Optimized Undersampling and Explainable AI for Photovoltaic Fault Detection and Diagnosis: A Tree-Based Multi-Tier Framework

Accurate fault detection and diagnosis (FDD) in photovoltaic (PV) systems is essential for ensuring reliable energy generation and efficient maintenance. This study proposes a hierarchical machine-learning framework for FDD in a 4-kW rooftop PV system operating in a tropical climate. To address the inherent class imbalance in PV fault datasets, a novel clustering-based undersampling strategy is introduced to optimize the selection of non-fault instances while preserving representative diversity. The proposed three-tier architecture sequentially performs fault detection, classification, and severity estimation. Tree-based ensemble models, including Light Gradient Boosting Machine (LGBM), Extreme Gradient Boosting (XGB), Random Forest (RF), and Extra Trees (ET), are evaluated for predictive accuracy, robustness, and computational efficiency. Extensive exploratory data analysis and sensitivity studies highlight the impact of dataset size, feature selection, decision thresholds, and sensor noise on model performance. The framework consistently achieves over 98% accuracy across all tiers, with LGBM demonstrating the optimal balance between accuracy, stability, and inference efficiency. Explainable artificial intelligence (XAI) using SHapley Additive exPlanations (SHAP) values provides both global and local interpretability, revealing that solar irradiance, array voltage, current, and grid frequency are the most influential variables across detection and diagnosis stages. Sensitivity analysis under realistic sensor perturbations confirms the model’s resilience, with minimal degradation in accuracy at moderate noise levels. Compared to single-tier and traditional oversampling approaches, the proposed method improves diagnostic precision, reduces false positives and negatives, and enables scalable deployment for real-time PV monitoring. These findings demonstrate that the proposed framework delivers a reliable, interpretable, and computationally efficient solution for practical PV fault management.