Subseasonal forecasting Forecasting system Design Frdric Vitart European
Sub-seasonal forecasting, Forecasting system Design Frédéric Vitart European Centre for Medium-Range Weather Forecasts Slide 1 © ECMWF
Sub-seasonal forecasting Slide 2 © ECMWF
Pioneers in subseasonal predictions Pioneering and challenging work of Miyakoda et al. (1983), Spar et al. (1976), Shukla (1981) opened the door for subseasonal predictions. These studies • explored the predictability at a subseasonal time-scale (beyond deterministic predictable Dr. Kikuro Miyakoda limit), • recognized that the subseasonal prediction Source: Princeton Univ. webpage can be seen as an initial value problem with external forcings (boundary value problem). “Predictability In the Midst of Chaos” Shukla (1998), Palmer (1993) Miyakoda et at. (1983) Simulation of a blooking event in January 1977. MWR Spar et al. (1976) Monthly mean forecast experiments with the GISS model. MWR Spar et al. (1978) An initial state perturbation experiment with the GISS model. MWR Shukla (1981) Predictability of time averages. Part I. Dynamical predictability of monthly means. JAS Slide 3 © ECMWF 3
January 1977 Slide 4 Source: NOAA/NWS © ECMWF http: //www. srh. noaa. gov/images/mfl/news/Snow. South. Florida 35 th. pdf 4
January 1977 T 850 Forecast (Day 10 -30) Miyakoda et al. 1983 Slide 5 © ECMWF 5
First report to the international community • Cubasch, Tibaldi, Molteni: Deterministic extended-range forecast experiments using the global ECMWF spectral model • Molteni, Cubasch, Tibaldi: Experimental monthly forecasts at ECMWF using the lagged-average forecasting technique • • • 4 case studies in winter 1983/84 9 -member lagged-average forecasts I. C. from operational analysis at 6 -hour interval T 21 and T 42 spectral model Fixed SST, persisted from I. C. (no cheating!) Correction for systematic error, based on 10 30 -day integrations in winters 1981/82 and 1982/83, started at 10 -day intervals • Comparison w. r. t. deterministic forecast from last I. C. and persistence Slide 6 © ECMWF
Sub-seasonal forecasts Sub-seasonal forecasting is still in its infancy. 10 years ago, only a couple of operational centres were producing sub-seasonal forecasts. Now most of the Global producing Centres are producing and issuing or experimenting sub-seasonal forecasts. From A. Kumar (NCEP/CPC) Slide 7 © ECMWF
Experimental Week 3+4 outlook From A. Kumar (NCEP/CPC) Slide 8 © ECMWF
Since 1983, most producing centres have developed sub-seasonal forecasts Timerange Resol. Ens. Size Freq. Hcsts ECMWF D 0 -32 T 639/319 L 91 51 2/week On the fly UKMO D 0 -60 N 216 L 85 4 daily NCEP D 0 -44 N 126 L 64 4 4/daily Fix EC D 0 -35 0. 6 x 0. 6 L 40 21 weekly CAWCR D 0 -60 T 47 L 17 33 JMA D 0 -34 T 159 L 60 KMA D 0 -60 CMA Hcst length Hcst Freq 2/weekly 11 4/month 3 1999 -2010 4/daily 1 On the fly Past 15 y weekly 4 weekly Fix 1981 -2013 6/month 33 50 weekly Fix 1979 -2009 3/month 5 N 216 L 85 4 daily On the fly 1996 -2009 4/month 3 D 0 -45 T 106 L 40 4 daily Fix 1992 -now daily 4 Met. Fr D 0 -60 T 127 L 31 51 monthly Fix 1981 -2005 monthly 11 CNR D 0 -32 0. 75 x 0. 56 L 54 40 weekly Fix 1981 -2010 6/month 1 HMCR D 0 -63 1. 1 x 1. 4 L 28 20 weekly Fix 1981 -2010 weekly 10 Slide 9 © ECMWF Past 20 y Hcst Size On the fly 1989 -2003
Sub-seasonal Forecast Configuration Different strategy for sub-seasonal forecasting: § In some centres, sub-seasonal forecasts use the same forecasting system as the seasonal forecasting system (e. g. UKMO, NCEP). More frequent start date or larger ensemble size. § In other centres, it is an extension of medium-range weather forecast (e. g. ECMWF/EC) § In other centres it is a separate system which contains characteristics from both medium-range and seasonal forecasting (e. g. ECMWF before 2008/JMA) Slide 10 © ECMWF
Sub-seasonal Forecast Configuration Very different configurations of the sub-seasonal forecasting systems (much more than for medium-range or seasonal forecasting). There is currently no consensus on the optimal configuration. Differences in configuration include: § Frequency of forecasts (daily/weekly/monthly) § Ensemble size: e. g. large ensembles run once a week (burst sampling) vs small ensembles run daily (lag ensemble approach) § Model resolution: currently from about 250 km to 50 km § Time range: between 32 and 60 days § Different model set-up: Ocean atmosphere coupling/active sea-ice Slide 11 © ECMWF
Main contribution to YOPP: S 2 S database Models Time-range Freq. Hcst length Hcst Freq Ocean coupling Active Sea Ice ECMWF D 0 -46 2/week Past 20 y 2/weekly YES Planned UKMO D 0 -60 daily 1996 -2009 4/month YES NCEP D 0 -44 4/daily 1999 -2010 4/daily YES EC D 0 -35 weekly Past 15 y weekly NO NO Bo. M D 0 -60 2/weekly 1981 -2013 6/month YES Planned JMA D 0 -34 weekly 1981 -2010 3/month NO NO KMA D 0 -60 daily 1996 -2009 4/month YES CMA D 0 -45 daily 1992 -now daily YES Met. Fr D 0 -60 monthly 1993 -2014 monthly YES ISA-CNR D 0 -32 weekly 1981 -2010 6/month YES NO HMCR D 0 -63 weekly 1981 -2010 weekly NO NO Slide 12 © ECMWF
Example. The new ECMWF Ensemble fc. system The ECMWF ensemble prediction system for the medium and sub-seasonal range IFS 41 r 1 32/64 km grid (T 636/319) 91 levels Coupling in single executable NEMO 1/1 -0. 3 d. lon/lat 42 levels H-TESSEL Initial conditions 4 -D variational d. a. 3 -D v. d. a. (NEMOVAR) Slide 13 © ECMWF perturbations EDA pert. sing. vectors 5 ocean analyses Ens. Forecast CGCM 51 runs T 639 to 10 d T 319 to 46 d
Sub-seasonal forecast products Short and medium-range forecasts: instantaneous/daily values Seasonal forecasting: Main products are seasonal or monthly means. Sub-seasonal forecast: Beyond 2 weeks, there is little predictability in the day to day variability, but there is some skill in predicting weekly mean anomalies. Slide 14 © ECMWF
ECMWF sub-seasonal forecasts Anomalies (temperature, precipitation. . ) - Slide 15 © ECMWF
Probabilities (temperature, precipitation. . ) Slide 16 © ECMWF
Weather Regimes Slide 17 © ECMWF
Tropical cyclone activity Slide 18 © ECMWF
MJO Forecasts Slide 19 © ECMWF
Sub-seasonal forecasts and re-forecasts Model systematic errors grow during the model integrations and after 2 weeks can be as big as the signal we want to predict. Two options: 1. Make corrections during the model integrations (bias or flux correction) (popular in the climate simulations) 2. Make a-posteriori corrections. The coupled ocean-atmosphere model is run freely and the model systematic errors are estimated from a set of model re-forecasts (same technique as for seasonal forecasting). Implicit assumption of linearity. We implicitly assume that a shift in the model forecast relative to the model climate corresponds to the expected shift in a true forecast relative to the true climate, despite differences between model and true climate. Most of the time, assumption seems to work pretty well. But not always. Slide 20 © ECMWF
Biases (eg 2 m. T as shown here) are often comparable in magnitude to the anomalies which we seek to predict Slide 21 © ECMWF
Since 1983, most producing centres have developed sub-seasonal forecasts Timerange Resol. Ens. Size Freq. Hcsts ECMWF D 0 -32 T 639/319 L 91 51 2/week On the fly UKMO D 0 -60 N 216 L 85 4 daily NCEP D 0 -44 N 126 L 64 4 4/daily Fix EC D 0 -35 0. 6 x 0. 6 L 40 21 weekly CAWCR D 0 -60 T 47 L 17 33 JMA D 0 -34 T 159 L 60 KMA D 0 -60 CMA Hcst length Hcst Freq 2/weekly 11 4/month 3 1999 -2010 4/daily 1 On the fly Past 15 y weekly 4 weekly Fix 1981 -2013 6/month 33 50 weekly Fix 1979 -2009 3/month 5 N 216 L 85 4 daily On the fly 1996 -2009 4/month 3 D 0 -45 T 106 L 40 4 daily Fix 1992 -now daily 4 Met. Fr D 0 -60 T 127 L 31 51 monthly Fix 1981 -2005 monthly 11 CNR D 0 -32 0. 75 x 0. 56 L 54 40 weekly Fix 1981 -2010 6/month 1 HMCR D 0 -63 1. 1 x 1. 4 L 28 20 weekly Fix 1981 -2010 weekly 10 Slide 22 © ECMWF Past 20 y Hcst Size On the fly 1989 -2003
Sub-seasonal Re-forecasts Two strategies for re-forecasts in S 2 S database: § Fixed re-forecasts (e. g. NCEP/Bo. M/JMA) The model version used to produce the sub-seasonal forecasts is “frozen” for a number of years (e. g. CFS 2). The re-forecasts have been produced once for all before the system became operational. Advantage: More user friendly. The user can compute skill and calibration once for all. § “on the fly” re-forecasts (e. g. ECMWF/UKMO/EC. . ) The model version changes frequently (at least once a year). Therefore re-forecasts have to produce regularly since the model version of the re-forecasts has to be the same as the real-time forecasts. Advantage: This methodology ensures the best model version has been used to produce the sub-seasonal forecasts. Slide 23 © ECMWF
The ENS re-forecast suite to estimate the M-climate … 28 6 13 20 27 March … 51 T 639 L 91 2014 5 2013 5 2012 20 y 5 2011 5 51 T 319 L 91 5 5 5 5 5 5 5 Slide 24 ERA Interim+ ORAS 4 ocean Ics+ Soil reanalysis Perturbations: SVs+EDA(2015)+SPPT+SKEB …. . 1994 Initial conditions: 5 © ECMWF 5 5 5 5
Why not using a 5 -week window? Week 0 Week +1 Week -2 Week +2 Example: Climate of 06/06 day 26 -32: 1 -week climate – 5 -week climate Slide 25 © ECMWF
Re-forecast strategy Re-forecasts are used for model calibration and also for skill assessment. § A large reforecast database is needed for calibration to distinguish between random error and systematic errors and also to estimate flow dependent errors. § A large reforecast database is also needed for verification and for flow dependant skill assessment, like assessing the concurrent impact of ENSO and specific phases of the MJO on the forecast skill scores. Signal to noise ration is also improved in long reforecast datasets (Shi et al, 2014) § Large ensemble size is also important for skill assessment , since some probabilistic skill scores are impacted by the ensemble size. However § Large re-forecast datasets with large ensemble size are often not affordable. Not clear what is more important: ensemble size, number of years? § Long re-forecasts suffer from inconsistent quality in the initial conditions (pre-satellite period). Slide 26 © ECMWF
VERIFICATION Slide 27 © ECMWF
ECMWF Extended-range forecasts Slide 28 28 © ECMWF
Precip anomalies : 26 July 2010 – 01 August 2010 Slide 29 © ECMWF
ECMWF Monthly Forecast Skill scores ROC area – Probability of 2 mtm in upper tercile Slide 30 © ECMWF
Skill of the ECMWF Monthly Forecasting System 2 -meter temperature in upper tercile - Day 12 -18 ROC score Reliability diagram Persistence of day 5 -11 Day 12 -18 Monthly forecast day 12 -18 Day 19 -32 Persistence of day 5 -18 Monthly forecast day 19 -32 Slide 31 © ECMWF
Skill can be flow dependant – Windows of opportunity Impact of MJO on forecast reliability T_850 > upper tercile, fc. day 19 -25 Blue line: no MJO in IC Red line: MJO in IC Slide 32 © ECMWF
Linkage with SNAP Impact of SSWs on forecast skill scores From Om Tripathi Slide 33 © ECMWF 33
Model development Slide 34 © ECMWF
Resolutions of One-month EPS at JMA GSM 0801 C GSM 0603 C km GSM 1304 h. Pa Grid resolution GSM 0103 Wave number GSM 9603 Num. of vert. lev. Model top Ensemble size Year x 3 horizontal resolution, x 1. 5 vertical levels, x 5 ensemble size * Slide 35 © ECMWF Indicates changes with resolution/ensemble size upgrades, only 35
Evolution of the ECMWF sub-seasonal ensemble forecasts Mar 2002 Frequency Oct 2004 Every 2 weeks Horizontal resolution Vertical resolution Feb 2006 40 levels Top at 10 h. Pa Ocean/ atmosphere coupling Every hour from day 0 Re-forecast period Past 12 years Re-forecast size Slide 36 © ECMWF Jan 2010 Nov 2011 Nov 2013 Once a week T 159 day 0 -32 Initial conditions Mar 2008 T 319 day 0 -10 T 255 day 10 -32 Twice a week T 639 day 0 -10 T 319 day 10 -32 62 levels Top at 5 h. Pa T 639 day 0 -10 T 319 day 10 -46 91 levels Top at 1 Pa Every 3 hours from day 10 Past 18 years Every 3 h from day 0 Past 20 years 5 members, once a week ERA 40 May 2015 11 members, twice a week ERA Interim 36
A success story: forecasting the Madden-Julian Oscillation Wheeler – Hendon (2004) MJO metric based on composite EOFs Slide 37 © ECMWF
MJO skill scores Slide 38 © ECMWF
MJO teleconnections in October-March 500 h. Pa height, MJO phase 3 + 10 days Slide 39 © ECMWF
Skill scores are improving! Slide 40 © ECMWF
Grid mesh/resolution and sp. harmonic truncation in spectral models Linear grid: Quadratic grid: Cubic grid: Slide 41 © ECMWF spectral truncation N-1, 2 N grid points at the equator spectral truncation N-1, 3 N grid points at the equator spectral truncation N-1, 4 N grid points at the equator “Reduced” grid: Octahedral grid: No. of points in longitude decreases in steps decreases continuously October 29, 2014 41
Impact of resolution upgrade on sub-seasonal scores Slide 43 © ECMWF October 29, 2014 43
Impact of resolution on track probability- Tropical cyclone PAM, 9 -15/03/ 2015 Day 12 -18 Day 19 -25 Oper TL 639/319 High Tco 639/31 9 Tco 639 Observed track Slide 44 © ECMWF
Slide 45 © ECMWF
Sea Surface Temperatures MJO Bivariate Correlation RPSS over NH U 50 T 850 Coupled Obs SSTs WEEK 1 WEEK 2 WEEK 3 Obs SSTs WEEK 4 Coupled Pers SSTs 80 case, starting on 1 st Feb/May/Aug/Nov 1989 -2008 Slide 46 © ECMWF
Correlations for week 4 Northern Hemisphere Winter Current system With seaice model (LIM 2) Slide 47 © ECMWF Summer
Active sea ice model: Z 500 Forecast Skill (weeks 1 -4) SEA ICE Control 80 cases – The vertical bars represent the 95% level of confidence Slide 48 © ECMWF
New Higher-resolution Ocean Reanalysis Slide 49 © ECMWF
New higher-resolution ocean model 1/4 vs 1 degree – Z 500 skill scores -NH Slide 50 © ECMWF
Conclusions Sub-seasonal forecasting is still in its infancy. There is no consensus on the optimal forecasting system. S 2 S database will help compare the various forecasting systems. S 2 S forecasts need calibration. Flow dependant calibration however would need more re-forecasts than currently produced. Sub-seasonal forecasts have improved over the past 10 years, but skill at week 4 is still marginally better than climatology. Model are getting more complex, with higher resolution and more components of the earth system. Slide 51 © ECMWF
- Slides: 50