|
Titel |
Toward post-processing ensemble forecasts based on hindcasts |
VerfasserIn |
B. Van Schaeybroeck, S. Vannitsem |
Konferenz |
EGU General Assembly 2012
|
Medientyp |
Artikel
|
Sprache |
Englisch
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 14 (2012) |
Datensatznummer |
250060050
|
|
|
|
Zusammenfassung |
Having in mind an operational implementation of post-processing at the Royal
Meteorological Institute of Belgium (RMI), we study possible approaches of correcting the
ECMWF ensemble forecast for stations in Belgium using the ensemble hindcast data set.
This data set is each week enlarged by eighteen independent five-member ensemble forecasts
using the current operational system. Therefore, the hindcasts constitute an ideal basis for the
training cycle of post-processing. Combined with a forward predictor-selection procedure we
propose to use a post-processing technique called error-in-variables model output statistics or
EVMOS. This technique was recently proposed and is based on linear regression and suited
for correcting ensemble forecasts. The corrected forecasts are produced for nine synoptic
stations in Belgium.
Different factors which influence the correction quality and which we aim to optimize are
the number of weeks of training data, the number of predictors and the clustering of daily
training data. We also investigate the influence of the training period, that is, the
period of days over which the training forecasts are initialized. More specifically, we
compare a training window which is centered around the forecast day with the
case where the days of training precede the forecast day. Different results for the
different training periods arise due to seasonal effects. We validate the different
approaches against the bias-corrected forecasts using observations at nine stations for the
ten-meter zonal and meridional wind speed and the two-meter temperature and for lead
times up to one week. This is performed by cross-validation for a period of fourteen
weeks.
For the inland stations and for all lead times, a mean-square-error (MSE) improvement of
around 1.5 m2/s2 and 0.5Â (°C)2 for wind and temperature, respectively, is obtained. The
MSE gain for wind at the two coastal stations is lower, especially for the meridional wind.
Systematic biases are negligible for wind and thus most of the EVMOS post-processing is
obtained by a variability correction. For two-meter temperature, on the other hand, systematic
biases dominate the EVMOS corrections evidencing the correct variability representation of
the model. For forecasting the day-time (12h) two-meter temperature, the best forecast is
obtained by a simple bias correction whereas EVMOS post-processing turns out most
effective for predicting the night-time (0h) forecast. In order to utilize EVMOS
post-processing operationally, we propose the use of three predictors, a training period of at
least seven weeks and preferably a training period centered around the forecast
date.
Initially more than eighty candidate predictors are considered from the hindcast data set,
based on which we construct eleven additional predictors. From the set of selected predictors
during validation we isolate the most prominent ones. Except for the corresponding variables,
by far the most frequently-used predictors for the ten-meter wind are North-South and
East-West surface stresses as well as the boundary layer height. For the two-meter
temperature, temperature at 850 hPa and maximal temperature in the last 6 hours are the most
crucial predictors. |
|
|
|
|
|