Developing Gridded Forecast Guidance for Warm Season Lightning over Florida Using the Perfect Prognosis Method and Mesoscale Model Output
Phillip E. Shafer Henry E. Fuelberg Florida State University April 4, 2007
Dept. of Meteorology
Accurate Lightning Forecasts Are Important to FPL
Lightning leads to outages FPL crews should be ready to respond Don’t want un-needed crews
Dept. of Meteorology
Project Timeline
• Phase I: 1 August 2002 – 31 July 2003 • Phase II: 1 June 2003 – 31 May 2004 • Phase III: 1 June 2004 – 31 May 2005 • Phase IV: 1 June 2005 – present
Dept. of Meteorology
Phase I—Develop Lightning Climatologies
Southeast Flow Flow Southwest
Flashes/km^2/Regime Day
Dept. of Meteorology
Detailed Climatologies For Dispatch Centers
Dept. of Meteorology
Detailed Climatologies For Dispatch Centers
Dept. of Meteorology
Phases II & III
• Equations derived for 11 FP&L service areas. • Morning radiosonde parameters used as predictors for afternoon lightning in each area. • Miami, Tampa, Jacksonville, Cape Canaveral • Generally, the sounding closest to each forecast area was used.
Dept. of Meteorology
Phase 3 Sample Forecast
Dept. of Meteorology
Forecasts Useful to FP&L Dispatchers:
―I can state, unequivocally, that SSD uses the forecasts daily as an integral part of the resource and switching decision making process.‖ ―On many occasions, we may have not held resources based on the projected weather levels only but the lightning forecasts were solid enough to override that decision- to our advantage I might add.‖
Dept. of Meteorology
Phase IV—Space and Time Varying Guidance for all of Florida
Dept. of Meteorology
Presentation Outline
1. 2. 3. 4. 5. 6. 7. 8. Motivation and Objectives Background Data Model Development Model Parameters Results for Dependent Data Results for Independent Test Summary & Conclusions
Dept. of Meteorology
1. Motivation and Objectives
Dept. of Meteorology
Motivation
• Lightning is one of the leading causes of weather related fatalities in the U.S. • Lightning can cause damage to trees and utility lines, leading to disruptions in power and communications. • Florida is the lightning capital of the U.S. • Many heavily populated areas are vulnerable. • Skillful probabilistic guidance in the 3-12 h time frame would have many potential societal benefits.
Flashes km-2 warm season-1 1989-2006 (May-September)
Dept. of Meteorology
Objectives
1. Use the perfect prognosis (PP) method to develop a highresolution gridded forecast guidance product for warm season cloud-to-ground (CG) lightning for all of Florida: -- Equations to produce spatial probability forecasts for one or more CG flashes, and the probability of exceeding various flash count percentile thresholds. -- 10 x 10 km grid, 3-h intervals 2. Evaluate the utility and skill of the PP scheme when applied to forecast output from several mesoscale models during an independent test period (2006 warm season).
Dept. of Meteorology
2. Background
Dept. of Meteorology
Myriad Factors
• • Sea breeze usually the dominant forcing mechanism over Florida during the warm season. Interactions between the sea breeze, the prevailing wind, and coastline curvature have been shown to influence lightning patterns (e.g., Lopez and Holle 1987; Hodanish et al. 1997; Camp et al. 1998; Lericos et al. 2002). Other myriad factors influence timing and location of convection and lightning: -- Local thermal circulations (e.g., water conservation areas, lakes, rivers, etc.) -- Urban effects (e.g., Westcott 1995; Steiger et al. 2002)
•
-- Thunderstorm outflows
Dept. of Meteorology
Cloud Microphysics
• • • Lightning ultimately is governed by cloud microphysical processes that are poorly resolved by NWP models. Factors influencing cloud electrification are poorly understood. Several hypotheses have been proposed: -- Precipitation hypothesis (Reynolds et al. 1957) -- Convection hypothesis (Vonnegut 1963) -- Non-inductive ice-ice collision mechanism (Williams 1985) • • Hypotheses depend on a vigorous updraft and robust ice phase for charge generation (Price and Rind 1992, 1993). But, advances have been made in our understanding of the factors influencing lightning production.
Dept. of Meteorology
Statistical Studies
• A variety of statistical techniques has been used to develop forecast models for thunderstorms and lightning: -- Multiple linear regression often employed in earlier studies (less computationally demanding). -- Binary logistic regression more appropriate when predictand is ―yes‖ or ―no‖ (e.g., Mazany et al. 2002; Bothwell 2002; Lambert et al. 2005; Shafer and Fuelberg 2006). -- Classification and regression trees (e.g., Burrows et al. 2004). • Many studies have used data from morning soundings to forecast afternoon lightning.
•
Data from NWP models is more location and time specific.
Dept. of Meteorology
Model Output Statistics
• Model Output Statistics (MOS): Objective forecasting technique in which statistical relationships are determined between a predictand and variables forecast by an NWP model. Advantage: Model biases and local climatology are automatically built into the equations. Usually the method of choice when practical. Drawback: NWP models are constantly changing. Any modifications to the NWP model that change systematic model errors require redevelopment of the MOS equations.
•
•
Dept. of Meteorology
Perfect Prognosis
• Perfect prognosis (PP): Statistical relationships are determined between observations of the predictand and observed atmospheric predictors.
•
Advantages: -- Equations are developed without NWP forecasts (i.e., they are model independent). -- Equations can be used with any NWP model and forecast projection, even as the models change.
Drawback: Assumes a ―perfect‖ forecast of the predictors by the NWP model and thus, does not account for model biases. Bothwell (2002): Used PP method to develop forecast equations for CG lightning over the western U.S.
• •
Dept. of Meteorology
3. Data
Dept. of Meteorology
Study Domain
Grid spacing = 10 km Only land grid points used in model development
Dept. of Meteorology
Lightning Data
• • • • The dependent variable National Lightning Detection Network System wide upgrades in 1995 & 2002 1995-2005 warm seasons used to develop climatological predictors. • 2002-2005 warm seasons used in equation development. • Data quality controlled for duplicate flashes and non-CG discharges. • Flashes summed within a 10-km radius of each grid point during each 3-h period (e.g., 0000-0259 UTC, …, 2100-2359 UTC).
Dept. of Meteorology
Lightning Predictands
• • 3-h flash totals transformed into binary variables. ―1‖ if one or more flashes or ―0‖ if no lightning
•
Binary variables also assigned based on whether the flash total exceeds the 50th, 75th, 90th, and 95th percentiles for a given 3-h period:
Dept. of Meteorology
Rapid Update Cycle (RUC)
• • ―Observed‖ atmospheric predictors derived from RUC analyses during 2002-2005 warm seasons (May-Sept). RUC data sources: 1. Atmospheric Radiation Measurement (ARM) Program (http://www.arm.gov/xds/static/ruc.stm) 2. National Climatic Data Center (http://nomads.ncdc.noaa.gov) 20-km, 50 level, hourly version (RUC20) implemented at NCEP during April 2002, with improvements in the analysis/physics. 13-km version (RUC13) implemented at NCEP on 28 June 2005 with further improvements in the analysis/physics. ~ 1.2 TB of RUC grib data was acquired and processed!
• • •
Dept. of Meteorology
Model Analyzed Predictors
• • Plethora of RUC-analyzed predictors investigated for possible inclusion in candidate predictor pool. Parameters found useful in previous studies were examined: -- Temperature (layer thickness, temperature advection, cold cloud thickness, etc.) -- Moisture (moisture flux convergence, theta-e advection, PW, layer mean RH, etc.) -- Stability (most unstable CAPE in various layers, CIN, best lifted index, Showalter Stability Index, TT, KI, temperature and theta-e lapse rates, etc.) -- Wind (wind divergence, vorticity, vorticity advection, layer mean U and V components, layer mean speed, layer shear).
Dept. of Meteorology
Model Analyzed Predictors
• All parameters calculated from the RUC 0-h temperature, dew point, wind, height, and surface pressure fields valid every three hours (e.g., 0000 UTC, 0300 UTC, …, 2100 UTC).
•
• •
Fields interpolated to array of 10 km grid points and transformed into a vertical sounding.
RUC cloud hydrometeor profiles found to be unusable. Assumption: The model analyses give the best estimate of the state of the atmosphere at the analysis time, and thus, can be treated as ―observations‖ for purposes of developing the PP equations. We focused mainly on parameters that are well handled by today’s NWP models.
•
Dept. of Meteorology
Statistics Software
• S-PLUS version 6.1 for Windows.
•
•
Statistical Package for the Social Sciences (SPSS) version 11.5 for Windows.
Both are state-of-the-art software packages with a wide range of analysis and modeling capabilities.
Dept. of Meteorology
4. Model Development
Dept. of Meteorology
Map Type Predictors
• • Pattern type lightning frequencies developed and used as candidate predictors. Capture local enhancements due to interactions between the low-level wind, thermal circulations, and coastline topography, which are not well resolved by NWP models. 3-hourly observed sea-level pressure fields used for pattern classification- implies direction and speed of low-level flow. SLP fields obtained from RUC analyses spanning the 19982005 warm seasons (~1224 days). Simple correlation technique used to develop the map types (e.g., Lund 1963, Reap 1994).
• • •
Dept. of Meteorology
Map Type Predictors
SLP fields interpolated to array of 100 km grid points.
Dept. of Meteorology
Map Type Predictors
• 5 map types developed using correlation threshold of 0.70. • 2 dominant types (A and B) comprise ~44% of the sample. • ~22% unclassified at a threshold of 0.70. • • Relative lightning frequencies and the unconditional mean number of flashes calculated for each map type and 3-h period. When developing equations, all unclassified maps were assigned the type with which they were most correlated.
Dept. of Meteorology
Map Type Predictors
Type A composite Mean no. flashes: 1800-2059 UTC
• High northeast of Florida • Prevailing E-SE flow
Most lightning confined to West Coast and east of Lake Okeechobee.
Dept. of Meteorology
Map Type Predictors
Type B composite Mean no. flashes: 1800-2059 UTC
• Ridge over South Florida • SW flow across the state
Lightning focused along East Coast and Big Bend region.
Dept. of Meteorology
Map Type Predictors
Type C composite Mean no. flashes: 1800-2059 UTC
• Transition between A and B • SE flow over South Florida, S-SW flow across the north.
Lightning maxima evident along both coasts.
Dept. of Meteorology
Map Type Predictors
Type D composite Mean no. flashes: 1800-2059 UTC
• High north of Florida, lower pressure to the SE. • Most common after cold frontal passage.
Dry NE flow confines most lightning to South Florida.
Dept. of Meteorology
Map Type Predictors
Type E composite Mean no. flashes: 1800-2059 UTC
• Variation of type B- lobe of high pressure over Gulf. • W-NW flow across the state.
Lightning confined to East Coast and Big Bend, with less coverage than type B.
Dept. of Meteorology
Generalized Linear Models (GLMs)
• MLR assumptions of constant variance and Gaussian residuals rarely are met with count data- can lead to undesirable and sometimes nonsensical results.
•
We considered several regression methods: -- Forecasting one or more flashes: Binary Logistic Regression -- Forecasting the amount of lightning: Poisson and Negative Binomial Regression GLMs can be used for response variables that follow any probability distribution in the exponential family (e.g., Normal, Binomial, Poisson, Negative Binomial, etc.).
GLMs accommodate non-Gaussian distributions of residuals and non-constant variance.
•
•
Dept. of Meteorology
Binary Logistic Regression
pi b0 b1 x1 ... bK xK ln 1 p i
exp( b0 b1 x1 ... bK x K ) pi 1 exp( b0 b1 x1 ... bK x K )
• •
• •
Most appropriate when predictand is ―yes‖ or ―no‖ Log link function relates odds ratio to linear combination of predictors. Probabilities bounded on the interval [0,1] Accommodates Bernoulli distribution of residuals.
Dept. of Meteorology
Poisson Regression
ln( [ xi ]) b0 b1 x1 ... bK x K
[ xi ] exp( b0 b1 x1 ... bK x K )
exp( ) y Pr( y | ) y!
• • • More appropriate model for count data Log link function linearizes the expected value () of the dependent variable (y) Poisson probability model assumes that events occur randomly and at a constant average rate () with Var(y) = , where is a dispersion parameter. Poisson model assumes = 1
Dept. of Meteorology
•
Count Distribution
• • Strongly skewed distribution Many cases with 10 or fewer flashes, few with 100 or more. Very large variance (~80 times greater than the mean) Data significantly overdispersed with respect to Poisson model ( >> 1) Likely cause: Counts generated by an inhomogeneous Poisson process- count rates vary in space and time.
•
•
•
1800-2059 UTC period Cases with one or more flashes
Dept. of Meteorology
Negative Binomial Regression
ln( [ xi ]) b0 b1 x1 ... bK x K
[ xi ] exp( b0 b1 x1 ... bK x K )
( y ) Pr( y | [ xi ], ) [ x ] y!( ) i
• • •
[ xi ] [ x ] i
y
Alternative probability model with shape parameter Var(y) is a quadratic function of : Var(yi | [xi] ) = ( [xi] + -1[xi] 2 ) More accurately characterizes the uncertainty in the predicted count than does the Poisson model.
Dept. of Meteorology
Poisson vs. Negative Binomial
Flash Count Probability Distribution Implied from Poisson and Negative Binomial Models 1800-2059 UTC period
0.26
Neg Bin (Null Model)
0.24 0.22 0.20 0.18
Poisson (Null Model) Observed Frequency Mean = 23.15 Theta = 0.3420
Probability
0.16 0.14 0.12 0.10 0.08 0.06 0.04 0.02 0.00 0 10 20 30 40 50 60
(Flash Count - 1)
• •
Poisson model poorly represents count distribution Negative binomial captures large number of cases with 10 or fewer flashes.
Dept. of Meteorology
Negative Binomial Regression
• • • • • NB distribution has been used in previous studies to model thunderstorms at KSC. No known study has used the NB as the probability model for lightning counts. Since the count distribution is left-truncated at y=1, we can treat y-1 as having a NB distribution. Probabilities for each y-1 must sum to 1. Probability of exceeding any count threshold T :
T 1 Pr( y T ) Pr( y) 1 Pr( y) y T y 1
Dept. of Meteorology
Regionalized Approach
• Domain first divided into 9 areas- separate models developed for each. Best results achieved by consolidating 9 areas into 4 larger regions: -- East Coast -- West Coast -- Panhandle -- Alabama & Georgia Regions overlap to minimize problems at regional boundaries.
•
•
Dept. of Meteorology
Final Candidate Predictors
• • Long list of candidate predictors contains redundant information. Principal component analysis used to select subset of predictors with less mutual correlation. Correlations with lightning predictands are low- no single observed predictor is good indicator of lightning. Power terms and cross products (interactions) also calculated and included in final predictor pool. 3-h change in each parameter also included (trend indicators). Map type predictors Climatological predictors
•
•
• • •
Correlations for East Coast region 1800-2059 UTC period
Dept. of Meteorology
Equation Development
• • • • Combination of forward stepwise selection and cross-validation used to develop BLR and NB models. Even years (2002 and 2004) used as ―learning‖ sample. Odd years (2003 and 2005) used as ―evaluation‖ sample. Procedure identifies best combination of predictors that is most likely to generalize to independent data, and not over-fit the dependent sample. Models containing only climatology and persistence (L-CLIPER) also developed as a benchmark for assessing forecast skill.
•
Dept. of Meteorology
5. Model Parameters
Dept. of Meteorology
Model Parameters
Most important parameters for 1800-2059 UTC
BLR models for one or more CG flashes
NB models for the amount of lightning
Dept. of Meteorology
Model Parameters
Most important parameters for 1800-2059 UTC
BLR models for one or more CG flashes
NB models for the amount of lightning
Moisture
Dept. of Meteorology
Moisture
Frequency of One or More Flashes vs. Precipitable Water 1800-2059 UTC period
0.40
10
Unconditional Mean Number of Flashes vs. K-index 1800-2059 UTC period
Frequency of One or More Flashes
0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.00 1.00
Unconditional Mean No. of Flashes
9 8 7 6 5 4 3 2 1 0 -10.0
2.00
3.00
4.00
5.00
6.00
7.00
0.0
10.0
20.0
30.0
40.0
Precipitable Water (cm)
K-index (deg C)
• • •
Deep layer moisture most favorable large-scale environment for Florida thunderstorms. Peak in FREQ1 for PRECPW ~ 5.5 cm. Largest PRECPW usually associated with tropical influences.
Dept. of Meteorology
Model Parameters
Most important parameters for 1800-2059 UTC
BLR models for one or more CG flashes
NB models for the amount of lightning
Instability
Dept. of Meteorology
Instability
Frequency of One or More Flashes vs. Best Lifted Index 1800-2059 UTC period
0.35
Unconditional Mean Number of Flashes vs. Best Lifted Index 1800-2059 UTC period
10
0.30 0.25
Unconditional Mean No. of Flashes
Frequency of One or More Flashes
9 8 7 6 5 4 3 2 1 0 -12.0
0.20 0.15
0.10 0.05
0.00 -12.0
-10.0
-8.0
-6.0
-4.0
-2.0
0.0
2.0
4.0
6.0
-10.0
-8.0
-6.0
-4.0
-2.0
0.0
2.0
4.0
6.0
Best Lifted Index (deg C)
Best Lifted Index (deg C)
• • •
Likelihood of one or more flashes and the amount of lightning increases with decreasing BESTLI. Sufficient instability leading to a strong updraft is necessary for charge generation. Layer CAPE and SSI selected for other periods.
Dept. of Meteorology
Model Parameters
Most important parameters for 1800-2059 UTC
BLR models for one or more CG flashes
NB models for the amount of lightning
Near surface (1000 hPa) forcing
Dept. of Meteorology
Near Surface Forcing
0.35
Frequency of One or More Flashes vs. 1000 hPa Moisture Flux Convergence 1800-2059 UTC period
9
Unconditional Mean Number of Flashes vs. 1000 hPa Moisture Flux Convergence 1800-2059 UTC period
Frequency of One or More Flashes
Unconditional Mean No. of Flashes
8 7 6 5 4 3 2 -10.00
0.30
0.25
0.20
0.15
0.10 -10.00
-5.00
0.00
5.00
10.00
15.00
-5.00
0.00
5.00
10.00
15.00
1000 hPa Moisture Flux Convergence (x 10 -7 kg kg-1s-1)
1000 hPa Moisture Flux Convergence (x 10 -7 kg kg-1s-1)
• •
Low-level moisture flux convergence due to the sea breeze, lake/river breezes, outflows, etc. Weak non-linear effect--lightning that occurs in divergent stratiform regions.
Dept. of Meteorology
Model Parameters
Most important parameters for 1800-2059 UTC
BLR models for one or more CG flashes
NB models for the amount of lightning
Prevailing low-level wind
Dept. of Meteorology
Prevailing Low-Level Wind
0.40
Frequency of One or More Flashes vs. 1000-700 hPa Mean U-wind Component 1800-2059 UTC: East Coast Region
12
Unconditional Mean Number of Flashes vs. 1000-700 hPa Mean U-wind Component 1800-2059 UTC: East Coast Region
0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.00 -12 -10 -8 -6 -4 -2 0 2 4 6
-1
Unconditional Mean No. of Flashes
8 10 12
Frequency of One or More Flashes
10
8
6
4
2
0 -12 -10 -8 -6 -4 -2 0 2 4 6
-1
8
10
12
1000-700 hPa Mean U-wind Component (m s )
1000-700 hPa Mean U-wind Component (m s )
• • •
Highly non-linear relationship. Peak lightning for offshore speeds between 2 - 4 m s-1 Offshore flow produces better developed sea breeze and greater convergence.
Dept. of Meteorology
6. Results for Dependent Data
Dept. of Meteorology
Reliability
Reliability Diagram Logistic Models for Prob(>= 1 flash) 1800-2059 UTC: All Regions
0.90 0.80
•
Observed Relative Frequency
0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00 0.00
•
•
•
Reliability is a measure of the quality of probabilistic forecasts. Forecast probabilities correspond well with observed frequencies. The forecasts ―mean what they say.‖ Reliability for other time periods also is very good.
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
Forecast Probability
Forecasting one or more flashes All regions combined
Dept. of Meteorology
Reliability
Reliability Diagram Negative Binomial Prob(>= 50th percentile) 1800-2059 UTC: All Regions
0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00 0.00
Reliability Diagram Negative Binomial Prob(>= 75th percentile) 1800-2059 UTC: All Regions
0.55 0.50
Observed Relative Frequency
Observed Relative Frequency
0.10 0.20 0.30 0.40 0.50 0.60 0.70
0.45 0.40 0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.00 0.00
0.10
0.20
0.30
0.40
0.50
Unconditional Probability
Unconditional Probability
Forecasting ≥ 50th percentile All regions combined
Forecasting ≥ 75th percentile All regions combined
Dept. of Meteorology
Reliability
Reliability Diagram Negative Binomial Prob(>= 90th percentile) 1800-2059 UTC: All Regions
0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.00 0.00
0.30
Reliability Diagram Negative Binomial Prob(>= 95th percentile) 1800-2059 UTC: All Regions
Observed Relative Frequency
Observed Relative Frequency
0.05 0.10 0.15 0.20 0.25 0.30 0.35
0.25
0.20
0.15
0.10
0.05
0.00 0.00
0.05
0.10
0.15
0.20
0.25
0.30
Unconditional Probability
Unconditional Probability
Forecasting ≥ 90th percentile All regions combined
Forecasting ≥ 95th percentile All regions combined
Dept. of Meteorology
Verification Scores
• • Probabilistic forecasts converted to deterministic for verification. Optimum probability thresholds determined for each model and time period based on several verification scores from 2 x 2 contingency tables:
•
Critical Success Index- hit rate after removing correct ―no‖ forecasts from consideration.
•
We chose to maximize PSS- contribution to PSS by a correct forecast increases as the event becomes less likely.
Dept. of Meteorology
Verification Scores
Summary for Dependent Data
• • • • Scores for forecasting one or more flashes are very good--81% of events correctly forecast, with reasonable FAR and BR. Forecasting the amount of lightning is more difficult, with low predictability for 95th percentile events. Model CSI and PSS scores for all time periods are an improvement over L-CLIPER and persistence alone. Percent improvement in CSI greatest during most active periods.
Dept. of Meteorology
Example Probability Forecast
Prob(≥ 1 flash) Lightning Strike Verification
1200-1459 UTC 4 June 2004
1200-1459 UTC 4 June 2004
• • •
Forecasts based on RUC analyses. Maps show expected diurnal trend in lightning. Good agreement between forecast and verification.
Dept. of Meteorology
Example Probability Forecast
Prob(≥ 1 flash) Lightning Strike Verification
1500-1759 UTC 4 June 2004
1500-1759 UTC 4 June 2004
• • •
Forecasts based on RUC analyses. Maps show expected diurnal trend in lightning. Good agreement between forecast and verification.
Dept. of Meteorology
Example Probability Forecast
Prob(≥ 1 flash) Lightning Strike Verification
1800-2059 UTC 4 June 2004
1800-2059 UTC 4 June 2004
• • •
Forecasts based on RUC analyses. Maps show expected diurnal trend in lightning. Good agreement between forecast and verification.
Dept. of Meteorology
Example Probability Forecast
Prob(≥ 1 flash) Lightning Strike Verification
2100-2359 UTC 4 June 2004
2100-2359 UTC 4 June 2004
• • •
Forecasts based on RUC analyses. Maps show expected diurnal trend in lightning. Good agreement between forecast and verification.
Dept. of Meteorology
Example Probability Forecast
Prob(≥ 1 flash) Lightning Strike Verification
0000-0259 UTC 5 June 2004
0000-0259 UTC 5 June 2004
• • •
Forecasts based on RUC analyses. Maps show expected diurnal trend in lightning. Good agreement between forecast and verification.
Dept. of Meteorology
7. Results for Independent Data
Dept. of Meteorology
Independent Test
• Equations were applied to forecast output from several mesoscale models during the 2006 warm season: 1. 1500 UTC NCEP 13-km RUC (RUC13):
-- Runs at the highest frequency of any NCEP model
2. 1200 UTC NCEP 12-km NAM-WRF:
-- Transitioned from Eta to WRF-NMM in June 2006
3. 1500 UTC High Resolution (4 km) WRF:
-- South Florida domain (NWS Miami) -- Initialized with NCEP 1/12th degree SSTs and data from the Local Analysis and Prediction System (LAPS)
•
Results are for 1 August - 30 September 2006.
Dept. of Meteorology
What is LAPS?
• • Diagnostic tool in AWIPS--produces a high resolution 3D analysis of the atmosphere. Combines a background field (1-h forecast from AWIPS RUC40) with local data from a variety of observing systems: -- Surface observations (e.g., ASOS, local mesonetworks) -- Doppler radar reflectivity (Miami and Key West radars) -- Satellites (for cloud hydrometeors) -- Wind and temperature profilers -- Aircraft Produces 3D diabatic analysis grids for WRF initialization. WRF initialized with LAPS produces better forecasts of the sea breeze and surface parameters (Bogenschutz 2004; Etherton and Santos 2006).
• •
Dept. of Meteorology
WRF-LAPS
• • • WRF Environmental Modeling System (―Workstation-WRF‖) Runs for 19-30 September provided by Dr. Pablo Santos Runs for 1 Aug - 18 Sept produced at FSU using WRFEMS package
•
• • • 4-km resolution 1500 UTC initialization Forecasts out to 12 h •
Same model configuration as used at NWS Miami.
We did not compare results using different physics options or cumulus schemes (future research???)
Dept. of Meteorology
Skill Scores
Forecasting One or More Flashes Critical Success Index
• • • •
Scores are very encouraging! All models show positive skill through 2100-2359 UTC. WRF-LAPS generally outperforms RUC13 and NAM-WRF during most active period (1800-2059 UTC). Expected degradation in skill at longer forecast projections.
Dept. of Meteorology
Skill Scores
Forecasting One or More Flashes Peirce Skill Statistic
• • • •
Scores are very encouraging! All models show positive skill through 2100-2359 UTC. WRF-LAPS generally outperforms RUC13 and NAM-WRF during most active period (1800-2059 UTC). Expected degradation in skill at longer forecast projections.
Dept. of Meteorology
Skill Scores
Forecasting ≥ 95th percentile Peirce Skill Statistic
• • •
RUC13 is best performer through 6 h for forecasting 95th percentile events. All models show positive skill relative to persistence alone through 2100-2359 UTC. 1800 UTC cycles (not examined) likely would show positive skill for the 0000-0259 UTC period.
Dept. of Meteorology
Example Probability Forecast
Prob(≥ 1 flash) Prob(≥ 90th percentile) Verification
0-h WRF-LAPS forecast: 1500-1759 UTC 16 August 2006
• • •
Forecasts based on 1500 UTC WRF-LAPS. Timing and placement of features is not perfect, but… Generally good agreement between the forecasts and the verification.
Dept. of Meteorology
Example Probability Forecast
Prob(≥ 1 flash) Prob(≥ 90th percentile) Verification
3-h WRF-LAPS forecast: 1800-2059 UTC 16 August 2006
• • •
Forecasts based on 1500 UTC WRF-LAPS. Timing and placement of features is not perfect, but... Generally good agreement between the forecasts and the verification.
Dept. of Meteorology
Example Probability Forecast
Prob(≥ 1 flash) Prob(≥ 90th percentile) Verification
6-h WRF-LAPS forecast: 2100-2359 UTC 16 August 2006
• • •
Forecasts based on 1500 UTC WRF-LAPS. Timing and placement of features is not perfect, but... Generally good agreement between the forecasts and the verification.
Dept. of Meteorology
Example Probability Forecast
Prob(≥ 1 flash) Prob(≥ 90th percentile) Verification
9-h WRF-LAPS forecast: 0000-0259 UTC 17 August 2006
• • •
Forecasts based on 1500 UTC WRF-LAPS. Timing and placement of features is not perfect, but... Generally good agreement between the forecasts and the verification.
Dept. of Meteorology
8. Summary & Conclusions
Dept. of Meteorology
Summary & Conclusions
• Four warm seasons of RUC analyses and NLDN data were used to develop a perfect prognosis scheme for forecasting CG lightning over Florida:
-- Binary logistic regression was used to develop equations giving the probability of one or more CG flashes.
-- Negative binomial models were used to forecast the amount of lightning, conditional on one or more flashes occurring. • • Pattern type predictors were developed to capture enhancements due to local forcing. Deep layer moisture, instability, near-surface forcing, and the prevailing low level wind were found to have the greatest influence on the likelihood of one or more flashes and amount of lightning.
Dept. of Meteorology
Summary & Conclusions
• PP equations show forecast skill over L-CLIPER and persistence alone when applied to the dependent sample of RUC analyses, and during an independent test period using model forecasts.
•
• •
Results demonstrate that the scheme is model independent!
WRF-LAPS generally performs the best during the most active lightning period (1800-2059 UTC). Exact timing and placement of lightning maxima is not perfect, but there generally is good agreement between the forecasts and the verification. Methodology is an enhancement to schemes already in use.
•
•
Scheme will be used operationally by FP&L beginning in May.
Dept. of Meteorology
Future Work
• • Temporal resolution can be increased by developing separate equations for each hour--can be applied to hourly RUC forecasts. Plans are in place to incorporate the guidance into the Interactive Forecast Preparation System (IFPS) Graphical Forecast Editor (GFE) for use by forecasters at all Florida WFOs: -- A forecaster can use output from one NWP model or a blend of two or more models to generate the lightning forecasts. -- Forecasts can be accessed by the public through NWS web sites. Scheme may be expanded to other parts of the country. Bayesian framework may produce better results? Availability of higher resolution analyses and larger developmental sample should result in more robust parameter estimates.
• • •
Dept. of Meteorology
Questions?
Dept. of Meteorology