Ensembles, Forecasts and Predictability
141
bilities by Farrell (1988) are very suited in this case (Molteni et al., 1996) but other
choices are possible (see the breeding technique by Toth, 1993). Whatever the
technique to generate the perturbations, the net result is that the single forecast is
substituted by a bundle of forecasts, the ensemble, that takes its place. Ensembles
can give more informations, because they not only provide the deterministic forecast but also probabilistic information about the expected reliability and can give a
sense the possible alternative evolutions.
The need for ensembles is also found in longer climate simulations that do not
have the goal of practical forecasts. Usually these kinds of simulations have the
objective of reproducing the statistical properties of the atmospheric and ocean
dynamics, rather than a point by point time evolution. In this regard one might
argue that they allow a lower standard, but on the other hand they require a much
higher standard on the budget of quantities, like energy, radiation and salinity that
result from delicate balances. Small errors in the radiation balance can be tolerated
for a model designed for ten-days forecasts, but they are not acceptable in a model
designed to simulate for hundreds ofyears.
The longer simulations have also been crucial to reveal another source of variability and statistical behaviour. These kind of simulations are usually used to simulate the general circulation and the variability of the climate system in general.
The objective is here to reproduce as faithfully as possible the statistics ofthe system, yielding the right seasonal averages and an accurate balance of conserved
quantities. In this regard it is expected that initial conditions are not very important
because the model wiH produce its own equilibrium state, in practice initial conditions corresponding to random dates are often used. For instance, for a simulation
starting in winter an arbitrary winter day is selected. Coupled models are often used
in these experiments, but sometimes atmospheric GCM forced by prescribed Sea
Surf ace Temperatures (SST) are also used.
Very long simulations performed with prescribed SST revealed a dynamical
complexity that was somewhat unexpected. The atmospheric variability was found
to be sensitive to the external forcing caused by the varying distribution of SST, but
the response was not strictly consistent. Simulations that differed only in a choice
of initial condition, surely well beyond the deterministic range and therefore forgotten after months and years of simulations, were stiH producing different variability events. The reason lies in the turbulent structure of the atmosphere.
Referring to Fig. 8.3 it is possible to see how the response to an equatorial SST
usually consists in upward motion that is strongly correlated with local precipitation. The detailed shape and intensity of the vertical motion is a function of the
basic state according to a mechanism first proposed by GiH (1980). The response
can be different even ifthe same SST is used for different realizations. An adequate
number of realizations is necessary to identify the part of the response that is obviously linked to the SST. Simple ways include averaging over the ensemble (ensembIe mean), separation of variance (Rowell, 1997; Ward and Navarra, 1997),
methods based on the decomposition of variance (Wallace and Gutzler, 1981;
141
bilities by Farrell (1988) are very suited in this case (Molteni et al., 1996) but other
choices are possible (see the breeding technique by Toth, 1993). Whatever the
technique to generate the perturbations, the net result is that the single forecast is
substituted by a bundle of forecasts, the ensemble, that takes its place. Ensembles
can give more informations, because they not only provide the deterministic forecast but also probabilistic information about the expected reliability and can give a
sense the possible alternative evolutions.
The need for ensembles is also found in longer climate simulations that do not
have the goal of practical forecasts. Usually these kinds of simulations have the
objective of reproducing the statistical properties of the atmospheric and ocean
dynamics, rather than a point by point time evolution. In this regard one might
argue that they allow a lower standard, but on the other hand they require a much
higher standard on the budget of quantities, like energy, radiation and salinity that
result from delicate balances. Small errors in the radiation balance can be tolerated
for a model designed for ten-days forecasts, but they are not acceptable in a model
designed to simulate for hundreds ofyears.
The longer simulations have also been crucial to reveal another source of variability and statistical behaviour. These kind of simulations are usually used to simulate the general circulation and the variability of the climate system in general.
The objective is here to reproduce as faithfully as possible the statistics ofthe system, yielding the right seasonal averages and an accurate balance of conserved
quantities. In this regard it is expected that initial conditions are not very important
because the model wiH produce its own equilibrium state, in practice initial conditions corresponding to random dates are often used. For instance, for a simulation
starting in winter an arbitrary winter day is selected. Coupled models are often used
in these experiments, but sometimes atmospheric GCM forced by prescribed Sea
Surf ace Temperatures (SST) are also used.
Very long simulations performed with prescribed SST revealed a dynamical
complexity that was somewhat unexpected. The atmospheric variability was found
to be sensitive to the external forcing caused by the varying distribution of SST, but
the response was not strictly consistent. Simulations that differed only in a choice
of initial condition, surely well beyond the deterministic range and therefore forgotten after months and years of simulations, were stiH producing different variability events. The reason lies in the turbulent structure of the atmosphere.
Referring to Fig. 8.3 it is possible to see how the response to an equatorial SST
usually consists in upward motion that is strongly correlated with local precipitation. The detailed shape and intensity of the vertical motion is a function of the
basic state according to a mechanism first proposed by GiH (1980). The response
can be different even ifthe same SST is used for different realizations. An adequate
number of realizations is necessary to identify the part of the response that is obviously linked to the SST. Simple ways include averaging over the ensemble (ensembIe mean), separation of variance (Rowell, 1997; Ward and Navarra, 1997),
methods based on the decomposition of variance (Wallace and Gutzler, 1981;