CONCENTRATIONS INSTANTANEOUS DENSE GAS DATA- 123docz.net

IGROl.JP 4 ) AND X 2. ;..:il Y

b. Instantaneous dense gas

from Thorney Island

Figure 12. Model performance measures, Geometric Mean Bias MG = exp(4nCo - tnC 1

predictions and observations at distances greater than or equal to 200 m. a: Continuous dense gas group of datasets (Burro, Coyote, Desert Tortoise, Goldfish, Maplin Sands, Thorney Island).

b: Instantaneous dense gas data from Thorney Island. 95 percent confidence intervals on MG are indicated. The solid line is the

"minimum VG' curve, from Equation (331. The dashed lines represent

"factor of two" agreement between mean predictions and observations.

P and Geometric Variance VG = exp[(hCo - &IC 1 2 I for concentration

132

--`,,-`-`,,`,,`,`,,`---

I A P I PUBL*454b 9 2 9 0 7 3 2 2 9 0 0 5 0 5 5 8 5 - _. 4 7 9 ~~

consistent with removing an overprediction tendency on the monitoring arcs in the near-field, where the measured concentrations may not represent peak concentrations, Analysis of the fractional bias in Figures 12a and 12b leads to the following conclusions for the dense gas data sets with the closest monitoring arc excluded:

Models that Overpredict by More Than a Factor of Two:

Continuous Release NONE

Instantaneous Release INPUFF, AFTOX

Models that Overpredict by about a Factor of Two:

Continuous Release NONE

Instantaneous Release

FOCUS, DEGADIS, TRACE, PHAST Models that Overpredict by Less Than a Factor of Two:

Continuous Release Instantaneous Release

FOCUS, TRACE BM

Models with Insignificant Overprediction or Underprediction:

I Continuous Reléase

DEGADIS, GASTAR

Instantaneous Release AIRTOX

Models that Underpredict by Less Than a Factor of Two:

Continuous Release Instantaneous Release PHAST, HEGADAC, BM, SLAB, AFTOX GASTAR

Models that Underpredict by about a Factor of Two:

Continuous Release Instantaneous Release

GPM, CHARM SLAB, CHARM

Models that Underpredict by More Than a Factor of Two:

Continuous Release Instantaneous Release OB/DG, AIRTOX, INPUFF NONE

The results for the geometric variance in Figures 12a and 12b are similar to those in Figures 10a and 11, since the only difference is the

133

--`,,-`-`,,`,,`,`,,`---

AP5 PUBL*qSqb 9 2 0732290 0 5 0 5 5 B b 305

removal of the monitoring arcs with x < 200 m. Vlth t h i s change. most variances were reduced slightly. The largest variance in Group 1 is still given by the AIRTOX model, and the largest variance in Group 4 is still given by the AFTOX nodel.

for Group 4 (Thorney island], although the AIRTOX nodel has "moved up" into one of the top three positions.

m e GASTAR and EM models still show good performance

e. Groups 1 and 4 (Dense Gas Releases), Distances < 200 m

In order t o assess the differences between model performance at far and near monitoring arcs, the data for x < 200 m are presented in Figure

13.

However, the observations may not indicate the true maximum concentration, because of inadequate horizontal and vertical resolution of the monitoring net work.

Any dense gas effects will be amplified at these close distances.

Comparing Figures 12 and 13, it is seen that there indeed are more cases of model overprediction at close distances. Because of the shifts in the points, some of the models (for example, SLAB, GPM, TRACE, CHARM,

AIRTOX) demonstrate improved performance at close distances f o r the continuous sources (parts a of the figures). Shifts also occur for the instantaneous sources (parts b of the figures), with the performance of some models (for, example, DEGADIS) deteriorating at the close distances, while the performance of other models (for example, PKASTI improves.

f. Groups 3 and 5: Passive Gas Releases

The statistics f o r the passive gas releases in Group 3 (continuous passive-gas releases) and Group 5 (instantaneous passive-gas releases) are tabulated in Appendices D-6 and D-7, and the plots of geometric mean bias MG versus geometric variance VG are shown in Figures 14a and 14b, respectively. Note that statistics for the continuous releases are dominated by the Prairie Grass datacet, while those for instantaneous releases are derived solely from the Hanford dataset.

The confidence limits on the geometric mean bias, MG, for the continuous releases of passive gases shown in Figure 14a are small, because

--`,,-`-`,,`,,`,`,,`---

A P I P U B L * 4 5 4 b 7 2 0 7 3 2 2 7 0 0 5 0 5 5 8 7 2 4 1

$ 8

CONCENTRATIONS

CONTINUOUS DENSE GAS DATA WITH

SHORT AVERAGING TIME (GROUP 1 ) AND X < 200 M

D E W I S PHhl \ b

o.us o 7 5 o 5 1 2 4 8

(0vERPREI)IcTION) O E R P m I m O N )

a. Continuous dense gas group datasets (Burro, Coyote, Desert Tortoise, Goldfish, Maplin Sands, Thorney Island)

3 8

CONCENTRATIONS

INSTANTANEOUS DENSE GAS DATA (GílOUP 4 ) AND

X < 200 M

\ I i DECADIS l I

0.125 02.5 0 5 1 2 4 8 MG

(OYERPREDICTIOM (UNDERPREDICllON)

b. Instantaneous dense gas dataset from Thorney Island

Figure 13. Model performance measures, Geometric Mean Bias MG = exp(hC - LnC

O P

and Geometric Variance VG = exp[(hCo - h C 1 2 1 for concentration P

predictions and observations at distances less than 200 m.

a: Continuous dense gas group of datasets (Burro, Coyote, Desert Tortoise, Goldfish, Maplin Sands, Thorney Island). b:

Instantaneous dense gas data from Thorney Island.

confidence intervals on MG are indicated. The solid line is the

"minimum VG" curve, from Equation (33). The dashed lines

represent "factor of two" agreement between mean predictions and observations.

95 percent

135

--`,,-`-`,,`,,`,`,,`---

APT P U B L * 4 5 4 6 9 2 0732290 05055ôô 188

g s

CONCENT3AT I ON (GROUP 3 )

CONTINUOUS PASSIVE RELEASES

0.125 o z 0.5 I 2 4 8

(0vERPREDIcTx0N) 0 E R P R A ) I c r r O N ) MO

a. Continuous passive gas group of datasets (Prairie Grass and Hanford-continuous)

Figure 14. Model performance measures,

g s

C O N r n ~ A T I O N

INSTANTANEOUS PASSIVE RELEASES (CROUP 5 1

b. Instantaneous passive gas dataset from Hanford

Geometric Mean Bias MG = exp(.!nCo - tnC 1

- P

and Geometric Variance VG = e x p [ ( h C o - dnC 1 2 I for concentration predictions and observations. a: Continuous passive gas group P of datasets (Prairie Grass and Hanford-continuous). b: Instantaneous passive gas dataset from Hanford. 95 percent confidence intervals on MG are indicated.

from Equation ( 3 3 ) .

agreement between mean predictions and observations.

The solid line is the "minimum VG" curve, The dashed lines represent "factor of two"

136

--`,,-`-`,,`,,`,`,,`---

- - _ _ __--- _ _ _ _ _ - A P I P U B L * q 5 4 6 9 2 0732290 0 5 0 5 5 8 7 014

the Prairie Grass dataset provides many aata-?oints. The GASTAR, TRACE, and CHARM models have relatively large variances. The geometric mean biases for Group 3 can be summarized as follows:

Models that Overpredict by about a Factor of Two:

TRACE, GACTAR

Models that Overpredict by Less Than a Factor of Two:

AFTOX, DEGADIS, HEGADAS, SLAB

Models with No Significant Overprediction or Underprediction:

INPUFF, GPM, OBM;, PHAST, FOCüS, AIRTOX

The models with the lowest variance (VG - 1.5) for Group 3 are the HEGADAS and SLAB models.

slightly less than the mean value. The good performance of the HEGADAS model is surprising and probably fortuitous since that model is being initialized assuming a small area source, whereas the actual release was a small point source. A group of other models (AIRTOX, DEGADIS, OBDG, FOCUS, GPM, INPUFF, PHACT, and AFTOX) have relatively low VG values in the range from about 1.6 to 2.2, indicating that their scatter is approximately equal to the mean.

The magnitude of the scatter for these models is

The Hanford dataset (Group 5 ) in Figure 14b has few numbers, leading to a large span in 95 percent confidence limits for the geometric mean bias, MG. Even so, all of the models tend to overpredict the peak

concentrations on average. The GASTAR, AIRTOX, PHAST. INPUT, and CHARM

models have the best performance, with mean overpredictions of about 10 to 50 percent and scatters approximately equal to the mean. The TRACE model is unique in its very large degree of overprediction.

g. Analysis of Differences among Models

Up to this point we have characterized the tendency of each model to either overpredict or underpredict peak concentrations, based on the statistical measure, but we have not selected a “best“ model. One way to characterize a “best” g r o w of models is to identify the models with the

137

--`,,-`-`,,`,,`,`,,`---

A P I PUBL*UCYb 92 073Z290 0505sqo 83b

smallest mean bias and the smallest scatter, and then ask the question: which other models have a bias or a scatter wnich is not significantly different from that cf the "best" model? 'ihe answer provides one basis for defining the

"best" group of models.

Appendix D-8 contains tabulations showing whether or not the difference in the geometric variance between pairs of models is significantly different from zero, at the 95 percent confidence level. Consider first the results for the continuous releases of dense gas shown in Figure 12a (Group 1, for distances greater than 200 rnl.

performance, but we see that its variance is not significantly different from the variance for HEGADAS.

found for GASTAR & significantly different from and closer to zero than the bias for the HEGADAS model, although there is no difference between the biases of the GASTAR and DEGADIS models. We conclude that, in general, this group of three models does a better job than the others of matching the peak observed concentrations at distances of 200 m or greater for continuous releases of dense gases.

GASTAR appears to have the best overall However, we see that the geometric mean bias MG

A summary of model performance for the better performing models at distances greater than 200 rn is given in Table 13 for Data Groups 1, 3, 4, and 5 . There are no models that appear on the list of better models for all four data groups.

h. Analysis of Model Performance for Stable Ambient Conditions Another facet of model performance that can be evaluated with these data is the question of how the models perform for the subset of the dense-gas data for which the atmospheric stability class is either E or F

(that is, stable ambient conditions). Because "worst-case" dispersion conditions are usually found for these stabilities, many model applications focus on these stable ambient conditions.

geometric mean bias (MGI and geometric variance (VGI results, and the statistics themselves are tabulated in Appendices D-9 and D-10 for the continuous and passive dense-gas releases, respectively.

Figures lCa and 15b show the

138

--`,,-`-`,,`,,`,`,,`---

I O O N

A VI u u ổ

I- VI

a n -

I- a

VI ổ O I- <

I-

ổ w u I:

Y U u1

œ 2

u) I- 2

w œ

a a ổ

I m

' O

-1 m c

L u)

œ -

; r 9

a w

w w o

2 L

al >

i O n 5

k1 d

Y a x al

+ ?

n m

m n -

.- Y

(u a

c m m .

u m

m o

1 0 O N C A I

u x c -

+ U VI

I- õ

VI a

ul m

.- A x O

o - o .

? -

.C Y

n N

m r

m a

m u

u o -

a n u al

a5 7 m - I ul

$ 2 0 0 . - al x N m .

0 2 2 '

- E a m

m e

O C a l

E A l

C O Y

Y a = - - Y

Y b

C O

E - u

u m

u 4 u 2

E V I

? .-

n v)

.- a o O c

9 -

C .

- Y

139

m m w o * O : .

.- Y 9

m n

m - r

u -

1 0

o n 0

C L

(o o II Y u -

c c x

a a u

Ym Eu a l u m

- C

- 0

o r

!+G o 3 L o -

m a

m u

c >

a o

2 -

.- z u

c c

--`,,-`-`,,`,,`,`,,`---

A P I PUBL*lt5Yb 92 = 0732290 050559i! b 0 ợ =

CONCENTRATIONS

CONTINUOUS DENSE GAS DATA WITH SHORT A E R A G I N G TIME (GROW 1) X 2 200 M,

AND STABLE AMBIENT CONDITIONS

0.m o 2 5 os 1 2 1 8

ìvlû

( O Y E R P ~ I C I I O N ) (UNDERPrnIcIION)

a. Continuous dense gas group of datasets (Burro, Coyote, Desert Tortoise, Goldfish, Maplin Sands, Thorney Island)

CONCl3U'ủ.ATIONS

INSTANTANEOUS DENCE GAS DATA (GWJF 41 X 2 200 M, AND STABLE

AMBIENT CONDITIONS

3 s

1 I I

b. Instantaneous dense gas data from Thorney Island

Figure 15. Model performance measures, Geometric Mean Bias MG = exp(h Co - tnC 1

CONCENTRATIONS INSTANTANEOUS DENSE GAS DATA

ACMR (Anaiysis of Consequences of Toxic Releases)

SECTION VI SENSITIVITY ANALYSIS USING MONTE CXRLO PRCEDURES