Abstract: The monitoring effect value is an objective index that characterizes changes in the service performance of a dam and reflects changes in its working behavior. For an embankment dam, safety monitoring and behavior prediction by using this index are of great significance for the operation and risk management and control. Most of the previous methods focused on time-series modeling for a single measuring point, that is, developing a prediction model for a single location, and left room for improving the modeling of spatial correlation characteristics and the analysis of environmental driving mechanisms. This paper constructs a spatiotemporal fusion multi-measuring-point prediction model with a feature extraction mechanism integrated, starting from the characteristics of a multi-dimensional spatiotemporal distribution field, i.e., effect field, composed of the multi-point monitoring effect quantities, and considering time development similarity at these points and differences in spatial distributions. Through a parallel branch network, this model uses the Gated Recurrent Unit (GRU) to capture causal time-series characteristics driven by the environment, and combines with the Convolutional Neural Network (CNN) to mine spatial distribution patterns at the multiple measuring points. It can achieve the collaborative feature fusion of multi-dimensional information by introducing an adaptive feature fusion strategy, so that it succeeds in synchronous and high-precision prediction of the seepage flows in an embankment dam at multiple measuring points. A case study of the Huairou reservoir seepage, based on the monitoring data nearly 30 years long, shows our new model effectively balances its capability of representing spatiotemporal feature while maintaining computational feasibility. It has a synchronous multi-point prediction accuracy significantly higher than traditional methods, and advances behavior evolution analysis for the entire dam section.

Key words: embankment dam, behavior prediction, spatial-temporal fusion, multiple points, prediction model