Early Anomaly Detection for Power Systems Based on Kullback-Leibler Divergence Using Factor Model Analysis

Abstract

Real-time anomaly detection is a critical monitoring task for power systems. Most studies of power network detection fail to identify small fault signals or disturbances that might lead to damages or system-wide blackout. This work presents a methodology for analyzing high-dimensional PMU data and detecting early events for large-scale power systems in a non-Gaussian noise environment. Also, spatio-temporal correlations of PMU data are explored and determined by the factor model for anomaly detection. Based on random matrix theory, the factor model monitors the variation of spatio-temporal correlations in PMU data and estimates the number of dynamic factors. Kullback-Leibler Divergence is employed to measure the deviation between two spectral distributions: the empirical spectral distribution of the covariance matrix of residuals from online monitoring data and its theoretical spectral distribution determined by the factor model. Using IEEE 57-bus, IEEE 118-bus, and Polish 2383-bus systems, three different case studies demonstrate that the proposed method is more effective in identifying early-stage anomalies in high-dimensional PMU data collected from large-scale power networks. Performance evaluations validate that this method is sensitive and robust to small fault signals compared with other statistical approaches. The proposed method is a data-driven approach that doesn’t require any prior knowledge of the topology of power networks.