Station keeping control for aerostat in wind fields based on deep reinforcement learning

In this paper, a stratospheric aerostat station keeping model is established. Based on Markov decision process, Double Deep Q-learning with prioritized experience replay is applied to stratospheric aerostat station keeping control under dynamic and non-dynamic conditions. Ultimately, metrics like the average station keeping radius and the station keeping effective time ratio are used to assess the effectiveness of the station keeping control approach. The simulation analysis results show that: under the mission the station keeping radius is 50 km and the station keeping time is three days, in the case of no power propulsion, the average station keeping radius of the stratospheric aerostat is 28.16 km, the station keeping effective time ratio is 83%. In the case of powered propulsion, the average station keeping radius of the stratospheric aerostat is significantly increased. The powered stratospheric aerostat can achieve flight control with a station keeping radius of 20 km, an average station keeping radius of 8.84 km, and a station keeping effective time ratio of 100%. © 2024 Beijing University of Aeronautics and Astronautics (BUAA). All rights reserved.