Abstract: For the purpose of predicting the short-term output power of concentrating solar power（CSP）station,firstly introduce adaptive idea to improve the training algorithm of the recursive deep belief network,and establish a direct normal irradiance short-term prediction model. Secondly,a method for forecasting the short-term output power of CSP station combined with its static model is proposed. Finally,a performance test is carried out to verify the feasibility of improved recursive deep belief network and the effectiveness of the CSP station short-term output power prediction method. The research results show that the established improved recursive deep belief network can improve the prediction accuracy and training speed. Also,the proposed CSP station short-term output power prediction method can predict its output power more accurately.