Grated Recurrent Unit Network Quantile Regression for Silicon Content Prediction in Blast Furnace
Shihua Luo, Dong Wang, Yufang Dai, Xiaohui Liu
Ensuring the safe and steady operation of a blast furnace hinges on accurate predictions of silicon content. However, these predictions often fall short, proving unreliable and imprecise, particularly in unstable furnace conditions and in the face of substantial data swings. The wide variances and low reliability of silicon concentration predictions make them unsuitable as a reference for daily blast furnace maintenance and adjustments. To counter this engineering challenge, we introduce a novel silicon content prediction technique: the Gated Recurrent Unit Network Quantile Regression (GRUQR). This method amalgamates the Gated Recurrent Unit Network with quantile regression to refine the silicon content prediction model. Our approach first leverages GRUQR to anticipate the silicon content in molten iron across various quantiles. Subsequently, we scrutinize the patterns of silicon content changes and identify the optimal quantile for silicon content under different furnace conditions. We also discuss the reliability of our silicon concentration forecasts and present confidence intervals at various levels. To validate the effectiveness of the proposed GRUQR method, we employ real-world data from a Chinese blast furnace ironmaking process. These prediction results serve as a reliable reference for furnace operators, enabling them to determine the furnace temperature under challenging conditions.