Abstract
Despite their significant environmental benefits, solar photovoltaic (PV) systems are susceptible to malfunctions and performance degradation. This paper addresses detecting and diagnosing faults from a dataset representing a 250 kW PV power plant with three types of faults. A comprehensive dataset analysis is conducted to improve the dataset quality and uncover intricate relationships between features and the target variable. By introducing novel feature importance averaging techniques, a two-phase fault detection and diagnosis framework employing tree-based models is proposed to identify faults from normal cases and diagnose the fault type. An ensemble of six tree-based classifiers, including decision trees, random forest, Stochastic Gradient Boosting, LightGBM, CatBoost, and Extra Trees, is trained in both phases. The results show 100% accuracy in the first phase, particularly with the Extra Trees classifier. In the second phase, Extra Trees, XGBoost, LightGBM, and CatBoost achieve similar accuracy, with Extra Trees demonstrating superior training and convergence speed. This study then incorporates Explainable Artificial Intelligence (XAI), utilizing LIME and SHAP analyzers to validate the research findings. The results highlight the superiority of the proposed approach over others, solidifying its position as an innovative and effective solution for fault detection and diagnosis in PV systems.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call