Final Sulfur Content Prediction Model Based On Artificial Neural Network For Hot Metal Pretreatment

With the development of metallurgical industry and the improvement of steel quality, pre-desulfurization of hot metal has become an important task for steel production. In order to meet the demand of hot metal pretreatment processing with the character of fast rhythm and high efficiency, the predicting of the final sulfur content with computer model are put forward by the predecessors. However, it is a complex process to predict the final sulfur content. It is unsuitable for the features of multiparameter, nonlinear and uncertainty to modeling with traditional theroy model, and then the artifical intelligence method is applied to predict recently.Based on the productive practice of Meishan Steel Co. Ltd. and Benxi Steel Co. Ltd., adopted the improved BP algorithm, and used Visual Basic 6.0 programme software, the prediction model of final sulfur content during hot metal pretreatment processing is established. During modeling process, normal BP algorithm is analysed and improved for overcome its disadvantages of overmuch iterative repetition and slow convergence. All kinds of parameters in the model are elaborated. In the view of thermodynamic, kinetics and combining the characteristic of the field datas, the factors affecting final sulfur content during hot metal pretreatment processing are detailedly investigated. At the same time, the network configuration, input and output parameters are established.Training samples of Meishan Iron and Steel Co. Ltd. are 1154 heats, and which are 1900 heats for Benxi Iron and Steel Co. Ltd. 100 heats datas are randomly selected as the test samples respectively, and which are different from the above heats. The model is seperately trained and tested using the selected samples. And then the reasons resulting in error are analysed and discussed. The following conclusions are drawn:(1) The improved BP algorithm which is adaptive to this subject is put forward by adjusting study rate, adding momentum coefficient and employing the learning method of maximal error. The new study rate is as follows:...