Abstract: The measured deformation of a concrete faced rockfill dam is highly nonlinear and complicated, owing to a variety of influential factors and the collinearity among them. To improve the deformation prediction in dam analysis, this paper develops a deformation prediction model of concrete faced rockfill dams based on the factor fusion. First, we use the variational mode decomposition to decompose a deformation time series so as to effectively reduce its complexity and enhances feature extraction. Next, we employ the partial least square regression to reduce and fuse the influential factors of deformation, reducing the impact of multicollinearity between independent variables on model construction and enhancing model interpretability. Finally, we reconstruct and predict the subsequences using a one-dimensional convolutional network fused with a gated recurrent unit neural network. Analyses of certain real projects show our model greatly improves the efficiency and accuracy of deformation prediction for concrete faced rockfill dams, and is also useful for deformation monitoring and analysis of similar dams.

Key words: deep learning, dam deformation prediction, concrete faced rockfill dam, variational mode decomposition, partial least squares method