A Neural Network Model Predicting The Silicon Content In Hot Metal At No.3 Blast Furnace Of No.2 Ironworks TangShan Iron And Steel Co.

The key point ensuring the output and quality of the hot metal is the in-order-state of blast furnace. While the important index predicting the state of the blast furnace is the heat state of the hearth. Because of the running process complexity of the blast furnace, it is impossible for us to master the distribution of the temperature in the furnace. Thus the silicon content in hot metal indirectly reflects the temperature in the hearth and its variation tendency generally. In the iron-making process, we could not taking the exactly adjusting step, stabilizing the temperature in hearth, and cutting down the abnormal state of the furnace in time until the silicon content in hot metal and its changing tendency were predicted accurately. Therefore, establishing the prediction model of the silicon content in hot metal possesses profoundly academic and realistic meaning. According to the actual operational condition and technology levels on No. 3 furnace in No.2 Ironworks of Tangshan Iron and Steel Co., a neural network model for predicting the silicon content in hot metal was set up in this study using BP neural network integrated with partial expert knowledge.The following content were conducted and conclusions were drawn in this paper: (I) the comparison between mathematic model, expert system, and neural network model. There would be wide foreground for prediction model of silicon content in hot metal established through combining the neural network with partial expert knowledge closely. (II) Establishing of the prediction model of BP neural network. The structure of network was selected as 24-14-1. 24 input parameters in network model were decided by Delphi expert investigation. That the input parameters were introduced time order enforced the online prediction ability of the model. (III) Offline prediction of neural network model. The neural network model was trained using 150 groups of samples. The power values matrix gained from training process made up the neural network prediction model of silicon content in hot metal. Offline prediction results using 70 successive samples were suggested that the target ratios of the fixed pattern and amendatory pattern prediction were 100% when the error was below 0.1%, and that the target ratios of fixed pattern and amendatory pattern prediction were 82.9% and 84.3%, respectively, when the error was below 0.05%. (IV) Online prediction of neural network model. The amendatory pattern prediction model was used at No.3 furnace. The online prediction results using 210 group samples suggested that the target ratio was above 85%. According to the analysis to neural network model, we found that it could not control the silicon content in hot metal only by the neural network prediction model. Therefore, we introduced temperature rising and decreasing expert rules in the controlling system of silicon content in hot metal, and utilized the improved system to provide operational guidance locally. (V) Realization of the model function. Using VB language program, we exploited software of the prediction model of silicon content in hot metal. The software was composed of training, prediction, guidance, and help, which have some functions, such as self setting up of the model, offline and online prediction, providing operational guidance locally, etc.