Uncertainty Quantification-Based Multi-Objective Optimization Design of Electrical Machines Using Probabilistic Metamodels

In the design optimization stage of an electrical machine, metamodeling has been widely used to calculate key performance indicators such as electromagnetic torque and torque ripple. It is a data-driven approach that replaces computationally expensive simulations by creating metamodels with lower computational costs. However, in practical scenarios, there exists inherent uncertainty in the modeling process. Current methods with metamodels offer deterministic estimations, thereby failing to provide further information to support the optimization. In this paper, an uncertainty quantification-based optimization design of electrical machines is proposed. First, the Bayesian neural network and Gaussian process regression as probabilistic metamodels are trained on a dataset generated by finite element analysis. Subsequently, the metamodel showing better performance is selected to aid in the following optimization process. At the end, we combine the acquired uncertainty information with multi-objective optimization and present two optimization strategies: 1) Separate optimizations using the confidence bound values or the point estimation value. ... mehr2) Highly uncertain individuals are constrained during the optimization. The results show that both strategies allow a strategic decision regarding the predictive uncertainty during the optimization. Moreover, strategy II is able to lessen the uncertainty of generated designs while reducing prediction error.