Novel method for power system fault diagnosis based on Bayesian networks

Abstract

Three element-oriented Bayesian networks models are built to estimate the fault section of a power system. Each of them is composed of noisy-or and noisy-and nodes. The three models are used to locate three types of fault elements: transmission lines, transformers and bus bars respectively. The learning algorithm for network parameters is analogous to the back propagation algorithm of neural networks. Taking the sum of the mean-squared error between the expected values and the computed results of target variables as the minimizing optimization function, it adjusts the network's parameters continuously. According to the operation information of protective relays and circuit breakers, fault credibility of elements in the blackout area is calculated based on the structure of the Bayesian network. By comparing the resultant beliefs of possible fault elements, the fault element(s) is identified. The proposed approach can deal with uncertainties in fault section diagnosis, and the models have clear semantics, rapid reasoning, etc. The testing results for a real power system have shown that the fault diagnosis models are correct, efficient and are promising to be used in a large power system for on-line fault diagnosis.