Preference-Conditioned MADDPG for Risk-Aware Multi-Agent Siting of Urban EV Charging Stations Under Coupled Traffic-Distribution Constraints

The public deployment of electric vehicle charging stations must simultaneously balance construction economics, user accessibility, queueing pressure, feeder security, tail risk under demand uncertainty, and spatial fairness. These criteria are strongly coupled, yet most existing studies either rely on static optimization with limited behavioral realism or use multi-agent reinforcement learning for short-term charging operation rather than for long-term siting. This paper proposes a preference-conditioned multi-agent deep deterministic policy gradient (PC-MADDPG) framework for the urban charging station siting problem in a coupled traffic–distribution environment. Candidate charging sites are modeled as cooperative agents under centralized training and decentralized execution. Each agent outputs a continuous pile-allocation action, which is repaired into an integer expansion plan under a budget constraint. The environment evaluates each plan through attraction-based demand assignment, queue approximation, LinDistFlow-style feeder analysis, and a six-objective performance vector, including annual net cost, travel burden, service inconvenience, grid penalty, CVaR of unmet charging demand, and equity loss. On a reproducible benchmark with 12 demand zones, 10 candidate sites, an 11-bus radial feeder, and 16 stochastic daily scenarios, the proposed framework generates a non-dominated archive with 42 unique feasible plans. A representative PC-MADDPG solution opens 5 of 10 candidate sites and installs 20 fast-charging piles, achieving 99.88% mean demand coverage with an annual profit of 2.083 M$ and a maximum line utilization of 0.999. Relative to the NoGrid ablation, the selected full model reduces grid penalty by 23.87% and equity Gini by 51.08%, with only a 0.35% profit concession. Relative to the NoRisk ablation, the CVaR of unmet demand is lowered by 69.70%. Compared with a demand-greedy baseline, the proposed method reduces grid penalty by 11.72% and equity Gini by 25.19% while preserving similar demand coverage. These results provide proof-of-concept evidence, on a reproducible coupled benchmark, that preference-conditioned multi-agent learning can serve as a practical many-objective siting engine for charging-infrastructure planning when coupled traffic and feeder constraints are explicitly modeled.