Multi-station Joint Runoff Forecasting Using Graph Neural Network Coupled with Spatial Connectivity of Hydrological Stations

Accurate runoff prediction is crucial for urban planning and environmental protection, land use management, and flood prevention and mitigation. Given that meteorological changes and human activities in upstream basins significantly affect downstream flow variations, we propose a novel runoff prediction method. This method combines long short-term memory networks with a graph convolutional neural network framework and integrates physical hydrological information to effectively capture spatial dependencies. A weighted directed graph is employed to construct a neighbor matrix, which learns the complex topology and spatial correlations among nodes. The proposed method was rigorously tested across multiple basins and benchmarked against state-of-the-art models. The results indicate a substantial enhancement in runoff prediction performance. Notably, the Nash–Sutcliffe model efficiency coefficient improved by up to 30.68%, the mean absolute error decreased by up to 44.41%, the mean relative error reduced by up to 55.05%, and the root mean square error decreased by up to 30.43%. These improvements, however, varied across different hydrological stations. The model serves as an effective large-scale runoff forecasting tool by integrating global hydrological feedback, thereby improving the accuracy of hydrological forecasts for both the main streams and tributaries within the basin.