VC-TGCN: A Novel Soft-Sensor Method of Silicon Content in Blast Furnaces

The soft sensor of silicon content in molten iron is crucial for the ironmaking process. Based on the temporal characteristic of the silicon content, several methods utilizing recurrent neural networks have been proposed and validated. Besides the temporal feature extraction in soft sensor, the inter-correlation of the variables in the ironmaking industry can be used to further improve the accuracy. But the existing methods for constructing graph structures using a single correlation metric limits the accuracy. To get the holistic correlation, we proposed a novel method called Vine Copula-Temporal and Graph Convolutional Network (VC-TGCN). We introduced the R-vine copula to model the correlations in multivariate joint distributions. The model combines the temporal and spatial feature extraction based on graph convolutional networks, time convolutional networks and the vine copula, which can obtain more comprehensive dynamic relationships and correlations between variables of the ironmaking process. Experiments validate the superiority of VC-TGCN over other models, demonstrating its feasibility in soft sensor of the silicon content.