History-Dependent Prandtl–Ishlinskii Neural Network for Quasi-Static Core Loss Prediction Under Arbitrary Excitation Waveforms

Magnetics in power electronics applications often experience complex excitation waveforms containing high frequency harmonics and/or dc bias. Such a complexity challenges conventional magnetic core loss modeling methods, which are often prone to serious errors or only applicable to certain types of excitation waveforms. This article presents History-Dependent Prandtl-Ishlinskii Neural Network (HDPI-NN) for accurate quasi-static core loss prediction under arbitrary excitations. The high modeling accuracy and generality of HDPI-NN, is achieved by accounting for a more complete set of magnetization mechanism (i.e., domain wall motion, domain rotation, and memory effect) of magnetic materials, with model parameters determined by a specialized neural network. Quasi-static magnetic loss measurements for silicon steel B35A270 and ferrite N87 were acquired to support model training. The reliability of the proposed model has been confirmed through rigorous verification and comparative experiments. The results show that the proposed HDPI-NN model, even when trained only with 200+ data containing symmetric and low-frequency harmonics, is able to accurately predict the hysteresis behavior of the magnetic material excited under dc bias and/or higher frequency harmonic conditions.