Analysis of Machine Learning Models for Anomaly Detection Using PMU data

The growing complexity of modern power systems, coupled with the integration of Phasor Measurement Units (PMUs), necessitates robust and efficient anomaly detection mechanisms to ensure grid reliability and security. This study evaluates various machine learning models, including traditional methods (Logistic Regression, SVC), ensemble methods (Random Forest, AdaBoost, CatBoost, LightGBM), and deep learning architectures (CNN, LSTM, CNN-LSTM), for multi-class anomaly detection using PMU data. A dataset, generated from IEEE 39-bus test systems, simulates realistic scenarios with physical and cyber anomalies augmented by Gaussian and Laplacian noise to reflect real-world conditions. The models were assessed based on metrics such as F1 score, precision, recall, accuracy, and computational efficiency. Results show that ensemble methods, particularly CatBoost and LightGBM, achieved the best balance of classification accuracy and efficiency, making them ideal for real-time applications, while deep learning models like CNN-LSTM excelled at capturing complex patterns but were computationally intensive. Traditional models offered interpretability but struggled with higher-dimensional data. This research underscores the trade-offs between performance and efficiency across methods and highlights the potential of ensemble models as scalable solutions for grid monitoring.