A Novel Neural Network-Assisted Finite-Set Subspace Predictive Control Method for 3L-NPC Converters

Finite control-set model predictive control is widely employed in electric drive and power converter control due to its fast dynamic response and ability to handle multiple objectives. However, its performance is highly sensitive to system parameters, resulting in significant degradation when the model is mismatched. To address this limitation, this article proposes a novel parametric robust predictive control scheme that combines subspace predictive control with neural networks (NNs). Specifically, the nonlinear dynamics of the system captured by the NN predictor are incorporated within the real-time updated subspace predictor to accurately predict the state of the controlled system. This approach allows the controller to derive the optimal control law by utilizing only historical input and output data, enhancing the robustness of the system to parameter variations. Furthermore, we propose a capacitor voltage balancing strategy for a three-level neutral-point-clamped converter that does not require knowledge of capacitor capacitance or weighting factor adjustment. Finally, the availability of the proposed scheme is corroborated by a comprehensive comparison with existing control schemes in simulation and experiment.