Neural network-based reference generator for capacitor voltage control in modular multilevel cascaded converters

Read full text Summarise with VIRAT
Authors: Pablo Jiménez Cavallar
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Modular Multilevel Cascaded Converters (MMCC) have emerged as a key technology for the transition to low carbon emissions energy sources and are suitable for applications ranging from HVDC transmission to medium-voltage drives due to their scalability and voltage quality, enabling the integration of photovoltaic and wind energy farms. However, the high number of internal elements requires various complex nested and parallel control layers. A critical limiting factor is the intercluster energy balance (ICB), since operation at near equal input-output frequency magnitudes induces large imbalances, often necessitating oversized, expensive capacitors. While injecting circulating currents (CC) and common-mode voltage (CMV) can mitigate this, CCs increase resistive losses, and high CMV could damage motor loads. Consequently, determining optimal references to perform the ICB without compromising performance remains a difficult nonlinear problem. This work proposes and implements a feedforward neural network (FNN)-based ICB strategy that approximates a long-horizon model predictive control (MPC). Although MPC is essential for handling the nonlinear, multi-input multi-output (MIMO), and time-varying nature of the system dynamics and constraints, its computational burden is often prohibitive for long control horizons. To address this, a block coordinate descent (BCD) nonlinear optimization algorithm is used to decompose the optimal control problem into a sequence of constrained quadratic programming (QP), providing an optimal policy and acting as an expert. Then, a simple FNN is trained to imitate this decision-making process, guided by the structural properties of the control policy, effectively reducing real-time computational demand. This approach fully utilizes the converter’s internal degrees of freedom with a control horizon that would otherwise be non-real-time realizable, generating improved control references that outperform state-of-the-art methods regarding power losses and energy oscillations. The scheme is evaluated through simulations in Simulink and real-time hardware-in-the-loop (HIL) testing under demanding variable-frequency drive maneuvers. Results demonstrate that the proposed method successfully reduces energy imbalances and circulating current-induced losses, improving the closed-loop performance of this control task.

Fields of Study & Subjects

HVDC Systems and Fault Protection Multilevel Inverters and Converters Advanced DC-DC Converters

Publication details

Journal / source
CRIS
Publication date
25 March 2026
Language
EN