Comparing Linear and Nonlinear Time Series Models in River Flow Forecasting (Case Study: Baranduz-chai River)

Document Type : Research Paper


1 Msc student, Tabriz University, Tabriz, Iran.

2 Assosicate Professor, Tabriz University, Tabriz, Iran.

3 Professor, Tabriz University, Tabriz, Iran.

4 Assistant Professor, Urmia University, Urmia, Iran.


     Forecasting river flow using current models like time series, Physical-theoretical and regression models are very important in water resources management. In this study monthly and daily discharge of Barandouz-Chai River for period of 37 years have been used in order to linear ARMA and Nonlinear Bilinear models. ADF test has been used for stationarity test. Result showed removing nonstasionarity components cause to river series became stationary. For nonlinearity rate of monthly and daily streamflow series BDS test has been employed. As results showed monthly series detect low nonlinearity but daily series has high nonlinearity dependence. For monthly series AR(8) and BL(8,0,1,1) selected as the best models. Regression coefficient and root mean square error of these models are 0.906 and 2.892 (m3/s) for linear model and 0.930 and 2.436 (m3/s) for nonlinear model respectively. For daily streamflow series, AR(25) and BL(25,0,1,1) were best models with regression coefficient and root mean square error equal to 0.873 and 4.116 (m3/s) for linear model and 0.923 and 3.085 (m3/s) for nonlinear model respectively. According to results using bilinear nonlinear model cause to reduction error of daily model up to 33 percent and increased regression coefficient to 5.7 percent. Because of high nonlinearity rate of daily streamflow series the bilinear model accuracy increased.


  • Receive Date: 23 October 2012
  • Revise Date: 04 October 2014
  • Accept Date: 06 May 2013
  • Publish Date: 23 September 2014