High-Frequency Core Loss Modeling Based on Knowledge-Aware Artificial Neural Network

High-frequency core loss modeling plays a critical role in the magnetics design of power electronics. However, existing modeling tools fail to achieve both high speed and high precision. The conventional analytical approach enables fast estimations but performs poorly in accuracy. Magnetic loss mode...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE transactions on power electronics 2024-02, Vol.39 (2), p.1968-1973
Hauptverfasser: Deng, Junyun, Wang, Wenbo, Ning, Zhansheng, Venugopal, Prasanth, Popovic, Jelena, Rietveld, Gert
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:High-frequency core loss modeling plays a critical role in the magnetics design of power electronics. However, existing modeling tools fail to achieve both high speed and high precision. The conventional analytical approach enables fast estimations but performs poorly in accuracy. Magnetic loss models aided by loss maps feature high accuracy, but their model parameterization relies on large data. The emerging approach of artificial neural networks (ANNs) provides a promising alternative since it can achieve high speed and accuracy. However, conventional implementations of ANN require a large and accurate dataset for training, which is hard to achieve in magnetic loss modeling. To solve this problem, a knowledge-aware artificial neural network (KANN) is proposed that can achieve high accuracy with small training datasets. After introducing the principle of the proposed KANN, it is applied to high-frequency core loss modeling using the improved generalized Steinmetz equation as additional knowledge. To validate the performance of the proposed KANN-based design method for core loss modeling, it is applied to predict the losses of two ferrite cores in the frequency range of 50-450 kHz. The results show that the proposed method greatly outperforms present loss modeling approaches in accuracy and speed, requiring only a limited training dataset. An automatic loss modeling tool based on the new method is provided together with its open-source code.
ISSN:0885-8993
1941-0107
DOI:10.1109/TPEL.2023.3332025