One of the methods to forecast peak discharge among black box method is artificial neural network.
Lack of emphatic regulation to architect network is one of disadvantages in neural networks. There is not
a suitable criterion to determine number of layer and number of neuron in hidden layer, type of activation
function for hidden and output layers and assay and error method is the only solution. In this research,
numbers of suitable neurons in this layer and activation function type in output layer for two sub basin,
Gatehdeh and Gelinak, situated in Taleghan basin were investigated by perception neural network with
three layers and preservation of type of activation function for hidden layer. 20 percentage data for testing
stage, 65 percentage data for training stage and 15 percentage data for validation stage were entered in
MATLAB software and sigmoid and linear activation functions with the most application in hydrologic
subjects were selected for output layer. Linear function with less RMSE in both stations recognized suitable
and the numbers of adequate neurons in Gatehdeh and Gelinak stations were five and six neurons,
respectively.