Mining And Application Of Transformer DC Resistance And Oil Chromatogram Data

As an important part of a strong smart grid,transformers are particularly important for preventive testing,monitoring and diagnosis.In the context of increasing power data,for the preventive tests and online monitoring data of transformers,the data mining theory can be used to statistically analyze the threshold value of direct current(DC)resistance change of transformer winding,predict the development trend of dissolved gases in oil,diagnose the transformer fault,discover and eliminate hidden safety risks early.This paper takes transformer DC resistance data and oil chromatographic data as the research objects,and combines the data mining methods to carry out the following three aspects of research:(1)In view of the current test regulations,the threshold value of the transformer winding DC resistance change is too general,and there is a lack of related difference analysis.In this paper,DC resistance test data in a certain area has been collected and the variations in each following year have been calculated and compared.Further,by successfully combining graphic method and hypothesis test method,the distribution of DC resistance variation is explored.Then,the differences of DC resistance variations are analyzed by using the confidence interval in different dimensions.The analysis results show that,according to the voltage level,capacity,operating time and operating environment humidity,the DC resistance changes tend to be normally distributed and there are certain differences,which can supplement the transformer maintenance regulations in this area.(2)Aiming at the problem that the single prediction method of dissolved gas concentration in transformer oil is not accurate enough,a prediction model based on empirical mode decomposition(EMD)and random forest(RF)is proposed in this paper.The EMD method is used to decompose the dissolved gas concentration sequence into a series of sub-sequences with little mutual influence.The sub-sequences are used as input to establish different random forest prediction models,and the predictions of each sub-sequence are superimposed as the final prediction result.The analysis results show that the EMD method can effectively split the original characteristic gas sequence into sub-sequences with weak mutual influence.Compared with a single RF prediction model,the EMD-RF model can more accurately predict the development trend of the dissolved gas concentration in oil.(3)For the problem that the adjustment of parameters based on experience leads to a random forest model to diagnose transformer faults,the accuracy rate is not high enough,a fault diagnosis model based on particle swarm optimization(PSO)to optimize random forest parameters is proposed in this paper.The feature vector is formed as the model input with non-code ratios of dissolved gases in the oil.Further,a PSO algorithm is used to search two optimal parameters that are number of trees and number of splitting features of the RF model.Then,the PSO-RF model is established to diagnose the fault type.The effectiveness of the proposed diagnosis model and feature selection is verified by the diagnosis results of specific examples.The analysis results show that the feature selection based on the non-code ratios gets more fault information and PSO optimizes RF parameters to help improve the model diagnosis performance.Furthermore,as the sample space increases,the effect of the fault diagnosis model is better.