ICCAIS 2013: Main Track ICCAIS 2013

CO2 Vehicular Emission Statistical Analysis with Instantaneous Speed

and Acceleration as Predictor Variables
S. D. Oduro, S. Metia, H. Duc, and Q. P. Ha

Abstract— Models for predicting vehicular emissions of car- Australia. On an annual basis this is 2.3 higher than the
bon dioxide (CO2 ) are usually insensitive to vehicle modes of number of registrations for 2011 and an increase of 13.3%
operation (such as cruise, acceleration, deceleration, and idling) since 2007, when there were 14.8 million vehicles registered
as they are based on the average speed of motor vehicles. In
the present study, real world on-road second-by-second data in Australia. Due to a rapid increase of the road transport
are used to improve the accuracy of air quality models by fleet, there is also a growing source of greenhouse gas
considering modal emissions of CO2 in terms of vehicles’ emissions. Between 1990 and 2007, emissions from road
instantaneous speed and acceleration. A regression analysis transport increased by 26.1 per cent or 14.2 million tonnes. In
approach is used with speed and acceleration as the predictor 2007, road transport was responsible for 68.5 million tonnes
variables while CO2 emission factor as the outcome variable for
vehicles manufactured in 2002 and 2008. The results show that of greenhouse gas emissions. Passenger vehicles (cars) are
there is significantly a linear relationship between CO2 , speed the largest source of greenhouse gas emissions from road
and acceleration/deceleration in which speed, as compared to transport and accounted for around 61% in 2007 [5]. On
acceleration, has a stronger correlation with respect to the average, for every litre of petrol used in a motor vehicle,
CO2 emission factor. Also, for 2002 and 2008 vehicles, every 2.3 kilograms of carbon dioxide (CO2 ) is released from the
1m/s increase in speed will emit respectively 0.041g/s and
0.034g/s CO2 , whereas an increase in acceleration by 1m/s2 exhaust [6].
will produce 0.025g/s and 0.008g/s of CO2 emission in the case The on-road emission reduction in Australia has improved
of constant predictors. While speed and acceleration are all dramatically in recent years due to performance of new
significant predictors of CO2 emission, it is concluded from the vehicles by adopting more stringent fuel and emission stan-
magnitude of the t-statistics that speed has a greater impact dards. However, there remains the need to develop vehicle
than acceleration in predicting CO2 emission.
emission and air quality models that could predict emission
I. I NTRODUCTION factor and pollution concentration accurately. For this, The
Air Pollution Model (TAPM), developed by CSIRO Marine
Road transportation greatly contributes to the economic
and Atmospheric Research, has been used to predict the
activity of all nations, and thus constitutes one of the most
three-dimensional meteorology and pollutant concentration
significant indicators of the economic situation of a country.
profiles for a range of air pollutants. Although TAPM is
Nevertheless, it implies a series of negative effects on so-
very suitable for predicting large scale emissions such as
ciety: pollution, noise, accidents, etc. The recent evolution
for estimating the spatial distribution profiles of air pollutant
of pollution management has been greatly influenced by the
levels [7], the photochemistry incorporated in the model
successive introduction of new emission directives, focusing
may not be suitable for examining small perturbations in
on the reduction of vehicle pollutant emissions ([1], [2]).
emissions inventories particularly in vehicular emission such
On one side, the combined efforts of vehicle manufacturers,
as ramps, traffic and intersections traffic-signal control, and
oil operators and the administrative institutions of different
intelligent transportation systems since it does not address
countries have reduced local emissions to a great extent.
driving dynamics [8]. Another model is the MOBILE 6.2, de-
On the other side, the constant growth of the proportion
veloped by the US Environmental Protection Agency (EPA).
of modern vehicles has not been sufficient to constrain the
The most important input parameters are vehicle miles
emission of global contaminants emitted in direct relation to
travelled (VMT, or vehicle kilometres travelled, VKT) for
fossil-fuel consumption, which is the main energy source in
various vehicle categories and average speed. The European-
vehicles. According to a world report of historical vehicle-
based Computer Programme to Calculate Emission from
population trends [3], global registrations jumped from 980
Road Transport (COPERT) is another model which is widely
million units in 2009 to 1.015 billion in 2010, the 3.6%
accepted and used in most countries and regional emission
rise in vehicle population was the largest percentage increase
inventories [9]. Although this model is rather simple, it also
since 2000. According to the Australian Bureau of Statistics
uses an average speed for driving dynamics when calculating
[4], the total motor vehicles, including motorcycles was
vehicle emissions. As a matter of fact, driving dynamics
16.7 million during 2012 Motor Vehicle Census (MVC) in
cannot be appropriately characterized by the average speed
since a large number of different driving patterns can result
S. D. Oduro, S. Metia, and Q. P. Ha are with Faculty in the same average speed.
of Engineering, University of Technology, Sydney, Australia The driving mode is more suitable to represent driving
[Link]@[Link]
H. Duc is with Department of Environment and Conservation NSW, dynamics, given that emission rates largely depend on the
Sydney, Australia characteristics of traffic, vehicles and type of road inter-

978-1-4799-0572-0/13/$31.00 ©2013 IEEE 158

 ICCAIS 2013: Main Track ICCAIS 2013

sections. For example, factors like the type, size, age of a

vehicle, and condition of its engine, type and condition of
emission control equipment, engine characteristics, vehicle
maintenance and weight can be correlated to the emissions.
Engine size also affects the functioning of emission control
equipment [10]. Age of a vehicle and the poor maintenance
adds to the emissions of all classes of vehicles. Further,
the fuel quality has a direct effect on the vehicular exhaust
emission [11].
Due to difficulties of incorporating acceleration or
deceleration in a macro scale emission model, e.g., the
MOBILE and COPERT, there is therefore a need to develop
a small scale model which can predict vehicular emissions
during driving modes such as change in ramps, intersections
and traffic-signal controls. A good understanding of the
driving mode impact on emissions would improve existing
vehicular emission models [12]. This has motivated an effort
for modelling emissions as a function of vehicular speed Fig. 1. Trend of speed, acceleration/deceleration and CO2 emission factor.
and acceleration instead of average speed. Here, real-world
vehicular emission data collected second by second for
on-road traffic flow and at traffic intersections are used in and pressure; and also ambient temperature, pressure and
a statistical analysis to predict the CO2 emission factor. humidity on a second-by-second basis. The instrument used
Instantaneous speed and acceleration are used in this study has a special data logging software to log the correct values
to represent driving modes instead of the average speed. of the measured emissions and other required parameters. In
The main objective of this research focuses on developing addition, it was estimated that there would be a 1 second
a micro scale emission model that can show an explicit delay in the logging of the NOx data and 3 seconds for CO2
relationship between CO2 emissions and instantaneous data. The instrument manaufacturer attributed this delay to
speed and acceleration as predictor variables. Such a CO2 the time it took to convert the measured concentration from
emission model may provide input to strategic policies analog to digital output on the laptop. The data analysis
towards the greenhouse gase emission reduction. spreadsheets were adjusted to account for the delay in data
processing.
The scope of the study was restricted to the CO2
II. D ESCRIPTION OF THE CASE STUDY emission factor. As shown in Fig. 1, speed and
The study utilized secondary data corrected on a second acceleration/deceleration are plotted against time to
by second basis by the New South Wales (NSW) Road and study the trend of the predictor variables (speed and
Maritime Service (RMS), Department of Vehicle Emission, acceleration/deceleration) and the response of CO2 emission
Compliance Technology and Compliance Operation. The factor. The graph clearly shows that, the emission factor
vehicles included in the test were Toyota, Ford, Holden of CO2 emitted from the exhaust tail pipe is consistent
and Nissan. A total of ten (10) vehicles were selected from with the increase and decrease in speed and acceleration.
the total fleet with five (5) each manufactured in 2002 and According to [13], as the speed of the vehicle increases
2008, respectively. These represent the bulk of motor vehicle due to an acceleration, more fuel is burnt in the combustion
fleet in Sydney in particular, and in Australia generally. The chamber which emits more carbon dioxide from the engine.
2002 vehicles test was conducted in June 2008 while the
2008 model vehicles were tested in May 2012. The vehicle
III. S TATISTICAL A PPROACH
specifications include engine capacities, from 2494 cm3 to
2498 cm3 , multi-point fuel injection system, and 4-cylinder In this paper, multiple linear regression (MLR) is em-
engine with a fuel tank of 70 litres using unleaded petrol. ployed with speed and acceleration/deceleration as predictor
The instrument used for the measurement was an Horiba variables and CO2 emission factor as the outcome vari-
on-board system (OBS) measurement instrument OBS-1300, able. MLR is a statistical procedure that determines the
which consists mainly of a MEXA-1170 HNDIR analyzer, best fitting regression function (a regression surface or a
data integration unit which houses a MEXA-720 NOx an- response surface) among several variables. In general, the
alyzer, a power supply unit, a data logger PC and other linear regression model can be represented as follows:
accessories. The OBS and its accessories were installed in y = β0 + β1 X1 + β2 X2 + ... + βnX XnX + ε, (1)
the test vehicles. The NOx sensor was attached to the tail
pipe while the test vehicle is being driven down the road, where, for a set of observations, Yi is the outcome variable;
the OBS system logged concentrations of HC, CO, CO2 , and β0 , β1 , β2 ,...βnX are coefficients; X1 , X2 ,...XnX independent
NOx as well as data of vehicle velocity, exhaust temperature variables (predictors) and ε is the residual error (difference

978-1-4799-0572-0/13/$31.00 ©2013 IEEE 159

 ICCAIS 2013: Main Track ICCAIS 2013

between observations and predicted values). The linear re- σ 2 constant variance. The goal of the regression analysis is
gression model can be predicted on a relation of the form: to determine the values of the parameters of the regression
nX equation and then to quantify the good fit with respect to the
y = β0 + ∑ β j X j + ε, (2) dependent variable y. The completion of the data collection
j=1 phase results in the accumulation of a large amount of CO2
where ε is a random error term with an expected value of emission data, which is used as the response variable to speed
zero (i.e., E(ε) = 0). The approximate form of the relation and acceleration/deceleration representing the predictor (in-
in (2) is dependent) variables. A statistical package for social science
nX (IBM SPSS version 20) is employed for the data analysis.
ŷ = b0 + ∑ b j X j , (3) The model estimation processes involve:
j=1
(a) Fitting a preliminary model, then plotting the re-
where b j are typically estimated with least square procedures
sponse variable against each predictor variable.
from observations of the form [xi j , yi] =1, 2, ...nS , where
(b) Checking model assumptions by using the plots of
nS is the size of the data samples. In turn, the preceding
residuals vs. predicted values to verify that the error
approximation is often algebraically reformulated as,
variance is constant.
nX
(c) Using the plots of the residuals versus predicted
(ŷ − ȳ)/ŝ = ∑ (b j ŝ j )(x j − x̄ j )/ŝ, (4)
j=1
values to verify that the current model form is
appropriate, and
where (d) Using a normal probability plot of the residuals to
v
nS u nS
u verify that the errors are normally distributed.
ȳ = ∑ yi /nS , ŝ = t ∑ (yi − ȳ)2 /(nS − 1),
i=1 j=1 To find the unbiased estimators (Regression Coefficient,
v
u nS b1 ) for unknown parameters β1 , the least squares method
nS u is used to minimize the residuals error, i.e., εi = yi − ŷi ).
x̄ j = ∑ xi j /nS , ŝ j = t ∑ (xi j − x̄ j )2 /(nS − 1).
i=1 j=1
After regressing the outcome yi , here for the Carbon
Dioxide concentration, on x1 for speed, and x2 for
The coefficients b j ŝ j /ŝ in (4) are called standardized re- acceleration/deceleration, the SPSS results obtained are
gression coefficients. When x j are independent, b j ŝ j /ŝ can presented and discussed in the following section.
be used as measure of variable importance. Specifically,
a b j sˆj /ŝ indicates the effect of moving a variable away
from its expected value by a fixed fraction of its standard IV. R ESULTS AND D ISCUSSION
deviation while holding all other variables fixed at their The independent variables used for model identification
expected values. Applying a least-square estimated procedure are obtained through a sequential regression analysis. The
the following identity holds in relation to (2), CO2 regression statistic summary for 2002 and 2008 vehicles
nS nS nS is presented in Table I. Models 1 refer to the first stage in the
∑ (yi − ȳ)2 = ∑ (ŷi − ȳ)2 + ∑ (ŷi − yi )2 , (5) sequential when only speed was used as a predictor variable.
i=1 i=1 i=1
Models 2 are when both speed and acceleration/deceleration
where ŷi denotes the estimate of yi obtained from the regres- were used. The model summary compares the statistical
sion model. In order from left to right, the three summations charateristics of the two models in predicting CO2 emission
in the preceding equation are referred to as the total sum of from the vehicular exhaust tail pipe.
squares SStot , the regression sum of squares SSreg and the An evaluation of the linear relationship between y = CO2
residual sum of squares SSres . The latter, SSres , provides a concentration, speed (x1 ) and acceleration/deceleration (x2 )
measure of variability about the CO2 regression model, while was measured using Pearson’s correlation. Table I shows that
the R2 defined by, there exists a significant linear relationship between the CO2 ,
nS C speed and acceleration/deceleration. With nS =1200, nX =2,
R2 = SSreg /SStot = ∑ (ŷi − ȳ)2 / ∑ (yi − ȳ)2 , (6) and the probability of obtaining a statistic test for occurring
i=1 i=1 by chance no input/output relationship from the measured
provides a measure of the extent to which the regression data, or the p-value, chosen as 0.005, we obtained a corre-
model can match the observed results. Specifically, R2 is lation coefficient R=0.73 for 2002 vehicles and R=0.79 for
close to 1 when the variation about the regression model 2008 vehicles, i.e. a very strong positive linear relationship.
is small which indicates a good fit of the model and how The R2 value is provided here as a measure of how much
the regression model is successfully in matching with the of the variability in the CO2 vehicular emission is accounted
observed results. The hypotheses required to apply multiple for by the predictors (speed and acceleration/deceleration).
CO2 linear regression are that (i) the predictor variables As this measure tends to over estimate the strength of the
must be independent, and (ii) the residual errors εi must be association, we also used the adjusted square, R2ad j , for the
independent and normally distributed with zero mean and number of predictors in the model:

978-1-4799-0572-0/13/$31.00 ©2013 IEEE 160

 ICCAIS 2013: Main Track ICCAIS 2013

TABLE I
CO2 REGRESSION MODEL SUMMARY STATISTICS
  
k n−1
Rad j = R2 −
2
, (7)
n−1 n−k−1
where n = nS is the number of participants and k = nX is the Year Model R R2 Adjusted Standard Error
R2 of Estimation
number of predictors in the model.
2002 1 0.73 0.52 0.52 1.14
Here for models 1, R2 value for 2002 and 2008 were 2 0.77 0.60 0.60 1.37
0.52 and 0.61 respectively, which means that speed ac- 2008 1 0.79 0.61 0.61 1.13
counts for 52%, and 61% of the variation in CO2 emis- 2 0.88 0.73 0.73 1.23
sion. However, when speed and acceleration/deceleration was a. Predictors: constant speed
used as predictors as in models 2, R2 increased to 0.60 b. Predictors: constant Acceleration/Deceleration
(60%) and 0.73 (73%) of the variance in CO2 emission. c. Dependent variable CO2
Therefore, if speed accounts for 52% and 61%, we can
conclude that the other predictor accounts for additional 8% TABLE II
in 2002 vehicles and 12% in 2008 vehicles which means STANDARDIZED AND UNSTANDARDIZED PARAMETERS ESTIMATE
the inclusion of acceleration/deceleration has contributed to
a certain amount of the variation in CO2 emission with Year Model Unstandardized Unstandardized Standardized
2008 vehicles having the highest contribution. The results Coefficients Coefficients Coefficients
clearly shows that the vehicle age plays an inportant role bj Standard Error Beta
in affecting the emission rates. The older vehicels emit 2002 [Link] 0.578 1.46
higher emissions than the newer ones due to recent strict Speed 0.041 0.003 0.717
[Link] 0.569 0.467
emission controls like catalytic converter, on-board vapour Speed 0.041 0.002 0.720
recovery systems and improve fuel efficiency. This result is Acceleration 0.025 0.016 0.204
in line with [14], where, by applying microscopic models 2008 [Link] 0.730 1.32
to quantify the impact of vehicle speed and acceleration on Speed 0.034 0.002 0.622
the fuel consumption and emission rates, it is found that [Link] 0.729 0.181
Speed 0.034 0.002 0.693
the vehicle fuel consumption and emission level depends Acceleration 0.008 0.001 0.212
on the vehicular operation conditions. Thus, the changes Note:
in the vehicle fuel consumption and the emission rates are p<0.005
mostly associated with changes in the vehicle speed and
the driver acceleration aggressiveness. Similarly, as reported
in [15] from field data gathered under real-world driving
conditions, CO2 emissions usually increase with increasing As reported in [13], for complete combustion, about 99%
engine speed and acceleration. This is further confirmed of the carbon in the burnt fuel is emitted as CO2 . There is
in [16] that driving pattern is generally represented by a thus a strong linear relationship between the fuel flow and
speed profile of the vehicle but can be expanded to include CO2 emissions. Hence, both fuel flow and CO2 emission
also driving behaviours such as acceleration, load and gear increase with speed increase due to an acceleration. The cor-
change. All these studies justify for the incorporation of relation between CO2 and speed and acceleration is therefore
speed and acceleration as good predictor variables for the noticeable through linear trends showing the relationships.
vehicular CO2 emission factor. The b-values also explain to what degree each predictor
Table II summarizes the coefficients of the regression affects the outcome in the case when all other predictors are
model parameters for both steps in the sequential order. held constant. Here in 2002 vehicles, the coefficient for speed
Accordingly, the estimated regression functions for 2002 and (b1 =0.041) could be interpreted that increasing the vehicle’s
2008 vehicles are, respectively, speed by 1m/s will incur a rate 0.041g/s of CO2 whereas
for 2008 vehicles, increasing the vehicle’s speed by 1m/s
y = 0.578 + 0.041x1 , (8) will incur a rate of 0.034g/s of CO2 emission. Similarly,
an increase in acceleration for 2002 and 2008 vehicles by
y = 0.730 + 0.034x1 , (9)
1m/s2 will produce b2 =0.025g/s of CO2 and b2 =0.008g/s
for Model 1. These functions for Model 2 are of CO2 emission. As given in Table II, each of the beta
coefficients has an associated standard error indicating if
y = 0.569 + 0.041x1 + 0.025x2 , (10) the corresponding b-value differs significantly from zero.
This is related to the fact that at intersections and traffic
y = 0.729 + 0.034x1 + 0.008x2 , (11)
signals, acceleration events are likely to occur with a higher
also for 2002 and 2008 vehicles, respectively. Since all frequency.
the two predictors have positive b-coefficients, a positive In most cases, driving through an intersections, speed
relationship is indicated. So, as the speed increases due ramps or traffic signals entails stopping or slowing down,
to increase in acceleration, the CO2 emission factor also and resulting in acceleration events that are associated
increases. with higher emissions [17]. Previous researchers have also

978-1-4799-0572-0/13/$31.00 ©2013 IEEE 161

 ICCAIS 2013: Main Track ICCAIS 2013

TABLE III
t- STASTISTICS AND CORRELATION SUMMARY

Correlerations
Model t-value significance zero-order partial part
[Link] 4.21 0.000
Speed 16.2 0.000 0.74 0.74 0.74
[Link] 4.76 0.000 0.71
Speed 17.5 0.000 0.74 0.76 0.73
Acceleration 6.98 0.000 0.29 0.39 0.27
Note:p<0.005

Fig. 3. 2008 CO2 normal P-P plot of distributed residual.

Fig. 2. 2002 CO2 normal P-P plot of distributed residual.

compared emissions at roundabout, all-way-stop, and traf-

fic signals [18]. However, predicting emission as explicit
functions of speed and acceleration will provide a direct Fig. 4. 2002 vehicle’s standardized residuals against standardized predicted
values
way for estimation of vehicular emissions. Table III presents
the t-test, which characterizes the difference in significant
contributions made to the model from each of the predictors straight line in this plot represents a normal distribution with
(speed and acceleration). The measures include t-value as the points therein representing observed residuals. From 2008
function of n − k − 1 as in (7), statistical significance as per vehicles CO2 data, the distribution is roughly normal than
p-value, as well as zero-order, partial and semi-partial (part) the 2002 vehicles (although there is a slight deficiency of
correlations. residuals exactly on zero) the observed residuals are very
From model (10) and (11), with nS =1200, nX =2 and close to the line, indicating just a slight deviation from
p<0.001, the statistical t-value for speed is 17.5, and for normality and not a skewed distribution of data. Hence, the
acceleration is 6.98, indicating that the predictors of CO2 model could predict the emission factor quite well.
emission here are all significant. From the magnitude of An evaluation of the multi collinearity between the CO2 ,
the t-statistics, the speed has a greater impact predicting speed and acceleration can be obtained by using Variance
CO2 emission than acceleration. In order to cross-validate Inflation Factor (VIF). For an estimated b j , VIF(b j ) is
the model to see how well the model can be generalized defined as 1/(1 − R2j ), where R2j is correspondingly the
or predict the outcome of different sample, we refer to (7) coefficient of determination for X j with respect to another
for adjusted R-squared. The value obtained, R2ad j = 0.72, is predictor variable. It was observed that there is a strong
almost similar to the observed value of R2 (0.73), indicating linear relationship between the CO2 emission factor and
that the cross-validity of the model is very good. speed whereas that and the acceleration exhibits a rather
For further cross validation of the model to assess the variables have nonlinear relationship. The obtained VIF value
accuracy, Figs. 2 and 3 show a normal probability plot for of 1.021 confirms that there is no collinearity between the
2002 and 2008 vehicles. In Fig. 2, the probability-probability vehicle’s speed and acceleration. Figs. 4 and 5 show the
(P-P) plot shows a greater deviations from normality. The plots of standardized residuals against standardized predicted

978-1-4799-0572-0/13/$31.00 ©2013 IEEE 162

 ICCAIS 2013: Main Track ICCAIS 2013

rates are found to be higher during acceleration, which means

more fuel is required during a sudden depression on the
accelerator pedal.

VI. ACKNOLWLEDGEMENTS
All the data used for this study were supplied by Road and
Maritime Service, Department of vehicle emission, Compli-
ance Technology & Compliance Operations and NSW Office
of Environment & Heritage. Assistance provided by Paul
Walker and Gareth Jones is gratefully acknowledged.

R EFERENCES
[1] H. C. Frey, A. Unal, N.M. Rouphail, and [Link], “On-Road Mea-
surement of Vehicle Tailpipe Emissions Using a Portable Instrument”,
Journal of the Air & Waste Management Association,Vol. 53, pp. 992-
Fig. 5. 2008 vehicle’s standardized residuals against standardized predicted 1002, August 2003.
values [2] EEA, “Gasoline Evaporation from Vehicles”, EMEP/EEA emission in-
ventory guidebook, 2002,version [Link]
[3] [Link] vehicle population 110815
values of 2002 and 2008 vehicles CO2 emission factor to [4] Australian Bureau of statistics, “Motor Vehicle Census,” Australia
2012. Available: [Link]
assess how the data were dispersed. It can be seen that the [5] Australian Government, Department of Climate Change, “Emissions
points are randomly but evenly dispersed throughout the plots projections,” 2009. Available: [Link]
for 2002 and 2008 vehicles. The pattern clearly favours the [6] Australian Government, Department of Environment, Water,
Heritage and the Arts, “Annual Report,” 2009. Available:
assumptions of linearity and homoscedasticity in the regres- [Link]
sion. This means that the model appears to be both accurate 10
for the sample and generalizable to the population with the [7] H. Wahid, Q. P. Ha, H. Duc, and M. Azzi, Neural network-based meta-
modelling approach for estimating spatial distribution of air pollutant
same variance of dependent variables. This also suggests levels, Applied Soft Computing, 2013, doi 10.1016/[Link].2013.05.007
that using speed and acceleration as predictor variables can [8] P.J. Hurley, “TAPM V4. Part1: Technical Description” CSIRO Marine
improve accuracy in prediction of vehicles’ CO2 emission and Atmospheric Research Paper No.25.59pp, 2008.
[9] [Link] and Z. Samaras, “COPERT III: computer programme
factor. Furthermore, the developed models, containing an to calculate emissions from road transport-methodology and emission
additional variable, the acceleration, also accounts for the factors”. European topic centre on air emissions, EEA, Copenhagen,
vehicular dynamics involved. To improve the accuracy of air Denmark, 2000
[10] S. Samuel, L. Austin and D. Morrey, “Automotive test drive cycles
quality models, future work will incorporate other driving for emission measurement and real-world emission levels– a review,”
modes such as engine power, idling, time and gear changing Proceedings of the Institution of Mechanical Engineers, Part D,
data in the model calculations. Similarly, other emissions Journal of Automotive Engineering 216,555-564. 2002.
[11] H.C. Frey and K. Kim, “Comparison of real-world fuel use and
such as NOx, CO, HC will also be considered as they are emissions for dump trucks fueled with b20 biodiesel versus petroleum
important in the study of air pollution dispersion model. diesel.”Transportation Research Record 1987, 110-117, 2006.
[12] J. Rong, K. Mao, and J. Ma, “Effects of Individual Differences on
V. C ONCLUSION Driving Behavior and Traffic Flow Characteristics.” Annual Meeting of
the Transportation Research Board Washington, D.C.: Transportation
A number of vehicular emission models have been devel- Research Board, pp. 1-15, 2011
oped in recent years to estimate vehicular emission factors. [13] [Link], A. Unal, J. Chen, S. Li and C. Xuan, “Methodology for
Developing Modal Emission Rates for EPAs Multi-Scale Motor Vehi-
Most studies, concentrated on the average speed as the only cle & Equipment Emission System.” Publication EPA420-R-02-027,
predictor variable for a macro-scale emission factor model, Prepared by North Carolina State University for US Environmental
cannot predict vehicular emission in smaller areas such Protection Agency, Ann Arbor, 2002.
[14] H. Rakha and Y. Ding, “Impact of stops on vehicle fuel consumption
as traffics, intersections and ramps. The present study has and emission.” Journal of Transportation Engineering 129, 2332,
improved the CO2 emission modeling by using an statistical 2003.
analysis for second-by-second data collected in real-world [15] I. El-Shawarby, A. Kyoungho, and H. Rakha, “Comparative field
evaluation of vehicle cruise speed and acceleration level impacts on
conditions to predict the emission factor as explicit functions hot stabilized emissions.” Transportation Research Part D 10, 13-30,
of speed and acceleration using 2002 and 2008 vehicles. The 2005.
results presented for 1200 samples and significance p<0.005 [16] E. Ericsson, “Variability in urban driving pattern.” Transportation
Research Part D 5, 337-354, 2000.
have shown that vehicle speed is strongly correlated to the [17] N. M. Rouphail, H. C. Frey, and J. D. Unal, “Vehicle Emissions
CO2 emission factor with positive relationships. From the and Traffic Measures: Exploratory Analysis of Field Observations at
statistic measures it is observed that the speed of 2002 and Signalized Arterials.” Annual Meeting of the Transportation Research
Board. Washington D.C: Transportation Research Board, 2012.
2008 vehicles accounts for 52% and 61% variance in CO2 [18] K. Ahn and H. Rakha, “The Effects of Route Choice Decisions on Ve-
emission whereas with speed and acceleration, this increases hicle Energy Consumption and Emissions.” Transportation Research
to 60% and 73%. However, both contributions are all signif- Part D, 151-167, 2008.
icant to the CO2 emission model. Also, the CO2 emission

978-1-4799-0572-0/13/$31.00 ©2013 IEEE 163