A hybrid approach combining data-driven and mathematical models to predict the endpoint carbon content and temperature in BOF

Accurate endpoint prediction is critical in the basic oxygen furnace (BOF) steelmaking process. This paper proposes a hybrid approach that integrates data-driven methods with mathematical models to predict the endpoint carbon content and temperature in BOF. The proposed model is based on the Just-In-Time Learning (JITL) framework, which uses similarity functions to select historical heats that closely resemble the smelting conditions of the target sample, thereby constructing local models. The Trust Region Reflective (TRR) algorithm and Multiple Linear Regression (MLR) algorithm are independently used to calculate the decarburization parameters and heating parameters for these local models. These parameters are then applied in mathematical equations to predict the carbon content and temperature during the second blowing process of BOF. By integrating theoretical analysis with the Random Forest (RF) algorithm, it was determined that the key variables influencing both the decarburization and heating parameters were the temperature and carbon content measured by the TSC sub-lance, along with the slag weight. The study validated the JITL model with silicon steel, achieving an 85.83% hit rate for carbon content prediction (±0.02 wt%) and 83.42% for temperature prediction (±15 °C), demonstrating superior accuracy over the traditional sub-lance model.