**AIRMoN Technical Support Document
Comments**

Annapolis, MD, 24-27 October 1994

Roland R. Draxler

__Purpose__

To confirm as quickly as possible that a reduction in sulfur dioxide emissions results in an improvement of environmental quality as evidenced by a measured reduction of sulfur deposition.

__Problem__

The daily, seasonal, and even annual
variations *(give some numbers and references)* in
measurements of air quality and deposition are so large that the
small emissions reductions (10%) mandated by the Clean Air Act may
be obscured by the natural variability.

The primary component of the variability in the measurements is due to temporal variations in meteorological conditions (daily and seasonal cycles) as well as the more complicated meteorological variations that are introduced as a pollutant is transported from its source to the measurement location.

__Solution__

Although the chemical processes may be complex, meteorological variability can be modelled with sufficient accuracy to be able to use a model prediction to sort the measurement data into smaller data groups that have less variability within each group. Differences in measurement means between groups can the be explained with greater precision. For example measurements may indicate that over a 5 year period there was a 20% reduction in measured pollutant levels. However if the measurement data are segregated by upwind transport direction, one of the most simple model predictors, we may find that due to climatic reasons there was an slight increase over the 5 year period in the number of days that the airflow was from cleaner source regions. When pollutant measurements are compared for only those days when the flow was from polluted regions, we may find that there was no decrease overall pollutant levels. Essentially this is the methodology we propose to determine if Clean Air Act reductions are improving the environmental quality and why it is insufficient to rely only on measurement data to make this determination.

__Approaches__

The example given above was only a very
simple approach, in fact we plan to test several methods at
reducing the variance in grouped measurement data. These will
be called *Sectorial Sampling*, *Prediction
Differencing*, and *Source Fingerprinting*. Each
method differs in complexity and approach. At this stage we
do not know which is best. Regardless of which method is
selected, all approaches require the best and most detailed
meteorological data base from which these calculations can be
performed. In conjunction with NMC, we plan to archive the
analysis, forecast, and assimilation data fields from the most
advanced primitive equation models that are operationally
available, with a resolution of at least every 2h and 40km.
Additional meteorological measurements taken at the monitioring
sites will also be incorporated. Local meteorological
measurements may explain some of the variance introduced into the
measurements not evident from regional meteorological
data.

**Sectorial Sampling** - This is
the approach as noted in the example where it is presumed that
local concentration measurements will be higher when the airflow is
from pollutant source regions. Although these methods have
been used in many studies (Ref .... ) and have worked reasonably
well, it may not be sufficient to reduce the group measurement
variance. More objective methods, such as cluster analysis
(Ref ), can reduce the positional variance of the upwind
source region, over more simpler sectorial approaches, however
there may still be a significant contribution to the variance in
the measurements due to differences in meteorogical processes,
although the flow is still within the same sector. *(Are there
any references or numbers to indicate intra-sector
variability?)* This problem can be reduced to a large
extent by using a source-receptor matrix approach, in which
weighting factors (similar to a normalized concentration) are
computed for all the source-receptor combinations. These
weights are then used to group the measurement data. For
example comparing two different measurements, one high and one low,
although the model indicates are from the same source region, one
may have experienced significant precipitation prior to arrival at
the measurement site, and hence has a lower weight.

**Prediction Differencing** - In
this method we use a meteorological model, which includes a
chemical model component, to actually compute the air
concentrations and deposition corresponding to each
measurement. The model uses the best known emission inventory
in its calculation. The model emissions remain constant,
while the real world is subjected to the reduction. Will the
differences between the model prediction and measurements reflect
the emission reduction by showing a larger difference? The
time series of prediction-measured differences in theory will have
much of the meteorological variability removed. Therefore it
is only necessary to show that the mean of the differences prior to
reduction is significantly different than the mean of the
differences after the emission reduction. The following data
sample can be used to show that we need at least 225 independent
measurement points to reduce the standard error of the sample mean
to less than 0.05. However this is assuming that the standard
deviation will remain at 0.60 to show a pH reduction of 0.15.
As we accumulate more data points the sample deviation will go
up. This approach is simple in concept, however it needs
further investigation.

**Source Fingerprinting** - The above
approaches relied upon meteorological modeling techniques to group
measurement data. However there are other approaches that can
be used to aggregate time series and spatial observations.
The purpose of the groping remains the same, to reduce variance
within the group. One approach is the use Principal
Components to identify spatial and temporal coherence in the
measurements. The principal component approach can be used to
show the spatial elements that have the greatest contribution to
the measured variance. The pollutant source regions should
show a reduction to the variance contribution with time.
However it is not clear how many sampling stations are required for
this approach to prove satisfactory. A variation of this
option is develop time series analytical techniques at individual
stations (such as wavelet transforms) in conjunction with standard
meteorological measurements to deduce which meteorlogical variances
are contributing to the variance in the measurements. Another
independent method would be identify source regions through an
additonal conservative pollutant species, that is not subjected to
a reduction. Ratios of reduction/non-reduction pollutants in
conjunction with multiple regression methods can then be used to
provide a source region weighting factor to the measurements for
grouping.

Email me the full report