FlowGATFormer: A streamflow prediction model based on spatiotemporal dual attention

Achieving high accuracy in long-term river streamflow prediction is essential for simulating the terrestrial hydrological cycle and managing water resources. However, traditional hydrological models are often computationally expensive and require complex parameter estimation and calibration. Deep learning models, particularly long short-term memory (LSTM) networks, have shown promise in streamflow forecasting, but their predictive accuracy tends to decrease at longer forecast horizons. To address these issues, this study proposes FlowGATFormer, a streamflow forecasting model that integrates a graph attention network (GAT) with the Informer architecture. The model introduces a dual spatiotemporal attention to capture spatial dependencies and temporal dynamics in river systems. FlowGATFormer is applied to predict streamflow at five forecast horizons including 1, 6, 12, 16, and 24 days ahead. Specifically, the GAT models spatial dependencies among hydrological stations located upstream and downstream of a river network, while the Informer captures temporal dependencies within long-term streamflow series derived from the same hydrological station. Based on data from 11 stations in the CAMELS-CH dataset (1999–2020), FlowGATFormer significantly outperforms baseline models by achieving a Nash–Sutcliffe efficiency (NSE) of 0.5616 (NSE > 0.5) for the 12-day-ahead prediction. Compared to the LSTM baseline, the NSE improvement gradually increases with the prediction horizon, rising from 3.09% for the 6-day-ahead prediction to 10.98% for the 24-day-ahead prediction. We believe that the FlowGATFormer has the potential to be a robust data-driven approach for high-accuracy, long-term river streamflow prediction.