Synchrophasor Forensics: Tracking Spatiotemporal Anomalies and Diagnosing Grid Frequency Events with Machine Learning for Enhanced Situational Awareness

The large-scale integration of microphasor measurement units ( $\mu $ PMUs) is inevitable in smart grids due to the enhanced demand response strategies of active distribution networks and renewable energy sources. As synchrophasor data are crucial, with higher data volume and complexity, a comprehensive method is essential to decode real-time spatiotemporal anomalies and grid events precisely. This paper presents a comprehensive methodology for real-time synchro phasor forensic analysis by leveraging historical frequency and spectrum data. The proposed multi-stage framework investigates the integration of statistical time-frequency analysis with Autoregressive Integrated Moving Average (ARIMA) and machine learning (ML) to address the challenges in grid management. The first stage focused on preprocessing raw data from the New England ISO and European continental datasets, as it contained multiple events. After denoising and feature extraction by multiresolution analysis, the ARIMA model measures the rolling standard deviation (RSD) for the frequency components within different sliding window frames. Using the proposed event detection index, the synchrophasor event detection algorithm (SPEDA) detects frequency excursions, oscillatory events, and other event categories using statistical thresholds. During the second stage, the retrieved features and intensities of the detected events were cross-validated using ML models, thereby enhancing the overall effectiveness of the study. A meticulous scrutiny of ML classifiers and their evaluation metrics was performed, and it was found that Extreme Gradient Boosting (XGB) performs best overall. The proposed method has been investigated and proven effective in this integrated approach with advanced computational facilities to expedite complex computations and reduce training time.