Integrated control of the longitudinal acceleration oscillations and longitudinal tyre slip through the correction of the powertrain torque

The rapid diffusion of electric vehicles (EVs) has stimulated significant advancements in vehicle control system research, aiming to enhance both performance and comfort. One particularly promising approach involves the use of four on-board electric motors that can independently control the torque at each wheel. This provides high flexibility in managing the vehicle’s dynamics such as traction, stability, and handling capabilities. However, it also introduces complex challenges in real-time control, especially when aiming to mitigate longitudinal vibrations, prevent jerk effects and maintain optimal traction across a range of road conditions. The problem of real-time control can be anyway managed implementing innovative technologies and control methodologies such as artificial intelligence, which are less computationally expensive in terms of time and hardware needed. This thesis focuses on the design and implementation of a Nonlinear Model Predictive Control (NMPC) strategy that integrates three essential functions: vibration reduction, anti-jerk control, and traction control. The NMPC controller’s main objectives are reducing longitudinal vibrations, which can compromise ride comfort and vehicle stability, while also preventing unwanted jerk (sudden changes in acceleration). Additionally, the system incorporates traction control to maintain optimal tire-road interaction and ensure stability under varying driving conditions. A key innovation of this work lies in the adaptive nature of the NMPC controller's cost function. In conventional NMPC designs, the weight parameters that balance the importance of vibration reduction, jerk minimization, and traction control are typically fixed. However, in this approach, these weight parameters are dynamically adjusted based on an offline-trained reinforcement learning (RL) algorithm. Specifically, an actor-critic method is employed, where the actor learns to select the optimal weight parameters for the control objectives, and the critic evaluates the effectiveness of these decisions. In this project, the reinforcement learning training is conducted offline. The actor-critic framework is used to explore various control policies by adjusting the weight parameters, with the critic providing feedback on the effectiveness of these policies based on a predefined set of performance metrics. The actor then refines its policy over several iterations to converge on a set of weight parameters that optimize the NMPC controller’s performance on a set of different driving conditions. Once trained, the learned weight parameters are used to inform the real-time operation of the NMPC controller, allowing it to adapt to varying conditions without requiring further online learning. The offline nature of the RL training ensures that the system can learn an optimal control policy without the computational overhead of online learning during actual vehicle operation. This thesis aims to develop a comprehensive NMPC-based control system that integrates offline reinforcement learning to dynamically adjust the weight parameters of the cost function. The result is a highly adaptive and robust controller capable of improving vehicle stability, ride comfort and overall performance in electric vehicles with multi-motor drivetrains. By combining the real-time optimization capabilities of NMPC with the offline-learned adaptive weight adjustments, this research represents a significant step toward more intelligent and efficient vehicle control systems.