WRF Model Forecast Track Sensitivities of Tropical Cyclone Ernesto (2006) to Various Parameterizations, Grid Spacings, and Initial Conditions

1
WRF Model Forecast Track Sensitivities of
Tropical Cyclone Ernesto (2006) to Various
Parameterizations, Grid Spacings, and Initial
Conditions
Nick P. Bassill Michael C. Morgan 2006
This work was supported by the National Science
Foundation Grant ATM-0125169
2
Tropical Cyclone Ernesto Overview

• Ernesto originated from an African easterly wave
  on 24 August 2006 over the southern Leeward
  Islands
• While in the eastern Caribbean, Ernesto was
  classified a tropical storm, and was forecast to
  become a powerful hurricane
• However, Ernesto strongly deviated from its
  forecast track, surprising forecasters in the
  process
• This deviation began while Ernesto was south of
  the Dominican Republic

Above National Hurricane Center (NHC) Best
Track
3
National Hurricane Center Discussion for Ernesto
from 15 UTC 26 August 2006

• THE FORECAST TRACK IS MORE PROBLEMATIC AFTER 72
  HR. THE LARGE-SCALE MODELS AGREE THAT THE
  MID-LEVEL RIDGE OVER THE GULF OF MEXICO WILL
  WEAKEN AS A SHORTWAVE TROUGH DIGS SOUTHEASTWARD
  THROUGH THE MISSISSIPPI VALLEY. HOWEVER ... THERE
  ARE DIFFERENCES IN HOW MUCH WEAKENING WILL OCCUR.
  THE ECMWF AND NOGAPS CALL FOR ERNESTO TO RECURVE
  OVER THE EASTERN GULF OF MEXICO ... WHILE THE
  UKMET SHOWS ENOUGH RIDGE TO KEEP ERNESTO MOVING
  WEST-NORTHWESTWARD

Above NHC Five Day Forecast From 15 UTC 26
August 2006
4
WRF Model Configuration

• ARW core (January 2006 release)
• YSU boundary layer scheme (bl_pbl_physics1)
• 2nd order diffusion (diff_opt1, km_opt4)
• Rapid radiative transfer model (RRTM) longwave
  radiation scheme (ra_lw1)
• Dudhia shortwave radiation scheme (ra_sw1)
• Five-layer thermal diffusion land-surface model
  (sf_surface_physics1)
• Monin-Obukhov similarity theory based surface
  layer (sf_sfclay_physics1)

5
Experiment 1

• Changeable WRF Variables
• Grid Spacing
• Horizontal 30km, 45km, or 60km
  
• Vertical 31 or 54 levels
• Cumulus Parameters (CPs)
• (1) Kain-Fritsch (cu_physics1)
• (2) Betts-Miller-Janjic (2)
• (3) Grell-Devenyi ensemble (3)
• Microphysics Parameters (MPs)
• (1) Kessler (warm rain) (mp_physics1)
• (2) Eta-Ferrier (time efficient) (5)
• (3) WSM6 (most complex of these) (6)

6
Exp. 1 Continued

• 180 hour forecasts are created with every
  possible combination (48 forecasts)
• Forecasts are initialized with GFS 0.5 x 0.5
  forecast valid 12 UTC 26 August 2006
• The goal is to attempt to determine the relative
  importance of modifying grid spacings and
  parameterizations
• Any combinations including 54 vertical levels
  and the Betts-Miller-Janjic scheme fail due to
  unknown boundary condition problems

7
There are two primary tracks
North across Cuba and Florida
West across the Yucatan Peninsula
8
Initial Observations

• This track ensemble reproduces the problem
  observed in the previous NHC forecast discussion,
  in that there are two track scenarios (1) north
  across Cuba or (2) west across the Yucatan
  Peninsula
• All simulations predict a cyclone of
  approximately the same intensity when the split
  occurs
• It was presumed that some type of large-scale
  steering difference is the cause of the split
  (ergo the split is not a result of differences in
  the inner core of the cyclone)
• To examine why this split occurs, comparisons
  between a representative simulation of each was
  conducted

9

• Both are 30km x 30km x 31 levels and use
  Kain-Fritsch CP
• Blue run uses WSM6 MP
• Red run uses Kessler MP
• Future plots show blue minus red (i.e. west -
  north)

180
150
120
90
150
60
120
180
90
60
30
30
10
300-700 hPa Thickness Differences (Blue - Red)
TC John
11
West minus North 90 hour total precipitation in
inches
12
Difference in 300-700 hPa steering flow (blue -
red)
Average magnitude is 5 knots
13

• Experiment 2
• Is this simply the result of forcing by the
  boundary?
• 9 additional runs were done using
• This larger domain
• 60 km grid spacing
• 31 vertical levels

- All possible combinations of previously used
microphysics and cumulus parameterizations
Answer No, the results show tracks which are
approximately identical to their respective
smaller domain versions
14
Observations From Experiments 1 and 2

• Choice of MPs and CPs are most important in
  determining forecast track
• Choice of horizontal and vertical grid spacing
  modifies intensity, but not track philosophy
• Larger domain does not alter forecast track
  philosophy
• At these grid spacings, the Kessler MP scheme
  performs best
• Westward moving TCs prevent realistic formation
  of Hurricane John
• Eta/Ferrier microphysics scheme is generally
  unable to intensify Ernesto
• Example Average minimum central pressure of
  storms entering the Bay of Campeche

Eta/ Ferrier WSM6 ?SLP
30 km 989 952.5 36.5mb
45 km 1001 981.3 19.7mb
60 km 1002.7 991.3 11.4mb
15

• Experiment 3
• Will a smaller grid spacing allow
• the Eta/Ferrier scheme to intensify the cyclone
  similarly to the other schemes?
• any models the potential to more accurately
  forecast the future path of the cyclone?

16

• 4 km grid spacing over this domain
• 31 vertical levels
• No cumulus parameterization

17
Observations (2)

• Smaller grid spacing allows WSM6 MP scheme to
  perform best
• This could be a result of the ability to resolve
  graupel producing updrafts with this grid
  spacing, which should favor WSM6 over other MP
  schemes
• Smaller grid spacing does not allow Eta/Ferrier
  scheme to comparably deepen the cyclone

18
Experiment 4

• Experiment 1 was redone using
• - All combinations of CPs and MPs
• - The same domain
• - A grid spacing of 30 km with 31 vertical
  levels (for a total of 9 simulations)
• - All else equal, except
• These forecasts were initialized with GFS 1.0
  data from 12 UTC 26 August (instead of 0.5 data)
• Theoretically, 0.5 data is more accurate than
  1.0 data, so it would be expected to produce
  more realistic results

19
The actual track is shown in red

• Each color represents a particular CP-MP
  combination
• - Forecast tracks for 1.0 are dashed and solid
  for 0.5 (from Exp. 1)

20
Observations (3)

• Some runs have an increased likelihood of
  dissipating Ernesto when using 1.0 data
• There is a distinct northeastward shift of the
  forecast tracks with the 1.0 data
• This would result in smaller forecast errors,
  even though 0.5 data theoretically has better
  initial conditions

21
Experiment 5 As in previous experiment, except
now using the smaller domain shown in experiment 3
22
0.5 Data Simulations
23
Observations (4)

• Unlike Exp. 4, a smaller grid spacing in
  conjunction with 1.0 data prevents the storm
  from intensifying or taking a more realistic
  track
• This results in a larger grid spacing (30 km vs.
  4 km) providing a more accurate forecast!

24
Difference of 800 hPa vorticity (1.0- 0.5 s-1),
and location of Ernesto at initial time (Red for
1.0, Blue for 0.5)
25
(No Transcript)
26
Observations (5)

• 0.5 data initializes Ernesto as a more compact
  system, whereas 1.0 data smears out initial
  vorticity
• More intense storms as in Exp. 3 are steered to
  the northwest whereas weaker storms in Exp. 5 are
  steered by the low level flow to the west
• To examine the effect of using these different
  initial conditions, 9 hour forecasts were
  compared using the Kessler MP simulations

27
Zoomed in 9-hour accumulated rainfall difference
(1.0-0.5 in inches), sea level pressure
difference at 9 hours (1.0-0.5 in mb), and
location of Ernesto at 9 hours (Red for 1.0 ,
Blue for 0.5) using Kessler MP runs
28
Observations (6)

• A more compact initial system allows for more
  organized initial convection (with 0.5 data),
  which produces more latent heat release and
  allows the storm to intensify more realistically
  from the beginning
• Unlike runs using larger grid spacings, in this
  case inner core dynamics do influence forecast
  track
• Using a smaller grid spacing makes the
  differences between 1.0 and 0.5 data more
  important

29
Experiment 6

• Experiment 4 was redone using
• - All combinations of CPs and MPs
• - The same domain
• - A grid spacing of 30 km with 31 vertical
  levels (for a total of 9 simulations)
• - All else equal, except
• These forecasts were initialized with GFS 1.0
  data from 00 UTC 26 August (instead of 12 UTC 26
  August)

30
- Tracks for 00 UTC 1.0 data (solid) and 12 UTC
1.0 data (dashed) - Track points are normalized
to be at the same forecast time (ex. 42 hours
from 00 UTC, 30 hours from 12 UTC)
31
Observations (7)

• Generally similar forecasts, although some 00 UTC
  forecasts are better able to forecast the initial
  jump to the north
• Similar intensity issues when using the
  Eta/Ferrier MP scheme

32
General Conclusions

• Perhaps modeling focus should be on improving
  parameterizations, for three primary reasons
• 1) These experiments demonstrate that choice of
  MP and CP influence track and intensity more
  than choice of horizontal or vertical grid
  spacing (at least for this case at the times
  chosen)
• 2) Using a smaller grid spacing can be
  detrimental to model forecast with certain
  initial conditions
• 3) More advanced parameterizations would almost
  certainly be less computationally expensive than
  using a drastically reduced grid spacing

33
Future Work

• Test additional microphysics schemes
• Test these results with other TCs
• Try other initial conditions (besides GFS)