图学习与零序分量相结合的风电场集电线单相接地故障定位

A Fault Localization Scheme for Single-Phase-to-Ground Faults on Collecting Lines in Wind Farms Combining Graph Learning and Zero-Sequence Components

由于风电场集电线线路短、连接风机多且风机间距小,传统行波测距等故障定位方法难以适用。为此该文提出一种图学习与零序分量相结合的新型层次化单相接地故障定位方案。首先,计及集电线拓扑信息构建图数据,应用图卷积神经网络进行图分类建模判断集电线故障区域;然后,为克服样本不平衡的影响,结合迁移学习进行模型串联,并引入代价敏感机制后处理分类结果;最后,利用风电场接地方式与箱变接线特点,基于双端零序分量推导与风机无关的故障区域测距公式。所提方案可跨越风机实现定位,并能克服数据不同步的影响,所需测点少,适用于含分支集电线。仿真表明,图卷积神经网络相对传统深度网络判断正确率更高,模型串联结合代价敏感机制可有效克服样本不平衡影响;故障区域判断与测距效果不受故障位置、过渡电阻和风速的影响。

In wind farms,collecting lines are short with a large number of wind turbines connected to it,and fault characteristics of wind turbines are different from those of traditional power sources,so dense connection of special power sources is the typical characteristic of collecting lines.The elimination of influence of wind turbines,applicability of collecting lines with line branches and reduction of measurement points are key challenges faced by the fault localization for collecting lines.Given these above challenges,it is difficult for the fault localization method in transmission and distribution networks to apply in wind farms,and the existing fault localization method of collecting lines cannot effectively address all these challenges.Therefore,a new hierarchical fault localization scheme for single-phase-to-ground faults on collecting lines combining graph learning and zero-sequence components is proposed,which can achieve precise fault localization across wind turbines with the applicability of collecting lines containing line branches considered and a few measurement points used.Firstly,the graph data is constructed according to electrical information at measurement points and collecting line topology,and GCN(graph convolution neural network)is used as base-models for graph classification modeling to obtain the identification model of faulty areas which consists of a primary classifier and a secondary classifier;in the process of identification model construction,in order to overcome the impact of sample imbalance,multiple classifier models are connected in series with transfer learning combined,and cost sensitive learning based on Bayes minimum risk is also introduced to post-process classification results.Then,considering the neutral-point grounding type of wind farms and the winding connection characteristic of box-type transformers,a wind-turbine-independent fault distance calculation formula is derived based on zero-sequence components,and thus the fault point in faulty areas can be obtained by zero-sequence components.Besides,a misidentification processing mechanism for faulty areas is also proposed to correct the misidentification which occurs rarely.This above proposed fault localization scheme does not require strict data synchronization,can take into account both of structural and attribute information of collecting lines to identify the faulty area with limited fault data provided by a few measurement points,and can overcome influence of wind turbines to achieve the fault distance calculation.Verification results show that with the graph data constructed,faulty areas can be effectively identified by GCN,and the identification accuracy of all faulty areas reaches over 99%,which is higher than traditional DNN(deep neural network);besides,in the process of dealing with sample imbalance,model series connection increases the total average accuracy of branch faults by 12%,and further this above accuracy achieves another 6.1%improvement with cost sensitive learning introduced,which verifies the feasibility of the sample imbalance treatment strategy in this paper.The fault distance calculation method proposed for faulty areas can overcome the influence of fault locations,fault resistance,wind speed and data asynchronization,and errors of fault distance is no more than 150 m.In addition,the proposed misidentification processing mechanism is also proved to be effective and efficient in correcting misidentification of faulty areas.