Die Präsentation wird geladen. Bitte warten

Die Präsentation wird geladen. Bitte warten

Post-processing methods for probabilistic convection forecasts based on the limited-area ensemble COSMO-DE-EPS of DWD Lars Wiegand, Christoph Gebhardt.

Ähnliche Präsentationen


Präsentation zum Thema: "Post-processing methods for probabilistic convection forecasts based on the limited-area ensemble COSMO-DE-EPS of DWD Lars Wiegand, Christoph Gebhardt."—  Präsentation transkript:

1 Post-processing methods for probabilistic convection forecasts based on the limited-area ensemble COSMO-DE-EPS of DWD Lars Wiegand, Christoph Gebhardt German Meteorological Service (DWD), Germany

2 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Content  set-up COSMO-DE-EPS  EPS convection project  methodology  observation  forecast – probabilistic products (case study)  Bayes theorem  LASSO  data  result  further researches  product design  SESAR (Single European Sky)

3 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 model chain at DWD COSMO-DE: 2.8 km convection-permitting forecast model 50 vertical levels modelrun every 3 hours: + 27 h GME: 20 km COSMO-EU: 7 km set-up COSMO-DE-EPS further details: see talk SCI-PS Susanne Theis: Tue 13:30, room 524A

4 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 COSMO-DE-EPS GME IFS GSM GFS 20 members set-up COSMO-DE-EPS further details: see talk SCI-PS Susanne Theis: Tue 13:30, room 524A further details: see talk SCI-PS Susanne Theis: Tue 13:30, room 524A

5 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 project „EPS Convection“  predictability of small scale processes with non-linear and stochastic processes (e.g. convection) is strongly limited  i.e. leads to strong uncertainty already at short lead times  severe events and high impact weather are highly important for warnings in general or in aviation in particular  “charakteristics of HIW” and “limited predictability” leads to use of probabilistic estimation of high-resulotion forecasts of deep convection based on COSMO-DE-EPS  aims at supporting aviation weather forecasts and general weather warning process at DWD

6 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 methodology probabilistic products for convective parameters from COSMO-DE-EPS  DMO (direct model output) variables as well as direct calculatable variables (e.g. KO-index) from DMO  IMO (indirect model output), e.g. thunderstorms produced with regression methods  requirements for IMO (thunderstorm) forecasts:  observation of IMO (thunderstorm) as predicand  radar + lightning  EPS DMO forecasts as predictor(s)

7 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 observation thunderstorm combination of radar reflectivity and lightning  RX product  advantages: warning criterias are known within DWD (28, 37, 46, … dBz), high spatial/temporal resolution  adaptions:  conversion into COSMO-DE grid  lightning from NCM network  very accurate observations – only 0,02% of all lightnings have errors >2,8km  adaptions:  every grid point within 3km gets a distance weighted amount of a lightning measure

8 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 case study – thunderstorms 28 th July 2013

9 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 case study – thunderstorms 28 th July 2013

10 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 case study – thunderstorms 28 th July 2013

11 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 case study – thunderstorms 28 th July 2013

12 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 case study – thunderstorms 28 th July 2013

13 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 case study – thunderstorms 28 th July 2013

14 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Bayes theorem  variables: CAPE, CIN, TWATER, OMEGA, DBZ_CMAX, TOT_PREC  period: summer (Apr- Sept) 2012  forecast: 00UTC + 0/6/12/18h  based on grid points  1/0 event occurs/does not occur  X = variable from COSMO-DE(-EPS)

15 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Bayes theorem  variable: TWATER (total water content)  forecast: 00UTC + 0h  period: summer (Apr- Sept) 2012

16 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Bayes theorem  variable: TWATER (total water content)  forecast: 00UTC + 18h  period: summer (Apr- Sept) 2012

17 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Bayes theorem  variable: DBZ_CMAX (radar reflectivity column maximum)  forecast: 00UTC + 12h  period: summer (Apr- Sept) 2012

18 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Bayes theorem  variable: OMEGA (vertical velocity)  forecast: 00UTC + 0h  period: summer (Apr- Sept) 2012

19 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 LASSO – least absolute shrinkage and selection operator (Tibshirani 1996)  search for suitable predictors  comparison of predictors (variables (DMO) and their probabilistic products)  choose of predictors, which depict the observation best  tool: R statistic software (package glmnet + dependences)  logistic regression:  optimal for extreme values  Input/output can be probabilities  error measure: RMSE

20 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014  COSMO-DE-EPS  forecast: daily 03UTC + 21h  40 days in summer 2012 (22 th July – 30 th August)  93 variables (CAPE, CIN, T2m, TWATER, TQ, TI, …)  EPS products: mean, minimum, maximum  5 days, i.e. 8 calculations  observation: 1h radar maximum and lightning sum  radar: 16:30 – 17:25 UTC (available every 5 minutes)  lightning: 16:30 – 17:29 UTC (exact to the second) LASSO summer 2012 (40 days) data basis

21 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014  mean TWATER  maximum TQ (Graupel)  maximum radar reflectivity (maximum RR in atmospheric column)  all 3 variables are amongst the first 5 predictors for the 8 calculations  maximum TWATER in 5 out of 8 calculations within the first 5 predictors  to check: stability of predictors for longer time periods  result just shows the predictors for 8 x 5 days in summer 2012 result - predictors LASSO summer 2012 (40 days)

22 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 further research  LASSO  longer time periods  statistical robustness  quantiles and probabilities as predictors  time offset  neighborhood method  different synoptical regimes (convective time scale)  generation of thunderstorm forecast product from 3 or more predictors

23 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014  Objectives To develop ensemble post-processing techniques in order to provide consistent short-range probabilistic NWP products of convective risks across Europe, at the highest possible NWP resolution  combination of three convection-permitting ensembles systems. AROME-EPS (MF), COSMO-DE-EPS (DWD) and the UKV-EPS(UKMO) Super-Ensemble Mesoscale Forecast of Convection (SESAR-JU WP11.2.1, lead Meteo France)  generation of consistent, blended probability products for ATC  2 data phases  Summer 2012 (mid July – end August)  Spring 2014 (mid April – end June)

24 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Example: combined mean radar refelectivity

25 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Thanks for your attention! Comments? Questions?

26 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Supplementary slides

27 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 product design

28 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Direct Model Output (DMO)  Zusammenhang Wahrscheinlichkeit(DMO>Schwellenwert)  Ereignis  Wahrscheinlichkeiten dieser DMO-Variablen nicht zwingend gut kalibriert  Ansatz: Vorhersage “ja” für Wahrscheinlichkeiten(DMO>Schwelle) > A%  Bestimme hit rate/ false alarms für verschiedene A  Optimales A ist nutzerabhängig! (hit rate/ false alarms)  Datenlage der Beobachtungen von DMO oft flächig nicht beobachtbar (CAPE, TWATER, …)  Ereignis: Gewitter aus Radar + Blitz

29 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Weitere Arbeiten in EPS Konvektion  WX getestet:  Vorteil: größeres Gebiet: keine Radarmessungen werden verworfen  Nachteil: nicht sicher ab wann operationel, wird aller Vorraussicht nicht nachberechnet  Programme für abgeleitete Variablen aus DMO Variablen  KO-index  Convective time scale – Klassifizierung in synoptische/Luftmassengewitter Situationen  Erstellung des technischen und fachlichen Rahmens des Fachkonzeptes  Studien zu statistischen Eigenschaften der gewählten ‘high-priority’ Variablen  CAPE, CIN, dBz_cmax, TWATER,

30 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Beobachtung Gewitter Schwellenwerte für Gewitterklassifikation classradar reflectivity [dbz] lightning strikes [no./15 minutes] associated weather moderate>371moderate rain wind gust up to 7 Bft strong>46tbdheavy rain (10-25 l/m² in 1h, l/m² in 6h) wind gusts 8-10 Bft hail possible (Ø <1,5cm) severe>53tbdvery heavy rain (>25 l/m² in 1h, >35 l/m² in 6h) wind gusts >11 Bft large hail possible (Ø >1,5cm)

31 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 COSMO-DE

32 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Beispiel – Gewitterlage 28. Juli 2013

33 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Super-Ensemble Mesoscale Forecast of Convection ( SESAR-JU WP11.2.2, lead Meteo France)  Ziel: To develop ensemble post-processing techniques in order to provide consistent short-range probabilistic NWP products of convective risks across Europe, at the highest possible NWP resolution  Kombination dreier konvektionserlaubender Ensemblesysteme.  AROME-EPS (MF)  COSMO-DE-EPS (DWD)  UKV-EPS(UKMO)  Erstellung von konsistenten (räumlich verschnittenen) Produkten für die Flugsicherung (allerdings nur post-processing)

34 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Modellgebiete

35 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 SESAR domain Gemeinsames SESAR ‘Modellgebiet’  Dünkirchen (2.38E, 51N) als gemeinsamer Gitterpunkt  Dünkirchen kein GP in originalem COSMO-DE-Gitter  Auflösung 0.027°/0.022° (lon/lat – reguläres Gitter) Anpassungen:  Interpolation von rotiertem Gitter (0.025° lon/lat) auf SESAR-Gitter  Variablenanpassung: z.B. Windstärke auf SESAR-Gitter aus staggered grid  Korrekte Einstellungen der grib2 header

36 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Zwei Phasen der Datenarchivierung  1. Phase (Sommer 2012): 22. Juli – 30. August 2012  93 Variablen  20 Member  21h Vorhersage (stündlich)  03UTC Vorhersage  2. Phase (Frühling 2014): 1. April bis 10. Juni 2014 (71 Tage – 40 ausgewählte)  Selbe Spezifikationen wie in Phase 1  Vorhersage bis 27h neu

37 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 Observation thunderstorm

38 WWOSC 2014 Montréal, CanadaLars Wiegand, DWD18 th August 2014 COSMO-DE COSMO-DE-EPS „variations“ within the system ensemble members set-up COSMO-DE-EPS further details: see talk SCI-PS Susanne Theis: Tue 13:30, room 524A


Herunterladen ppt "Post-processing methods for probabilistic convection forecasts based on the limited-area ensemble COSMO-DE-EPS of DWD Lars Wiegand, Christoph Gebhardt."

Ähnliche Präsentationen


Google-Anzeigen