A temperature-coupled sequence learning model for high-frequency magnetic core loss prediction

Abstract

High-frequency magnetic components impose stringent requirements on both efficiency and thermal reliability in solid-state transformers. Accurate core loss estimation of soft magnetic materials under complex excitation conditions remains challenging due to unclear loss mechanisms, strong temperature dependence, and waveform diversity. To address these challenges, this paper proposes a data-driven magnetic core loss calculation framework based on a temperature-conditioned Transformer sequence prediction model. The proposed approach formulates magnetic core loss estimation as an end-to-end sequence-to-sequence learning problem, where the magnetic flux density waveform is mapped to the corresponding magnetic field strength waveform under given operating temperature and frequency. A conditional attention mechanism is introduced by explicitly injecting temperature information into the self-attention process, enabling the model to adaptively capture temperature-dependent hysteresis dynamics. In addition, learnable positional embeddings are employed to enhance phase-awareness within each excitation cycle. Based on the predicted B–H hysteresis loops, the core loss density is calculated through numerical integration, ensuring direct physical interpretability of the model outputs. Experimental results on a public magnetic material dataset demonstrate that the proposed method achieves high prediction accuracy under varying temperatures and frequencies, with an average relative error of 3.85% and a 95th-percentile error of 12.37% on the test set. The proposed framework provides an effective and computationally efficient solution for high-frequency magnetic core loss modeling and thermal-aware design of power electronic systems.