基于物理嵌入神经网络与记忆增强强化学习的分布 式电动汽车转矩分配方法

For the problems of insufficient environmental-model accuracy, limited training-condition representativeness, and weak temporal awareness in distributed-drive energy management, this study proposes an energyoptimal torque distribution method based on a high-precision environmental model and memory-enhanced reinforcement learning. Firstly, a Physical Embedded Neural Network （PENN） energy model for distributed powertrain systems is developed, which combines physical modeling with neural networks to improve energy-consumption estimation under complex driving conditions. Secondly, an online training framework is established using the PENN model and real-world vehicle-road input, where an LSTM-TD3 network is adopted to enhance temporal memory. The results show that the proposed PENN model yields significantly lower energy-consumption RMSE than pure neural-network models and reduces RMSE by 26.45% compared with pure physical models. Under complex conditions, the proposed strategy achieves 94.33% of the dynamic-programming reference performance, while reducing energy consumption by 1.3% and 12.54% compared to the conventional TD3 and rule-based strategies, respectively.