Regression And Classification Of Hot Metal Silicon Content In Blast Furnace Based On Elman-Adaboost Model

During the blast furnace ironmaking process,it is of vital importance that the reduction of silicon be controlled strictly and the variation of silicon content be mastered completely.According to the nonlinear and dynamic characteristics of the blast furnace,this paper had some deep research on the regression and classification prediction of silicon content with the Elman neural network and Adaboost algorithm as the main black-box modeling method,and then merged the results of regression and classification prediction by information fusion.Firstly,the regression prediction for silicon content was researched.In this section,Multi-variable Elman-Adaboost Strong Predictor(MEASP)and Single-variable Elman-Adaboost Strong Predictor(SEASP)were studied respectively.For MEASP,the average regression hitting rate reached 94.80%.And for SEASP,the value was90.21%.In conclusion,the reasons why the average hitting rate of MEASP was higher than SEASP were analyzed and revealed.Secondly,the classification prediction for the variation direction of silicon content was researched.In this section,Multi-variable Elman-Adaboost Strong Classifier(MEASC)and Single-variable Elman-Adaboost Strong Classifier(SEASC)were studied respectively.For Standard MEASC,the average classification hitting rate was 69.80%,thus the Improved MEASC was introduced.The Improved MEASC was modified from three aspects:the training model,the input training set and the training parameters during training process.It turned out that the average classification hitting rate was 89.19%,which was much more satifactory.For Standard SEASC,the average classification hitting rate was 71.70%.After modification,the average misclassified rate reached 85.32%.In summary,the analyzation and comparison between MEASC and SEASC were given in detail,as well.Thirdly,the combination of regression and classification prediction results by means of information fusion to improve the practical value of two prediction methods was discussed.The fuzzy logic inference was chosen as the fusion method.The membership degree was set by hitting rate.Through Mamdami fuzzy implications,it became the final membership degree which was the basis of following inference.Then the input of fuzzy control rules table could be obtained to shed light on the control measures to be taken.At last,the main research results were summarized,and outlook of further research was discussed.