Abstract
The integration of distributed energy resources into power grids fosters the development of precise monitoring, protection, and control applications by employing immense spatiotemporal data from micro phasor measurement units ( \mu PMUs). For enhanced situational awareness, a comprehensive methodology is required for real-time synchro phasor forensic analysis using advanced machine-learning techniques to detect and classify anomalies in grid events. This paper presents a twostage analytical framework that combines waveletautoregressive integrated moving average (ARIMA)-based analysis with a machine learning approach to enhance the identification and classification of events by leveraging historical frequency and spectrum data. The raw data from the New England ISO and European Continental Split datasets were preprocessed in the initial phase, as they included multiple events. The process involves a Stationary Wavelet Transform (SWT) for denoising, and sliding window ARIMA models to identify the Rolling Standard Deviation (RSD) for feature extraction and threshold setting. The frequency excursions and oscillations are classified based on the Synchro phasor Event Detection Algorithm (SPEDA) as per statistical thresholds. The retrieved features of the detected and localized events were cross validated using machine learning classifiers in the next stage, enhancing the overall efficiency and effectiveness of the study. The study demonstrates that advanced computing facilities accelerate sophisticated calculations and reduce model training time.