Development of Multi-model Ensemble technique and its application Daisuke Nohara APEC Climate Center (APCC), Busan, Korea 2007/2/21 JMA
Contents 1. Introduction of APCC 2. Seasonal forecast based on multi-model ensemble 1. Process of seasonal forecast 2. Deterministic seasonal forecast 3. Probabilistic seasonal forecast 4. Verification 3. Application using APCC forecast 1. index forecast 2. statistical downscaling 4. Summary
Asia Pacific Economy Cooperation (APEC) Asia Pacific Economy Cooperation (APEC) To meet the Bogor Goals of free and open trade and investment in the Asia-Pacific region • Trade and Investment Liberalisation • Business Facilitation • Economic and Technical Cooperation
History History Proposed APCN : 3 rd APEC S&T Ministers’ Meeting Oct. 1998 Approved APCN : 17 th APEC ISTWG Meeting Aug. 1999 … Jan. 2003 Established at the Korea Meteorological Administration Proposed the establishment of APCC : 4 th APEC S&T Minister’s Meeting Mar. 2004 Supported the establishment of APCC : 27 th APEC ISTWG Meeting Sep. Endorsed the establishment of APCC : 1 st APEC Senior Official’s Meeting Mar. 2005 Welcomed the establishment of APCC : 17 th APEC Ministerial Meeting Nov. Opening Ceremony : 13 th APEC Economic Leader’s Meeting
Goals of APCC Goals of APCC • Facilitating the share of high- cost climate data and information • Capacity building in prediction and sustainable social and economic applications of climate information • Accelerating and extending socio-economic innovation
Functions of APCC Functions of APCC • Developing a value-added reliable climate prediction system • Acting as a center for climate data and related information • Coordinating research toward the development of an APEC integrated climate- environment-socio-economic system model
Multi- -Institutional cooperation Institutional cooperation Multi Meteorological Service of Canada, Environmental Canada Beijing Climate Center China Meteorological Administration International Atmospheric Physics, International Research Institute Chinese Academy of Sciences for Climate and Society Central Weather Bureau Chinese Taipei
Seasonal forecast based on Multi-Model Ensemble (MME)
Schedule of APCC seasonal forecast Schedule of APCC seasonal forecast 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 DJF MAM JJA SON DJF MAM 2007MAM forecast Feb. 22 Precipitation JJA T850 Z500 U850, U200 SON V850, V200 SLP DJF T2m OLR APCC provides 3 month forecast in each season.
Procedure of seasonal forecast Procedure of seasonal forecast Collection : 17 model datasets Quality check Composite : Deterministic Forecast (4 kinds of schemes) Probabilistic Forecast Graphic Verification: Application: - previous prediction - index forecast - hindcast - statistical downscaling Outlook : Interpretation and Description of Global/Regional Prediction Dissemination : - Web Information Up-load - Backup of the Data and Documents
APCC Participating Models APCC Participating Models Member Acronym Organization Model Resolution Economies Australia POAMA Bureau of Meteorology Research Centre T47L17 1.875 ° × 1.875 ° L50 Canada MSC Meteorological Service of Canada NCC National Climate Center/CMA T63L16 China 4 ° × 5 ° L2 IAP Institute of Atmospheric Physics Chinese CWB Central Weather Bureau T42L18 Taipei Japan JMA Japan Meteorological Agency T63L40 GDAPS/KMA Korea Meteorological Administration T106L21 Korea GCPS/SNU Seoul National University T63L21 4 ° × 5 ° L17 METRI/KMA Meteorological Research Institute MGO Main Geophysical Observatory T42L14 Russia 1.12 ° × 1.4 ° L28 HMC Hydrometeorological Centre of Russia IRI International Research Institute T42L19 COLA Center for Ocean-Land-Atmosphere Studies T63L18 USA NCEP NCEP Coupled Forecast System T62L64 National Aeronautics and Space 2 ° × 2.5 ° L34 NSIPP/NASA Administration
APCC Deterministic MME Schemes APCC Deterministic MME Schemes APCC Deterministic MME Schemes ′ 1 ∑ = P F 1. SCM: Simple composite of individual forecast i M i with equal weighting. 2. CPP – Coupled Pattern Projection Method : ′ 1 ∑ = ˆ Simple composite of individual forecasts, after P F i M correction by statistical downscaling using CPPM i 3. MRG – Multiple Regression: ′ = ∑ P a i F Optimally weighted composite of individual forecasts. i i The weighting coefficient is obtained by SVD based regression. ′ 1 ∑ 4. SSE – Synthetic Multi-Model Super Ensemble: = α ˆ P i F i M i Weighted combination of statistically corrected multi model output
APCC Deterministic MME Schemes APCC Deterministic MME Schemes APCC Deterministic MME Schemes 2. CPP – Coupled Pattern Projection Method : Simple composite of individual forecasts, after correction by statistical downscaling using CPPM Y(t) = α X P (t) + β Predictor field X P (t) X(i, j, t) ( model forecast) Predictand Y(t) Y(t) ( observation) 1. determination of spatial pattern which is related to ★ the target region. 2. calculation of regression coefficient in training period. 3. composition of all models
APCC Deterministic MME Schemes APCC Deterministic MME Schemes APCC Deterministic MME Schemes 3. MRG – Multiple Regression: Optimally weighted composite of individual forecasts. The weighting coefficient is obtained by SVD based regression. O P Actual Data Set ( N ) Prediction Regression SVD technique
APCC Deterministic MME Schemes APCC Deterministic MME Schemes APCC Deterministic MME Schemes 4. SSE – Synthetic Multi-Model Super Ensemble: Weighted combination of statistically corrected multi model output E ( ε 2 ) –Minimization using EOF Actual Data Set ( N ) Synthetic Data Set ( N ) Superensemble Prediction
APCC Deterministic MME Forecast APCC Deterministic MME Forecast APCC Deterministic MME Forecast
APCC Deterministic MME Forecast APCC Deterministic MME Forecast APCC Deterministic MME Forecast
APCC Probabilistic MME Schemes APCC Probabilistic MME Schemes APCC Probabilistic MME Schemes � Combine different models Na : num. of above-normal Nn : num. of near-normal Combine : according to each model’s square root of ensemble size Nb : num. of below-normal � Merged 3-category distribution χ 2 (Chi-square) TEST Above-normal k ∑ − 2 3- 3 -Categorical distribution Categorical distribution ( O E ) i i χ = = 2 i 1 Near-normal E i O : Forecast frequencies E : Hindcast frequencies Under 0.05% significance level Below-normal 3
APCC Probabilistic MME Schemes APCC Probabilistic MME Schemes APCC Probabilistic MME Schemes • Normal fitting method APCC Multi-Model Probabilistic Forecast Temperature at 850 hPa (2006JJA) • Probabilistic forecast Climatological Target year PDF PDF P A : Prob. of Above-Normal P N : Prob. of Near-Normal P B : Prob. of Below-Normal
Dissemination Dissemination Dissemination http://www.apcc21.net
Verification for Seasonal Forecast
Verification of Deterministic MME (model performance) Verification of Deterministic MME (model performance) Verification of Deterministic MME (model performance) Taylor Diagram : 1983~2003, JJA Hindcast Taylor Diagram : 1983~2003, JJA Hindcast individual model
Verification of Deterministic MME Verification of Deterministic MME Verification of Deterministic MME (spatial distribution of MSSS) (spatial distribution of MSSS) (spatial distribution of MSSS)
Verification of Deterministic MME ( Verification of Deterministic MME (regionality) Verification of Deterministic MME (regionality regionality) ) 0.6 0.5 0.4 0.3 ACC 0.2 0.1 0 -0.1 -0.2 AUSTRALIA EAST ASIA NORTH RUSSIA SOUTH SOUTH AMERICA AMERICA ASIA Anomaly Correlation Coefficient for hindcast in each region (precipitation, JJA, 1983-2003) multi-model individual model ensemble schemes
Verification of Deterministic MME (ACC skill score) Verification of Deterministic MME (ACC skill score) Verification of Deterministic MME (ACC skill score) individual model individual individual model model 2004DJF forecast 2005DJF forecast
Verification of Probabilistic MME Verification of Probabilistic MME Verification of Probabilistic MME Globe, Above- -Normal, Precipitation Normal, Precipitation Globe, Above Reliability Diagram ROC Curve 1.0 1.0 CWB 0.9 0.9 GCPS Observed Relative 0.8 0.8 GDAPS_F GDAPS_O 0.7 0.7 Frequency HMC 0.6 individual 0.6 IRI model 0.5 IRIF 0.5 JMA 0.4 0.4 METRI Hit rate 0.3 0.3 MGO 0.2 NCC 0.2 NCEP 0.1 0.1 PMME 0.0 계열 14 0.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 False alarm rate Forecast Probability Forecast skill of probabilistic multi-model ensemble is better than that of individual model.
Application using APCC prediction
Application using APCC prediction Application using APCC prediction • Statistical downscaling using MME prediction – Thailand, Philippine, China • Index forecast – El Nino, PNA, AO, NAO, Monsoon (figures: from NWS CPC, http://www.cpc.ncep.noaa.gov/)
Recommend
More recommend
