Study On Condition Assessment And Fault Diagnosis Approaches For Power Transformers

A power transformer is the core of the energy conversion and transmission grid. Itis key hub power equipment in the first line of defense of the grid security. At present, ithas been more transformers in China whose operation periods are over20years. Thoserunning transformers are facing increasingly serious problems of such as equipmentfailure and insulation aging, and at the same time have increasing probabilities of anaccident. Transformer failure may cause huge losses of equipment assets and blackout,and even serious social impact. Therefore, the effective condition assessment and faultdiagnosis analysis for power transformers, to guide the operation and maintenance ofpower transformers and prevent and reduce the failure probability, have an importanttheoretical and practical significance.This dissertation collects a large number of technical standards, regulations,expertise, and actual state information considering power transformers. Stating with that,the dissertation studies the condition assessment index system, assessing methods,decision-making criteria and fault diagnosis approaches based on support vectormachine theory with intelligent optimization algorithms for power transformers. Thetransformer assessing decision-making models, based on set pair analysis theory and afuzzy with evidential reasoning integrated approach respectively, are studied in thisdissertation. And the research on transformer fault diagnosis based on DGA has beenmade some breakthroughs. The main innovative achievements are obtained asfollowing.Considering the uncertain problem that the transformer state information is fuzzyand incomplete, a condition assessment strategy based on set pair analysis theory isproposed on the basis of condition grade division and index parameter extraction. Andthe set pair algorithm and implementation steps are also constructed. The connectiondegrees and mathematical expressions are used for describing the uncertainty of states,and then combination of confidence criteria, the results of transformer conditionassessments are achieved. This method also offers a new way of condition assessmentfor power transformers.Aiming at the problem that there are so many assessing factors and indices, whichmay reflect the different aspects of transformers and have different assessing weights,the decision-making assessment model based on fuzzy and evidence reasoning for transformer insulation condition is proposed in this dissertation. A fuzzy membershipfunction is constructed to describe the factor layer of evaluation model. According tothe fuzzy evaluation results, the original basic probability assignment, which is used fordecision-making model of evidential reasoning, is determined. Thus the basicprobability assignment is obtained by evidence reasoning, and finally the assessmentresults are determined based on the decision rules of the maximum basic probabilityassignment function.The multi-classification least squares support vector machine (LS-SVM) is appliedto transformer fault diagnosis in this dissertation. And the multi-class classificationscheme is achieved by constructing more binary LS-SVM classifiers using combinationencoding. The optimal parameters of the LS-SVM classification model are obtained byusing particle swarm optimization (PSO) algorithm, and the overall generalizationperformance of the classification algorithm is improved by application of the idea ofcross validation (CV). The benchmark data sets in UCI machine learning database areemployed for validation. The cases of transformer fault diagnosis show that theproposed approach based on PSO and LS-SVM is accurate and effective. And theproposed approach has higher accuracies both in training and testing phases, comparedwith the transformer diagnosis methods of IEC three-ratio method, the back propagationneural network (BPNN), radial basis function neural network (RBFNN) and thestandard support vector machine (SVM).For the shortcomings that the classical PSO algorithm is easy to fall into localoptimum in practical applications, the PSO with time-varying acceleration coefficients(PSO-TVAC) is proposed to optimize the SVM model. Through introducing of dynamicinertia weights and acceleration coefficients, the ability of exploitation and explorationcan be controlled, and the PSO performance of global search and local search can bebalanced. The cases show that the improved approach has a faster convergence speed,higher accuracy and better diagnosis result.This dissertation studies the forecasting approaches based on support vectormachine regression (SVR) theory. The forecasting models of dissolved gases intransformer oil based on PSO-TVAC considering least squares support vector machineregression (LS-SVR) and wavelet least squares support vector machine regression(W-LSSVR) are established, which avoid the expansion of the number of unknownvariables in the traditional SVR method while simplify the parameter optimization forsupport vector machine regression. The case studies show that the proposed forecasting models both have a greater advantage in terms of prediction accuracy and stability thanthat of BPNN, RBFNN, generalized regression neural network (GRNN) and ε-SVRapproaches.In order to achieve accurate trend forecasting of gas contents in oil-immersedtransformers, a fuzzy information granulated particle swarm optimization-supportvector machine regression model is firstly proposed in this dissertation on the basis ofresearching on forecasting the dissolved gases in transformer oil. The time seriesprediction model of fuzzy information granulation is established, which simplify themanifestation for the time series model, without loss of the main information in timeseries. The granulated sample sets are trained by SVR model with PSO-TVAC. And onthe basis of the obtained prediction intervals of the information granules, the maximum,minimum and average levels of the dissolved gas contents are given by the forecastingmodel, which consistent with the actual information.