Abstract: Hybrid models are often used to predict the overall deformation of dams. However, most previous studies have focused on concrete dams, lacking an effort into the widespread masonry dams. This paper considers the time-dependent characteristics of nonlinear dam material, and constructs a corresponding multi-point statistical model of the hydrostatic component by introducing a time parameter and the relative coordinates for the observation points. Generally, difficulties exist in optimizing the coefficients of a hydrostatic component model if it is equipped with multiple parameters; thus, we proposed an improved particle swarm optimization (IPSO) algorithm to enhance the particle randomness and interactivity and accelerate the search for optimal model coefficients. By combining finite element method (FEM) and a Kalman filter (KF), a FEMK model is constructed and used for predictions. And a deep learning algorithm LSTM model is used to train the temperature and aging factors after PCA dimensionality reduction and to predict the corresponding deformation. The resultant model, namely the spatiotemporal hybrid model FEMK-LSTM-PCA jointly constructed by these two models, has a high accuracy in the overall prediction of dam deformation, which is verified by its application to engineering examples.

Key words: mortar masonry dam, time-dependent factor, multipoint prediction, hybrid model, deep learning