POWER SYSTEM TRANSIENTS CHARACTERIZATION AND CLASSIFICATION USING WAVELETS AND NEURAL NETWO

- 66 -
POWER SYSTEM TRANSIENTS CHARACTERIZATION AND CLASSIFICATION USING WAVELETS AND NEURAL NETWORKS
Fan Mo <email: yingfan@ee.umanitoba.ca> Electrical and Computer Engineering, University of Manitoba, Winnipeg, Manitoba, R3T 3M2 Advisor: Witold Kinsner <email: kinsner@ee.umanitoba.ca> Abstracts This paper presents a framework for transients classification using wavelet transform and two specific artificial neural networks, propabilistic neural networks (PNN) and resource allocating neural networks (RAN). One significant feature of wavelet for transients characterization is its time-scale two dimensional but compact representation. One distinguishing feature of PNN and RAN is the ability to adjust their architecture automatically to adapt to new environment quickly and accurately, which makes them the promising candidates for transients classification in power system. Some experimental results have indicated the suitability of this framework for transients classification. I. INTRODUCTION Electromagnetic transients in power systems result from a variety of disturbances on transmission lines, such as switching, lightning strikes, faults, as well as from other intended or unintended events. Such transients are extremely important, for it is at such times that the power system components are subjected to the greatest stresses from excessive currents or overvoltages[KiSh83][MoKi96]. In recent years, an increasing demand has been observed for power system monitoring and transient recording (including faults) combined with analysis, classification, and reporting. One important component of such system is the digital transient recorder, which has been available in some wellequipped substations to record various transients and provide assistant information about the transients for system operators to analyze them and take actions whenever a fault occurs. However, present transient recorders lack classification capability to distinguish “interesting” events from “trivial” events. The desired intelligent transient recorder should have the ability to distinguish and classify different kinds of transients, provide a means to filter out meaningful information needed to assess system performance, classify faults and other disturbances to help identify stability, fault or other power quality issues, and even more importantly to provide a means to look automatically for early anomalies that could cause catastrophic failure. Transients are signals with a finite life, that is, a transient dies to zero in a finite time. Frequency based analysis has been common since Fourier’s time; however, frequency analysis is not ideally suited for transient analysis, because Fourier based analysis is based on the sine and cosine functions, which are not transients. This results in a very wide frequency spectrum in the analysis of transients. A solution to the problem is to appeal to the use of transients to analyze transients. Wavelet transform is the right tool for transients analysis using a small transient - wavelet. The capability of artificial neural networks (ANN) to learn complex mappings, linear or nonlinear, from the input space to output space, while remaining robust in a noisy environment, has resulted in their application to power systems, ranging from fault classification and localization to power load prediction. The fundamental feature of neural networks is the ability of learning, however, in many applications of neural networks in power systems, the learning ability has not been fully realized. Such limitations and corresponding improvements are introduced in part III. II. WAVELET CHARACTERIZATION OF TRANSIENTS Wavelet theory is the mathematics which deals with building a model for non-stationary signals, using a set of components that look like small waves, called wavelets. It has become well known as a useful tool in several fields, including signal processing, image compression, biomedical applications, speech processing, acoustics, numerical analysis[Kins97] [KiLa93]. Given a time-varying signal x(t), a wavelet transform consists of computing coefficients that are inner products of the signal and a family of “wavelets”. In a continuous wavelet transform (CWT), the wavelet corresponding to scale a and location b is 1 t – b - ψ --------ψ ( a, b ) = --------(1) a a where ψ ( t ) is the wavelet “prototype”, usually called a “mother wavelet”. Each wavelet, ψ ( a, b ) , is a scaled (compressed or dilated) and translated version of the same original function ψ ( t ). The DWT of a simulated single-phase fault waveform using Daub4 is shown in Fig. 1. The diagrams show that at scales 1, the instantance of important events (fault occurring, circuit breaker opening and reclosuring) are clearly localized, while at scales 5, the local variations of the fundamental frequency are clearly shown. The tilling of timescale plane by wavelet transform for these two types of transients are shown in Fig. 3 [MoKi97].