Automatic efficiency detection of solar panels using deep learning-based segmentation approaches

Energy production and consumption in the world has become an important issue with the rapidly increasing population and developing technology. The rapid depletion of fossil energy resources and the negative effects of the use of these resources on human health lead to an increase in the amount of carbon dioxide (CO2) emissions. For this reason, today countries are turning to efficient and sustainable energy sources. Interest and investments in renewable energy sources have increased significantly in this direction. Among these sources, solar energy systems are one of the rapidly developing and promising applications of energy production technology. These systems consist of photovoltaic panels that convert sunlight directly into electrical energy. Our country has high solar potential due to its geographical location, with 9,979 plants and 10,048.512 MW installed capacity. However, there are errors such as cell error, module error, and panel error that negatively affect the efficiency of the photovoltaic systems in these plants. These errors decrease the power generation performance of the panel and reduce its efficiency. Human resources are needed to detect and diagnose failures in photovoltaic systems. However, due to the lack of manpower needed to detect these defects in the panels and the difficulty of transporting solar panels, maintenance, repair and inspection times of the photovoltaic systems are delayed. As a result, both the material losses are increasing and the energy efficiency of the panels is significantly reduced. In this study, it is proposed to detect and diagnose cell faults (hotspot failures) and module faults (bypass diode failures) that negatively affect the efficient operation of photovoltaic systems using a deep learning-based segmentation approach with the aid of a thermal camera. The recommended system has been tested for the efficiency of hotspot failure using YOLO algorithms and bypass diode failures using the U-Net algority. The performance of four different YOLO algorithms has been compared and the model with the best results has been identified. As a result of experiments, the YOLOv8x algorithm produced the best performance compared to other models with 88.7% specificity, 80.5% sensitivity and 83.8% mAP values. Four different U-Net models with different layer depths and filter numbers have been implemented for bypass diode failure segmentation. Based on experimental results, 87.79% AUC, 82.97% F1-Score and 70.89% IOU values were obtained with the highest performance U-Net-V2 architecture. Keywords: Photovoltaic systems, hotspot fault detection, bypass diode fault detection, YOLO algorithm, U-Net algorithm, image processing.