Spatiotemporal feature fusion modeling for multistep prediction of strip temperature in continuous annealing processes

The multistep prediction of strip temperature is essential for maintaining stable production and improving strip quality in the Continuous Annealing Process (CAP). The spatiotemporal coupling characteristics in CAP reveal the dynamic interactions among process parameters over time. However, existing methods fail to capture the joint influence of spatial coupling interactions and long-term temporal dependencies in multistep prediction tasks, limiting their ability to characterize the overall evolution of strip temperature. This study presents a spatiotemporal feature fusion modeling approach for multistep strip temperature prediction. The contributions are threefold: (1) a spatial–temporal encoder–decoder framework is designed for multistep temperature prediction, providing sufficient trend information; (2) a spatial causal graph construction approach is proposed to create a reliable digraph structure, reflecting the causal relationships among different parameters; (3) an attention-enhanced temporal modeling method is devised to accurately capture the long-range dependencies and trend patterns in the heating sequences. Experiments based on actual process data demonstrate that the proposed approach significantly reduces the Mean Absolute Error (MAE) by approximately 16.35%–28.34% compared to all baseline methods. This advancement establishes an accurate and reliable foundation for multistep strip temperature prediction in CAP. • An encoder–decoder framework is designed for multistep temperature prediction. • A spatial causal graph construction approach is proposed to create a digraph. • An attention-enhanced model is devised to capture long-range temporal dependency.