Assessment of climate change impact on floods using weather generator and continuous rainfall-runoff model

Authors

  • Mohammad Reza Khazaei,

    Corresponding author
    1. School of Civil Engineering, Iran University of Science and Technology, Tehran, Iran
    2. Department of Civil Engineering, Kermanshah Branch, Islamic Azad University, Kermanshah, Iran
    • College of Civil Engineering, Iran University of Science and Technology, Tehran, Iran.
    Search for more papers by this author
  • Bagher Zahabiyoun,

    1. School of Civil Engineering, Iran University of Science and Technology, Tehran, Iran
    Search for more papers by this author
  • Bahram Saghafian

    1. Technical and Engineering Department, Science and Research Branch, Islamic Azad University, Tehran, Iran
    Search for more papers by this author

Abstract

One of the potential impacts of climate change (CC) is the change on flood frequency and magnitude. Assessment of impact on floods involves many uncertainties. Difficulty in preparing reliable continuous time series of climate variables with fine time step and capturing extreme rainfalls for future climate scenarios has led to limited studies aimed at investigation of the change of flood regime due to CC. Weather Generators (WGs) have the potential for downscaling of GCMs (General Climate Models) outputs. However, WGs inabilities in relation to producing extreme rainfalls and low-frequency variabilities have caused WGs to receive limited attention in flood studies. In this study, a new WG model, that removes the aforementioned WGs' inabilities, is coupled with a continuous daily rainfall-runoff model to assess the CC impacts on floods of a basin in Iran. Changes in different characteristics of climate variables are applied in the downscaling procedure. Results indicated that, aside from the large uncertainty of emission scenarios, the climate change will lead to considerable increase in flood magnitude in the study basin. Copyright © 2011 Royal Meteorological Society

1. Introduction

Recent human activities have increased concentration of greenhouse gases and, consequently, raised average temperature of the Earth's surface. Scenarios that are published by the Intergovernmental Panel on Climate Change (IPCC) indicate that temperature of the Earth will increase by 1.4 to 5.8 °C in year 2100 in comparison to year 1990 (IPCC, 2001). This increase will also influence climate variables and will result in climate change (IPCC, 1995). By changing climate variables including temperature, precipitation, and evapotranspiration, the hydrologic regime of rivers, and consequently flood frequency and magnitude, will also change (IPCC, 2001). For active adaptation strategy, it is necessary to assess the impacts of climate change on floods.

In studies involving climate change impact assessment on floods, present climate data and future climate scenarios are generally used as inputs to rainfall-runoff models, while present condition and future scenarios are simulated. Such studies comprises two main steps: preparing future climate data and hydrological simulation of streamflow. The most common tool for simulation of global climate response to concentration of greenhouse gases are GCMs. However, since spatial and temporal resolution of GCM outputs are coarse, it is necessary that GCM outputs be downscaled (Wilby, 2007).

For hydrological simulation of flood flows, in some studies (e.g. Roy et al., 2001; Muzik, 2002), single-event rainfall-runoff models are used. Climate inputs of single event models are mainly rainfall events. In most of the past flood studies, the reason for using single-event models was the difficulty in obtaining reliable continuous climate scenarios (Roy et al., 2001). Also, calibration of continuous rainfall-runoff models is difficult and time consuming. In spite of the ease in using single-event models, the initial soil moisture condition is unknown. Owing to high sensitivity of simulated floods to soil moisture initial condition, application of single-event models faces great uncertainty (Roy et al., 2001). Consequently, for climate change impact assessment on floods, continuous models are preferred (Prudhomme et al., 2002).

Some studies have directly dealt with the issue of climate change impact on floods (IPCC, 2001; Prudhomme et al., 2003). The main difficulties include producing a representative fine time scale continuous time series of climate variables for future climate scenarios. Normally, a long-term watershed-scale time series with fine time step for future climate is required for flood studies (Prudhomme et al., 2003). As such, rainfall and temperature data shall be downscaled, while it is important to preserve the correlation between the downscaled variables (Fowler et al., 2007).

There are two fundamental approaches for downscaling of GCM outputs. The first is dynamical downscaling using Regional Climate Models (RCMs), and the second is statistical downscaling. RCMs use outputs from GCMs as boundary conditions and simulate atmospheric processes for a limited area under finer resolution (typically daily time scale with 50 km spatial resolution) (Prudhomme et al., 2003) While correlation between variables is preserved in downscaling (Fowler et al., 2007), RCMs are computationally expensive and generally its output for many regions and scenarios are not available (Prudhomme et al., 2003). Comparatively, statistical downscaling that is computationally inexpensive and efficient is considered more suitable for rapid approximation of local impacts of climate change (Wilby, 2007).

Many statistical downscaling techniques have been developed for downscaling of GCM outputs onto a finer resolution. Among them, Change Factors (CFs) and Weather Generators (WGs), preserve the correlation between downscaled variables. The CFs is the simplest method of downscaling (Prudhomme et al., 2002). In this method, the differences between means of GCM outputs for control and future periods are applied to every baseline observed data series, either as summation or multiplication. Therefore, CFs may be rapidly applied to various GCMs to produce a vast range of climate scenarios. A limitation of CFs is that it only adjusts the mean of climate variables and does not apply changes in temporal variability (Fowler et al., 2007). However, for climate change impact assessment, especially on extreme values, changes in variability can be more important than changes in averages (Prudhomme et al., 2002). Besides, for daily rainfall, the length of wet series will be fixed, while change in the wet and dry spell's lengths is regarded as an important component of climate change (Fowler et al., 2007). Also, increase in monthly rainfall can either increase the number of wet days or increase the amount of intense rainfalls and/or extreme values, where each of these modes have different impacts on the simulated flows (Reynard et al., 2001). Meanwhile, only the amount of rainfall in wet days will be increased with the same ratio in the CFs method (Fowler et al., 2007).

Among statistical downscaling methods, WGs have unique advantages, including: (1) They produce long-term series that provide a wider range of feasible situations and decrease uncertainty of climate variability (Semenov et al., 1998). Uncertainty of natural variability has important and decisive role in climate change impact assessment on floods studies, and can even be more important than uncertainty in relation to emission scenarios and structure of GCMs (Kay et al., 2009); (2) In downscaling by WGs, correlation between variables is preserved that is essential for hydrological simulations (Fowler et al., 2007); and (3) In downscaling, besides changes in average, changes in variability and other important statistics are applied (Fowler et al., 2007). In climate change impact assessment on floods, changes in climate variability can be more important than changes in averages (Prudhomme et al., 2002).

In spite of the WGs advantages, a number of deficiencies in performance of WG models has limited their usage, especially for assessment of climate change impacts on floods. The deficiencies are as follows: (1) WGs cannot accurately reproduce extremes (IPCC, 2001; Srikanthan and McMahon, 2001; Caron et al., 2008); and (2) Low-frequency variabilities are not correctly reproduced (Semenov et al., 1998; Hansen and Mavromatis, 2001; IPCC, 2001; Mavromatis and Hansen, 2001; Dubrovsky et al., 2004; Fowler et al., 2007).

Past studies on climate change impact assessment on floods have either used RCM outputs with or without CFs method (Kay et al., 2006a, 2006b; Leander and Buishand, 2007; Linde et al., 2010) or used simple CFs method for downscaling of GCM outputs (Reynard et al., 2001; Loukas et al., 2002; Prudhomme et al., 2003; Mareuil et al., 2007).

Furthermore, in most previous studies, uncertainty related to natural variability and short length of the data were not considered; while in some studies (e.g. Leander and Buishand, 2007; Kay et al., 2009; Linde et al., 2010), resampling method was applied. A major limitation of resampling methods is that they do not produce new values of climate variables, such as precipitation amounts larger than observed, but merely reshuffle the historical data to generate synthetic weather data (Sharif and Burn, 2006). Application of such sequences as input for rainfall-runoff models for simulation of flood events could lead to under-exploration of possible effects of climatic variability, and more extreme flood events could be underestimated (Sharif and Burn, 2006).

In another study of climate change impact assessment on floods, Mareuil et al. (2007) used a WGEN-type WG (called WeaGETS) for long weather series generation as input to a rainfall-runoff model. They found that simulated floods were underestimated, which was mainly the result of underestimation of the extreme daily precipitation events by WeaGETS (Mareuil et al., 2007). Also, Caron et al. (2008) validated WeaGETS by conducting a series of hydrological modelling experiments and found that extreme flood events were not reproduced well (Caron et al., 2008). A weather generator capable of producing precipitation amounts larger than observed, and consequently simulating more extreme events, would be advantageous for correct simulation of probability distribution of floods (Sharif and Burn, 2006).

In this paper, using a continuous rainfall-runoff simulation model and a new WG, climate change impacts on floods in Pataveh basin located in southwestern Iran is assessed. Downscaling and generation of long-term daily rainfall and temperature series is performed via a newly developed WG, which is able to simulate low-frequency variability and extreme rainfalls more realistically. In comparison to CFs method as used in many studies for climate change impact assessment on floods, the uncertainty of natural variability will decrease by generating long series of climate variables while changes in mean and variability and other important statistics may be enforced. In comparison to resampling method, the WG used in this study can generate more realistic series that contain new values such as precipitation amounts larger than observed. Streamflow is simulated using the ARNO continuous rainfall-runoff model.

2. Methodology and data

2.1. Rainfall-runoff model

In this research, the ARNO continuous rainfall-runoff model is used to simulate daily runoff (Todini, 1988). This model is broadly used in water management, low flow, and real-time flood forecasting studies in different parts of the world. Variations of the model have been implemented in GCMs (Todini, 1996; Abdulla et al., 1999).

ARNO is a semi-distributed conceptual model. In ARNO, the basin is composed of an infinite number of elementary areas, each of which has a different soil moisture capacity. The spatial distribution of the soil moisture capacity is expressed in the form of a probability distribution function. Soil moisture content is fed by rainfall that infiltrates into the soil and is depleted by evapotranspiration, drainage and percolation. For each of the elementary areas with different soil moisture capacity and different soil moisture content, the continuity of mass is simulated over time. Basin runoff is the integral of the runoff of elementary areas, transferred to the outlet of the basin via a routing module (Todini, 1996). In Figure 1, ARNO model flowchart and its parameters are shown. More details on ARNO model are described by Todini (1996).

Figure 1.

Flowchart representing the ARNO model

2.2. Rainfall generator

In WGs, generally, rainfall is considered as the primary variable, while other variables are generated dependant on the occurrence of rainfall. Therefore, success of the procedure in producing weather series, is greatly dependent upon the method of rainfall generation (Kilsby et al., 2007). Also, in flood-simulation studies, rainfall is regarded as the main variable and generating extreme rainfall is of great importance.

Empirical models have been used for rainfall generation by most WGs. Empirical rainfall models often reproduce low order moments of daily rainfall, but are not able to reproduce extreme events and low order moments of other time scales, including monthly variance (Burton et al., 2008; Srikanthan and McMahon, 2001). The proposed WG used in this study applies a more sophisticated Neyman-Scott point process (NSRP) model (Cowpertwait, 1991) in order to produce rainfall. NSRP is a stochastic model in which physical process of rainfall structure is described by stochastic methods. Capability of this model for preserving statistics of the observed rainfall in a broad range of time scales as well as in extreme events has been demonstrated in several studies (Burlando and Rosso, 1993; Cowpertwait, 1994; Cowpertwait et al., 1996a, 1996b; Cowpertwait et al., 2002; Olsson and Burlando, 2002; Kilsby et al., 2004). The structure of NSRP is illustrated in Figure 2 and is briefly described as follows (Burton et al., 2008): (1) Storm origins arrive as a Poisson process with the arrival rate represented by a parameter λ; (2) Each storm origin generates a (Poisson) random number C, with mean value ν, of raincells that each follows the storm origin after a time interval that is exponentially distributed with parameter β; (3) Each raincell produces a uniform independent duration and intensity. The duration and the intensity of each raincell are exponentially distributed with parameters η and ξ, respectively; and (4) The intensity of rainfall in each moment is equal to the sum of the intensities of all the active cells at that instant.

Figure 2.

Schematic of the NSRP model (Burton et al., 2008)

This process is time continuous, so a time series is achieved by discretising the process into daily or desired time step.

NSRP model has five parameters (as in Table I) that are determined via calibration of the model. For approximating parameters of the model, at least five statistics from observed time series are calculated and subsequently made equal to their corresponding derived expressions for the model. Consequently, a system of equations will be obtained and solved by a numerical optimisation algorithm, in order to determine model parameters (Burton et al., 2008). To preserve seasonal characteristics of rainfall, the parameterisation for each calendar month is separately provided.

Table I. Parameters of the NSRP (Burton et al., 2008)
DescriptionUnitsParameters
Mean waiting time between adjacent storm originshrλ−1
Mean waiting time for raincell origins after storm originhrβ−1
Mean number of raincells per stormν
Mean duration of raincellhrη−1
Mean intensity of a raincellmm/hrξ−1

2.3. Minimum and maximum temperature generator

Seasonal cycle of daily means and standard deviations of minimum and maximum temperature are approximated by finite Fourier series of order 3. These statistics are separately calculated for wet and dry days. In the previous studies, the finite Fourier series of order 3 have demonstrated suitable performance for modelling of annual cycle of means and standard deviations of the minimum and maximum temperature (Semenov et al., 1998). After reduction of daily means and standard deviations cycle, time series of residuals is modelled as a first order auto-regressive process as follows:

equation image(1)

where xk(i) is a 2 × 1 matrix of residuals for day i (with k = 1 for maximum temperature and k = 2 for minimum temperature), and xk(i − 1) is a 2 × 1 matrix of residuals for day i − 1. A and B are 2 × 2 coefficient matrices, and εk(i) is a two-dimensional vector for independent standard normal variables.

For reproduction of low-frequency variability, the time series of monthly averages is generated using a monthly multivariate AR(1) model. Then, series of monthly means of minimum and maximum temperatures that are generated by the daily model fits the monthly series generated by the monthly model. Therefore, monthly statistics of minimum and maximum daily temperature series are the same as monthly series statistics that are generated by the monthly model such that low-frequency variability will be reproduced.

2.4. Downscaling procedure

The WGs parameters that are estimated based on the observed series are perturbed according to the GCM climate change scenarios. Then, future climate scenarios are generated by using perturbed parameters (Wilby et al., 2002; Dubrovsky et al., 2004; Richter and Semenov, 2005; Khan et al., 2006; Kilsby et al., 2007). These change fields are calculated from changes in the corresponding statistics of the GCM outputs for future scenario in comparison to the control period for cells of the study region. The following downscaling procedure is conducted: (1) By using daily and monthly series of rainfall, minimum and maximum temperatures of GCM outputs for the control period and for the future scenario, the required daily and monthly statistics for the WG are calculated; and (2) Change field of the GCM output statistics for the desired scenario will be applied to the corresponding observed statistics.

2.5. Study basin and data

The basin under study is the Pataveh Basin, located between 30°N and 31°N latitudes and 51°E to 52°E longitudes in southwest Iran. The area of the basin is 2800 km2 and its average altitude is 2277 m (a.m.s.l.). Geographical location of the basin and its meteorological and hydrometric stations are shown in Figure 3. Daily rainfall and temperature data at Yasuj meteorological station close to the centroid of the basin are accessible during the 1974–2002 period. Average annual rainfall over this period is 890 mm. Also, daily streamflow at the outlet of the basin is also available over the same period. Average streamflow at Pataveh station is 52.2 m3/s varying from 18.5 to 106.5 m3/s in different years.

Figure 3.

Boundary of Pataveh basin and the location of hydrometeorologic stations

In this study, daily climate scenarios of the Coupled Global Climate Model (CGCM3) model for A2, A1B and B1 emission scenarios, from SRES series are considered. Spatial resolution of CGCM3 is 2.8° lat/lon and 31 levels in the vertical.

3. Results and discussion

3.1. Rainfall simulation

NSRP model was calibrated for 27 years of observed daily rainfall series (1974–2000) at Yasouj meteorological station. This period is common between the observed data and the GCM control period. The stationarity of the observed rainfall series was checked. No trend was detected in the 95% confidence level. Next, seven statistics of the observed daily rainfall series were calculated for approximation of five NSRP parameters for each calendar month. The statistics include the mean, variance and skewness of daily rainfall, lag 1 daily auto-correlation coefficient, probability of dry days, and probability of dry–dry (and wet–wet) transition.

For assessment of the NSRP performance, the statistics of the generated series were compared with the corresponding statistics of the observed rainfall. For doing so, 100 daily rainfall series of length 27 years were generated. For each generated series, the important statistics were calculated and were compared with the corresponding statistics of the observed series (Kilsby et al., 2007; Holman et al., 2009). Falling observed statistics within a confidence interval of the corresponding generated statistics, indicates good performance of the model at that probability level. In Figures 4 and 5, respectively, the distribution of the maxima annual rainfalls with 1 and 2 days duration are compared with the 90% bounds of the corresponding values of the generated rainfall on a Gumbel paper. In spite of the fact that the extreme values were not explicitly used in the calibration of the model, the distribution of the observed extreme rainfalls is well reproduced. This implies an acceptable performance of the NSRP model for simulation of daily rainfall time series with realistic extremes. Figure 6 compares observed and simulated daily annual maximum rainfall distributions. Simulated extreme daily rainfall distribution contains wider range of feasible extremes that is not observed in the recorded data. Thus, in comparison with the distribution of observed extreme rainfalls, uncertainty attributed to the sampling variability and short length of the data decreased in the simulated extreme rainfall.

Figure 4.

Extreme value plot of annual maximum 1-day rainfall at Yasuj station

Figure 5.

Extreme value plot of annual maximum 2-day rainfall at Yasuj station

Figure 6.

Comparison of observed and simulated daily annual maximum rainfall distributions at Yasuj station

3.2. Calibration of rainfall-runoff model

Eight years of high-quality data were used for calibration and validation of the rainfall-runoff model. The average streamflow of the basin at Pataveh station in this period ranges from 23.2 to 85.3 m3/s, covering wet and dry years. For calibration of the ARNO model, at least two to three years of data is required (Todini, 1996).

The efficiency criterion (EC) of calibration and validation stages were turned out to be 0.85 and 0.78, respectively, while determination coefficient (R2) were 0.87 and 0.84, respectively. For the wettest water year of the period (October 2001–September 2002), with average streamflow of 85.3 m3/s, the values of EC and R2 were 0.89 and 0.89, respectively. For the driest year of the period (October 1983–September 1984), with average streamflow of 23.2 m3/s, the values of EC and R2 were 0.82 and 0.90, respectively. Thus, the performance of the model in both dry and wet years is similarly very good.

In comparison with other studies involving daily rainfall-runoff simulation, Zhang and Savenije (2005) adopted acceptable calibration when EC is greater than 0.6. Meanwhile, Kamali et al. (2007) in daily calibration of WATCLASS hydrologic model for Smokey River basin in Canada, accepted EC values of greater than 0.7. Evans and Schreider (2002) in rainfall-runoff simulation of 6 basins in Australia by using CMD-IHACRES model, used 4 years of observed data for calibration and obtained EC values between 0.67 and 0.78. Loukas et al. (2004) calibrated UBC for a basin in Canada by using 20 years of observed data with an EC of 0.93. Zhang and Savenije (2005) used REWASH model in a basin in Belgium and achieved EC of 0.68 and 0.65 during calibration and validation stage, respectively. Comparing the results of this research with those reported in previous daily rainfall-runoff simulation studies indicates that the performance of the calibrated ARNO model for simulating daily streamflow of the Pataveh basin is quite acceptable.

3.3. Long-term flow simulation of present climate condition

Calibrated WG model was run to produce 100 generated series, each 27 years long, of daily rainfall, minimum and maximum temperatures. Potential evapotranspiration of the basin was calculated using the temperature data. The generated temperature, rainfall and evapotranspiration were used as inputs to the ARNO rainfall-runoff model to simulate 100 new series (2700 years in total) of daily streamflow for Pataveh basin under present climate condition. In Figure 7, historical annual maximum daily streamflow distribution is compared with the 95% bounds of the simulated annual maximum daily streamflow distribution. It may be seen that the 95% bounds of the simulated extremes match the observed data well, while, the curvature, the curve slope, the median extreme value as well as the greatest extreme are generally consistent with the historical pattern. This concludes that the WG can simulate weather data as inputs of rainfall-runoff model to simulate realistic daily extreme floods.

Figure 7.

Compression of historical annual maximum daily streamflow distribution with the simulated 95% bounds of the annual maximum daily streamflow distribution

Figure 8 shows frequency distribution of the maxima of the observed daily streamflow recorded at Partaveh station, the simulated streamflow based on historical meteorological data and the long-term simulated streamflow based on generated data. It is seen that good performance is achieved by coupling the WG and the rainfall-runoff model for simulation of daily runoff time series with realistic extremes. Simulated daily runoff distribution contains wider range of feasible floods that is not observed in the recorded data. This synthetic distribution that was obtained from long-term simulated daily runoff data, should reduce uncertainty of natural variability and short length of the recorded data.

Figure 8.

Frequency distribution of the maxima of the observed daily streamflow record, the simulated streamflow based on historical meteorological data, and the simulated streamflow based on generated data

3.4. Downscaling and climate change impact assessment

In order to produce future climate series, WG parameters were modified according to each of the scenarios for future periods and then future long-term climate scenarios were generated. For doing so, the required statistics for calibration of the WG were computed from the CGCM3 data during control (1974–2000) and future (2007–2033, 2037–2063, 2067–2093) periods for each of the A2, B1 and A1B scenarios. Change fields for each of the future scenario statistics in comparison with the corresponding statistics of the control period were applied to the correspondent observed statistic. Then, the parameters of the WG for the perturbed statistics were estimated. By using the perturbed WG parameters for each of the future scenarios, long-term series of rainfall and temperature for future climate scenarios were generated.

Using the generated future climate data and the rainfall-runoff model, 2700 years of daily streamflow for each of the scenarios at each of the future periods were generated. Annual maxima daily streamflow were extracted from the daily streamflow series. Figures 9–11 show climate change impacts on floods distribution at Pataveh station under A1B, B1 and A2 scenarios for periods of 2007–2033, 2037–2063, 2067–2093, and Figure 12 shows the impacts of climate change on floods distribution under A1B, B1 and A2 scenarios for period of 2067–2093. Also in Table II, annual maxima daily flows at Pataveh station is presented for the present climate conditions and future scenarios in different return periods. Results indicate that despite the magnitude of the change is upon the choice of the scenario, the climate change will impose increase on floods under all the considered scenarios (expect for floods with return period of less than 10 years under the A1B scenario for period of 2007–2033). The largest increases are associated with the A2 scenario for period of 2067–2093 and B1 scenario for period of 2037–2063. For instance, under these scenarios flood magnitude with a return period of 50 years shows a 253 and 104% increase, respectively. On the basis of the results, despite the uncertainty of the emission scenarios, the climate change will impose considerable increase on floods of the Pataveh basin, approximately, under all the considered scenarios.

Figure 9.

Climate change impact on daily flood distribution under A1B scenario for 2007–2033, 2037–2063, 2067–2093 periods

Figure 10.

Climate change impact on daily flood distribution under B1 scenario for 2007–2033, 2037–2063, 2067–2093 periods

Figure 11.

Climate change impact on daily flood distribution under A2 scenario for 2007–2033, 2037–2063, 2067–2093 periods

Figure 12.

Climate change impact on daily floods distribution under A1B, A2, and B1 scenarios for 2067–2093 period

Table II. Annual maxima daily flows for present climate condition and future scenarios in different return periods (in m3/s)
T (years)1974–20002007–20332037–20632067–2093
 HistoricalB1A2A1BB1A2A1BB1A2A1B
2391403467316499439418452665450
1065871795965110448338148721348868
207638961161843132810251006109616881068
258059551243890145610831067115218361157
50926112615091101188912771303137323451432
1001019131218591374263215151542173228311734
2001162157922641621314016561942211634112016
5001311197028841999448922932412264244603031

4. Conclusions

For climate change impact assessment on floods, it is necessary to simulate rainfall-runoff processes in a continuous mode, while in downscaling, the correlation between climate variables (at lest temperature and rainfall) should be preserved. Among statistical downscaling methods, the WGs and CFs method preserve the correlation between the downscaled variables. In downscaling, CFs method only applies changes in averages, while in climate change impact assessment on floods, change in the variability can be more important than change in the averages. In contrast, besides preserving correlation of variables in downscaling, WGs can enforce changes in various statistics of climate variables according to the climate change scenarios. WGs also produce long term series of climate variables that may decrease uncertainty of natural variability.

In spite of major advantages of WGs, a number of deficiencies in performance of WG models have limited their usage, especially for assessment of climate change impacts on floods. In most previous studies, CFs method is used for downscaling. In this paper, a weather generator is used that can generate weather data as inputs of a rainfall-runoff model for climate change impact assessment on floods. Rainfall in this WG is generated using NSRP model, and temperature is modelled using mixing of daily and monthly multivariate AR(1) models. Using this WG model and the calibrated ARNO rainfall-runoff model, daily runoff time series with realistic extremes and desired length may be generated for assessment of possible changes in the flood regime. By synthetic simulation of flood data, a wider range of feasible situations can be produced. Besides, by using this model for downscaling, it is expected that changes in average and in other important statistics, such as variance and length of wet series, can be appropriately enforced.

Climate change impact on floods in Pataveh basin under A1B, A2 and B1 scenarios for 2007–2033, 2037–2063, 2067–2093 periods was assessed. On the basis of the results of this research, climate change will impose considerable impact on flood regime in Pataveh basin under each scenario. For instance, flood magnitude with a return period of 50 years for the 2067–2093 period shows a 48, 55 and 153% increase in comparison to 1974–2000 historic period under B1, A1B, A2 scenarios, respectively.

Ancillary