Classification of high frequency transient events buried in low frequency sampled PMU data

This paper presents a deep-learning based method for the classification of transient events of the power system using the data collected by a standard Phasor Measurement Unit (PMU). Although transient power system events are high-frequency variations of the voltages and currents, it leaves a trace in the PMU time-series data, which are collected at relatively much lower-frequencies. The main motivation of this study is to use low-frequency data recorded by PMU devices, widely used in the industry, for the classification of transient events, which would normally require the use of high-cost power quality analyzers. In the proposed method, PMU data are converted into 2D data by Gramian Angular Field (GAF) image transformation method to increase the classification performance. Then, the GAF matrices corresponding to PMU time-series data are classified by a Convolutional Neural Network (CNN) based deep learning method. Transient event data used in this research work includes different numbers of frequency and angle time series of three different events, Line Trip, Generation Trip and Load Shedding, obtained from the Frequency Disturbance Recorders (FDR) developed at Virginia Tech. FDRs are connected to the FNET/GridEye network hosted by the University of Tennessee. Classification performances on the three power system events show that using only the frequency data gives the best performance results compared to that using angle only and both frequency and angle. Impact of the training data amount is examined for all cases and future work is proposed. Although there are similar studies in the literature, contribution of this research work is revealing the effect of PMU data type and size of training set in detail.