Bayesian Convolutional Neural Networks for Anomaly Detection in Power Systems

John McClinton, Patrick McClure, Preetha Thulasiraman

Abstract: Power systems are vulnerable to internal failures and physical tampering. To address these vulnerabilities, this research explores Bayesian machine learning methods for anomaly detection in power systems. By analyzing data such as current and voltage from non-invasive load monitoring (NILM) equipment, the proposed model aims to detect anomalies that signal potential failures or malicious intrusions before they lead to catastrophic system breakdowns. This project builds on prior work that used supervised learning methods for similar tasks. This research demonstrates a significant improvement in anomaly detection, particularly through the use of Bayesian Convolutional Neural Networks (CNNs) to detect anomalies in power systems. In particular, Monte Carlo Dropout (MCDroput) and Spike-and Slab Dropout, implemented with Flipout, were evaluated. Through testing on real-world data from a testbed operational technology (OT) system, the results show Bayesian CNN approaches, especially Spike-and-Slab Dropout, demonstrate an ability to detect known and unknown anomalies, as well as prediction mistakes, within a power system using electrical load monitoring. This presents a potentially scalable tool kit to enhance security of power systems and reduce maintenance cost and risk.