Prediction and regulation of blast furnace iron output based on iterative feature optimisation and interpretable analysis

To address the challenges of strong temporal dependencies, multivariable coupling, and nonlinearity in blast furnace hot metal output prediction, this study proposes an interpretable prediction framework that integrates a feature selection module based on Light Gradient Boosting Machine and Recursive Feature Elimination with Cross-Validation (LightGBM_RFECV) with a hybrid LSTM-Transformer model. First, an iterative two-stage feature optimisation framework is adopted: feature importance is evaluated using LightGBM (based on information gain), and redundant features are eliminated through RFECV with recursive cross-validation. Then, a combined prediction model that integrates temporal modelling and attention mechanisms (LSTM-Transformer) is constructed, achieving high prediction accuracy (MAE = 5.47, R 2 = 0.8747, and a ± 20-ton accuracy rate of 96.17%), significantly outperforming baseline models. Furthermore, the Captum toolkit is introduced to perform feature attribution, quantifying the influence mechanisms of process variables on output and forming an intelligent ‘prediction–attribution–regulation’ closed loop. Experiments based on real production data verify the effectiveness of the proposed method, providing data-driven decision support for energy conservation, emission reduction, and stable blast furnace operation, thereby contributing to the intelligent transformation of the steel industry.