End sulfur content prediction method of molten iron in KR based on Kmeans–BP neural network

Abstract

In the steel manufacturing process, an accurate prediction of end sulfur content in KR is crucial for steadily controlling sulfur content in molten iron and improving steel properties. Regarding the end sulfur content prediction in the KR process, an integrated modeling method based on Kmeans clustering analysis and the BP neural network (BPNN) is proposed in this paper. As an unsupervised learning method, Kmeans clustering analysis can complete data classification according to the similarity of influencing factors instead of depending on target values. The BPNN, as a supervised learning method, can effectively explore the correlation between influencing factors and target values. The integration of these two methods can realize information exploration of data from different dimensions. Based on this understanding and the actual production data in one steel plant, the prediction model of end sulfur content in KR based on Kmeans–BPNN is studied. First, datasets of different operating conditions are constructed according to the pattern recognition and classification of production data in the KR process through Kmeans clustering. By establishing the relation curve between the number of clustering centers and the mean error of clustering results and selecting the adjacent positions to 10% of the maximum mean error difference, the number of Kmeans clustering centers is confirmed as five. Then, the prediction model is trained by different datasets based on the BPNN. The input layer and hidden layer have five nodes, and the output layer has one node in the BPNN-based prediction model of end sulfur content in KR. A piecewise linear function is selected as the activation function, and the maximum number of training is fixed at 1,000. Finally, the prediction models of different datasets are integrated and formulated in the final prediction model of end sulfur content in molten iron, realizing the prediction of different molten iron conditions and operating conditions. To test and verify the effectiveness and accuracy of the prediction model based on the Kmeans–BPNN method, the end sulfur content prediction of molten iron in KR is performed by applying prediction models based on desulfurization reaction kinetics, routine BPNN, and Kmeans–BPNN using the same training and testing datasets. The prediction results indicate that the end sulfur content prediction in KR based on the Kmeans–BPNN method is significantly more accurate than that of the prediction model based on the desulfurization reaction kinetics and the routine BPNN model.

Keywords

• kr
• sulfur content
• prediction
• kmeans clustering
• bp neural network
• desulfurization reaction kinetics