随着节能环保法规的不断加严，插电式混合动力汽车（Plug-in Hybrid Electric Vehicle，PHEV）已成为一条重要的新能源汽车技术路线。与此同时，在车辆智能化和网联化的发展趋势下，PHEV将融合实时交通信息与智能控制方法，实现高效节能运行。本研究针对混合动力汽车能量管理的动态最优控制与策略迁移应用问题，按照数据驱动的PHEV动力系统建模、基于强化学习（Reinforcement Learning，RL）的能量管理策略优化以及控制策略的仿真-现实迁移三个层次展开研究。针对强化学习策略所需的高保真度训练环境，首先进行了混合动力系统的优化匹配以确定最优的硬件参数和基准控制策略，接着提出了结合机理模型和数据驱动模型的增量式误差标定方法。利用实车转毂测试数据，建立了具有误差矫正能力的高保真度PHEV模型，为基于RL的混合动力能量管理策略开发提供了高保真度训练环境。此外，在强化学习策略的车辆集成阶段，该方法可基于实车试验数据与模型仿真之间的差异进行模型误差的快速校正。为实现强化学习型能量管理策略的工程应用，以等效消耗最小策略为基础构建了构建了强化学习与最优控制相结合的能量管理策略架构，具有抗扰动能力，能够在周围环境与自身动力系统状态不准确的条件下，实现稳定的控制效果。针对动力系统机-电-热-交通信息耦合的高维复杂模型难以在同一时间尺度内进行最优控制的问题，提出了基于多时域强化学习（Multi-Horizon Reinforcement Learning，MHRL）的动态能量管理策略，将复杂神经网络分解为不同时间尺度任务的子网络进行协同训练，有效提升了时变工况与不同温度条件下的能量管理效果，并通过硬件在环测试验证了MHRL算法的实时性。最后，在能量管理系统的跨平台迁移应用与实车集成方面，针对强化学习型能量管理策略的仿真-现实迁移问题，提出了模块化智能体设计方法以及元强化学习策略训练机制，分别实现了对车辆动力系统模型以及驾驶工况的迁移。基于虚拟驾驶仿真的实车转毂测试验证了所提出的能量管理系统，增强了控制系统对不同车型平台和多变驾驶场景的适应性，提升了实车燃油经济性。

With the increasing stringency of energy-saving and environmental protection regulations, Plug-in Hybrid Electric Vehicle (PHEV) has become an important technology route for new-energy vehicles. At the same time, with the development of automotive connectivity, PHEVs will integrate real-time traffic information and intelligent control to improve system operation efficiency. Aiming at the application of dynamic optimal control and control strategy transfer for energy management of PHEVs, this research is carried out in three levels: data-driven PHEV powertrain modeling, Reinforcement Learning (RL)-based energy management strategy optimization, and control strategy simulation-to-reality transfer.To provide the high-fidelity training environment required for reinforcement learning strategies, the study optimized the system matching of the hybrid powertrain to determine the optimal component parameters and baseline control strategies. It then proposed an incremental error calibration method that combines mechanistic models with data-driven models. Using real vehicle hub testing data, a high-fidelity PHEV model capable of error correction was developed, providing a high-fidelity training environment for the development of RL-based energy management strategies. Additionally, in the integration phase of reinforcement learning strategy, this method allows for rapid model error correction based on the difference between real vehicle test data and model simulations.To facilitate the real-world application of RL-based energy management strategies, an energy management strategy architecture combining RL and optimal control is constructed based on the Equivalent Consumption Minimization Strategy (ECMS), which has the ability of anti-disturbance and can achieve stable control with inaccurate environment states. The intricate coupling of mechanical, electrical, thermal states and driving cycle results in a high-dimensional complex control problem for PHEVs, which is challenging to optimally solve within the same time scale. Therefore, a dynamic energy management strategy based on Multi-Horizon Reinforcement Learning (MHRL) is proposed, which decomposes the complex neural network into sub-networks with different time scales for collaborative training, which effectively improves the energy management effect under time-varying driving conditions and different ambient temperatures, and the real-time performance of the MHRL algorithm is verified by hardware-in-the-loop test.Finally, regarding the cross-platform application and real-vehicle integration of the energy management system, a modular agent design method and a training mechanism of meta-reinforcement learning strategy are proposed to realize the migration of vehicle power system model and driving conditions respectively. A real-vehicle integration method and online update architecture for the MHRL strategy are designed. Based on real-vehicle hub testing in a virtual driving simulation, the designed MHRL energy management system is validated, demonstrating that the method can effectively complete the strategy transfer from simulation to reality, achieve cross-vehicle platform application, and enhance the strategy adaptability to different user driving cycles.