Integrating Mechanistic Modeling with Attention-Enhanced GRU Networks to Predict Silicon Content and Slag Basicity in Blast Furnaces

Accurate prediction of hot metal silicon content and slag basicity during blast furnace operations is critical for process optimization and production control. Traditional mechanistic models, though grounded in solid physical principles, struggle to capture the complex dynamics and non-equilibrium factors under real-world conditions. Meanwhile, purely data-driven models lack explicit physical interpretability. To address these limitations, this study proposes a hybrid prediction framework that integrates mechanistic models with data-driven methods. First, the final silicon content in hot metal and slag basicity are predicted using Aspen Plus® based on thermodynamic equilibrium principles. Subsequently, a Gated Recurrent Unit network combined with a multi-head attention mechanism is used to correct the residuals between the mechanistic model predictions and actual plant data. Cross-validation ensures model robustness, and evaluation metrics such as RMSE, MAE, and HR are utilized to assess performance. The method proposed in this study shows the smallest deviation from actual measurements, exhibiting the best overall performance. This approach preserves the physical interpretability and theoretical foundation of the mechanistic model while overcoming its computational and predictive limitations. Moreover, the trained model operates efficiently, enabling real-time predictions and dynamic optimization, providing substantial practical value for industrial decision-making and process adjustments.