A Machine Learning Framework for Online PMU Data Classification, Recovery and Robust Modal Estimation in Sustainable Energy Systems

Abstract

In modern power systems, Wide-area monitoring systems (WAMS) play a critical role in maintaining the stability of large-scale power systems. However, the reliability of synchrophasor measurements is often compromised by missing values and anomalies, which significantly affects classification and mode estimation tasks. A thorough framework is suggested in this study for both classification and data recovery under distorted PMU measurements. To address this challenge, a new method integrating Local Outlier Factor (LOF) and Agglomerative Clustering has been developed to classify PMU signals into regions such as ambient and ringdown, enabling targeted analysis. The TLS-ESPRIT algorithm is then applied within each region for robust mode estimation. By combining LOF for outlier detection, Piecewise Cubic Hermite Interpolating Polynomial (PCHIP) interpolation for data imputation, and TLS-ESPRIT for mode extraction, this approach proves more resilient to anomalies and missing values than conventional techniques. This combined approach shows a strong solution for better mode estimation and increased sustainability in power systems. The suggested method has been tested on two types of data: real-time data gathered from the IEEE 39-bus system and information sourced from the Western Electricity Coordinating Council (WECC).