Tapped Alloy Mass Prediction Using Data-Driven Models with an Application to Silicomanganese Production

The accounting mass balance on pyrometallurgical plants, to which the tapped masses of alloy and slag are essential inputs, forms an integral part of the process control and planning of any smelter. Thus, it is desirable to be able to predict tapped mass ahead of time. This paper examines three data-driven models that aim to predict tapped alloy mass for a submerged arc furnace producing silicomanganese. All the models are linear and based on lagged data and, thus, can be described as autoregressive models with exogenous inputs (ARX). They differ in the selection of the predictors. Ordinary least squares (OLS) model uses all of the available predictors, whereas least absolute shrinkage and selection operator (LASSO) model selects most important predictors and partial least squares (PLS) model finds best predictors in the latent space. Feature selection analysis is performed on the model results. It is shown that the model based on OLS with a reduced number of the predictors slightly outperforms other models. It is shown that power input is the strongest predictor of the tapped alloy mass, confirming current industry practice. Tap duration, energy input corresponding to the previous tap, and the previously tapped alloy mass are shown to be weak but statistically significant predictors. Using mutual information, it was shown that it was not possible to improve tapped mass prediction accuracy using tapping and recipe data alone.