Adaptive physics-informed graph convolutional network for flow prediction in the downstream river network of the dongjiang river

Accurate flow prediction in tidal–sluice river networks remains challenging due to nonlinear hydrodynamics, bidirectional flow propagation, and complex multi-source interactions. This study develops an Adaptive Physics-Informed Graph Convolutional Network (PI-GCN) to achieve physically consistent flow forecasting in the downstream river network of the Dongjiang River, China. The model embeds node-level water-balance equations into the graph-learning process and introduces an adaptive weighting mechanism that dynamically adjusts the contribution of the physical loss term based on training convergence. By integrating the spatial learning capability of GCN with physics-guided constraints, the PI-GCN enables simultaneous optimization of data fidelity and mass-conservation consistency. Using hourly discharge data from a calibrated HEC-RAS model covering 235 cross-section nodes, the PI-GCN achieved a mean R 2 of 0.87, outperforming conventional GCN and 1D-CNN models by 25–55% in MAE reduction. The node-level mass-balance residuals (water-balance equation errors) approach zero, indicating near-perfect mass balance across the river system. The model maintains stability under tidal backwater and sluice-operation conditions and demonstrates an optimal configuration with a 7–12 h history window and 3 h forecast horizon, balancing accuracy and temporal adaptability. This study provides a unified and generalizable framework for integrating adaptive physical constraints into graph-based hydrodynamic prediction, offering a physically interpretable and robust approach for data-driven river-network forecasting.