Post-Disaster Multi-UAV Task Planning Based on Graph Neural Network Decoupling

This paper addresses multiple unmanned aerial vehicles (UAVs) task planning for post-disaster supply delivery, a problem that comprises two coupled multi-objective optimization problems (MOPs): task allocation and path planning. These sub-problems are highly coupled, which therefore requires a layered decoupling strategy. To this end, a multi-UAV task planning method that combines graph neural network (GNN) with an improved multi-objective particle swarm optimization (MOPSO) algorithm is proposed in this paper. A GNN is employed to model the complex relationships between UAVs and tasks, extracting node and graph embeddings through message-passing mechanisms to provide shared, high-quality global prior information. A multi-objective optimization model is constructed for task allocation and operation planning. Given its high-dimensional, nonlinear and strongly coupled characteristics, an MOPSO algorithm with adaptive dynamic weight is designed to enhance convergence efficiency. Simulations show the proposed method outperforms benchmark methods in task completion rate (TCR), task completion time (TCT), energy consumption and MOP performance metrics across varying scales.