Daily natural gas consumption forecasting based on a structure-calibrated support vector regression approach

Abstract

An accurate forecast of natural gas (NG) consumption is of vital importance for economical and reliable operation of the distributive NG networks. In this paper, a structure-calibrated support vector regression (SC-SVR) approach is proposed to forecast the daily NG consumption, which is correlated with the past time series using the SVR model. To better accommodate the dynamic nature of the NG consumption, the structural parameters of the SVR model are online calibrated in response to the receding horizon of the NG consumption series. The calibration of the structural parameters for the next-day forecast is performed by extended Kalman filter. The proposed SC-SVR approach is evaluated using real data collected from a NG company in the period from January to December 2012. The results indicate that the mean absolute percentage error and the root mean squared error are 2.36% and 3913.88m3/d, respectively. To show the applicability and superiority of the SC-SVR approach, two peer methods, i.e., least squares SVR model and dynamic back propagation neural network are also employed for comparison. The results show that, thanks to nonlinear mapping capability of the SVR and dynamic nature of the online calibration for the model structure, the proposed SC-SVR method is capable of improving the forecast accuracy for the daily NG consumption.