Learning Power System’s Graph Signals for Cyber and Physical Stress Classification

Abstract

Situational awareness towards various types of cyber and physical stresses in power systems is critical for the reliable operation of these critical infrastructures. Identifying the type of stress that has occurred in the system is particularly crucial for deciding the corrective measures for mitigating the stress and also for future preventive planning. In this paper, a two-stage stress classification framework based on the learning of the power system’s graph signals has been proposed. Specifically, graph signal processing (GSP) has been utilized to extract features from the power system’s graph signals for building the models. Using GSP allows for capturing information about the interconnections and interactions among the components of the grid along with its spatio-temporal dynamics. It has been shown that this machine learning-based classification with GSP-based features is effective for classifying between cyber and physical stresses as well as further classifying among different types of cyber and physical stresses. Abrupt changes in the load demand and tripping of a transmission line are considered as examples of physical stresses, while five types of cyber attacks with no abrupt onset on the PMU time-series are considered as cyber stresses. Various GSP-based features are evaluated and a dimensionality reduction technique based on down-sampling in the graph-frequency domain is proposed. The classification performances have been evaluated across various classifiers using data under different noise levels.