Physics-guided attention neural network for interpretable and robust photovoltaic fault classification with strict external validation

Accurate and interpretable fault diagnosis in photovoltaic (PV) systems is essential for improving operational reliability and reducing energy losses. This study proposes a Physics-Guided Attention Neural Network (PGANN) for three-class PV fault classification, including Open-circuit, Short-circuit, and Partial Shading conditions. The proposed framework integrates physics-guided feature engineering into the input representation through calculated power, power error, and normalized voltage ratio, thereby embedding electrical-domain knowledge without imposing unstable physics-based penalty terms in the loss function. A feature-wise attention mechanism is incorporated to enhance nonlinear discrimination among fault patterns. Model interpretability is evaluated using SHAP and LIME analyses. The model is assessed using an internal stratified test set and a strictly non-overlapping external validation set drawn from the same Mendeley simulation pool. The final reported results are based on the external validation dataset, where the proposed PGANN achieved 93.90% overall accuracy, 93.81% macro F1-score, and high class-wise AUC values, reaching 0.995 for Open-circuit faults. Although the external validation set is independent at the sample level, it originates from the same simulation domain; therefore, the results demonstrate robust same-generator generalization rather than full real-world domain transfer. The findings show that the proposed PGANN framework provides a stable, interpretable, and physically consistent solution for simulation-based PV fault diagnosis.