Deep Learning Implementation of Model Predictive Control for Multioutput Resonant Converters

Abstract

Flexible-surface induction cooktops rely on multi-coil structures which are powered by means of advanced resonant power converters that achieve high versatility while maintaining high efficiency and power density. The study of multi-output converters has led to cost-effective and reliable implementations even if they present complex control challenges to provide high performance. For this scenario, model predictive control arises as a modern control technique that is capable of handling multivariable problems while dealing with nonlinearities and constraints. However, these controllers are based on the computationally-demanding solution of an optimization problem, which is a challenge for high-frequency real-time implementations. In this context, deep learning presents a potent solution to approximate the optimal control policy while achieving a time-efficient evaluation, which permits an online implementation. This paper proposes and evaluates a multi-output-resonant-inverter model predictive controller and its implementation on an embedded system by means of a deep neural network. The proposal is experimentally validated by a resonant converter applied to domestic induction heating featuring a two-coil 3.6 kW architecture controlled by means of a FPGA.