摘要
在钢包炉(LF)精炼过程中,准确预测合金元素收得率对于控制钢水成分、提高合金利用率及降低冶炼成本具有重要意义。近年来机器学习方法被广泛应用于冶金过程建模,但多数机器学习模型在实际应用中通常依赖复杂的超参数调优过程,且引入新数据后往往需要重新调优超参数,建模效率有待提高。针对上述问题,首先,结合LF精炼实际生产数据,构建了基于表格先验数据拟合网络(TabPFN)的Si元素收得率预测模型;然后,利用多种模型评价指标,将TabPFN模型与已有研究的参考炉次法、多元线性回归模型以及多种机器学习模型进行了对比分析;最后,融合沙普利加性解释(SHAP)方法对TabPFN模型进行了全局与局部层面的解释分析。结果表明,TabPFN模型在无需大量超参数调优的情况下,在拟合优度(R^(2))、平均绝对误差(E_(MA))、均方根误差(E_(RMS))、命中率和模型推理时间等关键性能指标上均优于已有模型,各项指标分别达到了0.83、1.59、2.03、98.4%和0.430 s。同时,融合SHAP分析从全局层面揭示了各输入特征变量对Si元素收得率的影响大小,从局部层面量化了各输入特征变量对Si元素收得率预测值的影响程度,实现了LF精炼合金元素收得率的高效、高精度和可解释性预测,为钢铁工业在智能制造背景下的冶金过程建模提供了新的研究思路与技术路径。
In ladle furnace(LF)refining process,accurately predicting alloying element yield is of great importance for controlling the chemical composition of molten steel,improving alloy utilization efficiency,and reducing smelting costs.In recent years,machine learning methods have been widely applied to metallurgical process modeling.However,most machine learning models typically rely on complex hyperparameter tuning and often require re-tuning hyperparameter when new data are introduced,limiting modeling efficiency.To address these challenges,a tabular prior-data fitted network(TabPFN)-based prediction model for Si element yield was established using actual production data.And this model′s performance was evaluated using multiple metrics and compared with the reference heat method,multiple linear regression model,and various machine learning models reported in previous studies.Shapley additive explanations(SHAP)was then employed to conduct both global and local interpretability analysis.The results show that the TabPFN model is superior to the existing models in key performance indicators such as R^(2),E_(MA),E_(RMS),hit rate and model reasoning time without a lot of hyperparameter tuning.The indicators reach 0.83,1.59,2.03,98.4%and 0.430 s,respectively.Meanwhile,the SHAP analysis reveals the influence of each input feature on the Si element yield at the global level,and quantifies the influence of each input feature on the predicted Si element yield at the local level,so as to realize the efficient,high-precision and interpretable prediction of alloying element yield,offering new research ideas and technical pathways for metallurgical process modeling in the context of intelligent manufacturing in the steel industry.
作者
信自成
张江山
张军国
刘青
XIN Zicheng;ZHANG Jiangshan;ZHANG Junguo;LIU Qing(State Key Laboratory of Advanced Metallurgy,University of Science and Technology Beijing,Beijing 100083,China;School of Automation and Electrical Engineering,University of Science and Technology Beijing,Beijing 100083,China;Tangshan Branch,Hebei Iron and Steel Co.,Ltd.,Tangshan 063000,Hebei,China)
出处
《中国冶金》
北大核心
2025年第11期178-186,共9页
China Metallurgy
基金
国家自然科学基金面上项目(52374321)
国家重点研发计划资助项目(2024YFB3713602)。
