Transformer Fault Diagnosis Based On Machine Learning Algorithm

Oil-immersed transformers are widely used in 220 kV and above substations of Shanghai Power Grid.They use transformer oil as insulation and cooling medium,and convert the high voltage in the high-voltage transmission network into a low voltage that meets the industrial and agricultural production conditions through the principle of electromagnetic induction.It is the most important electrical equipment in a substation.As the internal insulation of the transformer is gradually depleted,the hydrocarbons in the insulating oil are decomposed,and the decomposed ions are recombined into a gas and dissolved in the transformer oil.When the loss accumulates to a certain extent,it may cause serious consequences such as internal faults and insulation breakdown of the transformer,resulting in a large-scale blackout of the power grid.The traditional transformer fault diagnosis technology is mainly the threeratio method or the modified three-ratio method.This method requires periodic sampling of the transformer insulating oil.Then the oil chromatographic analysis technique is used to analyze the content and ratio of the characteristic gases dissolved in the insulating oil for fault diagnosis.Due to the operation management mode of the unmanned station,the oil sampling period of the transformer is generally three months to one year,which causes the transformer fault diagnosis cycle to be very long.The result of transformer fault diagnosis is lagging,and some development defects cannot be identified and tracked in time.In recent years,the extensive application of online monitoring technology for transformer oil chromatography has expanded the sample of dissolved gas data in transformer oil,which provides a possibility for transformer fault diagnosis based on machine learning algorithm.In view of this situation,this paper studies the transformer fault diagnosis technology based on machine learning algorithm,uses the logistic regression algorithm,support vector machine algorithm and eXtreme gradient boosting(XGBoost)algorithm to model and simulate the dissolved gases in transformer oil.Then we compare their performance.Since the dissolved gas data in the oil obtained at the engineering site are unlabeled samples,and the same transformer does not change significantly in a period of time,it makes the logistic regression algorithm and the support vector machine algorithm unable to perform the task of classification well and both of them have a problem of overfitting.The XGBoost algorithm improves the training accuracy of the system by continuously classifying the error.And the performance of the simulation experiment is obviously better than the conventional machine learning algorithm,which makes it suitable for the transformer fault diagnosis.Therefore,considering the results of simulation experiment,we combine the XGBoost algorithm with PMS and create a transformer fault diagnosis system based on machine learning algorithm.During the trial operation in Shanghai power grid,some hidden faults of the transformer were found successfully,which ensured the safe and stable operation of the power grid.