Surrogate-assisted Multi-objective Design Optimization Method And Its Applications For Electrical Machines

Electrical machines are widely employed as essential energy conversion components in many advanced industry fields including aerospace,electrified transportation and industrial robotics.The increasingly high requirements of machine characteristics have posed strong challenges on conventional design optimization methods that prominently rely on numerical simulations and manual design experience due to the demerits including low computation efficiency,poor optimization accuracy,and weak generality,which make them difficult to fulfill the requirements of multi-objective design optimization for machines.Based on the data-driven concept,the surrogate-assisted optimization method that can comprehensively integrate knowledge from multiple disciplines provides an effective approach for solving multi-objective optimization issues in advanced machines.However,the existing surrogate-assisted optimization methods are difficult to achieve high accuracy and efficiency simultaneously,as the results of complex geometrical topologies of machines,non-linearity of electromagnetic materials,and strong coupling between machine characteristics,which reduces their practical applicability.This thesis focuses on novel multi-objective surrogate-assisted design optimization techniques for electric machines with particular emphases on design of experiment(DOE),surrogate model generation and multi-objective optimization methods.The goals of the research include improving the suitability of the surrogate-assisted optimization method for electric machine design and expanding the application scope of this method through the comprehensive applications of new techniques,which theoretical and practically eases the application issues of surrogate-assisted optimization methods in machine designs.Firstly,a novel DOE method based on constrained Latin hypercube design technique is proposed.In the non-hypercubic constrained space formed by complex constraints between variables,the continuous local enumeration method and K-means clustering algorithm are employed to improve the utilization of design points.Subsequently,the spatial distribution uniformity and orthogonality of design points are improved based on a multi-criteria optimization criterion,which secures the accuracy of surrogate models.The proposed method can substantially enhance the adaptability of DOE methods to machine design optimization problems.Secondly,the thesis provides a novel surrogate model generation strategy using transfer learning of the analytical model.By reducing the difference between regression hyperplanes corresponding to the data from analytical models and that from the finite element analysis(FEA)in the transfer learning process,the required cases of FEA for training surrogate models with high global generalization can be significantly reduced with the assistance of computationally efficient analytical model.The proposed method has been validated by the torque characteristics optimization of a surface-mounted permanent magnet synchronous motor(PMSM)with resolvable optimization targets that provides a new optimization concept of combining model-driven and data-driven methods.Then,an online surrogate-assisted optimization strategy based on local additive points of the Pareto frontier has been given.By performing local sensitivity analysis and adding sampling points near the Pareto frontier obtained from different rounds of optimization,the local generalization ability of the surrogate model is improved with online cyclic updating,which enhances the optimization accuracy of robustness indicators sensitive to local features.The method eases the difficulty of robust optimization for high quality motors considering various manufacturing errors.It has been validated by the robust optimization of an interior PMSM that with high requirements for performance robustness.The performance robustness and theoretical yield rate of products can be significantly improved compared with the deterministic optimal design ignoring manufacturing errors.Finally,an integrated technique that employs all proposed methods is applied to deal with the challenge of multi-objective optimization of large-scale synchronous machines for transient parameters.An arbitrary rotor position standstill time-domain response transient parameter identification method is proposed to identify two-axes transient parameters within a single simulation test case,which enables efficiently obtaining training samples.By using transient circuit models and sensitivity analysis,a multilevel optimization strategy for transient parameters is proposed for decoupling optimization objectives and reducing difficulties in generating high-precision surrogate models.The proposed technique provides an effective solution for transient parameter optimization of large-capacity machines including both design optimization and test verification methods.It has also been experimentally validated in a small-capacity dynamic characteristics simulation prototype in which the transient parameters match well with a large-scale synchronous generator.