Application of neural network techniques for modeling of blast furnace parameters

Abstract

This thesis discusses the predictions of various output variables in a blast furnace. It compares the ability of multi-layer perceptron neural networks for prediction with other blast furnace prediction techniques. The output variables: Hot Metal Temperature, Silicon Content, Slag Basicity, RDI, and +10 are all modeled using the MLP networks. Different solutions are proposed for preprocessing the original data and finding the most relevant input variables. The NNRUN software is used to find the best MLP neural network. Finally, methods to control the output variables in the blast furnace are examined and a derivative-based sensitivity analysis is discussed.