Uncertainty quantification of displacement prediction for arch dams with cracks: fusion of deep stacked residual shrinkage network and conformal Monte Carlo quantile regression
Lists
Only you see this
Abstract
Abstract To enhance the reliability of displacement prediction for arch dams with cracks while quantifying the uncertainty in predictions, this paper innovatively proposes a novel method integrating deep learning and uncertainty quantification. Initially, the elastic net is employed to accurately identify the key factors affecting displacement. Subsequently, soft thresholding denoising technology and residual shrinkage block groups are introduced to construct the deep stacked residual shrinkage network (DSRSN), with the effectiveness of each component validated through an ablation study. Subsequently, conformal prediction, Monte Carlo dropout, and quantile regression are combined to propose the conformal Monte Carlo quantile regression (CMCQR) method, which comprehensively quantifies the aleatoric uncertainty of the data and the epistemic uncertainty of the model. Finally, the superiority of the proposed method is demonstrated through a case study of an arch dam with cracks. The case study shows that the DSRSN model outperforms traditional deep learning models in point predictions, capturing both the overall trend and local variations in dam displacement more accurately. The CMCQR method offers significant advantages in quantifying uncertainty, providing more comprehensive evaluation information. This research not only enriches the theoretical foundation of dam displacement prediction and uncertainty quantification but also offers novel technical support for dam safety management.
BibTeX
@inproceedings{xu2025uncertainty,
author = {Bo Xu and Wenjie Zhang and Hu Zhang},
title = {{Uncertainty quantification of displacement prediction for arch dams with cracks: fusion of deep stacked residual shrinkage network and conformal Monte Carlo quantile regression}},
booktitle = {Measurement},
year = {2025},
doi = {10.1088/1361-6501/ae2b2e},
}Loading related content...