Forecasting natural gas consumption in residential and commercial sectors in the US

ABSTRACT The paper proposes a parallel forecasting approach for weekly natural gas consumption in the US residential and commercial sectors, which models scrape data and ratio data separately and then combines the outputs to generate the forecasts. To improve forecasting accuracy, both semi-parametric and nonparametric models, including dynamic linear regression model and dynamic semi-parametric model, are adopted to model the effects of weather variables, and time series techniques are employed to address the serial correlation exhibited by the data. An algorithm focusing on forecasting accuracy is proposed to select the smoothing parameter for serially correlated data. The proposed model is empirically tested using data in the New England area from 2013 to 2018 and benchmarked against some deep learning approaches including Deep Neural Network, Long Short-Term Memory Neural Network, and Gated Recurrent Unit Neural Network methods. Overall, the results show that the proposed approach performs well in generating accurate forecasts.