Deep Reinforcement Learning Based Energy Management Strategy for Fuel Cell and Battery Powered Rail Vehicles

Fuel cell system has broad prospects of application for vehicles due to its zero-emission property. This paper focuses on energy management strategies (EMS) for hybrid rail vehicles powered by fuel cells and batteries on remote non-electrified routes. One of the core issues of hybrid rail vehicles is the power distribution to achieve the fuel economy and the battery's charge sustaining. Due to reinforcement learning's (RL) model-free feature, it is regarded as a novel method to obtain EMS without prior knowledge as a prerequisite. In this paper, twin delayed deep deterministic policy gradient (TD3) reinforcement learning is introduced to solve the energy management problems. The TD3 agent is trained in a stochastic training environment considering passenger flows using the measured data from real rail routes. Subsequently, a new reward function term concerning the battery's charge sustaining is introduced, which leads to a good performance in terms of convergence. Finally, the TD3-based EMS is validated and compared with a benchmark result. The results indicate that the TD3-based EMS can achieve near-optimal hydrogen consumption without prior knowledge of driving cycles.