A Robust Machine Learning Framework for Partial Discharge Diagnosis in Diverse Power Substations

Partial discharge (PD) detection is essential to avoid failures in electrical substations. This paper focuses on classifying PD types in both Air-Insulated (AIS) and Gas-Insulated (GIS) substations. Support Vector-Machine (SVM) and Random Forest (RF), two artificial intelligence (AI) methods, were created for this task. The paper provides the complete mathematical framework for two distinct machine learning (ML) classifiers specifically designed for subtraction analysis. The three primary PD-sources: corona, surface, and internal discharge were recognized by the models during training. The study validates the PD-features by analyzing and presenting distinct time-domain waveforms and frequency-domain spectra for each discharge type under noisy-conditions, providing a strong physical basis for the AI-classification. The RF-classifier achieved perfect accuracy of 99.6%. The SVM-classifier also showed high accuracy of 97.78%. The results demonstrate that AI can provide reliable early warning for substation maintenance. This helps improve the safety and reliability of power networks. The research offers a practical framework for applying these methods in Egypt's energy sector.