Rough neural fault classification for HVDC power systems

Abstract

This Ph.D. thesis proposes an approach to classify faults that commonly occur in a High Voltage Direct Current (HVDC) power system. These faults are distributed throughout the entire HVDC system. The most recently published techniques for power system fault classification are the wavelet analysis, two-dimensional time-frequency representation for feature extraction and conventional artificial neural networks for fault type identification. The main limitation of these systems is that they are commonly designed to focus on a group of faults involved in a specific area of a power system. This thesis introduces a framework for fault classification that covers a wider range of faults. The proposed fault classification framework has been initiated and developed in the context of the HVDC power system at Manitoba Hydro, which uses what is known as the TranscanTM system to record and archive fault events in files. Each fault file includes the most active signals (there are 23 of them) in the power system. Testing the proposed framework for fault classification is based on fault files collected and classified manually over a period of two years. The fault classification framework presented in this thesis introduces the use of the rough membership function in the design of a neural fault classification system. A rough membership function makes it possible to distinguish similar feature values and measures the degree of overlap between a set of experimental values and a set of values representing a standard (e.g., set of values typically associated with a known fault). In addition to fault classification using rough neural networks, the proposed framework includes what is known as a linear mean and standard deviation classifier. The proposed framework also includes a classifier fusion technique as a means of increasing the fault classification accuracy.