Comparing and Combining the EnVar and EnKF Methods in a Limited-Area - - PowerPoint PPT Presentation

Comparing and Combining the EnVar and EnKF Methods in a Limited-Area Deterministic (and Probabilistic) Context

Jean-François Caron, Seung-Jong Baek and Peter Houtekamer

Meteorological Research Division Environment and Climate Change Canada

2019 ISDA, Kobe, Japan, 24 January 2019

ECCC NWP systems in 2019

x 20 x 20 Global (2x per day) Regional (4x per day) Convective- scale Deterministic Ensemble EnKF 256- member 4DEnVar 4DEnVar

GDPS RDPS HRDPS GEPS REPS Dx=10km Dx=2.5km Dx=39km Dx=10km

Dxa=39km Dxa=39km Dxa=39km

x 20 Perturbation recentering Dx=15km

(4x per day) Limitation: Regional systems rely on global 4DEnVar with low-resolution ensemble covariances, making short- term convective-scale prediction challenging.

x 20

x 20 x 20 Global Deterministic Ensemble EnKF or VarEnKF 4DEnVar 4DEnVar

GDPS RDPS GEPS REPS Dx=2.5km Dx=25km Dx=10km

Dxa=25km Dxa=10km Dxa=25km

x 20 Dx=10km

10-km EnKF

x 20

Regional (24x per day)

Off-Topic

ECCC NWP systems in 202?

Major changes:

• Introduce limited-area 4DEnVar for regional system with hourly cycling and high

resolution limited-area ensemble covariances.

• Facilitates assimilation of high-resolution radar, cloud and surface observations.

Regional configuration

• As in Bédard et al (2018, MWR)

the experimental RDPS use a model with same resolution as ensemble and analysis increment (10 km, instead of 2.5 km)

4

• Limited-area 4DEnVar follows our global version, except that

1) Spectral decomposition based on a bi-Fourrier representation instead of spherical-harmonic is used for modelling the climatological B (Bnmc) and the localization in Bens 2) Localization was adapted: hloc = 1400 km instead of 2800 km; vloc = 1 unit of ln(p) instead of 3. 3) Hybridization was adapted: 87.5% Bens + 12.5% Bnmc instead of 75.0% Bens + 37.5% Bnmc

topography

Regional configuration

5

• Limited-area EnKF follows our global (256 member) version,

except that

• Same model configuration for all members instead of multi-

physics approach.

• Additive inflation still based on a global Bnmc, but with a

reduced scaling factor: 0.252 instead of 0.332.

• Land-surface initial conditions provided by the RDPS instead
• f the GDPS.
• LBCs: from the GDPS for the RDPS, from the global EnKF for the

regional EnKF.

Note: Unlike the ECCC's EnVar system, our EnKF do not assimilate: ground-based GPS, radiances from geostationary satellites, SSMIS and many CRIS, AIRS and IASI channels.

M(Xa ) vs M(Xa )

EnKF RDPS-4DEnVar

July 2016 January 2017

Tested approaches to improve the ensemble mean analysis

• 4 flavours of ensemble mean recentering were tested, all

based on (as in Houtekamer et al 2018, QJRMS):

7

𝐘𝒃𝒋

𝒔 = 𝐘𝒃𝒋 + 𝒅 𝐘𝒃 𝑭𝒐𝑾𝒃𝒔 − 𝐘𝒃 1. Full recentering (c=1) 2. CMC hybrid gain: c=1 for half of the members, c=0 for the other half 3. Hybrid gain with c= 1/2 4. Hybrid gain with c= 2/3

EnKF analysis EnKF ensemble mean analysis Recentering coefficient 4DEnVar analysis using Xb = Xb

(1) Impact on the forecasts from the ensemble mean analysis

8

Impact of the full recentering

July 2016 January 2017

M(Xa ) vs M(Xa )

EnKF-fullRecentering EnKF

Reference : EnKF

EnKF configuration July January Altitude Surface Altitude Surface Full recentering on EnVar +2.32 +1.09 +0.50 +1.29 CMC Hybrid Gain +1.93 +1.06 +0.50 +0.99 Hybrid Gain 1/2 EnVar + 1/2 EnKF +2.08 +1.08 +0.43 +0.91 Hybrid Gain 2/3 EnVar + 1/3 EnKF +1.95 +1.06 +0.52 +1.14

Changes in NWP index (+3 to +48h, 6h)

All the approaches have significant positive impact on the quality of the ensemble mean analysis; full recentering is the best.

(2) Impact on the forecasts from the 4DEnVar-based RDPS

11

Reference : 4DEnVar using Bens(EnKF)

Flavors of EnKF July January Altitude Surface Altitude Surface Full recentering on EnVar

• 0.23
• 0.22

+0.03 +0.14 CMC Hybrid Gain +0.19 +0.26

• 0.01
• 0.05

Hybrid Gain 1/2 EnVar + 1/2 EnKF

• 0.17

+0.20 +0.01

• 0.02

Hybrid Gain 2/3 EnVar + 1/3 EnKF 0.00 0.00 0.00 +0.20

Changes in NWP index (+3 to +48h, 6h)

Unfortunately, no clear impact on the forecast performances of the EnVar-based RDPS were detected

(3) Impact on the regional EPS

(i.e. 72h forecasts from 20 members picked from the EnKF)

13

Impact of the full recentering

July 2016 January 2017

M(Xa ) vs M(Xa )

EnKF-fullRecentering EnKF

Reference : EnKF

EnKF configuration July January Altitude Surface Altitude Surface Full recentering on EnVar +1.58 +0.29 +1.16 +0.59 CMC Hybrid Gain +1.58 +0.27 +2.23 +1.41 Hybrid Gain 1/2 EnVar + 1/2 EnKF +1.58 +0.38 +0.94 +0.52 Hybrid Gain 2/3 EnVar + 1/3 EnKF +1.43 +0.26 +1.07 +0.64

Changes in EPS index - Overall

All the approaches have significant positive impact on the quality of the ensemble forecasts; CMC hybrid gain is the best.

Reference : EnKF

EnKF configuration July January Altitude Surface Altitude Surface Full recentering on EnVar +4.61

• 0.25

+3.80 +2.28 CMC Hybrid Gain +7.51 +0.50 +34.00 +10.04 Hybrid Gain 1/2 EnVar + 1/2 EnKF +4.69

• 0.05

+3.42 +1.09 Hybrid Gain 2/3 EnVar + 1/3 EnKF +4.88

• 0.20

+3.73 +1.74

Changes in EPS index - Reliability

Reference : EnKF

EnKF configuration July January Altitude Surface Altitude Surface Full recentering on EnVar +1.47 +0.17 +1.07 +0.12 CMC Hybrid Gain +1.36 0.00 +0.58

• 2.39

Hybrid Gain 1/2 EnVar + 1/2 EnKF +1.49 +0.27 +0.84 +0.35 Hybrid Gain 2/3 EnVar + 1/3 EnKF +1.31 +0.17 +0.96 +0.30

Changes in EPS index - Resolution

Summary and conclusions

• Inserting various amount of information from a limited area

EnVar analysis (made using Xb = ensemble mean forecast) improves significantly our limited area EnKF ensemble mean analysis.

• Complete/full recentering provides the best performances for

forecasts initialized from the ensemble mean analysis.

• Unfortunately, using the ensemble-derived covariances from

the various recentered EnKF has no significant impact on the forecast performances of the EnVar-based RDPS.

• All the recentered EnKF analysis improves significantly the

performances of our regional EPS.

• The so-called CMC hybrid-gain approach (recentering of only

half of the members) provides the largest improvements, due to the resulting modification of the initial ensemble spread.

18

Extra

19

Impact of the CMC hybrid-gain

July 2016 January 2017

M(Xa ) vs M(Xa )

EnKF-CMC-Hybrid EnKF

spread (mse-obserr)1/2 Temperature @ 850 hPa

M(Xa ) vs M(Xa )

EnKF RDPS-4DEnVar

July 2016 January 2017

As shown before

M(Xa ) vs M(Xa )

EnKF-fullRecentering RDPS-4DEnVar

July 2016 January 2017

July 2016 January 2017

M(Xa ) vs M(Xa )

EnKF-fullRecentering RDPS-4DEnVar

Same LBCs