Exploring alternate coupling inputs of a data-driven model for optimum daily streamflow prediction in calibrated SWAT-BiLSTM rainfall-runoff modeling

Accurate streamflow prediction in mountainous regions is vital for sustaining water resources in downstream areas, ensuring reliable availability for agriculture, energy, and consumption. However, physically based prediction models are prone to substantial uncertainties due to complex processes and the inherent variability in model parameters and parameterization. This study addresses these challenges by exploring alternative coupling inputs for data-driven (DD) models to optimize daily streamflow prediction in a calibrated SWAT-BiLSTM rainfall-runoff model within the Astore sub-basin of the Upper Indus Basin (UIB), Pakistan. The research explores two standalone models (SWAT and BiLSTM) and three alternative coupling inputs: conventional climatic variables (precipitation and temperature), cross-correlation based selected inputs, and exclusion of direct climatic inputs, in calibrated SWAT-BiLSTM model. The study spans calibration, validation, and prediction periods from 2007 to 2011, 2012 to 2015, and 2017 to 2019, respectively. Based on compromise programing (CP) ranking, SWAT-C-BiLSTM (Q P ) and SWAT-C-BiLSTM (T 1 Q P ) showed competent performances followed by BiLSTM, SWAT-C-BiLSTM (PTQ P ), and SWAT. These findings highlight that excluding climatic parameters alternative SWAT-C-BiLSTM (Q P ) enhances the couple model’s accuracy sufficiently and underscores the potential for this approach to contribute to sustainable water resource management.