基于混合ResNet-BiGRU的高压直流输电线路故障测距

Fault Location for High-Voltage DC Transmission Lines Based on Hybrid ResNet-BiGRU

中文关键词:故障测距  原始故障信号  时频灰度图  混合残差网络  多任务学习

keywords:fault location  original fault signal  time-frequency grayscale graph  hybrid residual network  multi-task learning

中文摘要:针对现有故障定位方法存在微弱故障（以高阻接地故障为主）测距精度不足的问题，基于融合特征输入和多任务学习思想，提出了基于混合残差网络-双向门控循环单元（residual network-bidirectional gated recurrent unit，ResNet- BiGRU）的高压直流输电线路故障测距方法。首先，将采集的一维电压行波与对其进行连续小波变换获得的二维时频灰度图分别送入混合ResNet模型的一维特征处理模块和二维特征处理模块，用来提取暂态行波的时域全局特征和频域局部特征，并对提取的特征进行拼接融合；其次，以混合ResNet作为多任务学习的参数共享层，通过连接Softmax分类器对故障类型进行判别，由故障类型选择对应BiGRU定位器，使BiGRU定位器更具有指向性。最后，在仿真软件中搭建四端柔性直流输电系统进行实验验证，仿真结果表明，所提方法抗噪能力强，受过渡电阻干扰小，在不同故障位置均可得到较高精度的测距结果。

Abstract:Aiming at the problem of insufficient ranging accuracy of existing fault location methods for weak faults (mainly high-resistance grounding faults)， a fault ranging method for HVDC transmission lines based on the hybrid Residual Network-Bidirectional Gated Recurrent Unit (ResNet-BiGRU) is proposed in light of the ideas of fused feature input and multi-task learning. Firstly， the collected one-dimensional voltage traveling waves and the two-dimensional time-frequency grayscale images obtained via continuous wavelet transform are fed separately into the one-dimensional and two-dimensional feature processing modules of the hybrid ResNet model， so as to extract the global time-domain features and local frequency-domain features of transient traveling waves， followed by the concatenation and fusion of the extracted features. Secondly， the hybrid ResNet is employed as the parameter-sharing layer for multi-task learning； a Softmax classifier is connected to identify the fault type， and the corresponding BiGRU locator is selected according to the identified fault type， which renders the BiGRU locator more targeted. Finally， a four-terminal VSC-HVDC transmission system is established in simulation software for experimental verification. The simulation results demonstrate that the proposed method boasts strong anti-noise capability and low susceptibility to transition resistance interference， and it can yield high-precision ranging results at different fault locations.