Impact of Sampling and Neuron Configurations on Neural Networks for PMSM Model Order Reduction in Field-Circuit Coupling Simulation

The field-circuit coupling model is widely used in studying motor drive systems due to its excellent capability to consider nonlinear factors. However, the finite element analysis (FEA) part of this model is highly time-consuming. To reduce simulation time while preserving the accuracy of the finite element model, this paper introduces a method to build a reduced order model (ROM) based on artificial neural networks (ANNs) by data sampled from FEA. Two sampling strategies are proposed and evaluated, which are current-intensive sampling (CIS) and angle-intensive sampling (AIS). Different ANN structures are constructed and trained to examine how neuron configurations affect the ANN's performance. It is found that the optimal neuron configurations under AIS mode demonstrates a better fitting capability for nonlinear factors, suitable for constructing ROM of neural networks.