compare learning function in neural networks for river runoff modeling

accurate prediction of river flow is one of the most important factors in surface water recourses management especially during floods and drought periods. in fact deriving a proper method for flow forecasting is an important challenge in water resources management and engineering. although, during recent decades, some black box models based on artificial neural networks (ann), have been developed to overcome this problem and the accuracy privilege to common statistical methods (such as auto regression and moving average time series method) have been shown. in these research only attended change number of hidden layer and number of neurons for to approach to the best structure in neural network, and complex in proper network designand can’t be simply used by other investigators. in this study examined 15 the neural network learning function and the results indicated in the structure of the network with one hidden layer (ann1),learnlv1, learnh and learnis by mse=0.000158, 0.000185 and 0.000188, have been better performance than the other learning functions. and in the structure of the network with two hidden layer (ann2),learnh, learnsomb and learncon learning function by mse=0.000154, 0.000173 and 0.000176 have been better performance than the other learning functions.but on the other hand by ten times run this two models, learnsom and learngdm learning functions in ann1 model and learnh and learnos in ann2 model had most frequency among the best learning functions and thus it is better that the number of hidden layer not more than one, when we use back propagation network (that its learning function is learngdm). because in this way we have more chance to achieve ideal response. but if we are going to increase network performance byincreasing the number of hidden layer, it is better that use the default of network and learngdm carefully.