Implementation of Neural Network for High Impedance Fault Detection

In the thesis the aim was to detect the high impedance fault occurring on radial distribution system using neural network. A multilayer perceptron was used for distinguishing the linear and nonlinear high impedance faults by taking the feature vector as input R.M.S value of third and fifth harmonic components of feeder voltage and feeder current were used as a feature vector obtained by applying the fast Fourier Transformation on the feeder voltage and feeder current.The values of feeder voltage and feeder current are obtained for two kinds of fault cases (i.e. linear and nonlinier) by simulating the model of high impedance fault system. The values of third and fifth harmonics were obtained by applying the Fast Fourier Transformation .RMS values of these harmonics were used to train the Multilayer Perceptron Neural Network for classification of these two type of faults. It consists of total three layers, two hidden layers and one output layer. Each hidden layer consists of four neurons and one output layer consists of two neuron. This network was trained by using the Back propagation algorithm .Many types of back propagation algorithms were tested and it’s found that trainlm and trainbr were classifying the two kinds of fault s more perfectly compared to other algorithms. As well as for selecting the no. of neurons the network is tested for different number of neuron in each layer and it’s found that the network consisting of four neuron in each hidden layer performing well. The network was tested for different transfer function and it was found that it’s performance is good when log-sigmoid transfer function is used in all three layers or when tan-sigmoid transfer function is used by the neuron in two hidden layer and linear transfer function is used by neuron in output layer.