A particular strength of NWP models is that their fields are internally dynamically consistent. Further, all such operational NWP systems include as an essential component a data assimilation system whereby observational data are incorporated into the model so as to provide the most accurate possible initial state for the next forecast. These analyses and model fields can be used as dynamically consistent data sets to diagnose the current and future kinematics and dynamics of the atmosphere. Suggested techniques where the numerical analyses and forecasts can be used to enhance the forecaster’s understanding of the numerical guidance follow Section 4.2.2.

4.2.1.1       Components of an NWP system

The essential components of a NWP system comprise:

·                         Forecast model;

·                         Data;

·                         Assimilation/initialisation phase.

A NWP model numerically integrates the momentum, thermodynamic, and moisture equations to produce a forecast of temperature, humidity, pressure, and the wind field, at a future time. Most models include what are termed the “dynamics” – that component of the model that integrates the equations – and the “physics” in which those processes not explicitly included in the forecast equations are “parameterised” in order to include their effects. These parameterised processes include radiation and its interaction with the Earth and atmosphere, convection, surface processes, and turbulent processes (see Haltiner and Williams, 1980). Moist processes (stratiform rainfall and convection) have been included in the physical parameterisations for many years, but explicit prognostic equations for moisture substance are increasingly being used.

In order for a NWP model to produce an accurate forecast, not only do the model numerics need to be well formulated, and the physical parameterisations represent as accurately as possible the processes occurring in the atmosphere, but also the initial state must be accurately specified. This requires not only that the three dimensional state of the momentum and thermodynamics of the atmosphere be specified, but also an accurate depiction the Earth’s orography, surface characteristics (roughness, snow cover, sea ice extent and lead fraction) and sea surface temperature consistent with the model formulation.

The data input typically comprise surface observations from land, and ocean–based platforms, rawindsonde profiles, temperature, humidity and wind data observed or diagnosed from sensors on board orbiting and geostationary satellites, and observations from aircraft. During the assimilation process the model state is corrected to “fit” the observations to within their observational error, so that the resulting fields retain the balance and information content of a short–term forecast based on previous information, but also reflects the information content of the most recent observations. (In this context, the short–term forecast used to provide the guess field can also be regarded as data.) Observations are subject to complex quality control procedures, usually including gross error checks followed by some form of “buddy checking”. The systems also include techniques for elimination of redundant data (highly correlated observations) and complex data selection algorithms so that dense observations of one type do not overwhelm complementary data of another type.

Unfortunately, the data available for use in the Antarctic is inhomogeneously distributed in space and time. The density of surface observations varies from land to sea and is highest around the coast of the continent. Typical data distributions available to an operational NWP system are shown in Puri et al. (1998). As the distribution of data, and types of data, are so inhomogeneous, it is necessary to include mass–wind coupling in the analysis so that the mass field is adjusted in the presence of wind only data and vice–versa so that information is inserted in a balanced way into the forecast model, and thus retained during the model’s integration. The appropriate statistics for background error correlations and observational error characteristics must also be included. Daley (1991) provides a comprehensive review of the topic. It should be noted that satellite sounder data are not generally used at tropospheric levels across the Antarctic because of the high orography of the continent and the problems of deriving sounding over an ice surface. In addition, the only part of the surface messages from staffed stations and AWSs that is used is the surface pressure data. This is because the wind and temperature data tend to be representative of the local area rather than the synoptic–scale environment.

Until recently, most major NWP centres used forms of multivariate statistical interpolation (Lorenc, 1981) to correct the short–term forecasts (also known as the background field), with data insertion typically every six hours (intermittent data insertion). Recently, though, there has been a trend towards 4–dimensional variational analysis (e.g. Derber, 1989, Rabier et al., 1998). In such a system, the model state is variationally adjusted to minimise its difference from the observations, subject to the strong constraint that the adjusted state is consistent with the model equations. Some of the advantages of variational data assimilation are that variables such as radiance in a given spectral band, as measured by satellite, can be treated as an observation, rather than needing to first have a temperature “retrieved” from these data, and the addition of the time dimension allows the time–tendency of the model state to match the implicit time tendency of data observed at through the assimilation period.

Initialisation schemes are used after the intermittent data insertion phase to “balance” the analysis fields – that is, to remove dynamic inconsistencies between the mass and the wind fields that may lead to spurious gravity waves (see Figure 4 of Puri et al., 1998) being generated early in the model forecasts. Their removal is desirable so that subsequent data insertions are not adversely affected, and to reduce the spurious prediction of rainfall early in the forecast period. Such schemes can be normal mode initialisation schemes (e.g. Ballish et al., 1992), digital filter schemes (Lynch and Huang, 1992) or use some form of Newtonian relaxation (eg Davidson and Puri, 1992). The variational assimilation schemes include constraints based on the model equations, and so initialisation is implicit in these schemes.

4.2.2                                Use of primary model output variables

Forecasters have for many years based their subjective forecasts on analyses of MSLP pressure and tropospheric–height and wind analyses in pressure coordinates. Implicit in this thinking are the simple geostrophic relations that result from this choice. The forecast is then based on the knowledge and experience, and on conceptual models of the relations between the isobaric patterns and observed and expected weather elements. The use of NWP model output has tended to follow the same path, with a model forecast MSLP pattern being used as a basis for a subjective forecast of weather elements. With the improving quality of operational NWP systems over the last 20 years, more direct use of model output is advantageous. The precipitation forecasts from NWP models are an example of direct model guidance of a critical forecast parameter, and while further improvement in the skill of these forecasts is needed, the forecasts do contain very useful information (Ebert and McBride, 1997).

The primary variables carried in an NWP model are pressure, temperature, wind components, and humidity. Further, most operational NWP models are formulated with some form of terrain–following vertical coordinate system. Thus it is more direct to use the forecast models’ low–level wind fields as direct guidance, rather than first use what may be an extrapolation of the model data to a constant pressure or height surface, and then make subjective estimates of wind based on geostrophic relations and knowledge of local effects. (Antarctic forecasters are perhaps less subject to this behaviour than forecasters who habitually deal with areas of lesser relief, as they are well acquainted with the problems of pressure reduction over regions of high orography.) It must also be remembered that many of the significant weather producing systems are highly ageostrophic, and so using geostrophic assumptions is most likely to lead to misinterpretation of model guidance in the most critical weather situations.

NWP models usually perform reasonably well over the Southern Ocean (although the fine–scale detail can often be in error (Turner et al., 1996a), but the forecaster should not expect them to out–perform their design. For example, global models of the 1980s did not have sufficient resolution to adequately resolve the orographic slopes of the Antarctic plateau. Therefore the forecast fields from these models did not resolve the distortion of the wind fields by these orographic barriers. As the resolution of the models and the sophistication of their physical parameterisations have increased, so the representation of the smaller–scale effects in these models has become more realistic. Forecasters should ensure that changes to operational models are communicated to them so that they can account for changes in the character of the forecast fields.

A simple but very powerful example of the direct use of the model fields is in the identification of cold and warm fronts over the ocean areas around the Antarctic. The MSLP field can be used to diagnose the locations of fronts from the position of pressure troughs in the forecast fields. However, the primary characteristic of a cold or warm front is a zone of enhanced temperature gradient in the atmosphere. Therefore it is far simpler to inspect fields of temperature near the surface to determine the location of fronts than to use the pattern of MSLP contours to infer their location. The wind field shows a cyclonic curvature at a frontal “discontinuity”, and so this field can be used to supplement the information from the thermal field. Figure 4.2.3.1 shows a MSLP forecast from the Australian Limited Area Prediction System (LAPS) model over the Southern Ocean. This should be compared with the fields of equivalent potential temperature (q_e) and streamlines on the model’s lowest sigma (ρ/ρ_surf) surface (~ 70 m) in Figures 4.2.3.2 and 4.2.3.3 respectively. The zone of enhanced gradient of q_e marking the position of the cold front and wind shear at the front mark the cold front more precisely than does the MSLP field.

A further benefit of this type of approach is to then use the model information about cold frontal structure to understand the patterns of cloud shown (independently) in the satellite imagery (Figure 4.2.3.4), rather than use the cloud imagery to attempt to diagnose the position of the surface front.

              Figure 4.2.3.1     NWP MSLP forecast for 1100 UTC 25 August 1998.

              Figure 4.2.3.2     NWP equivalent potential temperature (q_e) forecast

              for 1100 UTC 25 August 1998 on the model’s lowest sigma surface (~70 m).

              Figure 4.2.3.3     NWP streamline forecast for 1100 UTC 25 August 1998

     on the model’s lowest sigma surface (~70 m).

              Figure 4.2.3.4     GMS imagery for 1130 UTC 25 August 1998.

4.2.3                                Use of conceptual models  

Conceptual models of atmospheric systems are widely used to encapsulate the space–time–parameter evolution of weather systems. Perhaps the best known is the polar front theory of Bjerknes and Solberg (1922), but many others exist in the literature. These include isentropic relative flow theory (Green et al,.1966) and its applications (eg: Browning and Harrold, 1970 and Carlson, 1980), and isentropic potential vorticity (IPV) thinking as espoused by Hoskins et al. (1985).

The advantage of a good conceptual model is that it enables a meteorologist to visualise a complex atmospheric state. Such conceptual models, when applied to numerical model data, can aid the forecaster’s understanding of “why” the model is predicting a certain evolution of the atmosphere. However, forecasters should have a strong understanding of the assumptions or basis of the conceptual model they are using. For example, the conceptual model of cyclogenesis shown by Hoskins et al. (1985) (see their Fig 21) has implicit in it that the amplitude and horizontal scale of the “induced circulation” is dependent on the static stability of the atmosphere. Thus the forecaster should be aware of this additional parameter.

Forecasters should also be cautious in applying a conceptual model developed in one geographical area, based on what might have been a relatively small sample of events, to a wider sample of situations than is warranted. While the Norwegian cyclone model has been widely applied, and very successfully, for some 70 years, recent field experiments, such as the Genesis of Atlantic Lows Experiment (GALE) have revealed some distinct differences from this classic model (see Shapiro and Keyser, 1990). Although these conceptual models may be applicable over the Southern Ocean they are rarely of value over the interior of the Antarctic where few 'classic' depressions are found. In addition, recent research has shown that many of the low–pressure systems found within the circumpolar trough have non–standard frontal structure and the conceptual models used in mid–latitudes may not be applicable here.

Conceptual models should not be used to “second guess” the NWP model. An example of such a process would be for a forecaster to subjectively assess the location of vertical circulations in the model using quasi–geostrophic concepts (Bluestein, 1993) or by use of jet entrance circulation concepts (Keyser and Shapiro, 1986) and then decide that the model has forecast the vertical motion to be misplaced relative to its wind and temperature fields. The model may be in error, but its fields will be internally consistent, and so this form of thinking is flawed. A forecaster can, however, use conceptual models to understand which features resolved by the model are associated with the forecast vertical circulations. Qualitative techniques for separating the forcing of vertical motion in numerical model output into that associated with wave–scale (curvature) and jet streaks (linear acceleration) are described in Keyser and Shapiro (1986), and examples applied to real situations are described in Velden and Mills (1990) and Mills and Russell (1992). Such techniques can enhance the understanding of particular model developments.

4.2.4                                Use of independent data – assessing whether the model is “on track”

If the forecast is in error due to deficiencies in initial state specification, there is little a forecaster can do to compensate, due to the non–linearity of the models. There are cases where large errors in the analysis, due to quality control errors, result in imperceptible changes in the subsequent forecast. On the other hand, Rabier et al. (1996) describe a case where very small changes in the analysis over North America produced very large changes in the forecast over Europe in a medium range forecast. This will be especially the case in the Antarctic/Southern Ocean where the lack of data will produce analyses that are poorer than in mid–latitudes. The ensemble forecast approach (see Sivillo et al., 1997 and references therein) is being actively explored to account for uncertainty in initial conditions, but how to use this information to best assist the preparation of a forecast is yet to be determined. Therefore, if forecasters have valid doubts about the quality of the initial conditions, then their subsequent forecasts should represent this uncertainty. However, it may be very difficult to reliably assess the “quality” of the initial analysis.

One means by which the quality of a model forecast can be assessed is to compare the early fields (say up to six hours) against the locations of weather systems in satellite imagery. This will show whether the timing of the developments is essentially on track.

A numerical system will objectively produce a consistent product given the same input data. It can sometimes be difficult for forecasters to reconcile their subjective manual analysis with the numerical product, particularly with somewhere like the Antarctic where there are many local wind systems and mesoscale disturbances. However, it should be remembered that the NWP systems essentially produce synoptic–scale analyses and that few mesoscale features will be represented. This does not mean that the objective analysis is “wrong” – it has been deliberately designed this way to initialise a forecast model. The fact that an objective analysis has not directly taken into account a particular observation may be because an observation was rejected by quality control procedures, not selected, or because the three‑dimensional multivariate nature of the analysis has affected the way in which that observation was treated. Clearly a univariate analysis will fit observations more closely than a multivariate system. In addition, the background field will also affect the pattern of the analysis.

Some confidence in the quality of the analysis can be gained if the distribution of data available to the analysis is presented to the forecaster, and if those data rejected by the analysis during quality control are also flagged. This can assist a forecaster to understand why a feature is present or not present in a particular analysis. These products may not be operationally available in the Antarctic, but can and should be. Further, it is almost self–evident that if a significant component of the observational data–base is missing from a given analysis, then the forecast based on that analysis is less likely to be accurate.

A powerful way in which conceptual models can be used to infer whether a NWP model is “on track” is by developing an understanding of what atmospheric features should be present in a given situation. Mills (1997) shows that a small region of enhanced cumulus cloud that could be tracked over the Southern Ocean for some 48 h was related to a region of cyclonic vorticity advection resolved in the assimilated analyses of the Australian Bureau of Meteorology’s limited area data assimilation system. The relation between such cloud systems and cyclonic vorticity advection maxima (“CVA maxima”) has long been known (see Anderson et al.1969, Zillman and Price, 1972), and forecasters use them as surrogates for the differential vorticity advection term in the quasi–geostrophic omega equation (Bluestein, 1993). Using NWP models, this methodology can be inverted, and the CVA patterns in the NWP model can be compared with the “CVA clouds” in the satellite imagery. If the independent evidence of the satellite imagery confirms the NWP fields, as was the case in Mills (1997), then there can be some reasonable confidence in the subsequent evolution of the forecast. Another example of this type of technique is described by Mansfield (1994), where water–vapour imagery was used to qualitatively validate the IPV structure of the model atmosphere in the early stages of a forecast. In this case, correctly or incorrectly located jet streaks, as identified by the dry slots in the water–vapour imagery, were reflected in the accuracy of the later stages of medium‑range forecasts. Weldon and Holmes (1991) and Bader et al. (1995) contain many more examples comparing NWP data with satellite and radar imagery.

These two cases quoted above are only examples of this technique – all independent data can be used to validate the model in this way, and it is a powerful technique for gaining confidence in and understanding of the reasons for a NWP model’s prediction.

4.2.5                                NWP systems and the Antarctic

From the above it will be realised that while NWP systems have made great advances in recent years the emphasis has often been on improving their performance in the tropics and mid–latitude areas where most of the large centres of population are located. Indeed while there are a huge number of verification statistics on model performance in the extra–polar regions the data available for the Antarctic is relatively limited. Projects such as FROST (Turner et al., 1996a), which examined the quality of Antarctic analyses and forecasts, found that the models were fairly good over the ocean areas but lacked the mesoscale and small synoptic–scale detail that is often what concerns the forecaster. Other studies, for example Leonard et al. (1997) who examined the performance of the UK Meteorological Office model over the Antarctic, found that temperatures over the interior were too cold resulting in the katabatic winds being too strong. However, as models are developed and new parameterisation schemes introduced the nature of the errors can change since the model developers tend to check the results of their changes most carefully in the extra–polar regions.

While small, rapid developments in the Antarctic can easily be missed by the NWP systems it should not be forgotten that they can successfully predict some major storms up to a week in advance. For example, Pendlebury and Reader (1993) described the forecasting several days in advance by the ECMWF model of a very deep storm that affected Casey Station giving wind gusts up to 66.9 m s^–1 (130 kt).

In summary, a forecaster should always be aware of the strengths and weaknesses of NWP model output in the Antarctic (see, for example, Pendlebury et al. (2003). It can be extremely valuable in giving guidance on the broad–scale synoptic environment but should be used in conjunction with satellite imagery and in situ data within the forecasting process.