Magnetic core loss is an important indicator for describing the performance of magnetic elements. The traditional physical model has an insufficient performance for predicting the magnetic core loss of magnetic elements under complex conditions such as high temperature, non-sinusoidal waveform, and high frequency. To address this issue, this study proposes a physics-informed neural network (PINN)-based model for magnetic core loss prediction. In particular, this PINN-based model is constructed with a hybrid network architecture as a baseline algorithm, which combines a convolutional long short-term memory network (Conv-LSTM), power spectral density (PSD), and an ensemble learning method (including extreme gradient boosting (XGB), gradient boosting regression (GBR), and random forest (RF)). This design aims to address the complexity of magnetic core loss prediction. Moreover, the Steinmetz equation (SE) is improved to enhance the adaptability under complex conditions, and this improved Steinmetz equation (ISE) is integrated as physical constraints embedded in the neural network for magnetic core loss prediction. Based on the traditional data-driven loss term, the physical residual term is introduced as a regularization constraint to enable the prediction to satisfy both the observed data distribution and physical law. The experimental results show that the PINN-based model has a good prediction performance of magnetic core loss under complex conditions.
Shen et al. (Sat,) studied this question.