Satellite Task Planning Based on Graph Neural Networks and Deep Reinforcement Learning
Jinyi Hou, Guangxi Zhu, Yingte Chai, Xianming Peng, Ping Wang · 2025
On the low computational efficiency and susceptibility to local optima that traditional heuristic satellite task planning algorithms face when dealing with large-scale satellite scheduling tasks, this study introduces a satellite scheduling model based on graph neural networks and deep reinforcement learning. This model aims to achieve more efficient operations and create more reasonable satellite resource scheduling plans. The approach involves constructing a mathematical model for satellite task planning problems using dynamic directed acyclic graphs. It establishes a graph neural network representation learning model and incorporates residual connections into the conventional graph neural network model to mitigate over-smoothing issues. A decision model for satellite observation task scheduling is developed using reinforcement learning, establishing a Markov decision process for satellite observation task scheduling. By jointly constructing a satellite scheduling decision model with the graph neural network representation model, the proposed model demonstrates a task completion rate over 6% higher than heuristic algorithms in conventional scheduling scenarios, with total revenue exceeding 9% relative to heuristic algorithms. In emergency scenarios, the completion rate for emergency tasks exceeds 86%, with disturbance rates to the original planning solutions below 17%. The efficiency of this scheduling model significantly surpasses heuristic algorithms in all scenarios, with running times reduced to below 20% of heuristic algorithm times, presenting a more notable advantage in efficiency for large-scale scenarios.