Multistep Forecasting Models of the Liquid Level in a Blast Furnace Hearth

Abstract

The extraction of molten iron and slag in the liquid phase from the lower part of a blast furnace (hearth) is usually accomplished according to operational experience and involves a high degree of uncertainty, mainly because the liquid level cannot be directly measured. This study presents a methodology for obtaining multistep models to forecast the hearth liquid level by measuring a voltage generated on the blast furnace shell, which is strongly correlated with the hearth liquid level. The results show that this electrical signal is a nonstationary and nonlinear time-series that, after appropriate treatment, can be represented by a time-delay neural network (TDNN) model. Some comparisons are made with linear time-series models represented by an autoregressive moving average model and a seasonal autoregressive integrated moving average model, and the results indicate that the TDNN model provides better forecasting performance up to one hour ahead.