Novel time-lag informed deep learning framework for enhanced streamflow prediction and flood early warning in large-scale catchments

Abstract

Constrained by the sparsity of observational streamflow data, large-scale catchments face pressing challenges in streamflow prediction and flood management amid climate change. Deep learning excels in simulation performance while flow lag information in data-driven approaches is barely highlighted. In this study, we introduce a time-lag informed deep learning framework for large-scale catchments. Central to this framework is the utilization of flow time-lag information between upstream and downstream subbasins, enabling precise flood forecasting at outlet driven by upstream data. Taking the monsoon-influenced large-scale Dulong-Irrawaddy River Basin (DIRB) as study area, we determined peak flow lag (PFL) days and relative annual flow scale (RAFS) for defined subbasins. By incorporating this time-lag information with historical flow data at different time intervals, we developed the optimal model for DIRB. This model was then applied to evaluate the flood processes in 2008 and 2009, using selected flood indicators. The results indicate that the time-lag information led to significant performance improvements, notably in the LSTM_PFL_RAFS model driven by upstream Hkamti sub-basin data, which achieved a Kling-Gupta Efficiency (KGE) of 0.891 (Nash-Sutcliffe efficiency coefficient, NSE, 0.904), surpassing LSTM's 0.683 (NSE, 0.785). Further integration of historical flow data with specific interval, the optimal model, H(15)_PFL utilizes Hkamti sub-basin data, reached an impressive KGE of 0.948 (NSE, 0.940). This model outperformed standard LSTM in accurately simulating key flood characteristics, including peak flows, initiation times, and durations for the 2008 and 2009 flood events. Notably, H(15)_PFL provides a valuable 15-day lead time for flood forecasting, extending the window for emergency response preparations. Future research that incorporates additional essential catchment features into the framework holds great potential in unraveling the complex mechanisms of hydrological responses to human activities and climate change.