United States Air and Radiation EPA420-R-02-026
Environmental Protection October 2002
Agency
&EPA EPA's Onboard Analysis
Shootout: Overview and
Results
> Printed on Recycled Paper
-------
EPA420-R-02-026
October 2002
EPA's Onboard Emissions Analysis Shootout:
Overview and Results
Constance Hart, John Koupal, Robert Giannelli
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.
-------
ABSTRACT
The Office of Transportation and Air Quality (OTAQ) of the U.S. Environmental
Protection Agency (EPA) has started work on the next generation mobile source emission factor
model, termed the "Multi-scale Motor Vehicle and Equipment Emission System" (MOVES).1
MOVES will be the successor to the MOBILE6 highway vehicle emission factor model and the
NONROAD model and include ozone precursors, particulate emissions, toxics and greenhouse
gases. In the last few years advances in technology have made it possible to measure vehicle
tailpipe emissions during real-world vehicle operation using portable on-board instruments. This
on-board emissions data will be the basis for the MOVES model. Because the measurement and
interpretation of on-board data is a relatively new area, EPA sought input from external experts
to provide examples and recommendations of modeling approaches including steps to calibrate
and validate their models. Each of the three organizations worked simultaneously and
independently under the same scope of work and each prepared its own report. The three
organizations were North Carolina State University (NCSU), the University of California at
Riverside (UCR), and Environ. EPA also worked simultaneously on another approach. This
paper shows each approach as well as the results of the shootout and then discusses how EPA
will proceed with the exhaust emissions calculation methodology in the MOVES model.
INTRODUCTION
Under the Clean Air Act, EPA's Office of Transportation and Air Quality (OTAQ) is
charged with developing emission factors for on-road sources such as light and heavy-duty
vehicles and trucks, and off-road sources such as construction and agricultural equipment. This
has led to the development of a number of emission factor estimation tools such as MOBILE (for
on-road VOC, CO and NOx), PART (on-road particulate matter and SOx), MOBTOX (on-road
toxics), and NONROAD (all off-road pollutants). These tools have been focused on the
estimation of mobile source emissions based on average operating characteristics over broad
geographical areas. Examples of this scale of analysis are the development of SIP inventories for
urban nonattainment areas and the estimation of nationwide emissions to assess overall trends. In
recent years, however, analysis needs have expanded in response to statutory requirements that
demand the development of finer-scale modeling approaches to support more localized emission
assessments. Examples include "hot-spot" analyses for transportation conformity, and the
evaluation of the impact of specific changes in a transportation system (e.g. signalization and
lane additions) on emissions.
In response to the acknowledged limitations of the MOBILES model in addressing these
modeling needs, separate modeling initiatives have been undertaken to develop tools which
provide a better assessment of finer scale emissions. Three notable efforts are the Comprehensive
Modal Emissions Model (CMEM) developed by UC Riverside under NCHRP Project 25-112;
TRANSEVIS, under development by Los Alamos National Laboratory through the U.S.
Department of Transportation3; and MEASURE, developed by Georgia Tech under cooperative
agreement with EPA's Office of Research and Development4.
The increasing requirements of model users as well as external recommendations from a
October 10, 2002 1
-------
variety of sources have indicated the need for more emission research and improved modeling
methodologies. A comprehensive review of EPA's mobile source modeling program was
published by the National Research Council (NRC) in May 20005. It recommended that EPA
develop a mobile source emission modeling system that is capable of supporting the expanding
range of mobile source emissions analyses. EPA has just released the latest on-road emission
factor model MOBILE6, which represents a substantial improvement from MOBILES,
particularly for finer-scale modeling. We view the development of the "Multi-scale Motor
Vehicle and Equipment Emission System" (MOVES)1 as a logical next step in the continual
effort to improve mobile source emissions models to keep pace with new analysis needs, new
modeling approaches, and new data.
MOVES as proposed would employ three analysis scales termed macroscale, mesoscale
and microscale6'1, and a primary goal of the model will be to use a common set of emission rates
for each scale to enable consistent results from MOVES across analysis scales. An initial issue
paper was published in April 20016 which proposed the concept of an "Emission Rate Estimator"
which would process the same set of instantaneous exhaust emissions data into modal emission
rates for use in all scales of analysis. The issue paper outlined three possible approaches to
developing this Emission Rate Estimator: 1) develop a physical instantaneous emission model
which takes microscopic vehicle trajectory information and produces emissions aggregated to the
desired level; 2) generate modal emission rates directly by processing a database of instantaneous
emissions into modal bins (e.g., acceleration, deceleration, cruise, idle), applying these rates
directly for finer scales analyses and aggregating as necessary for macroscale analysis; and 3)
link directly to a database of raw instantaneous emission measurements, so that the emission rate
estimator would essentially query a database of these raw data.
The NRC recommendations also address the need for improved model science and
improved model structure, two key objectives of MOVES. An improved modeling structure will
allow better responsiveness to new data and enable model validation, which in turn will facilitate
improved science. Improved science is also a direct function of the quality of information feeding
the model. The recent emergence of on-board emissions measurement devices is revolutionizing
how emissions data are collected for on-road and off-road mobile sources. Several commercial
applications of this technology have begun to enter the marketplace or are under development. In
addition, EPA is undertaking a major effort to advance the development of on-board emissions
analysis equipment, termed Portable Emissions Measurement System (PEMS). PEMS will
ultimately allow the gathering of instantaneous exhaust emissions data for HC, CO, NOx,
particulate matter, toxics and greenhouse gases. It will also include a global positioning system
(GPS) to allow linkage of emission measurements with the location and speed of the vehicle.
Vehicle operating information will be monitored and for vehicles equipped with on-board
diagnostics systems (OBD), the OBD stream will provide engine and vehicle operation
information. This system will enable the gathering of in-situ emissions across all mobile sources
in several geographic locations, for relatively low cost. We envision that this technology will
become the focus of EPA's emissions factor testing program and will provide the opportunity for
a significant shift in how emissions modeling is approached.
While on-board emission measurement represents a significant shift in the ability to
October 10, 2002 2
-------
collect real-world data, incorporating these data into emission factor models poses a challenge,
particularly at multiple scales as proposed for MOVES. Data collected with on-board emission
measurement systems would depart from traditional laboratory data in many ways. Whereas
nearly all elements of vehicle activity (speed, acceleration, soak time, etc.) and ambient
conditions (temperature, humidity, road load, road grade) are controlled in the laboratory, all of
these factors will vary from trip to trip in the real world. A method which develops emission
rates from on-board data therefore must contend with far more variation in parameters which
typically do not come into play with lab data, or are varied "one at a time" to specifically gauge
the effect of a particular factor.
The "On-Board Emissions Analysis Shootout" was devised to evaluate potential methods
for using on-board emissions data to generate emission rates for MOVES. The participants were
supplied with on-board emission data gathered on light-duty vehicles, transit buses and nonroad
equipment. They were then required to develop a model calibrated with this data which could
predict total HC, CO, CO2 and NOX emissions from a separate sample of independent data. The
predictions from the various methods were then validated with measured emissions in order to
provide an initial sense for the promise of different approaches to applying on-board emission
data in MOVES. This was considered a pilot study in the sense that not all possible factors which
will ultimately need to be addressed in MOVES were included in the analysis; the study was
designed to hold certain parameters constant, such as fuel effects and vehicle technology. The
purpose of the study was to provide an initial sense of how different approaches could be used to
develop models from on-board emission data, based on a limited sample of on-board data.
The participants in the shootout were selected through a competitive process, based on
evaluation of solicited proposals: North Carolina State University (NCSU), University of
California at Riverside (UCR), and Environ Corporation. EPA also participated according to the
same parameters as these contractors. The three contractors provided full reports to EPA on their
analysis methods and results, which are merely summarized here7'8>9; the EPA methodology,
developed by the authors, is presented in full in this report.
DATA COLLECTION
On-Road Vehicles. On-board emission data was gathered on a sample of 18 light-duty
vehicles in the summer of 2001, and 15 heavy-duty diesel transit buses in Fall 2001 in the Ann
Arbor, Michigan area. The light-duty vehicles were selected from employees of EPA, and of
Sensors Incorporated, and rental vehicles. Sensors provided testing and data processing for the
project10. Because of the small sample size and pilot nature of the project, an attempt was made
to limit the number of vehicle-related variables in the vehicle sample. Thus, all vehicles used
were certified to Federal Tier 1 tailpipe standards, with a model year range of 1996 through 2000,
and had either 4 or 6 cylinder engines. The transit buses were all from the in-use fleet of the Ann
Arbor Transportation Authority (AATA); as such, all of the buses had the identical engine and
after-treatment system, with a model year range of 1995 through 1997, and had similar mileage
accumulation. Overall the bus sample was much more homogenous than the light-duty vehicle
sample. Detailed vehicle information is shown in Tables 1 and 2.
October 10, 2002
-------
Off-Road Vehicles Three nonroad pieces of equipment were used for the shootout: a
bulldozer, compactor, and roller/scraper. These vehicles were recruited from a rental company's
fleet of nonroad equipment in the area around Columbus, Indiana, and were used for a test
program to assess the feasibility of EPA's prototype on-board emission measurement system. See
Tables.
Light-Duty Vehicle Test Procedure. All light-duty vehicles were first brought into EPA's
National Vehicle and Fuels Emission Laboratory for a standard FTP prep, including draining the
vehicle's fuel and refueling with standard certification fuel. The vehicles were then tested over
the FTP and US06 cycles while collecting both bag and modal data, with the on-board
measurement device (SEMTECH-G) installed to allow for correlation testing. The vehicle was
then refilled with standard certification fuel and returned to the owner. The SEMTECH-G
remained on the vehicle for a period of 1-3 days, in order to gather a minimum of 3 hours' worth
of total operating time. During this period the owner was given no special instructions on how to
operate the vehicle to ensure representative driving and operation patterns. The owners were
given trip logs to record the number of passengers and estimated payload for each trip.
Heavy-Duty Vehicle Test Procedure. Instrumentation of heavy-duty vehicles occurred at
the AATA's main garage. To circumvent the possibility of passenger concern, the buses were
operated under normal routes and driving conditions, but without passengers (simulated stops
were still made). The centralized fleet fuel was sampled to gain information on fuel parameters.
The targeted sample period was four hours.
Nonroad Equipment Test Procedure. All vehicles were reviewed by the technician to
confirm that the vehicle was in good working order before installing EPA's device. Installation
usually took approximately one hour to install and was done before and after working hours at
the working site. The on-board measurement device used for nonroad testing was a prototype of
EPA's Simple Portable On-Board Test (SPOT) device. This device was being designed to handle
the harsh environment in which nonroad equipment operate in and has the capability to operate
non-attended, gathering both activity and emission data for up to a week.
Light-Duty Vehicle and Bus Instrumentation. Sensors Inc. provided the on-board
emission measurement and data collection services for this effort under contract to EPA. Two
prototype SEMTECH-G analyzers were used for gasoline powered passenger vehicles and one
prototype SEMTECH-D analyzer for the diesel powered buses. A thorough discussion of the
detectors and calibration procedures and equipment can be found in a separate report prepared by
Sensors10.
Except for the HC and brake-specific mass measurements the SEMTECH-G and
SEMTECH-D analyzers are essentially the same. They both measure raw vehicle exhaust, collect
vehicle engine command module (ECM) data, (e.g., real-time engine information for the on-road
vehicles, including engine speed, throttle position, coolant temperature and air conditioning
status for light-duty vehicles), and store the data on an internal data logger that is automated by
key-on/key-off events. A post-processing utility computes real-time fuel-specific and distance
specific mass emissions based on engine airflow computed from the ECM data. The SEMTECH-
October 10, 2002 4
-------
D has an additional post processing utility which computes brake-specific mass emissions and
uses a heated flame ionization detector (HFID) to measure total hydrocarbon emissions (The
HFID system is operated at 195 °C.) In contrast, the SEMTECH-G determines total hydrocarbons
via direct measurement of hexane with a non-dispersive infrared detector (NDIR) which does not
require a heated sample gas line.
The CO and CO2 analyzers measure on a dry basis with non-dispersive infra-red sensors
(NDIR). The CO analyzer in the SEMTECH-D prototype unit in this study was designed for
gasoline exhaust and has not yet been optimized for the low concentrations found in diesel
exhaust. The analyzer is calibrated between 0.5%, or 5000 ppm and 8%. The levels found in the
buses in this study were below the lower portion of this range. Correlation testing showed a
measurement uncertainty of 50 ppm.
The NDIR for CO2 and CO are both single range devices, 0 - 16% and 0 - 8%,
respectively. A manufacturer supplied multi-point calibration curve is hard-coded into each
analyzer. To account for short term drift and linearity changes a single point and dual point user-
calibrations are available to adjust the entire calibration curve to match the calibration gases.
Both oxides of nitrogen, NO and NO2, were measured simultaneously through unique
adsorption bands of NO and NO2 in the UV range with a non-dispersive ultra-violet (NDUV)
analyzer. The NDUV detectors are also single range devices. The NO range is 0 - 3000 ppm, and
the NO2 range is 0 - 500 ppm. The calibration procedure is exactly the same as for the NDIR
analyzers.
A global positioning system (GPS) was used to keep track of the route taken by the
vehicle with resolution of one meter and an absolute accuracy of 15 meters for latitude, longitude
and altitude. Road grade was computed with the GPS data and some instantaneous results had
significant uncertainty, especially at low speeds. Finally, a probe placed remotely from the
SEMTECH analyzer was used to measure ambient pressure, temperature, and humidity each
second.
Correlation results were compiled by Sensors, Inc. in this report. Overall the bag results
from the SEMTECH-G system used for the on-road testing were within 5 percent of the
laboratory bag analysis system for CO2, CO and NO. HC results were within 5 percent with the
SEMTECH-G system was fitted with a FID; NDIR technology, which does not detect all HC
species, underpredicted the FDD-based bag analysis system by approximately 35 percent. The
NDIR technology was used for the on-board testing; this did not pose an issue for the shootout
analysis, since both the "modeling" dataset and "validation" datasets were based on the NDIR.
Limited correlation testing was also performed with the SEMTECH-D system on 2 heavy-duty
vehicles, with similar correlation to the FID-equipped SEMTECH-G demonstrated.
Nonroad equipment. The SPOT device allows for real-time measurements and presently
consists of: a data logger for storing the data, a wide-lambda O2 and NOx sensor, a venturi mass
air flow sensor, batteries and communication system. Data is gathered on a second-by-second
basis, with date and time stamps. The O2 and NOx are measured directly within the vehicle's
October 10, 2002 5
-------
exhaust stream with CO2 being calculated from the O2 readings.
Data provided to shootout participants. The shootout participants were provided with a
"modeling" dataset of all trips on 12 light-duty vehicles and 12 buses, and 3 hours of operation
on the nonroad equipment pieces, upon which to develop their shootout models. All data fields
collected were provided to the participants, including (for light-duty) second-by-second fuel
consumption and emissions, ambient parameters (temperature, humidity), engine parameters (e.g.
engine speed, coolant temperature, throttle position, mass air flow), GPS coordinates (latitude,
longitude, altitude), road grade (calculated from GPS altitude), and (for most light-duty vehicles)
air conditioning compressor status. These data, which came from several instruments, had been
time-aligned by Sensors prior to delivery to EPA. Sensors' data preprocessing also derived
calculated fields prior to delivery to EPA. Nonroad data provided to the participants was limited
to second-by-second emissions (NOx and CO2, the latter derived from O2), exhaust flow
(provided as a surrogate for engine load), engine RPM, ambient temperature, barometric
pressure, and relative humidity.
ANALYSIS METHODS
Modal Binning
NCSU pursued a modal "binning" approach, in which they defined operational bins based
on changes in speed and power, and refined the estimates within each modal binning using
regression analysis. This proposed methodology falls under the second "Emission Rate
Estimator" approach identified in EPA's April 2001 issue paper, to process instantaneous
emissions measurements produced in the laboratory or in the field.
NCSU looked at the data in many ways before settling on their approach. After extensive
quality control checks and data preparation, they looked at individual trips and the variability of
those trips within the same vehicle and between vehicles. The next step in their visualization of
the data was to look at the importance different variables had on the emissions of this dataset,
which would also be included in the prediction dataset, or could be estimated from the available
parameters. Scatter matrices were prepared for all pollutants to look for relationships among the
explanatory variables, including speed, acceleration, ambient temperature, humidity, altitude,
grade, air conditioning (on/off), and power demand. This analysis showed that there is a
substantial amount of variability in the emissions data and that there is not any single variable
which directly explains a large portion of the variability.
A spatial analysis utilizing the GPS information from the on-board data was also
performed. They were able to get roadway functional class information and determine the effects
of different roadway types on emissions. Based on this dataset there were no clear distinctions in
emissions by roadway types for any of the pollutants. Effects on emissions due to intersections
were also looked at in this effort. There were some effects found which were not consistent over
functional classes. It was concluded that further study would be needed to understand the issues
involved which was beyond the scope of this project.
October 10, 2002 6
-------
Cold Start Identification. A non-linear regression was fit for each trip for each pollutant
by using time as the predictor variable, see Figure 1. The assumption was made that there was
decreasing emissions until the cold start period ended, then the emissions would stabilize at tc.
The upper limit of the confidence intervals for each pollutant of the duration of the cold start, tc,
were compared. In the decision for a duration for each trip, all three values of tc were taken into
account. If large discrepancies occurred, the soak time was also considered for that trip and
previous trips. It was found that 34 of the trips had cold starts ranging from 70 to 391 seconds.
The identification of cold starts was used for categorizing the data so that hot-stabilized driving
could be separated.
Methods used for Modeling. Time series analysis revealed that the data in this study have
autocorrelation, or history effects, and therefore need to be treated carefully. Some thought must
be given to choosing proper statistical methods since the autocorrelation may invalidate the
assumptions. After exploring some time series approaches, NCSU found them to be impractical
for a model such as MOVES, which will require input data from a large number of vehicles and
trips. Alternative approaches must destroy or at least reduce the autocorrelation in the data.
Although there may be some loss of explanatory power associated with ignoring or destroying
autocorrelation, they found it would be possible to use other modeling approaches that are more
practical for taking advantage of the variety of available data sources.
NCSU found that binning the data was a feasible approach for the MOVES model which
would reduce the influence of autocorrelation. They ended up employing a combination of
techniques based upon modal analysis, regression and time series methods.
Hierarchical Tree-Based Regression (HTBR) and Ordinary Least Squares Regression
(OLS). HTBR determines which variable in the model should be selected to produce the
maximum reduction in variability of the response. HTBR lacks some desirable properties of OLS
procedures, such as available statistical tests which might be used to test the differences in HTBR
model formulations. In this study HTBR and OLS regression methods were combined to use the
strengths of both methods. The data were stratified into smaller data sets by using HTBR and
then OLS regression was done to capture relationships within the data strata.
Modal Emissions Binning. The second-by-second emissions data were divided into four
modal categories after taking out the cold start defined earlier: idle, acceleration, deceleration and
cruise. Idle is defined as zero speed and zero acceleration, acceleration is defined as a minimum
acceleration of 2 mph/sec or 1 mph/sec sustained for 3 seconds or more, deceleration is similar to
acceleration with negative rates, and cruise is all other events not defined in the other categories.
Average emission rates for each mode were calculated for each trip, then an average of these
estimates were calculated for all trips. Figure 2 shows these averages for all four pollutants with
the 95 percent confidence intervals. Pairwise t-tests indicate that three out of twenty four
pairwise combinations are not statistically different from each other. Two occurred for
comparison of average emission rate between idle and deceleration mode for both CO and NO
emissions, and the other between acceleration and cruise for HC emissions. This suggests that the
modal definitions used are reasonable.
October 10, 2002
-------
The HTBR technique was used to improve the driving modes. Trip data for each vehicle
were combined together for analysis. Regression trees were formed for each mode for each
pollutant (except CO2) using explanatory variables related to vehicle operation and vehicle
characteristics such as speed, acceleration, power demand, grade and engine size. For CO2
emissions, the original modes were considered to be adequate for their explanatory power.
Power demand was found to be the most useful variable to improve the explanatory
power within most of the driving modes. A single cutoff point was determined for all the
pollutants for each mode. For the continuation of the analysis, the following cutoff points were
used for each mode: 100 mi2/h2.sec for acceleration mode, -100 mi2/h2.sec for deceleration mode,
60 mi2/h2.sec for the cruise mode, and no cutoff for the idle mode since the vehicle is not
moving. Average modal emission rates for the new modes were estimated, see Figure 3 for CO
emissions.
With this approach, NCSU found that cold start and acceleration modes account for
approximately half of the total emissions. More than 30 percent of emissions are emitted during
acceleration mode for Nox, CO and CO2 (and 25 percent for HC). 40 percent or more emissions
for all emissions come from the cruise mode. See Figure 4.
After the development of modal definitions, OLS regressions were fit for each mode
using the explanatory variables: speed, acceleration, power, engine size, ambient temperature,
humidity, altitude and road grade. Second and third powers of speed and acceleration were also
included in the regression analysis. Many terms zeroed out or proved insignificant for different
modes.
Cold Start Model. Since the cold start data is in consecutive seconds, this data was
autocorrelated and a time series model approach was used. It was assumed that all cold starts are
from the same process and the data were combined. A regression model with time series was fit
to the data using the explanatory variables, adding coolant temperature to the list for its known
effect on cold start emissions. Since coolant temperature was not a part of the prediction dataset,
the relationship between coolant temperature and soak time was investigated. A strong
relationship was found between soak time and cold start duration. The fitted model is as follows:
CO = 0.175 - 0.00083 x coolant + 0.0013 x speed + 0.0002 x power - 1.197 x £t_2
where:
coolant = coolant temperature, degrees F
speed = vehicle speed, mph
Power = power demand, mi2/h2.sec
Sk = error term
The R2 for the relationship between the predicted cold start CO emissions and observed
ones is 0.33, for NO is 0.53, and for HC is 0.09. The R2 for the relationship between the soak
time and cold start duration was 0.43. For the prediction dataset the cold start duration can be
determined using the regression:
y = 52.249Ln(x) - 74.771
October 10, 2002 8
-------
where:
y = cold start duration, seconds
x = soak time, minutes
With the cold start duration, the coolant temperature can be assumed for use in the above CO
emissions prediction model using similar trip data, including ambient temperature and soak time.
Uncertainty Analysis. Residuals from each regression equation fitted for the observed and
predicted data were obtained. Coefficient of variation of residuals were then determined by
dividing standard deviation of residuals with the average observed values of emissions.
Uncertainty in the observed data was also determined by estimating average and variation of
individual modes for each trip. Trip averages and variations were then estimated using weighted
averages. Weights were based on the time spent in each mode. Coefficient of variation was then
estimated by dividing standard deviation with the average value.
Heavy Duty. NCSU took similar steps as for the light duty, for QA/QC, preliminary
analysis and visualization of the bus data. Since there is no cold start for the bus data, there were
only four modes to separate: idle, acceleration, deceleration, and cruise. The modal emissions
analysis results suggest, as in the light duty, that the modal definitions assumed were reasonable.
HC shows little variability among the four modes. After analyzing the results from the HTBR
analysis it was determined that acceleration equal to 2 mph/sec should be used as a cutoff point
for the acceleration mode for CO emissions. No reasonable cutoff was found for the other
pollutants. The acceleration mode was found to contribute the most emissions for the buses as
well as for light duty, see Figures 5-7.
Nonroad. The five possible explanatory variables for the nonroad equipment were
exhaust flow, engine RPM, ambient temperature, barometric pressure, and relative humidity.
Exhaust flow is a surrogate for engine load and was found to be highly correlated with both
pollutants, NO and CO2 for all three pieces of equipment based on the data visualization
techniques used.
A regressions with time series model and a modal binning approach were both explored
for the nonroad data. In the modal approach the data was binned based on the exhaust flow rate.
OLS and multiple OLS regression models were also investigated, even though the data was
known to be autocorrelated. In the end the modal approach seemed to have the largest degree of
explanatory power, also having the advantage of being the simplest and reducing the influence of
autocorrelation in the data by dividing it into segments.
Database Lookup of Individual Vehicle and Trip Results
UCR College of Engineering-Center for Environmental Research and Technology (CE-
CERT), pursued a database approach, deriving separate emissions for macroscale, mesoscale and
microscale based on a database lookup of individual vehicle and trip results. This methodology is
a modification of the third suggested approach identified in EPA's April 2001 issue paper.
October 10, 2002
-------
CE-CERT explored three conceptual approaches for this project: multivariate statistical
equation-based, driving summary statistic, and a hybrid database modeling methodology. In the
multivariate statistical equation approach emissions are estimated using statistical relationships
between the measured variables such as speed and measured emissions. This approach was
dropped from consideration because it was found to have problems with prediction errors when
used on vehicles whose driving behavior was at or beyond the range of the behavior observed in
the training sample used to develop the model. In the driving summary statistic approach
emissions are estimated by correlating driving summary statistics with emissions. Driving
summary statistics are calculated from readily available trip information and are designed to
measure important trip characteristics, such as average speed. This approach was used in the
preprocessing of the macroscale predictions for the hybrid database model, however, it was not
used on its own because the precision of the estimates varied considerably between types of
vehicles. In the hybrid GIS/database approach emissions are estimated directly from data in a
database. Pre-processing, or hybridization, of the data is necessary to facilitate matching of the
driving segments to be predicted with the best available driving segment in the database. This
approach was selected for further development because it was simple in concept and would
provide the greatest ease of expansion and easy incorporation into a Geographical Information
System (GIS) framework.
The challenge for the database approach is the near infinite number of combinations of
conditions that must be matched for accurate emissions prediction if an exact match is required.
CE-CERT's solution to this problem was to conduct a hybridization of the basic approach. This
uses preprocessing of the data in combination with statistical "maps" to identify the closest
driving data to that to be modeled. The implementation of the model is conducted differently at
the micro-, meso-, and macro- scales because a greater degree of matching can be obtained
within the existing data for the smaller time-scale events. It is easier to match a particular modal
event than it is a portion of a trip or an entire trip.
This methodology depends on the matching of the existing emissions measurements to
the operating conditions in the prediction dataset. CE-CERT preprocessed the data then used the
GPS-based location data to identify roadway type and to obtain accurate grade information. They
then identified the factors about a driving segment which affect emissions, and then matched
these factors to the driving segments in the prediction dataset.
The data set used for matching was selected from the vehicles that most closely match the
vehicle to be predicted. The optimum situation would be for the database to contain several
vehicles having the same mileage and options as the vehicle to be predicted. In this pilot project,
vehicles were selected based on the judgment of the research team for those most likely to have
similar emission rates and emissions behavior over the observed operating conditions. In an
automated implementation the matching methodology would vary by vehicle technology type.
At the microscale level, the driving traces were disaggregated into modal segments
encompassing accelerations, decelerations, steady-state cruises, etc. At the mesoscale level, the
driving traces were disaggregated into roadway links. At the macroscale level, the driving traces
were used in a trip-based manner, see Figure 8. Consistency of emission rates is maintained through
October 10, 2002 10
-------
the use of the same basic data for each level of the model. Emission rates are estimated by querying
the database to find a driving condition similar to the one being estimated based on vehicle, roadway,
and driving behavior characteristics. For the macro-scale level, regression on principal components
analysis is used to identify groupings of variables that are correlated with emissions to simplify the
search process.
In this hybrid approach, data from the on-board emission measurement units were used to
build up a database of emissions traces in a spatial framework that can be used for on-road based
emissions estimates as well as for larger area estimates.
Microscale. The first step is dividing the trip to be predicted into seperate modal events,
see Figure 9. For this project they were visually divided, but if this methodology will be
developed for MOVES it would have to be automated. Initially the modes were divided at their
end points, however in the case for acceleration events the matching driving trace frequently did
not end at the correct speed, see Figure 10. Differences in emissions were found between
accelerations peaking and similar accelerations that did not peak at the end of the segment. As a
result, acceleration modes included the peak inflection point to ensure that the matching trace
was the closest event that ended at the same speed.
Each modal segment is matched for speed, acceleration, and power. CE-CERT's
methodology matches each modal segment to all possible length segments within the prediction
dataset. A moving window of the same number of seconds from the prediction data set is
compared with the driving trace to be matched. Match scores are calculated for each of the
matching segments using three weighted criteria: the sum of the squared difference in speeds for
each second, the sum of the squared difference in accelerations, and the sum of the squared
difference in grade across the modal event. A weighting of 80, 10, and 10 for speed, acceleration
and grade was used, which was determined empirically.
Trips were divided into cold and hot operation sections. Matching was done accordingly
using a soaktime regression, with the prediction data set. Once segments were matched, the
grams/second emission rates were summed for the trip to calculate total grams/mile.
Mesoscale. The limitations of the small dataset were magnified with the aggregation from
modal segments to roadway segments. There was much more variability introduced due to
driving behavior and terrain covered. A multivariate classification for each event based on the
physical characteristics of the roadway, such as length, average grade, maximum grade, posted
speed, etc., was established.
With the GPS information provided with the On-board emission measurement data, CE-
CERT was able to determine the location of road links. From this they were able to break down
the trips into roadway links and determine the roadway characteristics for the analysis. A
database of average speed, average grade and average total emissions in grams was created by
direction for each link. It will not be possible to collect multiple sets of On-board emission
measurement data for multiple vehicles on even a small portion of the roads in a particular area.
October 10, 2002 11
-------
It should be sufficient to find a similar set of driving conditions in the On-board emission
measurement database that does have sufficient emissions estimates for characterization.
Typically the emissions which are to be predicted are on a specific road link which does
not exist within the database. This requires matching the road link characteristics, in addition to
the appropriate driving behavior for those vehicles on the correct links. Figure 11 shows an
example, with the gray line representing links with existing data and the red line the trip to be
predicted. The emissions for portion of the trip that overlaps Links 9, 1 and 2 will be predicted by
using the values from a similar vehicle in the database with similar driving behavior over each
link. The remaining section of the trip would be divided by roadway type, and the database
would be searched to find a link that best matches the characteristics of each section with a
similar vehicle.
Macroscale. A similar approach of sorting driving characteristics is used at this level. For this
scale the whole trip is defined not just an individual roadway section. Because of the increased
aggregation, it is even more critical for the summary statistics to have descriptive power. Trip
summary statistics were calculated and used in a stepwise regression against all of the pollutants
to determine which were the best descriptives. This excercise identified ten significant driving
summary statistics, see Table 4. A principal components analysis was conducted on a subset of
the driving statistics, based on these results. Soak time, cycle length and driving distance were
added to the set of driving characteristics because of their effects on emissions.
This approach identified five significant factors within the data. The first factor accounts
for 49% of variability between trips. It is primarily a factor weighted high for cumulative
variables such as sum velocity > 0 and sum acceleration > 0, etc. The second factor, which
accounts for 11%, is more heaviliy weighted on the higher power and acceleration variables. The
third through fifth factors account for 6 or 7% each. The third is primarily a function of mean
acceleration and mean grade; the fourth - loads heavily on the summaries of the higher power
events; the fifth - a mixture of the deceleration summary, soak time and mean specific power.
Further analysis narrowed this to factors one and four for CO2 and NO, and factors one and two
for CO and HC, see Table 5. These were plotted on an XY plot of the corresponding principal
components scores.
The emission rates for each of the test trips were then calculated using the average
emisison rate of the three closest test trips on the appropriate plot. For this pilot study, all of the
vehicles were used to give a robust data set for the principal components analysis without
screening for similarity to the test vehicles. If this methodology were to be implemented it would
likely give better predictions if the matching is done only against similar vehicles.
Heavy Duty. The procedures for predicting the bus data set were similar to the light duty
for all three scales.
Nonroad. Only micro-scale predictions were made for the off-road equipment due to the
lack of road link and trip events. The modal events were matched using exhaust flow data instead
of speed, which was less smooth than the speed data but proved to be sufficient.
October 10, 2002 12
-------
Microtrips
ENVIRON's basic approach was to divide the second-by-second driving from the On-
board emission measurement data into a series of microtrips. The microtrips were intended to be
sufficiently short to describe driving events for micro-scale analysis, while providing a means to
scale upwards to meso-scale and macro-scale modeling. Their approach attempts to avoid or
minimize errors from timing offsets of recorded emissions and vehicle behavior. This approach
can be looked at as a method for data filtering or smoothing. The microtrips were expected to
reduce errors associated with single point estimates from the second-by-second On-board
emission measurement data. This methodology falls outside of the three basic approaches defined
by the April 2001 EPA Issue Paper.
Environ based their approach on a calculation of vehicle specific power (power per unit
mass, or VSP, Jimenez-Palacios11), aggregating results over driving events defined by periods of
stable operation. The first step in their process, they did include a mass term in the calculation of
power:
Power, = (a + b * Speed, + c * Speed/) * Speed, + 0.5 * Mass * (Speed,2 - Speed,2)
+ Mass * g * grade * Speed, + Auxiliary Power
The coefficients a, b, and c were supplied by EPA for the light-duty vehicles in this study
and those for a transit bus were supplied by West Virginia University (WVU)12. See Tables 6, 7.
The concern about second-by-second emissions correlations was that emissions and
vehicle behavior need to be exactly matched in time or errors could occur in determining average
emission rates. Depending on the operation of the vehicle, offsets in load and recorded emissions
could occur and be variable during driving operation. As shown in Figure 12, there may be a lag
between the calculated load determined from the vehicle speed and grade, and the emission
measurements. Even correcting for the lag may not entirely account for all conditions and lags in
the detector response labeled "distributed response" in Figure 12.
Another reason for using a microtrip approach was to avoid data smoothing requirements
of noise signals transmitted through various speed and grade signals of vehicle behavior and
emissions results. With or without a mircrotrip approach, data smoothing of the load and
emission signals should be considered when analyzing the data to avoid instances where noise or
outliers could affect the results.
ENVIRON developed a microtrip search program to determine microtrips during hot
running conditions. Start periods were selected out and evaluated separately. The microtrips
themselves were defined with minimum length and an end point criteria. Emissions and
explanatory variables were recorded and averaged over each microtrip.
Start Period. The start emissions were determined from the difference in the hot running
emissions and those measured within the first 200 seconds. The criteria of 200 seconds for start
emissions were determined from Singer13. After 200 seconds the vehicle operation was
October 10, 2002 13
-------
considered to be in a hot running mode. The emissions during the first 200 seconds of operation
after a start were predicted with the hot running emission estimates, and start emissions were
determined as the difference between the actual and predicted emissions. This difference was
then compared to the soak time. ENVIRON recommended that the start period end point be
defined with a direct measurement such as engine, exhaust, or catalyst temperature.
Microtrip Criteria. The criteria for choosing a microtrip is schematically shown in Figure
13 where the beginning and end points were determined using the criteria of constant load
(steady-state condition) so that small time differences and detector response between emissions
and load would not be carried into the emissions correlations. The load was specified to remain
constant to within, for example, +/- 5 - 15 hp over the course of 3 or 5 seconds. (The speed time
stamp was corrected by 3 seconds for this bus (Bus #10) in order to capture, for instance, the
large CO increases during the initial part of the acceleration event demonstrated in Figure 13).
Explanatory Variables. Average load over the microtrip was the primary variable used to
distinguish the microtrips and emissions, however other variables were tested to explain the data
variance. Averages of the following were compiled over the microtrips: 1) load increases over
the microtrip - to distinguish between highly transient driving and steady-state driving trips of
similar average load; 2) previous trip load - to determine "memory effects" associated with the
operating temperature of the engine and emission control system; 3) agressive load events
causing enrichment - did not occur in this data set therefore did not use; 4) malfunction indicator
light - only occurred on one vehicle which did not exhibit unusual emission behavior, therefore
not used; 5) ambient temperature and humidity - only a limited range for this dataset therefore
did not indicate importance and did not use; 6) load vehicle weight relationship - was not useful
for such a small dataset, but could be important when looking at a whole fleet of vehicles.
For light duty, buses, and nonroad equipment the hot running emissions were predicted
and compared with the start emissions to predict the start emissions. The hot running emissions
predictions were determined through stepwise regression using load as the primary variable and
adding additional variables as they improve the fit. For all cases the emissions were transformed
to make the variance evenly distributed across mean load.
Light Duty Vehicles. Two separate modes were defined, idle or no load conditions, and
the microtrips with positive mean loads. Load was found to be a reasonable predictive variable
for CO, NO and CO2. HC results indicated more scatter, possibly due to noise associated with the
lower detection limits. In order to put all vehicles in the same terms, specific load, or "spload" -
the wheel load divided by vehicle weight - was used to develop equations describing the
emissions during microtrips. There were no enrichment conditions found in the dataset which
would cause the emissions to spike, possibly due to limited driving or vehicles tested.
Below are the equations used to describe the emissions during the microtrips, with the
regression coefficients in Table 8. A weight term should have been used to describe CO2
emission predictions, but since the weights did not differ significantly in this dataset, the error
was ignored.
October 10, 2002 14
-------
Sqrt(mean.hc) = intercept + A * sqrt(mean.spload) + B * spload.changes +
C * sqrt(prev.mean.spload)+ D * mean, humid + E * mean, temp
Log(sqrt(mean.co)) =intercept + A * sqrt(mean.spload) + B * spload.changes +
D * mean, humid
Sqrt(mean.no) = intercept + A * mean.spload
Log(mean.co2) =intercept + A * log(mean.spload)
The relationship between start emissions and soak time was established, and used to
predict start emissions of the unknown validation vehicles.
Heavy Duty. Similar regressions were made with the bus data, as for light duty vehicles,
using load as the primary descriptive variable. Starts were not modeled separately since very few
buses had any data on start emissions.
Nonroad. The microtrips were defined according to the exhaust flow instead of load since
the equipment load and engine speed were included in the dataset to be predicted.
VSP Binning
EPA pursued an approach based on the binning of vehicle specific power on a second-by-
second basis using pre-defined equations developed for characterizing Vehicle Specific Power, or
VSP. The conceptual differences between the NC State method and the VSP binning method
relate only to how operating modes are defined; the conceptual difference between the Environ
approach and the second-by-second binning method are fundamentally the amount of time
emission results and VSP are aggregated over.
VSP is generally defined as power per unit mass of the vehicle. The calculation of
absolute power generally centers on the forces a vehicle must overcome when operating on the
road, including: acceleration, the force of gravity due to positive road grade, tire rolling
resistance, and aerodynamic drag. Normalizing this power by mass to calculate VSP allows for
this metric to be estimated based only on the instantaneous speed of the vehicle and road grade,
if assumptions for the coefficients of rolling resistance and drag are made and no wind speed is
assumed. Jimenez-Palacios developed the following equation for calculating VSP for light duty
vehicles using this approach:11
VSP (kW/Metric Ton) = v[l.la + 9.81 (atan(sin(grade)))+0.132] + 0.000302v3
where v in m/s
a in m/s2
1.1 = coefficient of equivalent mass for rotating components
9.81 = acceleration of gravity (m/s2)
October 10, 2002 15
-------
0.132 = rolling resistance coefficient
0.000302 = drag term coefficient
The basic concept of Vehicle Specific Power has been applied in numerous studies in
different forms, and has been shown to be a useful metric for characterizing vehicle emissions.
As discussed by Jimenez-Palacios, similar metrics have included "Positive Kinetic Energy"
(PKE) proposed by Watson et al.14, and "Specific Power" employed in studies including EPA's
FTP Revision Study15 and the development of MOBILE6 driving cycles16, which were based on
the product of speed and acceleration without consideration for road load effects. Both Environ
and North Carolina State University employed different forms of the specific power concept as
part of their shootout analysis.
For the bus analysis, the general concept of Vehicle Specific Power was used but was
based on work performed by West Virginia University (WVU) for heavy-duty vehicles.17 The
WVU work presented an equation for absolute power which took the same basic components of
acceleration, the force of gravity due to positive road grade, tire rolling resistance, and
aerodynamic drag. Coefficients for the road load components were presented specifically for
transit buses. To generate VSP from this absolute power equation, we divided through by the
curb weight of the buses (12.02 metric tons) to develop the following equation:
VSP (kW/Metric Ton) = v[a + 9.81 (sin(grade))+0.0094] + 0.42v3
where v in m/s
a in m/s2
9.81 = acceleration of gravity (m/s2)
0.0094 = rolling resistance coefficient
0.42 = drag term coefficient
The approaches for analyzing the light-duty and heavy-duty vehicles were essentially
similar, with a few minor differences. VSP was calculated using the equations above for every
second on data for the 12 vehicles in the modeling dataset using the second-by-second inputs of
speed, acceleration and road grade. For NOx, only 11 light-duty vehicles were used. Vehicle 12
was not included because the data was indentified by Sensors as erroneous. The raw VSP results
were then assigned to VSP bins in increments of 1 kW/ton from -15 kW/Ton to +30 kW/Ton for
light-duty, and -30 kW/Ton to +30 kW/Ton for heavy-duty. All VSP values outside of these
boundaries were assigned to the "boundary" bin. The boundary bin levels were chosen as they
were so that a large set of data seconds (nominally at least 100 seconds) were contained in the
highest and lowest bins, in order to reduce the noise of a low number of data points beyond these
boundaries.
Light Duty Analysis. For the light-duty analysis, because there was variability in vehicle
parameters and operating conditions, additional variables were evaluated in order to improve the
explanatory power of the model. In keeping with the MOVES design concept of fleet and activity
bins, these additional variables were characterized as "bins"; bins were developed for mileage
(less than 50,000 miles or greater than 50,000 miles), soak time (less than 1 hour, 1 to 4 hours,
October 10, 2002 16
-------
greater than 4 hours), number of cylinders (4, 6), transmission type (manual/automatic), air
conditioner compressor status (engaged or not engaged). An additional segregation was made in
the light-duty dataset between start and running mode; the first 505 seconds of each trip was
defined as a start, a number chosen for consistency with Bag 1 of the FTP.
A General Linear Model (Univariate) was run in the SPSS software package with gram
per second emission rates (CO2, HC, CO and NOx) as the dependant variables and the following
independent variables (specified as covariate): VSP bin, mileage bin (less than or above 50,000
miles), start or running mode, number of cylinders, transmission type (automatic or manual), A/C
compressor status (on or off). Because of the large sample size, all variables were determined to
be significant; therefore, partial R2 and parameter estimates were used as the primary screening
criteria for determining additional "binning" variables. Partial R2, calculated from Type I sum of
squares, is an indication of how much each variable contributes to the total explained variability.
The parameter "range" expresses how much the dependent variable varies from the mininum to
the maximum value of the independent variable; it therefore provides another means of
comparison the relative importance of each variable. The results are shown in Table 9. Using
these metrics, VSP was determined to be the most important explanatory variable of those
analyzed for CO2, CO and NOx; it was less effective for describing HC, but was included in the
analysis to provide continuity in the modeling approach. Other binning parameters chosen for
inclusion were number of cylinders for air conditioning status ( whether the compressor was
engaged or not) for CO2 and number of cylinders, mileage, and start/running mode for CO, NOx
and HC. For start modes, an additional bin parameter was added based on soak time (less than
one hour, between 1 and 4 hours, and greater than 4 hours).
Once the important binning variables were defined for each pollutant, emission rates in
grams per second were developed for each bin combination. For CO2, unique emission rates were
calculated in bins defined by VSP, cylinder and air conditioning status by calculating the average
value over all seconds of operation falling in a given bin, across each of the 12 modeling dataset
vehicles. For HC, CO and NOx separate emission rates were calculated for running and start.
Running emission rates were generated by the VSP, cylinder, mileage and soak time bins. Start
emission rates were generated by the VSP, cylinder, mileage and soak time bins, and are
expressed as the difference in emissions between start and running (i.e. start "increment") by
VSP bin. The resulting emission rates are shown in Figures 14 through 23, with the start
increments shown in Figures 24 through 33..
A Microsoft Excel spreadsheet was developed to enable the prediction of total trip
emissions for each of the light-duty validation trips, for each of the pollutants. This spreadsheet
required the necessary vehicle and operating information to choose the correct emission rates
according to bin. The vehicle information required as input are the number of cylinders and
mileage bin. Trip information also required include VSP bin distributions. The VSP information
needs to be as a total and for the first 505 seconds to compute start emissions. Other information
on trips required are the soak times, A/C on/off status, and the total trip time.
Heavy Duty Analysis. VSP was the only variable considered for the heavy-duty analysis.
This was considered appropriate because most of the vehicle trips were after very short soak
October 10, 2002 17
-------
periods, hence start emissions were not significant; all of the buses were the same in terms of
engine parameters and technology; and all buses were of similar mileage level (generally between
200,000 and 300,000 miles).
Using vehicle speeds and the vehicle position, the distance traveled per second could be
determined along with the grade. Data consistency and physicality checks could be made
between and on the measured variables. Once the data checks were made the power was
determined using the equation given above for each individual bus. Figures 34-41 display the
results for bus 2 and 10.
The power per unit mass relationships with each of the four pollutants were then
averaged together to produce final emissions per second functions for CO, CO2, NOx, and HC.
They are displayed in Figures 42-45. A key difference between the bus analysis and light-duty
analysis was that the bus emission factors were based on the average gram per second results
across each vehicle in a given bin; the light-duty results were based on the average gram per
second results across each second in the bin. The bus methodology did not account for how much
time a given bus spent in the bin; all vehicles were weighted equally.
Nonroad Analysis. Nonroad emission rates were developed simply by taking a straight
average of the three-hours worth of operation for each equipment piece, leaving out negative
emission values. This method does not account for any variability in activity during the time
period, and hence represents a grosser level of aggregation than the current NONROAD model,
which accounts for activity through the application of load factors and transient adjustment
factors. The resulting emission rates are shown in TablelO.
RESULTS AND CONCLUSIONS
Trip-by-trip results were compiled for the methods presented in the contractor reports and
those developed by EPA as presented in this paper. For the on-road analyses, these methods are
characterized as follows: the EPA approach is termed "VSP Bin", the NCSU approach is termed
"Modal Bins/OLS", the Environ approach is termed "Microtrips", the UC Riverside approaches
are termed "Microscale Database", "Mesoscale Database", and "Macroscale Database".
The complete trip-by-trip results for each source (light-duty, heavy-duty, off-road),
method, and pollutant are presented in Appendix A, Table 1, along with the results averaged
across the six trips by light-duty/heavy-duty and pollutant. The focus of MOVES will be to
predict emissions at a more aggregate level than individual trips; even emission produced at the
finest level, microscale, is defined for MOVES as 15 minutes at a specific location. Under this
approach, all emission predictions within MOVES would be based on multiple vehicles
operating at a given place for a given period of time. As a result, the analysis of results based on
trip averages is most pertinent, and our conclusions regarding model performance are based on
the more aggregate results.
On-Road Results
October 10, 2002 18
-------
Overall model performance was first judged by looking at all pollutants and all sources
simultaneously; a promising method for estimating emissions must be robust across pollutant and
vehicle type. The absolute percent difference in trip-average emissions was computed and
averaged across the four pollutants for light-duty and heavy-duty, with results shown in Figure
46. The green dots on the chart represent results averaged across light-duty and heavy-duty; the
bars are ranked from lowest to highest based on this average.
Individual pollutant results are shown in Figures 47 through 55. In the first chart for each
pollutant, the trip mean results for each method compared with the actual trip mean results, with
a 95 percent confidence interval determined based only on the individual trip results. This
confidence interval only accounts for the variability between trips, not the error in the prediction
methodologies used to estimate the individual trip results. The spread of these bands is therefore
smaller than if within-trip variability were also characterized. For the second chart for each
pollutant, the percent difference from the trip mean is shown.
The primary conclusions drawn from the results presented in these figures are the
following:
1) Uncertainty intervals for the trip mean predictions overlapped with the uncertainty bands of
the actual trip mean predictions for nearly all methods, pollutants and (on-road) sources. Because
these uncertainty bands were conservative in that they only accounted for trip-to-trip variability
but not within-trip variability, this suggests that all of the methods would likely be successful in
predicting the observed trip averages within the uncertainty bounds.
2) The VSP Binning and Microtrip approaches, which both employed vehicle specific power
(VSP) as the fundamental explanatory variable, performed the best based on the percent
difference from the trip average across all pollutants, and across each pollutant individually. In
particular the VSP binning approach had the most consistent performance when judged in terms
of the individual pollutants, predicting HC, CO and CO2 for both light-duty vehicles and buses
to within 10 percent or less. From the performance of these approaches we conclude that VSP is
an excellent metric for characterizing vehicle emissions.
3) The Modal Bins/OLS approach predicted within 20 percent of the trip means across all
pollutants, on average. From the performance of this approach and the VSP Bin approach, we
conclude that modal binning approaches are promising.
4) UC Riverside's database approach performed well for buses, in which there was little
variability in vehicle parameters; but for light-duty vehicles, performance suffered for lack of
data to provide adequate coverage across vehicle parameters such as mileage and engine size.
The one exception was light-duty NOx, as discussed in conclusion (5).
5) Light-duty NOx performance was generally not as good as the other pollutants, with the
exception of the database approaches. UC Riverside and NCSU determined in the course of their
analyses that the validation trips had trip characteristics on outer edge of trips contained in the
modeling dataset, and/or ambient conditions different from the modeling dataset. It is therefore
October 10, 2002 19
-------
hypothesized that the database approaches were able to more closely match anomolies in the
validation dataset using individual vehicles compared to the VSP Binning, Microtrip and Modal
Bins/OLS approaches, which rely on predictions of average emission rate. It is likely that model
performance for the non-database approaches would improve if the models were applied to
independent observations generated under conditions within the bounds of the modeling dataset.
Off-Road Results
The off-road methodologies were termed as follows: EPA, "Straight Average"; NCSU,
"Modal Bins"; Environ, "Microtrips"; UC Riverside, "Microscale Database". Similar to the on-
road results, the off-road models were judged based on the ability to predict total emissions for
the three hours of operation for the validation dataset, across the three equipment pieces for off-
road. Figure 56 shows the summary results, averaged across CO2 and NOx. CO2 and NOx results
are shown separately in Figures 57 and 58.
The primary conclusions drawn from these results are the following:
1) All approaches performed relatively well; within 10 percent for the average of CO2 and NOx,
and within 15 percent for each pollutant. This was likely because the validation dataset was
comprised of the same pieces of equipment as the modeling dataset.
2) In particular it is important to note that the relatively simplistic aggregate approach performed
well compared to the more sophisticated approaches. This suggests that for off-road there is less
need or less value added for adopting finer-scale modeling approaches.
APPLICABILITY OF SHOOTOUT APPROACHES IN MOVES
The predictive results of the shootout indicate that all of the methods could likely be
developed to produce accurate methods of modeling using on-board emission data. This is
underscored by the fact that nearly all of the methods predicted to within the uncertainty bounds
of the trip mean. With more data and additional refinement, it is likely that each of the methods
could be furthered to produce reasonably accurate emission rates for MOVES. Given this,
consideration of which method is "best" for MOVES centers on an evaluation of how well the
methods fit within the MOVES design objectives, summed up by the following questions: Can
the method be applied consistently across the analysis scales? Can it be easily updated as new
data becomes available? Can it adapt data from a number of sources, including possibly
laboratory second-by-second data, bag data, inspection/maintenance program data and remote
sensing data? Would a software implementation of the approach be efficient? Each of these
questions is considered in relation to the three basic approaches developed for the shootout
(modal binning, database, and microtrips).
Can the method be applied consistently across the analysis scales?
A primary goal of MOVES is to allow the prediction of emissions at multiple scales,
specifically the microscale (a specific location for a time interval as short as 15 minutes),
October 10, 2002 20
-------
mesoscale (link-level analysis for a specific region), and macroscale (county-level analysis for a
large region, up to the national scale). A key challenge to estimating emissions at multiple scales
is the desire for consistent emission rates across analysis scales. Without this consistency, we are
concerned that analysis at the different scales will produce fundamentally different results of
uneven quality.
Emission rates derived using the microtrip approach could be applied at either the
macroscale or mesoscale, but microscale applications could be problematic if the operation in the
area of study was shorter than the duration of a predetermined "microtrip" (stable period to stable
period) definitions, which was defined as 20 seconds or longer in Environ's analysis. The
maximum length of a microtrip could be very long if the vehicle did not achieve stable operation,
to the point were the period of time a vehicle spent in an intersection was shorter than a single
microtrip.
The database approach as pursued by UC Riverside generated separate emission rates for
each analysis scale, which does not meet the objective of consistency across scales. It is
concievable that the microscale database approach could be applied to all three scales, however
this could impose a computational burden which comprised the goal of software efficiency.
We believe that the modal binning approaches provide the most flexibility for application
across analysis scales. The characterization of activity as time spent in a given mode applies at all
scales. Using this approach, MOVES could provide the same set of modal emission rates without
needing to be "aware" which scale the model was operating on.
Can it be easily updated as new data becomes available?
A key recommendation of the NRC panel in "Modeling Mobile Source Emissions" was
the need for more frequent updates, to allow more responsiveness to new data as it became
available. Meeting this recommendation requires a method of producing emission rates which is
simple and can be automated as much as possible. The ultimate evolution of this concept would
take the form of a "data crank" which takes raw data and produces emission rates for MOVES
automatically.
We think all of the methods presented in the shootout could meet this criteria. However,
the purely "data-driven" approaches (database and modal binning) would likely enable this more
than any approach requiring the development of a statistical regression model, which requires
more subjectivity on the part of the model developer.
Can it adapt data from a number of sources, including possibly laboratory second-by-
second data, bag data, inspection/maintenance program data and remote sensing data?
No one data source can provide an accurate picture of emissions over the range of vehicle
operation, over the range of the in-use fleet. Typically, current emission models rely on a primary
source of data (i.e. laboratory bag) and do not take full advantage of other types (e.g. second-by-
second) or sources (e.g. I/M data) to improve the characterization of real-world emissions. A
October 10, 2002 21
-------
primary objective for MOVES is to take advantage of many types of data, from many sources, to
get a more accurate picture of in-use emissions.
Analyses have been performed which suggest that a binning approach could be applied to
any source of data; the Shootout methods which employed modal binning demonstrated how on-
board second-by-second data could be used to generate emission rates, and these methods could
easily be extended to second-by-second data sources such as lab and EVI240 data. NCSU's
shootout report also contained a proof-of-concept analysis for deriving modal emission rates
from bag data, and several remote sensing studies have used VSP binning to characterize remote
sensing data. It is not clear that the database and microtrip approaches could be applied across
this same range of data, and it is likely that emission rates based on these approaches would
require a more limited data set.
Would a software implementation of the approach be efficient?
Given the scope of MOVES, it is likely that performance will be an issue aside from the
calculation of emission rates. For example, the calculation of a county-level national inventory
over several vehicle classes, vehicle ages, roadway types, and pollutants would require millions
of calculations internal to the model. It is therefore desirable to minimize the processing of
emission rates in order to keep the performance time of MOVES manageable. The method which
causes the most concern for this is the database approach, which at the microscale level (as
discussed, the only level which could provide consistent emission rates across all scales) would
require extensive searches of second-by-second data in order to find the best "match". The modal
binning and microtrip approaches would considerably reduce the amount of data necessary for
emission rate processing compared with the database approach.
Feasibility Assessment and Next Steps
When model performance and feasibility are considered, we judge that the modal binning
approach best meets the MOVES design criteria and provides promise for model accuracy and
representativeness. It is simple, data-driven, can be extended to data from multiple sources, and is
easy to update with new data. In particular, vehicle-specific power, which combines speed,
acceleration, road grade and road load parameters, appears to be an excellent metric for
characterizing vehicle exhaust emissions. We therefore plan to pursue additional proof-of-concept
work looking specifically at the modal binning approach on a wide variety of data sources,
including laboratory second-by-second data, bag data, inspection/maintenance program data and
remote sensing data, using vehicle specific power as a means of defining modal bins. This "phase
2" work will focus on concepts presented in three of the approaches: use of VSP, which was
applied by EPA and Environ; use of binning, applied by EPA and NCSU; and evaluation of
different "averaging" times, which would take into account the basis of Environ's work on
Microtrips.
ACKNOWLEDGEMENTS
October 10, 2002 22
-------
David Brzezinski, James Warila, Carl Fulper, Carl Scarbro of EPA, OTAQ, Carl Ensfield of
Sensors Inc., Ann Arbor Transportation Authority for donating their buses for the study, the EPA
and Sensors employees who donated their vehicles for the study, and Peter McClintock of Applied
Analysis for suggesting the VSP binning approach which EPA developed.
REFERENCES
1. Koupal, I; Michaels,H.; Cumberworth, M.; Bailey, C.; Brzezinski, D.; EPA 's Plan for MOVES:
A Comprehensive Mobile Source Emissions Model U.S. Environmental Protection Agency, Ann
Arbor, MI, March 2002
2. Barm, et al., Development of a Comprehensive Modal Emissions ModelNCURP Project 25-11
Final Report, April 2000
3. Williams, et al., The TRANSIMS Approach to Emissions Estimation Los Alamos National
Laboratory Report LA-UR 99-471, 1999
4. Bachman, et al., Modeling Regional Mobile Source Emissions in a Geographic Information
System Framework Transportation Research Part C 8 (2000) 205-229.
5. National Research Council. Modeling Mobile Source Emissions. National Academy Press;
Washington, D.C., 2000.
6 EPA's New Generation Mobile Source Emissions Model: Initial Proposal and Issues. U.S.
Environmental Protection Agency, Ann Arbor MI, April 2001; EPA420-R-01-007.
7. Frey, H.C; Unal, A; Chen, J; Recommended Strategy For On-Board Data Analysis and
Collection for the New Generation Model, Prepared for EPA by North Carolina State University
Computational Laboratory for Energy, Air and Risk, Raleigh NC, February 2002.
8. Barm, M; Younglove, T; Malcolm, C; Scora, G; Mobile Source Emissions New Generation
Model: Using a Hybrid Database Prediction Technique, Prepared for EPA by University of
California Riverside College of Engineering - Center for Environmental Research and Technology,
February 2002.
9. Final Report: On-Board Emission Data Analysis andCollection for the New Generation Model,
Prepared for EPA by ENVIRON International Corporation, February 2002
10. Ensfield, On-RoadEmissions Testing of18 Tier 1 Passenger Cars and 17 Diesel Powered
Public Transport Buses Sensors, Inc. October, 2002
11. Jimenez-Palacios, J., Under standing and Quantifying Motor Vehicle Emissions and Vehicle
Specific Power and TILDAS Remote Sensing, MIT Doctoral Thesis, February 1999
12. Clark, N.N.; Kern, J.M.; Atkinson, C.M.; Nine, R.D.; Factors Affecting Heavy-Duty Diesel
Vehicle Emissions, J. Air and Waste Management Association 52:84-94, January 2002
13. Singer, B.C. et al. A Fuel-Based Approach to Estimating Motor Vehicle Cold Start Emissions,
J. Air and Waste, February 1999
14. Watson, H.C., E.E. Milkins, M.O. Preston, C. Chittleborough and B. Alimoradian, Predicting
Fuel Consumption and Emissions-Transferring Chassis Dynamometer Results to Real World
Driving Conditions, SAE Technical Paper Series, 830435, 1983
15. FTP Revision Study Federal Test Procedure Review Project, EPA Office of Air and Radiation,
EPA Report No. EPA420-R-93-007, May 1993
16. Carlson, T.R.; Austin, T.C.; Development of Speed Correction Cycles, prepared for EPA by
Sierra Research, Inc., October 1996
October 10, 2002 23
-------
17. Andrei, P.; Clark, N.N.; Lyons, D.W.; Thomson, G.; Real World Heavy-Duty Vehicle
Emissions Modeling, WVU Doctoral Thesis, 2001.
October 10, 2002 24
-------
TABLE 1 Passenger Vehicles in On-Board Emissions Study
Vehicle
Number
1*
2
3*
5
A* *
6
7
8*
11
10**
12
13
15
16
17
14
18
Oft ft
Vehicle
Company
GM
FORD
FORD
GM
CHRYSLER
GM
GM
FORD
FORD
NUMMI
FORD
FORD
GM
GM
FORD
FORD
FORD
FORD
Model
Year
1998
1997
1996
1998
1997
1999
1999
1999
1998
1999
1997
1998
1996
1998
1998
1998
1996
2000
Displacement
(liters)
3.1
3
3
1.9
2.5
3.1
1.9
2
3
1.8
2
3
2.2
2.2
2
3
3
2
Number of
Cylinders
6
6
6
4
6
6
4
4
6
4
4
6
4
4
4
6
6
4
Gross Vehicle
Weight (gvwr)
4473
4687
4707
3327
4122
4013
3237
3485
4721
3485
3485
4721
3670
0
4078
5166
4707
3485
Rated
Power
160
145
200
124
160
150
100
104
145
120
110
145
120
120
125
200
145
110
Odometer
44362
79984
96099
37278
54733
26288
43242
39429
78187
57000
71446
47439
86999
56803
29233
41319
94321
25486
'Vehicles in validation set
* * Vehicles were not used for quality
TABLE 2 Buses in On-Board Emissions Study
Contractor
Test ID
BUS380
BUS381
BUS382
BUS383
BUS384
BUS385*
BUS386
BUS379
BUS377
BUS363
BUS361
BUS375*
BUS360*
BUS372
BUS364
BUS352"
BUS404"
Enginie
Company
DETROIT
DETROIT
DETROIT
DETROIT
DETROIT
DETROIT
DETROIT
DETROIT
DETROIT
DETROIT
DETROIT
DETROIT
DETROIT
DETROIT
DETROIT
DETROIT
NAVISTAR
Vehicle
Model Year
1996
1996
1996
1996
1996
1996
1996
1996
1996
1995
1995
1996
1995
1995
1995
1992
2000
Engine
Model Year
1996
1996
1996
1996
1996
1996
1996
1996
1996
1996
1996
1996
1996
1996
1996
1992
2000
Displacement
alters)
8.5
8.5
8.5
8.5
8.5
8.5
8.5
8.5
8.5
8.5
8.5
8.5
8.5
8.5
8.5
0
0
Rated
Power
275
275
275
275
275
275
275
275
275
275
275
275
275
275
275
253
0
Odometer
223471
200459
216502
199188
222245
209470
228770
260594
252253
283708
280484
211438
270476
216278
247379
0
0
'Vehicles in validation set
* * Vehicles were not used for quality
TABLE 3 Nonroad Equipment in On-Board Study
Equipment
Type
Scraper
Compacter
Track Dozer
Vehicle
Number
BET00611
COMPACT20947
D8LANDFILL
Engine
Company
CATERPILLAR
CATERPILLAR
CATERPILLAR
Model
Year
2001
1980
1990
Number of
Cylinders
6
6
6
Ignition
Type
CI
CI
CI
Engine
Series
3406
3306
3406
Rated
Power
515
170
305
-------
CO=0.0023+0.00008*|tc-t| if ttc
r
100 200 300 400 500 GOO 700 800 900 1000 1100
Figure 1. Determination of Presence and Duration of Cold-Start using Non-Linear Regression
1000
100 -
1
H
I
.83 10
I
3
1 -
0.1 -I
HC (mg/sec)
CO (mg/sec)
rfi
NO (mg/sec)
CO2 (g/sec)
! Cold-Start
Acceleration i i Deceleration1 - ' Cruise
Figure 2. Average Modal Emission Rates for All Trips
-------
1000
Cold-Start Idle
High Low High Low High Cruise Low Cruise
Acceleration Acceleration Deceleration Deceleration
Figure 3. Improved Average Modal Emission Rates for All Trips for CO
100%
80%
60%
40% i
20%
0%
Time NO
HC
CO
C02
ID Cruise
D Deceleration
Acceleration
Dldle
H Cold-Start
Figure 4 Average Distribution of Time and Emissions with respect to Modes
-------
1000
w
fell
s
0.1
Figure 5. Average Modal Emission Rates for All Trips for HDDV
1000
idle low acceleration high acceleration decel
Figure 6. Improved Modal Emission Rate for All HDDV Trips for CO
-------
100%n
80%
60%
40%
20%
0%
D Cruise
D Deceleration
Acceleration
D Idle
Time
NO
HC
CO
C02
Figure 7. Average Distribution of Time and Emissions with respect to Modes for HDDV
Trip based
"acility specific link
)ased
VIode based
-------
Arterial
Arterid
Arterial
\
Macrosca/e
Mesoscale Microscale
Figure 8. Disaggregation of PEMS driving data for hybrid
emissions database/GIS approaches.
-------
80.
60.
Speed
(mph)
40.
20.
0
-
-
V
-
-
/\
\
i
X-\
\
\
i
1
1
i
\
1
\J
1
1 *
j
j
j
1
i
i
r\
i
j
j~~\
\
i
f
i
<-..
/
i
s.
\ ^
I
\
\
\
\
1
\
\
\
\
]
/"
1
J
\
\
\
1
1
j
1
1
1
1
1
/
1
/
-
-
-
-
-
220 240 260 280 300 320 340 360 380 400
Time (seconds)
Figure 9. Example speed trace divided into modal segments.
-------
Speed
(mph)
0 10 20 30
Time (sec.)
Figure 10. Example speed trace modally divided at the peak with matched speed segments
appearing in alternating colors.
7
Vehicle trip to be modeled Link 4
Existing Links
. L
LinkS
Figure 11. Example of link based mesoscale methodology, see text for details.
-------
Table 4 Significant driving summary statistics
Dependent Variable
CO,
CO
HC
NO
Significant Driving Statistics
Sum velocity > 0 mph
Sum velocity > 70 mph
Sum acceleration < -6 mph/sec.
Sum sp > 400
Sum grade < 0
Sum acceleration > 5 mph/sec.
Sum sp > 100
Sum sp > 50
Sum velocity > 45 mph
Sum velocity > 80
Table 5 Regression of principal components on emissions of training trips.
Regression
C02
CO
HC
NO
First PC
Factor 1
Factor 1
Factor 1
Factor 1
Second PC
Factor 4
Factor 2
Factor 2
Factor 4
R-square
.932
.625
.553
.631
Table 6. Road load coefficients for light-duty vehicles.
Vehicle
Number
1*
2
3*
4**
5
6
7
8*
9**
10
11
12
13**
14
15
16
17
18
aObf)
8.85
4.09
5.57
38.85
23.7
34.1
23.7
4.36
4.36
34.39
11.74
4.36
4.09
11.74
32.9
32.9
4.13
11.74
***
b (Ibf/mph) c
-0.05
0.09
0.01
-0.05
0
0
0
0.01
0.01
0
-0.14
0.01
0.09
-0.14
0
0
0.03
-0.14
* Validation vehicles
** No data
(M/mph2)
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
Faulty instrumentation
Table 7. Road load coefficients for a bus (WVU) and a truck (EPA).
9
-------
Source
(Hp/lbm/(ft/sec))
(Hp/(ft/sec)A3)
WVU (2000)
EPA (1995)
1.68985E-05
1.84520E-05
0
0
0.000130600
0.000124035
Table 8. Regression coefficients for light-duty microtrips above zero load and zero load rates
Pollutant
HC
CO
NO
CO,
Intercept
0.0474
-2.1729
0.0153
3.9911
A
0.2784
11.1279
12.2165
0.4864
B
1.1403
0.0615
...
C
-0.0486
...
D
-0.0003
-0.0203
...
E
-0.0005
...
...
Zero Load
(2/S)
0.000377
0.000000
0.828444
Table 9. General Linear Model, VSP Binning Approach
Variable
VSP Bin
Cylinder
Start / Running
Mode
A/C
Auto/Manual
Transmission
Mileage Bin
Temperature Bin
Humidity Bin
CO2
Partial
R2
0.6531
0.2051
0.0000
0.0004
0.0001
0.0001
0.0004
0.0003
Parameter
Range
10.24
0.78
0.04
0.19
0.00
0.07
-0.20
-0.23
HC
Partial
R2
0.0835
0.0663
0.0762
0.0000
0.0004
0.0026
0.0016
0.0001
Parameter
Range
0.0043
0.0013
-0.0035
0.0002
0.0003
0.0005
0.0007
0.0002
CO
Partial
R2
0.0854
0.0060
0.0177
0.0001
0.0001
0.0046
0.0029
0.0008
Parameter
Range
0.2411
0.0154
-0.0688
0.0063
0.0037
0.0215
0.0395
-0.0185
NO
Partial
R2
0.3301
0.0252
0.0097
0.0001
0.0000
0.0070
0.0000
0.0000
Parameter
Range
0.0260
0.0011
-0.0027
-0.0001
0.0009
0.0020
-0.0003
0.0001
10
-------
Load
Emission Response
Offset
Figure 12. Load and Emission Response Offset Diagram
Bus 10 Subset (908 - 950 seconds)
350
250
150
50
-50
-150
-250
Microtrip
Cntena of Microtnp End Point
Range of EndPoints (3-5 seconds)
Range of Load Variability (e.g. +/- 5-15 hp)
Minimum tiine of 20 seconds
12
-3
-6
-9
- Mph(Left)
Meas. Hp (Left)
- Calc. Hp (Left)
- Normalized NOx (Lft)
Normalized HC (Left)
- Normalized CO (Left)
- Fuel Rate (Rt. g/s)
Figure 13. Example Microtrip Determination
11
-------
4 Cylinder Light Duty Vehicles
c
o
i:
to
re
u>
CM
O
O
10.00
-6rSO-
4.00
-ee-
-20.00 -15.00 -10.00 -5.00 .00 5.00 10.00 15.00 20.00 25.00 30.00 35.00
Vehicle Specific Power
Figure 14
6 Cylinder Light Duty Vehicles
CO2 (grams/second)
nn
....;
" """
i'''
»*
» 6 Cyl AC Off
* * 6 Cyl AC On
I
1
-20.00 -15.00 -10.00 -5.00 .00 5.00 10.00 15.00 20.00 25.00 30.00 35.00
Vehicle Specific Power (VSP)
Figure 15
12
-------
4 Cylinder Light Duty Vehicles
o
c
o
u
50K
* _
-20.00 -15.00 -10.00 -5.00 .00 5.00 10.00 15.00 20.00 25.00 30.00 35.00
Vehicle Specific Power
Figure 17
13
-------
4 Cylinder Light Duty Vehicles
o
c
o
O
O
H»ee-
)r366-
0.200
0.100
>f*ftlft»i
»4Cyl<50K
4 Cyl AC >50K
» * »*.»*
-20.00 -15.00 -10.00 -5.00 .00 5.00 10.00 15.00 20.00 25.00 30.00 35.00
Vehicle Specific Power (VSP)
Figure 18
6 Cylinder Light Duty Vehicles
c
o
to
"in
E
ra
3
O
O
MlltJi
»6Cyl<50K
6Cyl>50K
»
..-"*'
!* * *
-20.00 -15.00 -10.00 -5.00 .00 5.00 10.00 15.00 20.00 25.00 30.00 35.00
Vehicle Specific Power (VSP)
Figure 19
14
-------
4 Cylinder Light Duty Vehicles
0.020
0.010
M6S-
..I'
-20.00 -15.00 -10.00 -5.00 .00 5.00 10.00 15.00 20.00 25.00 30.00 35.00
Vehicle Specific Power (VSP)
Figure 20
6 Cylinders Light Duty Vehicles
0.020
6 Cyl <50K
6 Cyl >50K
T3
C
o
8
0.010
-20.00 -15.00 -10.00 -5.00 .00 5.00 10.00 15.00 20.00 25.00 30.00 35.00
Vehicle Specific Power (VSP)
Figure 21
15
-------
4 Cylinder, <50K Start Increment
-20
i!:>:.5i^
00 -15.00 -10.00 -5.00 .(
»4CyK50K< 1hour
4CylAC<50K>4hours
.*v **** -, »:»»
- ,
0 5.00 10.00 15.00 20.00 25.00 30.00 35
o
Vehicle Specific Power (VSP)
Figure 22
4 Cylinders, >50K Start Increment
-20
'.I***. :
**" "I"f^"o'oo
00 -15.00 -10.00 -5.00 .(
»4Cyl>50K< 1hour
4CylAO50K1-4hours
4 Cyl >50K>4hour
«
»
n »»
* . :''
:*"
A _" .
r , - -. «\
*
0 5.00 10.00 15.00 20.00 25.00 30.00 35
o
c
o
o
Vehicle Specific Power (VSP)
Figure 23
16
-------
4 Cylinder, <50K CO Start Increment
O
O
-2C
*4 Cyl <50K< 1hour
4 Cyl AC <50K> 4hours
'.» ,
i
"*** .T***o*doo
00 -15.00 -10.00 -5.00 .(
'
. '
, -. - " -'"
* .*/
* «' .»*-,* *' |*
0 5.00 10.00 15.00* 20.00»*25»00 J30.00 35
«
00
Vehicle Specific Power (VSP)
Figure 24
4 Cylinders, >50K CO Start Increment
O
O
-20
1 000
QpnMnRp^^HY t»CTOQ» 4
00 -15.00 -10.00 -5.00 .(
4 Cyl >50K< 1hour
4 Cyl AO50K 1-4hours
4 Hyl >50K >4hnnr
»
...'
Urn " B
:-.. *. 5
:%n>"**,^-
*
0 5.00 10.00 15.00 20.00 25.00 30.00 35
Vehicle Specific Power (VSP)
Figure 25
17
-------
4 Cyclinders < 50K NO Start Increment
second)
NO (grams
fO
o
T?M:..,,IT.,f.
00 -15.00 -10.00 -5.00 .(
*4 CyK50K< 1hour
4 Cyl AC<50K> 4hours
" M
,-..-..,-.-"
> » « 4 '
0 5.00 «0.0£ B15^)0 »20*)0 25.00 30.00 35
*» "
£
^
,
*
Vehicle Specific Power (VSP)
Figure 26
4 Cylinders >50K NO Start Increment
-2C
0 050
00 -15.00 -10.00 -5.00 .(
*4 Cyl >50K< 1hour *
4 Cyl AC >50K 1-4hours
4 Cyl >50K >4hour
n
.
*** *
"A *« '"' » " *
..:.._.... ^|, * »
0 5.00 10.00 15^)0 20.00 25.00 *30.00 35
.
o
c
o
s
o
Vehicle Specific Power (VSP)
Figure 27
18
-------
6 Cylinders <50K HC Start Increment
o
c
o
o
-20
«i«B^B«.lt«*(r07ir
00 -15.00 -10.00 -5.00 .t
+
6 Cyl >50K < 1hour
6Cyl >50K 1-4hours
6 Cyl >50K >4hour
« «
* * *
*
"«.;.:'.;.-* *
0 5.00 10.00 15.00 20.00 25.00 30.00 35
Vehicle Specific Power (VSP)
Figure 28
6 Cylinders >50K HC Start Increment
o
c
o
o
0.020
0.005
6 Cyl>50K< 1hour
6 Cyl >50K 1-4hours
6 Cyl >50K>4hour
-.
-20.00 -15.00 -10.00 -5.00 .00 5.00 10.00 15.00 20.00 25.00 30.00 35.00
Vehicle Specific Power (VSP)
Figure 29
19
-------
6 Cylinders < 50K CO Start Increment
O
O
-20
-4-56
1.000
00 -15.00 -10.00 -5.00
6Cyl>50K< 1hour
6Cyl >50K 1-4hours
6Cyl>50K>4hour
*
/
,
CO
5.00 10.00 15.00 20.00 25.00 30.00 3500
Vehicle Specific Power (VSP)
Figure 30
6 Cylinders >50K CO Start Increment
o
c
o
o
o
0.400
366-
0.250
0.200
0.100
0.050
6 Cyl >50K < 1hour
6 Cyl >50K 1-4hours
6 Cyl >50K>4hour
-20.00 -15.00 -10.00 -5.00 .00 5.00 10.00 15.00 20.00 25.00 30.00 35.00
Vehicle Specific Power (VSP)
Figure 31
20
-------
6 Cylinders <50K NO Start Increment
3
o
-20
00 -15.00 -10.00 -5.00 .t
6 Cyl >50K< 1hour
6 Cyl >50K 1-4hours
6 Cyl>50K>4hour
* »
u /.....- ***. - *
* ,, *
0 5.00 10.00 15.00 B20.00 25.00 30.00 35
Vehicle Specific Power (VSP)
Figure 32
6 Cylinders >50K NO Start Increment
o
c
o
i:
JO
re
u>
O
0.050
0.04C
0.030
0.020
6 Cyl >50K < 1hour
6 Cyl >50K 1- 4hours
6 Cyl >50K >4hour
-20.00 -15.00 -10.00 -5.00 .00 5.00 10.00 15.00 20.00 25.00 30.00 35.00
Vehicle Specific Power (VSP)
Figure 33
21
-------
o
o
40 T
35--
30 --
25 --
20--
15 --
10 --
5--
rmt
*
Bus 2 CO2
M I I | I I I I |
-35 -30 -25 -20 -15 -10 -5
10 15 20 25 30 35
power per unit mass(kVWton)
Figure 34. CO2 results for bus #2
0.004 -,
0.0035 -
j/>
| 0.003 ;
0
^ 0.0025 -
0.002 -
» »0.0015 -
* ****
* * 0.001 -
* * *»»»»
* 0.0005 -
: Bus 2 HC «
:
^ ^
; ./- -".. .
;*** ^ .
: *
-
-35 -30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30 35
power per unit mass(kW/ton)
Figure 35. HC results for bus #2
22
-------
0.2
0.18 --
0.16 --
0.14 ::
0.12 --
0.1 --
'£ 0.08
0
O
O
0.06 --
0.04
0.02
*
Bus 2 CO
*
H
-35 -30
-20 -15 -10 -5
10 15 20 25 30 35
power per unit mass(kVWton)
Figure 36. CO results for bus #2
0.6 -,
S 0.5 -
c
.2
1
0 0.4-
X
O
0.3 -
0.2 -
**»*-
* '
* »»
Bus 2 NOx
: ***
; ^/'
;;*'**
. ,
-35 -30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30 35
power per unit mass(kW/ton)
Figure 37. NOx results for bus #2
23
-------
Bus10CO2
35 T
30 --
25 --
20--
15--
s
O
** **
i T + T + i T i i i
4-9-
I I I I I I I h-
H I I I I I I I I I
-35 -30 -25 -20
-15 -10 -50 5 10 15 20 25 30 35
power per unit mass(kVWton)
Figure 38. CO2 results for bus #10
0.0009 j
0.0008 ;
c 0.0007 -
0
0 0.0006 -
O
0.0005 ;
0.0004 ,
0.0003 -
0.0002 ;
* * »»V
* ^ »» 0.0001 -
** » * **
BuslOHC
: .
- »
> .
»» * * » * *
*»» *» * * * » ^
*
-35 -30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30 35
power per unit mass(kW/ton)
Figure 39. HC results for bus #10
24
-------
0.12 T
-35 -30
Bus 10 CO
-20 -15 -10 -5
10
H
15 20 25 30 35
power per unit mass(kVWton)
Figure 40. CO results for bus #10
0.18 j
,-, 0.16 -
s :
1 °14:
1 :
0 0.12-
6 :
"Z.
0.1 -
0.08 :
0.06 :
* * »*^4 "
^
» »,* * 0.02 :
i i i i I i i i i ^ i ^ i I i i i i I i i i i i i i i I i i 1 0
BuslONOx
: *
* * * *
: ******
i »
: ^
i »
: »»
» »
-35 -30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30 35
power per unit mass(kVWton)
Figure 41. NOx results for bus #10
25
-------
35 T
30 --
25 --
20 --
15 --
10 +
I i i lO
:O2 Average over all Buses
» »
H
-35 -30 -25 -20 -15 -10 -5 0
10 15 20 25 30 35
power per unit mass(kW/ton)
Figure 42. CO2 results for buses #'s 1,2, 4,5,6,7,8,9,10,11,14, and 15.
0.0025 -,
0.002 -
:/'
^ ^» 0.0015-
« ..» 3f *'
** s>
c
0
'(/)
#» -g 0.001 ;
0 .
* » * o
0.0005 -
HC Average over all Buses
: ,...... v ..
:*
* ^
*»
»
*
-35 -30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30 35
power per unit mass(kW/ton)
Figure 43. HC results for buses #'s 1,2, 4,5,6,7,8,9,10,11,14, and 15.
26
-------
0.12 T
0.1 --
0.08 --
D)
c/> 0.06
0
O 0.04
CO Average over all Buses
-30 -25 -20 -15 -10 -5 0
H
10 15 20 25 30 35
power per unit mass(kVWton)
Figure 44. CO results for buses #'s 1,2, 4,5,6,7,8,9,10,11,14, and 15.
0.4 -,
0.35 -
0.3 -
0.25 -
s
£ 0.2 -
.2
y, :
'E
o> 0.15-
X
o
0.1 -
NOx Average over all Buses
'--
- *****
- 4«
.S
:/
-35 -30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30 35
power per unit mass(kW/ton)
Figure 45. NOx results for buses #'s 1,2, 4,5,6,7,8,9,10,11,14, and 15.
27
-------
TABLE 10 Nonroad Emission Factors
Machine
Scraper
Compacter
Track Dozer
Mean
673
349
1894
Standard
Deviation
390.5
194.6
1001
N
10,695
10,542
10,800
Comments
Removed records where NOx
rate <=0
Removed records where NOx
rate <=0
28
-------
Figure 46: On-Road Summary Results - Absolute Percent Difference From Trip Mean,
Averaged Across All Pollutants
45%
40%
35%
D Light-Duty Vehicles Buses Average
VSP Bins Microtrips
Modal Macroscale Microscale Mesoscale
Bins/OLS Database Database Database
Figure 47 - On-Road HC Results: Trip Average Results w/ Confidence Intervals (Grams)
2.0
1.0
0.0
VSP Bins Microtrips Modal Macroscale Microscale Mesoscale
Bins/OLS Database Database Database
Actual
29
-------
Figure 48 - On-Road HC Results: Percent Difference From Trip Mean
50%
40%
-40%
DLight-Duty Vehicles HBuses
VSP Bins Microtrips Modal Macroscale Microscale Mesoscale
Bins/OLS Database Database Database
Figure 49 - On-Road CO Results: Trip Average Results w/ Confidence Intervals (Grams)
120
100
VSP Bins Microtrips Modal Macroscale Microscale Mesoscale Actual
Bins/OLS Database Database Database
30
-------
Figure 50 - On-Road CO Results: Absolute Percent Difference From Trip Mean
140%
VSP Bins Microtrips
Modal Macroscale Microscale Mesoscale
Bins/OLS Database Database Database
Figure 51 - LDV NOx Results: Trip Average Results w/ Confidence Intervals (Grams)
14
12
10
8
VSP Bins Microtrips Modal Macroscale Microscale Mesoscale Actual
Bins/OLS Database Database Database
31
-------
Figure 52 - Bus NOx Results: Trip Average Results w/ Confidence Intervals (Grams)
350
300
250
200
150
100
VSP Bins Microtrips Modal Macroscale Microscale Mesoscale Actual
Bins/OLS Database Database Database
Figure 53 - On-Road NOx Results: Percent Difference From Trip Mean
VSP Bins Microtrips Modal Macroscale Microscale Mesoscale
Bins/OLS Database Database Database
32
-------
Figure 54 - On-Road CO2 Results: Trip Average Results w/ Confidence Intervals (Grams)
35
30
VSP Bins Microtrips Modal Macroscale Microscale Mesoscale Actual
Bins/OLS Database Database Database
20%
15%
-5%
-10%
-15%
-20%
Figure 55 - On-Road CO2 Results: Percent Difference From Trip Mean
VSP Bins Microtrips Modal Macroscale Microscale Mesoscale
Bins/OLS Database Database Database
33
-------
Figure 56: Off-Road Summary Results: Absolute Percent Difference From "Trip" Mean,
Averaged Across CO2 and NOx
9%
8%
7%
6%
5%
4%
3%
2%
1%
0%
Modal Bins Microscale Database Straight Average
Microtrips
Figure 57: Off-Road CO2 Results: Percent Difference From "Trip" Mean
0%
-2%
-4%
-6%
-8%
-10%
-12%
Modal Bins Microscale Database Straight Average
Microtrips
34
-------
Figure 58: Off-Road NOx Results: Percent Difference From "Trip" Mean
10%
8%
6%
4%
2%
0%
-2%
-4%
-6%
-8% -L
Modal Bins Microscale Database Straight Average
Microtrips
35
-------
Appendix A:
Individual Validation Trip Results
Draft July 11,2002
-------
Trip ID
Observed Emissions (grams)
HC CO CO2(kg) NOx
HC
"VSP Bin" Predictions
CO CO2
NOx
"Modal Bins/OLS" Predictions
HC CO CO2 NOx
SRM089TR_2
3DAU86TR_3
3DAU86TR_5
RAK416TR_2
RAK416TR_4
RAK416TR_5
Trip Average
+/-
BUS385TR_3
BUS385TR_4
BUS375TR_2
BUS375TR_3
BUS360TR_2
BUS360TR_3
Trip Average
+/-
Compactor
Bulldozer
Scraper
Trip Average
2.93
3.02
6.24
1.56
1.296
0.630
2.61
1.61
3.264
2.52
2.52
2.66
2.12
2.03
2.52
0.35
36.82
35.08
45.10
36.25
19.789
10.438
30.58
10.27
51.91
40.81
58.49
52.57
54.29
42.41
50.08
5.58
10.5
5.2
8.1
2.9
3.5
3.4
5.60
2.45
17.1
10.10
16.7
18.90
17.6
12.6
15.50
2.71
89.0
90.0
73.0
84.0
5.09
7.40
8.29
10.77
4.411
4.513
6.75
2.03
252.26
143.20
253.43
290.09
286.05
203.29
238.05
44.77
334.32
1666.29
655.58
885.4
4.14
2.93
3.36
1.76
1.82
1.06
2.51
0.93
2.82
1.78
2.51
2.52
2.52
1.71
2.31
0.36
43.79
45.28
54.4
18.5
19.27
9.83
31.85
14.55
67.11
37.44
59.25
61.44
57.83
39.52
53.77
9.82
12.7
5.5
8.5
2.9
3.4
3.8
6.13
3.05
19
9.8
16
16.5
15.8
10.8
14.65
2.86
82
98
66
82.0
8.98
8.75
13.1
7.2
8.24
6.66
8.82
1.82
246
137
214
220
209
143
194.83
35.52
340
1890
670
966.7
2.62
1.55
1.99
2.65
1.34
1.06
1.87
0.53
2.9
2.3
2.5
2.6
2.8
1.9
2.50
0.29
35.2
20.2
22
58.4
24.2
15.2
29.20
12.61
87
62
80
84
83
59
75.83
9.70
12.9
5.5
8.2
3.4
4.0
3.9
6.32
2.93
17.1
11.4
15.8
16.8
15.1
11
14.53
2.15
85
84
68
79
6.1
3.4
3
6.7
4.6
3
4.47
1.30
257
164
282
328
275
210
252.67
46.29
362
1640
610
870.7
-------
Trip ID
Observed Emissions (grams)
HC CO C02(kg) NOx
HC
"Microtrips" Predictions
CO C02
NOx
"Database Micro" Predictions
HC CO C02 NOx
SRM089TR_2
3DAU86TR_3
3DAU86TR_5
RAK416TR_2
RAK416TR_4
RAK416TR_5
Trip Average
+/-
BUS385TR_3
BUS385TR_4
BUS375TR_2
BUS375TR_3
BUS360TR_2
BUS360TR_3
Trip Average
+/-
Compactor
Bulldozer
Scraper
Trip Average
2.93
3.02
6.24
1.56
1.296
0.630
2.61
1.61
3.264
2.52
2.52
2.66
2.12
2.03
2.52
0.35
36.82
35.08
45.10
36.25
19.789
10.438
30.58
10.27
51.91
40.81
58.49
52.57
54.29
42.41
50.08
5.58
10.5
5.2
8.1
2.9
3.5
3.4
5.60
2.45
17.1
10.10
16.7
18.90
17.6
12.6
15.50
2.71
89.0
90.0
73.0
84.0
5.09
7.40
8.29
10.77
4.411
4.513
6.75
2.03
252.26
143.20
253.43
290.09
286.05
203.29
238.05
44.77
334.32
1666.29
655.58
885.4
3.16
2.22
1.77
2.66
2.05
1.44
2.22
0.50
3.8
2.7
3.25
3.68
3.52
2.5
3.24
0.43
36.64
26.24
16.27
33.62
21.3
13.71
24.63
7.40
57.1
42.7
46.1
52.1
54.94
37.6
48.42
6.06
10.5
4.7
6.8
3.7
4
4.3
5.67
2.09
18.5
10.8
13.9
16.1
15.1
10.2
14.10
2.54
83.8
78.1
65.34
75.7
12.46
7.07
7.12
10.5
9
7.03
8.86
1.80
274.7
160.38
227.5
269.85
236
165
222.24
39.75
344
1547.5
597.3
829.6
1.91
2.75
4.97
5.27
4.68
2.02
3.60
1.23
2.78
2.23
2.74
3.09
2.67
1.91
2.57
0.34
12.98
45.56
40.75
137.8
98.41
32.88
61.40
37.60
69.82
52.7
46.66
46.29
68.96
40.99
54.24
9.85
11.5
6.6
9.1
3.6
4.3
4.4
6.58
2.51
18
11
14.6
15.7
16.9
11.6
14.63
2.26
83.9
84.3
67.4
79
111
3.94
3.8
11.15
10.82
6.39
7.31
2.58
211.04
127.8
152.7
189.4
165.4
119.4
160.96
28.25
371.99
1618.24
621.27
870.5
-------
Trip ID
Observed Emissions (grams)
HC CO C02(kg) NOx
"Database Meso" Predictions
HC CO C02 NOx
"Database Macro" Predictions
HC CO C02 NOx
SRM089TR_2
3DAU86TR_3
3DAU86TR_5
RAK416TR_2
RAK416TR_4
RAK416TR_5
Trip Average
+/-
BUS385TR_3
BUS385TR_4
BUS375TR_2
BUS375TR_3
BUS360TR_2
BUS360TR_3
Trip Average
+/-
Compactor
Bulldozer
Scraper
Trip Average
2.93
3.02
6.24
1.56
1.296
0.630
2.61
1.61
3.264
2.52
2.52
2.66
2.12
2.03
2.52
0.35
36.82
35.08
45.10
36.25
19.789
10.438
30.58
10.27
51.91
40.81
58.49
52.57
54.29
42.41
50.08
5.58
10.5
5.2
8.1
2.9
3.5
3.4
5.60
2.45
17.1
10.10
16.7
18.90
17.6
12.6
15.50
2.71
89.0
90.0
73.0
84.0
5.09
7.40
8.29
10.77
4.411
4.513
6.75
2.03
252.26
143.20
253.43
290.09
286.05
203.29
238.05
44.77
334.32
1666.29
655.58
885.4
2.3
2.72
3.63
2.83
1.26
0.99
2.29
0.80
2.9
1.63
3.16
3.13
1.47
1.05
2.22
0.76
128.66
39.78
146.66
38.76
19.04
35.45
68.06
43.79
75.43
46.34
77.79
85.79
47.23
39.2
61.96
15.92
7
3.4
7.9
3.8
2.9
3.2
4.70
1.74
17.4
11
20.3
24.8
11.8
11.9
16.20
4.49
12.33
3
15.86
2.83
2.94
2.55
6.59
4.74
203.12
118.66
216.65
254.31
125.65
102.88
170.21
49.96
2.85
4.3
4.06
1.42
1.71
2.98
2.89
0.94
2.87
3.12
1.84
2.83
3.14
3.49
2.88
0.45
126.12
31.1
74.7
17.5
21.16
42.7
52.21
33.33
65.64
25.81
59.56
59.02
62.68
21.89
49.10
15.80
8.1
3.9
7
3.2
3.9
4.6
5.12
1.57
19.7
16.7
17
17.7
19.1
12.9
17.18
1.92
13.06
8.82
15.85
5.09
6.15
6.3
9.21
3.47
196.6
222.9
130.35
176.8
227
142.4
182.68
32.35
-------
------- |