摘要
该文针对采用传统数值方法进行大型油浸变压器绕组温升仿真时间较长的问题,提出一种基于U-net神经网络训练的快速计算方法,可以迅速地预测变压器绕组温升及热点。首先,根据流热耦合原理筛选输入变量,并运用流热耦合方法计算不同工况下的输出结果,并将之制作成训练集和测试集。同时,详细讨论3个对网络训练影响最显著的超参数;其次,将归一化后的训练集输入U-net神经网络进行训练,并设置超参数最佳组合;最后,将预测集输入训练好的模型进行预测计算及反归一化操作,预测绕组热点与Fluent仿真结果相差仅0.44 K,单次仿真时间从200 s缩短为0.07 s。预测结果与实验温度平均误差最大为2.31 K,最小为0.98 K,预测方差为0.31左右。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及热点,可满足变压器温度热点数字孪生的实时性仿真要求。
In this paper,a fast calculation method based on U-net neural network training is proposed for the problem of long simulation time of temperature rise of large oil-immersed transformer winding by traditional numerical methods,which can rapidly predict transformer winding temperature rise and hot spot.First,the input variables are screened according to the flow-thermal coupling principle,and the output results under different operating conditions are calculated using the flow-thermal coupling method and made into a training set and a test set.Then,the three hyperparameters that have the most significant influence on the network training are discussed in detail;meanwhile,the normalized training set is input into the U-net neural network for training and the optimal combination of hyperparameters is set.Finally,the prediction set is input into the trained model for prediction calculation and anti-normalization operation.In conclusion,the difference between the predicted winding hot spot and the Fluent simulation result is only 0.44 K.The single simulation time is shortened from 200 s to 0.07 s.Moreover,the average error between the prediction result and the experimental temperature is 2.31 K at the maximum and 0.98 K at the minimum,and the prediction variance is about 0.31.The results show that the method can be used to obtain the temperature and hot spot of oil-immersed transformer winding quickly,and can meet the real-time simulation requirements of transformer temperature hot spot digital twin.
作者
刘云鹏
高艺倩
刘刚
胡万君
王文浩
王博闻
高成龙
LIU Yunpeng;GAO Yiqian;LIU Gang;HU Wanjun;WANG Wenhao;WANG Bowen;GAO Chenglong(Hebei Provincial Key Laboratory of Power Transmission Equipment Security Defense(North China Electric Power University),Baoding 071003,Hebei Province,China;State Grid Zhejiang Electric Power Company Electric Power Research Institute,Hangzhou 310000,Zhejiang Province,China)
出处
《中国电机工程学报》
EI
CSCD
北大核心
2024年第7期2897-2909,I0032,共14页
Proceedings of the CSEE
基金
国网浙江省电力有限公司科技项目(5211DS220005)。
关键词
U-net神经网络
流热耦合
绕组温升
快速计算
数字孪生
U-net neural network
flow-thermal coupling
winding temperature rise
fast calculation
digital twin
