Abstract: During the model switch process of plug-in hybrid electric vehicle, the fluctuation of exhaust temperature and intake flow velocity invoked by frequent engine start-stop is obviously aggravated, which causes the reduction of catalytic efficiency and increase of exhaust emission of SCR catalysts. Especially, the above result is extremely obvious during the cold start stage at low temperature. On the other hand, it is difficult to establish an accurate transient model of SCR catalyst, and the traditional model-based hybrid control strategy development method is not effective. Taking a P2 plug-in diesel electric hybrid vehicle as the research object, the longitudinal dynamic model including engine, battery and SCR aftertreatment system was established. On this basis, deep reinforcement learning was applied to the energy management of plug-in hybrid electric vehicles. DQN algorithm was used to solve the weighted objective function composed of fuel consumption and emission, and the control strategy was obtained with the state variables of required power, battery SOC and SCR temperature, and the optimal motor power as the output variable. The final experimental results were compared with that of dynamic programming (DP) algorithm. The results show that the fuel consumption is 2.623 L / 100 km, and the NOx emission at the outlet of SCR catalyst is 0.227 5 g/km, which is 10.12% and 25.69% lower than that of DP control strategy, which proves the effectiveness of the proposed control strategy.

Key words: vehicle engineering, deep reinforcement learning, control strategy, multi-objective optimization, plug-in hybrid electric vehicle, dynamic programming

摘要： 插电式混合动力汽车工作模式切过程中发动机频繁启停引起的发动机排气温度和进气流速波动明显，导致SCR催化器催化效率降低和排放恶化，尤其是低温冷启动阶段更为明显。另一方面，建立精确的SCR催化器瞬态模型较为困难，传统基于模型的混合动力控制策略开发方法效果较差。以某P2构型插电式柴电混合动力汽车为研究对象，建立了包括发动机、电池和SCR后处理系统的整车纵向动力学模型；在此基础上将深度强化学习应用于插电式混合动力汽车的能量管理问题，采用DQN算法对油耗和排放组成的加权目标函数进行求解，得到以需求功率、蓄电池SOC和SCR温度为状态变量、以电机最优功率为输出变量的控制策略；最后将测试结果与DP算法进行对比分析。结果表明，燃油消耗为2.623 L/100 km，SCR催化器出口NOx排放为0.227 5 g/km，与DP控制策略相比，分别下降10.12%和25.69%，证明了提出控制策略的有效性。

关键词: 车辆工程, 深度强化学习, 控制策略, 多目标优化, 插电式混合动力汽车, 动态规划算法