Efficient Data-Driven Modeling of Core Loss in Magnetic Materials for Power Electronics Systems

The accurate modeling and prediction of magnetic core losses are of great significance in the design of power electronic systems. Although traditional empirical equations (such as Steinmetz equation and its improved forms) are easy to calculate, their prediction accuracy is limited under multi material, multi temperature, and multi waveform conditions, making it difficult to meet the needs of high-frequency and high-power density applications. With the development of artificial intelligence, data-driven methods are gradually becoming more optimal solutions. Traditional machine learning models can achieve high accuracy in static features, but still rely on manual feature engineering for complex waveform sequences, resulting in information loss. The introduction of deep learning methods provides a new path for this. This paper proposes a CNN-FCNN architecture Mixed Neural Network (MNN) that can simultaneously process discrete, continuous, and waveform sequence features, achieving unified prediction across materials, temperatures, and waveforms. Based on the MagNet database, the results show that MNN significantly outperforms MLP-LSTM in terms of prediction accuracy and generalization. Further combining XGBoost to construct a weighted hybrid model achieved the highest predictive performance (R2=0.997). Data driven method can break through the limitations of empirical equation and single model, and achieve high-precision general modeling of core loss. This method not only reduces the cost of repetitive modeling for a single material or specific operating conditions, but also provides new ideas for building universal power electronics design tools.