Multimodel Predictions on Converter Raw Material Addition Under GRNN Optimization: A Comparative Study

作者：Wang, Jianhao; Fang, Qing^*; Zhu, Wanjun; Yang, Tengfei; Wang, Jiahui; Zhang, Hua; Ni, Hongwei^*

来源：Metallurgical and Materials Transactions B: Process Metallurgy and Materials Processing Science , 2024.

DOI：10.1007/s11663-024-03031-3

• MULTIOBJECTIVE COLLABORATIVE OPTIMIZATION
• METALLURGICAL PROPERTIES
• NEURAL-NETWORK
• STEELMAKING
• MODEL
• STEEL
• IRON
• TEMPERATURE

Due to the complexity of the steelmaking process, establishing predictive models for raw material addition has been a challenging problem. While statistical models and artificial neural networks have been extensively researched for simulating the intricate steelmaking process, it is still unclear which method is most suitable for predicting the addition of raw materials. In this paper, various machine learning tools, including backpropagation (BP) neural networks, radial basis functions (RBFs), generalized regression neural networks (GRNNs), random forests (RFs), and support vector regression (SVR), were initially compared to validate their applicability in predicting the addition of raw materials in converters. During the predictive experiments, it was observed that different predictive models exhibited significant errors for the same sample points. This result could be due to measurement errors causing these sample points to deviate from the actual predictive logic, thus affecting the overall predictive results of the models. Therefore, a multimodel based on generalized regression neural networks (GRNNs) was proposed to optimize the initial prediction data. By utilizing GRNN technology to optimize the initial samples, sample points with anomalous data were filtered out, thereby improving the accuracy and robustness of the multimodel predictions. Through the application of multimodel predictions for predicting the addition of lime and light-burned dolomite, it was demonstrated that the predictive performance for anomalous samples improved significantly after optimization. According to the comparative analysis of the multimodel predictions, the random forest (RF) model performed the best, followed by the SVR, BP and RBF methods. This result indicates that the predictive accuracy provided by the RF algorithm meets the actual production control requirements of the converter steelmaking process.