Elevator Supervisory Group Controlsystems

(1)

DEGREE PROJECT, IN COMPUTER SCIENCE , FIRST LEVEL STOCKHOLM, SWEDEN 2015

Elevator Supervisory Group Control systems

ANALYSING THE PERFORMANCE OF

ELEVATOR ASSIGNMENT ALGORITHMS IN VARYING ENVIRONMENTS

CHRISTOFFER BLADES & FILIZ BOYRAZ

KTH ROYAL INSTITUTE OF TECHNOLOGY CSC SCHOOL

(2)

Elevator Supervisory Group Control systems

Analysing the performance of elevator assignment algorithms in varying environments

CHRISTOFFER BLADES & FILIZ BOYRAZ

Degree Project in Computer Science, DD143X Supervisor: Jens Lagergren

Examiner: Örjan Ekeberg

CSC, KTH May 2015

(3)

(4)

Abstract

Elevators are used every day and that puts expectations on how fast an elevator can handle passenger requests. This report aims to describe different elevator assignment algorithms for Elevator Supervisory Group Control (ESGC) systems and compare them to each other, with an emphasis on the delay time.

To research the field of ESGC systems four different algorithms were chosen and implemented: Simple control, Zone control, Collective control and Fuzzy control. The algorithms were also chosen in such a way that their strategies differentiated from each other to some degree. The algorithms were then tested under conditions similar to an office building with interfloor traffic. Five different test cases demonstrated the performance and behavior of the strategies in varying environments.

While Collective control proved to have a consistent performance, the more calculated Fuzzy control had the best statistical results. Zone control was the most affected by challenging conditions such as high traffic in the building. Simple control divided the workload equally, but did not make suitable elevator assignments.

(5)

Referat

Hissar används varje dag och det skapar förväntningar på hur snabbt en hiss kan hantera passagerarresor. Den här rapporten ämnar beskriva olika hisstilldelnings-algoritmer för Elevator Supervisory Group Control (ESGC) system och även jämföra dessa med varandra, med tyngdpunkt på fördröjningstiden.

För att undersöka ämnesområdet ESGC system valdes och imple- menterades fyra olika algoritmer: Simple control, Zone control, Collecti- ve control and Fuzzy control. Algoritmerna valdes ut på ett sätt, så att deras strategier skilde sig åt till en viss grad. De testades sedan i en miljö som efterliknar ett kontorshus med trafik mellan våningar. Fem olika testfall demonstrerade prestationer och beteenden hos strategier- na i varierande miljöer.

Collective control hade bevisligen konsekvent prestation, medan den kalkulerande Fuzzy control åstadkom de bästa statistiska resultaten. Zo- ne control påverkades i högst grad av utmanande förhållanden, såsom hög trafik i byggnaden. Simple control delade upp arbetsmängden jämnt mellan hissarna, men lyckades inte med att göra lämpliga hisstilldelning- ar.

(6)

Introduction

Elevators come in many variations. They are used for more than simply carrying people and not all are machine driven. There are those which can only be called without having the option of selecting the direction or those where you can select the exact destination floor when calling the elevator, moving either horizontally, vertically or in multiple directions.

The importance of having an optimal control strategy grows as the number of building floors increases, since it makes the difference between elevator control strategies more noticeable. Different strategies are suitable for different purposes.

The strategy selection can be based on, for example, optimizing the time or energy consumption. A good control strategy would satisfy several criteria that measure the service quality. Among these criteria passenger satisfaction is central.

The strategy should not be dependant on what it is carrying or how it is driven.

Neither is the direction it is moving significant. However, the number of directions is relevant. The information the strategy has access to and can include in the calculations is also of importance. Furthermore a building can have more than one working elevator, in which case the multiple elevators might be controlled either by their own separate control system or by a shared one.

1.1 Purpose

Elevator algorithms being a common research field has led to several papers written on the subject, among which two also are from research done as Bachelor Thesis studies at KTH (1; 4). A usual approach has been to optimize algorithms for elevator control or implement more advanced calculations in the controllers. This has been done using machine learning, AI or a very detailed set of rules. The focus is often put on changing the algorithms to suit the environment rather than finding what environment would suit the algorithm.

Much research has been done to solve a problem where the environment is already known. It is therefore more natural to adapt the solution for its specific use. However the conclusions drawn can not be used to get the overview which

(9)

CHAPTER 1. INTRODUCTION

is sometimes needed. One would like to know the result of control strategies in different environment settings and use it as a basis for an analysis of the algorithm characteristics. Also, the algorithms in use are combinations of several strategies which can be evaluated separately to get a better understanding of the differences between the strategies.

The aim of this report is to summarize the result from an examination of how well elevator controllers with different elevator assignment strategies perform and behave. The main goal is to make several elevators in an elevator shaft collaborate.

The environment is a simulated office building. This project is tested under the assumption that the most basic property of an office building is its constant interfloor traffic. Other aspects of the building such as its size, elevator number and passenger amount are kept interchangeable to draw conclusions of how the results change based on these aspects in the environment.

The conclusions made can be used to find weaknesses and improvement possibil- ities for each strategy. The question of whether the number of elevators in a building is of importance, or if there is a limit to how heavy traffic the controllers can handle will also be discussed. The estimation of how well the elevator strategy works will be based on the average delay time for all passengers, as well as some emphasis on the number of trips the strategies manages to complete. All included elevators will be of the vertical transportation type used for passenger transportation in an office building.

This problem is important and of interest because humans spend time travelling in elevators on a daily basis. To make this experience as painless as possible, the elevator has to use some kind of prioritizing algorithm when moving people between floors and, if there are several elevators, they have to communicate with each other.

Waiting a long time for the elevator to show up, waiting inside the elevator while it takes long detours or passing by the destination floor without stopping are the main things that should be avoided.

1.2 Used Concepts

Elevator request: An elevator request is sent to the elevator control when a passenger presses the call button in the desired direction from a building floor.

Elevator assignment: An elevator assignment is made by the elevator controller after a new request has arrived. The controller then selects the most suitable elevator to answer the request.

Destination request: A destination request is sent after the passenger has entered the elevator and chosen the desired destination floor.

Waiting time: The time the passenger spends outside of the elevator after an elevator request until it gets picked up by an elevator.

Service time: The time the passenger spends inside the elevator after it has been picked up by an elevator until it reaches the destination floor.

Optimal time: The time it takes for an elevator to travel between the start floor 2

(10)

1.2. USED CONCEPTS

and destination floor of a passenger.

Delay time: The summation of the waiting time and service time subtracted by the optimal time.

Elevator capacity: The number of passengers the elevator can carry.

(11)

(12)

Chapter 2

Background

The field of ESGC systems have been researched in many studies before and have examined a broad variation of similar problems. These previous studies shows that current elevator control strategies requires that a solution has to be given within less than half a second, otherwise the elevator will not be able to act upon a solution for example until it has passed a floor where the elevator was supposed to stop. A solution approach surpassing that time limit, no matter how well its performance is, would not be of any practical use. It has also been stated that the only way to calculate the average simulation time of a control strategy is through simulation and that the average waiting time is the factor that represents the elevator capacity the best since it is the factor which is affected the most in heavy traffic (7).

2.1 Elevator Dispatching Problem (EDP)

The Elevator Dispatching Problem is a stochastic and dynamic problem where you only know of starting set of request when the dispatch decisions is made. Possible future inputs may then change the planned route. These problems are typically solved by making a prior optimization of the initial state and then adapting the solution to later changes.

In a research done by Siikonen (5) a variation of this problem was given a solution where the algorithm would calculate not only were the elevator would be dispatched next, but also the complete route from the initial set of requests. The same algorithm was then used to generate new optimal decisions every time a new request appears. This was done as a lower level Travelling Salesman Problem for each elevator.

2.2 Strategies

There are a number of strategies an elevator control system can use to transfer passengers between floors. Research done on elevator group control systems has produced methods that use different approaches to this problem. Genetic algorithms

(13)

CHAPTER 2. BACKGROUND

use evolution to optimize its strategy and tries to make it better with each iteration.

It has a number of variables that it modifies after each successful iteration depending on the variables fitness.

Another approach is that the control system use a predefined algorithm that always handles elevator calls the same way. This type of algorithm may not be the most efficient approach but it is always consistent in the way it shuttles passengers.

A predefined algorithm has to figure out a clever strategy for calculating solutions based on the data it has access to. These algorithms will generally take the same approach to every problem instance. A simple strategy could be to circulate in the building and pick up passengers when they pass by a request or to divide the responsibility of all elevators between the building floors.

2.2.1 Strategies adapted to environmental conditions

If the control system is implemented for a known source the algorithm can be adapted to special circumstances that will exist in the environment it is used. For instance, if it is known that floor 3 has most of the incoming traffic, idle elevators can always return to floor 3. It will shorten the waiting time for the majority of the passengers. However it could also cause a longer waiting time for the few other passengers travelling from other floors.

Zone control is therefore an alternative to prefer if the building has multiple elevators. In Zone control the building floors are divided into zones of different sizes depending on the expected amount of traffic from each zone. The elevators are then assigned to one zone, and one zone might have more than one responsible elevators. The elevator can only pick up passengers from the floors inside the zone but is able to leave passengers at the correct destination floor even outside the zone.

This approach is proven not to be a suitable solution in up-peak traffic due to the elevator request being made from the same zone all the time (3).

2.2.2 Collective operation

The most traditional strategy is to use collective operation or duplex collective operation for two elevator shafts. It works by collecting requests in only one direction and when finished collect the requests in the other direction. Elevator systems with this kind of strategy have two buttons for the passenger to call the elevator by, one for each direction (3).

The problem with duplex Collective control is that if there is a lot of traffic in a short period of time, the elevators "bunch". That means they arrive at the same floor together and leave again thus acting as a single elevator instead.

2.2.3 Fuzzy Logic Group Controller (FLGC)

When simulating the outcome of an elevator strategy performance one also has to include factors of the humane behavior, how they move around in buildings and any possible delays caused by their actions. In problem approaches were the

6

(14)

2.3. PASSENGER FLOW

solution heavily depends on the predicted pattern of its passengers the accuracy and correctness is vital. But the predictions are seldom a good representation of real traffic flows which is why fuzzy logic is a good alternative. In fuzzy logic different solutions are given depending on the current traffic situation, the calculations are based on several conditions.

A variation of this approach was implemented in the study later described in the paper IEEE (6). The research further improved the performance of their strategy by letting the algorithm self-tune while executing. AI methods were used for that purpose but resulted in far too long process times. The performance was measured as the difference between the desired and the actual output of the strategy. However letting the the algorithm self-tune by selecting a fuzzy rule set proved to be successful and could be used for actual application. (6)

2.2.4 Evolution strategy

Another strategy for creating a controller that is adaptable to the current situation is the strategies which include evolving. The evolving genes of the strategy is selected in between executions to improve the performance in the next execution.

In one study that attempted the evolution strategy the simulation was done on destination call based elevators, meaning elevators where the passenger selects not only the moving direction but also the destination floor. The performance was measured as the number of passengers the elevator could transport within an hour with less than 30 seconds waiting time. To make sure that the selected genes of the algorithm actually were the better performing ones, an extra constant was added to the performance value in the comparison. For example the x genes are only considered better than the y genes if p(x) > p(y) + c (2).

2.3 Passenger flow

To be able to create a simulation the human factors used have to be accurate. Data, such as the passenger traffic could be collected through making the right measure- ment and gathering statistics. However it has also become a general assumption that the passenger arrival follows a Poisson process, somewhat depending on peak hours (5).

Passenger flow can be divided into 3 phases during a day. A typical office building has a time period with up-peak traffic where passengers arrive to the building through the entrance floor and travel to their work area. Later a time period with down-peak traffic when passengers go from their respective floors to the entrance to leave the building. The rest of the time interfloor traffic is more evenly distributed between all building floors (7).

(15)

(16)

Chapter 3

Problem definiton

Conclusions drawn from our examination are how the elevator assignment affects the elevator strategies performance and behavior. Comparing the results from the performances in five different settings, all different from one another in one attribute.

All environments are possible representations of an office building with interfloor traffic. Results from the investigation gives a detailed description of how different elevator control strategies behave and perform under situations where controllers are able to control several elevators in a shaft.

The alternating factors in the building environment are the number of building floors, density of the passenger arrival, number of elevators and elevator capacity.

The performance of the elevators are judged by the average passenger delay time.

(17)

(18)

Chapter 4

Method

Algorithms and strategies for solving the same or similar problems were selected from the literature study and then implemented in a program. Their performance were observed and the necessary statistics were obtained to make a comparison between all strategies. Each strategy has rules of prioritizing which elevator should be assigned for answering a request. For instance, a naive strategy could be to always pick up passengers in the same order that requests are made, only stopping for other passengers if a floor with a waiting passenger is passed, or to go in the direction from highest to the lowest floor requested. The varying factors in the simulation are how many floors the building has, the number of elevators, how often passengers are using the elevators and what the capacity of the elevator is.

Performance of the strategies was measured as the time they take to complete passenger requests from the point when an elevator request is made until the passenger reaches the destination. It is more appropriate to measure the time since the distance would not consider all possible delays. For example the number of stops the elevator has to make would not be noticeable when measuring the distance. The elevator will be assumed to always travel at a constant speed, but adding some time for the deceleration when stopping on a floor.

4.1 Implementation

The programming language used for the implementation is Java, it is a object- oriented language and the object-based approach suits the problem description.

4.1.1 Assumptions & Constraints

Number of passengers in the simulation is 20 and a new elevator request is sent once every 7th to 10th iteration. The simulation is limited to run for a constant 10000 iterations. The number of floors are always more than the passengers, between 40 and 70. The passengers are kept less than the floors in the building to avoid having a situation where every floor has a waiting passenger, since this would prevent the

(19)

CHAPTER 4. METHOD

different abilities of the strategies to be shown properly. The number of elevators in the shaft should be less than the number of passengers, between 2 and 4, otherwise the passengers could simply use one elevator each. Each elevator will have the capacity to transport 3 to 7 passengers at a time.

The values are chosen to get distinct results from each strategy. Having any smaller variation between the values would not be able to show a significant difference in the results. However, a larger variation would cause the strategies to have equally low performance.

When calling the elevator the passenger has to chose the direction they wish to travel in, up or down, and after entering an elevator the specific floor is chosen. The elevator has no knowledge of the number of passengers it is currently transporting or the number of people which wishes to get on or off the elevator on specific floors.

That information can not be included in the calculation of the control strategies.

Also the elevators can not change its movement direction in between floors or pass a passengers destination floor without stopping. Each passenger request should only be answered by one elevator to avoid unnecessary extra work.

The constraints are set with the passenger experience in mind. It is assumed in the project that passengers rather have a longer service time, time spent travelling in the elevator, than a long waiting time. Shortening the waiting time will thus be of greater importance when comparing the control strategies.

4.1.2 Passenger flow and time

The data needed for this report is mainly concerned with the flow of passengers in an office building and it was created by randomizing all passenger actions which is not necessarily realistic. The data was used for creating typical or edge situations where the elevators have to perform well.

The time the elevator takes to travel between two floors, when already in movement, will be measured as 1 time unit. Every stop the elevator makes will add another 3 time units for the elevator to decelerate before the stop and the accel- erate when it continues its trip. Depending on the number of people getting on or off the elevator at that floor another 1 to 2 time units will be added since it is relevant for the time that is needed. The total stop time might also be extended by 1 time unit if the doors are blocked by a passenger. The simulations for this project did not consider abnormal delays caused by passenger actions. More precisely the calculation of the stop time is the total integer value in equation 1.

[stopT ime() = 3 + RandomInt[0, 1] + (N umberOf P assengers ∗ 2)/CapacityOf Elevator]

(eq. 1)

The summation of the stop time will thus be in the interval [3,6].

4.1.3 The ElevatorControl class

In the simulation one of several control strategies can be used to decide on the elevator actions. All control strategies extends the same ElevatorControl class which

12

(20)

4.1. IMPLEMENTATION

operates the dispatching of each elevator.

The main methods in ElevatorControl consists of callElevator and nextAction.

callElevator is called for every new elevator request and adds incoming calls in a list deciding where the elevators will be dispatched next. While nextAction is called for every iteration in the simulation. The strategies are not strictly implemented by the definitions. Instead they have the biggest emphasis on the elevator assignments.

The callElevator method is therefore implemented separately for every strategy while nextAction is a shared method. Only Collective control has a slightly modified version of the nextAction method to better fit with the implementation of Collective control.

nextAction does a series of checks to decide on the elevator actions. If the elevator has stopped or if there are no request to be answered the elevator should not move. Else the elevators direction of movement has to be decided. The direction will generally stay the same as long as there are more requests ahead. But if there are many waiting destination requests, they will be of higher priority than the elevator requests.

4.1.4 Implemented strategies SimpleControl

Simple control divides all the elevator requests equally between the elevators in the shaft. A new incoming request will be assigned to the elevator with the least current requests. If there are several elevators with least amount of request, the first found one among them will be selected. Code 1 below describes how the elevator request is assigned to a specific elevator during the simulation.

Elevator min =

building.elevators[random.nextInt(Building.NUMBER_OF_ELEVATORS];

for (Elevator e : building.elevators) {

if ((min.elevatorRequest.size() + min.destinationRequest.size()) >

(e.elevatorRequest.size()+e.destinationRequest.size())) { min = e;

} }

min.elevatorRequest.add(startFloor);

(code 1) ZoneControl

The second implemented strategy is a zone-based algorithm. The implementation differs from the definition in the behavior that idle elevators do not return to the zones they are responsible of and the zones are not overlapping each other. There is only one elevator assigned to every zone and the building is divided into zones of equal size. The elevator requests are assigned to the elevator assigned to that zone.

(21)

CHAPTER 4. METHOD

Requests from other zones can only be handled by the elevator responsible for that zone, but elevators can drop off passengers in zones that are not their own. When the callElevator method is called the zone of the starting floor for the new passenger is calculated and then added to the request set of the corresponding elevator, as seen in code 2 below.

float i = startFloor / (numberOfFloors/numberOfElevators);

building.elevators[(int) i].elevatorRequest.add(startFloor);

(code 2)

CollectiveControl

A third elevator control attempts an approach on the collective control strategy. It differs the most from the rest of the implemented strategies because the elevator requests are not directly assigned to an elevator but instead stored in the controller as a collection. There are two different collections depending on the passengers direction, the request is added to a collection after the callElevator method is called in code 3. Elevators will only stop on floors with passengers waiting to travel in the same direction as the elevator. Each time the elevators are moved, the direction is checked to make sure that there still are waiting passengers in the direction the elevator is moving. If not, the direction is changed. Once a passenger has gotten on an elevator the destination request is made and added to a collection inside the elevator class.

if(direction == Passenger.UP) upCalls.add(startFloor);

else

downCalls.add(startFloor);

FuzzyControl

In the fourth strategy, Fuzzy control, the most suitable elevator is chosen for the elevator assignment of an incoming request. The suitability of an elevator is calculated by checking its current condition, such as how close the elevator is to the passenger, what stops are already planned, how many passengers it is carrying and if it is moving towards the passenger. A value is added for every condition that is met and the most suitable elevator can then be selected. Code 4 is from the calcSuitability method in Fuzzy control and is called once for every elevator when the callElevator method is called. The elevator with the highest sum value will be assigned to answer the request.

int sum = (Elevator.CAPACITY - elevator.destinationRequest.size())*8;

if (elevator.elevatorRequest.contains(startFloor) ||

elevator.destinationRequest.contains(startFloor)) { sum += 32;

14

(22)

4.2. INTEPRETATION OF DATA

}

if (isClose(elevator)) { sum += 32;

}

if (direction == Passenger.UP && elevator.direction == Elevator.UP &&

elevator.currentFloor < startFloor) { sum += 16;

} else if (elevator.direction == Elevator.DOWN &&

elevator.currentFloor > startFloor) { sum += 16;

}

if (elevator.direction == Elevator.WAIT) { sum += 8;

}

if (isNotBusy(elevator)) { sum += 8;

} return sum;

4.2 Intepretation of data

The performance of the strategies was compared foremost on the average delay time the elevators take to complete requests. But also on average waiting time as well as the number of passenger trips. Passengers might accept more time travelling in the elevator than having to wait for it to appear on their start floor. Therefore strategies should emphasize minimizing the waiting time for every passenger and be evaluated accordingly. In the report from IEEE (6) there was a claim that the time a passenger has to wait for the elevator to arrive can be perceived to be longer than the time they spend travelling in an elevator, even if that is in fact false.

The waiting time is therefore increased by a factor of 1.2 to better represent the experienced delay time.

delayT ime() = (1.2 ∗ waitingT ime + serviceT ime − optimalT ime (eq. 2)

The total delay time of the elevator is calculated as a weighted sum between the waiting and service time as shown in equation 2. The result can be translated to real time by assuming that the elevator speed will be adapted to the building size, elevators should be able to travel through the building in 30 seconds. For instance 40 time units will be 30 seconds in a building with 40 floors (8).

The statistics from the simulation is collected while the program is executing.

The number of trips is increased for every passenger which is picked up by an elevator, at the same time is the waiting time for that passenger stored. The number of floors a passenger travels during a trip as well as the service time is added only after the passenger has reached its destination.

(23)

(24)

Chapter 5

Results

The elevator strategies were tested on their performance in five different environments. The first environment being a basic case with intermediate difficulty for the elevator strategies, where they are expected to perform well. The other five cases are the same as the first case where only one variable has been altered.

The environments were chosen with the purpose of getting results from varying situations. Higher or lower traffic with fewer or more elevators which will sometimes forces the elevators to travel long distances to pick up passengers and other times the amount of passengers will result in many stops for the elevators. This is meant to push the limits of the strategies and get concrete statistics of how the changes in the environment affects the performance and behavior of the strategies.

5.1 Test cases

5.1.1 Case 1

Floors: 40 Elevators: 2

Elevator capacity: 7 Passenger density: 1/10

(25)

CHAPTER 5. RESULTS

Test results for each strategy from the simulation with the basic test values are shown in fig 1.

5.1.2 Case 2

Floors: 70 Elevators: 2

Elevator capacity: 7 Passenger density: 1/10

18

(26)

5.1. TEST CASES

Test results for each strategy from the simulation with taller building are shown in fig 2.

5.1.3 Case 3

Floors: 40 Elevators: 4 Elevator capacity: 7 Passenger density: 1/10

(27)

CHAPTER 5. RESULTS

Test results for each strategy from the simulation with more elevators are shown in fig 3.

5.1.4 Case 4

Elevator capacity: 3 Passenger density: 1/10

20

(28)

5.1. TEST CASES

Test results for each strategy from the simulation with less elevator capacity are shown in fig 4.

5.1.5 Case 5

Elevator capacity: 7 Passenger density: 1/7

(29)

CHAPTER 5. RESULTS

Test results for each strategy from the simulation with higher traffic are shown in fig 5.

22

(30)

5.2. NR OF COMPLETED TRIPS FOR EACH CASE

5.2 Nr of completed trips for each case

The number of trips for each strategy and test case are shown in fig 6.

(31)

(32)

Chapter 6

Discussion

All elevator control strategies except Collective control could pass by a waiting passenger if the elevator has not been assigned to stop at that floor. On the other hand the probability of bunching is rather low for all strategies except Collective control. In the other strategies there is always one elevator which is assigned to a request and stops at a certain floor. Only if two requests are made from the same floor in different directions two elevators could be called to that floor. In the case of Collective control, an elevator will not stop at the same floor as another, unless there is a destination request to that floor, but they could be moving in the same directions acting practically as the same elevator.

No elevator control strategies except Fuzzy control and Collective control includes the information of the passengers’ direction of movement when assigning elevators. Also no elevator control strategies except Fuzzy control and Simple control includes the information of the the number of requests it is currently handling when assigning elevators.

6.1 Evaluation of the method

Since only a limited part, the elevator assignment, of the algorithm was to be examined it was more convenient to implement new versions of each strategy. Also all algorithms had to be adapted to the elevator control class. Isolating the testing to focus on only the elevator assignment made the results clearer and easier to connect to each strategy.

6.1.1 Passenger flow

The strategies have only been tested under interfloor traffic in the simulation. It is meant to be seen as a cut out from a day and does therefore not have a beginning or end. The passengers are assumed to already be inside the building and the elevators are spread out on all building floors. At the start of the simulation however there are no current passengers travelling. An option, to make the simulation better

(33)

CHAPTER 6. DISCUSSION

represent a cut out time period from an ordinary day, would be to have several requests placed out at the beginning of the simulation.

The simulation could also have been made more realistic by gathering existing data sources for information concerning passenger flow. Such data should already be available in previous studies. Also the possibility of an elevator going out of service might be included in the simulation to create more difficult scenarios for the elevator strategies. But the effects of having more or less elevators are tested in case 3. The results from case 3 can be used for drawing conclusions on what consequences an elevator out of order would have.

6.1.2 Passenger behaviour

The construction of the simulation makes it improbable for groups of passengers to travel together. Every passenger plans new elevator trips separately without influence of how other passengers travel. Another way to create the simulation could have been to add a numberOfPeople property to every passenger object, the passenger object could then have been either a single passenger travelling alone or a group of people.

The simulation also assumes that passengers always wait for an elevator to arrive no matter how long they have waited and board an elevator that has stopped if there is room left, without considering the direction the elevator is travelling in.

This could mean that two passengers, travelling in different directions from the same floor, could send elevator requests to two different elevators but still end up both boarding the first elevator that shows up on that floor which is not optimal. The second elevator to show up will then not have any passengers to pick up on that floor. It could be a realistic behavior if the passenger do not mind taking a small detour to avoid waiting a longer time for another elevator to arrive. But it also seems unrealistic that a passenger only travelling a few floors down would consider traveling a long distance up before reaching its destination.

6.1.3 Simulation construction

Decisions about the simulation construction that has to be made includes, for example, the question if an elevator that passes by a building floor with a waiting passenger always should stop even if the request has not been made in the same direction the elevator is travelling in. Neither was the length of the simulation obvious, one could either chose to let the simulation run for a fixed number of iterations or give it a number of task and let the simulation run until all task are completed.

Another aspect was to decide when and how the elevator control should make the calculation of the elevator movements. Either a new calculation is made for every step the elevator takes, only when a new elevator request or destination request is made or every time the elevator stops on a floor. The calculations made being either where the elevator should be dispatched next or a whole route.

26

(34)

6.2. OBSERVATION

In earlier testing an unwanted behavior of elevators keep their direction towards new elevator requests even when they are full were noticed. To make the strategies work more efficiently the elevators should start moving towards destination requests if they already have many waiting requests. An elevator can not know the number of passengers it is carrying; however the number of destination requests give an indication of the number of people the elevator is carrying. The destination requests are only as many as there are requests made from inside the elevator. If the amount surpasses the elevator capacity, and there are not any more destination floors ahead in the current moving direction, the elevator can change its direction.

Answering destination requests before picking up any new passengers would improve the performance. Since all elevators keep track of elevator and destination requests separately the information can be used in the controllers. Another solution of putting new requests in a queue to delay the elevator assignment if all elevators are busy was discussed but not implemented.

6.2 Observation

6.2.1 Simple control

Simple control is clearly dividing the workload between the elevators equally, but fail in making well calculated choices otherwise. It becomes apparent in the simulation that the elevators selected in the elevator assignment are not always the most suitable. Such as an elevator far away from a request being selected in the elevator assignment.

6.2.2 Zone control

Zone control has the most unbalanced division of workload between the elevators.

At the final state of the simulation it is common that one elevator is overloaded with requests while others are idle.

The problem also causes long waiting times which occurs when one elevator get more requests than it has the capacity to carry. When the elevator travels through the zone to pick up all passengers some will be left waiting not being able to board the elevator. Once the elevator has taken another turn in the building to drop off passengers and return to the assigned zone, it will again be completely filled up.

This makes it difficult for all requests to be answered.

6.2.3 Collective control

What could be observed when the simulation was running was that Collective control very often calls on more elevators than necessary. At the beginning of the simulation, when the first passenger arrives, all elevators start to move towards the request and the passenger then gets picked up by which ever reaches the passenger first. Another noticeable pattern was that the elevators often start moving in the

(35)

same direction, and since the strategy requires them to keep moving in the same direction until the furthest passenger is picked up, other passengers could be kept waiting.

6.2.4 Fuzzy control

The behavior of Fuzzy control is the most situation adapted. Selecting the elevator for an request by several criteria rather than only one. The observations from the simulation shows that the elevator assignment is not always the most obvious choice.

However in the long run the performance is acceptable.

At the end of the simulation with Fuzzy control there is usually very few passengers left in the building, however the elevators are still carrying many of the passengers. Since Fuzzy control does the elevator assignment directly when a request is sent and assigns the, at the time, most suitable elevator, most of the passengers will be located where the elevator soon can pick them up. This would results in shorter waiting time.

6.3 Statistical results

All test cases where simulated for the same number of iterations, 10000. The iteration limit results in a maximum of 1000 trips for the test cases where the passenger density is 1/10 new passengers every iteration. But considering a building size of 40 floors it is only possible for the elevators to take 125 complete tours in the building.

The number of completed passenger trips will be used as a reference of the results reliability. For example a short waiting time is not of importance if the algorithm has not been able to complete most of the passenger trips.

The worst possible waiting time would occur when an elevator has to take a whole tour in the building, 40 floors up and 40 floors down, with additional time for every stop it makes along the way. The waiting time can add up to about 120.

The aim is to have a low average delay time. But since the average waiting time and service time are the addends for the average delay time, the focus in the discussion will be on the average waiting time and average service time.

6.3.1 Case 1

Case 1 was the basic test. The performance was more or less the same for all strategies and the statistical results as expected. Case 1 will be used in comparison with the other test cases.

6.3.2 Case 2

Case 2 tested the performance of each strategy in a larger building which resulted in longer waiting time as well as service time, which is a direct consequence of the

28

(36)

6.3. STATISTICAL RESULTS

building size. When the building size is increased the average number of floors per trip is increased.

However the results for Zone control are far worse than the rest of the strategies and shows the weakness of this strategy. The elevators travel longer distances but once the elevator leaves the zone it can not pick up any new passengers, which makes the strategy inefficient since it can not do any work outside the zone.

6.3.3 Case 3

In case 3 the number of elevators was increased and the strategies clearly have a lower average waiting time and average service time. More elevators entails that passengers can travel in separate elevators and the number of stops is decreased.

The increase in elevators also help in keeping a larger area of the building covered and there will often be an elevator close by when a new call arrives. The effect of this results in improved waiting times, even though the worst possible waiting time remains unchanged.

The strategy to take least advantage of the number of elevators is Zone control since every request still only can be answered by the right elevator responsible for that zone. However the zone each elevator has to cover is smaller.

6.3.4 Case 4

In the fourth test case the elevators had lower elevator capacity. The waiting time is affected by the change while the service time remains unchanged. Since the elevators often will be full they can not pick up any new passengers even if the elevator stops at the right floor. This causes the increased waiting time when passengers have to call on a new elevator if the first one to arrive was full. The service time was not affected because of that the number of calls made to the elevators is not changed.

The elevators stop at as many floors as in the first basic test case.

The performance of Zone control and Fuzzy control, plummets in case 4. Lim- iting the elevator capacity clearly has a profound effect on the ability to finish trips for the affected strategies. Zone control is made inefficient when each elevator only can carry a small amount of passengers. Fuzzy control does include a value for how many calls each elevator has in the calculations, however it is not as heavily regarded as other criteria, making Fuzzy control less suited in this test case.

Simple control is separated from the other strategies with shorter waiting time and service time. When the elevator capacity is constricted the most efficient strategy is to divide the workload like in Simple control to avoid overloading elevators.

6.3.5 Case 5

Case 5 tested the strategies performance in high traffic. It had a negative impact on both the waiting time and service time. With the elevators having a higher workload each, the elevators were delayed in picking up waiting passengers as well as leaving

(37)

the passengers at their destination floor. The increased density of passengers caused more stops.

Zone control struggles with completing passenger trips for similar reasons as in case 2.

30

(38)

Chapter 7

Conclusion

The tested elevator assignment strategies each have one main priority. Simple control divides the requests equally. Zone control clusters the request in zones. Collec- tive control picks up passengers in the order they are passed by an elevator. Fuzzy control is a combination of the others, being able to adapt to the current situation.

7.1 Simple control

Simple control is the most reliable algorithm overall when it comes to finishing trips, according to the results. The Simple control divides the elevator requests based on each elevators total current requests, which sometimes leads to a situation where an elevator that is far away receives the request. This is not an optimal situation and that is one problem that decreases the chances of finishing trips at an acceptable time.

Another possibility is for the requests the elevators receives to be spread out in the building, thus forcing the elevators to travel long distances in vain. It would obviously be more efficient if the requests were clustered and assigned to the same elevator like in Zone control.

7.2 Zone control

Zone control is the strategy that is the most affected by changes in the environment, as seen in cases 2, 4 and 5. The biggest problem is that the number of trips are greatly decreased and the strategy becomes inefficient.

The results from Zone control are also inconsistent and highly affected by the randomized starting position of each passenger. Many passengers from one zone often result in a long waiting time. It becomes noticeable in large differences between average waiting time and worst waiting time which can be observed from the data tables attached in the appendix.

(39)

CHAPTER 7. CONCLUSION

7.3 Collective control

The greatest disadvantage is that all elevators often move towards the same request, but the results from this strategy are very consistent and the same for all passengers.

As seen in the result diagrams, Collective control has the worst average times in most of the cases, but still manages to have a higher average in finishing trips than Zone control and Fuzzy control. This shows that Collective control is a reliable strategy.

7.4 Fuzzy control

Fuzzy control has the statistically strongest performance, but is also the strategy with the highest potential for improvement. The criteria, which are checked in the algorithm, are changeable and adding or removing criteria helps the strategy make better choices.

32

(40)

Appendix A

Data tables corresponding to all test cases

A.1 Case 1

A.2 Case 2

A.3 Case 3

A.4 Case 4

A.5 Case 5

(41)

(42)

(43)

APPENDIX B. FINAL STATES IN CASE 5 FROM THE SIMULATION

Appendix B

Final states in Case 5 from the simulation

B.1 Simple control

36

(44)

B.2. ZONE CONTROL

B.2 Zone control

(45)

APPENDIX B. FINAL STATES IN CASE 5 FROM THE SIMULATION

B.3 Collective control

38

(46)

B.4. FUZZY CONTROL

B.4 Fuzzy control

(47)

(48)

Bibliography

[1] J. Axelsson S. Bernelind (2013), Elevator Control Strategies. Royal Institute of Technology.

[2] T. Beielstein, C.-P. Ewald S. Markon (2003), Optimal Elevator Group Control by Evolution Strategies. Universtität Dortmund.

[3] R. Caporale, G. Strakosch (2010), Vertical Transportation Handbook, 4th Edi- tion. John Wiley Sons.

[4] F. Ceder A. Nordin (2013), Elevator Control Strategies -Simulating different algorithms to find the most efficient strategy. Royal Institute of Technology.

[5] H. Ehtamo, M.-L. Siikonen, J. S. Sorsa, T. Tyni, J. Ylinen (2009), The Elevator dispatching problem. Transportation Science.

[6] W. P. Hew, J. Jamaludin N. A. Rahim (2010), An Elevator Group Control System With a Self-Tuning Fuzzy Logic Group Controller. IEEE Transactions on industrial electronics, Vol. 57, No. 12.

[7] M.-L. Siikonen (1993), Elevator Traffic Simulation. KONE Elevators Research Center.

[8] Toshiba elevator and building corporation (2015), Traffic Planning.[

url: http://www.toshiba-elevator.co.jp/elv/infoeng/products/pd08.jsp] (ac- cessed: 2015-05-08).

(49)

www.kth.se

Elevator Supervisory Group Controlsystems

Elevator Supervisory Group Control systems

ANALYSING THE PERFORMANCE OF

ELEVATOR ASSIGNMENT ALGORITHMS IN VARYING ENVIRONMENTS

Elevator Supervisory Group Control systems

Abstract

Referat

Contents

Chapter 1

Introduction

1.1 Purpose

1.2 Used Concepts

Chapter 2

Background

2.1 Elevator Dispatching Problem (EDP)

2.2 Strategies

2.3 Passenger flow

Chapter 3

Problem definiton

Chapter 4

Method

4.1 Implementation

4.2 Intepretation of data

Chapter 5

Results

5.1 Test cases

5.2 Nr of completed trips for each case

Chapter 6

Discussion

6.1 Evaluation of the method

6.2 Observation

6.3 Statistical results

Chapter 7

Conclusion

7.1 Simple control

7.2 Zone control

7.3 Collective control

7.4 Fuzzy control

Appendix A

Data tables corresponding to all test cases

A.1 Case 1

A.2 Case 2

A.3 Case 3

A.4 Case 4

A.5 Case 5

Appendix B

Final states in Case 5 from the simulation

B.1 Simple control

B.2 Zone control

B.3 Collective control

B.4 Fuzzy control

Bibliography