Machine Learning-Driven Model Predictive Control of Modular Multilevel Converters

The Modular Multilevel Converter (MMC) has garnered significant interest recently due to its superior harmonic performance and improved efficiency in high-voltage direct current electrical grids. Model Predictive Control (MPC) is widely adopted for the MMC applications, as it provides a straightforward control design, facilitates the inclusion of multiple control objectives through a flexible cost function formulation, and offers excellent control performance. An emerging and promising solution involves integrating MPC with machine learning (ML)-based models, in which neural networks learn MPC behavior and predict the results as the traditional MPC does.In this paper, a multi-layered neural network is designed to approximate the control behavior of MPC correctly, enabling a substantial reduction in computational effort during real-time operation and replacing the complex optimization routines of MPC with lightweight neural network regression models that are both efficient and decoupled from the algorithmic complexity of traditional MPC. The performance of controllers is evaluated under both small and large disturbances in active power and reactive power.