GAT-GRU Based Model for Water Network Flow Prediction

Hydrological prediction can make advance judgement on the occurrence of floods, which can effectively reduce the damage losses caused by natural disasters. This paper proposes a spatiotemporal water network flow prediction model that combines Graph Attention Network (GAT) and Gated Recurrent Unit (GRU). The Graph Attention Network is utilized to capture spatial dependencies such as upstream and downstream node relationships, while the Gated Recurrent Unit is employed to capture time dependencies related to rainfall intensity. A hydrodynamic model is established for the study area, and rainfall data is simulated using a rainfall generation formula for model validation. The predictive results are visualized for presentation. Experimental results demonstrate that the new model outperforms mainstream hydrological prediction models significantly.