Abstract: For the high precision control problem of large-scale satellite formation，an attraction-based deep deterministic policy gradient（ADDPG）was proposed.Firstly，topological configuration characteristics of a hypercube topological formation were formulated，and a satellite formation dynamic model was established.Then，the virtual center of hypercube satellite formation was designed to measure the overall flight state of the formation.In order to solve the problems of exploration and expansion balance of model-free deep reinforcement learning，the ε-imitation action selection strategy method was introduced.Lastly，the satellite formation control strategy based on ADDPG was proposed.The algorithm does not depend on the environmental model.With the existing information being optimized，the probability of blind trial and error in the initial exploration of the learning model would be decreased.The simulation results show that the ADDPG strategy enables higher precision as well as lower energy consumption.Compared with the well-known algorithm，the algorithm introduced in this paper not only accelerates the formation convergence rate，but also improves the control precision by 5% and reduces the energy consumption by 7%.Thus，the effectiveness of the algorithm is verified.

Key words: ADDPG strategy, virtual center, hypercube topology, satellite formation, deep reinforcement learning