Research On Modeling By Svm And Cbr And Applications In BOF Steelmaking Process

Steelmaking production is an important process in steel industry and basic oxygen furnace (BOF) steelmaking is the most popular mode. With the development of modeling and computer technology, BOF steelmaking production automation is continuously improved. Process control of BOF steelmaking reflects the automation level and affects the quality of steel. However, due to constraints of steel production process and quality of raw material, the imported model systems can not achieve the desired results. Therefore, modeling suitable models for national conditons has great significance to improve the level of automation of BOF steel production. Sub-lance based basic oxygen furnace is selected as the object of study and the models related to the process control of BOF are deeply researched in this paper. The main contents of this paper are as follows:1) Support vector machine (SVM) and case based reasoning (CBR) based control models in main blow phase are researched. For lime addition amount control, an alkalinity deviation estimation model based on SVM is proposed to evaluate the difference between actual alkalinity and aim alkalinity of slag. Further more, traditional lime addition calculation formula is modified by introducing the alkalinity deviation; for the problem of oxygen blow volume control, a case-based reasoning model with mixed case retrieve and case reuse is proposed, k nearest neighborhood and geometrical similarity methods are used in case retrieve step and correspond to the weighted sum and incremental regression in reuse step to adequately mine the information from different aspects.2) The control models based on hybrid intelligent methods in second blowing stage are researched. For the coolant addition amount, SVM based decision model is built to determine whether the coolant is needed. If needed, SVM and CBR based combine model is adopted to calculate the coolant addition amount. For oxygen volume control, a dynamic case base strategy is applied and Bayesian rewards and punishment are introduced when calculating the similarity between current case and history case to improve the accuracy of case retrieval.3) SVM and particle swarm optimal (PSO) based carbon content and temperature prediction models mainly contain the endpoint prediction and real time prediction are researched. Mechanism analysis and mutual information calculation based variable selection are proposed to choose appropriate input variables for endpoint carbon content and temperature prediction. Then the SVM method is used for modeling and predicts the endpoint. In main blow stage, bath carbon content and temperature real-time forecasting model is built based on theory decarburization model. The second blowing bath carbon and temperature prediction models are based on experience model and CBR model. The real-time bath carbon content and temperature prediction can also realize the endpoint control.