Concrete Dam Deformation Prediction Model Based on Attention Mechanism and Deep Learning

Accurate prediction of dam deformation is crucial for understanding future deformation trends, identifying potential hazards, and ensuring dam operational safety. Due to long-term exposure to complex and variable environmental conditions, concrete dam's structural safety faces numerous challenges. Traditional statistical methods based on hydrostatic-season-time (HST) theory, while having clear physical meanings and being easy to implement, are limited by their inherent linear assumptions, resulting in constrained prediction accuracy. Machine learning models such as random forest, support vector regression, and extreme learning machine (ELM) extend statistical approaches but still lack the ability to establish temporal dependencies due to their static input-output mapping relationships. To address the multi-factor coupling, temporal dependency, and nonlinear complexity characteristics of concrete dam deformation, this paper proposed a hybrid prediction model integrating attention mechanisms, bidirectional long short-term memory (BiLSTM) networks, and sparrow search algorithm (SSA). The model comprehensively matched the multiple requirements of information weighting, temporal dependency modeling, and model optimization in concrete dam deformation prediction, forming a synergistic enhancement effect among methods.The attention mechanism operates on both feature and temporal dimensions to comprehensively enhance the model's focus on critical information. The feature attention mechanism identifies and weights key factors, including water pressure, temperature, and time effects, based on their varying contributions to dam deformation prediction. Temporal attention mechanism addresses the unequal importance of historical data by assigning weights to different time moments according to their relevance to current predictions. Subsequently, the BiLSTM network established temporal dependencies of deformation data from both forward and backward directions, effectively capturing long-term dependencies in dam deformation processes through its dual-directional temporal modeling capability. Key hyperparameters, which significantly influence model performance, were automatically optimized using SSA to minimize subjective manual pre-setting and avoid the randomness of artificial parameter configuration. The proposed model was validated using radial displacement data from a high arch dam that has been in service since 1993, with data spanning from January 1, 2010, to December 31, 2019. The dam is located on the lower reaches of the Yalong River with a normal water level of 1 200 m and a maximum height of 240 m. A comprehensive comparative analysis was conducted against multiple linear regression (MLR), ELM, gated recurrent unit (GRU), and convolutional-long short-term memory (C-LSTM) methods. Results demonstrate that the proposed method can accurately predict displacement variation trends and exhibits superior prediction performance compared to both deep learning and non-deep learning comparative methods, with RMSE and MAE reduced by 10.73% and 4.37% respectively compared to the second-best method, achieving an <italic>R</italic>² of 0.987 1. Ablation experiments confirm the effectiveness of both feature and temporal attention mechanisms, with their synergistic effect significantly enhancing prediction accuracy. The model successfully captures nonlinear and time-varying characteristics in concrete dam deformation processes, showing high consistency with measured deformation patterns and demonstrating excellent engineering practicality. This research provides new insights for deformation prediction in related hydraulic engineering projects and establishes a foundation for developing real-time early warning methods based on deformation prediction for dam safety management.