Neural Network-Assisted Model Predictive Control of a Modular Multilevel Converter

Abstract

This paper describes an advanced algorithm for current control in a modular multilevel power converter. The proposed algorithm is based on deadbeat model predictive control, which is further enhanced by applying an artificial neural network. Rather than supplanting the existing predictive controller, the neural network is used to supplement it, thereby improving control performance. The reference current tracking capability is enhanced under all operating conditions, particularly in cases where the reactive elements of the converter exhibit a relatively high resistive component. By introducing the neural network, the inherent delay of model predictive control in such scenarios is virtually eliminated, which reduces the overall control error. The network is trained offline on a set of control inputs obtained by optimizing the current response in a wide range of operating modes. Algorithm verification is performed for a three-phase modular multilevel converter in different steady-state and transient conditions. A state-of-the-art high-fidelity real-time simulator is used in both the algorithm development and verification stages.