Design & Operating Point dependent Surrogate Models for PMSM

Abstract:
Surrogate models have become an essential part of the optimization of electrical machines. Current approaches mainly focus on the prediction of post-processed machine characteristics (direct surrogate models). This implies an enormous data loss during the interpolation. In this paper, operating pointdependent surrogate models are developed. These models are capable of mapping the flux linkage and loss maps of any machine in the design space without any further simulation. In order to identify suitable models, different machine learning methods and architectures are compared. The final models are able to predict 95% of the flux linkages with a relative absolute error of less than 5 %. Subsequently, a loss optimization is performed on the test data using the current dependent maps generated by the model. The resulting speed-torque curves achieve a mean relative absolute error of 0.31 %, and the efficiency maps a mean absolute error of 0.082 percentage points compared to the simulation.