Efficiency Map Prediction of Motor Drives Using Deep Learning

Abstract

In this article, a new method for predicting efficiency maps of electric motor drives is proposed using deep learning (DL). Since many operating points need to be simulated using finite-element (FE) analysis to estimate the efficiency map of a single motor drive topology with certain geometry dimensions and materials, incorporating the whole efficiency map into the design optimization process is an overwhelmingly time-consuming task and may be impossible, depending on the availability of computational resources. Therefore, two DL network architectures are employed in this work to quickly and accurately predict efficiency maps. In the first architecture, the two important stages of efficiency map calculations, i.e., the flux linkage maps and the torque-speed envelopes, are replaced by a combination of recurrent and feedforward neural networks to account for geometric and operating point variations. For the second architecture, an end-to-end DL model was trained to predict the same efficiency maps. The output of the proposed methods has a good match with that of the FE solution, indicating a high prediction accuracy as well as low run-time useful for design and optimization problems.