Predicting oxygen consumption in basic oxygen furnace based on metallurgical mechanism and explainable machine learning

Oxygen blowing plays a critical role in controlling the chemical composition and temperature of molten steel in a basic oxygen furnace (BOF). However, most studies have prioritized improving the prediction accuracy of oxygen consumption models, with limited attention given to model interpretability. In this study, a hybrid model integrating the metallurgical mechanism, machine learning (ML), and Shapley additive explanation (SHAP) analysis was established to predict the oxygen consumption in BOF. The isolation forest algorithm, Pearson correlation coefficient, maximal information coefficient, t-distributed stochastic neighbor embedding method, oxygen balance mechanism (OBM), ML method, and Bayesian optimization (BO), along with various performance metrics and SHAP analysis, were employed to eliminate outliers from the original production dataset, analyze inter-variable correlations, visualize high-dimensional data, develop a single metallurgical mechanism model, establish a single ML model, optimize the hyperparameters of the ML model, assess the model performance, and interpret the optimal model. The results showed that the BO-OBM-CatBoost model achieved the best overall performance, with coefficient of determination (R²), root-mean-square error (RMSE), and mean absolute error (MAE) of 0.7041, 1.24 m³/t, and 1.00 m³/t, respectively. Meanwhile, the prediction hit ratio for the oxygen consumption per ton of molten steel within the error range of −3, 3 was 99.06%. Finally, SHAP analysis was applied to the optimal BO-OBM-CatBoost model to reveal the influence of different input variables on the prediction results from both the global and local perspectives. This study contributes to the accurate prediction of oxygen consumption in BOF and supports the optimization of smelting processes using interpretability analysis.