Professional Documents
Culture Documents
Banks
by
James H. Banks
Civil and Environmental Engineering Department
San Diego State University
San Diego, CA 92182-1324
Phone: (619) 594-7051
Fax: (619) 594-8078
E-mail: banks@mail.sdsu.edu
J. H. Banks
ABSTRACT
One- and two-stage regression models linking high-volume pre-queue flow (PQF) and
queue discharge flow (QDF) at freeway bottlenecks to the geometric, vehicle-population,
and driver-population characteristics of the sites were developed and evaluated in an
attempt to explain variations in PQF and QDF among bottlenecks. One-stage models
linked site characteristics directly to PQF and QDF; two-stage models utilized the
average time gap in the critical lane and the ratio of critical lane flow to flow per lane as
intervening variables, where the time gap is the time separation between the rear of the
lead vehicle and the front of the following one, and the critical lane is that with the
highest flow. Explanatory variables included number of lanes, vertical alignment,
proportion of heavy vehicles, and population characteristics derived from census data.
Models were identified by stepwise regression analysis and evaluated by comparing
predicted values of PQF and QDF with the results of Highway Capacity Manual (HCM)
capacity calculations and average measured values of PQF and QDF at the sites used to
develop the models and at a second group of sites used for model verification. Initial
models developed by the study all performed similarly: in most cases they performed
better than the HCM (which tends to overestimate both PQF and QDF), but in no case did
calculated values of PQF and QDF correlate with measured flows for the verification
sites. After one site with apparently anomalous data was eliminated, the only significant
explanatory variable was the number of lanes. It is concluded that, although some
variation in PQF and QDF among bottlenecks is explained by the number of lanes, a
great deal more remains unexplained.
J. H. Banks
Capacity flows at urban and suburban freeway bottlenecks vary considerably from site to
site (1). The reasons for this variation are interesting from the standpoints of both theory
and practice, since accurate understanding of them can provide insight concerning
fundamental features of traffic flow and support for improved freeway design and
operation. As will be shown later, Highway Capacity Manual [HCM (2)] methods for
analyzing freeway bottleneck capacity do not explain this variation in most cases.
Moreover, considerable past research suggests that flows immediately prior to breakdown
usually exceed queue discharge flows (3 13), so that either this high-volume pre-queue
flow (PQF) or queue discharge flow (QDF) might be said to represent capacity,
depending on the circumstances. This paper reports on an attempt (ultimately not very
successful) to explain the variation in PQF and QDF among urban and suburban freeway
bottlenecks by means of the geometric, vehicle population, and driver population
characteristics of different bottleneck sites.
BACKGROUND
A previous paper (1) reported on the first phase of research to develop and evaluate an
alternative to current HCM methods for analyzing freeway bottleneck capacity. This
alternative approach begins with the observation that capacity flow rates result from the
average time gap and average passage time in the critical lane and the lane flow
distribution, as represented by the critical lane flow ratio (CLFR). The average time gap
is the time separation between the rear of the lead vehicle and the front of the following
one, the passage time is the time it takes a vehicle to pass a point, the critical lane is that
with the highest flow, and the CLFR is the ratio of flow in the critical lane to average
flow per lane. These characteristics, which may be computed by transformations of
volume and occupancy data typically produced by loop detectors, are related to one
another by
q=
(1)
rc (g c + pc )
where
q
gc
pc
rc
=
=
=
=
=
It was proposed to use these transformed characteristics as intervening variables in twostage models linking capacity flows to the geometric, vehicle population, and driver
population characteristics of bottlenecks. It was hoped that such two-stage models could
provide greater insight into the factors influencing bottleneck capacity and result in better
estimates of capacity at urban and suburban bottlenecks than could one-stage models (in
which bottleneck characteristics are linked directly to capacity flows) or existing HCM
methods.
J. H. Banks
The first phase of the research focused on analysis of interrelationships among the
transformed flow characteristics. Peak period flow and occupancy data from fifteen
bottlenecks located in the Minneapolis-St. Paul, San Diego, and Seattle areas were
collected over extended periods of time and reduced to provide site-mean averages of the
various flow characteristics for periods of PQF and QDF. Bottleneck flows were also
recorded for an additional six sites for which time gaps could not be determined because
there were no detectors in the bottleneck sections. The fifteen sites for which all flow
characteristics were available were used in the development of the models and are
referred to as calibration sites; the six additional sites were used for model verification
and are referred to as verification sites. Figure 1 shows schematic diagrams of the study
sites and Table 1 lists some of their characteristics.
Analysis of data from the calibration sites established that (a) PQF and QDF were
distinct flow periods; (b) average flow rates during PQF were typically higher than those
during QDF, confirming the results of past research; (c) the average time gap in the
critical lane was the dominant influence on both PQF and QDF, with CLFR having a
lesser but possibly significant influence and average passage time having little or no
influence; (d) the time gap and passage time were significantly correlated; and (e) the
relationships between average critical lane time gap and PQF and QDF might well be
linear, despite the form of Equation 1. On the basis of these results, linear regression
models relating the critical lane flow to the critical lane average time gap were calibrated
for PQF and QDF and a model of the form
q=
1
(a bg c )
rc
(2)
where a and b are regression coefficients was proposed as the basis for two-stage models
linking bottleneck site characteristics to PQF and QDF. Details of the study sites, data,
and analysis for the first phase of the research are described in (1) and in the project
report (14).
The main goal of the second stage of the project was to calibrate and evaluate
models linking PQF and QDF to the geometric, vehicle population, and driver population
characteristics of the bottlenecks. For purposes of comparison, both one-stage and twostage models were developed. Models were compared with one another and with the
current HCM procedures and were evaluated on the basis of how well they explained
variations in PQF and QDF at both the calibration and verification sites.
SELECTION OF POTENTIAL EXPLANATORY VARIABLES
The first step in the development of the candidate models was to identify potential
explanatory variables. The set of variables selected for testing included geometric
characteristics (vertical alignment and number of lanes), vehicle population
characteristics (proportion of heavy vehicles in the traffic stream) and a number of
variables derived from census data that were intended to represent driver population
characteristics. As may be noted from Table 1, the study sites included several different
types of bottlenecks. Ideally, separate models would have been developed for each type;
however, there were too few examples of each type (except possible merge bottlenecks)
to support meaningful models. Since the sample was already small, and the analysis of
J. H. Banks
TH 169
MN-02
Medicine Lake Rd.
Plymouth Ave.
I-35W
MN-08
Diamond Lake Rd.
I-94
46th St.
I-394
MN-18
TH 169
Penn Ave.
I-94
MN-22
CR 152
CR81
MN-21
Boone Ave.
TH 280
I-94
N
SD-05
I-15
SD-08
SD-07
I-405
N
NE 85th
N
Governor Drive
BothellEverett Hwy.
NE 116th
I-5
WA-03
th
WA-01
236th St.
I-405
WA-02
Exit 24
Highland Valley
Rd./ Pomarado Rd.
El Cajon Boulevard
I-805
CR 81
Rancho
Bernardo Rd.
I-805
University Ave.
MN-25
CR 152
MN-23
N
CR 81
I-94
Nobel Drive
I-94
MN-14
TH-100
Huron St.
CR 61
220th St.
I-405
WA-04
I-90
Verification Sites
N
I-15
SD-01
West Bernardo Dr.
Via Rancho Parkway
N
I-5
47th St.
I-805
SD-02
43rd St.
I-5
SD-03
Manchester Avenue
Lomas Santa Fe Dr.
SD-06
Via De la Valle
I-8
I-8
N
Fletcher Parkway
Lomas Santa
Fe Drive
SD-09
70th Street /
Lake Murray Blvd.
Waring Rd.
SD-12
N
College Ave.
J. H. Banks
Peak
Lanes Type
How used
MN-02
MN-08
MN-14
MN-18
MN-21
MN-22
MN-23
MN-25
SD-01
SD-02
SD-03
SD-05
SD-06
SD-07
SD-08
SD-09
SD-12
WA-01
WA-02
WA-03
WA-04
PM
PM
PM
PM
PM
PM
PM
PM
AM
AM
AM
AM
PM
PM
PM
AM
PM
PM
PM
PM
AM
2
3
3
2
2
2
3
2
4
4
4
4
4
4
4
4
5
3
2
3
2
calibration
calibration
calibration (initial models)
calibration
calibration
calibration
calibration
calibration
verification
verification
verification
calibration
verification
calibration
calibration
verification
verification
calibration
calibration
calibration
calibration
merge
merge
merge
merge
merge/horizontal curve
merge
merge/3-d curve
lane drop
merge/grade
merge
merge/grade
weave exit leg, grade
merge/grade
weave exit leg
weave exit leg, grade
merge
merge or grade
merge
diverge
merge
weave
traffic flow characteristics in the first phase of the study had not revealed any obvious
differences among the different bottleneck types, all were lumped together for purposes
of developing the models.
Current HCM methods relate variations in urban and suburban bottleneck
capacity to geometric and vehicle population characteristics. Capacity variations, as
computed by HCM procedures, are affected directly by a heavy vehicle factor and
indirectly by the number of lanes, which influences estimated free-flow speed. The heavy
vehicle factor depends on the proportion of heavy vehicles in the traffic stream and the
length and steepness of the grade. The finding that critical lane time gaps are the
dominant flow characteristic influencing PQF and QDF suggests that capacity flows are
primarily the result of fundamental features of driver behavior particularly the
minimum spacing that drivers will tolerate and consequently that driver population
characteristics, either alone or in combination with geometric and vehicle population
characteristics, may also be an important influence on bottleneck capacity. Note that
although the current HCM method does involve a driver population factor, this factor is
always taken to be 1.0 for urban and suburban commuter traffic.
In the case of geometric and vehicle population characteristics, the study followed
the lead of the HCM by considering the number of lanes, the length and steepness of the
grade approaching the bottleneck, and the proportion of heavy vehicles. In addition, the
J. H. Banks
HCM heavy vehicle factor was computed from the vertical alignment and vehicle
classification data; however, because all the sites were relatively flat, it proved to be
almost perfectly correlated with the percentage of heavy vehicles and was dropped from
further consideration.
Data on vertical alignments were obtained from as-built plans supplied by the
respective state departments of transportation. Heavy vehicle data for the Seattle sites
were obtained directly from the Washington State Department of Transportation, and
those for Minneapolis-St. Paul and San Diego from internet sites maintained by the
respective departments of transportation (15, 16). For each site, the vehicle classification
data used were those from the nearest available count station. In addition, hand counts
were conducted at the San Diego sites to provide additional data on heavy vehicle
presence during peak periods.
As it turned out, use of routinely-collected vehicle classification data resulted in
serious consistency problems. Each of the states concerned has a different vehicle
classification scheme and none is fully compatible with that used by the HCM.
Washington State classifies large vehicles by length and provides data by time of day on
an hourly basis. California classifies vehicles by the number of axles and provides only
annual average daily traffic volumes. Minnesota provides separate average daily traffic
volumes for all traffic and heavy commercial vehicles, which are defined as those with
six or more tires. In no case are recreational vehicles treated as a separate category, as in
the HCM. Although the capacity flows being modeled pertained to peak periods, daily
counts had to be used in the analysis because these were the only vehicle classification
data available for all sites. This obviously introduced some error: at most sites where it
was possible to make comparisons, peak period heavy vehicle percentages were
significantly less than daily percentages, but the relationship varied among the sites.
Consequently, the available data tended to overstate heavy vehicle presence during peak
periods, but by different amounts at different sites.
Selection of variables related to driver population characteristics was influenced
by theoretical considerations and the nature of the available data. Since the feature of
driver behavior most directly related to bottleneck capacity is the willingness to tolerate
close vehicular spacing, the driver population characteristics of primary interest were
those believed to be related to driver aggressiveness. It was surmised that aggressiveness
would correlate with age, gender, income, education, trip purpose, metropolitan area size,
and population density. Specifically, it was hypothesized that aggressiveness would
correlate positively with median income (because of the greater value of time) and with
metropolitan size and population density (because of the faster pace of life) and
negatively with age and educational level, once income effects were accounted for. It was
also surmised that drivers making work trips would be more aggressive than drivers
making other types of trips and that males would be more aggressive than females.
The major difficulty in quantifying these characteristics is that data related to
them are not commonly available for specific traffic streams. For instance, although
travel diary surveys provide information about traveler characteristics, they typically ask
for origins and destinations but not routes, so that it is not possible to identify the
respondents using a given bottleneck. In the absence of direct data, the obvious approach
is to try to estimate driver characteristics indirectly, using census data from the
surrounding area.
J. H. Banks
The major drawback of this approach is that it assumes that the census data are
representative of the actual driver population, although this is not known to be so. Drivers
using a bottleneck are typically drawn from a rather large area (the full extent of which is
not known), and the data are of necessity drawn from a more limited zone. There is no
way of knowing the extent to which the characteristics of the population in the zone are
representative of those of the population in the entire area, nor of knowing how the
characteristics of the overall population relate to those of the drivers. The best that can be
hoped is that the census data are reasonably representative of the bottleneck users. A
second important drawback is that some of the relevant data for instance, trip purpose
are not available through the census.
Census data (17) were downloaded for zones consisting of sets of census tracts
believed to represent reasonable commuter sheds for the study sites. These zones were
defined as follows: (a) zones for morning-peak bottlenecks were located upstream of the
bottleneck and those for evening peak bottlenecks downstream; (b) in the absence of
other constraints, zones were approximately 15 km long by 6 km wide, centered on the
freeway in question; (c) where there were parallel freeways less than 12 km away, the
zone boundary was placed approximately halfway between the two freeways; (d) where
applicable, zones were limited by major traffic barriers and by the edge of the urbanized
area; (e) if the upstream end of the zone, as determined by the 15 km rule, included a
central business district (CBD), the zone was terminated at the CBD; and (f) census tracts
were included if more than half of their area was within the zone boundary as determined
by rules (a) through (e).
The downloaded census data were then reduced to determine the median age,
median income, percent of males aged 18 to 24, percentage of college graduates, and
population density for each zone. The percent of males aged 18 to 24 was included as a
separate variable because the insurance industry tends to regard this group as posing a
particularly high risk of traffic accidents (and hence, presumably, to be particularly
aggressive). The other variables are also intended to represent driver characteristics
believed to be related to driver aggressiveness. Metropolitan area size was omitted as a
variable because only three areas were involved and they are all reasonably similar in
size.
IDENTIFICATION AND EVALUATION OF MODELS
Initial models were identified by using of stepwise regression analysis to isolate the
explanatory variables with the most influence on pre-queue and queue discharge flow
characteristics. In the case of two-stage models, the resulting regression equations were
then combined with equation 2. The initial models were evaluated in terms of their ability
to explain variations in PQF and QDF at the calibration and verification sites and were
compared with one another and the current HCM procedures. Modified models based on
the evaluation results were then proposed. As described in the preceding section, the
combined set of geometric, vehicle population, and driver population variables included
roadway grade (GRD), percent heavy vehicles in the traffic stream (PHV), median age
(AGE), median income (INC), percent of males aged 18 to 24 (YML), percentage of
college graduates (PCG), and population density (PDN). Response variables included
flow per lane, gap, and CLFR.
J. H. Banks
J. H. Banks
J. H. Banks
models had performed similarly, only the simpler one-stage models of PQF and QDF
were used for this analysis.
In this case, the only significant explanatory variable was the number of lanes.
The resulting models are:
PQF model:
(3)
QDF model:
qD = 1,775 + 68.23LNS
(4)
Correlation coefficients for estimated flow versus measured flow were +0.469 for the
PQF model and +0.629 for the QDF model; t-statistics for the slopes were 1.99 and 3.03
respectively. The correlation coefficient and t-statistic for the PQF model are not
significant at 0.05 (although they are at 0.10), but those for the QDF model are significant
at 0.01.
Table 2 compares the resulting estimates of PQF and QDF with the means of
measured flows for sites with different numbers of lanes and shows that the estimates
produced by the models are similar to the mean measured flows. At the verification sites
(all but one of which had four lanes) the models underestimated PQF but slightly
overestimated QDF. This result is largely explained by the fact that mean PQF for the
verification sites was about 6 percent above that for the four-lane sites used to calibrate
the model, while mean QDF was nearly equal for both groups of sites. Since all but one
of the verification sites had the same number of lanes, the models obviously cannot
explain the variation in flow among them. Table 3 gives the errors for individual sites and
compares the performance of the models with that of the HCM method.
Table 2 Estimated and Measured Flows for Sites with Different Numbers of Lanes
Number of lanes
Measure
Pre-queue flow
Model
Mean flow, calibration sites
Mean flow, verification sites
2065
2064
2097
2098
2129
2127
2257
1911
1908
1980
1991
2048
2038
2041
J. H. Banks
10
Table 3 Comparison of Errors at Individual Sites for the PQF Model, the QDF
Model and the HCM Method
PQF
Measured
flow
Site
Calibration
MN-02
MN-08
MN-18
MN-21
MN-22
MN-23
MN-25
SD-05
SD-07
SD-08
WA-01
WA-02
WA-03
WA-04
Year
2000
2004
2004
2000
2000
2000
2000
2000
2000
2004
2004
2004
2004
2004
2004
2004
2153
1999
2041
2043
2016
2047
2173
2059
2130
2095
2108
2179
2097
2055
2120
2064
Ave. error
Verification
SD-01
SD-02
SD-03
SD-06
SD-09
2004
2004
2004
2004
2004
Ave. error
2419
2416
2179
1982
2287
QDF
Error (estimated
measured)
Model
HCM
-89
66
56
21
49
17
-76
38
-66
34
21
-50
0
10
-23
0
+82
+237
+209
+154
+199
+153
+56
+170
+63
+145
+140
+81
+186
+199
+119
+194
+1
+149
-290
-287
+50
+147
-158
-222
-216
+88
+190
-73
-108
-47
Measured
flow
2033
1920
1936
1916
1842
1884
2046
2022
1940
1989
2043
2083
1986
1983
1966
1746
2175
2184
1926
1824
2094
Error (estimated
measured)
Model
HCM
-122
-9
+43
-4
+70
+27
-67
-42
-28
+59
+5
-35
-7
-72
13
+166
+202
+316
+313
+282
+373
+317
+183
+207
+254
+252
+205
+165
+297
+271
+273
+513
+276
-127
-136
+122
+224
-46
+73
+62
+349
+465
+172
+7
+222
The conclusion that the number of lanes is the only significant explanatory
variable that regression models may not be the best approach to the prediction of capacity
flows. A simpler (and possibly more accurate) approach is to treat the mean values of
PQF and QDF as the predicted capacities for sites with a similar number of lanes and to
use the standard deviations of the site-mean flows as an indicator of the precision of the
estimates. Table 4 gives this information, with values for four-lane sites computed using
data from verification sites as well as calibration sites.
J. H. Banks
11
Flow condition
Number of
directional
lanes
2
3
4
PQF
QDF
Mean flow
St. dev.
Mean flow
St. dev.
2064
2098
2208
53.1
52.3
155.6
1908
1991
2040
87.7
43.6
123.1
EVALUATION OF APPROACH
This study began with the hypothesis that two-stage models that explain PQF and QDF in
terms of bottleneck site characteristics might be an attractive alternative to current HCM
methods. For the most part, the results did not support this hypothesis: one-stage and twostage models performed similarly, and neither could reliably predict most of the variation
in capacity flows at the study sites. In the end, the only site characteristic found to
influence capacity was the number of lanes, and the significance of even this relationship
is doubtful in the case of PQF. Models based on the number of lanes explain some of the
variation among sites but leave a great deal more unexplained.
Although the performance of these models was disappointing, it was generally
better than that of the current HCM method. This not only failed to explain the variation
among the sites (correlations between estimated and measured flows tended to be
negative but not significant) but also resulted in large positive biases for all cases except
that of PQF at the verification sites. The large biases with respect to QDF underscore the
need to treat PQF and QDF separately. Although it appears that the capacity in the
HCM is intended to be PQF, this is not stated explicitly, and the failure to provide for
separate estimates of PQF and QDF can lead to serious errors if the HCM estimates are
used in calculations (such as those outlined in Chapter 22 of the manual) that involve
congested flow.
Two important features of the study approach proved to be vulnerable to the
unavailability or poor quality of data. First, this study, like most intended to support
practical capacity analysis, attempted to develop statistical models using data from a
relatively large number of sites. In such studies, the size of the sample and the degree to
which it is representative are of critical importance. A variety of data problems, including
the small number of metropolitan areas for which data are available, inappropriate
detector location, and long-term detector failures, led to a sample of bottlenecks that was
not adequately representative: different types of bottlenecks were mixed together in the
sample and only a very limited range of metropolitan area sizes was available.
In addition, a major goal of the study was to gain a better understanding of how
driver population characteristics influence bottleneck capacity. As it turned out,
relationships between population characteristics and capacity could not be established,
perhaps in part because of the lack of direct data about driver characteristics for particular
traffic streams. It comes as no great surprise that the expedient of substituting census data
J. H. Banks
12
for data specific to the drivers proved unsuccessful, since the characteristics of the
general population in the commuter shed zones may well have differed from those of the
drivers.
These data deficiencies are apt to remain barriers to better understanding of the
reasons for variations in PQF and QDF among bottlenecks. For the most part, problems
with automatically-collected traffic data stem from the fact that traffic surveillance
systems are not primarily intended to support research and often are not a high priority
for the agencies providing them. These problems could be remedied by the expansion of
traffic surveillance systems, better design of such systems in terms of things like detector
location, better maintenance, and more widespread posting of data on the internet. The
lack of data about the characteristics of drivers using particular bottlenecks might be
remedied by collecting such data as part of routine travel surveys; however, it is unlikely
that samples of the size commonly used in such surveys would provide enough data about
individual bottlenecks. For research purposes, it might also be possible to conduct
surveys of the users of individual bottlenecks. Models based on such surveys might be
hard to apply, however, since similar data (and hence similar surveys) would be needed
for each location analyzed
CONCLUSIONS
This study has considered the effectiveness of one- and two stage models that attempt to
explain variation in PQF and QDF among urban and suburban freeway bottlenecks in
terms of the geometric, vehicle-population, and driver-population characteristics of the
sites. Evaluation of the models was based on their overall accuracy and their ability to
explain variations in capacity among sites and included comparing them with one another
and with current HCM methods. Major conclusions are:
1. It is important that PQF and QDF be treated separately in capacity analysis.
Regardless of the definition of capacity adopted, both are important features of
bottleneck flow, and serious inaccuracy can result if they are confused.
2. Current HCM methods do not explain variations in capacity flows among
relatively flat urban and suburban bottleneck sections. In addition, estimates based on the
HCM methods are biased high for QDF; in most cases, such estimates are also biased
high for PQF.
3. The performance of one-stage and two-stage models is similar; hence the
simpler one-stage models are to be preferred.
4. Of the various site and population characteristics considered, the only one
found to influence capacity was the number of lanes, and the significance of even this
relationship is doubtful in the case of PQF; models based on the number of lanes explain
some of the variation among sites but leave a great deal more unexplained
5. Lack of data and poor data quality are apt to remain major barriers to
understanding variations in PQF and QDF among bottlenecks.
ACKNOWLEDGEMENTS
This work was performed as part of the California PATH Program of the University of
California, in cooperation with the State of California Business, Transportation, and
Housing Agency, Department of Transportation; and the United States Department of
Transportation, Federal Highway Administration. Special thanks are due to my research
J. H. Banks
13
J. H. Banks
14