Spatiotemporal Graph Neural Network for River Discharge Prediction

Discharge prediction is essential for water resource management, enabling a better understanding of hydrological variability and its response to environmental and human factors. We present a Spatio-Temporal Graph Geural Network (STGNN) model predicting hourly river discharge for the upper Weiße Elster, flowing into the Saale river a tributary of the Elbe. The STGNN is a novel graph-based Long Short-Term Memory (LSTM) neural network jointly modeling spatial connectivity and temporal dynamics. It uses elevation, land‑use and soil data as well as approximately one year of temporally aggregated weather variables, but no previous discharge to generate one prediction at the 8793 nodes of this catchment. It was evaluated against a traditional random forest regression model and a graph neural network without explicit temporal structure, outperforming them across multiple metrics, achieving a $R^2$ of 0.763, a Root Mean Square Error (RMSE) of 9.54e-3 mm/h and a Kling–Gupta Efficiency (KGE) of 0.753. The trained STGNN was applied to the nearby Schwarzwasser catchment and achieved a KGE of 0.618, highlighting is ability to generalize. These results show that data-driven modeling can profit from physical realism and offer an adaptable framework combining predictive accuracy and generalizability to support water resource management.