Abstract: To address the challenges faced in obtaining accurate meteorological data, and increasing uncertainty of photovoltaic power output during transitional weather, an ultra-short term interval prediction model for photovoltaic power was proposed. The methodology leverages the Sparrow algorithm to optimize variational mode decomposition（VMD）, which decomposes historical PV output into multiple sub-modes with strong temporal characteristics across different weather conditions. Secondly, each submode is predicted by LSTM, and the point prediction results are combined by superimposition. Finally, kernel density estimation was used to construct the error model and obtain ultra-short term interval prediction results for photovoltaic power. Simulation results illustrate that in all kinds of weather conditions, the proposed model has higher prediction accuracy and stronger adaptability than the prediction method using only meteorological factors, and can provide more accurate confidence intervals on the basis of point prediction.