Recent trends in dynamical mediumrange tropical cyclone track prediction and the role of resolution

1
Recent trends in dynamical medium-range tropical
cyclone track prediction and the role of
resolution v physics in the ECMWF model
Mike Fiorino michael.fiorino_at_noaa.gov Assimilatio
n and Modeling Branch Global Systems
Division Earth System Research Laboratory Boulder,
CO 4 March 2009
2
Why medium-range (72-h) track?

• Operational long-range warning is a
  medium-range track problem
• no skill vis-à-vis CLIPER in early 1980s
• high skill is required for using improved
  dynamical intensity guidance I cant believe a
  rapid intensification forecast if the track
  doesnt make (physical) sense
• Modeling ( HFIP)
• analysis v model error 72 h is when model
  errors become dominant
• multi-scale dynamics synoptic scales dominate,
  but vortex scales can have large impacts
• closely follows tropical wind score
• MRT is the 500 mb 5-d Anomaly Correlation score
  of the tropics

3
The importance of modeling
POD of NHEM TC in reanalysis v operations
TC POD
improvement in 1989 from physics change, ditto
for degradation in 1994
850 mb tropical wind score time when the
correlation of forecast and analysis vector wind
drops to 70
4
OBS (not) critical to TC analysis?

• POD of TC in reanalysis v operations
• ERA-15 uses model physics circa 1994 and achieves
  1994 operational model POD throughout the
  1979-1994 period
• TC POD follows the tropical wind score and the
  tropical wind score follows the physics
• score improves from 17 d
• when score reaches 3-4 d, global models began
  showing MRT skill
• model of 1st order importance in analysing the
  1st order TC vortex

5
History of Atlantic MRT (72-h) MFE Models v CON
medium-range track (72-h) mean forecast error
two models UKMO (global) and GFDL (limited-area)
with long history v best or baseline consensus
(BCON) 1992-2008
CLIPER trend? better databases error cut in half
from 280 ? 130 nm 1990s-gt 2008 BCON better than
models OFCL BCON
higher skill in 2008 even though CLIPER error
increased ? UKMO and GFDL models did very well
in 2008.
6
Atlantic gain()/loss(-) v BCON

• context for ECMWF results
• 2008 a good year for the models in the LANT
• models typically 20-25 worse than BCON in all
  basins, but more so in WPAC

7
ECMWF resolution v physics changes
8
ALL TCs improvement over BCON
ECMWF is 20 better than BCON when most models
are 20 worse at the medium range (72 h)

• MRT TC skill difference greater than change in
  tropical wind score
• increased resolution (red v orange) improves
  72-120 h error
• physics change improves forecast at all times
• ?skill relationship (physics gt res) consistent
  between basins, but stronger signal in WPAC and
  SHEM, more muted in LANT

9
Summary

• ECMWF has shattered both TC MRT predictability
  limit from the 1990s (130 nm) the skill of
  consensus and the (old) PACOM (50, 100) 150 nm
  requirement
• Updated results as of 20090303 in SHEM (TC
  activity 54 below normal) 72-h MFE
• CONW 158 nm
• ECMWF 108 nm
• 32 gain over BCON 61 cases

10
My personal (biased) takeaways

• How ECMWF achieved these results is of
  fundamental importance to HFIP
• improved TC v synoptic-scale forecast?
• deterministic (hi-res) v ensembles
• can ensembles work until the model(s) achieves
  the same skill as ECMWF? no, physics 1st
• TC vortex analysis
• 1st order structure (assimilate the working best
  track) 1stdetailed, hi-res obs 2nd
• global v limited-area models
• when do errors the global-scales become
  significant? 36 h ?