322 Martin Fischer
above. The simulated SST anomalies are corrected by this interface before they are
pas sed to the statistical atmosphere model.
The coupled model simulates a regular ENSO cycle with a period of approximately five years, which was inferred from a multi decadal control run (Latif and
Fliigel 1991, Fliigel 1994). This is only one possible approach for a statistical
atmosphere. Other statistical methods can be used and also non linear approaches
may be useful. This simple atmospheric model is then coupled to a limited domain
primitive equation model of the tropical Pacific. Useful skill is achieved for lead
times up to one year. A problem with this kind of models is that the atmospheric
feedback has to be determined from data, that means from the past. If the relation
between SST and wind stress anomalies changes for some reason, the prediction
skill of these models drops dramatically. This happens for instance during the early
nineties. It seems that during that time the mean state ofthe Pacific ocean is different from the state in the eighties, and thus also the feedback between ocean and
atmosphere changed. The parameters of the statistical atmosphere described above
were determined from data from the seventies and eighties. For this period the coupled model was successful in forecasting the evolution of the ENSO cycle. In the
nineties, however, the forecast skill at a lead time of one year dropped from about
0.7 for the eighties to about 0.2 to 0.3.
Changes of the mean state or of feedback mechanisms are a general problem for
every statistical method. To catch them sophisticated statistical methods would be
necessary and very long observational time series would be needed to obtain reliable estimates of the parameters. However, most ENSO related observations do not
go further back than 1950, and thus are too short. The only way to overcome this
problem is to use fully physical models which are presented in the following section.
Coupled general circulation models
Coupled general circulation models are the most complex models applied to
ENSO forecasting. Such models have been developed over the last 10 to 15 years
by scientific groups all over the world. Examples are the ECHO model (ECHAM
(ECMWF Hamburg Model) + HOPE (Hamburg Ocean Primitive Equation model))
(Frey et al. 1997), the GFDL model (Rosati et al. 1996), the COLA model (Kirtman et al. 1997), or the NCEP model (li and Leetmaa 1996) to name just a few
coupled general circulation models. Usually they consist of a global atmosphere
general circulation model and a global ocean general circulation model. The coupling is done by forcing the ocean model with momentum fluxes (wind stress), heat
fluxes, and fresh water fluxes as computed by the atmospheric model, and the SST
computed by the ocean model is used as lower boundary condition for the atmosphere. The time intervals to exchange informations between the two models typically range from a few hours up to one day. Thus, the only external forcing that is
needed to run a coupled general circulation model is the solar radiation which may
be easily computed even for the next millennium (Fig. 16.16). The advantage of
coupled general circulation models relative to statistical approaches, either in
above. The simulated SST anomalies are corrected by this interface before they are
pas sed to the statistical atmosphere model.
The coupled model simulates a regular ENSO cycle with a period of approximately five years, which was inferred from a multi decadal control run (Latif and
Fliigel 1991, Fliigel 1994). This is only one possible approach for a statistical
atmosphere. Other statistical methods can be used and also non linear approaches
may be useful. This simple atmospheric model is then coupled to a limited domain
primitive equation model of the tropical Pacific. Useful skill is achieved for lead
times up to one year. A problem with this kind of models is that the atmospheric
feedback has to be determined from data, that means from the past. If the relation
between SST and wind stress anomalies changes for some reason, the prediction
skill of these models drops dramatically. This happens for instance during the early
nineties. It seems that during that time the mean state ofthe Pacific ocean is different from the state in the eighties, and thus also the feedback between ocean and
atmosphere changed. The parameters of the statistical atmosphere described above
were determined from data from the seventies and eighties. For this period the coupled model was successful in forecasting the evolution of the ENSO cycle. In the
nineties, however, the forecast skill at a lead time of one year dropped from about
0.7 for the eighties to about 0.2 to 0.3.
Changes of the mean state or of feedback mechanisms are a general problem for
every statistical method. To catch them sophisticated statistical methods would be
necessary and very long observational time series would be needed to obtain reliable estimates of the parameters. However, most ENSO related observations do not
go further back than 1950, and thus are too short. The only way to overcome this
problem is to use fully physical models which are presented in the following section.
Coupled general circulation models
Coupled general circulation models are the most complex models applied to
ENSO forecasting. Such models have been developed over the last 10 to 15 years
by scientific groups all over the world. Examples are the ECHO model (ECHAM
(ECMWF Hamburg Model) + HOPE (Hamburg Ocean Primitive Equation model))
(Frey et al. 1997), the GFDL model (Rosati et al. 1996), the COLA model (Kirtman et al. 1997), or the NCEP model (li and Leetmaa 1996) to name just a few
coupled general circulation models. Usually they consist of a global atmosphere
general circulation model and a global ocean general circulation model. The coupling is done by forcing the ocean model with momentum fluxes (wind stress), heat
fluxes, and fresh water fluxes as computed by the atmospheric model, and the SST
computed by the ocean model is used as lower boundary condition for the atmosphere. The time intervals to exchange informations between the two models typically range from a few hours up to one day. Thus, the only external forcing that is
needed to run a coupled general circulation model is the solar radiation which may
be easily computed even for the next millennium (Fig. 16.16). The advantage of
coupled general circulation models relative to statistical approaches, either in