Magnetization Mechanism-Inspired Neural Networks for Core Loss Estimation

Fast, accurate, and data-light (small dataset) core loss modeling is critical to the design of high-density high-efficiency power electronics systems. However, conventional core loss modeling methods often demonstrate a significant lack of accuracy, design efficiency and/or cost-effectiveness especially under complex operating conditions. To tackle this problem, this article proposes a core loss modeling method, magnetization mechanism-inspired neural network (MMINN). MMINN is a hybrid data-driven and physics-driven model, which is designed to capture the fundamental magnetization process of magnetic materials at the micro-level. Therefore, MMINN can not only achieve high accuracy and computational efficiency as many other fully data-driven models do, but also enjoy the simplicity (fewer parameters to tune) and generality of a purely physic-driven model. The model training and testing are based on the open-source magnetic core loss database—MagNet, which includes dynamic hysteresis responses under various excitations at high frequencies. An abstract performance comparative analysis among a transformer-based neural network model, a long short-term memory (LSTM) feed-forward neural network model and the proposed model is discussed in the context of the core loss estimation. MMINN shows significant advantages in model accuracy and memory requirements, especially with small datasets, making it well suited for power magnetics design.