摘要
高炉出铁口处的铁水温度为表征铁水质量与炉热状态的核心工艺参数之一,其准确预测对冶炼过程控制具有重要意义。然而,高炉冶炼具有非线性、高动态、强时滞及时空多尺度耦合特征,导致基于传统冶金机理的建模方法难以进行实时温度预测与炉况诊断。针对此问题,构建了一种基于异构大数据的CNN-GRU-Attention深度学习预测模型,该模型将卷积神经网络(CNN)的局部特征提取能力、门控循环单元(GRU)的时序动态建模优势以及自适应注意力机制(Self-Attention)相结合,实现对冶炼过程多维工艺操作参数与铁水温度复杂关联的深度映射。重点探讨了基于冶金机理的数据预处理方法和网络架构优化对模型性能的增强机制,确立了由CNN-GRU协同特征提取层与自适应注意力权重分配层构成的优选架构。研究结果表明,经冶金机理指导的数据清洗方法可有效提升数据质量,优化后的模型在测试集上实现了±5℃范围内的铁水温度预测准确率为86%;工业应用阶段,模型对连续铁水温度的预测值与实际值的偏差在±10℃范围内的命中率可达88%。构建的模型显著提升了高炉炼铁过程的数字化表征能力,建立的预测系统已通过工业现场验证,具备良好的工程应用价值与推广前景。
The hot metal temperature at the blast furnace taphole is a core process parameter that characterizes both the quality of the iron product and the thermal state of the furnace.Its accurate prediction is of great significance for process control in ironmaking.However,due to the nonlinear,highly dynamic,strongly time-delayed,and spatiotemporally multi-scale coupled nature of the blast furnace smelting process,traditional mechanism-based metallurgical modeling methods are inadequate for real-time temperature prediction and furnace condition diagnosis.To address this issue,a CNN-GRU-Attention deep learning prediction model based on heterogeneous big data was constructed.This model integrates the local feature extraction capability of convolutional neural networks(CNN),the advantage of gated recurrent units(GRU)in temporal dynamic modeling,and the adaptive self-attention mechanism to achieve deep mapping of the complex relationships between multidimensional process parameters and hot metal temperature.Emphasis was placed on discussing data preprocessing methods based on metallurgical mechanisms and network architecture optimization enhance model performance,leading to the establishment of an optimized architecture composed of a CNN-GRU cooperative feature extraction layer and an adaptive attention weight allocation layer.The results show that data cleaning guided by metallurgical mechanisms significantly improves data quality.The optimized model achieved an accuracy of 86%within a±5℃error range for hot metal temperature prediction on the test set.During industrial application,the model attained a hit rate of 88%within a±10℃range between predicted and actual values of continuous hot metal temperature.The developed model markedly enhances the digital representation capability of the blast furnace ironmaking process.The prediction system has been validated in industrial field trials,demonstrating considerable engineering application value and promising potential for broad adoption.
作者
韩明博
王浩男
郑华伟
刘栋梁
李强
HAN Mingbo;WANG Haonan;ZHENG Huawei;LIU Dongliang;LI Qiang(School of Metallurgy,Northeastern University,Shenyang 110819,Liaoning,China;Ironmaking Plant,Wuhan Iron and Steel Co.,Ltd.,Wuhan 430080,Hubei,China;R&D Center of Wuhan Iron and Steel Co.,Ltd.,Baosteel Central Research Institute,Wuhan 430080,Hubei,China)
出处
《中国冶金》
北大核心
2025年第11期168-177,共10页
China Metallurgy
基金
国家自然科学基金资助项目(52274328,52074079)
中国宝武低碳冶金创新基金资助项目(BWLCF202403)。
关键词
循环神经网络模型
铁水温度预测
高炉
大数据
深度学习
recurrent neural network model
hot metal temperature prediction
blast furnace
big data
deep learning
