Multi-Sensor Partial Discharge Prediction in Switchgear Cabinets Using Machine Learning Models

This study develops a multi-sensor, data-driven approach for accurate and timely partial discharge (PD) prediction in switchgear cabinets, a critical factor for power system reliability. By integrating Gaussian Mixture Models (GMM) for anomaly detection, the model incorporates data-driven PD confidence directly into classification, effectively distinguishing true PD events from normal variations. Advanced feature engineering captures essential chemical dynamics during discharge, while physical knowledge is embedded as constraints in the model training process, aligning predictions with known PD behavior and enhancing learning efficiency. Experimental results show the GMM-restricted SVM model achieves 98% accuracy, 99% recall, a 0.95 AUC, and reduces alert time error to 3 seconds, compared to 8 seconds in baseline models. Ablation studies confirm that combining GMM-based confidence with physical constraints strengthens both precision and recall. This framework provides a robust and interpretable solution for predictive maintenance in power systems, with future work focusing on expanding sensor inputs, real-time monitoring, and finer discharge classification.