Application of Machine Learning to Oscillation Detection using PMU Data based on Prony Analysis

Abstract

Various types of oscillations could occur in the power grid from time to time. Most of them are harmless, while some could significantly impact the reliable power system operations. With increased penetration of renewable energy sources and the general transition to more complex power system operation comes the need for automated and accurate oscillation detection and classification methods. Such methods have been extensively studied in the past. Still, most of the earlier work was done for situational awareness purposes based primarily on simulated waveforms from synthetic power system models. This paper presents the results of an oscillation event detection method using machine learning algorithms trained on features extracted by Prony analysis from field-recorded PMU data. The unique experience of working with field-recorded historical synchrophasor data obtained from 38 PMUs located in the Western Interconnection of the US is shared. Four machine learning oscillation detection and classification models are trained using the results of Prony analysis as input features. The CatBoost classifier outperforms alternatives achieving 76.86% accuracy. An analysis of the data and related labels reveals several aspects of the event labeling that may have hindered the performance of the investigated detection and classification techniques. In the end, we suggest future event labeling approaches that might help avoid the challenges and limitations of current PMU recording practices.