It is of great significance to accurately optimize the design of high-frequency transformer (HFT) to improve the comprehensive performance of itself and its system. However

the existing HFT optimization design methods usually use simplified analytical formulas to calculate electromagnetic and thermal parameters such as core

winding loss and hot-spot temperature

although the calculation speed of these analytical formulas is fast

their accuracy is low

which leads to the low accuracy and reliability of the optimized design results. In order to take both the accuracy and speed of HFT optimization design into account

this paper introduces a surrogate model of HFT electromagnetic thermal parameters based on least squares support vector machines (LS-SVM) for the first time

and proposes a new HFT optimization design method based on non-dominated sorting genetic algorithm Ⅱ (NSGA-Ⅱ). In order to verify the superiority of the LS-SVM surrogate model proposed in this paper

the results show that the LS-SVM has higher accuracy compared with the recurrent neural network and deep neural network surrogate model. Finally

based on the proposed optimization design method

a 5 kHz/10 kW HFT is optimized using multiple objectives

and the optimal design scheme is verified by finite elements method

and the results show that the proposed surrogate model of core loss

winding loss and hot-spot temperature has lower errors compared to the corresponding traditional analytical ones

with errors of 2.77%

3.03% and 0.92% respectively. Thus the accuracy and reliability of the proposed optimization design method is verified.