Distributed hydrogen refueling stations enable the coupling of renewable generation, storage, and demand fulfillment; however, their performance depends on coordinated control under strict physical safety limits. Centralized controllers are often impractical due to privacy constraints and unreliable communication links, while unconstrained learning can reduce operating costs at the expense of unsafe pressure excursions. Therefore, this study evaluates safety-constrained coordination across multiple stations using federated learning-based distributed scheduling and benchmarks a non-federated Local Control baseline (local-only, no coordination). Using a feasibility-first rule with an acceptance threshold of τ=0.2 on the pressure violation metric (Vp≤0.2), the best feasible overall controller (Local Control) achieved a cost of 2131.83 with pressure violation Vp=0.172, representing a 37.22% reduction relative to a centralized reference cost of 3396.25. Federated training with Federated Averaging and a solar–wind mixing scheme produced the best feasible federated policy (cost 2423.72, Vp=0.163) with 866,688 transmitted bytes. Extensive simulations report cost, unmet demand, safety violations, and communication overhead, demonstrating that feasibility-first selection is essential because lower-cost policies can be unsafe (e.g., cost 1952.27 with Vp=2.63).
Al-Shammari et al. (2026) studied this question.