EPA/600/A-97/003
TRANSPORTATION RESEARCH RECORD 1472
Driving Pattern Variability and Impacts on
Vehicle Carbon Monoxide Emissions
David C. LeBlanc, F. Michael Saunders, Michael D. Meyer, and
Randall Guensler
An analysis of instrumented vehicle data revealed significant differ-
ences in operating mode profiles for vehicle operations in Atlanta, Geor-
gia; Baltimore, Maryland; and Spokane, Washington. Differences in
such operating mode characteristics as acceleration rates and cruise
speed distributions are important in the development of new emissions
models because certain vehicle and engine operating modes are prov-
ing to be significant sources of elevated emissions rates. Although not
conclusive, these data indicate that the variations in operating mode
fractions across cities may be related to differences in road network
characteristics. A simple predictive model, based on three operating
parameters (vehicle speed, engine speed, and manifold absolute pres-
sure) was developed from data collected from eight instrumented Gen-
eral Motors 3.1-L vehicles and is capable of predicting elevated carbon
monoxide (CO) emission rates for various vehicle and engine activities.
These emission results do not apply to hydrocarbons (HC) or oxides of
nitrogen (NO,), which behave differently. The modeling technique dis-
cussed has been developed exclusively for CO. The model is used to
estimate the relative CO emission differences associated with the dif-
ferences in operating profiles noted from city to city (and potentially
from driver to driver). This modeling approach appears capable of ade-
quately distinguishing the CO emission effects associated with varia-
tions in engine and vehicle operations for individual vehicle makes and
models. However, it should be noted that the large variability in vehi-
cle-to-vehicle CO emission response to changes in operating modes that
has been noted in ongoing studies indicates that a model based on vehi-
cle speed and acceleration profiles alone may not provide sufficient CO
emission rate predictive capabilities for the fleet.
Mobile emissions are known to be a significant source of air pollu-
tion in U.S. cities, typically accounting for more than 50 percent of
the ground-level ozone and 70 to 90 percent of the carbon monox-
ide (CO) (/). It is because of this role in air pollution that federal
legislation has focused on stringent motor vehicle emission stan-
dards and to a limited extent on the implementation of transporta-
tion control measures (TCMs) to control the levels of pollutants that
originate from mobile sources. With over 100 metropolitan areas in
violation of ozone standards and 60 in violation of CO standards (/),
there is a significant challenge facing the United States in attaining
and maintaining ambient air quality standards.
Of great importance in meeting this challenge is the development
and validation of a model that can accurately estimate changes in
pollutant emission rates associated with changes in transportation
network, vehicle, and driver characteristics. Although existing emis-
sions models have been in use for many years (with improvements
made in each new generation of model release), these models still
have serious deficiencies (2,3) that prevent their use in accurately
assessing emission rates at the corridor level (i.e., for transportation
system links). Ongoing research continues to add to an understand-
ing of the basic phenomena associated with emissions occurring
from components of the vehicle fleet. For example, several remote
sensing studies have shown that a small proportion of the fleet,
known as "super-emitters," may be responsible for a large propor-
tion of the excess emissions (4,5). The public perception is that these
super-emitters are either poorly maintained or very old vehicles.
However, recent studies have shown that new, properly maintained
vehicles can become high emitters under certain operating condi-
tions, such as high load conditions (6,7,8). Hence, a small fraction
of each vehicle's activity may be responsible for super-emissions,
or a large fraction of that vehicle's daily emissions (9). New models
must be capable of addressing the effects of both the presence of
super-emitters in the fleet and the occurrence of super-emissions
events associated with various vehicle operating modes.
Inherent in all the existing emissions models, and in most of the
new models being developed, is the assumption that there is an aver-
age driver, or at least that the variations in driver to driver behavior is
insignificant in the production of emissions from the vehicle. Aver-
age values for vehicle miles traveled and speed are used, resulting in
the loss of variation inherent from vehicle to vehicle and driver to dri-
ver. Much of the research related to developing new test driver cycles
(which may replace or supplement the federal test procedure cycle)
for emission rates assumes typical driving in urban areas (10). How-
ever, if the engine mode of operation is going to become an impor-
tant element of new models, there is clearly a need to better under-
stand how driver behavior can affect the frequency of these modes.
For example, given the same vehicle, are older drivers likely to drive
more conservatively than younger drivers, entering into engine
enrichment modes less often? Is there evidence to suggest that driving
patterns are indeed different from one city to another?
This paper examines instrumented vehicle data sets from Balti-
more, Maryland; Spokane, Washington; and Atlanta, Georgia, to
assess first the variation in driver behavior from one city to another
and to assess the potential impact this variation might have on CO
emissions estimation. After the sources and limitations of the data
used in this study are laid out, this paper examines the differences
in the frequency of activity that leads to high CO emissions among
the three urban areas. Then, two methods for estimating CO emis-
sions as a function of vehicle and engine operating modes are pre-
sented and used to assess the potential impacts that different driving
patterns may have on CO emissions estimation.
INSTRUMENTED VEHICLE DATA
School of Civil and Environmental Engineering, Georgia Institute of Tech-
nology, Atlanta, Ga. 30332.
A 1992 study in Spokane, Washington; Baltimore, Maryland; and
Atlanta, Georgia; instrumented approximately 350 vehicles with a

-------
46
TRANSPORTATION RESEARCH RECORD 1477
device that recorded data for ihree parameters: vehicle speed in
meters per second, engine speed in revolutions per minute (RPM),
and manifold absolute pressure (MAP) in kilopascals (kPa). The
three-parameter data set yielded 213 vehicles for which valid data
were recorded on all channels. In Baltimore and Spokane, a six-
parameter data base contained data from 79 vehicles for the fol-
lowing measures: vehicle speed in meters per second, engine speed
in RPM, throttle position in percentages, and one of three measures
of air flow, engine coolant temperature, and the output of a wide-
range oxygen sensor that monitored exhaust gas composition (i.e.,
air-to-fuel ratio). The six-parameter data set yielded 46 vehicles
with valid data on all channels. Both studies recorded each para-
meter once per second, and each device continuously recorded the
date and time of operation.
Each resulting data set was subject to strict quality control pro-
cedures. More than 15 error-detection measures were used to track
the wide variety of anomalous conditions that could be part of any
given data set. Many of the problems detected were transient and
were corrected by substituting the erroneous value with an interpo-
lated value. Only the vehicle records containing valid data on all
recorded channels for the entire study period were used in this
analysis.
To avoid the emissions modeling problems associated with ele-
vated emissions rates during vehicle warm-up (2), the research team
used data collected only from hot stabilized engines. Engines were
assumed to have achieved hot stabilized combustion, and catalytic
converters were assumed to have reached light-off temperatures
by the time the engine temperature reached 70°C. Thus, the six-
parameter data used in developing emission rate models excluded
all data from operations when the engine coolant temperatures were
lower than 70°C.
Each vehicle recorded data for approximately 1 week before the
instruments were removed. In the three-parameter data set used in
this study, Atlanta drivers recorded over 3.0 million sec of opera-
tion from 76 vehicles, Baltimore drivers recorded 2.5 million sec of
operation from 68 vehicles, and Spokane drivers recorded 1.9 mil-
lion sec of operation from 69 vehicles. The six-parameter data used
in this study recorded 1.6 million sec of operation from 46 vehicles.
Driver Selection
Baltimore and Spokane drivers were solicited at centralized emis-
sions inspection centers, with vehicles instrumented at the time of
solicitation (11). Atlanta has no centralized emission inspection.
Drivers were solicited at three driver's license stations; their vehi-
cles were instrumented later at remote sites.
Data Limitations
The six-parameter data hase was limited. The sample size was small
and appears biased in important respects. For example
•	Only fairly new vehicles (i.e., model years between 1989 and
1991) were instrumented,
•	A limited number of engine types and sizes were included,
•	Young drivers are poorly represented (only 1 of the 46 drivers
was under the age of 25,
•	Manual transmission vehicles were underrepresented, and
•	High-performance vehicles were not included in the sample.
For the three-parameter data sets in all three cities, efforts were
made to select a representative sample of drivers and vehicle types
from the target population. For example, the three-parameter data
set was not restricted to the type or age of vehicle instrumented.
Potential driver bias has not yet been examined in the three-
parameter data set. However, based on the preliminary analysis of
the Atlanta data set, the three-parameter data set appears less likely
to be biased than the six-parameter data set.
Both data sets are somewhat limited in their usefulness because
geographic positioning data or accelerometer data were not col-
lected for use in evaluating the impacts of grade on speed, acceler-
ation, and throttle position. Furthermore, without positional infor-
mation, the data could not be directly associated with the roadway
classification upon which the vehicle was operating. Hence, if the
noted speed was 56 km/hr (35 mph), it was not possible to deter-
mine directly whether the activity occurred on a congested freeway
or a free-flowing arterial.
Despite the potential biases and shortcomings in the data sets, the
six-parameter and three-parameter data sets from these three cities
still represent a rich source of information on vehicle activity. The
data serve as an excellent point of departure for preliminary discus-
sions of the potential impacts of variations in driving patterns.
DATA ANALYSIS
Vehicle speed distributions for the three-parameter data set for each
city are shown in Figure 1 and indicate the proportion of total driv-
ing time spent in each specific speed range. For example, approxi-
mately 15 percent of the total driving time in Spokane occurred in
the 48 to 56 km/hr (30 to 35 mph) speed range, compared with only
8 and 6 percent in Baltimore and Atlanta, respectively. If it is
assumed that the speed range from 25 to 40 mph (40 to 64 km/hr)
represents driving that would occur primarily on arterial highways
or congested freeways, Spokane has the highest percentage of such
activity in the three cities studied. In addition, Spokane has the low-
est percentage of activity above 60 mph. Atlanta drivers tended to
drive much faster than their counterparts in the other two cities.
If drivers in the different cities operated on uncongested free-
ways, the shape of the high end of the speed distributions should be
the same. Because they are different, it may be because (a) the dri-
vers in the different cities do not drive on uncongested freeways,
which means that they do not have freeways, they do not drive on
their freeways, or that their freeways are not uncongested, or (b) the
drivers in the different cities are driving differently on uncongested
freeways, which means that the freeways may be physically differ-
ent, causing different responses, or that the freeways are physically
similar but that there is a behavior difference between drivers in
various cities. Unfortunately, with the data collected, the reason
cannot be determined.
If it is assumed that the largest fraction of vehicle activity occurs
on arterial highways, this activity occurs in Spokane and Baltimore
in the 40 to 64 km/hr (25 to 40 mph) speed range. In Atlanta, this
fraction of activity occurs in the 56 to 80 km/hr (35 to 50 mph)
speed range. In addition, there appears to be a less distinctive break
between the arterial highway activity fraction and the freeway frac-
tion in Atlanta. However, depending on congestion conditions,
some of the data from congested freeways may overlap data from
arterial operations.
These results are perhaps not surprising given the different ter-
rain and road network characteristics of the three cities. Although

-------
LeBlanc et at.
47
-Atlanta
Baltimore
Spokane
120 km/hr
Vehicle Speed
FIGURE 1 Vehicle speed distributions for three-parameter data sets. (Bin marked "5-10"
refers to speeds £ 5 mph and < 10 mph.)
the Spokane metropolitan area includes an Interstate highway, the
Interstate serves primarily as a bypass link and does not serve as a
major transportation link for trips internal to Spokane. Baltimore is
a more densely developed, older city than Atlanta and Spokane,
with a freeway system that is more expansive than Spokane's but
not as large as Atlanta's. The freeway system in Baltimore is aug-
mented with a highly developed set of arterial highways that have
traditionally served many of the internal Baltimore trips. Atlanta, on
the other hand, has a newly expanded freeway system that many
drivers use as the major means of reaching destinations in the
Atlanta area. The freeway system is accessed by a large network of
major arterial roads, many with high levels of capacity that experi-
ence high-speed activity.
Although the reasons for the noted differences are as yet unclear,
the data in Figure 1 clearly indicate that there are substantial differ-
ences in vehicle speeds from one city to another. These differences
are statistically significant and were substantiated through discrim-
inant analysis, where a set of functions is derived that minimizes the
variance within a group and maximizes the variance between
groups (12), The discriminant analysis results are contained in
Table 1. In this case, two functions were needed to classify each
driver into the three groups. The proportion of total driving time in
each of 16 speed bins for each driver was used to predict in which
city the driver operated the vehicle.
If the speed profiles contained little information about the city in
which a driver operated, a discriminant analysis would misclassify
most of the drivers, with a success rate approaching that of chance
assignment. In this case, the speed profiles worked well in deter-
TABLE 1 Results of Discriminant Analysis Based on Speed Profiles
Actual Group
Predicted Group Membership
City Cases
Baltimore
Atlanta
Spokane
Baltimore 68
51
10
7

75.0%
14.7%
10.3%
Atlanta 76
13
63
0

17.1%
82.9%
0.0%
Spokane 69
13
2
54

18.8%
2.9%
78.3%
mining in which city the driver operated, with a success rate of 79
percent. Atlanta drivers were most frequently correctly grouped,
indicating that Atlanta drivers* speed profiles are more distinctive
than those for Baltimore or Spokane. Also, no Atlanta driver was
misclassified as Spokane drivers were. There is also some overlap
between the driving patterns found in Baltimore and the other two
cities.
However, Figure I does not allow for observations about the style
or aggressiveness of driver behavior (which could also be related to
the characteristics of the road network). For the purposes of this
paper, aggressive is defined to indicate higher acceleration rates. If
each data set was segregated into subsets by driver according to the
proportion of driving in arterial or highway modes, previous
research indicates that the acceleration distributions would not be
distinctly different for these subsets (13). That is, drivers who spend
most of their driving time at freeway speeds are not more likely to
drive more aggressively in any speed range than the drivers who
spend most of their driving time at arterial speeds. However, one
possible measure of driver aggressiveness is the distribution of
acceleration across all speed ranges.
To examine the potential differences in acceleration profiles
across cities, the standard deviation of the acceleration and deceler-
ation values was calculated for 8 km/hr (5 mph) bins for each of the
cities (Figure 2). A larger standard deviation implies a larger num-
ber of vehicles with greater acceleration or deceleration values, or
both, in that speed group, a phenomenon of great interest in esti-
mating emissions related to engine load or power enrichment. By
this measure, Atlanta drivers were more aggressive in most speed
ranges.
The acceleration profiles were also examined by using discrimi-
nant analysis, and the results are better than those obtained using
only the speed profiles. The analysis grouped the drivers into their
correct cities 85 percent of the time, with Atlanta drivers grouped
properly 88 percent of the time.
The results of the discriminant analysis clearly show that the
driving patterns are significantly different across the three cities. It
may be that particular transportation network characteristics are the
most important parameter. For example, higher levels of accelera-
tion changes could be explained by a larger number of opportuni-

-------
TRANSPORTATION RESEARCH RECORD N72
m/sec2 MPH/sec
3 ;
; 	—ASarta
0	30	60	90	120 km/fir
Vehicle Speed
FIGURE 2 Standard deviation of acceleration distributions for three-parameter data sets.
(Bin marked "5-10" refers to speeds > 5 mph and < 10 mph.)
ties in a road network to accelerate or decelerate (e.g., stop signs or
traffic signals). This could certainly be one explanation for the dif-
ferences in the lower speed ranges (such traffic controls are not
found on freeways). The greatest differences in acceleration stan-
dard deviations between Atlanta and the other two cities occur in
the 24 to 72 krn/hr (15 to 45 mph) speed ranges. This suggests that
the greatest variation in acceleration behavior may occur on arte-
rial highways (or congested freeways). One possible contributing
factor is that the Atlanta arterial road network covers a much larger
geographic area than that of the other two cities, often providing
drivers with greater distances before signal interruptions. This may
not only allow greater speeds but also account in part for the distinct
differences in acceleration activity. Perhaps the transportation sys-
tem characteristics have conditioned drivers to drive in the manner
noted for each city. That is, drivers may simply respond to various
infrastructure characteristics, such as lane width or presence of
highway barriers, in terms of speed and acceleration profiles.
However, demographic differences or vehicle sample could also
account for some of these characteristics. It may be that driver char-
acteristics are responsible for the differences noted across the cities.
Perhaps the age distribution or previous driving experiences play a
role in modal profiles. Perhaps the vehicles themselves are an
important explanatory variable or an interaction term with driver
characteristics. It may even be that local law enforcement habits
play a role in these differences. There are no clear reasons why such
differences in vehicle activity occurred. But the Georgia Institute of
Technology has undertaken additional studies to explain these dif-
ferences. In future studies, vehicle characteristics, driver character-
istics, and infrastructure characteristics will be controlled during
data collection so that statistical analyses are more likely to reveal
the factors thai appear to affect these activity differences {14).
In summary, an examination of instrumented vehicle data sets
from three U.S. cities indicated that there are significant differences
among the cities in vehicle activity profiles. These differences may
be caused by the characteristics of the road networks or the driving
styles found in separate regions of the United States. The impor-
tance of this finding is that it suggests the existence of potentially
substantial differences across cities in mobile emissions estimates,
depending on the relative contribution of modal emissions to the
overall emissions inventory. To examine the potential impacts of
these activity differences, two simple predictive models for CO
emissions, derived as functions of vehicle and engine parameters,
were developed from the data collected during the six-parameter
study. These models were then used to examine the relative CO
emissions from the cities, given the different speed-acceleration
distributions found in each city.
POTENTIAL EMISSIONS IMPACT OF
DRIVER BEHAVIOR VARIABILITY
CO emissions can be estimated from variables contained in the six-
parameter data set. By coupling the wide-range oxygen sensor
(which detects the exhaust air-to-fuel ratio) reading with mass air
flow and assumed catalytic converter efficiency, the CO emissions
rate can be estimated. The development of this method was covered
extensively in a previous work (15) and is not repeated here.
Using the methodology developed previously {15), two alterna-
tive engine-specific models were developed from the largest subset
of the six-parameter data available, eight General Motors vehicles
with 3.1 -L engines that were equipped with MAP sensors. The eight
vehicles in this subset made 350 trips, and emissions were modeled
on a per-trip basis. Two different models were considered: a speed-
acceleration model and a speed-MAP-RPM model. These parame-
ters were chosen because the three-parameter data included these
variables. These models could likely be improved if throttle posi-
tion were also used as a predictive variable, because many engine
control units base commanded enrichment logic on throttle position
(as well as other factors not included in the model) and because
throttle position is controlled directly by the driver.
The speed-acceleration model initially considered six zones of
operation (Figure 3). The characteristics of the zones are

-------
lA'HhltW V! (tl.
49
1.	Speed and acceleration combinations within the bounds of the
1-TP test.
2.	Speeds less than the maximum FTP speed of 92 km/hr (57
mph) and acceleration rates higher than the FTP maximum for any
given speed,
3.	Speeds less than the maximum FTP speed and deceleration
rates higher than the FTP maximum for any given speed,
4.	Speeds greater than the maximum FTP speed and acceleration
rates greater than 0.45 m/sec2 (I mph/sec),
5.	Speeds greater than the maximum FTP speed, acceleration
rates less than 0,45 m/sec2 (I mph/sec), and deceleration rates less
than 0.45 m/sec2 (1 mph/sec),
6.	Speeds greater than the maximum FTP speed of 92 km/hr (57
mph) and deceleration rates greater than 0.45 m/secJ (1 mph/sec).
Of these six zones, the two deceleration zones (Zones 3 and 6)
were not found to be statistically significant. Zones 3 and 6 were
merged with the FTP zone. The functional form of the regression
equation is
CO(g/sec) = 0.050514 + 0.094067 * (HI_SPEED) +
0.642077 * (LO_ACCEL) + 0.823341 * (HI_ACCEL)
where
HI_SPEED = the proportion of each trip with speeds greater
than 92 km/hr (57 mph) and acceleration rates
less than 0.45 m/sec2 (1 mph/sec) (Zone 5),
LO_ACCEL = the proportion of each trip with speeds less than
92 km/hr (57 mph) and accelerations greater than
those found on the FTP (Zone 2), and
H1_ACCEL = the proportion of each trip with speeds greater
92 km/hr (57 mph) and accelerations greater than
0.45 m/sec1 (1 mph/sec) (Zone 4).
The R2 for this model is fairly poor at 0.29, with an
F-statistic of 46.9 and a standard error of 0.035 g/see.
The speed-MAP-RPM model is based on the concept that engine
parameters govern commanded enrichment and will better predict
modal emissions for a single engine type. When the CO emissions
rate is plotted across MAP and Rl'M. four zones were defined to
account for different engine modes. These four zones were defined
arbitrarily as operations with
•	MAP S 70 kPa and RPM < 3,500, corresponding to normal
driving;
•	MAP > 70 kPa and RPM rs 3,500, corresponding to high-load
conditions, such as climbing a steep hill;
•	MAP £ 70 kPa and RPM > 3,500, corresponding to a high-
RPM, low-load condition, which rarely occurs; and
« M AP > 70 kPa and RPM > 3,500, corresponding to high-load
conditions that are often associated with both commanded enrich-
ment and high-mass air flows.
Each of these four zones was then examined for variance with
respect to vehicle speed, With the exception of the rare high-RPM,
low-load condition, the CO emission rates in each zone varied sim-
ilarly with speed. Each zone showed the lowest emission rates when
speed was less than 16 km/hr (10 mph). Emission rates then became
speed-invariant to approximately 113 km/hr (70 mph). In light of
this, each of the four engine zones was divided into three speed
zones: less than 16 km/hr {10 mph), between 16 and 113 km/hr (10
and 70 mph), and greater than 113 km/hr (70 mph).
This model required fine tuning as well. The high-RPM, low-load
zone had very little data and did not exhibit a clear relationship with
vehicle speed; thus, the three zones were merged into a single zone.
The high-load, zone with RPM less than 3,500 had insufficient data
to support separate groups for moderate and high speeds, and these
two zones were merged as well. No data points included activity at
speeds less than 10 mph and with both high MAP and RPM.
The resulting regression equation displayed much better results
than the model based only on speed and acceleration with an R2 value
of 0.56, an F-statistic of 62,8, and a standard error of0.029 g/see.
m/sec2 MPH/sec
1,5
1.0
0.5
o
ca
<5
< -o.s
-1.0
-1.5


HI_ACCEL
FTP
LO_ACCEl
j

HIJSPEED



I
HI_DECEL

LO_DECEL


10
20
—i	
30
30
40
50
60
70
00 MPH
60
Speed
SO
120 km/hr
FIGURE 3 Zones used in development of spccd-acceleration model.

-------
Ml
TRANSPORTATION RESEARCH RECORD 1472
The engine-basal model clearly explains more of the variation in
the CO emission rales of vehicles equipped with 3.1 -L engines than
does the spced-aceeleration model. This is because the engine con-
trol unit (on-board computer) commands enrichment based largely
on monitored engine parameters. A given speed-acceleration com-
bination does not directly determine these engine parameters.
(Note, however, that the linear acceleration variable, change in
speed per change in time, used in the model did not include accel-
eration due to grade.)
Throttle position is another potentially significant variable
because it is monitored by the on-board computer system and
appears to be used by many vehicles in commanded enrichment
algorithms. Preliminary analyses indicate that emission rate varia-
tion for specific vehicle makes and models can be comparably
explained by a model based solely on RPM and throttle position.
Because throttle position appears to be only partially correlated with
engine load expressed as MAP (a Pierson correlation coefficient of
roughly 0.75 for the six-parameter data examined), variation in
throttle position may be in part due to the individual differences
in how the driver uses the throttle to interface with the engine.
It is important to keep in mind, however, that other studies have
indicated that the vehicle-to-vehicle variations in emissions
response to various operating modes and loads (i.e., modes that may
cause commanded enrichment) appear to be large (9). Hence, an
engine-parameter model developed from single or limited engine
types may be inappropriate when applied to other vehicles.
The speed-acceleration and speed-MAP-RPM models were then
re-derived using the entire six-parameter data set. The three-para-
meter data set was not used because engine sizes were not recorded.
In the case of the Spokane and Baltimore data, even the vehicle type
was unknown. It is theoretically possible to derive engine size data
from the vehicle identification number, but this was not attempted.
In future analyses, it would be ideal to use some measure of differ-
ences among vehicles, particularly engine size, when extending this
type of model. As expected, the models did not perform as well
when they were derived from data collected for several vehicles
with different engine types and control strategies taken together. In
the case of the speed-acceleration model, the RJ dropped from 0.29
to 0.17, with the standard error rising from 0.035 to 0.056 g/sec. The
speed-MAP-RPM model did not suffer as severe a degradation,
with the R2 dropping from 0.56 to 0.37, which is a better fit than the
speed-acceleration model was able to manage over a single vehicle
type. However, the standard error is also fairly high at 0.050 g/sec.
The proportion of each trip spent at low speeds and normal engine
operation was taken as the regression constant because this region
would usually correspond to idle and was found to be statistically
insignificant. The regression equation, where the value of each vari-
able is the proportion of each trip that fell into a given zone is
CO(g/sec) = 0.029854 + 0.034631 (NOR_MED)
+ 0.196595(NOR_HI)
+ 1,304044(HI_RPM)
+ 0.029155(HI_MAP_LO)
+ 0.273061 (HI_MAP_MEDHI)
+ 3.228802(HI_LOAD_LOMFX>)
+ 22.74787(HI_LOAD_Hl)
where
NOR^MED = activity at speeds between 16 and 113 km/hr (10
and 70 mph) and normal engine parameters;
NOR_HI = normal engine parameters where speed > 113
km/hr (70 mph);
HLRPM = all activity at high RPM and MAP < 70 kPa;
HI_MAP= activity at high MAP, but RPM < 3,500, with the
speed divisions as above; and
HI_LOAD= activity where MAP and RPM are both high.
The smaller degradation of the engine model may be because any
engine is likely to be in enrichment at high MAP and RPM, and to
some degree at high MAP independent of RPM. However, the fre-
quency of high-load activity for any vehicle will vary as a function
of engine size and vehicle weight (i.e., load is associated with the
power-to-weight ratio). Engine size appears from other studies to be
an important causal variable (9). and engine size and vehicle weight
were not used as explanatory variables in the derived models. That
these variables are not included is a limitation in the derivation and
application of these two models. Note, however, that engine size
may not be a sufficient discriminant variable—the GM 3.1-L vehi-
cles equipped with MAP sensors behaved differently from the GM
3.1-L engines equipped with LV8 sensors, and there were signifi-
cant differences between these and the 3.0-L Ford.
It is important to note that the estimate of the CO emissions rate
does not include measure of startup or cold-operation emissions
because these data were not used in the analyses. In addition, as
noted earlier, the six-parameter data base is limited to only a few
engine types of a limited manufacture dale range. Large engines,
sports cars, manual transmissions, and young drivers are all under-
represented in this data set. Any values obtained by extrapolation to
the three-parameter data should not be considered an accurate esti-
mate of overall CO emissions for a particular vehicle, only as a pre-
liminary indication of emission rate differences associated with dif-
ferences in vehicle activities. A much larger study would be
necessary to obtain enough data to accurately predict the emissions
rates for the general population. Applying the models developed
using the six-parameter data to the three-parameter data sets is not
ideal. The application was intended to explore only the capability
of the three-parameter data to distinguish among different driving
patterns and to see whether the differences in speed and accelera-
tion behavior have a possible impact on emissions. As such, these
models are taken as a common metric that should be used for com-
parative purposes only.
These models based on the six-parameter data produce inter-
esting results when applied on a trip-by-trip basis to the three-
parameter data sets. There were 4,354 trips recorded in Atlanta,
3,701 trips in Spokane, and 3,641 trips in Baltimore. The speed-
acceleration model tended to predict a much smaller variability than
the speed-MAP-RPM model (Figure 4). The speed-acceleration
mode! yielded a median CO emissions rate (on a per-trip basis) of
0.078 g/sec for Atlanta drivers, 0.067 g/sec for Baltimore drivers,
and 0.064 g/sec for Spokane drivers. The speed-MAP-RPM model
yielded a median CO emissions rate of 0.102 g/sec for Atlanta driv-
ers, 0.087 g/sec for Baltimore drivers, and 0.079 g/sec for Spokane
drivers (Figure 5). These results are not surprising when compared
with the overall behavior patterns found using speed and accelera-
tion profiles. (Spokane drivers exhibited the lowest average speeds
and the lowest acceleration rates, the Baltimore drivers were in the
middle, and the Atlanta drivers exhibited the highest speeds and
acceleration rates.) This trend is replicated in the results of these two
models. Interestingly, the speed-acceleration model shows little dif-
ference between the median emissions rates of Spokane and Balti-
more drivers, in contrast with the results of the speed-MAP-RPM

-------
Is Blanc ct til.
51
% 10
* 0.9
JQ
tft
§ 0.8
I 07
Ui
O 0.6
O
° 0.5
c
3 0.4
| 0.3
&
.g 0.2
s
| 0.1
3
o
0.0
	*		

/ ~ /
/
/.* /
'' /
t ' /
I' /
$¦' /
	Atlanta j
	Spokane
1' I
f' /
t: /
	Baltimore
/
'• /
/- /

/. /
*• i
/
!

i' f
ill
i.f
V

v.
if
V

0.05	0.1	0.15	0.2
Predicted CO emissions (g/sec)
0.25
0.3
FIGURE 4 Cumulative distribution of predicted CO emissions by trip for speed-
acceleration model.
model, which predicts larger differences between median emission
rates in these cities.
At this time, it is unclear whether the relatively poor performance
of the speed-acceleration model is due to inherent flaws in attempt-
ing to model emissions based solely on speed and acceleration, lack
of control over the grade variable, or inadequate model specifica-
tion (i.e., only four activity zones were used in this model), or
whether poor performance is an artifact of the potential biases
within this particular data set. Nevertheless, this initial work indi-
cates that the speed-MAP-RPM model may provide greater sensi-
tivity to changes in driving patterns.
One factor that has not yet been discussed is long-term modeling
implications of speed-acceleration models versus speed-MAP-
RPM models. When a CO emission rate model is developed, the
challenge that remains is to quantify the activity that must be used
in the modeling process. That is, if a speed-acceleration model is
used, the vehicle activity on a transportation link must be quantified
in terms of speed and acceleration profiles. If a speed-MAP-RPM
model is used, the vehicle activity on a transportation link must be
quantified in terms of speed, MAP, and RPM profile. This is clearly
not a simple modeling issue. Whereas the identification of speed
and acceleration profiles is fairly straightforward and likely to be
independent of the vehicle subfleet characteristics operating on the
link, the RPM and MAP profiles arc totally dependent on the char-
acteristics of that vehicle subfleet. Hence, the potentially higher
explanatory power of engine-based models may be compromised if
0.05	0.1	0.15	0.2
Predicted CO emissions (g/sec)
0.25
0.3
FIGURE S Cumulative distribution of predicted CO emissions by trip for
speed-MAP-RPM model.

-------
TRANSI'ORTAHON RESEARCH RECORD N72
highly uncci tain vehicle MAP and RPM distributions are linked
with the emission rates. Clearly, in constructing long-term emis-
sions models, a difficult balance must be reached.
CONCLUSIONS
Mobile source emissions are dependent on vehicle type, vehicle
activity, and possibly transportation network or driver characteris-
tics, or both. Important and statistically significant differences in
vehicle activity profiles have been found among the three cities
studied. It is unclear from this data set whether network character-
istics explain these differences completely or whether other charac-
teristics of these cities also play a role. A study looking for differ-
ences and similarities between drivers in cities with similar
transportation networks would be necessary to test this hypothesis.
The differences noted in vehicle activity profiles suggest that
emissions models must adequately incorporate these variations into
the modeling regime if they are to be applied across a variety of
metropolitan areas. An emissions model using engine operating
parameters could provide a basis for newer, state-of-the-art trans-
portation models where fleets of vehicles are modeled based on the
characteristics of driving conditions and engine modal operations.
These models can account for differences in driving habits and pos-
sibly point out locations on the transportation network (such as on-
ramps) where high-emissions driving would occur. However, such
an application requires accurate vehicle and engine operating pro-
files to be developed for the vehicle fleet for the emission rate algo-
rithms to be applied. Note that these results should not be extrapo-
lated to HC or NOj.
A model that uses only the speed and acceleration distributions
for a given roadway segment can be developed and applied. How-
ever, this approach initially appears to be much less sophisticated
than the engine-based approach. It should be noted, however, that
the model tested in this research used linear acceleration and did not
account for grade effects. Once grade effects are included in net
acceleration, the speed-acceleration model may provide signifi-
cantly improved explanatory power. Also, the effects of grade may
be more significant at higher speeds than at lower speeds. In addi-
tion, the spccd-acceleration model developed used only four activ-
ity zones, and improvements in explanatory power may result from
a more refined model. Although a model based only on speed and
acceleration may not perform as well as an engine parameter model,
the activity data are likely to be more easily and accurately mea-
sured and modeled. Hence, the approach may simply be more
practical than an engine model.
Perhaps most important, this paper highlights the need for further
research on variation in driving behavior. As emissions modeling
research continues to develop new approaches on emissions pre-
diction based on engine modal operation, the transportation com-
munity needs to know more about the characteristics of drivers that
would cause these vehiclc-and engine-operating distributions to
occur. Driving patterns vary from one city to the next; hence, it is
not enough to collect statistically valid vehicle data within a single
city. At the very least, this would suggest that an important input
variable for emissions models may be a driving behavior factor that
represents the driving style and trip cycles found in that particular
city, perhaps as a function of infrastructure, fleet characteristic, and
demographics. Additional research is necessary to define better the
different characteristics of this driving factor.
ACKNOWLEDGMENTS
This paper is the result of research funded by the U.S. Environ-
mental Protection Agency (EPA). The authors acknowledge the
support and participation of Jim Markey and John German of the
Office of Mobile Sources, EPA. We also thank Ted Ripberger of
AEERL, EPA, for his continued assistance. We acknowledge the
contribution of Wang Shun (Ricky) Wu, whose assistance during
the collection and analysis of the Atlanta three-parameter data was
invaluable.
REFERENCES
1.	Beckham, B.. W. Reilly. and \V. Becker. Clean Air Acl Amendments
and Highway Programs. TR News. May-June 1990.
2.	Guensler, R., and A. B. Gcraghty. A Transportation/Air Quality
Research Agenda for the 1990's. Proc.. 84th Annual Meeting of the Air
and Waste Management Association, Vol. 8, Emissions (AM91-8):
Paper No. 87.2, Pittsburgh, Pa , June 1991.
3.	Gcrtler, A., and W. Pierson. Motor Vehicle Emissions Modeling Issues.
Proc., 84th Annual Meeting of the Air and Waste Management Associ-
ation, Pittsburgh, Pa., June 1991.
4.	Bishop, G. A., D. H. Sleadman, J. E. Peterson, T. J. Hosick, and P. L.
Guenther. A Cost-Effectiveness Study of Carbon Monoxide Emissions
Reduction Utilizing Remote Sensing. Journal of the Air and Waste
Management Association, Vol. 43, July 1993.
5.	Lawson, D. R., P. J. Groblicki, D. H. Stedman. G. A. Bishop, and P. R-.
Guenther. Emissions from In-Use Vehicles in Los Angeles: A Pilot
Study of Remote Sensing and the Inspection and Maintenance Program
Journal of the Air and Waste Management Association. Vol. 40. No. 8,
Pittsburgh, Pa., Aug. 1990, pp. 1.096-1,105.
6.	Kelly, N. A., and P. J. Groblicki Real-World Emissions from a Mod-
em Production Vehicle Driven in Los Angeles; Journal of the Air and
Waste Management Association: Vol. 43; Pittsburgh. Pa.. Oct. 1993.
pp. 1,351-1,357.
7.	LeBlanc, D. C., F. M. Saunders, M. D. Meyer, C. Ross, R. DuBose, and
C. T. Ripberger. Use of Wide Range Oxygen Sensors in Instrumented
Vehicles—Preliminary Studies. Presented at Air and Waste Manage-
ment Conference on Emission Inventory Issues, Durham, N.C., Oct.
1992.
8.	Meyer, M. D., C. Ross, G. T. Ripberger, and M. O. Rodgers. A Study
of Enrichment Activities in the Atlanta Road Network. Proc.. Interna-
tional Specialty Conference on Emission Inventory Issues, Durham,
N.C., Air and Waste Management Association, Pittsburgh, Pa., 1992.
9.	Guensler, R., D. C. LeBlanc. and S. Washington. Jcckyl and Hyde Emit-
ters; The Emission Inventory: Applications and Improvement. Proc.,
Fourth International Conference on the Emission Inventory. Air and
Waste Management Association, Pittsburgh, Pa., Nov. 1994.
10.	Guensler, R. Data Needs for Evolving Motor Vehicle Emission Model-
ing Approaches. In Transportation Planning and Air Quality II (P. Ben-
son, ed.), American Society of Civil Engineers, New York, N.Y.. 1993
11.	Federal Test Procedure Review Project: Preliminary Technical Report.
EPA Report 420-R-93-007, Environmental Protection Agency, Office
of Air and Radiation, 1993.
12.	SPSS for Windows: Advanced Statistics, Release 6.0. SPSS, Inc.,
Chicago, III., 1993.
13.	Verma, N., Characterization of Driving Behavior Based on the Atlanta
FTP Study. Special Research Problem Report. School of Civil Engi-
neering, Georgia Institute of Technology, Atlanta, 1994.
14.	Ross, C., R. Guensler, M. D. Meyer, and M. O. Rodgers. The Atlanta
1995 Driving Panel: Research Plan (draft). Dec. 1994.
15.	LeBlanc, D. C., M. D. Meyer, F. M. Saunders, and J. A. Mulholland.
Carbon Monoxide Emissions from Road Driving: Evidence of Emis-
sions Due to Power Enrichment In Transportation Research Record
1444, TRB, National Research Council. Washington, DC, 1994.

-------
A tp xpt> r D locn TECHNICAL REPORT DATA
L Jr ~ OU (Please read Instructions on the reverse before cample
EPFAT6r0f/A-97/003
3.
4. TITLE AND SUBTITLE
Driver Behavior Variability and Impacts on Vehicle
Emissions
5. REPORT DATE
6. PERFORMING ORGANIZATION CODE
7. author(s) d# c< LeBlanc, F. M. Saunders, M. D. Meyer,
and R. Guensler
8. PERFORMING ORGANIZATION REPORT NO.
9. PERFORMING ORGANIZATION NAME AND ADDRESS
Georgia Institute of Technology
School of Civil and Environmental Engineering
Atlanta, Georgia 30332
10. PROGRAM ELEMENT NO.
11. CONTRACT/GRANT NO.
CR 817732-02
12. SPONSORING AGENCY NAME AND ADDRESS
EPA, Office of Research and Development
Air Pollution Prevention and Control Division
Research Triangle Park, NC 27711
13. TYPE OF REPORT AND PERIOD COVERED
Published paper; 1-8/94
14. SPONSORING AGENCY CODE
EPA/600/13
is.supplementary notes^ppcd project officer is Carl T, Ripberger, Mail Drop 61, 919/
541-2924. For presentation at Transportation Research Board Annual Meeting, Wash-
ington, DC, 1/26/95.
is. abstract paper reports results of an analysis of instrumented vehicle data that
revealed significant differences in vehicle operating mode profiles for vehicle opera-
tions in Atlanta, GA, Baltimore, MD, and Spokane, V/A. Such differences in opera-
ting mode characteristics as acceleration rates and cruise speeds are important in
the development of new emissions models in that certain vehicle and engine operating
modes are proving to be significant sources of elevated emissions rates. Although
not conclusive, these data indicate that the variations in operating mode fractions
across cities may be related to differences in driver behavior, where driver behavior
is defined as the differences between individuals in their response to roadway charac-
teristics and conditions. A simple predictive model, based on three operating para-
meters (vehicle speed, engine speed, and manifold absolute pressure) and developed
from data collected from eight instrumented General Motors 3.1 liter vehicles, is
capable of predicting elevated emission rates for various vehicle /engine activities.
This model is used to estimate the relative carbon monoxide emissions differences
associated with the differences in operating profiles noted from city to city (and po-
tentially from driver to driver).
17. KEY WORDS AND DOCUMENT ANALYSIS
a. DESCRIPTORS
b.IDENTIFIERS/OPEN ENDED TERMS
c. cosati Field/Group
Pollution Acceleration (Physics)
Motor Vehicles Carbon Monoxide
Emission Variability
Estimating
Mathematical Models
Motor Vehicle Operators
Behavior
Pollution Control
Mobile Sources
Driver Behavior
Cruising Speed
13 B 20K
13F 07B
14G
12 A
051
05J
13. DISTRIBUTION STATEMENT
Release to Public
19. SECURITY CLASS (This Report}
Unclassified
21, NO. OF PAGES
20. SECURITY CLASS (This pa^e)
Unclassified
22. PRICE
EPA Form 2220-1 (9-73)

-------