Research On Intelligent Prediction Of Transformer Faults Based On DG

Power transformer is the key equipment of power system,which plays an important role in the transmission and distribution of electric energy.Its normal and stable operation will directly affect the reliability of power supply of the grid,and transformer failure will damage the stability of power system and produce huge social and economic losses.Therefore,the fault diagnosis and oil dissolved gas prediction of transformers can be used to accurately analyze the operating conditions of transformers in advance,to detect potential transformer failures and their development trends early,and to arrange maintenance and overhaul in a scientific manner.In this thesis,the following studies are carried out based on dissolved gas in oil data,combined with machine learning algorithms:(1)To address the problem that the feature parametres characterizing transformer faults are divergent and improper human selection may easily cause missing features or feature redundancy and thus degrade the performance of the classification model,a method is proposed to filter the 22-dimensional feature vectors characterizing transformer fault types using the random forest algorithm,and the results are validated with extreme gradient boosting tree(XGBoost)as the classification model.The algorithm results show that when comparing the commonly used five feature gases,triple ratios,and uncoded ratios,the average accuracy of fault diagnosis of four models,including BPNN,SVM,ELM,and XGBoost,is improved when the 10-dimensional features proposed in this paper are used as inputs,which verifies the feasibility of the proposed feature selection;with the selected input features,the average accuracy of diagnosis and the recall rate of XGBoost model for each type of With the selected input features,the average diagnostic accuracy and recall rate of XGBoost model for each type of fault are better than the other three models,which verifies the effectiveness of the models.(2)To address the problems of class imbalance in transformer fault samples and difficulty in determining key parameters of the XGBoost model,a transformer fault diagnosis method is proposed to pre-process the data with the adaptive integrated oversampling algorithm(ADASYN)and find the optimal hyperparameters of the XGBoost model with the Snake Optimizer(SO)algorithm.First,ADASYN is used to oversample the fault data set to reduce the class imbalance of the data set and the bias of the classification model;then the SO algorithm is used to iteratively find the optimal values of the important hyperparameters of the XGBoost model and establish the SO-XGBoost model to identify the transformer fault types.The results show that the SO algorithm performs better than the genetic algorithm and the particle swarm algorithm in terms of global search ability and convergence accuracy;the ADASYN algorithm can reduce the class imbalance of the training set and thus improve the correct diagnosis rate of the model;the generalization performance of the SO-XGBoost model is better than other models under different sample sizes of the training set.(3)A model based on convolutional neural network(CNN)and snake optimized bidirectional long and short-term memory network(SO-Bi LSTM)is developed for the low accuracy of dissolved gas concentration prediction in transformer oil.The multidimensional feature vectors characterizing the gas concentration variation are transformed by the convolutional neural network and constructed as time series input to the bi-directional long and short-term memory neural network(Bi LSTM),and the hyperparameters in the Bi LSTM network are iteratively optimized using the SO algorithm to build the CNN-SO-Bi LSTM prediction model.The results of the algorithm show that the proposed prediction model has higher accuracy compared with the general recurrent neural network prediction method.