Transcript presented

Estimates of worst 20% natural condition
deciview: application of the new IMPROVE
algorithm and a revised statistical approach
Rodger Ames, CIRA ([email protected])
Marc Pitchford DRI, NOAA
RPO Monitoring and Data Analysis Call
April 26, 2006
1
Reasons to revise
• new IMPROVE algorithm (NIA)
– new OC mass conversion factor, mass scattering
efficiencies, etc.
• more monitoring site provide better spatially resolved
statistical parameters
– 55 sites available for 1996-2000 analysis, 141 for
2000-20041
– old method based on simplistic east/west division of
frequency distribution standard deviation. In new
approach every site has statistical underpinning
• revised statistical approach uses new insights
– Gaussian dv frequency distribution assumption can
be improved upon
1. Five urban sites were not used for 2000-2004 statistical analysis. Results for four AK sites, two 2
Hawaii sites, and VIIS are not presented here. Two sites failed RHR criteria for baseline period.
Steps
1. Estimate aerosol light extinction (aerosolbext)
and deciview (dv) frequency distribution
parameters by simulating natural conditions
from current data.
2. Determine appropriate statistical approach to
convert natural condition mean aerosol mass
concentrations to worst 20% natural condition
dv.
3. Simulate worst 20% natural condition dv using
key frequency distribution parameters obtained
from 2000-2004 IMPROVE data.
3
Step 1: simulate natural conditions scenario
from current data
• roll back current sulfate and nitrate aerosol mass
concentrations to natural levels
– assume nitrates and sulfates are predominately
anthropogenic in origin, while carbon and crustal
species are predominately natural origins
– for each IMPROVE site (n obs > 300, 129 sites) during
2000-2004, daily nitrate and sulfate concentrations are
scaled so that the 5-year mean equals the estimated
annual average natural sulfate and nitrate levels for the
east and west U.S.1
4
Step 1: natural condition aerosol mass
concentrations, east/west U.S.
Table 1. Estimated particle mass concentrations (mg/m3)
for the east and west U.S. under natural conditions1
East
West
Error Factor
Ammonium bisulfate
0.2
0.1
2
Ammonium nitrate
0.1
0.1
2
Organic carbon mass
(1.5*OC)
1.5
0.5
2
Elemental carbon
0.02
0.02
2-3
Soil
0.5
0.5
1.5-2
Coarse Mass
3
3
1.5-2
East is defined as “basically up to
one tier of states west of the
Mississippi” and west as “basically
the desert/mountain areas of the
Mountain and Pacific time zones”1
Table 2. modifications to natural conditions mass and key differences in
new IMPROVE algorithm in revised approach
Mass conversion factor
East
West
Ammonium
sulfate
0.23
0.115
Organic carbon
1
0.333
Old IA
1.4
New IA
1.8
Mass scattering efficiency
Old IA
4
New IA
Small size
Large size
2.8
6.1
1. Trijonis J.C.; Malm W.C.; Pitchford M.; White W.H., Chapter 24 of Acidic Deposition: State of Science and
Technology, vol. 3 Terrestrial, Materials, Health and Visibility Effects, edited by P. M. Irving, U.S. Natl. Acid Precip. 5
Assess. Prog. (NAPAP)., Washington D.C., 1990.
Step1: roll back current sulfate and nitrate aerosol
mass concentrations to natural levels
Current and natural condition scenario daily dv frequency distribution
standard deviation using new IMPROVE algorithm and 2000-2004 data.
Figure 1. current dv standard deviation.
Figure 2. dv standard deviation for natural
condition scenario.
6
Step1: natural condition scenario and recap of old
default approach
• natural condition scenario dv standard deviation (s) is
between 3-4 for most of the continental U.S. IMPROVE
sites, with a region of lower dv s between 2-3 along the
Colorado Plateau
• regions of higher dv s, between 4-5, along the CA Sierra
and Northern Rockies. Other localized areas with dv s in
the 4-5 range occur, for example in FL at Everglades and
sites in the southwest and mountain regions
• old approach assumed east and west U.S. dv standard
deviations of 3 and 2, respectively to conform to
east/west mass concentrations. Also assumed Gaussian
dv frequency distribution (FD) to estimate distribution
tails from FD standard deviation and mean
7
Step 2: revise statistical approach - dv frequency
distributions
Examine natural condition scenario FD shape parameters. Natural condition
scenario dv FDs all have positive skew and kurtosis.
log (abext + R)
40
35
30
kurtosis
25
20
15
10
5
0
-5
0
1
2
3
4
5
skewness
Figure 1. FD kurtosis vs. skewness
Figure 2. map of FD skewness
Skew greater than 0 indicates the distribution is skewed to the right, or towards higher
values. A normal distribution has a skewness of 0. Positive kurtosis indicates distributions
8
with large tails. As presented, a normal distribution will have kurtosis of zero.
Step 2: revise statistical approach - log(aerosolbext)
frequency distributions
Natural condition scenario log(aerosolbext) frequency distribution skewness
and kurtosis.
log abext
18
16
14
12
kurtosis
10
8
6
4
2
-2
-1.5
-1
-0.5
0
-2 0
0.5
1
1.5
2
2.5
3
skewness
Figure 1. FD kurtosis vs. skewness
Figure 2. map of FD skewness
Skew is symmetric about zero and has less magnitude than in dv FDs. Some outliers
exist, possibly due to a few outlying datapoints.
9
Step 2: revise statistical approach - log(aerosolbext)
frequency distributions
Natural condition scenario
log(aerosolbext) frequency
distribution mean and standard
deviation
Figure 2. map of FD mean
log abext
0.6
0.5
std dev
0.4
0.3
0.2
0.1
0
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
mean
Figure 1. FD standard deviation vs. mean
Figure 3. map of FD standard deviation
10
Step 3: Simulate worst 20% natural condition
deciview using frequency distribution parameters
• generate random Gaussian (n=500) frequency
distributions to simulate natural log(aerosolbext)
FD.
– use natural condition aerosol mass (east/west), site specific fRHs,
site or regionally specific FD standard deviations.
• add site specific rayleigh to each (daily)
simulated abext FD value
• calculate daily simulated worst 20% dv
– FD(dv,nc)=10*ln(FD(FDbext/10)
• calculate mean of worst 20% dv
– P80 = floor(.8*n)+1
– G90= mean(FD(dv,nc(p80…n))
11
Step 3: natural condition scenario log(abext)
frequency distribution standard deviation
Figure 1. 2000-2004 log aerosol bext natural condition scenario frequency distribution standard
deviation, same as previous figure but scaled to observed range.
12
Step 3: natural condition scenario log(abext)
frequency distribution standard deviation
Natural condition scenario log(aerosolbext) frequency distribution standard
deviation at 3 and 2 step values.
Use site specific log(aerosolbext) standard deviation for NC simulations.
13
Recap: old default method, mean of worst 20%
natural condition dv
Figure 1. Ames and Malm (2001, Bend OR, AWMA conference proceeding). Modified to better estimate
14
worst 20% mean from normal distribution parameters (92 percentile rather than 90th).
Recap: old default method, mean of worst 20%
natural condition dv
Figure 1. Ames and Malm (2001 Bend OR, AWMA conference proceeding). Modified to better estimate
15
worst 20% mean from normal distribution parameters (92 percentile rather than 90th).
Step 3: new approach, mean of worst 20% natural
condition dv
Figure 1. Use NIA, site specific rayleigh, site specific standard deviation, Rand500 to generate normal
16
NC dv FD. Units are dv. Rand500 overestimating FD s by ~3%.
Step 3: new statistical approach vs. old default
method, natural condition mean of worst 20% dv
Regions in darker blue indicate ~< 10% change old default estimates. Positive values
indicate higher values from the new approach.
Figure 1. Fraction which the new NC G90 differs from the old approach (expressed as new/old-1)
17
Recap: glide slopes from current 2000-2004
conditions to default natural conditions
Figure 1. Glide slopes from the old default approach. Uses old IMPROVE algorithm for baseline and
18
natural condition values. Units are dv/10 years.
Step 3: glide slopes using new statistical approach
Regions in dark blue indicate current conditions are close to natural condition
estimates, or glide slope ~ zero.
Figure 1. Glide slopes from the new approach. Uses new IMPROVE algorithm for baseline and natural
19
condition values. Units are dv/10 years.
Step 3: new statistical approach vs. old default
method, glide slope comparison.
Regions in red indicate little modification to glide slope using new approach.
Negative values indicate decrease in glide slope magnitude for new approach.
Figure 1. Fraction which the new glide slope differs from the old approach (expressed as new/old-1) 20
Conclusions
• little change in NC estimates from NIA alone.
• log (aerosolbext) better approximates Gaussian
distribution than dv.
• new approach worst 20% dv higher by 30-80% than old
default approach along mountains of CA, northwest, and
northern U.S. Rockies.
• new glide slopes close to zero in some western regions,
suggesting current worst day conditions are near natural
estimates.
• using FD parameters derived from 2000-2004 data in
natural condition estimates makes sense because the
baseline data contain similar FD characteristics. That is,
sites with higher baseline dv (due to fire or dust) should
get higher natural conditions value. But do they?
21
Future work
Sensitivity analysis
• Gaussian FD assumption is a better approximation for log
transformed aerosolbext metric however, some deviations occur
– sensitivity to changes in natural condition mass concentrations
– sensitivity to changes in skew, kurtosis parameters
– examine sites with large magnitude FD skew, kurtosis - what is driving
negative skew at some sites?
Refinements to new natural condition estimation approach
• incorporate observed dist skew, kurtosis, and other distribution
parameters
• incorporate longer time period into NC scenario – better averaging for
fires, etc…
• how much can we rely on model data from 2000-2004 period for NC
estimates? (fire emissions in WRAP Plan02b, GEOS-Chem output)
Species specific worst 20% natural condition estimates
• decompose aerosolbext FD into component species
• species GP = (species base line abext- species NC abext)/60 yrs
22
2000-2004 log(abext) FD mean and standard
deviation
23
2000-2004 log(abext) FD mean and standard
deviation under natural condition scenario
24
Recap: mean of worst 20% natural condition dv
using old method and new IMPROVE algorithm
Figure 1. New IMPROVE algorithm used in default approach. Rayleigh = 10
25
Step 3: new statistical approach natural condition
mean of worst 20% dv
Figure 1. Uses NIA, site rayleigh, Rand10K to generate NC dv FD. Units are dv.
26
2000-2004 worst 20% dv baseline
Figure 1. Current conditions baseline for 2000-2004 worst 20% dv. Units are dv.
27