Abstract: Deformation is a crucial indicator of the structural behaviors of a water dam. To address the issues of uncertainty quantification and model interpretability in dam deformation prediction, this study presents a sparse Bayesian learning (SBL)-driven model for interval prediction of dam deformation, considering both data noise and parameter uncertainty. We adopt a parallel Rao-3 algorithm and a cross-validation strategy to optimize adaptively the parameters of the kernel function, and then construct an optimized sparse Bayesian learning model that accurately captures the nonlinear relationship between the input variables (i.e. reservoir water level, temperature, and time-dependent variables) and output variables (i.e. dam displacements). For the variables that influence dam deformation, we calculate their feature importance by integrating global sensitivity analysis with this new prediction model, and gain valuable insights into the impact of feature variables on deformation prediction. A case study is made on an EDF concrete arch dam originating from the 16th International Benchmark Workshop on Numerical Analysis of Dams. The results demonstrate our prediction model outperforms the multiple linear regression statistical models, radial basis function networks, and Gaussian process regression models in terms of point prediction and interval prediction accuracy while maintaining good interpretability.

Key words: dam deformation prediction, interval prediction, safety monitoring, sparse Bayesian learning, global sensitivity analysis, interpretability