Model-Free Predictive Controllers for Voltage Source Inverter: A Unified Experimental Benchmarking

Abstract

This article presents an extensive experimental comparison of different predictive current control strategies applied to a two-level voltage source inverter supplying a resistive–inductive load. All controllers were implemented on the same hardware platform and evaluated under identical operating conditions at a controller sampling frequency of $\text{20}\,\text{kHz}$. The evaluated methods include a conventional model-based (MB) predictive controller, an auto-regressive model with exogenous input and recursive least squares estimation (ARX–RLS), and three first-order ultra-local (UL) formulations: the linear extended state observer, the least absolute shrinkage and selection operator-based estimator and the extended-affine (EA). Experimental results confirm that the MB achieved the lowest computational time, whereas the ARX–RLS controller attained the lowest total harmonic distortion (THD). The EA-UL method provided the best overall tradeoff between harmonic performance, transient dynamics, and computational effort. The study highlights that model-free-predictive controllers (MF-PC), particularly ARX–RLS and EA-UL, can deliver performance comparable to or better than MB while significantly reducing dependence on system modeling. These findings confirm the potential of MF-PC approaches as efficient and robust solutions for next-generation power electronic converters.