Abstract: There are many factors affecting the atmospheric density of the thermosphere and the mechanism is complex.It is difficult to establish an accurate atmospheric model，resulting in the perturbation of atmospheric resistance，which has become one of the main factors affecting the orbit prediction accuracy of the space station.The orbit prediction method based on PPO2 algorithm of reinforcement learning was studied.The reinforcement learning network was used to modify the relevant parameters in the atmospheric model and improve the orbit prediction accuracy.Firstly，the orbital dynamics model of the space station was established，the error characteristics of the atmospheric model parameters were analyzed，and the orbital dynamics model modification scheme based on reinforcement learning was designed.PPO2 algorithm was selected as the reinforcement learning algorithm，the design of training parameters and reinforcement learning network model were completed，the training and test samples of PPO2 algorithm were generated，and the simulation training and test were completed.The simulation results show that the scheme could effectively compensate the orbit prediction error caused by the inaccuracy of atmospheric density model，and improve accuracy and efficiency of the orbit prediction of the space station.

Key words: atmospheric drag, space station, orbit prediction, orbital dynamics model modification, PPO2 algorithm