Probabilistic Lightning Forecasts Using Deterministic Data

1
Probabilistic Lightning Forecasts Using
Deterministic Data

Evan Kuchera and Scott Rentschler 16 Aug 2007
2
Motivation

• Air Force operators require skillful and
  objective probabilistic weather information to
  maximize efficiency and minimize loss
• Typically this is accomplished with ensembles for
  grid scale phenomena
• However, sub grid scale processes are
  probabilistic in nature even with deterministic
  data
• We believe that ensemble forecast skill will be
  higher if a probabilistic approach is taken with
  each ensemble member for sub grid scale phenomena
• Addresses both sub-grid scale and flow
  uncertainties

3
Motivation

• Examplelightning forecast with SPC SREF method
• 10 ensemble members
• CAPE values of 130,125,120,115,110,105,103,102,101
  ,101
• With a forecast threshold of 100 J/kg, this gives
  a 100 chance of lightning
• However, with values so close to the threshold,
  the true probability is likely much closer to 50
  than 100
• This can be accounted for somewhat with real-time
  calibration after the ensemble is created (as SPC
  does with success), but this is not necessarily
  an option for the Air Force (resource
  constraints, lack of calibration data)

4
Background

• Lightning background
• Need graupel and ice particle collisions to
  transfer negative charge to the larger particles
• Thunderstorm updrafts need to grow large graupel
  particles with enough fall speed to cause a
  separation of charge in the vertical
• The theoretical value of CAPE required to do this
  is only 25 J/kg

5
Background

• CAPE background
• Accepted parcel theory assumption is that as the
  parcel rises, all condensate is immediately
  removed, and that there is no latent heat of
  freezing
• However, lightning is caused by frozen
  condensates in an updraft!
• We decided to test CAPE both waysthe traditional
  way, and with condensates/latent heat of freezing

6
Image from NASA-GHCCWorldwide lightning
climatology
7
Traditional Lifted Index
8
TEST Lifted Index
9
Methodology

• Goal create a probabilistic lightning algorithm
  using a large set of CONUS observations and
  physical assumptions relevant worldwide
• 2006 3-hourly 20 km RUC analyses
• NLDN lightning in the RUC grid box (0-3 hr after
  analysis)
• 3 hour precipitation from METARS
• Find which forecast parameters are the best, then
  curve fit the probability of lightning given a
  binned value of that parameter

10
2006 Results
11
NLDN 3-hourly lightning climatology for a 16 km
grid box (2003-2006)
12
Results
GL CAPE is calculated from the LFC to -20C Set to
zero if equilibrium level is warmer than
-20C TEST is condensate and latent heat of
freezing included
13
CAPE 0, Precipitation 0.01
14
CAPE0, Precipitation 0.01
15
CAPE 0, Precipitation0
16
Results
Climatology0.155
17
Results
Perfect Reliability
18
Results
No Skill Forecast
19
Results
SPC method forecast 100 chance of lightning if
GL CAPE is greater than 100 J/kg and
precipitation is greater than 0.01 inches.
Forecast 0 otherwise. NULL method Always
forecast 0 chance of lightning. TEST method
Algorithm presented here. BSS Brier skill
score, compares mean squared error of forecast to
mean squared error of climatology. 1 is perfect,
0 is no skill, negative is worse than
climatology. ROC area Total integrated area
underneath ROC curve. 1 is perfect, 0.5 is no
skill.
20
Summary

• Algorithm has been developed to forecast
  lightning probability given observed instability
  (RUC analysis) and precipitation (METARS)
• Algorithm is somewhat sharp, reliable at all
  forecast probabilities, and has good resolution
  of events and non-events
• Buoyancy calculations probably need to account
  for condensate and latent heat of freezingbut
  our data are not conclusive on this point

21
Other/Future Work

• Equations have been developed (not shown here) to
  forecast strikes per unit area for application to
  any model resolution
• After knowing strikes per unit area, can forecast
  probabilities for smaller areas (i.e. Air Force
  base warning criteria area) based on downscaling
  climatologyequation has been developed for this
  purpose as well
• Just beginning to look at algorithm with model
  data and in ensemblesissues with model
  precipitation forecasts
• Acknowledgments ARM data archive, Dr. Tony
  Eckel, Stephen Augustyn, Bill Roeder, Dr. David
  Bright, Jeff Cunningham

22
Questions?
GFS 66 hour grid point lightning probability
forecast valid this afternoon
23
Backup Slides
24
Backup Slides

• Adjustments for changes in model resolution or
  area of interest
• First, re-calculate total number of strikes for
  the new model grid box area
• If model grid is finer than RUC, re-calculate
  probabilities using inverse of strikes equation
• If model grid is coarser than RUC, increase
  probabilities using special upscaling equation
• If area of interest is smaller than area of model
  grid, recalculate strikes and use downscaling
  equation to get probabilities

25
Backup Slides

• Downscaling equation details
• Inputs
• Strikes (S)
• horizontal resolution of coarse area in km (C)
• horizontal resolution of fine area in km (F)
• Equation 1-1-(F2/C2)(SA)
• Where A is a fudge factor depending on F
• A1-0.17LN(F-1)
• A equals unity when F is 2 km, and slowly
  decreases toward zero as F approaches 350 km
• In nature, lightning tends to be randomly
  distributed at 2 km (storm scale) but more
  clustered at higher resolutions. A attempts to
  account for this
• Best to use this equation from 2 to 128 km grid
  sizes
• If strikes is less than one, calculate equation
  using 1 strike, then multiply result times number
  of strikes

26
Backup Slides

• Upscaling
• Probability added to
• 1-probability1-(F2/C2)downscaled
  probability
• This ensures high probabilities will only occur
  when the original probability was high, or the
  area has increased substantially with moderately
  high initial probabilities
• No testing as to whether this is calibrated

27
NWS Topeka forecast taken from the web on 15
Aug Friday, August 17 at 7pmTemperature
89FThunder Backup Slides
28
Backup Slides
29
Backup Slides