Research on HEV energy management strategy based on improved deep reinforcement learning.

A parallel hybrid vehicle was studied to establish the demand power and power system model of the whole vehicle and proposed an energy management strategy based on improved deep reinforcement learning (DRL). The DRB-TD3 algorithm was proposed to improve the sampling efficiency of the original algorithm by improving the twin delayed deep deterministic policy gradient (TD3) algorithm in DRL and introduced dual replay buffers. A rule-based constraint controller was designed and embedded into the algorithm structure to eliminate unreasonable torque allocation. The performance of the dynamic planning (DP)-based energy management strategy was used as a benchmark for simulation experiment under UDDS driving conditions. The experimental results show that the DRB-TD3 algorithm has the best convergence performance compared with the deep deterministic policy gradient (DDPG) algorithm and the conventional TD3 algorithm, with 61.2$ \% $ and 31.6$ \% $ improvement in convergence efficiency, respectively, the proposed energy management strategy reduces the average fuel consumption by 3.3$ \% $ and 2.3$ \% $ compared with the DDPG- and TD3-based energy management strategies, respectively, the fuel performance reaches 95.2$ \% $ of DP-based, which with the best fuel economy, and the battery SOC can be maintained at a better level, which helps to extend the battery life. [ABSTRACT FROM AUTHOR]