Abstract: In view of the strong randomness of wind speed in wind energy,it is difficult for wind power to be connected to the grid,reliable and high-quality wind speed prediction results are very important issues for the planning and application of wind energy. In this research,a spatio-temporal fusion model（CSWLSTM）combining convolutional neural network（CNN）and shared weight long shortterm memory network（SWLSTM）is proposed to fully extract the spatial and temporal information contained in the wind speed sequence to improve prediction accuracy. In addition,in order to obtain reliable wind speed probability prediction results,a new hybrid model integrating CNN,SWLSTM and GPR is proposed,called CSWLSTM-GPR. CSWLSTM-GPR is applied to the case of wind speed prediction in Inner Mongolia,China. Comparing the wind speed prediction methods of CNN and SWLSTM models with the same structure in terms of point prediction accuracy,interval prediction applicability and comprehensive performance of probability prediction. The reliability test of CSWLSTM-GPR ensures the reliability and persuasiveness of the predicted results. The experimental results show that CSWLSTM-GPR can obtain high-precision point prediction,appropriate prediction interval and reliable probability prediction results in the wind speed prediction problem. It also fully demonstrates that the CSWLSTM proposed by this research has good application potential in wind speed prediction.