Machine Learning Approaches in Anomaly Type Detection and Localization in Distribution System

Determining the type and location of cyber-physical anomalies (i.e., faults or cyber attacks) in power distribution systems is critical to take adequate mitigation and restoration actions. Machine learning (ML) methods present advantages over classical methods to enhance anomaly detection in distribution systems. This paper reviews and presents different methods for ML-based anomaly type detection and localization in power distribution systems. Particularly, four main components have been identified in learning-based approaches: data aggregation, feature extraction, network structure, and classification task, each of which must be carefully selected according to specific application requirements and constraints. Various neural network structures are implemented and compared on a 13-bus test feeder, including deep neural, convolutional, and graph convolutional networks. Numerical studies illustrate the considerable impact of feature extraction, network structure and classification task on the performance of machine learning-based methods.