Some Thoughts on HFIP Bob Gall Thanks! For putting up with me for - - PowerPoint PPT Presentation

Some Thoughts on HFIP

Bob Gall

Thanks!

For putting up with me for the last six years , For helping to create a Hurricane Project that I believe has been very successful

• I will be mostly retiring at the end of the year
• Vijay will be taking over as Development Director

Rather than the usual outline of HFIP technical achievements in the past year—much of which you will hear from the team reports-- I thought I would outline some of my thoughts and comments about HFIP

Reasons for the success of HFIP

• It had significant funding—you can’t make much happen without

sufficient resources

• It brought together a broad community -- research to development

to operational implementation to work collectively on the hurricane problem.

• It facilitated communication within a large group

– Bi-Weekly telecons, annual meeting, workshops

• Community participation is developing project plans—the teams
• It developed facilities

– DTC for code access and testing—provides basic link R2O – A dedicated large computer facility

The community effort allowed a large group of scientists to focus on a single problem

Some general Comments

• Initialization remains HFIP’s biggest problem
• Don’t ignore the global model
• A question about statistical significance
• Don’t ignore Ensembles
• More community focus on developing physics

packages

• Comments on Recent HFIP Performance

Don’t ignore the global model

• You will hear a recommendation from the SRC to cease

any focus on the global model problem

– I am not sure that is good advice – SRC feels we should focus on initialization/physics and the first 1-3 days—focus on the regional model

• But the global model is still a central part of any

regional system—to the HFIP goals

• There is a sense within NOAA that the NGGPS project

will take care of the global model

• But there problems/issues with the global models that

are unique to the hurricane problem

– There needs to be some way to insure that they are appropriately addressed

6

The HFIP Project Vision/Goals

• Vision

– Organize the hurricane community to dramatically improve numerical forecast guidance to NHC in 5-10 years

• Goals

– Reduce numerical forecast errors in track and intensity by 20% in 5 years, 50% in 10 years – Extend forecast guidance to 7 days with skill comparable to 5 days at project inception – Increase probability of predicting rapid intensification at day 1 to 90% and 60% at day 5

6

NCEP vs ECMWF for Atlantic 2006-2008 % gain over HFIP baseline (track)

GFS ECMWF

7

NCEP vs ECMWF for Atlantic 2012 % gain over HFIP baseline (track)

GFS ECMWF

8

The Initialization Problem

• There is no doubt that initialization remains a major

problem for the program

– The problem pretty much eliminates the value of the regional model forecasts in the 0-2 day range (intensity) – Forecasts of RI by a model during these first 2 days have little reliability

• The problem is likely mostly related to initial conditions

that are inconsistent with the model dynamics/physics

• The initialization will likely ultimately be solved

through data assimilation

– But the resultant initial flow will need to be model consistent somehow

• Improved data will help but it isn’t the main problem

Stream 2.0 Skill (AL Intensity)

• Smaller

sample size

• Decay

SHIPS shows highest model skill

• APSI skill

gain is largest through 72 h

Stream 1.5 Skill (AL Intensity)

• Statistical-

dynamical configurati

• ns show

highest skill including SPC3

• Dynamical

models transition from (-) to (+) skill with lead time

• CXTI and

UW4I show lowest skill

Stream 1.5 Skill (EP Intensity)

• CXTI and

HWFI show highest skill

• Statistical-

dynamical configuratio ns generally lose skill with lead time

• HFIP 5-yr

skill goal met intermittent ly

Question about statistical significance

• Recently there has been a lot of emphasis on looking at

the statistical significance of error comparisons

– Such as the impact of some change compared to a control run

• There is no doubt that this is very important in some

settings

• But note that almost all tests of the impact of some

change in the model at NCEP are not statistically significant

– Yet they are used to make decisions on model changes – And the models get better.

Impact of Radar Data

Don’t Ignore Ensembles

• I probably don’t have to say this to ensemble people

– But they seem to want to focus on probabilities

• But most folks want to know when and where NHC thinks the

hurricane is going to hit and how strong it will be when it does— which is a deterministic forecast

• Ensembles give some information like that—ensemble mean

– But we are throwing away a huge amount of information that can be used to improve a deterministic forecast from both multi-model and single ensembles.

• This isn’t a criticism of forecasters

– It is a criticism of the project that hasn’t put enough emphasis on developing simple tools for extracting this information from the ensemble and presenting them to the forecaster

Emphasis on Physics Packages

• The two primary areas where we can improve the

models is Initialization and improved physics packages

• In my opinion we need more focus within the

broad community (outside the operational centers) on developing/testing/improving physics packages

– Particularly the university community – I am not sure why physics gets less emphasis in the research community than say data assimilation (and cores)

Comments on Recent HFIP Performance

Operational Intensity Forecast Trends* and HFIP Goals

*Courtesy NHC: 2013 results are preliminary, subject to revision

HWRF Intensity ATL Basin Cumulative Forecast Improvements

Improving 15-20% per year since 2011

2013 version is approaching 5 year goal

23

Stream 1.5 Skill (AL Intensity)

• Statistical-

dynamical configuratio ns show highest skill including SPC3

• Dynamical

models transition from (-) to (+) skill with lead time

• CXTI and

UW4I show lowest skill most lead times

Stream 2.0 Skill (AL Intensity)

• Smaller

sample size

• Decay

SHIPS shows highest model skill

• APSI skill

gain is largest through 72 h

Stream 1.5 Skill (AL Track)

• GPMI

lowest skill for most lead times, but still higher than HFIP 5-yr skill goal

• HWRF

highest skill among

• peration

al models

Stream 2.0 Skill (AL Track)

• Smaller

sample size

• HWRF and

FIM configurati

• ns show

highest skill most lead times

• HFIP 5-yr

skill goal surpassed for most lead times