Abstract: For the time series prediction issue of dam deformation, a spatiotemporal multi-dimensional input matrix of deformation is derived considering the correlation of deformation at multiple measuring points; an Informer-AD dam deformation prediction model is constructed that integrates multi-dimensional spatiotemporal information based on K-means clustering. We use the K-means clustering to partition rationally the deformation measuring points, then apply a panel data regression model to integrate the analysis of spatiotemporal dimensions and partition results. Finally, we develop an Informer-AD dam deformation prediction model to integrate multi-dimensional spatiotemporal information. This model is used to learn spatial feature sequences and integrate spatial features through a fully connected layer to output predicted dam deformation values. Its application to a concrete gravity dam shows that our prediction method, considering spatiotemporal correlation, can fully explore the relationship of the overall state of dam deformation versus the spatial distribution characteristics of measuring points. It better captures the spatiotemporal characteristics of deformation values and thus improves prediction accuracy, which implies that our model has a high accuracy and satisfactory applicability, useful for engineering application.

Key words: deep learning, dam deformation prediction, Informer-AD, spatiotemporal correlation characteristics, K-means clustering