A fusion energy management strategy with continuous self-evolution for hybrid electric vehicles

Conventional rule-based energy management strategies for hybrid electric vehicles exhibit limited adaptability under complex and highly dynamic driving conditions, while learning-based strategies still face challenges in terms of training–deployment feasibility, safety constraints, and engineering reliability. To address these issues, this paper proposes a continuous self-evolution fused energy management framework for hybrid electric vehicles. The proposed framework adopts a deterministic rule-based control strategy as the baseline, with a reinforcement learning agent serving as an adaptive correction layer. A layer-wise nested continuous strategy evolution mechanism is developed, following a coarse-to-fine optimization path characterized by “convergence freezing–new layer stacking–action scale contraction”, which enables sustained improvements in energy-saving performance. In addition, a shadow learning–based training–deployment decoupling mechanism is designed to ensure control safety and practical feasibility during online operation. The effectiveness of the proposed framework is validated under different RL agent algorithms. Simulation results demonstrate that the proposed fused strategy achieves effective energy savings under both Deep Deterministic Policy Gradient and Double Deep Q-Network algorithms, with fuel consumption reductions of 10.5% and 25.2% per 100 km, respectively, and corresponding reductions in equivalent fuel–electricity energy consumption of 22.5% and 25.9%. Furthermore, with the introduction of a three-layer continuous reinforcement learning evolution structure, the reductions in fuel consumption per 100 km further reach 25.2% and 26.2%, while total fuel–electricity consumption is reduced by 26.6% and 26.9%, respectively. In addition, fusion energy management strategy maintains low computational and memory overhead.