基于机器学习的连铸保护渣黏度与熔化温度预报

Prediction of viscosity and melting temperature for mold fluxes based on machine learning algorithms

连铸保护渣在保障连铸顺行和铸坯质量方面发挥着关键作用,其作用的发挥与保护渣的性能密切相关。保护渣的黏度和熔化温度是保护渣的关键性能参数,其测定所需的时间与物力成本较高。鉴于机器学习技术的发展,本研究基于K近邻算法(KNN)、核岭回归算法(KRR)与随机森林算法(RF)3种机器学习算法建立保护渣黏度和熔化温度预报模型,旨在准确预报保护渣的黏度与熔化温度,为保护渣的快速、便捷设计提供指导。结果表明,基于KRR算法的模型在保护渣黏度预报方面有着较好的表现,其决定系数(R^(2))、均方根误差(RMSE)和平均绝对误差(MAE)分别为0.983、0.023和0.014;基于RF算法的模型在预报保护渣熔化温度方面具有更高的可靠性,R^(2)、RMSE和MAE分别为0.823、14.004和8.974。输入特征的排列重要性分析表明SiO_(2)、F、Al_(2)O_(3)对黏度的影响依次降低,Na_(2)O对熔化温度的影响最大。与广泛应用的黏度和熔化温度预报模型相比,KRR算法黏度预报模型和RF算法熔化温度预报模型的平均相对误差(MRE)分别为6.26%和0.83%,远低于目前广泛应用的预报模型的平均相对误差,表明机器学习模型具有较高的可靠性。基于模型建立的黏度和熔化温度分布图可为保护渣的设计提供直观参考。

Mold fluxes play important role in ensuring the smooth casting and slab quality,which is closely related to the properties of mold fluxes.The viscosity and melting temperature of mold fluxes are key parameters in the design and production of mold fluxes;however,the determination of property is time-consuming and labor-intensive.In view of the development of machine learning technology,three machine learning algorithms including KNN,KRR and RF were employed to establish the prediction model to accurately predicting viscosity and melting temperature of mold fluxes,providing a guide to the fast and convenient design of mold fluxes.The experimental results show that compared to other two models,KRR-based model has the best performance in the viscosity prediction,in which the coefficient of determination(R^(2))is 0.983,the root mean squared error(RMSE)is 0.023,and the mean absolute error(MAE)is 0.014.The RF-based model is more reliable in predicting the melting temperature with R^(2)of 0.823,RMSE of 14.004,and MAE of 8.974.The ranked importance analysis of the input features shows that the order of influence of viscosity is SiO_(2),F,and Al_(2)O_(3)and Na_(2)O has the greatest influence on melting temperature.Compared with the widely used viscosity and melting temperature prediction models,the mean relative errors(MRE)of the KRR-based viscosity prediction model and the RF-based melting temperature prediction model are 6.26%and 0.83%,respectively,which are much lower than MRE of the widely used prediction models,indicating that the machine learning-based models have a high reliability.Moreover,the viscosity and melting temperature distribution maps established based on the models could provide intuitive reference for the design of mold fluxes.