Mathematical modelling of sinter quality and analysis of variables in a sinter plant

Abstract

Achieving consistent sinter quality is essential in steelmaking, despite variability in raw material inputs. This study analyzes 426 daily production records from an industrial sintering plant over 19 months to develop predictive models for four key quality indices: Shatter Index (SI), Reducibility Index (RI), Reduction Degradation Index (RDI), and Mean Particle Size (MPS). Modeling approaches included multiple linear regression and multilayer perceptrons (MLP) trained via gradient descent with momentum (MLP-GDM) and the Levenberg -Marquardt algorithm (MLP-LM). The MLP-LM models showed the best performance: for SI (R² = 0.7410, MAPE = 0.36%), MPS (R² = 0.6030, MAPE = 4.99%), and RDI (R² = 0.4478, MAPE = 5.79%). RI was less predictable (R² = 0.2660, MAPE = 2.29%). Results indicate that neural networks outperform linear models in capturing the process's nonlinearities, especially for mechanical strength and granulometry. A novel contribution of this work is the use of singular value decomposition (SVD) on hidden layer outputs to estimate the optimal number of neurons, significantly reducing hyperparameter search time. SVD-based estimates closely matched those found through exhaustive tuning -- differing by no more than four neurons -- streamlining architecture selection from hours to seconds. These findings support the use of data-driven modeling for process optimization in steelmaking. SI and MPS models show accuracy levels suitable for routine monitoring and decision-making, while RDI and RI models require further refinement before deployment.