Fault Location And Fault Diagnosis In Power System Based On Neural Networks

Fault location and fault diagnosis in power system are important guarantee of dealing with faults rapidly and accurately. They are also vital segments of assuring the safe and steady operation of power system.This paper analyzes the performance of two representational types of traditional neural networks that are back-propagation neural network (BPNN) and Kohonen neural network (KNN). The simulation results show that the BPNN has better self-study abilities and generalization properties than the KNN does. But the BPNN converges slowly and is easy to get into local minimum. For solving the problem a novel wavelet neural network (WNN) based on the sigmoid function is established to select fault phases in this paper. The performance of the WNN is compared with that of the BPNN. The results of comparison show that this WNN is better than the BPNN in performing fault phase selection.Because neural network hardly converge in the situation of mass samples, a block model of training WNN for fault location, which can greatly reduce the number of the training samples of each sub-module and make neural network converge quickly, is presented. Use the WNN of that block training model to perform fault location and analyze the results of locating fault for the foundation of taking that kind of WNN model into practice.Based on WNN, a fault diagnosis algorithm of power system is developed. Then, rough set (RS) theory, which is a novel tool for data mining, is introduced into the WNN to improve its capability in the fault diagnosis. Thus, a stratified WNN model based on RS theory is presented in this thesis, which is used to identify fault elements. The properties of three types of WNN models (the stratified WNN model, the pure WNN model, the WNN model combined with RS theory through a conventional method) have been compared. It can be concluded that the convergence of WNN and the ability to recognize variant samples are clearly improved using the stratified WNN model.