A sigmoid regression and artificial neural network models for day-ahead natural gas usage forecasting

Reliable and accurate day-ahead forecasting of natural gas consumption is vital for the operation of the Energy sector. Three different forecasting models are developed in this paper: The sigmoid function regression model, the feed-forward neural network, and the recurrent neural network model. The models were trained, compared, and validated using gas consumption data from 115 measuring stations in Slovenia and Croatia, which have been in operation for more than three years. The Genetic optimisation algorithm was used to train the neural networks and the Levenberg-Marquardt algorithm was used to obtain the parameters of the sigmoid model. The results show that both neural network models perform similarly, and are superior to the sigmoid model. The models were prepared for use in conjunction with a weather forecasting service to generate day-ahead or within-day forecasts, and are applicable to any geographical area. The neural network models achieve mean absolute percentage error between 5% and 10% in the entire temperature range. The sigmoid model reaches similar accuracy only for temperatures below 5°C, while for higher temperatures the error reaches up to 30%–40%. •Development and application of machine learning and regression models to the problem of forecasting natural gas demand.•Comparison of forecasting models taking into account different model designs and parameters.•Feed-forward and recurrent neural network models are able to achieve similar level of accuracy, both outperform sigmoid regression.•The models developed can be used in conjunction with the weather forecasting service to predict the day-ahead gas demand.