Optimal System Excitation of a Permanent Magnet Synchronous Motor Using Differentiable Model Predictive Excitation

Comprehensive models of nonlinear permanent magnet synchronous motors (PMSMs) are necessary to design control strategies that fully utilize the capabilities of such motors and to reliably monitor their performance. Traditionally, such models have been built through extensive system knowledge combined with parameter or flux linkage look-up table estimation. The trend towards data-driven machine learning is increasingly being adopted in PMSM-based applications, reducing the need for engineering knowledge and allowing natural incorporation of parasitic effects into the models. However, extensive data sets must be gathered from experiments to support the creation of such models. Additionally, the corresponding experiments should be kept minimal to reduce costs and time demand. The recently proposed differentiable model predictive excitation (DMPE) algorithm addresses this problem by generating well distributed data sets from unknown, dynamic systems. In this work, three variants of DMPE are applied to a numerical PMSM experiment and the resulting data sets are compared to the data sets generated by a state-of-the-art system excitation method and by two heuristic baselines. The results of the experiments show that, according to multiple coverage metrics, the DMPE variants outperform the other methods in generating balanced data sets over a variety of motor speeds.