Abstract
The increasing prevalence of Electric Vehicles (EVs) on highways is making energy-efficient control decision-making essential. However, optimizing regenerative braking (RB) improves energy recovery, extends driving range, and enhances safety and efficiency. This study proposes an explainable deep reinforcement learning (DRL) framework, Recurrent Proximal Policy Optimization based on Long Short-Term Memory (RPPO-LSTM), to optimize RB in real-world driving conditions. The proposed shared LSTM-based Actor-Critic architecture captures temporal dependencies in high-dimensional driving sequences, enabling efficient policy learning while enhancing regenerative energy recovery. Firstly, a custom OpenAI Gym environment was designed to model realistic EV braking dynamics. Secondly, the model was trained and validated on the publicly available, real-world, downloadable dynamometer dataset (D3). The proposed RPPO-LSTM framework demonstrated superior learning performance, with lower training loss than the PPO strategy, and achieved a 22% improvement in cumulative regenerative power compared with the baseline, while reducing comfort and friction losses and maintaining smooth braking control. RPPO-LSTM achieved the lowest average of total loss (0.0113), value function loss (0.0581), entropy loss (-1.188), Kullback–Leibler (0.00144) and Policy Gradient Loss (-0,005) compared to the PPO strategy. Finally, two complementary interpretability methods were employed: Mutual Information (MI) to select relevant input control signals prior to training, and SHAP to explain the trained model’s decisions post-training and quantify the influence of driving dynamics on energy regeneration. This approach outperforms conventional control strategies by enabling adaptive, real-time energy recovery while maintaining safety and mechanical integrity, showing that DRL can develop more intelligent RB systems for future and autonomous EVs.

This publication has 12 references indexed in Scilit: