Abstract: Constructing a high-accuracy deformation prediction model of dam structure is of great significance for dam risk assessment and formulation of preventive measures. Previous dam deformation prediction models lack an effective explanation of the time-lag characteristics, and ignore an influence analysis and evaluation of the deformation characteristic factors in model construction, thereby lowering prediction accuracy. This paper presents a Gated Recurrent Unit (GRU) architecture combined with a temporal attention mechanism to overcome these problems. First, a Kalman filter is used to denoise the original dam deformation data series and remove its outliers; then, Random Forest (RF) is used to analyze and evaluate the importance of different deformation characteristic factors, and pick out key model input factors. Finally, to consider the dam deformation lag, a time attention mechanism is used to further improve the attention of the GRU model to the time-dimension dynamic features and to enhance its adaptive learning capability to time-dimension information. This, through visualizing time attention, can further improve the interpretability of a prediction model for the dam deformation in the hidden state stage. The results of engineering case studies show our model, of higher prediction accuracy and strong explanatory power for hidden state levels, can reveal the long-term effects of temperature and water level factors on dam deformation. Thus, it is a new effective method for dam safety monitoring.

Key words: dam deformation prediction, deformation lag, time attention mechanism, gated recurrent unit neural network, deep learning