电工技术学报2025，Vol.40Issue(5)：1559-1574,16.DOI:10.19595/j.cnki.1000-6753.tces.240281
计及铁心非线性的变压器空间动态磁场加速计算方法
Accelerated Calculation Method of Space Dynamic Magnetic Field of Transformer Considering Core Nonlinearity
摘要
Abstract
The magnetic field is one of the key physical fields that reflects the operating state of the transformer.Quickly obtaining the dynamic distribution of the spatial magnetic field of the transformer is one of the foundations for constructing the digital twin of the transformer.The main acquisition methods of transformer magnetic field can be divided into sensor measurement,finite element simulation and algorithm inversion.The installation position of the magnetic field sensors are not flexible,and can only obtain the magnetic field at the measurement points;the finite element simulation can calculate the magnetic field at any positions,but it takes a long time and cannot meet the requirement of digital twin second-level simulation.The deep learning algorithm and structure must be selected and designed according to the characteristics of training data.The existing fast calculation methods are difficult to accurately obtain the magnetic field distribution under core saturation conditions. Based on the nonlinear characteristics of the transformer core and the differential distribution of the main and leakage magnetic flux,this paper proposed a magnetic field accelerated calculation method considering the nonlinearity of the core.Firstly,the field-circuit coupling simulation model of three-phase transformer was constructed,and the key variables were parametrically scanned.A large number of magnetic field data under different nonlinear working conditions were obtained by simulation,and the main flux and leakage flux data sets related to the nonlinear working conditions of the core were constructed.Secondly,a two-branch deep learning model combining convolutional neural network and long short-term memory network was proposed to train and extract the spatial and temporal characteristics of magnetic field data,and solved the model training problem caused by the obvious difference between the main and leakage magnetic flux.Finally,the nonlinear mapping relationship between the input voltage,current and the internal space magnetic field distribution was obtained by using the model,and the accelerated calculation of the spatial dynamic magnetic field was realized,which provided a fast method for obtaining magnetic field data for the construction of transformer digital twin. Considering three typical nonlinear working conditions of power frequency overvoltage,DC bias and inrush current,the magnetic field acceleration calculation model of three-phase transformer and test transformer was trained.It takes about 0.04s to calculate the magnetic field distribution of a single time step,which is greatly shortened compared with the finite element simulation.The results showed that for the three-phase transformer,comparing the calculated values of the magnetic field acceleration calculation model with the finite element simulation,the average absolute errors of the main magnetic flux and the leakage magnetic flux are 0.04 T and 0.9mT,respectively,and the relative error is less than 5%.For the single-phase test transformer,the root mean square error of the radial and axial leakage flux are about 0.1 mT and 0.05 mT,respectively,and the ratio to the peak is within 10%.Therefore,the magnetic field acceleration calculation model proposed can quickly and accurately calculates the spatial magnetic field distribution of the transformer,and provides a method for quickly obtaining magnetic field data for the construction of transformer digital twins.关键词
非线性/卷积神经网络/长短期记忆网络/磁场/加速计算Key words
Nonlinear/convolutional neural network/long short-term memory network/magnetic field/acceleration calculation分类
信息技术与安全科学引用本文复制引用
司马文霞,孙佳琪,杨鸣,邹德旭,彭庆军,王劲松..计及铁心非线性的变压器空间动态磁场加速计算方法[J].电工技术学报,2025,40(5):1559-1574,16.基金项目
国家自然科学基金资助项目(51977018). （51977018）
