A deep learning prediction and interpretation method for concrete dam displacement considering the lag effect and the impact of cracks

Abstract During the long-term operation of concrete dams, the instability of external environmental factors and the generation of cracks pose significant challenges to the safety of concrete dams. To enhance the accuracy of dam displacement prediction, this paper first constructs 4 different input variable sets based on the Hydrostatic-Temperature-Time (HTT) model, taking into account the lag effect of temperature and the impact of cracks on concrete dam displacement. Secondly, the Extreme Gradient Boosting (XGBoost) algorithm is utilized to perform feature selection on the input variable sets, thereby obtaining the optimal feature subset to enhance the precision of the prediction model. Concurrently, a deep learning prediction model for the displacement of concrete dams with cracks, namely CNN-GRU-Attention, is established by integrating Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), and Attention Mechanism (Attention). Finally, taking a concrete gravity arch dam as an example, using 5 input variable sets and 4 comparative models, the performance of the models in predicting concrete dam displacement with different input variable sets is analyzed and compared, along with the inherent prediction performance differences among the models themselves. The importance of dam displacement influencing factors is also interpreted using the Attention. The results indicate that it is necessary to consider the lag effect of temperature and the impact of cracks, XGBoost feature selection can significantly improve model prediction performance, and the CNN-GRU-Attention model can more accurately predict dam displacement with better interpretability. This study provides a novel strategy for the health monitoring of concrete dams.