Research on Energy-Efficient Multi-Objective Satellite Orbit Control Based on Deep Reinforcement Learning

Abstract

To address the challenges of multi-objective trade-offs, uncertainty, and nonlinearity in satellite orbit control, this paper proposes a data-driven control framework based on deep reinforcement learning. Methodologically, a novel composite reward function is designed that integrates fuel consumption, trajectory-tracking accuracy, control smoothness, and mission constraints, such as collision avoidance. An adaptive weighting mechanism is introduced to balance these competing objectives. The algorithm employs an enhanced Soft Actor-Critic architecture, in which both the actor and critic networks are constructed with deep residual networks and augmented with an attention mechanism to capture long-term dependencies in orbital dynamics. Experimental results demonstrate that, for low-Earth orbit transfer and proximity operations, the proposed approach reduces average fuel consumption by approximately 18.7% compared to conventional optimal control and baseline deep deterministic policy gradient methods, while meeting the same mission accuracy requirements. Additionally, control stability is improved by 22.4%. Under conditions of model parameter perturbations and measurement noise, the method achieves a success rate of 99.2%, confirming its strong robustness and adaptability. In conclusion, the developed deep reinforcement learning framework enables effective multi-objective coordination and long-horizon fuel-efficient planning, providing a viable pathway toward autonomous and intelligent orbital control.