An Improved Label Dataset Generation Method of Neural Network Based Finite Set Model Predictive Control For Two-Level Grid Connected Converter

This paper introduces an enhanced label dataset generation method for neural network-based finite set model predictive control (NNFS-MPC) applied to two-level grid-connected converters. The neural network emulates the classical FS-MPC and is trained offline to expedite the FS-MPC process by eliminating the need for online optimization. Conventionally, the label dataset for training the imitation neural network controller consists of one-hot encoded indices of the selected voltage vectors, turning the neural network training problem into a multi-categorical task. However, as demonstrated in this paper, such a label dataset loses significant system information and is inefficient, necessitating an excessively large training dataset. In contrast, an improved label dataset created using the probability distribution of the selected voltage vector is more informative. Rather than treating the imitation learning problem of FS-MPC as a multi-categorical task, it is more appropriate to consider it as a regression problem to learn the probability distribution. Simulation results indicate that the proposed improved label dataset generation method can substantially reduce the size of the training dataset while achieving performance comparable to the conventional FS-MPC controller.