Surrogate Model-Based Multiobjective Optimization of High-Speed PM Synchronous Machine: Construction and Comparison

This article employs surrogate models for largescale high-speed (HS) electrical machine optimization to reduce heavy computational burden caused by the finite element method (FEM) based on nondominated sorting genetic algorithm II. Three artificial neural networks, namely, multilayer perceptron (MLP), support vector regression (SVR), and generalized regression neural network (GRNN), and the classical Kriging model are developed based on the K-fold crossing validation method. To find out the most suitable surrogate model, a HS permanent magnet synchronous machine (PMSM) applied in electrically assisted turbocharger (EATC) considering its multiphysics characteristics is optimized by different models. A detailed comparison is provided in terms of prediction accuracy and modeling time consumption. The optimal design is verified to guarantee the accuracy of optimization results. The key contributions of this article include an automatic HSPMSM optimization process is proposed, where complete modeling and tuning method as well as a reasonable comprehensive evaluation metric of different surrogate models are conducted. It is found that invalid samples from the initial dataset contain useful information to improve prediction accuracy. Hence, an entire surrogate process should benefit from a pretrained classifier. Also, it reveals that MLP, SVR, and Kriging model can greatly improve the design optimization effectiveness with high accuracy, whereas GRNN is not suitable for this specific optimization scenario.

Automated summary

This article uses surrogate models to optimize large-scale high-speed electrical machines, reducing the computational burden caused by the finite element method. Four different surrogate models, including multilayer perceptron, support vector regression, generalized regression neural network, and Kriging model, are developed and compared using cross-validation. The results show that multilayer perceptron, support vector regression, and Kriging model greatly improve the effectiveness and accuracy of design optimization, while generalized regression neural network is not suitable for this specific optimization scenario.