Post on 07-Jan-2016
description
CMAQ PM2.5 Forecasts Adjusted to Errors in Model Wind Fields
Eun-Su Yang1, Sundar A. Christopher2 1Earth System Science Center, UAHuntsville
2Department of Atmospheric Science, UAHuntsville
Shobha Kondragunta3, and Xiaoyang Zhang4
3NOAA/NESDIS/STAR4Earth Resources Technology Inc. at NOAA/NESDIS/STAR
Presentation to the 9th CMAS Conference
October 11, 2010
OUTLINE
• Motivation- Georgia fires in 2007- large discrepancy between simulated and observed PM2.5
• Smoke position error due to model wind error- wind error with a random component- wind error with random & autoregressive (AR) components
first attempt to estimate transport error amplified by AR processes
• Suggestion to improve PM2.5 forecasts
• Summary
MOTIVATION: Georgia fires in 2007
(Left) Fire spots and fire plumes viewed from the MODIS Terra on May 22, 2007. Taken from http://rapidfire.sci.gsfc.nasa.gov/realtime.
Fire location and emission data from the GOES biomass burning emission product
(ftp://satepsanone.nesdis.noaa.gov/EPA/GBBEP)
Transport of fire plume from MM5 and CMAQ
MODIS AquaAOT
CMAQ FIRESurface PM2.5
AIRSTotal column CO
Overall Agreement
The CMAQ PM2.5 simulations (middle) reproduced well the observed pattern of smoke transport (see the map of AIRS total column CO).
MOTIVATION: Discrepancy
CMAQ missed the high PM2.5 episode
May 26, 2007
Uncertainties in fire emission estimates and plume injection height were not the primary cause.
Select the worst MM5 case example during May 20-31.
24-hour forward trajectories, starting at 925 hPa, 00Z May 26, 2007
Position error might be too underestimated
MM5 Wind Evaluation (1)
Comparison with the Automated Surface Observing System (ASOS) observations at 10 m for April 1 – May 31, 2007
MM5 Wind Evaluation (2)
Bias (m/sec)
(u / v)
RMSE (m/sec)
(u / v)
Birmingham -0.1 / 0.4 1.5 / 1.5
Atlanta -0.2 / 0.6 1.4 / 1.7
Tallahassee 0.0 / 0.1 1.4 / 1.4
Bias: -0.2 ~ 0.6 and RMSE: < 2 m/sec
• agree well with the previous works [e.g., Emery et al., 2001; Olerud and Sims, 2004]
• similar errors by comparing with the Rapid Update Cycle (RUC) analysis data at various pressure levels
MM5 Winds versus RUC Analysis Windsalong Trajectory Path
If the RUC wind is assumed true:
First-Order Autoregressive Process, AR(1)
0.75 < AR(1) < 0.95 for (MM5 – RUC) and (MM5 – ASOS) winds
The Statistical Perspective
Variance of position at time t = wind variance at time 0 * time interval + wind variance at time 1 * time interval
… + wind variance at time t-1 * time interval
VAR{Dt} = VAR{U0} + VAR{U1} + … + VAR{Ut-1}
= VAR{ (ε0 + φ ε-1 + φ2 ε-2 + φ3 ε-3 + φ4 ε-4 + φ5 ε-5 + …)
+ (ε1 + φ ε0 + φ2 ε-1 + φ3 ε-2 + φ4 ε-3 + φ5 ε-4 + …)
+ (ε2 + φ ε1 + φ2 ε0 + φ3 ε-1 + φ4 ε-2 + φ5 ε-3 + …)
…
+ (εt + φ εt-1 + φ2 εt-2 + φ3 εt-3 + φ4 ε-4 + φ5 ε-5 + …) }
where σε2/(1 - φ2) = σ2, εi and εj are independent if i ≠ j [see Yang et al., 2006, JGR]
σε2/(1-φ2) = σ2
σε2/(1-φ2) = σ2
σ2 (1+φ)/(1-φ)φ σ2 (1+φ)/(1-φ)
Variance of Position Error without AR(1)
t = 1, VAR = 1 ∙ σ2
t = 2, VAR = (2 + 2φ) ∙ σ2
t = 3, VAR = (3 + 4φ + 2φ2) ∙ σ2
t = 4, VAR = (4 + 6φ + 4φ2 + 2φ3) ∙ σ2
t = 5, VAR = (5 + 8φ + 6φ2 + 4φ3 + 2φ4) ∙ σ2
…
where φ is the AR(1) coefficient (-1< φ < 1).
If φ = 0, VAR ~ t & standard deviation ~ t1/2.
If φ 1, VAR t2, standard deviation t
t = 1, VAR = 1 ∙ σ2
t = 2, VAR = (2 + 2φ) ∙ σ2
t = 3, VAR = (3 + 4φ + 2φ2) ∙ σ2
t = 4, VAR = (4 + 6φ + 4φ2 + 2φ3) ∙ σ2
t = 5, VAR = (5 + 8φ + 6φ2 + 4φ3 + 2φ4) ∙ σ2
…
where φ is the AR(1) coefficient (-1< φ < 1).
If φ = 0, VAR ~ t & standard deviation ~ t1/2.
If φ 1, VAR t2, standard deviation t
Variance of Position Error with AR(1)
Position error with AR(1) in model wind error
t1/2
t1
The Physical Perspective
Map from http://www.hpc.ncep.noaa.gov
Surface map at 18Z May 26, 2007
Model wind errors have short-term memory.
Modeled or analyzed winds are subject to have the AR behavior due to the use of
incomplete modeling of wind that is aroused from dynamical
simplifications, incorrect physical parameterizations, incorrect initial and boundary conditions, among others.
17Z 18Z 19Z … days later
RUC ???
MM5 ???
Allow smoke plume in all grids within the range of (1-sigma) position error to not miss the observed-but-not-predicted case
NotObserved Observed
Predicted occasionally rarely
NotPredicted
most often occasionally
When smoke position error >> smoke width:
site
Think about Ensemble Approach
Initial perturbations (σ) are typically assumed Gaussian (normal)
Magnitude of the perturbations is given empirically
Our approach is analytical and needs model run only once
Comparison with AirNow PM2.5 Observations
Local emissions Local + fire emissionsLocal + fire emissions
with increased range of smoke
CMAQ: local emissions from the emission inventory model (SMOKE) fire emissions from the GOES biomass burning emission productObservations: daily surface-level PM2.5 observations at 17 AirNow sites
for April - May, 2007: 8 in Alabama, 6 in Georgia, and 3 in northern Florida.
missed alarm versus false alarm
EPA daily PM2.5 limit
SUMMARY
• Model winds are a key factor in predicting smoke position
• Model wind errors are positively autocorrelated
• Smoke position error tends to be underestimated without consideration of AR(1) in model wind error
• AR(1) term is included to capture high PM2.5 episodes and reduces the number of missed-alarm cases
• Ensemble approach could be improved by adding AR processes