A Concrete Dam Deformation Prediction Method Based on LSTM With Attention Mechanism

Dams are the main water retaining structures in the hydraulic engineering field. Safe operations of dams are important foundations to ensure the hydraulic functionalities of these engineering structures. Deformation, as the most intuitive feature of the dams' operation behaviors, can comprehensively reflect the dam structural states. In this case, the analysis of the dam prototype deformation data and the establishment of a real-time prediction model become frontier research contents in the field of dam safety monitoring. Considering the multi-nonlinear relationships between dam deformation and relative influential factors as well as the time lag effect of these influential factors, this article adopts long-short-term memory (LSTM) network algorithm in deep learning to deal with the long-term dependence existing in dam deformation and explore the deformation law. The method proposed in this work can effectively avoid the gradient disappearance and gradient explosion problems by using the recurrent neural network (RNN). In addition, this work adopts the Attention mechanism to screen the information that has significant influence on deformation, combining the Adam optimization algorithm that has high calculation efficiency and low memory requirement to improves the learning accuracy and speed of the LSTM. The model overfitting is avoided by applying the Dropout mechanism. The effectiveness of this proposed model in studing the long time series deformation prediction of concrete dams is confirmed by case studies, whose MSE (mean square error) and other 4 error indexes can be reduced.