The classification of power system disturbance waveforms using a neural network approach

Owing to the rise in power quality problems, the use of transient recorders to monitor power systems has increased steadily. The triggering strategies used by these transient recorders to capture disturbance waveforms are usually based on the violation of a set of predetermined measurement thresholds. Unfortunately, threshold based triggering strategies are difficult to apply in situations when only waveforms corresponding to a given class of disturbances need to be recorded. This inability of the recorder to automatically discriminate between waveform types tends to burden the user with the task of manually sifting and sorting through a large number of captured waveforms. This paper describes an artificial neural network methodology for the classification of waveforms that are captured, as part of a larger scheme to automate the data collection process of recorders. Two different neural network paradigms are investigated: the more common feedforward network (TDNN), and a modification of that, the time-delay network (TDNN), which has the ability to encode temporal relationships found in the input data and exhibits a translation-shift invariance property. Comparisons of both network paradigms, based on a typical distribution circuit configuration, are also presented