Differential Evaluation And Hierarchical Diagnosis Of Transformer Based On Dissolved Gas Monitoring

Power transformer is a very important node in the power system.Its performance directly affects the safe and stable operation of the power system.With the expansion of the power grid and a large number of transformer inputs,the number of fault transformers increases year by year.If not contained,it will lead to huge economic losses,so it is necessary to monitor the transformer operation state,and accurately identify the fault type and severity.Dissolving gas analysis in oil(DGA)can be used to identify the abnormal discharge and overheating in transformer.As the development of off-line chromatographic detection technology,on-line dissolved gas analysis in oil has been promoted by power grid enterprises to overcome the disadvantages of long off-line sampling period and complex detection procedures.With the widely introduction and application of dissolved gas in oil on-line monitoring devices,operators and maintenance personnel have obtained unprecedented massive data,providing rich information for the operation and maintenance of equipment.However,there are a large number of abnormal operation monitoring devices,which hinder the operation and maintenance personnel to master the transformer operation status.These anomaly monitors are numerous and difficult to identify.In order to realize the fast abnormal detection of on-line chromatographic monitoring devices,a dissolved gas in oil monitoring device abnormal identification method based on KPCA was established.The method uses kernel function to compress and transform on-line oil chromatographic data.Identify abnormal data by combining the judgment indexes constructed by the principal component analysis method.Extract abnormal data segments,compare abnormal working data features of monitoring device,identify abnormal monitoring device.The method can effectively identify step mutations with a minimum amplitude of 5%.The identification time of 1.2 million data was 10%of the existing method,so as to quickly and accurately identify abnormal monitoring devices.And 69 sets of abnormal oil chromatographic monitoring devices in 715 devices were detected,with 95.7%recognition accuracy.After abnormal monitoring device removed,abnormal transformers need to be screened.Quick locking of potentially defective transformers.Different from offline monitoring data,online data has strong fluctuation.To solve the problem that it is difficult to identify abnormal transformer states under severe fluctuation,a transformer anomaly identification method based on Canopy model was proposed.The fluctuation coefficient is introduced to quantify the change of characteristic gas.A variable weight high-dimensional space based on the fluctuation coefficient is established to weaken the influence of characteristic gas with large monitoring error.Eventually,Canopy model was used to identify transformer anomalies in a high dimensional space.This method is applicable to the field situation with data loss and large fluctuation,effectively restrains the data fluctuation at the boundary of state change.Compared with the existing methods,this method has better clustering effect,improves the clustering Silhouette Coefficient by 22%,and effectively identifies the overheating anomalies that do not reach the standard threshold.The time complexity function of the algorithm shows that the algorithm has higher efficiency,and the operation time is only 41%of k-means.The possible abnormal transformers are screened out and the threshold method is used to further determine whether there is an internal fault.In the past,due to the small amount of off-line data,it was necessary to calculate the threshold value for the chromatographic data of group transformer oil.However,each transformer has its materials,structures,loads and operating operation condition,and the characteristic gas content varies others’ greatly.The statistical data of transformer population cannot represent the single transformer status,and the threshold value in special cases may be wrong and misleaded.With the introduction and application of large mount of on-line monitoring devices for oil chromatography,the history online data volume of a single device has been able to calculate the threshold value,but the data distribution of a single transformer is unknown and the fault demarcation point is difficult to determine.To solve these problems,a normal-fault dissolved gas generation platform was set up to simulate the gas generation process under normal and discharge faults.It was found that the characteristic gas of normal transformer meets the weibull distribution model with three parameters.The defect models of pin plate and column plate were designed to study the variation relationship between the chromatographic content of transformer oil and the development of internal defects.It was found that there were three stages of characteristic gas change trends with the development of discharge defect:initial,development and severity stages according to discharge;At the initial stage of discharge,no characteristic gas is produced;In the discharge development stage,H2,C2H2 increase linearly;In the severe stage of discharge,all characteristic gases increase significantly.According to the discharge stage,the calculation method of defect rate and failure rate based on P-R curve is proposed,and the differentiated chromatographic threshold calculation method is established by combining weibull function,so as to judge whether the transformer has internal faults.The differentiation threshold of a 110kV transformer is calculated to effectively identify the spark discharge defects of the transformer.Aiming at the situation that it is difficult to fit some online equipment data due to the small amount of data,a differentiated classification method based on classification level is proposed to expand similar online data and improve data fitting degree.Compared with the current warning value,the differential threshold can effectively reduce the occurrence of false alarm and false alarm.It is necessary to accurately judge the fault type and severity of transformers with faults identified.Aiming at the inaccuracy of current diagnostic methods,a deep learning model is introduced to diagnose faults.With recall and accuracy as the misjudgment and confusion evaluation indexes of oil chromatography faults,the optimal input and parameter selection strategy of DBN fault diagnosis were established,and the diagnosis results of deep belief network on dissolved gas faults in oil was analyzed.It is found that DBN is insufficient in feature extraction for chromatographic multi-classification.To solve this problem,a fault diagnosis method of dissolved gas in oil based on combined DBN is established.The fault diagnosis process is divided into two steps:fault type identification and severity identification.The first layer of the network is used to identify fault types with 1 DBN,and the second layer is used to identify the severity of different fault types with 3 DBNS.Compared with single DBN,combined DBN can improve the recall rate and accuracy rate by 20%and 16.6%respectively.The recall rate and accuracy rate of spark faults were improved by 15.1%and 18.6%respectively.Low temperature fault recall rate increased by 11.5%,accuracy rate increased by 16.6%.Combined DBN can accurately determine the fault type,and the overall accuracy can be improved by 18.6%compared with the three-ratio method,and 9.1%compared with the single DBN,effectively identifying the overheating defects of 220kV transformer.