Toward Ecological Decision-Making on Highways: A Multiobjective Eco-Driving Framework for Connected FCHEVs via Offline-to-Online Reinforcement Learning

Leveraging vehicle-to-everything (V2X) and artificial intelligence (AI), fuel cell hybrid electric vehicles (FCHEVs) enable eco-driving to coordinate internal power distribution and external decision-making. However, existing research overlooks the coupling between longitudinal–lateral behaviors and energy-saving objectives, and underutilizes V2X capabilities. To address this, a multi-objective eco-driving framework is proposed, co-optimizing energy management strategy (EMS) and decision-making with durability considerations on highways. To overcome the limitations of traditional off-policy reinforcement learning (RL) and soft actor-critic (SAC) algorithm, an offline-to-online reinforcement learning (OTORL) framework is proposed. Incorporating behavior cloning (BC) and re-attentive experience replay (RAER), an offline-to-online reinforcement learning (OTORL) framework, comprising “SAC-BC offline robust training” and “SAC-RAER online efficient fine-tuning”, is established with a rule-based safety guardian. Ablation experiments show the proposed framework improves energy efficiency by 13.95%, safety by 4%, comfort by 9.1% and durability by average 10.36%, with minor compromises in travel efficiency and power performance. Tests under varying traffic densities and real-world NGSIM dataset scenarios further demonstrate its adaptability and continuous improvement in complex traffic environments.