| Blast Furnace(BF for short)ironmaking is the important production link of iron and steel industry and the most important way of modern ironmaking.For blast furnace ironmaking,it is necessary to accurately judge the operation situation of the whole blast furnace,timely adjust the operation system and technological parameters such as the distribution of gas in the furnace,full utilization of heat and smooth discharge of slag iron,so as to realize high quality,high yield,low consumption and long life operation of blast furnace production.However,there are complex physical and chemical reactions inside the blast furnace,high temperature and pressure,strong coupling,non-linearity,time-varying working conditions and the existence of solid,liquid and gas polymorphism,smelting environment is very bad,it is difficult to carry out realtime monitoring and effective control.At present,molten iron quality(MIQ for short)parameters at the outlet are widely used to indirectly reflect the operation state of the blast furnace,including molten iron temperature(MIT),silicon content(Si),phosphorus content(P)and sulfur content(S).However,the existing detection technology cannot directly measure the quality parameters of molten iron online,and the offline test takes a long time and lags behind.Therefore,it is particularly important to establish an accurate and reliable prediction model for the MIQ.The application of the data-driven modeling method represented by neural network in the industrial production process has achieved more and more gratifying results.However,the existing modeling of MIQ parameters still has problems in computational efficiency,robustness,online learning ability and data collinearity.To solve these problems,this thesis carries out research on robust Random Vector Functional-link Networks(RVFLNs)modeling for MIQ of blast furnace by using artificial intelligence and statistical learning methods,and conduct experiment on a 2650m3 BF of Liuzhou Iron and Steel Company in Guangxi province with the support of National Natural Science Foundations,"High-performance operation control and implement technology of large blast furnace"(61290323)and "Experimental verification platform construction and application verification of large blast furnace high-performance operation control"(61290321).The main contributions are given as follows:(1)Firstly,Considering that the existing MIQ parameter model only models a single index,and that the model lacks robustness to outliers,is susceptible to the collinearity of data,and has poor generalization ability,etc.,basic RVFLNs algorithm is improved,robust regularized RVFLNs modeling for MIQ parameter of BF is proposed.Using the regularization theory and the structural risk minimization principle,L2 norm and L1 norm are introduced into the loss function at the same time to construct the Elastic Net and sparse output weight matrix of network,solving the multicollinearity problem and preventing the model from overfitting.Combined with M-estimates,the adverse effects of outlier outliers existing in sample data on the model were solved,and the robustness and accuracy of prediction were improved.At the same time,the gaussian weight function was designed for determining the parameters independently according to the standardized residual distribution to improve the model robustness and computational efficiency.Finally,the actual data of blast furnace ironmaking process were compared and tested with different weight functions and different modeling methods.The results showed that the proposed modeling method has strong robustness,higher estimation accuracy,and sparse output weight matrix to solve the multicollinearity problem.(2)In view of the dynamic characteristics of the time-varying conditions in the blast furnace ironmaking process,on the basis of the modeling of the robust RVFLNs,the online sequential strategy is introduced,and the Robust Online Sequential Learning RVFLNs with Forgetting Factor(R-FF-OS-RVFLNs for short)modeling method of the MIQ parameters is proposed to improve the online learning and adaptive ability of the model.The online sequential strategy can update the model parameters in time according to the new batch of data.At the same time,the forgetting factor method is introduced to solve the phenomenon of "data saturation" that occurs with the increase of data,increase the effect of current data and improve the accuracy of the model.Regularization of the output weight to prevent overfitting and improve generalization ability.Finally,based on the data of numerical simulation and actual blast furnace ironmaking process,the comparison with other different modeling methods shows that the proposed modeling method has good online learning ability,and can effectively overcome the influence of the time variation of blast furnace operating condition on the robust modeling performance,thus ensuring the robustness and accuracy of the prediction of MIQ parameters.(3)Firstly,the RVFLNs modeling method based on Partial Least Squares(PLS for short)is proposed for the data noise and multicollinearity problems of blast furnace ironmaking process,and then the robust PLS-RVFLNs modeling algorithm based on Generalized Mestimates(GM-estimates)is proposed for the situation that the robust RVFLNs based on Mestimates is unable to resist the abnormal input of operating data samples in blast furnace ironmaking process.Through the residual size of the model and the distance information of the input vector in the high-dimensional space,the contribution of the sample to the model is determined respectively,so that it can resist the influence of outliers in the training data that both the input and output directions of the sample are abnormal.At the same time,PLS algorithm solved the influence of multicollinearity on the model by mapping high-dimensional data of hidden layer to low-dimensional space,which greatly improved the model’s robustness and prediction accuracy.Finally,sufficient data tests were carried out based on the numerical simulation and actual blast furnace ironmaking process data.The results show that the proposed method has stronger robustness and prediction accuracy. |
PDF Full Text Request