Towards a Power System Fault Classification System: A Rough Neurocomputing Approach

Abstract A rough neurocomputing approach to classifying power system faults is presented in this paper. Preprocessing fault data entails discretization of power system fault data obtained from the Transcan Recording System at Manitoba Hydro. An approach to discretizing power system fault data is briefly described in this article. After preprocessing, rough set methods are used to prepare fault decision tables and generate fault classification rules. Each condition vector contain values of attributes for a new power system fault becomes input to mixture of rough set-based expert networks, where each the processing performed by each “expert” is tailored to a particular fault type. A collection of rough expert networks are connected to what is known as gating network that selects the output of competing expert networks (the winner) as the output of the network. That is, the gating network implements derived fault classification rules and selects the expert network with the best classification (i.e., highest probability of correctly classified fault). The architecture of such a network is extensible inasmuch as a new rough expert network can be added to accommodate the discovery of a new type of power system fault. The basic architecture of a new fault neural classification system is presented. The contribution of this paper is an overview of the basic building blocks in a rough set-based power system fault classification system.