Identifying precursors to frequency fluctuation events in electrical power generation data

Abstract

To predict an occurrence extraordinary phenomena, such as failures and fluctuations in an electrical power system, it is important to identify precursor events that signal an impending fluctuation event. In this paper we integrate wavelet analysis with statistical inference methods to identify a precursor pattern for frequency fluctuation prediction. The frequency time series data was converted into the wavelet domain to extract the time dynamics after which change point detection methods were used to signal significant deviations in the wavelet domain. The change points extracted were taken as early indicators of a fluctuation event. Using historical data on known fluctuation events we trained a regression model to estimate the gap between a change point and its corresponding fluctuation point. Our results show that change points could be predicted a number of time steps in advance with a low false alarm rate.