Similar articles

Abstract: The siting of low-altitude takeoff and landing platforms is a fundamental challenge for the future development of Urban Air Mobility (UAM). We formulate the problem as a variant of capacitated facility location problem (CFLP) with flight-range and service-capacity constraints, and propose SPID, a deep reinforcement learning (DRL)-based solution framework. The problem is modeled as a Markov Decision Process (MDP). To address dynamic coverage, we design a DRL framework with a multi-head attention mechanism to capture spatiotemporal patterns and integrate dynamic and static information into a unified input state vector. A Gated Recurrent Unit (GRU) is used to generate the query vector, enhancing sequential decision-making. The action network is guided by a loss function combining unmet demand penalties with service-distance costs, enabling end-to-end optimization. Experimental results demonstrate that SPID significantly improves solution efficiency and robustness over traditional methods under flight and capacity constraints. It also outperforms heuristic algorithms in accuracy, with a performance gap of around 6% compared to exact solutions. This work provides an efficient, scalable approach to low-altitude site selection, supporting rapid decision-making in large-scale UAM scenarios.