Least square support vector machine network-based modeling for switched reluctance starter/generator

This paper presents the least square support vector machine (LS-SVM) network as a new tool to develop the model of the switched reluctance machine (SRM) and predict the dynamic performances of SRM system. The basic premise of LS-SVM regression is that it forms a very efficient mapping structure for the nonlinear SRM. By using the measured sample data of SRM, the LS-SVM is designed to learn the nonliear magnetization data with rotor position and phase current as input, and the corresponding flux linkage and torque as output. It has a good capability of generalization and is computationally efficient. With the developed modeling method, a LS-SVM current-dependent inverse flux linkage model and a LS-SVM torque model are used to simulate the dynamic performances of a 6/4 SRM operating as a starter/generator, and the accuracy of the model is tested via comparison to the measurements of steady state phase current characteristics.