Data-driven a convergence-enhanced fusion energy management strategy based on teacher agent guidance for hybrid electric vehicles

Learning-based energy management strategy is considered to be one of the most promising vehicle energy-saving technologies, but the problems of low sampling efficiency and slow convergence rate of its training process seriously restrict the large-scale development. To address this problem, this paper proposes a fusion energy management strategy that combines the prior knowledge with deep learning, which is guided by the teacher strategy (rule-based and equivalent consumption minimization strategy), to solve the problems of poor strategy convergence effect and slow convergence rate caused by inconsistent convergence direction of intelligent agents. In this paper, we present a fusion-based energy management strategy with teacher agent guidance, which combines the already mature rule-based and equivalent consumption minimization strategy and guides the convergence direction of the intelligent agent. By incorporating prior knowledge from the teacher agent, the proposed energy management strategy initially constrains the exploration direction of the agent with prior knowledge. During iterations, the influence of prior knowledge on the agent is adjusted through dynamic weight, forming three stages: “constraint - transition - autonomy”. This enables it to both improve convergence speed and achieve better fuel economy under different working conditions and expert experiences. Simulation results show that the method improves the convergence rate from 40.1 % to 42.7 % and the fuel economy from 1.2 % to 1.9 % compared to the traditional learning method, reaching 97.9 % to 98.7 % of the dynamic programming algorithm. In addition, the proposed method can be easily popularized for other types of hybrid vehicles.