Research On Single-Phase Ground Fault Location Method Of Distribution Network Based On Stack Auto-Encoder

Transmission network and electric load are connected by distribution network,which provides directly power for users.Moreover,people’s normal life and production are decided by its safe and stable operation.Due to its complex structure,flexible operation mode and a complex line condition,the fault proportion of distribution network accounts for 80% of that of the whole grid,which single phase grounding fault accounts for 70% of the total.Therefore,the operation and maintenance of distribution network has been always a major difficulty.Small current grounding mode is adopted in our country—not directly grounded or grounded through the arc suppression coil.The single phase fault can not be easy identified due to the weak fault feature as it occurs,and it is easily turns into a short-circuit fault and expands the scope of the fault.The fast fault location is the precondition to remove the fault and to regain electricity.Basing on the precondition,deep learning is used in this paper to digs out fault feature of the fault data aiming at the various failure date acquisition methods,and then distribution network fault is located.The details of this paper carried out as follows:Aiming at the research state about the distribution network fault location both at home and abroad,we state briefly the deficiencies and difficulties of the current approaches.Then based on the classification of existing topological structures in our country,the research status both at home and abroad and analyzed the technical difficulties in the Chine’s distribution network fault location progress are summarized,which include weak fault feature,a complex topological structure and the deficiencies of measuring installation.The feasibility of using deep learning application for distribution network fault location is elaborated and a introduction to the development status of deep learning and machine learning and to the BP neural network model and stacked auto encoder model which used in this paper will be given.According to the advantages of deep learning—self taught learning,a strong fitting of non-linear multi-power function and end-to-en learning—the advantages and feasibility of applying deep learning in fault location is analyzed.Based on faults extraction of reflected pulse feature,a approach to locate fault is proposed.In offline mode,Three same high-voltage pulse signal is injected into the bus end,then we know the first signal detected at the end of the line is by using the phase-mode conversion to analyze fault feature.The arrival time of the first pulse signal of aerial mode is tested and the fault feature is extracted by using auto-encoder(AE).Combining the fault feature extracted by AE with arrival time,a fault feature sequence is constituted,then a BP neural network ranging model is built between fault feature sequence and fault distance and a non-linear relationship is trained between the fitting feature sequence of neural network and fault distance.the fault branch is determined on the basis of signal arrival time of the first line mode and zero mode at the end of line.PSCAD is taken as the platform and this method is tested by the simulating distribution network,the reliability is verified finally.A location method basing on μPMU information is proposed.Based on identity fault segments with dissimilar waveforms and stacked auto-encoder(SAE),the fault ranging method is proposed by analyzing fault current and phase angle fault feature.A SAE ranging model between fault waveform and fault distance is built by comparing the zero-sequence current waveform on both sides of the fault area and by using the measurement date on both sides of fault area,then a end-to-end solution to locate the fault point I proposed by the use of voltage phase and current phase.Finally,PSCAD is taken as the platform,the performance of the approach is tested by the simulating distribution network.The results indicated that this method can locate the fault point effectively and can withstand the effects of transition resistance,fault type and noise as well.