Abstract: The errors of three types of LSTM power prediction methods were compared to evaluate the role of operational weather forecast in photovoltaic power prediction and the influence of different divisions of training and test sets on prediction accuracy. The three types of power prediction methods are: using photovoltaic power only, using photovoltaic power and meteorological observation, and using photovoltaic power and meteorological forecast. The meteorological variable used is total irradiance, which has the highest correlation coefficient with photovoltaic power. The analysis period is from January 1 to June 30, 2020, with weather forecasts from the ECMWF and NOAA/NCEP. The results show that, for the data with limited time length, the different division of the training and test sets will have influence on the accuracy of the prediction model. If total irradiance observation is used, the relative error of short-term power predictions can be reduced by about 2.3%. Compared with only using photovoltaic power, the relative error of short-term power prediction is reduced by about 2.1% by using both photovoltaic power and meteorological forecast. Compared with NOAA/NCEP weather forecast, ECMWF weather forecast significantly reduces the error of power prediction. Compared with only using photovoltaic power, using meteorological forecast data can enhance the accuracy of photovoltaic power prediction.