[Objective] To address the challenge of declining accuracy in photovoltaic power prediction caused by rapidly changing cloud conditions， this paper proposes an ultra-short-term photovoltaic （PV） power prediction method based on the cloud-irradiance-power coupling mechanisms. [Methods] First， a two-stream convolutional network is constructed to extract spatial and motion features of clouds， while a Transformer encoder is employed to capture the temporal features of irradiance sequences. Then， cross-attention mechanism is applied to fuse these features for multi-step ahead prediction of surface irradiance. Finally， to overcome the limitations of purely data-driven approaches， this study constructs a photovoltaic power prediction model by using irradiance forecasts as inputs and incorporating physical constraints governing the surface irradiance-to-power output conversion. [Results] Experimental results show that the proposed irradiance prediction method outperforms baseline models under various weather conditions. Furthermore， using the predicted irradiance as input significantly improves PV power forecasting accuracy compared to methods that directly incorporate cloud information. The introduction of physical constraints further enhances PV power prediction performance and improves the model’s generalization capability. [Conclusions] This study demonstrates that deeply mining the impact of dynamic 3D cloud information on irradiance and incorporating physical prior knowledge between irradiance and PV power represents an effective way to enhance the accuracy and model generalization capability of ultra-short-term PV power prediction.