Electric Machine Design by a Novel Fast Model Predictive Optimization Strategy Treating Dimension-Expensive Problem

As a crucial preprocessing for motor products, design optimization methodologies have been widely explored. However, existing studies always indicate some issues. Analytical model-based optimization involves complex mechanisms. The surrogate-based optimization can hardly satisfy the accuracy of multiple objectives or in scenarios with high dimensions. The numerical tool-based optimization is time-consuming even with the multilevel technique. Accordingly, a novel fast model predictive optimization (MPO) strategy is proposed. It follows a parametric analysis stage to quantify the intrinsic mechanisms between variable-objective pairs. The space-filling sampling technique and parameter correlation are applied. By searching in a library of machine learning (ML) algorithms, MPO constructs a trustable model with only limited case evaluations. To treat the objectives with different sensitivities, the proposed MPO strategy learns the algorithms, hyperparameters, and parameters. A novel ranking regularization is introduced to the ordinary loss function for examining the model accuracy or validity. Hence, a finite-element analysis (FEA) can be avoided in the subsequent optimization. After performing the sensitivity analysis, the optimal design of permanent magnet machine is prototyped and tested to verify the proposed design methodology. In short, MPO finds a more complete Pareto frontier more efficiently than the FEA-based optimization.