A YOLOv8-Based Model for Hot-Spot Detection in Photovoltaic Systems

Hot spots in photovoltaic (PV) power stations can damage PV modules, leading to a decline in power generation efficiency. To address the challenges of frequent misdetection and missed detection caused by the small size of hot spots, this study proposes a novel You Only Look Once version 8(YOLOv8)-based detection method, termed Dynamic Agent-attentive U-enhanced YOLOv8 Architecture (YOLOv8-DAU), for identifying hot spots in PV arrays. In response to the technical demands for rapid detection of small targets in complex scenarios, this paper innovatively introduces a new algorithm. Firstly, the Agent Attention mechanism is incorporated to enhance global modeling capabilities with linear complexity, thereby improving the robustness of detecting subtle hot spots. Secondly, to mitigate the impact of various noise interferences on hot spot detection, the DySample sampling module is introduced, effectively reducing both false positives and false negatives. Lastly, the UIoU loss function is optimized by dynamically allocating weights and integrating a dual attention mechanism, significantly enhancing detection accuracy. Experimental results demonstrate that the proposed YOLOv8-DAU outperforms YOLOv5, YOLOv7, YOLOv8, YOLOv8-Dysample, YOLOv8-Agent Attention, and YOLOv8-UIoU on the PV array hot spot dataset, achieving higher accuracy and detection speed. The precision and recall rates reach 91.6% and 87.41%, respectively.