| |
Ensemble Prediction |
 |
|
|
Numerical models and data assimilation systems have improved enormously
over recent years so that today's 3-day forecast is as good as a 1-day
forecast 20 years ago. Despite this, an NWP forecast looking a few days
ahead can frequently be quite wrong, and even 1-day forecasts can
occasionally have large errors. The reason for this lies in the chaotic
nature of the atmosphere, which means that very small errors in the
initial conditions can lead to large errors in the forecast, the
so-called butterfly effect. This means that we can never create a
perfect forecast system because we can never observe every detail of
the initial state of the atmosphere. Tiny errors in the initial state
will be amplified such that after a period of time the forecast becomes
useless. This sensitivity varies from day to day, but typically we can
forecast the main weather patterns reasonably well up to about 3 days
ahead. Beyond that uncertainties in the forecasts can become large.
To cope with this uncertainty, we use Ensemble Forecasts. Instead of
running just a single forecast, the model is run a number of times from
slightly different starting conditions. The complete set of forecasts is
referred to as the ensemble and individual forecasts within it as
ensemble members. The initial differences between ensemble members are
very small so that if we compared members with observations it would be
impossible to say which members fitted the observations better.
All members are therefore equally likely to be correct, but when
we look several days ahead the forecasts can be quite different. Some
days the forecasts from different ensemble members are all quite
similar, which gives us confidence that we can issue a reliable
forecast. On other days the members can differ radically and then we
have to be more cautious. Click here
for some further illustrations of the concept of ensemble prediction.
As an illustration of the sensitivity, the following charts show an
example of two equally valid 4-day forecasts of the surface pressure
(isobars) from an ensemble forecast. Differences at the start of the
forecast, in the top row, are so small that we cannot tell which is more
accurate, but the forecasts below are very different! (In reality, of
course, the weather systems over the British Isles at Day 4 would
probably have originated further west at the start of the forecast, but
it illustrates how very similar ensemble members grow apart during the
forecast.)
Forecast A on the left predicts a deep area of low pressure over
Ireland bringing strong winds and rain to much of the British Isles;
forecast B on the right predicts that the high pressure over the
Atlantic will be much stronger and does not develop the low at all, and
thus suggests fine weather although with a cool northerly wind and the
risk of showers in the S and E. Clearly in this situation a forecaster
who has access to only a single model forecast is in danger of issuing a
forecast which could go seriously wrong. By using this sort of
information from an ensemble with many members, Met Office forecasters
are able to assess the range of possible scenarios and issue
advice on the probabilities and risks associated with them. Ensemble
Prediction is thus all about Risk Management in weather forecasting.
|
|
|
The Met Office does not run its own medium-range
ensemble forecasts, but makes use of the Ensemble Prediction System
(EPS) run by the European Centre for
Medium-Range Weather Forecasts (ECMWF). ECMWF is an international
organisation supported by 25 European states, including the UK, and
specialises in NWP for medium-range prediction. ECMWF does not issue
weather forecasts itself, but distributes its products to the National
Meteorological Services of its member states, including the Met Office,
for use in production of weather forecasts.
Back to top
|
| The ECMWF
Ensemble Prediction System |
|
The ECMWF EPS consists of 51 forecasts run twice daily using the ECMWF
global forecast model with a horizontal resolution of around 80km. One
member, called the control forecast, is run directly from the ECMWF
analysis, our best guess at the initial state of the atmosphere.
Initial conditions for the other 50 members are created by adding small
"perturbations" to this analysis. These perturbations are designed to
identify those regions of the atmosphere which are most likely to lead
to errors in the forecast on each particular occasion. Small random
variations in the model itself are also introduced to allow for some of
the approximations which have to be made in how the model represents
the atmosphere.
The charts below show an example of surface pressures
(isobars) for all 51 members of the ECMWF ensemble for a sample 4-day
forecast from November 2003. On many occasions the atmosphere is much
more predictable than this, but this illustrates the level of
uncertainty that regularly occurs in forecasts only a few days ahead.
Clearly an ensemble forecast contains a huge amount of information
which we need to condense for both forecasters and end-users! Below we
describe some of the ways we can do this, including the use of Probability Forecasts. More about the EPS from
the ECMWF user guide
Back to top
|
| Describing
uncertainty in forecasts |
|
Ensemble prediction allows the uncertainty in forecasts to be assessed
quantitatively. This uncertainty can be passed on to users of the
forecast in several ways.
For example, we can provide a range of possible values for a forecast
parameter (such as temperature or windspeed) such that we know how
confident we are that the actual value will fall within that
range. For the scientist this is very similar to putting an error
bar on the forecast. The example below shows how maximum and minimum
temperatures for each day can be given a range of uncertainty. The full
length of each vertical line represents the 95% confidence range, while
the central bar represents a 50% confidence range. The horizontal line
across this bar is the mid-point of the distribution, and may be used to
estimate the most likely temperature.Thus for the first night we can be
95% certain the minimum temperature will be between 8 and 13 Celsius,
and 50% certain it will be between about 11 and 12 Celsius.
Alternatively we can estimate the probability of certain events
happening, for example of the temperature falling below freezing or the
wind speed reaching gale force. Probability forecasts can help
users to assess the risks associated with particular weather events
which are important to them.
Back to top
|
| The Met
Office Ensemble Post-Processing System (Previn) |
|
Ensemble forecasts from ECMWF are post-processed to produce a wide
variety of chart displays to aid forecasters, and also high-quality
probability forecasts which can be supplied to customers. Two examples
illustrate how these can be used.
A daily task for our weather forecasters is to assess the most likely
developments of major weather systems for several days ahead - most of
our forecast products depend on them getting this right! Among the most
important systems for NW Europe are Atlantic low pressure centres
(cyclones). To aid the forecasters in assessing the most likely
positions and movements of lows, cyclone tracks predicted by all the
ensemble members are plotted on a single chart. The example below for a
3-day forecast in February 2004 shows that it is most likely that a low
pressure will move northeastwards to the west of the British Isles.
However for anyone interested in risks, it is also worth noting that
there is a small chance that the low will pass further south over
Scotland, bringing rain and wind much further south across the country.
This is just one example of the types of charts used to summarise
ensemble forecasts for forecasters.
The ensemble provides useful estimates of probabilities for many
weather events, but these can often be further improved by statistical
post-processing. Calibrated probability forecast data are generated
daily for over 300 sites worldwide. The graph below shows an
example of a calibrated 5-day forecast of the relative probabilities of
different temperatures at Heathrow Airport for midday on 28th February
2004.
Clearly the most likely temperature is around zero Celsius, with a 27%
probability of temperature below freezing (right). However there is also
a real possibility of quite mild temperatures above 6 Celsius. On a
balance of probabilities, information like this was used by forecasters
on this occasion to issue early warnings of a cold spell of wintry
weather several days ahead. In the event Heathrow experienced
temperatures below freezing overnight with light snow, rising to a
maximum around 5 Celsius and falling rapidly to 1 Celsius in heavy
snow at 1720. On this occasion the balance of probabilities provided
good guidance, and this should normally been the case. However the lower
probability events should also be expected to occur on some, fewer,
occasions. Had the temperature actually been 8-12 Celsius the warning
would have seemed excessive, but would still have been fully justified
on the basis of the evidence available at the time it was issued.
These site-specific probability forecasts are verified routinely to monitor
performance and demonstrate the capability of statistical
post-processing.
Back to top
|
| Short-range
ensemble research |
|
Current operational use of ensembles is restricted to application of
the ECMWF EPS for medium-range prediction. Important uncertainty can
also occur in short-range forecasts. Usually at short-range, up to 3
days ahead the general weather pattern is well forecast by a single
model run, but there can still be uncertainty in the resulting fine
details of the weather, for example in the amount, location or timing
of rainfall. On rare occasions there can also be significant
uncertainty in the large-scale weather patterns, and these occasions
can be particularly important as they may be associated with severe
weather developments. Research is currently being undertaken to
investigate whether ensembles designed specifically for short-range use
can help in quantifying the uncertainty in these areas.
Back to top
|
|
|
|
Home 
