Explainable Machine Learning for Photovoltaic Fault Diagnosis: A Comparative Study

In recent years, photovoltaic (PV) systems have become a key component of the global energy transition. However, PV systems are exposed to different types of faults, such as physical, environmental, and electrical, which reduce energy production and pose safety risks if not properly diagnosed. Machine Learning (ML) models are able to classify faults with high accuracy but often act as black boxes, hindering the ability of the users to interpret the models’ predictions. Explainable AI (XAI) aims to address this issue by providing tools that generate explanations by analyzing the relations between input and output of the trained model. This study optimizes and compares a set of ML models to find the most performing one on a dataset collected from a real PV plant. Among them, LightGBM demonstrated the best overall performance. Then, a quantitative evaluation framework was used to compare two XAI methods: Local Interpretable Model-agnostic Explanations (LIME) and Shapley Additive Explanations (SHAP), which proved to be the most effective in the PV domain.