Author(s): Xi Zhou; Siyu Cai

Linked Author(s): Siyu CAI

Keywords: Heterogeneous graph; Spatial attention; Temporal attention; Runoff forecasting

Abstract: Due to the limited number of meteorological and hydrological stations in most watersheds and their uneven spatial distribution, using conventional data-driven models to predict runoff based on these observations faces problems of insufficient accuracy and an inability to account for the distribution of meteorological stations. To improve prediction capabilities in regions with complex terrain, this study developed a graph neural network model that integrates meteorological and hydrological observation information, aiming to more accurately model the interactions between different types of stations in the watershed through a heterogeneous graph structure. In the model, a graph convolutional network and graph attention mechanism are used to extract spatial relationships, and a temporal attention mechanism combined with a long short-term memory network (LSTM) is introduced to enhance the modeling of dynamic features in time series. The proposed model was evaluated using observed data from multiple hydrological stations, and the results show that it significantly outperforms traditional models in prediction accuracy, particularly demonstrating more stable performance in areas with complex meteorological drivers or heterogeneous terrain structures. The study highlights the potential of heterogeneous graph modeling and spatiotemporal attention integration in hydrological forecasting, providing a feasible solution for watershed runoff prediction under conditions of sparse data.

DOI:

Year: 2026