Physics-Informed Machine Learning for Intelligent Fault Detection and Classification in Photovoltaic Energy Systems: A Comparative Study

The rapid global expansion of photovoltaic (PV) systems has intensified the demand for reliable and intelligent fault detection frameworks. Conventional threshold-based monitoring approaches remain inadequate for large-scale installations, as they fail to adapt to the nonlinear relationships between environmental conditions and fault electrical signatures. This work presents a physics-informed machine learning framework for the automatic detection and classification of six PV operating conditions — normal operation, open-circuit, short-circuit, degradation, partial shading, and ground fault — using electrical parameters derived by numerically solving the single-diode model (SDM) equations across continuous fault severity ranges with realistic sensor noise injection.

Four supervised classifiers are systematically evaluated under identical preprocessing conditions: Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and XGBoost. The framework achieves near-perfect multi-class fault classification, with SVM reaching 100% accuracy and F1-score, and XGBoost attaining 99.88%, while maintaining above 99% accuracy under sensor noise levels up to 5% — encompassing the typical accuracy range of commercial PV monitoring sensors.

The central finding of this study is that the performance ceiling of approximately 84% commonly reported in the literature is not an algorithmic limitation but a data representation limitation. Physically correct fault modeling, in which each fault class occupies a distinct and non-overlapping region of the electrical feature space as governed by the SDM equations, is the decisive factor enabling reliable classification. This framework requires no additional sensing hardware and is directly applicable to utility-scale PV process monitoring and control systems.