Introduction of KMA statistic model and ensemble system

1
Introduction of KMA statistic model and ensemble
system

• Korea Meteorological Administration
• Numerical Weather Prediction Division
• Joo-Hyung Son

2
Statistical models

• PPM (Perfect Prognostic Method)
• Daily Max/Min and midnight temperature
• Probability of Precipitation
• MOS (Model Output Statistics)
• Digital Forecast
• KF(Kalman Filtering)/DLM(Dynamic Linear Model)
• Daily Max/Min Temperature
• 3 hourly temperature
• Daily Max/Min Temperature of 10 days

3
Statistical models
Max/Min Temp PoP
PPM
PPM
RDAPS
KF
Max/Min Temp
KF
DLM
3hr Temp
RDLM
GDAPS
Max/Min Temp
DLM
GDLM
4
PPM for Max/Min Temp

• Predictant
• 00 UTC 1(00UTC, Max/Min)
• 12UTC 1(Max), 2(00UTC, Min)
• Forecast regions
• 70 sites in Korea
• Model development
• May 1, 1988 Feb 28, 1992 (4 years)
• Regional reanalysis of JMA
• Climate data of 70 weather sites

5
PPM Model structure for Max/Min Temp
Forecast equation

• Temp (t) A Bobs(0) Cimodel
  predictori(t)
• A, B, Ci (i1,2,,n) fixed coefficients

predictor
predictor
1000, 850, 700,500,400,300hPa Wind speed,
direction, Temperature Dewpoint temp, Height et
al. from RDAPS Observation, climate
Forecast eqs for each season, sights
predictant
Max/Min and 00LST temperature of 70 sights
Predictant
6

• PPM Predictors
• select a group of predictors which explain
  predictant(temperature) well from 44 predictors
  ltmethod forward-backward selectiongt
• the number of the predictors of each seasonal
  and regional forecast equations are ranged from 5
  to 10

Main predictors for 12hr Max temp Main predictors for 12hr Max temp Main predictors for 12hr Max temp Main predictors for 12hr Max temp Main predictors for 24hr Max temp Main predictors for 24hr Max temp Main predictors for 24hr Max temp Main predictors for 24hr Max temp
Spring Summer Fall Winter Spring Summer Fall Winter
T850, CLMT, OBS, VOR8 PCWT OBS, CTOP VOR8 RH50, PCWT T85, CLMT, OBS, TTD8, VORS T85, CLMT, OBS, TTDB, VOR8 T85, CLMT, PCWT, VOR8, OBS OBS, VORS, T85, CTOP, RH70 T85, CLMT, TTD8, PCWT, VOR8 T85, CLMT, VOR8, TTD8, 70Q4
Main predictors for 12hr Min temp Main predictors for 12hr Min temp Main predictors for 12hr Min temp Main predictors for 12hr Min temp Main predictors for 24hr Min temp Main predictors for 24hr Min temp Main predictors for 24hr Min temp Main predictors for 24hr Min temp
Spring Summer Fall Winter Spring Summer Fall Winter
CLMT, OBS, T85, KYID, TTD8 OBS, T85, CLMT, VOR8, TTD8 CLMT, OBS, TTD8, VOR8, T85 T85, OBS, TTD8, TAD8, VOR8 CLMT, T85, KYID, TTD8, S70 OBS, T85, CLMT, VOR8, PCWT CLMT, TTD8, PCWT, T85, OBS T85, OBS, TTD8, VOR8, TAD8

• OBS observation, CLMT climate, PCWT virtual
  prediction, VOR vorticity, TAD temperature
  advection, KYID KY index

7
PPM Model structure for PoP
Forecast equation

• Temp (t) A Bobs(0) Cimodeli(t)
• A, B, Ci (i1,2,,n) fixed coefficients

predictor
predictor
1000, 850, 700,500,400,300hPa Wind speed,
direction, Temperature Dewpoint temp, Height et
al. from RDAPS Observation, climate
Forecast eqs for Each region according to warm
and cold season
predictant
PoP of 18 regions
Predictant
8

• PPM Predictors
• PoP
• the number of sites observed precipitation in
  the region
• Total number of sites in the region
• 18 regions
• 24 region by cluster analysis
• (Moon(1990))
• forecast experiment

• the forecast equations are developed according
  to the warm(April-September) and
  cold(October-March) season and each regions.

principle predictors for PoP principle predictors for PoP
Warm season DWL, VR850, QA700, VV700, S850, RH500
Cold season DWL, VV850, RH850, VR850, S850, 7Q4

• 18 regions for forecast of PoP

9
(No Transcript)
10
KF for Max/Min Temp

• Predictant
• 00 UTC 1(Min/Max), 2(Min)
• 12 UTC 1(Max), 2(Min/Max)
• Forecast regions
• 40 in Korea,
• 32 in North Korea, China, Japan

11
KF for Max/Min Temp

• Kalman Filter algorithm

vtN(0,Vt) observation noise
wtN(0,Wt) process noise
Gt 1
V0 2
4/365 0 0 W0 0
1/365 0 0 0 1/365
1 Ft RDAPS
Latest Obs temp
12
(No Transcript)
13
DLM(Dynamic Linear Model)

• DLM
• Improved Kalman Filter algorithm
• Weights(regression coefficient) are modified
  according to the prior condition with time.

14
DLM(Dynamic Linear Model)

• DLM

vtN(0,Vt)
wtN(0,Wt)

• Use the updating algorithm to estimate Wt with
  time
• Find appropriate Wt increasing discount
  factor(0ltdeltalt1) from 0.01 to 1 with interval
  0.01
• the discount factor is selected when RMSE between
  observation and forecast is the lowest

15
DLM(Dynamic Linear Model)

• RDLM(Regional DLM)
• 3hourly forecast up to 48hr
• RDAPS
• 38 sites
• GDLM(Global DLM)
• Max/Min temp for 10 days
• GDAPS
• 38 sites

16
(No Transcript)
17

• Ensemble Prediction System

18
KMA Ensemble Prediction System
GBEPS 1.1.1 GBEPS 1.2.1 GBEPS 2.1.1 GBEPS 2.3.1 GBEPS 2.3.1.1 GBEPS 2.3.1.2
Operation period 2001.3.12003.10.31 From 2003.11.1 From 2005.2.
Data assimilation 2dOI ? 3dOI 3dOI ? 3dVar 3dVar
Model GDAPS T106L21 GDAPS T106L30 GDAPS T106L30
Vertical resolution 21 levels 30 levels 30 levels
Perturbation method Breeding Breeding ? Breeding Factor Rotation Breeding ? Breeding Factor Rotation
Target area (BV) Global Northern Hemisphere Northern Hemisphere
Lead time 10 days 8 days 8 days
Ensemble members 17 (16 members 1 control) 17 (16 members 1 control) 17 members
19
Schematic diagram

• The global spectral model T106L30 with the
  slightly different initial conditions run 17
  times.
• Both perturbed analysis and control analysis
  are projected to 24hours with the model, and
  departures from the control analysis at 24hours
  are scaled down to the norm of initial
  perturbations

20
Schematic diagram
D day
D day 12hr
D1 day
D1 day 12hr

• 17members could be similar each other because
  they are generated from the identical model, so
  this is to make different perturbation among the
  members manually.
• In the new system, the factor rotation was
  added every alternative step.

21
2005. 6. 11
old(cray-before)
NEW (cray-frot)
22
2005. 6. 11
old(cray-before)
NEW (cray-frot)
23
EPS products (http//190.1.20.56)
24
Mean and Spread
25
Spaghetti
26
Spaghetti ( with ensemble spread)
5520m
5640m
27
Stamp map
display the global model, mean and standard
deviation and spaghetti as well as each member.
28
Categorical PoP

• 12-hour precipitation gt given thresholds
• 1, 5, 10mm for winter season
• 1, 10, 50mm for other seasons
• The probability
• These probability maps are used for the early
  warning guidance of severe weather.

29
Categorical PoP
30
Probability of Surface Max Wind

• Surface maximum wind gt 10m/s, 14m/s
• The probability
•  
• These probability maps are used for the early
  warning guidance of severe weather.

31
Probability of Surface Max Wind
32
Time series of Probability

• Precipitation
• 12hr accumul gt 1mm
• 12hr accumul gt 10mm
• 12hr accumul gt 50mm
• Surface Max Wind
• sfc wind gt 10m/s
• sfc wind gt 14m/s
• Principle cities
• Seoul, Daegu, Daejeon
• Busan et al.

33
EPSgram
Time series of primary cities
34
Ensemble Plumes
Time series of 8-day forecast at cities The
dispersion of members with forecast evolution
Variable Pmsl, 500H, 850 T
35
Hwangsa (yellow sand) trajectory
36
Typhoon Strike Probability Map by EPS
37
Thank you
38
Factor analysis

• Factor analysis
• Factor analysis is a statistical technique to
  explain the most of the variability among a
  number of observable random variables in terms of
  a smaller number of unobservable random variables
  called factors
• Factor rotation
• Factor rotation is to find a parameterization in
  which each variable has only a small number of
  large loadings. That is, each variable is
  affected by a small number of factors, preferably
  only one. This can often make it easier to
  interpret what the factors represent.