Uncertainty analyzing of Neural Network in downscaling of ‎HadCM3 data with bootstrap confidence interval method

In the statistical downscaling methods which is based on the relationship between AOGCMs data and ground based climatic variables (such as rain and temperature)، the future period of those variables are simulated. Since in the simulation، all effective parameters cannot be modeled، estimated values suffers from be uncertainty. The outputs of downscaling models are used as inputs to agriculture and water resources models; therefore، identifying the models inputs’ error or uncertainty is essential to realize the reliability of obtained results. In this research، an attempt is made to investigate the uncertainty of Artificial Neural Network (ANN) as a downscaling model in a case study in the northwest of Iran. For this purpose، precipitation، minimum and maximum temperature variables were used in the designed ANN model، and the NCEP data was employed for its calibration and validation. The HadCM3 was the selected AOGCM in this study. Observed daily time series were gathered at all stations in the study period and on the basis of bootstrap method the 99% confidence interval was calculated for all the variables. In the next step، the simulated (downscaled) mean and variance of the variables by the ANN model، compared to the calculated confidence interval. To compare the results، the criterion of the number of station-month was used. The results showed that the average maximum temperature at 14 station-months were within the confidence interval. The results of monthly analysis showed that the accuracy of ANN model in summer was low and its uncertainty is more than the other seasons. In the simulation of minimum temperature based on this criterion، 18 station-months were within the confidence interval. The accuracy of ANN to estimate the minimum temperature in summer was low with high uncertainty in almost all the stations. Moreover، in June and August in any of the stations estimated values were not within the confidence interval. Due to the high variability of rainfall in relation to temperature، confidence range was very high، and in some stations was more than 50% of average monthly precipitation. Because of the high confidence rang of precipitation، in 53 Stations-month cases were within the confidence interval.