Professional Documents
Culture Documents
O
`LJ
#b
`
CC#b
`
DOW#b
"
LOD
#b
`
NC#b
"
RF#b
`
TRAJ (1)
The dependent variable, O
`
, is the 1-h domain peak
ozone. The "rst explanatory variable, O
`
, is the predic-
tion from the nonlinear regression equation (Eq. (2)),
which is a function of temperature, surface wind speed,
and relative humidity.
O
`
"(
#(
`
#
`
TMAX#
"
TMAX`)
;exp(
`
WS))exp(
"
RH) (2)
W.G. Cobourn, M.C. Hubbard / Atmospheric Environment 33 (1999) 4663}4674 4667
Table 1
Predictors used in the regression equations
Predictor Units Symbol Timing
Air mass trajectory corridor * TRAJ Upwind, previous 36 h
Cloud cover tenths CC 9 a.m.}2 p.m. (average)
Day of week * DOW Saturdays
Length of day hours LOD Sunrise to sunset
Maximum temperature 3F TMAX Daily 24-h peak
Minimum temperature 3F TMIN Daily 24-h min.
Nighttime calms * NC 12 midnight}4 a.m.
Nonlinear model prediction ppb O
`
*
Rainfall in RF Daily 24-h total
Relative humidity % RH 9 a.m.}1 p.m. (average)
Surface wind speed mph WS 9 a.m.}3 p.m. (average)
Table 2
Regression coe$cients for the Standard model
Coe$cient Fitted value S.E. t-statistic
Constant !43.7 7.55 !5.79
O
`
0.800 0.0310 25.83
CC !0.732 0.199 !3.67
DOW 4.14 1.29 3.21
LOD 4.16 0.530 7.85
NC 1.55 0.566 2.74
RF !2.29 1.37 !1.67
TRAJ 11.3 1.24 9.06
Table 3
Regression coe$cients for the nonlinear regression model
Coe$cient Fitted value S.E. t-statistic
`
181 127 1.42
`
!9.26 3.09 !3.00
"
0.0933 0.0193 4.83
`
!0.115 0.0124 !9.32
"
!0.00654 0.000703 !9.30
The remaining explanatory variables } cloud cover, day
of week, length of day, number of calms, rainfall, and
trajectory corridor location, are summarized in Table 1.
The "tted coe$cients for the Standard model are shown
in Table 2 along with the standard errors and t-statistic.
Except for the rainfall variable, the t-values exceeded 2.7
in absolute value, indicating that all the explanatory
variables contributed signi"cantly to the regression. The
term representing the nonlinear regression was by far
the strongest contributor, with a t-value exceeding 25.
The overall explained variance (R`) was 0.790, the mean
absolute error (MAE) was 8.45 ppb, and the root mean
square error (RMSE) was 10.58 ppb.
In order to optimize the nonlinear model, many vari-
able transformations and groupings of terms were evalu-
ated. The best results were obtained by using a second
order transformation of temperature and by applying an
exponential transformation to wind speed and relative
humidity separately (Eq. (2)). The "tted coe$cients for
the nonlinear parameters are shown in Table 3 along
with the standard errors and t-statistic. The accuracy of
the nonlinear regression model is quite good considering
that only three explanatory variables are used. The ex-
plained variance, MAE, and RMSE are 0.724, 9.52 ppb,
and 12.12 ppb, respectively. This level of performance is
su$ciently good that the nonlinear model by itself could
be used as the primary forecast model for those commu-
nities that lack access to the other predictor variables.
The Hi-lo model was "t to the days with peak ozone
concentrations below 50 ppb and above 105 ppb using
the same methods and explanatory variables as for the
Standard model, except that the variables for 24-h rain-
fall and the number of overnight calms were removed
from the regression for lack of statistical signi"cance.
Though the sample size was much smaller (162 vs. 580
days) the retained regression variables were statistically
signi"cant.
3.3. Comparison with previous models
To demonstrate the improvements realized in the
latest generation of regression models resulting from
the expanded nonlinear regression term and inclusion
of the trajectory parameter, a stepwise comparison of
models is presented in this section.
The forecast model used during the 1997 forecasting
season (`Standard Ia)
>"
b
"
#b
TMAX#b
`
TMAX`#b
`
TMAX"
#b
"
exp(WS)#b
`
AT
#b
"
TMIN#b
`
CC#b
`
RF
#b
"
NC#b
"
DOW
`
(3)
4668 W.G. Cobourn, M.C. Hubbard / Atmospheric Environment 33 (1999) 4663}4674
Table 4
Comparison of mean absolute error, root mean square error,
and explained variance of various models for the period
1993}1997 (580 days)
Forecast model MAE
(ppb)
RMSE
(ppb)
R`
Hybrid II 8.35 10.50 0.844
Standard II 8.45 10.58 0.790
Standard II (w/o traj.) 8.99 11.31 0.760
Standard I (w/o traj.) 9.50 11.97 0.729
Fig. 3. Scatterplot of Hybrid II model hindcasts against obser-
vations for the period 1993}1997. The diagonal indicates the line
of perfect correspondence between hindcasts and observations.
di!ers from the 1998 Standard model (`Standard IIa) in
several ways: (1) The use of nonlinear regression was
limited to the wind speed term only. (2) Maximum tem-
perature was treated as a linear variable and was trans-
formed up to the fourth order versus second order for
Standard II. (3) The e!ects of humidity were captured via
minimum temperature instead of relative humidity. (In
e!ect, minimum temperature served as a proxy for the
dewpoint, but actually gave better results than dewpoint
with the Louisville data set.) (4) The proxy term for the
seasonal cycle was theoretical atmospheric transmittance
instead of length of day. (5) Standard I bene"ted from
a square-root transformation of the dependent variable
in order to stabilize variance. Standard II did not bene"t
from any stabilization transformation probably owing to
the greater use of nonlinear regression which better cap-
tured interactions between the variables. Finally, Stan-
dard I lacked the air mass trajectory term. When we re"t
the Standard II model to exclude the trajectory term, but
keeping everything else the same and using the same days
to "t both models, the Standard II (w/o traj.) had 5.4%
lower mean absolute error (8.99 vs. 9.50 ppb), and 4.3%
greater explained variance (0.760 vs. 0.729), as compared
to Standard I. The bulk of this improvement is attributed
to incorporating temperature, wind speed, and relative
humidity into a single nonlinear regression term. Inclu-
sion of the air mass trajectory provided similar improve-
ments. The mean absolute error decreased an additional
6.0% (8.45 vs. 8.99 ppb), and the explained variance
increased an additional 3.9% (0.790 vs. 0.760). To sum-
marize (see Table 4), the Standard II regression had
11.1% less mean absolute error and 8.4% greater ex-
plained variance than Standard I but used only one
additional explanatory variable.
3.4. Hybrid model
By using the 3S criteria to identify the probable high
ozone days, a strategy can be employed to switch in an
optimal way between the Standard model to the Hi-lo
model to improve forecast performance beyond what
could be achieved using either model alone. Combining
the Standard II and Hi-lo II models in this way gives rise
to the Hybrid II model. The following 3S criteria were
used to invoke the Hi-lo II model: (1) expected maximum
temperature greater than 87 3F, (2) expected wind speed
less than 6 mph, and (3) expected cloud cover less than 2.5
tenths. Relative to the Standard II model, this switching
strategy reduced the mean absolute error by 1.2% (8.35
vs. 8.45 ppb) and increased the explained variance by
6.8% (0.844 vs. 0.790). Inspection of the scatterplot of
ozone observations vs. Hybrid II hindcasts (Fig. 3) con-
"rms that the switching does not introduce systematic
overprediction or underprediction error anywhere in the
range: the scatter of points coalesces along a straight line
having a slope of 1 and an intercept of 0.
4. Prediction of high ozone concentrations
It stands to reason that on the few days of extremely
high concentration during the ozone season, all of the
factors which contribute to high ozone must be at or near
to their extreme values for the season. Good regression
models can account for a substantial portion of the daily
ozone variation, but the models cannot account for the
e!ects of factors not included in the model. Conse-
quently, regression models (and other types of statistical
models) will tend to underpredict on days when all ozone
related factors are nearly maximal. The prediction errors
tend to be substantially negative on these days, due to the
strong e!ect of the unexplained factors. There are at least
two possible means of improving the prediction errors on
high ozone days. First, if some separate method (indepen-
dent of the principal regression model) can be found to
identify probable high ozone days, then an adjunct re-
gression equation can be developed which is specially
"tted for high ozone. This is the hybrid model approach.
W.G. Cobourn, M.C. Hubbard / Atmospheric Environment 33 (1999) 4663}4674 4669
Fig. 4. Distribution of ranked model hindcast errors on days
during the 1993}1997 period when observed [O
`
]*120 ppb, by
model.
Table 5
Comparison of mean absolute error, root mean square error,
and explained variance of various models for the 3S-days (77
days) during the 1993}1997 period
Forecast model MAE
(ppb)
RMSE
(ppb)
R`
Hybrid II 9.07 11.30 0.633
Standard II 9.84 11.85 0.458
Standard II (w/o traj.) 11.67 14.15 0.362
Standard I (w/o traj.) 12.25 14.99 0.318
Second, if an additional ozone-related parameter (which
is su$ciently independent from the existing ones) can be
found, then this parameter can be added to the regression
in order to reduce the amount of unexplained variance.
Adding the trajectory parameter follows this approach.
This last development has been particularly useful, since
there is a strong association between high ozone days,
and the `in-corridora back trajectories. For example,
during 1993}1996, on the days where [O
`
]*120 ppb,
93% were `in-corridora days. By contrast, of the days
where [O
`
](60 ppb, only 10% were `in-corridora days.
These developments have made it possible to formu-
late a hybrid regression model which is much more
accurate in forecasting extreme ozone events than the
previously reported regression model. In most cases,
very little subjective adjustment to the model forecast is
necessary. Subjective or `experta skill is still required,
however, in evaluating all relevant meteorological in-
formation to make the best forecasts of the meteorologi-
cal input parameters. With the current hybrid model, the
high-end errors have been signi"cantly reduced, as com-
pared to previous models (Fig. 4). The graph shows the
successive improvements achieved by the updated regres-
sion equation (Standard II w/o traj.), the trajectory para-
meter (Standard II), and the hybrid technique (Hybrid
II). Based on the retrospective predictions, or `hindcastsa
of the 1993}1997 model calibration period, using ar-
chived meteorological data for the model input, the
Hybrid II model predicted 63% of the [O
`
]*120 ppb
days to within 15 ppb; and the MAE for those days was
12.1 ppb. In contrast, the Standard II model without
the trajectory parameter predicted only 36% of the
[O
`
]*120 ppb days to within 15 ppb, and the corre-
sponding MAE was 18.1 ppb. For the Standard I model,
re"tted to the 1993}1997 data, these statistics were re-
spectively 21% and 21.7 ppb. For the 77 days during the
1993}1997 period which were meteorologically prone to
high ozone, according to the `3Sa criteria, the Hybrid II
model also performed signi"cantly better than the other
models discussed above (Table 5).
The capability of predicting concentrations exceeding
the air quality standard, or other relevant threshold
levels, for purposes of issuing health advisories, etc., is
a critical aspect of OAD programs. The `detection ratea
(DR), de"ned as the fraction of observed threshold ex-
ceedances predicted by the model, is a useful indicator of
the model performance in predicting high threshold
values, such as the 1-h, 120 ppb standard still in e!ect in
the Louisville MSA as of this writing. The DR is also
referred to as the `probability of detectiona (e.g. Ryan,
1995). In simple terms, an event is detected if both the
model and observed values exceed the threshold. This
could be called the `zero tolerancea DR, because obser-
vations just over the threshold (e.g. by 1 ppb) would be
regarded as undetected, or `misseda, even if the predic-
tion was very close, but just under the threshold. It is
possible to assign a `tolerancea to the DR computation,
wherein exceedances would be regarded as detected as
long as the predictions exceeded an `alarm levela de"ned
as the threshold minus a certain tolerance (e.g. 15 ppb).
This scheme is rational in that it takes into account the
model imperfections and tendency to underpredict the
extreme values.
Another useful indicator of model performance is the
false alarm rate (FAR), nominally de"ned as the fraction
of predicted threshold exceedances in which the observed
values are below the threshold. It is appropriate to use
the same tolerance discussed above for computation of
the FAR, so that predicted exceedances would not be
counted as false alarms as long as the observed concen-
tration exceeded the alarm level.
Using the 1-h, 120 ppb NAAQS as the currently ap-
plicable threshold for Louisville, the theoretical DR and
FAR values were computed for the 1993}1997 model
calibration period. Both indicators were good to excel-
lent for the Hybrid II model, depending on the chosen
4670 W.G. Cobourn, M.C. Hubbard / Atmospheric Environment 33 (1999) 4663}4674
Table 6
Theoretical detection rates and false alarm rates for the Hybrid
II and Standard II (w/o traj.) models, based on a 120 ppb
threshold, for various tolerances during the 1993}1997 period
Tolerance
(ppb)
Hybrid II Std. II (w/o traj.)
DR (%) FAR (%) DR (%) FAR (%)
0 53 35 18 50
5 64 31 39 48
10 75 29 57 33
15 96 26 68 30
tolerance (Table 6). For example, at a tolerance of 15 ppb,
corresponding roughly to the 85th percentile of the
prediction errors for the period, the DR was 96%, and
the FAR was 26%. The DR and FAR indicators for the
Hybrid II model were signi"cantly better than those of
the Standard II model without the trajectory parameter
(Table 6), at several tolerance levels. These statistics are
based on hindcasts made for the 1993}1997 calibration
period, and so are theoretical values. In the forecast
mode, the values would be expected to di!er somewhat
(lower for the DR and higher for the FAR) for two
reasons. First, because the model input is meteorological
forecast data, and there are errors in these forecasts.
Second, because in the forecast mode the model is
being tested against new data, and the e!ect of the unex-
plained factors contributing to ozone may vary from year
to year.
The tolerance that leads to an acceptably high DR and
acceptably low FAR can be used to establish the alarm
level for issuing OADs in the community. Thus, if the
ozone forecast would equal or exceed the threshold
minus the tolerance value, the forecaster would recom-
mend that an ozone action day be called. During the
1998 ozone forecast season in Louisville, a 15 ppb toler-
ance was chosen, so OADs were recommended on days
when the model forecasts exceeded 105 ppb.
5. Model validation
5.1. Daily predictions for the 1998 ozone season
The Hybrid II model was deployed by the Air Pollu-
tion Control District of Je!erson County (APCDJC) in
cooperation with the Indiana Department of Environ-
mental Management (IDEM), during the period 14 May
to 16 September 1998. The o$cial forecasts for each day
were prepared by IDEMmeteorologists. Since these fore-
casts were geared primarily towards the identi"cation of
the days appropriate for designation as ozone action
days, here they are referred to as the `ozone actiona,
or OA forecasts. Normally, these were prepared each
Monday, Wednesday and Friday. Some OA forecasts
were updated on Tuesdays and Thursdays if high ozone
was expected. Thus, less than half of the OA forecasts
were next day forecasts, a shortcoming which was due to
sta$ng limitations. A second set of forecasts, called `raw
forecastsa, was prepared at the university. The raw ozone
forecasts were all next day forecasts, and were generated
by directly entering the unadjusted meteorological fore-
cast data obtained from the NGM numerical weather
prediction model (the NGM MOS output is available
twice daily from several internet sites) and the trajectory
parameter as determined by the HYSPLIT trajectory
model. In addition, the hindcasts from the Hybrid II
model were prepared and archived once the recorded
meteorological data became available from the National
Weather Service (NWS). The observed daily 1-h maxima
of ground level ozone were provided by the APCD.
The raw forecasts tracked the day-to-day ozone vari-
ation reasonably well (Fig. 5). Some of the high peaks
(e.g. [O
`
]'120 ppb) were seriously missed by the raw
forecasts, however. The o$cial forecasts were more accu-
rate on these days, particularly when next day forecasts
were made. The hindcasts of the high peaks were also
more accurate than those of the raw forecast. One prob-
lem with forecasting the high peaks is that in this realm,
the model predictions are much more sensitive to small
changes in certain meteorological parameters such as
temperature and wind speed. Therefore, careful review of
all pertinent meteorological information, taking into ac-
count expected errors in the raw forecast information, is
crucial to the process. Such expert skill requires experi-
ence with both meteorological and ozone forecasting,
and can be very e!ective in improving the next-day
forecasts when relatively high ozone concentrations are
expected.
The overall statistical comparison of the 1998 hind-
casts, raw forecasts, and OA forecasts is presented in
Table 7. The hindcast MAE was 11.0 ppb, which is about
2.7 ppb higher than the value from the hindcasts of the
1993}1997 model calibration period. This deviation was
higher than expected, since experience with using the
original model on the 1997 ozone forecasts led us to
expect only about 1 ppb added to the hindcast MAE. The
hindcast average error, or bias was !6.8 ppb, whereas
from the 1997 results we had anticipated a near zero bias.
However, the original model exhibited similar increases
in the MAE and bias in calculating a set of hindcasts for
1998. Part of the explanation for these anomalies is that
some is due to normal year-to-year statistical variation.
For example, the annual MAEvaried from 7.0 to 9.1, and
the annual bias varied from!1.3 to 1.0 for the hindcasts
of the 1993}1997 model calibration period. It is possible
also that uncharacteristic di!erences in meteorology and
ozone climatology between 1998 and the 1993}1997 cal-
ibration period played an important role. We note that
for Louisville, May and September of 1998 were much
W.G. Cobourn, M.C. Hubbard / Atmospheric Environment 33 (1999) 4663}4674 4671
Fig. 5. Time series of observed and predicted daily maximum 1-h ozone concentration for the raw forecasts during the 1998 ozone
season.
Table 7
Error statistics for the 1998 daily 1-h maximum ozone hindcasts, raw ozone forecasts, Ozone Action forecasts, and next day Ozone
Action forecasts
Index Hindcast Raw forecast OA forecast 1-d OA forecast
90th pct. (ppb) 7.8 21.8 27.0 31.1
10th pct. (ppb) !22.6 !20.3 !18.8 !22.1
Average (ppb) !6.8 0.9 4.1 3.4
MAE (ppb) 11.0 13.4 15.4 15.2
Table 8
Detection rates and false alarm rates for the 1998 daily 1-h
maximum ozone hindcasts, raw forecasts, ozone action forecasts
and next day ozone action forecasts
Index Hindcast Raw
forecast
OA
forecast
1-d OA
forecast
DR 1.00 0.40 0.60 1.00
FAR 0.20 0.50 0.47 0.33
warmer and drier than usual, and July was cooler and
rainier, as compared to the model calibration period.
The MAE of the raw forecast, at 13.4 ppb, was slightly
higher than that of the hindcast. The bias, at 0.9 ppb, was
actually a signi"cant improvement. This `boosta may
have come from the fact that for our area, the NGM
model tends to over-predict peak temperatures during
the summer months. The MAE of the OA forecasts was
greater than that of the raw forecasts by about 2 ppb, and
the bias was higher by about 3 ppb. This was mostly due
to the fact that the forecasters' "rst priority was to ident-
ify and recommend OADs, rather than to minimize the
overall forecast error. Therefore, uncertainties in the vari-
ous forecasted meteorological parameters were typically
hedged in the direction of higher predicted ozone, in
order to avoid a `missed exceedance.a The error statistics
for the next day OA forecasts (2-d and 3-d forecasts
removed) were only slightly improved. Probably, the
aforementioned `miss avoidancea technique contributed
strongly to the overall error statistics of the next day OA
forecasts as well. Nevertheless, the technique was success-
ful when measured against its objective, since OADs were
called on all days exceeding the 120 ppb threshold for
which next day forecasts were available.
5.2. Predictions of high ozone in 1998
The detection rate (DR) for the 1998 hindcasts, based
upon the nominal threshold value of 120 ppb, and a 15
ppb tolerance, was 5/5, or 100% (Table 8). The false
alarm rate (FA) was 2/10, or 20%. These values are close
to the expected values of 96% and 26% from the calib-
ration hindcasts. For the raw forecasts, the correspond-
ing DR was 2/5, or 40%, and the FAR was 6/12, or 50%.
The disparity in these "gures is due to the critical rela-
tionship between ozone and meteorological factors on
the days when conditions conducive to high ozone exist.
4672 W.G. Cobourn, M.C. Hubbard / Atmospheric Environment 33 (1999) 4663}4674
This makes the meteorological forecast errors more im-
portant on these critical days. For logistical reasons
associated with disseminating an ozone action alert for
the next day, the forecasts must be issued by 11 a.m. This
time is usually just prior to the release of the next up-
dated numerical weather model output, which is the 1200
UTC (8 a.m. EDT) model initialization run. As a result,
the available output is the 0000 UTC run, which is
about 15 h old at the time of the forecast. Since ozone
normally peaks near 3 p.m. EDT, the next day forecasts
are actually for about 42 h ahead. A skilled ozone
forecaster can take the critical ozone relationships and
forecast uncertainty into account, as well as departures
between the raw model forecasts and other expert fore-
casts, e.g. from NWS. Other information, such as the
location and strength of high pressure systems may also
be taken into account. Thus, the DR for the next day OA
forecasts was 3/3, or 100%, and the FAR was 4/12 or
33%. These indicators compare well to those of the
calibration hindcasts. When the 2-d and 3-d forecasts are
added in, the results are less impressive: 3/5 (60%) for
the DR and 8/17 (47%) for the FAR. Of the two missed
forecasts, one was for two days ahead, and the other was
for three days ahead. Both occurred in middle September,
near the end of the ozone forecast season. It appears
doubtful that forecasts of high ozone events beyond a next
day outlook can be made with any degree of reliability.
5.3. Application to forecasts of 8-h daily maximum ozone
An optimized version of an 8-h ozone forecast model
would involve re-examining all possible meteorological
input parameters, using various averaging periods for the
data, and combining these parameters into several di!er-
ent formulations with the aim of minimizing the model
error. As a prelude to this major undertaking, a rudimen-
tary 8-h forecast model was developed by using a correla-
tion equation to convert the 1-h forecasted ozone values
to equivalent 8-h forecasts. The correlation equation was
a second order polynomial, "tted to the 1-h and 8-h ozone
daily maxima from the 1993}1997 calibration period. The
purpose was to test the feasibility of using the same or
similar model formulation for an 8-h forecast model, and
to arrive at an upper bound estimate of the model MAE.
The resulting preliminary 8-h model predicted the 8-h
ozone daily maxima reasonably well. The time series of
predicted and observed values appeared much the same
as those in Fig. 5, when rescaled appropriately. As for the
overall forecast statistics, the MAEs for the converted
hindcasts, raw forecasts, and OA forecasts were respec-
tively 10.1, 10.8 and 11.9 ppb. The DR values, based on
the nominal threshold of 80 ppb, and using a 15 ppb
tolerance were respectively 84, 92 and 89%. Of course,
there were a far greater number of threshold exceedances
for the 8-h, 80 ppb threshold: 38 as compared to 5 for the
1-h, 120 ppb threshold. The corresponding FAR values
were 22, 24 and 29%. These results seem rather good for
such a crude estimation procedure. Thus, it seems likely
that it is feasible to develop an 8-h forecast model which
is at least as accurate as the 1-h forecast model.
6. Conclusions
The Hybrid II model was signi"cantly more accurate
than the original Standard I model for predicting 1-h
maximum daily ozone for the Louisville area during the
1993}1997 model calibration period. The improvement
was even more dramatic at the higher end of the ozone
distribution. The improved accuracy is attributed to
a more advanced regression equation, use of the hybrid
concept, and the addition of the trajectory parameter.
The Hybrid II model was "eld tested during the 1998
ozone season, and it performed reasonably well.
When the model was used by ozone forecasters during
the 1998 ozone season to predict the days that were
expected to approach or exceed the current 1-h ozone
standard, the model performed well when the forecasts
were next day forecasts. When the forecasts were made
two or three days out, the model did poorly, due to the
added uncertainty in the meteorological forecasts. It ap-
pears feasible to use the same or similar formulation for
an 8-h ozone forecast model. The accuracy would be
expected to be about the same, or slightly better than for
the 1-h model described herein.
Acknowledgements
This work was partially supported by the Air Pollu-
tion Control District of Je!erson County (APCDJC).
The authors wish to thank Art Williams, Arthur Chang,
Cynthia Lee, and other APCDJC o$cials for their en-
couragement and cooperation. LGE Energy Corporation
and the University of Louisville Research Foundation also
provided support. We wish also to thank meteorologists
Steve Sherman, John Welch, Mark Derf, and Mark Ney-
man of the Indiana Department of Environmental Man-
agement for their cooperation and technical assistance
throughout the model development and evaluation.
Graduate students Becky Hall, Richard Decker and
Chutikoon Gajaseni faithfully performed many data col-
lection and analysis chores required for this project.
References
Ashbaugh, L.L., Malm, W.C., Sadeh, W.Z., 1985. A residence
time probability analysis of sulfur concentrations at Grand
Canyon National Park. Atmospheric Environment 19,
1263}1270.
Bloom"eld, P., Royle, J.A., Steinburg, L.J., Yang, Q., 1996.
Accounting for meteorological e!ects in measuring urban
W.G. Cobourn, M.C. Hubbard / Atmospheric Environment 33 (1999) 4663}4674 4673
ozone levels and trends. Atmospheric Environment 30,
3067}3077.
Comrie, A.C., 1994. Tracking ozone: air-mass trajectories and
pollutant source regions in#uencing ozone in Pennsylvania
forests. Annals of the Association of American Geographers
84, 635}651.
Comrie, A.C., 1997. Comparing neural networks and regression
models for ozone forecasting. Journal of the Air and Waste
Management Association 47, 653}663.
Draxler, R.A., 1991. The accuracy of trajectories during
ANATEX calculated using dynamic model analysis versus
rawindsonde observations. Journal of Applied Meteorology
30, 1446}1467.
Holzworth, G.C., 1967. Mixing depths, wind speeds and air
pollution potential for selected locations in the United
States. Journal of the American Meteorological Association,
1039}1044.
Hubbard, M.C., Cobourn, W.G., 1998. Development of a regres-
sion model to forecast ground-level ozone concentration in
Louisville, KY. Atmospheric Environment 32, 2637}2647.
IMSL, 1992. FORTRAN Subroutines for Statistical Analysis.
User's manual-version 3.0. IMSL, Houston, TX, USA.
Moy, L.A., Dickerson, R.R., Ryan, W.F., 1994. Relationship
between back trajectories and tropospheric trace gas concen-
trations in rural Virginia. Atmospheric Environment 17,
2789}2800.
Niu, W., 1996. Nonlinear additive models for environmental
time series, with applications to ground-level ozone data
analysis. Journal of the American Statistical Association 91,
1310}1321.
NOAA, 1998. HYSPLIT4 (Hybrid Single-Particle Lagrangian
Integrated Trajectory) Model. Internet URL: http://
www.arl.noaa.gov/ready/hysplit4.html, NOAA Air Re-sour-
ces Laboratory, Silver Spring, MD, USA.
Poirot, R.L., Wishinski, P.R., 1986. Visibility, sulfate and air-mass
history associated with summertime aerosol in northern Ver-
mont. Atmospheric Environment 20, 1457}1469.
Ryan, W.F., 1995. Forecasting severe ozone episodes in the
Baltimore metropolitan area. Atmospheric Environment 29,
2387}2398.
Ryan, W.F., Luebehusen, E., 1996. Accuracy of ozone air quality
forecasts in the Baltimore metropolitan area. 96-TA45.03,
Presentation at 89th Annual Meeting and Exhibition, Nash-
ville, Tennessee, of the Air and Waste Management Associ-
ation, 23}28 June 1996.
Ryan, W.F., Doddridge, R.R., Dickerson, R.R., Morales, R.M.,
Hallock, A.H., Roberts, P.T., Blumenthal, D.L., Anderson,
J.A., 1998. Pollutant transport during a regional O
`
episode
in the mid-Atlantic states. Journal of the Air and Waste
Management Association 48, 786}797.
US EPA, 1994. The experimental UV index factsheet: explaining
the index to the public. EPA 430-F-94-019, US. Environ-
mental Protection Agency}National Oceanic and Atmo-
spheric Administration.
US EPA, 1996. Review of National Ambient Air Quality Stan-
dards for Ozone: Assessment of Scienti"c and Technical
Information. EPA 452/R-96-007, U.S. Environmental Pro-
tection Agency, Research Triangle Park, NC, USA.
US EPA, 1997. OTAG technical supporting document. Internet
URL: http://www.epa.gov/ttn/otag/"nalrpt/.
US EPA, 1998a. National Air Quality and Emissions Trends
Report, 1996. EPA 454/R-97-013, U.S. Environmental Pro-
tection Agency, Research Triangle Park, NC, USA.
US EPA, 1998b. NO
V
emissions coal-"red plants continental
U.S., Internet URL:http://www.epa.gov/acidrain/emission/
index.htm.
Wol!, G.T., Lioy, P.J., 1977. An investigation of long-
range transport of ozone across the Midwestern and
Eastern United States. Atmospheric Environment 11,
797}802.
Wol!, G.T., Lioy, P.J., 1978. An empirical model for forecasting
maximum daily ozone levels in the Northeastern U.S.
Journal of the Air Pollution Control Association 28,
1035}1038.
4674 W.G. Cobourn, M.C. Hubbard / Atmospheric Environment 33 (1999) 4663}4674