Supervised Imitation Learning of Finite-Set Model Predictive Control Systems for Power Electronics

In the past years, finite-set model predictive control (FS-MPC) has received a lot of attention in the power electronics field. Due to very simple inclusion of the control objectives and straightforward design, it has been adopted in a lot of different converter topologies. However, computational burden often imposes limitations in the control implementation if multistep predictions are deployed or/and if multilevel converters with many possible switching states are used. To remove these limitations, we propose to imitate the predictive controller. It is important to highlight that the imitator is not intended to improve the dynamic or steady-state performance of the original FS-MPC algorithm. In contrast, its key role is to keep approximately the same performance while at the same time reducing the computational burden. Our proposed imitator is an artificial neural network trained offline using data labeled by the original FS-MPC algorithm. Since the computational burden of the imitator is not correlated with the complexity of the FS-MPC algorithm it emulates, implementation of much more complex predictive controllers is made possible without prior limitations. The proposed method is validated experimentally on a stand-alone converter configuration and the results confirm a good match between the imitator and the predictive controller performance. Simulation models of both controllers are provided in the supplementary files for three different prediction horizons.