Extreme disasters endanger power grid safety. Water accumulation, a key risk factor, is crucial for power supply security but hard to identify due to blurry images and changing shapes during disasters. We propose WA-Yolov8, a method based on YOLOv8, introducing GAM to enhance feature extraction from blurry images and integrating ODConv for dynamic convolution kernel adjustment to cope with shape variations. Experimental results show that the precision, recall rate, mAP50 and MAP50:95 of our proposed method WA-Yolov8 on the WA-dataset are 90.2%, 88.9%, 94% and 56% respectively, which are improved by 1.4%, 1.8%, 1.3% and 2.2% respectively compared with YOLOv8. In summary, the method we proposed achieves good results and better meets the needs of risk factor identification.