基于U-Net的变压器磁场云图预测方法

Transformer Magnetic Field Cloud Map Prediction Method Based on U-Net

摘要: 有限元分析和计算已成为电磁装置或系统性能计算的主要工具，但由于传统有限元方法求解电磁场时面临建模复杂、计算资源消耗过大等问题，本文采用了一种U型深度卷积神经网络(U-Net)的磁场云图预测模型。以变压器作为研究对象，建立变压器电磁耦合有限元模型，通过改变变压器的几何结构参数、材料和激励信息，计算得到磁场云图作为神经网络训练的样本数据。为提高网络预测性能，通过田口法对U-Net模型进行优化，确定最优模型设置。将U-Net模型预测磁场云图与有限元计算结果对比，U-Net模型预测磁场云图中每个像素点的均方误差在0.3%~0.9%范围内，能够很好地学习到变压器数据集之间的映射关系，生成高分辨率的图像，从而减少了计算时间，对深度学习在预测磁场云图方向上有很大的实际意义。

Abstract: Finite element analysis and calculation have become the main tools for the performance calcula-tion of electromagnetic devices or systems. However, due to the complex modeling and excessive consumption of computing resources when solving electromagnetic fields by traditional finite element methods. This paper adopts a U-shaped deep convolutional neural network. Network magnetic field cloud map prediction model. Taking the transformer as the research object, the electromagnetic coupling finite element model of the transformer is established. By changing the geometric parameters, materials and excitation information of the transformer, the magnetic field cloud image is calculated as the sample data for neural network training. In order to improve the network prediction performance, the U-Net model is optimized by Taguchi method to determine the optimal model settings. Comparing the U-Net model’s predicted magnetic field cloud map with the finite element calculation results, the U-Net model predicts that the mean square error of each pixel in the magnetic field cloud map is in the range of 0.3%~0.9%, which can well learn from the transformer data set. The mapping relationship between the two can generate high-resolution images, thereby reducing the calculation time, which is of great practical significance for deep learning in predicting the direction of the magnetic field cloud map.