Stability Control and Energy Management of an All-Wheel Drive Electric Vehicle Using Artificial Intelligence Methods

This thesis presents the development of reinforcement learning (RL)-based control frameworks for torque vectoring and energy optimization in all-wheel-drive (AWD) electric vehicles (EVs). The proposed model-free controllers aim to enhance vehicle dynamic stability and energy efficiency simultaneously under nonlinear and uncertain driving conditions. A detailed vehicle model with seven degrees of freedom (7 DOF) and a Pacejka tyre formulation is employed to capture realistic vehicle behaviour. Three deterministic actor–critic algorithms, i.e., deep deterministic policy gradient (DDPG), twin delayed deep deterministic policy gradient (TD3), and curriculum learning-enhanced TD3 (CL TD3), are designed and compared against conventional model-based controllers. The RL agents are trained using a multi-objective reward function that incorporates yaw stability, sideslip angle regulation, tire slip limitation, and energy efficiency based on real electric machine maps. The simulation and verification processes are carried out using the IPG CarMaker environment, which provides a high-fidelity virtual test platform for realistic vehicle dynamics assessment. Simulation results demonstrate that the proposed TD3 and CL TD3 controllers achieve the best trade-off between dynamic stability and energy efficiency, with faster convergence and improved robustness across varying driving conditions. The findings confirm that RL-based control strategies provide an effective and computationally efficient alternative to traditional model-based approaches for real-time torque vectoring and energy optimization in EVs without the requirement for an explicit model of the vehicle.