Abstract
Accurate prediction of silicon content in blast furnace ironmaking is essential for optimizing furnace temperature control and production efficiency. However, large-scale industrial datasets exhibit complexity, dynamism, and nonlinear relationships, posing challenges for feature selection. Traditional methods rely heavily on static statistical techniques and expert knowledge, limiting their adaptability to dynamic operating conditions. This study proposes a Bayesian online sequential update and support vector regression recursive feature elimination (BOSVRRFE) algorithm for dynamic feature selection. By integrating Bayesian dynamic updating and recursive optimization, BOSVRRFE adjusts feature importance in real-time, efficiently optimizing input variables. Experiments with data from a large steel enterprise validate BOSVRRFE's performance in silicon content prediction. Results show that BOSVRRFE outperforms traditional static methods in prediction accuracy, real-time adaptability, and model stability. Additionally, it rapidly responds to operational changes, supporting real-time industrial prediction and optimization. This study provides theoretical and practical guidance for silicon content prediction and introduces an innovative approach to feature selection in complex dynamic industrial data.