Methodology for Identifying Soiling on PV Panels Using RGB Images and Deep Learning

Soiling is one of the major factors that can affect the performance of PV installations in many regions of the world. On the other hand, the variety of PV models, types of soiling, and climate conditions makes it very difficult to create a universal model. To address this problem, this study presents a methodology for the identification of soiling on PV panels via semantic segmentation, which can support the decision-making process in terms of surface cleaning and automation of the process. The methodology includes data collection, preparation of training and testing data, training of models, and application of the optimal one. Next, the methodology is demonstrated using a small dataset of clean and dirty PV panels, three neural network architectures (DeepLab v3, U-Net, and PSPNet), and two backbone models (ResNet34 and ResNet50). The obtained results show the feasibility of the methodology and allow highlighting the DeepLab v3 model with a ResNet34 backbone as the best-performing algorithm for identifying pigeon droppings. The second-best combination is the U-Net + ResNet34, which showed good efficiency for identifying smaller dirty areas. The proposed methodology could be useful for operators of large-scale photovoltaic installations by supporting the decision-making process when it comes to the timely cleaning of specific areas for performance improvement and lower costs.