Simulation and Model Sensitivity Analisys of a Wind Turbine Tower

Manufacturing Plant

Paulo Tomé

, Eduardo Teixeira

, Freddy Assunção

, Luís Marques

João C. P. Reis

and João M. C. Sousa

TEGOPI - Industria Metalomecânica, S.A., Vilar do Paraíso, Portugal

LAETA, IDMEC, Instituto Superior Técnico, Universidade de Lisboa, Lisboa, Portugal

Keywords: Wind Turbine, Manufacturing, Industrial Process.

Abstract: A modelling method for a complex wind turbine tower manufacturing plant is proposed, through the

specification of the major assumptions in the model. Using this methodology a DES model was developed,

and a sensitivity analysis to some of the main process variables accounted for in the model is presented.

From this study several versions of the model were developed, and their results are compared against real

data from a manufacturing company.

1 INTRODUCTION

With early applications for water pumping and

milling, the use of wind energy dates back to the 2nd

century B.C. (Brito, 2014) with the first wind mills.

In recent years, having developed greatly trough the

70’s as consequence of the oil crisis (Hemami,

2012), wind energy plays an important role in

today’s society. In Portugal, according to REN

(Redes Energéticas Nacionais, 2013), at the end of

2012 almost 23% of the connected power was

obtained using wind. Nowadays, most wind towers

in use can be described by the four major

components: tower, rotor (hub and three blades),

nacelle and foundation (Hemami, 2012). In this

paper the manufacturing process of a tubular

metallic tower is considered and modelled using

DES.

2 MODELLING

2.1 Simulation Methods

According to McHaney (McHaney, 2009), computer

simulation can be broadly defined as:

“Use of a computer to imitate operations of a real

world process or facility according to appropriately

developed assumptions taking the form of logical,

statistical, or mathematical relationships which are

developed into a model.”

Figure 1.1: Typical modern wind tower with tubular

metallic tower (Hemami, 2012).

A model is defined as the representation of a

system with the purpose of studying it. Only aspects

that influence the problem in study must be

considered and then represented in the model. By

definition the model is a simplification of the real

system (Banks, et al., Fourth Edition).

It is possible to classify computational simulation

according to three categories: Monte Carlo

Simulation (MCS), Continuous Simulation (CS) and

Discrete Events Simulation (DES). MCS uses

generation of random numbers to simulate, without

considering time explicitly as a variable. This

method of simulation is defined by Law and Kelton

(Law & Kelton, 2000) as being “a scheme using

random numbers that are used to solve deterministic

or stochastic problems where time plays no role”.

CS models use a set of equations in representation of

725

Tomé P., Teixeira E., Assunção F., Marques L., C.P. Reis J. and M.C. Sousa J..

Simulation and Model Sensitivity Analisys of a Wind Turbine Tower Manufacturing Plant.

DOI: 10.5220/0005040807250732

In Proceedings of the 4th International Conference on Simulation and Modeling Methodologies, Technologies and Applications (SIMULTECH-2014),

pages 725-732

ISBN: 978-989-758-038-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

a system evolution trough time. DES is

characterized by blocks of time when nothing

happens until an event changes the state of the

system.

For this study, DES methodology is considered.

The definition of important concepts as system,

system state, entity, attributes, events and others can

be found for instance in references (Banks, et al.,

Fourth Edition) and (Altiok and Melamed, 2007).

2.2 Case Study

The desired model is required to simulate to a

certain degree of accuracy the resource allocation

and production output of part of a Portuguese wind

turbine tower manufacturer, in order to help the

company evaluating the effect of altering key

variables in production time.

According to the company, a tower is produced

from smaller elements (in Portuguese named:

virolas) welded to each other in order to become a

section of a tower (in Portuguese named: tramos).

The sections are then assembled on sight for

obtaining the complete tower. A simplification of

the process flow is given in figure 2.2.

Figure 2.1.a: Virolas. Figure 2.1.b: Tramos.

Figure 2.2: Production flow.

The considered entities (Virolas) are assumed to

have an unlimited stock. Therefore the operations of

“Preparation” and “Assemble Virola” do not

interfere with the production and for this reason they

are not modelled.

A tower built from 34 Virola elements welded to

form 5 Tramos sections, according to figure 2.3 is

considered.

The production of a subsection consists of the

operations of transportation of the elements to the

production line, welding, inspection, assembling of

accessories and other operations. For each

production line, depending on the section in

production, around 65 to 90 different operations can

be executed in order to manufacture that section.

Figure 2.3: Sections of the considered tower (TEGOPI).

As an example, the simplified flow diagram for

the lower section S5 is presented in figure 2.4. The

simplification is due to some operations being

grouped, as is the case of “welding” that represents a

sequence of six different welding operations.

Since each process requires different machine

and worker resources, in order to be able to evaluate

the impact of changing the availability of any of

them in the model, the modelled system features:

• 10 production lines

• 20 vehicles

• 26 different resources

• 12 worker profiles

• Around 500 operations

2.3 Model Assumptions

With the purpose of modelling the studied system in

an objective way, a set of assumptions was defined,

such as:

i. The production orders are to be specified for

the sections (Tramos)

ii. The created entities are transported to the

production line by vehicles and workers

simultaneously

iii. The vehicles mentioned in ii travel in specific

networks and have a limit for the maximum

transporting weight

Prepar-

ation

Virola

Assembl

e Virola

Assemble

Tramo

Weldin

Tramo

Inspection

Tramo

Finish

SIMULTECH2014-4thInternationalConferenceonSimulationandModelingMethodologies,Technologiesand

Applications

726

Figure 2.4: Simplified flowchart of the manufacturing process of section S5.

iv. All vehicles are “driven” by two workers,

where the profile of the worker is depending on the

entity type to transport.

v. At the production line the entities are

processed according to the array of operations

specified by the company

vi. Whenever a shift is interrupted or finished,

the workers assigned to the running operations are

released

vii. At the beginning of a shift where pending

operations exist, those are resumed according to the

priority associated to the respective entities.

In addition, two specific production lines are

assigned to the production of each specific section

due to weight constraints of the vehicles.

Production orders are given for the sections by

specifying a start date and a destination production

line. As soon as one element reaches the assembly

stand of one production line, the next element of the

production sequence is created and sent to this same

production line. Initial element priorities can be

specified for each entity and destination production

lines.

If an order for the production of a section is

given and the destination production line is busy, the

model will create the entity, but this entity will not

be allowed to leave the stock area until the

destination line finishes the work in progress. Then

becoming available to processes the new ordered

section.

The model takes transportation times according

to the distances from stock to production line

locations into consideration. After completing the

transportation tasks, the vehicles stay idle at the

unload station. For accepting a transportation request

a condition is evaluated. This condition specifies

that the chosen vehicle is one of the available ones

and has a load capacity that is enough to carry the

element, but with the lowest possible carrying

capacity.

All operations are performed by at least one

worker sometimes using resources to perform the

task.

Having had access to the average times of the

different operations only but knowing that this times

are not constant, a time distribution is used to model

this behaviour. According to reference (Altiok &

Melamed, 2007), a triangular distribution can be

used when the actual time distribution is unknown

but it is reasonable to assume that minimum (a) and

maximum (b) possible values exist and that the most

likely value (c) is inside this interval. Triangular

distributions are defined by the following function:























2



, 





2







0, 

(1)

In the model the average value provided by the

company is considered to be the most likely value

(c). The minimum and maximum values (a and b)

are obtained applying a percentage deviation around

c. The model considers 5% deviation around c, but it

is possible to change this rule easily by using the

spread sheets attached to the model.

Multiple profiles of workers are defined (12 in

total) according to the company’s specifications.

These profiles are not flexible to perform a task from

another profile. The same applies for resources. This

implies that an interrupted operation can only be

resumed when the needed worker profile(s) and

resource(s) are available to attend this task. Workers

and resources “capacities”, costs and reliability logic

SimulationandModelSensitivityAnalisysofaWindTurbineTowerManufacturingPlant

727

(the last only for resources) are able to be feed easily

and changed into the model through the attached

spread sheets. For the later purpose of model

validation reliability logic is not considered since it

is not mentioned in the supplied data.

3 RESULTS ANASLYSIS

According to Kelton (Kelton, Smith, Sturrock,

2013), only for the simplest models one is able to

prove categorically the aspects of correct modelling.

Therefore, the model should be tested in a way that

is possible to verify that it behaves in the desired

manner. First the model sensitivity to changing key

variables is tested and analysed. Latter, data

provided from the company is used to evaluate

model performance. Multiple versions of the model

are used to verify this point.

3.1 Sensitivity Analysis

Since transportation of elements from the stock

location to the production line takes a considerable

amount of time, figure 3.1 shows the production

time using calculated distance (Model*1) or an

smaller average distance (Model*2) for the

production of section S5 in production line 8.

As expected, the net time needed for the

production of this section decreases when average

distances are considered. A decrease in occupation

times of both workers (PF1 and PF8) and vehicles

can be observed in figure 3.2 with a greater level of

detail.

Figure 3.1: Production times: section S5 figuin line 8.

In fact, when the complete model is considered

and production orders that lead to the manufacture

of 4 towers are given, Model*2 needs about 3 days

less of labour for finishing the production.

Figure 3.2: Resources occupation: section S5 in line 8.

The order in which operations are executed is

related to the priority given to the entities (Virola) in

processing. Once that the number of resources is

limited, situations can occur where not enough

resources are available to attend to all requested

operations. Therefore, the priority assigned to

entities plays an import role in the system and by

changing this parameter model results should be

affected. According to the company TEGOPI -

Industria Metalomecânica, S.A., Portugal, the

priority is defined to be highest for entities of section

S5 followed by the entities of sections S1, S4, S3

and therefore, S2 entities were given the lowest

priorities of the set. This configuration is used in

both Model*1 and Model*2.

A new set of priorities was defined for Model*3

by keeping the relative priority of the different

sections but increasing priority of the sections that

will form the two first towers. This way assuring

that priority is given to an older production order

and by doing so a better approximation to the real

system should be achieved. A summary of due dates

obtained with the models considered up to this point

can be observed in table 3.1.

Table 3.1: Due dates summary.

Due date

Model*1 Model*2 Model*3

tower 8-May-13 3-May-13 30-Apr-13

In the construction of the model, special attention

was placed into the allocation of workers to

operations. Since specific profiles of workers are

used to perform operations on the job-shop using

resources of some kind, varying the available

number of workers and their profiles should have a

major effect on the delivery dates. This effect can be

832

301

519

752

245

508

150

350

550

750

950

Total time Stopage

time

Net time

Time (hours)

Model*1 Model *2

242

111

21,0

236

105

9,6

100

150

200

250

300

PF1 PF8 Vehicles

Time (hours)

Model*1 Model *2

SIMULTECH2014-4thInternationalConferenceonSimulationandModelingMethodologies,Technologiesand

Applications

728

observed through analysis of the results obtained

with Model*4. Model*4 is built from Model*3 but

setting an unlimited number of both workers and

resources.

Table 3.2: Limited VS Unlimited resources.

Due date

Model*3 Model*4

1st tower 28-Feb-13 15-Feb-13

2nd tower 4-Apr-13 25-Feb-13

3rd tower 10-Apr-13 21-Mar-13

4th tower 30-Apr-13 26-Mar-13

As expected, the model reacted to the increase

of resources by reducing significantly the delivery

times. Note that for obtaining the results above the

number of workers used is much higher than the

ones scheduled by the company. Figure 3.3 and

figure 3.4 show the comparison of used resources.

Figure 3.3: Workforce: TEGOPI VS Model*4.

3.2 Model Validation

Validation of the model is achieved by comparing

delivery times of the different scenarios with data

provided from the company.

3.2.1 Single Production Line

With knowledge of the average processing times of

each operation to be performed in each section, it is

possible to obtain an estimated average time for the

total delivery time of each section. For this scenario

unlimited resources were considered so that delays

resulting from lack of resources do not occur.

Figure 3.4: Resources TEGOPI VS Model*4.

Figure 3.5 to figure 3.7 show the occupation of the

resources to manufacture section 5S for the cases:

Validation scenario and Model*2. Model*2 is

chosen because it considers average distances from

stock to production, thus better matching the

available data for validation.

Figure 3.5: Machine occupation: section S5.

Deviations between 0.2% and 2.1% when compared

with the validation data were obtained for machine

occupation. Workers utilization varies up to 4.8%.

Regarding transportation times there is a 15.2%

difference between the validation data and Model*2.

In case Model*1 was considered, around 21 hours of

total transportation time would be needed.

1st Shift 2nd Shift 3rd Shift

Number of Workers

Tegopi Model*4

Arc

ArcoSubmerso

Carro chanfrar

EasyLaser

Lix2

Lix4

Macarico

Rebarbadeira

RF1

Semi

UltraSom

SO213

SO257

SO283

SO294

SO296

Tegopi Modelo*4

100

ARC.

Arco-Submerso

Carro chanfrar

Easy-laser

Maçarico

Rebarbadeira

Semi

SO213

SO294

Ultra-sons

VS - Linha8

Time (hous)

Validation Scenario Model *2

SimulationandModelSensitivityAnalisysofaWindTurbineTowerManufacturingPlant

729

Figure 3.6: Workers occupation: Section S5.

Figure 3.7: Vehicles occupation: section S5.

Figure 3.8 compares the net production time used.

Figure 3.8: Production time: section S5.

3.2.2 Validation Scenario

Having already verified model sensibility to key

parameters, the considered models can now be

compared to the expected due dates from the

validation data.

Although the model results may seem to be poor

some considerations must be taken into account:

 According to the supplied data there are shifts

where worker profiles are not available (e.g.: PF6 is

not available on the third shift).

Table 3.3: Due dates: Multiple scenarios.

Validation Data Due Date

Production

Order

Due Date Model*1 Model*2 Model*3

1st tower 11-Jan-13 13-Feb-13 11-Mar-13 26-Feb-13 28-Feb-13

2nd

tower

18-Jan-13 22-Feb-13 22-Mar-13 21-Mar-13 4-Apr-13

3rd tower 23-Jan-13 15-Mar-13 29-Apr-13 1-May-13 10-Apr-13

4th tower 28-Jan-13 19-Mar-13 8-May-13 3-May-13 30-Apr-13

Meanwhile, on the data for validation this

situation appears not to occur since there is no

indication of delayed operations as result of

missing workers.

 In simulation it often occurs, that after the end

of a shift the next shift does not have enough

resources to resume all interrupted operations.

Priority operations are resumed first. This can

lead to delay due to lack of resources in low

priority operations.

 Model does not consider workers to be flexible.

Take for instance an operation that uses the worker

profile above (PF6) and that is interrupted in the end

of the 2

shift. This operation can only be resumed

after one shift of interruption, affecting the

production time due to this delay.

3.2.3 Unlimited Resources

A version of the model with unlimited resources and

average paths (Model*4) is used. This allows the

model to get closer to the data for validation,

according to table 3.4.

For the first and second towers results are

matching the expected since the production lines

start free and times are not affected by delay. For the

next towers delay occurs since the production lines

are occupied when the production order reaches the

system.

Table 3.4: Due dates: Model*4.

Validation Data

Due date

Model*4

Production

Order

Due Date

tower 11-Jan-13 13-Feb-13 15-Feb-13

tower 18-Jan-13 22-Feb-13 25-Feb-13

tower 23-Jan-13 15-Mar-13 21-Mar-13

tower 28-Jan-13 19-Mar-13 26-Mar-13

100

150

200

250

300

PF1

PF2

PF3

PF4

PF5

PF6

PF7

PF8

PF9

PF10

PF11

PF12

Time (hours)

Validation Scenario Model *2

Total Time

Time (hours)

Validation

Scenario

Model *2

150

250

350

450

550

650

750

850

Total time Stopage

time

Net time

Time (hours)

Validation Scenario Modelo*2

SIMULTECH2014-4thInternationalConferenceonSimulationandModelingMethodologies,Technologiesand

Applications

730

Table 3.5: Due dates tower 3 and 4: Model*4.

Validation Data

Due date

Model*4

Production

Order

Due Date

tower 23-Jan-13 26-Feb-13 27-Feb-13

tower 28-Jan-13 4-Mar-13 5-Mar-13

To eliminate this effect, a scenario for the

production of towers three and four only was

simulated. Results in table 3.5 show that the due

dates in this situation are inside the acceptable limits

for the model.

In order to achieve these results, Model*4 used

145 workers. The company actually scheduled 84

workers. A situation in which the number of workers

used is higher than the number scheduled is verified.

Figure 3.9: PF5 1st shift usage: Model*4.

The number of workers PF5 of the 1

shift

(PF5_t1) is chosen because it represents the bigger

deviation.

Four workers are scheduled by the company and

ten were used by Model*4. Figure 3.9 shows

detailed statistics regarding the workers of PF5_t1

used by Model*4. Hence profiles PF5_t1[7] to

PF5_t1[10] are used for less than a eight hours shift,

results appear to show that the model lacks some

logic for task selection preventing that, for example,

a task with more than eight hours duration starts

close to the end of a shift.

In a situation when more workers are available, a

larger number of machine resources is used

according to figure 3.10.

Figure 3.10: Allocated machines: Model*4.

3.2.4 Tuning of Model Assumptions

In this version of the model another approach is

used, using the knowledge that for completing the

scheduled production of the four towers the

company spent three months. The model will be

used to achieve this delivery time using a number of

workers similar to the one scheduled by the

company.

Model*4 is used as the starting point and then,

the number of workers is limited based on the

previous usage time. Workers with less than 10

hours of use are discarded for the next run. Using the

previous example of PF5_t1, in Model*5 the number

of workers of this profile is limited to 6. This way

Model*5 uses 96 workers but still more machines

than the number scheduled by the company. Now

limiting the number of machines according to

specification (Model*6) and then reducing five more

workers (Model*6.1) the model has provided the

following results.

Table 3.6: Due dates: Model*6 and Model*6.1.

Due date

Model*5 Model*6 Model*6.1

tower 15-Feb-13 19-Feb-13 21-Feb-13

tower 25-Feb-13 4-Mar-13 5-Mar-13

tower 21-Mar-13 22-Mar-13 28-Mar-13

tower 26-Mar-13 5-Apr-13 9-Apr-13

100

120

140

160

BusyTime (h)and nunber of occorences

BusyTime Occurrences

8h 16h

24h

Machine resources used

Tegopi Modelo*4

SimulationandModelSensitivityAnalisysofaWindTurbineTowerManufacturingPlant

731

With delivery time of approximately three months,

Model*6.1 uses a total of 91 workers meaning 8.3%

above the number of workers specified by the

company. Although more workers are used in

simulation, this value is according to the company’s

expectation for the model’s performance without

considering resource flexibility.

4 CONCLUSIONS

The proposed model allowed the analysis of the

behaviour of the studied system. Although the model

represents all the production process with a

reasonable degree of detail, special emphasis was

given to workers allocation to tasks. Nevertheless it

is easy to change other parameters and test more

scenarios.

Individual production lines were studied and

analysed in greater detail by verifying worker and

resources utilization and completion time to ensure

they could be used for correct modelling of the

overall plant.

It is possible to establish that transport traveling

distance and priority logic have a significant

influence in the response of the model. But more

important, the model is very sensitive to changes in

worker capacity and worker distribution by the

several specialization profiles. This is due to the

demand of at least one worker to execute each task

from the productive process.

The model results agree with the data supplied

by the company for the individual production lines.

For the overall plant the results showed some

discrepancies, but the sensitivity analysis allowed an

identification of the most significant problems of the

original model version. By redesigning some of the

model assumptions and logic that were originally not

part of the supplied data, it was possible to improve

the model results to within 8.3% of error relative to

the real data values. This result was considered as

acceptable by the industrial partner under the

assumptions that were considered.

ACKNOWLEDGEMENTS

This work was supported by FCT, through IDMEC,

under LAETA Pest-OE/EME/LA0022 and partially

supported by the project R046, IDMEC.

REFERENCES

Altiok, T. & Melamed, B., 2007. Simulation Modeling and

Analysis with Arena. United States of America:

Elsevier Inc.

Banks, J., II, J. S. C., Nelson, B. L. & Nicol, D. M., Fourth

Edition. Discrete-Event System Simulation. New

Jersey: PRENTICHEA L-L lN TERNATIONAL

SERIEINS lN DUSTRIAL AND SYSTEMENSG

INEERING.

Brito, S. d. S., 2014. CRESESB. [Online] Available at:

http://www.cresesb.cepel.br/content.php?cid=tutorial_

eolica.

Chance, F., Robinson, J. & Fowler, J., 1996. Supporting

manufacturing with simulation: model design,

development, and deployment. San Diego, CA., IEEE.

Hemami, A., 2012. Wind Turbine Technology. 5 Maxwell

Drive, Clifton Park, NY 12065-2919, USA: Cengage

Learning.

Kelton, W. D., Smith, J. S. & Sturrock, D. T., 2013. Simio

and Simulation: Modeling, Analysis, Applications -

Third Edition. s.l.:Simio LLC.

Law, A. M. & Kelton, W. D., 2000. Simulation Modeling

and Analisys, Third Edition. s.l.:McGraw-Hill Book

Company.

McHaney, R., 2009. Understanding Computer Simulation.

s.l.:Roger McHaney & Ventus Publishing ApS.

Redes Energéticas Nacionais, R., 2013. Energia Eólica em

Portugal 2012.

TEGOPI, n.d. [Online] Available at: http://www.tegopi.pt/

?mod=pdinamicas_pag&id=2&_item=1.

SIMULTECH2014-4thInternationalConferenceonSimulationandModelingMethodologies,Technologiesand

Applications

732