Aplicação de técnicas de aprendizado de máquina e estatística na previsão de demanda de biocombustível

Abstract

The analysis and forecast of strategic demand are fundamental in the planning of the production chain, being significant in several segments. The transport sector is a good product, a highly dynamic consumption, which is an example of demand on demand. The scenario provides an opportunity for the application of predictive configuration and automatic learning techniques, these resource opportunities being of great value for understanding the behavior of the term demand. With the development of computational capabilities, it was possible to improve mathematical modeling, used for time series forecasting. The present work shows different techniques of machine learning and statistics, to verify the performance of these techniques in forecasting the demand for biofuel. With the help of the Python programming language, the sales data of biofuels, ethanol, and biodiesel, were used for a modeling using three methods:ARIMA, Long Short-Term Memory - LSTM and Gradient Boosting. During the residual analysis, it was observed that the ARIMA models show higher quality in the configurations. However, from the results obtained and through the MAPE error metric (Mean Absolute Percentage Error), the LSTM method has the best performance, with a MAPE error of 11.1% for biodiesel and 11.3% for ethanol.