Urban Air Mobility (UAM) represents a promising solution to metropolitan transportation challenges, yet efficient scheduling and resource allocation in large-scale UAM networks remain computationally intractable for traditional optimization methods. This paper proposes a hybrid scheduling framework that couples graph neural network prediction with deterministic constraint repair for UAM scheduling and resource allocation. We first develop a heterogeneous graph representation scheme that captures diverse network elements including vertiports, waypoints, and aircraft with their distinct attributes. A spatiotemporal graph attention network (ST-GAT) architecture is then constructed, integrating spatial topology awareness with temporal dynamics modeling through dual attention channels. Furthermore, an adaptive multi-objective optimization strategy is designed to balance throughput maximization, delay minimization, and energy efficiency while satisfying operational constraints. Experimental results on a simulated metropolitan UAM network demonstrate that the proposed method achieves 94.2% throughput under peak demand conditions with average delay of 4.1 min, outperforming genetic algorithm and simulated annealing baselines by 15.9% and 19.9% respectively. The integrated pipeline—comprising neural scheduling prediction and post-processing feasibility repair—maintains sub-second inference latency up to 250 concurrent aircraft and 60 vertiports, confirming its suitability for real-time operational deployment within the tested scale.
Xu et al. (Fri,) studied this question.