Um Estudo Comparativo de Técnicas de Aprendizado de Máquina para a Predição de Temperatura em Regeneradores

Abstract

This article examines the application of machine learning models, specifically XGBoost (XGB) and LightGBM (LB), among others, for predicting hot blast stove temperatures in an integrated steel mill. Unlike previous studies, the dataset used in this investigation does not exhibit time series characteristics, complicating the use of algorithms that depend on such features. The dataset, provided by a steel mill in Brazil, underwent an initial feature pre-selection through expert interviews, followed by Principal Component Analysis (PCA). Preprocessing steps included outlier removal using the Interquartile Range (IQR) and missing value imputation with moving averages. The models were evaluated using metrics such as R2 and RMSE. After hyperparameter tuning via GridSearchCV, LB and XGB demonstrated superior performance across both metrics. The results indicate the significant effectiveness of these models, with LB showing a slight advantage.