A new spatiotemporal long-term prediction method for Continuous Annealing Processes

Abstract

Accurately and consistently predicting the heating state is essential to maintain stable operation in continuous annealing processes (CAPs). However, long-term prediction biases often arise due to unmodeled dynamics associated with high-dimensional, time-varying, and strongly coupled variables. This study introduces a spatiotemporal-based forecast model designed to extend the prediction horizon while significantly reducing bias accumulation. The model leverages the spatial characteristics derived from classified process parameters by analyzing the internal structure and dynamics of the process. Additionally, it captures the temporal features of each parameter type through deep learning techniques that preserve and learn from historical data, enabling the model to account for the autocorrelation of multiple variables, including the output, and their correlation with the output. We conducted experiments with real process data, confirming the model’s accuracy and consistency in real-world settings. Additionally, ablation experiments validated the need to integrate both temporal and spatial features for long-term prediction accuracy. Compared to existing methods, the proposed model significantly reduces prediction bias and enhances forecast robustness.