Data- and Model-Based Discharge Hindcasting over a Subtropical River Basin

Abstract

This study aims to evaluate the performance of the Soil and Water Assessment Tool (SWAT), a simple Auto-Regressive with eXogenous input (ARX) model, and a gene expression programming (GEP)-based model in one-day-ahead discharge prediction for the upper Kentucky River Basin. Calibration of the models were carried out for the period of 2002–2005 using daily flow at a stream gauging station unaffected by the flow regulation. Validation of the calibrated models were executed for the period of 2008–2010 at the same gauging station along with another station 88 km downstream. GEP provided the best calibration (coefficient of determination (R) value 0.94 and Nash-Sutcliffe Efficiency (NSE) value of 0.88) and validation (R values of 0.93 and 0.93, NSE values of 0.87 and 0.87, respectively) results at the two gauging stations. While SWAT performed reasonably well in calibration (R value 0.85 and NSE value 0.72), its performance somewhat degraded in validation (R values of 0.85 and 0.82, NSE values of 0.65 and 0.65, for the two stations). ARX performed very well in calibration (R value 0.92, NSE value 0.82) and reasonably well in validation (R values of 0.88 and 0.92, NSE values of 0.76 and 0.85) at the two stations. Research results suggest that sophisticated hydrological models could be outperformed by simple data-driven models and GEP has the advantage to generate functional relationships that allows investigation of the complex nonlinear interrelationships among the input variables.