Automatically Tuning of Weighting Factors for FCS-MPC in PMSM Drives Using Lightweight Neural Networkweight Neural Network

The deep neural network (DNN) has recently emerged as a compelling approach for predicting the weighting factors of finite control-set model predictive control (FCS-MPC). Nonetheless, the considerable size of DNN parameters necessitates additional computational resources. To address this concern, this paper proposes a lightweight network (L-DNN) strategy to reduce the number of parameters and computational demands for DNN. The proposed method utilizes TensorRT for tensor decomposition and quantization for the trained DNN. As a result, the proposed method occupies fewer logic resources in FCS-MPC, providing the desired behavior with fast dynamic response. Comparative simulations demonstrate that the designed lightweight strategy significantly reduces DNN parameter size without compromising the predictive accuracy. Finally, the dynamic adjustment of weighting factors is validated through simulations on the permanent magnet synchronous motor drives fed by a three-level neutral-point-clamped inverter. More demonstrations and experimental validations will be presented in the full paper.