Blast Furnace Molten Iron Quality Parameters Modeling Methods Based On Modified Random Vector Functional-link Networks

Blast furnace ironmaking is the most important section of the entire steel manufacturing process.In the production process,iron ore,coke and fluxes in a certain mixture ratio are charged from the top of the BF.At high temperature and high pressure,iron ores are reduced to molten iron after a series of complex reactions,which is discharged from the taphole.The molten iron quality(MIQ)has a big influence on the following converter steelmaking processing,thus it is essential to accurately learn the parameters of the of MIQ so as to ensure the smooth operating of the BF to produce the qualified molten iron.However,the BF ironmaking process is extremely complex including gas,solid and liquid transaction.The physical and chemical reactions are very complicated and inner conditions are extremely complex,which results that it is impossible for the operators to monitor the changes of its internal operation condition on real time.Thus,it is necessary to establish reliable mathematical models of MIQ to describe the current and designed internal temperature and parameters that can offer comprehensive information of BF and molten iron for operators.Currently,Random Vector Functional-link Networks(RVFLNs)is widely used to the modeling,regression and classification problems in practical engineering and also used to solve the modeling of MIQ.Compared with conventional BP,RBF neural network,it has faster speed and higher computational efficiency.However,RVFLNs has some shortcomings:instability in the results;poor generalization ability;convergence and it is difficult to choose the optimal number of hidden layer nodes.Considering these problems,this paper carry out research on multivariate molten iron parameters quality modeling based on modified RVFLNs and conduce experiment on No.2 BF of Liuzhou Iron and Steel company with the support of the National Science Foundation(61290323).The main contributions are given as following:(1)The traditional statistic modeling method and single content prediction of MIQ can not reflect the complicated and dynamic ironmaking process.This paper introduced Nonlinear AutoRegressive with exogenous inputs(NARX)model to the RVFLNs and used the improved RVFLNs to predict MIQ.The first step is to determine the 16 variables that influence the MIQ by analyzing the state variables and control variables and introduced principle component analysis(PCA)to pick a few key factors as the input variables of model to reduce the complexity of computing.In order to describe the dynamic of the production process that can better reflect the practical ironmaking process,the time delays of the relevant input and output variables were considered in the model and the optimal orders of the NARX is determined based on the AIC(Akaike’s Information Criterion)and SR(Schwartz-Rissanen).Finally,compared experiments were employed by using the improved RVFLNs and traditional RVFLNs and results demonstrated that the improved RVFLNs produced better estimating accuracy.(2)Aiming at the problems of overfitting and poor generalization capability for current RVFLNs,this paper proposed an improved RVFLNs:AE-P-RVFLNs based on the Autoencoder and principal component analysis(PCA)and established the NARX model for MIQ.The first step is to determine the 16 variables that influence the MIQ by analyzing the state variables and control variables and introduced principle component analysis(PCA)to pick a few key factors as the input variables of model to reduce the complexity of computing.In order to describe the dynamic of the production process that can better reflect the practical ironmaking process,NARX model were introduced and the optimal orders of the NARX is determined based on the AIC(Akaike’s Information Criterion)and SR(Schwartz-Rissanen).Then in order to find the useful information from the complex real industry data and hidden relationship of the input data,Autoencoder is introduced to train the input data and the calculated output weight is treated as the input weight of a fresh network.Then,the PCA is used to reduce the dimension of hidden layer output matrix so as to avoid the multicollinear problem in calculation and the number of hidden nodes is also greatly reduced,which simplify the network structure and avoid the overfitting problem cause by too many hidden nodes.Finally,actual industrial experiments and contrastive researches have been conducted on the 2#BF in Liuzhou Iron&Steel Group Co.of China using the proposed method.The results shows that the proposed method improves the calculation accuracy and calculation speed.What’s more,it effectively solves the overfitting problem of the conventional RVFLNs.(3)Aiming at the overfitting problem caused by improper number of hidden nodes and how to determine the optimal number of hidden nodes,this paper proposes a data-driven multivariable dynamic NARX modeling method to estimate the MIQ parameters using an improved incremental random vector functional-link networks(I-RVFLNs).First,the new modeling method improves the terminal condition for the multiple input and multiple output(MIMO)dynamic system,which suits the practical BF ironmaking process very much.Finally.Then,since the traditional I-RVFLNs has too many hidden nodes when meeting the termination conditions,the improved I-RVFLNs can ensures the errors decreases monotonically without adding any hidden nodes by sequentially changing the output weight of the fixed hidden nodes to approximate ideal output.Finally,comparative tests of the improved I-RVFLNs based NARX dynamic prediction model and other well-known modeling methods have been experiment on 2#BF in Liuzhou Iron&Steel company.Compared with other models,the new method has accuracy and speed.