Forecasting residential gas consumption with machine learning algorithms on weather data

¹ Faculty of Technology, Innovation and Society, The Hague University of Applied Sciences, Rotterdamseweg 137, 2628 CN, Delft, The Netherlands
² Faculty of Engineering, University of the Basque Country, Paseo Rafael Moreno 3 48013 Bilbao, Vizcaya, Spain
³ Escuela Politécnica Superior, University Francisco de Vitoria, Carretera Pozuelo a Majadahonda, Pozuelo de Alarcón, Madrid, Spain

^* Corresponding author: t.b.salcedorahola@hhs.nl

Abstract

Machine learning models have proven to be reliable methods in the forecasting of energy use in commercial and office buildings. However, little research has been done on energy forecasting in dwellings, mainly due to the difficulty of obtaining household level data while keeping the privacy of inhabitants in mind. Gaining insight into the energy consumption in the near future can be helpful in balancing the grid and insights in how to reduce the energy consumption can be received. In collaboration with OPSCHALER, a measurement campaign on the influence of housing characteristics on energy costs and comfort, several machine learning models were compared on forecasting performance and the computational time needed. Nine months of data containing the mean gas consumption of 52 dwellings on a one hour resolution was used for this research. The first 6 months were used for training, whereas the last 3 months were used to evaluate the models. The results showed that the Deep Neural Network (DNN) performed best with a 50.1 % Mean Absolute Percentage Error (MAPE) on a one hour resolution. When comparing daily and weekly resolutions, the Multivariate Linear Regression (MVLR) outperformed other models, with a 20.1 % and 17.0 % MAPE, respectively. The models were programmed in Python.