Detection of hot spot anomalies in photovoltaic power plants using yolo-based image processing algorithms

The efficiency and operational safety of photovoltaic (PV) power plants are directly related to the early detection of anomalies such as hotspots on the panels. The primary objective of this thesis is to conduct a comparative analysis of the performance of four different versions of the You Only Look Once (YOLO) algorithm (YOLOv9, YOLOv10, YOLOv11, and YOLOv12), a deep learning-based object detection method, for detecting hotspot anomalies in PV panels. Within the scope of this study, a dataset consisting of thermal images collected by an Unmanned Aerial Vehicle (UAV) from a solar power plant was utilized. The performance of the models was objectively evaluated using standard metrics, including precision, recall, F1-score, mAP50, and mAP50-95. The analyses revealed that the YOLOv9 m and YOLOv11 s models, in particular, produced more stable and highly accurate results in the hotspot detection task compared to the other models. The YOLOv11s model stood out with its high recall and F1-score, while the YOLOv9m model demonstrated the best performance in the mAP50-95 metric, which reflects challenging detection conditions. These findings indicate that the aforementioned models can be effectively and practically implemented in the maintenance and inspection processes of PV power plants. Furthermore, the methodology developed in this thesis, which involves mapping detected anomalies with their geographical coordinates, offers significant potential for the automation and optimization of maintenance operations.