Author(s): Gustavo Adolfo Atuncar Zevallos

Linked Author(s): Gustavo Adolfo Atúncar Zevallos

Keywords: Deep learning; LSTM network; Rainfall-runoff; SWAT model

Abstract: The accurate simulation of rainfall–runoff processes is essential for advancing hydrological forecasting and supporting evidence-based water resources management. This study evaluates and contrasts the performance of a physically based hydrological model (SWAT) and a data-driven deep learning architecture (Long Short-Term Memory, LSTM) for daily discharge simulation in the Vilcanota River Basin, a high-elevation Andean watershed characterized by steep topography, glacierized headwaters, and marked climatic variability. SWAT was parameterized using detailed physiographic, land-cover, soil, and meteorological datasets, whereas the LSTM model was trained using daily precipitation, temperature, and discharge with a 60-day input window to capture basin-scale hydrological memory. Model performance was assessed for the 2012–2016 evaluation period using RMSE, NSE, and KGE. The results demonstrate clear performance gains for the LSTM network (RMSE = 35.80; NSE = 0.91; KGE = 0.91) compared to SWAT (RMSE = 42.73; NSE = 0.87; KGE = 0.81). Furthermore, computational and setup efficiency differed substantially: SWAT required three days for configuration, calibration, and validation, while LSTM achieved complete training and optimization in approximately ten minutes.

DOI:

Year: 2026