Software Design On Self-organized Evolutionary Experience Prediction Of Silicon Content In Hot Metal Tapped From Blast Furnace

It is essential to keep reasonable thermal level in blast furnace so that iron-making process in it works at a steady and cost-effective way. Silicon content in hot metal tapped from blast furnace is an indirect criteria of the thermal level in blast furnace. However, existing models used for prediction of silicon content in hot metal can not meet the requirement of process control of blast furnace for the model. In this thesis, a new model used for prediction of silicon content in hot metal based on self-organized experience evolution approach has been investigated by developing prototype of the model with software engineering methodology, optimizing model parameters and testing it with process data of blast furnace in Tianjin Iron Plant.The prediction model of silicon content in hot metal tapped from blast furnace based on self-organized experience evolution is composed of following components: feature extraction and vector quantification of temporal process data, memory, storage, accumulation and evolution of the prediction experience. According to the requirements of operation of blast furnace and software engineering methodology, the properties and function of prototype of prediction system have been analyzed and hierarchical structure and data flow and data processing in the system have been outlined by using data flow diagram. As a result, the prototype, which is developed with OO methodology, function-satisfactory and user-friendly, is able to be transplant and embedded into Blast Furnace Guard, a type of software for intelligent supervisor control of blast furnace process.Eight of input and state variables of blast furnace, i.e. silicon content, irregular coke charge, coal injection, blast pressure, blast temperature, blast volume, batch of charge in a tapping period, ratio of ore-to-coke in charge, are considered as input of prediction model. With process data of blast furnace No.1 in Tianjin Iron Plant, the parameters, Ru and Sita, of model have been optimized based on criteria of predictability, hit ratio, direction hit ratio and mean sum of square error according to each variable respectively. In addition, time delay of impact of irregular coke charge to silicon content in hot metal has been analyzed statistically. It is shown that silicon content in hot metal of 60.24% tapping increases after six hours of irregular coke charging.From the results of optimizing parameters of the model, two cases of off-line prediction of silicon content in hot metal have been tested with plant process data. It isshown that prediction by model is better than these made by human expert according to the criteria of predictability, hit ratio and direction hit ratio, whatever if the prediction are made based single variable or multi variables and if with experience or without experience of machine system. When mean sum of square error is used for criteria, the prediction by using model with three or four variable without experience are better than these by expert. The prediction by using repeated learning of system with four variables makes a remarkable result. In this case, the predictability can reach to 77.4% on average. Even expert makes prediction from more information, the hit ratio and direction hit of prediction by model is higher 15% than that by expert. Furthermore, a few promising ways to improve prediction of model have been discussed.