144 Antonio Navarra
8.6 Forecasting with ensembles
The need for ensembles changed the usual notion of the forecast as a single
deterministic realization attempting to reproduce the time march ofthe atmosphere.
Since the beginning ofmodem weathernumerical forecasting in the 50's the development has been driven by a constant attempt to push the limit and the extension of
accurate forecasts as further in the future as possible. It appeared that the improvement in the numerical technology would guarantee a smooth progress of the predictability limit. In fact, in the late 60's the possibility of using global models,
higher resolution and detailed description of the physical processes description
allowed the extension ofthe predictability limit to the medium range (7 - 10) days.
The development of sophisticated techniques of data assimilation to generate good
initial condition bas strengthened the trend and allowed the routine realization of
medium range forecasts. In the 70's a determined attempt was made to extend the
range of useful predictions to the monthly time scales, but this time the task
appeared to be much more difficult. Results showed that monthly predictions were
an erratic business in which exceptional successes were mixed with clamorous failures. Cases could be found of particular situations in which the atmosphere was
predictable for more than 2 weeks, but in general the reliability of the predictions
was low. The main reason for this difficulty was the inability of the models to follow in a deterministic way the turbulent transitions that the atmosphere is performing on that time scale.
The interpretation of the setback was soon found in the Lorenz discovery. Small
errors would amplify and it was impossible to keep them under control for a time
10nger than a week or so. It appeared that a very g100m future was in store for forecasting and that a cloud of uncertainty was going to obscure forever the future of
the climate system. But the situation was not hopeless. The limitation on predictability observed in monthly forecasting was not completely consistent with the
results from long climate simulations with prescribed SST that showed a substantial reproducibility of the atmospheric variability at time scales of a season or more.
It was necessary to reconcile the predictability limit derived from Lorenz' discovery and the potential predictability implied in the long climate experiments.
In fact the apparent paradox can be solved if we realize that there is no reason
why Lorenz' theory, namely the sensitivity to initial conditions, should apply also
to long climate simulations with prescribed SST. The unpredictability of forecasts
stems from the nature of this problem as an initial value problem, for which one has
to worry about the development of unknown components of the initial conditions
as we have seen in the preceding sections. Long climate simulation do not feeI the
effect of the initial conditions, in fact they are mostly independent of the particular
initial condition chosen, because they are boundary value problems, in which the
statistics is driven by the prescribed SST forcing.
A tight window for the possibility of forecasting at the seasonal time scale then
appear, but it is linked to the availability of accurate SST. The atmosphere will
respond to an SST distribution in a coherent and reproducible way, opening the
8.6 Forecasting with ensembles
The need for ensembles changed the usual notion of the forecast as a single
deterministic realization attempting to reproduce the time march ofthe atmosphere.
Since the beginning ofmodem weathernumerical forecasting in the 50's the development has been driven by a constant attempt to push the limit and the extension of
accurate forecasts as further in the future as possible. It appeared that the improvement in the numerical technology would guarantee a smooth progress of the predictability limit. In fact, in the late 60's the possibility of using global models,
higher resolution and detailed description of the physical processes description
allowed the extension ofthe predictability limit to the medium range (7 - 10) days.
The development of sophisticated techniques of data assimilation to generate good
initial condition bas strengthened the trend and allowed the routine realization of
medium range forecasts. In the 70's a determined attempt was made to extend the
range of useful predictions to the monthly time scales, but this time the task
appeared to be much more difficult. Results showed that monthly predictions were
an erratic business in which exceptional successes were mixed with clamorous failures. Cases could be found of particular situations in which the atmosphere was
predictable for more than 2 weeks, but in general the reliability of the predictions
was low. The main reason for this difficulty was the inability of the models to follow in a deterministic way the turbulent transitions that the atmosphere is performing on that time scale.
The interpretation of the setback was soon found in the Lorenz discovery. Small
errors would amplify and it was impossible to keep them under control for a time
10nger than a week or so. It appeared that a very g100m future was in store for forecasting and that a cloud of uncertainty was going to obscure forever the future of
the climate system. But the situation was not hopeless. The limitation on predictability observed in monthly forecasting was not completely consistent with the
results from long climate simulations with prescribed SST that showed a substantial reproducibility of the atmospheric variability at time scales of a season or more.
It was necessary to reconcile the predictability limit derived from Lorenz' discovery and the potential predictability implied in the long climate experiments.
In fact the apparent paradox can be solved if we realize that there is no reason
why Lorenz' theory, namely the sensitivity to initial conditions, should apply also
to long climate simulations with prescribed SST. The unpredictability of forecasts
stems from the nature of this problem as an initial value problem, for which one has
to worry about the development of unknown components of the initial conditions
as we have seen in the preceding sections. Long climate simulation do not feeI the
effect of the initial conditions, in fact they are mostly independent of the particular
initial condition chosen, because they are boundary value problems, in which the
statistics is driven by the prescribed SST forcing.
A tight window for the possibility of forecasting at the seasonal time scale then
appear, but it is linked to the availability of accurate SST. The atmosphere will
respond to an SST distribution in a coherent and reproducible way, opening the