Bayesian Regularization-Based MPC for a Hybrid Modular Multilevel Converter

Modular multilevel converters (MMCs) have become one of the applicable structures for voltage source converters (VSCs) due to their salient features including their applicability to different voltage levels, low harmonics response, high efficiency, and fast transient response. However, one of the challenges of using MMCs is their controlling system. Conventional MMC controlling techniques are ineffective and exhibit delayed transient responses when a load changes or a fault occurs. The use of model predictive control (MPC) has been suggested as a viable solution to the problems with conventional controlling techniques. The MPC provides MMCs with a multi-objective controlling system and improves their dynamic response. However, the MPC-based methods have computational challenges in finding the optimum switching state. This paper develops a neural network-based MPC approach to tackle the time-consuming performance of the traditional MPC method. In this study, the Bayesian regularization (BR) back-propagation neural network is applied to train the algorithm. The trained machine is applied to control a hybrid MMC, which includes full-bridge and half-bridge submodules. To highlight the efficacy of the proposed BR-based MPC method, the results of circulating current, transient responses, and harmonic reduction, indicating a viable performance are compared with the traditional finite set MPC.