Time Series Prediction of Dissolved Gas Concentrations in Transformer Oil using ARIMA-NKDE Method

Abstract

The dissolved gas concentrations in oil are key indicators to characterize the condition and possible fault types of power transformer. It is important to accurately predict the dissolved gas concentrations in transformer oil, and hence to timely determine fault development of transformer. In this paper, we propose a so-called ARIMA-NKDE method to predict the dissolved gas concentrations in oil. Firstly, the time series of dissolved gas concentrations in oil are decomposed at different scales by discrete wavelet transform (DWT), and the wavelet coefficients can be obtained. Then, autoregressive integrated moving average (ARIMA) is applied to predict the future evolution of wavelet coefficients. The inverse transform of discrete wavelet is used to reconstruct the future evolution of wavelet coefficients. In order to further improve the accuracy of prediction, the probability density function of the dissolved gas concentrations prediction error is constructed via non-parametric kernel density estimation (NKDE), and the predicted time series of gas concentrations under different confidence levels are obtained by coefficient inversion. The results show that the prediction accuracy achieve 92%, which is significantly higher than ARIMA (85%) and LSTM (81%) methods.