Applying an outer-loop to the WRF-LETKF system for typhoon assimilation and prediction

1
Applying an outer-loop to the WRF-LETKF system
for typhoon assimilation and prediction

• Shu-Chih Yang1,Kuan-Jen Lin1,
• Takemasa Miyoshi2 and Eugenia Kalnay2
• 1 Dept. of Atmospheric Sciences, National
  Central University, Taiwan
• 2 Dept. of Atmospheric and Oceanic Science, Univ.
  of Maryland, USA

2
Motivation

• The regional EnKF needs a spin-up period when
  cold-starting the ensemble with global analyses.
• For typhoon prediction, the reliability of the
  mean state strongly affects the track prediction.
• The quality of initial mean state determines the
  spin-up length.
• During the spin-up, valuable observations (e.g.
  reconnaissance aircraft) cant be effectively
  used.
• The Running In Place method can serve as a
  generalized outer-loop for the EnKF framework to
  improve the nonlinear evolution of the ensemble
  (Kalnay and Yang, 2010, Yang et al. 2012).
• The RIP method is designed to re-evolve the whole
  ensemble to catch up the true dynamics,
  represented by observations.
• Both the accuracy of the mean state and structure
  of the ensemble-based covariance are improved.

3
Standard LETKF framework
LETKF (ti-1)
Obs(ti-1)
xa0 (ti-1)
Nonlinear model Mxa(ti-1)
Time
xb0 (ti)
LETKF (ti)
Obs(ti)
xa0 (ti)
4
Standard LETKF framework
LETKF (ti-1)
Obs(ti-1)
xa0 (ti-1)
Nonlinear model Mxa(ti-1)
Time
Adjust dynamical evolutions at an earlier time
xb0 (ti)
LETKF (ti)
Obs(ti)
xa0 (ti)
5
Running in place in the LETKF framework
LETKF (ti-1)
Random Pert.
Obs(ti-1)

?
Nonlinear model M xa(ti-1)
Time
?
xbn (ti)
no-cost Smoother
?
?
LETKF (ti)
Obs(ti)
Threshold gt e
xa(ti)
?
False
Re-evolve the whole ensemble to catch up the true
dynamics, represented by Obs
6
Increase the influence of observations

• ?Hard way
• Reduce the observation error and assimilate this
  observation once.
• Compute the analysis increment at once
• ?soft way
• Use the original observation error and assimilate
  the same observation multiple times.
• The total analysis increment is achieved as the
  sum of multiple smaller increments (advantageous
  with nonlinear cases).

7
Increase the influence of observations

• ?soft way RIP/QOL are advantageous corrections
  with nonlinear cases
• With a smaller ensemble spread, smaller
  corrections allow the increments to follow the
  nonlinear path toward the truth better than a
  single increment.

8
With RIP/QOL, filter divergence is avoided!!
Results from the Lorenz 3-variable model (Yang et
al. 2012a)
4D-Var LETKF LETKF LETKF
4D-Var standard QOL RIP
linear window 0.31 0.30 0.27 0.27
nonlinear window 0.53 (window75) 0.68 0.47 0.35

• With RIP/QOL, the LETKF analysis with nonlinear
  window is much improved, even better than 4D-Var!

• RIP and QOL use the observations more efficiently
  for the under-observed cases.

Assimilation with different obs. sets
x y z xy xz yz xyz
ETKF 2.9 1.67 7.16 1.01 1.53 0.78 0.68
QOL 1.98 1.23 5.94 0.82 1.16 0.60 0.47
RIP 1.57 0.97 3.81 0.56 0.66 0.40 0.35
Performance RIP gt QOL gt standard ETKF
9
With RIP/QOL, filter divergence is avoided!!
Results from the Lorenz 3-variable model (Yang et
al. 2012a)
4D-Var LETKF LETKF LETKF
4D-Var standard QOL RIP
linear window 0.31 0.30 0.27 0.27
nonlinear window 0.53 (window75) 0.68 0.47 0.35

• With RIP/QOL, the LETKF analysis with nonlinear
  window is much improved, even better than 4D-Var!

• RIP and QOL use the observations more efficiently
  for the under-observed cases.

Assimilation with different obs. sets
x y z xy xz yz xyz
ETKF 2.9 1.67 7.16 1.01 1.53 0.78 0.68
QOL 1.98 1.23 5.94 0.82 1.16 0.60 0.47
RIP 1.57 0.97 3.81 0.56 0.66 0.40 0.35
Performance RIP gt QOL gt standard ETKF
How about RIP for a dynamical complex model?
10
Application of LETKF-RIP to typhoon
assimilation/prediction

• Experiment setup
• Regional Model Weather Research and Forecasting
  model (WRF, 25km)
• RIP is used as an generalized outer-loop to
  improve the ensemble evolution during the spin-up
• Assimilation schemes LETKF and LETKF-RIP with 36
  ensemble members

• LETKF-RIP setup
• Computed the LETKF weights at analysis time
  (00,06,12,18Z)
• Use these weight to reconstruct the ensemble (U,
  V) at (03,09,15,21Z)
• perform the 3-hr ensemble forecasts
• Re-do the LETKF analysis (only one iteration is
  tested)

A
B
time
15
06
09
00
03
18
12
11
Observations used in the OSSE experiments
12
Results from OSSE Impact on analyses(Improving
the TYs inner core)
N-S vertical cross-section of wind speed
The covariance structure is strongly correlated
to error pattern.
RIP Effectively spins up the dynamical structure
of the typhoon!
Yang et al. 2012b
13
Results from OSSE Impact on typhoon prediction
(Improving the TYs environmental condition)
Capture the west-ward turning direction of the
typhoon track 12 hour earlier!!
F300hPa (2-day forecast)
Yang et al. 2012b
14
Application to real observations
2008 Typhoon Sinlaku
Observations Upper-air sounding, dropsonde,
surface station
09/08 00Z-09/09 12Z Tropical depression 09/12
18Z-09/11 00Z Typhoon 09/11 06Z-09/12 18Z Super
typhoon 09/13 00Z-09/14 06Z Typhoon
15
Experiment settings for the real case
Standard LETKF run
LETKF
With RIP
LETKF-RIP
With RIP
turn-off RIP
LETKF-RIPs
?
?
?
?
?
?
?
?
2008/09/08 00Z
09 00Z
10 00Z
12 00Z
11 00Z
? Dropsondes (reconnaissance aircraft) are
available

• Two cases are evaluated
• Typhoon structure is well observed by the
  dropsondes
• Typhoon is under-observed.

16
Dynamical adjustment
OBS (QuikSCAT)
(Wind Speed)suf vs. w (LETKF-RIP)

• The RIP scheme enhances the cyclonic flow and the
  vertical motion near the eyewall.
• The asymmetric structure with the strong winds
  is captured in the LETKF-RIP analysis.

Difference (RIP - LETKF)
(Wind Speed)suf vs. w (LETKF)
Time 2008/09/09 12Z
17
Observation Impact
Observation impact is computed according to Li et
al. (2010), Kalnay et al. (2012).
Anal time2008/09/09 06Z
The first aircraft data is available for the
typhoon structure
The effectiveness of the dropsonde data is
greatly improved by RIP and the negative impact
shown in the control run can be alleviated.
Reduce the 6-hr fcst error
18
Impact on Typhoon track prediction
4-Day track prediction initialized at 09/09 06Z
Anal time2008/09/09 06Z
C130 aircraft data
At early developing stage of the typhoon, the
aircraft data can provide more useful information
with RIP for improving track prediction.
19
Another case the typhoon is under-observed
Analysis time 09/10 12Z
No reconnaissance aircraft available near the
typhoon
For the under-observed case, the LETKF-RIP
analysis still leads to a better track
prediction.
20
Impact on Typhoon track prediction
Mean absolute track Error

• With RIP, the track prediction is significantly
  improved during the LETKFs spin-up period.
• RIP is especially useful for improving forecasts
  beyond 36 hours and the typhoon landfall location
  is better predicted.

36hr
Error of land fall location
Time LETKF LETKF-RIP
09 06z 130 (N) 87(N)
09 12z 87(N) 43(N)
09 18z 77 21(N)
10 00z 13 11
10 06z 47 87(N)
10 12z 47 32
10 18z 60 0
Mean cross track Error
N not landfall at Taiwan
Averaged during the spin-up period (first two
days)
21
LETKF-RIP vs. LETKF-RIPs (RIP is turned off
after spin-up)
Averaged for the last two days
Init 09/09 12Z
Cross-track error
LETKF-RIPs
LETKF
LETKF-RIP
Error of central Psfc
When RIP is turned off after the spin-up (2-day),
the performance of the track/ intensity
prediction is even better than the LETKF
prediction.
22
Summary

• RIP can be used as a generalized outer-loop to
  improve the nonlinear evolution of the ensemble
  during the spin-up.
• The RIP scheme accelerates the spin-up of the
  WRF-LETKF system and provides further adjustments
  for improving typhoon assimilation and
  prediction.
• The value of flight data during the developing
  stage of typhoon is effectively increased.
• Even when the typhoon is under-observed, the RIP
  analysis can still provide an improved
  prediction.
• The dynamical adjustments include both the inner
  core and environmental condition of the typhoon.
• RIP can be turned off after the system has spun
  up (2 days).