Core Loss Prediction Model Based on Waveform Classification

Core loss is governed by coupled factors, strong nonlinear behavior, and excitation waveform; traditional empirical models often fail to deliver high accuracy under complex operating conditions. Data heterogeneity across waveforms also limits the generalization ability of data-driven models. To address these issues, this paper proposes a waveform-classified Bayesian-optimized artificial neural network (WC-BOANN) core loss prediction model, which combines Random Forest waveform classification with a Bayesian-optimized artificial neural network. Features including form factor, skewness, total harmonic distortion, and differential variance are first extracted from magnetic flux density time series. The Random Forest waveform classification then identifies sinusoidal, triangular, and trapezoidal waveforms automatically. Separate artificial neural network models are built for each excitation waveform, and Optuna-based Bayesian optimization is used to tune the network structure and hyperparameters. The results show that the unclassified Bayesian-optimized artificial neural network (UC-BOANN) performs unevenly across materials, with MAPE values of 9.46%~19.65% and R2 values of 0.9737~0.9857. By contrast, WC-BOANN improves prediction accuracy: on the test sets of four magnetic materials, all R2 values exceed 0.99, all MAPE values are below 3%, the 95-percentile prediction errors are below 7.5%, and the ±10% prediction accuracy (±10% PA) exceeds 99%.