African Journal of
Agricultural Research

  • Abbreviation: Afr. J. Agric. Res.
  • Language: English
  • ISSN: 1991-637X
  • DOI: 10.5897/AJAR
  • Start Year: 2006
  • Published Articles: 6576

Full Length Research Paper

Comparison of artificial neural networks and stochastic models in river discharge forecasting, (Case study: Ghara- Aghaj River, Fars Province, Iran)

Mehrdad Fereydooni1*, Mehrdad Rahnemaei2, Hossein Babazadeh3, Hossein Sedghi4 and Mohammad Reza Elhami5
  Islamic Azad University, Science and Research Branch, Tehran, Iran. 2Department of Water Engineering, Islamic Azad University, Shiraz, Iran. 3Water Science Department, Science and Research Branch, Islamic Azad University, Tehran, Iran. 4Department of Water Engineering, Islamic Azad University, Science and Research Branch, Tehran, Iran. 5Department of Mechanic Engineering, Imam Hossein University, Tehran, Iran.
Email: [email protected]

  •  Accepted: 15 June 2012
  •  Published: 23 October 2012

Abstract

 

This study presents the application and comparison of artificial neural networks (ANN) and Stochastic models to predict the monthly flow discharge of the Ghara –Aghaj River in the southwest of Iran. The models used are, multiple perceptron using back-propagation algorithm (MLP/BP), recurrent neural network (RNN) and autoregressive moving average (ARMA). Artificial neural networks used different inputs including the monthly values of rainfall (mm), air temperature (°C), evaporation (mm), and the output of the model consists of monthly discharge (m3/s) in Band-e-Bahman climatology and hydrometric stations. Different topologies of neural networks were tested with changes in the input layers, nodes and hidden layers. Finally, MLP with two hidden layers and RNN with a hidden layer and one node were selected as the best ANN models. The finally selected ANN models (MLP-RNN) showed some disagreement with the observed values, especially during the peak discharge, but these models showed their ability to forecast low discharge with an acceptable accuracy. The neural networks (MLP-RNN) and ARMA models were validated in terms of correlation coefficients (R), root mean square errors (RMSE) and scatter indexes (SI). The results showed that the MLP (R = 0.92, RMSE = 2.4, SI = 0.56) and RNN (R = 0.9, RMSE = 1.95, SI = 1.25) techniques had better performance than the ARMA (1,13) (R = 0.68, RMSE = 5.37, SI = 1.25( model.

 

Key words: Artificial neural networks (ANN), back-propagation, prediction, autoregressive moving average (ARMA), recurrent neural networks, Mond basin.