Supervised imitation learning of FS-MPC algorithm for multilevel converters

Model predictive control (MPC) applications for multilevel power electronics converters are often facing problems of a high computation burden. By using supervised imitation learning, it is possible to synthesise computationally light controllers, which can capture the behaviour of computationally heavy MPC. To obtain a high performance controller, which can do the correct control actions, training data generation and pre-processing of the data are of high importance. This paper presents guidelines for training data generation and artificial neural network (ANN) controller design for a multistep-horizon finite set FS-MPC applied to neutral point clamped (NPC) converter. A particular challenge of the selected converter topology is that some control actions are used more often than others, thus the training data will be heavy skewed i.e. it will be difficult for the controller to learn when to apply these actions due to the lack of data. A workaround for solving this challenge is discussed in the paper. The performance and the robustness of the designed controller has been validated in a hardware in the loop (HIL) system, where the limitations of the synthesised ANN controller were explored. It was observed that ANN controller performance can match the performance of the FS-MPC algorithm when operating within the span of training data values and the computational burden was much lower.