Tendency-Based Approach for Link

Travel Time Estimation

Guojun Chen1; Jing Teng2; Shuyang Zhang3; and Xiaoguang Yang4
Downloaded from ascelibrary.org by Sree Vidyaniethan Engineering College on 11/09/13. Copyright ASCE. For personal use only; all rights reserved.

Abstract: From the historical bus trajectories, it was found that the headway bias amplified as buses travel on the route. When controls on
buses are unavailable, the following buses will maintain the movement tendency toward their previous one, whether being closer to, farther
away, or stable, judged by the running state in which buses fall. A tendency-based model for link travel time estimation was proposed, and
three tendency-based corrections were introduced in the model, which are the long-term tendency, the short-term tendency, and the combined-
term tendency. Then, contrast experiments were conducted in which the boundary of the running state is a control variable to show the
performance of the tendency-based model under different boundary values. The experiment results show that, with the increase of the boun-
dary value, the degree of improvement of the tendency-based mode to the historical data–based model first increases, then decreases, and
converges to zero finally. The optimal boundary value for the tendency-based model was calibrated, judged by the net number of trips
improved and net mean absolute error reduced, and the results show that the long-term and the combined-term tendency-based models
have a lower optimal boundary and higher optimization potential, and are faster to be steady enough, which made the short-term
tendency-based model less competitive. DOI: 10.1061/(ASCE)TE.1943-5436.0000486. © 2013 American Society of Civil Engineers.
CE Database subject headings: Buses; Travel time; Estimation.
Author keywords: Bus transit; Link travel time; Estimation; Movement tendency.

Introduction but the premise is the knowledge of the service performance in

advance.
Before the wide equipment of automatic vehicle location (AVL) The prediction of bus travel time can help measure the service
devices, it was almost impossible for transit agencies to automati- performance in the future, which is valuable for optimizing the con-
cally acquire bus travel time in real time, and buses were generally trol schemes. The accurate prediction of bus travel time is always a
assumed to be maintaining the service reliability through the trip. topic of concern because bus travel time is affected by a great many
From the 1990s, AVL devices have been widely equipped by transit factors. In general, those factors can be divided into five categories,
agencies all over the world. By the year 2000, 88 transit agencies in which include infrastructure (Alfa et al. 1988; Sun and Hickman
the United States had operational AVL systems, and 142 were plan- 2006; Tétreault and El-Geneidy 2010; Chen et al. 2005; Chen et al.
ning such system [Transit Cooperative Research Program (TCRP) 2012; Chien and Kuchipudi 2003), such as the number of stops,
2003]. A major application of the real-time location data is to opti- the number of intersections, and the length of segment; running
mize the control schemes for recovering the detected deviations; environment (Abkowitz and Engelstein 1982; Cathey and Dailey
nevertheless, the deviations sometimes may be so large that the 2003; Chen et al. 2012; Chien and Kuchipudi 2003; Sun and
generated control schemes are very costly for the operation and Hickman 2006; Tétreault and El-Geneidy 2010), such as the
cannot be applied in practice. Proactive control schemes are ex- weather condition, the time periods, and the traffic congestion; op-
pected to avoid the deviations and reduce the burden on operation, eration and management (Chen et al. 2005; Cathey and Dailey
2003; Lin and Zeng 1999; Sun and Hickman 2006), such as the
scheduled travel time; passenger demand and activity (Sun and
1
Ph.D. Candidate, Key Laboratory of Road and Traffic Engineering of Hickman 2006; Tétreault and El-Geneidy 2010), such as the board-
the Ministry of Education, Tongji Univ., 4800 Caoan Highway, Shanghai ing and alighting time and the passenger arrival rate; and driver
201804, P.R. China. E-mail: [email protected] behavior (Chen et al. 2005; Lin and Bertini 2004; Lin and Zeng
2
Associate Professor, Key Laboratory of Road and Traffic Engineering
1999), such as the schedule recovery.
of the Ministry of Education, Tongji Univ., 4800 Caoan Highway, Shanghai
201804, P.R. China (corresponding author). E-mail: tengjing@tongji The most stated models for bus travel time prediction/estimation
.edu.cn include the historical data–based model, the regression model, the
3
Master Candidate, Key Laboratory of Road and Traffic Engineering of kalman filter algorithm, the artificial neural network model, the
the Ministry of Education, Tongji Univ., 4800 Caoan Highway, Shanghai Markov chains model, and the support vector machine. Advanced
201804, P.R. China. E-mail: [email protected] models are proved to be superior to the historical data–based model
4
Professsor, Key Laboratory of Road and Traffic Engineering of the on prediction/estimation accuracy; on the other hand, the historical
Ministry of Education, Tongji Univ., 4800 Caoan Highway, Shanghai data–based model has advantages in its scalability to account for
201804, P.R. China. E-mail: [email protected] flexible routes (Ni and Wang 2008), ease of calibration, less com-
Note. This manuscript was submitted on December 19, 2011; approved
on August 7, 2012; published online on March 15, 2013. Discussion period
putation burden, and ease of program coding, which make it widely
open until September 1, 2013; separate discussions must be submitted for used in practice. The major drawback of the historical data–based
individual papers. This paper is part of the Journal of Transportation En- model is the assumption that travel time will remain constant during
gineering, Vol. 139, No. 4, April 1, 2013. © ASCE, ISSN 0733-947X/ the same time period, like the trip-specific or the day-specific (Sun
2013/4-350-357/$25.00. and Hickman 2006), which ignores the variation of travel time and

350 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / APRIL 2013

J. Transp. Eng. 2013.139:350-357.

 leads to estimation errors, especially in some special cases, like reliability, and the maintaining of it is also the objective of oper-
bunching. ation controls.
From the collected bus trajectories, it was found that the head- The headway bias, which is the time gap between the headways
way bias between trips amplified as buses travel on the route, which of adjacent trips at two bus stops, is introduced as a metric for
indicates that there is a movement tendency between trips. Correc- measuring the fluctuation of headway.
tions in accord with the movement tendency between trips were
introduced in the historical data–based model, which was expected
to improve the performance of estimation accuracy of the historical ΔhSi Sj ðk; k þ 1Þ ¼ hSj ðk; k þ 1Þ − hSi ðk; k þ 1Þ;
data–based model. The rest of paper is organized as follows: in the i; j ¼ 1; 2; : : : ; N and i < j ð1Þ
next section, the experimental route and collected data are intro-
duced; then the movement tendency between trips is defined
where ΔhSi Sj ðk; k þ 1Þ = headway bias between trips k and (k þ 1)
Downloaded from ascelibrary.org by Sree Vidyaniethan Engineering College on 11/09/13. Copyright ASCE. For personal use only; all rights reserved.

and a tendency-based model for link travel time estimation is pro-

posed with three tendency-based corrections; following are groups from stop Si to stop Sj ; hSi ðk; k þ 1Þ = headway between trips k
of contrast experiments in which the boundary is a control variable and (k þ 1) at stop Si ; and N = number of stops on the bus route.
to show the performance of the tendency-based model under differ- The headway bias from the departure terminal stop to the stop at
ent boundary values, and the optimal boundary of the running state which buses arrive is defined as the path headway bias, while the
for all trips is calibrated. headway bias from the previous stop to the stop at which buses
arrive is defined as the link headway bias. After buses depart from
the departure terminal stop, they are affected by a great many fac-
Data Collection tors, which lead to the fluctuation of bus headway between trips.
From the historical bus trajectories, it was found that the path head-
The test bed is the southbound corridor of bus route No. 49 located way bias between trips tends to be amplified (as shown in Fig. 3),
in Shanghai, China, which has 20 stops and is 12.17 km long (as and when the path headway bias is beyond a critical value, it seems
shown Fig. 1). The experimental route is frequently serviced and the impossible for buses to recover the deviation by themselves when
scheduled headway is between 5 and 10 min most of the time. No controls are unavailable, which finally leads to the service irregu-
time-check stops are located on the bus route and buses are running larity, like bunching or a long service interval.
on the route without control after departing from the terminal stop. To describe the characteristics of buses running, three running
AVL devices have been equipped on all buses in service, and states were introduced, the stable state, the closing state, and the
bus locations are collected and archived in real time. Three types departing state. If hΔ and h∇ are the upper and lower limits of the
of location data are collected—the location-at-time data (when stable state, then
buses have traveled for 30 s between stops), the location-at-distance • In the stable state, when the path headway bias falls in the stable
data (when buses have traveled 250 m within 30 s between stops), boundary, that is, h▿ ≤ ΔhAS1 Sn ðk; k þ 1Þ ≤ hΔ , the fluctuation
and the location-at-stop data (when buses arrive at a stop or depart
from it).
AVL records on May 7, 2010, which consists of 178 trips, were 49019
selected out from the database for data analysis and model evalu-
49016
ation (as shown in Fig. 2).
49013
Stop No.

49010
Movement Tendency between Buses
49007
For a frequently serviced bus route, the headway regularity is 49004
always considered as the most important component of the service
49001
4:00 6:00 8:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 0:00
N Hankou Road Stop Time of day (Date: May 7, 2010)
Shanxi South Road Stop
Xizang Middle Road Stop
Huangpi North Road Stop Fig. 2. Historical bus trajectories (178 trips)
S Shimen NO.1 Road Stop
Shanxi North Road Stop
Tongren Road Stop
15
Huashan Road Stop Departure Terminal
Changshu Road Stop Midway stop 10
Huaihai Middle Road Stop
Headway Bias/min

Destination Terminal
Hengshan Road Stop 5
Jianguo West Road Stop
Zhongshan Hospital Stop 0
Fenglin Road Stop
-5
No.2 Bus Company Stop
Cancer Hospital Stop
Shanghai Stadium Stop -10
Dongan Newly State Stop
Wanping South Road Stop -15
49001 49004 49007 49010 49013 49016 49019
Shanghai Swimming Gymnasium Stop Stop No.

Fig. 1. Illustration of the experimental route (southbound) Fig. 3. Path headway bias with buses traveling on the route

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / APRIL 2013 / 351

J. Transp. Eng. 2013.139:350-357.

 of bus headway is random and it is uncertain whether the path Historical Data–Based Model
headway bias will be enlarged or narrowed;
In the historical data–based model, it is assumed that link travel
• In the closing state, when the path headway bias is below the
time remains constant during the same period. In other words,
lower limit of the stable state, that is, ΔhAS1 Sn ðk; k þ 1Þ < h∇ ,
the link travel time of the following trip is assumed to be the same
the stable state is broken and the path headway bias will be con-
as that of the previous trip.
tinuously narrowed; and
• In the departing state, when the path headway bias is over the tESn Snþ1 ðk þ 1Þ ¼ tASn Snþ1 ðkÞ ð2Þ
upper limit of the stable state, that is, ΔhAS1 Sn ðk; k þ 1Þ > h∇ ,
the stable state is broken as well and the path headway bias will where tESn Snþ1 ðk
þ 1Þ = estimated travel time on the link Sn Snþ1
be continuously enlarged. of trip (k þ 1); and tASn Snþ1 ðkÞ = actual travel time on the link
When buses fall into the closing state or departing state, the ser- Sn Snþ1 of trip k.
Downloaded from ascelibrary.org by Sree Vidyaniethan Engineering College on 11/09/13. Copyright ASCE. For personal use only; all rights reserved.

vice regularity has been broken and it is hard for buses to recover
the deviations by themselves. Daganzo and Pilachowski (2011)
pointed out that buses may still lose ground to the buses they follow Tendency-Based Model
when the gaps are beyond a critical value, even if the schedule- Compared with the historical data–based model, corrections were
based or headway-based control schemes are available. introduced in the tendency-based model with respect to the move-
The movement tendency is defined as the following buses main- ment tendency-based trips. The path headway bias is the long-term
taining the running state to the preceding one as they travel on the fluctuation of bus headway, which stands for the long-term move-
bus route when controls are unavailable. In accord with the running ment tendency between trips, while the link headway bias is the
states, the movement tendency has three types as well: short-term fluctuation of bus headway and stands for the short-term
• Maintaining tendency in which the headway will remain stable movement tendency between trips. Three types of corrections were
as buses travel on the route if they fall in the stable state; considered in the tendency-based model—a long-term tendency-
• Closing tendency in which the headway will be narrowed based correction, a short-term tendency-based correction, or a
as buses travel on the route if they fall in the closing combined-term tendency-based correction.
state; and
• Departing tendency in which the headway will be enlarged as Long-Term Tendency-Based Model
buses travel on the route if they fall in the departing state. The long-term tendency-based correction method is based on the
assumption that the path headway bias will be further amplified on
the next link in proportion when buses fall in the closing state or the
Link Travel Time Estimation Model departing state, or remain the same when buses fall in the stable
state.
A tendency-based model to link travel time estimation was pro-
posed in which corrections were introduced with respect to the tESn Snþ1 ðk þ 1Þ ¼ tASn Snþ1 ðkÞ þ β LSn Snþ1 ðk þ 1ÞΔhS1 Sn ðk; k þ 1Þ ð3Þ
movement tendency between trips, which is judged by the running
state in which buses fall. The historical data–based model, in con- where β LSn Snþ1 ðk þ 1Þ = long-term tendency-based adjustment coef-
trast, was proposed as well. ficient, which is judged by the states in which buses fall.

8
< 0;
> ðh▿ ≤ ΔhS1 Sn ðk; k þ 1Þ ≤ hΔ Þ
β LSn Snþ1 ðk þ 1Þ ¼ 1 A ð4Þ
>
: tA ðkÞ tSn Snþ1 ðkÞ; ðΔhS1 Sn ðk; k þ 1Þ > hΔ or ΔhS1 Sn ðk; k þ 1Þ < h∇ Þ
S1 Sn

Short-Term Tendency-Based Model

The short-term tendency correction method is based on the assumption that the headway bias on the previous link will be propagated to the
next link in proportion when buses fall in the closing state or the departing state, or the headway bias on the next link will be equal to zero
when buses fall in the stable state.

tESn Snþ1 ðk þ 1Þ ¼ tASn Snþ1 ðkÞ þ β SSn Snþ1 ðk þ 1ÞΔhSn−1 Sn ðk; k þ 1Þ ð5Þ

where β SSn Snþ1 ðk þ 1Þ = short-term tendency-based adjustment coefficient, which is judged by the states in which buses fall.
8
< 0;
> ðh▿ ≤ ΔhS1 Sn ðk; k þ 1Þ ≤ hΔ Þ
S 1
β Sn Snþ1 ðk þ 1Þ ¼ ð6Þ
>
: tA tASn Snþ1 ðkÞ; ðΔhS1 Sn ðk; k þ 1Þ > h▿ or ΔhS1 Sn ðk; k þ 1Þ < h∇ Þ
Sn−1 Sn ðkÞ

Combined-Term Tendency-Based Model tESn Snþ1 ðk þ 1Þ ¼ tASn Snþ1 ðkÞ þ αβ LSn Snþ1 ðk þ 1ÞΔhS1 Sn ðk; k þ 1Þ
The combined-term correction method is based on the assumption
þ ð1 − αÞβ SSn Snþ1 ðk þ 1ÞΔhSn−1 Sn ðk; k þ 1Þ ð7Þ
that the headway bias on the next link is affected by the compre-
hensive effect of the long-term tendency and the short-term ten- where α = weight of the long-term tendency, which is assumed to
dency mentioned previously. be 0.5.

352 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / APRIL 2013

J. Transp. Eng. 2013.139:350-357.

 Model Validation and Evaluation Measures of Performance

As mentioned previously, the performance of the tendency-based All measures developed subsequently made use of the difference
models is determined by the boundary of the running state. It is between the estimated link travel time and the actual link travel
always the case that the path headway bias between trips at all stops time, which is known as the estimation errors.
is between the maximum and the minimum values.
εSn Snþ1 ðkÞ ¼ tESn Snþ1 ðkÞ − tASn Snþ1 ðkÞ ð9Þ

δ min ðk; k þ 1Þ ≤ ΔhS1 Sn ðk; k þ 1Þ ≤ δ max ðk; k þ 1Þ; n ¼ 2; : : : ; N where tESn Snþ1 ðkÞ = estimated travel time of trip k on the link Sn Snþ1 ;
tASn Snþ1 ðkÞ = actual travel time of trip k on the link Sn Snþ1 ; and
ð8Þ
εSn Snþ1 ðkÞ = travel time estimation error of trip k on the link Sn Snþ1 .
Downloaded from ascelibrary.org by Sree Vidyaniethan Engineering College on 11/09/13. Copyright ASCE. For personal use only; all rights reserved.

Accuracy is the closeness degree of the estimated value to the

where δ min ðk; k þ 1Þ = minimum path headway bias between trips true value. The mean absolute error (MAE) is the metric to measure
k and (k þ 1) at all stops; and δ max ðk; k þ 1Þ = maximum path the accuracy of the model, and a lower MAE value indicates a
headway bias between trips k and (k þ 1) at all stops. higher accuracy of the model.
The relative sizes between the lower and upper limits of the run-
ning state and the maximum and minimum values of the headway 1 X N−1
MAEðkÞ ¼ jε ðkÞj ð10Þ
bias play an important role in state recognition. When the upper N − 1 n¼1 Sn Snþ1
limit of the running state is larger than the maximum headway bias
and the lower limit is smaller than the minimum headway bias, the where MAEðkÞ = mean absolute error of the trip k; and N = number
tendency-based model will lose their sensitivity to the headway of stops, which equal 20 for the experimental route.
variation completely and will perform the same with the histori- To show the differences in estimation accuracy between the
cal-based model without improvement. A large boundary may lead tendency-based model and the historical data–based model, relative
to a tiny improvement of the tendency-based model to the historical mean absolute errors of trips were calculated by subtracting
data–based model. On the other hand, when the boundary of the the MAE of the historical data–based model from that of the
running state is quite small, the tendency-based model may mistake tendency-based model.
the occasional variation of headway bias as the running state and
amplify it in the link travel time estimation, which will cause large MAEr ðkÞ ¼ MAEt ðkÞ − MAEh ðkÞ ð11Þ
errors to the tendency-based model.
To analyze the effect of the boundary on the performance of the where MAEh ðkÞ = MAE of trip k of the historical data–based
tendency-based model, three groups of contrast experiments were model; MAEt ðkÞ = MAE of trip k of the tendency-based
conducted for the long-term–based model, the short-term tendency- model; and MAEr ðkÞ = relative MAE of trip k between the
based model, and the combined-term tendency-based model, historical data–based model and the tendency-based models. If
respectively. In each group of experiments, the boundary of the run- MAEr ðkÞ < 0, then the accuracy of the tendency-based model is
ning state is a control variable, which varies from
0% to
100%, superior to that of the historical data–based model, and vice versa.
increasing by 10%, of the maximum absolute values of the head- Additionally, the percentage of improvement on estimation
way bias between trips (as shown in Fig. 4). When the boundary of accuracy of the tendency-based model from the historical data–
the running state was equal to
100% of the maximum absolute based model is measured as well.
headway bias, the tendency-based model will completely lose its
MAEr ðkÞ
sensitivity to the variation of the headway bias, which results in MAEp ¼ × 100% ð12Þ
the same performance of the tendency-based model as the historical MAEh ðkÞ
data–based model.
In this way, 34 groups of estimation results on 19 links of 177
trips (the first trip will not be estimated) were obtained, with 33 Results Analysis
results for the tendency-based model and 1 for the historical
With the experiment results, comparisons were made on three as-
data–based model.
pects between the historical data–based model and the tendency-
based model under different trips and different boundary values
of the running state.
1200
Performance Comparison under Different Trips
Maximum Absolute Headway Bias/s

Different trips have different running states, judged by the variation

900 of headway bias between trips. The estimation results under differ-
ent trips can be put in to two categories, trips with constant running
state in which the tendency-based model always performs better
600 (constant closing or departing state) or worse (constant stable state)
than the historical data–based model regardless of the scale of the
boundary values, and trips with a fluctuating running state in which
300
the tendency-based model performs better than the historical data–
based model only under a certain boundary of the running state
0
(as shown in Fig. 5).
2 46 90 134 178 The long-term tendency-based model and the combined-term
Trip No. tendency-based model are superior to the short-term tendency-
based model on both number of trips improved and the degree
Fig. 4. Maximum absolute headway bias between trips
of improvement. Additionally, the combined-term tendency-based

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / APRIL 2013 / 353

J. Transp. Eng. 2013.139:350-357.

 30 30

Relative Mean Absolute Errors/s

Relative Mean Absolute Errors/s
20 20

10 10

0 0

-10 -10

-20
Downloaded from ascelibrary.org by Sree Vidyaniethan Engineering College on 11/09/13. Copyright ASCE. For personal use only; all rights reserved.

-20

-30 -30
2 46 90 134 178 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
(a) Boundary of the running state/percentage
(a) Trip No.
40
40

Relative Mean Absolute Errors/s

30
30
Relative Mean Absolute Errors/s

20
20
10
10
0
0
-10
-10
-20
-20
-30
-30 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
2 46 90 134 178 (b) Boundary of the running state/percentage
(b) Trip No.

30 30
Relative Mean Absolute Errors/s

Relative Mean Absolute Errors/s

20 20

10 10

0 0

-10 -10

-20 -20

-30 -30
2 46 90 134 178 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
(c) Trip No. (c) Boundary of the running state/percentage

Fig. 5. Relative MAE of the tendency-based model on trips: (a) long- Fig. 6. Relative MAE of the tendency-based model under percentage
term tendency-based model; (b) short-term tendency-based model; of boundary values: (a) long-term tendency-based model; (b) short-
(c) combined-term tendency-based model term tendency-based model; (c) combined-term tendency-based model

model seems to be more stable than the long-term tendency-based data–based model than the short-term tendency-based model. How-
model because the relative MAE values are more centralized. ever, the differences among them become less significant because
the distribution of the relative mean absolute errors narrowed
Performance Comparison under Different Boundary Values gradually with the increase of the boundary value, and all of them
(Percentage) converged to zero finally (as shown in Fig. 6).
Under different boundary values of the running state, the estimation A critical percentage value of the boundary of the running state
accuracy of the tendency-based model is different as well. The for all trips has not been observed because the relative MAE values
long-term tendency-based model and the combined-term tendency- of all trips under the same boundary are not consistent over or
based model achieves a greater improvement to the historical below zero.

354 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / APRIL 2013

J. Transp. Eng. 2013.139:350-357.

 30 180

Relative Mean Absolute Errors/s 20 150

Trips improved (long-term tendency)

10 120

Number of trips
Trips worsened (long-term tendency)

0 Trips improved (short-term tendency)

90
Trips worsened (short-term tendency)
-10 Trips improved (combined-term tendency)
60
Trips worsened (combined-term tendency)
-20
30
Downloaded from ascelibrary.org by Sree Vidyaniethan Engineering College on 11/09/13. Copyright ASCE. For personal use only; all rights reserved.

-30
0 300 600 900 1200 0
0 2 4 6 8 10 12 14 16 18
(a) Boundary of the running state/s (a) Boundary of the running state/min

40
60
Long-term tendency
30
Relative Mean Absolute Errors/s

30 Short-term tendency

Net improved number of trips

Combined-term tendency
20
0
10
-30
0
-60
-10
-90
-20
-120
-30
0 300 600 900 1200
(b) -150
Boundary of the running state/s 0 2 4 6 8 10 12 14 16 18
(b) Boundary of the running state/min
30
Fig. 8. Number of trips affected by the absolute boundary values:
Relative Mean Absolute Errors/s

20
(a) number of trips improved or worsened; (b) net improved number
of trips
10

-10 [as shown in Fig. 8(a)]. The net number of trips improved,
subtracting the number of trips worsened from the number of trips
-20 improved, achieves its highest value at 4.00 to 4.10 min for the
long-term tendency-based model (49 trips), at 5.80 to 5.85 min for
-30 the combined-term tendency-based model (34 trips), and at 4.38 to
0 300 600 900 1200
(c)
4.40 min for the short-term tendency-based model (45 trips) [as
Boundary of the running state/s
shown in Fig. 8(b)]. When the boundary is larger than 8.98 min,
Fig. 7. Relative MAE of the tendency-based model under absolute there will be no trips worsened in the tendency-based model, but
boundary values: (a) long-term tendency-based model; (b) short-term the trips improved are rare.
tendency-based model; (c) combined-term tendency-based model At the same time, the accumulated relative MAE values of the
trips improved and that of the trips worsened drops gradually with
the increase of the boundary values, but the former drops much
faster than the latter, especially when the boundary ranges from
Performance Comparison under Different Boundary Values 0 to 6 min [as shown in Fig. 9(a)]. The net MAE values of the
(Absolute) tendency-based model, pulsing the accumulated MAE values of
In contrast, the absolute values of the boundary were matched the trips improved with that of the trips worsened, achieves its
with the estimation results, and an amazing phenomenon was lowest value at 4.07 min for the long-term tendency-based model
found that when the absolute value of the boundary is beyond (−301 s), at 6.22 min for the short-term tendency-based
a critical value, the tendency-based model will always gain an model (−175 s), and at 4.38 min for the combined-term
improvement on estimation accuracy to the historical data–based tendency-based model (−277 s) [as shown in Fig. 9(b)].
model, which proved the assumption in movement tendency def- In fact, when the boundary is larger than 1.28 min for the long-
inition (as shown in Fig. 7). Additionally, the degree of improve- term tendency-based model, 3.82 min for the short-term tendency-
ment drops with the increase of the boundary values after it is based model, and 2.20 min for the combined-term tendency-based
beyond the critical value. model, the tendency-based model begins performing better than the
historical data–based model on estimation accuracy.
Boundary Calibration Judged by the net optimized number of trips and net improved
With the increase of the boundary values, the number of trips wors- estimation accuracy, the optimal boundary was calibrated at
ened drops at a faster rate than the number of trips improved 4.07 min for the long-term tendency-based model, 4.40 min for

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / APRIL 2013 / 355

J. Transp. Eng. 2013.139:350-357.

 1800 70

Accumulated Relative MAE of trips/s 60

MAE of improved trips (long-term tendency)
1200

Mean Absolute Errors/s

MAE of worsened trips (long-term tendency) 50
MAE of improved trips (short-term tendency)
40
MAE of worsened trips (short-term tendency)
600
MAE of improved trips (combined-term tendency) 30
MAE of worsened trips (combined-term tendency)

0 20

10 Long-term tendency Combined-term tendency

Downloaded from ascelibrary.org by Sree Vidyaniethan Engineering College on 11/09/13. Copyright ASCE. For personal use only; all rights reserved.

Short-term tendency Historical model

-600 0
0 2 4 6 8 10 12 14 16 18 2 46 90 134 178
(a) Boundary of the running state/min
(a) Trip No.

1500 30

1200
20
900
Net relative MAE/s

Relative MAE Values/s

10
600
0
300

0 -10
Long-term tendency
-300 -20 Long-term tendency
Short-term tendency
Short-term tendency
Combined-term tendency
-600 Combined-term tendency
0 2 4 6 8 10 12 14 16 18 -30
(b) Boundary of the running state/min 2 46 90 134 178
(b) Trip No.
Fig. 9. Estimation accuracy affected by the absolute boundary values:
60%
(a) relative MAE of improved or worsened trips; (b) net improved
relative MAE of all trips
Relative Percentage of MAE Values

40%

20%
the combined-term tendency-based model, and approximately
6 min for the short-term tendency-based model. The long-term 0%
tendency-based model is slightly better than the combined-
term tendency-based model at the optimal solution to the boundary -20%
of the running state, but both of them are significantly better than
Short-term tendency
the short-term tendency-based model, judged by the net improved -40%
Combined-term tendency
estimation accuracy and number of trips optimized. Long-term tendency
-60%
2 46 90 134 178
Model Evaluation
(c) Trip No.
With the boundary of the running state set at 4.00 min, estimation
on all trips was conducted to show the degree of improvement
Fig. 10. Comparison of the estimation accuracy between models with
of the tendency-based model to the historical data–based model
calibrated boundary: (a) MAE among models on different trips;
(as shown in Fig. 10).
(b) relative MAE of tendency-based models on different trips; (c) rela-
The improvement of the tendency-based model was mainly cen-
tive percentage of MAE of tendency-based models on different trips
tralized around the peak hours (from 7:00 to 10:00 and from 15:00
to 18:00), but also companied with failures in improvement. The
percentage of improvement can reach up to 40%, but that of
deterioration is at the same scale as well [as shown in Fig. 10(c)]. explanation for this phenomenon is that the running state is not
Fortunately, the worsened trips with a high percentage are mostly at completely stable, and buses are still able to transfer from the clos-
low MAE values. ing or departing state to the stable state by the occasional variation.
The failures in improvement of the tendency-based model de- Additionally, the short-term tendency-based model is much slower
crease with the increase of the maximum absolute headway bias. to be steady enough, and the performance is less competitive
The long-term and the combined-term tendency-based model will as well.
not fail at all in the improvement to the historical data–based model
on estimation accuracy, but the number of trips improved is limited
when the maximum absolute headway bias is larger than approx- Summary, Discussion, and Conclusion
imately 8.38 min (as shown in Fig. 11). However, when the maxi-
mum absolute headway bias is between 4.00 to 8.38 min, Focusing on the phenomenon that headway bias tends to be am-
both gains and losses in improvement are possible; a reasonable plified as buses travel on the route, a tendency-based model for link

356 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / APRIL 2013

J. Transp. Eng. 2013.139:350-357.

 30
Acknowledgments
Relative MAE values/s 20 This work was supported by the National Natural Science Founda-
tion of China (Grant No. 61174185). The authors wish to acknowl-
10 edge Shanghai Transportation Investment Group Co, LTD for
providing AVL data during the project.
0

-10 References

-20
Long-term tendency Abkowitz, M., and Engelstein, I. (1982). “Temporal and spatial dimensions
Short-term tendency
Downloaded from ascelibrary.org by Sree Vidyaniethan Engineering College on 11/09/13. Copyright ASCE. For personal use only; all rights reserved.

of running time in transit systems.” Transportation Research Record

Combined-term tendency
877, Transportation Research Board, Washington, DC, 64–67.
-30
0 2 4 6 8 10 12 14 16 18 20 Alfa, A., Menzies, W., Purcha, J., and McPherson, R. (1988). “A regression
Maximum absolute headway bias/min model for bus running times in suburban areas of Winnipeg.” J. Adv.
Transp., 21(3), 227–237.
Fig. 11. Relative MAE of the tendency-based model against the Cathey, F. W., and Dailey, D. J. (2003). “A prescription for transit arrival/
maximum absolute headway bias departure prediction using automatic vehicle location data.” Transp.
Res. Part C, 11(3), 241–264.
Chen, G., Yang, X., An, J., and Zhang, D. (2012). “Bus-arrival-time pre-
diction models: Link-based and section-based.” J. Transp. Eng., 138(1),
60–66.
travel time estimation was proposed with respect to the movement Chen, M., Liu, X. B., and Xia, J. X. (2005). “Dynamic prediction method
tendency between trips, which is judged by the running state into with schedule recovery effect for bus arrival time.” Transportation
which buses fall. The proposed tendency model is expected to be Research Record 1923, Transportation Research Board, Washington,
an improvement to the historical data–based model with three DC, 208–217.
tendency-based corrections introduced, the long-term tendency, Chien, S., and Kuchipudi, C. M. (2003). “Dynamic travel time prediction
the short-term tendency, and the combined-term tendency. with real-time and historical data.” J. Transp. Eng., 129(6), 608–616.
With the increase of the boundary of the running state, the Daganzo, C. F., and Pilachowski, J. (2011). “Reducing bunching with bus-
to-bus cooperation.” Transp. Res. Part B, 45(1), 267–277.
tendency-based model achieves a better performance of estimation
Lin, W., and Bertini, R. (2004). “Modeling schedule recovery processes in
accuracy than the historical data–based model when the boundary transit operations for bus arrival time prediction.” J. Adv. Transp., 38(3),
value is beyond a critical value, then achieves the best improvement 347–365.
to the historical data–based model when the boundary value is Lin, W-H., and Zeng, J. (1999). “Experimental study on real-time bus arrival
the optimal one, and finally converges to zero. The long-term time prediction with GPS data.” Transportation Research Record 1666,
tendency-based model achieves the best performance among the Transportation Research Board, Washington, DC, 101–109.
three tendency-based models, while the short-term tendency-based Ni, D., and Wang, H. (2008). “Trajectory reconstruction for travel time es-
model shows the lowest improvement to the historical data– timation.” J. Intell. Transport. Syst., 12(3), 113–125.
based model, judged by the net number of trips improved and net Sun, A., and Hickman, M. (2006). “Vehicle travel time prediction
relative MAE values improved. The combined-term tendency- using primitive AVL data.” Proc., 85th Transportation Research
based model, which is slightly worse than the long-term Board Annual Meeting (CD-ROM), Transportation Research Board,
Washington, DC.
tendency-based model, shows a higher stability because the relative
Tétreault, P., and El-Geneidy, A. (2010). “Estimating bus run times for new
MAE values are more centralized. limited-stop service using archived AVL and APC data.” Transp. Res.
The calibrated boundary of the running state is of great value to Part A, 44(6), 390–402.
the operation control as well because buses are faced with a higher Transit Cooperative Research Program (TCRP). (2003). “Real-time bus
risk of amplifying the headway irregularity when the headway bias arrival information systems: A synthesis of transit practice.” Synthesis
is beyond the stable boundary, which may lead to undesired bunch- 48, Transportation Research Board of the National Academies,
ing or large service interval conditions. Washington, DC.

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / APRIL 2013 / 357

J. Transp. Eng. 2013.139:350-357.