To address the limitations of traditional photovoltaic (PV) power prediction algorithms, particularly their inability to accurately capture cloud conditions and their low prediction accuracy, a PV power prediction algorithm based on the fusion of ground-based cloud images and dual-stream data is proposed. First, accurate cloud condition information from ground-based cloud images is utilized, and dense optical flow is employed to extract spatiotemporal and detail change features between adjacent image frames. Then, the advantages of convolutional neural network in feature extraction and residual network in suppressing information loss in model learning are combined to improve the learning ability of the prediction model on the long-term mapping relationship between PV power output and image data. In addition, an attention mechanism is introduced to compensate for the underutilization of critical information during model training. Experimental results indicate that integrating ground-based cloud images and optical flow data offers more spatiotemporal features under cloudy weather conditions. Compared with benchmark models, the proposed method reduces the root mean square error (RMSE) and the mean absolute error (MAE) by 15.50% and 11.65% under sunny conditions, and by 4.05% and 5.15% under cloudy conditions, respectively. This contributes to accurate and reliable forecasting of PV power output by effectively utilizing cloud motion information, thereby improving the timeliness and accuracy of scheduling operations in PV power stations.