Prediction of Converter Gas Recovery Flow based on A Dynamic Multi-graph Fusion Network

Considering the characteristics of the intermittence of the Linz Donaniz converter gas (LDG) recovery, the dynamic updating of converter plans, and the interplay between converters, this paper proposes a novel LDG recovery flow prediction approach based on a dynamic multi-graph fusion network (DMGFN). The converters are treated as the graph nodes in the proposed DMGFN, where the time convolution module and the graph convolution module are designed to leverage and extract the temporal and spatial connections and static and dynamic associations between different nodes. In graph convolution module, two dynamic graphs are constructed based on the converter plans and the historical production rhythm correlations of the converters at each time step, which are integrated with pre-defined static graph. These modules are trained simultaneously to extract the effective information contained in the related time series data. The DMGFN can effectively describe the real-time plans and process information from both upstream and downstream production processes to reveal the dynamic nature of real-time steelmaking plan alterations. Simulation experiments are performed utilizing the data from a steel plant in China, demonstrating that the proposed method significantly enhances the prediction accuracy of LDG recovery flows. In comparison to other methods, the proposed approach yields more precise prediction results in aspect of both the recovery timing and the total volume of LDG for each converter.