Research On Methods Of Fault Diagnosis And Prognosis For Power Transformer Based On Improved Support Vector Machines

Power transformers have a high place in power system. It plays an important role in changing voltage and transmission of electric energy. It also has severely affected the stability and security of the whole power system. Failure of a transformer could cause heavy losses. Fault diagnosis and prediction are the basis of normal-running and condition-based maintenance for power transformer. Aiming at characteristics of small samples and less information in power transformer fault diagnosis data, a series of researches on fault property diagnosis, fault position identification and fault prediction is made in the thesis, which is based mainly on support vector machine theory and combined with other intelligent methods. And an assistant decision system of condition-based maintenance for power transformer is also developed.Firstly, fault property diagnosis in power transformer is studied. The thesis had an in-depth research on multi-class support vector machines algorithm based on binary tree, analyzes and compares classifying accuracy and speed of multi-classification methods with different tree structures. Aiming at the problem that binary tree structure is determined by experiential knowledge subjectively in the area of incipient fault property diagnosis for power transformer, the method that binary tree structure is shaped according to data samples and clustering is discussed. A novel tree-built approach based on divisive hierarchical strategy and fuzzy c-means clustering is proposed. Combined with support vector machines classification, fault property diagnosis model based on fuzzy C-means and binary tree support vector machines is constructed. Based on references and fault samples, fault property is divided into seven classes. Characterized by five characteristic gases dissolved in transformer oil, a great deal of transformer fault diagnosis tests has been done. The results show that, the proposed method not only overcomes misidentification caused by non-coding region and absolute boundary of three ratio method, but also avoids unclassifiable region which the "one-against-one" and "one-against-rest" method has, and improves diagnosis accuracy.Secondly, fault position identification in power transformer is studied. The research on posterior probability estimates for support vector machine is made deeply. By utilizing "one-against-one" multi-class support vectormachines and pairwise coupling, coupling probability is transformed into multi-class probability. It solves probability outputs of multi-class support vector machine effectively. Aiming at the difficulty that evidence theory can hardly determine basic probability assignment in transformer information fusion diagnosis method, regarding probability outputs of support vector machines as basic probability assignment under frame of identification is proposed. It realized the objectivity of basic probability assignment valuation and avoided complexity of basic probability assignment valuation in information fusion. Integrating support vector machines and evidence theory, the thesis treat dissolved gases analysis data and routine electrical tests data with three different support vector machines, and synthesize evidence bodies determined by probability outputs of support vector machines. Finally, a fault position identification model for power transformer is proposed, which is based on multi-support vector machines and D-S evidence theory. Fault diagnosis examples show that fault positions can be identified effectively by support vector machines and D-S fusion model, and both accuracy and generalization are better than single characteristic support vector machines method and multi-neural network and evidence theory fusion method.Thirdly, concentration prediction method of dissolved gases in transformer oil is studied. The principle of grey relational analysis and fuzzy support vector machines regression theory is discussed in this paper. At present, existing concentration prediction models for dissolved characteristic gases lack considerations of the influence of oil temperature, loads and interaction among gases. To solve the problem, the thesis makes use of grey relational degree analyses recent dissolved gases analysis data to extract strong-correlated factors and eliminate weak-correlated factors. Considering different influence degree of recent samples and earlier samples, each input sample is assigned to different weights according to its sampling time. The samples from recent past is given more weighting than the samples far back in the past, which reflects the later data had a greater impact on the following prediction results than the earlier data. Combined grey relational analysis with fuzzy support vector machines, concentration prediction model of dissolved gases in transformer oil based on grey relational analysis and fuzzy support vector machines is proposed. The prediction results of actual cases show that prediction accuracy of the proposed model is better than that of the standard support vector machines, fuzzy support vector machines and grey relational upport vector machines model.Lastly, a decision support system model of condition-based maintenance for power transformer is developed.On the basis of analyzing process of decision-making, structure of assistant decision system is designed. Fault diagnosis model and fault prognosis model is also constructed,which is based mainly on aforementioned methods of fault property diagnosis, fault position identification, dissolved gases concentrations prediction, and abide by relevant guide for condition-based maintenance. Finally, adopting Eclipse platform, the assistant decision system is developed using the Java language and ORACLE database. This system is being used by electric power companies and is helpful to realize the standardization, scientification and institutionalization of condition-based maintenance for power transformer. requirements of assistant decision system of condition-based maintenance for power transformer on intelligence, distributivity, openness, security, adaptability and so on.