Incorporating Flow Direction as a Physical Prior in Graph Networks: A Directed Temporal Graph Convolutional Network for Enhanced Water Quality Forecasting in Lake Basins

ABSTRACT

Graph neural networks (GNNs) for water quality forecasting typically employ undirected graphs, failing to capture asymmetric flow‐driven transport. This study introduces a Directed Temporal Graph Convolutional Network (D‐TGCN)—a flow‐aware architecture that explicitly incorporates upstream‐downstream connectivity as a physical prior. Evaluated in the Poyang Lake Basin, D‐TGCN reduced RMSE by up to 19.1% for Total Nitrogen and averaged 13.4% for Dissolved Oxygen compared to undirected models, with maximal improvements in stable unidirectional river sections. The model also demonstrated effective cross‐basin transferability. These results confirm that embedding flow direction substantially improves watershed‐scale prediction accuracy and spatial explicitness. Our work establishes a physics‐informed, spatiotemporally coupled framework for intelligent water quality forecasting using Earth science big data.