Bayesian Monte Carlo analysis applied to a regional scale transport chemistry model Deguillaume L., Beekmann M., Menut L., Derognat C. Improving emission.

Bayesian Monte Carlo analysis applied to a Bayesian Monte Carlo analysis applied to a

regional scale transport chemistry modelregional scale transport chemistry model

Deguillaume L., Beekmann M., Menut L., Derognat C.

Improving emission uncertainties

Characterizing ozone production and

chemical regimes

Over the Ile-de-France region

13/10/2006 - Gloream

Context Photochemical air pollution

EMISSIONS

TRANSPORT

CHEMISTRY

O3OH HO2

NO NO2

NO2 NO

RO2 RO

Reduced VOC Oxidized VOC

ROOH

O3+h

HNO3

H2O2

RCHO +h

hh

+NO2

+ HO2

+ HO2

NOx limited

VOC limited

Context Modelling those processes

Development of a chemistry transport model CHIMERE

(IPSL/ INERIS/LISA)

Model domain : 150×150km

Horizontal resolution : 6×6km grid

Vertical resolution : 8 layers in hybrid pressure coordinates

Chemical mechanism : reduced Melchior

Meteorology: ECMWF

Anthropogenic emissions: EMEP, ARIA, AIRPARIF

Biogenic emissions: Simpson et al. (1999)

46

47

47

47

48

48

48

48

48

48

49

49

49

49

49

49

49

50

50

50

50

50

50

50

50

50

50

51

51

51

51

51

52

4445464748495051525354

48.5

49.0

1.5 2.0 2.5 3.0

Ile-de-France region

Airparif

Objectives

Observation : Simulated concentrations are very sensitive to emissions

Inverse modelling of emissions from observations (ground based and satellite)

(1) Uncertainties in emissions is always rather large and difficult to estimate

Uncertainties on activity factor, emission factor spatial distribution, temporal

variability...

(2) Evaluation difficult since emitted pollutants undergo chemical transformation

and are tranported away from sources

Problems

Objectives Improving emission uncertainties with a Bayesian approach

Application to semi-climatologic (summers 1998+1999) period for generalization

Adjoint model Kalman filter Bayesian Monte Carlo analysis

Methods : inverse modelling

To verify and improve available

estimates of atmospheric pollutants

emissions

Alternative to bottom-up

construction of emission cadastres

To improve performance of

atmospheric models, especially in

diagnostic studies

To develop a general

observation-based methodology for

estimating parameters of the

atmosphere that cannot be

observed directly

Real world Mathematical world

Concentrations(observations)

Concentrations(simulations)

TransportChemistryDeposition

Transport /Chemistry

model

Emissions Emission cadastre

Comparison

Improve, update

Principle of the Bayesian Monte Carlo analysis

A priori uncertainties

in emissions

A priori uncertainties of

input parameters

« Model uncertainties »

Monte Carlosimulations

A priori concentrations

without constraints

Weighting by observations

A posteriori

distributions of

emissions

Uncertainties

in observations

Correction on a priori distributions of emissions

(also on other perturbed input parameters)

(-) Single correction factor over the whole grid domain and time period of the simulations

(+) Information on the value but also uncertainty associated to the emissions

Mathematical formulation

2

j,

j,kj

j,e

N,1j

YO5.0exp

1

2

1

P(O|Yk)

For each kth Monte Carlo

simulation, the agreement

function:

« Probability to observe a vector of observations O given that the model output Yk is

the true value for the kth Monte Carlo simulation »

Hypothesis: Observations present a normally distributed errors ε

N independent observations Oj

Each simulation is weighted by P(O|Yk) Cost function

A posteriori probability density function vs. a priori one

Results ?

Cumulative probability density function (CPDFs)

(probability that a given model prediction Xk stays below the limit X)

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

Logarithmic variation

0

0.5

1

1.5

2

No

rm.

pro

ba

bili

ty

ENOx_Working days

Perturbed model input parameters A priori uncertainties of

input parameters

« Model uncertainties »Parameters 1 σ Uncertainty

Emissions

Anthropogenic VOCs + 40

Anthropogenic NOx + 40

Biogenic VOCs + 50

Rate constants

NO + O3 + 10

NO2 + OH + 10

NO + HO2 + 10

NO + RO2 + 30

HO2 + HO2 + 10

RO2 + HO2 + 30

RH + OH + 10

CH3COO2 + NO + 20

CH3COO2 + NO2 + 20

PAN + M + 30

Photolysis frequencies and radiation

Actinic fluxes + 10

J(O3 2 OH) + 30

J(NO2 NO + O3) + 20

J(CH2O CO + 2 HO2) + 40

J(CH3COCO …) + 50

Meteorological parameters

Zonal wind speed + 1

Meridional wind speed + 1

Mixing layer height + 40

Temperature + 1.5

Relative humidity + 20

Vertical mixing coefficient + 50

Others

Deposition velocity + 25

O3OH HO2

NO NO2

NO2 NO

RO2 RO

Reduced VOC Oxidized VOC

ROOH

O3+h

HNO3

H2O2

RCHO +h

hh

+NO2

+ HO2

+ HO2

Log-normal distribution

Uncertainty ranges uncertainty assessment studies and expert judgements

Measurement constraintsobservations

1.6 1.7 1.8 2.0 2.22.11.9 2.72.3 2.6 2.82.52.4 2.9 3.048

48.2

48.1

48.8

49.1

49.0

48.9

48.7

48.6

48.5

48.4

48.3

49.4

49.3

49.2

LATITUDE

LONGITUDE

PARIS

Nd

1k

Ns

1j

Nh

1i3 )k,j,i(obs

NhNsNd

1)O,NO(OBS

(1) Urban NO and O3

NO direct forcing for NOx emissions

O3 information on ozone precursor

emissions (VOC, NOx)

AIRPARIF NETWORK

(2) Rural O3 buildup

The two daily maximal [O3] O3 plume

The 3 lowest [O3] O3 background

For simulations and observations:

+(1) Days where the maximal [O3] is observed and

simulated at the same or neighbouring station

(2) [O3 max] - [O3 back] > 10 ppb for simulations

and observations

Results – Cumulative probability

Blue without constraints

Simulated urban NO, O3 and O3 production in the plume- summers 1998 + 1999

Uncertainties are reduced by a factor :3.2

2.4

1.7

Deguillaume et al., in press, JGR, 2006

60 80Urban NO (ppb)

0

0.2

0.4

0.6

0.8

1

Cum

ula

tive p

rob

abilit

y

40200

Urban NO climato 98+99

7020 30 40 50 60

Urban O3 (ppb)

0

0.2

0.4

0.6

0.8

1

Cum

ula

tive

pro

bab

ility

Urban O3 climato 98+99

40

O3 in plume (ppb)

0

0.2

0.4

0.6

0.8

1

Cum

ula

tive p

rob

abilit

y

0 10 20 30

O3 plume climato 98+99

2σ 2σ

2σ 2σ 2σ

2σ

Totalconstraints

Red Total constraints

Results – Probability density functions

Blue histogram a priori distribution with 40% 1σ uncertainty

Emissions of anthropogenic NOx and VOC – cumulative summers 1998 and 1999

1σ uncertainty : 22 % for NOx , 31% for VOC emissions

NOx emissions remain nearly unchanged - VOC emissions are enhanced (+16%)

Better fit the observations in 1999 vs. 1998 (in 1998, the constraints do not act in a similar way)

-1.2 -0.8 -0.4 0 0.4 0.8 1.2Log. variation

0

0.4

0.8

1.2

1.6

2

Norm

. p

rob

ab

ilit

y

ENOx_98+99_climato

-1.2 -0.8 -0.4 0 0.4 0.8 1.2

Log. variation

0

0.4

0.8

1.2

1.6

2

Norm

. p

rob

ab

ilit

y

EVOC_98+99_climato

-1.5 -1 -0.5 0 0.5 1 1.50

0.4

0.8

1.2

1.6

Log. variation

EVOC/ENOx_98+99_climato

Totalconstraints

Red lines a posteriori distribution

Conclusion & perspectives

Bayesian Monte Carlo uncertainty analysis Semi-climatologic period

Ile-de-France region

A posteriori PDF NOx emission unchanged average

reduced standard deviation (20% vs. 40%)

VOC emission enhancement (+16%)

reduced standard deviation (30% vs. 40%)

Uncertainty in the simulated urban NO, urban O3 and O3 production in the plume are

strongly reduced

indirect constraints on VOC emissions (urban O3 and O3 production) lower reduction

Adjustements in the other model input parameters (vertical mixing coefficient, 1998)

Better fit in 1999 than in 1998 because constraints in 1998 do not act in a similar way

Other region ? Marseille area

(ESCOMPTE, AIRMAIX network)Satellite measurements

Objectives Characterizing ozone production and chemical regimes

Better understand the buildup of pollution episodes around Paris region

Perspectives…

Application to semi-climatologic (summers 1998+1999) period for generalization

Direct simulation Bayesian Monte Carlo approach

Constraints by observations

Emission reduced scenario (-30% NOx and -30% VOC emissions)

Analysis of the chemical regime over the Ile de France region

2 approachesMethodology

(2) Characterizing ozone production and chemical regimes

Preliminary results ...

46

47

47

47

48

48

48

48

48

48

49

49

49

49

49

49

49

50

50

50

50

50

50

50

50

50

50

51

51

51

51

51

52

4445464748495051525354

48.5

49.0

1.5 2.0 2.5 3.0

47

48

48

48

48

49

49

49

49

49

49

50

50

50

50

50

5050

50

50

51

51

51

51

51

52

52

52

52

53

53

5353

54

54

54

54

4647484950515253545556

1.5 2.0 2.5 3.0

48.5

4344

45

46

46

46

46

47

47

47

47

474

7

47

48

48

48

48

4848

48

48

48

48

49

49

49 49

49

49

49

49

50

40414243444546474849505152

1.5 2.0 2.5 3.0

48.5

49.0

Anticyclone sur l’ocean mouvement aiguille montre vent vers le sud

NOx -30% VOC -30%Reference

Daily maximum of O3 averaged over summers 1998 and 1999