River flow prediction using hybrid models of support vector regression with the wavelet transform, singular spectrum analysis and chaotic approach


Baydaroglu O., Koçak K. , Duran K.

METEOROLOGY AND ATMOSPHERIC PHYSICS, cilt.130, ss.349-359, 2018 (SCI İndekslerine Giren Dergi) identifier identifier

  • Cilt numarası: 130 Konu: 3
  • Basım Tarihi: 2018
  • Doi Numarası: 10.1007/s00703-017-0518-9
  • Dergi Adı: METEOROLOGY AND ATMOSPHERIC PHYSICS
  • Sayfa Sayıları: ss.349-359

Özet

Prediction of water amount that will enter the reservoirs in the following month is of vital importance especially for semi-arid countries like Turkey. Climate projections emphasize that water scarcity will be one of the serious problems in the future. This study presents a methodology for predicting river flow for the subsequent month based on the time series of observed monthly river flow with hybrid models of support vector regression (SVR). Monthly river flow over the period 1940-2012 observed for the KA +/- zA +/- lA +/- rmak River in Turkey has been used for training the method, which then has been applied for predictions over a period of 3 years. SVR is a specific implementation of support vector machines (SVMs), which transforms the observed input data time series into a high-dimensional feature space (input matrix) by way of a kernel function and performs a linear regression in this space. SVR requires a special input matrix. The input matrix was produced by wavelet transforms (WT), singular spectrum analysis (SSA), and a chaotic approach (CA) applied to the input time series. WT convolutes the original time series into a series of wavelets, and SSA decomposes the time series into a trend, an oscillatory and a noise component by singular value decomposition. CA uses a phase space formed by trajectories, which represent the dynamics producing the time series. These three methods for producing the input matrix for the SVR proved successful, while the SVR-WT combination resulted in the highest coefficient of determination and the lowest mean absolute error.