Research on Reinforcement Learning Model Predictive Control for Diesel Engine VGT-EGR-TVA System

This paper proposes a novel weight coefficient self-learning expanded state model predictive control (MPC) algorithm for addressing the multi-objective tracking control problem in a diesel engine equipped with a Variable Geometry Turbocharger (VGT), Exhausted Gas Recirculation (EGR), and Throttle Valve Actuator (TVA). The algorithm incorporates a simplified operating mechanism to establish a linear time-varying prediction model and employs an expanded state observer to facilitate real-time disturbance observation and compensation. In addition, the Q-learning reinforcement learning algorithm is utilized to learn the weight coefficients, which results in dynamic weight coefficients that can effectively adapt to the system's time-varying and nonlinear features and changes over a long-time scale. Through simulation experiments, the algorithm demonstrates superior tracking control performance compared to traditional PID algorithms, with a significant reduction in the overshoot of pre-turbine pressure under load step conditions by 32.4%. Additionally, the algorithm reduces the tracking errors of intake manifold pressure and EGR rate in the dynamic driving cycle, as quantified by IAE (Integrated Absolute Error), by 31.2% and 21.5%, respectively, in comparison to traditional PID algorithms.