Deep Learning Modeling Method For Dissolved Gas In Power Transformer Oil

Power transformer is one of the key equipment in power system,its running state directly affects the safety and stability of the whole power system,so it is very necessary to judge the transformer status in real time.Dissolved Gas Analysis is one of the most effective methods to diagnose transformer internal faults.Although the traditional diagnosis method based on DGA is simple and convenient for practical engineering application,the accuracy of diagnosis is low due to the difficulty in determining the boundary,the incomplete coding system and the sensitivity to the uncertainty of actual DGA data.Although the shallow machine learning diagnosis method based on DGA has solved some problems existing in the traditional methods,the diagnosis effect is not ideal due to its poor generalization ability,easy overfitting,difficult feature extraction and other shortcomings.Moreover,it is difficult to analyze the DGA data of a single moment point,which leads to the difficulty of online modeling.To solve the above problems,this paper uses deep network to build a diagnostic model.Specific research contents are as follows:This paper proposes a transformer fault diagnosis method based on convolutional neural network,which uses convolution and pooling to extract sensitive fault features at a single time point,and uses the deep-level features excavated to solve the problems of difficult feature extraction and poor generalization ability.Firstly,the diagnostic model is built,and the relevant evaluation criteria are established.Then,the factors influencing the performance of the diagnostic model are tested and compared with the performance of the shallow network diagnostic model.Finally,the powerful feature extraction ability of the model is visualized.A transformer fault diagnosis method based on convolutional neural network and bidirectional long short term memory network is proposed to solve the problem of on-line modeling by considering the characteristics of complex fault time.Firstly,the diagnostic model was built and relevant evaluation criteria were established.Then,the influences of time step and bath size on the performance of the diagnostic model were tested and compared with other diagnostic models.Finally,the spatial and temporal features extracted from the model were visualized.