Seasonal Forecasts and Predictability Masato Sugi Climate Prediction
Seasonal Forecasts and Predictability Masato Sugi Climate Prediction Division/JMA 1
History of Seasonal Forecasts at JMA 1942 Statistical One-month and Three-month forecasts 1943 Statistical Warm/Cold season forecasts 1996 1999 2003 Dynamical One month forecast El Nino Outlook with Coupled Model Dynamical Three month forecast Dynamical Warm/Cold season forecasts 2
Operational models for seasonal forecasts at JMA One month forecasts : AGCM with persistent SSTA T 106 L 40 GSM 0103 26 member Three month forecasts: AGCM with persistent SSTA T 63 L 40 GSM 0103 31 member Warm/Cold season forecasts: Two tier method T 63 L 40 GSM 0103 31 member using SSTA predicted CGCM 02 3
Seasonal Forecasts Issuance time Lead time Forecast period Forecast range Lead time Forecast period 1 month 0 - 2 week 1 - 4 week 3 month 0 - 2 month 1 - 3 month 6 month 0 - 3 month 4
Analysis of Variance (ANOVA) Decomposition of meteorological variable: If and are statistically independent, then Variance explained by the i-th component : correlation between and 5
Decomposition of observed variable : predictable signal : unpredictable noise : variance of signal : variance of noise Potential predictability gives the upper limit of forecast skill. 6
: Predictable signal : Unpredictable noise Variance noise variance climatological total variance signal variance Forecast lead time 7
Predictable signal and unpredictable noise Predictable signal: - some low-frequency internal modes - externally forced slowly varying modes - decadal modes - trends due to global warming Unpredictable noise: - high-frequency internal modes - most low-frequency modes that have strong interaction with high-frequency modes In seasonal forecasts, most important predictable signal is SST forced variability. 8
Ensemble forecasts - starting from slightly different initial conditions - with the same boundary condition (SST) 9
Estimating potential predictability R from ensemble simulation : simulated variable : predictable signal : unpredictable noise : ensemble mean : deviation from potential predictability 10
Ensemble simulation experiment - MRI-JMA 98 AGCM T 42 L 30 - GISST 1949 - 1998 - 6 -member, 50 -year simulation 11
12
13
14
15
JJA DJF 16
Forecast PDF 17
Three-Category Forecast Climatological PDF PB 33% PN PA 33% - 0. 43 c 0 0. 43 c PA : probability of Above normal PN : probability of Normal PB : probability of Below normal 18
Probability of three categories Forecast PDF - 0. 43 c 0 xs 0. 43 c PA : probability of Above normal PN : probability of Normal PB : probability of Below normal 19
Deterministic category forecast Category of highest probability Forecast category Forecast PDF Percent Correct (Pc) : percentage of correct forecast 20
0. 01 0. 04 0. 09 0. 16 0. 25 0. 36 0. 49 0. 5 0. 64 0. 7 0. 81 0. 9 0. 0 0. 1 0. 2 0. 316 0. 447 0. 548 0. 632 0. 707 0. 775 0. 837 0. 894 0. 949 1. 0 0. 995 0. 980 0. 954 0. 949 0. 917 0. 894 0. 866 0. 837 0. 800 0. 775 0. 714 0. 707 0. 632 0. 600 0. 548 0. 447 0. 436 0. 316 33 % 36 39 42 43 46 47 49 51 54 55 58 59 63 65 68 73 74 82 21
Overall skill of seasonal forecasts for seasonal mean temperature over Japan Percent correct of three category forecasts: 40~50% This value corresponds to the correlation between ensemble mean and observation: 0. 23~0. 52 Even though the percent correct is 40~50% probability forecast is still useful. 22
For example, if percent correct is 47% , then correlation is 0. 44, s = 0. 44 c , n = 0. 90 c. If forecast ensemble mean Xs = 0. 4 c , then Climatological PDF Forecast PDF 23
If potential predictability is 50% , then correlation is 0. 707, s = 0. 707 c , n = 0. 707 c. If forecast ensemble mean Xs = 0. 7 c , then Climatological PDF Forecast PDF 24
Summary • In seasonal forecasts , it is important to understand the predictability and intrinsic uncertainty. • Potential predictability is generally high in the tropics but low in the extratropics. • Although there is a large uncertainty in seasonal forecasts, the forecast probability information is still potentially useful. • Application technology of probability forecast to agriculture, water management, health, energy, etc. , need to be developed. 25
Appendix 26
Estimation error in R due to model deficiency underestimated overestimated underestimated 27
A proposal for estimating model independent potential predictability 28
Ensemble mean for large ensemble size We further assume then 29
correlation RMSE 30
Perfect model Climatology forecast 31
Ensemble mean better skill because Perfect model 32
Multi model ensemble mean better skill when 33
Multi model ensemble mean If and for all i then 34
Multi model ensemble mean if but then weighted average improves the skill 35
Estimating ensemble simulations from multi model if 36
Summary By using multi-model ensemble simulations we can estimate 1) model independent signal variance and potential predictability, 2) signal amplitude and model error variance for each model, 3) optimum weight for multi-model ensemble 37
- Slides: 37