EUROPEAN STEEL TECHNOLOGY AND APPLICATION DAYS 2017 - ESTAD 2017

With increasing computing power, data storage capacity, advanced algorithms and innovative sensor technolo-gies it is today possible to approach the BOF process from another point of view: machine learning models and Data-driven Prediction Models (DdPM). These approaches process large amounts of data to predict the BOF process conditions, e.g. temperature, carbon and phosphorus content of the melt at the end of blowing (EOB). In a cooperative effort between SMS group and ArcelorMittal Gent, a detailed study based on approx. 10,000 BOF heats has been carried out. The target values of the investigation were chosen to be melt temperature TEOB and carbon content [%C]EOB. In an off-line analysis, different strategies for preprocessing and validation were employed in combination with several supervised learning approaches (e.g. Bayesian regression, Support Vector Machine (SVM), deep neural networks (DNN)) as well as different learning schemes (e.g. sliding learning). The data-driven methods can either directly predict the target values TEOB and [%C]EOB or predict deviations from the already existing metallurgical model to improve the prediction accuracy. The metallurgical model is based on known physical and chemical correlations, i.e. mass balance, energy balance, and statistical equations. It could be assessed that the DdPM model provides a higher prediction accuracy as compared to the conventional model. One aim of using offline DdPM approaches is to gain a better understanding of influence factors such as scrap type, lance pattern etc., to detect drifts or shifts in the process and to improve the metallurgical model. In a next step, the DdPM approach shall be incorporated in the online BOF process control in order to improve the model proposal and prediction accuracy. Due to the fact that SVM, DNN are not easy to interpret, the use of DdPM as a stand-alone unit might be an additional option. The paper summarizes fundamental R&D work of the partners, focuses on the applied mathematical models and shows DdPM potentials.