Professional Documents
Culture Documents
Joaquín R. Fernández
and Pablo Cortés
E
levators (also called “lifts” in some locales)
are installed in buildings to meet the verti-
cal transportation needs of the building oc-
cupants. The total vertical transport capacity
of a building is a crucial factor in its success
as a working, living, or service facility. Elevators must
be easily accessible, readily available, provide quality
service, and be reliable.
Since 1945, the number of buildings of medium to
high height has increased considerably (see Table 1),
providing a strong impetus to the elevator manufac-
turers, who have been receptive to the inclusion of
new technologies. First, automatic cabins were devel-
oped, followed by hydraulic cabins. More recently,
digital equipment has been included, and control sys-
tems and traffic management have been studied [2].
The total cost of an elevator, which includes the
direct installation costs, the maintenance service
costs, and the indirect costs resulting from the space
it occupies (in the layout of the building), is too high
to tolerate suboptimal elevator placement or ineffi-
cient operation. However, elevators must be designed
and positioned to provide the required service at a
minimum cost. These opposite viewpoints result in a
compromise between the quality of provided service
and economic restrictions. This tradeoff between
Program-
waiting time requirements are easily
Linear
ming
Methodology achieved as a collateral effect, even
when they are not properly opti-
Decision
mized. Although power consump-
Tree
tion is considered to be crucial in a
sustainable development, few ap-
Pattern Detection
Camera Scale
proaches currently include power re-
Intelligence
Artificial
Radar
duction [11]. Technology has evolved,
and new hoist elevator models allow
a portion of the energy that the eleva-
tor employed while moving up to be
regained. However, only a few stud-
Number of Passengers
Data Capture
Calculation of
Deck Position
regain of energy in their research.
on a Deck
Preferential
Hardware
Sensor
Software
Others have attempted to reduce
Service
Control
call allocation. Observers collected
data from the beginning of the last
century until the 1980s in manual sur-
veys [15]. By the end of the 1980s, the
Movement
Elevator
on the current traffic. Once an up-peak or down-peak ing weight between periods, and the traffic pattern detected
period is detected, the system checks the load of the decks in the previous interval. Finally, the anticipation of demand
for a predefined time interval. If the loads are greater than by incorporating historical data tables into the dispatching
a determined load value, then the up-peak or down-peak algorithm has been attempted [38].
condition was detected. Some basic approaches are used to
address this problem by using models based on basic learn- Approaches to Parameter Calibration
ing [30], which uses a historical and recent database to One of the main difficulties when designing an EGCS is the
determine if the traffic pattern that occurs is up peak, down need to a priori set the values of a wide variety of parame-
peak, or neither. ters that affect system performance [39]. These parameters
According to statistical learning theory, demand can be range from the specific settings of an optimization algo-
predicted by using a support-vector machine with a qua- rithm, to weights in heuristic dispatcher or PI controllers,
dratic loss function, which is called a least-squares support- or even to the parameters defining the linguistic regions of
vector machine [31]. In [32], a system was patented to detect fuzzy approaches. A discussion of the approaches to over-
up-peak patterns by using statistical data obtained from come this difficulty is given next.
the actual incoming demand. This problem requires expert knowledge, but if the
Fuzzy logic is especially suitable for detecting traffic system is not able to adapt correctly to the traffic pattern
patterns [33] due to its own definition and the nature of the changes, then the system will not maximize its perfor-
detection problem. Fuzzy-logic rules have been suggested mance. A control system in which the parameter values are
to classify traffic patterns, but relevant input data, such as calibrated during the night while accounting for the data
the time of day or the passenger rate of arrival or depar- gathered during the day has been suggested [40]. This
ture, are not considered [34]. Another model that is able to system conducts a calibration by using a genetic algorithm
detect traffic and allocate landing calls using a set of fuzzy that aims to provide robustness to the system in which the
rules has also been suggested. However, this suggested performance does not change between the optimization
model lacks any control strategy based on typical criteria, conditions and the actual work conditions. This change
such as the AWT, percentage of long waiting calls, or power between optimization conditions and actual work condi-
consumption [35]. The proposed control system can be tions may result from changes in the objective function
divided into three different modules, one for allocating (tenant preferences) or from different boundary conditions.
calls, one for generating the control strategy, and one for These changes are addressed by using a solution bank
traffic management, which gathers and memorizes the pas- memory or excluding some quality solutions that are found
senger flow with time [36]. This module can recognize dif- for different traffic patterns to obtain fast convergence,
ferent patterns by measuring a set of five features from respectively. However, the latest trend in calibration sys-
three different perspectives that constitute the system tems consists of solving the problem without inferring any
inputs. Based on the traffic, the features of up traffic and data from statistical memory [41]–[45].
down traffic can be distinguished. From the traffic percent- An EGCS that considers two different criteria simulta-
ages, the system can examine the centralized in-traffic per- neously along with AWT and power consumption and
centages and the distributed out-traffic percentages. optimizes them simultaneously by employing an evolu-
Finally, the time can be measured throughout the day. tionary standardized-objective weighted aggregation
From these five features, the system is capable of deducing method is suggested in [41]. Because it is mathematically
eight different traffic patterns. impossible to minimize the criteria simultaneously, the
Another possibility for detecting traffic patterns is to system balances the importance of each one by using a set
combine the fuzzy inference with statistics that were calcu- of weights that depend on certain features. This balance is
lated from a data-gathering system [37]. In this case, the achieved by using a PI controller that calibrates the set of
system can adapt smoothly to different situations. In [9], weights after each iteration and depends on the deviation
the detection of the traffic pattern using a fuzzy-logic of the anticipated AWT deviation. When the demand is
model solves the lack of memory in fuzzy-logic designs by heavy, the system attempts to minimize the AWT. How-
considering as the input values the total ascending and ever, when the demand is low, the system attempts to mini-
descending weights that are carried by the cabins during a mize power consumption. A similar EGCS has been
time interval, the variation in the ascending and descend- suggested [43], but the latter’s short evaluation function is
Technical
Services
Architectural Design Variations
for Cars and Shafts
This section reviews several design alternatives that are
used to increase the handling capacity and improve the
efficiency in the dispatching process. For this purpose,
these designs place more than one car in the same shaft,
Local either attached to each other or free to travel separately.
Elevators
This section first presents double-decker systems and then
Skylobby analyzes other multicar designs.
Technical
Services Double-Decker Systems
A double-decker system consists of two cabins that are
joined and travel together along the same shaft. In fact, a
Express Express double-decker is an elevator with two cars attached
Elevators Elevators
together, one on top of the other. This system allows pas-
sengers on two consecutive floors to use the elevator
simultaneously, significantly increasing the passenger
capacity of the elevator shaft. The main advantage of a
double-decker system is the elevator footprint reduction
Six Stories
Objective
Data Traffic-Pattern Landing-Call Parameter Double-Decker Multicar Optimization
Reference Gathering Detection Allocation Calibration System System Dynamic Set
[17]–[19]
[26], [37]
[35], [36], [74]
[34]
[13]
[40]–[42], [44]
[39]
[7], [50]–[52], [61]–[65], [69], [70],
[72], [73], [77], [78], [80]–[82],
[84], [86], [88]
[91], [92], [96], [100], [101]
[103]–[112]
[114]–[117], [119], [120]
Criteria
Maximum WT
+ Number Average Load
Uniform Distance of Stops + Number of Preference
% of Round- + Number Number of Passengers + Service Handling
Reference AWT LWT Energy ATT AJT Trip Time of Stops RGP Loops per Trip Preference Capacity Interval
[35], [40], [44], [50],
[51], [61]–[64], [69],
[72]–[74], [77], [78],
[80], [82], [86], [91],
[92], [106], [107]
[36]
[34], [52], [108]
[13]
[41]
[42]
[70], [88], [96]
[81]
[7]
[65]
[84]
[100]
[101]
[103], [109], [110]
[104]
[105]
[111]
[112]
[114]
[116]
[115], [117], [119], [120]
AWT: average waiting time
%LWT: percentage of long waiting calls
ATT: average trip time
AJT: average journey time
Max WT: maximum waiting time
RGP: genetic network programming
Interval: average time for having an idle elevator ready to serve at the main floor.
names the adjacent deck that was not directly considered Multicar Elevator Systems
by the algorithm for optimizing the landing-call assign- Multicar elevators (MCEs) consist of several cars that are
ment as “other-cage.” The system handles 12 different items placed in a single elevator shaft. This architectural variation
or features that represent the state of the system. has recently attracted significant attention among high-per-
The authors of [102] suggest another genetic network formance transportation systems. However, the accumulat-
programming approach. First, this algorithm optimizes the ed knowledge regarding single-car elevators is not readily
genetic network programming for solving the double- applicable to MCEs due to their specific operation and differ-
decker system problem. Next, its performance is compared ent performance. An MCE system’s main problem is avoid-
with the performance obtained by using the conventional ing conflict between the two cars. Therefore, an MCE group
methods for ratification. control method that can efficiently allocate service calls
Methodology
Particle
Historic Fuzzy Neural Genetic Ant Colony Tabu Multi- Search Petri Dynamic Immune Swarm Cellular
Reference Memory Logic Net Algorithm Optimization Search agent Tree Net Programming Algorithm Algorithm Automata
[26]
[37]
[74]
[13],
[34]–[36],
[40], [44],
[50]–[52]
[41],
[61]–[65],
[69], [70],
[80], [100],
[111], [112]
[42]
[77], [78]
[88]
[81], [82]
[7]
[72], [73],
[92], [109]
[86]
[84]
[101]
[115]
[116], [117]
[85], [119],
[120]
while preventing interference among the cars is required. In However, the main objective is usually to reduce the average
[103], an MCE group control algorithm is proposed that im- total waiting time, as is the case of [107], which proposed a
plements an interference risk evaluation, a schedule-comple- method that minimizes the waiting times based on an algo-
tion-time minimization function, and idle-car parking rithm that allocates requests to elevator cars by using a local
strategies. The algorithm achieves an efficient allocation of search method to reduce the wasted time of idle cars. Other
calls and prevents interference without restricting the opera- techniques, such as [108], consist of defining dynamic zones
tion of elevator cars. The simulation results showed good in which the elevators can operate without conflict. More
performance under low and medium passenger arrival complex approaches use other artificial intelligence tech-
rates. The study in [104] also proposes a method to avoid col- niques such as reinforcement learning in [109], where a con-
lisions between cars in an MCE system, but in this method, trol agent selects the best of four strategies, namely:
collision avoidance is treated as a dynamic optimization 1) transportation strategy, 2) passenger strategy, 3) zone strat-
problem (to determine which floor each car should visit next egy, and 4) difference strategy, according to the traffic pat-
under the constraint that collisions never occur). In this ap- tern. Similar approaches, such as [110], use an average
proach, the group controller gives basic car travel schedules. reward learning method to reduce service completion time
Other contributions, such as [105], follow a mathematical ap- and prevent interference. In [111], a continuously running
proach to study the technical conditions, relevant parame- real-time genetic algorithm is used, and in [112] a genetic
ters, and engineering implementation issues for handling programming network is selected to operate the cars at
and overstepping the cars in the shaft. Some approaches equal time intervals to obtain a better performance. This ge-
[106] are traffic sensitive and depend on the traffic pattern to netic network utilizes different setting parameters, the most
give preference to a set of different optimization criteria. relevant being the maximum predicted waiting time, the
number of passengers to be transported, the number of al- problem, which results in improved service quality. How-
located hall calls to the candidate car, and the predicted ever, EGCSs can perform other secondary tasks to improve
riding rate (number of passengers/car capacity) of the candi- dispatching, such as reducing the amount of uncertainty
date car when it arrives at the assigned hall call. by gathering proper data or by detecting traffic patterns,
and calibration for better algorithm performance. A trend
Dynamic Sectorization Management exists regarding the development of EGCSs that are able to
One of the most frequently employed and efficient meth- integrate and fulfill these entirely different tasks.
ods for improving the handling capacity of a vertical trans- Although the main aim of the research has been focused
port installation during severe up-peak traffic patterns is to on improving the quality of service, another aim also exists
divide the building into zones. In this way, each zone can to improve the quantity of service that is offered by im-
only be attended to by the elevator that is already assigned proving the definition of the boundaries in dynamic sector-
to it; see Figure 6. Roughly, the sectorization methods can ization during up-peak periods. Consequently, the elevators
be classified as static or dynamic when the boundaries of can carry more passengers in less time. The traffic detec-
the zones are fixed or flexible, respectively. Dynamic sec- tion function is not only relevant to reduce the amount of
torization performs better because it is capable of adapting uncertainty but helps the system to autotune its own pa-
to changes in traffic flow using, for example, artificial intel- rameters according to the traffic demand. This function is
ligence algorithms. generally easily identified in research proposals because
The efficiencies of the algorithms that define the vertical transport systems are designed to function under
dynamic sectorization are related to the performances of the maximum up-peak capacity. This design results in an
the traffic-detector systems. The problem can be solved oversized system for the remaining traffic patterns, which
using dynamic programming ([113] and [114]), and the can be advantageous for minimizing criteria other than the
authors of [115] designed an algorithm for the optimal defi- typical waiting time, such as power consumption. Howev-
nition of zones based on cellular automaton. In [116] and er, although the main objective in landing-call allocation is
[117], a control system is presented that (by means of an to reduce the AWT to provide optimum services for the
artificial immune algorithm) improves the handling capac- passengers, several articles have included the tenants’ pref-
ity of the vertical transport installations, minimizes the erences in the optimization process, such as giving service
round-trip time of each installation, and maintains a uni- to preferential floors or reducing power consumption.
form round-trip time (see [118] for background on artificial Regarding power consumption, although there is a lot of
immune algorithms). In [119] and [120], the optimization of interest in sustainable development, it has not been consid-
the zone definitions is suggested by using a particle swarm ered in many articles regarding vertical transport systems.
algorithm. For a wide background on particle swarm opti- The few references that have addressed the problem do so
mization algorithms and a guide to selecting algorithm using artificial intelligence heuristics. However, the com-
parameters for successful convergence, see [121]. putational load of each algorithm and the cost of the pro-
duction relative to the benefits of their implementation are
Relevant Literature Classification typically not addressed. Furthermore, no studies exist that
Tables 3–5 summarize and classify the most relevant litera- relate these factors for modern controllers or old-fashioned
ture discussed in this article. Moreover, the references are controllers from an economic viewpoint (those based on
classified according to their aims, methodology, and the simples logic rules). Another significant aspect in vertical
criteria used. transport research is the lack of benchmarks and the non-
existence of comparisons among the EGCSs that are sug-
Conclusions and Further Research gested in the different articles. This fact, together with the
This article contains a complete survey of the EGCS lack of data, such as the computational time or the environ-
research that has been undertaken in the past. These results mental conditions set for the simulations, complicate the
are represented and classified in Table 3. Furthermore, this task of properly evaluating all of the presented approaches.
article summarizes the literature related to EGCS tasks and The dynamic sectorization problem is typically solved on
notes the relationships among the different subjects in the the timescale of several seconds, so there is usually enough
vertical transport field. Consequently, this study can serve time to calculate and evaluate all of the possible solutions
as a starting point for new research. Currently, most before choosing the best one. In those cases, dynamic sector-
research has aimed to optimize the landing-call allocation ization problems can be solved with algebraic methods