Temporal correlation-aware prediction and early warning method for dissolved gases in transformer oil

Abstract

A method is proposed for predicting and warning the dissolved gas content in transformer oil. This method combines the Granger causality test and neural basis expansion analysis with exogenous variables, addressing dynamic coupling and complex temporal correlations in gas component data. First, the Granger causality test is used to analyze temporal correlations in gas component concentrations and their relationship with transformer load, to identify mutual influence between time-series data and achieve dynamic variable selection of the prediction model. Then a dissolved gas prediction model is established using neural basis expansion analysis with exogenous variables. Finally, transformer status warning is achieved under threshold constraints by combining confidence interval distribution analysis of predicted gas component concentrations. The results of a case study show that this model can achieve a mean absolute percentage error below 5.0%. Dynamic variable screening based on Granger causality verification can significantly enhance prediction accuracy, providing a more reliable reference for early warning of transformer status.