Comparative assessment of a physically based model (SWAT) and a deep learning model (LSTM) for streamflow simulation in the Abelti Watershed

Accurate streamflow simulation is fundamental for effective watershed management and sustainable water resources planning. This study is a novel application of the Long Short-Term Memory (LSTM) models for streamflow simulation in the Abelti Watershed, Omo-Gibe Basin, Ethiopia. It also compares the performance of the LSTM model with the physically based Soil and Water Assessment Tool (SWAT) model. The analysis covered the period 1992–2021, with monthly streamflow data from 1992 to 1994 used for warm-up period, from 1995–2015 used for LSTM training and SWAT calibration, and data from 2016–2021 used for testing and validation. The LSTM model was optimized through systematic hyper-parameter tuning and window size adjustment. The LSTM model attained R², RMSE, MAE, NSE, and PBIAS values of 0.996, 0.13, 0.072, 0.94 and 0.33 during training, and 0.991, 0.14, 0.09, 0.97 and 0.58 during testing, respectively. The SWAT model achieved R², NSE, and PBIAS values of 0.79, 0.75, and 0.54 during calibration, and 0.79, 0.77, and 0.23 during validation, respectively. The results demonstrate that the LSTM model outperformed the SWAT model in streamflow prediction accuracy and goodness-of-fit, reflecting its stronger ability to capture complex nonlinear rainfall–runoff relationships. While both models exhibited low bias, SWAT showed slightly lower PBIAS during validation. Overall, the findings suggest that LSTM is a promising alternative to process-based hydrological models for streamflow simulation and warrants further evaluation under diverse hydro-climatic conditions.