Scholarlink

<p>To address the delayed availability of sinter quality measurements and the limited accuracy and generalization ability of conventional prediction models, a sinter quality prediction method integrating thermodynamic calculations and ensemble learning is proposed. A total of 196 sintering-cup experimental datasets were used, including 156 datasets for model training and 40 datasets for testing. First, based on the melt-zone chemical composition of each experimental dataset, thermodynamic calculations are performed using FactSage to obtain characteristic parameters describing the high-temperature mineralization behavior during sintering. Next, a targeted data preprocessing scheme and a two-stage feature-selection strategy are employed to improve feature quality and reduce input dimensionality. A CatBoost-based model is then developed for sinter quality prediction, and its hyperparameters are optimized using the differential evolution algorithm. By incorporating thermodynamic mineralization parameters, the proposed model is evaluated against several representative ensemble learning methods. The results show that differential-evolution-based hyperparameter optimization significantly improves model performance. On the test set, the optimized model achieves <i>R</i><sup>2</sup> values of 0.9314 and 0.9553, MSE values of 0.0167 and 0.0187, and MAE values of 0.0779 and 0.0883 for the tumbler index and sieve index, respectively. Within a tolerance of ±&#xa0;0.2 percentage points, the hit rates reach 87.5 pct for the tumbler index and 90.0 pct for the sieve index. In addition, feature importance analysis provides further insight into the influence of high-temperature mineralization characteristics on sinter quality. The proposed method provides a physics-informed framework for sinter quality prediction and has potential engineering value for intelligent sintering production.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Prediction of Sinter Quality Indices Using Thermodynamic Calculations and Ensemble Learning

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Jiabao Wang,
  • Xiaohui Fan,
  • Xiaoxian Huang,
  • Xuling Chen,
  • Zhenxiang Feng,
  • Min Gan,
  • Shaochang Nie,
  • Jingjing Yang

摘要

To address the delayed availability of sinter quality measurements and the limited accuracy and generalization ability of conventional prediction models, a sinter quality prediction method integrating thermodynamic calculations and ensemble learning is proposed. A total of 196 sintering-cup experimental datasets were used, including 156 datasets for model training and 40 datasets for testing. First, based on the melt-zone chemical composition of each experimental dataset, thermodynamic calculations are performed using FactSage to obtain characteristic parameters describing the high-temperature mineralization behavior during sintering. Next, a targeted data preprocessing scheme and a two-stage feature-selection strategy are employed to improve feature quality and reduce input dimensionality. A CatBoost-based model is then developed for sinter quality prediction, and its hyperparameters are optimized using the differential evolution algorithm. By incorporating thermodynamic mineralization parameters, the proposed model is evaluated against several representative ensemble learning methods. The results show that differential-evolution-based hyperparameter optimization significantly improves model performance. On the test set, the optimized model achieves R2 values of 0.9314 and 0.9553, MSE values of 0.0167 and 0.0187, and MAE values of 0.0779 and 0.0883 for the tumbler index and sieve index, respectively. Within a tolerance of ± 0.2 percentage points, the hit rates reach 87.5 pct for the tumbler index and 90.0 pct for the sieve index. In addition, feature importance analysis provides further insight into the influence of high-temperature mineralization characteristics on sinter quality. The proposed method provides a physics-informed framework for sinter quality prediction and has potential engineering value for intelligent sintering production.