A multi-step time series prediction model for blast furnace permeability index based on deep learning

Abstract: The permeability index of a blast furnace serves as a key parameter reflecting the gas-solid flow balance, and its accurate prediction is crucial to ensure stable furnace operation. To address its inherent strong time-series dependency, this study proposes a deep learning framework based on a Bayesian-optimized Temporal Fusion Transformer (TFT) model for multi-step permeability index prediction. First, a high-quality dataset suitable for multi-step prediction is constructed using preprocessing techniques, integrated with feature selection methods including Spearman and Maximal Information Coefficient (MIC), thereby laying a robust foundation for model training. Second, a multi-step prediction model based on Bayesian-TFT is developed, wherein Bayesian optimization is employed to refine TFT hyperparameters, enhancing the efficiency of training convergence. Subsequently, a rolling window strategy combined with 10-fold cross-validation is adopted to improve the model's generalization capability under complex operating conditions, achieving prediction accuracies of 97.20% and 98.45% within acceptable error ranges. Finally, interpretability analysis is conducted to identify core control parameters, providing a basis for predictive regulation of permeability. This study enables one-hour-ahead multi-step prediction of the permeability index, offering a novel approach for intelligent blast furnace ironmaking and demonstrating significant engineering value.