A dual-approach machine learning and deep learning framework for enhanced fault detection in photovoltaic systems: Incorporating SDM parameter analysis and thermal imaging

Due to the frequent occurrence of faults in photovoltaic (PV) systems, effective and reliable fault diagnosis methods are essential to ensure optimal energy production. In this study, we propose two complementary approaches for fault detection in PV modules. The first method is a hybrid machine learning strategy combining Random Forest and Gradient Boosting classifiers, integrated via a Voting Classifier. This approach leverages key parameters from the single diode model (SDM) to evaluate the health status of PV modules. Our work stands out through a more intuitive and controlled methodology: SDM parameters are systematically varied based on realistic degradation scenarios inspired by actual physical causes. This allows us to generate a reliable synthetic dataset, suitable for training a deep learning model for automated fault classification. This method achieves a classification accuracy of 99.55%, demonstrating robustness in identifying rare and complex faults. The second method adopts a deep learning framework based on the Swin Transformer V2 architecture, specifically designed for the analysis of thermal infrared images. This enables the detection of subtle thermal anomalies indicative of potential defects, reaching a validation accuracy of 96.39%. Additionally, a dedicated web-based platform has been developed to facilitate the visualization and classification of PV faults. This tool supports rapid diagnostics and a better understanding of failure mechanisms, contributing to more efficient maintenance of photovoltaic systems.