An Interpretable Multi-Level Stacking Framework for Photovoltaic Power Prediction and Fault Analysis

Abstract

Accurate power prediction and early fault detection are critical for the reliability and efficiency of photovoltaic (PV) systems. This study proposes a multi-level stacking ensemble framework that integrates multiple regression models across hierarchical layers for PV power prediction and fault analysis. In the first level, ensemble algorithms including Extra Trees Regressor, Gradient Boosting Regressor, Histogram-based Gradient Boosting, and LightGBM generate base predictions. These outputs are refined by secondary models (AdaBoost, Decision Tree, and ElasticNet) in the second level, with Ridge Regression serving as the final meta-learner. The framework is evaluated using real PV system data encompassing normal operation, partial shading, and module faults. Model performance is assessed through coefficient of determination (R2), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). Explainable AI techniques, specifically LIME and SHAP, are employed to interpret model predictions and identify key influencing factors. The proposed approach demonstrates superior predictive accuracy and robustness compared to individual models, with statistically significant improvements confirmed by the Diebold-Mariano test. The results highlight current, fault conditions, and partial shading as the most influential variables. This framework offers an accurate, interpretable, and scalable solution for PV system monitoring and predictive maintenance.