Abstract: To solve the problem that the historical data of new photovoltaic power stations are limited and the distribution of photovoltaic data in different periods of time is quite different, a transfer ultra-short-term photovoltaic prediction modeling framework for single/multiple photovoltaic power stations is presented. First, a weighted rolling time window clustering method based on a weighted structure preserving dimensionality reduction algorithm with inside and outside features is presented to fully consider the uncertainty of photovoltaic series and the inherent bias of numerical weather prediction. Secondly, geodesic flow kernel is used to integrate infinite subspaces to simulate the gradient process of the distribution of photovoltaic data, thus completing the data distribution alignment.Finally, a photovoltaic power prediction model is built using gradient boosting decision tree. The validity of the proposed algorithm is verified by using the public dataset PVOD.