Rainfall-runoff modeling using least squares support vector machines

Abstract

Over the past decade, artificial neural networks (ANN) have been widely used in the runoff modeling studies. In spite of a number of advantages, ANN models have some drawbacks, including the possibility of getting trapped in local minima, over training, subjectivity in the determining of model parameters, initialization of the weights in each simulation randomly, and the components of its complex structure. In the past decade, a new alternative kernel-based technique called a support vector machine (SVM) has been found to be popular in modeling studies because of its advantages over ANN. Least squares version of support vector machines (LS-SVM) provides a computational advantage over standard support vector machines by converting quadratic optimization problem into a system of linear equations. The LS-SVM method is preferred in this study. The main purposes of this study are to examine the applicability and capability of LS-SVM for the prediction of runoff values of Tahtali and Gordes watersheds, which are the major surface water resources for the city of Izmir in Turkey, and to compare its performance with ANN and other traditional techniques such as autoregressive moving average and multiple linear regression models. For these purposes, meteorological data (rainfall and temperature) and lagged data of runoff were used in modeling applications. Some favorite statistical performance evaluation measures were used to assess models. The results in study indicate that the LS-SVM and ANN methods are successful tools to model the monthly runoff series of two study regions and can give better prediction performances than conventional statistical models. Although these two methods are powerful artificial intelligence techniques, LS-SVM makes the running time considerably faster with the same or higher accuracy. In terms of accuracy, the LS-SVM models, which involve different normalization types, resulted in increased accuracy to that of the ANN models.