Multi-period drone network planning with topological redundancy: a preference-conditioned deep reinforcement learning approach

Planning resilient drone logistics networks requires balancing economic efficiency and topological redundancy during multi-period expansion. Physical redundancy, such as backup bases, can mitigate disruptions but requires additional infrastructure investment that can constrain cost-effective phased deployment. We formulate this planning problem as a dynamic multi-objective combinatorial optimization problem incorporating synergistic coverage. Coordinating take-off and landing depots creates overlapping elliptical service zones that provide alternative service configurations. We propose the preference-conditioned dynamic attention network (PC-DAN), an end-to-end deep reinforcement learning framework that injects tradeoff preferences directly into the encoder. A single trained policy generates phased deployment schemes for different cost–redundancy preferences without retraining. On 30-node instances, PC-DAN achieves a 99-fold speedup over Gurobi while retaining 83.0% of its hypervolume. At the same scale, it increases hypervolume from 0.476 to 0.831 compared with the non-dominated sorting genetic algorithm II. The policy also generalizes without retraining from 100-node training instances to unseen 150-node networks. A case study involving 129 logistics nodes in Suzhou reveals forward-looking spatial deployment strategies. Rather than concentrating bases to maximize local overlap, redundancy-oriented plans establish spatially dispersed coverage anchors to extend city-wide coverage and improve structural redundancy. These findings highlight the importance of the spatial distribution of service configurations and provide a framework for phased drone logistics infrastructure planning within intelligent transportation systems.