Neural Network-Based Optimized Control for Enhancing Voltage Support of Grid-Forming MMCs Under Voltage Sags

Abstract

With the integration of renewable energy and power-electronic devices, grid-forming modular multilevel converters (GFM-MMCs) play a critical role in active grid support. An AC grid voltage sag can trigger a large support current, which may cause large voltage fluctuations in submodule capacitors and arm overmodulation, thereby threatening system safety. This paper proposes a multidimensional collaborative method to improve the support capability of grid-forming MMCs under severe grid voltage sags. The multidimensional physical constraints of internal energy fluctuation during fault transients are clarified. The corresponding safe operating boundaries are then established, after which a coordinated optimization strategy is developed. This approach integrates second-harmonic circulating current and zero-sequence voltage injections. Offline optimization utilizes a particle swarm optimization (PSO) algorithm across the full operating range. Expanding the safe P–Q operating region requires no extra hardware costs. A neural network enables a millisecond-level direct mapping control architecture. This architecture addresses the long online computation time of traditional heuristic algorithms by embedding offline optimization data into the network weights. The trained network performs rapid forward computation to generate optimized commands, which is verified by a hardware-in-the-loop (HIL) experiment. The experimental results verify the effectiveness of the proposed method, with clear performance improvements being observed. The strategy suppresses capacitor-voltage peak and prevents overmodulation. This directly improves the MMC support capability during severe faults.