A Linear Machine Learning-Based Model Predictive Control With Repetitive and PI Elements for a Three-Phase Inverter

Abstract

Existing machine learning (ML) based model predictive control (MPC) methods are either inferior to the online optimized with quadratic programming (QP) MPC or have high computational complexity and cannot be implemented in the resource-limited digital signal processor (DSP). This paper solves these two issues by using linear ML methods and adding extra interpretable features. First, the intrinsic linearity of the training data generated by the QP-MPC has been theoretically proved such that the linear ML methods, e.g., linear neural network (LNN) and linear support vector regression (LSVR), can be used to capture the linearity characteristics of the training dataset. The linear operation significantly reduces the computational complexity from O(2ⁿ) to O(1) so that they can be implemented in the DSP. Second, extra features with the repetitive and proportional integral (RPI) elements are added as input to the linear ML-based MPCs. Experimental studies with QP-MPC, LNN-MPC, and LSVR-MPC with RPI elements are carried out under linear and nonlinear load conditions. The results show that linear ML-based MPCs are superior to the QP-MPC in power quality and tracking errors. Moreover, the linear ML-based MPCs outperform the QP-MPC under the parameter mismatch and two-degree-of-freedom (2DOF) controllers, demonstrating their adaptive capabilities. This article is accompanied by a video demonstrating the real-time operation.