Previsão de séries temporais de consumo diário de gás natural no Brasil

Abstract

This work aims to define a short-term forecast model for Natural Gas (NG) consumption in Brazil. NG plays a crucial role in Brazil's energy matrix, impacting several sectors such as industry, electricity generation and residential heating. Forecasting short-term natural gas consumption is essential for regulatory compliance and economic efficiency, especially for distribution companies. The complexity of NG consumption, influenced by economic parameters, meteorological data and calendar factors, presents challenges for accurate predictions. The aim is to fill a gap in the literature of Brazilian studies and limited research on forecasting in Brazil, which highlights the need for localized studies and models, to which it's intended to contribute. The methodology is a comparison of several models such as Regression, ARIMA, Random Forest, Neural Networks and Fuzzy Time Series. The proposed model combines the Yamakawa Neo-Fuzzy-Neuron with the Unscented Kalman Filter (UKF) for daily and weekly prediction of NG consumption. The experimental results highlight the model's improvements over existing approaches. Finally, it is concluded that Yamakawa's adaptive model is always the most efficient, and that it can be used whenever there are two or more time series, since each of them advances independently, then everything comes together in a UKF, which filters the means and standard deviations.