Abstract: Due to the uncertainty of cloud movement,it is difficult to accurately estimate the output power of photovoltaic systems,thereby impacting the reliability and cost-effectiveness of renewable energy integration. To effectively utilize satellite cloud observation data,this paper proposes an ultra-short-term photovoltaic power forecasting model based on cloud features.Convolutional neural networks are used to extract features from satellite cloud images, which are then fused with four meteorological features through correlation analysis and used as inputs to the photovoltaic power forecasting model. On this basis, the spatial relationships between these features are analyzed using convolutional neural networks,and long shortterm memory networks are employed for time series prediction of photovoltaic output power. Furthermore,considering that data from different time periods within a single calendar day may have varying impacts on the prediction,we introduce a multihead attention mechanism to identify crucial time points and significant features. This further enhances the accuracy of the proposed model in this study. The predictive performance of the proposed model is verified using actual data from photovoltaic power plants,along with corresponding satellite cloud images and weather data. The computational analysis results demonstrate high prediction accuracy and timeliness of the model,particularly during periods of high noon irradiation and intense cloud movement fluctuations,where the model still maintains a high level of prediction accuracy.