Abstract

Photovoltaic (PV) systems are increasingly vital to global energy transitions but remain vulnerable to surface anomalies such as dust, snow, bird droppings, and physical or electrical damage—that significantly reduce power yield and long-term reliability. Manual inspection is inefficient for large-scale solar farms, motivating the development of intelligent and automated fault detection systems. This study introduces a reproducible comparative deep-learning framework that systematically benchmarks four convolutional neural network (CNN) architectures ResNet50, EfficientNetB0, MobileNetV3Small, and DenseNet121 under identical preprocessing, training, and validation settings. The framework integrates runtime efficiency analysis, five-fold cross-validation, and a real-time GUI-based deployment interface, bridging the gap between academic benchmarking and field-level implementation. A six-class labeled dataset of 1,574 RGB images was expanded through extensive data augmentation (rotation, flipping, brightness adjustment, and Gaussian noise perturbation) to simulate diverse real-world conditions. Among the four tested models, DenseNet121 achieved the highest macro-averaged F1-score (≈ 0.96), followed by ResNet50 (0.93), EfficientNetB0 (0.92), and MobileNetV3Small (0.92), highlighting clear accuracy–efficiency trade-offs across architectures. The novelty of this work lies in its multi-model benchmarking design and transparent methodology, providing a standardized and reproducible reference for future PV image-based diagnostics. Practically, integrating the models into a real-time graphical user interface (GUI) demonstrates their feasibility for UAV-based or on-site PV inspection. From an operational and economic perspective, this approach supports cost-effective, scalable, and non-invasive monitoring solutions tailored for modern large-scale solar farms.