Machine-Learning Aided Multiobjective Optimization of Electric Machines— Geometric-Feasibility and Enhanced Regression Models
Adrian-Cornel Pop, Zhaofeng Cai, Johan Gyselinck
Abstract
This article deals with the optimization of electrical machines by means of artificial neural network (ANN)-based classification and regression models. Geometrically (or otherwise) unfeasible designs are detected with high accuracy during the optimization process by means of an <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ad hoc</i> classification model, whereas continuous targets are predicted through regression models. Training samples are generated with an expensive finite-element (FE) model, resulting in small training sets. Moreover, the design optimization normally involves multiple but correlated subobjectives. The correlation can be leveraged using chained regression or a multioutput ANN; it is shown that both methods can achieve higher predictive performance than predicting the targets separately. The developed methods are successfully applied to a permanent-magnet synchronous machine (PMSM) with 12 geometric parameters and four subobjectives, and considering both no-load and load operation. The results show very good predictive performance of the ANN models and a significant reduction of the computational effort. The optimization can thus be run several times, with, e.g., modified weighting of the subobjectives, with little extra cost.