can we predict the impact of observations on 3 to 6 day winter weather forecasts l.
Download
Skip this Video
Loading SlideShow in 5 Seconds..
Can we Predict the Impact of Observations on 3 to 6 day Winter Weather Forecasts? PowerPoint Presentation
Download Presentation
Can we Predict the Impact of Observations on 3 to 6 day Winter Weather Forecasts?

Loading in 2 Seconds...

play fullscreen
1 / 74

Can we Predict the Impact of Observations on 3 to 6 day Winter Weather Forecasts? - PowerPoint PPT Presentation


  • 235 Views
  • Uploaded on

Can we Predict the Impact of Observations on 3 to 6 day Winter Weather Forecasts? Masters Thesis Defense May 10, 2007 Kathryn J. Sellwood University of Miami, R.S.M.A.S., M.P.O. Dept. Why winter weather? It’s not all about Hurricanes!

loader
I am the owner, or an agent authorized to act on behalf of the owner, of the copyrighted work described.
capcha
Download Presentation

PowerPoint Slideshow about 'Can we Predict the Impact of Observations on 3 to 6 day Winter Weather Forecasts?' - Ava


An Image/Link below is provided (as is) to download presentation

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.


- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -
Presentation Transcript
can we predict the impact of observations on 3 to 6 day winter weather forecasts

Can we Predict the Impact of Observations on 3 to 6 day Winter Weather Forecasts?

Masters Thesis Defense

May 10, 2007

Kathryn J. Sellwood

University of Miami, R.S.M.A.S., M.P.O. Dept.

observations upstream in storm tracks used to improve forecast in downstream locations
Observations Upstream in Storm Tracks Used to Improve Forecast in Downstream Locations

Storm Tracks for 1994 winter season (left) 6 year average eddy kinetic energy (right) from James, 1994

influence of observations propagates eastward
Influence of Observations Propagates eastward

Poorly Observed

Region

Future Winter Storm

Location

slide6
IF

we can predict the impact of observations

on specific 3-6 day winter weather forecasts

THEN

Observations can be used to improve accuracy and extend effective time range of forecasts

signal from observations at t 0 left and t 80 hours right
“Signal” from Observationsat t=0 (left) and t=80 hours (right)

combined 200 h-Pa u,v and T squared signal at 0 hours (left) and 80 hours (right)

outline
Outline
  • Background:

Targeting

ETKF Technique

Results from Previous Targeting Field Programs

  • Research Methodology
  • Results
  • Conclusions
  • Future Work
operational targeting timeline
Operational Targeting Timeline

Targeting Method Future Analysis Verification

initialization time (targeting) time time

ti ta tv

Decision time

 36-60 hours  1-7 days 

slide10

Targeting in WSR

target regions identified for 2 day

east coast forecast

Observation time

A

Day 1

A

B

All WSR flight paths Operational flight path

Day 2 -Verification time

A

B

ensemble transform kalman filter etkf
Ensemble Transform Kalman Filter(ETKF)
  • Quantifies impact of observations
  • Estimates the forecast error covariance matrix from an ensemble
  • Assimilates observational data using a Kalman filter
  • Computes resulting reduction in forecast error variance

5400m and 5820m 500 h-Pa height ensemble

slide12
Forecast error covariance matrix computed from matrix of ensemble perturbations Z

Pf = Zf ZfT

Kalman Filter equation used to obtain new error

covariance matrix if observations are assimilated

Pq= Pr – Pr HqT (HqPrHqT+Rq)-1 HqPr

Difference is the Signal Covariance Matrix Sq

Sq=Pr–Pq

Signal Variance = reduction in forecast error variance

etkf issues
ETKF Issues
  • ETKF relies on linear theory
  • Depends on ensemble quality
  • Assumes Kalman filter data assimilation / operational scheme is 3D-Var
  • DA scheme introduces small scale noise that contaminates signal (Hodyss and Majumdar, 2007)
results from previous field programs
Results from previous Field Programs
  • Targeting is effective in reducing short term forecast errors (Langland et al, 1999, Szunyogh et al., 1999, 2000, 2002)
  • ETKF effective for short range 1-3 day targeting (Majumdar et al., 2002-a, Szunyogh et al., 2000, 2002)
  • Flow regime related to effectiveness of ETKF
  • Data “signal” propagates in the form of upper tropospheric Rossby wave packets (Szunyogh et al., 2000, 2002)
  • Downstream development maintains wave packets and influences signal propagation (Szunyogh et al., 2002, Majumdar et al., 2002-b)
  • The presence of coherent wave packets beyond operational lead times is evidence of data influence in the medium range
observations over pacific improve 3 day forecast over u s east coast and 6 day european forecast
Observations over Pacific improve 3 day forecast over U.S. east coast and 6 day European forecast

Improvement at 78 hours (u,v,T) 144 hour improvement (u,v,T)

main objectives
Main Objectives
  • Quantify ETKF’s ability to predict signal variance in the medium range
  • Determine scales at which ETKF is effective
  • Explore influence of flow regime
  • Determine whether ETKF can distinguish between promising targeting cases and those where observations would have minimal effect
2 comparison methods
2 Comparison Methods
  • Method 1: measures the spatial correlation between the ETKF predicted signal variance and the squared GFS signal
  • Method 2: makes a quantitative evaluation of ETKFs’ skill in predicting signal variance
the data set
The Data Set
  • Data is from the 2006 Winter Storm Reconnaissance Program
  • 19 individual cases
  • Forecast variables are 200 h-Pa winds and temperature
  • ETKF signal variance derived from a 50 member ECMWF ensemble
  • Signals calculated as the difference between 2 forecasts that are identical except 1 omits the WSR observations
  • Forecast produced using NCEPs Global Forecast System Model (GFS)
  • Both fields at 1 degree resolution
methodology
Methodology
  • Spatial fields of ETKF “predicted” signal variance and GFS squared-signal (“verification”) are smoothed by averaging over lat-lon grid cells.
  • Field domain includes 180° W to 20° E from 20 to 80°N
  • Correlation coefficients between these smoothed spatial fields are calculated, at all lead times, for various grid spacings.
  • Correlation coefficients for actual case-specific predictions are then assessed relative to a “no-skill” baseline, constructed by randomizing the predictions of all 19 weather cases in the sample.
the randomized baseline
The Randomized Baseline
  • Baseline random correlations are computed for all lead times from 0 to 144 hours
  • The ETKF predicted signal variance for each of the 19 cases is compared to the squared GFS signals from the 18 different cases
  • Results in a distribution of 342 random correlations
  • This baseline captures the (non) skill of case-independent spatial structure (like climatological storm tracks)
  • ETKF’s skill for the individual cases is compared against this baseline
case specific vs random correlations
Case Specific vs. Random Correlations

Same Case

Similar Pattern

Randomly Selected Case

Less Correlated

correlation skill of case specific vs randomized predictions 2 grid
Correlation skill of case-specific vs. randomized predictions (2° grid)

Blue = Random Red = ETKF

slide29
Average ETKF Correlations (solid) and Random Correlations (dashed) as a Function of lead time and resolution
significance test
Significance test
  • Are the ETKF correlation coefficientssignificantlygreater than the random distribution?
  • Use Kalmogorov-Smirnov test for the difference of two PDFs.
slide32

13/20 data points

= Maximum difference between

distributions

4/ 16 data points

the kolmogorov smirnov test
The Kolmogorov – Smirnov test
  • Compares cumulative distribution function (CDF)
  • Produces 2 statistics based on “D” value
  • H statistic tests the null hypothesis that the 2 distributions are equal

H = 0 cannot reject null hypothesis

H = 1 can reject with 95% confidence

  • P statistic gives probability that the 2 distributions are indistinguishable
  • Test applied for all lead times and resolutions
5 grid
5° grid

H = 1 for all

lead times > 0

10 grid36
10° grid

H = 1 for all

lead times > 0

15 grid37
15° grid

H = 1 for all

lead times > 0

20 grid38
20° grid

H =1 for all

lead times > 0

30 grid39
30° grid

H = 1 for all

lead times > 0

etkf significantly beats random for all grid spacing and lead times 0
ETKF significantly beats random for all grid spacing and lead times > 0
  • At 0 hour leads ETKF predictions are not significantly better than random climatology
  • ETKF case-specific predictions exhibit significantly better than random skill for the time ranges (3-6) days that we are interested in
  • Skill tends to improve (relative to random) at longer lead times
  • Higher correlations at lower resolution (larger grid)
slide41
ETKF has been shown to have skill in predicting the general pattern of signal variance over a large domain but…
  • We want to apply ETKF to specific forecasts
  • Can the ETKF predict signal variance specifically in predetermined verification regions at 3-6 day lead times?
  • If so at what resolutions and for what size verification regions
verification regions
Verification Regions
  • Same methodology as full domain comparison
  • ETKF predicted signal variance compared to squared GFS signal over 20° x 20°, 40°x 40° and 60°x 60° verification regions
  • Verification regions selected using wave packet technique of Zimin et al., 2003
  • Verification regions placed at the leading edge of wave packet maximum in ETKF predicted signal variance
typical 120 hour v r s
Typical 120 hour V.R.s

60 X 60 40 x 40 20 x 20

influence of verification region size and resolution on etkf skill average correlation coefficients
Influence of Verification region size and resolution on ETKF skill (average correlation coefficients)

Verification Region Resolution

etkf significantly beats random for case specific verification regions
ETKFsignificantlybeats randomfor case specific verification regions
  • ETKF skill is significant even at 0 hour leads for 40° and 60° verification regions
  • Better correlations for 40° and 60° V.R.s
  • Higher correlations for lower resolution (larger grid)
  • Better skill relative to random at short leads for 20° V.R. and long leads for 40°

and 60° V.R.s

  • Skill on synoptic scales ~ 500-1000km
skill varies from case to case
Skill varies from case to case

Why does the ETKF perform so much better (or worse) in particular cases?

What makes a good or bad targeting case?

individual cases
Individual Cases
  • 9 best and 6 worst cases are identified
  • Flow regime is quantified by 3 indices
  • Flow regime evaluated over 3 regional areas

How Does Background Flow Regime

Effect The ETKF’s Performance?

flow regime regions
Flow Regime Regions

Continental

Atlantic

Pacific

flow regime indices
Flow Regime Indices
  • Zonal Index (Namias, 1947)

Ug at 700 h-Pa level between 35° and 55° N

  • Blocking Index (Tibaldi and Moltini, 1990)

/y at 500 h-Pa between 40° and 60° N

  • Ug Anomaly (Horel, 1985)

Ug – Ug(climo) at 55°N 165°W

  • PNA Index (Hanson et al., 1993)
  • NAO Index (Liu et al., 1995, Benedict et al., 2003)
low frequency non zonal patterns
Low Frequency Non-Zonal Patterns

PNA(positive phase) NAO(negative phase)

evaluation of flow indices best cases
Evaluation of Flow Indices:Best Cases

Zonal Index & Blocking Index:

Zonal Flow Regime all cases

Ug Anomaly:

1 case local blocking all others zonal

PNA:

negative 5 cases positive 4 cases

NAO:

positive 5 cases negative 4 cases

evaluation of zonal indices 6 worst cases
Evaluation of Zonal Indices6 worst cases

Zonal Index & Blocking Index:

4 cases zonal flow 2 cases blocked flow

Ug Anomaly:

local blocking in all cases

PNA:

negative all cases - zonal

NAO:

negative 5 cases blocked

positive 1 case - zonal

influence of flow regime
Influence of Flow Regime
  • Best cases characterized by zonal flow especially in Pacific Observation region
  • PNA and NAO likely to be a factor only when associated with blocking
  • Worst cases characterized by blocked flow in the Pacific
  • 2 Worst cases are during Atlantic blocking
  • ETKF predicts poorly for blocked regimes but does not seem to be negatively affected by waves (meridional flow)
can etkf distinguish between good and bad cases
Can ETKF Distinguish Between Good and Bad Cases?
  • ETKF initialized for 3-6 day Atlantic forecasts
  • ETKF target regions compared to operational flight paths
  • 9 best, 6 worst cases evaluated

Were Flight Paths Deemed Sensitive to

Observations for these Forecasts?

did etkf predict sensitivity in flight regions for 3 6 day lead times
Did ETKF predict sensitivity in flight regions for 3-6 day lead times?
  • Sensitive regions (shaded areas) for 4 day forecast in East Atlantic and WSR flight tracks for 2 of the best cases
similar results for all good bad cases
Similar Results for all good/bad cases
  • The best cases have observational sensitivity for 3-6 day forecasts in the WSR flight region
  • In the worst cases the WSR flight path misses the target region for the longer term forecast
  • Results suggest ETKF is capable of identifying sensitive regions for targeted observations on these time scales
  • ETKF is apparently able to distinguish good and bad targeting cases
conclusions
Conclusions
  • ETKF is effective out to 6 days
  • ETKF has skill in case-specific verification regions
  • ETKF is effective on synoptic scales
  • Best prediction at low resolution
  • Optimal verification region size is 40°
  • ETKF can identify good and bad cases
  • Blocked flow regime degrades ETKF skill

THE ETKF TARGETING METHOD CAN BE

EXTENDED IN MANY INSTANCES TO

MEDIUM RANGE FORECASTING

future related work
Future / Related Work
  • Determine if the ETKF predicted signal variance is associated with a reduction in forecast error variance in the medium range
  • Explore the energy dynamics associated with signal propagation
  • Take a closer look at how downstream development influences targeting
i would like to express my great appreciation to
I would like to express my great appreciation to …

my committee members Brian Mapes, Istvan Szunyogh

and especially my advisor, Sharan Majumdar

the many experienced people who provided valuable

advice and information including Dan Hodyss, Yucheng

Song, Fuqing Zhang and many others

my family (even though they still can’t figure out my schedule)

my RSMAS friends and colleagues

the dolphins at the Seaquarium

Thank you all !