Imitation Learning-Based Efficient Energy Management for Fuel Cell Electric Vehicles Using Data Aggregation

Emerging artificial intelligence (AI) learning technologies provide new insights into energy management strategies (EMSs) for fuel cell hybrid electric vehicles (FCHEVs). However, existing learning-based EMSs are obtained through inefficient trial-and-error learning, leading to low training efficiency. To address this issue, this paper proposes an EMS based on imitation learning (IL). Specifically, the EMS is trained using the behavior cloning (BC) algorithm to mimic the behaviors of dynamic programming (DP), avoiding the time-consuming trial-and-error and improving training efficiency. Furthermore, to further improve the adaptability and fuel economy of the strategy, a data aggregation mechanism is introduced to the BC-based EMS. This mechanism aggregates historical expert data with real-time expert data collected during driving to promptly update the parameters of the EMS. Simulation results show that the proposed BC-based EMS saves 70.48% of the training time cost compared with popular reinforcement learning-based EMS and its fuel economy reaches 96.15% of the DP-based EMS. Additionally, the proposed data aggregation mechanism enhances the adaptability of the BC-based EMS, resulting in a 2.96% improvement in fuel economy under unfamiliar driving conditions constructed from real-world velocity data. This paper provides a new solution to the energy management problem by adopting a novel AI learning approach.