Use of Ensemble Forecasting at Météo-France

advertisement
USE OF ENSEMBLE FORECASTING AT MÉTÉO-FRANCE
(J. Nicolau, Météo-France)
Météo-France (SCEM/PREVI)
42, Avenue G. Coriolis , 31057 Toulouse Cedex
Phone : 33 5 61 07 85 32; Fax : 33 5 61 07 84 53: E-mail : jean.nicolau@meteo.fr
Summary
Intuitively, the idea that a weather forecast does not have the same reliability of one day to the
other, according to the weather situation, is well known as well by our users as by meteorologists.
For a few years, methods of numerical prediction have made possible quantitatively to estimate
this reliability (one takes often the opposed formulation, that of the estimate of uncertainty) at the
time where the forecast is performed and not only a posteriori. The main tool allowing estimating
such information is the Ensemble Forecasting. In the first part, we are presenting how and why
such tool is used at Météo-France in the medium-range weather forecasting, while insisting on the
influence which it can have as well on the manner of working of the forecaster as on the products
of the prediction. In a second part some other applications of the Ensemble Prediction System are
described.
I.
Use of EPS in the Medium-Range weather forecasts
The limit of the synoptic methodology
The weather forecast is at once based on the identification of meteorological patterns.
Traditionally, the forecaster is using the synoptic scale characterised by well defined structures
such as disturbances or fronts. The forecaster has to predict the evolution of such structures in
order to give sensitive weather. This prediction is based on the use of numerical model. These
models regularly controlled, show that the forecast skill decreases with the range. Up to Day 4
less than 50% of forecasts are good and up to Day 6 almost 90% are wrong. In this control, the
synoptic scope is judged in its consequences on the sensitive weather over France. The first point
to note is that models are inconsistent from each other and changeable from one day to another.
The supra-synoptic scale
As predictability is directly linked to the range and to the spatial scale of the prediction, the synoptic
guidance does not seem to be relevant up to Day 4. A new scale, the supra synoptic scale, is then
considered. In this scale, synoptic features are not taken into account as complex meso-scale
features for the synoptic scale. For instance, the occurrence time and the precise location of a
perturbation can't be realistically predicted. For that reason this new scale will be characteristic of
the global circulation.
Weather types
In order to elaborate the weather forecast, it is necessary to associate for each "supra" pattern an
ensemble of weather parameters. Of course traditional synoptic weather parameters such as
precipitation can't be associated to such pattern. With the supra-synoptic scale, a certain number
of flow types slowly variable over a period of 2/3 days are identified. From these flow types, it is
possible to determine the main characteristics of the weather. Four main flow types have been
defined:

Straight Flow : rapid, straight, perturbed, rainy often windy.

Undulating Flow : alternating ridge and trough, unsettled weather, a rainy period is followed
by a sunny one. There is at least one rainy period.
2

Warm Blocked Flow : anticyclonic conditions. Dry weather, sunny in summer, sometimes
foggy or dull weather in winter.

Cold Blocked Flow : low-pressure conditions, cool or cold weather, often rainy.
Estimating the uncertainty : The Ensemble Prediction System (EPS)
The ensemble forecast is a set of integrations of a deterministic numerical weather prediction
(NWP) model; these integrations differ only in their initial states, reflecting uncertainties in such
starting conditions. In essence, the ensemble prediction is an attempt to estimate the nonlinear
evolution of the forecast error probability function (PDF) through a finite sample of deterministic
forecast integrations. One of the fundamental difficulties in this approach lies in the choice of initial
conditions for the individual members of the ensemble. From the definition above, these must
reflect an adequate sampling of the probability distribution function of analysis error. However, the
phase-space dimension of current NWP models is well in excess of a million. Actually only a few
perturbations (initial states modifications) will be retained. Hence, with a limited ensemble size, it
is important to sample those trajectories starting within the initial error balls, which evolve to
significantly different large-scale flow patterns, such as weather regimes.
The operational EPS at ECMWF is presently based on the use of 25 perturbations defined from the
Singular Vectors method. These perturbations are added and retracted from the initial state. Thus
a set of 51 predictions (50 perturbed forecasts and one unperturbed, the so-called control forecast)
is obtained. This system is running daily to a 240 hours range.
The EPS classification : the tubing
The huge quantity of information in the Ensemble Prediction System makes its direct use very
difficult. Tools have been developed in order to condense the information coming from the 51
ensemble forecasts : the tubing (figure 1) is a method designed to classify ensemble forecasts.
Forecasts that are similar to the ensemble mean are grouped into one cluster, the so-called central
cluster. The other forecasts are grouped into a few tubes indicating different deviations from the
ensemble mean. Each tube is represented by its extreme, i.e. the forecast which is the most
different from the ensemble mean in the direction of the tube. The extremes of the tubes are not
alternatives; they are extreme representatives of the tubes, almost caricatures, allowing to better
visualise the different tendencies present in the ensemble, by contrast with the central cluster
mean forecast.
The 7 days weather forecast
Every day, the medium range forecaster of the Central Forecast Department in Toulouse, is
provided with three sets of the tubing classification of the EPS, representing the central cluster
mean and a variable number of tubes (generally less than 5 or 6). Each set are representative for
a period of 72 hours, based on +96h, +144h and +192h forecasts. The fields used are 500hPa
height and temperature, mean sea level pressure and 850 hPa temperature, on North Atlantic and
Western Europe area. Other deterministic fields and probabilistic fields of weather parameters like
temperature anomalies and precipitation amounts are also used. The forecaster describes the
most likely forecast evolution by interpreting the large scale patterns of the central cluster mean.
He(she) supplies a "supra-synoptic scenario" for D4/D5 (D+5/D+6 model range) and the evolution
expected for D6/D7 (D+7/D+8 model range), with associated comments, and a graphical
document, the so-called PRESSYME (PREvision Supra SYnoptique Moyenne Echéance) for each
of these two periods. The association between flow types and weather characteristics is based on
the expertise of the forecasters (more or less a climatological knowledge). A weather types
automatic classification of all of the ensemble members is provided to the forecaster. The kind of
weather expected for France is then defined in the technical guidance, with more precision for
parameters like wind, temperatures and rain intensity. The precise description given on a daily
basis for the weather characteristics may vary for a same flow type, depending on the season, the
3
region (quarters of France typically) and on the different flow types that have been identified (it is
not an automatic transcription of the flow type analysis). It is here necessary to underline the way
of interpreting the central cluster mean. This field represents more or less predictable features. It
plays the role of a clever filter: some details described by T159 members are rubbed by this way
and that means that such details are not predictable. Of course the predictability is not the same
one day to the other and the filtering degree is varying with the spread of the ensemble. If the
spread is small, then the mean will describe precise details (in practice the equivalent truncation of
the central cluster mean will vary between T20 and T10).
Recently a new graphical product has been added to the technical medium range guidance, for the
national media, the so-called "Nébule MEDIA" (figure 2). This product includes information on
weather tendencies and some risks (strong winds, storms, snow).
In addition, a confidence index is defined by the forecaster. This confidence index is linked to the
number of tubes, which are, in term of weather parameters over the area of interest, significantly
different from the central cluster mean. It is then evaluated in a subjective way, on a scale of 1 to
5:



4 if the confidence is strong (0 or 1 tube)
3 if the confidence is normal (2 or 3 tubes)
2 if the confidence is weak (more than 3 tubes)
Values of 1 and 5 are rarely used and only in very specific cases: very strong confidence, which is
more often very difficult to appreciate, or on the opposite, very weak confidence due to a great
number of tubes with significant variants over the country. Finally, the confidence index can be
modified by the forecaster according to the kind of weather expected. For example, the forecaster
may have a very strong confidence for a warm block, but a lower confidence for the corresponding
kind of weather (generally fair), due to a possible presence of low clouds, especially in winter.
The medium range weather forecast evaluation
Subjective evaluation of the forecaster weather flow analysis is conducted in a daily basis. Four
marks are used :

A: very good forecast.

B: good forecast, large scale circulation well forecasted, small differences on chronology or
spatial localization of forecast flow patterns.

C: rather bad, not necessarily a bad forecast of large scale circulation but differences are
too important on France.

D: bad forecast of large circulation.
If A and B forecast are considered to be "good" forecasts, the percentage of good forecasts on the
past year was : 87 % for D4-D5 and 59% for D6/D7. Considering the correlation between
confidence index and skill forecast, it appears that when the confidence index is high the forecast
is good, while when the confidence index is low, the correlation with bad forecasts is not so
evident. As a consequence it can be said that forecasters are rather inclined to "overprotect" but
also that there is still too many bad forecasts which are not detected.
Finally the 7 days weather forecasts give good results in terms of scores but the public perception
for such forecast is not really positive. Firstly the information given seems to be too vague,
secondly the uncertainty quantification (confidence index) is perhaps not the appropriate way to
give the weather predictability because there is a lot of bad interpretations about it.
4
II.
Use of EPS for professional applications
EPS allows also probabilistic prediction for weather parameters. The probability of an event can
be directly calculated as being the proportion of members of the ensemble for which this event
occurs (figure 3). However it appears that such probabilities need to be readjusted (we will speak
about "calibration") in reliability. For instance, for the event: "24h-precipitation 1 mm exceeding
over Toulouse", for the cases where the probability is 80%, the effective frequency of this event is
only 60%. It thus will be required to gauge these probabilities according to a "climatology" of the
forecasted probabilities, particularly since the request for these products is increasing on behalf of
the professionals (agriculture, buildings, transport, energy...).
Calibration
Generally, raw probabilities are not reliable. Model bias (for example rainfall over-estimating) or
the lack of statistical consistency of the ensemble (weak spread) can be at the origin of such
problem. This poor reliability is linked to the fact that the hypothesis is equi-probability of all of the
ensemble members. A very simple method to calibrate probabilities is based on the use of the
reliability diagram. The calibrated probability can be obtained by the observed frequency (y-axis)
corresponding to the raw probability (x-axis) on the reliability curve. This method has a strong
disadvantage: it uses the reliability diagram, which depends on the threshold.
The method used here is valid whatever the threshold. It was already used by Hamill et Colucci
(1997, 1998) and Eckel et Walters (1998). This method is based on rank diagrams (also called
"Talagrand diagrams", figure 4). The 51 values of the EPS are ordered from the lowest to the
highest. It is then possible to determine the bin including the considered threshold. The calibrated
probability corresponds to the cumul of all of the left side bins of the rank diagram before the
threshold's bin. This method allows increasing efficiently the reliability for parameters such as
precipitation (figure 5).
Statistic adaptation
The calibration method is not really satisfactory for such parameters as minimum and maximum
temperatures. For these particular parameters, a statistical adaptation method based on a multilinear regression is presently used. This method increases the results in term of reliability but also
in term of resolution, but acts in reducing the spread of the ensemble. In order to solve this
problem a first trial was undertaken. This trial consisted in applying a translation based on the
previous spread (before regression) to all the regressed members. Finally the best results were
obtained in combining the regression and the calibration methods (figure 6).
Products
With such methods, Météo-France can now supply professionals with probability products. One of
the most popular is the so-called Meteogram which provides the evolution of Maximum and
Minimum temperatures on one point with the ensemble mean value, the minimum and maximum
values given by extreme members and the uncertainty given by the interval including 75% of the
members. Of course, at Medium-Range such point is representative of a region. The risk of
extreme events such as heavy rain can also be evaluate by probability charts.
III.
Special applications of EPS
Two experiments based on the use of the EPS are presented here.
Low pressure centers tracking
The first one concerns the tracking of the cyclone "Connie", which passed close to the French
island of La Réunion (Indian Ocean) on the 25th of January 2000. A selection on low pressure
5
centers was done over the ensemble members in order to see the evolution of the cyclone from the
early to the medium range. The trajectory was well seen in the first steps of the forecast but
rapidly lost by the ensemble (figure 8).
Météo-France is waiting for the promising diabatic singular vectors computed over the tropical
areas to improve such trackings.
Model coupling
The second one is related to the dramatic Erika oil slick. On December 12, the tanker Erika was
wrecked near the Brittany coast. In order to track the pollutant, Météo-France used its model of
pollutant drift ("Mothy"). This model uses usually the wind speed and direction from a global
atmospheric model. The idea was here to get the more pessimistic evolution in terms of wind
speed and directions (i.e. towards the French coast) from the ensemble in order to evaluate the
earlier date of the pollutant arrival on the coast (figure 9). With this kind of forecast, it was
therefore possible to announce that there was no risk for the pollutant to arrive on the coast before
the 23th of December. This information was a posteriori verified and allowed the authorities to take
some measures (such as floating dams). The pollutant arrived on the coast on the 25th of
December after 2 weeks of daily tracking of the trajectory by Météo-France.
References
Eckel F. A. and M. K. Walters, calibrated probabilistic quantitative precipitation forecasts based on
the MRF ensemble. Weather and Forecasting, vol 13, 1132-1147, 1998.
Hamill T. S. and S. J. Colucci, verification of ETA-RSM short-range ensemble forecasts. Monthly
Weather Review. vol 125, 1312-1327, 1997.
Hamill T. S. and S. J. Colucci, evaluation of ETA-RSM ensemble probabilistic precipitation
forecasts. Monthly Weather Review. vol 126, 711-724, 1998.
Download