Article

Transportation Research Record

2018, Vol. 2672(8) 108–120
Ó National Academy of Sciences:
Bus Capacity Estimation using Stochastic Transportation Research Board 2018
Article reuse guidelines:
Queuing Models for Isolated Bus Stops in sagepub.com/journals-permissions
DOI: 10.1177/0361198118777358

China journals.sagepub.com/home/trr

Chao Wang1,2,3, Zhirui Ye1,2, Jon D. Fricker3, Yunlong Zhang4,

and Satish V. Ukkusuri3

Abstract
When serving passengers at a bus stop, a dwelling bus may cause a bottleneck that constrains traffic flow near the bus stop.
A major element in bus system capacity is bus stop capacity (veh/h). Therefore, this study proposes a method estimating the
capacity of bus stops isolated from the influences of traffic signals and other bus stops. Data collected for the seven most
common bus stop configurations in China were used to validate the proposed model. The results indicated that the arrival
process obeys a Poisson distribution and the service times fit a lognormal distribution. Stochastic queuing models for both
single-berth and multi-berth stops were developed to estimate bus stop capacity and bus delay time. To enable comparison,
the Highway Capacity Manual (HCM) model for bus stop capacity estimation was also used. The results showed that the pro-
posed method was more accurate and reliable, with a mean absolute percentage error (MAPE) of 8.44% compared with a
MAPE of 17.42% from the HCM method. In addition, the proposed method reduced root mean squared error (RMSE) by
58.8%, mean absolute error (MAE) by 57.9%, and variance of absolute percentage error (VAPE) by 52.6% compared with the
values from the HCM model. Sensitivity analyses were also conducted to investigate the effects of bus arrival rate, bus service
time, and the number of bus berths on the capacity of bus stops. The results indicated that the proposed method can be used
as a guide for designing bus stop facilities under specified conditions.

With the advance of modern technologies and the place. Under light ridership conditions, bus stops seem to
requirements to improve levels of service for public be a rather simple mechanism to model. Passenger inter-
transportation, there has been a transition from ‘‘vehicle- arrival times and bus headways are large, so the number
oriented’’ (private transportation) to ‘‘people-oriented’’ of bus berths is enough for dwelling buses. In high
(public transportation) in the development of urban demand situations, however, this is not the case. For
transportation networks (1–3). Urban public transporta- example, at peak periods at a bus stop, if there is no
tion systems perform an essential function in the mobi- unoccupied berth for an arriving bus, the bus must wait
lity of citizens in metropolitan areas around the world on the road, blocking traffic and impeding the progress
(4–6). Recently, more and more researchers have focused of passengers wishing to board and alight (10). Thus,
on studying the role of bus capacity, which is relevant to transit system capacity is controled by stop capacity in
optimizing resources on the basis of ridership. Many public transportation systems (13). The Highway
studies have found that the bus stop, which serves as a
crucial part in the bus system, has a great effect on traffic
1
flow (7, 8). When serving passengers at a busy stop, Jiangsu Key Laboratory of Urban ITS, Southeast University, Jiangsu
buses can interact in ways that limit their discharge flows Province, China
2
Jiangsu Province Collaborative Innovation Center of Modern Urban Traffic
(i.e., bus capacity at stops) (9). Thus, a dwelling bus may Technologies, School of Transportation, Southeast University, Jiangsu
become a bottleneck that constrains traffic flow near the Province, China
3
bus stop, and may degrade the bus system’s overall ser- Lyles School of Civil Engineering, Purdue University, West Lafayette, IN
4
vice quality (10–12). Texas A&M University, Zachry Department of Civil Engineering, College
Station, TX
In general, any bus stop consists of a stop area, where
vehicles stop, and an adjacent platform, where passengers Corresponding Author:
wait for vehicles and boarding/alighting operations take Address correspondence to Satish V. Ukkusuri: sukkusur@purdue.edu
Wang et al 109

Capacity Manual (HCM) reports that the capacity of a method that takes into account the statistical test for bus
bus stop is inversely proportional to the average bus ser- arrival process and service time and an analytical model
vice time and a second term that accounts for the varia- for single-berth and multi-berth stops is developed. In
tion in this service time such as delay time at bus stop this study, the bus stops are isolated from the influences
(14). Delay at stop is the sum of two average delays: one of traffic signals and other bus stops, where sufficient
in the entry queue, and the other in the berth after the space exists for accommodating the bus queues.
bus has finished serving its passengers (15). Gu et al. (16) The remaining sections of the paper are organized as
found that bus delays could be increased by placing the follows. The second and third sections describe the data
bus stop upstream of its neighboring intersection, rather collection and proposed methodology to model bus stop
than downstream of it. In contrast, car delays were capacity. The fourth section validates the proposed
increased by placing the stop some distance downstream method and evaluates the performance by analysis and
of the intersection, rather than upstream. comparison. The findings and conclusions are provided
Two categories of methods for modeling stop capacity in the final section.
have been reported in the literature. The first is the use
of analytical models to estimate capacity with a steady
state formula (17). The best known is the HCM formula Data Collection
(14). The parameters in this model include green time Based on the right-of-way, bus lanes at bus stops can be
ratio, clearance time, average dwell time, standard nor- divided into three categories: grade-separated bus lanes,
mal variable corresponding to a desired failure rate, and at-grade bus lanes, and non-exclusive bus lanes (23). The
coefficient of variation of dwell time. Fernández et al. bus dwell time will be affected by the layout of the bus
(18) introduced a concept called capacity of divided bus stop (24). In general, the more exclusive the design, that
stops. A divided bus stop contains berths that are sepa- is, the less interaction that a transit vehicle has with other
rated to reduce bus interference and increase bus capac- traffic, the fewer impacts there will be on bus dwell time.
ity. Jaiswal et al. (19) introduced bus lost time as an In terms of the form, bus stops can be classified into two
additional component of the minimum headway to cal- categories: on-line and off-line (23). Compared with an
culate bus platform capacity. Accordingly, a busway on-line bus stop, additional time is required for buses at
loading bus capacity model was developed with lost time an off-line stop to find an acceptable gap between con-
variables. The second category is the use of simulation secutive vehicles to rejoin the traffic stream. It can thus
models to calculate capacity as well as other performance be concluded that the form of a bus stop has an impact
measures of stop operations (10). Two specially-designed on the bus dwell time. According to the location of the
simulation models of stop operations can be found in the cross-section, bus stops can be divided into the two cate-
literature: IRENE (12) and PASSION (11, 20, 21). gories of median and curbside. Moreover, based on the
PASSION is developed as part of broader studies on the station design, bus stops can be classified into two cate-
interactions between buses, passengers, and traffic at bus gories: enclosed bus stop (having gate controled access
stops. to bus doors) and non-enclosed bus stop. According to
Although a variety of methods have been developed the above classifications, seven types of bus stop are
to estimate bus stop capacity, the performance of these most commonly observed in China (25), as shown in
models varies. For instance, Fernández and Planzer (11) Figure 1.
reported that the HCM formulas tended to under-predict
the stop capacity because the stochastic processes at bus (a) Type 1: The at-grade bus lanes are separated
stops may limit the usefulness of HCM formulas. In pre- from motor vehicle lanes by traffic markings.
vious studies, the arrival rate of a bus stream and the ser- Bus stops are on-line and on the curbside.
vice time of a bus stop were considered to be constant, (b) Type 2: There is no exclusive bus lane. Bus stops
neglecting the variation of bus arrival headway and the are on-line and on the curbside.
differences in bus stop designs (17, 20, 22). For a multi- (c) Type 3: The at-grade bus lanes are separated
berth stop, the HCM formula calculates the capacity as from motor vehicle lanes by traffic markings.
the product of the single-berth capacity and the number Bus stops are off-line (bay-style) and on the
of effective berths, ignoring the influence of the bus arri- curbside.
val process on stop capacity (9). (d) Type 4: There is no exclusive bus lane. Bus stops
In addition, most of the existing methods have relied are off-line (bay-style) and on the curbside.
(to greater or lesser degrees) on computer simulation. (e) Type 5: The grade-separated bus lanes are sepa-
The dearth of analytical models for evaluating capacity rated from motor vehicle lanes by separation
has thus far hampered the understanding of cause-and- strips. Bus stops are on-line and in the median of
effect relations at bus stops. In light of this, an estimation the cross-section.
110 Transportation Research Record 2672(8)

Figure 1. Schematic diagram of seven most common bus stops.

(f) Type 6: The at-grade bus lanes are separated In this study, data were collected for these seven dif-
from motor vehicle lanes by traffic markings. ferent types of bus stops in the cities of Nanjing,
Bus stops are on-line and in the median of the Changzhou, and Guangzhou, China. The data were col-
cross-section. lected under good weather conditions between May 19,
(g) Type 7: There is no exclusive bus lane. Bus stops 2014 and June 15, 2014 to exclude potential influences of
are on-line and on the curbside. Buses pull over adverse weather. In addition, there was no curb parking
to the curbside and occupy bicycle lanes to dwell. around the stops.
Wang et al 111

Table 1. Site and Traffic Characteristics of the Bus Stops and Service Times

No. Bus stop Type BN TC SS V C AST STV SVC

1 Gulou North Type 1 4 Peak 67 2677 4500 30.13 83.69 0.30

2 Non-peak 47 1723 4500 32.38 79.37 0.27
3 Beiji Huitang Type 2 2 Peak 113 3017 3900 23.90 82.98 0.38
4 Non-peak 108 2155 3900 19.01 44.53 0.35
5 Public Transport Corporation Type 3 5 Peak 96 3378 4500 33.64 97.10 0.29
6 Non-peak 101 2286 4500 30.80 71.32 0.27
7 Xuanwuhu Park Type 4 5 Peak 40 3850 4500 35.83 145.43 0.33
8 Non-peak 56 3054 4500 40.96 156.25 0.30
9 Renmin Park Type 5 3 Peak 34 2398 4000 34.37 71.35 0.25
10 Non-peak 40 1755 4000 32.86 40.27 0.20
11 Gangding Type 6 4 Peak 51 2848 4500 31.06 56.26 0.24
12 Non-peak 52 2205 4500 26.77 61.36 0.29
13 Danfeng Street Type 7 1 Peak 39 2078 3000 26.33 63.49 0.30
14 Non-peak 41 1386 3000 24.95 42.65 0.26

Note: BN = number of bus berths at the bus stop; TC = traffic condition (peak or non-peak period); SS = sample size (buses); V = traffic flow rate
(veh/h); C = capacity (veh/h); AST =average service time of buses (s), denoted by m; STV = variations in bus service time (s2), denoted by s2 ;
SVC = coefficient of variation in bus service time, defined as s divided by m.

Three video cameras were used at each stop to record It has been assumed that bus arrival at stops follows a
traffic data. One camera was set up at a high location Poisson process (26, 27). The interarrival time is an inde-
and two other video cameras were in front of and behind pendent, exponentially distributed random variable with
the bus stop, respectively. The recorded videos were arrival rate of buses l. Some studies also show that bus
reviewed by several trained graduate students to obtain service time at stops follows some general distribution,
traffic volume, bus average service time, variances in bus and the distribution of an individual bus service time is
service time and coefficient of variation in bus service independent of the number of berths (9, 28, 29). To
time (defined as standard deviation in bus service time examine and explore further the distributions of bus arri-
divided by average service time of buses). The site and vals and service time, we collected data on 885 buses
traffic characteristics of the bus stops and service times from the 14 sites, as shown in Table 1.
are shown in Table 1. It should be noted that the service
time at several stops, such as Gulou North and
Xuanwuhu Park, is higher during non-peak than peak
Bus Arrival Process
time. This is because longer departure intervals result in Different bus routes have different departure intervals
more boarding and alighting passengers during non- (i.e., pre-set headways). The distribution of bus arrivals
peak. with multiple routes would also be influenced by buses
on other bus routes. In addition, many stochastic vari-
ables have effects on bus arrivals, for example, traffic
Bus Stop Capacity Model conditions (30). Thus, to validate the Poisson process for
bus arrivals, the Kolmogorov-Smirnov test (K-S test) is
In operational terms, a bus stop can be considered as a introduced as a criterion. The K-S test can be used to
concurrent queuing system between buses and passengers compare a sample with a reference probability distribu-
(10). It is assumed that bus stops operate in a steady tion, or to compare two samples (31). The K-S statistic
state, such that the long-run average bus arrival rate quantifies a distance between the empirical distribution
never exceeds the stop’s capacity (9). Bus stop capacity is function of the sample and the cumulative distribution
closely related to the bus average service time at bus function of the reference distribution, or between the
stops (14). In addition, a bus halts at the stop area for a empirical distribution functions of two samples. The null
certain duration, which depends on the number of wait- distribution of this statistic is calculated under the null
ing passengers. If there is no unoccupied berth for an hypothesis that the samples are drawn from the same dis-
arriving bus, the bus may have to wait on the road, tribution (in the two-sample case) or that the sample is
blocking the traffic and impeding the progress of board- drawn from the reference distribution (in the one-sample
ing and alighting (10). Consequently, the arrival rate of case). In this section, the Poisson process for bus arrivals
buses can affect the bus service time, and thus affect the is a one-sample K-S test. The results of K-S tests show
bus stop capacity. that all of the P values for the 14 sites are much greater
112 Transportation Research Record 2672(8)

Exponential
distribution
than 0.05 (varying between 0.58 and 1.00). Hence, the

0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
tests show that the results are consistent with previous
studies that bus arrivals follow Poisson distribution for
isolated bus stops.

distribution
Weibull

0.051
0.552
0.016
0.119
0.082
0.036
0.405
0.456
0.909
0.786
0.269
0.461
0.258
0.421
Bus Service Time

P value (P . .05)
In this section, three common distributions including
normal, lognormal, and Weibull distribution are selected

distribution
to fit the bus service time. In addition, the exponential

Lognormal

0.362
0.994
0.363
0.461
0.607
0.297
0.955
0.985
0.997
0.960
0.491
0.960
0.439
0.864
distribution is also introduced as a comparison. Similar
to the way of validating bus arrival process, distribution
for bus service time is examined using the K-S test. In
general, the smaller the K-S value and the larger the P

distribution
value, the better the fitting performance. As shown in

Normal

0.108
0.587
0.013
0.064
0.074
0.114
0.536
0.488
0.964
0.879
0.211
0.572
0.266
0.542
Table 2, among the four distributions, the lognormal dis-
tribution (boldface in Table 2) has the best performance
for all bus stops during peak and non-peak periods.
In order to provide visual comparison of the fitting

Exponential
distribution
results for bus service time, the probability density func-

0.457
0.444
0.393
0.384
0.433
0.484
0.438
0.430
0.470
0.487
0.472
0.431
0.470
0.421
tion of normal distribution, lognormal distribution,
Weibull distribution, exponential distribution, and sam-
ple data for Beiji Huitang bus stop (Type 2) and Public
Transport Corporation bus stop (Type 3) in peak and K-S value (the smaller, the better)

distribution
Weibull
non-peak periods are chosen to be presented in Figure 2.

0.163
0.113
0.146
0.114
0.127
0.141
0.137
0.113
0.109
0.099
0.137
0.115
0.158
0.134
The results show that the sample data follow the prob-
ability density function of lognormal distribution better.
The distribution parameters of the lognormal distribu-
tion for each bus stop in peak and non-peak periods are
distribution
Lognormal

0.113
0.061
0.087
0.082
0.078
0.097
0.081
0.061
0.082
0.080
0.117
0.070
0.139
0.094
also displayed in the figure.
Based on the above testing results, an M/G-L/1 queu-
ing model for single-berth stops and an M/G-L/S queu-
ing model for multi-berth stops are developed to estimate
distribution

bus stop capacity, where M represents Poisson arrivals,

Normal

0.148
0.113
0.149
0.126
0.131
0.119
0.127
0.112
0.102
0.093
0.148
0.109
0.161
0.125
G-L represents lognormal bus service time, and 1(S) rep-
resents the number of bus berths. Because of the stochas-
tic nature of bus arrivals at bus stops, the proposed
models are called stochastic queuing models (SQMs).
Sample

113
108

101
size

67
47

96

40
56
34
40
51
52
39
41

Model for Single-Berth Stops

Non-peak

Non-peak

Non-peak

Non-peak

Non-peak

Non-peak

Non-peak

The Highway Capacity Manual (14) reports that the

Table 2. K-S Test for Bus Service Time

State

capacity of a bus stop is inversely proportional to the

Peak

Peak

Peak

Peak

Peak

Peak

Peak

average bus service time and a second term that accounts

for the variation in this service time. In this study, the
expected time spent in the queuing system Ws is used to
Public Transport

Xuanwuhu Park
Corporation

Danfeng Street
Gulou North

represent the sum of average bus service time and varia-

Renmin Park
Beiji Huitang
Bus stop

tion in this service time. Thus, the capacity for single-

Gangding

berth stops can be expressed as follows:

Cs = 3600=Ws ð1Þ
Type 1

Type 2

Type 3

Type 4

Type 5

Type 6

Type 7

where Cs is the capacity for single-berth stops (veh/h); Ws

Type

is the expected time spent in the system (s).

Wang et al 113

Figure 2. Probability density function of four distributions and sample data at selected bus stops: (a) Beiji Huitang (peak); (b) Beiji
Huitang (non-peak); (c) Public Transport Corporation (peak); (d) Public Transport Corporation (non-peak).

The input into the system is formed by the buses Cs = ð3600 3 2 3 ð1  l 3 E(Tn )ÞÞ=
approaching from upstream, and there are Poisson arri-   ð5Þ
2 3 E(Tn )  l 3 E2 (Tn ) + l 3 D(Tn )
vals with a mean headway of 1=l s. The service time at a
bus stop is expressed by generally distributed variables
with mean 1=u s. The service counter is the bus berths. Model for Multi-Berth Stops
According to Little’s Law (32, 33), Ws in the M/G-L/1
system is given by: In real-world bus stops, a bus’s entering and exiting
maneuvers will be affected by other buses dwelling at
Ws = Ls =l ð2Þ neighboring berths. Thus, the concept of the number of
effective berths is introduced for multi-berth stops. The
where Ls is the expected number of buses in the system capacity for multi-berth stops can be expressed by:
(veh); l is the arrival rate of buses (veh/s). The expected
number of buses in the system Ls can be represented as: Cm = ð3600 3 Nel Þ=Ws ð6Þ
  where Cm is the capacity for multi-berth stops (veh/h) and
Ls = r2s + l2 3 D(Tn ) =ð2 3 ð1  rs ÞÞ + rs ð3Þ
Nel represents the number of effective berths. According
rs = l=u = l 3 E(Tn ) ð4Þ to HCM (14), the first berth is 100% efficient in terms of
stop capacity. If a second berth is added, it increases the
where rs is the traffic intensity for single-berth stops; capacity by 75% (Nel =1.75) and a third berth may add
D(Tn) represents the variance of service time for lognor- 70% extra capacity (Nel =2.45). However, the fourth and
mal distribution (s2); u is the service rate (veh/s); E(Tn) fifth berths only increase capacity by 17% (Nel =2.62)
represents the mean service time for lognormal distribu- and 13% (Nel =2.75), respectively.
tion (s). Substituting Equations 2 to 4 into Equation 1, As for the multi-berth stops, similar to single-berth
the capacity for single-berth stops can be expressed by: stops, there are Poisson arrival headways with mean 1=l
114 Transportation Research Record 2672(8)

s and generally distributed service time with mean 1=u s. time (25). Average delay for re-entering the traffic stream
Explicit exact solutions for the M/G-L/S system have is a function of the capacity and the degree of saturation
not been obtained except for some special cases, for in the vicinity of a bus stop (14, 17). The analytical model
example, the M/M/S system. It is, however, possible to used to estimate average delay assumes that the demand
obtain approximate solutions for a relatively wide class is less than capacity for the period of analysis. According
of service time distributions (34, 35). Thus, the expected to the HCM (14), in most cases, the recommended analy-
time Ws in the M/G-L/S system can be approximately sis period is 15 minutes.
0 s
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
2
3600 V V ð 3600=C Þ 3 ð V =C Þ
tad = + 900 3 T 3 @  1 + 1 + A ð14Þ
C C C ð450 3 T Þ

given by: where tad is average delay for re-entering the car stream,
and T represents the analysis time period, T=0.25 for a
Ws = Wq + E(Tn ) ð7Þ 15-minute period.
S 2
rm 3 ðD(Tn ) + E (Tn )Þ In addition to average delay for re-entering the traffic
Wq = stream, there are two additional causes of delay. One is
ðS  1Þ! 3 2 3 E(Tn ) 3 ðS  rm Þ2
!1 boarding lost time, the time spent waiting for passengers
ð8Þ
X
S 1 j
rm S
rm to walk to the bus door(s) from their waiting position at
3 + the stop. When passengers wait at bus stops with multiple
j=1
j! ðS  1Þ! 3 ðS  rm Þ
loading areas, such as high-volume stations served by mul-
rm = l=ðS 3 uÞ = ðl 3 E(Tn )Þ=S ð9Þ tiple routes, they do not know in advance at which loading
area the bus will stop when it arrives. When a bus arrives,
where Wq is the expected time spent in the queue (s); rm there is typically a delay from when the bus doors open to
is the traffic intensity for multi-berth stops; and S denotes when the first passenger arrives to board the bus (25).
the number of bus berths. Substituting Equations 7 to 9 According to the field investigations, the amount of board-
into Equation 6, the capacity for multi-berth stops can be ing lost time was found to vary by different types of bus
expressed as follows: stop design, with median values ranging from 2.9 s to 4.1 s
 2
n)
3600 3 Nel 3 ðS  1Þ! 3 2 3 E(Tn ) 3 S  l 3 E(T
S
Cm =  S  2 ð10Þ
l 3 E(Tn ) 2 (T )Þ 3 P + ðS  1Þ! 3 2 3 E(T ) 3 S  l 3 E(Tn )
S 3 ð D(T n ) + E n n S 3 E(T n )

where

S 1 
X   S   !1
l 3 E(Tn ) j l 3 E(Tn ) l 3 E(Tn )
P= =j! + = ðS  1Þ! 3 S  ð11Þ
j=1
S S S

In addition, the mean service time E(Tn) and the var-

iance of service time D(Tn) for lognormal distribution and interquartile range values ranging from 0.4 s to 0.9 s.
can be expressed as: The second cause of delay is bus stop failure time. If bus
arrives at a stop to find all the loading areas are occupied,
 
s2 it is forced to wait until other buses leave the stop. In addi-
EðTn Þ = exp m + ð12Þ tion, when a bus is ready to depart from a stop, it also has
2
      to wait if it is blocked by other buses at the stop. In this
DðTn Þ = exp 2m + s2 exp s2  1 ð13Þ case, the bus will be delayed (i.e., bus stop failure time)
while waiting for all of the buses at the stop to finish ser-
ving their passengers (25). According to the field investiga-
Bus Delays at Stops
tions, the amount of bus stop failure time was also found
In this study, bus delays at stops consist of average delay to vary by different types of bus stops, with median values
for re-entering the traffic stream and other additional ranging from 4.9 s to 6.4 s and interquartile range values
delays such as boarding lost time and bus stop failure ranging from 0.5 s to 0.9 s.
Wang et al 115

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Model Validation u N
uP 2
u
Comparison of Results ti = 1 ð^yl  yi Þ
RMSE = ð18Þ
In addition to the SQM method, the model based on the N
HCM formula (14), which is one of the most well-known P
N
analytical models, is used for bus stop capacity estima- j^yl  yi j
i=1
tion and comparison. According to the HCM, the capac- MAE = ð19Þ
N
ity is computed as follows: vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u 2  P 2
u P N  N
3600 3 ðg=cÞ uN j^yl yi j
 j^yl yi j
Cs = ð15Þ u
tc + ðg=cÞ 3 td + Z 3 Cv 3 td t i = 1 yi i=1
yi
VAPE = ð20Þ
N (N  1)
3600 3 ðg=cÞ
Cm = Nel 3 ð16Þ
tc + ðg=cÞ 3 td + Z 3 Cv 3 td where yi and ^yi are actual and estimated capacity,
respectively.
where g=c is the effective green time of a downstream To evaluate the performance of the two methods, the
traffic signal (as for the isolated bus stops in this study, comparison data of MAPE at seven bus stops in peak
g=c = 1); tc is the clearance time between successive buses and non-peak periods are presented in Table 3.
(s); and td is the average dwell time (s). In this study, the According to the results, the SQM method has more
bus service time includes tc and td; Cv is the coefficient of accurate and reliable estimation (with 8.09% of MAPE
variation in bus service time; and Z is the standard nor- on average) than the HCM method (with 16.53% of
mal variable corresponding to a desired failure rate. The MAPE on average). In addition, the MAPE values of
above values are provided in Table 1. the SQM method at all bus stops are below 15%, with
The SQM method and HCM method are compared the lowest value of 0.53% at Gangding bus stop in non-
using field data at the seven bus stops. The bus flow that peak period. The performance of the HCM method,
can be served by the stop is equivalent to the flow that however, is inferior to the SQM method, with almost
can arrive at and depart from it. The literature commonly half of the MAPE values over 20%.
refers to this maximum bus flow as the bus stop capacity Table 4 summarizes all of the values of the MOEs for
(15). According to relevant references (10, 15, 36), actual the two methods. Overall, the results clearly show that
capacity (i.e., the number of buses that could be served the SQM method has more accurate and reliable estima-
by a stop) is defined as the maximum hourly equivalent tion. Compared with the values of the HCM model, the
of a period under saturated conditions, such as 5-minute, SQM method reduced MAPE by 51.1%, RMSE by
10-minute, or 15-minute. Instruments and equipment 58.8%, MAE by 57.9%, and VAPE by 52.6%.
used for the data collection included digital video cam- In addition, the results of bus delays for different types
eras, tripods, stopwatch, and computer. Between May of bus stops are shown in Figure 3. From Figure 3a and
and June 2014, around 4 hours of data collection was b, cumulative curves for bus delays at Type 1, Type 3,
conducted to observe the number of buses at each bus Type 5, and Type 6 bus stops were invariably to the left
stop for both peak and non-peak periods. Taking 5 min- of the curves for other types of bus stops during peak
utes, 10 minutes, and 15 minutes as analysis intervals and non-peak periods. This is because the buses in grade-
respectively, the maximum number of buses for these separated busways (Type 5) and at-grade busways (Types
periods from all the three types of intervals was selected. 1, 3, and 6) cannot be disturbed by other non-bus vehi-
The number of buses was then derived from an hourly cles. Overall, bus delay times at stops during peak period
volume, which is the actual capacity of a bus stop. were longer than the values during non-peak period.
Several measures of effectiveness (MOEs), including
mean absolute percentage error (MAPE), root mean
squared error (RMSE), mean absolute error (MAE), and Sensitivity Analyses
variance of absolute percentage error (VAPE), were cho- In order to analyze fully the proposed SQM method, we
sen to evaluate the performance of the SQM and HCM use sensitivity analyses to investigate the effects of bus
methods (37). These measures are calculated by: arrival rate, service time, and the number of bus berths
on the capacity of bus stops. Differences in the arrival
N
1X ^yl  yi rate of the bus stream affect bus capacity at single-berth
MAPE = ð17Þ
n i = 1 yi stops, as shown in Figure 4. At the same bus service
time, the probability of no stopped bus at the stop
116 Transportation Research Record 2672(8)

Table 3. Comparison of MAPE at Bus Stops in Peak and Non-Peak Periods

SQM method HCM method

No. Bus stop/ number of berths State Field data (veh/h) Capacity (veh/h) MAPE (%) Capacity (veh/h) MAPE (%)

1 Gulou North/ 4 Peak 316 314 0.63 242 23.42

2 Non-peak 270 290 7.41 225 16.67
3 Beiji Huitang/ 2 Peak 282 245 13.12 201 28.72
4 Non-peak 274 315 14.96 247 9.85
5 Public Transport Corporation/ 5 Peak 318 296 6.92 240 24.53
6 Non-peak 362 319 11.88 259 28.45
7 Xuanwuhu Park/ 5 Peak 264 275 4.17 218 17.42
8 Non-peak 223 241 8.07 187 16.14
9 Renmin Park/ 3 Peak 231 256 10.82 220 4.76
10 Non-peak 234 267 14.10 238 1.71
11 Gangding/ 4 Peak 335 356 6.27 254 24.18
12 Non-peak 379 377 0.53 287 24.27
13 Danfeng Street/ 1 Peak 103 111 7.77 107 3.88
14 Non-peak 122 114 6.56 113 7.38
Average MAPE 8.09 MAPE 16.53

veh/h to 170 veh/h at the arrival rate 0.06 veh/s, as bus

Table 4. Comparison of MOEs
service time increases from 20 s to 50 s. In addition, as
Measures of effectiveness shown in Figures 5a and b, at a same bus service time,
the increase in arrival rates of the bus stream significantly
Method MAPE (%) RMSE MAE VAPE reduces the bus stop capacity at two-berth and three-
SQM method 8.086 24.668 20.786 0.045 berth stops. However, the trend does not occur at the
HCM model 16.527 59.918 49.357 0.095 stops whose number of bus berths is more than three.
This is reasonable because the probability of unoccupied
berth(s) at the stop increases with the increase of the
number of bus berths. Hence, as for the four-berth
decreases with the increasing arrival rates of the bus (Figure 5c) and five-berth (Figure 5d) stops, the bus stop
stream, and further reduces the bus stop capacity. In capacity is not sensitive to arrival rates. For instance,
addition, the service time of buses at the stop affects bus when the number of bus berths is equal to four, the maxi-
stop capacity. At the same arrival rate, the increasing mum reduction of bus stop capacity is merely 7 veh/h
bus service time increases the probability of stopped bus when the arrival rate of buses changes from 0.01 veh/s to
at the bus stop. The stopped bus blocks bus streams near 0.15 veh/s.
the stop and impedes the progress of boarding and To understand the degree to which bus berths and bus
alighting, and further reduces bus stop capacity from 217 service time affect capacity at bus stops, the SQM
veh/h to 67 veh/h at the arrival rate 0.01 veh/s, from 204 method is further used to explore the change of bus stop
veh/h to 51 veh/h at the arrival rate 0.015 veh/s, and capacity when arrival rate of buses is 0.01 veh/s (36 veh/
from 190 veh/h to 29 veh/h at the arrival rate 0.02 veh/s, h) and bus service time is between 15 s and 45s. The
respectively. In this study, when the arrival rate is 0.02 curves in Figure 6a display the differences in how the
veh/s and bus service time is more than 35 s, the bus stop number of bus berths affects bus stop capacity. When the
capacity is reduced to fewer than 50 veh/h. This is in bus service time is short, additional berths could produce
accordance with the actual bus operations at bus stops, increasing returns in capacity. For example, Figure 6a
which becomes unsteady when bus arrival rate is near to shows that when bus service time is equal to 20 s, two-
or exceeds the service rate, that is, when the traffic inten- berth stops increase bus stop capacity by 141 veh/h over
sity is close to or exceeds 1.0. one-berth stops. This favorable trend does not continue,
The results of capacity estimation for multi-berth however, as bus service time increases. In addition, at the
stops are shown in Figure 5. Similar to single-berth stops, same bus service time, the incremental changes in bus
at a same arrival rate, the bus stop capacity reduces as stop capacity diminish dramatically as the number of bus
the bus service time increases. For example, as for the berths increases (Figure 6b). When bus service time is 25
three-berth stops, bus stop capacity is reduced from 439 s, the increasing returns in capacity are reduced from 134
Wang et al 117

Figure 3. Bus delay time at different types of bus stops for (a) peak period and (b) non-peak period.

veh/h to 19 veh/h for the first berth through the fifth Conclusions
berth, especially for the fourth and fifth berths, with
This study investigated bus arrival process and bus ser-
increases of 25 veh/h and 19 veh/h, respectively. This
vice time distribution, and applied a stochastic queuing
finding is in accordance with the actual bus operations at method to estimate the capacity for isolated bus stops.
bus stops; current practice states that it is not efficient to The method not only takes into account the statistical
build bus stops with more than three adjacent berths (9). test for bus arrival process and bus service time, but also
If more than three berths are necessary, the stop should considers the analytical model for single-berth and multi-
be split in two multi-berth stops (e.g., a four-berth stop berth stops.
can be split into two two-berth stops). The results further In this study, data collected from seven bus stop con-
indicate that the method is well validated and could be figurations in China were used to validate the SQM. A
used in practice for the analysis and estimation of bus K-S test was performed to explore the bus arrival process
stop capacity in China. and bus service time distribution. According to the P
118 Transportation Research Record 2672(8)

8.44% compared with an average MAPE of

17.42% from the HCM model. Compared with
the values of the HCM model, the SQM method
reduced MAPE by 51.1%, RMSE by 58.8%,
MAE by 57.9%, and VAPE by 52.6%.
(2) Bus delay times at Type 1, Type 3, Type 5, and
Type 6 bus stops were shorter than for other
types of bus stops during peak and non-peak
periods. This was because the buses in grade-
separated and at-grade busways cannot be dis-
turbed by other non-bus vehicles. Moreover,
Figure 4. Bus stop capacity with different arrival rates of bus bus delay times at stops during peak period
stream and bus service time at single-berth stops. were longer than the values during non-peak
period.
(3) Sensitivity analyses were conducted to investi-
values and K-S values, bus service times followed the log- gate the effects of bus arrival rate, number of
normal distribution, and bus arrival process obeyed the bus berths, and bus service time on the capacity
Poisson distribution. An M/G-L/1 queuing model for of bus stops. For single-berth stops, at the same
single-berth stops and an M/G-L/S model for multi-berth bus service time, the probability of no stopped
stops were developed to estimate bus stop capacity and bus at the stop decreases with the increasing arri-
bus delay time. To evaluate the SQM’s performance, a val rates of the buses and further reduces the bus
model based on the HCM formula was selected for com- stop capacity. At the same arrival rate, increas-
parison. According to the analysis results, the five main ing bus service time increases the probability of
findings were as follows: stopped bus at the bus stop, and further reduces
bus stop capacity from 217 veh/h to 67 veh/h at
(1) The SQM method had more accurate and reli- the arrival rate 0.01 veh/s, from 204 veh/h to 51
able estimation, with an average MAPE of veh/h at the arrival rate 0.015 veh/s, and from

Figure 5. Bus stop capacity with different arrival rates of bus stream and bus service times at multi-berth stops: (a) two berths; (b)
three berths; (c) four berths; (d) five berths.
Wang et al 119

weather conditions. To adapt the method, the research

team will need to collect traffic data in other weather
conditions to obtain the corresponding characteristics,
such as bus service time, and conduct further research to
investigate their impacts.

Acknowledgments
This study was sponsored by the Scientific Research
Foundation of the Graduate School of Southeast University
(No. YBJJ1633), the Fundamental Research Funds for the
Central Universities, the Scientific Research Program of the
Graduate School of Jiangsu Province (No. KYLX16-0280),
and the China Scholarship Council.

Author Contributions
The authors confirm contribution to the paper as follows: study
conception and design: Chao Wang, Zhirui Ye, Satish V.
Ukkusuri; data collection: Chao Wang, Zhirui Ye; analysis and
interpretation of results: Chao Wang, Jon D. Fricker, Satish V.
Ukkusuri; draft manuscript preparation: Chao Wang, Yunlong
Zhang, Jon D. Fricker, Satish V. Ukkusuri, Zhirui Ye. All
authors reviewed the results and approved the final version of
the manuscript.
Figure 6. Bus stop capacity and incremental changes in capacity
versus bus service time for different berths: (a) bus stop capacity
References
versus bus service time; (b) incremental changes in capacity versus
bus service time. 1. Yu, B., Z. Yang, P. Jin, S. Wu, and B. Yao. Transit Route
Network Design-Maximizing Direct and Transfer Demand
Density. Transportation Research Part C, Vol. 22, 2012,
190 veh/h to 29 veh/h at the arrival rate 0.02 pp. 58–75.
2. Bagloee, S. A., and A. Ceder. Transit-Network Design
veh/s, respectively.
Methodology for Actual-Size Road Networks. Transporta-
(4) Similar to single-berth stops, at a same arrival
tion Research Part B, Vol. 45, 2011, pp. 1787–1804.
rate, the bus stop capacity for multi-berth stops 3. Das, S., and D. Pandit. Determination of Level-of-Service
reduces as the bus service time increases. At the Scale Values for Quantitative Bus Transit Service Attri-
same bus service time, the increase in arrival butes Based on User Perception. Transportmetrica A:
rates of the bus stream significantly reduced the Transport Science, Vol. 11, No. 1, 2015, pp. 1–21.
bus stop capacity at two-berth and three-berth 4. Gonzalez, E. M., M. G. Romana, and O. M. Alvaro. Bus
stops. However, this trend does not occur at Dwell-Time Models of Main Urban Route Stops. Trans-
stops where the number of bus berths is more portation Research Record: Journal of the Transportation
than three. This showed that the probability of Research Board, 2012. 2274: 126–134.
unoccupied berth(s) at the stop increased with 5. Bordagaray, M., L. dell’Olio, A. Ibeas, and P. Cecin.
Modeling User Perception of Bus Transit Quality Consid-
the increase of the number of bus berths.
ering User and Service Heterogeneity. Transportmetrica A:
(5) In addition, the results show that it is not effi-
Transport Science, Vol. 10, No. 8, 2014, pp. 705–721.
cient to build bus stops with more than three 6. Chang, H., D. Park, S. Lee, H. Lee, and S. Back. Dynamic
adjacent berths. If more than three berths are Multi-Interval Bus Travel Time Prediction Using Bus
necessary, the stop should be split in two multi- Transit Data. Transportmetrica, Vol. 6, No. 1, 2010, pp.
berth stops (e.g., a four-berth stop can be split 19–38.
into two two-berth stops). 7. Ding, H., W. Wang, T. Luo, Z. Yang, Y. Li, and Z. Li.
Cellular Automata Based Modeling for Evaluating Differ-
This study explores the capacity of isolated bus stops ent Bus Stop Designs In China. Discrete Dynamics in
Nature and Society, Vol. 2015, 2015, pp. 1–10.
in general weather conditions. In different weather con-
8. Tirachini, A. Bus Dwell Time: The Effect of Different Fare
ditions, such as inclement weather, the bus average ser- Collection Systems, Bus Floor Level and Age of Passen-
vice time and delay time at bus stop may be different. gers. Transportmetrica A: Transport Science, Vol. 9, No. 1,
Thus, the proposed method can be applied in different 2013, pp. 28–49.
120 Transportation Research Record 2672(8)

9. Gu, W., Y. Li, M. J. Cassidy, and J. B. Griswold. On the 24. TCRP Report 19. Guidelines for the Location and
Capacity of Isolated, Curbside Bus Stops. Transportation Design of Bus Stops. http://www.gulliver.trb.org/pub-
Research Part B, Vol. 45, No. 4, 2011, pp. 714–723. lications/tcrp/tcrp_rpt_19-a.pdf. Accessed January
10. Fernández, R. Modeling Public Transport Stops by Micro- 16, 1996.
scopic Simulation. Transportation Research Part C, Vol. 25. Wang, C., Z. Ye, Y. Wang, Y. Xu, and W. Wang. Model-
18, No. 6, 2010, pp. 856–868. ing Bus Dwell Time and Time Lost Serving Stop in China.
11. Fernández, R., and R. Planzer. On the Capacity of Bus Journal of Public Transportation, Vol. 19, No. 3, 2016, pp.
Transit Systems. Transport Reviews, Vol. 22, No. 3, 2002, 55–77.
pp. 267–293. 26. Danas, A. Arrivals of Passengers and Buses at Two Lon-
12. Gibson, J., I. Baeza, and L. G. Willumsen. Bus Stops, Con- don Bus Stops. Traffic Engineering and Control, Vol. 21,
gestion, and Congested Bus Stops. Traffic Engineering and No. 10, 1980, pp. 472–475.
Control, Vol. 30, No. 6, 1989, pp. 291–302. 27. Kohler, U. Capacity of Transit Lanes. Proc., International
13. Vuchic, V. R. Urban Transit Operations Planning and Eco- Symposium on Highway Capacity, Karlsruhe, Germany,
nomics. Wiley, Hoboken, NJ, 2005. 1991.
14. Highway Capacity Manual. Transportation Research 28. Gross, D., J. F. Shortle, J. M. Thompson, and C. M. Har-
Board, Washington, D.C., 2000. ris. Fundamentals of Queueing Theory, 4th ed. John Wiley
15. Gu, W., M. J. Cassidy, and Y. Li. Models of Bus Queueing & Sons, Inc., Hoboken, N.J., 2008.
at Curbside Stops. Transportation Science, Vol. 49, No. 2, 29. Gu, W., and M. J. Cassidy. Maximizing Bus Discharge
2015, pp. 204–212. Flows from Multi-Berth Stops by Regulating Exit Maneu-
16. Gu, W., V. V. Gayah, M. J. Cassidy, and N. Saade. On the vers. Transportation Research Part B, Vol. 56, 2013, pp.
Impacts of Bus Stops near Signalized Intersections: Models 254–264.
of Car and Bus Delays. Transportation Research Part B, 30. Yu, B., W. H. K. Lam, and M. L. Tam. Bus Arrival Time
Vol. 68, 2014, pp. 123–140. Prediction at Bus Stop with Multiple Routes. Transporta-
17. Yang, X., Z. Gao, X. Zhao, and B. Si. Road Capacity at tion Research Part C, Vol. 19, 2011, pp. 1157–1170.
Bus Stops with Mixed Traffic Flow in China. Transporta- 31. Corder, G. W., and D. I. Foreman. Nonparametric Statis-
tion Research Record: Journal of the Transportation tics: A Step-By-Step Approach. Wiley, Hoboken, N.J.,
Research Board, 2009. 2111: 18–23. 2014.
18. Fernández, R., V. Burgos, and C. E. Cortés. Results of the 32. Little, J. D. C. A Proof for the Queuing Formula: L=lW.
Microscopic Modeling of Traffic Interactions at Stops, Operations Research, Vol. 9, No. 3, 1961, pp. 383–387.
Junctions and Roads for Designing of Bus Rapid Transit 33. Sztrik, J. Basic Queueing Theory. University of Debrecen:
Facilities. In: European Transport Conference 2007, Associ- Faculty of Informatics, 2011.
ation for European Transport and Contributors, Noord- 34. Kimura, T. Diffusion Approximation for an M/G/m
wijkerhout, Netherlands, 2007, pp. 17–19. Queue. Operations Research, Vol. 31, No. 2, 1983, pp.
19. Jaiswal, S., J. Bunker, and L. Ferreira. Modeling the Rela- 304–321.
tionships between Passenger Demand and Bus Delays at 35. Zhang, W., L. Chen, Y. Huang, L. Niu, M. Huang, D.
Busway Stations. Presented at 88th Annual Meeting of the Zhang, and W. Shi. Application of M/G/K Queuing
Transportation Research Board, Washington, D.C., 2009. Model in Queuing System of Electric Taxi Charging Sta-
20. Fernández, R. A New Approach to Bus Stop Modeling. tion. Power System Technology, Vol. 39, No. 3, 2015, pp.
Traffic Engineering and Control, Vol. 42, No. 7, 2001, pp. 724–729.
240–246. 36. Sun, F., D. Wang, D. Ma, and X. Ma. Method for Calcu-
21. Fernández, R., and N. Tyler. Effect of Passenger–Bus-Traf- lating the Capacity of Bus Bay. Journal of Southeast Uni-
fic Interactions on Bus Stop Operation. Transportation versity (English Edition), Vol. 28, No. 4, 2012, pp. 485–489.
Planning and Technology, Vol. 28, No. 4, 2005, pp. 273–292. 37. Wang, C., and Z. Ye. Traffic Flow Forecasting Based on a
22. Koshy, R. Z., and V. A. Arasan. Influence of Bus Stops on Hybrid Model. Journal of Intelligent Transportation Sys-
Flow Characteristics of Mixed Traffic. Journal of Transpor-
tems, Vol. 20, No. 5, 2016, pp. 428–437.
tation Engineering, Vol. 131, No. 8, 2005, pp. 640–643.
23. KFH Group. Transit Capacity and Quality of Service Man-
The Standing Committee on Bus Transit Systems (AP050) peer-
ual, 3rd ed. Transportation Research Board, Washington,
reviewed this paper (18-03627).
D.C., 2013.