Runoff is one of the major components of calculating water resource processes and is the main issue in hydrology. Many concept models are used to predict the amount of runoff, which in most cases depend on topographical and hydrological data. Conventional models are not appropriate for areas in which there is little hydrological data. Changes in runoff are nonlinear, meaning it is time & space independent. Therefore it is not easy to simulate the runoff by simple models. Nowadays an appropriate method used in cases where there is a lack of data, is ANN (Artificial Neural Network). The precipitations, temperatures and flows of KAN watershed station between the years of 1996 to 2006 and physiographic characteristics were used as input data for the Artificial Neural Network to predict runoff. 80% of the data is randomly input into the program and the remaining 20% is used to check the accuracy of the result. For the purpose of determining an optimal network, two types of transfer functions, 12 types of training functions and between 1 and 9 kind of hidden neurons are used. After analyzing the hidden layers and various training functions, the results show that the best structure for estimating the runoff is using the precipitation, temperature, flow, LM training function and Tansig transfer function and 4 of the hidden neurons as input data. The results indicated that a Neural Network with such a structure can accurately estimate the runoff. (0.78≥ R2 ≥ 0.68 and 0.03 ≤ RMSE ≤0.53). Keywords: Estimation of Runoff, Artificial Neural Network, Back Propagation Algorithm, Kan basin
)