Adaptive Barzilai--Borwein-Based Separable Spline Kernel Neural Network Predictive Control for Three-Level NPC Converters

Abstract

The finite control-set model predictive control (FCS-MPC) deploys an accurate mathematical model to describe the system dynamics and achieve output trajectory tracking, but this method has inherent limitations, including strong dependence on the model and sensitivity to parameter. In response to these challenges, a separable spline kernel-based predictive control strategy is developed in this study that employs the B-spline basis functions to construct a neural network control model in place of the explicit mathematical model and incorporates an adaptive Barzilai–Borwein-based gradient update to ensure stable system dynamics while reducing computational burden. The proposed method achieves superior performance in metrics such as computation time and output harmonics compared with the conventional FCS-MPC method, while maintaining robustness under load variations, offering a new control paradigm for power electronic converters that alleviates dependence on accurate parameters. Finally, the Lyapunov stability of the proposed scheme is provided, and extensive simulations and experiments are conducted for three-level neutral-point-clamped (3L-NPC) converters, which validate the feasibility and interest of the proposed design.