基于S变换相关度和深度学习的配电网单相接地故障选线新方法

Novel method for single-phase grounding fault line selection in distribution network based on S-transform correlation and deep learning

针对配电网发生单相接地故障时特征信息不明显,且现有选线方法易受故障条件和环境噪声影响的问题,基于S变换相关度和深度学习,提出一种具有强抗噪声能力和高泛化水平的配电网单相接地故障选线新方法。首先,利用S变换获取零序电流时频信息,基于各线路零序电流的全频段信息计算线路故障信息相关度;其次,为提高故障特征的可辨识度和抗干扰性,提出一种S变换相关度图形(SCF)构建方法;在此基础上,建立含SCF层的卷积神经网络深度学习模型(S-CNN),并利用Simulink仿真模型生成的故障数据对其结构参数和超参数进行分步训练;最后,通过S-CNN提取配电网故障零序电流深层特征,实现故障选线,并测试了S-CNN在配电网不同运行状况和故障条件下的选线效果。仿真结果和实际配电网故障数据测试表明:在强噪声干扰场景中,基于S-CNN的故障选线模型在不同故障位置、故障相角、过渡电阻条件下可实现高正确率选线,且在各线路零序电流采样不同步条件下仍具有较强的鲁棒性。

The fault characteristics of single-phase grounding faults in small current grounding systems are not obvious,and the existing line selection methods are susceptible to fault conditions and environmental noise.A novel line selection method based on S-transform correlation and deep learning is proposed,which has strong anti-noise ability and high generalization level.Firstly,the time-frequency information of the zero-sequence current is obtained by S-transform,which is used to calculate the fault characteristic information correlation between each line.Then,a construction method of SCF(S-transform Correlation Figure)is proposed to improve the identifiability and anti-interference of the fault characteristics,based on which,the S-CNN(Convolutional Neural Network deep learning model with SCF construction layer)is constructed,and its structural parameters and hyperparameters are trained step by step with the fault data generated by the Simulink simulation model.Finally,S-CNN is used to extract the deep features of the fault zero-sequence current to select the fault line,and the effect of S-CNN under different operating conditions and fault conditions are tested.Simulative results and actual distribution network data test show that the fault line selection model based on S-CNN can achieve high accuracy under different fault locations,fault phase angles,transition resistances,load fluctuations or strong noise interference conditions,and still has strong robustness under the condition of unsynchronized zero-sequence current sampling of each line.