Study Of Transformer Fault Diagnosis Based On ISMOTE-MIAO-LightGBM Under Unbalanced Data

Under the background of "double carbon",the new power system with new energy as the main body is moving towards intelligence and digitalization,and this rapid development has provided new solutions and technical means for condition monitoring and fault diagnosis of power equipment.As the core equipment of power system,power transformer undertakes the important task of power distribution and voltage conversion,and it is important to accurately evaluate its operation status and formulate reasonable operation and maintenance strategy to ensure safe and reliable operation of power system.At present,the intelligent diagnosis method based on DGA technology is gradually replacing the traditional transformer fault diagnosis method,but the actual collected DGA data is a kind of small sample imbalance data,which does not meet the data requirements of intelligent diagnosis technology,which seriously affects the fault diagnosis accuracy and the application of intelligent diagnosis model on the ground.For this reason,this paper proposes a transformer fault diagnosis model combining ISMOTE oversampling algorithm and MIAO-LightGBM to effectively improve the influence of unbalanced data on intelligent diagnosis technology and enhance the fault diagnosis accuracy of transformers.Firstly,to address the problems that the actual collected DGA data do not meet the data requirements of intelligent diagnosis technology,the traditional oversampling algorithm SMOTE is prone to synthesize noisy samples and the existing improved SMOTE algorithm rarely considers intra-class imbalance,an ISMOTE oversampling algorithm is proposed,aiming to solve the problem of the impact of imbalanced DGA data on intelligent diagnosis technology.The synthesis process is demonstrated on a homemade dataset,and the results show that the new samples synthesized by ISMOTE are highly similar to the original data and can take into account a few class sparse regions.The results are simulated and tested on SVM and RF classification models with publicly available datasets,and show that the AUC,G-means and F-measure metrics of the ISMOTE algorithm outperform the commonly used SMOTE,KMeansSMOTE,BorderlineSMOTE oversampling methods and the original data,which validates the effectiveness of ISMOTE algorithm.Secondly,the fault diagnosis model of MIAO-LightGBM is established to address the problem that the parameters of the Light Gradient Boosting Machine(LightGBM)model are difficult to determine and affect the diagnosis accuracy.The model adopts the LightGBM algorithm with excellent performance to solve the problems such as high-dimensional nonlinearity of DGA data,while the Aquila Optimizer(AO)is introduced to solve the problem of difficult to determine parameters,and three improvement strategies are proposed for its defects:(1)chaotic backward learning strategy combining Circle mapping and reverse individuals to improve the initial population quality;(2)nonlinear adaptive control The results show that the overall performance of the MIAO algorithm is better than that of PSO,GWO and the original AO.The MIAO algorithm is used to optimize the LightGBM hyperparameters and establish the optimal parameters of MIAO-LightGBM fault diagnosis model,and simulation experiments using balanced DGA data,the results show that the diagnosis accuracy of MIAO-LightGBM model is as high as 91.54%,which is 5.6%,9.85%and 5.6%higher than the accuracy of PSO,GWO and AO optimized models,respectively.Finally,the ISMOTE-MIAO-LightGBM transformer fault diagnosis model was established to address the effect of unbalanced DGA data on the application of intelligent diagnosis technology.Using ISOMTE to equalize the unbalanced DGA data,the results show that the synthetic DGA data are basically similar to the original data,and the samples of the sparse region within the class are increased,which further proves the effectiveness of ISOMTE algorithm.The simulation tests on the unbalanced DGA data show that the diagnostic model accuracy and Kappa coefficient of the equalized DGA data with the ISMOTE algorithm are higher than those of SMOTE,KMeansSMOTE and BorderlineSMOTE and the original data;the ISMOTE-MIAO-LightGBM model has the highest diagnostic accuracy and Kappa coefficients are the highest,reaching 91.13%and 0.8904,respectively.The results demonstrate that the method in this paper can effectively improve the impact of unbalanced data on intelligent diagnosis techniques,and provide some application value for diagnostic prediction and intelligent operation and maintenance of transformers.