Abstract: Traditional photovoltaic power prediction methods predominantly rely on physical or statistical models, which struggle to adequately characterize the complex nonlinear relationship between meteorological factors and photovoltaic power output, resulting in insufficient prediction accuracy. To address this, we propose an ultra-short-term photovoltaic power prediction correction method based on a cloud-edge collaborative architecture. At the edge end, a meteorological feature database is constructed through path analysis to extract key influencing factors. At the cloud end, a hybrid approach integrating BP neural networks and grey system theory is employed to achieve preliminary power prediction and dynamic correction. Experimental results demonstrate that after correction, the average absolute error between BP neural network predictions and actual values decreases to 0.58%, representing reductions of 0.41% and 0.76% compared to control methods using RBF neural networks and regularized extreme learning machines, respectively. During periods of significant daily power fluctuations, the corrected prediction curves exhibit significantly improved fit with actual values, effectively enhancing the accuracy and robustness of ultra-short-term photovoltaic power prediction. This method provides more reliable data support for grid dispatching.