Identification and Classification of Fault in an EHV Transmission line using S-Transform and Neural Network

Size: px

Start display at page:

Download "Identification and Classification of Fault in an EHV Transmission line using S-Transform and Neural Network"

Eleanore Henderson
5 years ago
Views:

1 I J C International Journal of lectrical, lectronics ISSN No. (Online) : and Computer ngineering 2(2): 80-87(2013) Special dition for Best Papers of Michael Faraday IT India Summit-2013, MFIIS-13 Identification and Classification of Fault in an HV Transmission line using S-Transform and Neural Network N. Roy* and K. Bhattacharya** *Assistant Professor, MCKV Institute of ngineering, Liluah, Howrah (WB), India **Professor of lectrical ngineering Department, Jadavpur University, Jadavpur, Kolkata, (WB), India (Received 15 October, 2013 Accepted 01 December, 2013) ABSTRACT: This paper presents a technique for diagnosis of the type of fault and the faulty phase on overhead transmission line. The proposed method is based on the multiresolution S-transform and Parseval s theorem. S-transform is used to produce instantaneous frequency vectors of the voltage signals of the three phases, and then the energies of these vectors, based on the Parseval s theorem, are utilized as inputs to a Probabilistic Neural Network (PNN). The power system network considered in this study is an HV Transmission line simulated in the PowerSim Toolbox of MATLAB. The fault conditions are simulated by the variation of fault location, fault resistance, fault inception angle. The training is conducted by programming in MATLAB. The robustness of the proposed scheme is investigated by synthetically polluting the simulated voltage signals with White Gaussian Noise. The suggested method has produced fast and accurate results. stimation of fault location is intended to be conducted in future. Index Terms: Fault, Fault identification and Classification, Probabilistic Neural Network (PNN), Signal nergy, S- Transform. method presented in [1] depends on the parameters, I. INTRODUCTION such as fault resistance, fault location and fault inception angle (that are not accessible). Wavelet transforms in conjunction with AI/fuzzy/expert system based techniques have been established to be inherently fast and accurate. A combined method of Discrete Wavelet Transform (DWT) and Fuzzy logic is implemented in [2] for fault A complete power system network involves a large capital investment. To maximize the return on this outlay, the system must be utilized as much as possible within the applicable constraints of security and reliability of supply. Hence, the fundamental parameter of the design of a power system network is that it should operate in a safe manner at all times. No matter how well designed, faults will occur in power system and these faults may cause a risk to life and property. The destructive power of a fault arc carrying a high current is very great. The provision of detection of fault, followed by its subsequent measure of protection, is therefore an integral part of power system design. The detection and classification of a particular fault has become an important area of research work. Soft computing techniques involving Wavelet Transform and S-Transform have become widely popular tools in detection of transients of power system network, diagnosis of power quality,classification of faults and identification of the affected phase. Wavelet Transform is mainly used for feature extraction from the current/ voltage signals. These features are subsequently used for training a neural network or configuring a Fuzzy-expert system. A fault classification technique for distribution systems is proposed in [1] by the modeling of sequence networks. The results have been shown with satisfactory accuracy but the speed of the given method has not been mentioned. It is not clear whether the identification on single circuit transmission lines. An adaptive wavelet algorithm (AWA) is presented in [3] for classifying transitory events caused by faults in transmission lines. The proposed algorithm generates the wavelets on the basis of the fundamental definitions of discrete wavelet transform (DWT) using a classification method based on probability such as the Bayesian linear discrimination analysis. The results have shown a high level of success in the classification, even higher than the approach using mother wavelets pertaining to known families, such as Daubechies. Bayesian linear classifier is once again employed to design a high-speed transmission line protection scheme in [4], where DWT is used to extract features from the transient current signals. Based on wavelet transform and approximate entropy, a novel algorithm to detect and identify faults during power swings is proposed in [5]. Genetic Algorithm (GA) has been established to be an excellent classifier of faults in transmission lines. A new scheme to enhance the solution of the problems associated with parallel transmission line protection is presented [6].

2 This paper demonstrates a novel application of wavelet transform to identify faults in parallel transmission lines. The discrimination scheme which can automatically recognize the type of fault is proposed using GA-ANN. A Wavelet-based protection algorithm is demonstrated in [7] for multi-terminal high voltage DC (MTDC) systems. Wavelet Packet Decomposition gives more refined features from a signal. A fast, accurate and new approach for the protection of thyristor-controlled series-compensated (TCSC) line using wavelet packets transform (WPT) is presented in [8]. Compared to discrete wavelet transform (DWT), continuous wavelet transform (CWT) allows performing a more detailed and continuous analysis of the spectrum energy of the fault transient. Such a feature can be used to detect individual frequencies that characterize the voltage transients generated by the fault. The S-transform is an extension of the wavelet transform and provides excellent time localisation of voltage and current signals during fault conditions. It has satisfactorily and accurately identified power system disturbances in [9-12] and has given promising results even under noisy environment. A new approach of Power Transformer protection strategy is demonstrated in [13] based on S-Transform. A method for distinguishing internal incipient faults in transformers during impulse test is proposed in [14]. The suggested technique is also investigated in noisy condition. This article presents a probabilistic neural network (PNN)] classifier based on ST, where only three features ( i.e. one feature/phase) are required for detecting a type of fault and the affected phase. The voltage signals of the three phases are processed through ST to generate complex S-matrix. nergy of voltage signal of the three phases A, B and C (A, B and C) is calculated from the absolute value of S- matrix. This simple feature extraction is done by programming in MATLAB. Since only three features are required, the memory requirement and computation time will significantly reduce. Moreover, using ST instead of WT will avoid the requirement of testing various families of wavelets in order to identify the best one for detection. The features extracted from ST are given to PNN for training, and subsequently, it is tested for an effective classification. The PNN can function as a classifier [15] and has the advantage of being a fast learning process, as it requires only a single-pass network training stage without any iteration for adjusting weights. Further, it can itself adapt to architectural changes. As the structure of the PNN is simple and learning efficiency is very fast, it is suitable for signal detection problems. Decision Tree (DT) based classifier has shown excellent performance in [16] where a new approach for fault zone identification and fault classification for thyristor Roy and Bhattacharya 81 controlled series compensator (TCSC) and unified power flow controller( UPFC) line using DT is presented. It has been established in [17] that PNNbased method is superior in distinguishing the fault transients than the Hidden Markov Model (HMM) and DT classifier with higher accuracy. The suggested PNN classifier is also tested with simulated voltage signals contaminated with synthetic noise. The scope of the present article is limited to the application of standard ST using Gaussian window. During handling of large size of data the method may be unsuitable due to high computational burden of standard ST. To improve its computational efficiency, the discrete orthonormal Stockwell transform (DOST) is proposed in [18]. A computationally fast version of discrete ST is presented in [19] where cross- differential protection scheme for power transmission systems is proposed. Another disadvantage of standard ST is its poor time resolution during the onset of a transient event. It suffers from poor energy concentration in the timefrequency domain. A fast adaptive discrete generalised ST (FDGST) algorithm is presented in [10] in which the algorithm optimises the shape of the window function for each analysis frequency to improve the energy concentration of the time frequency distribution. Hyberbolic S-Transform (HST) has shown promising results in [20] where a method is proposed to discriminate most important transient fault currents of transformers in which ST uses a hyperbolic window function. The suggested method in this article is capable of fault classification and identification accurately and fast. A fast version of discrete ST remains to be implemented in future scope of work and the results may be compared with those obtained in the present study. The rest of the paper is organized as follows. The simulated power system network is discussed in section II. The features of classification have been described in Section III. Section IV explains the PNN network in detail. The magnitudes of the classification-features have been tabulated in section V. The effect of noise in the suggested technique is studied in section VI. Section VII gives the conclusion of the present work. II. THR PHAS LONG TRANSMISSION SYSTM A simulated model of 735 kv long transmission system from the powsim toolbox of MATLAB-7 is used for the simulation of different types of faults. Fig. 1 shows the simulation model in which a three-phase, 50 Hz, 735 kv power system is transmitting power from a power plant consisting of six 350 MVA generators to an equivalent network through a 600 km transmission line. The transmission line is split in two 300 km lines connected between buses B1, B2, and B3.

3 Roy and Bhattacharya 82 Fig.1: Single-Line Diagram of three phase HV transmission system. The generators are simulated with a simplified Synchronous Machine block. Universal transformer blocks (two -windings and three-windings) are used to model the two transformers. Saturation is implemented on the transformer connected at bus B2. The phase voltage signals measured at the bus B1 have been considered for feature extraction both under normal and fault conditions. The sampling times of all the signals have been taken to be µs and the time period of simulation in MATLAB has been taken up to 0.16 secs. The sampling frequency is 3.2 khz. All the ten types of faults have been initiated at 59 different locations starting from B1, each being 10 km apart. The fault resistances considered for the simulation are 0Ω, 5Ω, 10Ω and 15Ω. The fault inception angles (θ) are 0 0, 45 0 and 90 0 and they correspond to three different time instants of fault initiation in the voltage waveform, i.e. 20ms, 22.5ms and 25ms. The total number of fault simulations made in this system are =7080. Figs. 2 shows that there are very fast highfrequency oscillations in the voltage profile at the onset of the short circuit between phases A, B and C. The amplitude of the shorted phases at the instant of fault occurrence, however, depends on the time of initiation of the fault. The amplitudes of sending end phase voltages of A, B, and C at θ = 0 0 are 145 kv, -425 kv, and 284 kv, respectively. Fig. 2: Voltage profile of three phases for LLL fault occurring at a 150 km from B1 with RF = 0Ω, θ = 0 0. III. FATUR XTRACTION BY S- TRANSFORM In the present analysis, ST has been employed to extract energy values of the voltage signals measured at sending end (B1) transmission line. ST of the voltage signal of each phase would generate a complex matrix. Signal energy is obtained from the absolute value of the S-matrix.

4 Roy and Bhattacharya 83 nergy of a signal is calculated based on Parseval s Theorem. This theorem states that the energy of a signal remains the same whether it is computed in a signal domain (time) or in a transform domain (frequency). 1 T 2 N () [ ] 2 signal= v t dt V n T = (1) 0 n= 0 Thus, based on Parseval s Theorem, the energy of a distorted signal can be given as N n n = S kt, ST NT k = 1 NT 2 (2) where, n =1 N/2, N is the signal length, and n ST NT is the energy vector of the instantaneous frequency at frequency n/nt. In this article, the maximum energy of each instantaneous frequency is calculated using equation (3) M 2 = Y () i (3) i= 1 Where Y(i) = peak amplitude of STA matrix at ith row of STA matrix and M = No. of sampled frequencies in the ST- matrix. ach row of the STA matrix represents the instantaneous frequency. IV. PNN BASD FAULT CLASSIFICATION Fault classification has been achieved by using probabilistic neural network (PNN) architecture specifically designed for classification problems. A useful interpretation of the network outputs under certain circumstances is to estimate the probability of class membership, in which case the network is actually learning to estimate a probability density function (pdf). This is the case of the PNN, a special type of neural network using a kernel-based approximation to form an estimate of the pdfs of categories in a classification problem. This particular type of ANN provides a general solution to pattern classification problems by following the probabilistic approach based on the Bayes decision theory. The network paradigm basically uses the Parzen Cacoulos estimator to obtain the corresponding pdfs of the classification categories. The PNN uses a supervised training set to where T and N are the time period and the length of the signal, respectively, and V[n] is the Fourier transform of the signal. In the case of the ST, the raw signal is decomposed in terms of its frequencies, and thus, a set of decomposed signals at each of the instantaneous frequencies in the raw signal can be obtained from the ST matrix. develop pdfs within a pattern layer. The PNN model is extremely fast and accurate, making it suitable for fault diagnosis and signal classification problems in real time. The structure of the PNN used for fault classification consists of two hidden layers. The first hidden layer is based on the radial basis transfer function, and the second hidden layer is based on the competitive transfer function. The size of the input layer for the PNN is 3 11, where 3 represents the input features (i.e., A, B,and C of the three phases) and 11 is the number of fault types, including the no-fault condition. The key advantages of the PNN are its fast training process; its inherently parallel structure, which is guaranteed to converge to an optimal classifier as the size of the representative training set increases; and its ability to add or remove training samples without extensive retraining. The normal signal and different faulty signals have been categorized as AG (1), BG (2), CG (3),AB (4), BC (5), CA (6), ABG (7), BCG(8), CAG (9),LLL (10) and normal (11) respectively. The size of the output layer of the PNN is Since a particular type of fault is likely to occur at one location at a given point of time, the output of the PNN is the category of that particular fault, i.e., one numeric integer ranging from 1 to 11. Out of 7080 cases of simulations, the features of 10 faulty voltage signals for each type of fault were used for training and the rest were used for testing purposes. V. RSULTS Table 1 shows the values of signal energies of all the fault types that have occurred at distances of 100 km and 200 km from the sending end of the transmission line (B1). The results of the PNN classifier are shown in Table 2. It is noticed in Table 2, that the total no. of misclassified cases is 141. The average accuracy of classification in the present study is 97.9%.

5 Roy and Bhattacharya 84 Table 1: Magnitudes of Signal nergy for different types of fault. D = 100km from B1, RF = 5 Ω, θ = 0 0, A (p.u.) B (p.u.) C (p.u.) Normal AG BG CG AB e e BC e e+00 4 CA e e+00 3 ABG e e BCG e e+00 4 CAG e e+00 3 ABC e e e+00 4 D = 200km from B1, R F =10 Ω, A (p.u.) B (p.u.) C (p.u.) θ = 0 0, Normal AG BG CG AB e e BC e e+00 7 CA e e+00 6 ABG e e BCG e e+00 7 CAG e e+00 6 ABC e e e+00 7

6 Roy and Bhattacharya 85 Table 2: Classification results from PNN. Type pf Fault No. of events PNN Output % correct No. of Correct predictions No. of Wrong predictions AG % BG % CG % AB % BC % CA % ABG % BCG % CAG % ABC % VI. FFCT OF SIGNAL NOIS In order to analyze the performance of the classifier with noisy input signals, simulations were performed with White gaussian noise added to the simulated voltage signals by considering a noise level of 20dB Table 3: Values of Signal nergy for noisy signal. SNR It is noticed from the results that the PNN classifier can classify the faults with a mean accuracy of 96.8%. As an illustration, Table 3 shows the values of signal energies for a particular fault condition in noisy environment. D=100km from B1, R F =5 Ω, θ = 0 0, SNR=20dB A (p.u.) B (p.u.) C (p.u.) AG BG CG AB e e BC e e+ CA e e+ ABG e e BCG e e+ CAG e e+ ABC e e e+ VII. CONCLUSION The feature extraction from the current or voltage signals is the most important part in fault detection using signal analysis. In this paper, the number of features used for detection of is only 3. Since only three, the memory requirement and computation time will significantly reduce. The features can be conveniently obtained from the absolute value of the S- matrix by programming in MATLAB. The type of fault and the affected phase can be accurately identified using the proposed classification scheme.

7 The PNN based classifier has been rigorously tested by simulating fault conditions with different values of fault location, fault resistance, fault inception angle. It has been found that the fault classifier can accurately classify the different types of faults with an average accuracy of 97.9%. The effect of noise on the simulated voltage signals has been also studied and the corresponding mean accuracy is 96.8%. The results obtained indicate that the proposed method is capable of detecting the type of fault and the faulty phase with high accuracy and speed. For a larger system with enormous data, a fast version of discrete ST is recommended to be implemented in future. RFRNCS [1] M. Abdel-Akher and K. Mohamed Nor, Fault analysis of Multiphase Distribution Systems Using Symmetrical Components, I Trans. On Power Delivery, vol. 25, No. 4, pp , Oct [2] P. Chiradeja, C. Pothisarn, Discrete wavelet transform and fuzzy logic algorithm for identification of fault types on transmission line, 8th International Conference on Advances in Power System Control, Operation and Management (APSCOM 2009), 2009, p [3] F.. Pérez,. Orduña, G. Guidi, Adaptive wavelets applied to fault classification on transmission lines, IT Generation, Transmission & Distribution, Volume 5, issue 7, 2011, p [4] F.. Pérez, R. Aguilar,. Orduña, J. Jäger, G. Guidi, High-speed nonunit transmission line protection using single-phase measurements and an adaptive wavelet: zone detection and fault classification, IT Generation, Transmission & Distribution, Volume 6, issue 7, 2012, p [5] L. Fu, Z.Y. He, Z.Q. Bo, Wavelet transform and approximate entropy based identification of faults in power swings, IT 9th International Conference on Developments in Power Systems Protection (DPSP 2008), 2008, p [6] R. Rajeswari, N. Kamaraj, Application of genetic algorithm in artificial neural network for fault classification in parallel transmission lines, IT- UK International Conference on Information and Communication Technology in lectrical Sciences (ICTS 2007), 2007, p [7] K. De Kerf, K. Srivastava, M. Reza, D. Bekaert, S. Cole, D. Van Hertem, R. Belmans, Wavelet-based protection strategy for DC faults in multi-terminal VSC HVDC systems, IT Generation, Transmission & Distribution, Volume 5, issue 4, 2011, p Roy and Bhattacharya 86 [8] S.R. Samantaray, P.K. Dash, Wavelet packetbased digital relaying for advanced series compensated line, IT Generation, Transmission & Distribution, Volume 1, issue 5, 2007, p [9] F. Zhao and R. Yang, Power-Quality Disturbance Recognition using S-Transform, I Trans. On Power Delivery, vol. 22, No. 2, pp , April [10] M. Biswal, P.K. Dash, stimation of time-varying power quality indices with an adaptive windowbased fast generalised S-transform, IT Science, Measurement & Technology, Volume 6, issue 4, 2012, p [11] S. Mishra, C.N. Bhende, and B. K. Panigrahi, Detection and Classification of Power Quality Disturbances Using S-Transform and Probabilistic Neural Network, I Trans. On Power Delivery, vol. 23, No. 1, pp , Jan [12] Ameen M. Gargoom,Nesimi rtugrul, and Wen. L. Soong, Automatic classification and characterisation of Power Quality vents, I Trans. On Power Delivery, vol. 23, No. 4, pp , Oct [13] R. Samantaray, B.K. Panigrahi, P.K. Dash, G. Panda, Power transformer protection using S- transform with complex window and pattern recognition approach IT Generation, Transmission & Distribution, Volume 1, issue 2, 2007, p [14] A. Ashrafian, M.S. Naderi, G.B. Gharehpetian, Characterisation of internal incipient faults in transformers during impulse test using index based on S matrix energy and standard deviation, IT lectric Power Applications, Volume 6, issue 4, 2012, p [15] Tripathy, M., Maheshvari, R. P., and Verma, H. K., Probabilistic neural-network-based protection of power transformer, IT lect. Power Appl., Vol. 1, No. 5, pp , [16] S.R. Samantaray, Decision tree-based fault zone identification and fault classification in flexible AC transmissions-based transmission line, IT Generation, Transmission & Distribution, Volume 3, issue 5, 2009, p [17] N. Perera,, and A. D. Rajapakse, Recognition of Fault Transients Using a Probabilistic Neural- Network Classifier, I Trans. On Power Delivery, vol. 26, No. 1, pp , Jan [18] Y. Wang and J. Orchard, Fast Discrete Orthonormal Stockwell Transform, SIAM. J. Science Computation, 31(5): , 2009.

8 [19] Krishnanand K. R. and P. K. Dash, A New Real- Time Fast Discrete S-Transform for Cross- Differential Protection of Shunt-Compensated Power Systems, I Trans. On Power Delivery, vol. 28, No. 1, pp , Jan Roy and Bhattacharya 87 [20] A. Ashrafian, M. Rostami, G.B. Gharehpetian, Hyperbolic S-transform-based method for classification of external faults, incipient faults, inrush currents and internal faults in power transformers, IT Generation, Transmission & Distribution, Volume 6, Issue 10, October 2012, p

Wavelet Transform for Classification of Voltage Sag Causes using Probabilistic Neural Network

International Journal of Electrical Engineering. ISSN 974-2158 Volume 4, Number 3 (211), pp. 299-39 International Research Publication House http://www.irphouse.com Wavelet Transform for Classification