A Study on Self-Learning Energy Management Strategies for Personalized Driving in Plug-in Hybrid Electric Vehicles Based on Deep Reinforcement Learning

To address the issue of reduced fuel efficiency in plug-in hybrid electric vehicles (PHEVs) caused by differences in driving styles, this paper proposes a personalized, self-learning energy management strategy (EMS) that integrates multi-feature driving style analysis with the Deep Deterministic Policy Gradient (DDPG). The study focuses on incorporating driving-style information into energy-management decisions. Firstly, based on CAN bus data, a Gaussian mixture model (GMM) is constructed using the expectation-maximization estimation method. By using weighted aggressive scores and preset thresholds, driving behaviors are classified into conservative, moderate, and aggressive types, thereby establishing personalized driving style representations. Second, a longitudinal dynamics model of the hybrid electric vehicle is established, incorporating the Pacejka tire model and a braking system model. Finally, a reward function is designed to minimize fuel consumption while penalizing excessively low battery state of charge (SOC). The driving-style code is included in the controller state, and DDPG jointly determines motor torque and the continuously variable transmission (CVT) ratio. Gaussian noise is added to the Actor output during training to explore the action space. A rule-based control strategy is adopted as a comparison strategy, and the proposed strategy is evaluated using the MATLAB (R2024b) platform. Over the 1980’s composite driving trajectory, equivalent fuel consumption decreases from 0.853 liters (L) to 0.705 L, a reduction of 17.34%. These results support the feasibility of the integrated strategy under the evaluated simulation conditions.