Application of XGBoost and kernel principal component analysis to forecast oxygen content in ESR

Abstract

A model combining kernel principal component analysis (KPCA) and Xtreme Gradient Boosting (XGBoost) was introduced for forecasting the final oxygen content of electroslag remelting. KPCA was employed to reduce the dimensionality of the factors influencing the endpoint oxygen content and to eliminate any existing correlations among these factors. The resulting principal components were then utilized as input variables for the XGBoost prediction model. The KPCA-XGBoost model was trained and proven using data obtained from companies. The model structure was adapted, and hyperparameters were optimized using grid search cross-validation. The model performance of the KPCA-XGBoost model is compared with five machine learning models, including the support vector regression model. The findings demonstrated that the KPCA-XGBoost model exhibited the highest level of prediction accuracy, indicating that the incorporation of KPCA significantly enhanced the regression prediction performance of the model. The accuracy of the KPCA-XGBoost model was 82.4%, 97.1%, and 100% at errors of ± 1.5 × 10⁻⁶, ± 2.0 × 10⁻⁶, and ± 3 × 10⁻⁶ for oxygen content, respectively.