Advancing Daily Streamflow Forecasting in Snow-Fed Mountainous Region Using a Novel Hybrid SWAT-BiLSTM Approach

Abstract

Abstract Precise prediction of streamflow ensures reliable planning and management of water resources. Physical-based prediction models are prone to significant uncertainties due to the complexity of processes involved as well as due to the uncertainties in model parameters and parameterizations. This study evaluates the performance of daily streamflow prediction in Astore a snow-fed mountainous region, by coupling physical-based semi-distributed hydrological Soil and Water Assessment Tool (SWAT) with data-driven (DD) Bidirectional Long Short-Term Memory (BiLSTM) model. Firstly SWAT and BiLSTM models are calibrated individually then coupled in three modes; SWAT-D-BiLSTM: flows obtained from SWAT with default parameters values used as one of the input in BiLSTM, SWAT-T-BiLSTM: flows obtained from SWAT with three most sensitive parameters values used as one of the input in BiLSTM and SWAT-A-BiLSTM: flows obtained from SWAT with all sensitive parameters values used as one of the input in BiLSTM. Input selection for DD model was carried out by cross correlation analysis of temperature, precipitation, and total rainfall with streamflow. The calibration, validation, and prediction of coupled models are carried out for periods 2007–2011, 2012–2015 and 2017–2019, respectively. Prediction performance is evaluated based on Nash-Sutcliffe Efficiency (NSE), coefficient of determination (R2), and Percentage Bias (PBIAS). Temperature showed greater correlation of 0.7 at 1-day lag as compared to precipitation and total rainfall with streamflow at daily time scale. The results showed that integrated model SWAT-A-BiLSTM outperformed SWAT-T-BiLSTM followed by SWAT-D-BiLSTM, BiLSTM and SWAT respectively. This study recommends coupling of hydrological models facing uncertainties with DD models.