Fast Nonlinear Model Predictive Control Algorithm with Neural Approximation for Embedded Systems: Preliminary Results

Abstract

This work presents preliminary results of research concerned with a fast nonlinear Model Predictive Control (MPC) algorithm implemented in an embedded system. In order to obtain a computationally efficient solution, a linear approximation of the predicted trajectory of the controlled variables is calculated for each sampling instant on-line which leads to a quadratic optimisation problem. Furthermore, the matrix of derivatives, which defines the linearised trajectory, is not determined analytically, but it is calculated (approximated) by a specially trained neural network. In order to show effectiveness of the discussed approach, a dynamic process with two inputs and two outputs is considered for which not only simulation results, but also results of real experiments performed in an embedded system based on a microcontroller are given.