Enhanced solar panel segmentation and hotspot recognition using U-Net: a multiclass semantic segmentation approach / Mohd Zulhamdy Ab Hamid ... [et al.]

Maintaining and optimising photovoltaic (PV) systems requires accurate segmentation and detection of thermal hotspots in solar panels. This study present a novel multiclass semantic segmentation approach based on a U-Net deep learning model to help solar panel and hotspot analysis. Utilising the U-Net architecture, solar panels, hotspots, and background components can be classified with high fidelity. A large dataset of thermal images with multiple class labels was rigorously trained and evaluated on the model. The U-Net model also achieved a very impressive overall accuracy of 97.96% and an average Intersection over Union (IoU) of 0.7246 on all classes. In particular, it recorded an IoU score of 0.9485 for background, 0.9677 for the solar panels, and 0.2578 for the hotspots. The model does well at separating background from solar panels, but lower IoU for hotspots suggests that defining areas with solar panels is more challenging, as they are smaller and less obvious. The results show how the U-Net model increases the fault detection accuracy in PV systems by accurately segmenting the components of the solar panel and the hotspots. Insights from these studies will lead to improved maintenance practices that can increase the operational lifespan of solar installations. By doing so, this study highlights the potential of deep learning models, particularly UNet, to facilitate solar panel analysis and ultimately contribute to more reliable and sustainable energy production through the automation of monitoring and maintenance in solar power plants, with scalability and efficiency.