Interpretable stacking ensemble learning for predicting the BOF oxygen blowing amount for making high-carbon steel with state-of-the-art machine learning and deep learning models

Abstract

Eight state-of-the-art machine learning and deep learning models designed for tabular data were developed to predict the basic oxygen furnace oxygen blowing amount for making high carbon steel. These models include extreme gradient boosting (XGboost), light gradient boosting machine (LightGBM), categorical boosting (CatBoost), tabular attention network (TabNet), neural oblivious decision ensembles (NODE), gated additive tree ensemble (GATE), gated adaptive network for deep automated learning of features (GANDALF), and deep abstract networks (DANets). The data with 20 tabular factors (information about hot metal, scrap, additives and endpoint values) were collected from 8300 heats. After tuning with 5-fold validation, in the test set, the LightGBM demonstrated the optimal prediction accuracy with R² = 0.796, mean absolute error (MAE) = 71.873 m³, and rooted mean square error (RMSE) = 93.013, and mean relative error (MRE) = 1.412%. The overall and individual differences in performance among the various models were analysed using the Friedman test and the post-hoc Nemenyi Test. Then, the model stack, comprising the top three performing models, along with linear regression as the meta-learner, jointly implemented the stacking ensemble learning. The MAE and MRE were improved to 71.863 m³ and 1.411%, respectively. In addition, SHapley Additive exPlanations (SHAPs) were implemented to quantify the exact impact of each factor by interpreting the LightGBM model. The results revealed the differences in the oxygen blowing amounts caused by each factor across various heats. Additionally, an oxygen blowing amount prediction software was developed based on the LightGBM and the SHAP analysis model.