Model ensembles for the simulation of air quality over Europe

1
Model ensembles for the simulation of air quality
over Europe

• Robert Vautard
• Laboratoire des Sciences du Climat et de
  lEnvironnement
• And many colleagues from IPSL, LISA, INERIS,
  EURODELTA and TRANSCOM projects

2
Why air quality modelling?

• Short-term forecasts (0-3 days)
• Long-term predictions of emission scenarios
  (climate?) ?2010 or 2020 or more
• Increase knowledge on processes together with
  observations

3
Air Quality forecastingPrevention

• 10 Years ago statistical models, actions based
  on observations
• Now many deterministic forecasting systems
• Data assimilation in some cases
• In France, PREVAIR system
• European GEMS/MACC projects (GMES)

4
What are regional AQ models?
Transport
Chemistry
Many many uncertainties
5
Regional air quality forecastnot really an
initial value problem
without assimilation
with assimilation
Blond et al 2004
Assimilation experiments
6
What are the skill of regional AQ forecasts?
PREVAIR Operational AQ forecasts (3
Summers) Average skill over gt200 stations in
Europe Honoré et al. 2008
7
Ensembles with perturbed meteorology (ARPEGE),
chemistry

• Carvalho et al., in preparation

8
Emissions controlAction

• Loss in life expectancy attributable to PM2.5,
  and 2020
• simulation with current legislation, Amann et al
  2005

9
But some species are very poorly simulated
PM Episode intercomparison Stern et al. 2008
10
Hopes from ensembles

• Represent the  unpredictable part  of the
  system
• Meteorological/emission  noise , knowledge gaps
• Provide better deterministic predictions by
   error cancelation 
• Delle Monache and Stull 2003 Galmarini et al.,
  2004 McKeen et al., 2005
• Predict the uncertainty (in forecasts, in
  scenarios), using the range
• Using one perturbed mode Hanna et al., 2001
  Mallet and Sportisse 2006, Deguillaume et al.,
  2008, or a model ensemble Vautard et al.,
  2006

How to evaluate ?

• Easy for deterministic predictions
• More difficult for uncertainty tools borrowed
  from ensemble weather forecasting

11
EuroDelta Experiment

• Regional, european scale evaluation of emission
  scenarios for 2010 or 2020
• Control experiment simulation of Year 2001
• 7 models CHIMERE, DEHM, EMEP, LOTOS-EUROS,
  MATCH, RCG, TM5,
• Comparison with rural stations (EMEP or AIRBASE)
• Results in
• Van Loon et al., 2007 (Atmos. Env.)
• Schaap et al., 2008 (in revision)
• Vautard et al., 2006 (Geophys. Res. Lett.)
• Vautard et al., 2008 (AE, submitted)

12
Example of improvement by ensemble averaging
Mean diurnal cycles
Ozone
OxO3NO2
13
Seasonal skill scores for ozone
Table 5 Correlation coefficients for daily
average and daily maximum O3.
  daily average daily average daily average daily average daily average daily maximum daily maximum daily maximum daily maximum daily maximum
  year DJF MAM JJA SON year DJF MAM JJA SON
EMEP 0.72 0.67 0.55 0.50 0.55 0.75 0.60 0.59 0.61 0.53
LOTOS 0.70 0.49 0.54 0.49 0.43 0.76 0.47 0.70 0.66 0.48
MATCH 0.80 0.68 0.66 0.60 0. 0.81 0.58 0.68 0.7 0.61
CHIMERE 0.76 0.62 0.58 0.64 0.60 0.84 0.62 0.71 0.77 0.62
RCG 0.71 0.58 0.59 0.52 0.36 0.76 0.56 0.70 0.61 0.44
DEHM 0.64 0.45 0.41 0.56 0.31 0.75 0.45 0.60 0.68 0.45
TM5 0.67 0.69 0.44 0.35 0.62 0.72 0.63 0.47 0.51 0.58
Ensemble 0.79 0.74 0.66 0.68 0.58 0.84 0.69 0.76 0.78 0.59
14
The skill of the ensemble mean

• Perfect ensemble Assume that the ensemble of K
  values xk is drawn from a distribution of
  physically possible states
• ? Then the observation xa has the same
  statistical properties than any member of the
  ensemble, and the RMSE of the ensemble average
  can be written
• b is the ensemble bias, s is the ensemble
  spread (standard deviation)
• ? The RMSE is a decreasing function of the
  number of members K
• ? The RMSE (ensemble skill) is linearly linked
  to the ensemble spread

,
15
Evaluation of uncertaintyConcepts and tools
borrowed from ensemble weather forecasting

• Reliability observation could be one of the
  members
• Observation compatible with predicted
  distribution
• Rank histogram count the times the rank is 1, 2,
  , n
• frequencies should be equal
• But predicted distributions can have no
  information content (random or climatological)
• Resolution the smaller the ensemble spread, the
  higher the resolution

16
Examples time series
Too large spread
Too small spread
17
Mean Rank Histograms
Stability of the ensemble
18
Reliability and Resolution
Resolution index Normalized spread
spread/stdev Reliability index (extreme counts
central counts) / total counts
19
Spread - Skill relation
20
CO2 Modelling TRANSCOMWork in progress

• CO2 modeling important for understanding and
  inverting fluxes
• TRANSCOM ensemble (Law et al., 2008) Evaluation
  of model ability to simulate CO2 at regional
  scale
• 2 Simulation Years 2002 and 2003
• 17 atmospheric models/model versions differing by
  resolution, input biospheric fluxes (2),
  anthropogenic CO2 fluxes (2)
• 6 monitoring sites from CARBOEUROPE-IP

21
Lack of spreadModel or/and data
representativeness problems?
22
Origin of ensemble spread and skill
23
Conclusions

• Develop methods to evaluate uncertainty
  prediction
• European ensemble displays relatively
  complementary aspects
• For ozone, poor resolution in Atlantic areas,
  poor reliability in complex terrain, balanced
  ensemble in Northern Europe.
• For NO2, poor reliability, for secondary
  inorganic aerosols reliable ensemble. For
  nitrate, poor reliability in gaz/solid balance.
• For CO2 model ensemble mean spread too small.
  Analysis coming soon.

24
European papers on evaluation and AQ model
ensembles(several missing, most probably!)

• Many individual model evaluations (to be
  reviewed)
• EUROTRAC reports
• Tilmes et al., 2002 Forecasts over 1 month of
  ozone in Germany
• Galmarini et al., 2004a,b 2007 (ENSEMBLE
  project, dispersion models, ETEX)
• EMEP review report Van Loon et al., 2004
• Vautard et al., 2007, AE (CityDelta project)
  City-Scale (5 EU cities, 1 year), eulerian
  approach
• Thunis et al., 2007, AE (CityDelta) Scenario
  ensembles at city scale
• Van Loon et al., 2007, AE (EuroDelta project)
  Regional scale, Eulerian, ozone, 1 year
• Vautard et al., 2006, GRL (EuroDelta, ozone)
  Ensemble uncertainty
• Schaap et al., 2008, AE (EuroDelta) PM10 and
  components evaluation