Attention-Mamba cooperative prediction and regulation model for blast furnace hot metal output based on time-lag features

Abstract

The blast furnace is a high-temperature closed black-box reaction vessel. Its internal physical and chemical processes are characterized by strong coupling, large time delay, and multivariable heterogeneous time lags, which severely restrict the accurate prediction and closed-loop control of hot metal output. Based on real industrial data, this paper proposes a complete time-lag feature-based attention-Mamba collaborative scheme for blast furnace hot metal output prediction and control. First, an adaptive correction self-attention-based imputation for time series adaptive anomaly correction model is constructed, which accurately quantifies the dominant time lags and effective action intervals of each process parameter on hot metal output by virtue of the diagonal masked causal attention mechanism. Then, a comprehensive importance evaluation index is extracted based on the attention weight distribution to screen core feature parameters. On this basis, an attention-Mamba hot metal output prediction model is proposed, which integrates the global attention importance vector as prior knowledge into the selective scanning mechanism of Mamba. Finally, an attention-Mamba-based genetic algorithm-model predictive control multivariable collaborative control framework is established, which deeply embeds the extracted time-lag parameters into the control logic and achieves global optimization under physical safety constraints and state violation penalty mechanisms. The experimental results show that: (1) the dataset processed with adaptive correction self-attention-based imputation for time series time-delay weighted alignment significantly reduces the root mean square error (RMSE) of the proposed attention-Mamba model by 37.2% compared to the original unaligned data (from 14.215 to 8.923 t); (2). The single-step prediction RMSE of the attention-Mamba model is as low as 8.923 t, with an R 2 of 0.931 and a ±15 t prediction hit rate of 93.2%; (3) industrial field tests demonstrate that the proposed control framework reduces the average error of hot metal output per cast by 38.9% and the maximum error by 48.2%. This study realizes the deep integration of high-precision hot metal output prediction and multivariable collaborative control for blast furnaces, and provides reliable theoretical and technical support for the intelligent closed-loop control of the blast furnace ironmaking process.