Variable Memory Sequence Neural-Network-Based Data-Driven Predictive Control for Power Converters

doi:10.1109/jestpe.2026.3677121

Variable Memory Sequence Neural-Network-Based Data-Driven Predictive Control for Power Converters

2026-03-24

0

PRE

AI

DOI：10.1109/jestpe.2026.3677121

Original

Original request for help

Save

Abstract

En 中文

Model predictive control (MPC) is an effective solution for handling nonlinearity and multiobjective problems in power converters. However, the conventional MPC method is prone to uncertainties caused by parameter variation and external disturbances. To address this challenge, in this article, a novel ultralocal (UL) model-based variable memory sequence online learning predictive control scheme is proposed. To be specific, the original system model is replaced by a first-order UL model, where the disturbance term is estimated by a neural network with a variable memory training architecture. This proposed variable memory training architecture, endowed with a combination of historical and current operating information, can fully leverage the temporal data to enhance the estimation accuracy of the disturbance term. In view of this merit, the improved UL model is further employed in the subsequent control process, thereby eliminating the need for an explicit model. Moreover, this proposal also incorporates the merits of a multivector mechanism, which allows for low current ripple, along with a fixed switching frequency. Finally, the stability analysis of the proposal is manifested, and the performance of the proposed method is validated for a three-level neutral-point-clamped (3L-NPC) converter through both simulation and experiment.

Keywords:

Finite-control-set model predictive control (FCS-MPC)

model-free predictive control (MFPC)

power converter

robustness

Variable Memory Sequence Neural-Network-Based Data-Driven Predictive Control for Power Converters

Abstract

Journal

Organization