• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2020, Vol. 56 ›› Issue (16): 181-192.doi: 10.3901/JME.2020.16.181

• 运载工程 • 上一篇    下一篇

扫码分享

基于机器学习速度预测的并联混合动力车辆能量管理研究

胡晓松1,2, 陈科坪1,2, 唐小林1,2, 王斌3   

  1. 1. 重庆大学机械传动国家重点实验室 重庆 400044;
    2. 重庆大学汽车工程学院 重庆 400044;
    3. 北京新能源汽车股份有限公司 北京 100176
  • 收稿日期:2019-10-28 修回日期:2020-03-30 出版日期:2020-08-20 发布日期:2020-10-19
  • 通讯作者: 唐小林(通信作者),男,1984年出生,博士,副教授,博士研究生导师。主要研究方向为混合动力汽车NVH与能量管理,路径规划。E-mail:tangxiaolin6@126.com
  • 作者简介:胡晓松,男,1983年出生,博士,教授,博士研究生导师。主要研究方向为电池储能系统、混合动力系统控制、车联网系统。E-mail:xiaosonghu@cqu.edu.cn;陈科坪,男,1994年出生。主要研究方向为混合动力系统控制,路径规划。E-mail:kepingchen@cqu.edu.cn
  • 基金资助:
    国家自然科学基金(51875054,51705044)和重庆市杰出青年基金(cstc2019jcyjjq0010)资助项目。

Machine Learning Velocity Prediction-based Energy Management of Parallel Hybrid Electric Vehicle

HU Xiaosong1,2, CHEN Keping1,2, TANG Xiaolin1,2, WANG Bin3   

  1. 1. The State Key Laboratory of Mechanical Transmission, Chongqing University, Chongqing 400044;
    2. Department of Automotive Engineering, Chongqing University, Chongqing 400044;
    3. Beijing New Energy Automobile Co., Ltd., Beijing 100176
  • Received:2019-10-28 Revised:2020-03-30 Online:2020-08-20 Published:2020-10-19

摘要: 针对当前混合动力汽车模型预测控制(Model predict control,MPC)能量管理中预测域速度预测不精准问题,进行基于机器学习的速度预测研究。首先建立基于无极变速(Continuously variable transmission,CVT)的单轴并联混合动力汽车模型,其次采用机器学习方法对未来时间窗内的行车速度进行预测,得到三种不同预测方法下的方均根误差(Root mean squared error,RMSE)精度,其中长短期记忆(Long short term memory,LSTM)神经网络最佳,前馈神经网络次之,支持向量机最差。然后,利用模型预测控制策略对车辆进行能量流分配,验证不同预测方法对燃油消耗、SOC的影响,且分析对比不同预测时间窗长度下的能量管理性能,找到不同预测方法下的最小油耗预测域。最后将预测控制与传统动态规划(Dynamic programming,DP)、等效燃油消耗(Equivalent consumption minimization strategy,ECMS)能量管理策略进行性能比较,发现机器学习预测控制对减少油耗具有良好的潜力,同时对模型预测控制算法预测域内的扰动量预测具有重要指导意义。

关键词: 并联混合动力汽车, 机器学习, LSTM, 预测能量管理, 模型预测控制

Abstract: To resolve the problem of inaccurate prediction-horizon speeds in model predictive energy management algorithms for hybrid electric vehicles, the speed prediction based on machine learning is examined. First, a single-shaft parallel hybrid electric powertrain model equipped with a continuously variable transmission(CVT) is established. Then, the machine learning algorithms are utilized to predict the vehicle velocity in the future time horizons, and root mean squared error(RMSE) values of three different prediction methods are obtained, where the performance of the LSTM-NN is the best, followed by the feedforward neural network, and the support vector machine is the worst. Model predict control(MPC) is subsequently deployed to manage energy flow distributions. The effects of different prediction methods on fuel consumption and SOC are verified, and the effects of the prediction horizon size on the energy management performance are comparatively analyzed. Consequently, the minimum fuel consumption prediction horizon is determined. Finally, the performance comparisons between the predictive controller and traditional Dynamic programming(DP) and equivalent consnmption minimization strategy(ECMS) are made, illustrating that the machine learning driven predictive control is promising for reducing fuel consumption. It also facilitates realizing disturbance quantity prediction in the prediction horizon for model predictive control algorithms.

Key words: parallel hybrid electric vehicles, machine learning, LSTM, predictive energy management, model predictive control

中图分类号: