Real-time MPC Design Based on Machine Learning for a Diesel Engine Air Path System

Abstract This study investigated a control design method based on machine learning to achieve non-linear model predictive control (MPC) at a low computational load. In addition, we describe examples of the application of this method to a diesel engine air path system. The solution to the optimal control problem determined at each point in time by MPC depends on several parameters at that time. Thus, if the relationship between the solution and the parameters could be approximated in advance using machine learning, solving this problem online would become unnecessary, and the control computation time could be reduced. We designed a controller that operates the valves of the air path system using this method and used a simulation to verify that this resulted in a favorable tracking performance of the target values. The computation time of the approximated MPC controller was 0.022 ms.