Using nonlinear autoregressive neural network for estimation daily evaporation: a comparison of neural networks with different algorithms

Abstract

AbstractIn this research a model of Dynamic NN(NARX) was applied to estimate the daily Evaporation of Mosul cityusing certain climate parameters(the maximum and the minimum temperature ,rain ,relative humidity ,windspeed and the sun shine )for any day in the year , and comparison for Static NN like FFBPNN, CFBPNN . Eachof these networks has two architecture: an architecture with four layers and five cells in hidden layers from onehand, and an architecture with five layers and five cells in the hidden layers from the other.Different algorithm were used for the training like: Levenberg-Marquardt algorithm (LM), Quasi-Newtonalgorithm (BFGS), Conjugate Gradient algorithm (CFG), Gradient Descent algorithm (GD) and GradientDescent with Momentum algorithm (GDM). Data was obtained from the forecast Directorate in AlRashedeyyahdistrict in Nineveh Province for the period (1995-2008) and used in the research. Data of ten years for the period(1995-2004)was employed to develop the models and the data of four years was used to evaluate the models andto compare their outputs with the data measured for the period(2005-2008). Moreover; determination coefficientR_square (R2) and the Root Mean Square Error ( RMSE) methods were used to estimate the level ofcorrespondence for the measured data and NN outputs to select the best prediction model from the modelsapplied.Results showed that the NARX with(LM) algorithm is efficient in improving a prediction model to estimate thedaily Evaporation as the value of coefficient estimation was 0.99, and this is considered the best and the fastestalgorithm if temperature, rain, relative humidity ,wind speed and sunshine data available for any day in the year