Accurate and Efficient Magnetic Modeling of Medium-Frequency Transformers Based on Artificial Neural Network

Artificial Neural Network (ANN) technology provides a promising solution to balance model precision and compute cost in the magnetic modeling of Medium-Frequency Transformers (MFTs). However, because of the complex geometry feature and material property in MFTs, the existing ANN architecture requires vast amounts of training data to describe the mapping relationship between dimension parameters and magnetic data. In this work, a physics-informed ANN-based modeling framework is proposed to accurately predict the leakage inductance and winding losses in MFTs. The model inputs, outputs, and model structure are optimized to improve the model's physical interpretability, which further alleviates the compute burden at the data generation stage. Notably, the proposed framework enables independent loss outputs for each winding layer, benefiting the thermal optimization of MFTs. According to evaluation results, with a processing time ten times less than the finite element simulation, the maximum prediction error of this proposed magnetic model is around 5%.