Support Vector Learning" Cambridge, MA: MIT Press, [20] Chih-Chung Chang and Chih-Jen Lin, LIBSVM: a library for support vector machines,