Development of molten steel end-point temperature prediction model in electric arc furnace steelmaking process using data mining methods

Abstract

In the electric arc furnace (EAF) steelmaking process, ​accurately predicting​ the end-point temperature of molten steel (EPT-MS) ​is​ critically beneficial for ensuring efficient and stable production. ​This study proposes​ a comprehensive data mining strategy for EPT-MS prediction ​that systematically integrates anomaly detection, feature selection, and machine learning algorithms. To validate this approach, a practical dataset ​consisting of​ 1235 heats ​with​ 32 process features was utilised. Four supervised learning algorithms including logistic regression, k-nearest neighbours, decision tree, and extreme gradient boosting (XGBoost) were employed to develop prediction models. The prediction performance was rigorously evaluated by considering multiple factors, including​ data partitioning (with train/test ratios varying from 9:1 to 1:9), anomaly detection methods (auto-encoder (AE) and principal component analysis), feature selection techniques (permutation importance analysis), and algorithm performance. The results demonstrated that prediction accuracy is influenced by data quality, feature selection, and algorithm choice. Specifically, AE-based anomaly detection significantly improved data quality and enhanced prediction accuracy, while permutation importance analysis effectively reduced model complexity without sacrificing performance. Among all algorithms, XGBoost delivered the best performance, achieving hit rates exceeding 40%, 80%, and 95% within error ranges of $\pm 5\circ \mathrm{C},\pm 10\circ \mathrm{C}$ , and $\pm 15\circ \mathrm{C}$ , respectively.