Research On Ensemble Learning Of Fault Diagnosis And Prediction And Maintenance Decision-Making Models For Transformers

The transformer is one of the core equipments of the power system, which operating state has close contact with power system safe and reliable operation. With the development of UHV/EHV and the expansion of interconnected power system, the transformer faults will cause great damage and impact on the power system. Therefore, it is significant for improving the safety and reliability of power system to effectively predict and diagnose the potential faults of transformers and develop scientific and rational transformer maintenance strategy. In this dissertation, based on the classification of the main transform fault modes, the key issues are researched in depth including the relevance amount between transformer features and failure modes, the generalization ability of transformer fault prediction models and fault diagnosis models, and the failure impact assessment and the failure probability estimate required by the risk assessment for the transformer faults. The main research work in this dissertation includes:①The method to quantify the relevance between the failure modes and the fault features of transformers is researched, and the association rules analysis method to obtain the confidence value of fault features of transformers is proposed. The Apriori algorithm for extracting the association rules between the failure modes and the fault features of transformers is studied. The multi-valued discretization method of sample data is researched. Then the role and influence of the multi-valued discretization method for the association rules analysis are analyzed. Example results show that, after the multi-valued discrete processing, association rules analysis has the capacity to quantify the relationship between the failure modes and the fault features of transformers.②The faults diagnosis methods for transformer of multi-features by regarding the confidence values of the fault features of the transformer as a priori knowledge is studied. Then a transformer fault diagnosis model is proposed integrated the clustering analysis with the support vector machine. The method to construct the decision-making tree for the faults classification is analyzed. Examples show that, for the fault diagnosis of multi-features on the transforms, the accuracy of diagnosis based on the combination model is higher then the single clustering or support vector machine.③The ensemble learning algorithm for fault diagnosis of transformers is researched. The method for representing the reference value of the samples by sample information entropy is described. A number of training samples subsets are constructed by the resampling process and the Information Entropy-Based Bagging algorithm is proposed for ensembling those single fault diagnosis models. Example results show that, IE-Bagging is able to improve the generalization ability and diagnostic accuracy of the single fault diagnosis models.④Methods to improve the generalization ability of the feature prediction model for fault features are studied. Sample probability is calculated by combining the sample entropy with its subjective information; then, an improved Bagging algorithm (E-Bagging) is proposed based on comprehensive sample entropy. Moreover, E-Bagging integrated with both support vector regression machine and combinatorial forecasting algorithms are also presented. Results indicate that this E-Bagging integration strategy can improve accuracy and generalization ability of solo or combinatorial forecasting models.⑤An evaluation method for fault features prediction model is proposed based on the Bayesian information criterion (BIC). Model accuracy and its complexity index as well as the relationship between sample number and complexity factors are then analyzed, followed by comparative study of BIC evaluation index for integrated algorithm in different sample scenarios. Experiments indicate that the E-Bagging integrated with combinatorial forecasting is the optimal model for large sample sets whereas the E-Bagging incorporated with SVRM outperforms other forecasting models for small sample sets.⑥Aiming at risk evaluation of transformer faults, methods to estimate faults'consequence and probabilities are studied. The fault loss is classified into four kinds of monetary loss including transformer's own damage, personnel loss, grid loss and social loss; total loss is used to evaluate the fault consequence quantitatively. Information entropy is introduced as the priori-knowledge to measure distributions of samples and objective function of fuzzy clustering is refined, then an improved fuzzy clustering algorithm for fault probability estimation is proposed accordingly. Examples show that the above methods for consequence and probability estimation can obtain more accurate results than that of expert score and Weibull distribution, respectively.⑦Based on risk evaluation, fundamental procedure and functional units of a decision-making model for transformer maintenance are established. First, strategy of using expectation of maintenance expense to quantitatively guide the maintenance plan is analyzed; then, this expected maintenance expense is used as the objective function of decision-making to establish the optimal maintenance model. Finally, examples illustrate the detailed steps of model establishment.