 |
| Titel |
Decadal Climate Predictions Using Sequential Learning Algorithms |
| VerfasserIn |
Ehud Strobach, Golan Bel |
| Konferenz |
EGU General Assembly 2015
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 17 (2015) |
| Datensatznummer |
250103015
|
| Publikation (Nr.) |
 EGU/EGU2015-2416.pdf |
|
|
|
|
|
| Zusammenfassung |
| Ensembles of climate models may improve future climate predictions and reduce their
uncertainties. We adopt two Sequential Learning Algorithms (SLAs) in order to establish a
weighted ensemble of models based on their past performance. The method has several
advantages, including the lack of a priori assumptions regarding the ensemble members; the
weight of the ensemble members can be updated upon the arrival of new measurements; and
it can be proved that the prediction of the weighted ensemble is at least as good as the
prediction of the best model if the learning period is long enough. Using this method, we aim
to achieve two main goals - provide better future climate predictions and reduce their
uncertainties.
The goal of SLAs is to minimize the cumulative regret with respect to each
one of the climate models. This is defined, for expert E, after n time steps, by the
quantity:
/n
RE,n ≈¡ l(pt,yt)- l(fE,t,yt) ≈¡ Ln - LE,n
t=1
(1)
where, l is the loss function - a measure of the difference between the predicted and the true
values; yt is the true value at time t; pt is the predicted value by the algorithm at time t; fE,t
is the predicted value by the expert E at time t; and L is the cumulative loss function. The
outcome of the algorithm is a weighted average of the climate models in the ensemble, that
is:
/N
pt ≈¡ WE,t- 1(RE,t-1)/
fE,t
E=1
(2)
where, WE,t-1 is the weight of expert E based on the regret up to time t - 1.
We adopted one 30-year CMIP5 experiment (between 1981-2011) to demonstrate the
ability of the SLAs to fulfill the two goals. Four variables were considered – monthly
averages of surface temperature, U and V components of 10m wind speed, and surface
pressure. The CMIP5 hindcast data was divided to a learning period in which the weight of
the models was updated in each time step and a verification period in which the weights of
the models are constant and the weighted average of the models was tested against
measurements.
It will be shown that one of the two algorithms is more suitable for decadal climate
predictions. This algorithm has the ability to improve climate predictions by a few
percent, compared to the predictions made according to a reference climatology
as well as predictions generated by the the commonly used regression method.
Also, the algorithm has the ability to reduce the uncertainties associated with the
predictions by more than 50% compared to the simple average of the model results. |
|
|
|
|
|
|