Transformer Enhanced Multi Agent Reinforcement Learning for Joint EV and Hydrogen Charging Infrastructure Planning

The rapid expansion of electric and hydrogen powered vehicles is reshaping urban mobility, but it introduces new challenges for charging and refueling infrastructure planning, grid stability, and low carbon energy utilization. This paper proposes a transformer enhanced multi agent reinforcement learning framework to coordinate heterogeneous actors, including electric vehicles, hydrogen trucks, charging stations, and grid operators, under a centralized training and decentralized execution paradigm. Agents learn policies through a joint objective that balances grid stress mitigation, equitable access, and emission reduction while respecting station and feeder constraints. A proof of concept simulation is evaluated on a smart city setting with 10,000 EVs, 2,000 hydrogen trucks, and 150 stations under peak demand surges, renewable intermittency, and station outage conditions. Compared to a MILP scheduler and a greedy decentralized scheduling baseline, the proposed approach reduces peak to average grid load ratio by 23% and 37%, respectively. It also lowers average waiting time by 18% versus MILP and 41% versus greedy scheduling, and improves fairness by 26%. When aligned with renewable availability windows, the framework achieves a 15% reduction in CO₂ emissions. Under outages affecting 10% of stations, it restores stable operation in 15 steps, compared to 34 for MILP and more than 50 for greedy scheduling. These results indicate a scalable and adaptive solution for sustainable smart city charging ecosystems.