Continuous learning energy management strategy design based on EWC-DDPG for electric vehicles

Appropriate torque distribution strategy is essential for the dual-motor electric drive powertrain, which provides strong and efficient power to the heavy-duty or performance electric vehicle (EV). Although various reinforcement learning-based algorithms have been adopted to optimize the energy management strategy (EMS), the EMS performance is subject to the data size and quality for algorithm training. In this study, a continuous learning algorithm is proposed to improve the adaptability of EMS to unknown scenarios from time to time to further enhance the energy efficiency. Specifically, the Elastic Weights Consolidation (EWC) mechanism is introduced into a Deep Deterministic Policy Gradient (DDPG) algorithm-based EMS of a dual-motor EV, which enables the EMS to learn the different characteristics of various unknown environments to further improve the energy efficiency. The simulation results show that, compared to traditional reinforcement learning EMS, the proposed EWC-DDPG algorithm achieved superior energy performance in different driving cycles with different time lengths. Specifically, compared with the baseline energy consumption, the agent incorporating the EWC mechanism consumed 1.77 %–5.38 % extra energy, while the DDPG-based agent exhibited higher consumption 6.17 %–14.4 %. The results also indicate that the adoption of the EWC mechanism to DDPG can effectively improve its robustness and generalization ability under complex and unknown driving cycles, which provides an effective way for the continuous optimization of the EMS of multi-power EVs.