基于改进DDAE的风电场集电线单相接地故障测距

为解决风电场混合接线的集电线短路后难以精确定位的问题,提出基于改进深度自编码网络的故障测距方法。分析集电线故障零序电流可知,暂态电流值、稳态电流幅值、稳态电流相位与故障距离呈现强非线性关系,借助深度学习挖掘这一复杂关系实现集电线精确定位。在深度去噪自编码框架上添加距离回归输出端口,采用联合训练以提升定位网络的准确性、抗噪性和鲁棒性。其过程为：首先,借助PSCAD/EMTDC搭建集电线模型,将给定时窗内故障零序电流序列和对应距离作为故障样本,仿真不同情况故障生成样本集。然后,在训练集上训练改进深度自编码网络得到最优网络用于精确测定故障距离。借助各测点零序电流幅值关系可先确定故障区域,将故障信号送入已训练好的网络即可确定故障所在精确位置。本文方法对集电线多分支、混合短线路有着良好的适应能力；定位性能明显优于传统机器学习算法,且受过渡电阻、采样率、噪音、故障相位角影响较小。

Single-phase ground fault location of wind farm collector based on improved DDAE

In order to solve the problem that it is difficult to locate accurately after short-circuit in the collection line of hybrid connection of wind farm collection line, a fault location method based on improved depth AutoEncoder network is presented.By analyzing the zero sequence current on power line faults, it is known that the transient current value, steady-state current amplitude, steady-state current phase and fault distance are strongly non-linear, and the precise location of power line is achieved by deep learning mining this complex relationship.A distance regression output port is added to the deep denoising self-coding framework, and joint training is used to improve the accuracy, noise resistance and robustness of the positioning network.The process is as follows: Firstly, the hub model is built with PSCAD/EMTDC, and fault zero sequence current sequence and corresponding distance in a given time window are used as fault samples to simulate different failure cases to generate sample sets.Then, an improved depth AutoEncoder network is trained on the training set to obtain an optimal network for precisely measuring the fault distance.With the help of the zero sequence current amplitude relationship of each measurement point, the fault area can be determined first, and the precise location of the fault can be determined by feeding the fault samples into the trained network.This method has a good adaptability to multi-branch and hybrid short lines of power collector.Location performance is significantly better than traditional machine learning algorithms, and is less affected by transition resistance, sampling rate, noise, fault phase angle.