Abstract: Aiming at the problems of complex modeling and low accuracy of existing short-term solar irradiance prediction methods,a GRU-RF dynamic weight combination prediction method based on deep learning is proposed. The atmospheric factors and solar irradiance data are fused,and the random forest（RF）model with fast operation speed and low model complexity is integrated with the gated recurrent unit（GRU）neural network with time sequence memory for dynamic weight weighting. The variation characteristics of solar irradiance,surface air temperature,relative humidity,surface wind speed and relative pressure received by the surface are predicted respectively. Through the comparison and analysis of several models,the results show that the GRU-RF model can be used to predict the solar irradiance in short time（9 h）with faster running speed,and can be used to predict the solar irradiance data well in different time intervals（5,10 and 15 min）.