Short Term Trend Forecast of On-Line Monitoring Data of Dissolved Gas in Power Transformer

Abstract

As the transformation equipment of electrical energy, transformer is the key equipment in power system. When it comes to breakdown, there will be a great loss. Dissolved gas analysis (DGA) online testing is an important indicator of transformer health assessment which is widely used in insolation testing as it is sensitive to discharge defect. Nowadays, most of researches on DGA online testing analysis are aimed at the faults diagnosis. However, in some conditions, the fault may develop very rapidly. The operation and maintenance personnel don't have enough time to figure the problem out before the breakdown occur. This problem can be solved by forecasting the DGA on-line testing data effectively. With the reveal of deterioration trend, the serious failure will be avoided and the reliability of transformer will be improved. This paper proposes a short term trend forecast method based on online data optimization for dissolved gas in oil, which is a time series forecast. This method is made up of five parts: data optimization, related gases selection, the orders selection, model parameters estimation and model checking, multi-step forecast. With the field interference and DGA online testing device status error, the DGA online data's quality can't be assured. To improve the accuracy of forecast model, the transformer online testing data needs to be optimized in first step, including Pauta criterion removing for singular value and linear interpolation for missing data. The second step, select related gases, as different gases have strong relationship. The third step is to build forecaster model based on Auto-Regressive and Moving Average Model (ARMA), using Akaike information criterion (AIC) to select the model orders. The forth step, estimate the unknown parameters by least square method. After that, the model should be verified by residual error testing to make sure the effective information of the time series is fully extracted. The final step, use the forecast model to get the DGA forecast value by multi-step forecast. In this way, the short term deterioration trend can be reveal. About 323 normal transformers' one-year data and an overheat case's data are used to test the method, with research findings: 1) the forecast method has good short term forecasted accuracy, forecast error less than 10%. It reveals that the model can be used in the short-time dissolved gases forecast. However, if the value is too small as C2H4 or strong volatility as CO2, the ARMA forecast accuracy decreases sharply. 2) The longer time span, the larger forecast error will be, especially when it comes to the changes in condition. It's supposed that the model's response time influence the forecast error greatly. The further step is to reduce the method's response time.