Spatiotemporal Attention-Augmented Vehicle Interaction Modeling for HEV Energy Management with Maximum Entropy Reinforcement Learning

Hybrid Electric Vehicles (HEVs) play an important role in sustainable transportation, with en-ergy management strategies (EMS) being crucial to their fuel efficiency. However, the complex and dy-namic urban driving environments present significant challenges to EMS adaptability. This study addresses multi-vehicle interactions by modeling the system as an undirected graph and proposes a deep reinforcement learning-based energy management method augmented with a spatiotemporal attention mechanism. The frame-work first employs temporal encoding to capture se-quential observations of graph nodes, followed by spatial encoding to extract feature of node attributes and interactions. The latent representation is then inte-grated into a maximum entropy reinforcement learning framework to optimize EMSs. Simulation results show that the SAT mechanism can improve the reward return during training and enhance fuel economy and adaptability. The proposed EMS achieves an average fuel efficiency improvement of 5 % compared to the baseline method without structured interaction modeling, and 6.4 % over those without temporal encoding.