Lifespan-consciousness and minimum-consumption coupled energy management strategy for fuel cell hybrid vehicles via deep reinforcement learning

Energy management strategy (EMS) based on optimized deep reinforcement learning plays a critical role in minimizing fuel consumption and prolonging the fuel cell stack lifespan for fuel cell hybrid vehicles. The deep Q-learning (DQL) and deep deterministic policy gradient (DDPG) algorithms with priority experience replay are proposed in this research. The factors of fuel economy and power fluctuation are incorporated into the multi-objective reward functions to decline the fuel consumption and extend the lifetime of fuel cell stack. In addition, the degradation rate is introduced to reflect the lifetime of fuel cell stack. Furthermore, compared to the referenced optimally energy management strategy (dynamic planning), the DQL-based and DDPG-based EMS with prioritized experience replay (DQL-PER, DDPG-PER) are evaluated in hydrogen consumption and cumulative degradation of fuel cell stack under four driving cycles, FTP75, US06-2, NEDC and LA92-2, respectively. The training results reveal that the DQL-PER-based EMS performances better under FTP75 and US06-2 driving cycles, whereas DDPG-PER-based EMS has better performance under NEDC driving cycle, which provide a potential for applying the proposed algorithm into multi-cycles.