United States Air and Radiation EPA420-R-01-042
Environmental Protection July 2001
Agency M6.SPD.001
&EPA Development of
Speed Correction Cycles
Printed on Recycled
Paper
-------
EPA420-R-01-042
July 2001
Development of Speed Correction Cycles
M6.SPD.001
Original Report by Sierra Research
April 30, 1997
Response to Comments, Corrections and Clarifications
Megan Beardsley
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.
-------
Development of Speed Correction Cycles
Response to Comments, Corrections and Clarifications
The Sierra Research report "Development of Speed Correction Cycles," SR97-04-01,
April 30, 1997, was posted on the MOBILE6 web site as M6.SPD.001 in June 1997 for
stakeholder comment. It was also sent for independent peer review.
While EPA used the original report to design driving cycles that were used to collect data
for developing speed and roadway type corrections used in MOBILE6, review of the comments
on the original report suggested the need for a number of corrections and clarifications to the
report. Rather than revising the body of the completed contractor report, EPA has chosen to
append a list of comments and corrections to the report.
We also have appended a complete list of comments received on the report and our
response to those comments. These are divided into three sections: stakeholder comments, peer
review comments from Dr. John Warner, and peer review comments from Professor
H.Christopher Frey.
EPA Corrections and Clarifications
1. The title and body of the report refers to "speed correction cycles." While the results of
testing on these cycles will be used to adjust emissions based on estimated average
driving speed, these corrections are also meant to be used to correct for differences in
driving on different roadway types ("facilities") under different conditions. Thus, it
would be more precise to replace the term "speed correction cycles" with a term such as
"facility- and speed-specific driving cycles."
2. On page 2 and again on pages 25-27, the report describes the development of an area-
wide cycle that was constructed from several freeway segments and a non-freeway cycle.
The report does not adequately explain the purpose of the area-wide cycles. These cycles
were developed for eventual comparison with MOBILE6 area-wide results and for
comparison with other area-wide speed cycles. The area-wide cycles have not been used
for testing vehicle emissions. While the facility-and-speed specific cycles discussed in
Chapter 3 were developed from the chase car data, the area-wide cycles were developed
from the instrumented vehicles. Since the instrumented vehicle data did not distinguish
arterial and local driving, a single "non-freeway" sub-cycle was developed.
3. On page 4, the report states that, in MOBILE5, speed correction factors are used to adjust
the emissions measured using the Federal Test Procedure (FTP). More specifically, the
1
-------
M0BILE5 speed correction factors are applied to the basic emission rate, a user-specified
weighting of the three "bags" of the FTP.
4. The reference to "bags" on page 6 may be confusing. The city-specific area-wide
emissions are simply a weighted average combination of various cycles as described on
page 26 of the report.
5 On page 10, the report says that similar speed/acceleration frequency distributions for
LOS A-C indicated that this range of LOS could be represented by one cycle. An
additional reason for grouping these LOS together is that the chase car data indicates that
these LOS are characterized by similar power frequency distributions. Since power-
frequency distributions are highly correlated with emission differences, their
consideration was an important factor in the decision to combine these LOS into a single
cycle. Similarly, power-frequency distributions were considered in the decision to
develop separate cycles for LOS D, E and F and to create a new "LOS G".
6. On page 12 there is a discussion of the "DiffSum" statistics used to compare driving
cycles to the target driving population. "DiffSum" is calculated as the sum of the
absolute values of the differences in the frequency distribution for each
speed/acceleration bin.
7. On page 12, the report mentions that one criterion for evaluating cycles was the amount
of operation in high specific power modes. The two modes (200-299 mph2/sec and >300
mph2/sec) were selected because EPA research conducted during the development of the
Supplemental Federal Test Procedure (SFTP) indicates that in these ranges, some
vehicles are designed for "commanded enrichment" which can significantly increase
emissions. We felt it was important that these modes not be under- or over-represented in
the new cycles.
8. In the report, "segment" has different meanings. In Chapter 3, "segment" refers to
continuous driving data on the same roadway type and congestion level by a particular
vehicle, as defined on page 11. In Chapter 4, "segment" refers to a driving cycle, in
particular, a sub-cycle of the area-wide driving cycle. In Chapter 5, "segment" is used in
both ways—the meaning should be inferred from the context.
9. The footnote on page 25 should read: "Large differences occur at specific congestion
levels..."
10. The report does not make clear the units for the equations on page 26. Composite
Emissions and Emissions are in g/mile. Weighting factors are a unit-less fraction. The
speed/acceleration frequency distribution (SAFD) and travel fraction (TF) are both unit-
less fractions. Note that, while the description of the equation for composite emissions
says the weighting factor represents the fraction of travel "in miles," for convenience the
2
-------
sample area-wide cycles described in this report actually use time-based weighting factors
derived from instrumented and chase car data as described on page B-2. Similarly, time-
based weighting factors were used for the Los Angeles area-wide weighted cycle
described on page 33. If area-wide cycles were to be constructed for another city,
distance-based weighting factors from local transportation model outputs would probably
be used.
11. Page 26 mentions that travel mix was adjusted to account for "short trip bias." This
adjustment is detailed in Appendix B.
12. The equation for the non-freeway speed/acceleration frequency distribution (SAFD) as
printed near the bottom of page 26 is incorrect. The correct equation (and the equation
actually used in this analysis for the calculation of the non-freeway SAFD) is:
SAFDn/-TFxSAFDP rr
SAFD„ „ = ^
Nonrwy ^
13. Page 28 incorrectly describes Figures 13 and 14, although the figures themselves are
labeled correctly. Figure 13 is a three-dimensional SAFD (Watson Plot) of the
Baltimore instrumented vehicle data. Figure 14 is a Watson Plot of the new area-wide
cycle with weighting factors for Baltimore applied.
14. The information in Tables 7 and 8 on pages 32 and 33 comes from several different
sources. The "population" values for facility-specific driving cycles in Table 7 are from
3-city (Baltimore, Los Angeles, Spokane) chase car data, while the population values for
the LA-92 are from the Los Angeles chase car data only. In Table 8, the Non-Freeway
Area-Wide segment is compared to the Baltimore weighted instrumented-vehicle data
and the LA-92 is compared to the Los Angeles chase car data.
15. Page B-3 of Appendix B describes the calculation of a "bump-up" factor to ensure that
travel fractions adequately represent short trips.. Because Los Angeles and Atlanta
lacked instrumented and chase car data, respectively, the bump-up factors for these cities
were calculated using data from Baltimore to replace the missing data. For example, it
was assumed that the fraction of short trips in Los Angeles was equal to that observed in
the Baltimore instrumented data. The Los Angeles bump-up factor was then calculated
by dividing this assumed instrumented-vehicle short trip fraction by the actual (but under-
represented) short trip fraction observed in the Los Angeles chase car data.
Note, the Atlanta bump-up factor was calculated and listed in Appendix B for
completeness but was never used.
3
-------
Stakeholder Comments on M6.SPD.001 and EPA Response
Comments received from North Carolina Department of Transportation (NC DOT) (Comment
#25), New York City Department of Environmental Protection (NYCDEP)(Comment #27), and
American Automobile Manufacturers Association (AAMA ) (Comment #37).
The following section reproduces stakeholder comments in plain text and intersperses
EPA's response in indented italic. Commenters are identified by acronym at the end of
each comment.
Facility specific drive cycles are perhaps the most significant of the proposed improvements to
the MOBILE Model. North Carolina approves of the move in this direction. The current model
uses an average drive cycle to represent all possible driving conditions. This leads to
counterintuitive results in some cases. However, we have some concerns based on the amount of
aggregation and disaggregation in the supporting materials. Is the variability of stop/delay time
implicit in the drive cycle that will be used to develop the basic emissions rates for each facility
type? Our experience is that stop/delay time varies across facility types. We believe that any
future version of the MOBILE model should account for this variation. An alternative method
would be to allow the user to specify stop/delay time for each facility type. (NC DOT)
We agree that stop/delay time is important. The frequency of stops and delays were
included in the speed/acceleration frequency distributions (SAFDs) that were used to
develop the testing cycles, which were then used to develop MOBILE's facilty-specific
speed correction factors. While users can not explicitly specify stop and delay time, they
can specify differing fractions of VMT on different roadway types at different speeds-this
effectively varies the stop/delay time included in the calculated emissions.
We also note that arterials and collectors will share a driving trace. As noted above our
experience indicates the existence of significant differences in stop/delay time and start mode
between facility types. Collectors resemble locals streets more than arterial streets. (NC DOT)
Due to sample size and testing cost concerns, we were not able to develop cycles for all
roadway classes at all speeds and LOS conditions. Instead it was necessary to group
some roadway types together. Because chase car data (and hence, speed and LOS
information) was not available for local roadways, grouping collectors with locals would
mean that we could no longer distinguish collector driving by speed and LOS. For this
reason we decided to group collectors and arterials. While there may be some differences
between travel on these two roadway types, we believe that the ability to specify the
average speed is sufficient for the most important distinctions between the two.
Also, note that "start mode " is handled separately in the model and is not directly linked
to roadway type. Varying start fractions by roadway type could be done but would
require multiple runs of the model.
4
-------
The proposed freeway drive cycles also provided some surprises. The proposed drive cycles
include: High Speed, LOS A-C, LOS D, LOS F, and LOS G. We recommend that the High
speed drive cycle and the LOS A drive cycle be combined, and that the drive cycle for LOS B-C
be kept together. Our understanding of the Highway Capacity Manual indicates that high speed
driving occurs under LOS A. We also propose that LOS F and LOS G be combined. To the best
of our knowledge, the Highway Capacity Manual does not recognize a LOS G. From the
associated driving trace, this drive cycle represents breakdown conditions and might best be
consolidated into the drive cycle for LOS F. (NC DOT)
It is true that our driving cycles do not match the Highway Capacity Manual. Our high
speed cycle is a subset of driving under LOS A, B and C conditions, while our "LOS G "
is a subset of driving under LOS F. These additional cycles were created because we
needed to test vehicles at the low and high-speed extremes of highway driving. Tests at
these points help determine the speed correction factors outside the range of more typical
driving.
In particular, the development of the high speed and LOS G driving cycles and the
grouping of the A, B and C driving cycles was partially based on the power frequency
distributions for these cycles. LOS A freeway driving has a power frequency distribution
similar to LOS B and C, while LOS G is signficantly different than LOS F and is useful
for modeling the most congested conditions.
Have the new facility-specific cycles been reviewed by DOT/FHWA personnel? We are
especially concerned about in-City roadways (arterials/collectors and local) where there are speed
limits that may only allow 30-35mph. The maximum and average speeds for the bottom 3 cycles
on Table 1, for congested in-City arterials and local roads, may be too high for many congested
New York City streets during peak hours. New York City is likely to have a traffic control sign
and signal density which is at the extreme end of the range in the nation. Frequency of starts and
stops, and therefore of acceleration/deceleration, will not only affect average speeds but also the
emissions associated with a given speed. We support any efforts by EPA to develop operating
mode data that would allow us to project the impact on emissions of a high density of traffic
signals on local streets as well as on arterials. (NYCDEP)
FHWA personnel have been involved in the discussions for how speed will be handled in
MOBILE6 and have not expressed concerns about these cycles.
As will be described in the speed correction factor report M6.SPD. 002, we have
continued to test vehicles on the New York City Cycle and will use this data to project
emissions at the low speeds that characterize New York City at peak hours.
Will speed corrections for arterials/collectors also utilize data from the NYCC and FTP cycles?
Will the speeds on local streets be adjusted? If yes, what cycles will be used other than the
5
-------
NYCC? How will idle CO emissions be calculated? Will they be calculated from the 2.5 mph
emissions estimates? If yes, will the 2.5 mph emissions for local streets be the same as the 2.5
mph emissions for arterial/collectors? If they are not the same, how will they be calculated?
How different can we expect the low speed correction factors to be in the new model compared
to those used in MOBILE 5? Is there any reason why idle emissions data is not directly collected
to use in the model instead of adjusting the 2.5 mph emissions? (NYCDEP)
These questions are to be addressed inM6.SPD.002, "Facility-Specific Speed Correction
Factors."
This approach is somewhat different than the ARB's approach, in which a single, self-weighted
inventory cycle was developed (the LA92, or Unified Cycle (UC)) and a significant number of
cars and trucks were tested on this cycle. Also, ARB developed Unified Correction Cycles for
developing speed correction factors for the UC. AAMA is unsure if EPA can devote enough
resources to make their approach more accurate than ARB. Concern stems from EPA's desire to
make one model fit all modeling purposes. The Unified Cycle approach is certainly more simple,
and has the advantage that only a single, self-weighted cycle needs to be run for area-wide
modeling. EPA's approach requires significantly more testing per vehicle, consequently, fewer
vehicles can be tested. There is also an issue with respect to whether vehicles can be maintained
at proper temperatures throughout the duration of the EPA cycle testing. AAMA recommends
that EPA also have all of the vehicles tested on ARB's Unified Cycle as well as the other cycles,
so the Unified Cycle can be compared to a weighted average of EPA's cycles. AAMA will
reserve further comments on both ARB's approach and EPA's approach until it evaluates the data
from EPA's test program, and particularly how EPA compares the data on the Unified Cycle to
the data from the EPA's test cycles. (AAMA)
EPA has tested all vehicles tested using the EPA cycles on the LA92 (California Unified)
cycle as well. A comparison of the LA92 results and the EPA cycle testing has been deferred
until a final methodology for use of the EPA cycle testing is determined.
Significant steps were taken as part of the EPA cycle testing to minimize the effects of
duration on emission results. All of the EPA cycles are much shorter in duration than the
LA92 cycle. The longest cycle is 73 7 seconds (slightly over 12 minutes) long. No more than
four cycles were run in series at one time and the order in which the cycles were done was
randomizedfor each vehicle tested. Coolant temperature was monitered during testing.
EPA decided to pursue a facility based approach to evaluate the effects of driving on
emissions when it became clear that an internally consistent method was necessary to allow
the calculation of both area-wide inventories and the calculation of conformity scale
emission analysis. EPA concluded that the incremental approach proposed by California
was no more likely to succeed in satisfying both parts of the emissions estimate challenge
(area-wide and conformity) than the approach taken by EPA. If properly conducted, both
6
-------
methods should agree. Further, EPA believes that their approach will provide valuable
insights into the needfor and design of future testing programs and models. Future testing
may require far fewer cycles than were tested for the current study, allowing for larger
sample sizes.
If EPA's approach of many factor-specific correction cycles remains unchanged for M0BILE6, it is
essential that the model contain default (nationwide) statistics to develop average emission rates for
a nationwide inventory. (AAMA)
National default estimates ofVMT by speed and facility type will be available inMOBILE6.
The development of these default values is described in M6.SPD.003, "Development of
Methodology for Estimating VMT Weighting by Facility Type. "
EPA must also allow users to output emissions based solely on current FTP certification test results,
for ready comparison with the current and historical emission standards. (AAMA)
We considered an "FTP output" option when developing MOBILE6, but it was quite
complicated to implement. We will keep this feature in mind for future versions of the
model.
7
-------
EPA Response to Peer Review Comments from John Warner,
Center for Statistical Consultation and Research, University of Michigan, April 1,1998
The following document reproduces Dr. Warner's comments in plain text and
intersperses EPA 's response in indented italic. Note, this is a scanned version of Dr.
Warner's comments. Not all mathematical symbols are correct.
This document contains a review of the report Carlson and Austin (1997) which was
prepared by Sierra Research, Inc. for the United States Environmental Protection Agency. This
review is divided into four sections. In Section 1,1 give a brief synopsis of the report. In Section
2,1 list eight specific recommendations for improving the statistical methodology of the report.
In Section 3,1 give detailed explanations of each of the eight recommendations. In section 4,1
list a few miscellaneous (that is, not specifically methodological) criticisms of the report. The
review closes with the list of the technical reports on speed correction cycles that I have
consulted in the process of conducting this review and a list of other references cited in the text.
In general, my conclusions are as follows: It appears that Carlson and Austin (1997) have
done a good job in collecting and tabulating data from a variety of sources. The basic model
(that is, the method for estimating area-wide emissions) is plausible and interesting. I am not
enthusiastic, however, about the proposed method for extracting speed correction cycles from
samples of speed traces. It seems likely to me that better methods can be found, particularly if a
more specific criterion is given forjudging the quality of a speed correction cycle. Finally, I feel
very strongly that the statistical analysis of the speed trace data could be improved substantially.
Indeed, Carlson and Austin (1997) present no inferential statistics of any kind, that is no
p-values, no confidence interval, and no standard errors. I recommend that EPA obtain a new
analysis of the speed trace data and I make a number of specific recommendations regarding the
methods which should be employed in this analysis. In particular, the group or individual who
carries out the new analysis should be familiar with random effects modeling and density
estimation. The issues raised in Subsection 3.1 should be settled by EPA before this new
analysis is attempted.
1 Synopsis of the report
Carlson and Austin (1997) present a new set of speed correction cycles (SCCs) for automobile
emissions. SCCs are selected speed traces (or plots of speed against time). When automobiles
are mounted on a dynamometer and run through a SCC, emissions are supposed to mimic the
emissions of the population of comparable automobiles traveling on a road of a given type
(facility type) under a given intensity of traffic congestion, or level of service (LOS). The model
proposed by Carlson and Austin (1997) supposes that the aggregate driving conditions in all
major US cities can be approximated (for the purpose of estimating auto emissions) by taking
8
-------
city specific weighted averages of SCCs from the proposed set of SCCs. If this method works as
intended, one could accurately estimate emissions in all major US cities by 1) estimating the city
specific proportion of travel on each facility at each LOS and 2) measuring the emissions on a
representative sample of cars on all cycles in the set. Urban planners could then evaluate the
effect of changing the mix of models in each city's passenger car fleet, or of changing the pattern
of facility use, without undertaking any new emissions testing.
Carlson and Austin (1997) make use of three types of data to fit their model: speed traces
sampled from chase cars (CCS), speed traces sampled from instrumented vehicles (IVs), and
weighting factors obtained from network based transportation models. (A fourth type of data,
laboratory emissions data from selected vehicles running along a specific SCC, would be used in
future stages of model development). In the above, weighting factors are estimates of the
proportion of all traffic in a given geographical location that travels on facilities of a given type at
a given LOS. A population of speed traces is the set of all speed traces which occur for a given
set of car/driver combinations occurring in a given geographical location under given restriction
on facility type and LOS. (For example, the population of Ford cars driven by middle aged
women on busy freeways in New York City). Chase cars can be used to sample from a
population of speed traces as follows: First, a set of characteristic routes and starting times are
chosen to represent the full range of driving patterns that obtain in a given city. Second, chase
cars, outfitted with radar, are sent out to travel these preassigned routes at the appropriate times
of day. While in transit, the chase cars are instructed to record speed traces from randomly
selected vehicles. Instrumented vehicles (IVs) are randomly selected vehicles that are equipped
with instrumentation for measuring and recording speed on a continuous basis. IVs produce
samples of speed traces when they are driven by their owner's for a fixed period of time usually
eight days (Defries and Kishan, 1992, p 3-21).
Carlson and Austin's (1997) report consists of 1) an executive summary and introduction,
2) a rough description of the methods by which SCCs are extracted from a sample of speed
traces, 3) plots of the SCCs, 4) a description of the method for combining data of the four types
describe above to estimate emissions in a given city, 5) a description of methods for evaluating
the fit between a SCC and a corresponding sample of speed traces. A series of appendices is also
provided. These contain tables of summary statistics, tabulations of empirical speed acceleration
frequency densities (SAFDs) for various samples of speed traces and SCCs, and details
concerning the calculation of area-wide emissions estimates. All this material is reviewed here.
Carlson and Austin (1997) also presents a section on intersection analysis, which will not be
reviewed because it is only tangentially related to the issues that I have been asked to consider.
2 Specific methodological recommendations
In this section I outline eight specific recommendations for improving the statistical methodology
in Carlson and Austin (1997). Each specific recommendation will be discussed in a separate
subsection below.
1. More care should be taken to define a specific set of summary statistics that can be used
to determine when a sample of speed traces is a good representation for an underlying
9
-------
population of speed traces. The same set of summary statistics should be used to
determine when a SCC is a good representation of a sample of speed traces. Carlson and
Austin (1997) use the SAFD for this purpose but Cohen, J. et al (1993) indicate that the
SAFD is not adequate. This issue needs to be settled before the underlying statistical
issues can be addressed in detail.
The U.S. EPA initially requested more detailed summary statistics from the
contractor. However, given the time and resource constraints on this work
assignment, we agreed to accept a more limited statistical description and
comparison.
2. Standard errors should be presented for the sample summary statistics discussed above.
T-tests and ANOVAs (and their nonparametric analogues) should be used to determine if
summary measures differ between populations. The unit of analysis for evaluating the
summary statistics should be the estimated speed trace of an individual car/driver
combination.
As mentioned above, we requested more detailed summary statistics from the
contractor. However, given the time and resource constraints on this work
assignment, we agreed to accept a more limited statistical description and
comparison.
3. A smoothing technique, such as kernel density estimation, should be used to estimate
car/driver specific SAFD's.
We will consider using such techniques in the future when developing new cycles.
4. Alternate methods should be considered for extracting a SCC from a sample of speed
traces. I suggest (tentatively) a method for obtaining a SCC as a solution of a
mathematical programming problem.
We will consider using such techniques in the future when developing new cycles.
5. Claims made in the text of the report should be supported with summary tables including
standard errors and p-values whenever these are available.
As mentioned above, we requested more detailed summary statistics from the
contractor. However, given the time and resource constraints on this work
assignment, we agreed to accept a more limited statistical description and
comparison.
6. Standard errors for the weighting factors should be obtained, if possible, from the
network based transportation models.
10
-------
We will consider this approach in the future when developing new cycles.
7. The equation defining SAFDNonFwy (near the bottom of page 26) is incorrect and needs to
be modified.
The equation on page 26 is listed incorrectly. We will make a note of this on the
cover page accompanying the document. However, the actual computations used
the correct equation as is clear from examining the non-Freeway SAFD which
does sum to one.
8. Random effects modeling of car/driver specific SAFDs (or other summary measures)
should be considered.
We will consider this approach in the future when developing new cycles
3 Discussion of the methodological recommendations
This section provides a detailed discussion of the recommendations listed in the previous section.
3.1 Criteria for extracting speed correction cycles
An SCC is a short selected speed trace which is derived from, and intended to represent, an entire
sample of (usually much longer) speed traces. More specifically, the average of any specific type
of emission (per mile) for any car, traveling on a straight line with constant grade along a SCC, is
supposed to be close to the average of the same emission (per mile) for the same car under
identical conditions, traveling along the all speed traces in the sample. I will refer to this notion
of agreement between SCC and sample as "total emissions agreement". This definition can be
extended in an obvious manner to define a measure of agreement between two populations of
speed traces, between a population and a sample of speed traces, etc. Carlson and Austin (1997)
argue that their new SCCs will be representative of the samples from which they were derived by
showing that the SCCs are close to the samples on a number of summary measures, (see pages
32-33). It appears that all of the stated summary measures are functions of the (sample or SCC)
SAFD. This includes the PKE and the specific power distribution. As a result, agreement
between the SAFD of the SCC and the SAFD the sample of speed traces implies agreement on
all of the above summary measures. It is also apparent that the method for finding SCCs,
proposed by Carlson and Austin (1997), explicitly attempts to minimize the difference between
the SAFD of the SCC and the SAFD of the sample. Finally, Carlson and Austin (1997) display
many pages of SAFD plots, apparently believing that these plots capture something fundamental
about the relationship between the SCC and the sample. This leads me to define SAFD
agreement between a SCC and a sample of speed traces as agreement between their underlying
SAFDs. SAFD agreement between a population and a sample of speed traces can be defined in
an analogous manner.
11
-------
It is tempting to identify SAFD agreement with total emissions agreement. Such an
identification, if it were permissible, would allow us to regard probablistic statements (i.e.
confidence intervals) about estimators of the SAFD as if they were statements about
corresponding SCCs. This is useful because, as I will show below, confidence intervals for
estimators of the SAFD are easy to obtain. In addition, the identification of SAFD agreement
with total emissions agreement has implications for the design of algorithms for extracting SCCs
from samples of speed traces.
The crucial question now becomes: Is it permissible to identify total emissions agreement
with SAFD agreement? I will refer to the hypothesis that the above two forms of agreement are
(for all practical purposes) identical as the total emissions hypothesis. More specifically, I will
define the total emissions hypothesis as the proposition that the instantaneous rate of emissions
for any fixed car, moving in a straight line on a constant grade, is a function of speed and linear
acceleration alone. In symbols:
e=f(s, a)
where e is the instantaneous rate of emissions, s is the speed, a is the linear acceleration, and f is
an unknown function. There would seem to be good theoretical reasons for supposing that the
total emissions hypothesis is, at least approximately, true. An argument for this would go as
follows: Emissions result from the burning of fuel and as such they ought to be proportional to
the rate at which the automobile dissipates energy. It would seem that energy is dissipated by 1)
changing the car's kinetic energy, 2) overcoming drag, and 3) heat loss. It seems clear the first
two of these energy sinks' are functions of speed and acceleration. It is less clear, however, that
heat loss, particularly in the engine, could be fully accounted for by speed and acceleration alone.
In fact, Cohen et al (1993), page 7, cites studies which indicate that speed and acceleration along
cannot account for all emissions. I recommend that this matter be settled definitively (if it has
not already been settled) either by equipping instrumented vehicles with emissions detection
equipment (as in the studies cited by Cohen) or by means of computer simulation models using
programs such as PC-VEHSIM, Carlson and Austin (1992). If the total emissions hypothesis is
not even approximately true, then all of the evidence given in Carlson and Austin (1997) in
support of the new SCCs, is suspect. If, on the other hand, the total emissions hypothesis is
substantially true (as I suspect) then we should proceed to make full use of the SAFD in
developing SCC extraction algorithms and in obtaining confidence regions.
3.2 Standard errors for the SAFD and summary statistics
Appendices A and C contain many tables of empirical SAFDs from various populations of speed
traces. Summary statistics derived from these SAFDs are also presented. No standard errors or
confidence intervals are presented for any of these parameters. This makes it impossible to judge
the variability of these estimates. I recommend that standard errors for the above quantities be
formed by treating each individual car/driver combination as an independent observation.
Estimates of population wide summary measures can then be obtained by taking means of
summary statistics calculated on each car/driver combination. Standard errors for these estimates
12
-------
can be calculated in the usual manner. For example, suppose one wanted an estimate the
population SAFD at a point (s0, a,J. I will call this quantity S (s0, a,J. To compute an estimate
for S(sO, aO), begin by estimating the SAFD of each car/driver combination at (s0, a,J. These
estimates will be called (s0, aJ, i = 1,..., N, where A' is the number of car/driver combinations.
The sample mean and standard error for {'Si (s0, aJ, i = 1,..., N} can then be used as a point
estimate and standard error for S (s0, a,J. More complicated, and potentially more accurate,
methods for estimating S (s0, a(t), could be developed by using random effects models (see
below).
3.3 Kernel density estimators of car/driver specific SAFDs
I suggest that kernel density estimators be used to estimate the car/driver specific SAFDs
(Silverman, 1986). These estimators will generally have better statistical properties then the
histogram estimators used by Carlson and Austin (1997). Unlike histograms, kernel density
estimators do not rely on arbitrary cutpoints for distinguishing between neighboring bins.
Instead, kernel density estimators use an empirically determined bandwidth (or smoothness)
parameter to produce estimates of a continuous density curve at any given point by taking
weighted averages of observations found near that point. Kernel density estimators will
outperform histograms most dramatically when one is attempting to estimate a density based on a
relatively small number of observations. This is important because (as discussed in Subsection
3.2 I am recommending that, for IV data, a separate estimate of the SAFD be computed for each
individual car/driver combination and, for CC data, a separate estimate of the SAFD be
computed for each car/driver combination on each facility type at each LOS. Under these
conditions some SAFDs may need to be estimated from relatively short driving segments. Kernel
density estimators are available, and easy to compute, using the S-PLUS statistical software
package (Statistical Sciences 1995). Some issues that arise in density estimation include the
choice of an appropriate band width and the overcoming of edge effects, which result from the
restriction that speeds may not be negative. A full resolution of these issues is beyond the scope
of this review, and could not, in any case, be accomplished without access to the speed trace data.
I recommend that the EPA contract with a competent statistical consulting organization to help
with this project.
3.4 A mathematical programming algorithm for extracting a SCC from a
sample of speed traces
The methods used by Carlson and Austin (1997) for extracting SCCs from a sample of speed
traces has a number of apparent shortcomings. First, the method does not guarantee that an
optimal fit (in any sense) to the sample SAFD has been found. Second, the method is subjective
in the sense that it requires human intervention to edit prospective speed correction cycles.
Third, the requirement, that the SCC be composed of segments of speed traces which actually
occur in the sample, greatly complicates the process of finding optimal SCCs and is never
justified rigorously. I suggest that the task of extracting a SCC from a sample of speed traces be
13
-------
formulated as a constrained optimization problem. To illustrate the process, suppose that we are
given a smooth estimator of a SAFD, which we will call 'S (s, a). 'S(s, a), will be a density
function defined for non-negative speed s and acceleration a. In practice, "S (s, a) would be a
(possibly weighted) average of car/driver specific estimators of SAFDs (of the form ~S;(s0, a0),
/= 1 N, described in Subsection 3.2).
Our goal is to find a speed trace with a SAFD which is very close to "S (s, a). I will
suppose that this speed trace is to be 100 seconds long, and that acceleration is to be piecewise
constant with jumps at each whole number time point. All speed traces of this kind can be
represented by an ordered pair (V, A), where V is a scalar initial speed and A is a 100 dimensional
vector with entries Ap i = 1,..., 100. At will be the acceleration between time i - 1 and time i. I
will denote the kernel density estimator of the SAFD of the speed trace (V, A) by ~S(s, a)[V, A],
Consider the measure of agreement between 'S (s, a) and ~S(s, aj [V,AJ given by
For fixed k > 1 we would like to find a value of (V, A) that minimizes (1). Side conditions like
for some C > 0, could also be imposed. The inequality in (2) insures that the acceleration does
not change too rapidly. (Changes in acceleration that are too rapid may result in SCCs that cannot
be driven). The integrand in (1) is differentiable for k > 1 and subdifferentiable for k> 1. If the
integral in (1) is approximated by a double summation, the resulting optimization problem can be
solved by standard methods. See Fetcher(1987) and NAG (1997) for a discussion of
optimization algorithms which assume differentiability. Hiriart-Urruty and Lemarchechal
(1993), and Schram and Zowe (1993), discuss bundle trust region algorithms, which apply to the
subdifferentiable case. If we accept the total emissions hypothesis, as defined in Subsection 3.1,
then an algorithm of the form described above may produce SCCs that are better matches to the
target SAFD than those proposed by Carlson and Austin (1997) It is likely that the improvement
would be most pronounced for freeway ramp SCCs, which appear to have poor fits to their target
SCCs (see Table 7 on page 32 of Carlson and Austin (1997)). If, on the other hand, the total
emissions hypothesis cannot be accepted, then the measure of agreement as given in (1) will need
to be modified.
3.5 More and better summary tables
Arguments such as those presented in the first paragraph of page 25 need to be supported with
additional summary tables. Standard errors and p-values should be included in these tables
j JS(.s\a) -S(s,a)[V,A~\kdsda
(2)
\Ai-Ai-l |
-------
whenever appropriate. It is not sufficient to point the reader to the lengthy undigested computer
output in the appendices with statements such as "By studying the detailed matrices the
differences in freeway operations between the cities become apparent".
3.6 Standard errors for the weighting factors
Standard errors for the weighting factors, which appear in the definition of composite emissions
at the top of page 26, should be obtained (if possible) from network based transportation models.
Error propagation formulas can then be used, in conjunction with standard errors from laboratory
emissions testing, to estimate the standard errors for the composite emissions. It may be possible
to adjust the standard errors for the composite emissions to account for variability in the
underlying SCC.
3.7 Defining SAFDNonFwy
The equation that defines SAFDNonFwy near the bottom of page 26, is incorrect as it stands. It
should read
SAFDIV-TFxSAFDFwyCC
SAFDNonFwy- l w ^ (4)
This modification is required to insure that SAFDNonFwy integrates to 1. Even (3) needs to be
used with care when estimates for SAFDIV, TF, and SAFDFwyCC are used in place of their true
population values (because the resulting estimate for SAFDNonFwy can assume negative values). I
suggest that one uses an estimator which rounds negative values of SAFDNonFwy (in (3)) up to
zero. SAFDNonFwy should then be renormalized so that it integrates to 1.
3.8 Random effects models
In some instances, the summary statistics from car/driver specific estimates of the SAFD may not
be independent. For example, when one compares the SAFDs from chase car data across LOS
groups, many car/driver combinations are likely to contribute to more than one LOS group. That
is, the same car/driver will be followed across several LOS boundaries. In such a case, car/driver
should be regarded as a random effect, leading one to consider mixed effects models. Mixed
effects models can be fit using SAS PROC MIXED (SAS, 1997; Little, Milliken, Stroup, and
Wolfinger, 1996). On a more abstract level, when one is estimating population-wide densities
of the SAFD from car/drive specific SAFDs, it is intuitively clear that more weight should be
given to estimates from car/driver combinations that have been observed for a longer period of
time. To address this issue, one could use weighted least squares in the mixed modeling
described above. Weighting each observation by the square root of the time observed might be a
reasonable option. Estimating the standard error of the SAFD summary statistic would lead to
15
-------
better weights.
4 Other comments
1. A more detailed description of the methods used for calculating "bump up' factors should
be given. In particular, it needs to be explained how the equation in the middle of page
B-3 could be used to compute "bump up" factors in the absence of data from both chase
cars and instrumented vehicles, i.e. for Atlanta and Los Angeles.
Because Los Angeles and Atlanta lacked instrumented and chase car data,
respectively, the bump-up factors for these cities were calculated using data from
Baltimore to replace the missing data. The Atlanta bump-up factor was never
used.
2. The hypothesis that there are differences in speed traces for different classes of car, i.e.
sports cars as opposed to economy cars, should be tested.
The hypothesis that different kinds of cars are driven differently is intuitive, and in
an ideal world we would test this assumption. However, given the time and
resources available for testing, we think it is reasonable to treat all cars as an
aggregate group, especially since we do not expect MOBILE users to have fleet
and activity input data that specifies the "class " of car.
3. The hypothesis that SAFDNonFwy is city independent is stated as an assumption (see bottom
of page 26). This should be tested.
Graphical and statistical cross-city comparisons of driving on specific
facility/LOS combinations are provided in Appendix A pages A-40 through A-41
and pages ASS through A-67. While these comparisons are not a conclusive
verification of the assumption that driving cycles are not city-dependent, they are
consistent with this assumption. Given the time and resources available, we felt
the assumption was a reasonable one.
Driving Cycle Literature Reviewed
Austin T et al (1997). "Methodology for Generating Driving Cycles for Inventory Development".
Report No. SR95-09-02, Sierra Research Inc.
Austin T et al (1992). "Characterization of Driving Patterns and Emissions for Light-Duty
Vehicles in California". Report No. SR92-11-02, Sierra Research Inc.
Carlson T and Austin T (1997). "Development of Speed Correction Cycles". Report No.
SR97-04-01. Sierra Research.
16
-------
Carlson T and Austin T (1994). "Evaluation of New Driving Cycles", Report No. SR94-0903
Sierra Research Inc.
Cohen, J. et al (1993). "Methods for Driving Cycle Development and Validation." Systems
Applications International.
Defries T. and Kishan 5. (1992). "Light-Duty Vehicle Driving Behavior. Private Vehicle
Instrumentation, Volume I", Technical Report. Radian Corporation.
Other Reference
Fletcher R (1987). Practical Optimization. Wiley New York.
Hiriart-Urruty, J-B. and Lemarchechal, C. (1993). Convex Analysis and Minimization
Algorithms. Springer-Verlag, Berlin.
Little, R.C., Milliken, G.A, Stroup, W.W., and Wolfinger R. D. (1996). SAS Systems for Mixed
Models. The SAS Institute, Cary, NC.
NAG Ltd. (1997). The NAG FORTRAN Library Manual, Mark 17. Programming. The
Numerical Analysis Group, LTD., Oxford, UK.
SAS Institute (1997). The MIXED Procedure. SAS/STAT changes and enhancements through
release 6.12 pp. 571-702, Cary, NC.
Schram, H., and Zowe, J. (1993). A version of the bundle idea for minimizing a nonsmooth
function: conceptual idea, convergence analysis, numerical results. SIAM Journal on
Optimization, 2, 121-152.
Silverman, B.W. (1986). Density Estimation for Statistics and Data Analysis. Chapman and
Hall, New York.
Statistical Sciences (1995). S-PLUS Guide to Statistical and Mathematical Analysis, Version
3.3. StatSci, a division of MathSoft Inc., Seattle.
17
-------
18
-------
EPA Response to Comments from Prof. H. Christopher Frey,
North Carolina State University, March 1998
The following document reproduces Prof. Frey's comments in plain text and intersperses
EPA's response in indented italic.
The purpose of this review is to provide comments on the methodology and validity of the
assumptions used in the report "Development of Speed Correction Cycles." The comments are
organized based upon the major sections of the report.
Title:
The title does not accurately convey the content of the report. The report describes development
of new facility-specific driving cycles, which is a significant departure from the speed correction
cycle approach currently used. A better title might be "Development of Facility-Specific and
Weighted Average Area-Wide Driving Cycles"
While the title has no impact on the conclusions of the study, clear writing is important to us.
However, changing the report would require a new contractor work assignment. We will
communicate this suggestion to the contractor as feedback and try to be clearer in other
references to this work.
Chapter 1 (Summary)
The use of facility-specific cycles that represent operation under different conditions is a
significant advancement over the current speed correction cycle approach, and it holds promise
for substantial improvement in highway vehicle emission factor estimation.
A key issue for future work is to assess whether the emissions for the new cycles are significantly
different from previous cycles. Furthermore, it is also important to determine whether emissions
on the new cycles are significantly different from each other. If two or more cycles produce
similar emissions, then it may be possible to reduce the number of cycles used for additional
vehicle testing. While not part of the scope of this report, recommendations should be made to
EPA in this regard.
This is a useful suggestion for future work.
The non-freeway segment of the area wide cycle (this is not a "speed correction cycle", so the
name should be changed) shown in Table 2 has a higher maximum acceleration than the arterial
and local roadway cycles shown in Table 1. Somewhere this should be explained. It is not clear
why the arterial/LOS and local roadway cycles were not weighted in some manner as input to the
development of an area-wide non-freeway cycle.
19
-------
We will add a cover sheet explaining unclear references, including what is meant by "speed
correction cycle. " The question on area-wide non-freeway cycles is addressed below.
Chapter 2 (Introduction)
The first paragraph should be double-checked for accuracy. It is unclear from other
documentation as to whether the MOBILE5 model is fundamentally based upon the entire FTP.
cycle, or whether it is actually based on Bag 2 of the FTP cycle, which is the hot stabilized
portion. The speed correction cycles mentioned in the first paragraph are all, as I understand it,
in hot stabilized mode. Thus, there are other correction factors aside from that used for average
speed, to adjust for differences in operating mode, ambient temperature, etc. Presumably the
intent here is that all proposed cycles would also be in hot stabilized mode, although this could
be an issue for the local cycle and some of the arterial cycles. This should receive some
discussion.
We will add a cover sheet explaining unclear references, including the discussion of the
Mobile5 model.
A key deficiency of the speed correction factor approach, which perhaps is implied but is not
explicitly stated, is that it involves an "interpolation" from one average speed to another.
However, since average speed is a meaningless representation for a driving cycle, as described in
the report, the speed correction factor actually facilitates extrapolations to nonexistent situations
for which no data were collected for any cycle.
While "average speed" may not be the best representation of a driving cycle, we do not
believe it is meaningless. In particular, average speed and roadway type are likely to be the
most detailed information available to modelers trying to characterize driving in their
metropolitan area, thus it is important to describe driving cycles by their average speed as
well as the roadway type on which such cycles may be driven. It is true that MOBILE often
extrapolates from available data to situations for which no data is available. This is one of
the basic purposes of the model.
In the second paragraph, it is probably not correct to say that it is assumed that the speed
correction factors apply universally, at least not in all cases, in the sense that many of the speed
correction factors are clearly intended to represent particular types of roadways under particular
traffic conditions. For example, the high speed cycles clearly represent freeway driving under
low congestion conditions. In practice, the Mobile5 model may be misused, but in that case the
assumptions leading to misuse reside with the model users, not with the model itself.
While it is true that high speed correction factors only make sense for low-congestion
freeway driving, the fact remains that MOBILE5 does not provide a mechanism for
distinguishing typical driving on different types of roadways at the same average speed.
20
-------
Page 6: in the description of Task 2 it is mentioned that city-specific bag weighting factors can
be obtained from transportation planning model outputs. Without more detail given, the
reviewer is highly skeptical of this claim. Transportation planning models typically do not
produce the microscale data required to develop speed-acceleration frequency distributions.
Transportation simulation models may be able to produce such output, but require considerable
"calibration" and tweaking, and are probably not at a state of development or practice where they
can be reliably used for this purpose.
The reference to "bags " on page 6 may be confusing and we will clarify this in our cover
note for the report. The city-specific area-wide emissions are simply a weighted average
combination of various cycles as described on page 26 of the report. Guidance for deriving
area-wide emissions from transportation models is provided in the report M6.SPD. 004,
"Guidance for the Development of Facility Type VMT and Speed Distributions. "
The discussion of Tasks on pages 5-7 is probably useful from a contractual viewpoint, but is
distracting to the reader in terms of following the main ideas. A preview of the main activities of
the project in the order that they are given in the following chapters would be more useful to the
reader.
Clear writing is important to us. However, changing the structure of the report at this point
would require a new contractor work assignment. We will communicate this suggestion to
the contractor as feedback.
Chapter 3: (Facility -Specific Cycle Development)
A key assumption mentioned at the bottom of page 9 and top of page 10 is that driving data
recorded for a given facility and LOS are not dependent upon the city in which the driving was
performed. This assumption must be verified.
Given the time and resources available, we felt the assumption was a reasonable one.
Graphical and statistical cross-city comparisons of driving on specific facility/LOS
combinations are provided in Appendix A pages A-40 through A-41 and pages ASS through
A-67. While these comparisons are not a conclusive verification of the assumption that
driving cycles are not city-dependent, they are consistent with this assumption.
When deciding upon grouping of LOS categories, it is mentioned that LOS A, B, and C for
freeways had similar speed/acceleration profiles. However, the more relevant criteria would be
whether emissions differ significantly for these LOS categories. More justification needs to be
given as to why LOS A, B, and C were sufficiently similar that it was judged that emissions
would be similar.
21
-------
Similarly, it was decided not only to keep LOS D, E, and F as separate cycles, but to add a fourth
LOS "G" cycle. Even though these cycles have different speed and acceleration profiles, they
may or may not have significantly different emissions. As an aside, we have found that in some
cases the emissions for the low speed cycles used for the speed correction factor in Mobile5 are
not significantly different (Kini and Frey, 1997).
An added cover sheet will explain unclear references, including the selection of LOS
categories for cycle development.
The discussion of the High-Speed Freeway Cycle appears to be similar to that for LOS-Based
Freeway cycles, with the exception of a minimum speed cut-off. Some discussion of speed
ranges should be given for the LOS-based Freeway cycles for LOS A, B, and C to clarify the
differences of this cycle compared to the high speed freeway cycle.
Table 3 on page 13 provide a summary of the speed and acceleration characteristics of the
two cycles. Furthermore, the driving traces in Figures 1 and 2 provide a graphical depiction
of the differences in the two cycles. In particular, the high-speed cycle lacks the low-speed
driving of the LOS A-C Cycle and includes less acceleration.
The discussion of creation of three arterial/collector cycles on page 10 is not well-motivated. As
discussed in the introduction, average speed is not a good way to represent a driving cycle. Thus,
it is not a sufficient rationale for grouping data. Most likely, the discussion in the text is not fully
representative of the thinking that went into development of the groupings, and additional
discussion would solve this problem.
As noted above, facility-specific average speed is an important driving cycle characteristic
because it is one of the few cycle characteristics that can be determinedfor local areas.
Average power for each LOS was also considered, but the LOS groupings were based
primarily on average speed and sample size (e.g., there weren 7 many LOS F observations
for arterials and collectors, so sample size constraints led E and F observations to be
grouped together).
The approach to using "segments" seems reasonable, but does raise issues regarding possible
autocorrelation of emissions with the speed/acceleration history of the vehicle, which could lead
to some unintentional discontinuities when segments are spliced together in terms of trends with
which speed and acceleration may be changing at the end of one segment and the beginning of
the next. However, the segment approach seems to be more realistic than the microtrip approach
previously used.
The procedure used for splicing segments together, based upon matching of endpoint speed and
accelerations of one segment to those of the adjacent segment, appears to be reasonable.
However, the tolerances used (plus or minus 0.5 mph for speed and plus or minus 0.5 mph/s for
acceleration) appear to be arbitrary unless they are more fully discussed. If there is some basis
22
-------
for these tolerances, they should be explained. For example, it would appear that 0.5 mph
change in speed from one second to the next is reasonable, since this implies an acceleration of
only 0.5 mph/s. A change in acceleration from, say, 0 mph/s to 0.5 mph/s implies a change in
speed of only about 0.5 mph from one second to the next. However, if one segment has rapidly
changing accelerations, and is spliced to a segment with constant acceleration, this might lead to
a discontinuity that could influence emissions. Thus, another criteria to consider would be the
rate of change in the acceleration at the end points of the segment.
We agree that there are situations where discontinuities between segments could
theoretically create cycles with emissions different than the emissions from the underlying
segments. While it is not clear that such discontinuities have a significant impact on the
cycles developed here, matching the rate of change in acceleration at the end points of the
segment would be helpful in avoiding these situations in the future, and we will consider
including this criterion when developing future cycles.
In the meantime, the selection of 0.5 mph and 0.5 mph/s as tolerances for matching was
based on engineering judgement since there was limited second-by-second data to determine
exactly what tolerances would be best.
The description of "DiffSum" needs to be clearer. Presumably, the absolute values of the
differences are used. Why not use a square error? It would help to show an equation for DiffSum
so that there is no ambiguity regarding how it is defined.
DiffSum will be described in the cover sheet.
One potential problem with the DiffSum approach is that it is equally sensitive to many small
deviations and to a few very large deviations. For example, if there were 10 bins of data and
each was in error by one unit, then DiffSum would be 10. However, if eight of the bins were
exactly correct, but two were in error by 5 units each, DiffSum would also be 10. It is likely that
many would prefer the former case to the latter case, since the former case would seem to have a
small random error while the latter case has a large error for some cases (perhaps indicative of a
systematic error). If a squared error approach were used, the latter case would receive a much
higher squared error (50) than the former case (10) and hence would be penalized more heavily.
The DiffSum approach may also depend upon the size of the bins used for the SAFDs. The
sensitivity of the DiffSum statistic to bin size should be discussed.
These are useful criticisms of the absolute value approach to comparing cycles. In future
work we will seriously consider using a squared value approach instead. However, we do not
believe that the difference between the two approaches is likely to have significantly skewed
the selection of segments for the cycles developed here.
23
-------
An explanation of the emissions significance of the high specific power modes that are
mentioned on the bottom of page 12 would help. Specifically, on what basis are emissions
expected to be significantly different for the range of 200-299 mph2/s versus >300 mph2/s?
These two modes (200-299 mph2/sec and >300 mph2/sec) were selected because EPA
research conducted during the development of the Supplemental Federal Test Procedure
(SFTP) indicates that in these ranges, some vehicles are designedfor "commanded
enrichment" which can significantly increase emissions. We felt it was important that these
modes not be under- or over-represented in the new cycles. We will add a cover sheet
explaining unclear passages, including a more detailed description of the choice of high
specific power modes.
It should be noted that the speed/acceleration distribution may not be the only main consideration
that affects emissions. It is possible that autocorrelations (or the time series of
speed/acceleration) are important as well. This possibility should be acknowledged and
recommended for investigation when emissions data are collected for these cycles. To the extent
that the times series or time history of speed and acceleration influences emissions, the procedure
used here may require modification in the future.
The second-by-second emissions data collected on the cycles described here does allow the
analysis of autocorrelation effects. While modeling such effects is beyond the scope of
MOBILE6, current work elsewhere to develop "modal models " is beginning to investigate the
time history of emissions, particularly for catalyst effects.
The description of the basic characteristics of each cycle can encourage some of the pitfalls of the
previous speed correction cycle approach. It would be useful to move away from average speed
as a description of a driving cycle. In Table 3 information for average speed, maximum speed,
and maximum acceleration are presented. However, it would be more useful to know something
about the joint distribution of speed and acceleration. While it may be easier to show this type of
information graphically, using an approach such as in Figure 13, it would also be possible to
create some large bins (each with a speed range and an acceleration range) and to indicate what
portion of the total cycle time is contained in each bin. Even if only four bins were given (e.g.,
low speed and high speed, each with low acceleration and high acceleration) this might give a
more meaningful indication of each cycle. Such information could be tabulated. Alternatively,
information regarding the standard deviation of speed and acceleration, and regarding correlation
among speed and acceleration (or speed and positive accelerations) would be useful.
If it is desired to use average speed to describe a cycle, then also include the standard deviation of
speed so that one can gain some idea of the variability in speed associated with the cycle.
We initially requested more detailed summary statistics from the contractor. However, given
the time and resource constraints on this work assignment, we agreed to accept a more
limited statistical description and comparison.
24
-------
Chapter 4: (Area-Wide Cycle Development)
In Chapter 3 cycles are presented for Arterial (LOS A-B, C-D, and E-F), and local roadways.
These are not mentioned in Chapter 4. Why? Why is it not possible to develop an area-wide
cycle by weighting of the freeway cycles with these additional cycles for arterials and local
roadways? Or, stated more specifically, why was a single non-freeway cycle developed instead
of weighting the arterial and local cycles? If a single non-freeway cycle is to be used, why were
arterial/local cycles developed?
The report does not adequately explain the purpose of the area-wide cycles. These cycles
were developedfor eventual comparison with MOBILE6 area-wide results andfor
comparison with other area-wide speed cycles. While the facility-and-speed specific cycles
discussed in Chapter 3 were developedfrom the chase car data, the area-wide cycles were
developed from the instrumented vehicles. Since the instrumented vehicle data did not
distinguish arterial and local driving, a single "non-freeway " sub-cycle was developed. This
explanation will be added to the cover sheet
What is the precision and accuracy of the measurements for speed made using the chase car?
Are measurement errors negligible? Was acceleration calculated using a finite difference based
upon second-by-second speed data? What is the error in the acceleration data/estimates? Some
discussion of this is needed.
Sierra Research Report SR92-02-01, "Design and Operation of an Instrumented 'Chase Car'
for Characterizing the Driving Patterns of Light-Duty Vehicles in Customer Service, "
preparedfor EPA, February 29, 1992, describes the chase car methodology; however, the
report does not quantify measurement errors. Acceleration is calculated each secondfrom
successive speed measurements.
Speed estimates of instrumented vehicles are described in EPA report 420-R-93-007, May
1993, "Federal Test Procedure Review Project: Preliminary Technical Report. " Many
instrumented vehicles have a speed resolution of only 1 mph.
In Chapter 3 a key assumption mentioned was that it has been assumed that driving on a
particular facility and a particular LOS is similar in any of the cities. Has that assumption been
verified by comparing chase car data from different cities for the same facility type and same
LOS? This pertains to Table 4, in which it is assumed that it is possible to compare facility/LOS
from one city to another. There is a brief mention of this on page 25, but more discussion is
needed.
A similar comment was made on Chapter 3. See our response above.
Table 4 is somewhat confusing because of all of the percentages shown. It would be helpful to
show check sums of 100% at the bottom of each set of categories merely for clarity, or to
25
-------
separate this into three different tables. Presumably, the middle set of rows (labeled by LOS
only) is for driving on the given LOS on all facility types. However, since data for arterials were
developed based upon LOS designations, it would be useful to also show the breakdown of
Arterial Total by LOS categories.
We agree this table could be clearer. However, given the resource constraints on this
project, at this time we have decided not to ask the contractor to revise the table or to
generate the travel distributions for arterials by LOS.
The discussion in the paragraph on the middle of page 24 would be strengthened by mentioning
not only average speed, but also standard deviation of speed for Spokane and Los Angeles.
As mentioned above, we initially requested more detailed summary statistics from the
contractor. However, given the time and resource constraints on this work assignment, we
agreed to accept a more limited statistical description and comparison.
The presentation of Table 5 and the association discussion on pages 24 and 25 is difficult to
follow. It might be more meaningful to show 3D graphs of the joint distributions for
speed/acceleration for at least a few cases, so that the reader can gain some insight into what the
"difference" represents. For example, a 3D graph of the joint distributions for speed/acceleration
for each of LA and Baltimore, accompanied by a 3D graph of the distribution of "differences"
between the two, would clearly indicate what the values in Table 5 represent. All three graphs
could be placed on one page. Presumably the data used to develop Table 5 are based upon all
LOS categories for freeways and non-freeways, respectively.
We would like to make the report easier to follow. However, the changes suggested here
would require a new contractor work assignment. We will communicate this suggestion to
the contractor as feedback.
The footnote on page 25 seems to be missing some words. It states "Large differences occur at
the congestion levels..." Which congestion levels? Is this meant to refer to a particular LOS?
The footnote should read: "Large differences occur at specific congestion levels... " This
will be noted in the cover sheet.
The word "segment" is used on the last paragraph on p. 25, but appears to have a different
meaning than on pages 11-12. The terminology should be more clearly defined. Would it be
correct to say that the area-wide cycle is comprised of facility/LOS-specific subcycles?
We will add a cover sheet explaining unclear passages, including the use of the term
"segment."
The equation on the top of page 26 appears to be missing the LOS "G" cycle for freeways.
26
-------
No, the area-wide cycle does not have an explicit LOS G sub-cycle. LOS G is included in the
area-wide cycle as a subset of LOS F.
The weighting factors are said to be the fraction of travel in miles. This presumes that the
emissions are in units of g/mi? The units are not defined. Please make sure units are correctly
specified for the weighting factors, emissions, and composite emissions.
We will add a cover sheet explaining unclear passages, including units for the equation on
page 26.
On page 26, it is stated that "short trip" bias was accounted for by adjusting the travel mix. How
was this adjustment made? The answer is in Appendix B but some description of the approach
would help. It is especially useful to mention that the adjustment procedure is based upon
measured values from both the instrumented vehicles and the chase car data sets.
We will add a cover sheet explaining unclear passages, including a reference to Appendix B.
Why would the non-freeway target (speed/accel?) driving distribution be city-independent? Was
the back-calculated non-freeway driving distribution compared to the non-freeway driving
distribution obtained from the chase car in different cities to check for expected differences or
expected similarities?
Because the chase car data underrepresents short trips, it was not possible to make a direct
comparison of the non-freeway driving SAFD with the chase car data from the different
cities. And, because the area-wide cycle is meant only for comparison and not for city-
specific emission calculations, the assumption that the non-freeway driving distribution is
city independent is appropriate for MOBILE6 purposes.
On top of page 27, more discussion is needed regarding the approach to development of the non-
freeway area-wide cycle. It is stated that this cycle is based upon instrumented vehicle (IV) data.
However, earlier (on p. 9) it was stated that facility type could not reliably be inferred from the
IV data. Thus, what quality assurance was performed to assure that the data used from the IV
data set truly represents non-freeway facilities?
The DiffSum procedure was used to select instrumented vehicle segments that represent non-
freeway cycles. These are the segments that best match the characteristics of the non-freeway
driving. Thus, while there is a small chance that some of these segments were actually
driven on freeways, the segments match the speed and acceleration profiles of non-freeway
driving and can be expected to produce a good representation of non-freeway emissions.
When comparing figures (i.e. Figures 13 and 14) it would help to have both on the same page or
on the same spread (facing pages).
27
-------
We would like to make the report easier to read. However, the change suggested here would
require a new contractor work assignment. We will communicate this suggestion to the
contractor as feedback.
Presumably, the cycle presented in Figure 13 is a weighted area-wide cycle for Baltimore that
includes the four freeway LOS cycles, the freeway ramp cycle, and the non-freeway cycle. If so,
it would help the reader to have more information regarding the specific example for Baltimore.
Specifically, it would help to show the weighting factors used for each of these six constituent
cycles. Note that in Appendix B time-based weights are given, whereas on page 26 distance-
based weights are given in the equation for emissions. The discrepancy between these two needs
to be corrected or explained.
While the equation for composite emissions on page 26 says the weighting factor represents
the fraction of travel "in miles, "for convenience, the sample area-wide cycles described in
this report actually use time-based weighting factors derivedfrom instrumented and chase
car data as described on page B-2. Similarly, time-based weighting factors were usedfor the
Los Angeles area-wide weighted cycle described on page 33. If area-wide cycles were to be
constructedfor another city, distance-based weighting factors from local transportation
model outputs would probably be used.
The text on page 28 is incorrect. Figure 13 is the Baltimore instrumented vehicle data.
Figure 14 is a weighted area-wide cycle. This correction will be noted in the cover sheet.
Chapter 5: (Evaluation of Driving Cycles)
There is a general problem in this report with the flow of ideas from Chapter 3 to Chapter 4 to
Chapter 5. Chapter 4 almost seems to be a separate report on a different topic, since it deals with
a non-freeway cycle and seems to ignore the arterial/local cycles. Chapters 3 and 5 appear to
deal with the same set of cycles, but Chapter 5 includes the non-freeway cycle at the end. More
transition is needed from one chapter to the next, since these changes in focus are confusing to
the reader.
We would like to make the report easier to read. However, the change suggested here would
require a new contractor work assignment. We will communicate this suggestion to the
contractor as feedback.
The comparison of facility-specific cycles to the "population" is interesting and informative.
However, the first comparison made is for average speed. As previously noted in the report, this
is not the most important figure of merit when comparing cycles. Thus, it should be emphasized
that not only are there minor differences in average speeds when comparing cycles with the
28
-------
population data, but that average speed is not by itself a meaningful descriptor of a driving cycle.
It would be useful to include standard deviation of average speed and standard deviation of
acceleration when comparing the cycles and populations. It would also be useful to consider
factors such as correlations or covariance between speed and acceleration.
The discussion of maximum speed and acceleration is too brief. While it is not surprising that
any particular sample from the population fails to capture the maximum value from the
population, it would be useful to have some idea of how close the sample is to the population.
For example, one could calculate the percentile of the population distribution for speed
corresponding to the maximum value for the cycle maximum speed (e.g., is the cycle maximum
speed at the 98th percentile of the population distribution for speed, or is it at the 63rd
percentile?). If the maximum speed in the cycle were at a low percentile of the speed distribution
from the population, that would suggest a possible inadequacy in the cycle, whereas if it were at a
high percentile then there would be increased confidence that the cycle is representative of the
population. Without this type of information, the comparison of maximum speed and maximum
acceleration between the cycles and populations is not very helpful.
As mentioned above, we initially requested more detailed summary statistics from the
contractor. However, given the time and resource constraints on this work, we agreed to
accept a more limited statistical description and comparison.
It would also be useful to explain a few cases. For example, for Arterial LOS A-B a maximum
acceleration of 14.9 mph/sec is indicated in the population, but the maximum used in the cycle is
only 5 mph/sec. Was the population maximum due to only one vehicle, with the next highest
acceleration closer to 5 mph/sec? Or is there a large fleet of high performance vehicles that are
not represented by the driving cycle.
The SAFD tables in Appendix C provide the speed and acceleration distributions for the
target populations. Clearly, the driving cycles do not reproduce the most extreme
accelerations and speeds. Significant additional analysis would be required to evaluate how
many vehicles are responsible for the extremes in the population distributions. However, for
emission estimates, the number of vehicles driven at these acceleration and speed extremes is
less important than the fraction of miles driven.
What was the LA92 cycle compared with? It is stated that the LA92 cycle was compared with its
target driving population— what population was actually used in the comparison? This is
important since the total SAFD difference and high power difference for the LA92 are used as
figure of merits for evaluation of the other cycles. It is important to clearly convey that the LA92
cycle is not being used as a strawman, but that the comparison of LA92 to the "target" population
is fair and realistic.
The "population " values for facility-specific driving cycles in Table 7 are from 3-city
(Baltimore, Los Angeles, Spokane) chase car data, while the population values for the LA-92
29
-------
are from the Los Angeles chase car data only. In Table 8, the Non-Freeway Area-Wide
segment is compared to the Baltimore weighted instrumented-vehicle data and the LA-92 is
compared to the Los Angeles chase car data. Since the LA92 cycle was designed to represent
driving in Los Angeles, these comparisons should be favorable to the LA92. This
information will be repeated in the cover sheet.
At the bottom of page 33 it is stated that "segment" (again, this word is used in different ways in
different places in the report) weighting factors for Los Angeles were used to develop an area-
wide cycle for Los Angeles. How were these factors developed? What were they? They should
be summarized in a table.
The use of the term "segment" is described in the cover sheet. The weighting factors for
Baltimore, Spokane and Los Angeles are explained in Appendix B and listed on page B-2. As
described earlier for Baltimore, the time-based weighting factors were used.
The comparison of the Area-Wide and the LA92 cycles with the Chase Car data might be
presented in a different order. It seems useful to compare LA92 with the chase car data, and the
area-wide cycle with the chase car data. These two comparisons illustrated that the area wide
cycle compares more favorably with the chase car data than does the LA92 cycle. The third
comparison, of areawide vs. LA92, seems unnecessary and potentially confusing, but if it is
desired to include it, make it the last one.
We would like to make the report easier to read. However, the change suggested here would
require a new contractor work assignment. We will communicate this suggestion to the
contractor as feedback.
For all of the comparisons, a key point to mention is that whether the error measures (total SAFD
% difference and high-power difference) are useful or appropriate will not be fully known until
emissions data are collected for these cycles. If there is any previous work in which these two
error measures were used and found to be good predictors of agreement or lack thereof with
respect to emissions, it would be useful and important to note that and discuss. In the absence of
such empirical verification of the utility of these error measures, at best these error measures can
be said to be based upon hypotheses regarding factors that are most influential with respect to
emissions, and that new error measures may be proposed after emissions data are collected and
evaluated. If new error measures are later proposed (based upon an empirical foundation), then
new cycles may have to be developed based upon those measures.
The contractor who prepared this report also did analysis for EPA's Supplemental FTP
study. The study found these error measures to be good predictors of emission agreement.
Therefore, they continued using them in this cycle development work.
30
-------
Chapter 6 (Intersection Influence Analysis)
Signalized intersections are of great importance. However, some discussion of non-signalized
intersections (e.g., four way stops with stop signs) would be useful, at least for context. If any
data are readily available on the number of signalized intersections in the U.S. versus non-
signalized intersections, this would provide some context.
As stated on page 35, intersections with stop signs are considered signalized intersections.
We do not know of readily available data on the fraction of intersections that are non-
signalized.
Some more transition is needed, since the third paragraph refers to second-by-second data from
speed correction cycles. This catches the reader off-guard, since the report states early on many
of the problems with such cycles and then proposes new cycles.
We would like to make the report easier to read. However, the change suggested here would
require a new contractor work assignment. We will communicate this suggestion to the
contractor as feedback.
In the 2nd full paragraph on page 36, in the last sentence, there is mention of "new speed
correction cycles." This is a misnomer, since the new cycles are not intended as speed correction
cycles. They are intended for use in developing area-wide cycles based upon weighting of
multiple cycles. The text in this paragraph should be revised.
We will add a cover sheet explaining unclear references, including what is meant by "speed
correction cycle. " However, it should be noted that the cycles discussed here are not meant
to only be used in developing area-wide cycles. They are also used to simulate driving at
specific average speeds on specific roadway types.
Since only a sample of data were used to create the information in Tables 10 and 11, it is
important to make a notation regarding the total amount of driving time and driving distance
upon which the fractions shown in the tables are based, especially for Table 11. In addition, for
each a case a note should be made regarding the percentage of the actual population that was
reviewed (there is one example of this for local roadways, but it is not stated whether the
percentage is on a time or distance basis).
The distance and percentage numbers requested are not readily available. For this reason
and because the intersection analysis was not used in MOBILE6, we have decided not to
follow up on this comment at this time. However, if the results of this analysis are used at a
later date, it would be worthwhile to ask the contractor to generate the requested data.
Since the comparisons of the cycles to the target population are of interest, it would be helpful to
present the data in a form that facilitates comparisons. A landscape table with two columns for
31
-------
each cycle (one for the cycle and one for the population) would help.
We would like to make the report easier to read. However, the change suggested here would
require a new contractor work assignment. We will communicate this suggestion to the
contractor as feedback.
Other Comments
A summary should be given of the recommended uses for the new facility/LOS-specific cycles,
and for the areawide cycle approach, and for emissions data collection.
We would like to make the report easier to read. However, the change suggested here would
require a new contractor work assignment. We will communicate this suggestion to the
contractor as feedback.
The report should conclude with a brief section highlighting the key assumptions underlying the
development of the new cycles, and making recommendations for future work pending the
outcome of emissions testing with the new cycles. Two key issues are:
- Whether there are significant differences in emissions among the new cycles. If two or
more cycles have similar emissions, then it may be possible to reduce the number of cycles used
for large scale testing efforts.
Now that we have emission results from these cycles, we agree that a statistical comparison
of emissions on the new cycles would be a useful tool in selecting which cycles should be
used in future vehicle testing.
- The adequacy of the objective function/error measures used in constructing the new cycles.
If the SAFD % Difference and high-power difference measures, and any others used, are found to
not adequately capture sensitivities between speed/acceleration and emissions, then new cycles
may have to be developed based upon new evaluation approaches.
We agree that tests of these correlations would be useful. Currently we are not planning to
develop any new cycles, but we will keep this suggestion in mind for the future.
Another factor to consider in emissions testing would be road grade, which is not discussed in
the report. However, engine power demand and emissions could differ for the same speed trace
depending upon the road grade. If road grade data are available from the three city study, this
information would be useful to include as part of recommendations for new driving cycles.
The effect of road grade on emissions is beyond the scope of this study and the MOBILE6
model. However, we are considering ways to include road grade in future research.
32
------- |