You are on page 1of 10

Computer-Aided Civil and Infrastructure Engineering 28 (2013) 594603

Freeway Travel Time Prediction Using


TakagiSugenoKang Fuzzy Neural Network
Yunlong Zhang* & Hancheng Ge
Zachry Department of Civil Engineering, Texas A&M University, TX, USA

Abstract: This article presents a TakagiSugenoKang


Fuzzy Neural Network (TSKFNN) approach to predict
freeway corridor travel time with an online computing algorithm. TSKFNN, a combination of a TakagiSugeno
Kang (TSK) type fuzzy logic system and a neural network, produces strong prediction performance because
of its high accuracy and quick convergence. Real world
data collected from US-290 in Houston, Texas are used
to train and validate the network. The prediction performance of the TSKFNN is investigated with different combinations of traffic count, occupancy, and speed as input options. The comparison between online TSKFNN,
offline TSKFNN, the back propagation neural network
(BPNN) and the time series model (ARIMA) is made to
evaluate the performance of TSKFNN. The results show
that using count, speed, and occupancy together as input produces the best TSKFNN predictions. The online
TSKFNN outperforms other commonly used models and
is a promising tool for reliable travel time prediction on
a freeway corridor.
1 INTRODUCTION
With an increasing need of accurate and reliable realtime traffic information for Intelligent Transportation
Systems (ITS), travel time as an important traffic variable has become increasingly crucial to traffic analysis and operations. For instance, as the input to
the Dynamic Route Guidance System (DRGS), travel
time information can be used to produce the shortest
path between an origin and a destination. Due to the
heterogeneity of different circumstances and nonlinear interactions between drivers and traffic facilities,
the prediction of traffic condition in the near future
is playing a more and more important role in many
ITS applications such as Advanced Traveler Informa*To whom correspondence
yzhang@civil.tamu.edu.

should

be

addressed.


C 2013 Computer-Aided Civil and Infrastructure Engineering.
DOI: 10.1111/mice.12014

E-mail:

tion Systems (ATIS), Advanced Traffic Management


Systems (ATMS), and Emergency Management Systems (EMS). As travel time is one of the most important measures of the traffic system and it is also critical to system users, its prediction has both theoretical
and practical significance (Chen et al., 1999). In this article we present a freeway corridor travel time prediction
model using TakagiSugenoKang Fuzzy Neural Network (TSKFNN) (Takagi and Sugeno, 1985) that not
only has the capabilities of neural networks but also
takes into account the human-like thinking and reasoning of fuzzy logic systems.

2 PREVIOUS STUDIES
There have been many methods used to predict travel
time. Earlier methods include time series models (Oda,
1990; Al-Deek et al., 1998; Anderson, 1994), machine
learning methods (You and Kim, 2000), and regression models (Zhang et al., 2003; Wu et al., 2004; Li
et al., 2008). In recent years, the applications of artificial intelligence, especially the techniques of neural networks, have garnered much attention and have been
considered as effective tools for travel time predictions
(Dharia and Adeli, 2003; Steven et al., 2002; Hoogendoorn et al., 2005; Park and Rilett, 1998; Rilett and Park,
1999; Park and Rilett, 1999; Van Lint, 2006; Krikke,
2002; Chien et al., 2002).
A hybrid model combining the use of the empirical
mode decomposition (EMD) and a multilayer feed forward neural network with back propagation was developed for travel time prediction (Hamad et al., 2009).
The key part of EMD in this study is the Hilbert
Huang transform to address highly nonlinear and nonstationary speed series. The prediction performance of
this proposed method was found to be superior to previous forecasting techniques based on loop data from I-66
in Virginia. A model (Yeon et al., 2008) was conducted

Freeway travel time prediction using TSKFNN

to predict travel time using Discrete Time Markov


Chains (DTMC). The model demonstrated that predicted travel time using DTMC does not differ from
the measured travel time at the 99% confidence level.
It was concluded that DTMC had a good ability to
predict travel time in both uncongested and congested
conditions.
Even though most of the studies on travel time prediction were on freeway segments, travel time characteristics on signalized networks were studied based on
loop and probe data (Bhaskar et al., 2011). Moreover,
traffic performance including travel time was explored
by the fusion of different data sources (Heilmann et al.,
2011). Neural network based models were often used in
travel time predictions on signalized arterials (Liu et al.,
2009; Singh and Abu-Lebdeh, 2007).
Generally neural networks have played an important role in the prediction of travel time due to their
strong ability of nonlinear approximation. However,
neural networks are difficult to design and implement
because the opaqueness of the trained networks cannot be easily understood. Importantly, neural networks
require a long training time and the values of parameters such as initial weights and training rate can significantly affect the performance. Another drawback of
neural networks is that the local minima are very difficult to be avoided because neural networks use the
gradient descent method, a technique that is very sensitive to the initial weights to obtain optimized weights.
Besides these, neural networks also have difficulties approximating human reasoning capabilities to humanlike thinking. A fuzzy system, which can model human
expertise, is potentially an effective tool to predict the
travel time, as it has a systematic calculus to deal with
such information linguistically. Because of this, fuzzy
representation and fuzzy systems are applied in many
recent studies in civil engineering (Hsiao et al., 2012;
Bianchini, 2012; Tagherouit et al., 2011). However, the
inadaptability to deal with changing external environment and the difficulty to model the complex system are
the main drawbacks of fuzzy system. Moreover, the design of fuzzy rules and membership functions is mainly
based on the experts experience without numerical criteria. Therefore, there is a need to create a more efficient tool combining the capabilities of neural networks
and fuzzy systems in many fields. In this case, the fuzzy
neural network (FNN), which can automatically generate a set of expert rules to model the problem and
subsequently use the rules independently, arouses researchers interest. The advantages of the FNN include
adaptability, parallelism, robustness, ruggedness, optimality, and the ability to solve system uncertainty with
the fuzzy set theory. The FNN is also adept at overcoming the local minima and fast converging (Jang and
Roger, 1993).

595

FNN models have been applied to several areas of


transportation engineering in recent years. The application areas included incident detection (Samant and
Adeli, 2001; Karim and Adeli, 2002), work zone capacity estimation (Adeli and Jiang, 2003), and signal
control (Srinivasan et al., 2006). However, there have
not been many FNN applications in freeway travel time
prediction.
Prasad et al. (1999) studied travel time prediction
with the fuzzy logic and neural networks based on the
loop detector data. They used fuzzy logic to convert
detector data to travel time and classified the detector data. The neural network predicted the travel time
based on the clustering of the data. The fuzzy logic and
neural network produced improved estimates of travel
time when compared with results from the linear regression model. However, they did not use a true sense of
FNNs to predict travel time. It is considered to be a regular neural network, and fuzzy logic is used to prepare
inputs to the neural network.
In this study, TSKFNN was developed to predict the
freeway corridor travel time with an online computing
algorithm. TSKFNN simultaneously adjusts parameters
of membership functions in the fuzzy logic part based
upon feedback from neural networks. In all, all parameters of fuzzy logic and neural network will be updated
simultaneously and interactively in the training process.
The relationships between traffic variables and the prediction performance were investigated. Volume, speed,
and occupancy were considered as input options and
ground-truth travel times were used in evaluation.
3 METHODOLOGY
3.1 TSKFNN model
TSKFNN embeds a TSK type fuzzy inference system in
a general structure of a neural network using a neural
network training method to find parameters of the fuzzy
inference system. The TSKFNN is one of the most frequently used FNN schemes developed due to its powerful representation and prediction. Moreover, as the
TSK model has an explicit analytical output, it is possible to incorporate mathematical knowledge into the
real-time control and easily combine it with optimization and adaptive theories.
The general structure of TSKFNN is given in
Figure 1 with multi-input and single output. TSKFNN
consists of two parts: one is antecedent network (AN)
and the other is consequent network (CN). The functions of AN are to cluster the input space and to match
the antecedent of fuzzy rules. Different kinds of specific
travel variable patterns are denoted by clusters. The
function of the CN is to generate the consequence of
fuzzy rules. The input of the model is the sampled traffic

596

Zhang & Ge

A1, A2, . . . , and An denote the antecedent fuzzy


rules expressed by natural language. y is the output of
the system, expressed by a linear function. The number
of An is determined by the number of clusters in the
membership function.
In this study a TSK type fuzzy system is embedded
into the structure of a neural network, which consists
of four layers as shown in Figure 1. The first layer, the
input layer, is used to connect each component of the
input vectors for transmitting the input to the next layer
directly, and the number of the node of the first layer
is the same as the number of the input. In the second
layer, each node represents a value of the linguistic variable, which is used to calculate the membership funcj
tion i (x) to which each component of the input vectors belongs. After testing three types of membership
functions such as Triangular function, Gaussian function, and Trapezoidal function, the following Gaussian
function is adopted as the general membership function in this article due to reported good performance
(Kreinovich et al., 1992).
j
i (x)

Fig. 1. The structure of fuzzy neural network.

variables data including volume, speed, and occupancy


at time k1, and the output of the model is the predicted
travel time of the same corridor at time k. The details
with respect to the structure of Figure 1 are stated in
the following sections.
3.2 Antecedent network
AN, simulating the antecedent fuzzy rules, is the main
part of a fuzzy system. The core of AN is a set of
IF- THEN rules with fuzzy implications, as well as the
membership function, a generalization of the indicator function in classical sets. The membership function A (x) expresses the degree that x belongs to fuzzy
set A. Therefore, the fuzzy set can be defined as A =
{(x, A (x))|x X } where X is a collection of objectives.
The general rule of TSK fuzzy inference is as following:
Input: x = [x1 , x2 , . . . , xn ]T where T is denoted as the
transpose of a matrix
(1)
Output: y

(2)

Rule: if X 1 is A1 AND X 2 is A2 . . . AND X n is An,


n

pi xi
then y = p0 +
1

(3)

=e

(xi cij )2
2
ij

(4)

where i = 1, 2, . . . , n, j = 1, 2, . . . , m i . n is the dimension of the input. m i is the number of the clustering data
xi . The parameters of cij and ij denote the center and
the width of the membership function, respectively. The
input can be directly transformed into linguistic information (fuzzification) by using the Equation (1). In the
third layer, each node represents a fuzzy rule to match
the antecedent of fuzzy rules and calculate the membership grade of the rules j .


(5)
j = min i11 , i22 , . . . , inn
2, . . . , m 2 }, . . . , in
where i 1 {1, 2, . . . , m 1 }, i 2 {1, 
n
m i . The num{1, 2, . . . , mn }, j = 1, 2, . . . m, m = i=1
ber of nodes in the third layer is the same as the number of fuzzy rules. The fourth layer is for normalization,
which has the same number of nodes as that of the third
layer.
j
j = m
(6)

i
i=1

It should be noticed that the initial weights of cij and


ij were determined by Fuzzy Subtractive Clustering
(FSC), which generates rules by enumerating all possible combinations of membership functions of all inputs.
FSC considers each data point as a potential cluster center and calculates the likelihood of being a cluster center based on the density of surrounding data points. The
data point selected as a cluster center has the highest

Freeway travel time prediction using TSKFNN

density in a certain area. Meanwhile, data points around


this potential cluster center point are excluded as a potential center. For the data points {q1 , q2 , . . . , qn }, the
density for each data point qi is defined as:
Di =

n


qi q 2j
(ra /2)2

iI

(7)

j=1

where ra is a positive constant which represents a neighborhood. It is obvious that the data point surrounded
by more data points in the area of the radius ra would
have the higher density value Di . After calculating the
density of each data point, the data point with the highest density Dc1 is selected to be the first cluster center
denoted by qc1 . Then, the density DiNew is recalculated
for all other data points excluding qc1 according to the
revised formula:
DiNew

= Di Dc1 e

(rb /2)

(8)

Thus, the influence of the data points near the first


cluster center will be significantly reduced because these
points are excluded in the further calculation after the
first cluster center is picked up. This process stated
above repeats until an adequate number of cluster centers is generated. In this study, the input data of volume,
speed, and occupancy are initially categorized by FSC.
When completing the subtractive clustering, each of
the cluster centers and radii can generate initial membership functions in the fuzzy inference system. The initial weights of cij and ij were determined using the following equations (Chiu 1994):
cij = xij

ij =

calculate the consequence of each fuzzy rule. The subnet of CN for the cluster c is shown in Figure 1.

wci qi (k 1)
(11)
yc =



q q 2
i c12

rij (qi jmax qi jmin )

597

(9)

(10)

where xij is the center of the cluster j of the input qi ,


rij is the radius for the cluster j of the input qi in the
calculation of subtractive clustering, qi jmax and qi jmin
are the maximum and minimum values in the cluster j
of the input qi , respectively. However, a large number
of inputs would result in a huge number of IF-THEN
rules.
3.3 Consequent network
The CN is used to generate the consequence of fuzzy
rules embedded in the structure of the neural network,
which consists of three layers. The first layer is used to
transmit the input vectors to the next layer. In the second layer, m nodes represent fuzzy rules and are used to

The third layer is used to calculate the output of


TSKFNN, the weighted sum of the consequence of
fuzzy rules. The predicted travel time at time interval
k is:

c yc
(12)
T (k) =
cC

The weight is the output of AN stated previously,


which is the normalized degree of membership.
3.4 Algorithm of learning process
The learning algorithm is developed and adjusted based
on the back-propagation algorithm of the BP neural
network because TSKFNN is essentially a feed-forward
network. Only the connection weight of CN wci , the
center cij and the width ij of the membership function
need to be adjusted. All parameters would be calibrated
at the same time during the process of the optimization.
The output of the network is
 

c yc =
c
wci qi (k 1)
(13)
T (k) =
cC

cC

iI

Minimizing the square of error represented by (11)


determines all parameters.
E=

1
(Tk T (k))2
2

(14)

where T (k) is the actual output and Tk is the observed


value. The derivation of the square of error is given as
following:
E
E T (k) yc
=
wci
T (k) yc wci

(15)

= (T (k) Tk ) c qi (k 1)
Then, the updated parameter wci can be calculated by
wci (k + 1) = wci (k)

E
wci

= wci (k) + (Tk T (k)) c qi (k 1)


(16)
where c is the number of cluster, i is the number of dimension of the input, and is the training rate. After
computing the predicted corridor travel time T (k), the
connection weights of CN should be updated based on
the Equation (13) above.

598

Zhang & Ge

3.5 Online computing algorithm for travel time


prediction
Online adaptive freeway corridor travel time prediction
demands a faster convergent speed. The online computing process consists of four steps: data input, online training, input data updating, and corridor travel
time prediction. The first step inputs data in previous
n intervals, including sampled traffic variables denoted
by qi (t), t = k n 1, . . . , k 2 and the corridor travel
times denoted by Tt , t = k n, . . . , k 1. The second
step trains the coefficients of TSKFNN including the
connection weights and parameters of the membership
function after inputting the data into the TSKFNN. In
the third step, the data set is updated by collecting the
new traffic and travel time data at time k, and then shifting the input data by one interval. The traffic variables
and corridor travel time data are then passed back to the
first step to update the coefficients of the system. The final step calculates the predicted value of corridor travel
time using the updated input data of the traffic variables. This online computing process is shown in Figure
2. Contrarily, without the iterative updating process, the
training process is considered offline. In offline training,
all training data items are used whereas online training
only uses the data from the most recent n intervals.

4 DATA DESCRIPTION
The actual data were collected on a freeway segment
located on US-290, in an urban area northwest of Houston, Texas shown in Figure 3. This corridor is on one
of the busiest commuting routes that connect the downtown commercial districts and suburb an residential areas. The study corridor, the section from node 31 to
node 32, is a 2.9-mile section. There are five Automated Vehicle Identification (AVI) stations denoted as
3997, 3890, 4026, 4010, and 4003 in this section. Speed,
volume, and occupancy data were collected with microwave detectors, and travel time data were from AVI
stations, all obtained from the Houston TranStar transportation management center. TranStar operates 24/7
and has been archiving AVI travel time and speed data
since October 1993. The data collected by recording vehicle toll tag IDs and the corresponding time stamps
when each time vehicles are passing the AVI stations
are used to determine the travel time for each vehicle
traveling on AVI segment (3231), as shown in Figure 3.
The data reduction and fusion is a necessary step in
selecting a study period, aggregating data, filtering out
false data entries, and interpolating missing data. Eventually, data from three Fridays, February 1st, 8th, and
29th in 2008, were reduced, and chosen as the base-

Fig. 2. The online computing process.

Fig. 3. The study corridor on US-290.

line data set to predict and validate TSKFNN. Fridays


were chosen because the travel time variations are typically the largest, making it most difficult for prediction.
The data of volume, speed, and occupancy were aggregated into 5-minute intervals. Within each 5 minutes,

Freeway travel time prediction using TSKFNN

the ground-truth travel time was determined by averaging all collected travel times for all vehicles passing the
AVI stations between 31 and 32. The data on February 1st and 8th were used to train TSKFNN, whereas
the data on February 29th were used to validate and
test TSKFNN. The ranges of all variables are from 49
to 445 vehicles for 5-minute count, from 14 to 72 mph
for speed, from 0.5 to 22.7% for occupancy, and from
133 to 839 seconds for travel time.

Table 1
Correlation analysis between traffic variables

Volume
Speed
Occupancy

k=1

where N is the number of time intervals, Tk is the observed travel time for interval k, and T (k) is the predicted travel time for interval k.

6 PRELIMINARY DATA ANALYSIS


The correlation analysis, as a preliminary analysis, was
used to identify the relationship among the collected
traffic data, and examine the reliability of the data to
see whether the variables have expected relationships.
In particular, the Pearson correlation shown in Equation (18) was adopted to analyze the relationship between traffic variables. The purpose of this analysis is to
choose proper traffic variables as the input of TSKFNN.
yi
xi 

yi

=


yi 2
xi
xi  

y

2
2 n
n

rx y

xi


(18)

i2

Here, r x y is the correlation coefficient, xi , yi are collected data points of two traffic variables, and n is the
number of observations. The result of correlation analysis is summarized in Table 1.
Based on Table 1, the correlation value between occupancy and volume is 0.73, the value between occupancy and speed is 0.95, and the value between
speed and volume is 0.54. Speed and occupancy had
a higher correlation with travel time than volume, naturally speed and occupancy should be considered as the
input. Volume as a variable could be left out of the fur-

Volume

Speed

Occupancy

Travel time

1
0.54
0.73

0.54
1
0.95

0.73
0.95
1

0.44
0.93
0.91

Table 2
RMSE (seconds) for training and testing TSKFNN with
different inputs

5 PERFORMANCE MEASURE
The root mean square error (RMSE) between the predicted and observed values of the travel time is used
to assess the performance of the TSKFNN model. The
RMSE is calculated according to Equation (17):


N
1 
(Tk T (k))2
(17)
RMSE =

599

Training
Testing

Input 1

Input 2

Input 3

16.2
32.4

15.4
28.4

14.6
25.1

ther analysis because of the low correlation with travel


time, however, because it is correlated with occupancy,
it can still be considered as one input item for travel time
prediction.

7 RESULTS AND ANALYSIS


Three different input options (Input 1: Volume and Occupancy, Input 2: Speed and Occupancy, and Input 3:
Volume, Occupancy, and Speed) were used to evaluate
the performance of TSKFNN with the online computing algorithm to predict corridor travel time. AVI travel
time data are used as ground truth for training and
testing.
It was found that the TSKFNN fit the training data
very well with a low RMSE as shown in Table 2, meaning that the TSKFNN can be developed to model the
training data very well. Moreover, the training error decreased very fast from the beginning of training then
stayed stable, indicating a very fast convergence with a
short training time. Input option 3 produced the lowest training RMSE when compared to the RMSEs of
the other options, indicating that TSKFNN can perform better in the training process with all three variables as input items. After the TSKFNN was trained,
the data set on February 29th was used for testing to
evaluate TSKFNNs ability to predict the corridor travel
time. The RMSE results for testing are also provided in
Table 2. The overall RMSEs are lower than 33 seconds,
indicating that the TSKFNN has a strong ability to accurately predict the tendency of the testing data.
The prediction results with testing data for all three
input options are demonstrated in Figure 4. As stated
previously, with the aggregation interval of 5 minutes,
the whole day (1,440 minutes from 00:00 to 24:00)
has 288 time intervals. Figure 4 shows that TSKFNN

600

Zhang & Ge

Fig. 4. Prediction results for different input options.

performed very well in prediction, as predicted


values match well with the observed values for the three
different input options. Although the three plots do not
have significant differences, TSKFNN did not perform
very well from time interval 170 to 180 and from 210 to
220 for Input 1, and also performed poorly in Input 2
from 180 to 210. On the contrary, the prediction with
Input 3 has better performance than with the other two
input options. This indicates that volume provided extra useful information for better TSKFFN predictions,
even though the correlation between volume and travel
time was not very high. This conclusion was supported
by results in both Table 2 and Figure 4.
We further extract and present the results by flow
condition in four categories: off-peak (Time 1 to 150 and
251 to 288), congestion buildup (Time 150 to 200), recovery (Time 201 to 220), and peak (Time 221 to 250).
RMSEs for different corresponding periods are summarized in the following table. Considering the travel times
in non-off peak periods have much higher values (illustrated in Figure 4), though elevated, these RMSE values
are all satisfactory. In particular, with Input scenario 3,
the RMSE for the peak is 40.1 seconds corresponding to
an observed travel time of about 850 seconds.
Based on Tables 2 and 3, the overall RMSEs indicate
that the TSKFNN has a strong ability to accurately predict the tendency of the testing data. The convergence
of the TSKFNN was also very fast. Moreover, it can be
found that Input 3 has the lowest RMSE in each sce-

Table 3
RMSE (seconds) of different periods for different input
patterns

Input 1
Input 2
Input 3

Off-peak

Buildup

Peak

Recovery

14.31
15.74
13.82

45.89
58.38
39.17

70.87
52.36
40.13

44.46
36.55
35.06

nario, proving that TSKFNN can predict the most accurate results with this combination of input variables.
We also evaluated the performance of TSKFNN using only the time series of travel time as input. The
travel time T(k-2), T(k-1), and T(k) of the corridor were
used as input of the TSKFNN to predict the travel time
T(k+1). The data set was the same Friday baseline data
set. The data of February 1st and 8th were again used
for training and the data of February 29th were used
for testing. After training the whole network the training RMSE was 15.3 seconds, which was low compared
to previous values in Table 2. Figure 5 also illustrated
that the performance of TSKFNN with the time series
of travel time as the input seemed to look even better
than some of the predictions in Figure 4. The RMSE of
27.76 seconds is better than those from prediction with
Input options 1 and 2, 28.4 and 32.4 seconds respectively, even though it is slightly worse than the RSME
of 25.1 seconds from Input 3. Overall, the TSKFNN

Freeway travel time prediction using TSKFNN

601

Fig. 5. Prediction results for inputs of past travel time.

prediction from the time series of past travel times can


produce good performance, however, the travel time
history we have from the AVI data is not commonly
available while speed, volume, and occupancy data are
commonly available from loop detectors that are widely
deployed.
Model performance was also investigated on an expanded data set. This expanded data set was collected
on weekdays from February 1st to 22th, 2008. The weekdays in the first two weeks are used for training, and the
weekdays in the third week are employed for validating the model. Count, speed, and occupancy are used as
inputs, and the results are summarized below.
RMSEs from Monday to Friday are 27.5, 28.7, 38.8,
33.5, and 24.1, respectively. The RMSE for Wednesday
prediction is somewhat higher than that for other days,
but the travel times during the peak on that day were
also significantly larger. An investigation of a separate
incident database revealed that there were several severe incidents during the peak period on that day. Overall, the prediction performance was very similar to that
of the single-day prediction reported previously from
the baseline data set. It should be noted the baseline
Friday data set was selected also because it was affected
by incidents to a lesser degree.

8 COMPARISON EXPERIMENTS
The comparison of performance in corridor travel
time prediction was made among online TSKFNN, off-

line TSKFNN, the back propagation neural network


(BPNN) and the time series model (ARIMA) using the
Friday baseline data set. Rilett et al. (1999), Van Lint
et al. (2005), and Guin (2006) indicated that Neural Network (NN) and ARIMA have good performance in the
prediction of travel time, we wanted to compare the performance of the TSKFNN with the results from these
two commonly used approaches. Online TSKFNN and
offline TSKFNN were compared to illustrate the advantage of the online computing algorithm. Speed, volume,
and occupancy are used as input to the TSKFNN.
The BPNN consisted of three layers with three input neurons and one output neuron. The sigmoid activation function was used for each neuron and the back
propagation training procedure was adopted. The performance of BPNN in the prediction is significantly affected by the number of neurons in the hidden layer and
the training rate. To minimize the RMSE between the
predicted and observed results, 10 neurons in the hidden layer and a training rate of 0.25 were found to produce the best results by the BPNN. Auto Regressive Integrated Moving Average (ARIMA) model is generally
referred to as ARIMA (p, d, q) where p, d and q are
the order of the autoregressive, integrated, and moving
average parts of the model respectively. After the optimization on parameters based on minimum resulting
AIC value, ARIMA (3, 1, 5) was adopted to predict the
corridor travel time.
The numerical result on RMSEs in Figure 6 confirmed that the two TSKFNN predictions have clearly
outperformed BPNN and ARIMA (reduced RMSE

602

Zhang & Ge

Fig. 6. RMSEs of the comparison experiment.

by 30%). It is also obvious that online TSKFNN outperformed offline TSKFNN. The comparison strongly
indicated that the TSKFNN performed better than commonly used BPNN and ARIMA and online computing algorithm enhanced the prediction accuracy for
TSKFNN.

The performance of the online TSKFNN can be further improved by changing the function of membership
or by increasing the number of clusters used. Testing
and validation of the online TSKFNN with other data
sets should also be conducted in the future.

REFERENCES
9 CONCLUSIONS
Based on the theory of intelligent control system, a
TSKFNN model was developed to predict the corridor
travel time on a freeway with an online computing algorithm that enhances the predictive ability. With the
structure of multi-input and single output, the TSKFNN
consists of two parts: AN and CN. The functions of
AN are to cluster the input space and to match the
antecedent of fuzzy rules. The function of CN is to
generate the consequence of fuzzy rules. Based on the
result of prediction based on the data collected from
US-290 it was found that online TSKFNN can accurately predict future travel time in the corridor. Moreover, it was demonstrated that the online TSKFNN performed better than a regular neural network (BPNN)
and the ARIMA model in the prediction performance
as online TSKFNN is able to adaptively adjust its coefficients with the recent training data prior to the prediction. TSKFNN with an online computing algorithm
performed best with volume, speed, and occupancy as
the input. This prediction performance is as good as
the TSKFNN prediction performance with past travel
time history as the input when the travel time history is
available.

Adeli, H. & Jiang, X. (2003), Neuro-fuzzy logic model for freeway work zone capacity estimation, Journal of Transportation Engineering, 129(5), 48493.
Al-Deek, H., DAngelo, M. &. Wang, M. (1998), Travel time
prediction with non-linear time series, Proceedings of the
ASCE 1998 5th International Conference on Applications of
Advanced Technologies in Transportation, Newport Beach,
CA, pp. 31724.
Anderson, J., Bell, M., Sayers, T., Busch, F. & Heymann,
G. (1994), The short-term prediction of link travel time
in signal controlled road networks, Proceedings of the
IFAC/IFORS 7th Symposium on Transportation Systems:
Theory and Application of Advanced Technology, Tianjin,
China, pp. 6216.
Bhaskar, A., Chung, E. & Dumont, A. G. (2011), Fusing loop
detector and probe vehicle data to estimate travel time
statistics on signalized urban networks, Computer-Aided
Civil and Infrastructure Engineering, 26(6), 43350.
Bianchini, A. (2012), Fuzzy representation of pavement condition for efficient pavement management, Computer-Aided
Civil and Infrastructure Engineering, 27(8), 60819.
Chen, P. S. T., Srinivasan, K. K., Mahmassani, H. S. (1999), Effect of information quality on compliance behavior of commuters under real-time traffic information, in Proceedings
of the 77th Transportation Research Board Annual Meeting,
National Academies Press, Washington DC, USA.
Chien, S., Ding, Y. & Wei, C. (2002), Dynamic bus arrival time
prediction with artificial neural network, Journal of Transportation Engineering, 128(5), 42938.

Freeway travel time prediction using TSKFNN

Chiu, S. (1994), Fuzzy model identification based on cluster estimation, Journal of Intelligent & Fuzzy Systems, 2(3), 267
78.
Dharia, A. & Adeli, H. (2003), Neural network model for
rapid forecasting of freeway link travel time, Engineering
Applications of Artificial Intelligence, 16(78), 60713.
Guin, A. (2006), Travel time prediction using a seasonal autoregressive integrated moving average time series model,
in Proceedings of the 2006 IEEE Intelligent Transportation
Systems Conference, Toronto, Canada, September 1720.
Hamad, K., Lee, E., Shourijeh, M. T. & Faghri, A. (2009),
Near-term travel time prediction utilizing Hilbert-Huang
transform, Computer-Aided Civil and Infrastructure Engineering, 24(8), 55176.
Heilmann, B., El Faouzi, N. E., de Mouzon, O., Hainitz, N.,
Koller, H., Bauder, D. & Antoniou, C. (2011), Predicting
motorway traffic performance by data fusion of local sensor data and electronic toll collection data, Computer-Aided
Civil and Infrastructure Engineering, 26(6), 45163.
Hsiao, F. Y., Wang, S. S., Wang, W. C., Wen, C. P. & Yu,
W. D. (2012), Neuro-fuzzy cost estimation model enhanced
by fast messy genetic algorithms for semiconductor hookup
construction, Computer-Aided Civil and Infrastructure Engineering, 27(10), 76481.
Jang, J-S. R. (1993), ANFIS: adaptive-network-based fuzzy inference system, IEEE Transactions on Systems, Man, and
Cybernetics, 23(18), 66585.
Karim, A. & Adeli, H. (2002), Comparison of the fuzzy
wavelet RBFNN freeway incident detection model with the
California Algorithm, Journal of Transportation Engineering, 128(1), 2130.
Kreinovich, V., Quintana, C. & Reznik, L. (1992), Gaussian membership functions are most adequate in representing uncertainty in measurements, in Proceedings of
NAFIPS92: North American Fuzzy Information Processing Society Conference, Puerto Vallarta, Mexico, Vol. II, pp.
61824.
Krikke, R. (2002), Short-range travel time prediction using
an artificial neural network, in Proceedings of 9th World
Congress on Intelligent Transport Systems, Chicago, Illinois,
USA.
Li, Z., Yu, H., Liu, Y. & Liu, F. (2008), An improved adaptive
exponential smoothing model for short-term travel time
forecasting of urban arterial street, Acta Automatica Sinica,
34(11), 14049.
Liu, H., Zhang, K., He, R. & Li, J. (2009), A neural network model for travel time prediction, in Proceedings of the
IEEE International Conference Intelligent Computing and
Intelligent Systems, Shanghai, China.
Oda, T. (1990), An algorithm for prediction of travel time using vehicle sensor data, in Proceedings of the IEE 3rd International Conference on Road Traffic Control, London, pp.
404.

603

Park, D. & Rilett, R. (1998), Forecasting multiple-period freeway link travel time using modular neural network, Transportation Research Record, No. 980743, TRB, National
Research Council, Washington DC.
Park, D. & Laurence, R. (1999), Forecasting freeway link
travel times with a multilayer feedforward neural network.
Computer-Aided Civil and Infrastructure Engineering, 14,
35767.
Prasad, V. P. & Nelson, P. C. (1999), Application of fuzzy logic
and neural networks for dynamic travel time estimation, International Transactions in Operational Research, 6, 14560.
Rilett, L. & Park, D. (1999), Direct forecasting of freeway corridor travel times using spectral basis neural networks, Presented at the 78th TRB Annual Meeting (CD-ROM), Washington DC.
Samant, A. & Adeli, H. (2001), Enhancing neural network
incident detection algorithms using wavelets, ComputerAided Civil and Infrastructure Engineering, 16(4), 23945.
Singh, A. & Abu-Lebdeh, G. (2007), Arterial network travel
time estimation using conditional independence graphs and
state space neural networks, in Proceedings of the 18th
IASTED Conference on Modelling and Simulation, Montreal, Quebec, Canada.
Srinivasan, D., Choy, M. & Cheu, R. (2006), Neural networks
for real-time traffic signal control, IEEE Transactions on
Intelligent Transportation Systems, 7(3), 26171.
Tagherouit, W. B., Bengassem, J. & Bennis, S. (2011), A
fuzzy expert system for prioritizing rehabilitation sewer networks, Computer-Aided Civil and Infrastructure Engineering, 26(2), 14652.
Takagi, T. & Sugeno, M. (1985), Fuzzy identification of systems and its applications to modeling and control. IEEE
Trans actions on Systems, Man, and Cybernetics, 15(1),
11632.
Van Lint, J. W. C. (2006), Reliable real-time framework for
short-term freeway travel time prediction. Journal of Transportation Engineering, 132(12), 92132.
Van Lint, J. W. C., Hoogendoorn, S. P. & Van Zuyle, H. J.
(2005), Accurate freeway travel time prediction with statespace neural networks under missing data, Transportation
Research Part C, 13, 34769.
Wu, C. H., Ho, J. M. & Lee, D. T. (2004), Travel time prediction with support vector regression. IEEE Transactions on
Intelligent Transportation System, 5(4), 27681.
Yeon, J., Elefteriadou, L. & Lawphongpanich, S. (2008),
Travel time estimation on a freeway using discrete
time Markov chain, Transportation Research Part B 42,
32538.
You, J. & Kim, J. (2000), Development and evaluation of a hybrid travel time forecasting model. Transportation Research
Part C 8, 23156.
Zhang, X. & Rice, J. (2003), Short-term travel time prediction.
Transportation Research Part C 11(34), 187210.

You might also like