AgentEMS: Integrating DRL and LLM-refined rules for hierarchical energy management of multi-stack fuel cell vehicles

Fuel cell electric vehicles (FCEVs) offer a promising pathway for decarbonizing transportation. Multi-stack FCEVs, in particular, provide the enhanced flexibility, redundancy, and scalability required for heavy-duty applications. However, their complex architecture introduces significant challenges for energy management. This paper proposes AgentEMS, a hierarchical energy management framework that explicitly decouples high-level decision-making from low-level control execution. The upper-level decision layer employs a deep reinforcement learning (DRL) agent to adaptively select optimal operating modes based on real-time driving conditions. The lower-level control layer then executes interpretable, rule-based power allocation strategies across the fuel cell stacks. To overcome the bottleneck of manual rule design, this framework integrates a large language model (LLM) as an offline rule synthesizer. A novel prompt engineering mechanism extracts structured control knowledge from dynamic programming optimal trajectories, guiding the LLM to automatically generate degradation-aware and energy-efficient control rules. By combining DRL adaptability with rule-based stability, AgentEMS ensures safe, interpretable real-time operation. Experimental results demonstrate enhanced system efficiency and significantly reduced fuel cell degradation. The proposed approach reduces fuel-cell degradation by more than 45% compared with conventional end-to-end DRL methods, indicating substantial potential for extending system lifetime.