Natural gas, forecasting, industrial, time series, consumption
Abstract
The forecast of natural gas consumption assumes crucial importance in the short and long-term planning of distributing companies. Anticipating the volume of gas to be distributed not only enables future investments in infrastructure and resource management policies, but also avoids potential economic losses. Despite the relevance of natural gas in the Brazilian energy matrix, particularly for the industrial sector, this significance is not adequately reflected in research in the field, creating a gap in published studies, especially regarding short-term forecasts. In this context, the objective of this study was to conduct a daily forecast study, with horizons of 1, 4, and 7 steps ahead, for three specific natural gas Receiving Stations. To this end, the treatment applied to the time series of each delivery point was presented, followed by the evaluation of classical, machine learning, and deep learning models. The results obtained show that forecasting natural gas consumption by receiving stations is not a trivial task and that machine learning models did not outperform classical forecasting methods.
