Bayesian networks-based approach for power systems fault diagnosis

In this paper, three element-oriented models based on simplified Bayesian networks with Noisy-Or and Noisy-And nodes are proposed to estimate the faulty section of a transmission power system. The three models are used to test if any transmission line, transformer, or busbar within a blackout area is faulty. They can deal with uncertain or incomplete data and knowledge relating to power system diagnosis, so they are flexible. The structures and initial parameters of the Bayesian networks depend on the prior knowledge of the domain experts. The parameters can be revised by using an error back propagation algorithm similar to the back-propagation algorithm for artificial neural networks. The fault diagnosis models do not vary with the change of the network structure, so they can be applied to any transmission power system. Furthermore, they have clear semantics, rapid reasoning, powerful error tolerance ability, and no convergence problem during the diagnosing procedure. Experimental tests show that the approach is feasible and efficient, so the prototype program based on the approach is promising to be used in a large transmission power system for online fault diagnosis.

I. Introduction

Fast and accurate fault section diagnosis is very important for power system restoration after a serious blackout. During past two decades, much research work has been done for estimating the fault section in a power system by using several approaches, such as logic based [1], rule based [2], fuzzy relation based [3], neural network based [4], [12], genetic algorithm (GA) based [5], etc. However, each above approach has its limitations. The logic- and rule-based approaches are difficult to deal with wrong and incomplete operational signals of an electric power system, and are hard to determine the faulty section among the inferred possible faulty sections; the fuzzy approach is very rough on reasoning and may lead to wrong conclusions; the neural and GAs have training problems and are difficult to be applied to a large real power system, because the connective parameters of a trained neural network for fault recognition have a big relation with the structure of the power system chosen as the sampling object.