TY - JOUR

T1 - Forecasting day-ahead natural gas demand in Denmark

AU - Karabiber, Orhan Altuğ

AU - Xydis, George

PY - 2020/4

Y1 - 2020/4

N2 - Natural gas demand forecasting is important for all players in the natural gas market. This work compares four possible day ahead natural gas consumption forecasting models in order to forecast the natural gas consumption of the four subnets in Denmark. The forecasts from the suggested model were used to regulate the linepack of the pipelines, which provides the stability and security of the natural gas transmission system. A detailed variable analysis, analysis of the exogenous variable error, and combination forecasts in order to maximize the forecasting accuracy are presented here. With the proposed models, a reduction in error, ranging from 34% to 72%, was achieved for each subnet in comparison to the current Energinet forecaster. Additionally, compared to a univariate model, the data rich models showed 20%–47% lower error. It was also seen that the exogenous variable error was negligible in comparison to the benefit of using variable rich models. Contrary to some of the recent studies, solar radiation was found ineffective in terms of predictive accuracy for the used data sets.

AB - Natural gas demand forecasting is important for all players in the natural gas market. This work compares four possible day ahead natural gas consumption forecasting models in order to forecast the natural gas consumption of the four subnets in Denmark. The forecasts from the suggested model were used to regulate the linepack of the pipelines, which provides the stability and security of the natural gas transmission system. A detailed variable analysis, analysis of the exogenous variable error, and combination forecasts in order to maximize the forecasting accuracy are presented here. With the proposed models, a reduction in error, ranging from 34% to 72%, was achieved for each subnet in comparison to the current Energinet forecaster. Additionally, compared to a univariate model, the data rich models showed 20%–47% lower error. It was also seen that the exogenous variable error was negligible in comparison to the benefit of using variable rich models. Contrary to some of the recent studies, solar radiation was found ineffective in terms of predictive accuracy for the used data sets.

KW - Artificial neural networks

KW - Combination forecasts

KW - Day ahead forecasting

KW - Natural gas consumption forecasting

KW - Variable analysis

UR - http://www.scopus.com/inward/record.url?scp=85079896039&partnerID=8YFLogxK

U2 - 10.1016/j.jngse.2020.103193

DO - 10.1016/j.jngse.2020.103193

M3 - Journal article

AN - SCOPUS:85079896039

VL - 76

JO - Journal of Natural Gas Science & Engineering

JF - Journal of Natural Gas Science & Engineering

SN - 1875-5100

M1 - 103193

ER -