Anomaly Detection for Primary Distribution System Measurements using Principal Component Analysis

The sensor measurements in power distribution systems can be potentially anomalous due to sensor malfunctions, communication failure, and cyberattacks. In this paper, a data-driven framework for network-wide anomaly detection in electric power distribution systems is proposed specifically to detect anomalies in primary distribution voltage magnitude measurements. The central hypothesis is that the voltage measurements at a given time-step for different buses of a phase in the power distribution system are correlated. This hypothesis is thoroughly validated for different distribution test systems. Upon leveraging this observation, the first stage of the proposed anomaly detection framework utilizes Batch Principal Component Analysis (PCA) - a dimensionality reduction technique - to reduce the dimensionality of the input data (voltage measurements). The second stage of the proposed framework entails the approach for anomaly detection and identification using the residuals computed from the PCA model. The applicability of the proposed approach is demonstrated using the IEEE 8500-node distribution test system to detect missing and bad data anomalies.