Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Environ. Sci. Technol.

2010, 44, 3594–3600

Emissions factor models should quantify the relation-


Vehicle-Specific Emissions Modeling ship between emissions, vehicle dynamics (e.g., speed and
Based upon on-Road Measurements acceleration), and roadway infrastructure (e.g., road grade)
at multiple scales (1). Since microscale emissions estimates
can be aggregated over space and time to produce meso-
H. CHRISTOPHER FREY,* or macroscale estimates, an ability to accurately estimate
KAISHAN ZHANG, AND microscale emission factors is a key step toward develop-
NAGUI M. ROUPHAIL ment of a highly flexible vehicle emissions estimation
Department of Civil, Construction and Environmental system.
Engineering, North Carolina State University, Currently, the most widely used emission factor model,
Campus Box 7908, Raleigh, North Carolina 27695-7908 MOBILE6 (1) in the U.S. and COPERT (4) in Europe, is
appropriate for macro- and mesoscale estimates of vehicle
Received September 18, 2009. Revised manuscript received fleet emissions but is not applicable at the microscale or to
March 7, 2010. Accepted March 24, 2010. individual vehicles. Alternative methods for emissions es-
timation applicable to the meso- and microscale have been
explored for a number of years, including stratification of
empirical data, statistical regression, neural network, and
Vehicle-specific microscale fuel use and emissions rate physically based approaches. Modal-based approaches in-
models are developed based upon real-world hot-stabilized volve stratification of data based on speed and acceleration,
tailpipe measurements made using a portable emissions vehicle specific power (VSP), or other factors and are derived
measurement system. Consecutive averaging periods of one either based on a priori definitions of modes or from a
to three multiples of the response time are used to compare two statistical analysis of data using methods such as hierarchical
semiempirical physically based modeling schemes. One tree-based regression (5). Examples of these types of models
scheme is based on internally observable variables (IOVs), include HBEFA (6), ARTEMIS (7), Ecogest (8) in Europe, and
CORSIM (9) in the U.S. Modal-based approaches are the
such as engine speed and manifold absolute pressure, while
foundation of the MOVES model that is under development
the other is based on externally observable variables (EOVs), to replace MOBILE6 (10). Statistically based approaches
such as speed, acceleration, and road grade. For NO, HC, include least-squares linear regressions and time series
and CO emission rates, the average R2 ranged from 0.41 to 0.66 modeling, among others (11-14). A neural network attempts
for the former and from 0.17 to 0.30 for the latter. The EOV to predict emissions by training the network with known
models have R2 for CO2 of 0.43 to 0.79 versus 0.99 for the IOV emissions and vehicle activity data sets (15). Physically based
models. The models are sensitive to episodic events in approaches typically attempt to explicitly model the physics
driving cycles such as high acceleration. Intervehicle and and chemistry of pollutant formation and control taking into
fleet average modeling approaches are compared; the former account detailed characteristics of the vehicle and its
account for microscale variations that might be useful for operation (16-18). Examples include the Comprehensive
Modal Emissions Model (CMEM) and Physical Emission Rate
some types of assessments. EOV-based models have practical
Estimator (PERE) models (17, 18).
value for traffic management or simulation applications since
The advantage of the modal approaches is that they do
IOVs usually are not available or not used for emission estimation. not require fitting of an analytical model and thus do not
introduce errors that might occur if the model is not a good
Introduction fit to the data. However, such models inherently involve a
The purpose of this paper is to develop an approach for significant amount of averaging of data, since a mode typically
microscale estimation of vehicle-specific emissions based represents the average emissions for a significant proportion
on real-world data, in order to support long-term future of the vehicle activity pattern. This may be adequate for
development of more accurate and robust approaches for mesoscale purposes but can reduce the model’s explanatory
emissions estimation. An accurate assessment of highway power and potentially limit its usefulness for microscale
vehicle emissions is essential for effective air-quality man- applications.
agement, and emission estimates are needed at various spatial Statistically based approaches typically involve fitting a
and temporal scales, including macro-, meso-, and microscale model to empirical data. For some types of models, such as
(1). Macroscale refers to urban or regional emission inven- higher order polynomials, the terms in the model may or
tories based on aggregated data. Mesoscale refers to sub- may not have any physical interpretation. For example, the
regional estimates, such as for urban corridors. Microscale product of speed (v) and acceleration (a) is a surrogate for
refers to nearly instantaneous vehicle emissions at specific power demand (av), while a3v3 does not have an explicit
locations, such as an intersection or segment of roadway. physical interpretation or relevance. However, if properly
These different scales support different types of analyses. formulated, statistically based models can be both physically
For example, many existing traffic demand models produce plausible while offering a useful degree of explanatory power.
mesoscale estimates of average vehicle speeds for roadway A key challenge, however, is whether such models can be
links (2). However, because vehicle emissions are highly developed appropriately for use in microscale emissions
episodic, there is a critical need to better quantify the local estimation.
effect of traffic control measures and other factors that The physically based approaches typically have sub-
contribute to variability in vehicle speed and acceleration stantial requirements for input data that limit their
(1). For example, microscopic simulation models such as practicality for general application. However, simplified
AIMSUN (3) need to be calibrated using microscale fuel and models that have a physical basis may offer a substantial
emissions data from real-world measurements. degree of explanatory power. For example, VSP, which is
a function of speed, acceleration, and road grade, is an
* Corresponding author e-mail: frey@ncsu.edu. excellent predictor of vehicle fuel use (16). Because
3594 9 ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 44, NO. 9, 2010 10.1021/es902835h  2010 American Chemical Society
Published on Web 04/08/2010
emissions of many pollutants have a significant relation- In some cases, a model with a physically based inter-
ship to fuel use, VSP is also a useful explanatory variable pretation cannot be simplified to a linear combination of
for vehicle emissions (5). basis functions and may be inherently nonlinear. In these
The objectives of this paper are to (1) characterize and cases, nonlinear regression is used. Nonlinear regression is
account for autocorrelation in vehicle fuel use and emission more sensitive to outliers (25).
rates; (2) compare and evaluate alternative microscale An assessment was made of the use of the Box-Cox
vehicle-specific models based on EOVs and IOVs; and (3) transformation (BCT) in order to eliminate heteroscedasticity
compare vehicle-specific versus fleet average models. in the residuals. However, when variables are back-
transformed, predictions of the model are biased and have
to be corrected. The bias-corrected back-transformed models
Experimental Section were found to perform more poorly than models fit directly
This section describes methodologies for (1) on-road data to the data without the BCT. Therefore, BCT was not further
collection; (2) development of data sets for model building applied.
and validation; (3) selection of vehicle-specific fuel use and Averaging Time. The assumption of iid residuals is often
emissions modeling approaches; and (4) model evaluation. not valid for models fit to raw second-by-second emissions
On-Road Data Collection. In order to capture a substantial data, since the vehicle activity and emissions in one second
range of intravehicle variability, data were collected exten- is dependent on activity and emissions in recent seconds.
sively for three “primary” vehicles using a portable emissions For example, an engine typically does not run at high RPM
measurement system (PEMS) and were supplemented with in one second unless it had been run at moderate to high
fewer data on each of seven additional secondary vehicles RPM in the previous second. The response time of the PEMS
(19). The measurements were made using the Montana emissions sensors are longer than the rate at which data are
system PEMS manufactured by Clean Air Technology In- reported (26, 27), which imposes a correlation in the
ternational, Inc. measured values even if the true emissions are changing
The primary vehicles were operated by three drivers on independently every second. A technique for reducing the
three alternative routes between each of two origin/destina- influence of autocorrelation and accounting for response
tion (O/D) pairs. On average, there were 65 h of second- time is to stratify the data and work with average, rather than
by-second data collected over a three week period for each continuous, values in each strata. Therefore, models were
primary vehicle, involving an average of 14 runs on each developed based on consecutive averages of vehicle activity
route and travel direction. The routes included a variety of and emission data that were equal to or greater than the
roadway facility types (i.e., feeder/collector street, minor response time of the emissions sensors and that include the
arterial, major arterial, freeway) and road grades (varying most significant autocorrelated lags.
between approximately plus or minus eight percent). Data Modeling Schemes. The vehicle-specific microscale emis-
were collected in both travel directions on each route for sions models were developed taking into consideration their
both morning and afternoon weekday time periods, thereby potential applications. One possible future application is to
capturing both peak and off-peak traffic flow. couple such models with transportation microsimulation
Based on analysis of data collected for the primary models that estimate externally (to the vehicle) observable
vehicles, an estimate was made of the minimum field data information such as speed and acceleration of individual
requirement necessary to adequately quantify intravehicle vehicles on a roadway network as well as characterize
variability in emissions. For each secondary vehicle, data infrastructure features such as road grade. Another is to use
collection occurred on preferred routes typically during one the models as part of an in-vehicle information system that
day by one driver. could report to a driver or a traffic management facility real-
Development of Data Sets for Model Building and time fuel use and emissions. Such information could guide
Validation. The field data were stratified into two data sets. real-time decision making regarding driver behavior, traffic
One was used to build models, and the other was used to control, or routing. An in-vehicle model could access
validate the models. In order for both data sets to have similar internally observable data from the on-board diagnostic
vehicle dynamics, trips from the same route were randomly (OBD) system, such as engine speed (ES), intake air tem-
but proportionally selected form each data set. Approximately perature (IAT), and manifold absolute pressure (MAP). A key
75% of all trips on a route were selected to populate the question is whether a model based on internally observable
modeling data set, and the other 25% were included in the variables (IOVs) is more accurate or precise than one based
validation data set. A Kolmogorov-Smirnov (K-S) two on externally observable variables (EOVs).
sample test (20) was used to confirm that the two data sets Internally Observable Variables Model (IOVM). ES and MAP
had nonsignificant differences in vehicle speed and ac- were used as explanatory variables because they are sig-
celeration distributions. nificant factors in fuel injection control (28) and are strongly
Selection of Vehicle-Specific Fuel Use and Emissions influenced by vehicle dynamics such as speed and acceleration.
Modeling Approaches. Model specification takes into ac- The formation of CO and HC are intimately coupled with
count the selection of a regression methodology, averaging the fuel combustion process whereas NO formation is
time, and explanatory variables. assumed to occur as part of postflame processes (28). These
Regression Approach. Both nonparametric and parametric processes account for engine-out emissions prior to the
regression approaches can be used for emissions modeling. catalytic converter. One modeling approach is used to
However, nonparametric regression (NPR) models do not estimate fuel use and emission rates of HC and CO, while
provide an explicit mathematical equation for model inter- another is used to model NO emissions. The former were
pretation (21-24) and thus were not used. modeled as (model 1 in Table 1)
A linear regression model can be a linear combination of
¯∆t
“basis functions” which may be nonlinear. In order to have m i ) aiP̄mapS̄engine∆t (1)
confidence in the predictive ability of a regression model,
each variable in the model should have a physical inter- where m j i∆t ) average mass flow rate for specie i (fuel, CO,
pretation. In order to produce unbiased estimates of the HC) for a consecutive averaging period ∆t, g/s, Pmap )
model parameters, the model residuals should have constant Manifold Absolute Pressure (MAP), kPa, Sengine ) Engine Speed
variance with a zero mean and be independently and (ES) in revolutions per minute, rpm, PjmapSengine∆t ) average
identically distributed (iid). of the product of Pmap and Sengine for a consecutive, averaging

VOL. 44, NO. 9, 2010 / ENVIRONMENTAL SCIENCE & TECHNOLOGY 9 3595


period ∆t, ai ) model coefficient for specie i, and ∆t )
duration of the averaging period (seconds). VSP ) v{a(1 + ε) + gr + gCR} +
2
Fv ( )
1 3 CDA
m
(4)
The NO formation rate in the engine is assumed to be
based on a simplified Zeldovich mechanism (28) where a ) vehicle acceleration (m/s2), A ) vehicle frontal
area (m2), CD ) aerodynamic drag coefficient (dimensionless),
CR ) rolling resistance coefficient (dimensionless, ∼ 0.0135),
d[NO] 6 × 1016 -69090 g ) acceleration of gravity (9.8 m/s2), m ) vehicle mass (in
dt
)
T0.5
exp
T (
[O2]e0.5[N2]e ) (2)
metric tons), r ) road grade, v ) vehicle speed (m/s), VSP
) Vehicle Specific Power (kw/ton), ε ) mass factor accounting
for the rotational masses (∼ 0.1), and F ) ambient air density
where (d[NO])/(dt) ) NO formation rate (mol/cm3-s), T ) (1.207 kg/m3 at 20 °C).
gas temperature in the cylinder (K), [O2]e ) equilibrium The aerodynamic drag term in eq 4 varies by vehicle. For
concentration of O2 (mol/cm3), and [N2]e ) equilibrium example, the difference could be a factor of 2 between a
concentration of N2 (mol/cm3). compact car and a full size passenger car (18). However, the
However, T, [O2]e, and [N2]e are not measured by PEMS aerodynamic drag term is relatively small for urban driving
or reported by OBD system, these are not observable. [O2]e compared to the other terms and does not significantly effect
and [N2]e are a function of the cylinder gas temperature and the estimates of VSP.
pressure. An assumption is made that the engine gas For each vehicle, individual terms in eq 4 were used as
temperature is proportional to fuel use. The resultant basis functions in a linear regression model (model 12 in
functional form is (model 2 in Table 1) Table 1)

( )
_ _ _ _ _
_∆t aNO -bNO mj∆t ) A0,j + Ajav∆t + Bjv∆t + Cjvr∆t + Djv3∆t + εj (5)
mNO ) exp (3)
(P̄mapS̄engine∆t)0.5 _∆t
PmapSengine where av j ∆t ) average of the product av, for a consecutive
averaging, period of duration ∆t (km2/h2s), εj ) model residual
for specie j, j ) specie (i.e., NO, HC, CO, CO2, and fuel use),
where m j NO∆t ) average mass emission rate for NO for a j j∆t ) average of mass emission rate specie j for a consecutive,
m
consecutive averaging period ∆t, g/s, and aNO, bNO ) model averaging period of duration ∆t (g/s), vj∆t ) average of v for
coefficients. a consecutive averaging period of duration ∆t (km/h), vr j ∆t
The “R” statistical package (29) was used to fit eqs 1 and ) average of vr for a consecutive averaging period of duration
3 to PEMS data for each vehicle. ∆t (km/h), vj3∆t ) average of v3 for a consecutive averaging
Externally Observable Variables Model (EOVM). VSP is a period of duration ∆t (km3/h3), and A0, j, A j, B j, C j, D j )
measure of load on a vehicle and is defined as the power per model coefficients for specie j. The model intercept A0,j
unit mass to overcome inertial acceleration (power demand), represents the fuel use or emissions rate when the vehicle
rolling resistance, road grade, and aerodynamic drag (16) is idling.

TABLE 1. Comparison of Goodness-of-Fit for Alternative Modeling Approachese


coefficient of determination (R2)
averaging intercept
model explanatory variables type of inputsd time (s) included NO HC CO fuel
a a
1 P̄mapSengine IOV 12/18 N 0.40 0.55 0.60 0.99
aNO
_
(PmapSengine) 0.5
exp
( -bNO
_
PmapSengine )
2a IOV 12/18a N 0.43 n/a n/a n/a
3b MERi,(i)1 to 14) EOV 1 NA 0.06 0.18 0.14 0.56
4 v,a,r EOV 1 Y 0.06 0.21 0.14 0.48
5 VSP EOV 1 Y 0.04 0.21 0.10 0.46
6 V̄SP EOV 5 Y 0.08 0.26 0.17 0.60
7 v, v3,av,r EOV 1 Y 0.06 0.24 0.16 0.54
8 v̄, v̄3, āv, r¯ EOV 5 Y 0.10 0.28 0.22 0.65
9 viaj(i,j)1,2,3) EOV 1 Y 0.07 0.18 0.16 0.56
10 VSPt, VSPt-1,VSPt-2,VSPt-3 EOV 1 Y 0.01 0.25 0.15 0.46
11c VSPt,εt EOV 1 Y 0.72 0.86 0.73 0.87
12a v̄, v̄3, āv, v̄r EOV 12/18a Y 0.16 0.22 0.32 0.71
13a v̄, v̄3, āv, v̄r EOV 12/18a N 0.35 0.48 0.58 0.89
14 v,a,r, Pmap,Sengine,CT,TP,IAT IOV and EOV 1 Y 0.17 0.46 0.22 0.95
a
Consecutive averages of both dependent and independent variables were used for modeling. The averaging time
interval is 18 s for NO and is 12 s for the others. b MER - modal emission rates. In this model, measured values were
compared with VSP-based modal emission rates, which are very similar to the MOVES model. The MOVES model uses
VSP-based modes that are also stratified by speed. The modal modeling approach in MOVES is based in part on a
comparison of modal modeling approaches evaluated by Frey et al. (5) that included the same VSP modal method as used
here, along with a method based on binning by both VSP and speed. The results for estimates of cycle average emission
rates are very similar for these two modal binning approaches. c This model is time series regression with serial errors
(TSRSE). The second term (vt) in this model is further modeled as a 5th order autoregressive (AR) model. Higher order AR
model did not significantly improve the explanatory power. d IOV - internal observable variables; EOV - external observable
variables. e Models were fit to PEMS data from a 2005 Chevrolet Cavalier using linear regression unless noted. “X j ” in this
table indicates that consecutive averages of both dependent and independent variables are used for modeling.

3596 9 ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 44, NO. 9, 2010


Model Evaluation. The purpose of model evaluation is to by considering lag effects in the explanatory variables, as
(1) compare and select from among several alternative embodied in model 10, produces a fairly poor goodness-
averaging times; (2) demonstrate the sensitivity of the model of-fit for all pollutants and thus is not an effective approach.
predictions to microscale features of driving cycles such as An alternative is to use consecutive averaging times in an
idle, acceleration, cruise, and deceleration; and (3) demon- attempt to “average out” autocorrelation effects. Models 1,
strate the application of the models to predict fuel use and 2, and 13 use consecutive averages and produce R2 values
emissions. that are quite high for fuel use (and CO2 emissions) and imply
Both IOVM and EOVM were developed from modeling that approximately 35 to 60% of the variability in the
data sets for each vehicle. These models were used to predict emissions of the other pollutants can be explained. Overall,
emissions and fuel use for comparison to validation values. for IOVMs, models 1 and 2 are the most practical basis for
The model development and comparisons were done for more extensive exploration; for EOVMs, the approach of
various averaging times. model 12 is chosen instead of model 13 because (1) both
In order to evaluate the sensitivity of the models to have the same modeling structure except model 13 does not
different driving cycles (intercycle variability) and to assess include an intercept and (2) excluding an intercept in a linear
the ability of the models to make microscale predictions for regression model provides a better goodness-of-fit but results
a given cycle (intracycle variability), the models were used in biased estimates of model coefficients. However, an
to predict fuel use and emissions for a selected set of standard intercept is not included in model 1 because ES and MAP
and empirical driving cycles. These cycles included four cannot be zero if the engine is on, whereas vehicle speed can
chassis dynamometer cycles and four representative real- be zero. For model 2, an intercept is not necessary because
world cycles from field data collection. The four dynamometer nonlinear regression is used.
cycles include Bags 1, 2, and 3 of the Federal Test Procedure Model Evaluation. IOVMs and EOVMs were developed
(FTP), US06, New York City Cycle (NYCC), and California for each of the three primary and seven secondary vehicles.
LA92. Each model was fit to the modeling data set. The goodness-
For the EOVMs, the model inputs are second-by-second of-fit is assessed based on the slope and R2 of a parity plot
speed profiles, from which acceleration can be inferred. Road of predicted versus actual average mass rates for each
grade was assumed to be zero. Average emissions for the consecutive averaging time period. The slope is an indicator
various cycles were estimated using the EOVMs. Emissions of model accuracy, while the R2 is an indicator of model
from the FTP cycle were selected as the benchmark for cycle precision. Furthermore, the models were applied to the
comparisons. Similar assessments for intercycle comparisons validation data sets to make independent predictions of
were done using MOBILE6. emissions and fuel use. The model predictions for the
For the IOVMs, only real world driving cycles could be validation data were plotted versus the actual values and a
used for assessment because vehicle-specific engine data slope, and R2 for the trend line was assessed. Ideally, the
were available for these but not for the dynamometer cycles. goodness-of-fit for the validation data set should be similar
to that of the modeling data set.
Results and Discussion The results of goodness-of-fit for the IOVMs are given in
Results are presented for (a) choice of averaging time; (b) Table 2. As expected, the slope in the modeling domain is
selection of modeling approach; (c) model evaluation; (d) unity for all vehicles. The slope in the validation domain
sensitivity of models to microscale events; and (e) comparison varies depending on the robustness of the model in explaining
of fleet-average models and vehicle-specific models. variability in the validation data. In most cases, the slope in
Averaging Time. The effect of averaging time on model the validation domain is close to one, with average values
performance and model response to changes in driving among the 10 vehicles for each pollutant ranging from 1.00
conditions was evaluated. An increase in averaging time leads to 1.05. In only four cases does the slope differ from unity
to an increase in R2, reduced influence of autocorrelation, by more than 20%, primarily for HC (in three cases) and CO
and a decrease in residual variance. However, a long averaging (in one case). For each pollutant, the average of the slopes
time will smear out the effect on emissions of microscale among all 10 vehicles does not differ significantly from unity.
events. Three multiples of the gas analyzer response time Thus, the IOVM approach is robust with respect to inter-
was selected for the averaging time. Hence, the averaging vehicle differences.
time for NO is 18 s, and for the other pollutants and fuel use However, the portion of intravehicle variability explained
it is 12 s (26). by the models varies among the vehicles. On average, 40 to
Modeling Approaches. Several modeling schemes were 99% of the variability can be explained for a given pollutant
compared with respect to goodness-of-fit using data for a or for fuel use. However, there are isolated cases in which
2005 Chevrolet Cavalier with 2.2 L engine. As shown in Table the R2 values are very low. For example, R2 values of less than
1, these include a VSP binning approach (model 3), linear 0.3 occur for 6 cases in Table 2, of which 4 are for CO, 1 is
and polynomial regression (models 4-10, 14), time series for HC, and 1 is for NO. The low R2 values for CO are likely
approaches such as moving averages (MA) and time series to be associated with vehicles that have highly pronounced
regression with serial errors (TSRSE) (30) (models 9 and 10, fuel enrichment characteristics, which in turn are associated
respectively), and physically based approaches (models 1 with short episodes of high CO emissions. The data were not
and 2). further stratified to attempt to isolate possible enrichment
In nearly all cases, the best goodness-of-fit for fuel use or episodes because the average duration of fuel enrichment is
a given pollutant emission rate was produced by models 1, less than 2 s, which is much shorter than the averaging time
2, 11, and 13. The lowest R2 among these four models for a used in the models. Furthermore, enrichment events are
given pollutant is typically higher than the R2 values for any sensitive to the response of the catalytic converter, whose
other model (except model 14 applied to fuel use). The highest performance (in terms of second-by-second control efficiency
R2 values for NO, HC, and CO were obtained with the TSRSE by pollutant) is mostly unobservable even when OBD data
model. However, because this model requires knowledge of are available.
prior model residuals, it cannot be used to make predictions. The goodness-of-fit is excellent for fuel use for both
It illustrates, however, that there is autocorrelation in the modeling and validation data sets for all 10 vehicles, with an
residuals of model predictions. This motivates the use of average R2 of 0.99 and an average slope of 1. Furthermore,
methods that either account for or reduce the effect of there is little variability in either of these values among the
autocorrelation. The attempt to account for autocorrelation 10 vehicles.

VOL. 44, NO. 9, 2010 / ENVIRONMENTAL SCIENCE & TECHNOLOGY 9 3597


TABLE 2. Goodness-of-Fit of Vehicle-Specific Internally-Observable Variable Models (IOVM) for NO, HC, CO, and Fuel Use for Ten
Vehicles Using Time-Averaged Engine Speed and Manifold Absolute Pressurea
NO HC CO Fuel
modeling validation modeling validation modeling validation modeling validation
b 2 2 2 2 2 2 2
vehicles Rm Slm Rv Slv Rm Slm Rv Slv Rm Slm Rv Slv Rm Slm R2v Slv
2005 Chevrolet Cavalier 2.2 L 0.43 1 0.43 1.02 0.54 1 0.56 0.99 0.61 1 0.59 1.05 >0.99 1 >0.99 1.00
2005 Dodge Caravan 3.3 L 0.10 1 0.10 1.16e 0.67 1 0.67 1.04 0.03 1 0.03 0.69e >0.99 1 >0.99 1.01
2005 Chevrolet Tahoe 5.3 L 0.50 1 0.50 1.02 0.59 1 0.59 1.20e 0.60 1 0.60 1.04 >0.99 1 >0.99 1.00
1997 Honda Accord 2.2 L 0.75 1 0.75 1.18e 0.70 1 0.69 0.78e 0.62 1 0.61 1.03 >0.99 1 >0.99 1.00
1998 Plymouth Breeze 2.4 L 0.75 1 0.74 0.99 0.44 1 0.43 1.00 0.48 1 0.49 1.02 >0.99 1 >0.99 0.99
2004 Dodge Stratus 2.7 L 0.75 1 0.75 1.19e 0.70 1 0.69 0.78e 0.62 1 0.61 1.03 >0.99 1 >0.99 0.99
1997 Dodge Caravan 3.3 L 0.92 1 0.90 1.01 0.88 1 0.88 1.05 0.67 1 0.67 1.00 >0.99 1 >0.99 1.00
2000 Dodge Caravan 3.3 L 0.84 1 0.84 1.00 0.06 1 0.10 1.07 0.05 1 0.05 1.09e >0.99 1 >0.99 1.00
2002 Dodge Caravan 3.3 L 0.72 1 0.68 1.00 0.74 1 0.81 1.49e 0.26 1 0.29 1.04 >0.99 1 >0.99 1.00
2000 Ford Crown Victoria 4.6 Lc 0.86 1 0.84 0.96e 0.82 1 0.83 1.01 0.12 1 0.12 1.08 >0.99 1 >0.99 1.00
average 0.66 1 0.65 1.05 0.60 1 0.62 1.04 0.41 1 0.41 1.01 >0.99 1 >0.99 1
0.10 0.10 0.99 0.060 0.10 0.78 0.03 0.03 0.69
range - 1 - - - 1 - - - 1 - - >0.99 1 >0.99 1
0.92 0.90 1.19 0.88 0.88 1.49 0.67 0.77 1.09
a
Unless otherwise noted, the functional forms of the models (model 2 for NO, and model 1 for HC, CO, and Fuel) are

NO:mNO )
_∆t aNO
_∆t
(PmapSengine) 0.5
exp
( -bNO
_∆t
PmapSengine )
¯∆t
HC, CO, Fuel:m i ) aiP̄mapS̄engine∆t

where aNO, bNO, and ai are fitted constants; m̄i and m̄NO are average mass rate, g/s Pmap is manifold absolute pressure, kPa; Sengine
is engine speed, rpm; P̄mapSengine is average of the product of Pmap and Sengine for a consecutive averaging period; and ∆t is the
duration for the averaging period (seconds). b Year, manufacturer, model, and engine size are indicated. c For this vehicle,
Pmap was not available from the on-board diagnostics (OBD) link. However, intake mass air flow, mIAF (g/s) is reported by
the OBD link. Hence, the model used here is

NO:mNO )
_∆t aNO
_∆t
(mIAF)0.5
exp
( ) -bNO
_∆t
mIAF
¯∆t
HC, CO, Fuel:mi ) aim̄IAF∆t

The averaging time for NO is 18 s and 12 s for other pollutants and fuel use. R2 and slope are based upon the model of actual
measurements vs model predicting values. R2m and Slm -R2 and slope using data from modeling databases; R2v and Slv - R2 and
slope using data from validation databases. e The numbers in italics indicate that the slope is significantly different from
unity.

The goodness-of-fit for the validation data is very similar Sensitivity of Models to Microscale Events. The sensitivity
to that for the modeling data, indicating consistent perfor- of cycle average gram per mile emission rates to eight selected
mance of the models. driving cycles, including five standardized and four empirical
MAP for the Ford vehicle was not reported by the OBD cycles, was evaluated for both the EOVMs and MOBILE6.
system; however, air intake flow rate was reported. Thus, for Relative trends in these normalized emission rates versus
this vehicle only, intake flow rate was used as an IOV. The average cycle speeds were compared between the EOVMs
R2 for fuel use and CO2 emissions is greater than 0.99 and and MOBILE6. In the EOVMs, second-by-second speed
ranges from 0.10 to 0.86 for other pollutants. profiles with corresponding average speeds are used as input.
The goodness-of-fit and model evaluation results for For MOBILE6, average speed is specified as an input. For
EOVMs are given in Table 3. For a given vehicle and pollutant CO, HC, and NO, the qualitative trends for the EOVM and
the R2 values are typically lower for the EOVM than for the MOBILE6 were similar for cycles with average speeds lower
IOVM. The average EOVM R2 values for a given pollutant are than that of the FTP: typically, the emission rate increases
0.17 to 0.59. Compared to IOVMs, the average decrease in as the average cycle speed decreases. The EOVMs predicted
the R2 values for EOVMs is -0.36, -0.37, -0.24, and -0.40 larger increases in emissions rates as cycle average speed
for NO, HC, and CO emissions and fuel use, respectively. was reduced. For example, for CO, the emission rate at an
The R2 values for EOVMs are similar for the validation and average speed of 7 mph is a factor of 1.4 greater than for the
calibration data. The slope of the validation cases is between FTP based on MOBILE6 and a factor of 2.1 greater based on
0.9 to 1.1 for 30 of the 40 pollutant/fuel and vehicle the average of the EOVMs for the 10 vehicles. At average
combinations. The average of the slopes does not significantly speeds higher than that of the FTP up to 65 mph, MOBILE6
differ from unity. predicts an increase in average emission rate for NOx and
Both IOVMs and EOVMs were developed based on tailpipe CO and a decrease for HC. On average over the 10 vehicles,
emissions. The tailpipe emissions are also influenced by the the EOVMs predict a decrease in average emission rate for
catalytic conversion efficiency (CCE). However, this influence all three of these pollutants, but there is substantial inter-
cannot be quantified because needed data such as engine out vehicle variability in which some vehicles have higher average
emissions and exhaust temperature are not observable with emission rates at the higher speeds. For CO2, MOBILE6 uses
OBD and PEMS data. One reason that R2 for fuel use is better a constant emission rate and therefore a comparison was
than for NO, CO, and HC is that the latter three are significantly not made. The EOVMs predict that the CO2 emission rate
influenced by CCE, whereas the former is not. decreases significantly as average cycle speed increases up

3598 9 ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 44, NO. 9, 2010


TABLE 3. Goodness-of-Fit of Vehicle-Specific Externally-Observable Variable Models (EOVM) for NO, HC, CO, and Fuel Use for Ten
Vehicles Using Time-Averaged Speed, Acceleration, and Road Gradea
NO HC CO Fuel
modeling validation modeling validation modeling validation modeling validation
2 2 2 2 2 2 2
vehicle Rm Slm Rv Slv Rm Slm Rv Slv Rm Slm Rv Slv Rm Slm R2v Slv
2005 Chevrolet Cavalier2.2 L 0.16 1.00 0.16 0.99 0.22 1.00 0.20 0.94b 0.32 1.00 0.31 1.00 0.71 1.00 0.70 0.99
2005 Dodge Caravan 3.3 L 0.05 1.00 0.05 1.04 0.32 1.00 0.33 1.01 0.02 1.00 0.02 0.94b 0.77 1.00 0.77 1.00
2005 Chevrolet Tahoe 5.3 L 0.24 1.00 0.23 0.98 0.20 1.00 0.19 0.98 0.24 1.00 0.24 1.00 0.62 1.00 0.61 1.00
1997 Honda Accord 2.2 L 0.11 1.00 0.15 1.26b 0.03 1.00 0.08 1.09b 0.06 1.00 0.07 1.56b 0.17 1.00 0.26 1.45b
1998 Plymouth Breeze 2.4 L 0.56 1.00 0.55 0.98 0.29 1.00 0.24 0.97b 0.33 1.00 0.32 0.98 0.79 1.00 0.80 1.00
2004 Dodge Stratus 2.7 L 0.44 1.00 0.41 1.00 0.46 1.00 0.47 1.02 0.41 1.00 0.42 1.15b 0.46 1.00 0.42 0.97b
1997 Dodge Caravan 3.3 L 0.31 1.00 0.35 0.88b 0.21 1.00 0.25 0.91b 0.14 1.00 0.12 0.90b 0.43 1.00 0.47 0.99
2000 Dodge Caravan 3.3 L 0.52 1.00 0.53 1.01 0.02 1.00 0.03 1.09b 0.04 1.00 0.04 1.19b 0.79 1.00 0.80 1.01
2002 Dodge Caravan 3.3 L 0.36 1.00 0.36 0.93b 0.28 1.00 0.29 0.85b 0.14 1.00 0.13 0.97b 0.71 1.00 0.74 1.00
2000 Ford Crown Victoria 4.6 L 0.27 1.00 0.20 1.16b 0.24 1.00 0.16 0.87b 0.02 1.00 0.03 1.25b 0.46 1.00 0.42 0.97b
average 0.30 1.00 0.30 1.02 0.23 1.00 0.22 0.97 0.17 1 0.17 1.09 0.59 1 0.60 1.04
0.05 0.05 0.36 0.02 0.03 0.85 0.02 0.02 0.94 0.43 0.35 0.97
range - 1.00 - - - 1.00 - - - 1 - - - 1 - -
0.56 0.55 1.16 0.46 0.47 1.09 0.41 0.42 1.25 0.79 0.80 1.45
a
The model form (model 12) is
_ _ _ _ _
mj∆t ) A0,j + Ajav∆t + Bjv∆t + Cjvr∆t + Djv3∆t + εj

The averaging time period for NO is 18 s and 12 s for the others with nonzero corresponding average speed. R2 and slope are
based upon the model of actual measurements vs model predicting values. R2m and Slm -R2 and slope using data from modeling
databases; R2v and Slv - R2 and slope using data from validation databases. b The numbers in italics indicate that the slope
is significantly different from unity.

to approximately 50 mph. At speeds higher than ap-


proximately 60 to 65 mph, the CO2 emission rate increases
with average speed.
MOBILE6 does not explicitly account for microscale
variation in acceleration. However, since both microscale
acceleration and speed are explicit inputs for EOVMs, EOVMs
are sensitive to microscale events in vehicle dynamics. For
example, as shown in Figure 1, a peak in the speed profile
corresponds to a peak in emissions. Although a longer
averaging time leads to lower spatial resolution, the NOx
EOVM with 18 s averaging time is still able to capture episodic
emissions events. A shorter averaging time produces a larger
shorter duration peak in emissions.
Fleet-Average versus Vehicle-Specific Models. Emissions
hotspots at specific locations identified for one vehicle were
not hotspots for another, and the trend in variation of
emissions over distance is not the same for each vehicle (31).
Thus, it is worthwhile to develop vehicle-specific models for
a variety of applications, such as near roadside human
exposure assessment.
Vehicle-specific models were compared to a “fleet” model
to assess the benefits of the former. Data collected from all
vehicles for this study were combined to form a “fleet” data
set from which fleet-based models similar to model 12 of
Table 1 were fit. Emission rates estimated by the fleet models
were compared to the weighted averages of emissions
predicted by vehicle-specific models for selected driving
cycles and for the same set of vehicles. The weighting was
based upon the amount of data for individual vehicles used
to develop the fleet model. Comparing the EOV fleet model
with the weighted average of vehicle-specific models, the
differences in average NO, HC, and CO emission rates ranged
from -37 to 83% depending on the pollutant and driving FIGURE 1. Microscale prediction of NO and HC emissions using
cycle. For fuel use, the differences were relatively small, externally observable variables models (EOVMs) (model 12) for
ranging from -5 to 7%. These differences indicate that there a 2005 Chevrolet Cavalier with 2.2 L engine.
is substantial intervehicle variability in emissions that is In general, IOVMs have higher explanatory power than
averaged out in the fleet model. For a large flow of vehicles, EOVMs. When IOVs are available, IOVMs are preferred.
this is appropriate. However, in order to capture episodic However, EOVMs may have more practical value for traffic
emissions at high spatial and temporal resolution, vehicle- management such as arterial signalization and level of service
specific models may be useful. The details are given in the determination (32) and simulation since IOVs usually are
SI. not available or not used for emission estimation. Over the

VOL. 44, NO. 9, 2010 / ENVIRONMENTAL SCIENCE & TECHNOLOGY 9 3599


entire trip, total fuel use and emissions estimated by IOVMs duty vehicles. Trans. Res. Part D: Transp. Environ. 2006, 11
and EOVMs are comparable. For example, for HC and fuel (5), 377–385.
use, the relative difference between IOVMs and EOVMs is (9) Halati, A.; Lieu, H.; Walker, H. CORSIM-corridor traffic simula-
tion model, Proceedings of the Traffic Congestion and Traffic
approximately 10% for all three primary vehicles over all Safety in the 21st century: Challenges, Innovations, and Op-
trips. For NO and CO, depending on averaging time and portunities, Chicago, IL,June 8-11, 1997.
vehicle, the difference varied from 9 to 65%. IOVMs predict (10) U.S. Environmental Protection Agency. A Roadmap to MOVES
more fuel use and emissions than EOVMs, except for NO. 2004; EPA420-S-05-002; Office of Transportation and Air Quality,
U.S. Environmental Protection Agency: Ann Arbor, MI, 2005.
(11) Sturm, P. J.; Almbauer, R. A.; Sudy, C.; Pucher, K. Application
Supporting Information Available of computational methods for the determination of traffic
Texts, figures, and tables pertinent to this paper. These emissions. J. Air Waste Manage. Assoc. 1997, 47 (11), 1204–1210.
materials include (1) summary of data collection activities; (12) West, B.; McGill, R.; Hodgson, J. W.; Sluder, S. Development of
(2) exploratory analysis of PEMS data; (3) documentation of data-based, light-duty modal emissions and fuel consumption
model parameters; (4) comparison of IOVMs and EOVMs; models; Paper 972910; Society of Automotive Engineers: War-
rendale, PA, 1997.
(5) evaluation of application of an EOVM model for microscale (13) Joumard, R.; Jost, P.; Hickman, J.; Hassel, D. Hot passenger car
prediction of emissions; (6) evaluation of Box-Cox trans- emissions modeling as a function of instantaneous speed and
formation; (7) identification and comparison of statistical acceleration. Sci. Total Environ. 1995, 169 (1-3), 167–174.
summaries for alternative average time intervals of activities (14) Ahn, K.; Hesham, R.; Trani, A.; Aerde, M. V. Estimating vehicle
and emissions; (8) time series regression with serial errors; fuel consumption and emissions based on instantaneous speed
(9) evaluation of vehicle-specific models for multiple averag- and acceleration level. Transp. Eng. 2002, 128 (2), 182–190.
(15) Thompson, G. J.; Atkinson, C. M.; Clark, N. N.; Long, T. W.;
ing times, vehicles, driving cycles, and comparisons to Hansevack, E. Neural network modeling of the emissions and
MOBILE6; (10) assessment of intervehicle variability in model performance of a heavy-duty diesel engine. J. Aut. 2000, 214 (2),
parameters and predictions; (11) assessment of “Open loop” 111–126.
fuel enrichment; (12) relationship between fuel use and CO2 (16) Jiménez-Palacios, J. L. Ph.D. Dissertation, Massachusetts In-
emissions; (13) VSP-based modal approach for emissions stitute of Technology: Cambridge, MA, 1999.
(17) Barth, M.; An, F.; Norbeck, J.; Ross, M. Modal emissions
modeling, and (14) minimum data requirement. This material
modeling: a physical approach. Transp. Res. Rec. 1996, 1520,
is available free of charge via the Internet at http:// 81–88.
pubs.acs.org. (18) Nam, E. K. Proof of Concept Investigation for the Physical
Estimation Rate Estimator (PERE) for MOVES; EPA 420-R-03-
Acknowledgments 005; Prepared by Ford Research and Advanced Engineering for
Office of Transportation and Air Quality, U.S. Environmental
This material is based upon research supported by the Protection Agency: Ann Arbor, MI, 2003.
National Science Foundation under Grant No. CMS-0230506. (19) Frey, H. C.; Zhang, K.; Rouphail, N. M. Fuel use and emissions
Any opinions, findings, conclusions, or recommendations comparisons for alternative routes, vehicles, road grade and
expressed in this document are those of the author(s) and time of day using in-use measurements. Environ. Sci. Technol.
do not necessarily reflect the views of the National Science 2008, 42 (7), 2483–2489.
Foundation. Drs. Bibhuti Bhattacharyya, David Dickey, and (20) Durbin, J. Distribution theory for tests based on the sample
distribution function; SIAM: 1973.
Jerome Sacks provided advice on statistical procedures. (21) Hastie, T. J.; Tibshirani, R. J. Generalized Additive Models;
Chapman and Hall: London, 1990.
Literature Cited (22) Simonoff, J. S. Smoothing Methods in Statistics; Springer-Verlag:
New York, 1996.
(1) National Research Council. Modeling Mobile Source Emissions; (23) Bowman, A. W.; Azzalini, A. Applied Smoothing Techniques for
National Academy Press: Washington, DC, 2000. Data Analysis: The Kernel Approach with S-plus Illustrations;
(2) Federal Highway Administration (FHWA). CORSIM User Guide; Oxford: New York, 1997.
U.S. Department of Transportation: Washington, DC, 1996. (24) Schimek, M. G. Smoothing and Regression; John Wiley and Sons:
(3) Transport Simulation Systems. AIMSUN Version 4.1 User New York, 2000.
Manual; 2002. (25) Bates, D. M.; Watts, D. G. Nonlinear Regression Analysis and Its
(4) Cloke, J.; Boulter, P.; Davies, G. P.; Hickman, A. J.; Layeld, R. E.; Applications; John Wiley & Sons: New York, 1988.
McCrae, I. S.; Nelson, P. M. Traffic management and air quality (26) Zhang, K.; Frey, H. C. Evaluation of response time of a portable
research programme. TRL Report 327; 1998; p 70. system for in-use vehicle tailpipe emissions measurement.
(5) Frey, H. C.; Unal, A.; Chen, J.; Li, S.; Xuan, C. Methodology for Environ. Sci. Technol. 2008, 42 (1), 221–227.
Developing Modal Emission Rates for EPA’s Multi-scale Motor (27) Arieli, R.; Van Liew, H. D. Corrections for the response time and
Vehicle & Equipment Emission System; EPA420-R-02-027; Pre- delay of mass spectrometers. J. Appl. Physiol. 1981, 51 (6), 1417–
pared by North Carolina State University for Office of Trans- 1422.
portation and Air Quality, U.S. Environmental Protection (28) Heywood, J. B. Internal Combustion Engine Fundamentals;
Agency: Ann Arbor, MI, 2002. McGraw-Hill: New York, 1988.
(6) Colberga, C. A.; Tonab, B.; Stahelb, W. A.; Meierc, M.; Staehelina, (29) The R Foundation for Statistical Computing. R; 2005. http://
J. Comparison of a road traffic emission model (HBEFA) with www.r-project.org (accessed December 16, 2005).
emissions derived from measurements in the Gubrist road (30) Shumway, R. H.; Stoffer, D. S. Time Series Analysis and Its
tunnel, Switzerland. Atmos. Environ. 2005, 39, 4703–4714. Applications; Springer-Verlag: New York, 2000.
(7) André, M.; Keller, M.; Sjödin, A.; Gadrat, M.; Crae, I. C.; Dilara, (31) Frey, H. C.; Zhang, K. Spatial and Temporal Analysis of Real-
P. the ARTEMIS European Tools for Estimating the Transport World Empirical Fuel Use and Emissions, Proceedings of the
Pollutant Emissions, the 18th Annual International Emission 100th Annual Exhibition & Conference of the Air & Waste
Inventory Conference, “Comprehensive Inventories -Leveraging Management Association, Pittsburgh, PA, June 2007.
Technology and Resources”, Baltimore, Maryland,April 14-17, (32) Unal, A.; Rouphail, N.; Frey, H. C. Effect of Arterial Signalization
2009. and Level of Service on Measured Vehicle Emissions. Transp.
(8) Silva, C. M.; Farias, T. L.; Frey, H. C.; Rouphail, N. M. Evaluation Res. Rec. 2003, 1842, 47–56.
of numerical models for simulation of real-world hot-
stabilized fuel consumption and emissions of gasoline light- ES902835H

3600 9 ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 44, NO. 9, 2010

You might also like