Hybrid Deterministic, Regression and Machine Learning Framework for Endpoint Temperature Prediction and Scrap Charge Optimization in BOF Steelmaking

Scrap charge selection has a significant influence on the thermal balance of the basic oxygen furnace (BOF) process and consequently on the final melt temperature. This paper presents a hybrid deterministic, regression, and machine learning framework for endpoint temperature prediction and steel scrap charge optimization in BOF steelmaking. The proposed methodology combines an existing deterministic BOF simulation model with regression analysis, machine learning surrogate models, and constrained nonlinear optimization. The dataset was constructed from operational records of 180 industrial BOF heats. The masses of seven scrap categories and the target endpoint temperature were obtained from these operational records, whereas the endpoint temperature used as the output for training the machine learning surrogate models was generated by the existing deterministic BOF process model. Three machine learning approaches, namely Support Vector Regression (SVR), Random Forest Regression (RF), and Gaussian Process Regression (GPR), were implemented and evaluated for endpoint temperature prediction using the masses of seven scrap categories and the target endpoint temperature as model inputs. Among the investigated surrogate models, Gaussian Process Regression achieved the best approximation performance, with a test MAE of 10.47◦C, RMSE of 16.38 ◦C, and R2 = 0.870, and was subsequently used in the optimization framework. In addition, the deterministic BOF simulation model was incorporated into a model-based optimization procedure using the same optimization objective. Both approaches were formulated as constrained optimization problems minimizing the deviation between the predicted and target endpoint temperature.