AI‐Based PV Fault Detection: Benchmarking Indicators and Feature Selection Methods

Abstract

This article presents a benchmarking of indicators for automated defect detection in photovoltaic modules based on I‐V curves. The methodology is structured in three stages. First, an ideal defect‐free I‐V curve is generated through simulation to provide a reference for comparison with experimentally measured I‐V data under standard test conditions. Second, several electrical and curve‐shape indicators are computed to quantify the deviation between simulated and measured curves. Electrical parameters are expressed as ratios between experimental and simulated values, whereas curve‐shape descriptors integrate the comparison directly into their formulation. Finally, a comprehensive feature selection strategy is applied; filter (Pearson correlation and Mutual Information) and wrapper methods (Recursive Feature Elimination, Permutation Importance, and Boruta) are tested across several machine learning classifiers, including Decision Trees, Support Vector Classifiers, K‐nearest neighbors, Random Forests, Gradient Boosting, XGBoost, and LightGBM. The approach is evaluated using an experimental I‐V curve database generated by the GdS‐Optronlab research group. Results consistently discard indicators such as the open‐circuit voltage ratio and short‐circuit current ratio, while highlighting the current at the maximum power point ratio, the fill factor, and the mutual information of the maximum power point.