Data-driven modeling for online predictions in steelmaking

Principal supervisor: Assoc. Prof. Björn Glaser, KTH

Co-supervisor: Docent Dr. Andrey Karasev, KTH

Chair at the defense: Prof. Mikael Ersson, KTH

Opponent: Prof. Qifeng Shu, University of Oulu, Finland 

Members of the grading committee         

Dr. Dmitrij Ladutkin, Saarstahl AG, Tyskland

Assoc. Prof. Valentina Colla, Scuola Superiore Sant'Anna, Italien

Prof. Dr. Stefan Sandfeld, Jülich Forschungszentrum, Tyskland

Substitute: Assoc. Prof. Stefan Schönecker, Materialvetenskap, KTH

Abstract

In this work, innovative data-driven process models were designed, developed, and examined online for their application in the steel industry. The objective was to help operators in decision making during calcium additions in liquid steel and casting of steel grades which are susceptible to submerged nozzle (SEN) clogging. The developed data-driven process models were examined in a real industrial environment to demonstrate the use of supervised machine learning (ML) and deep learning (DL) algorithms for online predictions of specific process parameters. The adaptation of two novel data-driven decision support systems in production helped steelmakers to address the critical challenge of minimizing costly production stops caused by SEN clogging. Calcium (Ca) additions during secondary steelmaking is a well-adopted practice to transform solid oxide non-metallic inclusions (NMIs) into globular shaped liquid oxides. This practice of Ca treatment helps to avoid SEN clogging. Hence, operators in steel plants follow standard operating procedures (SOP) that detail the use of static quantity of Ca wire additions. However, these SOPs, providing a baseline for production, do not account for the dynamic nature of steelmaking processes and the varying chemistry of NMIs for each 'heat' produced. To address this gap, the impact of varying CaSi wire additions, beyond the SOP's guidelines, on the transformation and behavior of NMIs in low-alloyed steel grades was explored by developing the 'ClogCalc' decision support system. The aim was to establish a more reliable and responsive approach to Ca treatment, potentially leading to more effective control in preventing SEN clogging. The implementation of 'ClogCalc' has demonstrated a significant 30% reduction in SENclogging, as evidenced by industrial trials at Voestalpine steel plant.

Recognizing 'castability' of steel by monitoring the conditions of SEN clogging is essential for uninterrupted continuous casting. With this information prior to casting, operators can take preventive action against possible SEN clogging occurrences, thus reducing unplanned downtimes. In response to the severe implications of SEN clogging, a novel approach to forecast 'castability' of steels was introduced by developing the 'Checkcast' decision support system. The adaptive neuro fuzzy inference system (ANFIS) and long short-term memory (LSTM) network model was used as a base algorithm for estimating 'castability' of steel grade. The output of 'Checkcast' helps to anticipate the event of clogging by analyzing both the past condition of the SENs and changes in the steel chemistry during the transport of the steel from refining ladle to the tundish. The verification was conducted at the Sidenor steel plant. While the primary focus of this study was to develop data-driven process models, efforts were also made to study the foundational principles governing the SEN clogging and evolution of NMIs in studied steel grade. Laboratory experiments were also conducted on liquid steel samples obtained from steel plants. Parametric liquid windows were derived using systematic thermodynamic assessments using FactSage software. In addition, to understand the wetting characteristics at interfaces of NMIs and the SEN refractory, interfacial properties were calculated using ab-initio calculations.

urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-354824