Modeling of watershed runoff using discrete wavelet transform and support vector machines

Abstract

In this study, hybrid models are proposed for watershed runoff forecasting by discrete wavelet transform (DWT) and support vector machines approaches. Standard support vector machines (SVM) and least squares version of standard support vector machines (LSSVM) are considered for the modeling studies. The study regions cover the Tahtali and Gordes watersheds which are located at the Aegean coast of Turkey. Meteorological data, which represent watersheds, were decomposed into wavelet sub-time series by DWT. Ineffective sub-time series were eliminated by using Mallows' C-p based all possible regression method to prevent collinearity. Then, effective sub-time series constituted the inputs of SVM and LSSVM. The mean squared errors, determination coefficient, and Nash-Sutcliffe coefficient statistics were used for the comparing criteria. The results of hybrid models were also compared with conventional SVM and LSSVM models. When the model statistics are examined, it has been observed that the DWT method has increased the performances of SVM and LSSVM. DWT combination has proved itself to be precise in predicting especially the peak and low runoff values. Among all the models, the combined use of DWT and LSSVM methods (DWT-LSSVM) has produced more sensitive runoff predictions for both basins.