1. Ensemble hindcasting of “Superstorm ‘93”
Robert Fovell
University of California, Los Angeles
Peter Dailey
AIR Worldwide
14th Cyclone Workshop
Ste.-Adele, Quebec, Canada

2. Motivation Superstorm ‘93: example of an extreme winter storm
APE far exceeded climatological mean (Bosart et al. 2000)
Risk assessment for insurance industry
Need to assess recurrence interval
Leveraging ~50 year record of cases (reanalysis) into 10,000 year catalogue
Ensemble hindcasting of historical cases
How much more intense could this storm have been?

3. Animation of SLP field (ECMWF reanalysis)
03/14/93 06Z
Storm Track
ecmwf_movie.GIF
03/13/93 06Z

4. Kocin-Uccellini Snow Footprint
NESIS = 12.52
NESIS
soc
Top 30 Events
NESIS = 12.5
NESIS (Kocin and Uccellini 2004):
Northeast Snow Impact Scale
• area-integrated
• population-weighted
Superstorm ‘93: #1 for ~100 year
period

5. Hypothesized distribution all East Coast winter storms
Need not be
Gaussian

6. Superstorm 93’s position among all EC winter storms

7. Superstorm 93’s own distribution
Storm itself could have been weaker or stronger
Even at weakest, likely still very significant event
How much more intense could it have been?
Implications for recurrence interval
Drew distrib w/ expectation that: (1) it’s normal; (2) event was in middle of its distribution

8. Ensemble design MM5 model ECMWF initialization
Perturb initial conditions
> 400 members
Purpose: retrieve event’s own frequency distribution
Overarching goal: not to bias the outcome
MM5 Model Domain Configuration
INNER DOMAIN (30km)
OUTER DOMAIN (90km)

9. Perturbation strategy
Goal: be unbiased
Random perturbations to T, Td
Columns perturbed to preserve
Lapse rate
Relative humidity
Objective analysis with large influence radius

10. Sample 300 mb T perturbation fields

11. All stats limited to common event period
Lead time strategy
Six hourly starting times
ranging from 0 to 90 hour leads
All stats limited to common event period

12. Storm’s own distribution
Control/seed/
benchmark case
Control/seed/
benchmark case
Default scenario
Perturbations as likely to weaken and strengthen storm
All quite intense
What if storm already at optimal intensity?
Most perturbations should weaken storm
Possible to retrieve this from a control run in the tail?
If only top distribution is possible, not that interesting. Width may be important, but shape is a given

13. Storm’s own distribution
Control/seed/
benchmark case
Control/seed/
benchmark case
Default scenario
Perturbations as likely to weaken and strengthen storm
All quite intense
What if storm already at optimal intensity?
Most perturbations should weaken storm
Possible to retrieve this from a control run in the tail?
If only top distribution is possible, not that interesting. Width may be important, but shape is a given

14. Results from 465 member ensemble

15. Kocin & Uccellini’s estimate: 12.52
NESIS - control runs
perfect model & data
Values range from
Kocin & Uccellini’s estimate: 12.52
NCDC’s estimate: 13.2
Both use
2000 census

16. (weakest would rank 10th among 20th century events)
NESIS - perturbed runs
Max 14.1 Average Min 6.1
(weakest would rank 10th among 20th century events)

17. Large and small snowfalls
12 h lead case h lead case
NESIS = NESIS = 7.7
Rank 1/ Rank 450/455

18. NESIS distribution for 465 cases
Which case best represents the actual event?

19. Selecting the benchmark case

20. RMS SLP error metric (with respect to ECMWF reanalysis)
Independent of snow information
More accurate, objective than snow metrics
Minimum error case is most similar to reanalysis with respect to location, timing and intensity
ECMWF analysis coarseness underrepresents actual cyclone intensity

21. SLP error 0 h lead control run: nominal benchmark
NESIS = 12.5
SLP min case: actual benchmark
NESIS = 13.7

22. Where benchmarks ranked (465 cases)
NESIS
Actual benchmark: ranked 3rd of 465
Nominal benchmark: ranked 33rd of 465 (exceeded by 7%)
Total snow over land
Actual benchmark: ranked 11th
Nominal benchmark: ranked 35th
Other metrics…

23. (weakest would rank 10th among 20th century events)
NESIS revisited
Max 14.1 Average Min 6.1
(weakest would rank 10th among 20th century events)

24. 0-24 h lead subset
310 cases
Little temporal trend - variance within lead >> among lead
NESIS ranks: actual benchmark 3/310; nominal 33/310

25. 0-24 h lead subset

26. Conclusion: Actual storm nearly as strong
0-24 h lead subset
Conclusion: Actual storm nearly as strong
as it could have been

27. 0-24 h lead subset NESIS control-relative
76% of perturbed cases
weaker than their control/seed

28. 0-24 h lead subset Total snow control-relative
Again, 76% of perturbed cases
weaker than their control/seed

29. 0-24 h lead subset Vorticity control-relative
56% of perturbed cases
stronger than their control/seed

30. Another case: December 1992 20th century rank: #42

31. 10-12 December 1992 Ranked 42nd among 20th century snowstorms
NESIS value influenced
by isolated snowbands

32. NESIS distribution for Dec 92 event

33. Future work Reconsider perturbation strategy, benchmark selection
Ensemble diversity sufficient?
Use ensemble output to identify best place(s) to perturb
More cases
Build a statistical model to create the yr catalogue
More from less…

34. end