A Neural Network-Based Dynamic Cost Function for the Implementation of a Predictive Current Controller

Abstract

Predictive current controllers based on finite control set-model predictive control (FCS-MPC) have been extensively used in power converters. One of the clear advantages of FCS-MPC is that several control targets and constraints can be included in a cost function and simultaneously controlled. In order to establish the importance of one controlled target in relation to the others, weighting factors are used. Once the weighting factors have been tuned, they remain unchanged. This paper presents a neural network-based novel approach to the problem, in which weighting factors are tuned online as a function of several merit figures and references. This adaptive method updates online the weighting factors in the cost function when either the merit figures or the references change to boost the performance of the controller. This strategy is validated through simulations and experiments are carried out on a three-level neutral point clamped converter. The results are compared with the conventional FCS-MPC, which is based on static cost functions.