Multi-Channel Graph Convolutional Network based End-Point Element Composition Prediction of Converter Steelmaking

Abstract

Abstract End-point composition is an important quality standard for the converter steelmaking process, which consists of multiple elements, including C, Si, Mn, etc. However, it is hard to measure the element composition online. Real-time and precise prediction for element composition is essential for the optimization of alloy addition so as to bring economic profits. Nevertheless, most conventional models neglect the correlations among element compositions and predict each element composition without the information from other elements. In this paper, a new multi-channel graph convolutional network is proposed to integrate these correlations with the process variables together for a more accurate prediction model. The proposed model uses graph structure to describe the correlations among element compositions. Specifically, through the multi-channel design, each element composition can be learned based on process variables in an independent channel. Element compositions and correlations among them are respectively described by nodes and edges in graph. With the constructed graph, the graph convolution across channels can fuse the features of correlated elements to explicitly exploit the correlation information for performance improvement. Besides, compared with conventional methods which learn relations among nodes based on distances, we take sparse representation learned by sparse coding as edges to describe the correlations among nodes. As strong correlations exist among element compositions, the consideration of correlation information can integrate the learning of correlated elements and bring performance improvement. Experiments based on the real converter steelmaking process demonstrate the superiority and effectiveness of the proposed model.