Research On The Method Of Single-Phase-to-Ground Fault Line Detection Based On Deep Learning

At present,the probability of single-phase-to-ground fault is as high as 83%when the small current grounding system of distribution network breaks down.Because the characteristics of single-phase-to-ground fault are not obvious and affected by various random factors,the accuracy of single-phase-to-ground fault line detection is still low.Improving the accuracy of single-phase-to-ground fault line detection can significantly improve the power supply recovery ability,improve customer satisfaction and reduce the loss caused by sudden power failure of single-phase ground fault.In order to solve the problem of low accuracy of single-phase-to-ground fault line detection caused by noise interference,fault types and the number of data samples,this paper proposes a fault line detection method based on improved stacked denoising automatic encoder(SDAE).Firstly,according to the steady and transient characteristics of zero-sequence current of single-phase-to-ground fault,a new data preprocessing method of signal-image conversion is proposed,which extracts the features of original fault data without predefined parameters,and avoids the difference of zero-sequence current under different fault conditions by normalization.Secondly,improved SDAE based on the method of metric learning constraint objective function,that is,the metric distance between data sample pairs is punished by a constant,which is used as a regularization term to restrict the objective function to overcome the intra-class diversity and inter-class similarity between single-phase grounding fault data samples and make it suitable for small sample cases.Finally,a small current simulation model is built based on Matble/Simulink,and fault samples are obtained by simulation under different fault conditions.At the same time,multi-layer unsupervised self-learning and supervised fine-tuning are used to realize fault feature extraction and line detection,and dropout is introduced to improve the generalization ability of the model.The simulation results verify the effectiveness of the improved SDAE model in different fault location,fault type,fault resistance,number of samples and neutral grounding mode,and the accuracy of line detection is higher than the traditional CNN and SDAE deep learning algorithms.