摘要
全并联AT牵引供电系统上下行线路并入自耦变压器,致使故障信号多路径传播,且牵引网导线阻抗不连续,传统方法很难实现牵引网故障准确定位。基于AT牵引网结构,推导牵引网上下行导线故障电流幅值与故障距离的非线性关系;基于改进的卷积神经网络(Transformer-based CNN,TCNN)和记忆注意力解耦长短期记忆神经网络(Memory Attended Decoupled LSTM,MADLSTM),通过增加注意力机制和残差连接,增强多导线电流幅值与故障距离的非线性函数关系,从而提高牵引网故障定位的精度;将前述方法与传统的卷积神经网络(CNN)和长短期记忆神经网络(LSTM)进行不同噪声条件下的对比验证。结果表明:基于TCNN+MADLSTM算法进行故障定位时,可自适应构建故障距离与多导线电流幅值的非线性函数关系,以及自适应计算故障距离,无须考虑波速影响;相较于传统的CNN+LSTM算法,TCNN+MADLSTM算法故障定位精度更高,故障区段识别精度可达100%,故障定位精度达72.100 m,均方误差为0.016 km^(2)。
The integration of the uplink and downlink circuits of all parallel AT traction power supply system into the autotransformer leads to the propagation of fault signals through multiple paths and the discontinuity of the impedance of the traction network conductors.Thus,it is difficult to achieve accurate fault location of traction network using traditional methods.Based on the AT traction network structure,the nonlinear relationship between the fault current magnitude on the uplink and downlink conductors of traction network and the fault distance is derived.Based on the improved transformer-based CNN(TCNN) and memory attended decoupled LSTM(MADLSTM) neural network,the nonlinear function relationship between the current magnitude on multiple conductors and the fault distance is enhanced through the addition of attention mechanisms and residual connections,thereby improving the accuracy of fault location of the traction power network.The proposed method is compared with traditional convolutional neural network(CNN) and long short-term memory neural network(LSTM) under different noise conditions.The results show that when the TCNN+MADLSTM algorithm is used for fault location,the nonlinear function relationship between the fault distance and the current magnitude on multiple conductors can be adaptively established,and the fault distance can be adaptively calculated without considering the wave velocity effect.Compared with the traditional CNN+LSTM algorithm,the TCNN+MADLSTM algorithm achieves higher fault location accuracy,with a fault section identification accuracy of 100%,a fault location accuracy of 72.100 m,and a mean square error of 0.016 km^(2).
作者
周欢
陈剑云
万若安
傅钦翠
李泽文
ZHOU Huan;CHEN Jianyun;WAN Ruoan;FU Qincui;LI Zewen(State Key Laboratory for Performance Monitoring and Guarantee of Rail Transit Infrastructure,East China Jiaotong University,Nanchang Jiangxi 330013,China;School of Transportation Engineering,East China Jiaotong University,Nanchang Jiangxi 330013,China)
出处
《中国铁道科学》
EI
CAS
CSCD
北大核心
2023年第4期206-218,共13页
China Railway Science
基金
国家自然科学基金资助项目(51467004)。
关键词
全并联AT牵引供电系统
故障定位
改进的卷积神经网络
记忆注意力解耦长短期记忆神经网络
All parallel AT traction power supply system
Fault location
Improved convolutional neural network
Memory attended decoupled long short-term memory neural network(MADLSTM)
