Comments
Description
Transcript
Verification of CFS forecasts Daily
Verification of Daily CFS forecasts Huug van den Dool & Suranjana Saha • CFS was designed as ‘seasonal’ system • Hindcasts 1981-2005, 15 ‘members’ per month • Here we look at CFS(T62L64) as NWP (never mind the delayed ocean analysis) • 25 years of forecasts (4500 forecasts out to 9 months) by a ‘constant’ T62L64 model !!!!!. Climatology of daily scores • • No ensemble average, no time average Compare to CDAS & CDC’s MRF : Cathy Thiaw (reruns) Acknowledgement and on CDAS info Bob Kistler/Fanglin Yang/Pete Caplan Day 5 AC-scores, using the harmonically smoothed model and observed climatologies (which are more ‘competitive’ than the old-old-old climo used on Pete Caplan’s page Variables: Extra-tropical Z500 (NH, SH) PSI200 and CHI200 Annual Mean CFS da Z500 - NH 1981-2005 Anomaly Correlation 80 70 60 180 forecasts per dot 50 40 30 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 YEAR Z500: 72.3 (4500 cases, grand mean) Congratulations with a constant system! From Pete Caplan’s EMC website: 0.7 Grand Mean NH over 1984-2005: 70.5 (CDAS1) ((CDAS statistics has several problems)) 0.7 Annual Mean CFS da Z500 - NH 1981-2005 (CD Anomaly Correlation 80 70 60 50 40 30 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 YEAR CDAS Annual Mean CFS da Z500 - NH 1981-2005 Anomaly Correlation 80 70 60 50 40 30 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 YEAR CFS Next Two Slides: Excursion to SH Z500 Annual Mean CFS da Z500 - SH 1981-2005 Anomaly Correlation 80 70 60 50 40 Scores ‘volatile’ in SH (still 180 per dot) SH (CFS) 30 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 YEAR Z500: 62.9 (4500 cases, grand mean) Scores go UP and UP! “Congratulations” to whom ??? Annual Mean CFS d Z500 - SH 1981-2005 (CD Anomaly Correlation 80 70 60 50 40 SH (CDAS) 30 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 YEAR Doing the best we can: Comparing CFS to CDAS day 5 Z500 scores) NH SH CFS 1981-2005 T62L64, coupled, R2 0.723 0.629 CDAS 1984-2005 T62L28, R1 0.705 0.623 Very flawed Warning: Number of forecasts per month differ. Climos differ!!! and change in ’96 for CDAS Prelim Conclusion • CFS is (slightly) better than CDAS. • NH much better than SH (typical for pre2000 technology) – this will change completely in next CFS • Trend-issues and non-constancy of system (will get worse) Next Two Slides: Excursion to TROPICS Annual Mean CFS da PSI200 - TROPICS 1981-2 80 Anomaly Correlation ENSO? 70 60 180 forecasts per dot 50 40 30 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 YEAR PSI200: 63.0 (4500 cases grand mean) No increase??? Annual Mean CFS da CHI200 - TROPICS 1981-2 80 Anomaly Correlation ENSO? 70 180 forecasts per dot 60 50 40 30 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 YEAR CHI200: 45.9 (4500 cases grand mean) This may be lowish for MJO type prediction. Definitely no increase over time. Perhaps a decrease!!! How about Bias Correction??? One of the claimed usages of hindcasts Daily CFS Scores Based on 375 forecasts. Anomaly Correlation Z500 1981-2005 NH July 100 Indisputable but very small improvement. 80 60 Is Z500 incorrigeable? 40 20 0 0 2 4 raw 6 8 10 12 forecast day 14 bias corrected The gain due to bias correction in a few selected months. Day 5 scores Z500 1981-2005 Raw Bias Corrected NH DEC 73.0 74.5 (+1.5) NH JUL 65.8 68.3 (+2.5) SH FEB 61.8 63.5 (+1.7) SH AUG 62.2 63.7 (+1.5) Is Z500 incorrigeable? Largely Yes, because the systematic error is small. Wait till you see CHI200 Daily CFS Scores Anomaly Correlation CHI200 1981-2005 TR December 100 Chi200 improves tremendously from cleaning up the bias, especially early on. 80 60 40 20 0 0 2 4 6 8 10 12 forecast day 14 375 forecasts Can we do MJO forecasts ?? raw bias corrected Daily CFS Scores Anomaly Correlation PSI200 1981-2005 TR February 100 80 60 40 20 0 0 PSI200 in Tropics does not improve very much from bias removal. !!!! 2 4 raw 6 8 10 12 forecast day 14 bias corrected Distribution of Skill in space is important and largely unexplored and unexplaind Skill as a function of (EOF) mode is interesting CFS Skill in Daily Z500 pred Anomaly Correlation 100 80 60 40 20 0 -20 10 8 6 4 2 0 -2 0 10 20 30 40 50 60 70 80 90100 EOF mode # skill_day-5 EV EV by Mode (%) February 1981-2005 NH 100 80 60 40 20 0 -20 February 1981-2005 NH 10 8 6 4 2 0 -2 0 10 20 30 40 50 60 70 80 90100 EOF mode # skill_day-5 EV Regular AC persistence persistence's regular AC EV by Mode (%) Anomaly Correlation CFS Skill in Daily Z500 pr Now: OUT TO 270 DAYS !! 100 80 Die-off curve, Z500, NH, 1981-2005, 4500 forecasts, all seasons aggregated 60 40 20 0 0 15 30 45 60 75 90105120135150165180195210225240255270 100 1.5 Detail: Month 2 to Month 8 80 1 Detail, first month. 60 NH 40 0.5 0 20 0 -0.5 0 5 10 15 20 25 30 30 45 60 75 90 105120135150165180195210225240255270 100 80 Die-off curve, Z500, SH, 1981-2005, 4500 forecasts, all seasons aggregated 60 40 20 0 0 15 30 45 60 75 90105120135150165180195210225240255270 100 2 80 1.5 60 SH 40 1 0.5 20 0 0 0 5 10 15 20 25 30 30 45 60 75 90 105120135150165180195210225240255270 FOCUS • • • • • Week 3 Week 4 = days 15-21 and 22-28 Physical basis of wk3/wk4 Ocean interaction is as much a liability as a promise at this point in time. DAY 15 DAY 28 CFS Pers CFS Pers NH 15.2 4.2 6.1 TR 36.2 29.3 29.6 27.5 SH 9.6 2.0 6.2 3.9 3.4 Anomaly correlation of CFS and ‘Persistence’ Z500 prediction in Feb 1981-2005 signal ______________________ ratio NOISE Improving signal to noise ratio: (often cosmetic) by reducing noise by applying some operator and, hopefully, not hurting the signal by this operator 1) Take a time mean (7 days) 2) Ensemble mean (not done here) 3) EOF filters (or better (maximally predictable modes)) Effect of a 7 day mean……… Day 15 Wk3 Wk4 Day 28 NH 15.2 15.7 9.2 TR 36.2 46.7 42.9 29.6 SH 9.6 10.6 5.2 6.1 2.0 Anomaly correlation of CFS Z500 prediction, daily as well as weekly, in Feb 19812005 Effect of EOF truncation…… Wk3 Wk3 WK3 WK4 EOF1 EOF1-10 WK4 WK4 EOF1 EOF1-10 15.7 22.0 23.9 18.7 9.2 12.0 So, overall, we went from daily scores (15.2….6.1) to weekly scores (15.7 and 9.2) to filtered weekly scores (22.0 and 23.9 at best) 15.7 9.2 Conclusions • CFS is (slightly) better than CDAS (Z500). • Over time (1981-2005), the CFS “system” appears quite constant with various qualifiers • If we need an as-constant-as possible in-house system, look no further (better than CDAS) • Bias correction: small +ve impact on Z500 and Ψ200 (1-2pnts) • Bias correction: large +ve impact on CHI-200 (15 pnts) • CFS loses 5-20% in terms of SD and eDOF in the first few days, then, admirably, stays nearly constant out to 270 days Conclusions • (Very) modest skill in wk3 and wk4 • Even with 2 of 3 signal to noise improvements in place AC is only 0.200.25. (No ensemble averagehere) • Waiting for the next CFS and CFSRR (higher Res, consistent IC) Conclusions: • The day 1 – 3 forecasts appear to be too damped, and damp faster than a regression would. Increasing anomaly amplitude as a postprocessor (undoing the sd decrease) actually improves the rms error early on. A curiosity? • Probably: initial conditions are damped as well. THE REST IS EXTRA CFS daily scores 1981-200 78 40 76 38 74 36 72 34 70 32 68 30 66 28 64 26 9 10 11 12 13 0 1 2 3 4 5 6 7 8 AC-NH at day 5 degrees of freedom Does the eDOF variation explain the AC variation???? CFS Skill in Daily Z500 pred Anomaly Correlation 100 80 60 40 20 0 -20 10 8 6 4 2 0 -2 0 10 20 30 40 50 60 70 80 90100 EOF mode # skill_day-5 EV EV by Mode (%) February 1981-2005 NH skill_day-10 CFS Skill in Daily Z500 pred 100 80 60 40 20 0 -20 10 8 6 4 2 0 -2 0 10 20 30 40 50 60 70 80 90100 EOF mode # skill_day-5 EV skill_day15 skill_day28 skill_day-10 EV by Mode (%) Anomaly Correlation February 1981-2005 NH Anomaly Correlation Daily CFS Scores Z500 1981-2005 NH May13/15 hurricane 100 80 60 40 20 0 It is not just resolution…. Member 5 0 2 4 6 8 10 12 forecast day 14 Only 25 forecasts T382('hurricane') CFST62 May13/15 CFS day Z500 - NH 1981-2005 T382 used GDAS Anomaly Correlation 90 80 70 60 50 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 YEAR CFS@T382 CFS@T62 RMSE growth CFS NOSEC RMSE (gpm) 160 140 120 100 80 60 40 20 0 Savijarvi Eq(5 CFS-jan 0 1 2 3 4 5 6 7 8 910111213141516 lead Initial error=6.1 gpm ; Systematic error growth s=13.5 gpm/day ; Small error amplification a=0.181/day ; e-infinity=157.7gpm ; NOSEC Savijarvi’s equation 5: RMSE growth CFS SEC RMSE (gpm) 160 140 120 100 80 60 40 20 0 Savijarvi Eq(5 CFS-jan 0 1 2 3 4 5 6 7 8 910111213141516 lead Initial error=4.7 gpm ; Systematic error growth s=13.1 gpm/day ; Small error amplification a=0.181/day ; e-infinity=155.7gpm ; SEC RMSE growth CFS SECSD RMSE (gpm) 160 140 120 100 80 60 40 20 0 Savijarvi Eq(5 CFS-jan 0 1 2 3 4 5 6 7 8 910111213141516 lead Initial error=4.7 gpm ; Systematic error growth s=12.6 gpm/day ; Small error amplification a=0.200/day ; e-infinity=158.3gpm ; SECSD Example: Systematic errors for mid-January at day 5. RMSE growth CFS NH 'jan 75 3 65 2 55 1 45 0 35 -1 25 -2 15 -3 0 raw 1 2 sec 3 secsd 4 sec-secsd Climatological Annual Cycle of day-5 scores Climo Day 5 Scores byNH: Month Best month is Feb (76.4), Z500 - NH 1981-2005 Worst month is July (67.7). Near sinusoidal variation. Range=8.7pnts Anomaly Correlation 80 70 NH 60 1 2 3 4 5 6 7 8 Month 9 10 11 12 Climo Day 5 Scores by M Z500 - SH 1981-2005 SH: Best month is Aug (64.9), Worst month is March (59.0). Typically 62-64, except Feb-Apr. Range=5.9pnts Anomaly Correlation 70 60 SH 50 1 2 3 4 5 6 7 8 Month Forecasts verifying in month shown Curious: March is best/worst in NH/SH. Aug is best/worst in SH/NH. 9 10 11 12 375 forecasts per month Climo Day 5 Scores by Month Z500 - NH 1981-2005 Anomaly Correlation 85 The NH has an annual cycle in skill, every year. 75 (Each color curve is a different year.) 65 55 1 2 3 4 5 6 7 8 Month 9 10 11 12 Climo Day 5 Scores by Z500 - SH 1981-2005 Anomaly Correlation 75 65 The SH does NOT have a clear annual variation in skill each year. 55 45 1 2 3 4 5 6 7 8 Month 15 forecasts per month 9 10 11 12 Climo Day 5 Scores by Month Climo Day 5 Scores by M Z500 - NH 1981-2005 Anomaly Correlation 70 Anomaly Correlation 80 Z500 - SH 1981-2005 70 60 60 50 1 2 3 4 5 6 7 8 Month 9 10 11 12 1 2 3 4 5 6 7 8 Month 9 10 11 12 Climo Day 5 Scores by Month Climo Day 5 Scores by M PSI200 - TROPICS 1981-2005 Anomaly Correlation 55 Anomaly Correlation 75 CHI200 - TROPICS 1981-20 65 45 55 35 1 2 3 4 5 6 7 8 Month 9 10 11 12 1 2 3 4 5 6 7 8 Month 9 forecasts verifying in month shown 375 forecast per dot(month) Large annual variation in the tropics! But volatile. 10 11 12 Day 5 scores • AC (obviously; already shown), But also: • SD • eDOF CFS daily scores 1981-200 standard deviation (gpm) 120 110 CFS misses 5-20 % of variability 100 90 80 70 60 d j f m a m j j a s o n d j sd (day 5)sd (OBS) Loss of variance does not increase beyond 10-15 days and is never more than 20% CFS daily scores 1981-200 effective degrees of freedom 44 42 40 38 36 34 32 30 28 26 CFS misses several degrees of freedom d j f m a m j j a s o n d j deg of freedom (OBS)degrees of freedom Loss of dof does not increase beyond 10-15 days and is never more than shown above CFS daily scores 1981-2005 Z500 NH standard deviation (gpm) 120 110 100 90 80 70 60 d j f m a m j j a s o n d Compare the SD reduction in the j sd (day 5)sd (OBS) NH CFS daily scores 1981-2005 Z500 SH To that in the standard deviation (gpm) 120 SH 110 100 90 80 70 60 d j f m a m j j a s o n d sd (day 5)sd (OBS) j CFS daily scores 1981-2005 Z500 NH effective degrees of freedom 44 42 40 38 36 34 32 30 28 26 d j f m a m j j a s o n d j Compare the DOF reduction in the deg of freedom (OBS)degrees of freedom (day 5) NH CFS daily scores 1981-2005 Z500 SH To that in the effective degrees of freedom 36 SH 34 32 30 28 26 d j f m a m j j a s o n d j deg of freedom (OBS)degrees of freedom (day 5)