Blast Furnace Molten Iron Quality Parameters Modeling Methods Based On Modified Neural Networks

Iron and steel is an important industrial raw material in the modern society,and blast furnace is an important production process of iron smelting.Blast furnace iron smelting is a complex physical,chemical reaction process where liquid iron is transferred from iron ore and other solid state iron compounds through high temperature and high pressure.During the smelting process in blast furnace,there occurs complex gas solid,solid-solid,solid-liquid multiphase reaction,accompanied with high temperature,high pressure,phase coupling,multi physical coexistence,chemical reactions and transport phenomena,with strong nonlinearity,multi variable coupled,time-varying,large time-delay complex dynamic characteristics.And because of the long time in the off-line laboratory testing,it is difficult to detect the quality parameters of hot metal namely temperature,silicon content([Si]),phosphorus content([P]),sulfur content([S])on line.This makes the quality operation of molten iron in the blast furnace is still relying on the manual experience of operators,thus restricting the development of operational control of blast furnace.Therefore,it is necessary to do the research about multivariate modeling of molten iron quality parameters,and achieve the online soft measurement of temperature,[Si],[P]and[S].It is also the key to realize operational control and optimization of blast furnace molten iron quality parameters.The general problem of the existing method for the modeling of molten iron quality parameters is the static modeling caused by single parameter modeling.Because the iron making is a complex process with multivariable,large time delay characteristics,the static modeling of single element is obviously difficult to meet the actual demand of operation and control.On the other hand,although good system identification and nonlinear approximation ability have been achieved by existed neural,there are some disadvantages like slow convergence speed,easy convergence to the local minimum points,the network parameters are not optimally determined,fitting,generalization ability is not strong,poor practicability etc..Considering these problems,this paper carry out research on multivariate molten iron parameters quality modeling based on modified neural network and conduct experiment on No.2 BF of Liuzhou Iron and Steel Company with the support of the National Natural Science Foundation(61290323).The main contributions are given as follows:1)Considering the static mapping problem and network parameters cannot be optimally determined by the traditional artificial neural network,we propose an artificial neural network modeling method which combines the advantages of adaptive genetic algorithm(AGA)parameter.By using AGA,the parameters of network are optimized and confirmed,and it effectively improve the algorithm and avoid the algorithm easily fall into local minima problems;in addition,by the introduction of principal component analysis(PCA),16 input variables of parameters that influence the quality of molten iron are conducted the dimension reduction,which greatly improves the accuracy and efficiency of modeling;at the same time,considering the dynamic of blast furnace,the self-feedback structure are built to establish a dynamic modeling with storage and having the processing ability to solve data in different time,which overcomes the limitations of the traditional neural network.In the end,the effectiveness of the proposed method are given by the practical industrial experiments.2)Since the gradient batch processing modeling method like BP neural network has disadvantages like slow learning speed,easily falling into local minimum value and cannot handle large amounts of data,a RVFLNs method with strong nonlinear function generalization ability and fast learning speed are combined with online sequential(OS)algorithm to establish an OS-RVFLNs dynamic soft measurement modeling method for molten iron quality parameters.Firstly,the single variable characteristic of the traditional random weight neural network is improved to multivariable version and OS-RVFLNs(M-OS-RVFLNs)algorithm is proposed.Secondly,the introduction of online sequential method has greatly enhanced the ability of RVFLNs network to deal with a large amount of data and the use of new data.And the advantages like good nonlinear generalization ability and fast learning rate of RVFLNs algorithm ensures the feasibility of blast furnace online soft measurement modeling,and makes a foundation of subsequent implementation of operational feedback control;in addition,in the practical application on No.2 BF,a method to the random parameters and the optimal hidden layer node in practical environment is proposed and the accuracy of the model is optimized.3)Considering the optimal number in the hidden layer of conventional RVFLNs and M-OS-RVFLNs is difficult to determine and the overfitting problems are not solved,a dynamic modeling method based on incremental RVFLNs for molten iron quality parameters is proposed.Firstly,in view of the single variable characteristics of traditional incremental RVFLNs,the multi input and multi output expansion is extended,and a multivariate incremental random vector functional-link neural networks(M-I-RVFLNs)is proposed.Secondly,the conventional incrementalalgorithm’s terminal conditions are modified to make it more conform to the practical engineering problems of blast furnace molten iron quality modeling.Compared with traditional method and sequential type random weight of neural network,the proposed M-I-RVFLNs solved the difficult problem of the optimal number of hidden nodes and overcame the fitting problem.Finally,industrial test and comparative study based on the proposed method are made on No.2 blast furnace of Liuzhou Iron and Steel Company.The results show that the measuring accuracy and universality of proposed method is better than the conventional method for measuring the quality of molten iron.