Investigating Machine Learning for Anomaly Detection in Phasor Measurement Unit Data

Abstract

This paper investigates the use of Machine Learning (ML) algorithms for Anomaly Detection (AD) in Phasor Measurement Unit (PMU) data. Specifically, this is an evolution of previous heuristic techniques which investigated constraint and temporal anomaly detection in PMU data. This work found that K-Means Clustering is suitable for AD in static PMU data, a set of PMU data at one time stamp, and Auto Regressive Integrated Moving Average (ARIMA) to detect anomalies in time series PMU data. A smart grid test bed with 10 PMUs and a physical power grid emulator was used to generate PMU data sets for this work. The K-Means and ARIMA ML models were designed, trained, and validated offline using a baseline dataset to build a representation of how the power grid performs under normal operating conditions. Once the models were validated, they were then subjected to a test data set to evaluate performance and suitability for this application. For static AD, the K-Means model evaluates distortion from new, or real-time, PMU data. For the temporal AD, the ARIMA model predicts the next PMU measurement and evaluates the delta between the prediction and actual data. Distortion values, or delta values, exceeding predefined thresholds are indicative of anomalies. In practical implementation this Anomaly Detection System (ADS) addresses the need to analyze large volumes of PMU data while also alerting system operators of significant deviations from normal behavior. This paper describes an implementation of ML in this ADS and presents initial results which highlight the feasibility of this approach.