Fault Diagnosis Method Of Power Transformer Based On The Dissolved Gas-in-oil Analysis

As one of the most key equipment,transformer is the core of energy conversion and transmission in the power grid,the operation state of transformer directly affects the safety of power network,and it is an important pivotal equipment in the network security defense system.Therefore,accurate diagnosis of transformer internal latent fault and further comprehensive study of transformer operating conditions can effectively guide the operation and maintenance of power transformers and state maintenance,significantly reduce the transformer failure rate,which is of great significance for ensuring the safe operation of the power grid.This paper mainly studies from the two aspects of power transformer fault diagnosis and fault prediction,the main research work is as follows:The state of the transformer can be basically understood by the content and composition of dissolved gas in its oil.Therefore,Dissolved gas-in-oil Analysis(DGA)becomes the important basis to determine whether the transformer is in safe and stable state.In view of the defects and deficiencies of the existing DGA based power transformer fault diagnosis methods,this paper proposes the stack Denoise auto-encoder Network and stack Sparse auto-encoder Network based on deep learning.These two methods are based on the network,according to the transformer DGA data characteristics and fault types,combined with softmax classifier to build transformer fault diagnosis model,and gave the detailed diagnosis steps.Numerical examples show that these two methods have better stability and higher accuracy.Transformer fault prediction is mainly based on the analysis of dissolved gas concentration in transformer oil to determine whether the transformer has potential faults.In this paper,based on the deep learning network,a sparse deep belief network is proposed to predict the concentration of dissolved gas in transformer oil.This model takes the concentration of characteristic gas as input,and through training the multi-hidden layer machine learning model based on restricted boltzmann machine,it can automatically extract the development law of gas concentration itself,activate the strong correlation between the influence of various gas components on gas concentration layer by layer,and suppress and weaken irrelevant and redundant information.The proposed method has higher prediction accuracy.Finally,case analysis verifies effectiveness and superiority of the proposed model.