Professional Documents
Culture Documents
Elevator Traffic Flow Prediction Using Artificial
Elevator Traffic Flow Prediction Using Artificial
Elevator Traffic Flow Prediction Using Artificial
net/publication/49910579
Article
Source: OAI
CITATION READS
1 276
1 author:
SEE PROFILE
All content following this page was uploaded by Choo Yong Lee on 08 March 2017.
APRIL 2008
iii
ACKNOWLEDGEMENT
ABSTRACT
Elevator traffic flow prediction is essential part of the modern elevator group
control system to enable controller apply the best dispatching strategy based on
predicted traffic flow data to achieve optimum operation with the aim to reduce
average waiting time of passenger for arrival of elevator to serve them. Generally,
elevator traffic flow has high complexity and passenger flow possesses nonlinear
feature which is difficult to be expressed by a certain functional style. In this thesis,
artificial intelligent technique radial basis function neural network (RBF NN) is used
to develop elevator traffic flow prediction model. RBF NN is selected because it is
suitable to model nonlinear system and can be trained using fast 2 stages training
algorithm assures fast convergence. The past interval traffic flow data and traffic
flow data at same time on previous days are used to train RBF NN so that it could
predict traffic flow ahead. Neural network toolbox that incorporates newrbe and
newrb functions in matlab software is employed to develop algorithm and program
of RBF NN. Optimum spread constant that will yield minimum mean square error is
obtained and become input to the RBF NN. Ten cases with different k and p are
studied to evaluate performance of RBF NN. Given training data collected from
field, RBF NN is able to predict elevator up peak traffic flow occur at 8:15 a.m. (in 5
minutes interval) which is short term traffic fairly accurate. Mean square errors from
simulation results are small and some of them could be neglected. The maximum
mean square error is 2.82 for case that use past 3 interval data on 4th day and past 3
days (1st,2nd and 3rd day) data to predict traffic flow on 5th day executed by using
newrb function. It is concluded that RBF NN is an effective artificial intelligent
technique to build elevator traffic flow prediction model.
v
ABSTRAK
Ramalan aliran traffik lif adalah bahagian penting dalam sistem kawalan lif
kumpulan moden supaya pengawal melaksanakan strategi pengangkutan terbaik
berasaskan data ramalan aliran traffik untuk mencapai operasi optimum dengan
tujuan mengurangkan masa tunggu min penumpang. Amnya, aliran traffik lif
mempunyai kekompleksan yang tinngi dan aliran penumpang memiliki cirri taklinear
yang sukar dinyata oleh corak fungsi tertentu. Dalam tesis ini, teknik kepintaran
buatan radial basis function neural network (RBF NN) digunakan untuk
membangunkan model ramalan aliran traffik lif. RBF NN dipilih sebab ia sesuai
untuk membentuk sistem taklinear dan boleh dilatih dengan algoritma perlatihan 2
peringkat cepat yang menjamin penumpuan cepat. Data aliran traffik selang masa
lepas dan data aliran traffik pada masa yang sama pada hari-hari lepas digunakan
untuk melatihkan RBF NN supaya ia dapat meramalkan aliran traffik akan dating.
Neural network toolbox yang mengandungi fungsi newrbe dan newrb dalam perisian
matlab digunakan untuk membangunkan algoritma dan aturcara RBF NN. Pemalar
menyebar optimum yang menghasilkan ralat kuasa dua min ( minimum diperolehi
dan dijadikan input kepada RBF NN. Sepuluh kes telah dikaji untuk menilai
percapaian RBF NN. Dengan membekalkan data latihan yang dikumpul dari tempat
kajian, RBF NN mampu meramalkan aliran traffic puncak atas yang berlaku pada
8:15 a.m. (dalam selang masa 5 minit) iaitu traffik jangka masa pendek dengan agak
tepat. Ralat kuasa dua min dari keputusan-keputusan simulasi adalah kecil dan
sebahagiannya boleh diabaikan. Ralat kuasa dua min maksimum ialah 2.82 untuk
kes yang menggunakan 3 data lepas pada hari keempat and data 3 hari lepas (hari
pertama, kedua dan ketiga) untuk meramalkan aliran traffk pada hari kelima
dilaksanakan dengan fungsi newrb. Ianya dirumuskan bahawa RBF NN ialah teknik
kepintaran buatan yang berkesan untuk membina model ramalan aliran traffik lif.
vi
TABLE OF CONTENTS
1 INTRODUCTION 1
1.1 Introduction 1
1.2 Objective of Project 3
1.3 Scopes of Work and Research Methodology 4
1.4 Thesis Outline 5
3 LITERATURE REVIEW 23
3.1 Radial Basis Function Neural Network (RBF NN) 23
3.2 Elevator Traffic Flow Prediction 26
3.3 Summary 38
vii
4 METHODLOGY 39
4.1 Introduction 39
4.2 Problem Statement 39
4.3 Elevator Traffic Flow Data Collection at Field 41
4.4 RBF NN Elevator Traffic Prediction Model Design 47
4.5 Software Implementation and Simulation 50
4.5.1 Simulation to Obtain Optimum Spread 52
Constant
4.5.2 Simulation of Network Performance and Goal 84
for Newrb Function
REFERENCES 119
APPENDIX 124
viii
LIST OF TABLES
5.3 Comparison of RBF NN output for newrb and newrbe and 105
real data (Case 3)
5.4 Comparison of RBF NN output for newrb and newrbe and 107
real data (Case 4)
5.5 Comparison of RBF NN output for newrb and newrbe and 108
real data (Case 5)
5.6 Comparison of RBF NN output for newrb and newrbe and 109
real data (Case 6)
5.7 Comparison of RBF NN output for newrb and newrbe and 110
real data (Case 7)
5.8 Comparison of RBF NN output for newrb and newrbe and 112
real data (Case 8)
5.9 Comparison of RBF NN output for newrb and newrbe and 113
real data (Case 9)
5.10 Comparison of RBF NN output for newrb and newrbe and 115
real data (Case 10)
x
LIST OF FIGURES
INTRODUCTION
1.1 Introduction
Local knowledge
module Non-local
Coordination
mechanism Schedule
elevator
The elevator traffic flow fluctuates significantly in morning, lunch hour and
evening. Generally, elevator traffic flow possesses non linear feature which is very
difficult to be expressed by a certain functional style. Various methods have been
employed as a fundamental to design elevator traffic flow prediction model. These
methods include Morlet wavelet function, support vector machine, least squares
support vector machine and wavelet support vector machine. Most of the articles
related to elevator technology are published by researchers from China due to strong
growth of elevator market as stated in 2006 annual report of Kone Corporation.
Nowadays, there are only limited research and development works related to
elevator technology being carried out in Malaysia. Tan Kok Khiang (Tan, 1997) and
Kumeresan (Kumeresan and Khalid, 2005) from Universiti Teknologi Malaysia have
developed elevator group control system based on ordinal structure fuzzy reasoning
method which performs better than conventional elevator group control system with
hall call assignment approach and it is believed that the system could be further
enhanced if traffic flow prediction module is incorporated to form a complete
intelligent elevator group control system.
Scopes of work for this project are to study elevator traffic flow and develop
RBF NN based algorithm by using matlab software to predict up peak traffic flow of
elevator. A case study is carried out to evaluate performance of RBF NN elevator
traffic flow prediction model. Traffic flow data is collected in one commercial
building as training data and testing data. Up peak traffic flow pattern at morning
will be studied and analyzed. Data is collected from 7:30 a.m. – 10:00 a.m (150
minutes) because it covers working hour starts at 8:00 a.m. to 9: 30 a.m. As
sufficient data is needed to train the RBF NN, data is collected for five working days
continuously. The overall scopes of work in sequence are listed as followings:
• To study fundamental of elevator traffic analysis.
5
• To acquire knowledge in radial basis function neural network (RBF NN) and
further understand its application especially in traffic flow prediction.
• To review past researches related to elevator traffic flow prediction.
• To acquire skill of using matlab software.
• To develop RBF NN based algorithm by using matlab software to predict up
peak traffic flow of elevator in office building.
• To collect data at field. The number of passengers that will use elevator in
five minutes interval shall be recorded as it represents traffic flow.
• To train the developed algorithm by using collected data.
• To simulate the developed algorithm by using collected data and verify the
simulated results by comparing with experimental result.
CHAPTER 2
The elevator traffic flows could be classified into three types namely
incoming, outgoing and inter-floor traffic as presented by G.C. Barney (1985) in the
book Elevator Traffic Analysis Design and Control which is the most important
reference for elevator traffic analysis. Elevator traffic pattern represents demand of
passenger to travel to particular floor at specific time. For example, majority of
passengers travel to upper floors cause up peak traffic to occur. The typical elevator
traffic pattern for a weekday is illustrated in Figure 2.1 while Figure 2.2 shows the
passengers arrival obtained from NESTE building. The differences of incoming,
outgoing and inter-floor traffic in building are illustrated in Figure 2.3.
8
Koyama, 2000). In some circumstances, up peak traffic could occur twice due to
working hour for offices in building starts at different time as shown in Figure 2.5.
For example, working hour for company A and company B in same building start at
8:00 a.m. and 8:30 a.m. respectively, therefore it is anticipated that up peak will
occur at 7:55 a.m. and 8:25 a.m.
\
Figure 2.5 Up peak traffic profile occurs twice.
11
Inter-floor traffic occurs when passengers travel randomly within floors. This
traffic is caused by the normal circulation of people around a building during the
course of their business. Sometimes this traffic is called balanced two way traffic as
it involves both up and down trips. It is balanced traffic because passengers usually
return to their original floor after finishing the business.
12
In 1995, fuzzy neural based elevator group control system was proposed by
Toshiba Corporation, Elevator and Escalator Division (Imasaki, N. et al., 1995). A
traffic forecasting module was incorporated into the EJ-1000FN control system so
14
that it could adapt itself quickly to changes of traffic in building. Fuzzy neural
network will learn to recognize traffic flow pattern in short term (hour order) and
store in memory to establish long term re-learning adaption. The block diagram of
performance tuning EJ-1000FN control system is illustrated in Figure 2.7. This
system is more intelligent than pure fuzzy logic elevator control system as it could
distribute elevators to particular floor with heavier traffic based on forecast.
Elevator
group
parameter
Traffic condition
1.0
Figure 2.9 Membership function for incoming, outgoing and inter-floor components.
1.0
0.0
0.0 50.0 100.0
Traffic Intensity
(% of Handling Capacity)
Figure 2.10 Membership function for incoming traffic intensity.
17
In 1998, fuzzy elevator group control system has been proposed by a team of
researchers from Korea (Kim et al., 1998). Further to previous research (Kim et al.,
1996), fuzzy logic is used as the core of the control system and traffic data
management is incorporated in the system. The structure of fuzzy elevator group
control system is illustrated in Figure 2.11.
19
Data management
Elevator management
Terminal management
Terminal
In the recent research to propose modern elevator group control system (Chen
et al., 2006), traffic flow prediction is again playing important role in the system.
Traffic prediction is realized by using support vector machine (SVM). The block
diagram of the proposed system is illustrated in Figure 2.12. The system consists of
20
five modules namely traffic data base, traffic prediction, traffic pattern recognition,
strategy generation and hall call assignment.
Traffic database
Elevator group
control system
Traffic prediction Traffic pattern Strategy generation
recognition
They proposed traffic data to be categorized into three parts namely number
of passengers in the entire building, number of passengers entering the main terminal
and number of passengers leaving the main terminal. The block diagram of traffic
predictor is shown in Figure 2.13. X[0], X[t], X[2t], …, X[(n-1)t] are traffic data
sampled every five minutes, t is the sampling interval.
21
X[0]
X[t]
X[(n-1)t]
Another passenger based elevator group control system with traffic database
module (traffic flow prediction is used to predict future traffic pattern) also showed
satisfactory result comparing with group control system without traffic database
module (Chiang and Fu, 2002). The simulation result of the research is illustrated in
Table 2.5.
22
Legend:
AWT represents average awaiting time.
ART represents average riding time.
2.3 Summary
From the researches and development works as discussed in chapter 1.1 and
chapter 2.2, it is proven that elevator group control systems with artificial intelligent
perform better than conventional system. It is also proven that the role of traffic flow
prediction module is important in the elevator group control system as it could
enhance the performance. Therefore, traffic flow prediction model shall be
developed by using suitable artificial intelligent technique to yield accurate values so
that the group control system could assign the best elevator dispatching strategy.
23
CHAPTER 3
LITERATURE REVIEW
The radial basis function neural network RBF NN is one of the architecture of
ANN (Abdullah, 2005). RBF NN offers one major advantage compared to multi-
layer perceptrons in that they can be trained using fast 2 stages training algorithm
without the need for time consuming non-linear optimization techniques.
Furthermore, RBF possess the property of best approximation. RBF NN may require
more neurons than standard feed-forward back propagation networks but it performs
24
better than back propagation networks when many training vectors are available.
The architecture of RBF NN is illustrated in Figure 3.1.
RBF network consists of three layers namely input layer, hidden layer and
output layer. The number of hidden is to be defined by network designer. If the
number of output Q = 1, the output of RBF NN is expressed as η(x, w) = Σ
w1kΦ( x – ck2). x is an input vector, x∈RRx1, Φ is basis function, •2
denotes Euclidean norm, w1k is weight in output layer, SI is the number of neurons in
the hidden layer
ayer and ck is RBF center in the input vector space. The equation could
be also expressed as
η(x, w) = ΦT(x)w
(3.1)
[ 1( x – c1) Φ2(x – c2) … ΦS1(x – cS1)]
where ΦT(x) = [Φ
wT = [w11 w12 … w1S1]
(3.2)
25
The basis function Φk(•) is the mathematical function. For example, Φk(x) =
x, Φk(x) = x3 or Φk(x) = exp (-x3/ζ2) which is Gaussian basis function. In order to
achieve good approximation results, large number of input vectors shall be provided
as the centers to ensure as adequate input space sampling. In the training phase, the
correct value of second layer weight need to be determined so that the weight sum of
basis function can approximate the given training output data. The training
procedure is as below:
1. Set the centers ck using all of the input values or clustering algorithm.
2. Set the required parameter using rule of thumb or algorithm.
3. Write equation of RBF output in matrix form.
η(x1, w
) φ1 (x1 , c1) • • • φS1 (x1 , cS1) w11
• • • • •
• = • • • •
• • • • •
( )
φ1 (x N , c1) • • • φS1 (x N , cS1) w1S1
η ,
xN w (3.3)
4. Define quadratic error e(w) = (y - Φw)T(y - Φw)
where y = [y1 … yN]
5. Solve for optimum set of second layer weight ŵ using least squares formula.
Clustering techniques are used to find a set of centers that more accurately
reflects the distribution of the data points. There are various clustering algorithms
like K-means algorithm, input clustering (IC) algorithm and dynamic clustering
algorithm (Qu et al., 2006). K-means clustering is the simple and common algorithm
used in RBF NN. The K-means algorithm is as below:
1. Assign the input data to random K sets.
2. Compute the mean of each set.
3. Reassign each point to a new set according to which is the nearest mean
vector.
4. Re-compute the mean of each set.
5. Repeat steps 2 and 4 until there is no further change in the grouping of data
points.
6. The mean of the sets will be the RBF centers.
26
RBF NN has been used as a basis to develop a method for approximation and
estimation of nonlinear stochastic dynamic systems (Elanayar and Shin, 1994). More
than one data set is used in order to ensure adequate learning by the neural network
over the domain of interest. A modified recursive least squares training algorithm is
employed to obtain the weights of neural networks. Once, the network will be able
to estimate the states of the unknown nonlinear system. RBF NN has been applied in
non-periodic prediction system. For instance, RBF NN is used to predict field
strength of mobile radio propagation based on collected and stored topographical and
morphographical data (Chang and Yang, 1997).
On the other hand, RBFF NN is also used to predict traffic flow which a
complex system (Qin et al., 2006) as RBF NN with local generalization abilities and
fast convergence speed could overcome shortcomings like slow convergence and
local minimum of back propagation neural network. A research has been carried out
to predict freeway traffic flow by using distributed RBF NN based on fuzzy c-means
clustering algorithm and simulation results show that it could capture the traffic flow
dynamic quite closely (Wang and Xiao, 2003). RBF NN is also being used to
develop a model to predict bus arrival time and it shows better results comparing
with historical data based model and regression model (Jeong and Rilett, 2004). As
conclusion, RBF NN is suitable to be used to develop a prediction model for
nonlinear system due to its advantages and examples as discussed above.
Combination of artificial intelligent techniques and forecasting model
model (Zong, 2000). Firstly, three layer feed forward neural networks with sigmoid
activation function could approximate whatever function. Secondly, neural network
has strong learning capability and adaptability. Thirdly, neural network could
overcome noise problem. In many researches, combination of artificial intelligent
techniques and forecasting model is best suited for short term traffic flow (for
example, elevator traffic flow) prediction. Neural network with advanced back
propagation algorithm has been built to predict elevator traffic flow (Wang and Dong,
2004). The predicted results of neural network ( ), exponential smoothing
( ) and real data ( ) are illustrated in Figure 3.2, 3.3, 3.4, 3.5 and 3.6.
Figure 3.5 The maximum proportion of inter-floor traffic flow on a working day.
The Morlet wavelet function has been used to develop wavelet neural
network (WNN) to predict elevator traffic flow (Huang et al, 2003). From the
wavelet theory, arbitrary function f(x) = L2(R2) can be approximated closely as
30
∑
(3.4)
where n is the number of network nodes, h(•) is basis wavelet, bt is translation factor,
at is dilation factor and wt is link weigh of the network. The architecture of WNN is
shown in Figure 3.7. WNN consists of three layers namely input layer, hidden layer
and output layer. The WNN could overcome the lag of statistical traffic data and has
properties like simple network architecture, fast convergence and higher precision.
The simulation results (consists of real traffic data, predicted data by WNN and other
method, Exponential Smoothing (ES)) of the research are shown in Figure 3.8, 3.9
and 3.10. From the results, it is concluded that WNN is able to predict elevator
traffic flow effectively. WNN with structural risk minimization algorithm (based on
statistical learning theory) is also developed to predict elevator traffic flow (Huang et
al., 2006). Wang et al. (2006) proposed to utilize non-parametric regression theory
and wavelet algorithm to decompose traffic flow data into different channels and
reconstructed prior to removal of noise in the traffic flow information. Then, nearest
neighbour method is employed to predict traffic flow. Therefore, it is anticipated
that wavelet theory will play essential role in elevator traffic flow analysis in future.
Figure 3.8 Elevator incoming traffic prediction on a working day by WNN and ES.
Figure 3.9 Elevator outgoing traffic prediction on a working day by WNN and ES.
32
Figure 3.10 Total number of passengers prediction on a working day by WNN and
ES.
Support Vector Machine (SVM) has been proposed to build elevator traffic
pattern recognition (Xu and Luo, 2005). Recently, SVM prediction models are also
being developed to predict elevator traffic flow (Luo et al., 2005). SVM theory was
developed by Vapnik of the AT & T Bell Laboratories in 1995. Like neural network,
33
SVM is also adaptive to complex system and robust in dealing with corrupted data.
Least squares support SVM (LS-SVM) which uses equality constraint and a least
squares error term in order to obtain a linear set of equations in the dual space. Same
as other prediction model, data is collected to train the SVM. It has two ordinal
phases namely training phase and prediction phase (Yan et al., 2006). The
simulation results of LS-SVM elevator traffic flow prediction model is shown in
Figure 3.11, 3.12 and 3.13. From the below figures, a LS-SVM is an effective
prediction model as it predicts the traffic flow quite close to actual traffic flow.
Hu, Yang and Zhu proposed an elevator traffic flow prediction model by
using wavelet support vector machine (WSVM) that combines the merits of support
35
vector machine and wavelet analysis theory (Hu et al., 2007). The linear prediction
function is defined as , where x R and y R , R
represents input space and Ø is non linear mapping function. The above equation is
based on data set T x! , y! #$! , where x! denotes the input vector, y! denotes the
desired value and n is the number of data points. In order to parameters w and b,
following objective function shall be minimized.
|| ' ∑
(|( ) ( , |* where C > 0is a constant and ε is a small positive
%
∑
(,( ) ,(- .( , (3.5)
The data points corresponding to the nonzero support vectors ,( ) ,(- . The
.( , is a Kernel function. The Morlet wavelet kernel function is used. It is
expressed as following.
B
.56 78,9 9 : A8,9 9 A
.( , ∏<= 012 3 ; >? @) %B
C (3.6)
where α denotes scaling factor. Figure 3.14 depicts the structure of WSVM
prediction model. The input vector x consists of two parts, the first part {x (t - i), i =
0, 1, …, k-1} denotes the traffic flow of the past k time intervals and the second part
{x ((t + 1) – 144i), i = 0, 1, …, p}denotes the traffic flow of the same time of past p
working days. The coefficient 144 denotes the sampled data points during a working
day (7:00 – 19:00), f denotes the predicted value at time interval t + 1 and N denotes
the number of support vectors.
36
F 1
,- ) ,
. ,
∑
,%- ) ,%
F ) G 1 k
.% ,
7F 1
k+1
,E- ) ,E
) 144:
.N ,
7F 1 ) 144?:
k+p
An experiment has been carried out to verify and validate the above model in
Shanghai with parameters of C = 1000, ε = 0.75, parameters of morlet wavelet kernel
function a = 7 and parameters k = 6 and p = 5 of input vector x. Figure 3.15
illustrates the comparison of prediction result and real traffic flow data. As viewed
from the figure, two lines tend to overlap each other implies that prediction result is
approximate closely to real data. Besides that, they also compare prediction results
of five different models as tabulated in Table 3.1 and it is concluded that WSVM
provides the best elevator traffic flow prediction.
37
Figure 3.15 Comparison of prediction result and real traffic flow data.
3.3 Summary
From the past researches, it is fair to say that RBF NN is suitable to be used
to develop elevator traffic flow prediction model since elevator traffic flow is short
time nonlinear process. Various elevator traffic prediction models show that training
and recall procedure are playing important role to ensure that error is minimized to
yield good approximation. Therefore, other methods discussed in this chapter shall
be carefully studied to develop good RBF NN elevator traffic flow prediction model.
39
CHAPTER 4
METHODOLOGY
4.1 Introduction
As a case study, Elevator traffic flow data are collected in Menara BHL
(Penang) in mid of December 2007. There are four elevators in four car group
operation provide service in that building. The traffic flow data has been collected
from 7:30 a.m. to 10:00 a.m. for one week. There are 23 stories in Menara BHL. It
is diversified office building.
The collected data at field shall be consistent so that it could be used to train
the network. Therefore, under few circumstances as listed below, data collection
shall not be carried out in order to eliminate data variation. The data collection has
been carried out for a week without any holiday and a day which special function is
conducted in building.
40
• During raining day, passenger tends to reach office late due to traffic jam
on road. This may cause the up peak traffic flow to shift right with respect
to time as shown in Figure 4.1.
• Number of incoming passengers will decrease and increase significantly
during holiday and special function is conducted in building. These may
cause the traffic flow pattern is difficult to be determined as shown in
Figure 4.2.
Time
Time
Figure 4.2 Up peak traffic flow pattern during holiday and when special function is
Data has been collected in mid of December, 2007. Figure 4.3, 4.4 and 4.5
show main entrance and side view and lift lobby of BHL Tower. Numbers of
passengers enter elevator(s) in 5 minutes intervals which represent traffic flow were
reordered. Collected data is tabulated in Table 4.1.
The graphs for traffic flow are plotted for easier interpretation. They are
illustrated in Figure 4.6, 4.7, 4.8, 4.9, 4.10 and 4.11.
44
Time (a.m.)
40
35
30
25
20
15
10
5
0
Time (a.m.)
Time (a.m.)
40
35
30
25
20
15
10
5
0
Time (a.m.)
Time (a.m.)
40
No. of Passengers Enter Lifts
35
30
25 Mon
20 Tue
15 Wed
Thu
10
Fri
5
0
07:30
07:40
07:50
08:00
08:10
08:20
08:30
08:40
08:50
09:00
09:10
09:20
09:30
09:40
09:50
Time (a.m.)
The algorithm of the RBF NN shall be developed and trained so that it is able
to predict elevator traffic flow fairly accurate. It is anticipated that algorithm may
not perform satisfactory first time and need to be fine tuned in order to obtain good
result. The flow chart in Figure 4.12 shows
ows the development of algorithm.
Figure 4.12 Flow chart to develop and fine tune the RBF NN based algorithm.
algorithm
x(t) 1
φ1
x(t- k + 1) y
k φ2 ∑
Output layer
x((t + 1) – 30) k+1
φS1
Input layer
The algorithm of RBF NN prediction model to predict traffic flow of p+1 day
is
for t = k to 29
set training data to x(t+1-k,p+1), x(t+1,p) and x(t+1,p+1)
train the RBF NN
for t = k to 29
set testing data to x(t+1-k,p+1), x(t+1,p)
test the RBF NN
obtain mean square error between RBF NN output and real data
end
As refer to example (k=4, p=4) illustrated in table 4.2, the first set training data
(t=k=4) is arranged in matrix as following
[ x(0,5) x(1,5) x(2,5) x(3,5) x(4,1) x(4,2) x(4,3) x(4,4) x(4,5) ]
which is
[ 6 1 4 5 11 10 12 12 12 ]
then, subsequent set of training data (t=k=5,p=4) is set as in following matrix
[ x(1,5) x(2,5) x(3,5) x(4,5) x(5,1) x(5,2) x(5,3) x(5,4) x(5,5) ]
which is
[ 1 4 5 12 10 10 11 12 13 ]
The same procedure will repeat until t=29. The first seven values in the matrix are
inputs to RBF NN while the last value is the target of the training.
newrb function iteratively creates neurons of radial basis network at a time. Neurons
are added to the network until sum square error falls beneath an error goal or a
maximum number of neurons have been reached (Demuth and Beale, 2003). The
newrb function has a format in matlab:
net = newrb (P, T, GOAL, SPREAD)
The above function takes matrices of input vector P and target vector T and design
parameters GOAL (error goal) and SPREAD (spread constant for the radial basis
layer) and returns the network with weights and biases in such a way that outputs are
T when the inputs are P. Goal is set to 0.05 in this elevator traffic flow prediction
model.
The newrbe function is used to create exact neural network. It could produce
a network with zero error on training vectors. It takes the format in matlab:
net = newrbe(P, T, SPREAD)
This newrbe function takes matrices of input vectors P and target vectors and a
spread constant SPREAD for radial basis layer. It returns a network with weights
and biases in such a way that output are exactly T for the inputs P.
The spread constant needs to be large enough so that active input regions of
radial basis neurons overlap enough to ensure several radial basis neurons always
have fairly large outputs at any moment. This makes network functions smoother.
However, spread constant should not be so large that all neurons respond in
essentially the same manner. As such, a program to evaluate the most suitable value
of spread constant for RBF NN prediction model that yields the minimum mean
square error shall be set up. Therefore, the most suitable spread constant could be
easily selected for all cases with different k and p. Performance of both newrbe and
newrb functions is compared and verified.
evaluate performance of RBF NN. The error is calculated as the difference between
target output (real data) and network output (predicted data) and it is stated as
following
Mean square error, ε = ∑N
LJt L ) nL N
%
(4.1)
N
The optimum spread constant for both newrbe and newrb functions is
obtained for various cases as followings.
• Case 1 : use past 4 interval data on 4th day and past 3 days (1st,2nd and
3rd day) data to predict traffic flow on 5th day.
• Case 2 : use past 4 interval data on 5th day and past 4 days (1st,2nd, 3rd
and 4th day) data to predict traffic flow on 5th day.
• Case 3 : use past 3 interval data on 4th day and past 3 days (1st,2nd and
3rd day) data to predict traffic flow on 5th day.
• Case 4 : use past 3 interval data on 5th day and past 3 days (2nd, 3rd and
4th day) data to predict traffic flow on 5th day.
• Case 5 : use past 4 interval data on 5th day and past 4 days (1st, 2nd, 3d
and 4th day) data to predict traffic flow on 5th day.
• Case 6 : use past 2 interval data on 5th day and past 4 days (1st, 2nd, 3d
and 4th day) data to predict traffic flow on 5th day.
• Case 7 : use past 2 interval data on 4th day and past 3 days (1st, 2nd and
3rd day) data to predict traffic flow on 5th day.
53
• Case 8 : use past 2 interval data on 4th day and past 2 days (2nd and 3rd
day) data to predict traffic flow on 5th day.
• Case 9 : use past 2 interval data on 4th day and past 2 days (2nd and 3rd
day) data to predict traffic flow on 4th day.
• Case 10: use past 1 interval data on 5th day and past 2 days (2nd and 3rd
day) data to predict traffic flow on 5th day.
Figure 4.14 Optimum spread constant for newrbe function (for case 1).
54
Figure 4.15 Optimum spread constant for newrb function (for case 1).
Figure 4.16 Optimum spread constant for newrbe function (for case 2).
55
Figure 4.17 Optimum spread constant for newrb function (for case 2).
Figure 4.18 Optimum spread constant for newrbe function (for case 3).
56
Figure 4.19 Optimum spread constant for newrb function (for case 3).
Figure 4.20 Optimum spread constant for newrbe function (for case 4).
57
Figure 4.21 Optimum spread constant for newrb function (for case 4).
Figure 4.22 Optimum spread constant for newrbe function (for case 5).
58
Figure 4.23 Optimum spread constant for newrb function (for case 5).
Figure 4.24 Optimum spread constant for newrbe function (for case 6).
59
Figure 4.25 Optimum spread constant for newrb function (for case 6).
Figure 4.26 Optimum spread constant for newrbe function (for case 7).
60
Figure 4.27 Optimum spread constant for newrb function (for case 7).
Figure 4.28 Optimum spread constant for newrbe function (for case 8).
61
Figure 4.29 Optimum spread constant for newrb function (for case 8).
Figure 4.30 Optimum spread constant for newrbe function (for case 9).
62
Figure 4.31 Optimum spread constant for newrb function (for case 9).
Figure 4.32 Optimum spread constant for newrbe function (for case 10).
63
Figure 4.33 Optimum spread constant for newrb function (for case 10).
As viewed from figure 4.13 to figure 4.32, variation of spread constant will
cause significant fluctuation in mean square error for cases 1, 3, 7 and 8. The mean
square error is almost constant for spread constant in the range for case 2. Mean
square error is very small and could be neglected for cases 4, 5, 6, 9 and 10.
Common feature for all cases with small mean square error is past k interval data on
mth day and past p days data of are used to predicted future traffic flow on the same
day. While for cases 1, 3, 7 and 8, past k interval data on (m-1)th and past p days
data are used to train the network. Obviously, RBF NN prediction model performs
better when same day data is used to train the network. The mean square error for 10
cases is tabulated in Table 4.3. to Table 4.12.
64
The mean square error and optimum spread constant for all cases obtained
from simulation are summarized in Table 4.13.
84
Table 4.13 Summary of mean square error and optimum spread constant.
Optimum Spread Constant Mean Square Error
Case
newrb newrbe newrb newrbe
1 15.6 7.1 0.0014 0.76
2 0.6 - 25.1 0.6 - 25.1 2.15 0.2605
3 20.6 5.1 2.82 0.00328
4 0.6 1.6 0 2.6295E-31
5 0.6 14.6 0 1.2136E-31
6 0.6 9.6 0 1.1269E-31
7 15.1 5.1 0.00044056 0.031554
8 15.1 11.1 1.107E-06 0.10656
9 6.1 5.1 1.1269E+31 1.6228E+29
10 0.6 2.1 1.0881E+31 6.9637E+30
In newrb function, a goal is set at 0.05. A graph shows training curve, goal
and epochs is generated when newrb function is debugged and executed. Graphs for
10 cases are illustrated in Figure 4.34 to Figure 4.43.
85
Figure 4.34 Training of network for newrb function (for case 1).
Figure 4.35 Training of network for newrb function (for case 2).
86
Figure 4.36 Training of network for newrb function (for case 3).
Figure 4.37 Training of network for newrb function (for case 4).
87
Figure 4.38 Training of network for newrb function (for case 5).
Figure 4.39 Training of network for newrb function (for case 6).
88
Figure 4.40 Training of network for newrb function (for case 7).
Figure 4.41 Training of network for newrb function (for case 8).
89
Figure 4.42 Training of network for newrb function (for case 9).
Figure 4.43 Training of network for newrb function (for case 10).
90
CHAPTER 5
5.1 Introduction
Figure 5.1 Simulation result for newrb function (for case 1).
Figure 5.2 Simulation result for newrbe function (for case 1).
93
Figure 5.3 Simulation result for newrb function (for case 2).
Figure 5.4 Simulation result for newrbe function (for case 2).
94
Figure 5.5 Simulation result for newrb function (for case 3).
Figure 5.6 Simulation result for newrbe function (for case 3).
95
Figure 5.7 Simulation result for newrb function (for case 4).
Figure 5.8 Simulation result for newrbe function (for case 4).
96
Figure 5.9 Simulation result for newrb function (for case 5).
Figure 5.10 Simulation result for newrbe function (for case 5).
97
Figure 5.11 Simulation result for newrb function (for case 6).
Figure 5.12 Simulation result for newrbe function (for case 6).
98
Figure 5.13 Simulation result for newrb function (for case 7).
Figure 5.14 Simulation result for newrbe function (for case 7).
99
Figure 5.15 Simulation result for newrb function (for case 8).
Figure 5.16 Simulation result for newrbe function (for case 8).
100
Figure 5.17 Simulation result for newrb function (for case 9).
Figure 5.18 Simulation result for newrbe function (for case 9).
101
Figure 5.19 Simulation result for newrb function (for case 10).
Figure 5.20 Simulation result for newrbe function (for case 10).
102
Simulation results are analyzed and discussed for 10 cases. Each case
evaluates performance of RBF NN traffic flow prediction model with specified k and
p parameters and the day which comparison would be made between real data and
simulated result. On the other hand, both newrbe and newrb functions in neural
network toolbox are used to predict and simulate traffic flow based on input vector
and target vector. Optimum spread constants are also obtained for all cases with
newrb and newrbe functions.
In case 1, past 4 interval data on 4th day and past 3 days (1st,2nd and 3rd day)
data are used to predict traffic flow on 5th day. From simulation results, RBF NN
prediction model with newrb and newrbe functions able to predict up peak traffic
accurately. There are only three continuous RBF NN output (5 minutes the peak) are
found deviated from real data quite significantly. The output of RBF NN prediction
model and real data are summarized in Table 5.1. The mean square error for newrb
function is 0.0014 while mean square error for newrbe function is 0.76131.
Performance of newrbe functionis better than newrb function. Overall performance
of RBF NN prediction model is considered satisfactory.
Table 5.1 Comparison of RBF NN output for newrb, newrbe and real data (Case 1).
RBF NN Output
Real Data
newrb newrbe
12 12.228 15.017
13 12.184 11.736
14 14.189 15.568
103
In case 2, past 4 interval data on 5th day and past 4 days (1st,2nd, 3rd and 4th
day) data are used to predict traffic flow on 5th day. From simulation results, RBF
NN prediction model with newrb and newrbe functions able to predict up peak traffic
accurately. Three continuous RBF NN output (5 minutes the peak) are found
deviated from real data but the deviation is not significant. The output of RBF NN
104
prediction model and real data are summarized in Table 5.2. The mean square error
for newrb function is 2.15 while mean square error for newrbe function is 0.2605.
Again, performance of newrbe function is better than newrb function. Generally,
performance of RBF NN prediction model is satisfactory.
Table 5.2 Comparison of RBF NN output for newrb, newrbe and real data (Case 2).
RBF NN Output
Real Data
newrb newrbe
12 12.267 12.197
13 13.353 13.209
14 14.486 14.34
24 24.674 24.51
21 20.639 20.677
34 33.705 33.395
42 42 42
34 33.517 33.601
38 37.296 37.54
36 31.706 32.949
31 29.616 29.927
26 26.469 26.395
32 32.067 31.941
20 20.405 20.327
15 15.289 15.289
20 19.854 20.05
19 18.716 19.011
16 16.115 15.66
11 10.192 11.113
26 27.566 27.436
9 8.8025 8.7643
20 20.157 20.267
10 9.373 10.082
14 13.649 13.557
14 14.351 14.615
12 12.088 11.521
105
In case 3, past 3 interval data on 4th day and past 3 days (1st,2nd and 3rd day)
data are used to predict traffic flow on 5th day. From simulation results, RBF NN
prediction model with newrb and newrbe functions able to predict up peak traffic
accurately which is same as previous cases. Three continuous RBF NN output (5
minutes the peak) are found deviated from real data quite significant which is very
similar to case 1. The mean square error for newrb function is 2.82 while mean
square error for newrbe function is 0.00328. Again, performance of newrbe function
is better than newrb function. Performance of RBF NN in case 3 is close to
performance in case 1 but not as good as case 2. This shows that RBF NN use past k
interval data on mth day and past p days data as training data to predict traffic flow on
same day performs better than RBF NN trained by using past k interval data on (m-
1)th day. The value of p either is 3 days (case 3) or 4 days (case 4) does contribute
significant effect in predicting traffic flow. Generally, performance of RBF NN
prediction model is considered satisfactory. The output of RBF NN prediction model
and real data are summarized in Table 5.3.
Table 5.3 Comparison of RBF NN output for newrb, newrbe and real data (Case 3).
RBF NN Output
Real Data
newrb newrbe
5 7.2314 13.779
12 11.513 13.411
13 12.512 12.233
14 11.573 15.196
24 23.874 22.68
21 21.574 21.904
34 31.204 25.472
42 41.925 39.967
34 34.924 33.371
38 33.478 32.587
36 28.923 24.402
106
In case 4, past 3 interval data on 5th day and past 3 days (2nd, 3rd and 4th day)
data are used to predict traffic flow on 5th day. For case 5 and 6, past 4 and 2 interval
data on 5th day and past 4 days data (1st, 2nd, 3d and 4th day) are used to train RBF
NN prediction model. From simulation results for case 4, 5 and 6, RBF NN
prediction model with newrb and newrbe functions able to predict traffic flow
accurately which is better than previous cases. The mean square error for both newrb
and newrbe functions are very small and could be neglected. This shows that RBF
NN use past k interval data on mth day and past continuous p days data as training
data to predict traffic flow on same day will yield very good result. RBF NN was
well trained due to data on 1st, 2nd, 3rd and 4th day are close to each other.
Performance of RBF NN prediction model is considered excellence. The output of
107
RBF NN prediction model and real data for case 4, 5 and 6 are summarized in Table
5.4, Table 5.5 and Table 5.6 respectively.
Table 5.4 Comparison of RBF NN output for newrb, newrbe and real data (Case 4).
RBF NN Output
Real Data
newrb newrbe
5 5 5
12 12 12
13 13 13
14 14 14
24 24 24
21 21 21
34 34 34
42 42 42
34 34 34
38 38 38
36 36 36
31 31 31
26 26 26
32 32 32
20 20 20
15 15 15
20 20 20
19 19 19
16 16 16
11 11 11
26 26 26
9 9 9
20 20 20
10 10 10
14 14 14
14 14 14
12 12 12
108
Table 5.5 Comparison of RBF NN output for newrb, newrbe and real data (Case 5).
RBF NN Output
Real Data
newrb newrbe
12 12 12
13 13 13
14 14 14
24 24 24
21 21 21
34 34 34
42 42 42
34 34 34
38 38 38
36 36 36
31 31 31
26 26 26
32 32 32
20 20 20
15 15 15
20 20 20
19 19 19
16 16 16
11 11 11
26 26 26
9 9 9
20 20 20
10 10 10
14 14 14
14 14 14
12 12 12
109
Table 5.6 Comparison of RBF NN output for newrb, newrbe and real data (Case 6).
RBF NN Output
Real Data
newrb newrbe
4 4 4
5 5 5
12 12 12
13 13 13
14 14 14
24 24 24
21 21 21
34 34 34
42 42 42
34 34 34
38 38 38
36 36 36
31 31 31
26 26 26
32 32 32
20 20 20
15 15 15
20 20 20
19 19 19
16 16 16
11 11 11
26 26 26
9 9 9
20 20 20
10 10 10
14 14 14
14 14 14
12 12 12
110
In case 7, past 2 interval data on 4th day and past 3 days (1st, 2nd and 3rd day)
data are used to predict traffic flow on 5th day. Like previous cases, the up peak is
predicted accurately.. The mean square error for both newrb and newrbe functions
are 0.00044 and 0.031554 respectively. One common feature of case 1, 3 and 7 is
past data on 4th day and past 3 days data are used to train RBF NN in order to predict
traffic flow on 5th day. The simulation result shows that its performance is better
than performance of case 1 and case 3. This shows that higher k value does not
guarantee better performance in predicting traffic flow. Generally, performance of
RBF NN prediction model is considered satisfactory. The output of RBF NN
prediction model and real data for case 7 is summarized in Table 5.7.
Table 5.7 Comparison of RBF NN output for newrb, newrbe and real data (Case 7).
RBF NN Output
Real Data
newrb newrbe
4 5.6832 9.8933
5 4.189 4.381
12 11.779 12.149
13 12.993 13
14 13.759 13.959
24 24.232 24.407
21 20.227 20.739
34 33.177 33.225
42 41.436 40.842
34 34.465 33.879
38 37.528 37.485
36 35.19 33.251
31 34.523 27.971
26 28.581 25.466
32 30.046 28.911
20 20.185 19.977
111
15 15.743 15.399
20 20.39 20.661
19 18.411 19.475
16 15.774 16.095
11 11.014 11
26 26.721 24.671
9 6.3568 13.097
20 19.375 20.067
10 10.02 10
14 14.429 15.243
14 13.945 14.07
12 11.717 11.745
In case 8, past 2 interval data on 4th day and past 2 days (2nd and 3rd day) data
are used to predict traffic flow on 5th day. From simulation result, RBF NN
prediction model with newrb and newrbe functions able to predict up peak accurately
with small mean square error. Case 8 is actually similar to case 1, 3 and 7 but with
the lowest k and p value (k=2 and p=2). This shows that higher k and p values do
not guarantee better performance in predicting traffic flow. Good performance of
RBF NN prediction model in case 8 is caused selection of optimum spread constant.
As such, it is concluded that spread constant plays important role in RBF NN
prediction model. Like case 4, 5 and 6, performance of RBF NN prediction model is
excellence. The output of RBF NN prediction model and real data for case 8 is
summarized in Table 5.8 and they are no difference because mean square error could
be neglected.
112
Table 5.8 Comparison of RBF NN output for newrb, newrbe and real data (Case 8).
RBF NN Output
Real Data
newrb newrbe
4 4 4
5 5 5
12 12 12
13 13 13
14 14 14
24 24 24
21 21 21
34 34 34
42 42 42
34 34 34
38 38 38
36 36 36
31 31 31
26 26 26
32 32 32
20 20 20
15 15 15
20 20 20
19 19 19
16 16 16
11 11 11
26 26 26
9 9 9
20 20 20
10 10 10
14 14 14
14 14 14
12 12 12
113
In case 9, past 2 interval data on 4th day and past 2 days (2nd and 3rd day) data
are used to predict traffic flow on 4th day. From simulation result, RBF NN
prediction model with newrb and newrbe functions able to predict up peak accurately
with small mean square error. It is expected because past interval data on 4th day is
used to train RBF NN even k and p values are small. Again, this shows that k and p
values are not major factor in RBF NN but optimum spread constant. The output of
RBF NN prediction model and real data for case 9 is summarized in Table 5.9 and
they are no difference because mean square error could be neglected.
Table 5.9 Comparison of RBF NN output for newrb, newrbe and real data (Case 9).
RBF NN Output
Real Data
newrb newrbe
5 5 5
5 5 5
12 12 12
12 12 12
14 14 14
23 23 23
22 22 22
33 33 33
42 42 42
35 35 35
35 35 35
29 29 29
28 28 28
28 28 28
34 34 34
20 20 20
17 17 17
21 21 21
114
19 19 19
16 16 16
14 14 14
23 23 23
9 9 9
20 20 20
12 12 12
14 14 14
13 13 13
10 10 10
In case 10, only past 1 interval data on 5th day and past 2 days (2nd and 3rd day)
data are used to predict traffic flow on 5th day. The 4th day data is not used to ensure
that is not continuity in p. From simulation result, RBF NN prediction model with
newrb and newrbe functions still able to predict up peak accurately with small mean
square error. It could be past interval data on 5th day is used even k=1 to train RBF
NN. Again, this proves that k and p values are not major factor in RBF NN but
optimum spread constant. Besides that, collected traffic flow data is also consistent
and no significant deviation among readings at particular time from Monday to
Friday. This cause discontinuity in previous days data training is not contributing to
significant deviation in prediction. The output of RBF NN prediction model and real
data for case 10 is summarized in Table 5.10.
115
Table 5.10 Comparison of RBF NN output for newrb, newrbe and real data (Case
10).
RBF NN Output
Real Data
newrb newrbe
1 1 1
5 4 4
5 5 5
12 12 12
12 13 13
14 14 14
23 24 24
22 21 21
33 34 34
42 42 42
35 34 34
35 38 38
29 36 36
28 31 31
28 26 26
34 32 32
20 20 20
17 15 15
21 20 20
19 19 19
16 16 16
14 11 11
23 26 26
9 9 9
20 20 20
12 10 10
14 14 14
13 14 14
10 12 12
116
5.3 Summary
CHAPTER 6
6.1 Conclusion
In this thesis, radial basis function neural network (RBF NN) which is one of
artificial intelligent technique widely used in pattern recognition is used to develop
elevator traffic flow prediction model. In elevator traffic flow prediction models
developed by employing artificial intelligent techniques (example: square support
machine), past k interval data and past p days data (k and p are positive integer) are
used to train the core network of prediction model. The same principle is adopted to
design and develop elevator traffic flow prediction model in this thesis. RBF NN
prediction model developed in this thesis is an effective model to perform traffic
flow prediction as it needs only one week data to train the core network to predict
short term up peak traffic with good results. Shorter training period is advantage of
prediction model as it would not occupy many memory slots in controller to store
training data. Performance of elevator group control system is determined by its
capability to handle up peak traffic in morning. Therefore, this RBF NN traffic flow
prediction model could be integrated into elevator group control system to enhance
its performance. Main controller could assign more elevators to serve the main floor
or command elevators return to main floor upon completion of service at the time
when up peak occurs during working days with reference to prediction result
118
feedback by RBF NN prediction model. This will benefit elevator users as average
waiting time at the up peak hour is being reduced.
Neural network toolbox with newrb and newrbe built in functions in matlab
software is good tool to develop RBF NN prediction model because it is user friendly
and simulates results in graphical format which leads to easier interpretation.
Optimum spread constant shall be determined so that RBF NN prediction model will
yield good results whether newrbe or newrb function is excuted.
RBF NN prediction model for elevator up peak traffic flow is also possible to
be modified so that it could be used to predict traffic flow in highway tolls. With the
predicted data, highway authority could introduce plan to reduce traffic congestion at
peak hour effectively.
119
REFERENCES
Barney, G.C. and dos Santos, S.M. (1985). Elevator Traffic Analysis Design and
Control. (2nd ed.) England: Peter Peregrinus Ltd.
Chang, B.R. and Tsai, S.F. (2002). Forecasting Non-Periodic Short-Term Time
Series Radial Basis Function Neural Network Approach. 2002 International
Conference on Systems, Mans, and Cybernetics. 6, 84 – 87.
Chen, Z, Luo, F., Xu, Y. and Cao, J. (2006). Design and Implementation of Modern
Elevator Group Control System. Proceedings of the Fifth International
Conference on Machine Learning and Cybernetics. 676 – 681.
Chiang, T-C. and Fu, L-C. (2002). Design of Modern Elevator Group Control
Systems. Proceedings of the 2002 IEEE International Conference on Robotics
and Automation. 2, 1465 – 1470.
120
Choi, D., Park, H.C. and Woo, K.B. (1993). Predictive Fuzzy Control for Elevator
Group Supervisory System. Fifth IFSA World Congress, 1374 – 1377.
Demuth, H. and Beale, M. (2003). Neural Network Toolbox User’s Guide. Ver. 4.
Massachusetts : The Mathworks Inc.
Elanayar, S.V.T. and Shin Y.C. (1994). Radial Basis Function Neural Network for
Approximation and Estimation of Nonlinear Stochastic Dynamic Systems.
IEEE Transactions on Neural Networks, 5(4), 594 – 603.
Hu, Z., Yong, G. and Zhu, C. (2007). Elevator Traffic-Flow Prediction Based on
WSVMs. Elevator World, 55(6), 94 – 96.
Huang, M., Xu, L., Wang, J. and Gu, S. (2003). Predictive Method for Traffic Flow
of Elevator Systems. Fourth International Conference on Control and
Automation. 683 – 87.
Huang, M., Cui, B. and Gu, S. (2006). Elevator Traffic Flow Prediction Based on
Wavelet Neural Network. Control and Decision. 25(5), 589 – 592.
Imasaki, N., Kubo, S. and Nakai, S. (1995). Elevator Group Control System Tuned
by a Fuzzy Neural Network Applied Method. Proceedings of International
Joint Conference of the Fourth IEEE International Conference on Fuzzy
Systems and the Second International Fuzzy Engineering Symposium. 4, 1735
– 1740.
Jeong, R. and Rilett, L.R. (2004). Bus Arrival Time Prediction Using Artificial
Neural Network. 2004 IEEE Intelligent Transportation System Conference.
988 – 993.
121
Kim C.B., Seong, K.A., and Hyung, L. (1993). Fuzzy Approach to Elevator Group
Control System. Fifth IFSA World Congress. 1218 – 1221.
Kim, C.B., Seong, K.A., Hyung, L., Kim, J.O. and Yong, B.L. (1995). A Fuzzy
Approach to Elevator Group Control System. IEEE Transactions on System,
Man, and Cybernetics. 2(6), 985 – 990.
Kim, C.B., Seong, K.A. and Hyung, L. (1996). Design and Implementation of
FEGCS: Fuzzy Elevator Group Control System. 1996 Biennial Conference of
the North American Fuzzy Information Processing Society. 109 – 113.
Kim, C.B., Seong, K.A. and Hyung, L. (1998). Design and Implementation of a
Fuzzy Elevator Group Control System. IEEE Transactions on Systems, Man,
and Cybernetics Part A: Systems and Human. 28(3), 277 – 287.
Kone Corporation. Annual Report 2006. Helsinki (Finland): Annual Report. 2006.
Li Z., Zhu, Y., Li, C. and Mao, Z. (2003). Elevator Traffic Flow Analysis Based on
Artificial Immune Clustering Algorithm. Journal of South China University of
Technology (Natural Science Edition). 31(12), 26 -29.
Luo, F., Xu, Y. and Cao, J. (2005). Elevator Traffic Flow Prediction with Least
Squares Support Vector Machines. Proceedings of the Fourth International
Conference on Machine Learning and Cybernetics. 7, 4266 – 4270.
Qin, W., Huang, Q., Yin, H. and Jia, L. (2006). Traffic Flow Forecasting Based on
RBF Neural Network. Journal of Tianjin Polytechnic University. 25(2), 71 –
73.
Qu, L., Chen, Y. and Liu, Z. (2006). Time Series Forecasting Model with Error
Correction by Structure Adaptive RBF Neural Network. Proceedings of the 6th
World Congress on Intelligent Control and Automation. 2, 6831 – 6835.
Siikonen, M-J., (1997b). Planning and Control Models for Elevators in High-Rise
Buildings. Research Report A68. Helsinki University of Technology System
Analysis Laboratory, Helsinki.
Tan, K.K. (1997). Fuzzy Reasoning by Ordinal Structure Model for Elevator Group
Supervisory Control. Master Thesis. Universiti Teknologi Malaysia, Skudai:
Wang, W. and Dong, Y. (2004). The Prediction Method of Short-Time Traffic Flow.
Journal of Shandong Jiaotong University. 12(2), 6 – 9.
Wang, X. and Xiao J. (2003), A Radial Basis Function Neural Network Approach
to Traffic Flow Forecasting. 2003 IEEE Proceedings on Intelligent
Transportation Systems. 1, 614 – 617.
Wang, X et al. (2006). Traffic Flow Forecasting Method Based on Wavelet Non-
parametric Regression Algorithms. System Engineering. 23(10), 44 – 47.
123
Xu, Y. and Luo, F. (2005). Traffic Pattern Recognition Method for Novel Elevator
System. Control Theory & Applications. 22(6), 900 – 904.
Yan, X., Liu, X., Mao, Z., Li, Z. and Tan, H. (2006). SVM-based Elevator Traffic
Flow Prediction. Proceedings of the 6th World Congress on Intelligent Control
and Automation. 2, 8814 – 8818.
Zong Q., Yue, Y., Cao, F. and Shang, X. (2000). Application of Intelligent
Forecasting Method in Traffic Analysis of EGCS. Transactions of Tianjin
University. 6(1), 18 -21.
Zong, Q., Wei, L. and Cheng, Y. (2004). Elevator Configuration Based on Markov
Network Queuing Model. Proceedings of the 5th World Congress on Intelligent
Control and Automation., 1, 236 – 239.
Zong Q., Dou, L. and Wang, W. (2006a). Elevator Group Control Scheduling
Approach Based on Multi-Agent Coordination. Proceedings of the 6th World
Congress on Intelligent Control and Automation. 7249 – 7253.
APPENDIX
MSEresult(jj,1)=(sum(rbfoutput'-x(5:30,6)).^2)./26;
end
figure; plot(MSEresult);
126