1 Introduction

Changing climate in different regions is due to increasing concentration of greenhouse gases and changing general circulation (Kaleris et al. 2001). Water resources have been facing severe challenges during the last decades all over the world. This condition is intensified by decreasing trend of precipitation and increasing trend of temperatures. It is expected that most of the regions in the world will experience a negative impact of climate change on water resources and freshwater ecosystems (Intergovernmental Panel on Climate Change, IPCC 2007). However, the characteristics of the impact may vary from one region to another. The prediction was done by the most general circulation models (GCMs) that indicated an increase in the number and occurrence of huge climatic variation (i.e., precipitation) in different regions over the world. Such climatic variations can affect the situation of water resources in the near future.

Various hydrological models have been applied for quantifying the impact of hydrologic climate change by using GCMs information. However, the spatial resolution between the output of GCMs and the required data for hydrological models is not sufficient, and it has always been a challenge. For addressing this type of problem, regional climate models (RCMs) and statistical downscaling methods (SDSMs) have been developed. The main problem of RCMs is that they are computationally expensive. So, they are applicable in some special regions. Furthermore, application of RCMs in a small watershed is problematic, and some scenarios defined at watershed scales are necessary.

The SDSM was developed to overcome the aforementioned problems. SDSM plays the role of an intermediate between GCMs or RCMs with other variables (represent the small scale such watersheds). SDSM is computationally cheap and simple runoff model (Srikanthan and McMahon 2001). Long Ashton Research Station Weather Generator (LARS-WG) is classified in the category of SDSM. Some successful applications of LARS-WG were reported by Semenov et al. (1998), Khan et al. (2006), Zarghami et al. (2011), Hassan et al. (2014), and Parajuli et al. (2016).

In recent years, many researchers focused on the impact of climate change on the runoff by using RCMs and SDSM in different scales. Chiew et al. (1995) investigated the impact of climate change on runoff and soil moisture of Australian basins. Matondo et al. (2004) evaluated the impact of climate change on hydrology and water resources in Swaziland. They used 11 GCMs for assessing the climate changes, and three of them were found to simulate the observed precipitation very well in the region. Jiang et al. (2007) compared the hydrological impact of climate change simulated by different hydrological models in the Dongjiang basin, China. Zheng et al. (2007) utilized an RCM for modeling water resources changes. Fujihara et al. (2008) applied dynamically downscaled data for hydrologic modeling and assessing the impact of climate change on water resources of a basin in Turkey. Fatichi et al. (2011) used a weather generator for modeling of future climate scenarios. Chang et al. (2014) investigated the impact of climate change and human activities on runoff in Weihe River Basin in China. Nkomozepi and Chung (2014) studied the impacts and uncertainty associated with climate change on water resources in the Geumho River Basin, Republic of Korea. The results indicated that climate change will lead to lower water resources levels in future compared to current situation. Lu et al. (2016) applied an integrated statistical and data driven framework for flood risk analysis under climate change. The results revealed that the maximum monthly and annual streamflows would both increase at the middle and end of this century. In another study, Vallam and Qin (2016) presented a combined LARS-WG and multi-site rainfall simulator RainSim approach to investigate flow regimes under future conditions in the Kootenay Watershed, Canada.

In the recent years, artificial intelligent (AI) paradigms including artificial neural networks (ANNs), adaptive neuro-fuzzy inference system (ANFIS), support vector machine (SVM), and gene expression programming (GEP) methods have been used in water science and hydrology (Savic et al. 1999; Firat and Gungor 2007; Wang et al. 2009; Nourani et al. 2014; Mehdizadeh et al. 2017). GEP, as a branch of genetic programming (GP), involves computer programs of different sizes and shapes encoded in linear chromosomes of fixed lengths. In the structure of GEP, the chromosomes are simple entities and the expression trees are exclusively the expression of their respective chromosomes (Ferreira 2006). Among AI, the GEP has the capability in presentation of relationship between dependent variable (target) and independent variables (inputs). This advantage is very important in modeling of various phenomena. Describing literature review related to GEP is beyond the scope of this study and only some studies are presented here.

Aytek and Alp (2008) utilized GEP for modeling rainfall-runoff process. Guven and Aytek (2009) introduced GEP for determining the stage-discharge relationship. Hashmi et al. (2011) applied GEP for downscaling of watershed precipitation. The results showed that GEP could be used in the process of precipitation downscaling. Kisi et al. (2013) compared the efficiency of ANNs, ANFIS, and GEP in rainfall-runoff modeling. Hashmi and Shamseldin (2014) used GEP in order to regionalize the flow duration curve. Zorn and Shamseldin (2015) estimated the peak of flood for Auckland region using GEP model. Zounemat-Kermani et al. (2017) applied GEP and decision tree methods for estimation of air demand in bottom outlet conduits of dams.

In this paper, the authors try to investigate the impact of climate change on runoff in two selected case studies from the northwest part of Iran. In the first part of the study, LARS-WG model was applied in order to downscale the climatic variables including temperature, rainfall, and solar radiation. In the second part of the study, the GEP model was used for prediction of runoff based on different input combinations extracted from the climate model.

2 Material and methods

The framework of this study includes two major steps. In the first step, the GCM data for two case studies under different emission scenarios downscaled. In the second step, the impacts of climate change on the runoff using GEP model were assessed. In the following sections, the study area, data, and methods are illustrated.

Fig. 1
figure 1

Location of the study area in Iran. a Overall position. b In respect to Lake Urmia

Full size image

2.1 Case studies and used data

For investigation on the impact of climate change on runoff, two case studies including Aji-Chay and Mahabad-Chay River basins are selected, which are located in the northwest part of Iran. Figure 1 represents the geographical location of the Aji-Chay River and Mahabad-Chay River basins and the gauge stations.

Fig. 2
figure 2

The flowchart of GEP (see Aytek and Kisi 2008)

Full size image

Aji-Chay River basin, as one of the main sub-basin of Lake Urmia, is located between 37° 42′ and 38° 30′ north latitude and 45° 40′–47° 53′ east longitude and has a total area of 12,790 km2. The altitude varies from 1270 to 3400 m. The variables of the daily rainfall, minimum temperatures, maximum temperatures, and sunshine hours were used as input of LARS-WG model. The model automatically converts the sunshine hours to solar radiation.

In Aji-Chay River basin, the climatic variable data obtained from Tabriz synoptic station and the hydrometric data of the Aji-Chay River at the Vanyar station for the period 1982–2010 were used.

Another case study, Mahabad-Chay River basin, the forth large sub-basin of Lake Urmia, is located in the south of Lake Urmia (between 36° 22′ and 37° 10′ N latitudes and between 44° 45′ and 45° 56′ E longitudes). The weather data related to Mahabad synoptic station and runoff data for Koutar hydrometric station was used in this case study. The historical data include a period of 23 years (1988–2010). For both case studies, the selected weather and runoff data have good quality.

Figure 1 shows the overall location of basins in Iran and the location of both stations respect to Lake Urmia. With considering the recent drought over Lake Urmia, the largest lake in the Middle East and the sixth largest saltwater lake on Earth due to climate change, dam construction over the upstream, agricultural consumption, etc., the results of current research can be useful in this matter. The statistical parameters of daily data sets (X) are presented in Table 1. In this table, X mean, X min, X max, S X, C V, and C SX denote the mean, minimum, maximum, standard deviation, coefficient of variation, and skewness coefficient values of each variable, respectively. It is apparent from the table that the daily rainfall and runoff data have highly skewed distributions and high variations for the both basins.

Table 1 Statistical parameters of the daily data sets for the two case studies
Full size table

2.2 Description of LARS-WG

The weather generator models simulate the climatic variables in two steps. In the first step, rainfall event and its intensity were simulated and then, other remaining variables such as minimum and maximum temperature, solar radiation, humidity, and wind speed were simulated (Johnson et al. 1996). LARS-WG as a downscaling model was introduced by Racsko et al. (1991). Daily time series of maximum and minimum temperatures, solar radiation, and precipitation can be generated by LARS-WG. Semenov et al. (2010) stated that the LARS-WG can be applied for observed daily climatic data for a special site in order to compute the parameters of fitted distribution to climatic data.

In the LARS-WG, rainfall modeling and probability of rainfall occurrence are done using semi-empirical method and Markov chain. Furthermore, the modeling of solar radiation and temperate is done based on semi-empirical method and Fourier series, respectively (Semenov et al. 2002; Zarghami et al. 2011).

In order to generate data using LARS-WG, firstly, the characteristics of each station (i.e., the name, geographical information, altitude, and the daily climatic data) should be defined as an inputs of model. Then, the model analyzed the data, and the outputs can be summarized in a text file, which including the statistical characteristics of historical data in the form of mean monthly and seasonal data. The model with the respect to the trend of observed data can be regenerated for a given period, and finally, by comparing the simulated and observed data, the performance of the model can be evaluated. After evaluation of model efficiency, for each station, the data set of future horizons based on climate change scenarios with respect to the output of GCMs can be generated (Semenov et al. 2002).

In this study, the latest version (5.5) of LARS-WG (http://www.rothamsted.ac.uk/mas-models/larswg/download) was applied for downscaling weather data. In this version, different information related to GCMs is provided. In the current study, the Hadley Centre coupled model (HADCM3) is utilized through the downscaling process. One attractive feature of the HADCM3 model is that, unlike the other models, HADCM3 requires no flux adjustments. Such adjustments could potentially impact aspects of climate variability, and so are undesirable (Dong and Sutton 2005). LARS-WG applies different statistical tests for comparing the data generated by the weather generator with historical data during the baseline period. The tests were implemented for the direct generation of synthetic observed data to test how well the LARS-WG is performing in reproducing the characteristics of the observed data. The t test was used for evaluating the monthly means of weather data. For each month, F tests were utilized on the variances of all the daily values for the month across all the year and on the variances of monthly mean values for the different years. The chi-squared test was utilized for comparing distribution of the wet and dry periods (Semenov et al. 1998). The value of “significance level” was chosen as 5%.

Bias and mean absolute error (MAE) were utilized for comparing the simulated data by LARS-WG and observed data (in the baseline period). Bias and MAE relationships are as follows

$$ \mathrm{Bias}=\frac{1}{N}\sum_{i=1}^N\left({S}_i-{O}_i\right) $$
(1)
$$ \mathrm{MAE}=\frac{1}{N}\sum_{i=1}^N\left|{S}_i-{O}_i\right| $$
(2)

where i is the number of month, and S i and O i are the simulated and observed values, respectively. More details about LARS-WG can be found in Semenov and Barrow (1997) and Semenov et al. (1998).

2.3 Impact simulation by gene expression programming

GEP was first introduced by Ferreira in 1999 (Ferreira 2001). GEP, as an extension of GP, is a search paradigm that involves computer programs such as mathematical expressions and decision tress. GEP is a genetic algorithm (GA) that applies populations of individuals and chooses them proportionally to fitness and introduces genetic variation by application of some genetic operators. In GEP, the individuals are encoded as linear strings with fixed length (the genome or chromosomes) which are then presented as nonlinear entities of different sizes and shapes (i.e., pars tress) (Ferreira 2001; Kisi and Sanikhani 2015). Figure 2 schematically presents the fundamental steps included in GEP.

In this study, a powerful soft computing model, GeneXpro Tools 4.0 was used to predict runoff using GEP. Runoff prediction by application of GEP includes some major steps (i.e., select appropriate fitness function, select terminal and function sets, choose chromosomal structure, and select the rate of generic operators). In the first step, fitness function was selected. The fitness (m i ) of an individual program (i) is described by Eq. 3.

$$ {m}_i=\sum_{j=1}^N\left( S-\left|{A}_{(ij)}-{B}_j\right|\right) $$
(3)

where S is the selection range, A (ij) is the value predicted by the individual program (i) for fitness case (j) (from between fitness cases), and B j is the target value for fitness case (j). In this study, root relative squared error (RRSE) and relative absolute error (RAE) were chosen as suitable fitness functions for Aji-Chay and Mahabad-Chay River basins, respectively. In the second step, terminal set (T) and function set (F) are selected. In this paper, the terminal set is including climatic variables (i.e., rainfall, maximum temperature, minimum temperature, solar radiation) in current and prior time steps. The selection of functions set is depending on the nature and complexity of the problem. In this study, different combinations of functions were used including basic arithmetic operators (+,  − ,  × , ÷) and some mathematical functions were as follows: \( \left( \ln x,{e}^x,{x}^2,{x}^3,\sqrt{x},\sqrt[3]{x}, \sin x, \cos x,\mathrm{arc} \tan (x)\right). \)

Selecting the chromosomal architecture is the third step. In the next step, the type of linking function (i.e., addition, subtraction, multiplication, and division) is selected. The final step is choosing the generic operators and their values (Ferreira 2006).

The values allocated to each parameter are summarized as follows: number of chromosomes, 30; head size, 8; number of genes, 3; mutation rate, 0.044; inversion rate, 0.1; one point recombination rate, 0.3; two point recombination rate, 0.3; gene recombination rate, 0.1; gene transposition rate, 0.1; IS transposition rate, 0.1; and RIS transposition, 0.1 (Kisi et al. 2015; Kisi and Sanikhani 2015). It is notable that these values are the default values provided by GeneXpro Tools.

2.4 Input combinations and performance criteria

In this study, different input patterns were considered for the investigation of the impact of climate change on runoff using GEP model. Various input combinations including rainfall (R), minimum temperature (T min), maximum temperature (T max), mean temperature (T mean), and solar radiation (SR) in current (t) and prior time (t − 1) were applied in GEP model. The input combinations are presented in Table 2.

Table 2 Different input combinations for runoff prediction by GEP
Full size table

For runoff prediction in different future horizons (i.e., 2020, 2050, and 2090) over the basins, historical data were used for training (calibration) of the models. Among the available data, 70% of data was used for training and the remaining 30% of data was used for validation of the models. Finally, the runoff prediction in future periods was made by application of the best input combinations during training and validation periods.

The performance of models was assessed with respect to determination coefficient (R 2) and root mean squared error (RMSE) statistics. RMSE can be defined as follows

$$ \mathrm{RMSE}=\sqrt{\frac{1}{N}{\sum}_{i=1}^N{\left({Q}_{\mathrm{o}}-{Q}_{\mathrm{e}}\right)}^2} $$
(4)

where N is the number of data sets, and Q o and Q e are the observed and predicted runoff values, respectively.

3 Application and results

3.1 Aji-Chay River basin

The monthly mean and standard deviation of observed and simulated weather data of Aji-Chay River basin for the period of 1982–2010 are compared in Fig. 3. The MAE and bias values for each data are also provided in Table 3. It is clear from the figure and table that the weather data are well simulated by LARS-WG, especially the minimum temperature, maximum temperature, and solar radiation. For rainfall, there are some deviations and significant differences in some months (e.g., May, July, October, and December). It is also clear from the Table 3 that the rainfall has the highest MAE (2.078) and bias (−0.212) values. This may be due to the fact that the strength of the correlation between successive values varies considerably for the rainfall data, while the auto-correlation of the temperature (or solar radiation) data is very strong. Also, highly skewed distribution of rainfall data may be another reason of this. From Fig. 3c, it is clear that the maximum difference between the observed and simulated rainfall is 7.5 mm in May.

Fig. 3
figure 3

Monthly mean and standard deviation of observed and simulated weather data (Aji-Chay River basin). a Minimum temperature. b Maximum temperature. c Rainfall. d Solar radiation

Full size image
Table 3 Bias and MAE values for weather data in Aji-Chay River basin
Full size table

The weather data were predicted by using three different scenarios of LARS-WG climate model such as B1, A1B, and A2 for the period of 2011–2030. The relative change in rainfall and absolute change in temperature (comparing to historical period) are shown in Fig. 4 for the Aji-Chay River basin. Solar radiation is not provided in this figure, because the variation of solar radiation is approximately linear and there is not have any increasing or decreasing trend. The relative change in solar radiation by application of different scenarios (i.e., A1B, A2 and B1) is about 0.97–1.01. From Fig. 4a, it is clear that the rainfall generally shows decreasing trend especially in the summer months in which more water is required by the region. This lacking in rainfall will cause water stress and reduce the per capita water access. Temperature is another important parameter which significantly affects droughts. From Fig. 4b, it is clear that the temperature shows significantly increasing trend (0.8 °C) in summer months. Generally, according to Fig. 4, the Aji-Chay River basin will have less rain and hotter summers in the period of 2011–2030. However, there is an increasing trend in rainfall and decreasing trend in temperature for the month of September, October, and December.

Fig. 4
figure 4

Relative change in rainfall and absolute change in temperature during 2011–2030 in respect to the baseline in Aji-Chay River basin using different scenarios (B1,A1B and A2) a) rainfall b) temperature

Full size image

Different GEP functions were tried in the study. The results of validation of used GEP functions are provided in Table 4. It is apparent from the table that the optimal GEP model was obtained by functions set of F5, which is default set in GeneXpro program and addition linking function. RRSE is found as the best fitness function in all GEP applications. Kisi et al. (2013) used GEP models in modeling rainfall-runoff process and they also found the RRSE as the best fitness function. The accuracy of the GEP models for each input combination (Table 2) is given in Table 5 for the both training and validation periods. According to Table 5, it is clear that the GEP model comprising inputs of R(t), R(t − 1), T max(t), T max(t − 1), T min(t), T min(t − 1), SR(t), and SR(t − 1) shows the best accuracy. Using only rainfall or together with temperature data gives poor accuracy in runoff prediction with respect to R 2 and RMSE. It is apparent from Table 5 that adding previous values of rainfall, temperature, and solar radiation data as inputs to the GEP model considerably improves the prediction accuracy. The relative RMSE and R 2 differences between the best (6th inputs) and the worst (1st input) input combinations are 52 and 98% in the validation period, respectively.

Table 4 Validation results of different GEP basic and linking functions for the parse tree (Aji-Chay River basin)
Full size table
Table 5 The performance criteria of GEP models using different input combinations in runoff simulation (Aji-Chay River basin)
Full size table

The equation of the optimal GEP model including sixth input combination for the Aji-Chay River basin is

$$ Q(t)= A+\frac{\sqrt{R\left( t-1\right)}}{ \sin \sqrt{\mathrm{SR}\left( t-1\right)}\sqrt{T_{\max}\left( t-1\right)-{T}_{\min }(t)}} $$
(5)
$$ A= B+\frac{\sqrt{R(t)} A \tan \sqrt{T_{\max }(t)}}{T_{\min}\left( t-1\right)-\mathrm{SR}\left( t-1\right)+\mathrm{SR}(t)} $$
(5a)
$$ B=\sqrt{R\left( t-1\right)+\frac{R\left( t-1\right)}{\mathrm{Ln}\ \left({T}_{\max}\left( t-1\right)\right)}} $$
(5b)

The expression tree of the optimal GEP model for Aji-Chay River basin is shown in Fig. 5. Figure 6 compares the observed and simulated runoff hydrographs in different future horizons for the Aji-Chay River basin. The hydrographs were predicted with respect to scenario A2 which assumes that the current socioeconomic situation will continue. Scenario A2 is better and more conservative than the A1B and B1. It is notable that the scenario B1 describes a convergent world with global population that peaks will occur in midcentury and decline thereafter. Also, A1B, as a sub-set of scenario A1, emphasizes on a balance on all energy sources (Intergovernmental Panel on Climate Change, IPCC 2007). It is clear from the Fig. 6 that the runoff mean and hydrograph peak significantly decrease by increasing time horizons. Table 6 gives the percent decrement in mean runoff and hydrograph peak for the future horizons of 2011–2030, 2046–2065, and 2080–2099. It can be seen that the decrements in mean and peak discharges will be as high as 58.7 and 50% at the end of this century. This result is parallel with the increasing trend in temperature and decreasing trend in rainfall shown in Fig. 4. These results imply that new and sustainable adaptation strategies are needed to rescue the water supply in this region in the future.

Fig. 5
figure 5

Expression tree obtained by GEP model (Aji-Chay River basin)

Full size image
Fig. 6
figure 6

The comparison of observed and simulated runoff hydrographs in different future horizons in Aji-Chay River basin

Full size image
Table 6 The percent of relative decrease of runoff in different future horizons comparing to baseline period (Aji-Chay River basin)
Full size table

3.2 Mahabad-Chay River basin

Figure 7 illustrates the monthly mean and standard deviation of observed and simulated weather data of Mahabad-Chay River basin for the period of 1988–2010. The MAE and bias values of each data are given in Table 7. Similar to the Aji-Chay River basin, here also, the weather data are well simulated by LARS-WG, especially the minimum temperature, maximum temperature, and solar radiation. Here also, some deviations and significant differences are clearly seen for rainfall in some months (e.g., January, February, March, and November). It is clear from Fig. 7 that the maximum difference between the observed and simulated rainfall is 8.3 mm in January. As seen from Table 7, the rainfall has the highest MAE (4.473) and bias (2.425) values.

Fig. 7
figure 7

Monthly mean and standard deviation of observed and simulated weather data (Mahabad-Chay River basin). a Minimum temperature. b Maximum temperature. c Rainfall. d Solar radiation

Full size image
Table 7 Bias and MAE values for weather data in Mahabad-Chay River basin
Full size table

The predictions of the rainfall and temperature data by three different scenarios of LARS-WG climate model such as B1, A1B, and A2 for the period of 2011–2030 are shown in Fig. 8. The figure indicates the relative change in rainfall and absolute change in temperature (comparing to historical period) for the Mahabad-Chay River basin. Solar radiation is not provided in this figure because the relative change in solar radiation by application of different scenarios (i.e., A1B, A2, and B1) is about 0.98–1.01. As can be seen from Fig. 8a, rainfall generally shows decreasing trend especially in the summer months alike to previous basin (Aji-Chay). From Fig. 8b, it is clear that the temperature shows significantly increasing trend in summer months. It is clearly seen from the Fig. 8 that the Mahabad-Chay River basin will have less rain and hotter summers in the period of 2011–2030 but not as much as Aji-Chay River basin. The validation results of different GEP functions are given in Table 8. From the table, it is clear that the optimal GEP model was obtained by function set of F3 and addition linking function. In Mahabad-Chay River basin, relative absolute error (RAE) is found as the best fitness function in all GEP applications. Table 9 gives the accuracy of the optimal GEP models for each input combination in the training and validation period. It is clear from the table that the GEP model comprising sixth input combination shows the best accuracy similar to the Aji-Chay River basin. Here also, using only rainfall and temperature data provides poor estimates for runoff prediction and adding previous values of rainfall, temperature, and solar radiation data as inputs to the GEP model considerably improves the prediction accuracy. The relative RMSE and R 2 differences between the best (6th inputs) and the worst (1st input) input combinations are 46 and 70% in the validation period, respectively.

Fig. 8
figure 8

Relative change in rainfall and absolute change in temperature during 2010–2030 in respect to the baseline in Mahabad-Chay River basin using different scenarios (B1, A1B and A2) a) rainfall b) temperature

Full size image
Table 8 Validation results of different GEP basic and linking functions for the parse tree (Mahabad-Chay River basin)
Full size table
Table 9 The performance criteria of GEP model using different input combination in runoff simulation (Mahabad-Chay River basin)
Full size table

The equation of the optimal GEP model in predicting runoff of Mahabad-Chay River basin is

$$ Q(t)=\left({T}_{\max }(t)-{T}_{\max}\left( t-1\right)\right)+\frac{T_{\max}\left( t-1\right)}{\mathrm{SR}(t)}+\frac{\mathrm{SR}(t)\left[ R(t)-\mathrm{SR}\left( t-1\right)\right]}{26.5+{T}_{\min}\left( t-1\right)} $$
(6)

The expression tree of the optimal GEP model for Mahabad-Chay River basin is shown in Fig. 9. The observed and simulated runoff hydrographs in different future horizons are compared in Fig. 10. Alike to the Aji-Chay River basin, here also, the mean runoff and hydrograph peak significantly decrease by increasing time horizons. The percent decrement in mean runoff and hydrograph peak is given in Table 10 for the future horizons of 2011–2030, 2046–2065, and 2080–2099. It is apparent from the table that the decrements in mean and peak discharges will be as high as 60.5 and 55.9% at the end of this century. This result confirms the increasing trend in temperature and decreasing trend in rainfall as shown in Fig. 8.

Fig. 9
figure 9

Expression tree obtained by GEP model (Mahabad-Chay River basin)

Full size image
Fig. 10
figure 10

The comparison of observed and simulated runoff hydrographs in different future horizons in Mahabad-Chay River basin

Full size image
Table 10 The percent of relative decrease of runoff in different future horizon comparing to baseline period (Mahabad-Chay River basin)
Full size table

4 Conclusion

In the current paper, the impact of climate change on runoff was investigated for two basins in northwest part of Iran. LARS-WG was utilized for downscaling the climatic variables including rainfall, solar radiation, and maximum and minimum temperatures. The results showed that LARS-WG could be successfully applied for downscaling of climatic variables. For both basins, the generated climatic variables in future horizon (2020) were compared with baseline period. The results indicated a decreasing rainfall trend, especially in the summer months in which more water was needed in the region. Furthermore, an increasing trend was found for temperature over the basins. In another part of the study, the GEP model was applied for prediction of runoff in the future horizons (2020, 2055, and 2090) by using A2 scenario. The GEP model whose inputs comprise the rainfall, solar radiation, and minimum and maximum temperatures in current and prior time was found to be the best one for runoff prediction. For Aji-Chay River basin, the mean and peak discharge were decreased by 58.7 and 50% in the 2090 horizon, respectively. For the Mahabad-Chay River basin, the decrements in mean and peak discharge will be as high as 60.5 and 55.9% at the end of this century, respectively. With respect to dramatic decrease of runoff in the future, it is necessary to find a solution to prevent or minimize the negative impact of climate change on runoff and the role of managers and decision makers in this matter is very essential in the region.