VMD-RF-DI-GRU: a deformation prediction method for concrete gravity dams based on feature decomposition and selection

Purpose Due to the influence of various time-varying factors, the deformation monitoring data of concrete gravity dams exhibit high nonlinearity and nonstationarity. The multitude of characteristic factors and redundant features affecting dam deformation can adversely impact prediction model accuracy. Design/methodology/approach This study proposes a deformation prediction method, termed VMD-RF-DI-GRU, for concrete gravity dams based on feature decomposition and selection. First, variational mode decomposition is employed to decompose dam deformation data into trend, periodic and random components, significantly reducing time series complexity and enhancing feature extraction. Next, the out-of-bag (OOB) error-based feature selection method from Random Forest identifies the most influential factors for each component. Finally, the DI-GRU model predicts each component by integrating the advantages of DenseNet and inception to effectively extract complex features. These features are fused and input into the gated recurrent unit model to achieve high-precision predictions. Findings The method is validated using deformation monitoring data from two typical measurement points of a concrete gravity dam in the Jinsha River basin. The research results demonstrate that the VMD-RF-DI-GRU model accurately simulates dam deformation evolution and achieves higher prediction accuracy than alternative models, providing valuable guidance for monitoring similar dam types. Originality/value These new findings highlight the model’s robustness and practical applicability.