United States Air and Radiation EPA420-R-03-005
Environmental Protection February 2003
Agency
oEPA Proof of Concept
Investigation for the Physical
Emission Rate Estimator
(PERE) to be Used in MOVES
Printed on Recycled Paper
-------
EPA420-R-03-005
February 2003
Proof of Concept Investigation for the Physical
Emission Rate Estimator (PERE) to be Used in MOVES
Assessment and Standards Division
Office of Transportation and Air Quality
U.S. Environmental Protection Agency
NOTICE
This technical report does not necessarily represent final EPA decisions or positions.
It is intended to present technical analysis of issues using data that are currently available.
The purpose in the release of such reports is to facilitate the exchange of
technical information and to inform the public of technical developments which
may form the basis for a final EPA decision, position, or regulatory action.
-------
Proof of Concept Investigation for the Physical Emission Rate Estimator
(PERE) for MOVES
By Edward k. Nam
Ford Research and Advanced Engineering
ABSTRACT
The new EPA mobile source inventory model, Multi-scale mOtor Vehicle & equipment
Emission System or MOVES, is an ambitious effort to model emissions from the micro to the
macro level. To maintain consistency across the scales, a road load based quantity known as
Vehicle Specific Power, or VSP, has been chosen as the primary causal variable in emissions
formation for modeling purposes. This work attempts to take the physical link between driving
activity and emissions one step further by introducing the Physical Emission Rate Estimator
(PERE). PERE is meant to supplement the data driven portion of MOVES and fill in gaps where
necessary. The model is essentially an effort to simplify, improve and implement the
Comprehensive Modal Emissions Model (CMEM) developed at University of California,
Riverside. PERE is based on the premise that for a given vehicle, (engine out) running emissions
formation is dependent on the amount of fuel consumed. As such, it models fuel rate as well as
CO2 generation with some degree of accuracy, which is the first step in the MOVES
development. Being a physically based model, it has the potential (with some modification) to
model new technologies (vehicles meeting new emissions standards), deterioration, off-road
sources, I/M programs, as well as being able to easily extrapolate to areas where data are sparse.
It may even save in data taking costs in the long run. Before the concepts in PERE can be
implemented however, it must be demonstrated to work in a limited case scenario. This paper is
an attempt to demonstrate the feasibility of PERE on a sample of (hot running and non hi-
emitting) Tier 1 vehicles. It is shown that the fuel consumption and CO2 predictions were
reasonably accurate for an independent driving sample, though uncertainty remains about a
possible 'speed' effect. Fuel and CO2 is the first installment of MOVES and is discussed in some
detail in this paper.
The second part of this paper deals mainly with criteria pollutants. The engine out behavior is
demonstrated to be relatively steady across engine families and for Tier 1 vehicles. This allows
the model, which is calibrated to only a few samples of engine data, to be generalized to other
data sets, where only tailpipe measurements are available. The engine out model is demonstrated
to be reasonably accurate; both in estimating VSP based as well as total cumulative emissions.
The catalyst (or tailpipe) portion of the model introduces some additional complexities, where it
is demonstrated that VSP or fuel rate is not the only explanatory variable in emissions formation.
A speed (or aggressivity) factor is encountered in the analysis, which needs to be investigated
further. It is possible that the emissions will need to be further subdivided into modes of driving.
This model is preliminary and will require more research, however, the initial validation looks
promising.
-------
TABLE OF CONTENTS
INTRODUCTION. 2
I - LOAD BASED APPROACH 4
Vehicle Specific Power 4
Physical Parameterized (Fuel Rate) Approach 6
A Comparison ofVSP and Fuel Rate 8
Binning vs Parameterized Approach 10
Comprehensive Modal Emissions Model (CMEM) 11
Aggregate Engine Out Trends 12
II- THE PHYSICAL EMISSION RATE ESTIMATOR (PERE) 19
Fuel Rate and CO2 19
Second by Second Engine-Out Emission 20
Engine Out CO 21
Engine Out HC 24
Engine Out NOx 26
Air Conditioner 29
Engine Out Calibration 29
Second by Second Catalyst Pass Fractions 33
CPF CO and HC 33
CPFNOx 44
Tailpipe Calibration 46
III - VALIDATION OF PERE 50
Fuel and CO2 50
Propagation of Uncertainty 52
Engine Out Validation 55
Tailpipe Validation 57
IV-CONCLUSION . 60
V- ACKNOWLEDGMENTS 62
VI - REFERENCES 63
APPENDIX A: Off Road Engines and Future Technologies 65
APPENDIXB: Fuel Types 67
APPENDIX C: CMEM Parameters: Simplification and Sensitivity 68
APPENDIXD: Second by Second Tailpipe Emissions 70
APPENDIX E: Sample Error Propagation Calculations 78
1
-------
INTRODUCTION
The next generation emissions EPA inventory model: Multi-scale mOtor Vehicle & equipment
Emission System or MOVES, is being designed with the recommendations of the National
Research Council report in mind (2000). The report made many suggestions for the improvement
of the MOBILE series of models. Subsequent EPA shootout projects proposed various
alternatives to the MOBILE approach (Hart, et al., 2002). Based on these shootout projects,
MOVES is tentatively designed so that emission rates would be derived from a road load based
criterion known as Vehicle Specific Power (VSP). This is an effort to include more of a physical
basis to the model, rather than relying on average driving cycle speeds, which may be too
aggregate to correlate to finer scale emissions. The emission rates are determined from a look-up
table of VSP bins, whose cells are occupied by various data sources (On-board emissions,
dynamometer, I/M, Remote Sensing, etc) statistically combined. The model therefore is
physically based but primarily data driven.
There are several approaches, which may be considered to supplement such a model. An
alternative to a data driven statistical model is one that is guided more on physical principles. In
contrast to binning data, a parameterized analytical function can be fitted to the data.
Additionally, in lieu of aggregating the emissions from normal and hi-emitters from the real-
world fleet into means or bins, the vehicle mix can be disaggregated and the emissions factors
can be adjusted by this mix (this is what MOBILE does presently). The approaches are
summarized in the table below: It is important to note that the approaches are not necessarily
mutually independent, and that a spectrum of methodologies may exist between the extremes. By
this nomenclature, the VSP approach is physical, binned, and fleet averaged. MOBILE, which is
a macroscopic model is more statistical, parameterized, with disaggregated normal/hi emitters.
Table 1: The contrasting modeling approaches to calculating microscopic emissions factors for a
fleet. The left side is generally more disaggregate or microscopic. The VSP approach is marked
with a 'V'.
Physical V
Statistical
Parameterized
Binned V
Normal/Hi Emitter Split
Normal/Hi Emitter aggregate V
There are two "modal" model systems, which enjoy some popular use. The first was developed
mostly at Georgia Institute of Technology in a series of papers by Fomunung, Washington,
Guensler, Bachman, and others. MEASURE is a model that predicts light duty vehicle emissions
based on their operating conditions all within a GIS framework. The operating conditions are
defined by their modes of operation, such as acceleration, cruise, deceleration and idle. The trip-
2
-------
based emission rates are determined using a statistical approach of weighting the causal factors
in a single equation for each criteria pollutant. These factors include average speed, proportion of
driving within the modes of driving, transmission type, odometer, injector type, catalyst type,
etc. The model is advantageous in many ways: it is capable of being calibrated with bag data, it
is statistically robust, validation has been performed, and it is incorporated into a GIS
framework. However, there are some disadvantages to this approach: it is limited in its ability to
model control strategies as well as future fleet emissions; being a statistical model, there are no
set physical limitations to the calibrated parameters. Physically, there are correlations between
some of the variables, which may be more easily 'teased out' using a more physical model.
However, there are many aspects of MEASURE, which influence the development of MOVES.
Another popular approach was developed mostly at the University of California, Riverside (CE-
CERT) in a series of papers by Barth, An, Ross, Scora, Younglove, Levine (Malcolm), Wenzel,
et al. The Comprehensive Modal Emissions Model, or CMEM, is based on physical principles,
which derived a fuel rate from road-load and simple powertrain model. Emissions rates are then
derived empirically from the fuel rate. The model has been validated in a number of studies for
typical (or composite) vehicles and has the potential to model a wide variety of light as well as
heavy-duty vehicles. CMEM will be described in greater detail in a later section.
This work is a proof of concept developing a physical parameterized model methodology from a
subset of the data, namely "normally emitting" Tier 1 light duty vehicles. The first section
describes the load-based emissions methodology, be it vehicle specific power, or fuel rate (both
very similar). It also discusses and develops the fuel rate based methodology based on CMEM.
This is the portion of the model that will output the fuel consumption as well as CO2 emissions.
Part II describes the model development process for the Physical Emission Rate Estimator
(PERE), based on CMEM principles, potentially to be used in some similar form in MOVES.
The section also describes the calibration of engine out (pre catalyst) and tailpipe (post catalyst)
emissions to parametric functions. All of the analysis is for hot running operation. Cold start may
have a significant impact on some of the trends and generalizations made in this paper, but will
be a subject of a future study. Part III describes the validation that has been accomplished
thusfar. A number of appendices supplement this paper, which explore tangential issues
important to the development of the modeling methodology.
3
-------
I - LOAD BASED APPROACH
Vehicle Specific Power
The current plan for MOVES is to use Vehicle Specific Power (VSP) as a variable on which
their emission rates can be based (Koupal et al, 2002). The VSP approach to emissions
characterization was developed by Jimenez-Palacios (1999). VSP is a measure of the road load
on a vehicle; it is defined as the power per unit mass to overcome road grade, rolling &
aerodynamic resistance, and inertial acceleration:
VSP = v * (a*(l+e) + g*grade + g*CR) + 0.5p*Co*A*v3/m (1)
where:
v: is vehicle speed (assuming no headwind) in m/s
a: is vehicle acceleration in m/s
e: is mass factor accounting for the rotational masses (-0.1)
g: is acceleration due to gravity
grade: is road grade
Cr: is rolling resistance (-0.0135)
p: is air density (1.2)
Cd: is aerodynamic drag coefficient
A: is the frontal area
m: is vehicle mass in metric tonnes.
The equation can also have an added vehicle accessory loading term (air conditioner being the
most significant) added to it. Moreover, higher order terms in rolling resistance can be added to
increase accuracy of the model (Gillespie, 1992). Using typical value of coefficients, in SI units
the equation becomes (CoA/m - 0.0005):
VSP (kW/metric Ton) = v * (1.04*a + 9.81*grade(%) + 0.132) + 0.00121*v3 (2)
If future technologies such as low rolling resistance tires, and aerodynamic designs, vary these
values significantly, that can be reflected with a change in the equation. It should be noted that
the while it may be reasonable to assume typical values for rolling and aerodynamic resistance
constants, it may pose a problem to assume a single mass for all cars (or vehicle types). There is
approximately a factor of 2 difference in CoA/m between an empty compact car and a full large
passenger car (Jimenez, 1999). Using a single value for all LDVs (for example) can result in a
significant error (in VSP) at high speeds when the aerodynamic resistance term dominates and
when feedgas emissions are relatively high.
In this form, the VSP equation is physically based, but the emissions generated from it would be
empirical. Looking at vehicle engine out emission trends as a function of this road load variable
seems to give reasonably good correlation, even among different vehicles, as we shall
demonstrate later.
This binning method easily lends itself to emissions factor generation. Data from Portable
Emissions Measurement Systems (PEMS) and other sources can be used to populate each bin.
4
-------
The main vehicle parameters required to obtain the emissions factor are (instantaneous) speed,
and acceleration. If the more detailed equation 1 is used, then the vehicle mass and road grade
are also required. As such, the model could also be used to quantify second-by-second emissions
(g/sec). It could also be integrated with a microscopic traffic model to quantify emissions from a
set of vehicles.
While it would be more accurate to fill these bins with engine out data, this is not often possible
due to the scarcity of such data. The original approach, as described in Hart et al (2002) and Frey
et al. (2002), fills these bins with tailpipe data. Throughout this paper, we will discuss both
means of modeling, and will specify whether engine out or tailpipe data are being used.
This approach to emissions characterization has been widely accepted in the Remote Sensing
Detection (RSD) community. This due to the fact that RSD measures road-side tailpipe
emissions at a particular moment in time, and since the VSP method can be independent of
vehicle mass, the calculations (and data taking) are simplified. Because of the static nature of
RSD, the emissions values tend to be correlated to a particular road-load or VSP point, which
can be directly measured. Quantifying the VSP allows for comparison between different
measurement sites as well as with other types of modal data, such as those obtained from some
dynamometer tests.
Figure 1 shows a sample NOx tailpipe emissions profile taken from a PEMS source. The curve is
familiar to emissions modelers, where low load driving produces less emissions than higher load
driving. One logical approach to modeling this effect would be to fit the emission curve to a
smooth analytical function such as the following:
NOx = a + 6* VSP + c*VSP2 + d*VSP3 + ... (3)
where a, b, c, d... are fitted parameters using perhaps least squares regression. With this
parameterized function, data can be more easily interpolated and extrapolated to regions where
data "holes" exist. This semi-empirical method of modeling emissions is approaching the
physically based method used by UC Riverside.
0JES
0JJ25
0J02
~~
0 JO 15
-------
Physical Parameterized (Fuel Rate) Approach
If the vehicle emissions system were modeled as a complete physical system, with as much
physical detail as existing data permits, this would presumably result in the most "robust"
prediction. Since tailpipe emissions formation is a physical process, one should be able to model
this system in this fashion. Therefore, starting with a microscopic basis is likely to lead to a more
robust model (though the results are not guaranteed to be more accurate in every case). It could
be argued that a detailed microscopic model could be used as the building blocks for more
aggregate macroscopic models, whereas one cannot easily derive a microscopic model based on
a macroscopic one. Of course, the limiting factor in each of these models is the uncertainty in the
activity input. The most microscopic modeling that occurs for vehicle emissions involves the
characterization of the chemistry and physics of in-cylinder combustion combined with thermal,
transport and catalytic chemistry aftertreatment processes. This level of modeling obviously
requires knowledge of engine, catalyst, fuel parameters, etc, which are unknowable for the
typical car in the fleet. It also assumes a level of knowledge of the manufacturer's engine
calibration strategies, which are highly proprietary.
The most microscopic emissions model, which can be practically applied to a typical car in a
technology class is exemplified by the Comprehensive Modal Emissions Model (CMEM)
developed at the UC Riverside (there are others). This model has been tested and validated under
a variety of conditions (Barth et al., 2000).
Quite simply CMEM takes the ubiquitous tractive road-load relation in equation 1 and combines
it with a simple engine/transmission powertrain model to calculate a fuel rate. Fuel rate, or fuel
consumption per unit time, forms the basis for CMEM, much as VSP forms the basis for the
previous method. The basic independent variables are still speed and acceleration. The tractive
(or brake) power (in kW) from equation 1 can be simplified to:
Pb(m,v,a) = A*v + B*v2 + C*v3 + m*v*(a* + gsinO) (4)
where
A, B, C, are rolling and aero resistive coefficients determined from dynamometer coast-
down tests. The A and B coefficients are equivalent to the rolling resistance terms of equation 1
and the C coefficient represents the air drag coefficients.
If we assume a mass, and typical vehicle coefficients (as was done in eq 2), we would get nearly
the identical results as equation 2. Hence these approaches are fundamentally similar. In fact a
very simple conversion equation can be derived:
Pb(m,v,a) ^ VSP * m (5)
The mass independence of VSP is one of the strengths of the variable, as we mentioned earlier.
However, vehicle mass (even among LDVs) may be a significant variable (Frey et al, 2002).
CMEM goes on to further calculate fuel consumption rate, FR (g/s):
FR =
-------
cp: is the fuel air equivalence ratio (mostly =1)
K(N): is the power independent portion of engine friction, dependent on engine speed.
N(v): is the engine speed
Vd: is the engine displacement volume
r|: is a measure of the engine indicated efficiency (~0.4).
Pacc(T,N): is the power draw of accessories such as air conditioning. This is a function of
ambient temperature, T, and humidity (and engine speed for AC). Without AC, it
is some nominal value ~ lkW.
LHV: is the factor lower heating value of the fuel (~44kJ/g)
Fuel rate is relatively insensitive to K. The engine displacement, Vd can be classified by vehicle
technology types; r| is relatively constant from car to car and from model year to model year; and
LHV is a known quantity, which can be adjusted (or assumed) based on the fuel type. Since this
is a model for a typical vehicle, transmission effects and gear shift points tend to average out, so
engine speed, N, can be very simply modeled as a function of speed and acceleration (Thomas
and Ross, 1997). For MOVES, the engine speed model should be even further simplified. Future
vehicles, (meeting new emissions standards) are not accounted for explicitly in this model, but
can be developed in a future version if there is enough information in the literature. KNVd is a
friction term, which has less of an effect on emissions than the tractive, or brake power term (the
inertial acceleration term dominates along with aero drag at higher speeds). As a result, it is not
as crucial to model these terms with as much care. During enrichment, the parameter cp, is
modified from Goodwin (1997):
cp = l+ 0.036*(FR-FRth) (7)
The enrichment threshold, FRth will be determined later. This is simply an approximation and the
coefficient (0.036) should be recalibrated to incorporate new data. We will not explicitly model
the enrichment fuel air equivalence ratio but it can be added later if necessary. A possible
advantage of modeling cp explicitly lies in the ability to simulate certain types of failure modes
such as when vehicles run rich, lean, or just have poor fuel control due to malfunctioning oxygen
sensors, etc. The model could thus be linked with Inspection and Maintenance (IM) programs
where the failure modes of On Board Diagnostic (OBD) parameters are tracked. A relatively
simple model of (p for CMEM can also be found in Barth et al. (2000), however, the model relies
on a simple model of the wide open throttle power curve, which may add unnecessary
complication and detail.
We confirm that tractive power (or VSP) remains the primary driver of engine out emissions
formation. This would show up as a first split in a hierarchical tree based regression (Frey et al,
2002). Finally, the fuel air equivalence ratio, cp, is almost always close to 1. Two notable
operating exceptions will change this value enough to significantly affect overall emissions: cold
start, and enrichment - both of which can and should probably be modeled explicitly due to their
significant effects on the inventory.
Henceforth, when we refer to "modeled fuel rate", we are using equation 6. Using this approach,
fuel rate has been modeled explicitly and accurately (including air conditioner usage) in the
literature knowing only v, a and a few vehicle and ambient parameters: m, Vd, T, and humidity,
7
-------
for generic vehicle types. An adaptation of this model for air conditioner was established by Nam
(2000). Estimating parameters, this can be modified to be consistent with RSD VSP reporting
schemes.
CO2 can also be estimated based on the fuel rate. Assuming that the hydrogen to carbon ratio in
fuels will also vary based on the fuel type, blend and cetane number. For typical gasoline, the
ratio is roughly 1.85. This gives the following calculation for CO2 based on a carbon balance:
C02 = 44[(FR-HC)/(12.0+1.85) - CO/28] (8)
Where each of the values are in mass units. This equation implies that a knowledge of CO and
HC emissions as well as fuel rate are necessary to calculate CO2. As we shall see later, HC and
CO are small correction terms. CO2 results obtained using this equation, are 'modeled' or
'calculated' values to be compared with direct measurements.
Another advantage to this approach of modeling engine load is that it allows for the modeling of
off-road engines. This is described in some detail in Appendix A. The proof of concept is
reserved for a future study.
A further advantage of this fuel approach is that it can potentially be used to model fuel effects
explicitly. Some details for this are described in Appendix B, but a proof of concept should be
conducted in the future.
A Comparison of VSP and Fuel Rate
There are several advantages to fuel rate over VSP:
• Fuel rate as a load surrogate is physically more causal to emissions production due to its
powertrain derivation (rather than road-load, which is a step removed)
• Fuel consumption (and effects) can be modeled explicitly (with no fitted parameters) as
opposed to empirically
• CO2 emissions can be approximated from the fuel calculations (carbon balance), so may not
require an empirical fit.
• Some future powertrain improvements can be modeled: improved engine efficiency, engine
downsizing, enrichment strategy, weight increase/reduction, catalyst improvement etc.
• Fuel based models can estimate area-wide fuel consumption, which can be compared with
fuel sales on a macroscopic level. CO2 estimated from this "fuel sales" based methodology
can be compared with that obtained from a more microscopic approach.
• The methodology is more microscopic, thus it can be used to build up more aggregate models
• The methodology is older and more developed than VSP
• Since the load surrogate is related to engine load rather than road load only, non-road modal
emissions may also be modeled using the same approach
• This fuel based methodology also aligns with the specific fuel based reporting of emissions
such as grams emissions per gram of fuel (or CO2)
8
-------
• Different fuel types can be modeled explicitly with physical parameters.
Both the fuel rate and the VSP approaches require a further empirical step of measuring
emissions and either binning or analytical function fitting. Dividing fuel rate by mass (similar to
VSP) has been shown to help some of the correlations (Nam, 1999). Thus we see that the
approaches are quite similar in some ways, yet CMEM is more physically rigorous, whereas VSP
is simpler. In fact, a CMEM-like model could be used to populate VSP bins, though the reverse
may not necessarily be true.
Figure 2 compares Fuel Rate and VSP more directly. The VSP was calculated using equation 2.
Fuel rate was calculated from a carbon balance (equation 8), rather than a direct measurement
from the fuel delivery system. Henceforth, this is what we refer to as "measured fuel rate" since
it is the closest value we have to a direct measurement. Note that weight splits the trends, such
that all other things being equal, heavier cars consume more fuel for a given VSP value. This is
not too surprising, and hints that in order to predict fuel consumption (and CO2) using only VSP
values, the weight distributions must be known. Even if using the general form of the equation
(1), engine displacement is not a term in the equation, as it is for fuel rate. Frey et al., (2002)
showed that engine displacement was one of the factors that affected emissions. Note that VSP
values can be negative during higher decelerations.
Fuel Rate compared to VSP for Tier 1 cars (10% sampling)
¦ >3125
0 <3125
8-
¦
¦ ¦
¦
¦
1
¦
¦ 11,
¦ «
0 ¦
¦
_ *
" ' ¦
0
0
¦
jHy
r $*«
0 oO
1
«J|
»
k c °
* ^
it
:8L
¦ _ ^ 0
0
0
-200 -150 -100 -50 0 50 100 150 200 250
VSP
Figure 2. Fuel Rate (measured) compared to VSP for a variety of normal emitting Tier 1 cars.
Each point is 1 second of data.
In conclusion, it has been demonstrated that basing hot-stabilized emissions (engine out or
tailpipe) on VSP alone can introduce a fair amount of variability due to an assumed vehicle
mass. Any estimate of fuel consumption or CO2 in particular will merely be an estimate based on
the initial values chosen for the VSP equation (2). In fact, there is a range of values one can
choose for this equation, each giving a different result. If some of the parameters in equation 1
were used, this would lead to more accurate value of VSP, and such a variable would be very
9
-------
similar to fuel rate, as we shall see later. Henceforth in this report, any calculation of VSP will be
based on equation 1, using assumed values from Thomas and Ross (1997). CMEM employs
dynamometer A, B, C coefficients to complete equation 1, but those coefficients are not always
easy to obtain. A fuel rate approach decreases some of the variability at the cost of adding
parameters, such as vehicle weight, engine size, powertrain efficiency etc., which are currently
available or estimated from the database.
The major disadvantage that CMEM is cited to have, is that it includes a catalyst model (since
fuel rate mainly gives empirical engine out emission trends), whereas VSP models tailpipe
directly. However, as we shall see later, this may not really be a disadvantage at all, but rather a
strength of the model. Since both are roughly based on road-load, and both ultimately predict
tailpipe emissions, the only difference is in essentially how many parameters, or what level of
complexity is used to obtain the final result. CMEM can be significantly simplified to nearly the
complexity of the VSP approach. However, by its nature, the physical approach is more
complex, and may require additional training for some users. Ultimately though, both are
calibrated to the same data, so the final tailpipe model should be similar. Before this is discussed
in more detail, we will first compare the binning to the analytical fit approach to modeling,
which is a separate (but related) topic.
Binning v.s Parameterized Approach
We have briefly mentioned the relationship between binning data and fitting a parametric
equation. Here, we compare the two methods in more detail, both assume a physical basis for the
model (VSP or fuel rate). Both are useful depending on the circumstance. Throughout this paper,
we will use both, where appropriate.
Binning
Advantages Disadvantages
1.
Simple
1.
A LOT of data must be taken
2.
Data driven: each bin relates real data
2.
Difficult to interpolate gaps accurately
3.
Bins can be sized to optimize fit
3.
Nearly impossible to extrapolate gaps
4.
Can fit analytical form to bin heights
beyond scope of data
5.
Does not require constant recalibration
4.
Difficult to incorporate data from many
6.
Easy to implement
sources (true for all approaches)
7.
Easy to update
5.
Can end up being a "black box" of
8.
Consistent across scales
statistics
9.
Uncertainty can be quantified
6.
Difficult to disaggregate to analytical
approach if needed
10
-------
Advantages
1. Fits interpolate excellently
2. Fits can extrapolate outside the bounds of
data to some extent
3. Data needs are fewer, due to above two
cases (cost is less)
4. Comprehensible/explainable
mathematical trends
5. Can populate data bins or cells- can
aggregate to bin approach if needed
6. Easy to implement
7. Easy to update
8. Consistent across scales
9. Uncertainty can be quantified
Disadvantages
1. Value is fitted thus may not necessarily
reflect measurement at any given value
2. Difficult to incorporate data from many
sources
3. Requires recalibration with each data set
4. More calculation intensive (may reduce
software efficiency)
5. The model may be more complicated,
and require additional training for users.
Comprehensive Modal Emissions Model (CMEM)
CMEM is designed to model emissions on a second by second basis using parameterized
functions. Since MOVES does not require predictions on this scale, or for specific vehicles,
CMEM can be simplified to be applicable to macro-level inventory analysis based on speed and
acceleration profiles or distributions. We have already presented some ways that the complexity
in CMEM can be reduced. There are quite a number of parameters used in the model that can
take on an assumed fixed value. If at some later time, technology forces a change or variation in
these parameters, then that can be reflected in the model (unlike in MOBILE, where new
correction factors must be introduced). This is a great strength of the combined physical-based
model and the parametric approach.
A list of the CMEM parameters, their applicability to MOVES and the model sensitivity to the
parameters is presented in Appendix C.
Cold Start/Soak
Cold start and Soak can be modeled with a correction factor approach. The first 100-200 minutes
of a car operation can be a separate mode of cold start. This can be modeled with an algebraic
HC, CO, NO "correction" factor. It is not necessary to have an integrated modal (or second by
second) cold start model at this time, since it is really its own separate mode of operation. This
improvement can be undertaken in a future study.
A Simpler Catalyst Model
Since MOVES does not need to accurately model second by second emissions, the goal is to
reduce the number of fitted parameters in a catalyst model. Engine (with fuel injection)
technology improvements leading to emission reduction (with the exception of EGR) have had a
relatively minor effect on (engine-out) emissions in the past few decades in comparison to
catalyst (and air fuel control) improvements. So an aftertreatment system modeled separately
would be able to model these technological improvements in isolation. However, as we will see
11
-------
later, a suitable aftertreatment model is difficult to design. The details are described in greater
detail in the next section.
Aggregate Engine Out Trends
In order to prove that a model like CMEM can be used to estimate emissions from a fleet of
vehicles, it may be necessary to split engine out from tailpipe emissions. The difficulty with this
approach is that in most cases, engine out data is not available. We hypothesize that engine out
emissions have not changed significantly since the introduction of fuel injection and oxygen
sensor feedback to maintain stoichiometric combustion conditions. These include vehicles
starting from the early 1990s. This section attempts to prove out this hypothesis.
Figure 3 shows the (modeled) engine-out CO trends as a function of fuel rate for the 11 UC
Riverside normal emitting composite (or average) cars. The slopes are determined from the
calibration parameters used in the Riverside model, which are made for the composite car
(second by second average of several vehicles). The categories are in the title of the figure.
Category 1 (no catalyst) has significantly higher emissions in the engine. It is likely that cars
without catalysts ran richer. The trend appears that the slope is dropping as new technologies and
standards are introduced. The Tier 0 to Tier 1 drop is significant (-50%). Within emissions
standards though, the emissions trends (sloped) are similar. Note that the tier 1 composite
vehicles are close to the average. The Tier 0 emissions are more spread out. This may be due to
deterioration effects or engine variation (carbureted vs fuel injected). Further investigation of this
may be required.
ECO for UCR: (8-11 = Tier 1, 4-7=Tier 0,cat/FI, 3=cat,carb, 2=2waycat, 1=nocat)
X
EC01
X
EC02
-
EC03
o
EC04
o
EC05
~
EC06
A
EC07
~
EC08
•
EC09
EC010
¦
EC011
~
ECOavg
¦Tier 1 average
o
~
..t
Fuel Rate (g/s)
Figure 3. Modeled Engine out CO in CMEM
Figure 4 indicates that Tier 1 trends are also tight with HC. From this, we might conclude that for
CO and HC, we can aggregate Tier 1 vehicles together from an engine out standpoint. We have
12
-------
yet to prove this for tailpipe results. There is somewhat more scatter for the Tier 0 composite
cars. Category 5 composite car (3way cat, fuel inject) emits the most, then the non-cat cars.
EHC for UCR: (8-11 = Tier 1, 4-7=Tier 0,cat/FI, 3=cat,carb, 2=2waycat, 1=nocat)
X
EHC1
X
EHC2
-
EHC3
EHC4
o
EHC5
~
EHC6
A
EHC7
~
EHC8
EHC9
EHC10
¦
EHC11
A
EHCavg
¦Tier 1 average
A
A
O
$
8 ~
e
~
~
A S-
4 5
Fuel Rate (g/s)
Figure 4. Modeled Engine out HC in CMEM
ENOx for UCR: (8-11 = Tier 1, 4-7=Tier 0,cat/FI, 3=cat,carb, 2=2waycat, 1=nocat)
o
m 0.15
X
ENOxl
X
ENOx2
-
ENOxS
ENOx5
ENOx4
~
ENOx6
A
ENOx7
~
ENOx8
ENOx9
ENOxl 0
¦
ENOxl 1
A
ENOxavg
Tier 1 average
I-'
.« B
' ¦ 8
s
-• *
..-I'
f *
,
-------
older cars may have run more rich. Except for the non-catalyst composite car, the Tier 0
composite cars are also very tight (and able to be aggregated).
The results above hint that, engine emissions (for properly functioning cars) are relatively
constant within an emissions standard, even when isolated by power to weight ratio and
odometer. However, these results were obtained by using modeled coefficients from composite
vehicles. To fully justify this, the variation in the individual vehicles (of different weights,
engine sizes, etc), should be examined. This is revisited in part 2.
Vehicle Data
The dataset for this study was taken from the set of normal emitting Tier 1 vehicles from the
NCHRP data set taken at UC Riverside. The data was isolated from the rest of the EPA dataset.
due to the second-by-second engine out emissions measured. Hi-emitters will be analyzed in a
separate document. The FTP bag emissions include bag 2, and the latter 475 seconds of bag 3
(after catalyst light-off). The data includes 41 vehicles, most having FTP as well as US06 tests.
The MEC (modal emissions cycle) tests were left out for validation purposes. The vehicles are
listed on table 2. Due to the limited number of vehicles, the sample is not representative of the
mix of (Tier 1) vehicles in the fleet.
The FTP and US06 bag emissions, used in the correlations in this section, are combined without
any weighting. The bag emissions are calculated by summing the second-by-second values and
then dividing by the distance traveled in the driving trace (g/mile). Due to this summation
approach, even though there are almost three times as many FTP points as US06 points, the
US06 cycle has a stronger influence on the aggregate emissions due to the much higher
emissions values. Due to the difficulty of working with large time trace data, this aggregation is a
first pass attempt at looking for obvious trends. Naturally, a modal or second by second analysis
is required once a hint of correlation is seen in the aggregate data. This is done in later sections.
It is important to note that the emissions figure include a US06 cycle, which can sometimes
exhibit emission spikes, so "high emissions" do not necessarily imply a necessity for a "hi-
emitter" classification. The known parameters for the vehicles include the following:
• Vehicle Make/Model
• Model Year
• Engine Size
• Number of Cylinders
• List Weight
• Odometer
• Rated Power
• YIN
14
-------
Table 2. The table of normal emitting Tier 1 vehicles in the NCHRP study.
Vehicle Make Model Model Yr Eng Sz NCyliListWt Wtton Odomete Rated Pwr Category
133
Honda
CIVIC
1995
1 . 5
4
2375
1.1875
52111
70
8
163
Honda
CIVIC
1994
1 . 5
4
2625
1.3125
78056
102
8
250
Oldsmobi
98
1994
3 . 8
6
3875
1.9375
54825
149 .
94
8
263
Honda
CivicEX
1995
1 . 6
4
2750
1.375
54843
102
8
266
Plymouth
Acclaim
1994
2 . 5
4
3125
1.5625
56936
104 .
99
8
284
Honda
Accord
1993
2 . 2
4
3500
1. 75
97869
140
8
302
Ford
Taurus
1995
3
6
3625
1.8125
63558
140
8
187
Toyota
Paseo
1995
1 . 5
4
2375
1.1875
56213
100
9
191
Saturn
SL2
1993
1 . 9
4
2625
1.3125
63125
124
9
192
Honda
CivicDX
1994
1 . 5
4
2375
1.1875
57742
102
9
199
Dodge
Spirit
1994
2 . 5
6
3000
1. 5
57407
141
9
201
Dodge
Spirit
1994
2 . 5
6
3000
1. 5
56338
141
9
229
Honda
CivicLX
1993
1 . 6
4
2625
1.3125
61032
125
9
242
Saturn
SL2
1994
1 . 9
4
2625
1.3125
64967
124
9
260
Toyota
CamryLE
1995
3
6
4000
2
51286
188
9
281
Honda
AccordEX
1993
2 . 2
4
3250
1.625
72804
140
9
17
Toyota
Tercel
1995
1 . 5
4
2250
1. 125
23249
93
10
75
Ford
Mustange
1995
3 . 8
6
3500
1. 75
28905
145
10
80
Ford
Taurus
1997
3
6
3625
1.8125
3415
140 .
19
10
92
Saturn
Saturn
1996
1 . 9
4
2625
1.3125
18000
100
10
105
Saturn
Saturn
1996
1 . 9
4
2625
1.3125
7107
100
10
129
Honda
Accord
1995
2 . 2
4
3250
1.625
37194
130
10
253
Toyota
Corolla
1996
1 . 8
4
2875
1. 4375
29480
100 .
11
10
282
Geo
Metro
1996
1
3
2000
1
32034
60.58
10
324
Chevy
Malibu
1997
3 . 1
6
3375
1.6875
3015
130 .
44
10
329
Plymouth
Breeze
1997
2
4
3250
1.625
23099
114 .
74
10
32
Cadillac
NS
1996
4 . 6
8
4000
2
13287
275
11
33
Buick
Lesabr
1996
3 . 8
6
3500
1. 75
22607
205
11
37
Honda
Civic
1995
1 . 6
4
2250
1. 125
49814
102
11
48
Infinity
G2 0
1995
2
4
3000
1. 5
21468
140
11
53
Plymth
Breeze
1996
2
4
3000
1. 5
15096
132
11
54
Chevy
Capri
1994
4 . 3
8
4250
2 . 125
43625
200
11
67
Ford
T-Bird
1996
4 . 6
8
4000
2
16390
205
11
74
Dodge
Neon
1996
2
4
2500
1.25
5312
132
11
102
Dodge
Spirit
1994
3
6
3000
1. 5
49492
141
11
104
Nissan
Sentra
1995
1 . 6
4
2750
1.375
35291
115
11
108
Toyota
Camry
1994
3
6
3625
1.8125
22258
188
11
116
Chevy
Cavalr
1996
2 . 2
4
2875
1. 4375
5690
120
11
120
Chevy
Camaro
1996
3 . 8
6
3625
1.8125
25877
200
11
123
Chry
Lebaron
1995
3
6
3375
1.6875
22197
141
11
286
Honda
AccordLX
1995
2 . 2
4
3000
1. 5
20606
130
11
15
-------
Weight, Engine Size and Power
Figure 6 shows the relationship between engine size, weight and rated power. Clearly all of these
quantities are correlated. A relationship based on one should be assumed to apply to the other. It
might be possible to use these relationships to fill in data holes when engine size or rated power
are needed, but are lacking in the data set.
4500
4000
3500
3000
s*
i:
2500
-------
Odometer
Figure 7 shows the relationships between model year and odometer reading for the vehicles. The
(California) vehicles appear to be driven at a rate of approximately 18,000 miles per year.
Odometer and Model Year for Tier 1 cars
110000 T 1 1 1
100000 -- # Odometer
Linear (Odometer)
90000 - J
80000 -: #
70000 -
<5 ; * ... t t
"5 60000
-------
Emission Rate Chooser Flow Chart for MOVES (excludes cold start)
#. Design
decisions
^ Output ^
CTIVITY INPUT
V.A
Fleet info:
M, Vd, Cr, CdA
Heat Index
Road grade
1.Standards
2.Odometer
3. Weight
N.categories
(N-l)2 bins
-~ Preference
VSP or
Power
Off-Road
Activity
Physical
Emission Rate
Estimator
P/T RPM
model
PERE: Physical
Emission Rate
Estimator
y
Enrichment
3. VSP or
FR based?
1. Physical or
Statistical?
Fuel Rate,
bhp, etc
Model TP
direct
Statistical
BER
estimator
Engine Out
(para/bin?)
Hi/Normal Emitter
I/M, RSD
Standard, deterioration
Parameter database
2. Bin or
parameter?
Catalyst Pass
Fraction
(para/bin?)
Parameterized
Fuel &TP model
Coefficients for
M, Odo, etc
VSP Binned
BER
based on data
5. Bin or
parameter?
/ Parameterized TP by
I FR: CO, HC, NOx,
V CH4, N20, etc
Binned TP by VSP
CO,HC,NOx,
CH4, N20, etc
6. "Hybrid"
approach?
Figure 8. Decision trees for the integration of a Physical Emissions Rate Estimator.
18
-------
II- THE PHYSICAL EMISSION RATE ESTIMATOR (PERE)
We have already outlined the advantages and uses for a physical model as a basis to fill
(extrapolate as well as interpolate) bins with few real data. The flow chart on the previous page
(figure 8) demonstrates some means by which this may be accomplished. The following sections
describe some of the decision-making processes. The data used in this section is from the
NCHRP data set (table 2) collected at UC Riverside.
Fuel Rate and CO2
Using equation 1 and 6, fuel rate is modeled with some degree of accuracy. The only addition
necessary was an engine speed model. This was obtained from Thomas and Ross (1997). The
results are not tremendously sensitive to shift points, so this simple model is sufficient.
Stationary off-road vehicles should have a more complex engine speed model. Figure 9
compares the model to measurement. R = 0.89 and the slope is 1.
~
~
~
L
~
~
~
~
^ .Pife
I D
~ ~
d
° Dn
~ ~
]
~
cP B
BT Q c
^ |
ID n °
h
]
~
np|H
SB 5. d
la" Dn
~
~
~ d
Hp" [
p ~
I
~
'Bu
3
2
Q
O 1
FR_MEAS
Figure 9. Fuel Rate measured and modeled (10% sampling of data). Each point is a second of
sampled data.
Using equation 8 and assuming that the fraction of CO and HC are miniscule compared to fuel
and CO2, the latter quantity can be calculated to fair degree of accuracy. Figure 9 shows the
comparison. R2 = 0.90 and the slope is 1. CO2 was calculated from basic physics, rather than
from an empirical relationship.
19
-------
25
20
XL
~
~
~
~
Q
O
CM
O
o
15
10
0
5
XL
~
0
5
10
15
20
25
TC02
Figure 10. C02 modeled and measured (x-axis).
We move on to the analysis of criteria pollutants: CO, HC, and NOx.
Second by Second Engine-Out Emissions
Direct modeling of tailpipe emissions from VSP or fuel rate is possible (decision node 4 in figure
8), but it is not recommended. An example of this type of analysis is provided in Appendix D.
We have already established the engine out emissions can be aggregated for all of the Tier 1
vehicles, we will now determine the significance of the differences between the vehicles from
observing the second by second data. The scatter or variability in this data should be isolated
first. This will help us to determine if different relationships or vehicle bins are required on this
scale. Since we are trying to minimize error, we will study emissions trends with VSP and with
measured FR, rather than with modeled FR. As we mentioned earlier, measured FR is calculated
from a carbon balance of the mass emissions (based on equation 8):
assuming a molar hydrogen to carbon ratio in the fuel of 1.85 for gasoline. Using this value of
FR rather than the one modeled from speed, accel, etc. leads to a better understanding of trends.
Later, we will use the modeled fuel rate to do the calibration and validation.
In the following study, we compare trends with both FR and with VSP to determine the better
independent (activity) variable. This is decision node 3 on the flow chart.
FR = (C02/44 + CO/28)*(13.85) + HC
(9)
20
-------
Engine Out CO
Figures 11 and 12 show 10% of the second by second points of the Tier 1 cars on the US06 cycle
(-7000 points). The FTP points were omitted since they would likely cluster at the low values of
CO and FR. The fuel rate relationship indicates that for some portion of the dataset, ECO is
linear with FR. This is consistent with CMEM and the literature. The relationship is linear with
VSP as well, but the scatter is somewhat less well defined with VSP. The correlation coefficient
is 0.65 for fuel rate and 0.46 with VSP:
Peco/fr - 0.65
Peco/vsp = 0.46
Engine out CO in NCHRP Tier 1 Cars
~ ECO
?! 3.0-
4.0 5.0
Fuel Rate (g/s)
Figure 11
sampled)
Engine out CO vs Fuel Rate for the NCHRP Tier 1 cars run on the US06 cycle (10%
21
-------
Engine out CO in NCHRP Tier 1 Cars
~ ECO
—6t0-
~
»
~
4.0-
~
•• t
~ •
V
~
~
~
~ ~
t ~
# ~
~~
~
~
%
»
~ ~
~ #
* * * * \
~~ ~
~
~
~
~ ~ ~
V
i ~~ ~ «
~ ~ ~~~
» t*
~
~
~
~
~
~
~
~
~
1 1 1 1
-10.0 0.0 10.0 20.0 30.0 40.0 50.0
VSP (kW/Tonne)
Figure 12. Engine Out CO vs VSP for the NCHRP Tier 1 cars run on the US06 cycle (10%
sampled).
Enrichment
Since CO emissions are extremely sensitive to enrichment, the scatter from the fuel rate plot is
assumed to come from enrichment. The ECO deviates from the simple linear relationship upon
enrichment, when the trend sharply increases (many references). The point at which the
enrichment occurs is referred to as the enrichment threshold. An example of this threshold can be
seen in the figure 13. Notice the sharp "knee" where the linear relationship starts to increase.
This can be modeled with a piecewise linear function.
8.0
~ 7.0
jo
3 6.0
O 5.0
•5 4.0
2 3.0
¦I) 2.0
£ 1.0
0.0
~
4.
~~
~
~ *
~~
~~:
~
,
•
<
~ ECO
A ~
~
1 1 1 1
1 1 11
111
111
0
345
Fuel Rate (g/s)
8
Figure 13. An example of ECO trends as a function of fuel rate for the Honda Accord (vehicle
284 on the NCHRP dataset). The "knee" that separates the two linear regions is the enrichment
threshold.
22
-------
While all the Tier 1 vehicles follow the same low FR trend, the location of the knee might
change, thus a family of enrichment curves would be visible (thus causing the scatter when
viewed all on a single plot). Nam (2000) determined that the enrichment threshold is dependent
on the product of weight and engine displacement. Using a larger data set, we investigate this
further. Table 4 shows the correlation of fuel rate threshold with various vehicle parameters.
Vehicle weight apparently has the best correlation, though they are all good. However, this table
does not include other light trucks, whereas the former study included some. Future analysis
should be conducted to include light trucks to see which of the variables gives the strongest
correlation.
Table 4: Correlations of enrichment threshold with vehicle parameters.
Parameter Correlation
Weight
Weight*Vd
Vd
Rated Power
0.875
0.863
0.861
0.756
Figure 14 demonstrates this trend with weight and displacement. The trend is less pronounced
when enrichment threshold is plotted as a function Power/weight. Truck thresholds are expected
to follow the trend, but this requires additional study.
Fuel Rate Enrichment Threshold for Tier 1 Cars
R = 0.7522
FRth
Linear (FRth)
1.5 2
Weight * Vd (tonnes*liter)
Figure 14. Enrichment threshold as a function of weight * engine displacement.
CMEM models ECO in the following manner:
ECO = [c0 * (1-cp-1) + aco] * FR
where
(10)
23
-------
c and a are fitted parameters
cp is the normalized fuel to air ratio.
This equation is satisfactory as it stands and can be used for MOVES. However, the enrichment
threshold relationship can be used to simplify the model and remove the dependence on cp
altogether.
If FR < FRth
ECO = aico * FR. (11)
If FR > FRth
ECO = a2Co * (FR-FRth) + aico * FRth, (12)
where
a's are fitted parameters, whose ratio may be defined by a simple constant. The latter part
of the equation is needed to ensure continuity. Also,
FRth = A*Wt*Vd + B (13)
is determined from the analysis above. The weight and engine displacement can either be
assumed or a distribution can be input into the model based on fleet information. A and B are
fitted coefficients using linear regression.
A = 1.26 ± 0.12, B= 1.82 + 0.16
This enrichment relationship will be used in other portions of the model where enrichment plays
a role.
Engine Out HC
The engine out relationships with fuel rate and VSP are shown in the following two figures. Both
for Fuel rate and VSP, the trends are reasonable. The correlation coefficients are:
Pehc/fr = 0.44
Pehc/vsp = 0.34
24
-------
Engine out HC in NCHRP Tier 1 Cars
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0
Fuel Rate (g/s)
Figure 15. Engine Out HC as a function of Fuel Rate for Tier 1 cars on the US06 cycle.
Engine out HC in NCHRP Tier 1 Cars
0t3G-h
0 25 -
~
¦ [~EHC
~
~
~
~
~
~
~
~ ~
t.
.~~
~ ~
• ~ ~
~
~
~ ~~ u-,u "
~
~ ~ ~ f* ~
.~~~% i + t
~
~1 ~ ~
**• ~~~~~*
¦ '-Z-—i—i—i—i—i—i—i—i—i—i—i—
iii
-10.0 0.0 10.0 20.0 30.0 40.0 50.0
VSP (kW/Tonne)
Figure 16. Engine Out HC emissions as a function of VSP for Tier 1 cars on the US06 cycle.
CMEM models EHC in the following manner:
EHC = aHC * FR + rHc (14)
where
a and r are fitted parameters.
This is a simple linear equation with an offset. The offset is typically quite small, but not
insignificant. The analysis above supports the use of the above parameterized equation to predict
engine out HC emissions. CMEM also adds modules for enleanment events, which complicate
the model. The enleanment HC puffs should only significantly affect emissions from vehicles
with malfunctioning catalysts, and should only add a small error to the model.
25
-------
Engine Out NOx
The following figures show the ENOx trends with fuel rate and VSP. The plots are both quite
scattered, though the fuel rate relationship is better. The correlation coefficients are:
Penox/fr - 0.81
Penox'VSp = 0.73
Engine out NOx in NCHRP Tier 1 Cars
0.35
0.30
— 0.25
«
3
O 0.20
0.15
c
'o>
0.10
0.05
0.00
I~ENOx I
»A.«
#~* ~~ ~ ~ ~
~ ~ .i* v ~~
• ~ ~
Stet/c -J* • -•
*1 • • ~
• ~ ,~ ~ *
«;
5.0
6.0
7.0
Figure 17
cycle.
Fuel Rate (g/s)
Engine Out NOx emissions as a function of Fuel Rate for Tier 1 cars on the US06
Engine out NOx in NCHRP Tier 1 Cars
3
o
ENOx
r:» s *
* ' *| 1 1 1 1 1 1
-10.0
30.0
40.0
50.0
VSP (kW/Tonne)
Figure 18. Engine Out NOx emissions as a function of VSP for Tier 1 cars on the US06 cycle.
26
-------
We need to study this scatter in more detail. The hypothesis is that, the effect is due to differing
levels of exhaust gas recirculation (EGR) strategy among the manufacturers (Nam, 2000). In this
study it was observed that NOx emissions was simply linear with fuel rate with vehicles that
used little to no EGR, but was non-linear (2nd order term) with engines that incorporated strong
EGR. The hypothesis indicates that the family of curves should look like the following figure.
No EGR
Strong EGR
Figure 19. Expected family of NOx curves for varying levels of EGR.
The following series of figures support this hypothesis.
|
HI
0.20
0.15
0.10
0.05
0.00
0
~ ~ 4 *
~ E
NOx
iV» *
~w
~
~ ~
~
~ ~
~ 4 ~
~
~
i
h. .
1
~ ~
~ ~
~ ~
~
FR 3
Figure 20. NOx trends without EGR (vehicle 213 category 8- 95 Civic)
27
-------
8
HI
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
«~
~
~ bNOx
.«
~
j> *
^ ~
~]
a***
r.
~
4
BP"
~ ~
*
0
2 FR 3
Figure 21. NOx trends with light EGR (vehicle 281, category 9- 93 Accord)
x
O
z
LU
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
~ ENO>
" 4
~
~ L
~
~
~
*
r*r«
~
~
f
* X
Lmm
I
I
0
FR
8
10
Figure 22. NOx trends with heavy EGR (vehicle 302, category 8- 95 Taurus)
Much of the high fuel rate scatter in the plots are due to enrichment events.
However, there are some vehicles that don't follow this trend of increasing non-linearity with
increasing EGR. One such example is presented in the following figure. Notice that the slope is
low compared to the previous plots. It is conceivable that the EGR ramps up even at high loads
in this vehicle.
~ ENOx
FR 3
Figure 23. NOx trend with Heavy EGR. (Vehicle 260, category 9, '95 Camry). It does not follow
the trends above.
28
-------
Engine out NOx has been modeled separately as a linear function in CMEM such that
ENOx = aNOx * FR (15)
where
a is a fitted parameter.
Due to EGR effects and the significance of low load emissions when the air conditioner is on,
Nam (2000) added a second order term, but we will simplify this further:
If FR < FRth
ENOx = amox * FR + a2Nox * FR2 (16)
where
FRth is the enrichment threshold (equation 17)
If FR > FRth
ENOx = a3NOx * (FR- FRth) + [amox * FRth + a2NOx * FRth2] (17)
The added terms are to ensure continuity. This is the form of the NOx model used for the rest of
this paper. While it is disadvantageous to complicate, or add parameters, to CMEM, this factor
may be necessary to improve the long-term predictability of the MOVES.
Air Conditioner
Excess emissions from air conditioner usage can also be modeled using a physical approach.
Nam (2000) modeled emissions employing the fuel rate approach. Basically, the air conditioner
seems to affect emissions linearly with the added load for CO and HC. By adding the excess load
into equation (6), the added emissions can be directly calculated using the same (engine out and
tailpipe) emissions parameters. Thus an increase in fuel rate results in a proportionate increase in
CO and HC emissions. The emissions from NOx, however was seen to be non-linear at low load
because of reduced exhaust gas recirculation effects. The details of the model can be found in the
reference. The activity portion of AC usage was determined from Koupal (1998).
Engine Out Calibration
Using the relations above, the parameters shown in table 5 were calibrated along with their
uncertainties. The fits were determined using a simple least squares fitting routine on the SPSS
software program. The uncertainties are 95% confidence intervals determined from the output of
the fit. The uncertainty in the parameter due to test variability is assumed to be follow a normal
distribution. However, from observations of the emissions scatter, it is likely that the estimation
of the uncertainties can be improved using lognormal, or some other asymmetric distribution
transformation. This task will be left to the next step of model development. The following table
shows the R fit to all of the second by second data in the Tier 1 NCHRP data set and the % error
to the total measured emissions (proportional to bag). This is the simplest level validation since it
is validating to the same data set used to calibrate the model. The bag comparisons and the fits
are quite good as well, with the exception of HC. This is due to the extremely low values of HC
in the data set as well as the omission of an enleanment (or "hydrocarbon puff) model.
However, we hypothesize, that the catalyst will eliminate most of the puffs.
29
-------
Table 5. Engine out emissions fitted parameters with high and low 95% confidence values.
Parameter
Value
St Dev
95% CI LC
95% CI HI
CO
0.4074
0.0072
0.3929
0.4219
aCO
0.1174
0.0004
0.1167
0.1182
aHC
0.0100
0.0001
0.0099
0.0101
rHC
0.0049
0.0001
0.0047
0.0050
a1NO
0.0043
0.0001
0.0041
0.0044
a2NO
0.0219
0.0001
0.0216
0.0222
a3NO
-0.0026
0.0017
-0.0060
0.0008
R2
% error
ECO
0.61
4.6
EHC
0.25
-0.6
ENOx
0.66
-9.3
One of the most stringent validation criteria for emissions modeling is to ensure that it matches
both the total (or bag) as well as the second by second emissions. A tool that is often used to
verify this is the cumulative emissions plot. Figures 24 and 25 show the cumulative emissions for
each of the species during the FTP and US06 driving respectively. Since this is a cumulative plot
over all (or most of) the vehicles in the data set, the model will undoubtedly overpredict on some
vehicles and underpredict on others. It should be reiterated that this model is not meant to be
used to predict emissions for a particular car, therefore a second by second validation may be too
stringent a criteria.
Comparison of this engine model calibration to the measurements indicate that HC is relatively
well across FTP and US06 type of driving. The model also indicates that the calibration for all
species is good for the US06 cycle, however CO and NOx overpredict at lower loads (FTP type
driving). It is possible that the catalyst model can compensate for this, but the discrepancy should
be investigated further. It is interesting to note that the US06 data is 'driving the calibration' due
to its relatively high emissions levels (compared to the FTP). It raises the question of whether
these are the appropriate driving cycles to calibrate the model. More will be discussed on this
later.
Using these parameters, we can calibrate a tailpipe emissions model, even if engine out data is
lacking. The variability with these coefficients is small enough to justify the use of these same
parameters for all similar engines. The majority of uncertainty and variability in emissions comes
from the catalyst, which we will discuss in the next section. However, with improved technology
changes expected for the future, it will be necessary to occasionally collect engine out emissions
data to keep the model well calibrated.
In conclusion, we have demonstrated that engine out emissions have a correlation with measured
fuel rate. It is even slightly better than the correlation with VSP. This is natural since fuel rate, as
an engine load, is more causal to engine emissions than a road load variable such as VSP.
However, the correlations are still good with VSP, and can be used also. The scatter that is
evident for CO and NOx is apparently due to vehicle variations of enrichment thresholds and
EGR rates respectively. The enrichment thresholds can be explicitly modeled as a function of
vehicle weight and engine size. The EGR scatter is somewhat more complicated to model, due to
lack of data. It is likely that the model may have to resort to a simple average for NOx, as we
have done above.
30
-------
FTP
3000.0
2500.0
<£. 2000.0
o3 1 500.0
O
O
LU
> 10 0 0.0
E
o
500.0
0.0
ECOmod sum
E CO su m
1 0000
5000
20000
tim e
5000
25000
30000
FTP
400.0
350.0
2 300.0
o 250.0
a> 200.0
O
n
« 150.0
5 100.0
50.0
0.0
EHCmod sum
EH C sum
450.0
5000
1 0000
1 5000
F1tpyie
20000
25000
30000
ENOmod sum
ENO sum
350.0
300.0
® 250.0
w 200.0
150.0
100.0
10000
15000
20000
time
5000
25000
30000
Figure 24. Measured and modeled cumulative engine out FTP emissions for CO, HC and NOx
31
-------
U S 06
5000.0
4500.0
ECOmod sum
E C Osum
4000.0
3500.0
9. 3000.0
2500.0
E 2000.0
1 500.0
1 000.0
500.0
0
0
2000
4000
6000
8000
1 0000
1 2000
1 4000
1 6000
1 8000
20000
tim e
U S 06
450.0
400.0
EHCmod sum
EH C sum
350.0
300.0
250.0
200.0
1 50.0
1 00.0
50.0
0
0
2000
4000
6000
8000
10000 12000 14000 16000 18000 20000
tim e
US06
1000.0
900.0
ENOmod sum
ENO sum
800.0
700.0
O
z
600.0
LU
<1>
>
¦43 500.0
_ro
I 400.0
3
o
300.0
200.0
100.0
0.0
0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000
tim e
Figure 25. Cumulative engine out measured and modeled CO, HC, and NOx emissions
from the US06 cycle.
32
-------
Second by Second Catalyst Pass Fractions
The aftertreatment system in an automobile (including catalyst and air to fuel ratio) is one of the
most complicated vehicle systems to capture in a model. This is because the conversion of
pollutants on the catalyst is highly sensitive to the air to fuel ratio of the combustion and the flow
rate of exhaust. The stoichiometric mass ratio of air to fuel ratio (A/F) required for complete
combustion in the engine is approximately 14.6. As little as a half of a percent lean of this value
could lead to significant NOx spikes (Nam, 1999). This is caused by an excess of oxygen in the
exhaust stream, which in turn prevents the reduction of NOx into its elemental parts: O2 and N2.
Alternately, rich excursions produce CO (and to a lesser extent HC) spikes, since a lack of
oxygen in the exhaust stream prevents the oxidation of CO into CO2.
Without accurate measurements of the A/F, catalysts can, at best, only be roughly approximated.
The data is not available, and the modeling of A/F is beyond the scope and use cases for
MOVES. However, we demonstrated that we can predict when vehicles will go into enrichment,
so CO and HC emissions should be somewhat easier to model than NOx.
The catalyst pass fraction is defined as the ratio of tailpipe to engine out emissions:
CPF = TP/EO (18)
or
TP = CPF * EO
When CPF = 0, the catalyst converts all (or 100%) of the pollutants. In more familiar terms, it is
related to the catalyst efficiency (%) by the following equation:
Cat Eff = 100% * (1 - CPF) (19)
It is easier to deal with CPF since we are searching for trends that increase with tailpipe
emissions. During extreme breakthrough events, it is sometimes possible for TP > EO (or for
CPF > 1) at any given second. Theoretically this is unphysical, but it is possible to measure this
as an artifact of timing mismatches between the second by second engine out and tailpipe
measurements.
The following sections describe a model for the catalyst pass fraction during hot running
operation. Cold start is substantially different and calls for a different approach. Since cold start
only occurs during an early portion of a trip, it will be modeled in a future report.
CPF CO and HC
Previous studies have found loose correlations between CPF and FR or with Engine out
emissions. CMEM, for example, employs the following equation for both CO and HC:
CPF = 1 - r*exp[(-b-c*(l-(p1)) *FR] (20)
where
ris the maximum CO (or HC) catalyst efficiency
b is the stoichiometric fitted parameter
33
-------
c is the enrichment fitted parameter
(p is the normalized fuel to air ratio.
At its asymptotes, this equation approaches 0 when FR=0, and 1 when FR=°o which roughly
says that as the exhaust volume flow increases, the catalyst ability to oxidize the pollutants
decreases due to the decreased exposure time. This is also known as "catalyst breakthrough".
Generally, this may be true, but the catalyst correlation (p) with fuel rate is quite poor, and it is
even worse with VSP:
Pcpfco/fr = 0.42
Pcpfco/vsp = 0.27
The hydrocarbon trends are not as clear as those of CO. The CPFHC correlation coefficients are
listed below:
PCPFHC/FR = 0.34
Pcpfhc/vsp = 0.23
Figure 26 shows the relationship between CPFCO and FR. The points have been separated by
stoichiometric/enrichment and light/heavy categories. While the scatter is excessive, there is still
a general increasing trend with FR, but there is no obvious weight trend discerned.
Catalyst PassFraction CO in NCHRP Tier 1 Cars
1.0
o CPFCO stoic <3500
A CPFCO stoic >3500
¦ CPFCO rich <3500
• CPFCO rich >3500
Fuel Rate (g/s)
Figure 26. CPFCO as a function of FR for Tier 1 cars.
The scatter in CPF plots is tremendous. HC, and NOx are no better in this regard. After a
significant effort to attempt to parameterize CPF behavior for all of the data points in this study,
it was deemed best to either complicate the model beyond the level of CMEM (by including an
air to fuel ratio algorithm, temperature model, and perhaps even "engine stress"), or greatly
34
-------
simplify it (sacrificing some accuracy). We chose the latter route. The approach selected is to bin
the CPF by VSP. In the flow chart (figure 8), this is just above decision 5. Figure 27 shows the
CPF binned by VSP. It is advantageous to bin data since it has an averaging effect over time,
smoothes out some very stochastic and "spiky" time series data.
si = 10847 5400 12026 10189 7473 5297 4144 3651 2859 2890 1735 802 695 363
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure 27. CPFCO binned by VSP and plotted logarithmically.
The plot appears well behaved but upon further investigation we encounter the following
difficulty. When the measured CPF is separated for the FTP and US06 cycles (figure 28), we
unexpectedly see a bifurcation in the data. An effect is also visible on HC, though it isn't as
significant. The effect is not as pronounced when plotted on a linear scale.
0.0
2.0
J
3
I
m
£ i m m hi
m ju
£
m
M
I
[i
1 1
m. at
w rn
TRIP
I
~ FTP
I
~ US0(
6909447(BB602t4ET948Q) 1555E 33818788329£0242'53232'2336594$686 1801 695 363
1 2 3 4 5 6 7 8 9 10 11 12 13 14
O) -3.0
8082293314718tS03H)493Z'1t2j14O7S2$9@37BO20O8j52379!3571349710 1810 701 364
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
VSPBIN
Figure 28. Measured CPF CO and HC binned by VSP but separated by driving cycle.
This separation in the data does not appear to be an engine out phenomena, but rather that of the
aftertreatment system. Figure 29 and 30 show engine out and tailpipe CO respectively
confirming this. There is a slight engine out discrepancy in bins 4 and 5, but it is much more
35
-------
pronounced in the tailpipe plot. Similar relationships occur when plotted as a function of fuel rate
instead of VSP.
~ FTP
O -2.0
~ US06
:M«mqi7iiqii4miH40T8HGrattros327aa7iaifl7ifli8i<) 701 364
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSP BIN
Figure 29. Measured ECO binned by VSP and separated by driving cycle.
~ HP
~ US06
i 6TOJMI*HBJITOHa5SI3atlflTI«E»BDBBC3Ja2aai5,JlflBa6iafl1 645 363
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure 30. Measured TPCO binned by VSP and separated by driving cycle.
The plots indicate that for the same road load conditions, tailpipe emissions are higher in the
more aggressive high-speed driving cycle. Younglove et al (1999) and Frey et al. (2002,
Appendix A-7) noticed this "speed effect" in their studies as well. The latter study observed it in
the RSD data. Their conclusion was that speed was the differentiating factor. Some driving has
higher speed activity at lower VSP. It is still possible to be driving at high speed yet have nearly
zero VSP if the vehicle is not accelerating. Figure 31 shows the speed distribution of the two
cycles as a function of VSP.
36
-------
8Oi22S4lBt72SB61D493?1H614O732690373O20O$S3792a7134971O 1810 701 364
1 2 3 4 5 6 7 8 9 10 11 12 13 14
TRIP
I
~ FTP
I
~ US06
VSPBIN
Figure 31. Speed on the FTP and US06 cycles within each VSP bin.
However, this alone is not likely sufficient to cause a significant increase in emissions since the
engine does not need to work hard during coasting and braking, thus it is not expected to have
higher emissions during these periods. There are many reasons why this may occur: throttle
dither, catalyst breakthrough, enrichment, other engine strategy differences, or something else
unforeseen. Further investigation is required to find the cause and then capture the phenomena in
PERE.
Separating out by modes, figure 32 shows CPFCO by VSP bin where only the accelerating data
points are employed. Note that the bifurcation is significantly reduced when compared to figure
31. Figure 33 shows the points where the vehicle is coasting or decelerating, verifying that these
modes are the major cause of the split. There is little difference in the coasting and the
deceleration profiles, so they are combined for this study. The definitions of the modes of driving
for this report are outlined in table 6. The enrichment mode is not included on this table.
Table 6. Modes of driving.
Mode Condition
Deceleration accel < -0.2 mph/s
Idle v < 2 mph
Cruise -0.2 < accel < 1 mph/s
Acceleration accel > 1 mph/s
It is likely that the definition of cruise (based only on acceleration) will change depending on the
speed. In a future study, it may be advantageous to better define these modes of driving. Cruise
driving implies that the driver is not driving at exactly the same speed or throttle. Every driver
dithers the throttle some amount. Even when using cruise control, road grade keeps the throttle
changing in order to maintain constant speed. Moreover, the added (minor) decelerations that
37
-------
occur during cruise driving, tend to reduce the instantaneous value VSP (as well as fuel rate).It is
this low acceleration type of driving we are defining as the "cruise" mode.
o
o
LL
CL
O
CD
_i
o
LT>
CT>
TRIP
~ FTP
US06
34233 102141207515 254W1110B58 50Q69 331371 178556 49779 1 475 498 342
3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure 32. CPFCO by VSP bin. Only the points where accel >1.0 mph/s are included.
Ell
~ FTP
US06
N = 6ASBMKUiSOHBEH858217S3II13SS151(1582)003 107 326 117 SI
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure 33. Cruise and decel CPF CO data as a function of VSP bin. Idle is omitted.
According to the above modal definitions, we can devise the following 'preliminary' model for
CPFCO.
During acceleration and enrichment (low accel) events:
CPF =(1-1) *exp(bj *FR)
(21)
38
-------
For higher emitter vehicles, it may be necessary to modify these algorithms so that they
maximize around 1.
During cruise and decelerations, the quantification is not as straightforward. The difference
between these two driving cycles is the aggressiveness of their cruise sections. Second by second
aggressiveness is very difficult to quantify without knowing the history. We will leave it up to a
future study to better quantify this. For this study, we will correct merely by speed as was
recommended by Frey et al. (2002). In Figure 34, the CPFCO is plotted by (instantaneous) speed
bin, where the bins are in 5 mph increments as defined in table 7.
Table 7. Speed bins
Bin Speed Range (mph)
1 0-5
2 5-10
3 10-15
10 45-50
15 70-75
16 75+
log CPFCO
1
~ 0-5mph
~ 5-10mph
~ 10-15mph
~ 15-20 mph
~ 20-25mph
~ 25-30mph
~ 30-35mph
~ 35-40-mph
II40-45mph
~ 45-50mph
~ 50-55mph
¦ 55-60mph
¦ 60-65mph
¦ 65-70mph
¦ 70-75mph
¦ 75+mph
VSP bin
70-75mph
60-65mph
50-55mph
40-45mph
30-35mph
20-25mph
10-15mph
0-5mph
Speed bin
Figure 34. CPFCO during cruise and deceleration plotted by VSP and speed bins.
From figure 34, it appears that cruise and decel CPFCO is relatively independent of VSP. Thus,
we can model it the following way:
39
-------
CPF = (1 - T)*exp(b2*v)
(24)
Completing figure 34, the next figure shows what it would look like if all the VSP, speed bins
were filled. Notice also that high speed, emissions seem to be relatively independent of VSP.
These cells tend to be highly populated by the high speed cruise portions of the US06 cycle.
log TPCO
Figure 35. Measured TP CO for the FTP and US06 driving cycles as a function of VSP and
speed.
Due to the specific nature of the two driving cycles used to calibrate, speed may only be an
incidental causal variable. It is unknown at this time, whether less aggressive high speed cruises
(e.g. using cruise control) would give lower emission rates. Looking more closely at the MEC01
cycle (the driving cycle used to validate this model) should shed more light on this since it has a
number of cruise events at varying speeds. Complicating matters, figure 36 indicates that speed
alone will not explain the cruise/decel difference between the two cycles.
~ 0-5mph
~ 5~10mph
~ 10~15mph
~ 15~20mph
~ 20~25mph
~ 25-30mph
~ 30-35mph
~ 35-40-mph
~ 40-45mph
~ 45-50mph
~ 50-55mph
¦ 55-60mph
¦ 60-65mph
¦ 65-70mph
¦ 70-75mph
¦ 75+mph
14 13 12 11 10 9
VSP bin
75+mph
60-65mph
45-50mph
30-35mph gpeed bjn
15-20mph
0-5mph
40
-------
-.5
-1.0
1
Jim n
CL
o
o
o
o
o
-1.5
-2.0
m m T m m a
[i
m
m
TRIP
I
~ FTP
I
~ US06
305QTI4Z304<8«ia7MB73Se37Hi2 990715«f 114S491283 9020429324622 28881825 363
1 3 5 7 9 11 13 15
2 4 6 8 10 12 14 16
VBIN
Figure 36. CPF CO during cruise and decels vs speed and separated by driving cycle.
Another limitation of this methodology is that equations 22 and 23 are not necessarily continuous
(temporally). However, this may not pose a serious problem as long as second by second
emissions are not being modeled.
It is clearly necessary to calibrate this model based on a more varied sample set. To get a sense
for the limitations of these two driving cycles, Figure 37 shows the relative activity of the FTP
and US06 driving cycles on a VSP and speed grid. Figure 38, shows the activity in a sample of
on-road driving (taken from the EPA PEMS data set) conducted in southeast Michigan. Note the
lack of driving within speed range 30 to 50 mph for the driving cycles. This indicates that the
driving cycles accelerate through this region, rather than cruising in it (hence the lack of
activity). This could potentially bias the CO model more toward a straight VSP approach due to
the lack of cruise driving.
However, it is clear that within the cruise mode, there is a speed or aggressivity affect on
emission, even within a VSP bin. The same effect is not evident for HC during cruise.
41
-------
0.05 Y\
0.0454^
0.04-K
0.035-V"
0.03-V
Activity 00254
Fraction 0.02-j
0.015-
0.01-
0.005
o
¦ 0-5mph
~ 5-10mph
~ 10-15mph
~ 15-20mph
~ 20-25mph
~ 25-30mph
~ 30-35mph
¦ 35-40-mph
¦ 40-45mph
¦ 45-50mph
¦ 50-55mph
¦ 55-60mph
¦ 60-65mph
¦ 65-70mph
¦ 70-75mph
¦ 75+mph
Figure 37. Activity fraction of VSP and speed bins for the FTP and US06 cycles. Idle is omitted.
~ 0-5mph
~ 5~10mph
~ 10~15mph
~ 15-20mph
~ 20~25mph
~ 25-30mph
~ 30-35mph
~ 35-40-mph
~ 40-45mph
~ 45-50mph
~ 50-55mph
¦ 55-60mph
¦ 60-65mph
¦ 65-70mph
¦ 70-75mph
¦ 75+mph
activity
70-75mph
60-65mph
50-55mph
40-45mph
30-35mph
20-25mph
10-15mph
0-5mph
VSP bin
Speed bin
Figure 38. Activity of On-board emissions data. This is a sample of driving in southeast
Michigan from 18 drivers.
42
-------
For CPF HC, the discrepancy originates from decelerations. Figure 39 shows the CPF HC points
so that only the cruise and acceleration points are plotted. Figure 40 shows only the deceleration
points. The discrepancy in deceleration is undoubtedly due to enleanment arising from
deceleration fuel shut off (An et al., 1998). Due to the positive terms in equation 1, it is possible
to be decelerating, yet have a positive VSP.
o
X
Ll_
Q_
O
O
_i
o
LO
CD
~ FTP
~ US06
237484939322657780385114719931910B20855224123Z673 49 702 1 810 701 364
3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure 39. Cruise and acceleration CPF HC as a function of VSP bin. Idle is omitted.
-1.0
-1.5
O
X
LL
Q_
O
O
_i
o
LO
CD
-2.0
-2.5
-3.0
J 5 i; m
-Km
I
I
I
IjT HE
TRIP
I
~ FTP
I
~ US06
N = 803K29539747261486 109810 536771 222079 111054 500 138 40 8
1 23456789 10 11
VSPBIN
Figure 40. Deceleration CPF HC as a function of VSP bin.
To better isolate this effect, we create deceleration bins:
43
-------
Table 8. Deceleration bins
Accel Bin Range (mph/s)
1
2
3
4
<-3
[-3,-2)
[-2,-1)
[-1,0)
Figure 41 shows that the hydrocarbon puffs are breaking through the catalyst in the positive VSP
values but moderate decelerations (accel bin 3, VSP bins 3-6). The large error bars indicate that
there are few data points with high decelerations in positive VSP bins (VSP bin >3).
A simple relationship that might work for HC during cruise, accelerations, and slow
decelerations, is as follows (same as for CO):
CPF =(1- rHC) *exp[b1HC *FRJ
During decelerations, it is similar to CO:
CPF = (1 - rHc)*exp[b2Hc*v]
(23)
(24)
o
X
LL
Q_
O
O
_i
o
LO
CD
-1.0
-1.5
-2.0
-2.5
-3.0
!>
I
ACCBIN
1.00
2.00
3.00
4.00
N = 38HB414t®3976 2556 4878 111123 571
1 2 3 4 5 6
VSPBIN
235 58
8 9
10 11
Figure 41. Deceleration CPF HC as a function of VSP bin split by binned acceleration.
Many previous publications employ the use of speed, acceleration, emissions (or even v*a, v and
emissions) "maps" to model emissions (Pischinger, 1998). The approach presented here maps
road load and speed with emissions for the some of the modes of driving.
CPF NOx
Historically, the catalyst behavior of NOx has always been extremely difficult to model. This is
due to many of the factors mentioned above. Just a simple study of the correlations gives the
following:
44
-------
PCPFNOx/FR = "0.01
PCPFNOx/VSP = -0.02
PCPFNOx/EINOx = "0.07
The CPF NOx trends as a function of FR are extremely scattered (not shown). The VSP, speed
CPF plot is shown on Figure 42. There is no simple monotonically increasing function that
would describe this relationship. The NOx curves have a local maxima or "islands" of emissions.
There appears to be two minima at VSP bin 5 at roughly 25 and 55 mph. There may also be one
at 65 (the cruise sections of the US06). Oddly, these roughly correspond to the activity peaks in
figure 38.
log CPFNO
VSP bin
70-75mph
l-65mph
l-55mph
40-45mph
30-35mph
20-25mph
10-15mph
0-5mph
¦ 0-5mph
~ 5-10mph
~ 10-15mph
~ 15-20mph
~ 20-25mph
~ 25-30mph
~ 30-35mph
¦ 35-40-mph
¦ 40-45mph
¦ 45-50mph
¦ 50-55mph
¦ 55-60mph
¦ 60-65mph
¦ 65-70mph
¦ 70-75mph
¦ 75+mph
Speed bin
Figure 42. CPF NOx as a function of VSP and speed. The curve is rotated in comparison to the
others seen so far to give a better indication of the texture.
To simplify matters, we first look to see if there is a fundamental difference in NOx profiles for
the two driving cycles in this data set. Figure 43 shows that the two cycles do not seem to be
significantly different (except VSP bin 4). "Significant difference", here is defined as non
overlapping confidence intervals.
45
-------
~ FTP
~ US06
^ O A
"2.4
72834«76E6OOCBB95S9l26-gB284S8S0231JSO@B5l52B34:268O4®69718O8 700 364
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure 43. CPF NOx as a function of VSP split by driving cycle.
Due to the low engine out emissions at the lower VSP bins (negative values), it is not as
important to accurately depict this region. We can use the same model as above:
CPF = (1 - rNOx)*exp[b1NOx*FR]
(25)
Given these simple algorithms, we can proceed to calibrate and validate the tailpipe portion of
the model.
Tailpipe Calibration
The engine out behavior is relatively straightforward and 'unscattered'. Because of this, a least
square regression gives parameters, which fit the second by second traces as well as giving
relatively accurate total emissions. However the catalyst module is somewhat more complicated.
For the catalyst model, it is advantageous to fit first to the low fuel rate regime (equation 21) and
then the high. It is determined that the best method is to calibrate the T parameter first, based on
idle (VSP bin 3) emissions. This is basically the minimum of the CPF emissions profile and sets
the baseline. For NOx, this minimum happens to fall in VSP bin 4, rather than 3. This parameter
can be read straight off of the plots.
For CO, table 9 indicates the values of the VSP bin 3 (and lowest speed) values of logCPF along
with their uncertainties (variabilities in this case). The uncertainties in Y are propagated using
standard error propagation equations (Taylor, 1982). For error (uncertainty) propagation, we
assume normal or log-normal distributions where appropriate. In this case, the uncertainty in
CPF was small enough to simply carry over (calculate) the uncertainty in catalyst efficiency
(sr).
46
-------
Table 9. Idle (or minimum) catalyst pass fractions (CPF0) and catalyst efficiencies (F) along with
their uncertainties (mean standard deviations).
logCPFO
d(logCPFO)
r
sr
CO
-2.18
0.006
0.993
0.007
HC
-2.57
0.005
0.997
0.003
NO
-1.55
0.012
0.972
0.028
For NOx, the idle VSP bin does not have the lowest CPF. That is in VSP bin 4 of the FTP cycle.
In that bin the T = 0.993. We will use this term in the calibration.
Since the catalyst pass fractions are so scattered (being a ratio of two measured quantities), it
makes much more sense to fit to the observable tailpipe, rather than the CPF. The question
remains whether the parametric fits should be conducted on a linear or logarithmic scale. For this
study, it turned out best to fit CO and HC logarithmically and NOx linearly. This will need to be
investigated further in a future study.
The next step in calibration is to perform a regression fit to the acceleration modes (equations 21,
23, and 25) using 67 as an adjustable parameter. EO and FR are modeled values. Then we can fit
to the cruise and or deceleration modes (equations 22 and possibly 24). One could either
calibrate to the second by second data directly, or to the aggregate data, or even to the
cumulative emissions. In this section, we calibrated to the second by second data, which can be
somewhat computationally intensive.
When this is done, the following coefficients along with their uncertainties are obtained:
Table 10. Fitted Catalyst parameters
Parameter
Value
St Dev
95% CI LO
95% CI HI
B1CO
0.749
0.012
0.725
0.773
B2CO
0.0216
0.0002
0.0212
0.0220
GAMMACO
0.993
0.004
B1HC
0.695
0.004
0.687
0.704
B2HC
0.0162
0.0002
0.0158
0.0167
GAMMAHC
0.997
0.001
0.997
1.001
blNO
0.456
0.003
0.451
0.461
GAMMANO
0.993
The following series of figures shows the calibration fits to the combined FTP and US06 driving
cycles. From observation, the fits are quite reasonable. A quantitative description of the goodness
of fit would require a proper analysis of the uncertainties, however this was not conducted due to
the preliminary nature of this model. This will take place in the next stage of model
development.
47
-------
0
-2
-4
O
10 c
CD -5
m
m b-
s
ID
® a
hb b
s
a
e s
N = 1OM8«S0ai2O2e2B109S373732329f1444863a528385a8389(l73S33BOaO26969536363
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure 44. Calibration of TPCO. "M2" is modeled.
LO r-n
CD -5.0
122BB606$3B3$38Bta222S80695SS9®3«333737l83829329S95t)7335S131170710136364
1 2 3 4 5 6 7 8 9 10 11 12 13 14
I
~ LGTCO
I
C LGTCO M2
I
~ LGTHC
I
~ LGTHCM2
VSPBIN
Figure 45. TP HC Calibration to the NCHRP Tierl data set.
48
-------
-1.5
O
-2.0
-2.5
-3.0
-3.5
-4.0
-4.5
c r,
CT> -5.0
m
Sffi
HJ3
em
i-in
Si
' s nt
nt in
: 11178ita«262aSBa®Bllia(rBi6a3a333382ra3B7371 zazaaaaasi S7«46080970ZOCB6364
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure 46. TP NOx Calibration to the NCHRP Tier 1 data set
49
-------
Ill - VALID A TION OF PERE
Fuel and CO2
Validation to MEC
Figure 47 shows the fuel rate comparison of the measured and the modeled. The model is based
on equation 6. The only significant deviation lies in VSP bin 14. It is likely that the enrichment
model is not capturing all of the vehicles. The error bars are 95% confidence intervals so only
reflect the variation in the model (from vehicle weight, etc.), and not the uncertainty. At lower
loads this uncertainty is relatively small, but under enrichment conditions, it would be necessary
to propagate the uncertainty terms from equation 13. For cumulative fuel consumption, the
model overpredicted the vehicle average by 1%.
FR MEAS
FR MODEL
12T2OT391970SOS68a6aOM5ISae40727l34Oa5BaS88831311ISe863923286S88I12616
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure 47. Fuel rate validation for the MEC data. Error bars are 95% confidence interval in the
variation.
Figure 48 shows the validation for average CO2 (calculated from all of the vehicles). This comes
directly from fuel rate equation 8. CO and HC terms can be added, but those are relatively small
correction terms. The comparison is excellent until the two highest VSP bins, where the model
overpredicts due to enrichment. During enrichment, a fraction of the CO and HC is prevented
from being oxidized in the catalyst, hence the combustion of fuel into CO2 and H2O is less than
ideal. The tailpipe CO and HC should correct the model at these higher values in the final form
of the model. It is impossible to quantify this uncertainty bar on these points without a CO
model, so it is omitted for the purposes of the CO2 model. It is presented here in this fashion to
demonstrate proof of concept for the first portion of MOVES (greenhouse gas). The cumulative
CO2 emissions overpredict by 5% (figure 49). With the measured HC and CO as a correction,
this differential drops to 2% (underpredicted).
50
-------
~ C02 meas
~ C02 mod
12T2233S319795096Se630«)525364B72^340«33262B8J131126BQ639282€68$8ai2lB16
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure 48. CO2 validation for the MEC data.
1%
FRmeas FRmod
C02 meas C02 mod C02 corrected
Figure 49. Validation for total FR and CO2.
Validation to UCC
Frey et al. (2002, p 230) reported a significant speed dependence on CO2 predictions. By
comparing the emissions from the vehicles driven on several CARB Unified driving cycles, it
was found that the model was overpredicting on the lower speed cycles and underpredicting on
the higher speed cycles.
As part of the proof of concept, 17 vehicles in the CARB dataset were simulated on the 8 UCC
driving cycles using PERE. Since the engine size was not supplied as part of the database, the
values were obtained from the Kelly Blue Book Website (www.kbb.com). Where there were
multiple values for engine size, the smaller value was assumed. However, if the fuel rate
51
-------
predictions were significantly underpredicting (-20% or more), the higher engine size was
chosen and recalculated. The second by second fuel rate and CO2 were calculated for each
driving cycle and car. This value was then "corrected" by subtracting the CO and HC
measurements to determine the limit of the accuracy of the model (as was done in the previous
section. Figure 50 shows the results from PERE compared to those obtained from NCSU. The
NCSU model overpredicted at low speeds but underpredicted at higher speeds. The results from
PERE were the opposite: it overpredicted as cycle speed increased. However, the speed effect in
PERE is less pronounced. It should be noted that the NCSU results only included the smaller
engine vehicles (<3.5L) whereas the PERE model used all the vehicles since it is able to
accommodate engine size explicitly in the model.
C02 Validation on UCC
0.15
0.1
0.05
15% or more for each of the cycles. More heavy
vehicles (including light trucks) should be examined in a future study to see if the overprediction
trend is real. Other than these vehicles, there was no significant weight effect discerned.
Propagation of Uncertainty
Before presenting the comparison of the engine out model, it is first necessary to discuss the
propagation of uncertainty in the model. It has been stated by many of the users of MOBILE,
that an estimate of the uncertainty in the model is a necessity. The following is a brief discussion
of some methods by which this can be accomplished.
52
-------
With any measured or modeled value, there is inherent variability and error. An example of
natural variability is the distribution of vehicle masses in the modeling data set. This is known
and accounted for in the model explicitly. An example of error is the uncertainty associated with
the enrichment threshold (FRth) of a given vehicle. This quantity was estimated from the scatter
plots as mentioned earlier, without much concern of the "measurement error" because it was
assumed that the unknown variability of the vehicles i.e. the difference in the manufacturer's
specific enrichment strategies (irrespective of their known weight and engine displacement)
would dominate the uncertainty. Thus in equation 13, each of the terms have both inherent
variability and error. These are estimated via the uncertainty in the A and B coefficients, which is
an output of the linear regression in most statistical software packages. For the purposes of this
report, we will use the general term "uncertainty" to denote both measurement error and
variability that is NOT known through natural variability. We discount the natural variability that
is known, such as the variability in vehicle weights in the study since they are explicitly included
in the calculations. We also do not account for systematic uncertainties, which are often
significant. Systematic uncertainties often come about due to a failure of the model to account
for certain physical effects. These important uncertainties are to be corrected (if significant)
when they are discovered.
The question now remains, how can we propagate uncertainties in equations such as 13, to the
final emissions calculations? There are a number of standard methods by which uncertainty in
the terms of equations 11-17 can be propagated through to the engine out calculation: Monte
Carlo, Bootstrap, and Error Propagation. For the sake of simplicity, we use the latter approach. If
the probability distributions are normally (or log-normally) distributed, then the error
propagation equations simplify since you only need to deal with means and standard deviations
(or mean standard deviations) about the mean. This is the most computationally simple of
methods, and the method used in this work. However, it may not be suitable for the final
MOVES product, and will depend on the circumstances.
Uncertainty in Fuel Rate
It is difficult to estimate the uncertainty in many of the variables in the fuel rate equation 6 (and
1). However, as we have already seen in figure 47, the model performs very well especially
when the car is assumed to be operating under stoichiometric conditions. During stoichiometric
operation, there are the following unknown uncertainties in K, N, e, Cr, and CdA. Thomas and
Ross (1997), indicate that some of these parameters are only "adjustable" within a narrow range,
such that the uncertainty associated with these variables are relatively small. These include some
of the variables comprising the simple engine speed sub model, and all of the above terms with
the exception of the engine friction term K. However, since the most of the fuel consumed is
needed to overcome road load (moreso than friction), the FR calculation is probably relatively
insensitive to uncertainties in KNVd. The accuracy and robustness of the fuel rate model attest to
this.
Thus the majority of the uncertainty lies in the enrichment portion of the model. However, due to
the proprietary nature of each of the vehicle manufacturer's enrichment strategies, the
uncertainties in these variables are unable to be estimated. A more thorough study of the
uncertainty in the simple models for 0, should be reassessed. However, due to the nature of the
53
-------
model to predict fuel consumption very accurately, even under enrichment conditions, we will
assume that the uncertainty in fuel rate is small in comparison to other uncertainties in the model,
i.e. 8FR ~0.
Uncertainty in Engine Out Emissions
There is a 1:1 correspondence between VSP and FR according to the fit in figure 47. The values
are shown in the following table for the average VSP value within the particular bin. The bin
definitions are from Frey et al. (2002) and the fuel rate value is the central value in the bin.
Table 11. VSP bins and central VSP and fuel rate values associated with the bins.
VSP bin
Avg VSP
FR
1
<-2
0.419
2
-1
0.339
3
0.5
0.317
4
2.5
0.590
5
5.5
0.857
6
8.5
1.152
7
11.5
1.431
8
14.5
1.643
9
17.5
1.908
10
21
2.100
11
25.5
2.385
12
30.5
2.988
13
36
3.773
14
>39
4.467
Using these bins, we can validate and estimate uncertainties. Some sample calculations are
shown in Appendix E.
54
-------
Engine Out Validation
The next three figures show the comparison of the engine out measurements and the model.
There was no effort to separate out speed effects in the engine out portion of the model. Note that
the modeled results only vary significantly from measurements at low VSP. Emissions tend to be
relatively low at low VSP values, so the error is not as pronounced when compared to total
emissions. Also, the regression tends to fit out the higher emission points. The plots are on a log
scale to give the reader an idea of the fit at across all VSP values. However, due to the relatively
small values of emissions occurring at lower VSP bins, it is probably more important look at the
middle and higher VSP bins to gauge the "goodness" of the model.
The model is indeed very good at the engine out level, i.e. most of the modeled points are
consistent to within the uncertainty with the measurements at each of the VSP bins. There are
some systematic overpredictions for HC in VSP bins 7-10. There is also a systematic
overprediction in bin 3 (idle). It is likely that engine strategy is quite different at idle than during
normal operation. If idle (light vehicle) emissions becomes an important factor to model in
MOVES, it will be necessary to modify or "correct" the idle emissions factors.
[
. 1
3
E
]
s
p
~ IgECOmeas
~ IgECOmod
~
<
3
~
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSP bin
Figure 51. A comparison of the engine out CO measurements from MEC cycle and the model
calibrated to the FTP and US06 cycles.
55
-------
-1.2
-1.4
j -1.6
r
u
o>
^ -1.8
-2
-2.2
~
t
-
< ~
<
I
:
-
„
<
~
~
~ IgEHCmeas
~ IgEHCmod
[]
•
6 7 8 9
VSP bin
10 11 12 13 14
Figure 52. A comparison of the engine out HC measurements from MEC cycle and the model
calibrated to the FTP and US06 cycles.
-0.5
-1
$ -1.5
05 „
o -2
-2.5
-3
a
\ I
1 |
E
4
0
3
~
II
ij E
i
' E
~
<
3
~
~ IgENOmeas
~ IgENOmod
< >
i
1
2 3 4 5 6 7 8
VSP bin
10 11 12 13 14
Figure 53. A comparison of the EO NO measurements from MEC cycle and the model.
The cumulative emissions are shown in figure 54. The error was not propagated to this value at
this time.
56
-------
2.5
'E
0)
o
« 1.5 -
0.5
0
2
ECOmeas/10 ECOmod/10
EHC meas ECO mod
ENOx meas ENOx mod
Figure 54. Measured and modeled cumulative emissions.
Tailpipe Validation
The tailpipe validation plots are shown in the next 3 figures. An uncertainty analysis of the
tailpipe calculations was not conducted in this study, due to the incomplete nature of the post-
catalyst model. However, qualitatively, the model appears to perform reasonably well. The HC
and CO consistently underpredict in the middle VSP regions and this must be considered before
proceeding.
Cumulative emissions did not compare as well as the engine out. In the next step of this study, it
will be important that both the load based behavior and the total emissions are consistent. In our
case, it was clear that the two driving cycles used to calibrate were insufficient to calibrate (or
develop) the model. Some of the cruise driving effects were not accurately captured.
57
-------
1
o
S "4
m
m
m -b
o s
UJ- %Bb
B
: 1O909M2-?2S4828a7803379392424235@8638S®®6Q6O3COQ32593938^816565212B13
1 2 3 4 5 6 7 8 9 10 11 12 13 14
I
~ LGTCO
I
~ LGTCOM
VSPBIN
Figure 55. MEC TPCO validation
6QP
~ LGTHC
~ LGTHCM
S -5.0
12T2Zaj3B197®EO36Ba6IOSa5252W0727l34O®I3a62a83131l126BS6I92a286«88212616
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure 56. MEC TPHC validation
58
-------
-1.5
-2.0
-2.5
-3.0
-3.5
-4.0
O "45
10 C n
CD -5.0
met
em
m
«B S _ ©
^ n &
i
~ LGTNOX
I
n LGTNOM
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure 57. MEC TPNOx validation.
59
-------
IV- CONCLUSION
We have demonstrated in this report that a physical emissions model can be a powerful addition
to MOVES. The advantages are as follows:
• Extrapolations to future model years and deterioration may be modeled in more detail
(though new standards and powertrain technologies may require a different approach)
• It might be capable of modeling some new technologies
• Some fuel effects can be modeled explicitly
• Off-road engines can be modeled (though data is needed)
• Fuel and CO2 are estimated from physical principles directly, rather than empirically
determined
• Many of the variations and anomalies in the data can be explained
• New data can be checked for quality or new trends
• A separate pre and post catalyst (engine out and tailpipe) model for criteria pollutants can
help isolate or pinpoint the source of emissions
• OBD and IM issues can be modeled explicitly
A purely empirical model is incapable of handling many of the above issues in a meaningful
manner without introducing a host of ad hoc correction factors. There are other physical
approaches that can be taken, but the one proposed in this paper, is complex enough to handle
many present and future emissions issues without overburdening the computational loads.
We first compared Vehicle Specific Power with fuel rate (or fuel consumption), and established
the link (and correlation) between the two. The fuel rate approach is the one taken by the CMEM
model developed at the University of California, Riverside (CE-CERT). A fuel rate based model
is not to be confused with other more macroscopic fuel-based models in the literature.
Before modifying, or simplifying CMEM, it was first necessary to demonstrate that engine out
emissions have remained relatively constant since the introduction of three way catalysts and
improved fuel control. This was necessary since engine out emissions are difficult to measure.
Based on the analysis performed, we can go on the assumption that engine out data will only be
required to be taken on occasion, to make checks on the engine model. Using a detailed data set
taken at UC Riverside, which includes engine out data, a Physical Emissions Rate Estimator
(PERE) was introduced. The model is a modification of CMEM in many respects. In some ways
it is simplified, in others, it is more detailed (out of necessity). The fuel rate, CO2 and engine out
models were developed and calibrated to the FTP and US06 cycles. A validation to a different
driving trace (the MEC01 cycle) was performed. A separate validation was performed for CO2
on some of the California unified cycles. The catalyst model was more of a challenge. Due to the
complication of a speed history effect (seemingly independent of VSP), a model was introduced
but not fully developed. Further analysis of the effects of driving cycles is required before the
final algorithms are established. However, the trends described herein, give some indications that
this can be accomplished.
In summary, some of the conclusions drawn in the report are:
60
-------
• Under many operating conditions, VSP and Fuel consumption are well correlated as long as
the full (rather than the approximated) VSP equation is used.
• There are strengths and weakness to both binning and parameterizing data. During the model
development process, it is often beneficial to parameterize to obtain the most information.
However, if the noise in the data is too large (as in catalyst modeling) binning provides a
convenient way to average (and filter) the data. In either case, for the final solution the data
must be binned so that PERE is compatible with the empirical arm of MOVES.
• The powertrain model was mainly captured from previous work in the literature. For the
purposes of MOVES, further analysis and simplification will be required. In particular, the
engine speed model will need to be simplified.
• For Tier 1 cars, engine out (pre-catalyst) emissions for criteria pollutants appears to be fairly
constant across manufacturers, age, and power to weight ratio.
• Therefore, an engine out model should be able to be calibrated with a relatively sparse (but
accurate) data set.
• Comparison of the engine model calibration to the measurements indicate that HC is
modeled accurately across FTP and US06 type of driving. The model also indicates that the
calibration for all species is excellent for the US06 cycle, however CO and NOx overpredict
at lower loads (FTP type driving). It is possible that the catalyst model can compensate for
this, but it should be investigated further.
• A significant speed effect on emissions was isolated to the catalyst. It was discovered that
emissions from the "cruising" and "deceleration" modes deviated from the expected VSP
dependence. For CO (at medium and low VSP bins) this deviation approached nearly an
order of magnitude. This will require further investigation.
• There is evidence that modes of driving affect emissions, such that when a vehicle is driving
in "cruise" mode, the catalyst behavior is dependent on speed as well as VSP, and when it is
in "acceleration" mode, the emissions are more road load (or VSP) dependent.
• It was found that the activity represented by the FTP and US06 cycles was insufficient to
calibrate PERE. There appeared to be a gap in (cruise) driving in the 40-55 mph range (based
on on-road driving of 18 drivers), thus artificially calibrating the model to low and hi speed
transients. It is recommended to calibrate future models using a wider driving range.
• During calibration of PERE, it was found that calibrating to second by second (or VSP bin)
measurements gave different results than when calibrating to cumulative (total or bag)
emissions. It is important that a calibration to second by second or VSP bin also give
accurate total emissions. If not, this could hint that the model is not capturing all the physical
effects, or that the activity is insufficient in certain operating regions (as mentioned earlier).
• The fuel consumption and CO2 portions of PERE were validated to an independent driving
cycle (MEC01).
• When validating to a series of CARB unified cycles (each having a different average speed),
it was found that PERE was overpredicting CO2 by 5-10% at higher average speeds
(>30mph). It is possible that some of the powertrain efficiency coefficients will need to be
updated. It is not certain whether speed is the causal variable in this case. The same
validation study indicated that the two heaviest vehicles overpredicted CO2 by about 15%.
Both of these anomalies will require further investigation.
• The engine out validation on an independent cycle was very good for HC and NOx for total
emissions. HC slightly overpredicted in VSP bins 7-10, however an improved quantification
of the uncertainty in these bins may show that the model is "consistent" with the
61
-------
measurements. There were discrepancies at very low VSP bins (decelerations), however,
these bins don't have a large effect on total emissions.
• The engine out CO model underpredicted total emissions by 23% though the VSP behavior
was excellent across all bins. This is likely due to more enrichment events in the MEC cycle.
It is hoped that the catalyst model can compensate for this underprediction.
• Idle emissions should be separated as a different "mode" since the model failed to accurately
capture VSP bin 3 (idle) properly.
• The tailpipe validation was rudimentary since the model is incomplete. However, it shows
promise.
The Physical Emissions Rate Estimator has the potential to be a powerful addition to MOVES in
its ability to supplement the empirical model. It can be crucial in 'filling in the holes' in the
dataset, as well as providing some means of forecasting emissions from vehicles meeting future
emissions standards. However this approach is still a proof of concept using a limited vehicle
data set. It has yet to be demonstrated for vehicles certified to standards more stringent that Tier
1, which are more difficult to model. For example, due to the extremely low levels of emissions
from SULEV and Tier 2 vehicles, a new paradigm might be required. While this is still a work in
progress, the initial studies show promise.
V-ACKNOWLEDGMENTS
The author is grateful to Ford Motor Company for permission to take part in this unprecedented
and productive collaboration. I would also like to express my appreciation to members of the
EPA modeling team for all their assistance and support: John Koupal, Connie Hart, David
Brzezinski, Chad Bailey, Carl Fulper and Carl Scarboro. From the academic community, I would
like to acknowledge feedback and assistance from Marc Ross, Matthew Barth, Ted Younglove,
Carrie Malcolm, and George Scora. Last but not least, I would like to acknowledge my
supervisor, David Chock and manager, Ken Hass.
62
-------
VI - REFERENCES
An, F., and Ross, M., "The Use of Fuel by Spark Ignition Engines," SAE Technical Paper, No
930329, 19937.
An, F., Barth, M., Scora, G., and Ross, M., "Modeling enleanment emissions for light-duty
vehicles," Transportation Research Record 1641, 1998.
Barth, M., An, F., Younglove, T., Scora, G. Levine, C., Ross, M., Wenzel, T., "The development
of a comprehensive modal emissions model (preliminary draft version 2)," Final report to the
National Cooperative Highway Research Program, April 2000.
Barth, M., An, F., Younglove, T., Scora, G. Levine, C., Ross, M., Wenzel, T., "Comprehensive
Modal Emission Model (CMEM), version 2.0 User's Guide," 1999.
Fomunung, I., Washington, S., Guensler, R., Bachman, W., "Validation of the MEASURE
Automobile Emissions Mode: A Statistical Analysis," Journal of Transportation and Statistics,
September, 2000.
Frey, C.H., Unal, A., Chen, J., Li, S., Xuan, C., Methodology for Developing Modal Emission
Rates for EPA's Multi-scale Motor Vehicle & Equipment Emissions System, prepared for Office
of Transportation and Air Quality, US EPA, 2002
Gillespie, T., "Fundamentals of Vehicle Dynamics," Society of Automotive Engineers, 1992.
Goodwin, R., "A Model of Automobile Exhaust Emissions During High Power Driving Episodes
And Related Issues," PhD Thesis, University of Michigan Physics Dept, 1996.
Guensler, R., Washington, S., Bachman, W., "Overview of the MEASURE Modeling
Framework," Transportation and Air Quality III, Emerging Strategies and Working Solutions,
Simon Washington ED., American Society of Civil Engineers, 1998.
Hart, C., Koupal, C., Giannelli, R., "EPA's Onboard Emissions Analysis Shootout:
Overview and Results," EPA report EPA420-R-02-026, October 2002.
Heywood, J., "Internal combustion engine fundamentals," McGraw-Hill, New York, 1998.
Jimenez-Palacios, J., "Understanding and Quantifying Motor Vehicle Emissions with Vehicle
Specific Power and TILDAS Remote Sensing," PhD Thesis, Massachusetts Institute of
Technology, Cambridge, MA, 1999.
Koupal, J., Cumberworth, M., Michaels, H., Beardsley, M., Brzenzinski, D., "Draft Design and
Implementation Plan for EPA's Multi-Scale Motor Vehicle and Equipment Emissions System
(MOVES)." EPA report 420-P-02-006, 2002.
Koupal, J., "Air Conditioning Effects in MOBILE6." EPA report M6.ACE.001, 1998.
63
-------
Nam, E. "Understanding and modeling NOx emissions from light driving vehicles during hot
operation," PhD Thesis, University of Michigan, Ann Arbor, MI, 1999.
Nam, E., Ross, M., "A Fuel Rate Based Catalyst Pass Fraction Model for Predicting Tailpipe
NOx Emissions from a Composite Car," SAE Technical Paper, No 1999-01-0455, 1999.
Nam, E., "Understanding and Modeling NOx Emissions from Air Conditioned Automobiles,"
SAE Technical Paper, No 2000-01-0858, 2000.
National Research Council, "Modeling Mobile Source Emissions," National Academy Press,
Washington D.C., 2000.
Pischinger R., "Instantaneous Emission Data and Their Use in estimating Passenger Car
Emissions", MEET Deliverable 6, COST 319 Final Report A2, Technische Universitaet Graz,
1998
Ross, M., "Fuel Efficiency and the Physics of Automobiles," Contemporary Physics, 1997,
vol38, number 6, pages 381-394
Taylor, J. R. "An Introduction to Error Analysis," University Science Books, 1982
Thomas, M, and Ross, M., "Development of second-by-second fuel use and emissions model
based on an early 1990s composite car," SAE Technical Paper, No 97-1010, 1999.
Younglove, T., Levine, C., Barth, M., Scora, G., Norbeck, J., "Analysis of Catalyst Efficiency
Differences Observed in an In-Use Light Duty Vehicle Test Fleet," Proceedings of the 9th CRC
On-Road Vehicle Emissions Workshop, San Diego, California, 1999
64
-------
APPENDIX A - Off Road Engines and Future Technologies
The fuel consumption mode described in this paper is fairly robust across engine types. Many
engines will follow the same fuel trends with power. It is hoped that simply adjusting the
powertrain and fuel parameters will be sufficient to model the different technologies and fuel
types. As we will show below, this is probably an oversimplification of the issue.
FR/NV
bmep
Figure Al. The Willans Line for an internal combustion engine.
Figure Al shows an example of what is referred to in the literature as a "Willans Line." The y-
axis is fuel consumption per stroke per displacement: FR/NVd. The x-axis is brake mean
effective pressure, which is the power per engine stroke (cycle) and displacement unit:
bmep = Pt,nR/NVd (Al)
where
nR =2 for 4 stroke engines and nR = 1 for 2 stroke engines.
This relationship could be described by a simple linear model such that
Pf = Pfo + aPb, (A2)
Where
Pro is the offset or idle engine friction term
a is the slope of the line and defines the efficiency of fuel conversion into work.
Quite simply this is the same form as the fuel rate equation 6. We can make the following
substitutions or approximations (Ross and An, 1993):
Pin = KNVd/LHV (A3)
a = 1/riLHV
This equation can be applied to all internal combustion engines as well as fuel types. Future
technologies may increase some of the coefficients in these terms (namely K, N, and r|).
65
-------
Of course, for off-road engines that are not mobile, the brake power needs to be defined by
torque and engine speed only:
Pb = 2ttNi (A4)
For off-road engines, it will be necessary to obtain a calibration of torque or power vs engine
speed. Engines that deviate from stoichiometry (such as diesels), may also need a simplified
model for cp or air-to-fuel ratio. It is likely that in many cases, a simple adjustment of the FR and
powertrain parameters will be insufficient to adjust the emission rates. For example, in our two
stroke engine case: the powertrain parameters can be adjusted to change the efficiencies, friction,
etc, but none of these changes will reflect the fact that 2 stroke engines inherently have higher
hydrocarbon emissions rates due to the scavenging process. Separate emission rates would have
to be determined. For the first stages of MOVES, it is recommended to determine the base modal
rates empirically until at some future date, an improved model can be developed.
66
-------
APPENDIX B - Fuel Types
One of the advantages of the fuel approach in this paper, is the possibility of modeling different
fuel types explicitly. The physical parameters: LHV (Heywood, p. 915), conversion efficiency
(r|), and engine friction may vary from fuel to fuel. The hydrogen to carbon ratio in fuels will
also vary based on the fuel type, blend and cetane number. For typical gasoline, the ratio is
roughly 1.85. The H/C for various fuel types can also be found in Heywood (p. 915).
It is possible that for some fuels, a simple substitution of these parameters will be sufficient to
quantify the change in emissions rates. However, it may also be necessary to determine new
emissions coefficients (similar to different technologies). There may be factors, which are
difficult to theoretically model at this time, which will affect the emissions rates. These factors
include, nitrogen content of fuel, oxygenates, other additives, etc. Modal toxic emissions will be
especially sensitive to fuel types.
In conclusion, this simple method of quantifying fuel consumption in internal combustion
engines may be sufficient to quantify the change in emissions in comparison to "standard"
gasoline, but there may also be many exceptions where separate emission rates (or corrections)
will be required. For future engine technologies, this technique will need to be reexamined on a
case by case basis. Efficiency improving mechanisms such as variable valve, cam, cylinder
number, and compression ratio may all affect the Willans line, but their affect on the emission
rates is unknown without more data.
67
-------
APPENDIX C - CMEM Parameters: Simplification and Sensitivity
This section describes the CMEM powertrain, engine-out and catalyst parameters, and lists
which may be omitted due to various reasons. A full statistical analysis was not accomplished in
this very brief study, and may perhaps be warranted in a future study.
Powertrain Parameters
The following are fixed parameters in the CMEM second by second powertrain model. They
only vary according to vehicle technology classes. The parameters obtained from readily
available databases are marked "known", the rest are estimated based on publications etc. The
parameters, which have little affect on emissions (as long as reasonable values are chosen), are
marked with an "x". Some parameters are only sensitive when enrichment is concerned. Those
are specifically marked.
Table CI. Powertrain Parameters
Parameter Known
M (mass) Yes
Vd (Engine Displ) Yes
Trlhp (dyno) Yes
S (N/v) Yes
Nm (rpmpeak) Yes
Qm (torquepeak) Yes
Pmax (peak power) Yes
Np (rmax) Yes
Nidle
Ng (Ngears) Yes
KO (friction)
el (efficiency)
82
Insensitive
x
Enrichment
Enrichment
Enrichment
x
x
X
X
X
X
Engine-Out Emission Parameters
The following tables lists fitted parameters in the CMEM engine-out second-by-second model.
The "Set Constant" column indicates which of the parameters could be fixed, based on previous
studies, which have demonstrated that these parameters do not seem to vary significantly from
model to model and year to year. However, they may vary from LDV to LDT, or very different
technology classes.
Table C2. Engine-Out Parameters
Parameter Set Constant
Co x
aCO
aHC
rHC
alNO
a2NO
FRNOl
Insensitive
68
-------
FRN02
hcmax x (enleanment)
hctrans x (enleanment)
tr x x (enleanment)
cpmin x x (enleanment)
5SPth x (enleanment)
r02 x (enleanment)
The enleanment portion of the model can be omitted since the catalyst takes care of these events
in most cases. In certain types of hi-emitters, this may have more of a significant effect, but it is
probably not worth the extra complication of the model to include this for only a very small
portion of the fleet. In this approach, only (the first) ~8 coefficients are necessary to fit.
Catalyst Parameters
The catalyst model is the weakest link in this or any emission model. An aggregate model such
as VSP subsumes a catalyst model since it is a modal tailpipe model. In this sense it is a
aggregate or "lumped parameter" model. It is necessary to greatly simplify a catalyst model in
order to be of use in MOVES. The "Soak" and "Cold Start" parameters are specified below.
These two regimes are more simply modeled in an aggregate fashion as we shall describe below.
Table C3. Catalyst Parameters
Parameter Set Constant Insensitive
CsoakCO Soak
CsoakHC Soak
CsoakNO Soak
pcatCO Soak
pcatHC Soak
pcatNO Soak
pCO Cold Start
pHC Cold Start
pNO Cold Start
Tel Cold Start
cpcold Cold Start
CSHC Cold Start
CSNO Cold Start
rco
rHC
rNO
bCO
cCO
bHC
cHC
bNO
cNO
y (tip in) x
(po (max) Enrichment
69
-------
Pscale (threshold)
Enrichment
The primary catalyst fitted coefficients are the b, c and T parameters. On a second-by-second
(sbs) correlation (as opposed to bag), these fits can be very poor for specific vehicles (even if the
validation is conducted on the same vehicles used to calibrate). We can conclude from this that
these parameters are not capturing the complexities of catalyst action. It is likely that it is
impossible to capture sbs behavior of a catalyst with only 9 fitted coefficients, or even 90!
Catalyst (and air fuel ratio) behavior is very stochastic in nature. For that reason, it is just as
reasonable to use a simpler lumped parameter model, this is outlined in the paper. Moreover,
since emissions are so sensitive to catalyst behavior, it is usually easier to model hi-emitters.
70
-------
APPENDIX D - Second by Second Tailpipe Emissions
We have seen that, to first order (for a given vehicle), engine out emissions are correlated with
fuel rate or quantities proportional to it: VSP, air flow, throttle, RPM*load, etc. In fact, the
correlations have been quite good with fuel rate in particular. Therefore, any analysis technique
that determines the significance of a causal variable (such as Hierarchical Tree Based
Regression, neural network, or design of experiment) should result in these "load type" variables
being the most important. Frey et al., (2002) also proved that VSP gives the first split for tailpipe
emissions using HTBR. However, the same correlation with load cannot be made with catalyst
behavior. The scatter in the second by second tailpipe emissions is most likely due to catalyst
effects. It is generally accepted by modelers that engine out emissions are relatively easy to
model, in comparison to catalyst pass fractions. The latter tends to be very sensitive to
parameters, which are very difficult for modelers to obtain, such as an accurate measure of air to
fuel ratio, oxygen storage rates, temperature of the catalyst, flow, precious metal loading,
honeycomb geometry, etc.
One of the goals of the present study is to determine if engine out and tailpipe should be
separated. Some of the examination above supports this need. However, it is interesting to
examine tailpipe trends directly for comparison. The following sections demonstrate the tailpipe
trends as a function of fuel rate and VSP. This discussion addresses decision node 4 on the flow
chart.
Tailpipe CO
The following figures show the trend of tailpipe CO as a function of FR and VSP. We have
already seen that some of the scatter at the higher CO values is due to enrichment events. The
correlation coefficients are:
Ptco/fr = 0.48, Ptco/vsp = 0.31
71
-------
Tailpipe CO in NCHRP Tier 1 Cars
6.0
5.0
_ 4.0
O
O
a>
a.
're
3.0
2.0
1.0
~ TCO
~
~
~ ~
jk_
-
~ $V«
$
~
Ti:i^ * * /• 1 ~ :~*
«iW» t« *
1.0
2.0
3.0
4.0
5.0
6.0
Fuel Rate (g/s)
Figure Dl. Tailpipe CO as a function of FR.
Tailpipe CO in NCHRP Tier 1 Cars
O
o
ai
| ~ TCO
~
~
4.0-
• ~
~
<
~ ~
~
~
~
~
2.0
~ 1
~
~
~
~ . «
# ~
%
~~
~ ~
~
# ~
• /V
~ ~
~
~
» ~
z* * ~ " • <
w
~ •
1 1 1
-10.0
0.0
10.0
20.0
VSP (kW/Tonne)
30.0
40.0
50.0
Figure D2. Tailpipe CO as a function of VSP.
Figure D3 shows the same except binned by VSP. The bin sizes are determined from Frey et al
(2002). Note that the bin sizes are not identical. The 95% confidence intervals are measures of
the confidence of the mean values within the bins, i.e. the mean is correct to within 95%. The
emissions are plotted logarithmically due to the skewness of the distribution. Figure D4 shows
the same data but with the error bars plotted as the standard deviation. This gives a measure of
the variability of the data (not the uncertainty of the mean). Figure D5 shows that much of the
variability originates from the two different driving styles. This is somewhat problematic, and
more will be discussed about this in the CPF section. It is not shown, but the fuel rate and the
CO2 do not demonstrate this (significant) split by driving cycle.
72
-------
N= 10847 5400 12026 10189 7473 5297 4144 3651 2859 2890 1735 802 695 363
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure D3. Tailpipe CO binned by VSP on a log scale.
1
CO
0)
^ -5
N = 10847 5400 12026 10189 7473 5297 4144 3651 2859 2890 1735 802 695 363
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure D4. Tailpipe CO plotted logarithmically as a function of VSP bin. Bars are 1 standard
deviation.
73
-------
0
o
o
i—
3 -4
O
10 c
o> -5
I
m
in
m
ffl
9 x
m m m m | i ~
n
. m
W
&
m
TRIP
I
~ FTP
I
~ US06
69094470980 02WS4S8D165H2338197883292 028276^272355949 6861801 695 363
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure D5. Tailpipe CO binned by VSP and separated by driving cycle.
Tailpipe HC
The following figures show the trend of tailpipe HC as a function of FR and VSP. The
correlation coefficients are:
Pthc/fr - 0.33,
Pthc/vsp ~ 0.23
The correlation is significantly worse than those obtained from engine out values.
Tailpipe HC in NCHRP Tier 1 Cars
0.040
0.035
0.030
¦2 0.025
5
£
V 0.020
a.
'a.
™ 0.015
0.010
0.005
0.000
~ THC
~
~
~
~ ~
~
~
~
~
~
~~
~
~
~
~
~
~
:
~
*
~ ~
~
~
~
~ ~
• ~
~* J
9 *~V* ~ "
l € %
~~ i
~~ ~
~~~~
~
~ *4» ~ \
«~
~
~
+ * ~ *
1 1 1 1
~
1 1 1 1
0.0
1.0
2.0
3.0 4.0 5.0
Fuel Rate (g/s)
6.0
7.0
Figure D6. Tailpipe HC as a function of fuel rate.
74
-------
Tailpipe HC in NCHRP Tier 1 Cars
0.040
0.035
0.030
0.025 --
0.020
0.015
•
0.010«
~ THC
• ~
H—
~ ~ ~ ~ ~
0-010*: ~ ~ * /~ ~ v. ~ ~ ~ ~ ~
~ • ~ */•~~~#*.~ ~ ~ •
* 0.0^5 ~
# «»<> i
-10.0
0.0
10.0
20.0
30.0
40.0
50.0
VSP (kW/Tonne)
Figure D7. Tailpipe HC as a function of VSP.
Figure D8 shows the same data binned by VSP. The variability and the discrepancy between the
cycles match that of CO (above).
-2.0 ¦
-2.5 ¦
-3.0 ¦
-3.5 ¦
O -4.0-
X
I—
o
z! -4.5 ¦
o
0s
c n
O) "5.0 ,
N
Figure D8.
Tailpipe NOx
The following figures show the trend of tailpipe NOx as a function of FR and VSP. The
correlation coefficients are:
= 12266 6138 13838 11223 8064 5599 4335 3784 2931 2950 1759 811 701 364
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Tailpipe HC binned by VSP.
75
-------
Ptnox/fr - 0.40
Ptnox/vsp = 0.35
The correlations have consistently been better for FR over VSP for all the pollutants. This
demonstrates again, that FR is a better independent variable for emissions characterization.
Tailpipe NOx in NCHRP Tier 1 Cars
~ TNOx
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0
Fuel Rate (g/s)
Figure D9. Tailpipe NOx as a function of FR.
Tailpipe NOx in NCHRP Tier 1 Cars
—fWJ9—
~ TNOx
0.08 -
-
~
~
;
~
~
0.06 -
0.05 -
0.04 -
;
~
—* ~-
~
~
~
;
~
~
;
~
~ ~ ~
~ t
~
. ~ ~
~ ~ «
~~ ~
~~ % ~
~ ~
* •»
~
ojfc -
*
~
t ~~
~ ~
~~
~
~
< W
~ ~
. *~
~
Li3«r» ~£ *
-10.0 0.0 10.0 20.0 30.0 40.0 50.0
VSP (kW/Tonne)
Figure D10. Tailpipe NOx as a function of VSP.
Figure D11 shows the same data binned by VSP. The variability and discrepancy between the
cycles is similar to that of CO (above).
76
-------
-2.0
-2.5
-3.0
-3.5
O "40
Z
I—
0
z! -4.5
o
oN
10 c n
CO -5.0
N = 11378 5494 11842 10301 7551 5333 4230 3717 2884 2916 1746 809 700 364
1 2 3 4 5 6 7 8 9 10 11 12 13 14
VSPBIN
Figure D11. Tailpipe NOx binned by VSP.
In conclusion, we have demonstrated that the correlations for the tailpipe emissions are
significantly worse than those of engine out. This is to be expected, but if separate further
physical analysis of engine and catalyst out trends can discover the cause of the tailpipe scatter,
the predictive capability of the model can be significantly improved. Furthermore, light trucks
and cars will meet the same emissions standards, it is likely that this will confound the tailpipe
trends presented in this paper. It is thus not recommended to model tailpipe directly using a
physical model. The separate modeling of engine out and catalyst would allow for the explicit
modeling of phenomena such as hi emitters, degradation, new technologies, etc, thus increasing
the versatility and power of the model. The added complexity would not compromise the
accuracy of the model at all, but rather can only increase it.
77
-------
APPENDIX E - Sample Error Propagation Calculations
The uncertainty in equation 13 is as follows (assume 8Wt = 8Vd = 0):
8FRth =
A* Wt*Vd*
A
+ 8B2
(El)
This turns out to be larger than the standard error of the measured FRth, so we should take the
more conservative estimate of error.
The uncertainty in equation 11 is simply:
8ECO = 5alco *FR
(E2)
This assumes that the fractional uncertainty in FR is small compared to that of a. Whether or not
this is strictly true, they are both small compared to the next term. Under enrichment, this
becomes:
81 TO
1
(8alcoFR)2 +a2coFRtl
8a
2 CO
y a2 CO y
+
SFR,
FR,
+ alcoFRth
8a
1 CO
v a\CO y
+
SFR,
FR.,
(E3)
The last additive term (in the square root) is very small compared to the other 2 terms and can be
ignored if simplification is desired. We propose that relative uncertainties can be ignored if they
are an order of magnitude smaller than the relative uncertainties of the other (larger) terms. Only
20% of the enrichment events (modeled) occur in VSP bins 11 or less, so we will use the latter
term only in bins 12-14. If there are more low power to weight ratio vehicles in a different study,
bin 11 will likely have a larger uncertainty, since they will tend to go into enrichment at a lower
VSP value. For the purpose of simplification, the FRth is taken to be identical to the FR at the
higher VSP bins.
The uncertainty in EHC is:
8EHC = yl(8aHCFR)2 + (8rHC )2
(E4)
The uncertainty in ENOx is simply:
8ENOx(FR) = ^{8amo*FR)2+(8a2NO*FR2)2
(E5)
Under enrichment, it is:
78
-------
dENOx =
1
(5a3NOFR)2 + a3NOFRt,
§a
3NO
\2 / _ \2
v a3 NO j
+
FRtl
\
+ {8EN0x(FRth)f
Once again, since the fractional uncertainty in a3 is nearly an order of magnitude larger than
of ai or a2, the last additive term can be ignored.
79
------- |