Machine learning based prediction and optimization of iron loss

Abstract

In this thesis, a machine learning-based method for predicting high-frequency iron core losses is proposed. To address the problem of limited accuracy of high-frequency sinusoidal excitation, a database of magnetic materials, different frequencies and temperatures is established. The key features extracted include excitation waveform, magnetic flux density, frequency and temperature. An integrated learning strategy is used to classify the excitation waveform and predict the iron loss by XGBoost. By analyzing the experimental results, it is concluded that the method has higher prediction accuracy, lower mean square error (MSE), and a coefficient of determination (R²) closer to one. Meanwhile, in order to achieve the dual-objective optimization of minimizing iron loss and maximizing magnetic energy transfer, this study also optimizes the magnetic element design using genetic algorithm (GA). The energy efficiency ratio of the magnetic element can be further improved by this method. Experimental results verify the effectiveness of the method under high-frequency conditions and provide reliable support for the optimal design of magnetic components for high power density power electronic systems. Future research can combine deep learning and migration learning to improve the adaptive capability of the model and extend it to a wider range of magnetic materials and application scenarios.