Extended Abstract
Introduction
As global warming and changes in global temperature are considered to be the most important instances of climate change in the present century, temperature can be introduced as an indicator reflecting the response and feedback of climate system to these changes. In this regard, climate forecasting is performed using "simulation" approach. Using atmospheric general circulation models such as RCPs and climate scenarios developed as their output is an accepted method of simulating climate variables, especially temperature. In each of these scenarios, radiative forcing changes at a certain rate until 2100. Downscaling is the main technique used in RCPs. Different methods are used for downscaling among which artificial neural network is more widely accepted due to its more accurate evaluations.
 
Materials & Methods
Data collected for the purpose of the present study include: 1) Daily maximum temperature recorded in Qazvin synoptic station during 1961-2005. These records were derived from Iran Meteorological Organization and used as an output for calibration, fitting, and finally selecting the best fit model for the observations, 2) Atmospheric observations including daily records of 26 atmospheric variables. These data were recorded by the United States National Centers for Environmental Predictions (NCEP) and the United States National Center for Atmospheric Research (NCAR) during 1961-2005 reference period and used as input or explanatory (predictor or independent) variables in the present study 3) Representative Concentration Pathway (RCP) extracted from atmospheric general circulation model (including the output of HadCM3 model) which is used to simulate 2006-2100 reference period.
Artificial neural network model was used to downscale atmospheric data and simulate maximum temperature recorded in Qazvin synoptic station. Using Pearson correlation coefficient, the correlation between maximum temperature recorded in Qazvin synoptic station and each of the 26 atmospheric variables was estimated. Then, forward selection and backward deletion, percentage decrease index, and stepwise methods were used to preprocess the variables, select the most appropriate predictor variables (input variable in the network) and perform statistical downscaling. Following the selection of suitable explanatory variables in each of the above mentioned methods, selected variables were separately given as input to the network to reach a proper design for the neural network architecture and perform the final simulation. In other words, the artificial neural network model was fitted four times with different input variables. Then, number of neurons and network layers were determined, a suitable weight was assigned to each variable and the neural network was trained to reach the most appropriate architecture for the neural network. Finally, emission scenarios (RCP2.6, RCP4.5, and RCP8.5) were given as input to the selected architecture, and maximum temperature was simulated for 2006-2100 reference period.
 
Results & Discussion
Appropriate explanatory variables were selected in the present study using four different preprocessing methods. Forward selection method with the lowest minimum mean square error (MMSE) of 6.7 and the highest correlation coefficient of 0.97 was selected as the most appropriate method. Therefore, variables obtained from this method, including average temperature near the surface, average pressure at sea level, and altitude at 500 and 850 hPa level, were selected as predictor variables entering the artificial neural network to simulate future temperature of the station. Finally, a neural network with 8 inputs, a hidden layer with 10 neurons and sigmoid transfer function, and an output layer with 1 neuron and Linear transfer function were confirmed using Levenberg-Marquardt algorithm. There were then used to simulate the future temperature of Qazvin synoptic station. The highest and the lowest temperature values were estimated for December with 9.9°C and January with 6.6°C, respectively. The lowest error rate also belonged to December (-1.5°C). Simulation results indicated that the highest increase in maximum temperature of Qazvin synoptic station in 2006-2100 reference period was observed in RCP8.5, RCP4.5 and RCP2.6 scenarios, respectively. The increasing trend in RCP8.5 scenario was estimated much higher than the base temperature. Moreover, the highest temperature increase (6.7°C) in RCP8.5 scenario belongs to June and the highest temperature decrease (3°C) in the optimistic scenario (RCP2.6) belongs to October. 
 
Conclusion
Selecting appropriate explanatory variables is an important step in the process of simulating future temperature. Various methods of variables selection, statistical downscaling and artificial neural network model were used to estimate and simulate temperature parameter. Then, RCP 2.6, RCP4.5, and RCP8.5 scenarios were simulated for the 2006-2100 reference period. Maximum temperature of Qazvin synoptic station in the simulated RCP scenarios (belonging to the reference period) was compared with maximum temperature in 1961-2005 period. Results indicate that the highest temperature increase in Qazvin station occurs in the pessimistic scenario (RCP8.5). The increasing trend of temperature begins with RCP2.6 scenario and reaches its highest level in RCP8.5 scenario. In these three scenarios, summer temperature of the next 94 years may increase at a higher rate as compared to other seasons in Qazvin. This means that not only Iran is located in an arid region, but also its temperature will be increasing in the future. Since temperature and precipitation in different parts of the world are considered to be among the most important indicators of climate change and global warming, various models designed to forecast and simulate these phenomena and the future climate suggest an increase in temperature during the coming decades.