Prediction of Dissolved Gas Content in Transformer Oil Using the Improved SVR Model

Abstract

Dissolved gas analysis in oil is an effective method for early fault diagnosis of transformers. Predicting future concentrations of characteristic gases can aid maintenance personnel in assessing the operational trends of transformers, thereby ensuring stable performance. To address the challenge of predicting dissolved gas content caused by inherent nonlinearity and non-stationarity, this paper proposes an ensemble empirical mode decomposition-cuckoo search-support vector regression (EEMD-CS-SVR) combined prediction model, utilizing ensemble empirical mode decomposition and support vector regression optimized by the cuckoo search algorithm. Firstly, EEMD is used to decompose the original dissolved gas content time series into a set of stationary modal components. Subsequently, SVR, known for its strong predictive performance, is employed to predict each modal component separately. Finally, CS is applied for global search to optimize and select SVR parameters, with the predicted dissolved gas content results being overlaid and reconstructed. Simulation experiments on H ₂ content show the mean absolute percentage error of 1.81% and the root mean square error of 0.707 µL/L, significantly enhancing prediction accuracy. Further validation through modeling and predicting CO and CH ₄ confirms the model's high accuracy and suitability for forecasting dissolved gas content in transformer oil.