Comparison of Artificial Neural Network (ANN) and SDSM Model to Downscaling of Temperature

In this study downscaling of temperature was down in Tajan Plain located in the province of Mazandaran. The result of atmospheric general circulation models was obtained with HadCM3 climate model under scenario A2. Since the output of atmospheric general circulation models has a low locative resolution, should be downscaled in the area or Basin level that it was conducted with statistical method. The statistical methods used included of downscaling SDSM5.5. And artificial neural network model. In this study, by using the average daily temperature data of Kordkheil Station during the30-year statistic Period (1971-2001) and the large-scale variables NCEP, as inputs to the neural network and SDSM model, simulation and downscaling was down respectively of the maximum and minimum temperature in the last period to determine models error. To this end were used of the features and functions available in the programming software MATLAB. Then To evaluate the performance of the models, were used the statistical criteria including of correlation coefficient, coefficient of declaration and root mean square error between observed and predicted values ​​of temperature. The obtained results show the appropriate performance of SDSM model for downscaling temperature Than the ANN model. So that the error percentage of SDSM model is lower and the correlation coefficient is more than the ANN model. The best Structure of neural network to simulate of maximum temperature is perceptron model with four hidden layer with the 5-5-6-6 architecture and for the minimum temperature Variable is perceptron model with three hidden layer with 5.3.1 architecture.