This paper proposes a lightweight deep learning model for high-accuracy core loss prediction under triangular and trapezoidal wave magnetic flux density excitations. The proposed model features a lightweight architecture of only 398 trainable parameters, significantly fewer than existing deep learning methods. In contrast to conventional deep learning approaches, the proposed model introduces a waveform segmentation method that decomposes the input waveform into a series of linear segments, independently predicts the core loss contribution for each segment, and integrates the final results through duty cycle weighting. Validation on the MagNet database demonstrates that the proposed model achieves competitive accuracy, with 95th percentile errors below 10% for four of the five tested materials (7.13%, 2.10%, 4.23%, and 9.94% for 3C92, T37, 3C95, and ML95S, respectively).
