Abstract: To address the missed and false detection problems caused by the complex background and the small size of targets in photovoltaic hot spot detection, an improved YOLOv8-PV hot spot detection algorithm is proposed. This algorithm introduces the BiFormer dynamic sparse attention mechanism, optimizes the CDM downsampling method, and improves the Bi-PAN-FPN feature fusion network, thereby effectively reducing the information loss of feature maps, enhancing the network resolution and receptive field, improving the model detection accuracy and generalization ability, and reducing the miss rate of small hot spot targets. The experimental results show that this algorithm achieves an average accuracy of 98.2% on the hot spot dataset, that is considerably higher than that of other mainstream object detection algorithms such as Cascade R-CNN, RetinaNet, and YOLOX, proving the effectiveness and superiority of our model.