Abstract
Reward function design is a critical issue in deep reinforcement learning (DRL)-based energy management, because manually selected weighting coefficients may introduce subjectivity and lead to suboptimal trade-offs among multiple objectives. In this study, direct hydrogen consumption and fuel cell degradation are expressed on a common basis by converting degradation loss into degradation-equivalent hydrogen consumption. Their sum is defined as total consumption and used as a composite reward feature. A maximum entropy inverse reinforcement learning (MaxEnt IRL) framework is developed to identify the relative weights between total consumption and SOC regulation from expert trajectories. The identified parameters are embedded into Twin Delayed Deep Deterministic Policy Gradient (TD3) and Deep Deterministic Policy Gradient (DDPG) agents. The framework is trained and evaluated over two consecutive WLTCs, with dynamic programming (DP), degradation-aware DP, TD3, and DDPG used as benchmarks. IRL-TD3 achieved hydrogen consumption of 273.6 g, which was only 4.19% and 1.90% higher than those of DP (262.6 g) and DP-deg (268.5 g), respectively. Compared with TD3, IRL-TD3 improved the hydrogen consumption by 2.43% and reduced the degradation-equivalent hydrogen consumption from 80.59 to 60.36 g, corresponding to a reduction of 25.10%. Tests under CLTC, UDDS, and HWFET further confirm the adaptability of the proposed reward-optimization framework.