Environmental Engineering Reference
In-Depth Information
important to dispersion modeling simulations, effectively overcoming the challenge
to this evaluation paradigm. Similarly, MM5SCIPUFF software developed by the
Pennsylvania State University was used to directly couple MM5 to the SCIPUFF
model.
2.4. Statistical evaluation methodology
Another major focus of this study was to develop an evaluation framework which
provided a consistent, analytical method for determining how well modeled and
observed concentrations fit. Previous EPA evaluations of long range transport
models (USEPA, 1998) largely were subjective in nature and lacked a common
foundation for model performance comparisons. The statistical framework
developed for the ATMES-II experiment (Mosca et al., 1998), as implemented by
Draxler et al. (2001) was chosen as the basis for this evaluation. The metrics for
the global statistical analysis are divided into four broad categories: (1) scatter
correlation in Eq. 1, (2) bias fractional bias in Eq. 2, (3) spatial coverage (figure of
merit in space in Eq. 3), and (4) the unpaired distribution (Kolmogorov-Smirnov
parameter in Eq. 4).
(
) ( )
∑
M
−
M
•
P
−
P
i
i
(1)
R
=
i
(
)
( )
⎤
⎡
⎤
⎡
2
2
∑
∑
M
−
M
P
−
P
⎣
⎦
⎣
i
i
(
)
FB
= 2
B
P
+
M
(2)
A
∩
A
FMS
=
M
P
×
100
%
(3)
A
∪
A
M
P
( ) ( )
KS
=
Max
C
M
−
C
P
(4)
k
k
The final score, model rank (
RANK
), provides a combined measure to facilitate
model intercomparison (Eq. 5).
RANK
is the sum of four statistical measures for
scatter, bias, spatial coverage, and the unpaired distribution (Draxler et al., 2001).
RANK
scores range between 0 and 4 with 4 representing the best model ranking.
(
)
)
(
RANK
=
R
+
1
−
FB
/
2
+
FMS
/
100
+
1
−
KS
/
100
(5)
