Research On State Diagnosis And Prediction Method For Power Transformer Based On Deep Belief Network Model

Power transformer is the key equipment of energy conversion and transmission in power system.In recent years,the unplanned shutdown accidents of power system caused by transformer faults have occurred frequently,which has seriously affected the development of national economy and the normal social order.Transformer fault is a process of slow development,which shortens the normal working life of the transformer and increases the risk of power outage.Dissolved gas analysis(DGA)is a reliable method for state estimation and fault diagnosis of oil-immersed transformers.which has been recommended as the main detection method by IEC and domestic standards.Therefore,it is necessary to explore the correlation between DGA data and transformer faults for transformer fault diagnosis and prediction.The above work has important theoretical value and engineering significance for supporting the state maintenance of power equipment and reducing the risk of unplanned outage of power grid.Specifically,the current power transformer fault diagnosis and prediction work has the following deficiencies:1)Since the transformer works in normal operation for a long time,it is difficult to accumulate a large number of fault case data,which is a typical scenario of small sample set.At the same time,the occurrence of faults is random,and there is a serious unbalance between the proportion of different faults in the fault case set,resulting in insufficient diagnosis accuracy of the conventional model.2)Transformer DGA time series is a complex nonlinear time series,which caused the prediction results of the conventional method with obvious "amplitude"error and "time-shift" error,especially the "time-shift" error caused the prediction results deviating from the actual situation.To fill the gap,this paper discussed the mathematical basic of fault diagnosis and DGA trend prediction firstly,and analyzed performance characteristics and applicability of conventional diagnosis model.Secondly,a state-oriented power transformer fault diagnosis was proposed,which can automatically switch between state-oriented and numerical-oriented error correction to avoid“over correction”or"owe correction".Thirdly,a deep recurrent belief network(DRBN)model for transformer state prediction was proposed based on time series theory and oil chromatography data characteristics.A self-adaptive delay network with timing correlation features was constructed,able to eliminate "time-shift" error in the prediction results.The iterative correction process of the error was updated so that the error flows simultaneously between and within network layers,thereby improving prediction accuracy.Finally,a large number of on-line monitoring cases are collected and sorted out to verify the advancement of the method.