Explainable Ensemble EfficientNet Deep Learning for PV Fault Diagnosis

Photovoltaic (PV) systems play a vital role in expanding clean and renewable energy generation, making reliable operation and efficient maintenance essential for improving energy security, reducing operational costs, and supporting sustainable energy infrastructure. Traditional fault inspection methods are labor-intensive and difficult to scale, creating a need for intelligent monitoring technologies that enhance the reliability and resilience of solar energy systems. In this study, we propose a robust PV fault classification framework based on an ensemble of EfficientNet models for automated vision-based monitoring of photovoltaic installations. The proposed approach integrates transfer learning, stratified five-fold cross-validation, and a soft-voting ensemble strategy to improve generalization and decrease model variance. It incorporates test-time augmentation (TTA) for enhanced robustness against environmental variations and addresses class imbalance through weighted sampling, evaluated using Accuracy, Macro F1-score, and Matthews Correlation Coefficient (MCC). Additionally, the framework integrates explainable artificial intelligence (XAI) with Grad-CAM to offer visual insights into model decision-making and promote transparency. The proposed system detects six common PV panel conditions (clean panels, bird droppings, dust accumulation, electrical damage, physical damage, and snow coverage) from RGB images. Experimental results show that the proposed framework achieves 92.09% classification accuracy, a macro F1-score of 0.926, and an MCC of 0.904 on a locked test set (20% held out prior to training), outperforming several widely used convolutional architectures including ResNet, DenseNet, and VGG16 trained under identical conditions. Rigorous cross-dataset validation employing SHA-256 hash-based overlap removal confirms generalization independence. Furthermore, large-scale auxiliary binary cross-domain validation on the DeepSolarEye dataset (45,721 images), using a practically motivated 5% power-loss threshold and evaluating sensitivity across alternative thresholds, provides additional evidence of robustness under substantial domain shift. By combining advanced deep learning, explainable artificial intelligence, and intelligent fault diagnosis, the proposed framework contributes to digital innovation for renewable energy systems, strengthens the resilience and efficiency of solar energy infrastructure, and supports the development of scalable and reliable approaches for renewable-energy infrastructure monitoring.