Reinforcement learning energy management strategy for HEVs based on structure-aware temporal prediction and information bridge mechanism

Traditional predictive energy management relies on accurate speed profile forecasting; however, subtle variations in speed inflection points may trigger large power fluctuations, thereby degrading the decision quality of the energy management strategy (EMS). To address this issue, this paper proposes a novel reinforcement learning (RL)-based EMS for hybrid electric vehicles (HEVs), leveraging structure-aware temporal forecasting and an information bridge mechanism. A power prediction model tailored for RL decision-making was constructed. A pioneering “Structure-Aware Temporal Loss” (SAT-Loss) function was designed to enhance the structural consistency of predicted sequences through robust weighted errors, time-step weighting, local statistical constraints, dynamic correlation constraints, and trend constraints. Simultaneously, a “Confidence-Gated Information Bridge” (CGIB) mechanism was embedded into the inputs of the DDPG Actor and Critic networks to enable the adaptive fusion of predicted hidden representations with the decision-making network. Experimental results demonstrate that SAT-Loss reduced local variance from 0.016 to 0.004 and increased the prediction sequence correlation coefficient from 0.88 to 0.95. Under the FTP-75, NYCC, and ECE driving cycles, the proposed LSTM-SAT + CGIB-DDPG strategy achieved reductions of 11.82%, 12.92%, and 8.27% in equivalent fuel consumption compared with the rule-based strategy, respectively, demonstrating improved energy-saving performance and generalization capability of the proposed energy management system. Hardware-in-the-loop experiments further validated the feasibility of real-time control using the proposed strategy.