Genetic Programming And Its Application In Pre-diction Of Silicon Content In Hot Metal Of Blast Furnace

"Prediction" is a timeless topic that is required in different areas. For the iron-making industry, prediction is applied on the content of silicon that reflects the temperature of blast furnace and the accuracy of which infects the whole process of iron-making. The fire work-ers requires to know the trend of silicon content, thus making appropriate operation. How-ever, the operation itself will infect the temperature, which should be considered by the prediction model. The so-called "closed-loop control" is making decision according to the output of prediction model and adjusting prediction model by such decision.The complexity, nonlinear, instability and the inaccuracy of the collection of data make big difficult for the prediction of silicon content. The models that based on mecha-nism do not have good result, while black-box models can not refresh itself fast and repress the prediction formula, thus they do not meet the requirement of closed-loop control. So this paper put attention on the study of "white-box models"."White-box models", in this paper, refer to the expert-knowledge-based models that can provide prediction formula. From the rules and formula provided by the model, we can infer the reason for the rise and decline of temperature, which help us to realize closed-loop control.In these models, Genetic Programming (GP) has almost the same nonlinear curve fit-ting capability and prediction ability as neural network, while it has two drawbacks, namely, the complexity of formula and instability. So this paper presents Bounded Genetic Pro-gramming (BGP) and Multi-function Regression to overcome these two drawbacks. BGP strengthens the stability of Genetic Programming while Multi-function Regression can de-compose the long function as a series of short functions.Depart from these improvements, the paper set up the theory of Extension Optimiza-tion and Abstract Learning, based on the Grammar Guided Genetic Programming. The the-ory allows computers to learn best white-box model according to the result of forecasting, that is, let computers try different kinds of prediction algorithms and select the best one, and learn expert knowledges at the same time.New algorithm for prediction of time-series. Genetic Series System (GSS) arose from this theory. It defined "Function Simulation Mode" and made decomposition for time-series prediction algorithms, which allows computers to design algorithms automatically. In ap-plication, we chose classification and regression to be its basic elements. The experiments proved that GSS has excellent predictive ability and it provided clear prediction formula as the same time.The study of this paper help to realize closed-loop control and it can be applied in other fields as well.