基于深度学习的混合动力汽车预测能量管理

为了优化混合动力汽车的能量动态分配过程，提升混合动力汽车的燃油经济性和动力电池荷电状态（SOC）平衡性，提高混合动力汽车能量管理策略的鲁棒性，以等效燃油消耗最小化策略为基础，结合对车辆未来行驶工况的预测研究，分析车辆未来行驶需求能量的变化，制定相应的动态调整策略。基于车联网通信技术，实时采集车辆的运行状态信息和交通信息，作为车辆未来工况预测模型的输入变量。以数据驱动为特征，基于混合深度学习建立工况预测模型。利用STL分解算法对各输入变量进行周期性、趋势性等特征分解，并对各输入变量的特征分量，使用混合深度学习网络从数据局部特征及时间维度依赖特征来深度挖掘目标车辆车速与外部信息及历史数据的关系，进而对车辆未来的行驶工况进行预测。利用预测的工况信息，分析车辆未来行驶需求能量的变化，应用于自适应等效消耗最小化策略等效因子的实时动态调整，从而实现对车辆的优化控制，并通过与传统自适应等效消耗最小化策略进行对比，验证该方法的有效性。研究结果表明：基于混合深度学习的工况预测模型预测精度比BP网络预测模型高44.72%；利用精确的预测工况信息预测能量管理，可以实时动态调整发动机和电机的功率输出，降低油耗并维持电池SOC平衡。

Predictive Energy Management Strategies in Hybrid Electric Vehicles Using Hybrid Deep Learning Networks

To optimize the dynamic energy distribution process, improve fuel economy and power battery state of charge (SOC) balance, and raise the robustness of hybrid electric vehicle (HEV) energy management strategies, corresponding dynamic energy management strategies were formulated based on the equivalent consumption minimization strategy (ECMS). Further, these strategies were integrated with predictive research concerning the energy demand of HEVs in the near future. Using the internet of vehicles communication technology, vehicle information using state and traffic information was collected, in real-time, as input variables for the vehicle's future road condition prediction model. A vehicle speed prediction model was then established based on the characteristics of data-driven and hybrid deep learning. The seasonal-trend decomposition using loess (STL) was used to decompose the periodic and trend features of each input variable. The hybrid deep learning network was, then, used to mine the relationship between the target vehicle speed, external information and historical data from the local and time-dependent features of the data, to complete the prediction of the future road condition of the vehicle. Based on the predicted road condition information, the change in the future driving demand energy of the vehicle was analyzed and was applied to the real-time dynamic adjustment of the equivalent factor of ECMS to realize the optimal control of the vehicle. The effectiveness of this method was verified by comparing it with traditional adaptive ECMS. The results show that the prediction accuracy of the hybrid deep learning model is 44.72% higher than that of the back propagation network model. The prediction of energy management is capable of dynamically adjusting the power output of the engine and motor in real-time, reducing fuel consumption, and maintaining the SOC balance of the battery.