汽车工程 ›› 2021, Vol. 43 ›› Issue (5): 683-691.doi: 10.19562/j.chinasae.qcgc.2021.05.006

• • 上一篇    下一篇

基于强化学习的多燃烧模式混合动力能量管理策略

张昊,范钦灏,王巍,黄晋,王志()   

  1. 清华大学,汽车安全与节能国家重点实验室,北京 100084
  • 收稿日期:2020-10-14 出版日期:2021-05-25 发布日期:2021-05-18
  • 通讯作者: 王志 E-mail:wangzhi@tsinghua.edu.cn
  • 基金资助:
    国家重点研发计划(2016YFB0101402)

Reinforcement Learning Based Energy Management Strategy for Hybrid Electric Vehicles Using Multi⁃mode Combustion

Hao Zhang,Qinhao Fan,Wei Wang,Jin Huang,Zhi Wang()   

  1. Tsinghua University,State Key Laboratory of Automotive Safety and Energy,Beijing 100084
  • Received:2020-10-14 Online:2021-05-25 Published:2021-05-18
  • Contact: Zhi Wang E-mail:wangzhi@tsinghua.edu.cn

摘要:

针对采用HCCI/SI多燃烧模式的功率分流型混合动力汽车,提出了一种基于深度强化学习(DRL)的能量管理策略。基于发动机台架试验和电机有限元分析建立了混合动力汽车模型。将整车作为环境,采用排序优先经验回放算法,训练基于深度Q网络(DQN)的能量管理智能体。在WLTC和NEDC工况下,与规则策略、自适应等效燃油消耗最小策略(A?ECMS)和动态规划结果进行对比,仿真结果表明:基于DRL的能量管理策略能在维持SOC的前提下,避免燃烧模式频繁切换,并且充分利用中小负荷HCCI燃烧,燃烧模式切换频率降低13%以上,燃油经济性提升6%以上。

关键词: 多燃烧模式, 混合动力汽车, 深度强化学习, 能量管理策略

Abstract:

For hybrid electric vehicles (HEVs) of the power split type with HCCI/SI multi?mode combustion, a deep reinforcement learning (DRL) based energy management strategy is proposed. Firstly, an HEV model is established based on engine bench test and motor finite element analysis (FEA). Then, considering the vehicle as the environment, a ranking prioritized experience replay algorithm is used to train the energy management agent based on deep Q network (DQN). Finally, the performance of the proposed strategy is verified by comparing with the results of the rule?based strategy, adaptive equivalent fuel consumption minimization strategy (A?ECMS) and dynamic programming under WLTC and NEDC driving cycles. The simulation results show that the DRL based energy management strategy can avoid frequent switching of combustion modes and make full use of HCCI combustion mode at small and medium loads while maintaining the level of SOC. The frequency of combustion mode switching is reduced by 13% or more, and fuel economy is improved over 6%.

Key words: multi?mode combustion, hybrid electric vehicle, deep reinforcement learning, energy management strategy