Long-Horizon Direct Model Predictive Control Based on Neural Networks for Electrical Drives
Issa Hammoud, Sebastian Hentzelt, Thimo Oehlschlaegel, Ralph Kennel
Abstract
In this work, the use of a multilayer perceptron feedforward neural network is proposed to capture the solution of the long-horizon finite control set model predictive control (FCS-MPC) problem in electrical drive systems. The motivation behind this research is based on treating the direct model predictive control problem of a power converter as a multi-class classification problem as it consists of a finite set of switching states, which can be seen as a finite number of different classes. By simulation results and hardware in the loop (HIL) test, it is proved that the solution of the long-horizon FCS-MPC can be captured by a real-time computationally implementable neural network that recognizes the converter switching states with an accuracy of 85 - 90%. Hence, it captures the performance enhancement of long horizon FCS-MPC in a computationally efficient manner (15.84 μs).