Research On Transformer Winding Frequency Response Defformation Identification Method Based On Improved Frechet Distance

Transformer is very important in the power system,once the transformer failure will seriously threaten the safety of the power system.Winding deformation is the main cause of transformer failure,so winding deformation detection can find hidden dangers in time,so as to prevent transformer failure,which is of great significance.Frequency response method is widely used in the field of transformer winding detection because of its good detection effect.However,the frequency response method is lack of numerical index to describe the difference between the two frequency response curves,and has low accuracy in diagnosing deformation types with small samples.At present,the numerical index used in power industry standard is correlation coefficient method,but this method has the disadvantages of low sensitivity and high misjudgment rate.So it is very important to study the data processing of frequency response method.Firstly,the distributed parameter model is selected to model the winding,and the parameter changes corresponding to the deformation are determined through the calculation formula of the parameters and different deformation conditions,and the differences of frequency response curves corresponding to different deformation are observed.Frechet distance is introduced into the field of transformer winding deformation diagnosis to analyze the frequency response data of winding.In order to prevent the high frequency stray capacitance interference,Gauss membership function is selected to adjust the weight of Frechet distance.Through the simulation,it can be seen that the sensitivity of the improved Frechet distance is significantly improved compared with the correlation coefficient method and mutual distance.The index difference between 10%winding deformation and severe deformation is increased by about 0.2,and the winding deformation can be judged without frequency division.Secondly,for the problem of low accuracy of classification caused by less samples of winding deformation,SOM network is proposed to classify the frequency response curve of winding,Kohonen algorithm is used to train frequency response data samples,and the winning neural element is determined by comparing the distance between test data and sample data,The straw hat function is used as the adjustment function to change the weight vector of the neighboring neurons around the winning nerve for training.The simulation results show that SOM network can effectively classify the data with alike trend without generous samples.In the case of the same number of samples as the training set,the diagnosis accuracy of SOM network is better than that of traditional neural network and support vector machine,and it can analyze much frequency response data simultaneously,and the response speed is faster.Finally,in order to further verify the effectiveness of the winding detection method mentioned in this paper,a transformer in a hydropower station was tested in situ.The test object is two transformers of the same model,one of which is deformed.The test results show that one of the transformers has obvious deformation,and the deformation type is whole compression deformation.The result is in line with reality,which proved the proposed method is effective.