Impacto da Ordem entre a Seleção de Características e a Janela Deslizante em Séries Temporais Multivariadas na Previsão do Consumo de Gás na Pelotização de Minério de Ferro

Abstract

This study explores machine learning techniques for predicting gas consumption in the iron ore pelletizing process, which constitutes a significant portion of the budget and increases the environmental impact due to its burning if used excessively. We used a set of data collected over 90 days in an industrial pelletizing plant, with 46 operational parameter variables. Our methodological approach transforms multivariate time series data into a tabular format using sliding window technique. Our primary objective is to ascertain the best sequence for applying feature selection—either before or after the sliding window method. We employed three feature selection methods: Pearson Correlation, AdaBoost, and Random Forests, and the regression models are AdaBoost, Random Forest, and Multilayer Perceptron (MLP), were evaluated using the Root Mean Square Error (RMSE) metric. The results based on RMSE and training time indicate that applying the sliding window followed by feature selection is the best order. In this approach, the best result in the case study obtained an RMSE of 0.39 with selection via Random Forest followed by MLP.