About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
Characterization of Minerals, Metals and Materials 2026 - In-Situ Characterization Techniques
|
| Presentation Title |
An XGBoost-SHAP based Model for Predicting End-Point Temperature in Converter Steelmaking |
| Author(s) |
Nanlv Liu, Yuchi Zou, Liwen Huang, Ling-zhi Yang |
| On-Site Speaker (Planned) |
|
| Abstract Scope |
Accurate prediction of converter endpoint temperature (EPT) is crucial for steel quality and production efficiency. This study proposes an XGBoost-SHAP framework to identify key factors and predict EPT. Trained on 1580 heats of production data, SHAP analysis identified scrap steel amount, hot metal amount and composition, blowing time, and oxygen flow rate as major EPT influencers. Feature selection based on SHAP results improved model performance, reducing RMSE by 10.3% (from 5.8°C to 5.2°C) and MAE by 9.5% (from 4.2°C to 3.8°C) compared to the model without selection. The XGBoost-SHAP model demonstrated superior performance, achieving an RMSE of 5.2°C (32% lower than BP neural network, 28% lower than SVM) and an MAE of 3.8°C (vs. 5.5°C for BP, 5.1°C for SVM). Results confirm SHAP's effectiveness in identifying significant features, and the proposed framework exhibits enhanced generalization over traditional methods. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Iron and Steel, Machine Learning, Process Technology |