Fault analysis on power systems using wavelet transformed transients and artificial intelligence

Abstract

This project involves an initial theoretical study followed by an analytical and industrial application for fault analysis on typical distribution ring network. Using a hybrid of expert systems [ES] and small modular neural networks, called target neural networks trained to identify one specific condition each time, the concept is to reduce the scale of the problem systematically by using many small nueral modules to improve the overall accuracy. It also proposes the theory of obtaining a recognisable fault signature for each type of component fault such as cable fault using wavelets and to use standard pre-trained neural modules to target trained them to recognise it. This increases the accuracy and reduce the cost of training. This was first tested on sample networks, and was subsequently applied in power systems and an actual distribution ring systems for fault location and identification. The inputs to the expert systems are the circuit breaker and relay status. The inputs to the neural network employ the waveform values of the transient current and voltages after discrete wavelet transformed (DWT) coefficients to obtain their fault signatures during a fault. The ES and the neural networks are independent of each other. They function only to reduce the overall response time and accuracy of the overall diagnostic problem.