Hybrid GNN and temporal encoder models for flood forecasting in the Vu Gia–Thu Bon River Basin

Abstract Flood forecasting in river networks requires models that can capture both spatial dependencies across river basins and the temporal dynamics of hydrological processes. This study proposes a hybrid approach combining Graph Neural Networks (GNNs) with temporal sequence models, such as Long Short‐Term Memory (LSTM) networks and Transformer encoders, to forecast water levels in the Vu Gia–Thu Bon (VGTB) River Basin, a highly flood‐prone catchment in Central Vietnam. The models are trained using multi‐source historical datasets, including water level observations, rainfall records, and reservoir release data, and are evaluated for their ability to predict water levels several hours in advance. Two model variants are developed: (1) a GNN‐LSTM model, which integrates GNN spatial learning with LSTM for temporal encoding, and (2) a GNN‐Transformer model, which leverages a Transformer encoder to effectively capture long‐range temporal dependencies. Results show that the hybrid GNN‐temporal models significantly outperform a conventional LSTM model, with the GNN‐Transformer model achieving a 20–30% reduction in root‐mean‐square error (RMSE). The GNN‐LSTM model also demonstrates notable improvements, highlighting the importance of incorporating river network connectivity into flood forecasting. The study demonstrates that the hybrid models are particularly effective during major flood events, as they better capture rapid water level rises and the spatial propagation of floods. The findings suggest that integrating graph‐based spatial learning with advanced temporal encoders offers a promising direction for improving flood forecasting accuracy, thereby contributing to more effective flood risk management and water resources planning.