Research on runoff prediction method based on multi-graph spatio-temporal feature fusion

To improve the accuracy of watershed runoff prediction and address the lack of interpretability in data-driven hydrological models, we propose a spatiotemporal prediction model, named the Multi-Graph Attention Mechanism with Gated Recurrent Unit (MGAT-MIC-GRU). This model leverages a multi-graph attention mechanism, combining Graph Attention Network (GAT) and Gated Recurrent Neural Network (GRU), for runoff forecasting. Firstly, employing hydrological data from the Qishui River Basin, a graph neural network is utilized to extract the topological structure of the watershed site and generate feature vectors; Secondly, according to the characteristics of hydrological time series data, a multi-graph attention runoff spatio-temporal prediction model based on MIC mechanism is established to predict the runoff of the basin; Finally, we conduct comparative experiments over multiple forecasting horizons to evaluate the predictive performance of various models, including the MGAT-MIC-GRU model. The research findings demonstrate that our proposed model improves the accuracy of runoff prediction.