Técnicas computacionais aplicadas à previsão de propriedades do sínter metalúrgico

Abstract

Both the need to produce metallurgical sinter with better and more stable properties, regardless of variations in the quality of the inputs used, such as iron ore and coke, as well as the need to ensure competitiveness in steel production in the face of a scenario of constant global economic crises, are motivating factors for the development of forecast models applied to the steel industry. This work proposes the development of computational tools to estimate metallurgical sinter quality indexes from their chemical characteristics and sintering process variables. The investigated indexes are Shatter Resistance Index (SI), Reducibility Index (RI), Degradation Under Reduction Index (RDI), and Average Particle Size (MPS). Investigating the inĆuence of input variables on the considered quality indicators, evaluating the quality of the estimating models developed and comparing them with results available in the scientiĄc literature, as well verifying the possibility of using a new hyperparametric optimization technique are other objectives of this work. Different algorithms were used to obtain the best prediction model for each of the studied responses, including multiple linear regression (MLR), stepwise regression (SR), multiple perceptron neural network with gradient descent with momentum learning algorithm (MLP-GDM), and multiple perceptron neural network with LevenbergŰMarquardt learning algorithm (MLP-LM).