The Analysis of Uncertainty of Climate Change by Means of SDSM Model Case Study: Kermanshah

Transcription

1 World Applied Sciences Journal 23 (1): , 213 ISSN IDOSI Publications, 213 DOI: /idosi.wasj The Analysis of Uncertainty of Climate Change by Means of SDSM Model Case Study: Kermanshah Ahmad Rajabi and Saeid Shabanlou Department of Water Engineering, Islamic Azad University, Kermanshah Branch, Kermanshah, Iran Submitted: Jun 12, 213; Accepted: Jul 21, 213; Published: Jul 28, 213 Abstract: SDSM downscaling model was used as a tool for downscaling weather data statistically. The prediction of climate change model was carried out according to 3 scenarios, A2, B2 and A1 based on 2 global circulation models, HadCM3 and CGCM1. In order to investigate uncertainty, bootstrap confidence intervals were calculated for estimated means and variances of daily weather parameters in every month. The predictor variables of CGCM1 model did not create an acceptable correlation with precipitation; therefore, precipitation data were downscaled only by HadCM3. The uncertainty of the means and variances in the minimum and maximum daily temperature of CGM1 model were less than HadCM3 and were closer to the observed data. The daily precipitation of the observed data in most months has less uncertainty than HadCM3. The precipitation in HadCM3 will increase 54% in A2 scenario and 51.5% in B2 scenario in time period. The minimum and maximum daily temperature in all scenarios will increase in time period. Key words: SDSM Downscaling Uncertainty Bootstrap Climate change Kermanshah (Iran) INTRODUCTION parameters such as precipitation and temperature in a long period with regard to the weather large-scale signals. Climate change is spreading slowly all over the world Since in statistical downscaling model, building and affecting the related parameters and it is predicted meteorological data is performed by combining the two that the process will continue in the future. According to probabilistic and regression methods, in classifying Global Circulation Models (GCM), which simulate the different downscaling models, statistical downscaling climate of the earth, the temperature of the earth will model is among the best models [4]. The prediction of increase about 1 to 3.5 C up to 21. The effects of climate climate change model was carried out according to 3 change on hydrological cycle are in the form of changes scenarios, A2, B2, A1, among the presented scenarios in in water level in ground waters, aquifers, lakes and also SRES [5]. change in the amount and temporal distribution of Lane et al., (1999) predicted in an analysis that precipitation and rivers runoff [1]. In IPCC s fourth climate change can increase the runoff in regions with assessment report (AR4), climate experiences in several high latitude due to the increase in precipitation and snow GCMs and different circulation scenarios were run melting. But in regions with low latitudes it is expected according to many data sets of climate change projects in that the runoff decreases [6]. future by 18 global modeling groups [2]. Using Khan et al., (26) evaluated the uncertainties downscaling techniques, GCM outputs can be changed seen in the downscaled results of daily precipitation into surface variables at the scale of the basin under and minimum and maximum daily temperatures study. Downscaling, in general, is defined as a which were gained from three downscaling models, relationship creator factor between large-scale cycles namely SDSM statistical downscaling model, (predictors) and the climate variables at the local scale LARS-WG and ANN Artificial Neural Network. The (predictands) [3]. results of the evaluation of uncertainty indicates SDSM (Statistical DownScaling Model) has been that SDSM was capable of maintaining the different developed by Wilby et al., (22) as a tool for statistical characteristics of observed data at the 95% downscaling statistically. The basis of this model is confidence level for downscaled data better than other multiple regression and is used to predict climate models [7]. Corresponding Author: Ahmad Rajabi, Department of Water Engineering, Islamic Azad University, Kermanshah Branch, Kermanshah, Iran. 1392

2 Hreiche et al., (27) investigated the climate change conditional process; it means that the precipitation effects on water resources of Lebanon catchments. amount will have correlation with the wet days. These wet They concluded that by using MEDOR rainfall runoff days have correlation with weather predictors at the local model and climate models, for every 2 increase in the scale. Temperature is modeled in the form of an environment temperature, the maximum discharge flow unconditional process. In such situation, a direct relation would happen 2 months sooner and that the droughts is considered between the predictors and the also would happen 15 days to a month sooner [8]. predictands [3]. Semenov (28) studied the ability of Lars-WG weather generator model for simulating the extreme weather events Evaluation of Uncertainty in Downscaled Data: Efron using 2 stations data all over the world with different presented bootstrap method for estimating the accuracy climates. The 95% confidence intervals (CI95) were amount and sample distribution of the statistics in calculated for the statistics under study. The maximum It is because of the idea of resampling of the data that this yearly means and the amounts of the return period of the method has developed by using computers in recent synthetic daily precipitation were placed in the 95% years. If x 1,,x n, iid are from F and are estimated by ˆF, by confidence intervals (CI95) of the observed data. But the replacing ˆF by F of the bootstrap estimator, the variance maximum daily temperature data were produced with a less of T n is calculated as follows: accuracy and for about half of the stations the synthetic data means were placed out of the 95% confidence 2 var [ Tn( X1,..., Xn) X1,..., Xn = E [ Tn E ( Tn)] interval [9]. MATERIALS AND METHODS n n ˆ 2 = [ T( ) ( ) ( )] ˆ n x Tn yd Fyi d Fx ( i), i= 1 i= 1 Data: There are 23 meteorological stations in the region * * in which {X 1,...,X n is a iid sample from ˆF and is called * under study. Regarding the fact that weather parameters the bootstrap sample and var [. X 1,...,X n] is a conditional of the base period in were calculated by variance with x 1,,x ncondition. circulation scenarios, the data of the stations which The above equations are usually complicated and * * possess daily data in this period are usable. Thus, the var (T n ) is not an explicit function from x 1,,x ntherefore, only station which posses enough data in this period is Monte Carlo method is used for approximation. Thus, Kermanshah synoptic station which is regarded as the bootstrap method is a combination of two methods: index station. The data of the predictors were received Substitution Principle and Monte Carlo approximation. * * from Canadian Climate Change Impacts Scenarios Group When var (T n ) is an explicit function of x 1,,x n, bootstrap (CCISG) site and all the data are normalized. equates exactly with classical substitution principle, otherwise bootstrap uses Monte Carlo approximation for Stochastic Downscaling Model (SDSM): SDSM is a approximation [1]. statistical weather generator which is used for simulating Confidence intervals present some information about the weather data in a given station in the current and the uncertainty of estimating mean and variance in future conditions under the effect of climate change. Its estimating means and variances. The present study makes data are in the form of daily time series for some weather use of the most common nonparametric method, bootstrap variables such as precipitation (mm), minimum and for finding the confidence intervals of the means and maximum temperature ( C) and other weather parameters. variances. The idea of bootstrapping is to resample a large The parameters of the regression equation are estimated number of new data sets with replacement from the by dual simplex algorithm. Suitable large-scale predictors original data set. This method starts with a sample size of are selected by correlation analyses, partial correlation n and its algorithm is as follows: and also by regarding the physical sensitivity between the predictors and the predictands, in the catchment Creating a new sample of size n with replacement under study. from the original sample. This model is produced in the form of a monthly Calculating the mean or variance of the new sample model for daily precipitation data which is produced in which is called m regression equations for the 12 months of the year. In Repeating steps 1 and 2, 1 times and the ith new this model, precipitation is modeled in the form of a sample mean or variance is called m. i 1393

3 Plotting the distribution of these 1 sample means or variances. Producing the 95% confidence interval for the mean or variances by finding the 2.5th and 97.5th percentiles of this produced distribution. where z is the a quantile of standard normal distribution, z and a, namely bias-correction and acceleration, are two parameters to be estimated. RESULTS AND DISCUSSION Bootstrap confidence intervals are calculated for the Prediction of Precipitation and Temperature by SDSM: daily precipitation and daily minimum and maximum In order to select the most effective predictor variables temperature estimated means and variances in every among the 26 effective predictor variables in the area, the month. For instance, if the daily precipitation data length correlation between the predictors and predictand was in January is 4 years, an observed sample of size =124 calculated for time interval. The variables of (4 years observed daily precipitation data of January CGCM1 did not create an acceptable correlation with the gives 31*4=124) has been used. Then 1 new precipitation of the site. Therefore, in order to downscale samples, each of the same size as the observed data, are the precipitation data just HadCM3 was used. Regarding drawn with replacement from the observed data. The 1 the correlation coefficients and P-Value, the 3 predictor resamples are drawn because this is the recommended variables (Table 1) among the 26 predictor variables minimum for estimating percentiles, required for estimating have the most correlation coefficient with precipitation. confidence interval for the means and variances [7]. The time series of simulated daily precipitation data were The mean is first calculated using the observed data carried out according to the selected weather predictor and then recalculated using each of the new samples, variables regarding HadCM3. The means and standard yielding a bootstrap distribution of the statistics of mean. deviation of the observed and simulated monthly From this distribution, the bias-corrected and accelerated precipitation are presented in Table 2. (BCa) percentiles are estimated. The BCa percentile is The amounts of direct and partial correlation between more accurate than the empirical percentile. The empirical the meteorological variables and daily temperature of percentiles are simply the percentiles of the empirical CGCM1 are presented in Table 3 and of HadCM3 in distribution of the replicates while the BCa method Table 4. Regarding the fact that the correlation amounts transforms the specified probability values to determine are investigated at daily scale and in a 4-year time which percentile of the empirical distribution most interval, the point that the calculated correlation amounts accurately estimate the percentiles of interest and then are small, is acceptable. Two 5 hpa geopotential height applying corrections for bias and standard error, BCa (P5) and the mean temperature (Temp) variables from confidence interval is estimated as follows [1]: CGCM1 have an acceptable correlation with daily temperature. These variables which have acceptable th. percentile =.975 z + z P-Value, were considered as daily temperature predictor ( z + ) a( z + az ) variables in CGCM1. In HadCM3, regarding the correlation coefficients and the calculated P-Value, these predictor variables have the most correlation coefficients 2.5 th. percentile = ( z + 1 ( + ) ) with daily temperature: mean sea level pressure (MSLP),.25 z + z.25 a z az Table 1: Amounts of the predictor variables correlation of HadCM3 with daily precipitation Predictor Variable Direct Correlation Partial Correlation P value 1 Vorticity p z hpa vorticity p8_z Relative Specific humidity at 5 hpa height r Table 2: Mean and standard deviation of observed and simulated monthly precipitation by HadCM3 (mm) JUN JUL ANN Mean S.D Simulated Mean S.D

4 Table 3: Amounts of the predictor variables correlation of CGCM1 with daily temperature Predictor Variable Direct Correlation Partial Correlation P value 1 p Temp Table 4: Amounts of the predictor variables correlation of HadCM3 with daily temperature Predictor Variable Direct Correlation Partial Correlation P value 1 Mslp p Temp Table 5: Mean and standard deviation of observed and simulated minimum and maximum temperature ( C) JUN JUL ANN Min. T. Mean S.D Max. T. Mean S.D CGCM11 Min. T. Mean S.D Max. T. Mean S.D HadCM3 Min. T. Mean S.D Max. T. Mean S.D hpa geopotential height (P5) and mean temperature also, CGCM1 had less uncertainty than HadCM3 in (Temp). These variables were considered as daily different months and were closer to the observed data. temperature predictors. The mean and the standard deviation of observed and simulated monthly Climate Change Effect on Weather Parameters: The precipitation are presented in Table 5. global circulation models under investigation were HadCM3 with A2 and B2 scenarios and CGCM1 with Bootstrap Confidence Intervals in Means and Variances A1 scenario. According to this, precipitation and minimum Estimates: The 95% confidence intervals were determined and maximum daily temperature parameters were for every month. Their confidence intervals are presented calculated in time period. The estimated in Fig. 1. The uncertainty of precipitation in HadCM3 was monthly precipitation mean in the present time and future almost similar to the observed data. But, totally, the is presented in Fig. 3. Besides, the precipitation changes uncertainty of CGCM1 was closer to the observed data. compared to the base period can be observed in Table 6. About the minimum daily temperature, also, the As it was mentioned above, the predictor variables of uncertainty in HadCM3 was more than the observed data CGCM1 did not have any acceptable correlation with in most months. But, the uncertainty of CGCM1 was precipitation in the region under study and it is because closer to the observed uncertainty in each month. of this fact that precipitation was not calculated by this Bootstrap confidence intervals for the variances of model. every month were determined too and their confidence In both HadCM3 scenarios, precipitation will increase intervals are presented in Fig. 2. The observed data in most months of the year. Precipitation in this model will precipitation had less uncertainty than HadCM3, except increase 54% in time period in A2 scenario, in dry months of the year. About the minimum daily but will not change considerably in summer months. temperature, also, the observed data had less uncertainty But in B2 scenario precipitation will increase 51.5% and than the models under investigation. The uncertainty of the precipitation change will be significant in summer; for the dry months of the year in CGCM1 was closer to the instance, the precipitation in June and July will be 5.8 observed data. About the maximum daily temperature, and 24 mm respectively. 1395

5 Table 6: Monthly precipitation change of different climate change models (GCM) in in Kermanshah GCM JUN JUL ANN HADCM3-A2 Change Percentage HADCM3-B2 Change Percentage C.I. (mm) Daily Precipitation (HadCM3) C.I.(mm2) Daily Precipitation SDSM(HadCM3) C.I.( C) Daily Min. Temperature Obsrev ed (CGCM1) (HadCM3) Daily Min. Temperature SDSM(CGCM1) SDSM(HadCM3) Daily Max. Temperature (CGCM1) (HadCM3) C.I.( C2) C.I.( C) Daily Max. Temperature SDSM(CGCM1) SDSM(HadCM3).5 C.I.( C 2 ) Fig. 1: Bootstrap 95% confidence intervals estimates of the weather parameters mean The estimated minimum monthly temperature in the present time and future is presented in Fig. 4. Besides, the minimum temperature changes compared to the base period are presented in Table 7. The minimum yearly temperature mean in both HadCM3 scenarios, will increase about.2 C. The minimum temperature will decrease in summer and autumn and increase in winter and spring. In CGCM1 the minimum yearly temperature will increase about.6 C which will occur in summer and autumn months. The estimated maximum monthly temperature in the present and future is presented in Fig. 5. Its changes compared with the base period are presented in Table 8. Fig. 2: Bootstrap 95% confidence intervals estimates of the weather parameters variance Precipitation 12 Current Simulated 1 HADCM3-A2 8 HADCM3-B Fig. 3: Estimated monthly precipitation mean of different models under investigation 1396

6 Table 7: Minimum monthly temperature change of different climate change models (GCM) in in Kermanshah GCM ANN HADCM3-A HADCM3-B CGCM1-A Table 8: Maximum monthly temperature change of different climate change models (GCM) in in Kermanshah GCM JUN JUL ANN HADCM3-A HADCM3-B CGCM1-A Min. Temperature CONCLUSION 15 The present study reports an investigation of climate change effect by SDSM abd by A2, B2 and 1 A1 scenarios, according to 2 GCMs: HadCM3 and CGCM1 on daily weather parameters. The results are 5 Current Simulated HADCM3-A2 presented for time period. Regarding HADCM3-B2 bootstrap confidence internals and the calculated -5 CGCM-A1 precipitation variance, the uncertainty of HadCM3 was closer to the observed data, although, in general the observed data had less uncertainty. About the minimum and maximum temperature, also, the Fig. 4: Estimated minimum monthly temperature mean of uncertainty in CGCM1 was less than HadCM3 and different models under investigation was closer to the observed data. By calculating Max. Temperature weather parameters for time period, 5 precipitation increase in most months of the year, 45 4 which will be more significant in dry months of 35 the year. The minimum temperature will decrease in 3 summer and autumn and will increase in winter and 25 Current Simulated spring. The maximum temperature will increase in 2 HADCM3-A2 15 most months, especially in summer and autumn HADCM3-B2 1 CGCM-A1 months. The maximum temperature will have an 5 insignificant increase or decrease in winter and spring. ACKNOWLEDGEMENTS Fig. 5: Estimated maximum monthly temperature mean of different scenarios of the models under This study was supported by Islamic Azad investigation University, Kermanshah Branch, Kermanshah, Iran. The maximum yearly temperature mean in A2 REFERENCES scenario of HadCM3 will increase about 1.8 C and 1.9 C in B2 scenario. This temperature increase will not be 1. IPCC, 2. Climate Change: Impacts, Adaptation significant in April and May. It will vary between 1 to and Vulnerability. Available at: C. In CGCM1 the minimum yearly temperature will pub/ tar/ wg2/index.htm. increase about.6 C which will occur mostly in summer 2. Semenov, M.A. and P. Stratonovitch, 21. The use and autumn. In winter and spring the maximum of multi-model ensembles from global climate models temperature will not have a significant increase or for impact assessments of climate change. Climate decrease. Research, 41:

7 3. Wilby, R.L. and C.W. Dawson, 24. Using SDSM 7. Khan, M.S., P. Coulibaly and Y. Dibike, 26. Version 3.1- A decision suport tool for the Uncertainty Analysis of Statistical Downscaling assessment of regional climate change impacts, User Methods. Journal of Hydrology, 319: Manual. 8. Hreiche, A., W. Najem and C. Bocquillon, Wilby, R.L., C.W. Dawson and E.M. Barrow, 22. Hydrological Impact Simulation of Climate Change SDSM- A Decision SuportTool for the Assessment on Lebanese Coastal Rivers, IAHS Pub., of Regional Climate Change Impacts. Journal of 52(6): Environmental Modeling and Software, 17: Semenov, M.A., 28. Simulation of extreme weather 5. IPCC., 27. Intergovernmental Panel on Climate events by a stochastic weather generator. Climate Change Fourth Assessment Report Climate Change: Research, 35: Synthesis Report. Available at: DiCiccio, T.J. and B. Efron, Bootstrap ipccreports/ar4-syr.htm. confidence intervals (with discussion). Stat. Sci., 6. Lane, M.E., P.H. Kirshen and R.M. Vogel, : Indicators of impact of global climate change on U.S. Water resources. ASCE, Journal of Water Resources Planning and Management, 125(4):