Deep reinforcement learning-based optimal decision-making framework for eco-driving in connected hub-motor electric vehicles

The integration of V2X communication in electric vehicles introduces new opportunities for energy-efficient and intelligent control, but also poses challenges for battery management and powertrain coordination. This paper presents a co-simulation framework of a connected hub-motor electric vehicle (CHMEV) in Simulink and SUMO, incorporating detailed models of the battery, in-wheel motors, and V2X modules. A multi-step optimal-based decision-making architecture is developed, combining deep reinforcement learning (DRL) with model predictive control (MPC) to compute optimal acceleration profiles for adaptive cruise control in V2X-enabled environments. A key contribution is a novel methodology to embed environmental constraints directly into the DRL training process via early-stage constraint-aware training, ensuring strict compliance with operational and safety boundaries. A torque allocation module was established, and a comprehensive analysis was conducted comparing three approaches to demonstrate the impact of different driving styles on battery and energy savings in both V2V and V2I scenarios. Simulation results demonstrate up to 22.2% improvement in energy efficiency under the proposed method in a V2I scenario, along with significant potential to extend battery life and vehicle driving range by applying a blended brake operation. The findings highlight the advantage of merging data-driven and model-based approaches for connected EV control.