All Times ET
Keywords: nonparametric model, multi-scale model, time series, deep learning
In the U.S., the agency of Energy Information Administration (EIA) provides monthly natural gas usage data for each state which are believed to accurately represent the natural gas consumption. However, the releasing date of such data lags by three months and the monthly aggregation tends to mitigate the weather sensitivity and smooth out the fluctuations in daily or weekly natural gas consumption. It is of great interests to develop forecasting models based on weather variables to predict natural gas consumption on a continuous week to week basis.
We propose a parallel forecasting strategy of weekly natural gas consumption that can be practically implemented by energy companies within the constraints of their data resources. Using the proposed method, various nonparametric models and deep learning architectures are employed for forecasting the weekly consumption of natural gas in the residential and commercial sectors of the New England area based on the data from January 7th 2013 to February 4th 2018.