Transcript 1 - Copernicus.org

A Study on the Ensemble MOS
for Medium Range Prediction in Korea Meteorological Administration
JunTae Choi ([email protected])
Korean Meteorological Administration, Rep. of Korea
1 Introduction
4 Result
3 Experiment
KMA will extend medium Range forecast from 7day to 10 day on Oct. 2014
A post processing technique, Ensemble Model Output Statistics (EMOS), was
developed to remove systematic error of ensemble dynamic model and to
provide weather forecaster with accurate guidance with uncertainty.
 Property of median and control member of EPSG prediction
 Verification : 6 fold cross validation, Jun. 2011 ~ May 2014
10 day
7
day
Forecast
Forecast
※ Status of Statistical Guidance in KMA
projection
short range
(upto 3day)
medium range
(upto 12day)
NWP model
method
Element
remark
Regional Model
MOS
Max/Min T, Pop, etc
MLR, for digital forecast
(12km, 87hr)
Kalman filter
Max/Min/spot T
SSPS(UKPP)
Spot T, RH, etc
mountain, adjust
Global Model
MOS
Max/Min T
MLR
(N512, 288hr)
Kalman filter
Max/Min T
EMOS
Max/Min T, cloud
MLR
MOS (SVR)
Spot T
SVR
EPSG (N320)
2 Method
 Ensemble MOS’s Equation
- Deriving ONE equation for all ensemble members and Applying it to
each member
Median of
24 members
MLR
with Stepwise
24 members from
EPSG(NWP)
Statistical Func.
Verification of independent variable
VIF, t-s, p-v, weight < 1δ of obs.
24 ensemble MOS
prediction
24 members from
EPSG(NWP)
 Definition of element and observation data
- Daily MAX/MIN temperature
: 135 station(synoptic, meta, AWS)
- Total cloud amount(12hrly mean) : 45 station(manned synoptic, meta)
 NWP data : Ensemble Prediction System for Global(EPSG)
- Model : Unified Model introduced from Met Office(N320)
- Archived period : Jun. 2011 ~ (more 3 years)
- Ensemble size : 24 members (1 control + 23 perturbed members(ETKF))
 Statistical method to derive equation
- Multiple Linear Regression(MLR) screened by stepwise selection
- Point equation : one equation for one station
※ CRPS : Continuous Rank Probability Score
※ BIAS is calculated with median of EMOS predict
 Enlargement of sample size to derive Eq.
- According to above figures(red line), the statistics between D day and
D+1 day prediction is very similar
(Apr. 2014)
two sample t-v
 Assumption
- The ensemble mean may represent the property of ensemble prediction
observation
- mean and standard deviation between median of all members and control
member are similar (t statistic value is 0.4 for temp and 0.9 for wind)
 Eq. from the median can work on prediction of individual member
Tsfc
RHsfc
W. SPDsfc
cloud amt
0.15
0.40
0.60
0.42
 Eq. for D day prediction is derived with D-1, D and D+1 day prediction
data, instead of only D day prediction data
 the size of sample to deriving eq. can be triple
- BIAS of EPSG was successfully removed by MOS.
- CRPS was decreased 0.6℃ for MAX/MIN temp. and 0.6 for cloud,
- Spread of EMOS prediction being equal or greater than that of EPSG
 Comparing the other models
- RMSE of median of EMOS prediction according to the sampling way
Sampling way
for D day prediction
MAX. Temp.
MIN. Temp.
cloud amount
D day prediction
2.08 ℃
2.17 ℃
2.62
D-1, D, D+1 day
prediction
2.07 ℃
2.15 ℃
2.27
(6 fold cross validation, Jun. 2011 ~ May 2014)
 Great improvement in cloud amount prediction,
and slight improvement for temperature
 Additional post-process for cloud amount MOS
- Theoretically, MOS prediction is closer to climatic value, longer prediction.
But cloud 0 and 10 is most frequently observed.
 Need post process(Y.K., Seo and J.T. Choi, 2013 ECAM)
- post process : Fitting the percentile between MOS and OBS. distribution
(Jun. 2013 ~ May 2014, 46 point)
GDAPS : Global Model(N512)
- The MOSs are better than direct output of ECMWF model
- Ensemble method is more important than resolution of NWP model
5 Discussion
POST P.
 EMOS with simple statistical method can provide reasonable guidance
in form of uncertainty.
 EMOS could be more accurate than direct output of ensemble model,
while its ensemble spread being not reduced.
 Ensemble MOS based on low resolution model is better than deterministic
style MOS based on high resolution model.