Majid Khazaei; Mohammad Reza Mirzaei
Abstract
Prediction in hydrology is as estimation of hydrological and meteorological conditions in a specific interval time. In this regard, understanding the relationship between precipitation and runoff is necessary for water resources optimal management. The purpose of this study was to compare different models ...
Read More
Prediction in hydrology is as estimation of hydrological and meteorological conditions in a specific interval time. In this regard, understanding the relationship between precipitation and runoff is necessary for water resources optimal management. The purpose of this study was to compare different models of artificial neural networks (two type of ANNs: RBF and MLP) and time series models (ARMA) to discharge estimation in a part of the Taleghan watershed, using monthly flow discharge data for a period of 30 years between 1977 and 2007. Among the different ARMA models, a model with a lowest error and akaike (AIC) criterion was selected as an optimal model. Using trial and error method, ANNs were designed by specifying the number of hidden layers and neurons in each layer, sigmoidal transfer function, training function, weight/bias learning function and performance function. Using trend analysis, Halt-Winters and Box-Jenkins (ARMA) methods, time series analysis showed that ARMA (2, 2) (R= 0.77) and Halt-Winters (R=0.72) presented more accurate results. In general, it could be concluded that ANNs models produced more accurate predictions of flow discharge than time series approaches. Also, the results revealed that the MLP model (average R=0.83) produced more accurate predictions of flow discharge than RBF model (average R=0.81). Assessment of accuracy of all models based on RMSE and R showed that the model 1 (with RMSE= 6.45 and R= 0.86) obtained with a network architecture of 4-20-1 configuration. Model 1 used the input vector consisting of antecedent monthly discharge with one to four time lag.
