Streamflow prediction in the Danube River Basin using a multi-source graph-integrated GCN-LSTM model

Danube River Basin Streamflow prediction is essential for water resources management and flood-risk mitigation. To better exploit spatial dependencies among gauging stations, we use a Graph Convolutional Network–Long Short-Term Memory (GCN–LSTM) framework for regional streamflow forecasting. The model integrates static and dynamic predictors from ERA5-Land and is evaluated at both monthly and daily temporal resolutions, with benchmarks against MLP, RNN, GRU, LSTM, and biLSTM. We further assess robustness and transferability through station-thinning experiments and cross-station validation at ungauged sites. GCN–LSTM consistently outperforms the conventional baselines and maintains strong predictive skill across station densities, indicating that high accuracy can be achieved without an extremely dense gauging network, although performance drops when fewer than ∼30 training stations are available. Daily inputs substantially improve the representation of short-term fluctuations and extreme flows, highlighting the importance of high temporal resolution for capturing rapid hydrological processes. Cross-station tests show a clear accuracy loss at ungauged locations that increases with distance from the training stations, underscoring that spatial generalization remains a key limitation in hydrologically heterogeneous regions. • Streamflow prediction in the Danube Basin using GCN-LSTM at monthly and daily resolutions. • Daily-scale inputs enhance model performance in extremes and short-term dynamics. • Station-thinning experiments show that model performance is largely stable as station numbers decrease. • Cross-validation confirms spatial generalization but highlights peak-flow underestimation. • Provides insights for refined basin management and extreme event risk assessment.