Under Irregular Sampling: Si-Enhanced Dynamic Relation GNN for Silicon Content Prediction in Blast Furnace

Accurate real-time prediction of silicon content (Si) in hot metal is critical for ensuring product quality and optimizing blast furnace (BF) operations. However, this task remains challenging because process variables exhibit heterogeneous sampling frequencies, and hot-metal Si measurements are collected at irregular, low-frequency intervals. To address these challenges, we propose a Si-enhanced dynamic relation graph neural network (Si-DRGNN) that explicitly models irregularly sampled sensor data and complex interdependencies among process variables. A learnable temporal embedding aligns asynchronous variables, while a masking mechanism and dynamic relation generator adaptively handle missing values and temporal gaps. A hierarchical patch-based graph structure is introduced to model cross-scale dependencies between high-frequency process variables and low-frequency Si. In addition, a temporal smoothing strategy captures gradual variations in hot-metal Si, enabling stable multi-step forecasting. Industrial experiments demonstrate that Si-DRGNN achieves high prediction accuracy but also enhances measurement information extraction and dynamic trend monitoring, providing actionable insights for BF process control.manuscr.