An integrated deep learning model for predicting concrete dam deformation with multi-point spatiotemporal correlation

Abstract

Concrete dam structures exhibit complex, nonlinear responses to various influencing factors, making accurate deformation prediction of concrete dams both crucial and challenging. Most monitoring models for dam deformation prediction primarily emphasize temporal features and the correlation between environmental factors and dam deformation, often neglecting the spatial dependencies within the data. Even multi-point models that account for spatial coordinates still struggle to effectively capture the time-varying spatial correlations between different measurement points. This study presents an integrated deep learning model for deformation prediction that incorporates multi-point, time-varying spatiotemporal correlations to overcome the aforementioned challenges. The proposed model leverages a combination of long short-term memory (LSTM) neural network and Kalman filter (KF), further enhanced by K-means clustering and an attention mechanism. K-means clustering identifies each target point’s associated measurement points; Kalman filter-estimated deformation values from the associated measurement points serve as additional inputs to the LSTM model to capture their time-varying dynamic influences; and the attention mechanism enhances the interpretability of the LSTM model. The proposed model is subsequently employed to predict and analyze the deformation of the concrete dam, and its performance is compared against five other prediction models, including multiple linear regression, a standalone LSTM model and other LSTM-based models. Results from six measurement points show that incorporating spatiotemporal correlations increases R² of the proposed model by an average of 11.0 % over the standalone LSTM model and 7.3 % over the attention-based LSTM model, which did not account for spatial correlation. The proposed model reduced RMSE by 47.1 % relative to the standalone LSTM model and by 38.3 % relative to the attention-based LSTM model. The MAE decreased by 45.8 % versus the standalone LSTM model and 24.0 % versus the attention-based LSTM model. Moreover, the proposed model offers meaningful interpretability, making it a practical and forward-looking approach for structural health monitoring of concrete dams.