You are on page 1of 5

3rd International Conference on Electric and Electronics (EEIC 2013


Study of a New Approach about Locating Single-phase-to-ground Fault
Bing Jiang, Chuan Zang, Zhiguo Xiao
Jiangsu provincial Key Laboratory of Power Transmission Equipment Technology HoHai University Changzhou, China
Abstract —— First this paper briefly analyzes the electric power transient signal of single phase grounding fault, and then the fault characteristics of zero sequence current and bus voltage is obtained by wavelet transformation. Finally by using genetic algorithm to optimize the original weights of back-propagation neural network, as well as taking the fault characteristics as the input characteristic vector of optimized back propagation neural network, the network is trained and tested. The simulation result shows that the prediction effects and convergence rate of optimized back-propagation network have better performance than traditional back-propagation network. It’s relatively error were less than 3%, independent of fault distance, power supply phase angle and the impact of transient resistance. Keywords--single phase grounding fault; wavelet transformation; optimized back-propagation network; genetic algorithm; fault location;

Jie Yin
East China Power Grid Electric Co. Ltd Suzhou, China


In isolated neutral system, grounding fault current grounds through imaginary earthing neutral of power transformer that provides single phase grounding fault current with pathway. In Fig. 1, grounding fault current is determined by per-phase capacitance to ground. Assuming that Phase A of Line 2 occurs grounding fault, A phase voltage is zero. Then, for non-fault Line 1, A, B, C three phase currents of capacitance to ground are:

I C A1 = jU ' A ωC01 ,



(1) (2) (3)

I CB1 = jU ' B ωC01 ,
I CC 1 = jU ' C ωC01 .
. .



I. INTRODUCTION In power system operation, the probability of single phase grounding fault in small current grounded system is up to 80%. The trend of planning and transforming 10kV power network in many economically developed regions is that the power is supplied mainly by cable, secondarily overhead line. When single phase grounded fault occurs, grounding residual current is very small. And it does not affect the load power supply. Therefore, the system does not immediately trip. But by now, healthy phase to ground voltage is line voltage in small current grounding system and grounding capacitance current has also increased. As a result, intermittent arc occurs easily at fault point and develops into a permanent grounding fault. It can lead that overvoltage reaches values over the normal phase voltage of 6 to 8 times. Its high amplitude and long duration can bring about a serious threat to existing overvoltage protection. Therefore, the basic guarantee to improve reliability of power network and reduce the accident rate is to determine the single phase grounding fault quickly and accurately. When single phase grounding fault occurs, it will generate transient fault current and fault voltage. This paper uses wavelet transform to identify transient fault signals. Genetic algorithm has good global search ability. The capacity of non-linear fitting of neural network is well. These studies establish correspondence between features and fault location, and realize the fault location.

Figure1. Single phase grounding fault in isolated neutral system

Fundamental steady-state zero sequence current in non-fault Line 1 is:
. . . ⎛ . ⎞ 3 I 01 = ⎜ I C A1 + I CB1 + I CC 1 ⎟ ⎝ ⎠


= j 3U 0 ωC01.
For the fault Line 2, A, B, C, three phase currents of capacitance to ground are:


I C A 2 = − j 3U o ω (C01 + C02 ),



(5) (6) (7)

I CB 2 = jU ' B ωC02 ,
I CC 2 = jU ' C ωC02 .
. .



© 2013. The authors - Published by Atlantis Press


k = 1. its Fourier transform is φ (ω ) . .2. x) = f ( x) * φs ( x) = ∫ f (t )φ ( )dt. Nodes from the front layer to the rear layer are linked by the weight coefficients.L . WAVELET TRANSFORM Assuming that φ (t ) ∈ L2 ( R ) ( L2 ( R ) is square integrable real space.   x  z  In Fig.L m. and its direction is opposite non-fault line. 2.2.1. If f (t ) is the periodic signal. y2 . The input signal transfers from the input layer through the hidden layer to the output layer (forward propagation).It can be obtained by formula 10 that wavelet transform value of signal is proportional to its amplitude. IV. and modulus maxima of wavelet transform can be used to determine the signal singularity. If the output layer gets the desired output. (15) X = ( x1 . i = 1. the learning algorithm ends. z are the network input and output vectors and each neuron is expressed by a node. When φ (ω ) is satisfied with the condition: Cφ = ∫ R φ (ω ) d ω < ∞.2. zero sequence current and zero sequence voltage of bus. In formulas. NEURAL NETWORK A. generated by different fault distance. Furthermore. L m. 2. let f (t ) = Af 0 (t ) . if x locates in the neighborhood of x0 and w f ( s0 . (14) ΔV = η (δ y X T )T . The current flowing through the fault point is the sum of non-fault phase currents of capacitance to ground. . dl )T is expected output. ∂w jk Δvij ∝ − ∂E . Under the principle that the errors continue to decrease. these two variables can be considered as characteristic quantities. the line from 1km to 9km. are both different. x0 ) is modulus maxima of wavelet transform.L . and hidden layer may be one layer or a multilayer (in Fig. x)∞A . in increments of 1km. then. bus voltage will have a change in the fault time. o2 . xn )T is input vector. (2) As shown in TABLE I. take three cases that are fault resistance of 50 ohm with phase angle of 0°. it is a single hidden layer). (10) sR s In the above formula. (13) ∂vij By derivation. Back Propagation Neural Network BP(Back Propagation) neural network is one of the most intensively studied and widely used artificial neural model. in other words. 2. The following are specific processes of weight values of each layer (BP learning algorithm): 1 1 l E = (d − O)2 = ∑ (d k − ok ) 2 . W = [ w jk ]m×l is the weight matrix from hidden layer to the output layer and V = [vij ]n×m is from input layer to the hidden layer. from line to bus.L n. . the adjustment formulas of weight values of each layer can be obtained. is signal space of finite energy.Zero sequence current of fault Line 2 can be obtained from the above formula: . O = (o1 . adjustment of weight values is proportional to the negative gradient of the errors.Ll .L .). The network is made up of the nodes of input layer. d 2 . the error signal (difference between the sample output and internet output) adjusts the weight coefficients of the neurons with gradient descent method to decreases the error signal. ( s0 . x. ω 2 (9) φ (t ) is called basic wavelet or mother wavelet. the amplitude of electrical quantities (voltage. its structure shown in Fig. fault d = (d1 . That is to say. The convolution type of wavelet transformation of f (t ) is: 1 x−t w f (s. j = 0. Y = ( y1 . The steps of fault location prediction model of BP neural network are as follows: (1) Use a three-layer BP network. III. x2 . x0 ) . Figure2. current) go through upheaval (partial phases of electrical quantities also change). 2. The learning process of BP neural network is as follows. ol )T is output vector. initialize the network and assign the network parameters and the weight coefficients which should take random numbers. The structure of BP neural network 356 . (8) ⎝ ⎠ Therefore. the zero sequence current of fault line is the sum of all zero sequence currents of non-fault line. (11) 2 2 k =1 E is defined as the output errors of the network and is expanded to the hidden layer and the input layer. ⎛. x) ≤ w f ( s0 . hidden layer and output layer. W f ( s. ⎞ 3 I 02 = ⎜ I CA 2 + I CB 2 + I Cc 2 ⎟ = −3 I 01 . written in vector form: ΔW = η (δ oY T )T . goes backward propagation that through the back-calculation of original join pathway.L . Otherwise. j = 1. ym )T is output vector of hidden layer. Therefore. namely: ∂E (12) Δw jk ∝ − . When the power system fault occurs.

a failure prediction model A can be obtained. and a new fault localization prediction model B will be obtained. genetic algorithm is a global optimization approach. the weights are random. take mutation operation to generate new individual G ' j .5 3 3.5 1 1. and the individuals without crossover operator replicate directly. John Holland introduced and established genetic algorithm at University of Michigan. Then.After achieving the required performance.5 1 1.5 3 Figure4. Practice shows that weights of BP algorithm are random. Here.5 2 4 Epochs 2. In coding. IV. k is the number of nodes in the output layer. i (16) fi is the match value of i , measured by error square sum E. find other ways to optimize the weights. . Therefore.5 3 Epochs 2 2. At each point. Optimized BP neural network When the failure prediction model A is established. (4) For G j with probability Pm. (3) For individuals Gi and Gi +1 with probability pc .E ( i ) 2 (17) E ( i ) = ∑∑ (Vk − TK ) . 36 data as input training samples. it can be concluded that to achieve the same accuracy. SIMULATIONS AND ANALYSIS In this paper. (18) p k Where i = 1.6 times just after optimization. decode the best individuals in the group to get the connection weights of optimized network [8]. before optimization require 3. use genetic algorithm.4. crossover. adjust weight coefficients between hidden layer and input layer. The steps of fault location prediction model of Optimized BP neural network are as follows: (1) Use a three-layer BP network. 10 0 Ps = fi ∑f i =1 N . The trained neural network converges faster. Select individual network by probability of formula: (5) Insert the new individuals into the population P and calculate their evaluation function. go to step (3). N . 2. Fig. such as . take crossover operator to generate new individuals G 'i and G 'i +1 . mutation and so on to evolve from generation to generation. Under each standard.5. After building the model B.resistance of 500 ohm with phase angle of 0°and fault resistance of 50 ohm with phase angle of 45°as horizontal standards. construct single phase grounding fault model of 10kV isolated neutral system with Simulink in MATLAB. crossover probability Pc. Otherwise. The optimized training error curve Contrasting Fig. encode with real numbers and initial population. Best Training Performance is NaN at epoch 4 Train Best Goal 10 Mean Squared Error (mse) -1 10 -2 10 -3 10 -4 10 -5 10 -6 0 0. as shown in Fig. with low learning efficiency. slow convergence rate. hidden layer and output layer. 357 . Therefore. shown in Fig. that the training error curve under the standard of the phase of 45°and fault resistance of 50 ohm can be archived. In the mid-1960s. initialize population P including cross scale. (2) Calculate and sort the evaluation function of each individual. Based on natural selection and genetic theory. based on the research of others. (4) Repeat step (2) and (3). and ultimately get the optimal solution. follow four steps of model A. namely: f (i ) = 1 . (6) If find the satisfactory individuals. hidden layer and output layer. According to the values calculated by fitness function. easily failing into local optimum and other shortcomings. Based on the above steps. Possible solution of the problem space is seen as the individual that is similar to the chromosome in the group and each individual is encoded into a string form. Training error curve before optimization B. (3) Based on error back propagation rule. 9 lines distance as output vector. 3 is a training error curve by the standard of the phase of 45°and fault resistance of 50 ohm.3 and Fig. genetic algorithm is applied to train the learning of neural network weights.5 4 Figure3. get zero-sequence current of the fault line and the zero-sequence voltage of bus and the real and imaginary parts of modulus maxima by equation (10)(Morlet complex wavelet) after normalization. until the prediction error satisfying the condition or the training reaching the required number of times. and predict the location by fault localization prediction model B.L. is the chromosome number. calculate the predicted values for each layer and compares the values with the true values to obtain the error of network output. simulate biological evolution that contains selection.4 Best Training Performance is NaN at epoch 3 10 -1 Train Best Goal Mean Squared Error (mse) 10 -2 10 -3 10 -4 0 0. p is the number of learning sample and Tk is the teacher signal.2 times. mutation probability Pm and weights between input layer and hidden layer. and complex network better reflects the superiority of genetic algorithm. then end.

Brouaye F. it avoids to trap into a local minimum, accelerates the speed of network processing and provides new impetus for intelligent diagnosis. Meanwhile.011 Relat ive error (%) 1. Transactions of China Electrotechnical Society. use testing samples for testing and get the test results of the fault line by anti-normalization to calculate the percentage of relative error. Analysis of Real Wavelet and Complex Wavelet Transform in Extracting PD Signal Feature[J]. 100km.05 The lengths of three lines are respective 20km.44 5. worthy of further study. Singularity Detection and Processing with Wavelets[J].21 0 5.026 0. Simulation model of 10kV single phase grounding fault 输出误差 0. Mallat S. Cui X M.33 4 2.47 5 3.941 0.000 0. International Journal of Intelligent Systems.mat file. the testing distances of fault line with A model are obtained. 11(3): 1301-1308.994 0. 1992. Dong T Z. 8(4):539-567.948 2. Wavelets: A New Tool For the Resonant Grounded Power Distribution Systems Relaying[J]. Pan H.11 7 7.73 4. 27(2): 156-161. get zero-sequence current of the fault line and the zero-sequence voltage of bus.01 7 8. Combined with wavelet transform,improved BP network which has good nonlinear mapping capabilities of input and output. Li X J.988 0. that is. The results are shown in TABLE I. 0. Computer Application.000 0.08 9 5. Application of BP neural network based on genetic algorithm[J]. REFERENCES [1] Chaari O. Hwang W L.998 0. Meunier M.15 -0. Guo J H.25 9 7. very suitable for the analysis of power system transient process.05 -0. in increments of 1km. Yao X. which the line is from 1km to 9km.14 7 2. At each point.020 0.996 0. These three criteria are standard 1 and standard 2. fault resistance of 500 ohm with phase angle of 0°and fault resistance of 50 ohm with phase angle of 45°as horizontal standards. is able to complete the fault classification and location. fault resistance and phase angle. Lv J. the horizontal axis is the fault line distance and the vertical axis is the output error (the difference between testing distance and fault line distance).17 0 -0.54 2.6. Take three cases that are fault resistance of 50 ohm with phase angle of 0°. TABLE I Testing results and relative errors under various standards Fault Line dista nce 1 4 2 3 3 4 4 6 5 3 6 9 7 3 8 0 9 0 8.15 优化前的误差 优化后的误差 Figure5. After training the network. And it can be drawn by the horizontal standards that this locating method is [7] [8] 358 .009 0.002 0. A Review Of Evolutionary Artificial Neural Networks[J].985 0.17 7 1.22 7 7.13 1 9. Normalize these data as training samples and put them into neural network that is optimized by genetic algorithm.989 0.984 0. 2009.It can be drawn that the output error of optimized BP neural network is closer to zero. the testing results is closer to the sample values.13 4 5.994 0. Lv Y Z.001 Relat ive error (%) 0. Under three standards.980 0.10 8 8.03 2.04 Test distanc e of standar d1 1. 1993. Journal of System Simulation. 100km.995 0. combined with the improved BP network input and output good nonlinear mapping ability.870 1. 38(2): 617-643. Testing results shows that this locating method is basically not affected by fault position.1 -0. Li X.08 6 6.006 0. 21(15): 4903-4907. genetic algorithm can optimize the learning rules and structure of neural network. Research of Fault Detection in Transmission Line based on Neural Network[J].14 8 4.06 Test distanc e of standar d2 1. 1996.58 Test distanc e of standar d2 1.001 0.15 3 3. The compare results are shown in Fig.021 Relat ive error (%) 2. Interharmonic Detection Algorithm Based on Amplitude and Phase Information of Morlet Complex Wavelet [2] [3] [4] [5] [6] TABLE I illustrates that wavelet transform and optimized neural network are good for fault location. At the same time.30 3. IEEE Transactions on Power Delivery.08 1 4. 2005. Take the first line for experiment. fault resistance and phase angle. In addition. 19(7): 90-94.000 0. put the file into workspace of MATLAB and get the real and imaginary parts of modulus maxima by Morlet complex wavelet.24 0 7. Sun C X. 2009.2 1 2 3 4 5 6 7 8 9 样本 Figure6. 25(12): 2777-2779.989 0. relative errors of the testing distances are less than three percent. IEEE Trans on Inform Theory. Power System Faults Detection Based on Singularity Detection of Wavelet[J]. Kou Y S.1 0.993 0. to complete the classification and location of the fault. Errors in the testing results before and after optimization V. 2004. CONCLUSION Wavelet transform as a modern signal processing. convert them into . Journal of Jilin University (Information Science Edition).A B C a b c Distributed Parameters Line1 Three-Phase Series RLC Load1 A B C A B C Three-Phase Fault Three-Phase V-I Measurement1 A B C A B C a b c A B C a b c Distributed Parameters Line2 Three-Phase Series RLC Load2 A B C Three-Phase Source Three-Phase V-I Measurement Three-Phase V-I Measurement2 A B Continuous powergui C a b c Distributed Parameters Line3 Three-Phase Series RLC Load3 A B C Vabc From Add Scope Three-Phase V-I Measurement3 Scope1 Scope4 1 Multimeter basically not affected by fault position.2 0. Standard 3.

Nanning: Guangxi University.Chinese Journal of Scientific Instrument. 28(11): 2103-2107. 2007. 31(10): 49-53. Li J K. Computer engineering & Software. Tan A L. Wang Z L. Proficient in MATLAB Neural Network[M]. 2009: 317-325. 2009. 2010. the Research of Single-Phase-Ground Line Selection and Fault Location based on Wavelet Packet Neural Network in Distribution System[D].[9] Transform[J]. Zhao Y. Study on Genetic Wavelet Neural Network based Multi-sensor Information Fusion Technique[J].Beijing: Publishing House of Electronics Industry. [10] Zhu K. [11] Gao M J. 359 .