Abstract: Dam deformation behavior is a consequence of long-term interaction of many factors, and its evolution pattern usually involves two dimensions: time and space. However, previous intelligent modeling of dam deformation lacks a comprehensive consideration of time and space variations, and a large amount of spatiotemporal information needs to be further excavated from the prototype observation data. This paper develops a dynamic monitoring model for dam deformation with spatiotemporal coupling correlation characteristics from two view angles: time-series correlation for a single measurement point, and spatial correlation of multiple measurement points. This model takes the gated recurrent unit (GRU) neural networks as core layers to model and learn the inherent time-varying patterns in a historical deformation data series, and constructs the features of spatial variations through iterative extraction of effective deformation factors. It uses a soft attention mechanism to improve the probability weight allocation rule of the GRU hidden states, thus achieving adaptive learning of key information. Its effectiveness is verified in a case study of the Fengman concrete gravity dam. The results show that this monitoring model can accurately simulate the dynamic deformation evolution of a dam, and are more accurate in extrapolation prediction than conventional monitoring models.

Key words: dam deformation monitoring, spatiotemporal correlation characteristics, dynamic modeling and learning, gated recurrent unit neural networks, attention mechanism