Recent Development of the JMA Global Spectral Model

1
Recent Development of the JMA Global Spectral
Model

• Masayuki Nakagawa
• JMA/NPD, visiting NCEP/EMC
• Nov. 10, 2009

2
Outline of the Presentation

• Overview of JMA
• Operational NWP models at JMA
• Recent development in global NWP
• Global Spectral Model
• Ensemble Prediction System
• Future plan

3
Overview of JMA
4
Structure of Central Government of Japan
JMA is placed as an extra-ministerial bureau of
the Ministry of Land, Infrastructure, Transport
and Tourism. Total staff 5700 Budget approx.
700 million/yr
5
Organizational Structure of JMA
6
Observation Networks (1)

• Surface observations
• 156 manned weather stations
• 1337 automatic weather stations
• Radars
• 11 Doppler radars
• 9 conventional radars

7
Observation Networks (2)

• Upper air observations
• 16 radiosonde stations
• 31 wind profilers
• Satellite observations
• Geostationary meteorological satellite
  (MTSAT-1R)

picture from the WMO homepage (modified)
8
(No Transcript)
9
Organization of NPD

• Numerical Prediction Division (74)
• Administration Section (5)
• Programming Section (11)
• Management of NWP system
• Development of data decoding system
• Numerical Analysis and Modeling Section (46)
• Development of NWP models and analysis systems
• Chief (1)
• Global Modeling Group (17)
• Mesoscale Modeling Group (13)
• Observation Group (15)
• Application Section (12)
• Development of applications (guidance, graphics,
  )

10
Operational NWP models at JMA
11
Operational NWP Models at JMA (1)

• Mesoscale model
• Horizontal Resolution 5 km
• Updates 8 times a day
• Forecast domain
• Japan and its surrounding areas

• Global model
• Horizontal
• Resolution 20 km
• Updates 4 times a day
• Forecast domain
• Global

12
Operational NWP Models at JMA (2)
Global Model (GSM) Mesoscale Model (MSM) Typhoon Ensemble Model One-week Ensemble Model One-month Ensemble Model Three-month Ensemble Model Warm/Cold season Ensemble Model
Purposes Short- and medium-range forecast Warnings and very short- range forecast Typhoon forecast One week forecast One month forecast Three month forecast Warm/Cold season outlook
Forecast domain Global Japan and its surrounding areas Global Global Global Global Global
Grid size/ Number of grids 0.1875deg./ 1920x960 (TL959) 5km/ 721x577 0.5625deg./ 640x320 (TL319) 0.5625deg./ 640x320 (TL319) 1.125deg./ 320x160 (TL159) 1.875deg./ 192x96 (TL95) 1.875deg./ 192x96 (TL95)
Vertical levels/ Top 60 / 0.1hPa 50 / 21800m 60 / 0.1hPa 60 / 0.1hPa 60 / 0.1hPa 60 / 0.1hPa 60 / 0.1hPa
Forecast hours (initial time) 84 hours (00, 06, 18 UTC), 216 hours (12 UTC) 15 hours (00, 06, 12, 18 UTC), 33 hours (03, 09, 15, 21 UTC) 132 hours (00, 06, 12, 18 UTC) 11 members 9 days (12 UTC) 51 members 34 days (12 UTC Wed. Thu.) 25 members x2 120 days (12 UTC once a month) 31 members 150-210 days (12 UTC 5 times a year (Feb., Mar., Apr., Sep. Oct.) 31 members
Analysis 4D-Var 4D-Var Global analysis with ensemble perturbations Global analysis with ensemble perturbations Global analysis with ensemble perturbations Global analysis with ensemble perturbations Global analysis with ensemble perturbations
13
Framework of GSM

• Resolution TL959, reduced Gaussian grid
• 0.1875 deg. / 1920 (equator)
• 6 deg. / 60 (closest to pole) x 960,
  roughly 20km
• 60 unevenly spaced sigma-p hybrid levels
• (surface to 0.1 hPa)
• Dynamics 2-time level, semi-Lagrangian time
  integration
• Time step 600 sec
• Cumulus Prognostic Arakawa-Shubert
• Cloud Prognostic cloud water
• PBL Mellor and Yamada level II
• Radiation(L) k-distribution method and table
  look-up method
• Radiation(S) Lacis and Hansen (1974)
• Gravity wave o(1-10km), o(100km)
• Land SiB
• Assimilation 4D-Var

14
Operational Global Objective Analysis
Cut-off time 2h20m for early run analyses at 00, 06, 12 and 18 UTC, 11h35m for cycle run analyses at 00 and 12 UTC, 5h35m for cycle run analyses at 06 and 18 UTC
Initial Guess 6-hour forecast by GSM
Grid form, resolution and number of grids Reduced Gaussian grid, 0.1875 degree, 1920x960 for outer model Standard Gaussian grid, 0.75 degree, 480x240 for inner model
Levels 60 forecast model levels up to 0.1 hPa surface
Analysis variables Surface pressure, temperature, winds and specific humidity
Methodology Four-dimensional variational (4D-Var) scheme on model levels
Data Used SYNOP, SHIP, BUOY, TEMP, PILOT, wind profiler, AIREP, SATEM, ATOVS, SATOB, surface wind data from scatterometer on the QuikSCAT satellite and MODIS wind data from Terra and Aqua Typhoon bogussing applied for analysis
Initialization Non-linear normal mode initialization and a vertical mode initialization for inner model
Early Analysis
Analysis for weather forecast. The data cut off
time is very short.
Cycle Analysis
Analysis for keeping quality of global data
assimilation system. This analysis is done after
much observation data are received.
15
Roles of GSM

• Basic information for a short- and medium-range,
  one week, one month and seasonal forecasts
• Basic information for typhoon track and intensity
  forecasts
• Assist of aviation and ship routing forecasts
• Provision of lateral boundary condition for
  Mesoscale Model
• Input data for ocean wave model
• Input data for ocean data assimilation
• Wind information for input of chemical transport
  model

16
Recent development in global NWP - GSM -
17
JMA/NWP Update Plan
Major Forecast Models in JMA
FY2003
FY2004
FY2005
FY2006
FY2007
FY2008
FY2009
FY2010
GSM(TL319)
GSM(T213)
60km
Horizontal Resolution
GSM(TL959)
RSM
20km
(NH)MSM
10km
MSM
Extend Forecast Time
(NH)MSM
5km
Data Assimilation Systems
FY2003
FY2004
FY2005
FY2006
FY2007
FY2008
FY2009
FY2010
GSM
4DVAR
3DVAR
4DVAR
(T159)
(T63)
(T106)
(TL319)
(T106)
Objective Analysis for
RSM
RSM operation was finished
4DVAR(40km)
MSM
(NH)4DVAR(10km)
4DVAR(20km)
HPC System Upgrade
Japanese Fiscal Year Start from April and End
in March
18
Upgrade of GSM in Nov. 2007
previous current
Forecast time 36(06,18)/ 90(00)/ 216(12) 84(00,06,18)/ 216hours(12UTC)
Horizontal resolution Approximately 60 km(TL319) Approximately 20 km(TL959)
Vertical resolution 40 layers(highest 0.4 hPa) 60 layers(highest 0.1 hPa)
Time integration 3-time level(?t900 sec) 2-time level(?t600 sec)
orography/ mask Equivalent to 60 km resolution Equivalent to 20 km resolution
Sea surface temperature Daily analysis (1 degree resolution) Daily analysis (0.25 degree resolution)
Sea ice concentration Climatology (1 degree resolution) Daily analysis (0.25 degree resolution)
Snow depth Daily analysis (1 degree resolution) 6 hourly analysis (higher resolution over Japan area)
19
Simulated Infrared Image
GMS-5 observation 00UTC Jul. 10 2002
20
Orography of Operational Models at JMA
GSM TL959 (20km)
MSM (5km)
Orographic effects are better captured by
higher resolution models. The surface parameters
such as temperatures and winds, might be
predicted more realistically by those models.
GSM TL319 (60km)
21
Sigma-P Hybrid Vertical Level of GSM
0.1 hPa about 65 km
Stratosphere (25 layers)
finer in lower atmosphere
Troposphere (35 layers)
lowest level about 20 m
22
Introduction of Reduced Gaussian Grid
A reduced Gaussian grid was implemented in GSM as
a new dynamical core in August 2008. On the
standard Gaussian grid, the longitudinal interval
between two grid points at the high latitudes is
smaller than that at the low latitudes. Hence,
it is redundant to use an equal number of grid
points for all given latitudes in global
model. The total number of grid-points is reduced
by about 30 in the reduced Gaussian grid, thus
saving the computational throughput.
Miyamoto (2007)
23
Moist Parameterization in GSM

• Cumulus convection
• Arakawa-Schubert scheme (Arakawa and
  Shubert 1974 Moorthi and Suarez 1992 Randall
  and Pan 1993)
• Convection triggering mechanism proposed by Xie
  and Zhang (2000) (DCAPE) was introduced to
  improve the rainfall forecast
• Clouds and large-scale precipitation
• Prognostic cloud water scheme (Sommeria and
  Deardorff 1977 Smith 1990)
• Marine stratocumulus
• Stratocumulus scheme (diagnostic) (Slingo 1980,
  1987 Kawai and Inoue 2006)

24

• Convection Triggering Mechanism

Xie and Zhang (2000) defined DCAPE (dynamic CAPE
generation rate) as (T, q) are (T, q) plus
the change due to the total large-scale advection
over a time interval ?t (integration time step
used in the model). They are equal to (T, q)
just after the calculation of model dynamics. Xie
and Zhang (2000) showed a strong relationship
between deep convection and positive DCAPE. In
TL959L60 GSM, deep convection (cloud top lt
700hPa) is assumed to occur only when DCAPEgt
-1/300 (J/kg/s) , which corresponds to dynamic
warming or moistening in the lower troposphere.
25
Precipitation (Typhoon)
T0610
TL959L60
TL319L40
Radar
6 hour accumulated precipitation valid at 12UTC
18 August 2006. The initial time of the
forecasts is 12UTC 17 August 2006. The gray area
in right panel indicate an absence of analysis.
Typhoon T0610 (WUKONG) was moving northward over
Kyushu Island. Both models predicted its
position well. TL319L40 GSM could not predict the
detailed distribution of precipitation and strong
rainfall over land. TL959L60 GSM simulated the
distribution and intensity of precipitation
better then TL319L40 GSM, including orographic
precipitation and heavy rainfall near the center
of the typhoon.
26
RMSE and Bias of Typhoon Central Pressure
TL319L40 GSM predicted weak typhoons compared to
the best track analyzed by RSMC-Tokyo Typhoon
Center because of its low horizontal
resolution. TL959L60 GSM predicted the typhoon
intensity better then TL319L40 GSM.
0 24 48
72 Forecast time (hour)
TYM 24-km resolution regional model covering a
tropical cyclone and its surrounding areas. Its
operation was terminated in November 2007.
27
Precipitation Scores against Raingauge
Observation (Aug. 2004)
Bias score
Threat score
Threshold mm/12h
Threshold mm/12h
FT3648 hrs, 80 km grid average over Japan
TL959L60 TL319L40 RSM (retired)
GSM tends to overestimate week precipitation
areas and to underestimate strong precipitation
areas in summer.
28
Precipitation Scores against Raingauge
Observation (Aug. 2004)
Bias score
The Introduction of convection triggering
mechanism proposed by Xie and Zhang (2000)
(DCAPE) reduced the tendency of GSM to
overestimate weak precipitation areas especially
from local noon to late afternoon.
0 12 0 12
JST
Forecast hour h
80 km grid average over Japan Threshold 1mm/3h
29
Northern Hemisphere RMSE
Aug. Sep. 2004
TL959L60 TL319L40
RMSE of Psea and z500 decreased slightly in both
summer and winter season.
Psea
z500
Dec. 2005 Jan. 2006
TL959L60 TL319L40
Psea
z500
30
Verification Score
RMSE of 24, 48 and 72 hour forecasts by GSM for
500 hPa geopotential height against analysis in
NH (20N 90N). Curves monthly means, horizontal
lines yearly means.
31
Pie chart showing the relative cost of various
components for 84 hours forecast
Resolution TL959L60 Computer HITACHI SR11000
70nodes(140MPIs) Real Time
31min24sec (fastest case 29min39sec)
Disk access (20)
Calculation (44)
Communication (36)
After Miyamoto (2008)
32
Recent development in global NWP - EPS -
33
Upgrade of 1W-EPS in Nov. 2007
previous current
Horizontal resolution Approximately 120km(TL159) Approximately 60km(TL319)
Vertical resolution 40 layers(highest 0.4hPa) 60 layers(highest 0.1hPa)
Time integration 3 time level(?t1200sec) 2 time level(?t1200sec)
orography/ mask Equivalent to 120km resolution Equivalent to 60km resolution
Method to make initial perturbations Breeding of Growing Mode method Singular Vector method
Perturbed area Northern hemisphere and tropical zone (20S 90N) Northern hemisphere and tropical zone (20S 90N)
Ensemble size 51 members 51 members
34
Specification of Typhoon EPS (Feb. 2008)
Purpose Improve both deterministic and probabilistic forecasts of tropical cyclone (TC) movement
Forecast domain Global
Grid size/ Number of grids 0.5625 deg./ 640x320 (TL319)
Vertical levels/Top 60 / 0.1 hPa
Forecast hours 132 hours (00, 06, 12, 18 UTC) Runs when TCs of TS/STS/TY intensity exist in the responsibility area of RSMC Tokyo - Typhoon Center (0N-60N, 100E-180E) or are expected to move into the area within the next 24 hours
Ensemble size 11 members
Method to make initial perturbations Singular Vector (SV) method Linear combination of SVs targeted on both TCs (up to three TCs in one forecast event) and a mid-latitude region
It is possible to obtain reliability of typhoon
track forecast from the ensemble spread of
typhoon track forecasts by Typhoon EPS. In
addition, alternative track scenarios to an
ensemble mean track are available.
35
Example of Typhoon Ensemble forecasts (1)
T0607 (MARIA)
Typhoon Ensemble forecasts (11 members blue
line control run)
Forecast by GSM
Analyzed track
Possibility of recurvature of the typhoon is
represented in Typhoon Ensemble forecasts.
Ensemble spread is large, which indicates the
reliability of the forecasts is relatively low.
36
Example of Typhoon Ensemble forecasts (2)
T0416 (CHABA)
Typhoon Ensemble forecasts (11 members, blue
line control run)
Forecast by GSM
Analyzed track
Ensemble spread is quite small, which indicates
the reliability of the forecasts is relatively
high.
37
Future plan (GSM)
38
Focus of NPDs recent efforts

• Model bias
• Temperature, moisture,
• Spin-up
• Precipitation,
• Land-sea contrast in precipitation
• Precipitation over tropical eastern Pacific
• Global circulation
• Formation of Typhoon
• Size of Typhoon
• Maximum wind radius
• Intensity of Typhoon
• Ocean mixing layer model

39
Future Resolution Upgrade Plan(next
supercomputer system)

• Deterministic forecast
• TL959L60 ? TL959L100
• Upgrade model dynamics and physics
• Introduce new satellite data
• Probabilistic forecast
• 1WEPS TL319L60M51 ? TL479L100M51
• Improve representation of smaller scale
  phenomena
• Improve forecast skill of severe weather
• TEPS TL319L60M11 ? TL479L80M25
• Improve probabilistic forecast skill of tropical
  cyclone movement
• Improve forecast skill of severe weather
  associated with tropical cyclones

40
Thank you!
Hare-run JMAs mascot Hare Japanese word for
fine weather.
41
Replacement of JMA Supercomputer
Current System
Previous System
Mar 2005-
Mar 2006-
Mar 2001-Feb 2006
80nodes
50nodes
HITACHI SR8000E1-80nodes
80nodes
HITACHI SR11000J1 -210nodes
768Gflops
27.5Tflops
42
Early Analysis and Cycle Analysis
Early Analysis
Analysis for weather forecast. The data cut off
time is very short.
Cycle Analysis
Analysis for keeping quality of global data
assimilation system and for supplying the first
guess to early analysis. This analysis is done
after much observation data are received.
84 hour forecast
Early Analysis
Ea00
84 hour forecast
Ea06
in hurry to issue forecast
The first guesses for Ea06 and Ea18 are supplied
from Ea00 and Ea12, respectively.
Da00
Da06
Da18
Cycle Analysis
Da12
in hurry to issue forecast
216 hour forecast
Ea12
Ea18
84 hour forecast
Early Analysis
43
Numerical/Dynamical Properties (1)

• Horizontal representation
• Spectral (spherical harmonic basis functions)
  with transformation to a reduced Gaussian grid
  for calculation of nonlinear quantities and most
  of the physics.
• Horizontal resolution
• Spectral triangular TL959 (deterministic), TL319
  (EPS)
• Vertical representation
• Finite differences in sigma-pressure hybrid
  coordinates.
• Vertical domain
• Surface to 0.1 hPa.
• Vertical resolution
• There are 60 unevenly spaced hybrid levels.

44
Numerical/Dynamical Properties (2)

• Time integration scheme
• A two-time level semi-implicit semi-Lagrangian
  scheme is used for the time integration.
• A constant time step length 600 sec. is used for
  the deterministic (TL959) model.
• Equations of state
• Primitive equations for dynamics in a spectral
  semi-Lagrangian framework are expressed in terms
  of wind components, temperature, specific
  humidity, cloud water and surface pressure.
• Diffusion
• A linear fourth-order horizontal diffusion is
  applied on the hybrid sigma-pressure surfaces in
  spectral space.

45
Physical Properties

• Cumulus Prognostic Arakawa-Shubert
• Cloud Prognostic cloud water
• PBL Mellor and Yamada level II
• Radiation(L) k-distribution method and table
  look-up method
• Radiation(S) Lacis and Hansen (1974)
• Gravity wave o(1-10km), o(100km)
• Land SiB

46
Reduced Gaussian Grid (Aug. 2008)
There are a large number of redundant grid-points
and insignificant wavenumber components in the
standard Gaussian grid. The total number of
grid-points is reduced by about 30 in the
reduced Gaussian grid.
After Miyamoto (2007)
The number of longitudinal grid points
must be the multiples of the number of
longitudinal sub-domains. must be the
composite numbers of the radices of FFT
kernels. should be the multiple numbers of
the longitudinal interval of the
radiation process.
47
Convection and precipitation

• deep convection - Arakawa and Schubert 1974
• conversion of cloud droplets to precipitation
• moisture detrainment from top of the cumulus
• re-evaporation of stratiform precipitation

Short-wave radiation
Long-wave radiation

upward mass flux
detrainment
condensation evaporation
Water vapor
Cloud water
Cumulus convection
Conversion from cloud droplets
re-evaporation
entrainment
convective downdraft
precipitation
compensative downdraft
48
Simple Biosphere model
lowest level of the atmospheric model
sensible heat
latent heat
canopy
sw rad.
lw rad.
bare ground
grass
thin skin layer
Snowmass is not treated explicitly and is
regarded as an iced water on the grass or bare
ground. Upper 5cm snow is accounted in heat
budget
soil layer
conductive heat (evaluated with force restore
method)
49
Transition Steps

• Algorithm development
• Preliminary testing
• Low resolution (TL319L60) forecast/assimilation
  experiment, summer and winter
• High resolution (TL959L60) single forecast
  experiment (no assimilation)
• Pre-Implementation testing
• High resolution (TL959L60) forecast/assimilation
  experiment, at least summer and winter
• Systematic error, RMSE, anomaly correlation,
  typhoon track and intensity, precipitation,
• Implementation

50
Introduction of new convection triggering
function to Arakawa-Schubert scheme
51
Moist parameterization in GSM

• Cumulus convection
• Arakawa-Schubert scheme
• Convection triggering function
• Rainwater and cloud water budget
• Clouds and large-scale precipitation
• Cloud water scheme
• Marine stratocumulus
• Stratocumulus scheme

52

• Convection triggering function (1)

Radar observation
GSM tends to predict convective precipitation too
early with too wide areas in summer daytime. In
order to improve the rainfall forecast, a new
convection triggering mechanism is
introduced. Xie and Zhang (2000) showed a strong
relationship between deep convection and positive
DCAPE (dynamic CAPE generation rate) which is
determined by the large scale advective
tendencies.
GSM forecast
6 hour accumulated precipitation, 12UTC 18 July
2005 initial, FT18 (15 local time).
53

• Convection triggering function (2)

Xie and Zhang (2000) defined DCAPE (dynamic CAPE
generation rate) as (T, q) are (T, q) plus
the change due to the total large-scale advection
over a time interval ?t (integration time step
used in the model). They are equal to (T, q)
just after the calculation of model dynamics.
54

• Convection triggering function (3)

Observed precipitation, CAPE and DCAPE by Xie et
al. (2004) . CAPE is almost always positive
during the day, while most of precipitation
occurred in late evening and early
morning. Precipitation and positive DCAPE
correlates strongly. They showed the introduction
of DCAPE improved the performance of their model.
Xie et al. (2004)
55
Precipitating area is closely related to the area
where DCAPEgt0, which suggests the capability of
DCAPE as the triggering function of deep
convection. In TL959L60 GSM, deep convection
(cloud top lt 700hPa) is assumed to occur only
when DCAPEgt -1/300 (J/kg/s) , which corresponds
to dynamic warming or moistening in the lower
troposphere. The threshold value depends on
horizontal resolution.
40 10 1
0.1
0
DCAPE
Radar obs.
GSM with DCAPE
GSM w/o DCAPE
6 hour accumulated precipitation and DCAPE valid
at 12 UTC 18 July 2005. Initial time of
forecasts is 12UTC 17 July 2005.
56
Case study (thunderstorm)
Radar obs.
GSM with DCAPE
GSM w/o DCAPE
6 hour accumulated precipitation valid at 12 UTC
9 August 2004. Initial time of forecasts is 12
UTC 8 August 2004.
GSM without DCAPE predicts too weak and wide
precipitation. GSM with DCAPE simulates the areas
and the intensity of thunderstorm better than
that without DCAPE.
57
Case study (Typhoon T0416)
Radar obs.
GSM w/o DCAPE
GSM with DCAPE
T0416
6 hour accumulated precipitation valid at 00 UTC
30 August 2004. Initial time of forecasts is 12
UTC 28 August 2004.
GSM without DCAPE predicts too weak
precipitation. GSM with DCAPE simulates the areas
and the intensity of heavy precipitation better
than that without DCAPE.
58
Statistics
Bias and equitable threat scores of 3 hour
accumulated precipitation forecasts against
raingauge observation over Japan for August
2004. Horizontal axis forecast time.
Bias score for weak precipitation (1mm/3hour) of
GSM without DCAPE (blue) is larger than 1 and
shows strong diurnal variation. The variation is
reduced substantially in GSM with DCAPE (red),
though the bias is still large.
59
Summary

• The convection triggering mechanism proposed by
  Xie and Zhang (2000) (DCAPE) was introduced to
  the A-S scheme to improve the rainfall forecast.
• GSM with DCAPE simulated the area and the
  intensity of heavy precipitation associated with
  thunderstorm and typhoon better than GSM without
  DCAPE.
• The tendency of GSM to overestimate weak
  precipitation areas especially from local noon to
  late afternoon is also reduced.
• DCAPE is implemented to the operational GSM in
  November 2007.