An Adaptive Weight Update Based Supervised Learning Predictive Control Method for PMSM

The existing learning-based predictive control methods rely on large amounts of data for offline training. To address this issue, this article proposes a supervised learning-based continuous control set predictive method, which only requires the sampling data at the current time to achieve online updates of the controller, significantly reducing implementation complexity and data requirements. Specifically, the proposed method approximates the ideal input using a radial basis function neural network and updates the weight matrix via gradient descent. Through the derivation of the relationship between state error and tracking error, the tracking error term in the update law is replaced, enabling the online update of the approximator. Considering the inherent approximation error of the neural network and external disturbances, a robust control term is introduced to correct the approximated input. Finally, the proposed method is validated on an experimental platform.