 |
| Titel |
Probabilistic forecasts of extreme local precipitation using HARMONIE predictors and comparing 3 different post-processing methods |
| VerfasserIn |
Kirien Whan, Maurice Schmeits |
| Konferenz |
EGU General Assembly 2017
|
| Medientyp |
Artikel
|
| Sprache |
en
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 19 (2017) |
| Datensatznummer |
250142025
|
| Publikation (Nr.) |
 EGU/EGU2017-5596.pdf |
|
|
|
|
|
| Zusammenfassung |
| Statistical post-processing of deterministic weather forecasts allows production of the full
forecast distribution, and thus probabilistic forecasts, to be derived from that deterministic
model output. We focus on local extreme precipitation amounts, as these are one predictand
used in the KNMI weather warning system. As such, the predictand is based on the
maximum hourly calibrated radar precipitation in a 3x3 km2 area within 12 large
regions covering The Netherlands in a 6-hour afternoon period in summer (12-18
UTC).
We compare three statistical methods when post-processing output from the operational
high-resolution forecast model at KNMI, HARMONIE. These methods are 1) extended
logistic regression (ELR), 2) an ensemble model output statistics approach where the
parameters of a zero-adjusted gamma (ZAGA) distribution depends on a set of covariates and
3) quantile random forests (QRF). The set of predictors used as covariates includes
model precipitation and indices capturing a variety of processes associated with deep
convection. We use stepwise selection to select predictors for ELR and ZAGA based on
the AIC. Predictors and coefficients are selected in a cross-validation framework
based one two-years of training data and the skill of the forecasts are assessed on
one-year of test data. The inclusion of additional predictors results in more skilfull
forecasts, as expected, particularly for higher precipitation thresholds and for forecasts
using the QRF method. We also assess the value of using a time-lagged ensemble.
Forecasts derived from ZAGA and QRF are generally more skilfull, as defined
by the Brier Skill Score, than ELR and lower precipitation amounts are skillfully
predicted. |
|
|
|
|
|
|