Previsão de propriedades físicas e metalúrgicas do sínter através de redes neurais artificiais e regressão linear múltipla.

Abstract

Due to the latest technological developments and high investments in research, steel processes have become increasingly competitive, through the construction of robust and sophisticated systems to map and model the processes involved. Parallel to this, the quality of raw materials, especially iron ore, decreases over time, increasing the need to develop statistical and mathematical models capable of controlling the process, in order to optimize the consumption of these resources, coupled with the reduction of costs. Therefore, this paper aims to evaluate the performance of two statistical methodologies - Artificial Neural Networks (ANN) and Multivariate Linear Regression (MLR) - in the prediction of physical properties (Degradation Index and Mean Particle Size) and metallurgical (Reduction Degradation Index and Reducibility Index) product of an iron ore sintering plant in order to improve process control and understanding the complex relationships between input and output variables. To achieve this paper’s objectives, a database of process variables, chemical and physical analysis of the sinter produced by Sinter Plant of the Brazilian steel company under study was used to feed the algorithms based on Artificial Neural Networks and Multivariate Linear Regression, developed in Matlab and Microsoft Excel, respectively, in order to obtain, as main results, the regression coefficient (R2) and absolute mean percentage error (AMPE) for each statistical modeling. The model in general obtained good results for the Degradation Index in both techniques, with R2 ranging from 0.43 to 0.71 with AMPE below 0.52%, which demonstrates its applicability at the industrial level. However, for the other properties, the models were unable to provide satisfactory results, with R2 ranging from 0.14 to 0.62 and AMPE ranging between 2 and 7%, which indicates that for such variables, the model serves only as trend guides, unless the models are refined by improving sinter sampling processes and implementing new input variables.