DFL-PVNet: A Lightweight Photovoltaic Hotspot Detection Network Based on Dynamic Feature Fusion

Photovoltaic (PV) panels are vital for renewable energy, but are prone to hotspot faults during operation, which can reduce dependability and cause safety risks. Conventional hotspot detection techniques are often hindered by high costs, intricate operations, and limited detection capabilities. In recent years, deep learning-based detection methods, particularly those leveraging infrared imaging, have demonstrated substantial promise. However, current models encounter challenges related to parameter complexity and limited flexibility in feature fusion, which impede their performance across varying operational conditions. Therefore, this study proposes the Dynamic Feature Fusion and Lightweight Photovoltaic Hotspot Detection Network (DFL-PVNet), which is based on an improved YOLOv8 structur. DFL-PVNet aims to reduce model parameters and computational complexity while optimizing feature fusion, thereby enhancing the accuracy and applicability of hotspot detection. Experimental results demonstrate that DFL-PVNet achieves superior performance in the hotspot detection of PV panels. Compared with baseline, the number of parameters and computation of DFL-PVNet decrease by 56.7% and 65.4%, and the recall, mAP@50 and mAP@50-95 increase by 14%, 11.6% and 5.4%, respectively, which are all better than the existing mainstream target detection methods.