Abstract:

To improve the efficiency of power converters， it is necessary to conduct a correlation analysis of the factors affecting the magnetic core loss of magnetic components in power converters. The core loss under the influence of specific factors can be predicted using regression methods. To improve the accuracy of core loss prediction， a data-driven method was adopted， employing a decision gradient boosting model and out-of-bag error variation to independently analyse the influencing factors. The K-means clustering method combined with the silhouette coefficient method was used to cluster the influencing factors and analyse the synergistic effects of factor combinations on core loss. Based on the importance analysis of these factors， the GhostNet neural network was used for prediction. A multi-objective genetic algorithm was used to explore the conditions under which magnetic components achieved maximum transmitted magnetic energy while minimizing core loss. Simulation results demonstrated that the proposed GhostNet-based core loss prediction method achieved excellent accuracy and strong generalization， with an coefficient of determination of 0.986 5， mean absolute error of 2.154 9×10⁴， and mean bias error of 3.418 2×10⁶ on the test set. Furthermore， the proposed multi-objective genetic algorithm exhibited excellent global search capabilities， effectively avoiding local optima and identifying a smaller Pareto front.

Key words: data-driven, magnetic core loss modeling, multi-objective genetic algorithm, core loss factor analysis, deep neural network, K-means clustering, silhouette coefficient method, power converters