Adaptive Dilation Attention and Disentangled Mixture‐of‐Experts for Hot Metal Silicon Prediction in Blast Furnace Ironmaking Process

Accurate prediction of hot metal silicon content is essential for stable and efficient blast furnace operations. However, this task presents significant challenges due to the nonlinear and multiscale dynamics of the smelting process and frequent sensor noise. To address these issues, this paper proposes an Adaptive Dilation Mixture‐of‐Experts (AD‐MoE) framework for silicon content prediction in the blast furnace ironmaking process. The model integrates adaptive dilation attention and a disentangled mixture‐of‐experts encoder to capture multiscale temporal dependencies, while a lightweight temporal smoothing strategy enhances stability against measurement fluctuations. The proposed framework is evaluated using mean squared error, mean absolute error, and multilevel hit rates within ±0.1%, ±0.05%, and ±0.03% tolerance intervals. Results on real blast furnace data show that AD‐MoE achieves superior accuracy and higher hit rate compared with several state‐of‐the‐art deep learning models, demonstrating its strong potential for industrial deployment.