ECMWF 20th Century Reanalysis Using Surface -Only Observations - - PowerPoint PPT Presentation

Specification of Background Errors in the

ECMWF 20th Century Reanalysis

Using Surface-Only Observations (ERA-20C)

Paul Poli, Dick Dee, David Tan, Hans Hersbach, Elias Hólm, Massimo Bonavita, Lars Isaksen, and Mike Fisher

2013 WMO Symposium on Data Assimilation; Poli et al.

World rld Map ap, , A.Orteli rtelius us , ci circa a 1570

World map of weather uncertainty in 1900, ERA-20C, circa 2013

1

Reanalysis: Dealing with an uncertain past

2013 WMO Symposium on Data Assimilation; Poli et al.

[Pa]

Surface pressure at Montreal, Quebec Observations from ISPD 3.2.6, collection #3004 (Canadian Stations Environment Canada )

2

ERA-20C system

 10-member Ensemble of Data Assimilation  24-hour window 4DVAR  ICOADS 2.5.1 and ISPD 3.2.6 provided observations: surface pressures and marine surface winds (a few of those

• bservations were recovered with funding from ERA-CLIM)

 Variational bias correction of surface pressure obs.  Model: IFS CY38R1, T159L91 or approx. 125 km hor. res. atmosphere, land, and waves  Forcings: HadISST2 ensemble, otherwise as CMIP-5. Used also to force a model-only integration (ERA-20CM)  All outputs: 3-hourly, 75 surface fields, 14 fields / model levels (91) and selected pressure levels (37))  Archive: 700 Tb of fields and 10 Tb of observation feedback  Data to be copied to http://apps.ecmwf.int/datasets/

2013 WMO Symposium on Data Assimilation; Poli et al. 3

ERA analysis window configurations

2013 WMO Symposium on Data Assimilation; Poli et al. 4

ERA-Interim: 12-hour 4D-Var ERA-20C: 24-hour 4D-Var ERA-40: 6-hour 3D-Var

More observations Smaller ensemble spread

2013 WMO Symposium on Data Assimilation; Poli et al. 5

1ox1o data count Ensemble spread [hPa] Surface pressure 1ox1o data count Ensemble spread [m/s] Marine winds

1900

1960

How ERA-20C estimates and uses background errors

 Every ry 10 days

• From past 90-day differences between 3-hr forecasts ensemble members
• System generates global background error covariances
• For each control variable

 Every ry da day

• The analysis uses these covariances
• Modulating locally the vorticity variance by the local ensemble spread

(after some rescaling, although this scaling is constant and very close to unity in ERA-20C)

• There is no time-varying, manual adjustment

 Bottom ttom line: e: Only y observati rvation

• n errors

rs are manua uall lly y specifie fied

• The model stochastic physics (model-specified) and
• The several SST ensemble members (data provider-specified)
• Are supposed to represent all the other sources of uncertainties *

* Terms and conditions apply. See full report for details.

2013 WMO Symposium on Data Assimilation; Poli et al. 6

Checking the error assumptions & spread-skill

2013 WMO Symposium on Data Assimilation; Poli et al. 7

Assumed med Actual ual measure ure of

• f skill

ll Fr From

• m the ensemble

mble spread d and d lo local l da daily ly mod

• dula

lation tion

More observations Smaller background errors

Sharper horizontal correlations Changing vertical correlations

2013 WMO Symposium on Data Assimilation; Poli et al. 8

Single observation analysis increment gradually smaller, affecting smaller areas ERA-20C system adapts itself to the information available Vertical cross-correlations between Ps and Temperature evolve also. …Haven’t yet made sense of these…

NWP: satellites, radiosondes, aircraft,… (for comparison)

January July

Vorticity bkg. error std. dev. Vorticity horizontal correlation

Correlation

Cross-correlation between temperature and Ps / sigma_Ps * sigma_T

Model bias exposed … by the vertical correlations

2013 WMO Symposium on Data Assimilation; Poli et al. 9

1979 2007 1979 2007

Temperature Anomalies Temperature Analysis Increments

2013 WMO Symposium on Data Assimilation; Poli et al. 10

Conclusions about ERA-20C & Outlook

 Inno novati ative ve componen

• nents

ts

• 24

24-ho hour ur 4DVAR analys ysis is

• Self-upd

updatin ting g backgro ground und error r global bal covarianc iances es, , with h local, l, cycling ing adjustmen ustment t of varianc ances es =Testbed bed for NWP developments

• pments
• Varia

iati tion

• nal

al bias corre recti tion

• n of surfac

ace e pressure sure observatio rvations

 Ense semb mble le produc ucti tion

• n essentia

ntiall lly y comple lete te

• ~700Tb

b dataset t produc duced ed in ~200 days

• GOOD

OD: : Seems to repre resent sent fairl rly known wn extreme eme events, s, provided vided they were observe rved d (in spite e of low hori rizon zontal tal resolution)… very likely ely thanks ks to the ensemb mble, le, flow- and time-depe depende ndent nt background ground error

• rs,

s, and 24-hou

• ur

r 4DVAR

• BAD: Trends

ds are contami amina nated ted by systematic matic analysis sis increments ements

• We think

nk we need to be more re aggress ssive ive with h the bias correc ection, tion, to present the analysis with unbiased departures (…à la 20CR)

• If this is confir

firmed, d, we would ld redo a singl gle member ber reanaly alysi sis, , prob

• babl

ably y with a 100-day “blitz run”, fixing also a few other problems

2013 WMO Symposium on Data Assimilation; Poli et al. 11

Thank you for your attention -- for more details:

ERA Report 14 available from the ECMWF website >> Publications >> ERA Reports >> ERA Report Series

http://www.ecmwf.int/publications/library/do/references/show?id=90833

Published at the same time as the production completed.

2013 WMO Symposium on Data Assimilation; Poli et al. 12

ERA-20C: 1899-2009 in 200* days

2013 WMO Symposium on Data Assimilation; Poli et al. 13 * approximately  Speed: between ~30-40 (2 nodes) and 60-80 (4 nodes) days/day/stream. Still missing Oct 2009-2010  During production 3.5 Tb/day, 350 million of meteorological fields, 2000 24-h 4DVAR assimilations run daily  A failure rate as low as 0.1% would have implied 2 manual interventions per day.  Home-grown solution to automatically detect model explosion, stop production, halve the model time-step, etc…

1914 1930 1942 1989 2012

Forecast scores

2013 WMO Symposium on Data Assimilation; Poli et al. 14

• N. Hem. extratropics: 1 day of forecast gain
• S. Hem. extratropics: 1.5 day of forecast gain
• Tropics: brings 12h forecast skill above 60%

Fit to observations

2013 WMO Symposium on Data Assimilation; Poli et al. 15

Southern mid-lat. Northern mid-lat.

Before assimilation After assimilation

Differences at stream boundaries

2013 WMO Symposium on Data Assimilation; Poli et al. 16