5:15 PM - 7:15 PM
[AHW29-P03] Application of Attention-Based Graph Neural Networks for Spatial Distribution Prediction of Streamflow
Keywords:Flood, Graph Neural Network, Deep learning, Streamflow
Accurate streamflow prediction is critical for water resource management and flood forecasting. Traditional physics-based hydrological models face challenges in responding quickly to rapid hydrological events due to inefficiencies in model calibration and high computational costs, especially for large-scale catchments. While deep learning models have been widely used, they often fail to capture the spatial and temporal dependencies of runoff dynamics. In this study, we present a novel hybrid model combining Graph Attention Networks (GAT) to capture the spatial topology of runoff transfer and Long Short-Term Memory (LSTM) networks, enhanced with an attention mechanism to simulate the temporal dynamics between upstream and downstream runoff. The model was applied to the Yangtze River Basin, the largest river basin in China, to predict streamflow at a 10 km spatial resolution. Validation results show that the model achieves a median Nash-Sutcliffe Efficiency (NSE) of 0.783 at secondary outlet stations and effectively simulates streamflow peaks during flooding events. Furthermore, the model can simulate the spatial distribution of daily streamflow for an entire year in just 10 seconds, offering a significant speed advantage over traditional physics-based models. This work marks a significant advancement in streamflow prediction using deep learning, offering both higher efficiency and improved predictive accuracy.