Research On Fault Diagnosis Method For Transformers Using DGA Data

As we all know,power transformers occupy a pivotal position in the whole power system.At present,the domestic large power transformer is mainly oil-immersed,its internal in some cases there may be air into,for example,in some production factory or high load operation occasions,when the power transformer in the insulating oil and some insulation materials for a long time and air contact,may occur after the deterioration phenomenon,resulting in the potential for failure,so inevitably there will be some failure.When a fault occurs,some fault gases will be produced in the insulating oil of power transformer,however,different fault gases often correspond to different types of faults,and there is a certain relationship between different faults,so it is necessary to analyze the dissolved gases in the insulating oil of power transformer periodically,in order to have a complete grasp of the operation status of power transformer and respond to the faults that occur in time.This thesis analyzes the current mainstream dissolved gas in oil analysis(DGA),and proposes an improved density-based clustering algorithm with noise(DBSCAN)for fault diagnosis of oilimmersed transformers by combining the adaptive capability of clustering algorithms in terms of data processing with the advantage of relative insensitivity to noise for the collected DGA data of different states;in addition,the use of support vector machines(In addition,to further improve the accuracy of fault prediction and the convergence speed of the model,a cross-validation,grid-seeking approach is used to find the optimal parameters of the optimal Gaussian radial basis kernel function in order to achieve the optimal the classification effect and accuracy.This thesis mainly does the following work:(1)Analyzed the advantages and disadvantages of the current mainstream dissolved gas in oil analysis method,combined with the highly adaptive capability of the DBSCAN algorithm,improved the traditional DBSCAN algorithm so that it can discover clusters of arbitrary shape,no bias on the clustering results and other features;proposed an optimized density-based clustering algorithm with noise.The accuracy in the number of clusters is improved and the accuracy of fault diagnosis is also enhanced.(2)The advantage of support vector machine in dealing with classification in terms of minimizing structural risk is utilized,and at the same time,a fault diagnosis method(MPA-SVM)based on the combination of marine predator algorithm and support vector machine is proposed for the limitation problems such as slow convergence speed and easy to fall into local optimum that are common in some traditional optimization algorithms.The method is a new idea to solve the small sample classification,optimizes the convergence speed and the optimization seeking ability of the system,improves the robustness and the accuracy of the fault diagnosis.(3)For the problem of selecting and setting the parameters of SVM kernel function and penalty parameters,there is no unified method at home and abroad,and a cross-validation grid-seeking method is proposed to obtain the best parameters;by setting the variation range of parameters,the SVM classifier is trained for classification,which simplifies the steps of parameter finding.The classification capability of the SVM model is improved and the correctness of fault diagnosis is obtained.