Abstract: In this paper, a hybrid predicion modelling method based on neural network is introduced for navigation satellite. The new model is based on dynamical model and the neural network model is adopted to modify it. During the training phase, neural network model tries to approach the difference of dynamical model prediction product, then compensate it in the prediction phase. In this paper, the characteristics of DMM prediction error for GPS are explored, and then predition strategy for long term (longer than 15days) is designed, at the last, a group of tests on GPS satellites are processed to prove and evaluate the performance of the new method, the test results show that the improvement for 15-40days prediction is obvious.