关键词: 路径规划, 多无人机, 应急通信, 多目标优化, 深度强化学习

Abstract: With the advancement of unmanned aerial vehicle (UAV) technology, UAV-enabled aerial communication base stations have increasingly become a crucial means of rapidly restoring emergency communications in post-disaster scenarios. In post-disaster emergency communication scenarios, multiple UAVs need to jointly optimize communication coverage maximization, coverage fairness, and energy consumption minimization for moving users in the affected area under limited battery capacity and partial observability, through distributed collaborative decision-making on UAV charging and flight. To address this, we formulate a cooperative multi-agent scheduling model based on a decentralized partially observable Markov decision process (Dec-POMDP). We then propose a predictive-assisted algorithm termed the Predictive Double Deep Graph Network (PDDGN). Specifically, PDDGN incorporates a prediction network to estimate future state observations and employs a deep recurrent graph network to integrate current observations with predicted information, thereby learning the observation–action value functions for UAVs and adapting to environmental dynamics. In addition, a double-target-network architecture is introduced; during training, differentiated soft updates between the two target networks are used to enhance training stability and improve learning efficiency. Simulation results demonstrate that, for dynamic communication coverage tasks under partial observability, the proposed PDDGN effectively improves communication coverage and fairness while reducing energy consumption. Compared with conventional deep reinforcement learning baselines, PDDGN achieves up to a 7% improvement in the composite Coverage–Fairness–Energy (CFE) score and a maximum reduction of 60.9% in convergence episodes, while exhibiting good stability and scalability.

Key words: Path planning, Multi-Unmanned Aerial Vehicle(UAV), Emergency communications, Multi-objective optimization, Deep Reinforcement Learning(DRL)