Evaluating Monthly Flow Prediction Based on SWAT and Support Vector Regression Coupled with Discrete Wavelet Transform

Abstract

Reliable and accurate streamflow prediction plays a critical role in watershed water resources planning and management. We developed a new hybrid SWAT-WSVR model based on 12 hydrological sites in the Illinois River watershed (IRW), U.S., that integrated the Soil and Water Assessment Tool (SWAT) model with a Support Vector Regression (SVR) calibration method coupled with discrete wavelet transforms (DWT) to better support modeling watersheds with limited data availability. Wavelet components of the simulated streamflow from the SWAT-Calibration Uncertainty Procedure (SWAT-CUP) and precipitation time series were used as inputs to SVR to build a hybrid SWAT-WSVR. We examined the performance and potential of the SWAT-WSVR model and compared it with observations, SWAT-CUP, and SWAT-SVR using statistical metrics, Taylor diagrams, and hydrography. The results showed that the average of RMSE-observation’s standard deviation ratio (RSR), Nash–Sutcliffe efficiency (NSE), percent bias (PBIAS), and root mean square error (RMSE) from SWAT-WSVR is 0.02, 1.00, −0.15, and 0.27 m3 s−1 in calibration and 0.14, 0.98, −1.88, and 2.91 m3 s−1 in validation on 12 sites, respectively. Compared with the other two models, the proposed SWAT-WSVR model possessed lower discrepancy and higher accuracy. The rank of the overall performance of the three SWAT-based models during the whole study period was SWAT-WSVR > SWAT-SVR > SWAT-CUP. The developed SWAT-WSVR model supplies an additional calibration approach that can improve the accuracy of the SWAT streamflow simulation of watersheds with limited data.