Hydrological Time Series Forecasting Using Three Different Heuristic Regression Techniques

Abstract

The ability of three different heuristic regression methods, least-square support vector regression (LSSVR), multivariate adaptive regression spline (MARS), and M5 model tree (M5Tree), was investigated in forecasting hydrological time series. In the study, daily streamflow data from two stations in Turkey were used by employing cross validation method. Models were compared with each other with respect to root-mean-square error, mean absolute error and determination coefficient. The LSSVR was found to be superior to the MARS and M5Tree models in daily streamflow forecasting. The effect of periodicity component and log transformation was also examined in forecasting daily streamflows by adding day number of each streamflow data and log transformated data as inputs to the applied models. It was observed that the accuracy of LSSVR model was slightly increased by adding periodicity component in both stations while the accuracy of M5Tree model was decreased in forecasting streamflows of Derecikviran station. The accuracy of all the three models was not increased by using log transformation of the daily streamflow data as inputs. © 2017 Elsevier Inc. All rights reserved.