Research on a wide applicability prediction model for embankments with different settlement patterns in permafrost regions

In permafrost regions, due to climate warming and other factors, subgrade deformation may persist for extended periods. Accurately predicting the settlement of frozen soil subgrades is crucial for the stable operation of transportation infrastructure. However, the deformation rules of frozen soil subgrades are diverse, influenced by complex factors, and challenging to monitor. Settlement monitoring data can comprehensively reflect the effects of multiple influencing factors. Against this background, this study focuses solely on historical deformation monitoring data and compares three commonly used types of embankment settlement prediction methods, including curve fitting, grey models, and machine learning approaches. Based on this, a Stacking ensemble algorithm was employed to integrate different categories of prediction models, validated using settlement data from eight monitoring sites, and a Stacking-based model for embankment settlement in permafrost regions was developed and compared with traditional models. The results demonstrate that individual prediction models tend to exhibit inconsistent performance across different monitoring sites and working conditions, often lacking sufficient generalization capability. In contrast, the Stacking Hybrid Ensemble Model effectively leverages the strengths of multiple models, significantly improving overall prediction accuracy while maintaining stable and reliable performance across diverse conditions and locations. This highlights its superior adaptability and generalization ability, underscoring its potential for practical engineering applications in cold-region infrastructure monitoring and maintenance.