Aplicação de modelos de séries temporais na previsão da produção, importação e consumo de fertilizantes no Brasil

Abstract

Fertilizers, in Brazil, are inputs of great importance in agricultural productivity. In this context, this work aims to compare time series models to forecast the production, import and consumption of intermediate fertilizers in Brazil. To carry out the comparison, between the models, historical series of the production, import and consumption of fertilizers in Brazil were used, in the period between 1998 and 2022. Forecast models, based on Linear Regression algorithms (LR), Neural Networks (MLP) and Support Vector Machine for Regression (SVR), were developed using the WEKA software. Results obtained from the three models were compared using the metrics RSME (Root Mean Squared Error), MAE (Mean Absolute Percent Error) and MAPE (Mean Absolute Percent Error). It was found, for a short-term horizon, that the MLP model performed better.