Prediction Model of Phosphorus Content at the End Point of Electroslag Remelting Based on MI and XGBoost Algorithms

Abstract

This study proposes a phosphorus content prediction model for the endpoint of electroslag remelting （ESR） refining process based on Mutual Information （MI） method and XGBoost. The MI method is utilized for feature selection and assessment of factors affecting the endpoint phosphorus content. The dataset after feature selection serves as the input variables for the model.The MI-XGBoost model is trained and validated using production data. Grid search cross-validation is employed for model structure adjustment and hyperparameter optimization. It is compared horizontally with MI-RR， MI-RF， MI-GBDT， and MI-KNN models. The results demonstrate that the MI-XGBoost model exhibits the highest prediction accuracy. The incorporation of MI and GridSearchCV enhances the model's predictive performance and fitting ability.Validation of the test set shows that the MI-XGBoost model achieves R2， Mean Absolute Error， Explained Variance Score， and Maximum Error values of 0.889 4， 0.000 4， 0.897 2， and 0.004 1， respectively， all superior to MI-RR， MI-RF， MI-GBDT， and MI-KNN models. The MI-XGBoost model effectively predicts the endpoint phosphorus content， providing valuable reference for endpoint control and determination in the ESR refining process. It presents a new perspective for realizing the intelligence of the ESR refining process.

Keywords

• electroslag remelting； mutual information method； xgboost algorithm； phosphorus content； machine learning