Soft rock within embankments is prone to continuous particle breakage， migration， and rearrangement due to rain infiltration and traffic loads， leading to uneven subsidence. subsidence deformation is a key indicator of embankment stability and safety， making accurate prediction essential for preventing road defects and instability. However， traditional single prediction models often lack generalizability and are not suitable for varying conditions in embankment engineering. This study collected and analyzed subsidence data from 18 soft rock embankments in highways and railways， which exhibited distinct subsidence patterns， including wave-like， broken line， and parabolic trends. Based on these data， using the Stacked Generalization （SG） ensemble algorithm， an SG fusion model predicting soft rock embankment subsidence was developed combining the prediction models from three different fields. The model avoided the hyperparameter tuning process， allowing for direct application in engineering practices. Besides， a Blocked K-Fold training strategy was employed to improve robustness. In comparison with traditional models， under conditions of limited monitoring data， the SG fusion model demonstrated significantly lower error rates and higher prediction accuracy across various projects. The findings suggest that the SG model is more applicable and robust for predicting soft rock embankment subsidence. This research provides theoretical and technical support for evaluating the service performance and post-construction maintenance of soft rock embankments.