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

Natural gas is emerging as the preferred source of energy at domestic level owing to its lower environmental impact. As such, there is a global focus on the largescale adoption of natural gas as policies are being put in place in its favor. For natural gas producers, forecasting the residential natural gas demand of a large group of buildings is vital for efficient logistics in the growing energy sector. In particular, forecasting on consumption is an integral part of energy management and transportation. In temperate regions, statistics have it that 76% of the gas supplied is used for space heating, while 19% and 5% is consumed for water heating and cooking, respectively. These types of figures are crucial when developing a model for forecasting purposes.

Unfortunately, there is no widely accepted model that enables the inclusion of such socio-economic factors. Therefore, it is imperative that a gas demand forecast model that is general enough to capture the theoretically unknown behavior of the system be developed. Such a model would indulge the application of sophisticated computer models that are normally found in machine learning, statistics and artificial intelligence.

In a recent research paper published in Journal, Energy, Jožef Stefan Institute scientists: Dr. Rok Hribar, Dr. Primoz Potocnik, Dr. Jurij Silc, and Dr. Gregor Papa presented a study in which they focused on finding an easily trainable and accurate models for forecasting residential gas demand of an urban area with an hourly resolution for up to 60 h into the future. In particular, they aspired to implement and compare different forecast models for residential natural gas demand of an urban area with a goal of optimizing efficiency of logistics for the entire energy sector. Their work is currently published in the research journal, Energy.

The team compared different forecast models in order to find the most appropriate one. In their quest, three machine-learning models were applied, namely: the linear regression (LR), the kernel machine (KM) and the recurrent neural network (RNN). The team implemented the LR and RNN model using advanced techniques, new to the energy-modeling sector. Finally, three empirical models, whose structure was based on historical data analysis and theoretical considerations, were developed.

The modeling approach employed revealed that the forecasts presented were based on past temperatures, forecasted temperatures and time variables, which included markers for holidays and other occasional events. Remarkably, the two most accurate models were found to be recurrent neural network and linear regression model.

In summary, the study by Slovenia scientists presented highly accurate models for forecasting natural gas demand based on social economic considerations. Generally, a less obvious trait of the models trained as per the study was their ability to model socially driven gas and in turn obtain heat consumption within buildings. Altogether, in realistic settings, such trained models could be used in conjunction with a weather-forecasting service to generate forecasts for future gas demand.