Research On Comdlexity And Extended Prediction Based On SVM In Blast Frunace Ironmaking

Metallurgical industry is the pillar industry of national economy. Blast furnace (BF) ironmaking consumes the largest part of energy. Thus every technological progress in BF ironmaking will bring enormous economic and social benefits. By investigating into the actual situation of BF ironmaking, the dissertation conducted a study on the complex ironmaking process and temperature prediction of blast furnace ironmaking with data collected at the scene of expert system for BF ironmaking, which provided important ideas and technological foundation for the realization of closed-loop control and thus was meaningful both in theory and practice.In the efforts to achieve closed loop control of BF ironmaking, the key problem is accurate prediction of furnace temperature, thus realizing the effective control of silicon content in hot metal, whose complexity is reflected not only in complexity of ironmaking mechanism but also in complexity of kinetics in ironmaking process. An accurate and effective predictive model will provide important guidance to the practical production. Chapter2discussed the complexity mechanism during the ironmaking process from the perspective of technique complexity, complexity of chemical reaction and complexity of the goal and operation. Then, on the basis of the previous non-linear and non-stationary qualitative study, the dissertation proposed the application of complexity measures, such as Shannon entropy, approximate entropy. Lempel-Ziv complexity and fractal dimension etc. to conduct the complexity measure analysis of the data collected from No.6BF at Baotou Steel, No.1BF at Laiwu Steel, No.6BF at Linfen Steel and No.7BF at Handan Steel. By the complexity analysis, the study disclosed the complexity feature in ironmaking process, which laid an important foundation for establishing furnace temperature prediction model. Chapter3proposed the chaotic prediction model based on phase space reconstruction and support vector machine (SVM) for the nonlinear feature of silicon content series. Its prediction accuracy did apparently improve, compared with RBF predicton model, adding-weight one-rank local-region method, Volterra adaptive prediction model based on phase space reconstruction. The simulation results in different production periods in the same BF or another BF showed some relation between model prediction accuracy and BF’s complexity. And the highly accurate prediction result can not be achieved from the same model of high complexity BF.Chapter4took the other variables affecting the ironmaking process into consideration, applied rank increase method to obtain all the variables affecting the current temperature of BF and drew an important conclusion that the current BF temperature was affected by the previous two furnaces. For multivariate input has lots of redundancy and noises, it can’t be applied in the practical process. Thus, the study proposed the method of feature extraction and feature selection to construct optimal input features to decrease dimensions of input features. In the aspect of feature extraction, the simulation result showed that extraction based on KPCA had more practical value, which not only decreased the input dimension but maintained higher prediction accuracy. In the aspect of feature selection, the study applied SVM-RFE model to initial feature sets to obtain the optimal features and nested sets, and hit rate reached85%based on optimal features sets. For the numerical prediction, the study proposed feature selection by mutual information, and obtained the optimal feature series. The model also achieved better prediction effects. To further improve the prediction accuracy and practical usability, Chapter5proposed multiple models based on the features of various conditions and parameters. Multiple SVMs based on fuzzy C-means clustering method improved the prediction accuracy. For the uncertainty of parameters of common clustering methods, this study proposed the improved affinity propagation clustering (AP). AP didn’t need to determine the clustering number in advance, and in this study Euclidean distance was improved by the use of weighting based on mutual information. Based on the improved AP, multiple SVM model was set up for prediction. In the non-stationary period, the hit rate reached83%while in the relatively stationary period, the hit rate89%.Chapter6made a further study of the rule extraction of SVM prediction model for the "black box" feature of SVM, which is beneficial for the control of the process of BF ironmaking.Temperature prediction model based on the inference rules of decision tree and SVM was proposed. The simulation result showed that when the model achieved the prediction accuracy, the model changed the "black box" of SVM into the understandable "IF-THEN" rules. So long as the BF operator adjusted the rules based on his experience, they can be applied to the practical application of BF. Chapter7summarized the whole content and innovation points, namely, it is a study of practical BF temperature prediction models based on complexity analysis, including tendency prediction model, numerical prediction model and rule extraction model. Finally, the prospects to the further research were given.