基于GA-BP神经网络的铁水硅含量预测方法  

Prediction method of hot metal silicon content based on improved GA-BP neural network

作　　者：何奕波 郭辉 张冰谦 朱强 汤海明 李怡宏 HE Yibo;GUO Hui;ZHANG Bingqian;ZHU Qiang;TANG Haiming;LI Yihong(School of Materials Science and Engineering,Taiyuan University of Science and Technology,Taiyuan 030024,Shanxi,China;Australia Institute for Innovative Materials,University of Wollongong,Wollongong NSW2522,New South Wales,Australia;Shanxi Taigang Stainless Steel Co.,Ltd.,Taiyuan 030003,Shanxi,China)

机构地区：[1]太原科技大学材料科学与工程学院,山西太原030024 [2]澳大利亚伍伦贡大学创新材料研究所,澳大利亚新南威尔士州伍伦贡市NSW2522 [3]山西太钢不锈钢股份有限公司,山西太原030003

出　　处：《钢铁研究学报》2024年第3期309-318,共10页Journal of Iron and Steel Research

基　　金：中央引导地方科技发展资金资助项目(YDZJSX2022C028);山西省基础研究计划面上资助项目(20210302123218,202203021211187);山西省高等学校大学生创新创业训练计划资助项目(202210109006)。

摘　　要：长期以来在高炉炼铁过程中铁水硅含量一直作为代表高炉热状态的重要指数。然而由于高炉具有动态特性,内部化学反应十分复杂,是一个典型的黑箱模型,因此对高炉铁水硅含量进行实时预测十分困难。针对这一问题,利用遗传算法(Genetic Algorithm, GA)对传统BP(Back Propagation)算法进行改进,构建GA-BP神经网络预测铁水硅含量。首先将高炉炼铁过程中的13个参数(如风量、风压等)进行特征提取,并利用遗传算法全局搜索BP神经网络最优的初始权值和阈值,接着利用前向传播算法(Forward Propagation, FP)在三层神经网络中传递筛选出来的特征并计算出预测值,其中三层神经网络每层神经元个数分别为7、50、1。最终将铁水硅含量预测值与真实值进行误差分析,利用梯度下降(Gradient Descent, GD)的原理不断更新神经元的权重,直到预测值与真实值之间的误差达到所给定的阈值。相比于传统BP神经网络,GA-BP神经网络改善了BP神经网络权值、阈值难定,学习速度慢且易陷入局部最优等缺点。将某钢厂生产过程中实时采集到的数据经过预处理之后,输入到神经网络中进行训练并且利用测试集来验证该模型的精度。最终该模型在测试集上取得了92%的正确率且均方误差(Mean Square Error, MSE)稳定在0.001,证明了该模型的有效性。选取了50组数据集之外的新数据来进行预测,结果验证了该模型具备指导生产实践的能力。The silicon content in molten iron during the ironmaking process in a blast furnace has been an important indicator of the furnace's thermal state for a long time.However,predicting the silicon content in real-time is extremely difficult due to the dynamic nature of the blast furnace and the complex internal chemical reactions that occur within it.To address this issue,a GA-BP neural network for predicting the silicon content in molten iron by improving the traditional BP algorithm using genetic algorithms(GA)was proposed.Firstly,feature extraction is performed on 13 parameters(such as air volume,air pressure,etc.)during the blast furnace ironmaking process.A genetic algorithm is used to globally search for the optimal initial weights and thresholds of the BP neural network.Then,the forward propagation algorithm(FP)is used to transmit the selected features in the three-layer neural network and calculate the predicted values.The number of neurons in each layer of the three-layer neural network is 7,50,and 1,respectively.Finally,an error analysis was conducted between the predicted silicon content in molten iron and that of the actual value,using the principle of Gradient Descent(GD)to continuously update the weights of neural network until the error between the predicted value and the actual value reached the given threshold.Compared to traditional BP neural networks,GA-BP neural networks improve the shortcomings of BP neural networks,such as difficult to determine weights and thresholds,slow learning speed,and easy to fall into local optima.After preprocessing the real-time data collected during the production process of a certain steel plant,it is input into a neural network for training and the accuracy of the model is verified using a test set.In the end,the model achieved an accuracy of 92%on the test set and a stable Mean Square Error(MSE)of 0.001,proving the effectiveness of the model.New data outside of 50 datasets for prediction were selected,and the results verified that the model has the ability to guid

关 键 词：高炉 炼铁 硅含量 BP神经网络 遗传算法 特征工程 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TF325.69[自动化与计算机技术—控制科学与工程]