Hierarchical energy management for fuel cell buses: A graph-agent DRL framework bridging macroscopic traffic flow and microscopic powertrain dynamics

Despite advances in connected vehicle technologies, fuel cell buses (FCBs) still face critical challenges in energy management: inefficient utilization of multi-source traffic data and suboptimal coordination between ecological driving and powertrain optimization. This study addresses these limitations through a hierarchical reinforcement learning framework that synergistically optimizes eco-driving patterns and energy allocation. A spatial-topological graph architecture explicitly models FCB interactions with dynamic traffic elements, while an edge-enhanced graph convolutional network (EGCN) extracts hierarchical spatial-temporal features from heterogeneous traffic data. By integrating EGCN with deep reinforcement learning, the framework improves eco-driving policy performance while considering both hydrogen consumption and powertrain degradation costs at the energy management layer. Results indicate that the proposed strategy reduces travel time by 4.76 % and energy consumption by 3.37 % compared to the intelligent driver model (IDM), and achieves 3.84 % and 5.98 % reductions, respectively, compared to the reinforcement learning strategy without EGCN enhancement. The energy management module achieves 97.84 % economic efficiency relative to dynamic programming (DP) benchmarks. This work uniquely leverages EGCN to resolve high-dimensional traffic-state representations in FCB operations, while developing a hierarchical DRL framework for energy-efficient optimization that bridges macroscopic traffic dynamics with microscopic powertrain control.