Data assimilation and air quality - Air Pollution Modeling and Its Application - page 280

are good at predicting extreme values. But the limitation is that they do not
include information about concentrations upstream, therefore they are often not so
good in predicting the beginning and end of a smog episode. We have investigated
the possibility to use the chemistry transport model LOTOS-EUROS to improve
on the statistical models.

2. Model Description

LOTOS-EUROS is a Eulerian regional model (10W-40E, 35N-70N, grid resolution
0.5 × 0.25 lon × lat with zooming possibility) with three dynamical vertical layers
and a surface layer, covering the lowest 3.5 km of the atmosphere. It uses an
emission database with simple time-dependency, ECMWF meteorological fields
and it includes chemistry for ozone formation (CBM-IV) and the formation of
secondary inorganic aerosols. A complete description can be found in Schaap
et al. (2008). Data assimilation with an ensemble Kalman filter (EnKF) is optional.

3. PM10 Forecasting

A test was done for 2004-2006, using LOTOS-EUROS on a zoom domain with
nesting. The concentrations of secondary inorganic aerosols are in good agreement
with observations, and also the concentrations of black carbon and sea salt aerosol
can be predicted reasonably well, which gives the model credit (Manders et al.,
2009). However, the model underestimated the total observed PM10 concentrations
substantially, like most CTMs. This is caused by components for which the source
processes are not well quantified, like mineral dust and secondary organic aerosols.
To compensate for this, a simple bias correction was proposed, which depends on
the season and the PM10 concentration:
PM10biascor = 1.54 * PM10 + 8.1R 2 = 0.52 Winter (DJF)
PM10biascor = 1.42 * PM10 + 7.5R 2 = 0.47 Spring (MAM)
PM10biascor = 0.76 * PM10 + 13.5R 2 = 0.26 Summer (JJA)
PM10biascor = 1.31 * PM10 + 9.1R 2 = 0.51 Fall

(SON)

The model results are shown in Fig. 1 and Table 1 , together with the observations
and the results form the statistical model. The bias-corrected LOTOS-EUROS has
a better time correlation than the statistical model, which is only marginally better
than a persistence model. The drawback is that LOTOS-EUROS is still under-
predicting the highest concentrations. This can be improved by using an alternative
bias correction forced through the origin, which enhances the extremes, albeit this
goes slightly at the expense of the general performance.