Hybrid GR4J-LSTM modeling for streamflow prediction of extreme events in data-scarce regions: Upper Blue Nile Basin, Ethiopia

This research was conducted in the Upper Blue Nile Basin of Ethiopia, focusing on the Gilgel Abay, Ribb, and Gumara watersheds. The study evaluates DeepGR4J, a hybrid rainfall-runoff model that combines the GR4J conceptual framework with a Long Short-Term Memory (LSTM) neural network. The model was trained using data from the Gilgel Abay watershed between 1994 and 2009, and its transferability was tested in the Ribb and Gumara watersheds. Outputs from GR4J, including net precipitation (Pn), evapotranspiration (En), production store content (Ps), and percolation (Perc), were used as inputs to the LSTM. Model performance was assessed using Nash-Sutcliffe Efficiency (NSE), Kling-Gupta Efficiency (KGE), Percent Bias (PBias), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Squared Error (MSE). DeepGR4J outperformed both standalone GR4J and LSTM models, achieving an NSE of 0.87 and KGE of 0.89 in Gilgel Abay. In the Ribb and Gumara watersheds, the model achieved NSE values of 0.91 and 0.92, respectively. The model also demonstrated excellent reliability in predicting extreme events, including the 1996 flood and the 2005 drought. These results demonstrate that DeepGR4J is a robust and transferable tool for streamflow simulation and flood prediction in regions with limited hydrological data. • DeepGR4J combines GR4J outputs with LSTM to improve streamflow prediction in data-scarce basins. • DeepGR4J combines GR4J outputs with LSTM to improve streamflow prediction in data-scarce basins. • DeepGR4J simulates floods and droughts reliably, outperforming standalone models in extreme flow events.