Reward Function Design for Reinforcement Learning based Energy Management Strategy for a Fuel Cell/Supercapacitor Hybrid Electric Vehicle

This paper presents a Reinforcement Learning (RL) based Energy Management Strategy (EMS) for fuel cell hybrid electric vehicles (FCHEVs) with supercapacitors (SCs). The use of SCs compared to batteries raises other challenges, due to their low energy density. It highly requires an appropriate EMS to anticipate future behavior to properly manage the power split between powertrain components. For that, a precise design of the reward function is introduced in this paper to enhance the EMS performance, based on the Soft Actor Critic technique. This allows a robust RL-based EMS for FC/SC vehicle with high learning ability. The EMS is trained under recorded drives and validated on the test cycles such as WLTP, all on a simulated environment.