Demand Forecasting Using Decomposition and Regressors of Natural Gas Delivered to Consumers in the U.S

Abstract

Natural gas is simple to use and easy to convert to different types of energy. Thus, its usage is increasing every year and this increase also validates itself in consumption. In this study, demand forecasting of natural gas delivered to U.S. consumers is studied. In the study, the time series decomposition method (TSD) and estimation of residuals after TSD method are performed with independent variables have been studied. Thus, more accurate model is tried to be obtained by combining residual estimates with TSD. Meteorological data, natural gas variables such as price, storage capacity and economic indicators are used in the residuals estimation. Sixteen variables are observed as effective in total thirty-eight independent variables. Residual modeling is performed by multiple linear regression where the insignificant variables are removed from the model. It is seen that standard precipitation index of 24 months (SP24), natural gas sold to commercial consumers (PNSCC), total natural gas underground storage capacity (NUSC) are effective independent variables on residual forecasting. Since the predictability is challenging, different models have been built according to whether SP24 is included in the model or not. The 24-month forecasts are made with TSD and 5% mean absolute percent error (MAPE) is obtained. The proposed models estimated 1.5% and 4.5% MAPE respectively for SP24 and not SP24 cases. The models decreased error on average 35.9%, and 15.4% respectively. The approach used in the study improved forecasting results for natural gas delivered to consumers in the U.S.