See

discussions, stats, and author profiles for this publication at:

https://www.researchgate.net/publication/23527776

Comparison of delay estimates at under-

saturated and over-saturated pre-timed
signalized intersections

Article in Transportation Research Part B Methodological · February 2004

DOI: 10.1016/S0191-2615(03)00003-1 · Source: RePEc

CITATIONS READS

167 1,056

3 authors:

Francois Dion Hesham Rakha

University of California, Berkeley Virginia Polytechnic Institute and State…
36 PUBLICATIONS 1,560 CITATIONS 336 PUBLICATIONS 4,767 CITATIONS

SEE PROFILE SEE PROFILE

Youn-Soo Kang
The Korea Transport Institute
11 PUBLICATIONS 270 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Tools for Tactical Decision-Making/Advancing Methods for Predicting Performance View

project

Multi-Modal Intelligent Traffic Signal Systems (MMITSS) Impacts Assessment View project

All content following this page was uploaded by Hesham Rakha on 16 October 2017.

The user has requested enhancement of the downloaded file.

 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

Transportation Research Part B xxx (2003) xxx–xxx

www.elsevier.com/locate/trb

F
Comparison of delay estimates at under-saturated and

OO
over-saturated pre-timed signalized intersections
a,*
Francois Dion , Hesham Rakha b, Youn-Soo Kang c

PR
Virginia Tech Transportation Institute, 3500 Transportation Research Plaza (0536), Blacksburg, VA 24061, USA
b
Charles Via Department of Civil and Environmental Engineering, 3500 Transportation Research Plaza, Blacksburg,
VA 24061, USA
c
Korean Transportation Institute, South Korea
Received 20 March 2002; received in revised form 28 April 2002; accepted 30 October 2002

Abstract
D
TE
11 Delay is an important parameter that is used in the optimization of traffic signal timings and the esti-
12 mation of the level of service at signalized intersection approaches. However, delay is also a parameter that
13 is difficult to estimate. While many methods are currently available to estimate the delays incurred at in-
14 tersection approaches, very little research has been conducted to assess the consistency of these estimates.
EC

15 This paper addresses this issue by comparing the delays that are estimated by a number of existing delay
16 models for a signalized intersection approach controlled in fixed-time and operated in a range of conditions
17 extending from under-saturated to highly saturated. Specifically, the paper compares the delay estimates
18 from a deterministic queuing model, a model based on shock wave theory, the steady-state Webster model,
RR

19 the queue-based models defined in the 1981 Australian Capacity Guide, the 1995 Canadian Capacity Guide
20 for Signalized Intersections, and the 1994 and 1997 versions of the Highway Capacity Manual (HCM), in
21 addition to the delays estimated from the INTEGRATION microscopic traffic simulation software. The
22 results of the comparisons indicate that all delay models produce similar results for signalized intersections
23 with low traffic demand, but that increasing differences occur as the traffic demand approaches saturation.
CO

24 In particular, it is found that the delay estimates from the INTEGRATION microscopic simulation model
25 generally follow the delay estimates from the time-dependent models defined in the 1997 HCM, 1995
26 Canadian Capacity Guide, and 1981 Australian Capacity Guide over the entire range of traffic conditions
27 considered.
28 Ó 2003 Published by Elsevier Science Ltd.
UN

29

*
Corresponding author. Tel.: +1-540-231-9619; fax: +1-540-231-5214.
E-mail addresses: fdion@vtti.vt.edu, fdion@ctr.vt.edu (F. Dion), hrakha@vt.edu (H. Rakha).

0191-2615/03/$ - see front matter Ó 2003 Published by Elsevier Science Ltd.

doi:10.1016/S0191-2615(03)00003-1
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

2 F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx

Nomenclature

Ai area of region i in time–space diagram (km h)

c s
ge =C ¼ capacity of intersection approach (veh/h)
C traffic signal cycle length (s)

F
d average delay per vehicle (s/veh)
D total delay incurred on intersection approach (veh s)

OO
do deterministic overflow delay (s/veh)
d1 uniform delay (s/veh)
d2 incremental delay accounting for randomness of vehicle arrivals and over-saturation
delay (s/veh)
d3 residual delay for over-saturation queues that may have existed before the analysis

PR
period (s/veh)
fPF adjustment factor accounting for the quality of progression in coordinated systems
fr adjustment factor for residual delay component
fp adjustment factor for situations in which the platoon arrives during the green interval
(0.9–1.2)
ge
k
efective green interval duration (s) D
incremental delay factor accounting for pre-timed or actuated signal controller set-
TE
tings
ka density of approaching traffic (veh/km)
kd density of discharging traffic (veh/km)
ki traffic density in zone i (veh/km)
EC

kj jam density (veh/km)

I adjustment factor for upstream filtering/metering
n; m capacity guide model parameters
N number of cycles over which calculations are performed
RR

P proportion of vehicles arriving during effective green interval

re effective red interval duration (s)
s saturation flow rate (veh/h)
SWij speed of shock wave between traffic zones i and j (km/h)
tcðuÞ time to clear the queue of vehicles in under-saturated conditions (s)
CO

tmðuÞ time to maximum extent of queue in under-saturated conditions (s)

T evaluation period (h)
TT travel time of vehicle during trip
TTo total travel time on intersection approach during evaluation period, over-saturated
UN

conditions (veh s)
TTu total travel time on intersection approach during signal cycle, under-saturated con-
ditions (veh s)
uf vehicle speed under free-flow conditions
ui traffic speed in zone i (km/h)
uðtÞ instantaneous speed at time t
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx 3

v vehicle arrival flow rate (veh/h)

vi traffic flow rate in zone i (veh/h)
xcðoÞ distance over which the residual queue grows over a signal cycle in over-saturated
conditions (km)
xmðoÞ maximum extent of the queue within a signal cycle with respect to queue size at the

F
beginning of the cycle in over-saturated conditions (km)
xmðuÞ maximum extent of queue within signal cycle in under-saturated conditions (km)

OO
X v=c ¼ volume-to-capacity ratio
Xo volume-to-capacity ratio below which the overflow delay is negligible in capacity
guide models

PR
30 1. Introduction

31 Vehicle delay is perhaps the most important parameter used by transportation professionals to
32 evaluate the performance of signalized intersections. This importance of vehicle delay is reflected
33 in the use of this parameter in both design and evaluation practices. For example, delay mini-
34
35
36
D
mization is frequently used as a primary optimization criterion when determining the operating
parameters of traffic signals at isolated and coordinated intersections. The Highway Capacity
Manual (HCM) further uses the average control delay incurred by vehicles at intersection ap-
TE
37 proaches as a base for determining the level of service provided by the traffic signals located at the
38 downstream end of the these approaches (TRB, 1997).
39 The popularity of delay as an optimization and evaluation criterion is attributed to its direct
40 relation to what motorists experience while attempting to cross an intersection. However, delay is
EC

41 also a parameter that is not easily determined. Teply (1989), for instance, indicated that a perfect
42 match between field-measured delay and analytical formulas could not be expected. The difficulty
43 in estimating vehicle delay at signalized intersections is also demonstrated by the variety of delay
44 models for signalized intersections that have been proposed over the years.
RR

45 Despite differences between the proposed delay models, very little research has been concerned
46 with the consistency of delay estimates from one model to the other. This paper addresses this
47 problem by comparing the delays that are estimated by a number of analytical delay models,
48 including deterministic queuing, shock wave, steady-state stochastic, and time-dependent sto-
CO

49 chastic delay models, and by further comparing these estimates to the delays that are produced by
50 a microscopic traffic simulation model. To achieve this goal, the paper first presents some
51 background material on vehicle delays at signalized intersection, followed by a description of the
52 various delay models that are being compared. Evaluations of the consistency of delay estimates
53 from these models are conducted by using them to evaluate delays on both under-saturated and
UN

54 over-saturated signalized intersection approaches.

55 2. Delay at signalized intersections

56 Delay at signalized intersections is computed as the difference between the travel time that is
57 actually experienced by a vehicle while going across the intersection and the travel time this ve-
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

4 F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx

Distance
Total delay

F
Stopped delay

OO
Acceleration
Deceleration delay
delay
Vehicle path without stop
Vehicle path with stop and gradual

PR
deceleration and acceleration
Vehicle path with stop and instantaneous
deceleration and acceleration

Time

D
Fig. 1. Definition of total, stopped, deceleration and acceleration delays.
TE
58 hicle would have experienced in the absence of traffic signal control. The diagram of Fig. 1 further
59 indicates that the total delay experienced by a vehicle can be categorized into deceleration delay,
60 stopped delay and acceleration delay. Typically, transportation professionals define stopped delay
61 as the delay incurred when a vehicle is fully immobilized, while the delay incurred by a deceler-
EC

62 ating or accelerating vehicle is categorized as deceleration and acceleration delay, respectively. In

63 some cases, stopped delay may also include the delay incurred while moving at an extremely low
64 speed. For example, the 1995 Canadian Capacity Guide for Signalized Intersections (ITE, 1995)
65 defines stopped delay as any delay incurred while moving at a speed that is less than the average
RR

66 speed of a pedestrian (1.2 m/s).

67 Fig. 2 illustrates in more detail the distinction between deceleration, stopped and acceleration
68 delay. The figure illustrates the simulated trajectory and speed profiles of a number of vehicles
69 arriving at a signalized intersection during a single cycle. These simulated trajectories were ob-
CO

70 tained using the INTEGRATION microscopic traffic simulation model (Van Aerde and Asso-
71 ciates, 2001). In the figure, it is first observed that only the first eight vehicles reaching the
72 intersection come to a complete stop. These vehicles need to stop either as a consequence of their
73 arrival during the red interval or during the green interval when the queue of vehicles that had
74 formed during the previous red interval has not yet fully dissipated. It is further observed that the
UN

75 following three vehicles only experience deceleration and acceleration delay, as these vehicles
76 reach the intersection when all previously queued vehicles have already started to move and
77 therefore only need to slow down to maintain a safe distance with the vehicles ahead of them.
78 While most of the delay incurred at signalized intersections is directly caused by the traffic
79 signal operation, a fraction of the total delay is attributable to the time required by individual
80 drivers to react to changes in the signal display at the beginning of the green interval, to me-
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx 5

2.15

2.10

2.05
Distance (km)

F
2.00

Fr
on
to

OO
fq
1.95 Bac
k of

ue
que
ue

ue
1.90

1.85

PR
1 2 3 4 5 6 7 8 9 10 11 12
1.80
890 900 910 920 930 940 950 960 970
Time (sec)

Vehicle 1 Vehicle 7 Vehicle 9 Vehicle 11

70
Speed (km/h)

60
50
40
30
20
D
Sp

10
TE
0
850 900 950 1000 850 900 950 1000 850 900 950 1000 850 900 950 1000
Time (s) Time (s) Time (s) Time (s)

Fig. 2. Simulated time–space diagram for typical traffic signal cycle.

EC

81 chanical constraints, and to individual driver behavior. In ideal situations, vehicles queued at an
82 intersection would start moving at their ideal speed immediately following the display of a green
83 signal. However, the first few drivers usually hesitate a few seconds before starting to accelerate,
RR

84 thus causing additional delays to all queued vehicles. This delay in the start of the queue dissi-
85 pation process may even cause additional vehicles to join the queue before its complete dissipa-
86 tion. When acceleration occurs, the rate at which vehicles accelerate also depends on mechanical
87 constraints dictating the maximum feasible acceleration rate and on the rate at which individual
driver chose to accelerate.
CO

88
89 As an example, Fig. 3 illustrates the simulated headways between successive stop line cross
90 times for the 12 vehicle arrivals of Fig. 2. As can be observed, the first vehicle crosses the stop line
91 4.3 s after the green initiation, while the second, third, and fourth vehicles follow with respective
92 headways of 3.0, 2.7 and 2.4 s. In this case, the larger headways observed at the beginning of the
UN

93 interval are entirely caused by acceleration constraints. In the absence of such constraints, all
94 vehicles would have crossed the intersection with average 2-s headway. This would have resulted
95 in a 24-s reduction in the total delay incurred during the simulated signal cycle.
96 To account for the additional delays due to driver reaction time and vehicle acceleration
97 constraints, the operation of a signalized intersection is usually defined in terms of effective signal
98 intervals instead of actual intervals in delay estimation models, as shown in Fig. 4. Instead of
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

6 F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx

4.5
4.0
3.5

Headway (s)
3.0
2.5
2.0
1.5

F
1.0
0.5
0.0

OO
1 2 3 4 5 6 7 8 9 10 11 12
Departing vehicle

Fig. 3. Simulated headway distribution at beginning of green interval.

PR
Lost Lost
time Effective green time
Vehicles discharge rate

Saturation
flow rate

D
TE
Red Green Amber Red
EC

0 6 12 18 24 30 36
Time from start of green (s)

Fig. 4. Queue modeling under deterministic queuing analysis.

RR

99 explicitly considering green, yellow and amber intervals and attempting to model variable de-
100 parture rates, delay calculations are typically performed by dividing the signal cycle into effective
101 periods of stopped and moving traffic within which constant traffic characteristics can be assumed.
102 The amount of difference between the actual and effective timings will thus depend on the as-
CO

103 sumptions regarding driver reaction time at the beginning of the green interval and vehicle ac-
104 celerations.
105 A final element that may affect the delays incurred at intersection approaches is the randomness
106 in vehicle arrivals. If vehicles were to arrive at uniform intervals, the delays incurred by vehicles
107 within successive signal cycles would be identical, as there would then be an exact replication of
UN

108 the arrival and departure patterns. However, under random arrival patterns, the number of ar-
109 rivals may fluctuate from one cycle to the other, thus resulting in different queue lengths. This may
110 in turn result in arrival demands that occasionally exceed the approach capacity, and therefore, in
111 higher delays. Finally, platooned arrivals may also occur in coordinated traffic signal systems. In
112 this case, the delay incurred by vehicles will depend on the degree to which the signals at successive
113 intersections are timed to provide a green indication during the periods of high arrival flow rate.
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx 7

114 3. Delay models for signalized intersections

115 3.1. Deterministic queuing model

116 Classic, deterministic queuing models can predict delay at signalized intersections for which the
117 number of vehicles that can be served during a green interval is greater than the number of arrivals

F
118 per cycle. These models view traffic on each intersection approach as a uniform stream of arriving
119 vehicles seeking service from a control device that provides a high service rate, but that also

OO
120 periodically stops servicing vehicles to accommodate traffic on a conflicting movement.
121 To illustrate how deterministic queuing models predict delays, consider the upper diagram of
122 Fig. 5, which illustrate cumulative arrivals and departures at an under-saturated intersection.

PR
Cumulative vehicles

Cycle 1 Cycle 2 Cycle 3

Time spent
by a vehicle
in queue D Area:
TE
Total delay

Cumulative a rrivals
A(t) Number of
vehicles in queue Cumularive departures
EC

D(t)

Time
Cumulative vehicles

RR

Overflow
Number of queue
vehicles in
queue
Area:
CO

Overflow
delay
Time spent
by a vehicle
in queue

Cumulative a rrivals
UN

Cumularive arrivals
Area: at capacity
A(t)
Uniform Cumularive departures
delay
D(t)

Time

Fig. 5. Idealized cumulative arrivals and departures for under- and over-saturated conditions.
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

8 F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx

123 From the figure, it can be determined that the area between the arrival and departure curves
124 represents the total uniform delay incurred by all vehicles attempting to cross an intersection
125 within a signal cycle. Assuming uniform arrivals and service times, i.e., a queuing system of the
126 type D/D/1, Eqs. (1) and (2) can then be derived to calculate the average uniform delay incurred
127 every signal cycle by vehicles attempting to cross the intersection. Eq. (2) is identical to what is
128 used in the HCM and the Canadian Capacity Guide for Signalized Intersections, as will be

F
129 demonstrated in the following sections:
r2
s 

OO
d¼ e
ð1Þ
2
C sv
 2
C 1  gCe
d¼
 ð2Þ
2 1  X gCe

PR
132 The model of Eq. (1) was generated by first assuming that vehicles arrive at a uniform and
133 constant rate. A consequence of this assumption is that the queue of vehicles that form at an
134 intersection operating in under-saturated conditions can always be cleared before the return of the
135 red signal. In reality, the randomness of traffic may cause some vehicles to remain queued at the
136 end of the green interval, especially at intersections operating near saturation. A second as-
137
138
D
sumption is that vehicles decelerate and accelerate instantaneously. As was illustrated in Fig. 1,
this assumption converts all deceleration and acceleration delays into equivalent stopped delay,
TE
139 and thus allows a direct estimation of the total delay incurred by vehicles attempting to cross an
140 intersection. This assumption also implies that all drivers follow average driving patterns, in
141 addition to assigning all incurred delays to the intersection approach, even though some delay
142 occur in reality on the exit link when vehicles are accelerating. A final assumption is that vehicles
EC

143 queue vertically at the intersection stop line. While this assumption does not represent a normal
144 queuing behavior and may not accurately represent the exact number of queued vehicle at a given
145 instant, it does not bias the delay estimation process over an entire queue formation and dissi-
146 pation process and is therefore a valid simplification when only considering delay estimations.
RR

147 In over-saturation conditions, the number of vehicles reaching the intersection exceeds the
148 number of vehicles that can be served by the traffic signal. This causes a growing residual queue to
149 occur, as illustrated in the lower diagram of Fig. 5. The overflow delay associated with this sit-
150 uation corresponds to the area between the line representing the arrivals that can be served at
151 capacity and the line representing the actual arrivals. In this case, Eqs. (3) and (4) can be derived
CO

152 to express the average delay over the number of vehicles discharged during the evaluation period
153 T.
3600T
v 
do ¼ 1 ð3Þ
2 c
UN

 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
do ¼ 900T ðX  1Þ þ ðX  1Þ2 ð4Þ

156 This model is time dependent, as the overflow delay increases with any increase in the evalu-
157 ation period. This is logical, as the residual queue keeps growing through the period. However,
158 while the formation of Eq. (4) expresses the total delay incurred during the time period T , it does
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx 9

159 not include the delay that is incurred by the vehicles that remain in queue after the conclusion of
160 T . It should be noted that Eq. (4) provides a general relationship that is valid for volume-to-
161 capacity ratios that are less than and/or greater than 1.0.

162 3.2. Shock wave delay model

F
163 Traffic flow can be characterized using flow, density and speed through an analogy with fluid
164 dynamics. Lighthill and Whitham (1955), as well as Richards (1956), made the first successful

OO
165 attempts at such a description. They both demonstrated the existence of traffic shock waves and
166 proposed a first theory of one-dimensional waves that could be applied to the prediction of
167 highway traffic flow behavior. Eqs. (5) and (6) represent their model. The first equation defines the
168 relation between volume, density, and speed that has been developed from the application of fluid
dynamics theory. Using Eq. (5), Eq. (6) was then developed to describe the speed at which a

PR
169
170 change in traffic characteristics, or shock wave, propagates along a roadway.
vi ¼ ki
ui ð5Þ
vj  vi
SWij ¼ ð6Þ
kj  ki
173
174
D
Using the model of Eqs. (5) and (6), Rorbech (1968) investigated the queue formation at in-
tersection approaches at the beginning of red intervals. Stephanopoulos and Michalopoulos
TE
175 (1979) further investigated the dynamics of queue formation and dissipation at isolated inter-
176 sections using the flow conversation principle of Eq. (5). In another study, Michalopoulos et al.
177 (1980) analyzed traffic dynamics between signalized intersections and demonstrated the existence
178 of shock waves propagating downstream of an intersection caused by the periodic operations of
EC

179 traffic signals. Michalopoulos et al. (1981) and Michalopoulos and Pisharody (1981) further de-
180 veloped a real-time control algorithm based on shock wave theory that minimizes total delay at
181 isolated intersections subject to constraints regarding maximum queue lengths on individual
182 approaches.
RR

183 The main difference between shock wave and deterministic queuing models is in the way ve-
184 hicles are assumed to queue at the intersection. While queuing analysis assumes vertical queuing,
185 shock wave analysis considers that vehicles queue horizontally. As illustrated in Fig. 6, the
186 consideration of the horizontal extent of a queue enables the capturing of more realistic queuing
187 behavior and the determination of the maximum queue reach. This is not possible with deter-
CO

188 ministic queuing models, as these models only track the number of queued vehicles, not their
189 spatial location.
190 In Fig. 6, the total travel time spent by all vehicles going through the intersection can be esti-
191 mated using the density and flow rate associated with each region. Since delay represents the added
UN

192 travel time caused by the traffic signal operation, the total delay incurred by the traffic within one
193 signal cycle can be estimated by calculating the difference between the total travel time with traffic
194 signals and the total overall travel time without traffic signals, as demonstrated in Eq. (7).
" #
X X
D ¼ TTuðsignalsÞ  TTuðno signalsÞ ¼ 3600 Ai ki  Ai kC ð7Þ
i¼A;B;C i¼A;B;C
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

10 F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx

Distance Shock waves

Xm Maximum SWI = Arriving vehicles /
Queue Reach Back of queue
Area C SWD = Front of platoon leaving
Xr Maximum intersection at saturation
Queue Size
SWR = Queued vehicles / Vehicles
SWI leaving queue at saturation

F
SWN SWN = Vehicles leaving queue at
SWR saturation / Arriving vehicles
Area B

OO
Area A

r g Time
SWD

tm(u) tc(u)
tr

PR
Fig. 6. Shock wave analysis for under-saturated approach.

196 Finally, Eq. (8) can be derived to compute the average total delay incurred by individual ve-
197 hicles due to the operation of traffic signals.
jxmðuÞ j
d ¼ 3600
½re
ðkj  ka Þ þ ðtmðuÞ þ tcðuÞ Þ
ðkd  ka Þ ð8Þ

199 with:
2
v
C


D
TE
1 v
re
s
xmðuÞ ¼
ð9Þ
3600 sðkj  ka Þ  vðkj  kd Þ
v
re
ðkj  kd Þ
tmðuÞ ¼ ð10Þ
EC

sðkj  ka Þ  vðkj  kd Þ

xmðuÞ ka  kd
tcðuÞ ¼ 3600 ¼ 3600jxmðuÞ j
ð11Þ
SWN vs
RR

203 Similar to the deterministic queuing models, the shock wave delay model of Eq. (8) assumes
204 that vehicles follow a non-random and consistent path and that all vehicles accelerate and de-
205 celerate instantaneously. These two elements are represented in Fig. 6 by the constant interval
206 between the lines representing the trajectories of individual vehicles and by the sharp angles in the
207 trajectories when a vehicle passes from one traffic zone to another. Similar to Eq. (2), the con-
CO

208 sequences of these two assumptions are that Eq. (8) only estimates uniform delay and assumes
209 that all delays are incurred on the approach side of an intersection. Finally, another common
210 element with Eq. (2) is the use of the effective signal interval durations to account for the start loss
211 and end gain.
UN

212 Shock wave theory can also be used to estimate approach delays in over-saturated conditions.
213 In this case, delay estimation is similar to the under-saturated case. As an example, Fig. 7 illus-
214 trates the spatial and temporal evolution of a queue in the first two cycles of operation of an over-
215 saturated signalized approach as seen from the shock wave analysis. In the figure, the shock waves
216 labeled SWi , SWR and SWN have the same origin as the similar shock waves shown in Fig. 6. In
217 this case, however, the over-saturation creates a new shock wave, SWS , locating the back of the
218 area containing fully stopped vehicles after the return of the red indication.
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx 11

Maximum queue reach

SWN
Distance within each cycle Shock waves
SWS SWI = Arriving vehicles / Back of
SWR queue
SWI SWR = Queued vehicles / Vehicles
SWN Area 5
leaving queue at saturation
SWR SWS Area 3 kd
xm SWN = Vehicles leaving queue at
SWI kj saturation / Arriving vehicles
2xc(o)

F
Area 2 SWs = Vehicles departing at saturation
Area 1 kd Area 4 Area 6 / Back of residual queue
kj kj kd
Xm(o) Xc(o)

OO
re ge re ge Time

Tm(o) Tm(o)

Fig. 7. Shock wave analysis for over-saturated approach.

PR
219 In Fig. 7, the areas 1, 3 and 4 all contain vehicles queued at the approach jam density, while the
220 areas 2, 5 and 6 are characterized by vehicles moving at saturation flow and with a density
221 corresponding to the discharge density. Based on these observations, it can be determined that the
222 area of regions 3 and 5 is identical to the area of regions 1 and 2. If the area of regions 4 and 6 is
223 known, the estimation of the delay incurred by vehicles on the approach is then only a matter of
224
225
knowing the number of cycles over which the estimation is performed. D
Similar to the under-saturated scenario, Eqs. (12) and (13) can be derived from Fig. 7 to es-
TE
226 timate the total travel time incurred by vehicles traveling on the over-saturated intersection ap-
227 proach with and without the traffic signals.

   X
xmðoÞ  xmðoÞ þ xcðoÞ N 1
TToðsignalsÞ ¼ N re
kj þ ge
kd þ i
xcðoÞ
ðre
kj þ ge
kd Þ ð12Þ
2 2
EC

i¼1
  X N 1
xmðoÞ ðxmðoÞ þ xcðoÞ Þ
TToðno signalsÞ ¼ N
ka re þ ge þ i
xcðoÞ
ðre þ ge Þ
ka ð13Þ
2 2 i¼1
230 with:
RR

 
C
ðv  sÞ
xcðoÞ ¼ ð14Þ
3600kj
232 Finally, the average over-saturation delay is computed as the difference between Eqs. (12) and (13)
CO

233 divided by the number of vehicle departures within the analysis period T , as indicated in Eq. (15).
n o
x ðxmðoÞ þxcðoÞ Þ
N mðoÞ 2

r e
ðk j  k a Þ þ 2
ge
ðkd  ka Þ
d¼
s
ge
N=3600
PN 1
i
xcðoÞ
fre
ðkj  ka Þ þ ge
ðkd  ka Þg
UN

þ i¼1 ð15Þ
s
ge
N=3600

235 3.3. Steady-state stochastic delay models

236 While deterministic queuing and shock wave delay models both assume uniform arrivals,
237 stochastic delay models attempt to account for the randomness of vehicle arrivals. One of the
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

12 F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx

238 fundamental and most often quoted models in the literature is the Webster model (Webster, 1958).
239 This model, which is expressed by Eq. (16), is comprised of three terms. The first term estimates
240 the average approach delay assuming uniform arrivals, which is consistent with Eq. (2) that was
241 derived earlier in the paper. The second term considers the additional delays attributed to the
242 randomness of vehicle arrivals. The third term is an empirical correction factors that reduces the
243 estimated delay by 5–15%, to be consistent with simulation results.

F
 2
c 1=3
C 1  gCe X2 ge
d¼
þ  0:65 2 X 2þ C ð16Þ

OO
2 1  X ge C 2vð1  X Þ v

245 Following WebsterÕs work, other stochastic models were proposed. These include the models by
246 Miller (1963), Newell (1960, 1965), McNeil (1968), and Heidemann (1994). These models all share
247 the same general basic assumptions. First, they all consider that the number of arrivals in a given

PR
248 time interval follows a known distribution, typically a Poisson distribution, and that this distri-
249 bution does not change over time. This implies that these models could not be applied to estimate
250 delays at intersections within a coordinated system, where arrivals are platooned as a result of
251 upstream traffic signals. Second, they all assume that the headways between departures from the
252 stop line follow a known distribution with a constant mean, or are identical. Third, while it is
253
254
D
recognized that temporary over-saturation may occur due to the randomness of arrivals, it is
assumed that the system remains under-saturated over the analysis period. Fourth, the system is
TE
255 assumed to have been running long enough to allow it to have settled into a steady state. Fifth, all
256 these models still consider that vehicles decelerate and accelerate instantaneously and thus, that all
257 drivers behave similarly.
EC

258 3.4. Time-dependent stochastic delay models for under-saturated and over-saturated conditions

259 A main consequence of the stochastic delay modeling described in the previous section is that
260 the estimated delays tend to infinity as traffic demand approaches saturation (v=c ratio of 1.0).
RR

261 This was considered as a weakness by many researchers (Akcelik, 1988; McShane and Roess,
262 1990; Fambro and Rouphail, 1997). Fig. 8 illustrates how delay estimation models should the-
263 oretically behave under varying demand levels. For low v=c ratios, the models should produce
264 delay estimates that are similar to those produced by deterministic queuing delay models as-
265 suming constant uniform arrivals. As the load increases, a larger proportion of the delay is caused
CO

266 by the randomness of vehicle arrivals and attributed to the inability to clear all queued vehicles in
267 some cycles. As the v=c ratio approaches 1.0, the models should finally not tend to infinity, but
268 should instead produce estimates that become tangent to the deterministic over-saturation model
269 of Eq. (2).
UN

270 The concept of a general time-dependent delay model was originally conceived by Robertson
271 (1979) and further enhanced by Kimber and Hollis (1979) using the coordinate transformation
272 technique illustrated in Fig. 8. This technique transforms the equation defining a steady-state
273 stochastic delay model so that it becomes asymptotic to the deterministic over-saturation model of
274 Eq. (4). Although there is no rigorous theoretical basis for this approach (Hurdle, 1984), empirical
275 evidence indicates that these models yield reasonable results. This explains why numerous time-
276 dependent delay formulas based on the coordinate transformation technique have been proposed
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx 13

Fitted time-
dependent
Average approach delay

delay model

F
Steady-state
stochastic delay

OO
model Deterministic
oversaturation
delay model

PR
Deterministic uniform delay model
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3
Volume-to-capacity ratio

Fig. 8. Stochastic time-dependent delay model concept.

277
278
D
over the years (Brilon and Wu, 1990; Akcelik, 1981, 1988; Akcelik and Rouphail, 1993; Fambro
and Rouphail, 1997) and have been incorporated into a number of capacity guides, such as those
TE
279 from the United States (TRB, 1994, 1997), Australia (Akcelik, 1981) and Canada (ITE, 1995).
280 The capacity guide delay models currently used in the United State, Australia and Canada are
281 all similar. The general form of these models is expressed by Eqs. (17)–(20), with Table 1 indi-
282 cating the specific values assigned to the parameters in each model.
EC

d ¼ d1
fPF þ d2 þ d3
fr ð17Þ

284 with:
RR

Table 1
Capacity guide delay model parameters
Model Parameters
CO

fr N m k I T Xo fPF
a
Australian (1981) 0 0 6 or 12 n=a n=a Variable 0:67 þ sge =600 1.0
Canadian (1995) 0 0 4 n=a n=a Variable 0 1.0 or Eq. (20)b
HCM (1994) 0 2 4 to 16c n=a n=a 15 min 0 1.0, 0.85, or Eq. (20)d
HCM (1997) 1 0 8 0.04–0.50e 1.0f Variable 0 Eq. (20)
UN

a
12 for random arrivals; 6 when platooning occurs.
b
1.0 for isolated intersections; Eq. (20) in other cases.
c
Function of arrival type (16 for random arrivals, 12 for favorable or non-favorable progression, 8 for very poor or
highly favorable progression, 4 for very unfavorable progression).
d
1.0 for pre-timed, non-coordinated signals; 0.85 for actuated, non-coordinated systems; Eq. (20) for coordinated
systems.
e
0.50 for pre-timed signals; 0.04–0.50 for actuated controllers.
f
1.0 for isolated intersection only.
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

14 F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx

 2
1  gCe
d1 ¼ 0:50C
 ð18Þ
1  gCe
minðX ; 1:0Þ
" rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi#
mkI
d2 ¼ 900X n T ðX  1Þ þ ðX  1Þ2 þ
ðX  Xo Þ ð19Þ
cT

F
ð1  P Þfp
fPF ¼ ð20Þ

OO
1  gCe
288 The models defined in Table 1 all assume steady-state traffic conditions. They estimate delays
289 under stochastic equilibrium conditions, when the arrival and departure flow rates have been sta-
290 tionary for an indefinite period of time. They also assume that the number of arrivals in a given
291 interval follow a Poisson distribution that remains constant over time, and that the headways be-

PR
292 tween departures have a known distribution with a constant mean value. Finally, these models do
293 not include all of the delay incurred by either arriving or discharged vehicles during the control
294 period considered. These models only include the delays incurred up to the end of the control period
295 and do not consider the additional delay that is incurred as the queue is being served. However, it
296 can be demonstrated that accounting for the additional delay that is incurred during the decay of the
297
298
D
queue and averaging over all the vehicle departures (in this case the arrival rate), the average delay is
the identical. Specifically, if it is assumed that there are no vehicle arrivals at the conclusion of the
TE
299 analysis period, then the average delay can be computed using Eq. (21), which is identical to Eq. (4).
h i
0:5ðv  cÞ
3600T
T þ ðvcÞT  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
c 1800ðv  cÞT
d¼ ¼ ¼ 900T ðX  1Þ þ ðX  1Þ2 ð21Þ
T
v c
EC

301 Similar to the models presented in the previous sections, the models of Table 1 were developed
302 assuming instantaneous accelerations and decelerations to simplify the delay estimation. In this
303 case, however, a number of relations have been proposed to estimate the proportion of stopped
304 and acceleration/deceleration delays. For instance, the 1994 HCM estimates the stopped delay at
RR

305 76% of the total estimated delay. Both Teply (1989) and Olszewski (1993) agreed that this factor
306 was incorrect for very low and very high signal delays and proposed alternative evaluation
307 methods. Teply first recommended a multiplicative adjustment factor that varies 0.36–0.83 and is
308 a function of the duration of the red interval, which was later adopted in the 1995 Canadian
309 Capacity Guide for Signalized Intersections, while Olszewski recommended a subtractive ad-
CO

310 justment factor varying with the approach speed.

311 The main differences between the various capacity guide models are in the way they attempt to
312 consider non-Poisson arrivals:
UN

(a) In the Australian model, adjustments for the type of arrivals are considered by simply altering
the value assigned to the parameter m. A value of 12 is used for random arrivals, while a value
of 6 is used for platooned arrivals.
(b) In the 1994 HCM model, both the parameter m and the progression factor fPF are modified.
In this case, values ranging from 4 to 16 are used for the parameter m, while alternative sug-
gested values for the progression factor fPF consider not only varying arrival patterns, but also
the use of pre-timed or actuated controllers at coordinated and non-coordinated intersections.
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx 15

(c) In the 1995 Canadian model, the type of arrivals only affects the progression factor fPF as the
parameter m is assigned a fixed value of 4.
(d) The 1997 HCM model fixes the parameter m at 8, but introduces the parameters k and I to
explicitly consider the effect of signal controller types and of arrival patterns affected by up-
stream signalized intersections. As a result, the product mkl, which is always equal to 4.0 in
the Canadian model, could vary in this case from 0.32 to 4.0.

F
326 In addition to the above differences, the 1994 HCM model is the only one to assign a non-zero

OO
327 value to the parameter n. This parameter was added to compensate for the modelÕs assumptions of
328 a zero initial queue and fixed 15-min analysis period. While the 1997 HCM model reassigns a
329 value of zero to the parameter n, this model is also the only one to define a non-zero value to the
330 parameter fr and explicitly considers the residual delays for over-saturation queues that may have

PR
331 existed before the analysis period.

332 3.5. Microscopic simulation delay models

333 Microscopic traffic simulation models have the ability to track individual vehicle movements
334 within simulated street networks. Vehicle tracking is usually done using car-following, lane-
335
336
D
changing, and gap-acceptance logic. This allows such models, among other things, to consider
virtually any traffic conditions, ranging from highly under-saturated to highly over-saturated
TE
337 conditions.
338 Because of their ability to track the movements of individual vehicles, microscopic simulation
339 models can determine the delay incurred by an individual vehicle while traveling a network of
340 links with different characteristics by comparing simulated and ideal travel times. No specific
EC

341 formulas are therefore required to evaluate uniform and overflow delay, or delays in under-sat-
342 urated and over-saturated traffic conditions, thus allowing for the evaluation of complex traffic
343 situations. In addition, the ability to record vehicle speed and position on a second-by-second
344 basis further allows the recording of speed profiles and the direct estimation of deceleration,
RR

345 stopped and acceleration delays.

346 For this study, the INTEGRATION microscopic traffic simulation software is used to obtain
347 simulated delay estimates. This simulation model features an integrated, dynamic traffic simula-
348 tion and traffic assignment model. Within the model, delay is estimated for each individual vehicle
349 by calculating, for each traveled link, the difference between the vehicleÕs simulated travel time and
CO

350 the travel time that the vehicle would have experienced on the link at free speed, as expressed by
351 Eq. (22) (Rakha et al., in press).
Z T 
uf  uðtÞ
D¼ dt ð22Þ
uf
UN

t¼0

354 4. Test scenario

355 To evaluate the consistency of delay estimates from the various models presented in this paper,
356 delay evaluations were carried out for the sample network of Fig. 9 using the INTEGRATION
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

16 F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx

Simulation parameters Signal operation

• Simulation time: 15 min • Cycle time: 60 s
• Saturation flow: 1800 veh/h • Effective red interval : 30 s
• Free-flow speed: 60 km/h • Effective green interval: 30 s

F
OO
PR
2 km 1 km

Fig. 9. Delay evaluation scenario.

357
358
359
D
microscopic traffic simulation software and the deterministic, shock wave, Webster and capacity
guide delay models defined by Eqs. (2), (4), (8), and (15)–(17), respectively. The example of Fig. 9
features a single-lane intersection approach at a fixed-timed traffic signal operating with a 60-s
TE
360 cycle length and a 30-s effective green interval. A 1-km exit link is also included to allow the
361 INTEGRATION model to capture the delays incurred by vehicles accelerating as they leave the
362 intersection and to compile total approach delay, as is done in the analytical models.
363 For each model, delay evaluations were specifically carried out for v=c ratios varying between
EC

364 0.1 and 1.4. This allowed evaluations to be conducted for a range of traffic conditions extending
365 from highly under-saturated to highly over-saturated conditions. For each scenario, vehicle ar-
366 rivals were further assumed to follow a random process with a constant average arrival rate. For
367 the analytical models, this only required a single application of the equations defining each model.
RR

368 For the INTEGRATION model, however, 10 replications were made of each scenario to account
369 for the stochastic nature of the modelÕs simulation process. Thus, unless otherwise noted, the
370 results from the INTEGRATION simulation model are an average of 10 simulations, while the
371 results from the analytical models are the delays reported by Eqs. (2), (4), (8), and (15)–(17).
CO

372 Finally, to ensure that appropriate comparisons are made for all v=c scenarios considered,
373 particularly in the over-saturation range, simulation were only carried out for a 15-min control
374 period. This 15-min period was imposed by the fact that some of the time-dependent delay models
375 considered, namely the 1994 HCM model, require a 15-min analysis period.
UN

376 5. Test results

377 Figs. 10 and 11 illustrate the results of the delay estimations that were carried out for the
378 example of Fig. 9. Fig. 10 illustrates the delays that were estimated over the entire range of v=c
379 ratios considered, while Fig. 11 provides a more detailed look at the delay estimates for v=c ratios
380 below 1.0.
TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx 17

180
Webster (1958)
Deterministic queuing / Shock wave
160
Deterministic over-saturation / Shock wave
Australian capacity guide (1981)
Average approach delay (s\veh)

140
Highway Capacity Manual (1994)
Canadian capacity guide (1995) / Highway Capacity Manual (1997)
120

F
INTEGRATION

100

OO
80

60

PR
40

20

0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4
Volume-to-capacity ratio (v/c)
D
Fig. 10. Delay estimates for under- and over-saturated conditions.
TE
EC

50
Webster (1958)
45 Deterministic queuing / Shock wave
Australian capacity quide (1981)
40 Highway Capacity Manual (1994)
Average approach delay (s\veh)

RR

Canadian capacity guide (1995) / Highway Capacity Manual (1997)

35 INTEGRATION - Average results

30

25
CO

20

15

10
UN

0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Volume-to-capacity ratio (v/c)

Fig. 11. Delay estimates for under-saturated conditions only.

 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

18 F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx

381 5.1. Consistency of delay estimates for under-saturated conditions

382 In Fig. 11, it is first observed that there is a general agreement between all the analytical delay
383 models considered when these are applied to the analysis of signalized intersections with very low
384 v=c ratios. However, it is also observed that this agreement tends to decrease with increasing v=c
385 ratios. For example, the difference between the minimum and maximum delay estimates from the

F
386 various analytical models does not exceed 6.0% for v=c ratios below 0.2. For a v=c ratio of 0.4, the
387 maximum difference increases to 14.4%, while differences exceeding 30% are observed for v=c

OO
388 ratios above 0.6.
389 Fig. 11 further illustrates that the deterministic queuing and shock wave models always produce
390 the lowest estimates. This is due to the fact that these two models consider only uniform arrivals.
391 As a result, these models cannot consider the potential for additional delays that arise from the
probability of having temporary over-saturation delays due to surges of arriving vehicles. It is also

PR
392
393 observed that the deterministic and stochastic models produce relatively similar delay estimates at
394 very low v=c ratios. This is an indication that the randomness of vehicle arrivals can be neglected
395 when estimating delays for highly under-saturated conditions and that the use of either deter-
396 ministic or stochastic models is valid in such conditions.
397 A more detailed analysis of Fig. 11 further reveals that the delays predicted by the Webster,
398
399
D
1995 Canadian Capacity Guide and 1997 HCM models are virtually identical for v=c ratios below
0.8. Above this ratio, the delays from the Webster model tend towards infinity for v=c ratios
TE
400 approaching 1.0. This is a well-known behavior of the model and other similar steady-state
401 stochastic models that invalidates their use when the v=c ratio tends to 1.0. In comparison, the
402 delays estimates at a v=c ratio of 1.0 from the various capacity guide models oscillate between 43.7
403 and 45.0 s per vehicle, while the average delays estimated by the INTEGRATION model ap-
EC

404 proach 38.8 s per vehicle.

405 It should be noted at this point that although the vehicles are generated randomly at their
406 origin within the INTEGRATION software, the level of randomness during a simulation is re-
407 duced as a result of the car-following behavior that is applied while vehicles are traveling along a
RR

408 link. On one hand, the car-following logic allows a vehicle with a large distance headway with the
409 vehicle in front of it to speed up to reduce the headway. On the other hand, the logic forces a
410 vehicle that has been generated close to its predecessor to slow down to maintain a safe car-
411 following distance. Consequently, it is expected that the simulation delays would be slightly less
412 than the random delay estimates.
CO

413 While a general agreement exists between the various capacity guide delay models in the trend
414 of increasing delays with increased v=c ratios, differences are also observed. The main differences
415 are more particularly found with the Australian Capacity Guide and 1994 HCM models, which
416 both consistently produce lower delay estimates than the 1997 HCM and 1995 Canadian Capacity
UN

417 Guide models. These differences are primarily due to the coordinate transformation technique
418 that was applied in each case to produce a model providing delay estimates that are asymptotic to
419 the deterministic over-saturation delay model of Eq. (2). The 1995 Canadian Capacity Guide and
420 1997 HCM models both produce identical results, as these two models are similar in this case. As
421 explained earlier, the main differences between these two models are in the values assigned to the
422 parameters fr , m, k, and l in Eqs. (17) and (19). In this case, however, the parameter fr has no
423 impact since it is assumed that no over-saturation queue exists before the start of the analysis
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx 19

240
Australian capacity guide (1981)
Highway Capacity Manual (1994)
200 Canadian capacity guide (1995) / Highway Capacity Manual (1997)
INTEGRATION - Average results
INTEGRATION - Individual simulation results
Average approach delay (s\veh)

F
160

OO
120

80

PR
40

0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4
D
Volume-to-capacity ratio (v/c)
TE
Fig. 12. Comparison of simulated and capacity guide delay estimates.

424 period. Both models also exhibit the same value for the mkl product in Eq. (19). While the Ca-
425 nadian model fixes the value of m at 4 and ignores the parameters k and l, the HCM model assigns
EC

426 a value of 8 to the parameter m, a value of 0.5 to the parameter k to account for pre-timed signal
427 operation, and ignores the parameter l.
428 The analysis of Fig. 11 further reveals that the average delay estimates from the INTEGRA-
429 TION simulation model are in general agreement with the estimates from the various capacity
RR

430 guide models. This agreement is particularly evident in Fig. 12, which superimposes the delay
431 estimates from the individual INTEGRATION simulation runs to the corresponding capacity
432 guide delay estimates. As can be observed, the estimates from the four capacity guide delay
433 models all fall within the range of delays obtained from the individual simulation runs. This result
434 thus either validates the under-saturated delay models defined in the various capacity guides, or
CO

435 the INTEGRATION simulation process, depending on the point of view taken.
436 In Figs. 11 and 12, while the estimates from the various analytical models generally increase
437 monotonically with higher v=c ratios, it is observed that there is more variability in the delay
438 estimates from the INTEGRATION simulation model than from the analytical models, partic-
UN

439 ularly at low v=c ratios. A portion of this variability is attributed to the stochastic nature of the
440 INTEGRATION simulation process, while another portion is attributed to the discrete nature of
441 the simulation process, which only considers complete vehicles and not fractions of vehicles.
442 A main effect of considering discrete vehicle departures is to increase the sensitivity of delay
443 estimates to specific vehicle arrival times. As an example, Table 2 computes the delay associated
444 with three arriving flows for the example of Fig. 9. In all cases, the vehicles arrive at the inter-
445 section with a constant 5-s headway and depart at a constant 2-s saturation flow headway. The
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

20 F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx

Table 2
Sensitivity of delay estimates to arrival patterns
Vehicle Cycle 1 Cycle 2 Cycle 3
Arrival Departure Delay Arrival Departure Delay Arrival Departure Delay
1 0 30 30 1 30 29 4 30 26
2 5 32 27 6 32 26 9 32 23

F
3 10 34 24 11 34 23 14 34 20
4 15 36 21 16 36 20 19 36 17

OO
5 20 38 18 21 38 17 24 38 14
6 25 40 15 26 40 14 29 40 11
7 30 42 12 31 42 11 34 42 8
8 35 44 9 36 44 8 39 44 5
9 40 46 6 41 46 5 44 46 2
10 45 48 3 46 48 2 49 49 0

PR
11 50 50 0 51 51 0 54 54 0
12 55 55 0 56 56 0 59 59 0
Total delay 165 Total delay 155 Total delay 10.5
Average delay 13.8 Average delay 12.9 Average delay 126

446
447
D
only difference between the arrival patterns is in the time at which the first vehicle arrives relative
to the start of the effective green. In the second cycle, vehicle arrivals are all offset by 1 s relative to
TE
448 the first cycle, while arrivals in the third cycle are offset by 4 s. As can be observed, a 1-s offset is
449 sufficient in this case to decrease the estimated average delay per cycle from 13.8 to 12.9 s, while a
450 four-second offset further results in a 10.50 s average delay. If delay calculations were made using
451 the deterministic queuing or shock wave models of Eqs. (2) and (8), the estimated average delay
EC

452 would then be 12.5 s for all three cycles. In the case of the Webster and capacity guide models,
453 similar estimates would also be obtained from all cycles, since these models only consider average
454 flow rates and traffic patterns.
455 Fig. 13 provides a more detailed look at the variability of simulated delay estimates. The figure
RR

456 plots for each v=c ratio the standard deviation and coefficient of variation (COV) (standard de-
457 viation divided by mean) of delay estimates that are obtained from the ten repetitions per v=c
458 ratio. As can be observed, significant variability as measured by the COV in delay estimates is
459 observed both at low v=c ratios and around a v=c ratio of 1.0. While the literature documents the
460 higher variability in delay as the v=c ratio approaches 1.0, it does not describe the high variability
CO

461 that is observed for extremely low traffic demands.

462 The two observed peaks of high COV that are observed in Fig. 13 can be attributed to two
463 factors. First, at very low v=c ratios there are larger temporal headways between successive ve-
464 hicles and thus the standard deviation is higher than the case of higher v=c ratios. This provides
UN

465 greater opportunities for vehicle spacing to be varied, as was demonstrated earlier in Table 2. As
466 traffic demand increases and the average temporal headway between successive vehicles decreases,
467 the impact of the vehicle departure randomness on delay estimates decreases, thus explaining the
468 reduced variability that is observed for v=c ratio less than 0.7. In addition, given that the mean
469 delay at low v=c ratios is very small and that the COV is computed as the ratio of the standard
470 deviation to the mean, it is not surprising to observe higher coefficients of variation at very low v=c
471 ratios.
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx 21

18 0.30

16
0.25
14

Coefficient of variation
12 0.20

F
standard deviation

Coefficient of variation
10

OO
0.15
8

6 0.10

PR
Standard deviation 0.05
2

0 0.00
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5
Volume-to-capacity ratio (v/c)
D
Fig. 13. Coefficient of variation of simulated delay estimates over range of traffic conditions considered.
TE
472 Another element to consider when interpreting the above results is the fact that the INTE-
473 GRATION car-following logic reduces the level of randomness of vehicle arrivals. As was ex-
474 plained earlier, the logic reduces differences in vehicle spacing. On one hand, the logic allows a
EC

475 vehicles with large distance headway with the vehicle in front of it to speed up to reduce the
476 headway. On the other hand, the logic forces a vehicle that has been generated close to its pre-
477 decessor to slow down to maintain a safe distance.
478 For v=c ratios above 0.7, signal cycle over-saturation is the main factor responsible for the
RR

479 increase in delay variability. As demand approaches saturation, there is an increased probability
480 that surges in vehicle arrivals may cause temporary over-saturation of a signal cycle. Since higher
481 delays typically result from over-saturated cycles, an increased variability in delay estimates thus
482 results as the probability of having over-saturated cycles increases. This variability reaches its
483 peak at a v=c ratio of 1.0, as demand fluctuations then create equal probabilities for over-satu-
CO

484 rated and under-saturated cycles. Finally, the variability decreases again as the v=c ratio increases
485 past 1.0 as the increasing congestion and number of over-saturated cycles constrain the freedom
486 of movements of vehicles and lead to more uniform flow patterns.
UN

487 5.2. Consistency of delay estimates for over-saturated conditions

488 For the over-saturated domain, the results of Fig. 10 indicate that there is a general agreement
489 between the INTEGRATION simulation model, the 1981 Australian Capacity Guide, the 1995
490 Canadian Capacity Guide and the 1997 HCM delay models. For these four models, the delay
491 estimates gradually approach the delays predicted by the deterministic over-saturation model of
492 Eq. (2) as the v=c ratio increases. The only exception to this trend is for the 1994 HCM model,
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

22 F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx

493 which does not produce a delay curve that is asymptotic to the deterministic over-saturation
494 model. This behavior is explained by the fact that this model is the only one to include the X 2
495 adjustment parameter in Eq. (19), as summarized in Table 1. As explained earlier, this parameter,
496 which was later removed in the 1997 model, was introduced to compensate for the assumption of
497 zero initial queue and fixed 15-min analysis period.
498 In this case, Fig. 12 again emphasizes the consistency of delay estimates from the INTE-

F
499 GRATION, the 1981 Australian Capacity Guide, the 1995 Canadian Capacity Guide and the
500 1997 HCM delay models. Similar to the under-saturation case, it is observed that the delays es-

OO
501 timated from these three capacity guide models fall within the range of delays estimated by the
502 INTEGRATION simulation model. In this case, it is interesting to note that the INTEGRA-
503 TION model always produces the lowest estimates and that the difference between the various
504 models dissipates as the v=c ratio increases. The lower delay estimates from the INTEGRATION
software can be attributed to the lower level of randomness that results from travel from the entry

PR
505
506 node to the traffic signal stop line. More consistent results between the various models are ob-
507 served as the v=c ratio increases since the randomness in vehicle arrivals is then gradually ab-
508 sorbed by the queue formed upstream of the traffic signal, thus creating more uniform traffic
509 conditions across the various models.

510 6. Conclusions
D
TE
511 This paper compared the delays predicted by the INTEGRATION microscopic traffic simu-
512 lation software and a number analytical delay models on a one-lane approach to a pre-timed
513 signalized intersection approach for traffic conditions ranging from under-saturation to over-
EC

514 saturation. The analytical models that were compared were representative of deterministic
515 queuing, shock wave, steady-state stochastic and time-dependent stochastic delay models. For the
516 steady-state stochastic models, the Webster model was used as an example. For the time-de-
517 pendent stochastic models, the models defined in the 1981 Australian Capacity Guide, the 1995
RR

518 Canadian Capacity Guide and the 1994 and 1997 HCM were considered.
519 The delay estimates predicted by each model were compared over a range of v=c ratios ex-
520 tending from 0.1 to 1.4 to assess their consistency. Over this range, the delay models from the 1981
521 Australian Capacity Guide, the 1995 Canadian Capacity Guide, the 1997 HCM, and the IN-
522 TEGRATION microscopic traffic simulation model produced delay estimates that generally agree
CO

523 with each other. Depending on the point of view considered, it can be concluded that these results
524 either validate the delay models currently used in the capacity guides from the United States,
525 Canada and Australia, or validates the INTEGRATION microscopic simulation software. In
526 addition, it was further determined that all the delay models considered in this paper produce
UN

527 relatively consistent delay estimates when applied to the analysis of under-saturated signalized
528 intersections with v=c ratios below 0.6, thus indicating the validity of the use of all these models in
529 such conditions.
530 While the study indicates a strong consistency between the time-dependent stochastic delay
531 models used in the more recent capacity guide for signalized intersections and the delays estimated
532 by the INTEGRATION microscopic traffic simulation model for the case considered, efforts
533 should be made to evaluate this consistency for more complex situations. In particular, the
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx 23

534 consistency of delay models should be evaluated for multi-lane approaches, intersections con-
535 trolled by actuated controllers, and intersections where non-random vehicle arrivals occurs as a
536 result of signal coordination with upstream intersections. The impact of varying driver behavior
537 should also be investigated, as this may impact the saturation flows used in the various delay
538 models, as well as the approach speeds and amount of time lost every cycle due to driver reaction
539 time.

F
OO
540 Acknowledgements

541 The authors would like to acknowledge the effort of the late Dr. Michel Van Aerde in the
542 realization of this study. Dr. Van Aerde acted as a supervisor for Dr. Youn-Soo Kang during the
initial stages of his doctoral studies at Virginia Tech.

PR
543

544 References

545 Akcelik, R., 1981. Traffic Signals: Capacity and Timing Analysis. Research Report 123, Australian Road Research
546
547
548
Board, Melbourne, Australia. D
Akcelik, R., 1988. The highway capacity manual delay formula for signalized intersections. ITE Journal 58 (3), 23–27.
Akcelik, R., Rouphail, N.M., 1993. Estimation of delays at traffic signals for variable demand conditions.
TE
549 Transportation Research, Part B 27 (2), 109–131.
550 Brilon, W., Wu, N., 1990. Delays at fixed-time traffic signals under time-dependent traffic conditions. Traffic
551 Engineering and Control 31 (12), 623–631.
552 Fambro, D.B., Rouphail, N.M., 1997. Generalized delay model for signalized intersections and arterial streets. In:
553 Transportation Research Record, No. 1572, TRB. National Research Council, Washington, DC, pp. 112–121.
EC

554 Heidemann, D., 1994. Queue length and delay distributions at traffic signals. Transportation Research B 28B, 377–389.
555 Hurdle, V.F., 1984. Signalized intersection delay models––A primer for the uninitiated. In: Transportation Research
556 Record, No. 971, TRB. National Research Council, Washington, DC, pp. 96–105.
557 ITE, 1995. In: Teply, S. (Ed.), Canadian Capacity Guide for Signalized Intersections, second ed.. Institute of
RR

558 Transportation Engineers, District 7, Canada.

559 Kimber, R.M., Hollis, E.M., 1979. Traffic queues and delays at road junctions. Laboratory Report 909, Transport and
560 Road Research Laboratory, Crowthorne, UK.
561 Lighthill, M.J., Whitham, G.B., 1955. On kinematic wave II, a theory of traffic flow on long crowded roads.
562 Proceedings of Royal Society of London 229 (1178), 317–345.
563 McNeil, D.R., 1968. A solution to the fixed-cycle traffic light problem for compound poisson arrivals. Journal of
CO

564 Applied Probability 5, 624–635.

565 McShane, W.R., Roess, R.P., 1990. Traffic Engineering. Prentice Hall, Englewood Cliffs, NJ.
566 Michalopoulos, P.G., Stephanopoulos, G., Pisharody, V.B., 1980. Modelling of traffic flow at signalized links.
567 Transportation Science 14 (1), 9–41.
568 Michalopoulos, P.G., Pisharody, V.B., 1981. Derivation of delays based on improved macroscopic traffic models.
UN

569 Transportation Research, Part B 15B (5), 299–317.

570 Michalopoulos, P.G., Stephanopoulos, G., Stephanopoulos, G., 1981. An application of shock wave theory to traffic
571 signal control. Transportation Research, Part B 15 (1), 35–51.
572 Miller, A.J., 1963. Settings for fixed-cycle traffic signals. Operations Research Quarterly 14, 373–386.
573 Newell, G.F., 1960. Queues for a fixed-cycle traffic light. Annals of Mathematical Statistics 31 (3), 589–597.
574 Newell, G.F., 1965. Approximation methods for queues with application to the fixed-cycle traffic Light. SIAM Review
575 7 (2), 223–240.
 TRB 396 No. of Pages 24, DTD = 4.3.1
21 January 2003 Disk used ARTICLE IN PRESS SPS, Chennai

24 F. Dion et al. / Transportation Research Part B xxx (2003) xxx–xxx

576 Olszewski, P.S., 1993. Overall delay, stopped delay, and stops at signalized intersections. Journal of Transportation
577 Engineering, ASCE 119 (6), 835–852.
578 Rakha, H., Dion, F., Sin, H., in press. Field evaluation of energy and emission impacts of traffic flow improvement
579 projects using GPS data: issues and proposed solutions. Transportation Research Record.
580 Richards, R.I., 1956. Shock waves on the highway. Operations Research 4, 42–51.
581 Robertson, D.I., 1979. Traffic models and optimum strategies of control. In: Proceedings of the International
582 Symposium on Traffic Control Systems, pp. 262–288.

F
583 Rorbech, J., 1968. Determining the length of the approach lanes required at signal-controlled intersections on through
584 highways. Transportation Research 2, 283–291.

OO
585 Stephanopoulos, G., Michalopoulos, P.G., 1979. Modelling and analysis of traffic queue dynamics at signalized
586 intersections. Transportation Research, Part A 13, 295–307.
587 Teply, S., 1989. Accuracy of delay surveys at signalized intersections. In: Transportation Research Record, No. 1225,
588 TRB. National Research Council, Washington, DC, pp. 24–32.
589 TRB, 1994. Highway Capacity Manual. Special Report 209, third ed. Transportation Research Board, National
590 Research Council, Washington, DC.

PR
591 TRB, 1997. Highway Capacity Manual. Special Report 209, third ed. Transportation Research Board, National
592 Research Council, Washington, DC.
593 M. Van Aerde and Associates, 2001. INTEGRATION Rel. 2.30 for Windows––UserÕs Guides. Volumes I and II, M.
594 Van Aerde and Associates, Ltd., Kingston, Ontario, Canada.
595 Webster, F.V., 1958. Traffic signal settings. Road Research Technical Paper No. 39, Road Research Laboratory, Her
596 Majesty Stationary Office, London, UK.
D
TE
EC
RR
CO
UN

View publication stats