Candidate: Master of Science Tero Vuolio
Place: L10, Linnanmaa,
Remote connection: zoom
Date & time: 29.01.2021 12:00
Topic: Model-based identification and analysis of hot metal desulphurisation
Opponent: Adjunct professor Mikko Helle
Custos: Professori Timo Fabritius
Sulphur is considered one of the main impurities in steel. Hot metal desulphurisation serves as the main unit process for sulphur removal in the production of steel. The main objective of this thesis is to identify the relevant phenomena and attributes needed to construct a mathematical model suitable for online use. The study also includes a detailed literature review on the modelling of hot metal desulphurisation, which considers a categorisation of the existing models for the process, but also outlines the main uncertainties in the process that may decrease the prediction performance of the existing models.
In this study, model-based process identification techniques are studied. More specifically, the objective is to study different techniques, both to explain the variance and to predict the end content of sulphur in the process. To do this, a modelling framework exploiting data-driven and mechanistic modelling techniques is proposed. The model identification procedure is divided into variable construction, variable selection, model structure selection, and model parameter identification steps. The model identification procedure considers both manual and automatic model identification techniques. The thesis focuses on grey box and black box model structures. In automatic model identification, the focus is on evolutionary search strategies, particularly genetic algorithms.
The results of this study show that in the case of lime-based hot metal desulphurisation, the major factors inducing variance in the end content of sulphur are related to the properties of the reagent, i.e. to the rate of the transitory contact reaction. If the particle size distribution is known a priori or can be assumed constant, the prediction accuracy of the models can be improved considerably. In addition, the parameterisation of the reaction models improves the prediction performance. It was also found that physically meaningful descriptions for the uncertain phenomena may help to constrain the search of parameters. In addition, in-depth phenomena-based analysis and automatic model identification strategies may assist in model selection.
Last updated: 2.2.2021