Algorithm Study And Application Of Transformer Condition Analysis And Maintenance Decision-making Base On Bayesian Control Chart Model

At present, in the power system, oil-immersed transformers are commonly used in industrialand mining enterprises and civil supplying system. It operates safe or not concern thedevelopment of the national economy and entire electric grid. Because transformer itselfcomplex internal structure and the complex and changing environment, there are a lot ofsecurity risks in the long-running process. Therefore, taking full use of existing historicalmonitoring data to analyze transformer fault state, it means a lot for carrying out the work ofelectric grid security. It helps to save costs and increase economic and social benefits.Transformer monitoring includes chemical and electrical monitoring. chemical monitoringdata have more reference value, because its data are more and that is convenient to researchand analyze. The most intelligent methods of transformer fault diagnosis are based on thechemical monitoring data. Chemical data mainly consist eight characteristic gases: hydrogen,methane, ethane, ethylene, acetylene, carbon monoxide, carbon dioxide and totalhydrocarbons.The main research contents can be concluded as follows:(1)Preprocessing the transformer monitoring data and extracting characteristics, respectively,using the Vector Autoregressive model with exogenous variable (VARX) and Partial leastsquares regression (PLS regression) to "clean" monitoring data, which can reflects thetransformer fault information best.(2) Using Bayesian formula continuous iterative to calculate the posterior probability, and useBellman dynamic programming equation to calculate the optimal alarm threshold. Thismethod breaks the limitations of the conventional method which uses a fixed threshold values,and the result is more scientific and reliable. According to the experts’ experience and powercompanies need to adjust the model sensitivity of the model to achieve the best effect. Everytime we get monitoring data, the model real-time updates posterior probability, compere thealarm threshold, then make the decision: continue to run or outage to maintain.(3) The research results of the model in conjunction with other traditional four comprehensivediagnostic applications, respectively, to the various methods give different weights and molecular composite score, according to the empirical formula to determine the severity ofthe equipment failure. This method has been run in the practical application.