Early attempts of long-term climate prediction focused only on the extrapolation of climate trends and cycles (Dyer and Tyson 1977, Currie 1993). In recent decades, however, the use of GCM-predictors from global climate models has become the common method of choice for long-term climate and subsequent hydrological forecasting, which is of particular interest for water planners, (Goddard et al. 2001). These general circulation models (Phillips 1956) or global climate models (GCMs) generate the future climate by simulating climate time-series at the global-scale on a rather coarse grid over up to a century-long period, assuming various CO2 - emission (SRES) scenarios for the 21st-century. For a meaningful application of a global-scale climate model (GCM) it must be assured that the latter is able to predict well enough the climate boundary or forcing conditions used, for example, in a subsequent hydrological model.
A GCM is a numerical model which solves the full physical equations describing mass- and energy transfer in the coupled ocean-atmosphere system. For this reason, a GCM is also referred to as a coupled general climate circulation model (CGCM)(e.g. Mechoso et al. 1995, Stockdale et al. 1998a, Fischer and Navarra 2000, Gualdi et al. 2003) in which an atmospheric model (AGCM) (e.g. Bengtsson et al. 1993, Cubasch et al. 1995, Hunt 1997, Mason and Graham 1999, Coppola and Giorgi 2005) is merged with an ocean climate model to reduce the problems with the specification of the appropriated boundary in an atmospheric AGCM alone (Delecluse et al.
1998).
Global climate models (GCM) (Gates 1984, Robinson and Finkelstein 1990, Lamb 1987) perform climate simulations on a rather coarse horizontal grid which, depending on the GCM, ranges from a size of 2.5°x2.5° ~(250km x 250 km) to, nowadays, in some cases down to 0.5°x0.5°
(50km x 50km). In spite of these ongoing improvements in the geometrical resolution as well as in the physical representation of the complex climate processes involved, as the power of modern computers is steadily increasing, the GCMs climate predictions are most of the time still not good enough for small-scale regional climate analyses.
Therefore, to link the open gap between the coarse grid-resolutions of GCM models and the fine-grid resolution required for understanding changes in the local climate, techniques of so-called “downscaling” must be applied. In impact assessments of climate change, namely, on water resources at the basin scale, downscaling is indispensable for the application of hydrological models (Fowler et al. 2007, Teutschbein et al. 2011). Although there have been countless applications of downscaling, in general, up-to-date (Giorgi and Mearns 1991, Robock et al. 1993, Hewitson and Crane 1996, Wilby and Wigley 1997, Wilby et al. 2004, Teutschbein et al. 2011) to obtain higher-resolution information about the future-climate from GCMs, many open questions and problems remain to be solved, as will be detailed throughout this chapter.
Regarding the particular methodologies used in general downscaling, these can essentially be divided in two categories (Fowler et al. 2007). The first one is called “dynamical downscaling”
which is essentially a small, but high-resolution, regional climate model (RCM) embedded in large, coarse-grid GCM, whereby the former employs as its boundary conditions the predictions of the GCM there, to simulate highly resolved climate variations in a local region (Wilby et al.
1998, Hay et al. 2000, Wilby et al. 2000, Gao et al. 2001, Fowler et al. 2005a, Frei et al. 2006).
However, the dynamical downscaling technique is often limited by the errors in the parent GCM that drives the large-scale fields (Mearns et al. 2003). Another limitation is the problem of internal variability from non-linear internal dynamics that conveys the uncertainty in regional climate simulations (Giorgi and Mearns 1991, 1999, Christensen et al. 2001). Since the dynamical downscaling technique requires huge computer resources and time-dependent boundary conditions from a parent GCM (Fowler et al. 2007), both of which are not available here this downscaling tool is not used in this study.
On the other hand, for some regions in the world local climate predictions from regional climate models (RCM) have already been archived and can then be used directly for further local climate impact studies. Such is the case, for example, for central Europe and Germany, in particular (Jacob et al. 2007), where Fink (Fink 2011) used the 21st-century regional climate predictions of the REMO-model (Jacob 2000) to analyze their impacts on the future hydrological regime in the Fulda catchment.
The second, more widely used class of downscaling tools is called “statistical downscaling”. In this category large-scale GCM-predictors are transferred to a finer grid-scale by using empirical (regression) relationships between the various coarse-resolution GCM predictors and the local climate variables (Karl et al. 1990, Wigley et al. 1990, Bretherton et al. 1992). Various sub-categories of statistical downscaling tools, where the SDSM (statistical downscaling method) is the most widely known representative (Wilby and Dawson 2013), in general, exist which include canonical correlation analysis (e.g. Karl et al. 1990, Wigley et al. 1990, Busuioc et al. 1999, Widmann et al. 2003), weather classification (typing) (e.g. Hay et al. 1991, Corte-Real et al. 1999a, Huth 2000), decomposition downscaling (e.g. Bretherton et al. 1992, Widmann et al. 2003, Liu and Fan 2012), the analogue method (e.g. Zorita et al. 1995, Zorita and Storch 1999), stochastical generation (e.g. Bardossy 1997, Bates et al. 1998, Semenov and Barrow 2002, Mehrotra and Sharma 2006, Vrac and Naveau 2007), artificial networks (e.g. Wilby et al. 1998, Schoof and Pryor 2001, Trigo and Palutikof 2001, Reusch and Alley 2002), semi-empirical downscaling (e.g.
Mearns et al. 1999), rescaling (e.g. Widmann and Bretherton 2000, Salathé 2003) and linear
regression (e.g. Ramírez et al. 2006, Fan and Wang 2009, Goyal and Ojha 2011, Sachindra et al.
2012).
For the country Thailand as a whole, downscaling of climate output of the GFDL-GCM by means of multi-linear regression and a k-nearest neighbor approach has recently been endeavored by Aksornsingcha and Chutime (2011), Singhrattna et al. (2011) and Singhrattna et al. (2012). For the present study region located in the eastern seaboard of Thailand, on the other hand, climate downscaling has not yet been carried out properly.
Among these afore-mentioned downscaling techniques, the regression model, as implemented in the SDSM-downscaling tool (Wilby et al. 2002) is probably the most widely used approach for regional climate prediction (Goddard et al. 2001). Also, by a successful combination of the stochastic and regression approaches - as in the LARS-WG (weather generator) (Semenov and Barrow 2002) - the general statistical downscaling can be improved for use in comprehensive future climate impact analyses under various SRES- scenarios (Wilby et al. 1998).
Methods and experiments of long-term climate predictions 3.2
In this study, a new combination of statistical regression and stochastical generation techniques is proposed to downscale the large-scale GCMs climate prediction output onto a regional
"climate grid" in the study area. More specifically, a multiple linear regression (MLR) downscaling method that sets up linear relationships between GCM- predictors and locally observed climate predictands (Chen et al. 2010b) is developed in this chapter. This is done for monthly climate data, as this the time-scale provided by many of the most reliable GCMs today.
The extension of this method that includes stochastic generation to resample the monthly climate to daily climate sequences, and so providing a multi-realization of climate events, will be further presented in Chapter 5.
Moreover, the two well-known conventional statistical downscaling techniques, i.e. the SDSM regression model and the LARS-WG stochastic method are likewise applied in this chapter. In fact these two methods will be employed as the references for the classes of statistical (Wetterhall et al. 2006, Chen et al. 2010b, Teutschbein et al. 2011) and stochastic (Khan et al. 2006, Hashmi et al. 2011) downscaling tools, respectively, and their results will be compared with those of the MLR-method.
Table 3.1. List of experiments of long-term prediction for examining the performance of various downscaling- and autoregressive models with predictors from various GCMs.
experiment downscaling
model domain GCM resolution number of
predictors grid size time
1.conventional model SDSM single ECHO-G 2.5°x2.5° daily 58
LARS-WG single ECHO-G 2.5°x2.5° daily 58
2. daily MLR MLR-daily single ECHO-G 2.5°x2.5° daily 58
3. monthly MLR MLR-monthly single ECHO-G 2.5°x2.5° monthly 21
4. multi-domain MLR MLR-monthly single HiRes/CRU 0.5°x0.5° monthly 5 MLR-monthly multi ECHO-G, BCCR, PCM
ECHAM5, GISS 2.5°x2.5° monthly 256 MLR-monthly multi ECHO-G, BCCR, PCM
ECHAM5, GISS, HiRes/CRU
2.5°x2.5°+
0.5°x0.5° monthly 256+5 5. autoregressive
models AR - - - - -
ARIMA - - - - -
ARIMAex multi ECHO-G, BCCR, PCM
ECHAM5, GISS, HiRes 2.5°x2.5°+
0.5°x0.5° monthly 256+5
A summary of the various experiments with the corresponding GCM/downscaling combinations used in this study is provided in Table 3.1, which also indicates the spatial and time resolutions of the different GCMs. Thus, the SDSM and LARS-WG models are applied to single-domain GCM-daily data, while the newly proposed MLR- downscaling technique uses both single- and multi-domain as well as both daily and monthly data. Additionally, for a test of its suitability, the group of autoregressive models which comprises the AR-, ARIMA- and ARIMAex- model, which are developed particularly for the use in short-term prediction in Chapter 4, is also applied here for long-term prediction. The theoretical bases of these downscaling models and a description of their main features are provided in the following sections.