Extended Abstract
Introduction
Temperature is a significant atmospheric element that manifests climate change, specifically global warming resulting from an increase in greenhouse gas concentration. Atmospheric simulation is a critical tool in studying changes in atmospheric-climatic elements, particularly temperature.
The most commonly used tool for simulating the responses of the climate to greenhouse gas increases and examining future temperature changes is the use of climate variables simulated by coupled atmosphere-ocean models (AOGCMs). General circulation models (GCMs) are powerful tools aimed at generating climate scenarios. However, GCMs cannot provide effective information on climate simulation at local and regional scales. Therefore, the downscaling method is used to bridge the gap between local and global scales.
The current research aims to simulate maximum temperature using an artificial neural network model that adopts data from the atmospheric general circulation model (HadCM3) under RCP8.5, RCP4.5, and RCP2.6 scenarios for the Yazd synoptic station from 2006 to 2095. The independent variable, as the input to the artificial neural network, was selected for statistical downscaling using four statistical criteria: Percentile Reduction, Backward Variable Elimination, Forward Variable Selection, and Stepwise Variable Entry. Finally, the maximum temperature of the Yazd synoptic station for the next century was simulated.
Data and Methodology
The present study aims to investigate the maximum temperature of Yazd's synoptic station in the context of climate change based on valid scenarios until 2095. To achieve this, three sets of data were used: average daily maximum temperature data from Yazd's synoptic station, observed atmospheric data for the period of 1961 to 2005 (NCEP data), and simulated data from 2006 to 2095 based on release RCP scenarios. The NCEP data from 1961 to 2005 included 26 atmospheric variables that will be used as independent or predictor variables.
Modeling, simulating, and forecasting temperature based on nonlinear and chaotic time series is a challenging task. Prior studies have shown that artificial neural networks (ANNs) are suitable for simulating and predicting basic processes that are not well known. It is crucial to select the correct input variables intelligently and according to the purpose of the artificial neural network's design for prediction and simulation. Accordingly, in this study, the most suitable atmospheric parameters as the input of the artificial neural network were selected by pre-processing and selecting the atmospheric variables for the base period (1961-2005) to simulate with four statistical criteria (Percentile Reduction, Backward Variable Elimination, Forward Variable Selection, and Stepwise Variable Entry). The resulting mean square error (MSE) obtained from the statistical criteria was compared, and the correlation coefficient and the similarity of the monthly time series trend of the simulated values with the target values were also analyzed. The best network architecture was selected to simulate the maximum temperature of Yazd's synoptic station from 2006 to 2095 under different RCP emission scenarios.
Discussion
The selection of explanatory variables for downscaling was based on four statistical methods: Percentile Reduction, Backward Variable Elimination, Forward Variable Selection, and Stepwise Variable Entry. After analyzing the mean square error (MSE), correlation coefficient, monthly average values of the maximum temperature of Yazd station, and estimated values from 1961 to 2005, the probability density function, cumulative probability function, and monthly time series trend obtained from all four methods, the explanatory variables were selected. These variables include mean sea level pressure, the divergence of 1000 hPa, zonal wind component, zonal wind intensity of 850 and 500 hPa, altitude and vorticity of 500 hPa, average temperature, and relative humidity at a 2 m height.
The structure and architecture of the neural network were designed based on these selected variables. The network consisted of a two-layer feedforward, with a sigmoid transfer function in the hidden layer, a linear function in the output layer, an input layer with eight variables, eight neurons, and the Lunberg-Marquardt training algorithm. This architecture was used to simulate the maximum temperature of Yazd's synoptic station under RCP2.6, RCP4.5, and RCP8.5 scenarios for two periods of 2050-2006 and 2095-2051.
Comparing the monthly average values of RCPs (RCP2.6, RCP4.5, and RCP8.5) in the first statistical period (2050-2006) with the base period (1961-2005), the maximum temperature of Yazd station indicates an increase in temperature in winter, spring, and summer, and a decrease in the autumn season under all three RCPs.
Comparing the monthly mean values of RCPs (RCP2.6, RCP4.5, and RCP8.5) of the second period (2051-1995) with measured mean maximum temperature (2005-1961) showed that temperature will increase the most in winter, spring, and summer, similar to the first period of the RCP8.5 scenario. In this scenario, unlike the other scenarios, the increase in temperature is evident in both subperiods for the autumn season. Finally, in the second period (2051-1995), the increase in the average maximum temperature of Yazd station in winter, spring, and summer, and the decrease in the average maximum temperature in autumn will be more significant.
Conclusion
The increase in greenhouse gas concentration resulting from human industrial activities is expected to cause global and regional warming in the future. The current study's findings indicate that the average maximum temperature of Yazd station will rise between 0.4 to 6.9  in winter, 0.2 to 8.1  in spring, and 1.1 to 7.7 in summer from 2006 to 2095. However, a decrease in the maximum temperature between 0.6 and 1.4  is expected in autumn. These results are consistent with those of other researchers.