Improving Ensemble QPF in NMC

37
Improving Ensemble QPF in Improving Ensemble QPF in NMC NMC Dr. Dai Kan National Meteorological Center of China (NMC) International Training Course for Weather Forecasters 11/1, 2012, Kunming

description

Improving Ensemble QPF in NMC. Dr. Dai Kan National Meteorological Center of China ( NMC ) International Training Course for Weather Forecasters 11/1, 2012, Kunming. Outline. QPF operations in NMC Improving QPF by ensemble Improving PQPF. - PowerPoint PPT Presentation

Transcript of Improving Ensemble QPF in NMC

Page 1: Improving Ensemble QPF in NMC

Improving Ensemble QPF in NMCImproving Ensemble QPF in NMC

Dr. Dai KanNational Meteorological Center of China (NMC)

International Training Course for Weather Forecasters

11/1, 2012, Kunming

Page 2: Improving Ensemble QPF in NMC

OutlineOutline QPF operations in NMCQPF operations in NMC Improving QPF by ensembleImproving QPF by ensemble Improving PQPFImproving PQPF

Page 3: Improving Ensemble QPF in NMC

WFO-- subdivision of NMC (National Meteorological

Center)

NMCNMC

Administrative Office

Personnel and Staff Education Division

Division of Operational & Reach Management and S. T. development

Integrative Office

Weather Forecasting Office

NWP Operating and Developing Division

Typhoon and Marine Met. Division

Applied Met. Services Division

NMC Agricultural Met. Center

Met. Service for Decision-making Office

Open Forecast System Laboratory

Retirees Office

Operations( 8)

Management( 5)

Severe Weather Prediction Center

Page 4: Improving Ensemble QPF in NMC

QPE

QPF (no PQPF)

Early warning of heavy rain

Precipitation phase in Winter

Total process precipitation forecast

QPF’s duties

Page 5: Improving Ensemble QPF in NMC

7-Day 24Hour Precipitation Forecast:

Day1-3: Updated Twice a day, at 00,12UTC

Day4-7: Updated Once a day, at 00UTC

00UTC

12UTC

Threshold: 0.1, 10, 25, 50, 100, 250mm

Page 6: Improving Ensemble QPF in NMC

Various observation dataOperational

determinate model Ensemble model

Distinguish w

eather system

QP

F verification

Ensem

ble QP

FE

nsemble Q

PF

QPF Products

Multi-m

odel ensemble Q

PF

Multi-m

odel ensemble Q

PF

Point to point forecast

Synoptic situation forecast

Grid editing technique

QPF revise

Blending m

ethod QP

F

Historical data query

QP

F G

ridding

Key method

QPF technical support and operational process

Page 7: Improving Ensemble QPF in NMC

Ensemble systemEnsemble system

T213-GEPS, 10 days, 15 mem.T213-GEPS, 10 days, 15 mem. WRF-REPS, 60 hours, 15 mem.WRF-REPS, 60 hours, 15 mem. ECMWF, NCEP GEPSECMWF, NCEP GEPS TIGGE dataset (not real-time, 3 TIGGE dataset (not real-time, 3

days-delay)days-delay)

Page 8: Improving Ensemble QPF in NMC

Ensemble analysis and visualization systemEnsemble analysis and visualization system

Ensemble Predication Toolkits V0.3

probability spaghetti Stamp

Box-plot

Page 9: Improving Ensemble QPF in NMC

OutlineOutline QPF operations in NMCQPF operations in NMC Improving current QPF by ensembleImproving current QPF by ensemble Improving PQPF Improving PQPF

Page 10: Improving Ensemble QPF in NMC

Ensemble outputs as a single forecast

Mean and spread

Max, middle, min

%10, %25, %75, %90 quantile

Probability-matching ensemble mean (PM)Compared with deterministic forecast

Advantages and disadvantages of each product

How to improve current operational QPF by ensemble

Page 11: Improving Ensemble QPF in NMC

Observations:

Longitude: 110~122E

Latitude: 28~38N

Covering Huaihe catchment

745 observation stations

~0.4 degree space

Forecasts:

ECMWF global EPS

2007~2012, summer

Verification

Page 12: Improving Ensemble QPF in NMC

Verifications resultsModel forecast to stations, 1-day 24h rain rate ~ frequency

deterministic forecast, PM approximate to

ensemble member

Compared with observation curve:

— <33mm, over-forecast

— >33mm, under-forecast

Page 13: Improving Ensemble QPF in NMC

Verifications results

Mean and middle forecast:

More under-forecasts for heavy rain

No improvement for light or moderate

rain

Page 14: Improving Ensemble QPF in NMC

Verifications results

Max forecast:

More over-forecast

Close to observation for heavy rain (>150mm)

Min forecast:

More under-forecast

Close to observation for light rain (<10mm)

Page 15: Improving Ensemble QPF in NMC

Verifications results

10% 25%

75% 90%Close to obs. for different precipitation amount

Page 16: Improving Ensemble QPF in NMC

Except PM, no statistic products close to deterministic forecast

Each product has advantages and disadvantages

Can we construct a single product which fuse advantages of each product?

Page 17: Improving Ensemble QPF in NMC

Fusing productFor each grid point, there are 51 member

forecast MF.

Set fusing value FP = :(1)If max(MF) >= 100mm, then

FP=max(MF);

(2)If %90(MF) >= 50mm, then FP=

%90(MF) ;

(3)If %75(MF) >= 25mm, then FP=

%75(MF) ;

(4)If middle(MF) >= 10mm, then FP=

middle(MF) ;

(5)Else FP= %10(MF)

Page 18: Improving Ensemble QPF in NMC

Verifications results

FP approximate to observations

for different precipitation amount

Page 19: Improving Ensemble QPF in NMC

Verifications results

FP has higher Threat score than

deterministic forecast for each

precipitation amount

Threat score

Page 20: Improving Ensemble QPF in NMC

Fusing product( 1 ) Good for short-range (0~72h) QPF,

higher TS than deterministic forecast for

different amount rain.

( 2 ) Easily implemented in QPF

operations.

( 3 ) Risk of high false alarm ratio,

special for medium-range

( 4 ) Threshold decided roughly and

subjectively.

( 5 ) In future, use frequency match

algorithm to precisely calibrate frequency

error.

Page 21: Improving Ensemble QPF in NMC

OutlineOutline QPF operations in NMCQPF operations in NMC Improving QPF by ensembleImproving QPF by ensemble Improving PQPF Improving PQPF

Page 22: Improving Ensemble QPF in NMC

Verifications results2007~2012, Summer, 1day precipitation – station obs.

Under-dispersiveness:Under-dispersiveness:U shape of Talagrand histogram

Page 23: Improving Ensemble QPF in NMC

Verifications results2007~2012, Summer, 1day precipitation – station obs.

Lack of reliability:Lack of reliability:

Reliability curve not on the diagonal

•0.1mm/1Day, Overforecasting (wet bias)

•25mm/1Day, Poor resolution (overconfident)

0.1mm/1Day 25mm/1Day

Page 24: Improving Ensemble QPF in NMC

Verifications results2007~2012, Summer, 1day precipitation – station obs.

Low accuracy for high thresholds:Low accuracy for high thresholds:ROC area 0.74 < 0.8 for thresholds > 50mm/1Day

50mm/1Day

Relative operating characteristic

Page 25: Improving Ensemble QPF in NMC

Post-processingTo provide reliable forecasts

Logistic regression approach

Choice of predictors x. Estimation of the b0 and b1 over a training period. Calibrated probabilities p for a threshold T directly addressed.

Page 26: Improving Ensemble QPF in NMC

Post-processing

Logistic regression approach

Predictors: ensemble mean and spread with 1/3 power transformation

Training period: latest 30 days ; or 2007~2011 5 years summer history forecast (from TIGGE archive )

Forecast period: 2012 summer

Page 27: Improving Ensemble QPF in NMC

Post-processing0.1mm/1dayOriginal

0.1mm/1dayCalibration(history forecast)

0.1mm/1dayCalibration(30 train days)

Page 28: Improving Ensemble QPF in NMC

Post-processing10mm/1dayOriginal

10mm/1dayCalibration(30 train days)

10mm/1dayCalibration(history forecast)

Page 29: Improving Ensemble QPF in NMC

Post-processing25mm/1dayOriginal

25mm/1dayCalibration(history forecast)

50mm/1dayOriginal

50mm/1dayCalibration(history forecast)

Page 30: Improving Ensemble QPF in NMC

Logistic Regression PQPF( 1 ) Calibrate ensemble PQPF

effectively

( 2 ) More training samples, more better

results

( 3 ) History forecast errors may change

with model updating, which influence the

calibration.

( 4 ) Reforecast can offer a better way,

which we can not gain these dataset.

Page 31: Improving Ensemble QPF in NMC
Page 32: Improving Ensemble QPF in NMC

No product close to deterministic forecast Each product has advantages and

disadvantages

Can we get a statistic product which close to deterministic forecast or member forecast

Can we construct a product which fuse advantages of each product

Page 33: Improving Ensemble QPF in NMC

Probability matching

1. Rank the gridded rainfall from all n QPFs from largest to smallest, the keep every nth

value starting with the n/2-th value.

2. Rank the gridded rainfall from the ensemble mean from largest to smallest.

3. Match the two histograms, mapping rain rates from (1) onto locations from (2).

(from Beth Ebert )

……………

… … … … …

1~51

52~102

Rank form largest to smallest

Ensemble mean

Ensemblemember

Page 34: Improving Ensemble QPF in NMC

Verifications results

PM approximate to ensemble

member or deterministic forecast

Page 35: Improving Ensemble QPF in NMC

QPF Products

Day1: 6-h QPF, updated 3 times a day at 00, 06, 12 UTC

Page 36: Improving Ensemble QPF in NMC

Winter: Day1-3 24-h QPF updated twice a day

Including the snow, freezing rain, sleet.

24h 48h 72h

Precipitation phase forecast:

Page 37: Improving Ensemble QPF in NMC

Total process precipitation (for the whole life of a synoptic system)