362
DAVID ANDERSON ET AL.
forecast time or the ensemble size. Thus, for example, the deterministic
forecast is made using a horizontal resolution of T511 (~40km), the EPS
uses T255, the monthly uses T159, and the seasonal uses T95.
A further complication is that to make forecasts beyond a week or two
requires information on the state of the ocean as well as that of the
atmosphere. Forecasts of the monthly and seasonal range are made with
coupled atmosphere-ocean models. The longer the range, the less important
are the atmospheric initial conditions, but the more important the ocean
initial conditions become (generally speaking). There might be a case for
having an active ocean even for the short to medium range forecasts, as sea
surface temperature changes associated with tropical cyclones and storm
tracks might be important. This latter point requires confirmation, but
regardless of whether the ocean impacts the atmosphere on the medium
range, it is necessary to perform real-time ocean analyses for the monthly
and seasonal forecast systems. At these longer forecast ranges, model error
is sufficiently large that it cannot be ignored. One way to account for model
error is to run the model for many realizations in the past to estimate the
model climate and climate drift. This in turn brings in a need for extended
analyses into the past as well as the present. ECMWF is thus heavily
involved in reanalysis, both atmospheric and oceanic. It will be argued later
that in fact reanalysis is an integral part of a forecast system, (at least for
monthly and seasonal timescales and possibly also for medium range).
2.
Atmospheric analyses
The atmospheric analysis is done using 4d-var (as discussed by F. Rabier
in this volume). The 4 indicates the use of time as the 4th dimension. So the
analysis is performed over a space-time box. It is performed over the
highest resolution possible (this will be discussed later), producing an
analysis which has the resolution of the first guess. The ECMWF model is
formulated in spectral space rather than physical space i.e. in terms of
spherical harmonics rather than grid points. In spectral space the resolution
is T511 (about 20 km). The effective resolution is not this high, however.
The T511 analysis is then used to provide initial conditions for the EPS,
monthly and seasonal forecasts by truncating the analysis in spectral space to
the appropriate resolution.
In addition to analyses made in near-real-time for the purpose of
generating forecasts, analyses are made of past events as well, called
reanalyses. Extended atmospheric reanalyses are major undertakings
requiring years of dedicated work. So they are not undertaken lightly: in
fact, at ECMWF, there have been only two to date. The first, denoted ERA15 spanned the 15 years 1979-1993, and the second, denoted ERA-40,
spanned the 40+ years Sept 1957- Aug 2002 (Uppala et al 2005). The
analysis system used in these reanalyses should be the same throughout. For
ERA-15, the scheme was OI, while for ERA-40 it was 3dvar. The expense
DAVID ANDERSON ET AL.
forecast time or the ensemble size. Thus, for example, the deterministic
forecast is made using a horizontal resolution of T511 (~40km), the EPS
uses T255, the monthly uses T159, and the seasonal uses T95.
A further complication is that to make forecasts beyond a week or two
requires information on the state of the ocean as well as that of the
atmosphere. Forecasts of the monthly and seasonal range are made with
coupled atmosphere-ocean models. The longer the range, the less important
are the atmospheric initial conditions, but the more important the ocean
initial conditions become (generally speaking). There might be a case for
having an active ocean even for the short to medium range forecasts, as sea
surface temperature changes associated with tropical cyclones and storm
tracks might be important. This latter point requires confirmation, but
regardless of whether the ocean impacts the atmosphere on the medium
range, it is necessary to perform real-time ocean analyses for the monthly
and seasonal forecast systems. At these longer forecast ranges, model error
is sufficiently large that it cannot be ignored. One way to account for model
error is to run the model for many realizations in the past to estimate the
model climate and climate drift. This in turn brings in a need for extended
analyses into the past as well as the present. ECMWF is thus heavily
involved in reanalysis, both atmospheric and oceanic. It will be argued later
that in fact reanalysis is an integral part of a forecast system, (at least for
monthly and seasonal timescales and possibly also for medium range).
2.
Atmospheric analyses
The atmospheric analysis is done using 4d-var (as discussed by F. Rabier
in this volume). The 4 indicates the use of time as the 4th dimension. So the
analysis is performed over a space-time box. It is performed over the
highest resolution possible (this will be discussed later), producing an
analysis which has the resolution of the first guess. The ECMWF model is
formulated in spectral space rather than physical space i.e. in terms of
spherical harmonics rather than grid points. In spectral space the resolution
is T511 (about 20 km). The effective resolution is not this high, however.
The T511 analysis is then used to provide initial conditions for the EPS,
monthly and seasonal forecasts by truncating the analysis in spectral space to
the appropriate resolution.
In addition to analyses made in near-real-time for the purpose of
generating forecasts, analyses are made of past events as well, called
reanalyses. Extended atmospheric reanalyses are major undertakings
requiring years of dedicated work. So they are not undertaken lightly: in
fact, at ECMWF, there have been only two to date. The first, denoted ERA15 spanned the 15 years 1979-1993, and the second, denoted ERA-40,
spanned the 40+ years Sept 1957- Aug 2002 (Uppala et al 2005). The
analysis system used in these reanalyses should be the same throughout. For
ERA-15, the scheme was OI, while for ERA-40 it was 3dvar. The expense