Evaluation of a multimodal large language model for fault detection and classification in photovoltaic modules

Abstract

As Photovoltaic (PV) solar deployment expands, accurate automated fault detection is increasingly important to avoid module degradation and energy losses. This study evaluates Multimodal Large Language Models (MLLMs) for detecting visual and thermal faults in PV modules using two open datasets. Results show promising performance: 80% accuracy in multi-class visual classification on the UTFPR dataset with zero-shot instructional prompting, and 43% accuracy in thermal fault classification with few-shot descriptive inference on the Infrared Solar Modules dataset, surpassing previous work by 19 percentage points. Although Convolutional Neural Networks (CNNs) still achieve higher raw accuracy, MLLMs offer flexible and interpretable textual diagnoses without extensive retraining, highlighting their potential for intelligent solar maintenance.