Generalized MPC-DSVPWM Methods: Reduction Techniques and Explainable Machine Learning With Conformal Prediction for PMSM Drives

Addressing computational time challenges of model predictive control with discrete space vector pulse width modulation (MPC-DSVPWM) poses a significant hurdle. MPC-DSVPWM can emulate desired multilevel inverter performance by generating virtual voltage vectors (VVs). Different reduction methods have been used to choose optimal VVs; however, they are limited to low level inverters and require large initialization of VVs. Additionally, machine learning methods have been implemented; however, they require complex optimization methods, as the classification number increases exponentially with the level of the inverter. Moreover, ANNs lack explainability in choosing optimal features and confidence in their predictions. To solve the above mentioned problems. We eliminate the need for initializing and optimizing a large number of virtual VVs and propose reduction method to select optimal switching states. Secondly, we propose robust machine learning algorithm by employing explainable machine learning techniques to select important features. Then, we implement parallel neural networks to handle computations efficiently. Then, conformal prediction technique is implemented to ensure trustworthy and confident performance. The proposed methods are validated by both simulation and experimental results on permanent magnet synchronous motor (PMSM) for two-level inverter. The proposed method shows both excellent steady-state and fast dynamic performance.