基于U-net神经网络的35 kV油浸式变压器绕组温度快速计算

Fast Calculation of 35 kV Oil-immersed Transformer Winding Temperature Based on U-net Neural Network

针对采用传统数值方法进行油浸变压器绕组温升仿真时间较长的问题,提出了一种基于U-net神经网络训练的快速计算方法,以迅速地获得变压器绕组温升。针对1台35 kV油浸式变压器,利用Fluent软件生成了不同工况下深度学习所需的训练集,在确定超参数的最佳组合后,变压器温度场的计算效率得到显著提高,最后建立光纤试验测温平台对算法的有效性进行了验证。以Fluent软件得到的结果为参考,B相低压绕组内外侧和高压绕组内侧U-net神经网络的相对误差在0.24%、0.21%和0.39%左右,单次计算时间从10854s缩短到0.05s,且预测结果与试验温度平均误差最大为4℃,最小为2℃。研究结果表明,该方法可用于快速获得油浸式变压器绕组的温度,可以满足油浸式变压器温度及热点数字孪生技术的实时性仿真要求。

In this paper,a fast calculation method based on U-net neural network training was proposed to solve the problem of long time for calculating the temperature rise of oil immersed transformer windings using traditional numerical methods,which can quickly predict the temperature rise and hot spots of transformer windings.For a 35 kV oil-immersed transformer,the training sets required for deep learning under different operating conditions was generated by using Fluent software.After determining the optimal combination of hyperparameters,the computational efficiency of the transformer temperature field was significantly improved.Finally,a fiber-optic test temperature measurement platform was established to verify the effectiveness of the algorithm.By using the results obtained from Fluent software as reference,the relative errors of the U-net neural network for the inner and outer B-phase low-voltage winding and the inner high-voltage winding were around 0.24%,0.21%and 0.39%,and the single calculation time was shortened from 10854 s to 0.05 s,and the average error between the prediction results and the test temperature was 4℃at the maximum and 2℃at the minimum.The results show that the method can be used to obtain the temperature of oil-immersed transformer windings quickly,and it can meet the real-time simulation requirements of oil-immersed transformer temperature and hot spot digital twin technology.