Deep neural networks are proposed for short-term natural gas load forecasting. Deep learning has proven to be a powerful tool for many classification problems seeing significant use in machine learning fields such as image recognition and speech processing. We provide an overview of natural gas forecasting. Next, the deep learning method, contrastive divergence is explained. We compare our proposed deep neural network method to a linear regression model and a traditional artificial neural network on 62 operating areas, each of which has at least 10 years of data. The proposed deep network outperforms traditional artificial neural networks by 9.83% weighted mean absolute percent error (WMAPE).

In this paper, we first propose a modified definition for timescale separation ratio for the discrete-time system. Then, based on this new timescale separation ratio, a timescale analysis is presented to decompose the system into fast and slow subsystems. Next, using the difference (delta) operator, a timescale analysis is presented using a unified formulation (applicable to both continuous-time and discrete-time systems). Furthermore, it is shown that by using simple transformations, one can obtain the discrete-time results we developed earlier, and by similarly setting the sampling interval to zero, we obtain the continuous-time result developed previously by others. Thus, with this unified approach to a timescale system, a single technique is sufficient to develop the techniques for the continuous-time and discrete-time systems. Having a pair of routines for each task emphasizes the differences between discrete-time and continuous-time theory rather than the similarities. An F-8 fighter aircraft is given as an example to illustrate the methodology.