Article ID: ISIJINT-2025-155
The permeability index directly influences the gas-solid equilibrium within blast furnace, affecting the reduction reactions and overall furnace operation. Accurate prediction of permeability index remains challenging due to the system inherent complexity, time delays, and noise interference. This research proposes a hybrid TCN-GRU deep learning framework enhanced by variational mode decomposition (VMD) and error compensation (EC) correction for permeability index. First, a multi-stage feature selection method combining LightGBM and Spearman's rank correlation analysis identifies key predictor variables while addressing time-lag effects. The permeability index series is then decomposed into intrinsic mode functions (IMFs) via VMD to mitigate non-stationarity. Each IMF is modeled using a TCN-GRU architecture that captures multi-scale temporal dependencies through dilated causal convolutions and recurrent gating mechanisms. To further refine results, prediction errors from IMF components are recursively fed back into the model for error compensation. Tested on production data from a steel plant in Southern China, the framework demonstrates exceptional performance in one-hour-ahead permeability index prediction, achieving a RMSE of 0.201, a R2 of 0.965, and a remarkable hit rate of 98.039% within ±0.5 error margin. Crucially, it maintains strong multi-step prediction capability, delivering considerable hit rates for two-hour and three-hour predictions. These results underline the model's ability to handle complex blast furnace dynamics, providing a robust tool for proactive process optimization. This approach advances intelligent ironmaking by enabling precise permeability index prediction, supporting energy conservation, and enhancing operational stability in industrial applications.
