IET Intelligent Transport Systems

Research Article

Bus frequency optimisation considering user ISSN 1751-956X

Received on 8th February 2018
Revised 15th August 2018
behaviour based on mobile bus applications Accepted on 29th October 2018
E-First on 20th November 2018
doi: 10.1049/iet-its.2018.5012
[Link]

Nannan Lin1,2, Weimin Ma2 , Xiaoxuan Chen3

1School of Urban Rail Transportation, Shanghai University of Engineering Science, Longteng Road 333, Shanghai, People's Republic of China
2School of Economics and Management, Tongji University, Siping Road 1239, Shanghai, People's Republic of China
3Transmart Technology Inc., 411 S Wells St, Chicago, USA

E-mail: yiyi880922@[Link]

Abstract: In this study, a model for bus frequency optimisation was presented, which determined the optimal bus frequencies
aiming to minimise the users’ total travel cost. Unlike past researches, where the impacts of real-time information were ignored,
the user behaviours affected by mobile applications were considered in this study. The effects of bus real-time information on
user behaviours had been proved by some other researches. By assigning a user's trip to three stages (pre-trip, on-board and
end-trip), four different user behaviours were taken into account in the model building process: shortest total travel time, shortest
pre-trip time, shortest walking time and shortest riding time. In addition to the object bus line and stops on it, other bus lines and
stops serving the same traffic zones were also considered. The demand was given in the form of an origin–destination (OD)
matrix. Each value in the OD matrix represented the passenger amount from one zone to another. Finally, the illustrated model
was tested by four cases and the results showed a good performance.

Nomenclature 1 Introduction
r* objective round-trip bus line One of the major factors in determining transit service is the
f r* frequency of line r* selection of the most suitable frequency (runs/h) for each route in
n one-way stops amount of line r*, including the first the system. Frequency optimisation is a necessary product for the
and the last stop public transit system, and it is also true that a transit agency will
K set of OD pairs with a generic element k often evaluate and optimise frequencies on routes. Bus frequency
Qk amount of OD pairs in set K design plays a vital role in intelligent public transit system, which
Ok set of origin nodes of OD pair is also a major component of setting timetables.
k(k ∈ K)Ok = {o1, o2, …, oδk} Real-time information systems, especially mobile bus
Dk set of destination nodes of OD pair applications are very common in modern public transportation. For
k(k ∈ K), Dk = {d1, d2, …, dδk} example, Google Transit, which is started as a Google Labs project
δk trips amount of OD pair k in December of 2005, is now directly integrated into the Google
Maps product and provides interfaces to Google Transit which
Ak set of boarding stops in Ok, Ak = {ak1, ak2}
exist on a variety of mobile devices, making use of location sensors
Bk set of alighting stops in Dk, Bk = {bk1, bk2} such as GPS and Wi-Fi localisation on the device to improve the
Tk set of travel arcs from Ok to Dk, with a generic usability of the transit application. These mobile bus applications
element α provide real-time information, including bus schedules, routes and
Rk set of bus lines from Ak to Bk, Rk = {rk*, rk12, rk21, rk22} real-time arrival information, to enable users to make smarter
rk* k segment of line r* travel choices. Lots of scholars, such as Katrin Dziekan and Karl
Kottenhoff [1], had addressed that dynamic real-time information
Gk graph representing the transit system of OD pair k, had adjusted travel behaviours such as better use of wait time or
Gk = (Ok, Ak, T k, Rk, Bk, Dk) more efficient travelling. Consequently, the competitive forces of
ckα cost of arc α the public transit have been enhanced greatly. Meanwhile, real-
Cr* total travel cost of line r* time information systems have put forward to some new challenges
Ck travel cost of line rk* and requirements for public transit agencies in information age.
S set of users’ selection strategies, S = {S1, S2, S3, S4} In Fig. 1, some interfaces of one mobile bus applications are
PS set of probabilities of user selecting strategy S, shown. Specifically, when planning a trip, a user can first search
PS = {PS1, PS2, PS3, PS4} the boarding stops, alighting stops and the destination on the map,
and choose his/her starting time secondly. Then all the available
PY probability of users using mobile bus applications
routes are displayed on the next interface. The optimal trip plan is
PN probability of users not using mobile bus applications made personally.
Vk set of trips through rk*, V k ∈ T k Generally, setting frequencies concerns passengers’ route
vk trips amount through rk*, vk ≤ T k choices serving their origins and destinations. In the age of
β1, β2, β3, β4 variables that indicated whether a user j( j ∈ V k) took intelligent information, user choice behaviours in frequency
rk* optimising process should be considered. This is a worthy topic for
research.
N maximum user capacity of one bus Therefore, this paper addressed the topic of optimising
lbr* lower bound of the objective bus line's frequency frequencies for a single bus line, a problem that must receive
ubr* upper bound of the objective bus line's frequency attention, either explicitly or implicitly. The research provided
insights and a solution for intelligently integrating cost saving and

IET Intell. Transp. Syst., 2019, Vol. 13 Iss. 4, pp. 596-604 596
© The Institution of Engineering and Technology 2018
 17519578, 2019, 4, Downloaded from [Link] by Algeria Hinari NPL, Wiley Online Library on [03/07/2024]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
the bus frequency. Yan and Tang [10] provided an integrated
framework that could systematically analyse the planning and real-
time stage together and help carriers plan bus routes and schedules
for stochastic bus travel time situations. However, the user demand
was assumed to be fixed.
More recently, Yu et al. [11] presented a bi-level programming
model for bus frequency design, aiming to minimise the total travel
time of users subject to the constraint on the overall fleet size of
each company by accounting for the route choice behaviours of the
users. dell'Olio and Ibeas [12] proposed a bi-level optimisation
model with constraints on bus capacity to determine the optimal
size and operating frequencies. It took into account the variations
in cost due to variations in demand caused by changes in the
frequencies and bus sizes. The latest work about frequency
optimisation was done by Martínez et al. [13]. They studied the
Fig. 1 Interfaces of a mobile bus application transit frequency optimisation problem to determine the time
interval between subsequent buses for a set of bus lines. A mixed
integer linear programming formulation for Constantin and
Florian's [6] model was proposed. A new metaheuristic approach
was used to solve the problem.
In addition, Grotenhuis et al. [14] showed that the public
transport travellers had many desires for mobile real-time
information, which were expected to affect travellers’ choices.
Fig. 2 Round-trip bus line Katrin Dziekan and Karl Kottenhoff [1] presented two studies to
support and prove dynamic real-time information could reduce
an effective level of service. The basic premise here was that such users’ perceive wait time and affect their travel mode choices.
integration was achievable. The purpose of this paper, by Watkins et al. [15] and Ferris et al. [16] also obtained similar
considering of users’ behaviour affected by mobile bus conclusions to support that mobile real-time information could
applications, was mainly aiming to obtain the optimal frequency of affect users’ behaviour, with the help of a mobile application
the objective bus line. The goal of modelling was to minimise the named ‘OneBusAway’.
users’ total travel cost of the objective bus line. Overall, although different bus frequency optimisation models
The paper was structured as follows. In Section 2, existing had been proposed in the literature, they had very similar
researches on bus frequency optimisation models were optimisation objectives, i.e. company's operation cost and/or users’
summarised. In Section 3, the problem was described and a travel cost, and considered similar factors, i.e. user demand,
mathematical model was established. A metaheuristic approach crowding, fleet size, travel time uncertainty and so on. However,
was introduced in Section 4 and the method was implemented in this is already an era of information. Inevitably, users’ travel
Section 5. Finally, several remarks were concluded in Section 6. behaviours are affected by intelligent information, especially the
fast development of mobile bus applications. To the best of our
2 Literature review knowledge, no existing study had optimised bus frequency
considering users’ behaviours affected by mobile bus applications.
In this section, a review of representative studies in the field of
frequency optimisation for public transport system was provided.
Some meaningful researches about the impact of mobile real-time 3 Mathematical model
information on users were also summarised. The model proposed in this paper mainly aimed to obtain the
Earlier researches on frequency design were devoted to satisfy optimal frequency of the objective bus line, considering user
the maximum user demands without exceeding the capacity or behaviour affected by mobile bus applications. The goal of
some threshold value. Schéele [2] determined bus line frequencies modelling was to minimise the total travel cost (from the user
considering the distribution of trips between different zones. The point) of the objective bus line.
behaviour of the users was embedded into the model: given an OD In order to present the model, first the problem was described,
pair, the demand was divided among the different lines according including the lines, the demand and the user behaviour. Then the
to entropy and a bus capacity constraint. Furth and Wilson [3] model for frequency optimisation was presented. To ease the
formulated the problem for determining route frequencies that reading of the formulations, a glossary of the main symbols used is
aimed to maximise the network social benefit subject to constraints provided in Nomenclature.
on total subsidy, fleet size and levels of vehicle loading. Han and
Wilson [4] proposed a model to set frequencies on heavily utilised 3.1 Problem description
lines. The objective function stated the minimisation of the
occupancy level at the most heavily loaded point on any route in Consider a public transport system with an integrated round-trip
the system. Ceder [5] described several appropriate data collection bus line r* (objective line) and other lines which partially shared
approaches. Four different methods were presented to derive the users with the objective line. The frequencies fr* of the objective
bus frequency: two were based on point check (maximum load) line would be determined by the optimisation model. The travel
data and two proposed the use of ride check (load profile) data. information of the lines, such as current location and travel time
The frequency optimisation problem was presented as a non- from one stop to another, could be obtained from mobile bus
linear bi-level problem by Constantin and Florian [6]. In this applications. On the demand side, the total amount of users that
model, the objective functions of both levels were the same, which need to perform trips between different points of the system was
was to minimise the total travel time. Tom and Mohan [7] known. The users could search information about those lines,
presented a model to optimise transit routes and frequencies based including walking time to stops, from mobile bus applications.
on genetic algorithm, aiming to minimise both operating costs and Also, the users should decide how to perform such trips according
users’ total travel time. Gao et al. [8] formulated a multi-objective to the information. No transfer was considered.
bi-level programming model for transit network design problem.
The problem modelled the supply side in form of the optimal 3.1.1 Lines and demand: Fig. 2 was a round-trip bus line r* with
frequency setting. The salient characteristic of this work was the n stops.
internalisation of the congestion in the behaviour of the users. Chen It was assumed that the demand between different zones of the
[9] computer multiple headways for a single bus line to maximise city was given in the form of an origin–destination (OD) matrix;
the expected daily profit, which was discontinuous at changes in
IET Intell. Transp. Syst., 2019, Vol. 13 Iss. 4, pp. 596-604 597
© The Institution of Engineering and Technology 2018
 17519578, 2019, 4, Downloaded from [Link] by Algeria Hinari NPL, Wiley Online Library on [03/07/2024]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Table 1 Value matrix of OD pairs K
O D
1 2 … n−1 n Σ
1 δ12 … δ1(n − 1) δ1n δ1
2 δ21 … δ2(n − 1) δ2n δ2
… … … … … …
n−1 δ(n − 1)1 δ(n − 1)2 … δ(n − 1)n δ(n − 1) ⋅
n δn1 δn2 … δn(n − 1)1 δn ⋅
Σ δ⋅ 1 δ⋅ 2 … δ ⋅ (n − 1) δ⋅ n

Fig. 3 Graph model

3.1.2 User behaviour: When planning and undertaking trips,

users had different purposes to fulfil. These purposes were
presumed to affect the tasks and decisions, and hence the trip of a
user could be assigned to three different stages, roughly in
conformity with three location types: pre-trip, on-board and end-
trip.

(a) Pre-trip: The pre-trip stage was essentially the trip planning
step. Pre-trip planning took place at the origin of a trip, e.g. at
home or at office, including the information searching time,
walking time from origin to boarding stop and waiting time at the
boarding stop. With mobile bus applications, users could know the
walking time from origins to the boarding stops, travel time from
boarding stops to alighting stops of all the selectable bus lines, and
also the waiting time for the nearest bus of each line. In contrast to
a car journey, a travel by bus needed virtually all the relevant
information beforehand. In pre-trip stage, users chose the optimum
bus lines according to the information.
(b) On-board: On-board stage was mainly the riding time in a
vehicle, which was known from mobile bus applications in the pre-
Fig. 4 Time-space trajectory of a trip trip stage. There was no transfer considered in this paper.
(c) End-trip: End-trip stage was mainly the walking time from
each element different from zero in this matrix was called an OD alighting stops to destinations. This information could also be
pair. In this paper, represent the demand as a set of OD pairs K (see known from mobile bus applications in the pre-trip stage.
Table 1). Qk was the amount of OD pairs in set K. Ok(Dk) were sets
of origin (destination) nodes, respectively, of OD pair k(k ∈ K) and A time-space trajectory was drawn to analyse a complete trip,
the non-negative value δk represented the value of OD pair k, i.e. its as shown in Fig. 4. There were four types of arcs: searching arc,
corresponding amount of trips. Ok = {o1, o2, …, oδk}, walking arc, waiting arc and riding arc. A pre-trip stage included
Dk = {d1, d2, …, dδk}. Table 1 was the value matrix of OD pairs K searching arc, walking arc from origin to boarding stop and waiting
and Qk = n(n − 1). arc at boarding stop. Assume all the users had planned out their
trips and arrived 1 min earlier at boarding stops than the nearest
Consider an OD pair k separately. δk was the total trip amount
buses, i.e. waiting time was equal to 1 min. Tk was the set of all
of OD pair k. Represent the trips in OD pair k as a directed graph
arcs of OD pair k, with a generic element α. ckα was the cost of arc
Gk = (Ok, Ak, T k, Rk, Bk, Dk), where Ok(Dk) were sets of origin
(destination) nodes, respectively, Ak(Bk) were sets of boarding α. In Fig. 4, Cksearching, Ckwalking, Ckwaiting and Ckriding, respectively, were
(alighting) stops in Ok(Dk), respectively, and T k was set of travel the cost of searching arc, walking arc, waiting arc and riding arc,
arcs from Ok to Dk. Define Rk as set of bus lines from Ak to Bk. i.e. searching time, walking time, waiting time and riding time.
Assume all the users had planned out their trips and arrived 1 min
Suppose there were at most two boarding stops ak1 and ak2 in Ok earlier at boarding stops than the nearest buses, i.e.
and at most two alighting stops bk1 and bk2 in Dk, four bus lines ckwaiting = 1(min). Define ckpre − trip as the cost of pre-trip stage. f r*
(rk*, rk12, rk21, rk22) from Ak to Bk, rk* was the k segment of objective was frequency, then 1/ f r* was the headway of line r*. Assume
bus line r*, then Ak = {ak1, ak2}, Bk = {bk1, bk2} and passengers arrived randomly, then passengers’ waiting time
Rk{rk*, rk12, rk21, rk22}. Tk included the walking arcs from followed uniform distribution in (0, 1/ f r*). Thus, there was
ckpre − trip = 1/(2 f r*).
Ok{rk*, rk12, rk21, rk22} Ok (Bk) to Ak (Dk) and riding arcs from Ak to
Bk. Fig. 3 illustrated the structure of the graph Gk.

598 IET Intell. Transp. Syst., 2019, Vol. 13 Iss. 4, pp. 596-604
© The Institution of Engineering and Technology 2018
 17519578, 2019, 4, Downloaded from [Link] by Algeria Hinari NPL, Wiley Online Library on [03/07/2024]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 5 Logical relationship among the inside factors

The purpose of this paper was to optimise the frequency f r* (see (3))
considering different users’ behaviour, in order to minimise the (see (4))
total travel cost Cr* of the objective bus line r*. In actual life, it Then, the total trips vk through line rk* could be formulated
was impossible that all the users following mobile bus applications.
Define PY and PN as the probabilities of using and not using mobile 4
bus applications separately. Note that the decisions of the users vk = ∑ δkPY PSl βl + δkPN β 2 (5)
depend on the line frequencies. For example, in Fig. 3, passengers l=1
had four optional routes, and one of the important factors to affect
their decisions was the waiting time, which was closely related to The travel cost Ck of line rk* was as follows: (see (6))
bus frequencies. For mobile bus applications users, the frequency Thus, the total travel cost Cr* of the objective bus line r* was
of all lines could be known beforehand, and then the pre-trip time presented as follows: (see (7))
could be computed. The model assumed that a given mobile bus
application user selected the strategy for one of the following four 3.2 Frequency optimise model
principles. For users not using mobile bus applications, it was
assumed that they took the first arrived bus at the fixed stop. To For minimisation problems, the objective function (7) could then
some extent, it could be assumed that these users followed the be modified (see (8)) . In addition to the aforementioned major
second principle, i.e. shortest pre-trip time. Fig. 5 stated the logical elements, other issues on the objective bus line were necessary to
relationship among the inside factors. be considered. Examples include (i) the number of users using
mobile bus applications should not exceed the total number of
S1: Shortest total travel time; i.e. minimise users in each OD pair, and (ii) each bus could be served at a
{ctravel = (cpre − trip + con − board + cend − trip)}. maximum of one time. Consequently, two types of side constraints
were necessary: (i) the sum of probabilities of user selecting one
S2: Shortest pre-trip time; i.e. minimise {cpre − trip}.
strategy should be less than or equal to one, and (ii) the trips
S3: Shortest walking time; i.e. minimise {cwalking}. amount through the objective bus line should be less than or equal
S4: Shortest riding time; i.e. minimise {criding}. to the maximum bus capacity. The frequency optimisation model
could be formulated as follows:
Define S as the set of users’ selection strategies and PS as the
4
set of corresponding probabilities. S = {S1, S2, S3, S4},
PS = {PS1, PS2, PS3, PS4}. Define Vk as the set of trips through minimise Cr* = ∑ ∑ δkPY PSl βlck riding
+ δkPN β2ckriding
k∈K l=1
rk*, V k ∈ T k. vk was the amount of trips through rk*, vk ≤ δk. β1, β2, β3 (9)
and β4 were the variables that indicated whether a user j( j ∈ V k) + ∑ [1/(2 f r*) + ckend − trip( j)]
j ∈ Vk
took rk*, with the following expressions:
(see (1))
(see (2))

1 if ctravel(rk*) < ctravel(rk12, rk21, rk22) or ctravel(rk12, rk21, rk22) = 0

β1 = (1)
0 otherwise

1 if cpre − trip(rk*) < cpre − trip(rk12, rk21, rk22) or cpre − trip(rk12, rk21, rk22) = 0
β2 = (2)
0 otherwise

1 if cwalking(rk*) < cwalking(rk12, rk21, rk22) or cwalking(rk12, rk21, rk22) = 0

β3 = (3)
0 otherwise

1 if criding(rk*) < criding(rk12, rk21, rk22) or criding(rk12, rk21, rk22) = 0

β4 = (4)
0 otherwise

Ck = vkckriding + ∑ (ck pre − trip

( j) + ckend − trip( j))
j ∈ Vk
4 (6)
= ∑ δkPY PSl βlck riding
+ δkPN β2ckriding + ∑ [1/(2 f r*) + ck end − trip
( j)]
l=1 j ∈ Vk

IET Intell. Transp. Syst., 2019, Vol. 13 Iss. 4, pp. 596-604 599
© The Institution of Engineering and Technology 2018
 17519578, 2019, 4, Downloaded from [Link] by Algeria Hinari NPL, Wiley Online Library on [03/07/2024]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Cr* = ∑ Ck
k∈k
4 (7)
= ∑ ∑ δkPY PSl βlck riding
+ δkPN β2ckriding + ∑ [1/(2 f r*) + ck end − trip
( j)]
k∈K l=1 j ∈ Vk

minimise Cr*
4
(8)
= ∑ ∑ δkPY PSl βlckriding
+ δkPN β2ckriding + ∑ [1/(2 f r*) + ck end − trip
( j)]
k∈K l=1 j ∈ Vk

4 4.2 Tabu search

subject to ∑ PSl ≤ 1 (10)
l=1 Tabu search was a metaheuristic that guided a local procedure to
explore the solution space beyond local optimality. In order to do
PY + PN = 1 (11) that, the search may accept moves which worsen the current
solution and may avoid moves which were performed recently.
vk ≤ N, k∈K (12) One of the main components of Tabu search was the use of
memory. This paper focused on the short-term memory strategies.
lbr* ≤ f r* ≤ ubr* (13)
4.2.1 Memory, tenure and tabu classification: The most
vk ≤ δk, k∈K (14) common type of short-term memory, namely, recency-based
memory was used in this paper. A tabu list called TabuStart
δk > 0, k∈K recorded the last iteration when the frequency was changed. A
(15)
move TextX was tabu active if Iter ≤ TabuStart (TextX) + Tenure. In
this expression, Iter indicated the current iteration number while
ckriding ≥ 0, ckend − trip ≥ 0, k∈K (16) Tenure indicated the number of iterations where a move remained
tabu active. In general terms, the search mechanism usually avoids
The objective function (9) denoted the minimisation of the total moves which are tabu active. However, if the number of non-tabu
travel cost of the objective bus line. Constraint (10) ensured the active moves is small, additional mechanisms are adopted, among
amount of mobile bus applications users did not exceed the total them aspiration criteria is one.
number of users in each OD pair. Constraint (11) meant both the
passengers using mobile applications or not were considered in the
4.2.2 Aspiration criteria: Aspiration criteria were introduced to
model. Constraint (12) ensured that each bus was served at its
determine when tabu activation rules could be overridden. In the
maximum of one time. Constraint (13) held the frequency of the
paper, given that the computation of the objective value of a
objective bus line in a certain range. Constraint (14) ensured the
solution is costly, only neighbour solutions without tabu active
trips amount through the objective bus line was not more than the
status were evaluated. When the possible number of neighbour
total number of user in each OD pair. Constraints (15) and (16)
solutions obtained exclusively by means of non-tabu active moves
were used to calculate the optimisation model. In addition, some
was beyond a minimum threshold Θmin (a parameter of the
explanations about the variables were made. f r* was the decision
algorithm), the aspiration criterion implemented removed the tabu
variable. PSl was an uncertain variable. Its value directly affected active status with lowest TabuStart (TextX) + Tenure until reaching
the objective function and the objective value. βl(l = 1, 2, 3, 4) were the minimum number of neighbours stated by Θmin. This was
determined by (1)–(4), which were used to realise users’ choice known as aspiration by default.
behaviour. δk was known in advance. Values of the travel cost of all
arcs, i.e. ckriding and ckend − trip, could be obtained from mobile bus 4.2.3 Neighbourhood exploration: As a systematical way to
applications. explore the neighbour solutions, the aspiration plus strategy was
implemented. A number plus of solutions were explored after
4 Metaheuristic approach finding a solution which objective value improved an aspiration
value. To keep bounded the number of neighbour solutions to visit,
The proposed optimisation problem was expected to be hard to two values were used that indicated the minimum and maximum
solve, especially for large size instances. Therefore, a metaheuristic number of solutions to explore (parameters Min and Max,
approach based on Tabu search [17] was developed in this respectively). More specifically, if the aspiration value was reached
research. This method was based on a local search in the range of or surpassed during the exploration of the ith neighbour, then the
frequency. In order to avoid getting trapped in local optima, tabu search explores:
concepts were applied to the base local search mechanism. The
main concepts involved in the developed metaheuristic were a. i + plus neighbours if Min ≤ i + plus ≤ Max,
presented in the following paragraphs. b. Min neighbours if i + plus < Min, or
c. Max neighbours if Max < i + plus.
4.1 Local procedure
The local procedure was a search that used an operation called The candidate neighbours resulting from the application of the
move to define the neighbourhood of any given solution. The strategy explained above were evaluated in a random order.
solution of the proposed model was the frequency of the objective
bus line. Θ was assumed to be the set of possible frequencies, 5 Numerical tests
ordered from lowest (lbr*) to highest (ubr*) value. In this way, the In this section, a test bus line was selected and four cases of user
line increased (decreased) its frequency when it changed to the behaviour were considered to validate the proposed model. The test
next highest (lowest) value in Θ. The move operation that defined line and traffic zones were first introduced and then the numerical
the neighbourhood structure was then a compound change on a results were obtained from the application of the exact model.
couple of frequencies of the solution. Some analyses were discussed according to the numerical results.

600 IET Intell. Transp. Syst., 2019, Vol. 13 Iss. 4, pp. 596-604
© The Institution of Engineering and Technology 2018
 17519578, 2019, 4, Downloaded from [Link] by Algeria Hinari NPL, Wiley Online Library on [03/07/2024]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
5.1 Test bus line and traffic zones questionnaires. By sorting out and analysing these questionnaires,
some data could be collected, including: the OD matrix, the
The NO. J116 line in Jiading District, Shanghai, China, was proportion of using mobile bus applications, each user's time of
selected because it provided opportunities to test diverse situations. leaving home/office, walking time, waiting time, time of leaving
This line was 14.6 km long, with 19 stops and ran from Anting bus stop and riding time. Among these data, each user's walking
North Railway Station to Huangdu Bus Station. This line was not time and riding time were searched through mobile bus
too long but there were diverse traffic conditions and user demand applications. The behaviour of each user could also be gathered,
variations in route. which would be the basis of following model test results and
The traffic zones were the sets of origins and destinations of comparison analysis.
users. The service area of J116 was divided into 19 zones There were 19 traffic zones for J116, so the number of OD pairs
according to the stops setting, as shown in Fig. 6. Qk was 342. Considering the current maximum dispatching interval
Table 2 listed the codes, names of traffic zones in Fig. 6, and the
for J116 was 30 min, the paper set the policy minimum frequency
stops and bus lines within 5 min walking distance from these zones
lbr* was 1.5 run/h. Assume the accepted minimum dispatching
(see Table 2).
Also, the number of bus lines between each OD pair was interval for J116 was 10 min, then the maximum frequency ubr*
counted for a more detailed and direct display of the links between was 6 run/h. The type of vehicle for J116 was large bus, which has
each OD pair and the competitive relation between other bus lines 28 seats and a size 10,499 × 2500 × 3150 (mm). The standing area
and NO. J116, as shown in Table 3. It could be found that the NO.6 was about 8 m2. Suppose each square could hold seven users, then
and NO.7 zones were mainly passenger flow distribution centres the transport capacity per vehicle N could be 84 users.
(see Table 3).
5.3 Test results and comparison analysis
5.2 Data and parameters applied in model test
The all-day passenger flow distribution on the 19 stops of J116 was
For the model application, single-day data for 17 April 17 2017 shown in Fig. 7. As shown in Fig. 7, morning peak and evening
(Tuesday) was collected at every bus stop listed in Table 2, from peak were much in evidence. The main passenger distribution
05:00am to 19:00pm (service time of NO. J116). One researcher centres could also be found clearly. Therefore, the service time of
was assigned to each stop, who issued a survey questionnaire to J116 could be divided into six time periods according to passenger
users. The researcher needed to record the survey stop, arrival time, flow distribution: 05:00–07:00, 07:00–09:00, 09:00–11:00, 11:00–
departure time and bus line for each user. Besides, the survey 14:00, 14:00–17:00 and 17:00–19:00.
mainly contained the following six simple questions. About users’ behaviours, Fig. 8 illustrated the current
proportion of using mobile bus applications and the strategies of
① Your origin point. users choosing bus lines.
② Your destination point and alighting stop. As shown in Fig. 8, there were about 64% of users using mobile
③ If you used any mobile bus application before travel? (YES/NO) bus applications when planning trips, and the other 36% travelled
④ Your time of planning a trip and leaving home/office. following experiences or else. Among these mobile bus
applications users, 24% adopted strategy S1, i.e. shortest total travel
⑤ Which bus line will you take?
⑥ Why will you take this one? time, 10% adopted strategy S2, i.e. shortest pre-trip time, 16%
adopted strategy S3, i.e. shortest walking time, and the left 14%
a. Shortest total travel time (S1); adopted strategy S4, i.e. shortest riding time.
b. The earliest one (S2); In order to have more quantitative elements regarding the user
behaviour and application possibilities of the proposed method,
c. Shortest walking time (S3);
another three hypothetic cases were performed, according to the
d. Shortest riding time (S4). tendency of mobile bus applications. In other words, four cases
were considered to test the model. Case 1 was using the real data
Passengers were canvassed in 19 stops and they answered 2348 from actual survey. Case 2 and case 3 adjusted the users’
questionnaires totally, in which there were 1988 valid behaviours while keeping the proportion of using mobile bus

Fig. 6 J116 Line, stops and traffic zones code

IET Intell. Transp. Syst., 2019, Vol. 13 Iss. 4, pp. 596-604 601
© The Institution of Engineering and Technology 2018
 17519578, 2019, 4, Downloaded from [Link] by Algeria Hinari NPL, Wiley Online Library on [03/07/2024]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Table 2 Code and name of traffic zones, stops and bus lines within 5 min walking distance
Traffic zone Traffic zone name Stops within 5 min walking distance Bus lines within 5 min walking distance
code Name Number Name Number
1 Anting North Railway Anting North Railway Station 1 J116, J69 2
Station
2 Qinfu Zone Qinfu Jiayuan 1 J116, J55, J69 3
3 Heting Zone Minfeng Road-Xinyuan Road, Xinyuan 2 J102, J116, J55, J108, J69 5
Road-Hejing Road
4 Qinle Zone Minfeng Road-Moyu Road 1 J102, J116, J55 3
5 Yulan Zone Yulan Xincun, Changji Road-Moyu Road 2 J55, J67, J102, J103, J108, J116 6
6 Anting Station Anting Bus Station, Zepu Road-Moyu 2 J55, J67, J102, J103, J112, J114, J116, 9
Road Beian Line, Qingan Line
7 Jiatinghui City Life Anting Subway Station, Anting Station 2 J55, J67, J69, J102, J103, J112, J114, 9
Square J116, Anhong Line
8 Shanghai Auto Museum Moyu Road-Boyuan Road, Boyuan Road- 2 J112, J116, Anhong Line 3
Moyu Road
9 Wanli City Boyuan Road-South Miquan Road 1 J112, J116, Anhong Line 3
10 Shanghai Auto-expo Park Boyuan Road- South Yutian Road, 2 J116, Anhong Line 2
Boyuan Road-Yutian Road
11 SAIC Motor Boyuan Road-Anyan Road, Anyan Road- 2 J116, Anhong Line 2
Antuo Road
12 Automotive Innovation Antuo Road-Anhong Road 1 J116 1
Port
13 Anting Golf Club Anhong Road-Boyuan Road, Boyuan 2 J116, Anhong Line 2
Road-Anhong Road
14 Zhujia Village Boyuan Road-North Jiasong Road, North 2 J116, Anhong Line, Jiasong Line, Lu'an 4
Jiasong Road-Boyuan Road Highway Line B
15 Nigang Village Boyuan Road-Xinhuang Road 1 J109, J116, Anhong Line 3
16 Qianjia Village Boyuan Road-South Lv'yuan Road 1 J109, J116, Anhong Line, Lu'an Highway 4
Line B
17 Xujia Village Xujia Village, Boyuan Road-Xujia Village 2 J109, J116, Anhong Line 3
18 Lv'yuan Zone Lv'yuan Road-Xinhuang Road 1 J109, J116, J63, Lu'an Highway Line B 4
19 Huangdu Bus Station Huangdu Bus Station, Huangdu 2 J109, J116, J63, J123, Lu'an Highway 7
Line B, Jiasong Line, Qinghuang Line

Table 3 Number of bus lines between each OD pair

OD 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
1 — 2 2 1 1 1 2 1 1 1 1 1 1 1 1 1 1 1 1
2 2 — 3 2 2 2 3 1 1 1 1 1 1 1 1 1 1 1 1
3 2 3 — 3 4 3 4 1 1 1 1 1 1 1 1 1 1 1 1
4 1 2 3 — 2 3 3 1 1 1 1 1 1 1 1 1 1 1 1
5 1 2 4 2 — 5 5 1 1 1 1 1 1 1 1 1 1 1 1
6 1 2 3 3 5 — 7 2 2 1 1 1 1 1 1 1 1 1 1
7 2 3 4 3 5 7 — 3 3 2 2 1 2 2 2 2 2 1 1
8 1 1 1 1 1 2 3 — 3 2 2 1 2 2 2 2 2 1 1
9 1 1 1 1 1 2 3 3 — 2 2 1 2 2 2 2 2 1 1
10 1 1 1 1 1 1 2 2 2 — 2 1 2 2 2 2 2 1 1
11 1 1 1 1 1 1 2 2 2 2 — 1 2 2 2 2 2 1 1
12 1 1 1 1 1 1 1 1 1 1 1 — 1 1 1 1 1 1 1
13 1 1 1 1 1 1 2 2 2 2 2 1 — 2 2 2 2 1 1
14 1 1 1 1 1 1 2 2 2 2 2 1 2 — 2 3 2 2 3
15 1 1 1 1 1 1 2 2 2 2 2 1 2 2 — 3 3 2 2
16 1 1 1 1 1 1 2 2 2 2 2 1 2 3 3 — 3 3 3
17 1 1 1 1 1 1 2 2 2 2 2 1 2 2 3 3 — 2 2
18 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 3 2 — 4
19 1 1 1 1 1 1 1 1 1 1 1 1 1 3 2 3 2 4 —

applications same. Case 4 assumed a higher proportion of using case 2, the probability of S1 was decreased by 5% and the
mobile bus applications but same probabilities of users’ probability of S2 was increased by 5% correspondingly. In case 3,
behaviours, as shown in Table 4. conversely, the probability of S1 was increased by 5% and the
In case 2 and case 3, only the probabilities of S1 and S2 were probability of S2 was decreased by 5%.
adjusted, considering that of each user's walking time and riding The test results are shown in Table 5.
time were fixed and not about the frequency, while the pre-trip Fig. 9 showed the optimised frequencies of each time period in
time and total travel time were closely related to the frequency. In different cases.

602 IET Intell. Transp. Syst., 2019, Vol. 13 Iss. 4, pp. 596-604
© The Institution of Engineering and Technology 2018
 17519578, 2019, 4, Downloaded from [Link] by Algeria Hinari NPL, Wiley Online Library on [03/07/2024]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 7 All-day passenger flow distribution on stops of J116

Fig. 8 Current proportion of using mobile bus applications and the users’ choice behaviours
(a) Proportion of using mobile bus applications, (b) Strategies of users choosing bus lines

Table 4 Parameter setting of four cases

Proability Case
Case 1, % Case 2, % Case 3, % Case 4, %
S1 24 19 29 30
S2 10 15 5 12.5
S3 16 16 16 20
S4 14 14 14 17.5
others 36 36 36 20
Σ 100 100 100 100

Table 5 Test results of three cases

Time period Optimal service frequency, run/h Total travel cost, min
Case 1 Case 2 Case 3 Case 4 Case 1 Case 2 Case 3 Case 4
05:00–07:00 2.00 2.07 1.94 1.82 7364 7590 7147 6889
07:00–09:00 3.75 4.29 3.33 3.00 17,072 17,607 16,368 15,285
09:00–11:00 3.00 3.16 2.86 2.61 11,989 12,518 11,063 10,418
11:00–14:00 2.40 2.61 2.31 2.14 11,233 11,869 10,817 9789
14:00–17:00 2.73 3.00 2.61 2.40 13,968 14,150 13,018 12,098
17:00–19:00 4.00 4.62 3.53 3.16 18,045 18,607 17,321 16,733
total 17.88 19.74 16.57 15.13 79,671 82,340 75,734 71,212

Result analysis when increased the users’ will for shortest pre-trip time, the
As shown in Table 5 and Fig. 9, some interesting conclusions values of optimised frequencies were also increased
could be found. accordingly. If users paid much more attention on shortest total
travel time, the optimised frequencies were lower.
i. Overall, the bus frequencies in peak hours were obviously iv. Compare case 1 and case 4 at last. Case 4 increased the
higher than other hours, which was closely to the truth and proportion of using mobile bus applications while keeping the
consistent with the passenger flow distribution in Fig. 7. same user behaviours. It was an interesting phenomenon that
ii. By vertical comparing in Table 5, in peak hours, the the optimised frequencies decreased obviously. Although there
differences of optimised frequencies in four cases were much was only one test case for this assumption, the comparison
bigger than other periods. In trough hours, no matter adjust the partly reflected that the widespread use of mobile bus
users’ behaviours or increase the proportion of using mobile applications could save not only users’ travel cost, but also the
bus applications, the optimised frequencies were little changed. bus operating cost. This conclusion was consistent with part of
While in the peak hours, the optimised values changed greatly. Watkins et al. (2011), Ferris et al. (2011) and Dziekan and Karl
In other words, the higher passengers flow, the greater Kottenhoff's (2007) researches.
differences. In addition, the bus frequency optimal value was
more sensitive to changes of users’ behaviours and the
proportion of using mobile bus applications when there was a 6 Conclusions
higher passengers flow.
iii. Horizontally, compare case 1 with case 2 and case 3 first. Case In this paper, a bus frequency optimisation model was presented.
2 decreased and increased the proportions of S1 and S2, User behaviours affected by mobile bus applications were
respectively, and case 3 did conversely. It could be found that considered as major variables in this model. By assigning a user's
trip to three stages (pre-trip, on-board and end-trip), four different
IET Intell. Transp. Syst., 2019, Vol. 13 Iss. 4, pp. 596-604 603
© The Institution of Engineering and Technology 2018
 17519578, 2019, 4, Downloaded from [Link] by Algeria Hinari NPL, Wiley Online Library on [03/07/2024]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 9 Optimised frequencies of each time period in different cases

user behaviours were taken into account in the model building 8 References
process: shortest total travel time, shortest pre-trip time, shortest
[1] Dziekan, K., Kottenhoff, K.: ‘Dynamic at-stop real-time information displays
walking time and shortest riding time. In addition to the object bus for public transport: effects on customers’, Transp. Res. A, Policy Pract.,
line and stops on it, other bus lines and stops serving the same 2007, 41, (6), pp. 489–501
traffic zones were also considered. The demand was given in the [2] Schéele, S.: ‘A supply model for public transit services’, Transp. Res. B,
form of an OD matrix. Each value in the OD matrix represented the Methodol., 1980, 14, (1), pp. 133–146
[3] Furth, P.G., Wilson, N.H.M.: ‘Setting frequencies on bus routes: theory and
passenger amount from one zone to another. The goal of modelling practice’, Transp. Res. Rec., 1981, 818, (1981), pp. 1–7
was to obtain the optimal frequency of the objective bus line, with [4] Han, A.F., Wilson, N.H.M.: ‘The allocation of buses in heavily utilized
a function to minimise the users’ total travel cost. networks with overlapping routes’, Transp. Res. B, Methodol., 1982, 16, (3),
Numerical tests, using survey data from a set of traffic zones pp. 221–232
[5] Ceder, A.: ‘Bus frequency determination using passenger count data’, Transp.
corresponding to the objective bus line, in Shanghai, were Res. A, General, 1984, 18, (5–6), pp. 439–453
performed to validate the model. In order to describe the effect of [6] Constantin, I., Florian, M.: ‘Optimizing frequencies in a transit network: a
user behaviour changes on frequency optimisation, four cases were nonlinear bi-level programming approach’, Int. Trans. Oper. Res., 1995, 2,
calculated. Comparative results showed, with the use of mobile bus (2), pp. 149–164
[7] Tom, V.M., Mohan, S.: ‘Transit route network design using frequency coded
applications, bus frequencies were affected not only by the amount genetic algorithm’, J. Transp. Eng., 2003, 129, (2), pp. 186–195
of passengers, but also from the changes of user behaviours. In the [8] Gao, Z., Sun, H., Shan, L.L..: ‘A continuous equilibrium network design
information age, if users paid much more attention on shortest pre- model and algorithm for transit systems’, Transp. Res. B, Methodol., 2004,
trip time, there was a need to increase the frequency. In addition, it 38, (3), pp. 235–250
[9] Chen, H.: ‘Stochastic optimization in computing multiple headways for a
was also found that the widespread use of mobile bus applications single bus line’, J. Chin. Inst. Ind. Eng., 2007, 24, (5), pp. 351–359
could save both users’ travel cost and agencies’ operating cost. [10] Yan, S., Tang, C.H.: ‘An integrated framework for intercity bus scheduling
Nevertheless, the proposed model and the test results in this under stochastic bus travel times’, Transp. Sci., 2008, 42, (3), pp. 318–335
research should be useful as a reference for the development and [11] Yu, B., Yang, Z., Yao, J.: ‘Genetic algorithm for bus frequency optimization’,
J. Transp. Eng., 2009, 136, (6), pp. 576–583
further testing. [12] dell'Olio, L., Ibeas, A., Ruisánchez, F.: ‘Optimizing bus-size and headway in
transit networks’, Transportation, 2012, 39, (2), pp. 449–464
[13] Martínez, H., Mauttone, A., Urquhart, M.E.: ‘Frequency optimization in
7 Acknowledgments public transportation systems: formulation and metaheuristic approach’, Eur.
The authors express their gratitude to those who helped them J. Oper. Res., 2014, 236, (1), pp. 27–36
[14] Grotenhuis, J.W., Wiegmans, B.W., Rietveld, P.: ‘The desired quality of
during the writing of the paper. They specially thank the students integrated multimodal travel information in public transport: customer needs
who did great work in the data collection and arrangement. Finally, for time and effort savings’, Transp. Policy, 2007, 14, (1), pp. 27–38
they thank the friends who gave a hand in spelling and grammar [15] Watkins, K.E., Ferris, B., Borning, A., et al.: ‘Where Is My Bus? Impact of
correction. This paper is supported by the National Key Research mobile real-time information on the perceived and actual wait time of transit
riders’, Transp. Res. A, Policy Pract., 2011, 45, (8), pp. 839–848
and Development Plan of China (Grant No.2017YFC0804903) and [16] Ferris, B., Watkins, K., Borning, A.: ‘Onebusaway: behavioural and
the National Natural Science Foundation of China (Grant satisfaction changes resulting from providing real-time arrival information for
No.71601110). public transit’. Proc. of the 2011 Transportation research board annual
meeting, Washington, D.C., United States, January 2011, pp. 1807–1816
[17] Glover, F.: ‘Tabu search-part I’, ORSA J. Comput., 1989, 1, (3), pp. 190–206

604 IET Intell. Transp. Syst., 2019, Vol. 13 Iss. 4, pp. 596-604
© The Institution of Engineering and Technology 2018