Abstract: Deformation behavior of a dam usually features significant spatiotemporal correlation during its collaborative load-bearing process. However, existing methods suffer from drawbacks such as feature redundancy and inadequate utilization of residual information in multi-point collaborative inversion and prediction. This paper describes a spatiotemporal hybrid deformation model for gravity dams that integrates feature selection and residual correction techniques. First, mechanical parameters of the dam body are inversed using a seepage-stress coupled finite element model combined with the multi-objective grey wolf optimizer (MOGWO). Then, we construct a spatiotemporal input set from the data of monitoring point coordinates, and select key features using an improved BorutaShap algorithm, so that a FEM-BorutaShap-MLR prediction model can be built. Finally, the iTransformer architecture is adopted to capture nonlinear spatiotemporal dependencies in the residuals to correct the results predicted. Validation results at three monitoring points of a gravity dam show this prediction model achieves determination coefficients (R2) greater than 0.98 at all points, and all its evaluation metrics outperform the models compared. The numerical model, making use of physical mechanisms and data-driven approaches, improves its accuracy in multi-point deformation prediction for gravity dams.

Key words: gravity dam deformation prediction, spatiotemporal hybrid model, feature selection, residual correction, multi-objective optimization