Previsão computacional do teor de silício no ferro-gusa

Abstract

The silicon content of pig iron is one of the main indicators of its quality and the thermal state of the blast furnace. Furthermore, a high silicon content can damage industrial equipment, leading to the need for maintenance and, consequently, a loss of process efficiency. For these reasons, several studies have been developed over decades to predict and monitor the silicon content in pig iron, suggesting the use of data-driven models. In this context, this work tested models such as the logistic perceptron, multilayer perceptron artificial neural networks with up to two hidden layers, and different versions of support vector machines adapted for regression (SVR, TSVR, LSSVR) to predict the silicon content in pig iron. A technique for estimating the number of hidden neurons in neural networks based on singular value decomposition (SVD) was also investigated to reduce tuning time and computational cost. Among the neuron-based models, the neural network with one hidden layer presented the best balance between performance and computational cost, while the SVD-based technique provided a smaller hyperparameter testing window, therefore, it was used in a sensitivity analysis to study the influence of each input variable on the silicon content in pig iron.