Downscaling ensembles using forecast analogs

1
Downscaling ensembles using forecast analogs

• Jeff Whitaker and Tom Hamill
• tom.hamill_at_noaa.gov
• jeffrey.s.whitaker_at_noaa.gov

2
CDC MRF Reforecast Data Set

• Definition a data set of retrospective numerical
  forecasts using the same model to generate
  real-time forecasts
• Model T62L28 MRF, circa 1998 (http//www.cdc.noa
  a.gov/people/jeffrey.s.whitaker/refcst for
  details).
• Initial States NCEP Reanalysis plus 7 /- bred
  modes (Toth and Kalnay 1993).
• Duration 15 days runs every day at 00Z from
  19781101 to now. (http//www.cdc.noaa.gov/people/j
  effrey.s.whitaker/refcst/week2).
• Data Selected fields (winds, hgt, temp on 5
  press levels, precip, t2m, u10m, v10m, pwat,
  prmsl, rh700, heating). NCEP/NCAR reanalysis
  verifying fields included (Web form to download
  at http//www.cdc.noaa.gov/reforecast).

3
Applications

• Predictability studies
• Diagnosis of model error
• Statistical correction of real-time forecasts
• 6-10 day and week 2 CPC temp and precip tercile
  probabilities ? (now operational)
• Uses logistic regression at stations (Hamill et
  al, 2004, MWR, p. 1434)

4
HSS scores 9/10/03- 9/9/04 Week 2 Temp Official
14.74 CDC 16.80 Precip Official 10.27 CDC
8.09
5
But these forecasts are very coarse resolution

• Finer-scale detail is desirable, especially for
  precip.
• How can we take large-scale NWP/GCM output and
  downscale it to provide skillful
  higher-resolution forecasts?
• How to correct for regime-dependant errors?

6
Analogtechnique(pioneered by van den Dool,
Toth, von Storch, others)
Step 3 extract observed weather
Step 2 find dates of old analogs
Forecast analog 1, 2/12/95
Observed Wx, 2/12/95
TODAYS ENS MEAN PRECIP FORECAST
Forecast analog 2, 1/16/98
Observed Wx, 1/16/98
Step 1 make todays forecast
Forecast Analog 3, 3/1/83
Observed Wx, 3/1/83
BMA?
7
Local analogs are patched together

• Initial implementation very simple
• Single forecast field (precip).
• L2 norm (rms) using ens. mean fcst.
• Analog ensemble members receive equal weight.
• 50 analog members - NARR.

8
Example 4-6 day analog forecasts, valid 29-31
Dec 1996)
9
Skill of Analog Forecasts
10
Skill of Analog Forecasts
11
Skill of Analog Forecasts
3 days
12
Skill of Analog Forecasts
13
Application - Tercile Forecasts

• Prob of above normal for 2nd N days of forecast
  (N1 to 6).
• All JFMs 1979-20035 (no analogs within /- 45
  days of verifying analysis used).
• NARR precip over entire CONUS.

day 2
days 3-4
days 4-6
days 5-8
days 6-10
14
Analog Forecast Skill - Upper Tercile
15
Analog Forecast Skill - Upper Tercile
16
Analog Forecast Skill - Upper Tercile
17
Analog Forecast Skill - Upper Tercile
18
Analog Forecast Skill - Upper Tercile
19
Analog Forecast Skill - Upper Tercile
20
Analog Forecast Skill - Upper Tercile
21
Free parameters (WCoast, 4-6 day upper decile)

• Analog Size
• Analog Search Region (75 analogs)
• Finding analogs for each member 5 analogs per
  member, skill is degraded (BSS 0.183).
• Forecast variable, analog weighting?

of Analogs 25 50 75 100
BSS 0.2050 0.2179 0.2185 0.2168
Grid Points 4 16 36
BSS 0.1967 0.2185 0.2165
22
Conclusions

• Forecast analogs (using ensemble mean) hold great
  promise.
• preserves covariances.
• non-parameteric.
• corrects for regime-dependant errors.
• produces 3-day lead time improvement in PQPF
  skill relative to operational system run at twice
  the resolution.