Maximum entropy deep inverse reinforcement learning-based energy management strategy for hybrid electric logistics trucks

Considering external environmental factors in the energy management strategy (EMS) is vital for hybrid electric logistics trucks. In this regard, in contrast to traditional research focusing on modeling and analyzing one or several external factors, this paper directly investigates the underlying mechanism by which related state quantities influence energy allocation. A deep reinforcement learning-based EMS is proposed to treat the strongly coupled human–environment–vehicle system from a novel perspective, where all external environmental factors serve as inputs to the model, directly influencing the decision-making of the energy management module through complex internal couplings. First, expert actions are obtained through dynamic programming (DP) to generate a dataset of optimal state–action pairs to serve as reference training samples. Subsequently, a maximum entropy deep inverse reinforcement learning framework is developed to uncover the latent decision-making mechanisms of the expert policy. Within this framework, a reward function network based on gated recurrent units is designed to process sequential state information. The expert policy is modeled as a truncated Gaussian distribution to provide a soft learning target. The Kullback–Leibler divergence is minimized between the soft policy distribution induced by the reward function and the expert distribution, ensuring the approximate accuracy of the learned reward-guided policy while maintaining sufficient stochasticity and exploratory capability. Finally, by integrating multidimensional state information from the information layer, the trained reward function is used to achieve efficient energy allocation. The simulation results indicate that the proposed EMS closely approximates the global optimal solution of DP compared to existing approaches.