Lysbilde 1

Report
Applications of inverse modeling for
understanding of emissions and
analysis of observations
Rona Thompson, Andreas Stohl, Ignacio Pisso,
Cathrine Lund Myhre, et al.
1
Content of presentation
FLEXPART transport model
Statistical analysis of observation data: Methane results for Zeppelin
station
Inversion basics
Applications to halocarbon emissions
FLEXINVERT
2
The FLEXPART model
Model descriptions in Atmospheric Environment,
Boundary Layer Meteorology, Atmospheric Chemistry and Physics
Lagrangian particle dispersion model
Turbulence and convection parameterizations
Dry and wet deposition
Data input from ECMWF, GFS, MM5, WRF,…
Used at probably >100 institutes from
several dozen countries
Model set-up
Can be run both forward (from sources) or backward
(from measurement stations) in time, whatever is more
efficient
Here: Backward in time for 20 days
Model output: 4-dimensional emission sensitivity field
(3 space dimensions plus days backward in time)
Mixing ratio = emission sensitivity field x emission flux field
http://zardoz.nilu.no/~andreas/STATIONS/ZEPPELIN/Zeppelin_201001/ECMWF/polar_
column_t/Zeppelin_201001.polar_column_t_1.html
Transport climatology (2001-2012)
Footprint emission sensitivity maps averaged for
the four seasons (upper panels) and normalized
to annual mean
DJF
MAM
JJA
SON
Cluster analysis
Cluster analysis of trajectory output
(Dorling et al., 1992)
Cluster analysis can be used to
stratify measurement data
according to transport pathway
Disadvantage: no good control on
the ”shape” of the clusters, no
clear separation of sources, no
quantitative information on
emissions
6
Cluster analysis (2001-2012)
Siberia and Central Asia = SCA, Western
Arctic Ocean = WAO, Arctic Ocean = AO,
Canada and Greenland = CGA, North
Atlantic Ocean = NAO, East Asia and
North Pacific = EA, Europe and North
America = ENA, Siberia Northeast Asia =
SNEA
7
”Ashbaugh method”
Ashbaugh, 1983; Ashbaugh et al., 1985
Define a grid
Associate M measurements with trajectories and calculate total gridded residence
time ST from individual gridded residence times
where i, j are grid indices. Then, select subset with L=M/10 highest 10% measured
concentrations
To identify source/sink areas, calculate
If concentration not associated with transport: RP(i,j) = 0.1 everywhere
Where there is a source: RP(i,j) > 0.1
8
”Ashbaugh method”
Detrended and
deseasonalized
2001-2012 CH4 data
Highest 10%
Lowest 10%
log(s m-3 kg-1)
Emission sensitivity
Sp
Emission sensitivity
normalized by emission
sensitivity for all data
Rp
9
”Ashbaugh method” – local scale
Detrended and
deseasonalized 2001-2012
methane data
Highest 10%
Lowest 10%
log(s m-3 kg-1)
Emission sensitivity
Sp
Emission sensitivity
normalized by emission
sensitivity for all data
Rp
10
The inverse modeling problem
Needs a large set of atmospheric concentration measurements, ideally from many
stations and/or campaigns
Want to use these data to determine the emissions of the studied substance
Substance can be subject to removal processes (e.g., aerosols) or considered
(almost) passive on short time-scales (e.g., CH4)
To use inverse modelling, the underlying atmospheric transport model must be able
to account for these processes, i.e., it must be possible to establish quantitative
source-receptor relationships
Systematic errors in the model would (likely) cause bias in retrieved emissions
11
Bayesian inversion basics
Aim: Determination of the emission sources from air concentration measurements
M ... M x N matrix of emission sensitivities from transport model calculations
… often called source-receptor relationship
x ... Emission vector (N emission values)
y ... Observation vector (M observations)
Difficulty: poorly constrained problem; large spurious emissions can easily result to
satisfy even single measurement data points as there is no penalty to unrealistic
emissions
Solution: Tikhonov regularization: ||x||2 is small
Bayesian inversion basics
Slight reformulation if a priori information is available
yo ... Observation vector (M observations)
xa ... A priori emission vector (N emission values)
Tikhonov regularization: ||x-xa||2 is small
We are seeking a solution that has both minimal deviation from the a priori, and
also minimizes the model error (difference model minus observation)
Bayesian inversion basics
Minimization of the cost function
1
2
1. Term: minimizes squared errors (model – observation)
2. Term: Regularization term
x, o ... Uncertainties in the a priori emissions and the observations
diag(a) … diagonal matrix with elements of a in the diagonal
The uncertainties of the emissions and of the „observations“ (actual mismatch
between model and observations) give appropriate weights to the two terms
Halocarbon emissions
in China
Example: HFC-23
a by-product of HCFC-22 production
Black dots: 3 measurement stations
Top panel: emission distribution
available a priori
Bottom panel: inversion result
Asterisks: known locations of HCFC22 factories
New development by Rona Thompson:
FLEXINVERT
Description planned for Geosci. Mod. Dev.
• Planned as an open-source development
• Partly builds on Stohl et al. (2009) algorithm
• Algorithm specifically developed for long-lived greenhouse gases
• Allows coupling of 20-day FLEXPART backward runs with global model output
• Modular, so can be adjusted to different requirements (CH4, CO2, N2O, SF6, etc.)
• Allows flexible time resolution of the emissions (e.g., monthly)
• Facilitates error correlations of the prior emissions (spatially and temporally)
• Calculates posterior flux error covariances (i.e., errors in emissions)
First application to East Asia
Emission sensitivity log(s m3 kg-1)
Variable grid resolution
Application to East Asia (1)
Atmospheric observations in nested domain
Institute
Type
No. sites
CAMS
in-situ (CRDS)
4
NIES
in-situ (GC-FID)
2
NOAA
flask (GC-FID)
4
JMA
in-situ (NDIR)
3
KMA
in-situ (GC-FID)
1
NIER
in-situ (GC-FID)
1
TOTAL
15
Application to East Asia (2)
Prior emissions
Source
Dataset
Total (TgCH4 y-1)
anthropogenic
- rice cultivation
- waste
- fuels
- animal agriculture
EDGAR-4.2
331
natural wetlands
LPJ DGVM model
175
biomass burning
GFED-3
13
geological
based on Etiope et al. 2008
55
termites
Sanderson et al. 1996
19
wild animals
Olson et al. 1997
5
soils
Ridgewell et al. 1999
-38
ocean
Lambert and Schmidt, 1993
17
TOTAL
577
Results (1)
Annual mean fluxes for 2009
China a priori:
61.6 TgCH4/y
China a posteriori:
59.6 TgCH4/y
Results (2)
0.27
0.53
0.40
0.71
0.45
0.57
0.64
0.79
0.33
0.57
0.52
0.72
0.37
0.49
0.41
0.69
0.38
0.50
0.29
0.35
0.12
0.26
OBS
0.27 PRIOR
0.71 POST
BKGND
Conclusions
In MOCA, we will use inverse modeling as a tool to analyze CH4 data
using station network (Zeppelin, Pallas, etc.)
using campaign data
Algorithm (almost) ready but will need further development/testing
Will also utilize other means of analyzing data

similar documents