A Generic Single-Objective Machine Design Framework Utilizing Gaussian Process Regression and Bayesian Optimization

Abstract

Due to the complex rotor design of reluctance synchronous machines, a finite element analysis is essential for the accurate calculation of machine relevant performance objectives, like mean torque and torque ripple. This necessitates a large number of simulation steps, resulting in a high computational burden and a long simulation time per design evaluation. Therefore, an efficient optimization algorithm is required. This paper proposes a novel and generic framework for single-objective machine design optimization using Gaussian process regression (GPR) and Bayesian optimization (BO). Different kernel functions (squared exponential, Matérn, rational quadratic) and hyperparameter configurations are assessed to evaluate the regression accuracy of three optimization (performance) objectives: mean torque, torque ripple, and power factor. Bayesian optimization with the infill criterion Expected Improvement is finally applied to perform optimal machine design for a machine with 18 design variables. It outperforms the classical methods such as genetic or particle swarm algorithms as it results in (much) faster optimization results with better machine designs; even for such a high number of design variables.