Dynamic spatio-temporal graph network based on multi-level feature interaction for sinter TFe prediction

Sinter ore is one of the primary raw materials for blast furnace ironmaking, and the iron grade (TFe) is a crucial indicator for assessing the quality of sinter ore, as its concentration directly impacts the yield and quality of blast furnace production. The iron ore sintering is a complex industrial process characterized by inconsistent sampling frequency, variable time delays, and intricate spatio-temporal dependencies, making the prediction of TFe content particularly challenging. Previous research has often focused on extracting global features from the sintering process as a whole, neglecting the interactions between subprocess features and the integration of subprocess features with global features. To address these challenges, this paper proposes a dynamic spatio-temporal graph network for TFe prediction in sinter ore based on multi-level feature interactions (MLDSTGN). First, to effectively utilize high-frequency sampling data while accounting for variable time delays in the sintering process, a time delay feature reconstruction (TDFR) method is designed. This method serializes and extends the data through a delay window, facilitating alignment with low-frequency sampling data. Second, adaptive graph construction (AGC) employs an attention mechanism to learn the underlying spatial dependencies. This module constructs a global graph representing overall dependencies in the sintering process and a local graph for the dependencies within subprocesses. The proposed spatio-temporal graph learning (STGL) module captures long- and short-term temporal features at different levels, and the spatial information aggregation layer further extracts deep spatial features that encompass the synergistic effects among variables. Additionally, multi-level dynamic interactive learning (MDIL) is introduced to enhance information transfer between global and local features. Finally, simulation results based on actual operational data demonstrate that the proposed model outperforms all baseline methods in multi-step TFe prediction, with a reduction of at least 4.86% in MAE and at least 13.27% in RMSE.