A New Data Optimization Model For Enhancing The Accuracy Of Power Transformer Fault Classification

Electrical power transformers plays an essential role in electrical power systems.Its operation reliability can drastically affect electrical systems security.Defects in transformers generally result from the accumulative electrical and thermal stresses that can cause the destruction of the dielectric strength and insulator components.These defects are extremely dangerous and can cause permanent damage to the electrical system.The results of such failures may cause a massive loss in electricity production and distribution,valuable downtime,huge cost for maintenance,and interruptions to the community and city operations.When faults occur inside the transformers,they produce different types of gases.These gases are important when they are used in inspection and repair programs.By checking the concentration of these gases,we can determine the different types of faults and their severity inside the transformer.The interpretation of these gases can be conducted by different methods,such as dissolved gas analysis(DGA).This method plays a significant role in evaluating and analyzing the transformer faults and their health condition;it also reveals high capability in identifying possible faults and provides great assistance in finding the defective parts without the need to stop transformers in operation.The interpretation of DGA data can be made by different methods such as ratio methods,key gas methods,graphical methods,and artificial intelligence methods.These methods have drawbacks such as strict boundary ratios associated with conventional methods,the curse of dimensionality,unbalanced data set and outliers associated with Artificail Intelligence(AI)methods.Therefore,this thesis discusses two paths of data optimization that are used to enhance the performance accuracy of power transformers fault classification models and eliminate the drawbacks of conventional and AI methods.The first path include,investigating the effects of using data mining(DM)statistical methods on enhancing the diagnosis accuracy of faults in Power Transformers(PT).This path include solving the problems of unbalanced data sets and outlier removal from DGA data by using random over/under sampling algorithms for data balancing and the mean and standard deviation(Std)methods for outliers removal.After that,two case studies are developed to investigate the effects of the optimized data obtained by the random over/under sampling algorithms individually.The obtained data will be used to train the Logestic regression and K nearest neighbohoud models in every case.Then,the trained models will be tested using the testing data samples for the final evaluation.In the second path,the previously mentioned problems are solved by introducing a new model of data optimiztion.This model is based on implementing the combination of subset of set method(C-set)and modified fuzzy C-means algorithm(MFCM)to solve the quandaries of unbalanced data,strict boundary ratios,the curse of dimensionality,and outliers.This process starts by using the C-set method for creating unrepeated multiple subsets,which contains information with high relevance and helps secure the important characteristics from different areas in the input space,through which analyzing and managing each subset of the fault groups is facilitated.After that,the MFCM algorithm is used to cluster each subset individually by eliminating misclassified data samples or samples that have been assigned low membership value(outliers).This step is used to eliminate outliers that may result from inaccurate measurements or coding and can drastically affect the statistical analysis and blur the underlying trend of the corresponding data.Moreover,the problem of the curse of dimensionality can be solved using C-set&MFCM in which the total number of samples in each subset can be reduced while maintaining the integrity of the original data set.Besides,the use of MFCM can also solve the problem of strict boundary ratio by assigning fuzzy membership value to all data points so that each data point can belong to more than one cluster with some certainty level.Then,the optimized cluster centers obtained from each subset will be combined to form the balanced optimized data set in which all types of faults will have an equal number of samples.In this path,two phases of data optimization have been developed.Phase one includes using the C-set method with the standard fuzzy C-means algorithm to optimize the training data.This data was inputted to train the Multiclass Support Vector Machine MCSVM and obtain the final results.The final classification accuracy obtained for this phase was 88.9%.In phase two,the original data set has been divided into training and testing sets.The training data was divided into two groups.Then,the C-set&MFCM was applied to the two groups separately,in which two fault combinations were taken for one group,and three fault combinations were taken for the other group.The optimized data set from the two groups are combined to form the final set.This set is inputted to train different classification models such as the Optimizable Multiclass Support Vector Machine(MCSVM),Decision Trees,and Artificial Neural Networks.The final accuracy results obtained after testing these models were{(MCSVM)～90.9%,(DT)～88.64%,(ANN)～93.18%}.To show the superiority of the proposed data optimization models of the second path,the performance accuracy obtained from phase one and two(of the second path)was compared with the results obtained by the statistical methods in the first path as well as the other methods.All in all,the experimental results indicate that the proposed data optimization model of the second path has significantly improved the fault identification accuracy in power transformers.