Study Of Characteristic Parameters For Power Transformer Oil Chromatography Data

Power transformers(PowerTransformer)are the hub of the power supply system,and their timely detection and accurate diagnosis not only provide the basis for normal operation and implementation of condition maintenance of power transformers,but also have important significance for people’s life and social stability.At present,the common method for transformer online fault diagnosis is Dissolved Gas Analysis(DGA),but the traditional DGA ratio method has the problem that the established code combination table does not cover all the fault types,and some code combinations cannot find corresponding types in the code table in practical application as well as different literature for There are differences in the input features selected for the training of diagnostic models in different literature.For the shortcomings in the above studies,this paper starts from the input features and combines the preferred feature combinations with the support vector machine multi-classification model to select the optimal feature combinations to train the transformer fault model and further improve the accurate diagnosis of transformers.In this paper,the following elements are investigated.In response to the problem that the feature combinations used in different literature vary and ignore whether the selected feature combinations can effectively reflect the fault types,a DGA-based fault feature set is established by reading a large amount of literature and combining it with the Large Power Transformer Fault Diagnosis and Case Studies to provide a feature-seeking space for the subsequent methods.Second,two evaluation functions,information gain and F-Score,are combined to quantitatively calculate the fault features,and the calculated values are used to rank the fault features,and a support vector machine multi-classification model is introduced to verify the ranked features.The experiments were conducted to compare and validate different feature sets and common classifier methods,and the experimental results revealed that the performance of the preferred feature combination was better than that of the model trained with traditional feature parameters,which improved the recognition rate of the fault diagnosis model.Finally,considering that the Filter model is only used to filter features by quantitative calculation of features,the combination relationship between different features and the close integration with the training model is neglected.Therefore,the Particle Swarm Optimization(PSO)and Support Vector Machine(SVM)algorithms are introduced to design a transformer fault diagnosis algorithm that combines PSO and SVM for feature selection.The selected feature combinations are evaluated by using the standard particle swarm algorithm for optimizing the support vector machine parameters while using the discrete particle swarm algorithm for feature search.The experiments were analyzed from gas combinations,gas ratio combinations,fault examples and literature[16]comparisons,respectively,and the experimental results showed that the fault diagnosis rate of the subset after feature selection improved by about 11%to 23%compared to gas content and IEC ratio,which has better diagnostic performance compared to several commonly used fault features;the fault diagnosis rate improved by 8%compared to literature[16].Considering that missing data in actual field applications can affect the results of transformer fault diagnosis,after analyzing several commonly used interpolation methods,the K-Nearest Neighbor algorithm(KNN)is introduced to fill in the missing data.The experimental results show that the method has better performance capability for the presence of missing data and can provide some help for transformer fault diagnosis in the presence of missing data.