A hierarchical reinforcement learning framework for high-efficiency control of turbine-assisted fuel cell air supply systems

The integration of turbine expanders into fuel cell systems (FCS) can significantly improve system efficiency. However, it also increases the complexity of operating parameter optimization and real-time control of the air supply system. To address these challenges, this study proposes a hierarchical reinforcement learning framework, termed GA-RDDPG, which combines a genetic algorithm (GA)-based efficiency-oriented feasible control baseline with deep deterministic policy gradient (DDPG)-based online residual correction to improve transient performance while maintaining air-supply safety. A progressive curriculum-learning strategy is incorporated into the DDPG training process to improve learning efficiency and reduce exploration in unsafe operating regions. The results demonstrate that GA-RDDPG achieves coordinated control of compressor speed and turbine nozzle opening, while effectively balancing air supply energy consumption and stack performance to improve overall FCS efficiency. At 0.6 A cm −2 , the average cell voltage reaches 0.77 V and the system efficiency reaches 54.61 %, representing improvements of 5.5 % and 1.35 %, respectively, compared with the baseline strategy. Under a 70 kW s −1 power variation, the GA-RDDPG agent exhibits excellent transient response and robustness, with cathode pressure and airflow stabilizing within 2 s. Under demanding vehicular operating conditions, the GA-RDDPG agent also demonstrates strong generalization. Although moderate transient overshoots in turbine nozzle opening are observed, the controller maintains stable air-supply regulation.