using mac readers for real-time arterial emissions estimates 2

Bigazzi, Kothuri, Tufte, Monsere, and Figliozzi 1

Using MAC Readers for Real-Time Arterial Emissions Estimates 1

2 Alexander Bigazzi 3 Department of Civil and Environmental Engineering 4 Portland State University 5 P.O. Box 751 6 Portland, OR 97207-0751 7 Email: [email protected] 8 Phone: 503-725-4282 9 10 Sirisha Kothuri 11 Department of Civil and Environmental Engineering 12 Portland State University 13 P.O. Box 751 14 Portland, OR 97207-0751 15 Email: [email protected] 16 17 Kristin Tufte 18 Department of Computer Science 19 Portland State University 20 P.O. Box 751 21 Portland, OR 97207-0751 22 Email: [email protected] 23 24 Christopher Monsere 25 Department of Civil and Environmental Engineering 26 Portland State University 27 P.O. Box 751 28 Portland, OR 97207-0751 29 Email: [email protected] 30 31 Miguel Figliozzi 32 Department of Civil and Environmental Engineering 33 Portland State University 34 P.O. Box 751 35 Portland, OR 97207-0751 36 Email: [email protected] 37

38 39

Submitted to the 90th Annual Meeting of the Transportation Research Board, January 2011 40 41 July, 2010 42 43 44 7304 words [5,054 + 9 figures x250] 45

46


ABSTRACT 1 This paper demonstrates the use of matched-vehicle travel times generated from MAC readers for 2 emissions estimates on urban arterials. A conceptual model is described for applying MAC data for real-3 time arterial emissions estimates. The core of the estimation method is then applied to a case study using 4 a small set of probe vehicle trajectories with simulated MAC detectors and the MOVES 2010 emissions 5 model. Results show that the accuracy of the MAC-based emissions estimates – as compared to more 6 detailed emissions estimates using second-by-second speed profiles – varies greatly by vehicle and 7 pollutant. MAC reader spacing does not have a large impact on the accuracy of emissions estimates, 8 either in terms of changing corridor lengths or changing detector spacing over a fixed corridor length. 9 Similarly, average travel speed does not affect the accuracy of MAC-based emissions estimates. The key 10 issue for accuracy here is that MAC-based emissions estimates rely on the real-world relevance of a 11 library of drive schedules to be matched to the measured travel speeds. The research results suggest that a 12 simple factor adjustment to the emissions estimates from default drive schedules can improve MAC-13 based emissions estimates with only a few probe vehicle runs. Although this study is based on a small 14 number of probe runs at one location, it provides insights into the potential for applying MAC address 15 readings for arterial emissions estimates. 16

INTRODUCTION 17 The lack of data for arterial roadways is an increasingly recognized gap in our understanding of system 18 performance (1). With travel times and speeds not widely and accurately measured, secondary estimates 19 of emissions and fuel consumption are challenging and highly uncertain. Yet urban arterials are not only a 20 major locality for motor vehicle emissions, as activity centers they are also the setting for high human 21 exposure to traffic-related pollution. Urban arterial emissions estimates are important performance 22 metrics for public health, ecology, and climate change concerns. 23

Media Access Control (MAC) address readers are an emerging tool for directly collecting 24 roadway travel time data (2-5). MAC addresses are unique identifiers broadcast over a limited spatial 25 range, most commonly encoded in Bluetoothtm-enabled wireless devices. Detection and matching of a 26 MAC address at two locations in the roadway network allows calculation of the intervening travel time 27 for the wireless device – which is assumed to be accompanied by a traveler. Travel times from MAC 28 address readings can be used to calculate various arterial performance measures. One such application is 29 emissions estimates based on real-time, real-world travel speed data – something rarely undertaken in the 30 past but with potential uses in road-user information services and pollution-responsive dynamic traffic 31 management systems. 32

This research addresses the use of matched-vehicle travel times generated from MAC readers for 33 emissions estimates on urban arterials. Included in this paper are a discussion of applying MAC travel 34 time information for real-time emissions estimates and a case study using probe vehicle trajectories on an 35 urban arterial in Portland, Oregon with the MOVES 2010 mobile-source emissions model. The objectives 36 of this study are 1) to illustrate the use of MAC-based travel times for arterial emissions estimates and 37 quantify the accuracy of such an approach, and 2) to identify the impacts of MAC reader spacing on the 38 accuracy of arterial emissions estimates. The rest of the paper is organized as follows: background 39 information, a description of the conceptual model for real-time arterial emissions estimates, case study 40 methodology and results, and conclusions. 41

BACKGROUND 42 MAC address readings are an emerging transportation data source enabled by field-ready technology and 43 undergoing operational tests in various locations (2-5). While the MAC address detection technology is 44 readily accessible, the accurate estimation of travel times from MAC readings requires algorithms for 45 matching and filtering that are still being researched. Beyond travel times, MAC address readers are also 46 being deployed to collect data on network routing and origin-destination pairings. Several ongoing 47


challenges of applying MAC readers for travel information exist, including privacy concerns, discerning 1 the mode of travel, demographic bias, and siting and spacing issues for the devices themselves. 2 Three classes of MAC readers with varying powers and operating ranges are available. Class I 3 radios can detect Bluetoothtm-enabled devices up to 100 m away (2). The ability and accuracy of MAC 4 readers to detect devices depends on power and line of sight. Typical capture rates of the MAC readers lie 5 between 0.5 – 6% of average daily traffic (2-5). These previous studies have shown that the travel time 6 estimates from MAC Readers are comparable to the estimates obtained via other methods such as floating 7 car studies, automatic license plate recognition, etc. Thus, MAC address matching using Bluetooth 8 technology represents a cost effective and reliable technique to capture travel times. With accurate 9 knowledge of the roadway distance between MAC readers, these travel time measurements can also 10 generate average travel speeds. 11

Accurate emissions and fuel consumption estimates require some knowledge or assumptions 12 about the operating speeds of vehicles in a traffic stream. Microscopic emissions estimates employ 13 detailed second-by-second vehicle speed profiles for activity data on each vehicle (see, for example U.C. 14 Riverside’s CMEM model (6)). When detailed vehicle activity data are not available, average travel 15 speeds (estimated or measured) can be used for emissions estimates using macroscopic models. Rather 16 than assuming a constant speed, macroscopic emissions models typically use average travel speed to 17 select an appropriate speed profile (speed time-series) from archived drive schedules (also known as drive 18 cycles); see, for example, the MOVES 2010 model from the U.S. Environmental Protection Agency 19 (EPA) (7), which has both microscopic and macroscopic applications. Drive schedules are archetypal 20 driving patterns for different combinations of facility type, vehicle type, and average speed. The accuracy 21 of such an approach (as compared to a microscopic model) depends on the relevance of the drive schedule 22 to local, real-world driving behavior and the accuracy of the travel speed input data (8-13). 23

Due to the lack of arterial travel time data, most macroscopic emissions modeling currently uses 24 average travel speed estimates based on measured or modeled traffic flows and relationships between 25 flow and speed – e.g. (14-16). The resulting emissions estimates vary greatly depending on the travel 26 speed estimation method (17,18). Another option is the generation of speed profiles from traffic 27 microsimulation, though that approach has not been well validated for emissions estimates because of key 28 differences from real-world driving in vehicle accelerations and unsteady cruise speeds (19-21). 29 Alternatively, MAC address readers provide the opportunity for real-time arterial emissions estimates 30 based on directly measured travel times and speeds. Given the importance to emissions estimates of 31 having accurate travel speeds, the proposed approach is an advance over current macroscopic approaches 32 because it makes it possible (a) to provide real-time or offline corridor-level emissions estimates using 33 measured speeds and (b) to utilize a single factor to adjust default drive schedules to actual corridor traffic 34 conditions. Additionally, the proposed approach is not as onerous in terms of calibration or estimation 35 time as microsimulation models because it allows the efficient utilization of drive schedule libraries. 36 Although MAC travel times are still a macroscopic data source, they offer potential improvements over 37 current methods of arterial emissions estimation. 38

CONCEPTUAL MODEL FOR APPLYING MAC-ADDRESS TRAVEL TI MES FOR ARTERIAL 39 EMISSIONS ESTIMATES 40 In this section we propose a conceptual model for real-time arterial emissions estimates based on MAC 41 matched-vehicle travel times. The general strategy is to use MAC-detected average travel speeds with 42 other local data (stored or measured) and an emissions model to generate arterial emissions estimates. 43 Figure 1 illustrates the proposed data sources, models, and estimation steps. Emissions estimates can be 44 made for road sections down to the spatial resolution of the MAC readers. Most of this approach is 45 essentially the same as other macroscopic emissions estimates, with the key difference of incorporating 46 real-time, real-world travel speed information. Detailed steps are described below, along with a discussion 47 of emissions modeling and accuracy issues. 48


Figure 1. Data flow for real-time arterial emissions estimates using MAC readers

Data Sources 1 Applying MAC-based travel times for arterial emissions estimates requires a mix of data sources. Ideally, 2 input data for an emissions model are directly measured in real time, though that is infeasible for several 3 classes of data (fuel formulations, for example). Alternatively, appropriate average data can be applied 4 from local databases, as in (14,22,23). In the approach described here, travel speeds and traffic counts are 5 measured in real-time from the roadway and other inputs come from archived data. 6

Measured

travel speed

Microscopic

emissions modelRoadway emissions

(per road-mi per hr)

Marginal

emissions rates

(per veh-mi)

Measured

traffic flow

Drive schedule library*

0

20

40

60

0 200 400 600

Sp

ee

d (

mp

h)

Time (sec)

Drive Schedule #1054

0

20

40

60

0 200 400 600

Sp

ee

d (

mp

h)

Time (sec)


0

20

40

60

0 200 400 600

Sp

ee

d (

mp

h)

Time (sec)


MAC reader 1

Traffic

counts

Roadway

MAC reader 2

* Drive schedules

selected by avg. speedStored or measured

• Veh. types

• Weather

Stored local averages

• Fuels

• Veh. fleet age

• I/M program

• etc.

speed travelAverage timeTravel

Distance =

(a)

(b) (c)

(g)

(f)

(e)

(h) (i)

(d)

** interpolate, combine by vehicle

type distribution, and average

**


Estimation Steps (See Figure 1) 1 For each vehicle (a) traveling between MAC reader 1 and MAC reader 2, the time difference between 2 detections is used (with appropriate filtering) to calculate the travel time (as in (2,4)). The vehicle’s travel 3 time is combined with the roadway distance between MAC readers to calculate its average travel speed 4 (5). Additionally, inductive loop detectors or traffic cameras at signalized intersections provide counts of 5 vehicles on the corridor (24). As an unrestricted facility with entering and exiting vehicles, traffic counts 6 at multiple detection locations between MAC readers are combined for average traffic flow rate on the 7 corridor. The result is real-time measurements of average travel speeds (b) and traffic flows (c) on the 8 road section between two MAC readers. 9

Each measured travel speed is then used to mine an appropriate drive schedule from a drive 10 schedule library (d) by matching the average speed of the drive schedule to the travel speed – or by 11 selecting the two drive schedules with the closest average speeds to the measured travel speed and 12 interpolating the final emissions rate estimates between them (the internal approach of the MOVES 2010 13 emissions model (7)). In addition to average speed, drive schedules are selected by relevant facility type 14 and vehicle type. Since vehicle types are not associated with the measured travel speeds, drive schedules 15 are selected at (d) using the measured vehicle speed and the relevant facility type with all on-road vehicle 16 types, and a marginal emissions rate is computed for each speed-facility-vehicle type combination (in 17 steps d-h). The final marginal emissions rate for each vehicle (a) is then computed as the weighted 18 average of these emissions rates, weighted by the on-road vehicle type distribution. This parallels the 19 macroscopic approach of MOVES 2010 and other average-speed models, where composite emissions 20 rates are weighted averages from on-road vehicle type distributions. 21

The selected drive schedules (d) are one type of input to the emissions model (e), required to 22 determine the activity (operating mode) distribution of the modeled vehicles. Some of the other emissions 23 model inputs are mined as average values from a local database (f), including fuel formulation, vehicle 24 age distributions, and inspection and maintenance (I/M) programs (as in (22,23)). Other inputs can be 25 measured and input real-time if the system architecture permits (e.g. vehicle type distributions from 26 length-based vehicle class detectors, or temperature and humidity from a nearby meteorological station) – 27 or can be based on local averages as well (g). The use of license plate recognition cameras tied to a 28 vehicle registration database would allow even more detailed real-time vehicle fleet data – including 29 vehicle ages and fuel types (gasoline, diesel, etc.). 30

The output (h) from the emissions model is marginal emissions rates (in mass emissions per 31 vehicle-mile) for each measured vehicle (a). The final marginal rates are computed by interpolating 32 between drive schedules (if neighboring drive schedules are used), and taking the average of vehicle-type 33 rates weighted by the on-road vehicle type distribution (either directly measured or from a local average 34 (g)). In the final step, all computed marginal emissions rates in a time interval are averaged and multiplied 35 by the measured traffic flow (c) to generate a roadway emissions rate estimate in mass per mile of 36 roadway, per hour (i). The averaging of measured-vehicle emissions rates for application to all vehicles in 37 a time interval relies on unbiased MAC reader travel time measurements – or algorithms to compensate 38 for MAC bias. 39

Emissions Model 40 The emissions model (e) using drive schedules as input is a microscopic model, such as CMEM (6), 41 MOVES 2010 project-level (7), VERSIT (25), or others. Alternatively, a macroscopic, average-speed 42 emissions model can be applied that takes the average measured travel speed (b) as input and uses an 43 internal drive schedule library (essentially, (d) is embedded in (e) – as in MOVES 2010 regional 44 estimates). The breadth and depth of accompanying input data demands in (f) and (g) depend on the 45 emissions model selected. Since emissions models are often computationally expensive, the emissions 46 estimation can be executed and the outputs tabulated for each potential drive schedule using a range of 47 local scenarios for other input data (steps d-h). Then, the tabulated marginal emissions rates can be 48 quickly mined based on real-time measured speed data to allow faster execution and more feasible 49 application in a real-time estimation environment (similar to (26)). 50


Accuracy Issues 1 The essential point to the accuracy of this approach (as in all average-speed emissions modeling) is the 2 representativeness of the drive schedules for the real-world driving activity of on-road vehicles. 3 Depending on the microscopic emissions model methodology, the drive schedules are used to distribute 4 the vehicle’s activity into operating modes of some type (for example by speed bins, a combined speed 5 and acceleration matrix, or engine loads/vehicle specific power – see (27)). Error is introduced when the 6 operating mode distribution of the drive schedule doesn’t match the real-world distribution (8). Further 7 error can be introduced by the drive pattern for models (such as CMEM) that use a time-based approach 8 to estimation (meaning the model has memory of preceding operating modes, unlike MOVES 2010, 9 which distributes activity by operating mode without accounting for sequence). Research is underway to 10 improve estimates of vehicle specific power distributions based on facility-specific average speeds (27), 11 which would reduce the error of apportioning operating mode distributions in average-speed approaches 12 such as this. 13

A particular problem for MAC-based travel times is that the measured travel speeds are not 14 associated with specific vehicle types (though this problem is shared by other macroscopic emissions 15 estimation approaches). As such, the above method does not account for average-speed bias by vehicle 16 class – although it does account for speed profile differences by using vehicle class-specific drive 17 schedules. Research is needed to determine if facility-specific average-speed correction factors can be 18 applied to adjust for average-speed vehicle class bias in MAC travel time readings. 19

CASE STUDY METHODOLOGY 20 This case study is designed to assess the accuracy of MAC-based arterial emissions estimates. Probe 21 vehicle data are used to generate ground-truth emissions estimates and to simulate MAC readings for 22 average-speed emissions estimates. The key questions are how do MAC-based estimates compare to 23 estimates using detailed vehicle trajectory data, and what are the impacts of varying MAC reader spacing. 24

Real-world second-by-second speed profiles from probe vehicles on an urban arterial are used to 25 estimate “ground truth” emissions rates (as in (28)). MAC-based travel times are then simulated from the 26 trajectory data at a range of detector spacings, and emissions estimates made using the average travel 27 speed approach illustrated in Figure 1. Comparisons between ground-truth emissions and MAC-based 28 emissions assess the accuracy of MAC-based estimates as compared to estimates using idealized vehicle 29 activity data – not compared to directly measured emissions. Also, simulating MAC readings from the 30 probe vehicle data assumes accurate MAC address travel time calculations. The post-processing of MAC 31 address readings for accurate travel time estimates is outside the scope of this paper, while recognizing 32 that emissions estimates are subject to the same ongoing issues as travel time applications (filtering 33 stopovers, multiple readings from single vehicles, demographic bias, etc.). 34

Study Location 35 The study location is SE Powell Boulevard (U.S. 26 – an urban arterial in Portland, Oregon) between 12th 36 and 82nd avenues. This 6 km corridor is a four lane undivided arterial with two lanes in each direction, 37 with left turn bays present at intersections. There are eleven signalized intersections along the study 38 corridor (at SE 21st, SE 26th, Se 33rd, SE 39th, SE 50th, SE 52nd, SE 65th, SE 69th, SE 71st, SE 72nd and SE 39 82nd avenues). The corridor is heavily traveled and serves approximately 37,750 vehicles per day. Since 40 the corridor connects downtown Portland to the eastside suburbs, the westbound (WB) direction 41 experiences heavy traffic in the morning peak period and similarly the eastbound (EB) direction has 42 heavier traffic in PM peak period. 43

For this study 17 probe vehicle runs were collected during the AM and PM peak periods using 44 passenger cars traveling in both directions (with and against peak flows). The trajectory data were 45 gathered using in-vehicle GPS devices collecting location information and differentiating for speed at 1 46 second intervals. Probe vehicles followed an “average car” data collection approach, wherein the test 47 vehicles traveled at the average speed of the traffic stream. Travel lanes were not specified, with the 48 drivers being allowed to use their own judgment for lane changing maneuvers. 49


Figure 2 shows a map of the study corridor on the left side (west is up) with a sample WB peak-1 direction probe vehicle run where each point is a one-second speed reading; green dots are higher speeds 2 and red dots are lower. The right side of Figure 2 shows the same sample probe vehicle trajectory on the 3 space-time plane, where instantaneous speed is the slope of the trajectory line. Four simulated MAC 4 readers are also illustrated in Figure 2, with equal spacing of 2 km. Each simulated MAC reader is 5 assumed to detect the probe vehicle as it passes by, and travel times are calculated by the time elapsed 6 between successive MAC readings (illustrated here between readers #2 and #3). Combined with the 7 distance between detectors, the measured travel times provide an average speed estimate between 8 simulated detectors (which is then used as (b) in the emissions estimation process shown in Figure 1). 9

10

Figure 2. Map of study corridor with sample WB (peak direction) probe vehicle run from 4:30pm on 6/10/2010 (speeds as colored dots, where red is slow and green is fast), with simulated MAC

readers and the same probe vehicle trajectory on the space-time plane (map source: Google Earth)

0

3

6

0 100 200 300 400 500 600

Dis

tan

ce (

km

)

Time (sec)

Travel time

De

tect

or

spa

cin

g

(tra

vel d

ista

nce

)

MAC

reader #4

MAC

reader #3

MAC

reader #2

MAC

reader #1

Tra

ve

l Dir

ect

ion


Emissions Model 1 For consistency, all emissions modeling in this study (both microscopic and macroscopic) uses the 2 MOVES 2010 mobile-source emissions model from the EPA (7). MOVES allows for microscopic 3 emissions estimates using the project-level analysis with custom drive schedule inputs (second-by-second 4 speeds from trajectory data) – see MOVES documentation for details (29). These “ground-truth” 5 microscopic emissions estimates are made using full run trajectories and using trajectories broken into 6 shorter segments of approximately 600 meters (with each run broken into 10 run-segments) – for a total 7 of 170 run-segments. 8

In addition to the drive schedules, required MOVES inputs include fuel formulation, I/M 9 program, and meteorology (all taken from Multnomah County defaults for January 2010), and vehicle age 10 distribution by class (assumed using national average projections from the MOBILE 6.2 emissions model, 11 EPA’s predecessor to MOVES). Emissions rate estimates are performed for 7 pollutants: CO2e 12 (greenhouse gases in carbon dioxide equivalent units), CO (carbon monoxide), HC (gaseous 13 hydrocarbons), PM2.5 (particulate matter smaller than 2.5 microns), benzene, and NOx (nitrogen oxides). 14 Each pollutant is modeled for 4 vehicle types common on urban arterials: gasoline passenger car 15 (PassCar_gas), gasoline passenger truck (PassTruck_gas), diesel passenger truck (PassTruck_diesel), and 16 diesel light-duty commercial vehicle (LtCommTrk_diesel). The modeled emissions are for January 2010, 17 on an urban unrestricted roadway facility at 0% grade. Running, evaporative, and brake/tire emissions are 18 included, but start, refueling, and extended idle emissions processes are not. 19

MAC-based (average-speed) emissions estimates are made in MOVES using project-level 20 analysis with average “link” speeds, which MOVES uses to cull relevant drive schedules from its internal 21 library. Other than the substitution of average speeds for custom drive schedules, all other model inputs 22 are the same. In this way we can isolate the emissions estimation differences due to different levels of 23 vehicle activity data (detailed trajectories versus run-segment average travel speed). As stated above, the 24 average speeds are calculated from the trajectory data by assuming MAC detection as the probe vehicle 25 passes the simulated MAC reader locations. The MAC-based emissions estimates then connect steps (b) 26 and (h) in Figure 1 above. 27

As with the trajectory-based estimates, the average-speed emissions rate estimates are performed 28 for full probe runs (2 simulated MAC readers 6 km apart – 1 at each end of the corridor) and for run-29 segments of 600 meters (11 simulated MAC readers – at equally-spaced locations delineating 10 30 segments per probe run). Additionally, in order to investigate MAC-reader spacing effects, average-speed 31 emissions estimates are performed for a range of run-segment lengths (detector spacings) from 200 to 32 3,000 meters (up to 30 segments per probe run). The MAC-based emissions rate estimates (h) are then 33 compared with the ground-truth emissions rates estimates to generate the results presented in the 34 following section. 35

RESULTS 36 Figure 3 shows the variability of ground-truth CO2e emissions rates from PassCar_gas over the course of 37 each run, separated by direction. The vertical axis units are percent difference in a run-segment’s 38 emissions rate from the run’s average emissions rate. The box-plot for each run shows the spread of 39 emissions rates over 10 run-segments of 600m each, caused by different speed profile characteristics on 40 each run-segment. Overall, the run-segment emissions rate variability was greater in the peak flow 41 direction, which involved more stops due to congestion. This variability is important when considering 42 local pollution hotspots – either related to traffic characteristics or sensitive populations. Emissions rate 43 variability for other pollutant-vehicle combinations is similar, and excluded for space economy. 44


1

Figure 3. Variability in ground-truth PassCar_gas CO2e emissions rates over 10 run-segments of 600m each for 8 runs in peak flow direction and 9 runs in off-peak direction (with departure times)

2

3

Figure 4. Percent error in total corridor emissions from MAC-based estimates at a range of detector spacings (run-segment lengths); each line is a probe run

Runs in peak dir. Runs in off-peak dir.

06/1

0/20

10 1

6:10

06/2

2/20

10 0

7:47

06/2

2/20

10 0

8:18

06/2

2/20

10 1

6:49

06/2

2/20

10 1

7:13

06/2

2/20

10 1

7:37

06/2

2/20

10 1

8:03

06/2

2/20

10 1

8:32

-50

0

50

100

Seg

men

t em

. rat

e %

diff

. fro

m ru

n av

g.

06/1

0/20

10 1

7:30

06/1

0/20

10 1

8:09

06/2

2/20

10 0

7:37

06/2

2/20

10 0

8:09

06/2

2/20

10 1

7:02

06/2

2/20

10 1

7:25

06/2

2/20

10 1

7:53

06/2

2/20

10 1

8:17

06/2

2/20

10 1

8:46

-50

0

50

100

Seg

men

t em

. rat

e %

diff

. fro

m ru

n av

g.

1000 2000 3000 4000 5000 6000

-10

0

10

20

30

CO2e - PassCar_gas emissions

Segment Lengths/Detecor Spacing (m)

Tot

al e

mis

s. %

err

or

1000 2000 3000 4000 5000 6000

-10

0

10

20

30

HC - PassCar_gas emissions


Tot

al e

mis

s. %

err

or

1000 2000 3000 4000 5000 6000

-10

0

10

20

30

PM2.5 - PassTruck_diesel emissions


Tot

al e

mis

s. %

err

or

1000 2000 3000 4000 5000 6000

-10

0

10

20

30

NOx - PassTruck_diesel emissions


Tot

al e

mis

s. %

err

or


Figure 4 shows the percent error in total run emissions rates for MAC-based estimates using 1 detector spacings from 200 to 6,000 meters, with one line for each probe run. A selection of vehicle-2 pollutant combinations is presented, though the results from other combinations were consistent with 3 these. There is a slight decrease in MAC-based emissions rates at spacings shorter than 1km, but the 4 detector spacing does not have a large effect on the total corridor error. This suggests that errors in the 5 MAC-based emissions estimates are not due to the spatial resolution of the average-speed activity data. 6

Figure 5 compares the emissions rate versus average speed relationships of average-speed 7 modeled (lines) and full-corridor ground-truth (points) emissions estimates. The probe run average speeds 8 over the full corridor range from 15 to 30 mph. Throughout this range the ground-truth estimates and the 9 average-speed estimates both follow the same curve shape, though with a slight offset. This pattern is 10 consistent across pollutants and vehicle types, though the offset size and direction fluctuates. Since MAC-11 based emissions estimates are computed using the average-speed methodology, this result suggests that in 12 addition to not being affected by detector spacing (Figure 4), the MAC-based emissions error is not 13 skewed by congestion level (as indicated by average speed or delay). 14

15

Figure 5. Emissions rate vs. average speed curves; lines are average-speed modeled and points are full-corridor trajectory-based estimates

A similar situation is seen in Figure 6, which shows the average-speed modeled emissions rate vs. 16 speed curves for PassCar_gas only (line), along with ground-truth emissions rates from 600m run-17 segments on all runs (points). Although there is more spread with more data points, the ground-truth data 18 still match the shape of the average-speed modeled line, and over a wider range of speeds. The ground-19 truth data do not trend away from the MAC estimated rates at high or low speeds, indicating again that 20 congestion level does not have a major impact on the accuracy of the method. The average-speed 21

10 20 30 40 50

0

500

1000

1500

2000

CO2e Emissions vs. Speed

Average Speed (mph)

CO

2e R

ate

(g/m

i)

PassCar_gasPassTruck_gasPassTruck_dieselLtCommTrk_diesel

10 20 30 40 50

0.0

0.1

0.2

0.3

0.4

PM2.5 Emissions vs. Speed

Average Speed (mph)

PM

2.5

Rat

e (g

/mi)


10 20 30 40 50

0

2

4

6

8

NOx Emissions vs. Speed

Average Speed (mph)

NO

x R

ate

(g/m

i)


10 20 30 40 50

0.0

0.2

0.4

0.6

0.8

1.0

HC Emissions vs. Speed

Average Speed (mph)

HC

Rat

e (g

/mi)



approach, though it does produce error, is not noticeably poorer at estimating a slow run-segment with 1 stop-and-go conditions than a free-flowing run-segment with average travel speed over 35 mph. 2

Figure 6. Emissions rate vs. average-speed curves lines are average-speed modeled and points are 600m run-segment trajectory-based estimates

The varying error bias of the MAC-based estimates over all vehicle-pollutant combinations is 3 shown as mean percent error (MPE) in Figure 7. The MPE is calculated by comparing MAC-based and 4 ground-truth emissions rates over the full 6km corridor (Total_MPE) and over 600m run-segments 5 (Segment_MPE). There is a great deal of variability across vehicle-pollutant combinations, though in 6 general the MAC-based approach overestimates emissions rates on this corridor. The highest percentage 7 errors are for pollutants with very low emissions rates (PM2.5 and NOx from gasoline vehicles). The MAC 8 method is generally more accurate for diesel vehicles than gasoline. The corridor and run-segment MPE is 9 similar for all vehicle-pollutant combinations. Although the segment-level emissions can vary greatly (see 10 Figure 3), the average-speed approach to estimation is similar in accuracy for 600 and 6,000 meter road 11 sections. 12

10 20 30 40 50

0

200

400

600

800

1000

1200

CO2e Emissions vs. Speed

Average Speed (mph)

CO

2e R

ate

(g/m

i)

PassCar_gas

10 20 30 40 50

0

2

4

6

8

CO Emissions vs. Speed

Average Speed (mph)

CO

Rat

e (g

/mi)

PassCar_gas

10 20 30 40 50

0.0

0.5

1.0

1.5

2.0

NOx Emissions vs. Speed

Average Speed (mph)

NO

x R

ate

(g/m

i)

PassCar_gas

10 20 30 40 50

0.00

0.05

0.10

0.15

0.20

0.25

0.30

HC Emissions vs. Speed

Average Speed (mph)

HC

Rat

e (g

/mi)

PassCar_gas


1

Figure 7. Mean percent error (MPE) of all runs over the 6km corridor and over 600m run-segments, for all vehicle-pollutant combinations

The errors described here arise from speed profile differences between the measured probe 2 trajectories and the drive schedules mined from the MOVES database. Speed profile/drive schedule 3 differences lead to operating mode distributions within the emissions model that do not reflect the true 4 operating mode distribution of the probe vehicles. In order to look in more detail at the sources of error in 5 these estimates, we can mine the EPA drive schedules for a range of speeds, facilities, and vehicle classes 6 from the MOVES 2010 MySQL database. Figure 8 presents drive schedule comparisons using Gaussian 7 kernel density plots of the speed distributions of all probe runs combined (a) and for 3 individual probe 8 runs with different average speeds (b - d) along with the MOVES drive schedules for passenger cars on 9 urban arterials having the closest average speeds. In (a-c) the probe vehicle runs have more distinct 10 bimodal distributions around free-flow speeds and stops, while the MOVES distributions are more spread. 11 Both the probe run and MOVES schedule in (d) had more moderate and low speeds than the other speed 12 profiles – reflective of more congested driving at the lower average speed. 13

The speed profile differences reflected in this Figure lead to the emissions rate differences show 14 in Figures 4 through 7. While the drive schedule average speed might match closely for a given run, the 15

-20 0 20 40 60 80 100 120

CO2e

CO

PM2.5

NOx

HC

benzene

CO2e

CO

PM2.5

NOx

HC

benzene

CO2e

CO

PM2.5

NOx

HC

benzene

CO2e

CO

PM2.5

NOx

HC

benzene

Pa

ssC

ar_

ga

sP

ass

Tru

ck_

ga

sP

ass

Tru

ck_

die

sel

LtC

om

mT

rk_

die

sel

Mean Percent Error of All Runs

Total_MPE

Segment_MPE


distribution of speeds can be markedly different (e.g. Figure 8c). As we see in Figure 7, those speed 1 profile differences affect different vehicle types and pollutants in different ways. 2

3

4 (a) (b) 5

6 (c) (d) 7

Figure 8. Comparative speed distributions (as Gaussian kernel density) between MOVES drive schedules and (a) all probe vehicle runs and (b – d) for three probe runs with different average

speeds; MOVES drive schedules shown with closest average speeds to probe runs

Comparing the speed distributions of the MOVES drive schedules with the probe vehicle data 8 reveals unique local driving characteristics, as could be expected of any one locality (13,30). The more 9 spread nature of the MOVES schedules in Figure 8 reflects the fact that they are archetypal driving 10 patterns designed to represent a wide variety of urban arterial roads and drivers for aggregate emissions 11 estimates throughout the U.S. For average-speed arterial emissions estimates the database of drive 12 schedules would ideally be collected locally on the arterial of interest, to be more representative. But 13 recording a unique set of drive schedules at the full range of average speeds and vehicle types for every 14 arterial in a city would be a time and cost intensive process. If not collected on the exact roadway, a 15 compromise effort would be a database of drive schedules collected on other regional arterials with 16 similar characteristics (lane configuration, intersection density, commercial activity, etc.) – though this 17 could still be a costly undertaking. 18

As an alternative approach, we propose a simple method to partially localize the emissions 19 estimates using a small amount of probe vehicle data. From the results above, the error across runs for 20 each vehicle-pollutant combination tended to have a consistent bias unaffected by detector spacing or 21 average speed (Figures 4 through 7). Thus, the proposed method applies a correction factor that is simply 22 the mean ratio of ground-truth to average-speed total run emissions rates for each vehicle-pollutant 23

Speed (mph)

Den

sity

0.00

0.01

0.02

0.03

0.04

10 20 30 40 50

MOVES_1030_25mphMOVES_1041_19mphProbe_Runs_All_22mph

Speed (mph)

Den

sity

0.00

0.02

0.04

0.06

10 20 30 40 50

MOVES_1029_31.0mphProbeRun6_29.7mph

Speed (mph)

Den

sity

0.00

0.01

0.02

0.03

0.04

10 20 30 40 50


Speed (mph)

Den

sity

0.00

0.01

0.02

0.03

10 20 30 40 50



combination. This correction can be calculated with only a handful of probe runs on each arterial, and 1 need not span all congestion levels. 2

As an illustration, we perform a random draw of 8 probe runs from the 17 of this study to 3 calculate a correction factor, then apply that correction factor to the MAC-based emissions rates of the 4 other 9 runs. The resulting MPE and mean absolute percent error (MAPE) are computed with respect to 5 ground-truth rates for the adjusted subset of 9 runs. This process was repeated with 100 random draws to 6 generate the full-corridor and 600m run-segment percent errors shown in Figure 9 (averaged across all 7 pollutants and vehicles). The corrections greatly reduce error for these data, with similar results across 8 vehicle-pollutant combinations. Correction factors were consistently in the range of 0.5 to 1.1, and mostly 9 around 0.8 – as implied Figure 7. Although this method still requires some probe vehicle runs, it is less 10 costly than developing a full set of custom drive schedules and improves the accuracy of MAC-based 11 emissions rate estimates. The correction method works here on a small set of probe runs collected over a 12 short time range with only two different drivers. Further validation is required to see if the relationships 13 hold under more varying traffic and driver conditions. 14

15

16

Figure 9. MPE and MAPE for adjusted MAC-based emissions rates, averaged across all pollutants and vehicles; spread is from 100 random draws to calculate adjustment factors

CONCLUSIONS 17 This paper demonstrates the use of matched-vehicle travel times generated from MAC readers for 18 emissions estimates on urban arterials. A conceptual model is described for applying MAC readers for 19 real-time arterial emissions estimates. The core of the estimation method is then applied to a case study 20 using a small set of probe vehicle trajectories with simulated MAC detectors and the MOVES 2010 21 emissions model. Results show that the accuracy of the MAC-based emissions estimates – as compared to 22 more detailed emissions estimates using second-by-second speed profiles – varies greatly by vehicle and 23 pollutant. 24

MAC reader spacing does not have a large impact on the accuracy of emissions estimates, either 25 in terms of changing corridor lengths or changing detector spacing over a fixed corridor length. There are, 26 however, other potential benefits of increased detector density such as improved capture and match rates, 27 and more localized emissions estimates (which can vary greatly over a corridor). Similar to detector 28 spacing, average travel speed does not affect the accuracy of MAC-based emissions estimates. For the 29 probe runs of this study, the accuracy of MAC-based emissions estimates was better for diesel vehicles 30 than for gasoline. 31

The key issue for accuracy here is that MAC-based emissions estimates rely on the real-world 32 relevance of a library of drive schedules to be matched to the measured travel speeds. As shown in these 33 results, national default drive schedules (such as contained in the MOVES model) will not always apply 34 for an individual arterial of interest. While a complete set of locally-collected drive schedules for all 35

Corridor_MPE Corridor_MAPE Segment_MPE Segment_MAPE

-5

0

5

10

15

% E

rro

r


travel speeds and vehicle types would be ideal, these research results suggest that a simple factor 1 adjustment to the emissions estimates from default drive schedules can improve MAC-based emissions 2 estimates with only a few probe vehicle runs. 3

Although this study is based on a small number of probe runs at one location, it provides insights 4 into the potential for applying MAC address readings for arterial emissions estimates. Next steps include 5 validation of these results with a broader set of probe vehicle data, establishment of a data structure for 6 compiling a regional database of locally-collected speed profiles/drive schedules, and integration of real-7 time arterial emissions estimates into the PORTAL transportation data archive at Portland State 8 University (portal2.its.pdx.edu), which is currently in the process of incorporating MAC reader data from 9 SE Powell Blvd. 10

ACKNOWLEDGEMENTS 11 The authors would like to thank Bryan Hangartner for assistance in data collection, and OTREC and the 12 U.S. Department of Transportation (through the Eisenhower Graduate Fellowship Program) for their 13 support of this project. 14

REFERENCES 15 [1] Bonnenson J., M. Pratt, and M. Vandehey, Predicting the Performance of Automobile 16

Traffic on Urban Streets, Washington, D.C.: Transportation Research Board of the National 17 Academies, 2008. 18

[2] Quayle S., P. Koonce, D. DePencier, and D. Bullock, “Arterial Performance Measures 19 Using Media Access Control Readers: Pilot Study in Portland, Oregon,” 89th Annual 20 Meeting of the Transportation Research Board, Washington, D.C.: 2010. 21

[3] Wasson J.S., J.R. Sturdevant, and D.M. Bullock, “Real-Time Travel Time Estimates Using 22 Media Access Control Address Matching,” ITE Journal, vol. 78, 2008, pp. 20–23. 23

[4] Malinovskiy Y., Y. Wu, Y. Wang, and U. Lee, “Field Experiments on Bluetooth-based 24 Travel Time Data Collection,” 89th Annual Meeting of the Transportation Research Board, 25 Washington, D.C.: 2010. 26

[5] Haghani A., M. Hamedi, K. Sadabadi, S. Young, and P. Tarnoff, “Freeway Travel Time 27 Ground Truth Data Collection Using Bluetooth Sensors,” 89th Annual Meeting of the 28 Transportation Research Board, Washington, D.C.: 2010. 29

[6] Barth M., F. An, T. Younglove, G. Scora, C. Levine, M. Ross, and T. Wenzel, Development 30 of a Modal-Emissions Model, Transportation Research Board, 2000. 31

[7] U.S. Environmental Protection Agency (EPA), Motor Vehicle Emission Simulator 32 (MOVES) 2010 User's Guide, U.S. EPA, 2009. 33

[8] Nesamani K.S. and K.P. Subramanian, “Impact of Real-world Driving Characteristics on 34 Vehicular Emissions,” JSME International Journal Series B, vol. 49, 2006, pp. 19–26. 35

[9] Joumard R., F. Philippe, and R. Vidon, “Reliability of the current models of instantaneous 36 pollutant emissions,” Science of the Total Environment, vol. 235, 1999, pp. 133–142. 37

[10] Rakha H., K. Ahn, and A. Trani, “Comparison of MOBILE5a, MOBILE6, VT-MICRO, 38 and CMEM models for estimating hot-stabilized light-duty gasoline vehicle emissions,” 39 Canadian Journal of Civil Engineering, vol. 30, 2003, pp. 1010–1021. 40

[11] Smit R., A.L. Brown, and Y.C. Chan, “Do air pollution emissions and fuel consumption 41 models for roadways include the effects of congestion in the roadway traffic flow?,” 42 Environmental Modelling and Software, vol. 23, 2008, pp. 1262–1270. 43

[12] Bai S., D. Eisinger, and D. Niemeier, “MOVES vs. EMFAC: A Comparison of Greenhouse 44 Gas Emissions Using Los Angeles County,” 88th Annual Meeting of the Transportation 45 Research Board, Washington, D.C.: 2009. 46


[13] Lin J. and D. Niemeier, “Regional driving characteristics, regional driving cycles,” 1 Transportation Research Part D: Transport and Environment, vol. 8, Sep. 2003, pp. 361-2 381. 3

[14] Bachman W., W. Sarasua, S. Hallmark, and R. Guensler, “Modeling regional mobile source 4 emissions in a geographic information system framework,” Transportation Research Part 5 C, vol. 8, 2000, pp. 205–229. 6

[15] Hallmark S.L. and R. Guensler, “Comparison of speed-acceleration profiles from field data 7 with NETSIM output for modal air quality analysis of signalized intersections,” 8 Transportation Research Record: Journal of the Transportation Research Board, vol. 9 1664, 1999, pp. 40–46. 10

[16] Zhang H., “Link-Journey-Speed Model for Arterial Traffic,” Transportation Research 11 Record: Journal of the Transportation Research Board, vol. 1676, Jan. 1999, pp. 109-115. 12

[17] Timoshek A., D. Eisinger, S. Bai, and D. Niemeier, “Mobile Source Air Toxics Emissions: 13 Sensitivity to Traffic Volume, Fleet Composition, and Average Speed,” 89th Annual 14 Meeting of the Transportation Research Board, Washington, D.C.: 2010. 15

[18] Bai S., Y. Nie, and D. Niemeier, “The impact of speed post-processing methods on regional 16 mobile emissions estimation,” Transportation Research Part D: Transport and 17 Environment, vol. 12, Jul. 2007, pp. 307-324. 18

[19] Hirschmann K. and M. Fellendorf, “A toolbox to quantify emission reductions due to signal 19 control,” 89th Annual Meeting of the Transportation Research Board, Washington, D.C.: 20 2010. 21

[20] Jackson E. and L. Aultman-Hall, “Analysis of Real-World Lead Vehicle Operation for the 22 Integration of Modal Emissions and Traffic Simulation Models,” 89th Annual Meeting of 23 the Transportation Research Board, Washington, D.C.: 2010. 24

[21] Park J.Y., R.B. Noland, and J.W. Polak, “Microscopic Model of Air Pollutant 25 Concentrations: Comparison of Simulated Results with Measured and Macroscopic 26 Estimates,” Transportation Research Record: Journal of the Transportation Research 27 Board, vol. 1750, 2001, pp. 64–73. 28

[22] Bigazzi A. and R. Bertini, “Adding Green Performance Metrics to a Transportation Data 29 Archive,” Transportation Research Record: Journal of the Transportation Research Board, 30 vol. 2121, Dec. 2009, pp. 30-40. 31

[23] Anderson W., P. Kanaroglou, E. Miller, and R. Buliung, “Simulating Automobile 32 Emissions in an Integrated Urban Model,” Transportation Research Record: Journal of the 33 Transportation Research Board, vol. 1520, Jan. 1996, pp. 71-80. 34

[24] Klein L.A., Sensor technologies and data requirements for ITS, Artech House, 2001. 35 [25] Smit R., R. Smokers, and E. Rabe, “A new modelling approach for road traffic emissions: 36

VERSIT+,” Transportation research. Part D, Transport and environment, vol. 12, 2007, 37 pp. 414–422. 38

[26] Guensler R., K. Dixon, V. Elango, and S. Yoon, “MOBILE Matrix: Application of Georgia 39 Statewide Multimodal Transportation Planning Tool for Rural Areas,” Transportation 40 Research Record: Journal of the Transportation Research Board, vol. 1880, Jan. 2004, pp. 41 83-89. 42

[27] Song G., L. Yu, and Z. Tu, “Distribution Characteristics of Vehicle Specific Power on 43 Urban Restricted Access Roadways,” 89th Annual Meeting of the Transportation Research 44 Board, Washington, D.C.: 2010. 45

46


[28] Boriboonsomsin K. and M. Barth, “Evaluating Air Quality Benefits of Freeway High-1 Occupancy Vehicle Lanes in Southern California,” Transportation Research Record: 2 Journal of the Transportation Research Board, vol. 2011, Jan. 2007, pp. 137-147. 3

[29] U.S. Environmental Protection Agency, Technical Guidance on the Use of MOVES2010 for 4 Emission Inventory Preparation in State Implementation Plans and Transportation 5 Conformity, U.S. EPA, 2009. 6

[30] Ericsson E., “Variability in urban driving patterns,” Transportation Research Part D, vol. 7 5, 2000, pp. 337–354. 8

9

using mac readers for real-time arterial emissions estimates 2

Documents