An Interpretable Random Forest Approach for Predicting Specific Mould Flux Consumption in Continuous Casting of Steel Slabs

Mould powders play a crucial role in the continuous casting of steel by providing lubrication, regulating heat transfer, capturing inclusions, preventing re--oxidation and offering thermal insulation at the meniscus. Accurate prediction of mould powder consumption is a valuable indicator for process stability. Thus, numerous empirical relations were proposed but give only limited guidance, beyond casting speed and slag viscosity, and there is little consensus on the wider set of factors that govern powder behaviour across grades, sections and powders. This thesis deals with the development of a data driven model for predicting slab level specific mould powder consumption (kg/m²) and for identifying the principal factors that influence it. A curated dataset was prepared from Level-Two measurements at voestalpine Stahl GmbH, where each record represents averaged values for a slab. These inputs were augmented with engineered descriptors such as the surface to volume ratio and rheological properties supplied by powder manufacturers. A documented preprocessing pipeline removed non--steady operating periods and outliers and ensured that learning reflected routine production. A Random Forest regressor was trained and evaluated under cross validation and was compared against representative consumption relations from the literature. The final model achieved a coefficient of determination of 0.88 on held out data. Feature importance analysis showed that steel carbon content exerted the strongest influence on mould powder consumption, followed by casting speed, the surface to volume ratio, steel titanium content and slag viscosity. These results demonstrate that lubrication demand depends on a combination of steel chemistry, geometry and kinematic conditions rather than on speed and viscosity alone. The model provides reliable forecasts of mould powder consumption and supports improved decision making during sequence planning and operational reviews. More broadly, the study illustrates how industrial data combined with interpretable machine learning can generate practical insights into the process physics of mould lubrication and can contribute to more efficient continuous casting operations by enabling earlier identification of consumption deviations and more consistent operation across grades and sections.