A comparison of models for forecasting the residential natural gas demand of an urban area

• Published in: Energy (Oxford) 2019-01, Vol.167, p.511-522
• Main Authors: Hribar, Rok, Potočnik, Primož, Šilc, Jurij, Papa, Gregor
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Buildings; Data analysis; Data processing; Demand; Demand forecasting; Economic forecasting; Empirical analysis; Energy modeling; Forecast accuracy; Learning algorithms; Logistics; Machine learning; Natural gas; Neural networks; Recurrent neural networks; Regression analysis; Regression models; Residential areas; Residential buildings; Residential energy; Urban areas; Weather forecasting
• ISSN: 0360-5442; 1873-6785
• DOI: 10.1016/j.energy.2018.10.175

Forecasting the residential natural gas demand for large groups of buildings is extremely important for efficient logistics in the energy sector. In this paper different forecast models for residential natural gas demand of an urban area were implemented and compared. The models forecast gas demand with hourly resolution up to 60 h into the future. The model forecasts are based on past temperatures, forecasted temperatures and time variables, which include markers for holidays and other occasional events. The models were trained and tested on gas-consumption data gathered in the city of Ljubljana, Slovenia. Machine-learning models were considered, such as linear regression, kernel machine and artificial neural network. Additionally, empirical models were developed based on data analysis. Two most accurate models were found to be recurrent neural network and linear regression model. In realistic setting such trained models can be used in conjunction with a weather-forecasting service to generate forecasts for future gas demand. •Application of machine learning models to the problem of forecasting residential natural gas demand.•Development of empirical models that stem from the analysis of historical data.•Comparison of models considering various criteria for different forecasting horizons.•Linear regression and recurrent neural network are best under different criteria.•Models are usable in conjunction with weather-forecasting service for real world problems.