Abstract: The short-term power forecasting of distributed photovoltaic lacks the same spatio-temporal meteorological data.Traditional methods of power forecasting directly use the data of adjacent centralized photovoltaic stations, ignore the time shift of meteorological information caused by geographical location offset, which is difficult to meet the requirements of forecasting accuracy. A hybrid forecasting method considering the time shift of meteorological information is proposed. In the mechanismdriven model, the optimal time shift is used to modify the meteorological data offset. In the data-driven model, the temporal pattern attention（TPA） mechanism is introduced to weaken the impact of meteorological data offset. Then, the two methods are fused through the Stacking ensemble learning framework to form a hybrid mechanism-data-driven model to further improve the forecasting stability and accuracy. The case analysis based on the actual data of the distributed photovoltaic and public meteorological stations shows that the proposed method can effectively use the meteorological data of geographical location offset to achieve higher accuracy of power generation forecasting for distributed photovoltaics.