Traffic Analysis For Elevators PDF
Traffic Analysis For Elevators PDF
Traffic Analysis For Elevators PDF
Abstract
Vertical transportation refers to the problem that arises when a passenger wishes to
travel by lift between the floors of a building. Controllers are installed in lifts that aim
to maximise specific criteria such as passenger waiting time, the energy consumption,
traffic analysis must be previously carried out in the building. In fact, peak traffic can
lead to dramatic waiting times during busy hours in the case of no modification of the
car-call allocation rules. This paper presents a fuzzy logic based controller for peak
traffic detection in elevator systems. The controller is validated with the most
authorised traffic patterns in the industry for office buildings, namely the CIBSE Guide,
and the Strakosch and Siikonen traffic patterns. The fuzzy logic controller demonstrate
a suitable performance providing adequate traffic pattern identification for all cases.
The controller was proved to be robust and reliable when tested with patterns depicting
sudden or smooth changes, and with high or low maximum and minimum peaks.
1
1. Introduction
Vertical transportation refers to the problem that arises when a passenger wishes to
travel by lift between the floors of a building. The situation takes place when the
passenger requests a lift by pressing a landing call button installed on every floor or
selects its destination by pressing the corresponding button located inside every deck.
The elevator controller receives the call and identifies which car of the elevator group
would be most suitable for such call. The system goal to be solved through allocation is
to select a lift for each call that will minimise a preselected cost function representing a
criteria or group of criteria such as the passenger waiting time, the total percentage of
large waiting times, the ride times or the energy consumption amongst others.
Most of the controller algorithms are preconfigured and make use of performance rules
that are applied depending on the type of traffic in the building at that time. The
elevator group control system therefore needs to comprise some form of traffic-type
detector. This can be carried out from using a simple timer (which usually leads to a
bad performance due to the lack of flexibility) up to the use of most sophisticated IA
Traffic patterns represent the demand and behaviour of passengers wanting to use a lift
to travel from their origin to their destination. Depending on the specific form of the
typical traffic pattern in an office building, following the classic theory (Barney, 2003).
It shows the number of up and down landing calls that are registered during the working
day.
2
Traffic volume
Moving up traffic
uppeak lunchpeak
Time
lunchpeak downpeak
Traffic volume
Normally, at the start of the day there is a larger than average number of up landing
calls. These are due to the building’s workers arriving to start work. This stage is
called uppeak traffic. Later in the day there is the opposite phenomenon, and a larger
than average number of down landing calls takes place. This corresponds to the
building’s population wanting to go home after the working day. This traffic pattern is
called downpeak. In the middle of the day there are two joint phenomena, due to the
appearance of up and down peaks. Figure 1 depicts a situation of people leaving the
building for lunch and people coming back after lunch. This period has been called
midday or lunchpeak traffic. Finally, the rest of the day can be characterised for a
constant low demand (usually around 4% of the total population) in both directions.
This period has been called the interfloor traffic. Interfloor traffic can be either balanced
or unbalanced depending on whether the demand and/or the destination are heavily
3
The correct determination of the period of traffic pattern being experienced by the
building is a key factor, because most of the calculations to estimate the passengers’
average waiting time, the round trip time and other performance measurements, depend
on the identification of the corresponding peak period [references for a quick revision
on performance indexes in vertical transportation are (Barney, 2003) and the (CIBSE
For example, if uppeak traffic situations are not identified quickly, long queues can
build up in the main entrance floor of the building (typically the ground floor) and
passenger waiting times will become longer. Long waiting times cause dissatisfaction
with the operation of the lifts. However, the uppeak mode should not be activated
unnecessarily, as a real practice rule consisting of the direct return of lifts to the
entrance floors would be activated, resulting in long waiting times for passengers who
are waiting at landings other than the ground floor. In such a case, calls issued from
floors other than the main entrance floor obviously take longer to be served than during
normal traffic.
Most elevator designs are currently based on uppeak calculations because this is the
most problematic period in terms of handling capacity (Barney, 2003), that is in terms
the transportation system (waiting times, ride times, etc...) uppeak period is not the most
appropriate choice for the analysis because the lunchpeak traffic can in fact lead to more
complex situations and has been considered the most difficult period for dispatching
landing calls (see for example Barney and Peters, 2005; and Cortés et al. 2004).
4
For example, a classic widely implemented algorithm such as the Estimated Time of
situations, but its results are poor for lunchpeak or downpeak situations. Also, a
served by subgroups of the elevator group system. In this sense, the Adaptive Call
peak period. In summary, when the complexity of the algorithm is increased, a quick
and correct identification of the period becomes a key factor for the success of the
algorithm.
The paper continues with the second section, which includes a literature review of
methods used to detect different types of traffic both in industry (patents) and scientific
literature. The third section is devoted to the detailed description of the fuzzy logic
controller, and the fourth section presents the experimental results using real data. The
2. Literature review
The traditional procedure for traffic pattern identification has been the use of passenger
traffic surveys. In manual surveys, observers count passengers entering and exiting the
lifts. Manual surveys are normally based on one of two approaches. The first is a
5
Survey from main terminal, when the observers count passengers in and out of the lifts
of the main terminal floor (typically the ground floor). This method is only suitable for
uppeak designs that include the problems addressed previously. An alternative method
is In-car survey, when observers are situated in the lift car and count passengers in and
Manual surveys are discussed in detail in Barney and dos Santos (1985), and the
Elevator World’s Guide to Elevatoring (1992). The main problem of such methods is
the statistical nature of such data. Due to data being collected during one specific point
in time and the fact that complex real life buildings are continuously varying and their
population habits are always changing, static surveys do not provide reliable permanent
data. Similarly, the human observers’ method becomes impracticable for heavy traffic
An alternative to manual observers consists of the use of historical data in order to carry
out a statistical analysis together with a set of rules based on threshold values.
Thangavelu and Kandasamy (1993) make use of the load per car during the day to
predict the traffic behaviour of the following day. The data is managed and compared
to a set of threshold values that allow the peak period to be identified, taking into
account the relative population of the building. Data collection is carried out during
small time intervals. The proposal consequently consists of an expert rules system
6
KONE has been developing several approaches based on calculus with statistical series
for historical data. This is the case of the single exponential smoothing and the adaptive
response rate single exponential smoothing (ARRSES) that was used by the Traffic
Master System 9000. See Siikonen (1997), where the predictions are updated for each
floor and journey direction. More recently, Tyni (2005) has patented an algorithm for
KONE that includes statistical analysis and fuzzy rules to predict future traffic volumes.
In a similar way, Luo et al. (2005) presented a elevator traffic flow prediction based on
statistical learning theory. They predict elevator traffic flow using least squares support
vector machines, which are a kind of support vector machine with quadratic loss
function.
The statistical methods include the same problem: there is a delay between the data
collection and the prediction for following days or periods, meaning that no real-time
computer vision tracking systems allows the capture of data in real time and in a
dynamic manner. This data must however be real-time managed in order to identify the
traffic pattern and the peak period of the building. In this context, Fuzzy Logic plays a
relevant role when dealing with control methods in cases of uncertainty, whether for
input data or output decision. Siikonen and Leppala (1991) proposed a fuzzy logic
algorithm to classify the traffic pattern; however features such as the ratio of incoming
passengers and the ratio of outgoing passengers were not used as input variables in their
work. This paper presents a fuzzy logic controller that takes into account these aspects
7
and other significant characteristics of vertical transportation systems. The fuzzy logic
algorithm is capable of processing real-time data very quickly and identifying the
Moreover, the use of complex approaches providing outstanding simulation results are
never installed in real practice due to hard-designs difficulties because they are not
a real practice implementable solution. So, the lack of need for a complex memory and
the use of fuzzy logic are decisive factors that make the model easy and economically
worthy to be implemented.
In real life, traffic patterns can show variations and a certain level of inconsistency
between a priori similar days. Even rapid fluctuations from one instant to another
require fast peak traffic identification. Nowadays the peak traffic identification of
surveys or static statistical analysis. Even more, the use of a historical memory does not
it.
Here we propose the use of Fuzzy Logic (FL) for the dynamic identification of peaks.
FL is especially useful when quick and sudden fluctuations are merged with other slow
and smooth stages. In this sense, FL is used to deal with reasoning that is approximate,
rather than precise. In fuzzy logic the degree of truth of a statement can range between
0 and 1 and is not constrained to the two truth values {true (1), false (0)}, as in classic
8
predicate logic. When linguistic variables are used, these degrees may be managed by
specific functions.
FL is a robust method that does not need much input information. It can be described in
the following three steps: (i) fuzzification, where the values from the inputs are
converted into fuzzy values; (ii) inference process based on logic rules; (iii)
The FL model we developed follows the flowchart in figure 2. The controller provides
an accurate pattern forecasting just starting from some simple initial data:
The use of feedback increases the robustness of the controller by aiding it to recognise
more accurately the demand curve which leads to a better forecasting. It works as a
reinforced recent memory. Even more, using such simple feedback avoids tedious
The controller analyses the situation of the traffic in the building every ∆t seconds,
9
The controller receives information from input sensors related to the car load each time
an event (change of load or new call appears) is detected. Data related to the movement
mu' Variation of the weight (car load) moving up between ∆t and ∆t-1.
md Total weight (car load) moving down during the interval ∆t.
md' Variation of the weight (car load) moving down between ∆t and ∆t-1.
Values for mu and md are calculated by weighting the registered total car load in a trip
mu ∆t
= ∑ mui ⋅ ∆tui
i
md = ∑ md j ⋅ ∆td j (1)
∆t
j
Figure 2 illustrates a typical trip process in a vertical transport system where up calls
and down calls take place and the cars have to serve those calls. The case
10
mu1 mu 2 md 1 mu 3 md 2 mu 4
∆tu1 ∆tu 2 ∆td 1 ∆tu 3 ∆t d 2 ∆tu 4
t0 tT
∆t − 1 ∆t ∆t + 1
t0 tT
Where:
mu ∆t
= mu1 ⋅ ∆tu1 + mu2 ⋅ ∆tu2 + mu3 ⋅ ∆tu3 + mu4 ⋅ ∆tu4
(2)
md ∆t
= md1 ⋅ ∆td1 + md 2 ⋅ ∆td2
It has to be taken into account that the total weight must include the sum of the car load
11
The variation of weight being transported is calculated as the percentage variation
mu − mu
mu' = ∆t ∆t −1
∆t mu ∆t −1
(3)
'
md ∆t
− md ∆t −1
m d ∆t =
md ∆t −1
The car load moving up and down is caught by the sensors and is then grouped in one of
Similarly, the variation in car load between consecutive periods is classified according
to:
Triangular membership functions are constructed in order to state the ranges of variation
for each fuzzy variable. The membership function identifies the participation of each
input, associating a weight with each of the inputs processed, defining overlaps between
12
mu (or md) membership function
1.0
S M B
S&M M&B
0.0
M/3 M/2 2M/3
N Z P
N&Z Z&P
0.0
-0.3 0.0 +0.3
The expected maximum capacity moving up (or down) during the time interval is
calculated as:
∆t − t stops
M = 80% × ( car capacity ) × (4)
2
Formula (4) is calculated taking into consideration that rarely the total weight
transported by the car exceeds an eighty percent of the maximum allowed amount of
weight, and also it considers that a car that is not stopped spends half the time moving
upwards and half the time moving downwards if we consider a time interval long
enough.
13
3.4. Inference process
Fuzzy variables (input for the FL controller) are processed to get the corresponding
'
R3u: mus && mup THEN Output = UP_FLOW
'
R4u: mum && mun THEN Output = UP_FLOW
'
R6u: mum && mup THEN Output = UP_FLOW
'
R7u: mub && mun THEN Output = UP_FLOW
'
R9u: mub && mup THEN Output = UP_FLOW
'
R12u: mus && mup THEN Output = NON_UP_FLOW
'
R13u: mum && mun THEN Output = NON_UP_FLOW
'
R15u: mum && mup THEN Output = UP_FLOW
'
R16u: mub && mun THEN Output = UP_FLOW
14
'
R18u: mub && mup THEN Output = UP_FLOW
'
R3d: mus && mup THEN Output = DOWN_FLOW
'
R4d: mum && mun THEN Output = DOWN_FLOW
'
R6d: mum && mup THEN Output = DOWN_FLOW
'
R7d: mub && mun THEN Output = DOWN_FLOW
'
R9d: mub && mup THEN Output = DOWN_FLOW
'
R12d: mus && mup THEN Output = NON_DOWN_FLOW
'
R13d: mum && mun THEN Output = NON_DOWN_FLOW
'
R15d: mum && mup THEN Output = DOWN_FLOW
'
R16d: mub && mun THEN Output = DOWN_FLOW
'
R18d: mub && mup THEN Output = DOWN_FLOW
15
The input degree of membership is calculated by means of different subsets operated by
IF conditions together with the use of the logical sum OR (||) and the logical product
AND (&&).
The fuzzy inference gathers the conditional rules around four fuzzy variables meaning
how are the strength of the up and down flow, and the strength of the non-up and non-
down-flow. The rules include not only the former variables but also the previous pattern
Once the rules have been processed, the firing strength of each rule is calculated. The
logical products for each rule are then combined by means of the root-sum-square
(RSS) method, which combines the effects of all applicable rules, scales the function at
their respective magnitudes, and computes the fuzzy centroid of the composite area.
The RSS method was chosen to include all contributing rules, since there are few
member functions associated with the inputs and outputs for the respective
Thus:
"Non_Up_Flow" = ( R1u 2
+ R10u 2 + R11u 2 + R12u 2 + R13u 2 )
16
"Non_Down_Flow" = ( R1d 2
+ R10d 2 + R11d 2 + R12d 2 + R13d 2 )
3.5. Defuzzification
The next step consists of the defuzzification of the data into a crisp output by means of
the fuzzy-centroid algorithm, combining the results of the inference process and
To do so, the following output membership functions (for downpeak and uppeak) are
17
Uppeak membership function
1.0
NON-UPPEAK UPPEAK
N&U
0.0
M/6 M/2 5M/6 7M/6
NON-DOWNPEAK DOWNPEAK
N&D
0.0
M/6 M/2 5M/6 7M/6
Where M/6 and 7M/6 represent the non-down-flow and non-up-flow, and up-flow and
down-flow centroids respectively. Using the fuzzy-centroid algorithm, the up-flow and
down-flow analyses are set out below with the calculation of the crisp values (7).
crisp downpeak =
[No-down.center ] × [No-down.strength] + [Down.center ] × [Down.strength]
No-down.strength+Down.strength
(7)
crisp uppeak =
[No-up.center ] × [No-up.strength] + [Up.center ] × [Up.strength]
No-up.strength+Up.strength
18
Once the crisp numbers have been obtained, the final peak identification rules are
Finally, in order to select the peak proposal, the maximum method is used, which
selects the output at which the fuzzy subset has it maximum truth value (the higher
probability).
4. Experimental results
In order to test the performance of the proposed controller, simulations using ElevateTM
software have been carried out. ElevateTM is a vertical transport simulation software
release from Peters Research Ltd. (a detailed description about ElevateTM can be found
in Caporale, 2000) and is mainly based on Dr. Peters’ developments. Its main
The proposed building has been designed to cope with the demand in terms of handling
capacity, as it is described in Barney (2003). The building has five floors with 41 people
on each floor, so a single 600 kg capacity car ensures an acceptable transport capacity
by means that 15% of the population can be served during the busiest five minutes of
uppeak traffic.
19
The detection of the controller is analysed using different traffic patterns that have been
especially created by expert researchers in vertical transport. These three different types
of traffic are a well-known basis of analysis that is managed by the most relevant
vertical traffic planning and simulation tools (Peters, 1998; Carporale, 2000; Cortés et
al. 2006). The patterns employed for the simulations are the CIBSE Guide pattern
(2004), the Strakosch pattern (Strakosch, 1998) and the Siikonen pattern (see Siikonen,
1993 and Siikonen, 2000). They all correspond to the daily traffic in an office building,
from the start of the working day to the end, and consider all the possible types of
traffic.
The following figures represent the daily hour and the vertical transport demand (the
population). Different colours and tones represent the different traffic patterns
forecasted by the controller. With regards to the computational time, the FL controller
The CIBSE Guide traffic pattern in Figure 5 depicts a typical day in an office building,
20
Figure 5: CIBSE Guide traffic pattern and FL controller output
The CIBSE Guide model describes the typical expected behaviour, showing a clear
uppeak period at the start of the day and a downpeak period in the latter part of the day.
moderate uppeak.
The Strakosch pattern represents a more atypical pattern with smoother peaks, smaller
maximums and minimums, and light uppeak and downpeak periods taking place both in
the morning and in the afternoon. Figure 6 depicts the controller response when
21
Figure 6: Strakosch traffic pattern and FL controller output
Finally, we present the Siikonen pattern. This traffic pattern has been created by Kone
Corporation and is used by the company as a basis of analysis for traffic analysis in
office buildings. It is claimed to be the most realistic traffic pattern in literature. Figure
22
Figure 7: Siikonen traffic pattern and FL controller output
The three patterns displayed above show significant differences. This aspect provides
an idea about the robustness of the controller and its reliability, independently following
Occasionally, the controller shows a quick change in the identified peak (e.g.
Strakosch pattern between 11:15 and 11:30). These rarely fast variations on the
detected pattern occur because slight modifications in the real demand shake sometimes
the working point between the theoretical frontiers of the predefined patterns. However,
there is no need to provide the controller with a hysteresis mechanism: when the
demand is moving between the bounds of two different patterns, it is not very important
to classify the traffic pattern as one type or the other because in fact it is possible to state
23
Moreover, neither the selected car dispatcher nor the number of elevators in the system
condition the simulation results as long as the system is correctly designed to attend
passengers’ demand employing less time than the length of the predefined time interval
(∆t). Taking into account that the traffic controller is generally designed after the
elevator system and regarding the mentioned aspect, the time interval (∆t) should be
chosen at least greater than two times the Round Trip Time in order to make the
simulations were carried out concerning this aspect and it could be observed that for
time intervals lower than this threshold sometimes the behaviour was the not desired.
Note: Round Trip Time is the average time each lift employs since it opens its doors for
gathering the passengers at the main floor until it opens them again for gathering the
5. Conclusions
This paper presents a fuzzy logic based controller for traffic peak detection in elevator
systems. The controller has been tested with the most authorised traffic patterns in the
industry for office buildings, which are the CIBSE Guide, Strakosch, and Siikonen
(Kone Corporation) traffic patterns. The FL controller showed a more than suitable
and reliable when tested with patterns depicting sudden or smooth changes, and with
Changes are never detectable before the end of the period as the controller only puts
into exam the traffic happened once the time interval is over. Theoretically the average
time employed in the detection of a change in the traffic pattern is ∆t/2 minutes. E.g. in
24
a typical 5 minutes interval the average control reaction takes place every 2.5 minutes
once the traffic has changed. This figure gives an approximate idea of how fast the
controller detects a change in the traffic pattern. Besides, in this sense it also exists a
compromise between a fast detection and a reliable detection: Time interval could be
reduce so average detection is prompter but never less than twice the round trip interval
as working under this threshold could lead to undetermined and unreliable performance.
Fuzzy logic based controllers usually have a problem based on memory faults, affecting
the performance of such controllers. To avoid such problems, fuzzy logic controllers
are combined with neural networks, thus increasing the complexity of the controller
implementation. In the case of dispatching lifts this factor is crucial. The computation
availability does not allow complex implementations as neural networks could require it
when combined with fuzzy logic. The use of feedback allows avoiding such an
inconvenience: Through the parameter T_prev, which provides information about the
traffic pattern detected in the previous period, the system is able to “remember” what
Our further research is making use of this research in the attempt to further the design of
a new fuzzy logic controller that is integrated into a multiagent system, governing a sub-
Acknowledgements
The authors would like to thank MACPUAR, S.A. (Spanish elevator company) for the
support of this research line since 2000. Also, the authors acknowledge the financial
25
support given by the Consejería de Innovación, Ciencia y Empresa (Junta de Andalucía)
References
Barney, G., 2003. Elevator Traffic Handbook: Theory and practice, Taylor & Francis
Group.
Barney, G.C., dos Santos S.M., 1985. Elevator traffic analysis design and control
Barney, G. and Peters, R. 2005. Traffic planning and selection of lift equipment and
Caporale, R. S., 2000. Elevate traffic analysis software (eliminating the guesswork).
Cortés, P., Larrañeta, J., Onieva, L., 2004. Genetic Algorithm for Controllers in
Elevator Groups: Analysis and simulation during lunchpeak traffic. Applied Soft
Computing. 4, 159-174.
Cortés, P., Muñuzuri, J., Onieva, L., 2006. Design and analysis of a tool for planning
and simulating dynamic vertical transport. Simulation: Transactions of the Society for
Luo, F., Xu, Y-G., Cao, J-Z., 2005. Elevator traffic flow prediction with least squares
Peters, R.D., 1998. Simulation for control system design and traffic analysis, in: Barney,
26
Siikonen, M.L., 1993. Elevator Traffic Simulation. Simulation: Transactions of the
Siikonen, M.L., 1997. Elevator group control with artificial intelligence, Helsinki
Siikonen, M.L., 2000. On traffic planning methodology, in: Lustig, A. (Ed.), Elevator
Elevcon Berlin.
Siikonen, M.L., Leppala, J., 1991. Elevator traffic pattern recognition, in: Proceedings
Strakosch, G.R., 1998. The vertical transportation handbook 3rd Edition. John Wiley &
Sons, Inc.
system predicting "peak-period" times for elevator dispatching. Otis Elevator Company,
The Chartered Institution of Building Services Engineers London, 2004. CIBSE Guide
Tyni, T., 2005. Identification of incoming peak traffic for elevators. KONE Corporation
27