| Transformer is a key equipment in power system that undertakes electrical energy conversion,and its operation status directly affects the normal operation of power system,once the operation failure occurs,it may produce great threat to power system and cause huge economic loss.With the continuous promotion of smart grid construction,machine learning has shown powerful generalization ability in transformer fault intelligent diagnosis,which provides more scientific and reliable development paths and technical solutions for transformer fault diagnosis research.In this paper,based on machine learning algorithms,we focus on data pre-processing,sample balance processing,hyperparameter optimization,and propose to improve Stacking fusion model for transformer fault intelligent diagnosis.Firstly,to address the problem of insufficient feature characterization capability of the original data,this paper incorporates ratio calculation rules for feature derivation to e nhance the data processing and inferring capability of the machine learning algorithm.On this basis,the data are normalized,and the correlation between sample features and labels is calculated using mutual information to verify the feasibility of feature derivation,screen out important features,and reduce the computational complexity of the model.Due to the distribution characteristics of transformer fault states,there is a serious category imbalance problem in the dataset.In this paper,we design conditional generation adversarial network for data balancing processing to improve the defects caused by noise interference of traditional sampling algorithm,and use SMOTE,Borderline-SMOTE,and ADASYN in traditional sampling algorithm for comparison experiments to quantitatively assess the quality of generated samples using KS test.The results of the comparison experiments show that the conditional generative adversarial network can generate high-quality samples based on the multidimensional features of real data,and has strong stability and reliability.Secondly,for the complex hyperparameter problem of machine learning algorithms,this paper uses the TPE hyperparameter optimization algorithm to continuously search for the optimal parameter combinations.The balanced training set is input to four machine learning algorithm models,namely,Random Forest,XGBoost,Light GBM,and Cat Boost,for transformer fault diagnosis.Based on the results of the importance ranking analysis of hyperparameters,a more scientific parameter range and space are set to significantly improve the optimization search efficiency.After several rounds of iterative training,the hyperparameters taken for the best performance of the model are calculated,which substantially improves the search efficiency of hyperparameters and the overall performance of the algorithm model.The comparison experimental results show that before optimization,the fault diagnosis accuracy of Random Forest,XGBoost,Light GBM,and Cat Boost are73.33%,82.15%,82.58%,and 82.37%,respectively.After TPE hyperparameter optimization,the fault diagnosis accuracies were improved to 80.22%,88.39%,88.60%,and 88.82%.Finally,to address the problem that single algorithm model is easy to fall into local optimal solution and poor robustness,this paper uses four multi-model algorithm fusion strategies for comparison experiments.Among them,Stacking fusion performs the best,with an accuracy of 89.46%.Based on Stacking fusion,using the original characteristics of the data,this paper designs an improved Stacking fusion model to diagnose transformer faults step by step.The first layer of Stacking classifier is trained on three generalized labels of normal,discharge fault and overheating fault,and the output generalized label probabilities are fused with the original input for feature fusion and input to the next layer of Stacking classifier for training,and seven specific state types are diagnosed,and the accuracy rate is improved to92.26%.The performance of the improved Stacking fusion model designed in this paper is significantly improved as demonstrated by several sets of comparison experiments,which provides an effective solution to the problem of transformer fault diagnosis with complex hierarchical structure. |
