The objective of climate models is to describe long term changes in the atmospheric system, at the global or regional level. They make use of atmospheric models, which are basically the same type of models as used in operational weather forecasting. While weather forecasting attempts to predict the conditions of the atmosphere accurately for a period of about one week, using specific initial conditions given by assimilated observation data, climate models operate for much longer periods, generally from several decades up to the end of the century. For climate models, the initial conditions are less important compared to weather models because the objective is to simulate changes in average or extreme statistical properties of key meteorological phenomena rather than making detailed short-term forecasts.
Historically there were significant differences between the physics in climate and weather models. For example, early weather models ignored radiation in their implementation (Shea et al.1994). However, today the physical processes and their implementations in climate and weather models are nearly identical.
In fact recent literature shows many examples of using weather model systems to conduct climate type simulation experiments, for example Lo et al. (2008), who used the WRF model for climate downscaling. The main differences are the prediction duration of interest and the“type”of prediction that is being undertaken (Giorgi, 2005):
(1) Prediction of the evolution of the atmospheric state, given an initial condition: this is called predictability of the first kind.
(2) Prediction of the statistics of the response of the atmosphere for some external forcing (e.g. climate forcing by means of scenarios for GHG emissions or atmospheric concentrations): called predictability of the second kind.
Weather prediction is a clear example of predictability of the first kind. The weather prediction rarely shows high accuracy beyond a period of one week. In general, the prediction accuracy within a given system is a function of the temporal scale of the phenomenon that is considered. Typical atmospheric processes leading to rainfall are of relatively short time scales (e.g. cumulus formation: a few hours; synoptic systems: a day or two). Climate projections are essentially of the second kind: they largely“forget”the initial conditions of the simulations and respond to the external climate forcing, behaving much like a boundary value problem. The results of such attempts are inherently probabilistic and it is important to treat them as such in further analyses.
To make a clear distinction between predictions of the first kind and predictions of the second kind, predictions of the second kind are commonly calledprojectionsrather than predictions.
In between weather predictions (∼5–10 days) and climate projections (∼30–100 years) comes forecasts on seasonal to decadal time scales. Despite the chaotic nature of the atmosphere, predictability on longer time scales may exist through the large-scale, low-frequency climate oscillations discussed in Section 3.2., such as ENSO, which influences for example seasonal precipitation in most parts of the world (e.g.
Trenberth & Caron, 2000). Seasonal to decadal forecasting is performed using dynamical, statistical or a combination of methods. The forecasts generally exhibit good skill in the tropics and for monsoon rainfall. In other regions the skill is generally low but expected to improve with future development of observation systems as well as climate models (e.g., Smithet al.2012).
When the climate change projections are to be made at regional scales, we face further challenges.
Historically GCM projections, which cover the entire globe, have been largely limited to coarse resolutions, typically several hundred kilometres, primarily due to computational limitations. While this limitation is continually relaxed, there are other issues that are intrinsic to the atmospheric system that makes it challenging to use local scale projections. First, there are numerous local effects (e.g. geographical features) that modify the atmospheric flow in complex ways. Secondly, small horizontal shifts in storm movement can dramatically alter local precipitation patterns. Further there is evidence that the variability of atmospheric variables (including precipitation) increases with refined spatial scales (Giorgi, 2002).
While a simplification, it is generally true that the smallest space/time scale at which predictions have real information decreases both spatially and temporally with an increasing range of prediction (Figure 4.1). The figure shows that with increasing range of prediction, fine-scale (both time and space) information is lost. Long-range model results should be interpreted as space/time averages over a large range or as statistics, which can be seen as another way of aggregating. Even with sufficiently high grid resolution (which is far from the case today) a GCM will not be able to deterministically predict cumulus formation. However, it may be able to give: 1) rainfall averaged over a large region over a large time period; and 2) statistical information on cumulus formation. This is an intrinsic property of the atmospheric system–not a limitation in current models which can be overcome with better numerical schemes and refined resolutions.
Physical basis
The development of atmospheric models started from the late sixties to the mid-seventies, with the development of the theory of chaos by Lorenz (1975). This theory allowed describing the behaviour of
complex systems like the earth’s atmosphere. All physically based atmospheric models use some sort of numerical scheme to solve the following five equations in a three-dimensional space:
(1) Conservation of mass (2) Conservation of momentum (3) Conservation of heat (4) Perfect gas law
(5) Conservation (and phase changes) of water
Note that atmospheric models describe water in a different way than what is done in hydrology. Hydrologists typically model water as an incompressible fluid that does not change its properties significantly within the typical ambient range of temperatures (Chow, 1964; Chowet al.1988). Therefore the motion of water is generally described using the conservation laws of mass and momentum. In the case of gases, it is not realistic to assume incompressibility. They expand significantly with an increase of temperature. In addition to the conservation of mass and momentum, describing atmospheric motion requires a conservation of thermal energy (as described by the first law of thermodynamics) and a relationship between temperature, pressure and volume (as described by the perfect gas law). It is typical to consider water as a separate constituent having its own conservation equation.
Boundary conditions
Like any other type of numerical model, atmospheric models need initial conditions and boundary conditions to operate. The boundary conditions are often divided into top, bottom and lateral types (Figure 4.2).
In most atmospheric models the top boundary conditions use simple parameterizations for important quantities (e.g. radiation exchange). These parameterizations are often simplified and depend on known Figure 4.1 A simplified view on predictability of atmospheric conditions.
external processes (e.g. seasonal solar radiation). For example, many GCMs set the model top well above the troposphere making it possible to assume no air exchange over the model top. For instance, the NOAA GISS ModelE (Schimitet al.2006) sets the model top at 0.1 hPa level, which is approximately 64 km above the stratosphere. Modern atmospheric models use a simplified surface model (called Land Surface Model– LSM) to cover the bottom boundary. For example, ModelE uses an improved version of the Rosenzweig and Abramopoulos (1997) LSM, that can represent vegetation, snow and catchment hydrological processes.
When the model coverage is global, as is the case for a GCM in applications of climate change projections and a Global Weather Model (GWM) in weather prediction applications, the model does not have lateral boundaries. Therefore the need for lateral boundary conditions does not arise. This leads to an important ability for these models: given initial conditions they can simulate future atmospheric conditions without any future boundary conditions. Therefore global atmospheric models can make predictions in the real sense of the word.
Regional models
In regional studies, another type of model is used, namely aLimited Area Model(LAM). Such as model is called RCM when used for applications of climate change projections andRegional Weather Modelfor numerical weather prediction. As the name implies, regional models have lateral boundaries and therefore depend on the specification of the lateral boundary conditions. These boundary conditions can come from two sources:
– In the case of historical studies they can be based on past observations. However, in order to combine different types of observations with inevitable error sources in a physically consistent way, data assimilation is used with the help of prior knowledge of the state given by a numerical atmospheric model (Wanget al.2000). For historical studies suchreanalysis dataare used as lateral boundary conditions (often also for initial conditions). For more information on the basics of data assimilation and reanalysis the reader is referred to Sheaet al.(1994).
There are several publicly available reanalysis datasets:
○ One is the global reanalysis data from the National Centre for Environmental Prediction (NCEP), which are dating back to 1948 and continuously updated (Kistler et al.2001): http://www.esrl.
noaa.gov/psd/data/reanalysis/reanalysis.shtml.
Figure 4.2 Types of boundary conditions in atmospheric models.
○ Interesting is the NCEP FNL Operational Model Global Tropospheric Analyses ds083.2 dataset, a 6-hourly dataset covering the globe at 1×1 deg resolution (CISL, 2011) at 26+atmospheric layers from the earth’s surface to 10 hPa (about 30 km) height. They use the Global Telecommunications System (GTS) and other data sources and perform reanalysis using the Global Forecast System (GFS), which is an operational GWM.
○ Other global reanalysis datasets are the 40-year reanalysis data (ERA-40) from the European Centre for Medium-Range Weather Forecasts (ECMWF) at 2.5 deg resolution covering the period 1957–2002 (Uppalaet al.2005), and the continuously updated ERA-Interim dataset covering the period from 1979 onwards: http://www.ecmwf.int/products/data/archive/ descriptions/e4/index.html.
○ Next to these global datasets, some regional datasets are available, for example for Europe (E-OBS;
Haylocket al.2008), India and East Africa (Yatagaiet al.2009) and South America (Silvaet al.
2007). Another useful source of daily rainfall data is the Global Historical Climatology Network (GHCN)–Daily, which contains records from over 75000 stations in 180 countries and territories, with lengths that range from less than 1 year to more than 175 years: http://www.ncdc.noaa.
gov/oa/climate/ghcn-daily/
○ One has to note that these reanalysis/historical data are more reliable for more recent periods. A strong increase in the accuracy is expected from the 1980s due to the availability of satellite data. Serious inaccuracies may, however, remain in recent years. For example, Uppala et al.
(2005) found a serious bias in ERA-40 rainfall data due to the Pinatubo eruption of 1991, which caused significant artificial shifts in some of the assimilated satellite streams and reanalysed rainfall fields.
– For analyses involving future forecasts as in climate modelling but also in operational weather forecasting, the regional models (RCMs or RWMs) depend on predictions made by global models (GCMs or GWMs).
Figure 4.3 shows a taxonomy of atmospheric modelling applications. As outlined above, the two traditional classes of applications are weather modelling and climate modelling. They have little overlap in the temporal range, and strongly differ in spatial scales. Weather prediction modelling typically covers spatial scales from mesoscale (strict technical sense meso-gamma scale,∼10 km) all the way to global scale. Typical examples of models used for such short range predictions are RAMS (Pielke, 1992), MM5 (Grell, 1994) and WRF (Skamarock, 2005). The typical usage of these models is at the scale range from the meso-gamma scale to the synoptic-scale (∼1000 km) as they use specified lateral boundary conditions, with a forecast range of several days to weeks. In special research applications the models are sometimes employed at smaller spatial scales (e.g. ,1 km, microscale or smaller). They also can be used as global models with special setups. However, the most prevalent use of these models remains within the mesoscale range. This application therefore is also known asMesoscale Meteorological Modelling(MMM) (Pielke, 2002).
For climate models, because they consider long term changes in the atmospheric system, the spatial domains are large compared with the spatial domains covered by the weather prediction models, as also shown in Figure 4.3. They indeed go from global to synoptic-scale level. There also would be the option to use climate models for the smaller mesoscale range, hence to limited-area application domains. This would be useful in applications of urban hydrology. Most useful for these applications would be to restrict the spatial domain to the area of a single city or urban area. Whether this is practically feasible will be discussed in Chapter 5. Atmospheric models bounded to the smaller mesoscale will hereafter be referred to as LAMs. This name refers to the atmospheric model, independent on whether it is used for weather prediction or climate modelling.