You are on page 1of 57

DEGREE PROJECT IN ELECTRICAL ENGINEERING,

SECOND CYCLE, 30 CREDITS


STOCKHOLM, SWEDEN 2021

Evaluation of Energy Losses


in a Wind Farm
GABRIEL CLAESSON

KTH ROYAL INSTITUTE OF TECHNOLOGY


SCHOOL OF ELECTRICAL ENGINEERING AND COMPUTER SCIENCE
Abstract
In recent years, the power production from wind power and other
renewable energy sources has increased significantly. Arguably, this is due
to the greenhouse effect and the economic benefit of generating power from
the wind. Wind power has the potential to replace other non-renewable
power sources. Thus, it is of interest to maximize the efficiency of the wind
farms, to produce the most power to the grid. This implies that the power
losses within the wind farms needs to be minimized. Evidently, there are
expected and unavoidable power losses within the wind farm. However,
when there are unexpected power losses or the power losses exceed what
is expected, the question arises as to what the reason for this is and if it
is possible to avoid or decrease these power losses.
In this thesis, a case study is conducted in collaboration with Skellefteå
Kraft. An existing wind farm is studied, as the case company noted
inconsistencies in power losses. One section has a larger share of power
losses than the other section in the wind farm. Thus, it is of interest
to the case company to find out why this is, to find the source to these
power losses, as unnecessary power losses mean a loss of profit as well
as loss of useful power produced from renewable energy sources. This
suggests that if the sources of the power losses are identified and the
power losses are deemed unnecessary for the operation of the wind farm,
the case company can take actions accordingly. Hence, this thesis aims
to identify the sources of the power losses and to create basis to whether
these power losses are justified and necessary for the operation of the
existing wind farm or not.
To study the power losses in the existing wind farm, a model is devel-
oped utilizing load flow analysis. The load flow analysis is based on real
hour power production data of the year of 2019. Thus, several load flow
calculations are carried out to modify the system parameters to optimize
the accuracy, and to verify the model. The model is then used to estimate
and evaluate the power losses within the wind farm, and for identification
of the sources of the power losses.
The results of the case study prove that a rather accurate model was
successfully developed. The model indicates that, for the year of 2019,
the difference in power losses between the two sections of the wind farm
was primarily due to the de-icing systems. The de-icing system of one
section constituted for a significantly larger share of the power losses in
that section than what the de-icing system of the other section constituted
for the power losses in the other section. This suggests that the de-icing
system needs to be evaluated further. Due to the design of the wind farm,
there is an additional transformer in one of the sections. Through utilizing
the model, the power losses of the additional transformer were estimated.
For the year of 2019, the model indicated that the power losses of the
additional transformer contributed rather insignificantly to the difference
in power losses between the two sections.
Sammanfattning
Under de senaste åren har kraftproduktionen från vindkraft och andra
förnybara energikällor ökat avsevärt. Det beror bland annat på växthus-
effekten och de ekonomiska fördelarna med att generera el från vind-
kraft. Vindkraften har potential att ersätta andra icke förnybara ener-
gikällor. Därför är det av intresse att maximera vindkraftparkernas effek-
tivitet och maximera den genererade effekten ut på elnätet. Detta innebär
att effektförlusterna inom vindkraftsparkerna bör minimeras. Det finns
förväntade och oundvikliga kraftförluster inom vindkraftsparken. När det
uppstår oväntade effektförluster eller effektförlusterna överstiger det som
förväntas, blir frågan om vad orsaken är och om det är möjligt att undvika
alternativt minska på dessa effektförluster.
I detta arbete genomförs en fallstudie i samarbete med Skellefteå
Kraft. En befintlig vindkraftspark studeras, eftersom företaget har no-
terat oväntade effektförluster. En sektion har större andel effektförluster
jämfört med den andra delen av vindkraftsparken. Därför är det av in-
tresse för företaget att ta reda på anledningen till detta, eftersom onödiga
effektförluster innebär ekonomiska och effektmässiga förluster. Det bety-
der att om källorna till effektförlusterna identifieras och är onödiga för
driften av vindkraftsparken kan företaget vidta åtgärder därefter. Syftet
med detta arbete är därför att identifiera källorna av effektförlusterna
och identifiera om dessa effektförluster är nödvändiga för driften av vind-
kraftsparken eller inte.
För att studera effektförlusterna i den befintliga vindkraftsparken ut-
vecklas en modell med hjälp av lastflödesanalys. Lastflödesanalysen är
baserad på realtidsdata för kraftproduktion under år 2019. Således utförs
flera lastflödesberäkningar för att modifiera modellens systemparamet-
rar, för att öka noggrannheten och för att verifiera modellen. Modellen
används sedan för att uppskatta och utvärdera effektförlusterna inom
vindkraftsparken och för identifiering av källorna till effektförlusterna.
Resultaten från studien visar att en noggrann modell kunde utvecklas.
Modellen indikerar att skillnaden i effektförluster mellan de två sektioner-
na i vindkraftparken för år 2019 främst berodde främst på avisningssyste-
men. Avisningssystemet i den ena sektionen utgjorde en betydligt större
andel av effektförlusterna än i den andra sektionen. Detta tyder på att
avisningssystemet bör utvärderas ytterligare. På grund av vindkraftspar-
kens utformning finns det en extra transformator i en av sektionerna.
För år 2019 indikerade modellen att effektförlusterna för den extra trans-
formatorn knappt bidrog till skillnaderna i effektförluster mellan de två
sektionerna.
Acknowledgements

I would like to thank my supervisor, Stefan Stankovic at the Royal Institute of


Technology at the Division of Electric Power and Energy Systems, for guidance
and valuable feedback throughout the course of this thesis. I am grateful for
your time and support.

I would like to thank my supervisor, Anders Larsson at Skellefteå Kraft, for


many interesting discussions, feedback, and encouragement. Thank you for
making this thesis possible. I would also like to thank the people at Skellefteå
Kraft for the kind welcome and their generosity.

Many thanks to my examiner Lennart Söder, professor at the Royal Institute


of Technology at the Division of Electric Power and Energy Systems, for the
helpful comments and suggestions.
Contents

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Delimitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Covid-19 disclaimer . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Theoretical background 7
2.1 Grey box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Load flow analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Quasi-dynamic simulation . . . . . . . . . . . . . . . . . . . . . . 8

3 Methodology 9
3.1 Research process . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Information search . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3 Data collection and preparation . . . . . . . . . . . . . . . . . . . 10
3.3.1 Data collection and description . . . . . . . . . . . . . . . 10
3.3.2 Management of missing data . . . . . . . . . . . . . . . . 10
3.3.3 Data selection and filtration of bad data . . . . . . . . . . 11
3.3.4 Training, verification and validation data . . . . . . . . . 12
3.4 Model building . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.4.1 Software and configurations . . . . . . . . . . . . . . . . . 13
3.4.2 Model assumptions . . . . . . . . . . . . . . . . . . . . . . 14
3.4.3 Model training . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4.4 Model verification . . . . . . . . . . . . . . . . . . . . . . 17
3.5 Analysis of the final model . . . . . . . . . . . . . . . . . . . . . . 18

4 Case Study 19
4.1 Skellefteå Kraft . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 Case study wind farm . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2.1 Wind farm description . . . . . . . . . . . . . . . . . . . . 19
4.2.2 Area description . . . . . . . . . . . . . . . . . . . . . . . 20
4.2.3 Grid requirements . . . . . . . . . . . . . . . . . . . . . . 21
4.2.4 Wind turbine types . . . . . . . . . . . . . . . . . . . . . 21
4.2.5 Cable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.2.6 Three-winding transformers . . . . . . . . . . . . . . . . . 22
4.2.7 Two-winding transformers . . . . . . . . . . . . . . . . . . 22
4.3 Model of the existing wind farm . . . . . . . . . . . . . . . . . . 24
4.3.1 Case model description . . . . . . . . . . . . . . . . . . . 24
4.3.2 Case specific model assumptions . . . . . . . . . . . . . . 25
4.3.3 Characteristics of the wind farm sections . . . . . . . . . 26
4.3.4 Active power, reactive power, and voltage measurements . 27
4.3.5 De-icing data . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.3.6 Model setup . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5 Results 35
5.1 Framework for evaluating results . . . . . . . . . . . . . . . . . . 35
5.2 Parameter estimates . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.3 Model performance . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.4 Estimated power losses . . . . . . . . . . . . . . . . . . . . . . . . 38
5.4.1 The de-icing system . . . . . . . . . . . . . . . . . . . . . 39
5.4.2 The additional transformer . . . . . . . . . . . . . . . . . 41

6 Analysis and discussion 42


6.1 Scrutiny of results . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.2 Data quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.3 Source of error and model improvement . . . . . . . . . . . . . . 44
6.4 Sustainability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

7 Conclusion 46
7.1 Main findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
7.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
1 Introduction

This chapter aims to give an introduction to the topic of wind farm power losses
starting from a wider societal context. The research problem together with the
following research questions are then presented along with academic contribu-
tions and delimitations.

1.1 Background

Renewable energy sources for electricity production are growing. To limit the
amount of pollution and CO2 released, countries and international collabora-
tion have increased their focus in finding new efficient ways to produce fossil
free electricity [1]. In 2015, 195 countries agreed on 17 United Nations’ sustain-
able development goals for 2030 where number seven was ‘affordable and clean
energy’ [2] [3]. To transition from oil and coal as the primary energy resources
efficient alternatives are necessary. The change is powered by decreased costs
of wind turbines and solar panels [4]. To optimize for onshore wind power pro-
duction, several wind turbines are often organized together in a selected area
called a wind farm.

Wind power and specifically wind farms have seen a rapid increase in the last
decade. In Sweden, the installed capacity and realized electricity production
from wind power have increased by almost a factor of ten in the last ten years.
The share of wind power in Sweden’s total electricity production has subse-
quently increased to over 10% [5]. To contribute to a continued development
towards affordable and clean energy continuous investments are required to ex-
pand the current fleet to wind turbines and wind farms. However, increasing
installed capacity of wind power is not a definitive answer, focus must also be
directed towards maintaining and improving current wind farms efficiency.

Wind farms are clearly a case for economies of scale [6]. With respect to wind
farms some additional considerations need to be taken regarding power losses.
The potential energy losses reduce the wind farm’s overall efficiency and in
turn lead to less electricity produced. Power losses are not only bad in pure
economical terms, but it also decreases the share of renewable energy produced
to the electricity grid. Subsequently, the sources for power losses are important
to identify and mitigate.

This thesis is done in collaboration with Skellefteå Kraft. The company provided
the case study that is presented later in the report. The core reason for the case
study is that Skellefteå Kraft have found inconsistencies in power losses in one
of their wind farms. Thus, there are unexplained and unidentified power losses
within the wind farm that is of interest to examine. This thesis intends to
address the issue of unidentified power losses. Furthermore, it seeks to give
basis to explain the power losses in the case study.

1
Presumably, there are zero systems that are ideal. This applies to power systems
too. The power system is not a hundred percent efficient, likewise, the wind
power system is not either. The power in the wind cannot be utilized to a full
extent. It is not possible to completely extract the energy contained in the wind.
One can also expect other power losses in a wind power system.

The power losses within a wind farm can be categorized as internal power con-
sumption and internal power losses. The internal power consumption can be
defined as the power consumed related to the operation of the wind farm. This
includes the wind turbines, the cooling systems, de-icing system, the control sys-
tem of the individual wind turbines and the wind farm, and other local power
consumption such as maintenance buildings. In colder climates, the de-icing
system is expected to constitute a significant part of the internal power con-
sumption. It could constitute up to as much as about 20 percent of the power
losses [7]. The internal power losses can be defined as the power losses related
to the components within the wind farm. This includes the transmission losses
within the cables, the transformer power losses, and other power losses such as
the ones at the point of common coupling (PCC) within the switching station.
If there are unknown or unexplained power losses in the wind farm i.e. the
actual power losses in the wind farm is higher than the theoretical, the question
arises to where these power losses occur and if they are justified.

An individual wind turbine can both produce and consume active and reactive
power which depends on how the wind farm is controlled and the specifications
of the wind turbine. The active power is the real power dissipated in the elec-
trical circuit. The reactive power is the futile power which flows between the
power source and the load [8]. Through the control of the power factor of the
wind turbine, the active and the reactive power is controlled. The power factor
indicates the phase shift between the voltage and the current. In wind turbines
the power factor set point is most often set equal to one, or at intervals close
to one [9]. This is to minimize the losses and the heating in the converter.
Through the regulation of the consumption and the production of the reactive
power the wind turbines make it possible to manage the voltages throughout
the wind farm at desired levels. It depends on the control system of the wind
farm but often the voltage at the PCC is managed to stay at a specific level [10].
The wind farm may also be used to provide stability for the rest of the grid,
meaning that the wind farm can provide the grid with reactive power compen-
sation. Hence, usually there is a set requirement of an interval for the reactive
power at the PCC.

It is useful to model the power flows in the wind farm to identify the source of the
power losses and to find the internal power consumption that is unjustified. This
gives information as to where improvements are desired. Load flow analysis is
a numerical analysis tool of electric power flow in electric power systems, which
can be used to analyse steady state operations, to calculate voltages, current
flows, and power losses. In general load flow analysis is useful for investigations
of power systems with regards to fault analysis, system stability as well as
for future expansion plans of the power system or alternative designs of power
systems [11].

2
Wind power production fluctuates as the wind changes [12]. Therefore, load flow
analysis can be utilized to interpret and evaluate the wind farm power system
over a period with several different scenarios of power production. Which could
be useful to explain power losses in the wind farm. Quasi-dynamic simulations
carry out multiple load flow calculations for longer periods of time, which is
helpful to identify what happens in the wind farm [13]. This thesis aims to
investigate the factors which affect power losses within a wind farm by creating
and evaluating a load flow model.

1.2 Research problem

The purpose of a wind farm is to extract energy from the wind and to generate
power to be utilized. The amount of available power, generated by a wind farm,
depends on numerous factors, for instance the energy in the wind, the design of
the wind farm, the design of the control system, the components, the location,
the internal power losses, and the internal power consumption in the wind farm.
There are expected power losses and expected internal power consumption in a
wind farm. The main problem is if there are unnecessary power losses in wind
farms. These power losses have economic and environmental consequences.

There are some challenges with power loss identification and estimation in wind
farms. One aspect is that it is expensive to thoroughly measure specific power
losses and to gather extensive data within the wind farm. Specific measurements
are needed if the power losses are to be calculated. Thus, the estimated power
losses depend on the quality of the measurement data. The exact power losses
could be difficult to calculate due to redundancy of data, inaccuracy of data,
and/or the lack of data. Estimating power losses utilizing models would provide
information to where the power losses occur. This would give basis to what area
would need improvement to decrease the power losses in the power system, in
this case wind farms.

In this thesis, a case study is conducted regarding an existing wind farm. The
case company, Skellefteå Kraft, noticed irregularities in power losses in one of
their wind farms. The power losses were higher in one section of the wind farm
compared to the other section. This is curious and is of interest to investigate the
reasons for this. Thus, the thesis seeks to identify the source of the unexplained
power losses. There are several approaches to the problem that can be taken. It
is possible to set up measurement meters around the wind farm to gain power
consumption data of components. Subsequently, compare the measurements
acquired to the theoretical power consumption to see if there are excessive power
consumption or power losses. This would probably not be the most economical
approach, as measurement instruments are expensive. It also depends on the
size of the wind farm regarding how many measurement instruments are needed
and how many components within the wind farm are to be examined. Another
approach is to create a model of the wind farm to identify the source of the
power losses.

To create a representative model of the existing wind farm, there are certain
things required to be addressed, for example what information is available. In
this thesis, the existing wind farm can be interpreted as a ‘Grey box’. Where

3
the ‘Grey box’ contains the physical information and the structure of the wind
farm together. The parameter values of the components of the wind farm are
unknown. What goes on within the ‘Grey box’ is unknown [14]. However,
load flow analysis can be utilized to model and solve the ‘Grey box’ and create
an accurate representation of the wind farm. In other terms, the load flow
analysis provides a solution to the state of the power system with respect to
the active and the reactive power generation and demand and the voltage. The
model could give an indication of supposedly unnecessary power consumption
or power losses in the wind farm, by comparison of the model output result and
the actual output of the wind farm.

This means that if the load flow analysis is to be used to investigate the existing
wind farm, there must be a lot of data available. It depends on the set up of
the model, as the input and the output data of the model would be different.
In any case, the data needed are measurements of the active and reactive power
generation and demands as well as voltage data of the wind turbines and the
connection to the grid. The load flow analysis provides a single snapshot solution
of the state of the power system. Thus, multiple load flow analysis must be
done to evaluate the unknown parameters of the power system, the wind farm.
This is because otherwise the estimated unknown parameters would not be
correctly estimated and correspond to reality. Several parameters cannot be
estimated from a single state, a single solution, as the estimation would only
correspond to that specific state. There is also the possibility that there are
multiple estimations of the parameters that satisfies the solution. Thus, the
model would probably not represent other states of the wind farm accurately.
Additionally, the power losses and the power consumption of the wind farm
would not be represented accurately, and it would not be possible to properly
provide a basis to where the unnecessary power losses occur. Therefore, several
data measurements must be available, arguably, the more data available the
better representation of the wind farm can be modelled.

To make an accurate model of the wind farm, the output of the load flow cal-
culations is compared to the actual measured values of decided quantities. The
comparison between the result of the model and the actual measured values of
the wind farm gives an error of the model. This error can be estimated utiliz-
ing a couple of tools. For instance, the mean square error (MSE) is calculated
as the average squared difference between the estimated values and the actual
values, which provides an accuracy estimate of the model. This means that
the smaller the MSE value becomes, the more accurate to reality the model be-
comes [15]. Hence, to make the model more accurate to reality, the parameters
of the model are to be configured to minimize the error, the MSE. When the
error estimate is small enough, the model of the wind farm depicts reality close
enough. Thereafter, the model can be utilized to look for improvement options,
or areas of improvement. In other terms, to see where the power consumption
and the power losses occur, thus, appropriate actions can be implemented.

4
The research problem described above lead into the following research questions
for the thesis:

• What factors affect power losses in wind farms?


• How can power losses in wind farms be identified by utilizing a power flow
model?

If an accurate model of the wind farm is created, the source of the power losses
and power consumption could possibly be identified. Thereafter, these power
losses or power consumptions could be evaluated to whether they are necessary
to the operation of the wind farm or not. The expectation is that there are some
power losses in the existing wind farm that are unnecessary high in one of the
sections compared to the other. The internal power consumption of the wind
farm in both sections should be very similar to one another. If it proves that
the internal power consumption of for instance the wind turbines in the two
sections differ from one another, further studies should be conducted. This is
for justification and verification purposes, and if this is true appropriate actions
should be taken.

1.3 Contributions

The thesis contributes primarily with additional information and insights to the
case company, about the power losses within the existing wind farm studied.
The thesis investigates the production data of the wind farm for the year of
2019, and the power losses in the wind farm. Through examination of data and
a developed model, the sources of power losses were evaluated and estimated.
Hence, the thesis provides more information to the company of the source of
power losses and further areas to improve upon within the wind farm, to decrease
the power losses. Secondly, the thesis contributes with further research to power
loss identification in wind farms, utilizing load flow analysis. It provides further
knowledge of how the load flow analysis can be incorporated in power loss
estimation utilizing extensive data. It also provides further knowledge of the
sources of power losses in wind farms in subarctic climates.

1.4 Delimitations

The load flow model of the existing wind farm presented puts emphasis on
the identification of power losses during operation utilizing annual hourly data.
The data from the existing wind farm was acquired for the years of 2017, 2018
and 2019. However, the data from 2017 and 2018 was considered insufficient
and invalid to use due to inconsistency. Thus, the model development and the
evaluation of the model is limited to the data of the year of 2019. The developed
load flow model is designed to use as a basis for identification of the power losses
of the existing wind farm. It is meant to be used as a tool for identification and
estimation of the power losses.

5
1.5 Covid-19 disclaimer

This thesis was conducted during the spring of 2020, during the Covid-19 pan-
demic. The pandemic affected the thesis in various ways. A planned field trip
was cancelled abruptly.

6
2 Theoretical background

In the following chapter the relevant theoretical background is presented. The


theoretical background includes the information to the relevant concepts in the
thesis together with theory of the load flow model used in the case study.

2.1 Grey box

The concept of the ‘Grey box’ is used in this thesis. The concept is useful as
it provides information related to how to interpret the problem of power losses
in the existing wind farm and how to approach it. The information regarding
the existing wind farm is presented later in the thesis. The concept is utilized
as the physical structure of the existing wind farm is known but there are
uncertainties of the parameter values of components. In short, there is data of
the power production from the wind turbines and the power output to the grid,
but there is no data of what occurs in between the two. Hence, the system in
between can be viewed as the ‘Grey box’, which is to be evaluated [16].

The concept of the ‘Grey box’ is a combination of the ‘Black box’ and the ‘White
box’. The ‘Grey box’ is meant to be used as a tool to determine errors in systems
because of the operation and/or the structure of the system. The ‘Black box’ is
the representation of a system or a device or an object, where one does not know
what occurs within the ‘box’ or how the ‘box’ acts and works. The structure
of the ‘box’ is unknown. The ‘Black box’ is observed indirectly, one observes
the input and the output of the ‘box’, whereas one cannot observe how the
input is converted to the output [17]. The ’White box’ is the representation
of a system, currently most often used in software development. It is purely
theoretical. Where one knows what occurs within the ‘box’ and how the ‘box’
act and work. The structure of the ‘White box’ is known. The ‘White box’ is
observed directly, with the knowledge of what the software code does [18].

The ‘Grey box’ concept incorporates parts of both the ‘Black box’ and the
‘White box’. The ‘Grey box’ combines the two ‘boxes’ in the sense that the
structure of the model is known together with internal data as well as math-
ematical models used. However, certain parameter values of the model are
unknown [19].

2.2 Load flow analysis

The load flow analysis is a tool for examination of power systems. It computes
the steady state characteristics of a power system with defined power genera-
tion and determined transmission network structure. It solves the system of
nonlinear complex power balance equations with iterative numerical methods.
It provides a solution to represent the state of the power system in a particular
moment. It is utilized to examine the current state of power systems, addition-
ally it is used to provide basis for future expansion plans of power systems, and
to determine optimal operations [11]. Load flow calculations are usually solved
by iterative numerical methods. The solver used for the load flow calculations

7
in this thesis is the Newton-Raphson method.

The load flow analysis can be utilized to evaluate a power system in terms of
power losses. If there are several data points available, of power generation and
demand and voltages, several load flow analyses can be carried out. Thus, it is
possible to utilize the load flow analysis as a tool to estimate the parameters
of the power system for several scenarios, and to determine if the parameters
of the model are well corresponding to reality. It would thereby be possible
to find out if the components operate as expected regarding the power losses
providing reliable information of the sources and the magnitudes of the power
losses around the wind farm. This means that identification of excessive power
losses or excessive power consumption can be done. Therefore, with several data
points, the power losses of the power system can be investigated utilizing load
flow analysis.

2.3 Quasi-dynamic simulation

The quasi-dynamic simulation carries out multiple load flow calculations over
a certain time period. As the tool utilized time-based parameters, one evalu-
ates the system in several scenarios under normal operation conditions. One
compares the output of the quasi-dynamic simulation for the model with the
actual output of the electric power system. Thus, one can adjust parameters
to increase the accuracy of the model. Arguably, the more data points one has
access to, the more accurate one can evaluate and model the system [20].

8
3 Methodology

In the following chapter the methodology will be presented.

3.1 Research process

The thesis investigates the factors that cause the power losses in an existing
wind farm. As explained in chapter 1.1 the power losses in the wind farm can
be put into two categories: internal power consumption and internal power
losses. The internal power consumption refers to the power consumed related
to the operation of the wind farm. The different operations include the power
used for the wind turbines to function, the control system, the de-icing system
for the wind turbine blades in cold climates, running of local facilities, and
the cooling system for the converters. The internal power losses refer to the
power losses related to the components within the wind farm. This includes
the transmission losses of the cables, losses in transformers and the power losses
within the substations.

The research problem in question has to do when there are unidentified power
losses in the wind farm. If the sources of the power losses are unknown, there
could be excessive power losses within the wind farm. Some power losses could
be considered inevitable, however, it is of great interest to try and minimize
these losses. To minimize the power losses, the first step is to identify the
source. As stated, some power losses might be inevitable and justified for the
wind farm to stay operational. The study examines the option of how load flow
analysis can be utilized to identify unknown power losses in an existing wind
farm. The research process is displayed in figure 1. The research process with
reference to figure 1 is explained in detail in the following subsections.

Figure 1: Research process

3.2 Information search

The first block in the research process was the information search. The in-
formation search consisted of retrieving and reviewing the available knowledge
and literature within the field of power losses in wind farms. A literature study
was conducted. Additional sources of information came from the case company.
The sources from the case company included: test reports from the wind farms,
different databases of operation data covering the last five years, mapping of
wind farm structure and various components.

9
After the initial research which gave a basic understanding of the underlying
systems, discussions were held with the supervisor at the case company. The
discussions resulted in a research problem which was specific to the case study.
To broaden the scientific approach to the research problem, discussions were
held with the supervisor at the Royal Institute of Technology which led to the
formulation of the research questions and research problem presented in the
thesis.

Along with the rest of the thesis work, additional sources of information were
used to complement and deepen the analysis of the paper. This was done in
parallel to the following three blocks shown in figure 1. Notably, additional data
and test reports from the case company were studied to fully understand the
underlying system of the wind farm.

The result of the information search was not only the formulation of the research
questions and research problem, but also the approach to try to answer the
resulting research questions and problem formulated. The information search
suggested that in order to identify and explain the factors behind power losses
in wind farms a model for simulation and evaluation was needed. The model
chosen was a load flow model which together with historical data from a wind
farm could be simulated to identify potential power losses.

3.3 Data collection and preparation

The second block in the research process was the data collection and preparation.
The data collection and preparation consisted of several steps described below.

3.3.1 Data collection and description

The data used in the case study was collected from the case company’s internal
databases. The data consisted of generation data of the wind turbines, the
voltage data of the wind turbine buses, the load data of the different loads
throughout the wind farm and the output power data from the wind farm to the
grid at the PCC. There are also measurements of the hourly power consumption
of the loads in the wind farm. The available data was hourly values from 2017
to 2020. The data was downloaded from the databases to Python and the
model with Excel as an intermediary. In total roughly 26 000 data points were
collected.

3.3.2 Management of missing data

For different time periods between 2017 and 2020 there were missing data points
for several of the wind turbines. In these cases, it is important to deal with the
corresponding missing data points before feeding them into the model. There
are several different ways to deal with missing data and the appropriate method
depends on the characteristics of the data as well as the quantity available. In
cases where just a few data points from individual wind turbines are missing, the
chosen approach was to fill in the values with those of other nearby, correlated

10
wind turbines. As such, the information was not lost even though the particular
missing data point became an estimation rather than the measured occurrence.
The other nearby wind turbine which was used to fill in the missing data was
chosen based on geographical proximity.

For missing power consumption data of individual wind turbines, the same ge-
ographical proximity logic was used to fill in missing values. For other missing
values of different larger systems, the previous value has been used where ap-
plicable. For the instances where there were missing voltage data, the data
point was removed as the voltage data was considered the validation data for
the model to be evaluated against.

3.3.3 Data selection and filtration of bad data

After collecting the data and managing missing values of the data points there is
another step which needs to occur. The data needs qualitative work to evaluate
if it is useful for the model. By looking at the amount of missing data a decision
was taken to only use the hourly measurements from 2019 and remove the data
from 2017 and 2018. The data of 2019 was comparatively more consistent.
There was considerably more data missing from 2017 and 2018 than from 2019.
The resulting amount of data points (8760) was considered enough to perform
the load flow calculation on even after further filtration.

Another qualitative step included filtering out outliers as well as other data
points which logically did not add up in the different measurements. In this
case, a certain criterion was created to evaluate whether a data point was to be
included or not in the training and evaluation of the model. A data point which
did not fulfil the criterion was removed.

For a viable hour of the year, the individual data point should support the
following criterion:

• Each data point should contain sufficient information for the training of
the model as well as for the validation of the model. This means that there
must be enough data to run the load flow calculation and enough data to
validate the result of the load flow calculation. The validation data is used
to compare the result of the load flow calculation with. Thus, the data
needed for each data point is the generation data of the wind turbines,
both active and reactive power as well as data of the voltage at the wind
turbines. Furthermore, the data of the active power output at the PCC
is needed and the data of the power consumption of the loads in the wind
farm is needed. It depends on the model setup, what data is needed to
run and what data is needed to validate the load flow calculations, this is
further addressed in subsection 3.3.4: Training, verification and validation
data.

11
• The measured active power produced from all the wind turbines combined
should be close to the measured output active power to the grid, at the
PCC. It should be within a certain interval depending on the data and
the size of the wind farm. This is due to large variations of produced
active power during the year and to neglect outliers and data points where
the data is insufficient. The interval should ensure to not eliminate data
points excessively. The interval of acceptance chosen for the measured
active power produced and the measured active power output to the grid
was +/- 8% of the installed power of the wind farm;
• The loads within the wind farm that consume a combined amount of
power less than 0.1% of the power produced by the wind farm are to be
neglected. This criterion is not to be utilized if there is data of numerous
loads in the wind farm, suppose the loads power consumption would add
up to more than 0.1%;

After applying the criterion on the 2019 data a total of 7420 data points re-
mained. This was considered enough to proceed with the model building and
load flow analysis.

3.3.4 Training, verification and validation data

The available data has to be separated into two parts, two sets of data. The
two datasets are referred to as the input dataset and the validation dataset.
The input dataset consists of the data that is used as input to the model, to
run the load flow calculations. The validation dataset consists of the data that
is used for comparison with the output of the model to evaluate how good the
estimated model is, to compare with the result of the load flow calculations.
The data in the two datasets is dependent on the setup of the model.

To run the load flow calculations, as mentioned before, the types of buses in
the power system, the wind farm model, need to be identified. This defines
what type of input data is used to run the simulations and what type of data is
used as validation data. For the PQ-buses the input data needed is the active
and reactive power data for both load and generation. For the PU-buses the
input data needed is the active power data of the generation and load and the
voltage data. For the slack bus, the input data needed is the reference point
for the voltage together with the active and reactive power load data. As such,
the input dataset needs to include the measured data of these quantities for the
respective bus type.

The output of the load flow calculations depends on the types of the different
buses. The output of the PQ-buses is the voltage. The output of the PU-buses
is the reactive power. The output of the slack bus is the active and reactive
power generation data. The output quantities are compared with the validation
dataset. The validation dataset needs to include the measured data of the
quantities for the respective bus type. The comparison is used to calculate the
error estimates, which is used as the estimation criteria described in subsection
3.4.3: Model training.

12
Table 1: Datasets
Input data Validation data
Dataset
quantity quantity
Training dataset
Pg , Qg , Pld , Qld Ppcc Ug
(75% of available data points)
Verification dataset
Pg , Qg , Pld , Qld Ppcc , Ug
(25% of available data points)

Table 1 shows the setup of the datasets used in this thesis. As shown the input
data in this thesis is the generation data. This is the active power production
(Pg ) and the reactive power production (Qg ) of each wind turbine in the wind
farm. The input data also includes the load data, the active and reactive power
(Pld ), (Qld ). The output of the load flow model will be the voltage data at the
wind turbines (Ug ) and the active power data (Ppcc ) at the PCC. As such, the
validation data in this thesis will be the voltage data at the wind turbine buses
and the active power data at the PCC. The validation data is to be compared
to the output of the load flow model. This is explained further in the section
3.4: Model building.

3.4 Model building

The model building chapter presents the method used when configuring the
model and using the data prepared to perform load flow calculation. The de-
scriptions of the methods used for estimation of the model parameters are pre-
sented below. This includes the building process, the assumptions, the training
process, and the verification process of the model.

3.4.1 Software and configurations

The program DigSilent Power Factory was used for model building. The soft-
ware is used by the case company and was provided for this thesis. The program
was utilized to create a single line diagram of the wind farm with the main com-
ponents. It was also used to solve the load flow calculations within the wind
farm during operation.

The programming language Python was then utilized to communicate with


DigSilent Power Factory to run the quasi-dynamic simulations and to extract
the relevant results from the model and generate and calculate the error mea-
surements of the model. Python version 3.8 was used.

13
3.4.2 Model assumptions

To enable load flow calculation there are some assumptions that are inherent
to the model which exist to decrease the complexity of the problem. There is
a significant number of variables that affect the power system, and all cannot
be taken into account. There are variables that are more important and have a
larger effect on the power system, the wind farm, than others. The assumptions
presented below are all used in this thesis to decrease the complexity of the
wind farm. Furthermore, the additional assumptions specific to the case study
are presented in chapter 4.

The general assumptions of the analyses used in this thesis are:

• The system is assumed to have a constant frequency of 50 Hz;


• The system is assumed to have symmetrical operation. This implies that
the system is balanced across all three phases and a one-phase equivalent
model can be used;

• The wind turbines can both consume and produce reactive power. To
what extent depends on the wind turbine;
• The initial values of the model parameters to be estimated are obtained
from the literature. If there are test reports for the components, the
measured samples are to be considered. For instance, the load losses of
the transformer are not definite as they depend on the tap position. It
is of value to set a standard tap position of the transformer and consider
the load losses at that specific tap position. This is to consider the lack of
data of the tap position. As it is unknown when and on what tap position
is active, the standard is set to consider the average power losses for the
year.

3.4.3 Model training

One of the purposes of the model is to create an accurate representation of the


wind farm. The model training session is used to find out more information
about the wind farm. The wind farm is initially represented with the ‘Grey
box’ model because the physical information is known such as the layout to-
gether with the components within the wind farm. There are, however, certain
parameter values of the wind farm that are unknown. Going from a ‘Grey box’
that does not hold enough information of a model to an accurate representation
of the wind farm can be found by estimating the ‘Grey box’ model parameter
values. The error of the model can be measured as the difference between the
output of the model and the real system measurements. It implies that the error
of the model, the output results of the model compared to the actual measure-
ments at the wind farm is to be minimized for the model to improve and more
reliably depict the properties of the actual wind farm. The error is explained
further down below.

14
One of the purposes of the training session of the model is to adjust the model
parameters to minimize the overall error and thus create a more accurate repre-
sentation of the wind farm. The model training session uses 75% of the available
data points. The 25% remaining data points are used for model verification
against other scenarios and to see how the model performs in other scenarios.
The training and verification datasets were divided in chronological order as the
underlying data is time dependent.

Figure 2: Training Session of the Model

Figure 2 illustrates the training session of the model, which is to be explained


further in depth. The training session is an iterative process to estimate the
unknown parameters of the wind farm. As the figure shows, the input data
together with the initial parameters are fed into the model, the ‘Grey box’.
The ‘Grey box’ contains the physical information of the wind farm: the type of
components and the layout of the wind farm.

In the model training the wind turbine buses are considered PQ-buses which
has implications on which input and output data that is to be considered. The
input and output data of the PQ-buses is previously explained. The PCC is
considered as the slack bus. As such, the input data is the voltage reference
point.

The error of the model is estimated and evaluated through a couple of estima-
tors; the mean square error (MSE), the root-mean square error (RMSE) and the
weighted root-mean square error (wRMSE) [15] [21] [22]. The error estimates
can be interpreted as measurements of accuracy and quality. The lower the val-
ues of the error estimates, the better accuracy of the model. How the MSE, the
RMSE and the wRMSE are calculated is explained in more detail below. The
parameters of the model are then modified, and the simulations are run once
again. A new error is then evaluated. This is done until the MSE, the RMSE
and the weighted RMSE reach a plateau where they stop decreasing. As such,
a more accurate representation of the wind farm is created.

15
To summarize the steps, a list is provided below:

1. The single-line diagram of the wind farm is created with the essential
components, the transformers, generators and loads and the cables with
the corresponding parameters;

2. The initial conditions of the model are chosen, the input data and the
initial parameters of the components are chosen. The quasi-dynamic sim-
ulations are run. In other words, a number of load flow calculations are
carried out in sequence over the data points chosen as inputs to the model;
3. The results of all the load flow calculations are compared with the vali-
dation data, which gives an error of the model. The error of the model is
estimated through calculating three estimators: the MSE, the RMSE and
the wRMSE;
4. The model parameters are updated, to decrease the value of the error
estimates, to increase the accuracy;

5. Steps 2 to 4 are done repeatedly until the estimator values reach a plateau
where they stop decreasing. As such, the error of the model is considered
minimized.

The initial parameters of the components are assumed as the standard param-
eters mentioned in the model assumptions. The initial parameters are the ones
acquired from the literature and the test reports of the components. The pa-
rameters values are then updated in the iteration process of the training session
mentioned above.

The quasi-dynamic simulations carry out multiple load flow calculations. Uti-
lizing several data points to carry out several load flow calculations means that
the error estimates are calculated for several data points. Thereafter, the aver-
age of the MSE values of the same quantity are calculated. To clarify, the MSE
estimate of the voltage is calculated at the specific wind turbine and this is done
for all wind turbines. Thus, providing an error estimate value of the voltage at
all the wind turbines. The average of these MSE values becomes a measure for
the accuracy of the model.

The MSE is calculated first, the RMSE and the weighted RMSE are calculated
based on the MSE. Before calculating the error estimates it is important that
the data compared (the resulting output and the validation data) are in the per
unit system ([p.u.]). This is to make sure that all the parameters have the same
impact on the MSE value. The voltages and the active power should influence
the error estimation the same.

Equation (1) shows how the MSE is computed. Where Yi is the result value of
the output of the model, Ŷi is the actual measured value, the validation data,
and n is the number of data points.

n
1X
M SE = (Yi − Ŷi )2 (1)
n i=1

16
The MSE is calculated for the different quantities. The MSE values for the same
quantities are combined as an average. The average MSE for the voltage, the
active power and the reactive power are all calculated separately. To calculate
the average MSE value of the specific quantity, equation (2) is used.

n
1X
M SEq = M SE (2)
n i=1

The M SEq is to denote the quantity, voltage, active power or reactive power.
Here the n denotes the number of MSE values of the same quantity. For instance,
how many wind turbine buses where the MSE value is based on the voltage.

The root-mean square error and the weighted root mean square error are com-
puted from the MSE values. The RMSE is computed using equation (3), as-
suming that all weight coefficients are equal to one. The weighted RMSE is
computed using equations (3) and (4), where the weight coefficients are decided.
These two estimates are calculated the same way as the MSE. The purpose of
the weight is to control the importance of each quantity. Thus, the weight im-
pacts how the accuracy of the model is estimated. This also means that the
error estimates depend on the setup of the model, that some quantities have a
larger impact on the accuracy than other quantities.

Equation 3 and 4 displays how the RMSE and the weighted RMSE are com-
puted. Where wq is the weight of the quantity displayed in equation 4.

v
u n
uX
RM SE = t wq M SEq2 (3)
q=1

n
X
wq = 1 (4)
q=1

3.4.4 Model verification

The model verification process occurs after the training session has been per-
formed. In the training session of the model, the parameter values were fine
tuned to better represent the real system. The parameters investigated were
modified to increase the accuracy of the model and decrease the values of the
error estimates. This was done utilizing the load flow calculations of one dataset,
as mentioned previously, which consisted of 75% of the data points available. In
the verification of the model session, the performance of the model is evaluated
using the dataset which consists of the 25% data points remaining.

The dataset consisting of the 25% of the data points available is used to see how
well the model parameters were estimated in the training session of the model.
This is to see how the model performs with an independent dataset and to
evaluate the previous training estimations by again calculating error estimates,
namely MSE, RMSE and wRMSE. This provides an indicator of how the model

17
would respond to other data points, for example when faced with new future
scenarios.

To verify the model against the independent dataset, the power losses of the
model is expected to deviate within a boundary to the actual power losses
of the wind farm. As such, the load flow calculations are performed for the
independent dataset, and the simulated active power output to the grid at the
PCC is compared to the actual measured active power output to the grid at
the PCC. The limits to how much the simulated can deviate from the measured
active power at the PCC is not fixed for every model, as it depends on the size
of the power system as well as the quality of data.

In this thesis, the model is considered verified when the deviation is in between
+/- 0.04 in per unit at most for the simulated and the actual measured active
power output at the PCC, where the base power is the installed power of the
wind farm. Additionally, the simulated total amount of energy output to the
grid must be within the interval of +/-1% of the total measured energy MWh
output to the grid at the PCC.

The subsequent error estimate from the model verification further indicates
whether the model is a good representation of the wind farm. The model ver-
ification also evaluates the results of the model parameter estimates from the
training session. Therefore, it is important that the limitations of the compo-
nent parameters are taken into account in the training session.

3.5 Analysis of the final model

The fourth block in the research process is the analysis of the final model, as
the research process figure 1 displays. The model developed is analysed and
the output results are interpreted. The power losses of the developed model are
estimated using the results from the load flow calculations. For the analysis of
the power losses in the wind farm the load flow calculations are performed once
more but on the entire dataset (that is 100% of the filtered 2019 data). The
calculations result in an estimation of power losses in different components in
the system. Here it is important that the model is trained so that when the
final calculations are performed, it is based on the best available guess of the
system parameter. The power losses estimated by the model are compared and
analysed to identify the source/sources of the unexpected and excessive power
losses. If there are any unexpected or excessive power losses, it is important
to consider different possibilities to further investigate and potentially mitigate
the power losses.

18
4 Case Study

In this chapter, the information regarding the case study is presented. First, the
wind farm is described and then the model of the wind farm is presented.

4.1 Skellefteå Kraft

Skellefteå Kraft is a power company located in the northern part of Sweden,


in the province Västerbotten. The company is one of the five largest electric
power producers in Sweden, and the power produced is from purely renewable
energy resources. The bulk of the power is produced from hydro and wind
power. The company manages the transmission network of more than a 10000
km in the area. This includes maintenance, monitoring and development of
the transmission network. Currently, the company manages five wind farms
of a total of 130 wind turbines, and the largest wind farm consists of 99 wind
turbines.

As previously mentioned, Skellefteå Kraft noticed that within the wind farm of
the 99 wind turbines, there are inconsistencies in power losses. There is a larger
share of power losses in one section of the wind farm than in the other. It is
of interest to Skellefteå Kraft to find the reason for this, and if it is possible to
decrease the power losses in the wind farm. The thesis addresses the issue of
power losses in the existing wind farm and seeks to give basis to the differences
in power losses between the two sections.

4.2 Case study wind farm

In this subsection a more thorough description of the studied wind farm is


presented, together with component parameters and an illustrative model figure
of the wind farm.

4.2.1 Wind farm description

The existing wind farm consists of 99 wind turbines with the installed capacity
of approximately 247.5 MW, and an annual production of almost 700 GWh.
The wind farm is connected to the national transmission network, the Swedish
TSO company, at 400 kV. There are certain requirements for the connection
point of the wind farm which will be presented later in the report.

Table 2: Wind turbine distribution


Section I II
Cluster I II III IV
Number of wind turbines 30 30 21 18

19
The wind farm was built in two sections, four clusters of wind turbines. Table 2
displays how the 99 wind turbines are distributed in the four clusters. Section
I consists of the wind turbines in clusters I and II and section II consists of the
wind turbines in clusters III and IV.

The two sections are connected to the Swedish TSO company, by two three-
winding transformers with individual power monitoring systems for each of the
two sections. Within the wind farm there are control systems for safe operation,
control systems for the power production of the wind turbines, de-icing systems,
cooling systems and heating systems and other service facilities.

4.2.2 Area description

The location of the existing wind farm is in the northern part of Sweden. It is
an onshore wind farm at a low-fell zone. The low-fell landscape is rather flat
with its highest elevation point above the forest frontier, and it is more upland
than the surroundings. The wind farm lies on an altitude around 700 meters
above sea level. The map provided below in figure 3 is to give an indication of
where the wind farm is located.

Figure 3: Map of Sweden, with an approximate location of the analysed wind


farm

The climate is subarctic, specifically, it is cold, windy, and icy. Subarctic climate
is generally characterized by usually very cold winters, and short, cool to mild
summers. The temperature can vary quite vastly, between negative 50 to plus
25 degrees Celsius. The average annual wind speed in the area is estimated to
be around 8 meters per second.

20
4.2.3 Grid requirements

There are a couple of requirements that need to be met by the case company
to be connected to the national grid. This is in accordance with the national
network code, the Nordel grid. The company must follow the requirements
related to the frequency stability and voltage stability [23] [24]. According to
the Nordic grid code, the frequency on the grid should be kept between 49.9 and
50.1 Hz. In cases where the frequency deviates from this interval, the frequency
controlled disturbance reserve is activated to stabilize the frequency [25]. As
such, the wind farm needs to be able to operate connected to the grid within
the frequency interval. The highest operating voltage of the grid is at 420 kV
and the lowest is normally 90 percent of the nominal voltage, in this case 400 kV.
Additionally, there is a bilateral contract regulating the reactive power exchange
at the PCC [26].

4.2.4 Wind turbine types

There are two types of wind turbines within the wind farm. The two types are
referred to as Type A and Type B. There are a total of 60 wind turbines of
Type A all located in section I of the wind farm and there a total of 39 wind
turbines of Type B all located in section II.

Table 3: Wind turbine specifications


Parameter Type A Type B
Rated Power [MW] 2.5 2.570
Power factor +0.9/-0.95 ±0.95

Table 3 display a short summary of the parameters of the two types of wind
turbines. The Type A wind turbines are of type 3 wind turbine design and the
Type B wind turbines are of type 4 wind turbine design.

4.2.5 Cable

The internal grid of the wind farm is connected using underground cables. There
is only one type of cable in this wind farm. It is used throughout the wind farm,
in both sections. The parameters of the cable are presented below in table 4.

Table 4: Cable parameters


Parameter Value
Conductor material Aluminum
DC-Resistance (20◦ C) [Ω/km] 0.125
Capacitance (20◦ C) [µF/km] 0.26
Max. Operation Temp [◦ C] 90

21
The parameters are given in dc-operation conditions and as the wind farm op-
erates in ac conditions the parameters values should be adjusted for these op-
erating conditions. This suggests that the given parameters in table 4 are not
precise for ac operation. For instance, the cable resistance is larger in ac con-
ditions than in dc conditions. This is due to the skin effect and arguably the
proximity effect. The length of the cables between the connection points and
the wind turbines are known and varies as the distance between wind turbines
varies [27].

4.2.6 Three-winding transformers

There are several transformers within the wind farm, out of which two are
three-winding transformers. These transformers are of the same type.

Table 5: Summary three-winding transformers parameters


Parameter Value
Highest rated power 150 / 75 / 75 [MVA]
Voltage 415 ±1.33%/33/33[kV ]
Highest Rated Current 209 / 1312 / 1312 [A]
Connection Yo / d / d
Phase Displacement YNd11d11
Frequency 50 [Hz]
Load losses (Guaranteed) 365 [kW]
No Load losses (Guaranteed) 58 [kW]

Table 5 displays the specifications of the three-winding transformers. One of


the three-winding transformers connects section I to the grid and the other one
connects section II to the grid, at the PCC.

4.2.7 Two-winding transformers

There are several two-winding transformers within the wind farm. There is
one unique two-winding transformer. It is located in section II of the wind
farm. It connects three Type B wind turbines to the substation. It is the
only transformer in the wind farm that increases the voltage of the three wind
turbines from 22 kV to 33 kV, see table 6.

There are two other types of two-winding transformers in the wind farm, the
wind turbine transformers (WTT). The two-types are the WTTs in section I
and the other is the WTTs in section II. The WTTs in section I and section II,
increase the voltage from 0.660 kV to 33 kV and 0.690 kV, respectively. The
specifications of the WTTs in section I of the wind farm can be found in table
7 and the specifications of the WTTs in section II can be found in table 8.

22
Table 6: Summary Two-winding transformer parameters
Parameter Value
Highest rated power 10 [MVA]
Voltage 33 ±2x2.5%/22[kV ]
Highest Rated Current 174.95 / 262.43 [A]
Connection D / Yo
Phase Displacement Dyn11
Frequency 50 [Hz]
Load losses (Guaranteed) 33 [kW]
No Load losses (Guaranteed) 8 [kW]

Table 7: Summary Section I WTT transformer parameters


Parameter Value
Highest rated power 3.0 / 3.0 [MVA]
Voltage 33 / 0.66 [kV]
Highest Rated Current 55.25 / 2624.3 [A]
Connection D / Yo
Phase Displacement Dyn11
Frequency 50 [Hz]
Load losses 18.000 [kW]
No Load losses 5.360 [kW]

Table 8: Summary Section II WTT transformer parameters


Parameter Value
Highest rated power 3.280 / 3.280 [MVA]
Voltage 33 ±2.5%/0.690[kV ]
Highest Rated Current 57.39 / 2744.5 [A]
Connection D / Yo
Phase Displacement Dyn5
Frequency 50 [Hz]
Load losses 23.098 [kW]
No Load losses 4.950 [kW]

23
4.3 Model of the existing wind farm

In this chapter, the information related to the specific assumptions and data of
the existing wind farm together with how the model is set up is provided.

4.3.1 Case model description

The model of the wind farm is needed to visualize the system and to gain a better
understanding of how the wind farm is designed, to make a simplified version
of the system. Through a developed model it is easier to explain and illustrate
reality. The model provides information about the system that perhaps are
otherwise overlooked. It could help explain events that occur within the system.
In this case, the model can help provide information about the power flows
within the wind farm and perhaps give an indication to the source of the power
losses and the power consumption. Hence, the model can help with identification
of power losses that may be overlooked. Finally, the model can give some sort of
basis to predict outcomes and give reason to the outcomes. It can also help to
improve the explanation to the occurrences within the system, the wind farm.

Figure 4: Wind Farm Single-line Overview

Figure 4, illustrates a very simplified model of the wind farm. As explained


earlier, the wind farm is connected via two three-winding transformers to the
Swedish national transmission network, the TSO. The three-winding transform-
ers are connected to four substations that connect further to the wind turbines.
The substation also connects to four local transformers. The two local trans-
formers in section II are idle and the two local transformers in section II are not
idle. Which provides power to other facilities within the wind farm.

24
Figure 5: Wind turbine connection

Figure 5, illustrates more specifically an example of the setup of one connection


of the wind turbines that connect to the substations. This is an example, but
other wind turbine connections look similar. As mentioned earlier, the wind
turbines are divided into clusters as table 2 shows. Together with the overview
information of the wind farm and presented components parameters values, the
single line diagram was developed.

4.3.2 Case specific model assumptions

Additional assumptions are made for the specific environment surrounding the
case study wind farm. These assumptions are specific and separate from the
ones presented in the methodology section. The assumption made are listed
below:

• The wind farm operates in ac conditions;


• The line-to-line voltages across the wind farm are assumed to be symmet-
rical in order to decrease the complexity of the model. This means that
the operation of the wind farm is balanced over the three phases;
• The initial estimate of the resistance of the cable - The resistance of the
cable is assumed to be 10% larger than what is given in the specification
of it. Hence, the resistance of the cable is assumed to be 0.1375 Ω/km.
This is because the resistance of the cable in the cable specifications was
given in dc and not ac, this means that factors such as the skin effect were
not taken into consideration. Hence, the expected increased value of the
cable resistance;
• The cable operation temperature is set at 60 ◦ C, two thirds of its maximum
operating temperature. This is an initial estimate, which will be evaluated.
The resistance of the cable varies with the operation temperature [28];
• The load losses and the no load losses in the transformers assumed are
the ones mentioned in the test reports as the guaranteed losses, tables 5,
6, 7, and 8 for the assumed values. This is to simplify and decrease the
complexity of the model. As the test reports show, the measured power
losses within the transformers vary due to several parameters, for example
the tap position of the transformer as well as the voltage across the trans-
former. Without data it is almost impossible to predict the precise power
losses of the transformer at any given time. Thus, a standard is needed.

25
Therefore, the guaranteed values of the test reports are assumed. This
assumption should not impact the validity of the model. This is because
the power losses of the different states of the tap positions should cancel
out each other as several data points are evaluated.

4.3.3 Characteristics of the wind farm sections

The data for 2019 consisted of 8760 data points. As mentioned in the method-
ology the data was filtered and after filtering, 7420 data points remained. The
data for the wind farm came from two different sections of wind turbines ex-
plained previously. The two different sections had unique characteristics shown
below.

Table 9: Existing measured data - 2019


Section I: Section II:
Type of Data
at wind turbine at wind turbine
Active power [MW] X X
Reactive power [MVar] X X
Active power de-icing system [MW] - X
Voltage (line-to-line) [kV] X -
Local transformer load [MW] X X
Active power (at PCC) [MW] X X

Table 9 shows the data available for the year of 2019. The annual hourly data
from the wind turbines has been acquired from the certificate meters that are
located as close as possible to the generator, to exclude internal losses of the
wind turbine. Implicitly, the certificate meters are located on the 0.66 kV and
the 0.69 kV side of the WTT transformer of the Type A and the Type B
wind turbines, respectively. The annual hourly data from the grid is evidently
acquired from the meter at the PCC.

The measured active and reactive power at the wind turbines are the annual
hourly average measurements and so are the measured active power output
to the grid at the PCC and the active power loads of the local transformers.
Thus, there may be some inconsistencies as the production may differ at times
to a great degree from the average. Through examination of the active power
load data of the local transformers it was concluded that the loads were to be
neglected for the model. This was because the consumption consumed less than
0.1% of the power production combined. The voltage data acquired from the
wind farm was the measurements at the Type A wind turbines in section I. The
voltage data for the Type B wind turbines in section II does not exist.

26
Figure 6: Acquired data

Figure 6, displays a very simplistic model of where the data is acquired for the
model. Where Pgen and Qgen is the active and reactive power generation data,
respectively. Where Pload is the active power load data and where the Pgrid is
the data of the output active power supplied to the grid from the wind farm. Z
is the impedance of the cable, within the wind farm.

4.3.4 Active power, reactive power, and voltage measurements

The measurements for the active power, reactive power and voltage are illus-
trated and described below for the case wind farm. The figures include all data
for 2019.

Figure 7 displays the active power data of the wind farm. The graph shows the
sum of the active power production of all the wind turbines, together with the
active power output to the grid at the PCC for every hour of 2019.

27
Figure 7: The active power production of all wind turbines and the active power
output to the grid

Figure 8 displays the reactive power data of the wind farm. The graph shows
the sum of the reactive power production of all the wind turbines of section I
and section II for every hour of 2019.

Figure 8: The reactive power production of all wind turbines in section I and
section II

28
The voltage data at the Type A wind turbines in section I is the measured
annual hourly mean line-to-line voltage of each phase. Through examination of
the voltage data, it is noticed that each line-to-line voltage barely differs from
one another. Thus, it is assumed that the system is balanced between the three
phases, as the difference is very small, as figure 9 displays.

Figure 9: Voltage line-to-line percentage deviation

Figure 9 displays that the line-to-line voltages differ less than +/- 0.5 percent.
Therefore, the average of the line-to-line voltages is utilized to model the wind
farm, equal magnitudes of the three phases is assumed.

29
Figure 10 displays 200 data points of the active power and the reactive power
of both sections. The figure also includes the voltages of two wind turbines in
section I. This is to display the correlations between the quantities, within the
two sections. The voltage data is from section I, as there is no available data of
the voltages in section II.

Figure 10: The data for 200 data points; the active and the reactive power
produced and the voltage in per unit.

The base power assumed here, displayed in figure 10, is different for the two
sections. The power is normalized to the maximum power per section. Thus,
the per unit value of section I is not the same as the per unit value of section
II. To clarify, the active power production is not the same for the two sections.
The power shown in the figure, figure 10, is the normalized one for each section
of the wind farm.

4.3.5 De-icing data

In the case study wind farm, there is hourly active power consumption data
of the de-icing system of the Type B wind turbines, in section II. The data of
the de-icing system of the Type A wind turbines, in section I, does not exist.
Therefore, the data of the de-icing system of the Type A wind turbines was
derived for the purpose of accuracy of the model to reality. Furthermore, this
is also to comply with the argument that if the de-icing system of the Type B
wind turbines are to be taken into consideration, then the de-icing system of
the Type A wind turbines should also be taken into consideration.

30
The active power consumption data of the de-icing system of the Type A wind
turbines are created from the data of the de-icing system of the Type B wind
turbines. Therefore, some assumptions are to be made to be able to create fair
and viable data. The assumptions are listed below.

• The de-icing system of the Type A wind turbines at section I consumes


40% of the de-icing system of the Type B wind turbines at section II. This
is due to the design of the two different de-icing systems. Although, the
percentage is perhaps somewhat arbitrary, but it is an initial estimate of
the de-icing system at section I and will be evaluated further;
• The time differential of the Type A wind turbines and the Type B wind
turbines de-icing systems are the same i.e. as the de-icing systems are
either ON or OFF, the period of the de-icing systems are ON the same
amount of time each hour;
• The de-icing system of the Type A wind turbines work similar to the de-
icing system of the nearby Type B wind turbines. Thus, somewhat taking
into consideration similar weather conditions; temperature, humidity etc.
Arguably, this assumption is rather vague. However, this is to utilize the
data of the de-icing system of the wind turbine in the area that supposedly
is the closest approximate to the data missing.

The data of the de-icing system of the Type A wind turbines was created from
nearby Type B wind turbines. It was created from different Type B wind
turbines to make sure to ensure diversity to the de-icing system data for the
Type A wind turbines. This means that the data was not created from just
one wind turbine. Accordingly, the de-icing system is not ON for all the wind
turbines at the same time, which the created data now supports.

Figure 11: The de-icing data of section I and section II

31
Figure 11 displays the de-icing system power consumption data of both sections.
The figure includes all data for 2019. To be noted, the de-icing system power
consumption of section I was derived from the de-icing system of section II. This
means that the data of the de-icing system in section I is not a true represen-
tation of the actual de-icing system of section I, however it is an estimate. It
should also be noted that the displayed data of the de-icing system of section I
is the initial assumption that it is 40% of the section II de-icing system.

4.3.6 Model setup

The model of the existing wind farm is set up as a load flow model. The first step
to create the model is to set up the single-line diagram with the components,
loads and the connection points, accordingly. The parameters of the components
are defined using the data presented earlier in the component tables, tables 3,
4, 5, 6, 7, 8. The cable lengths between the connection points throughout the
wind farm are known.

After the single-line diagram was developed, the types of buses used for the
model were identified. The buses were defined as PQ-buses, as there was active
and reactive power production data and load data for all the wind turbines in
the wind farm, but there was missing voltage data for the Type B wind turbines
in section II. Therefore, to be consistent in the model the same buses were used
throughout, meaning that all the wind turbine buses were set up as PQ-buses.
The bus-bars were modelled as PQ-buses with zero active or reactive power
injection. The bus-bars with loads connected to them were also modelled as
PQ-buses. The slack bus was set up at the PCC of the wind farm and the grid.

Thus, the input data of these buses correspond to the defined types of buses.
Table 10 presents the data for the load flow model, together with the parameters
that are to be estimated.

Table 10: Model data


Parameter Cable resistance Operation De-icing
estimation (Rref ) [Ω/km] temperature, (T ) [◦ C] Section I [%]
Active power production data [MW]
Model
Reactive power production data [MVar]
Input data
De-icing consumption data [MW]
Model Active power output (PCC) [MW]
Output data Voltage data (wind turbines) [kV]
Validation Active power output (PCC) [MW]
data Voltage data (wind turbines) [kV]

As table 10 shows, the model input data is the power generation data of the
wind turbines, the active and reactive power production data, as well as the de-
icing system power consumption data at the wind turbines. This is the de-icing
system of both sections in the wind farm, where the de-icing data is the derived
data from section II. The model output data is the output data of the load flow
calculations. This includes the active power output at the PCC from section I

32
and section II of the wind farm at the high voltage side of the two three-winding-
transformers and the voltage data at the wind turbines. The validation data
is the data that the model output data is to be compared with. Hence, the
validation data consists of the measured data of the active power output at the
PCC and the voltage data at the wind turbines. This is the voltage data for
the wind turbines in section I, not for section II as this data does not exist.

Furthermore, the model of the wind farm is divided into two. One model for
section I and one model for section II. This is because there is a lack of validation
data for section II, as there is no voltage data. It is possible to divide the
wind farm in two because of the design of the wind farm and because it uses
the same components throughout the wind farm, other than the type of wind
turbines. Thus, the model parameters can be estimated from section I and then
the estimated parameters can be used for section II. The extra power losses due
to the additional transformer in section II is estimated using the data for section
II and the estimated model parameters from section I. Although the validation
data for section II only consists of the active power output to the grid at the
PCC, it can be used as an indication for the model.

The training session of the model is carried out and the named parameters in
table 10 are to be estimated using section I. The de-icing system data of the
wind turbines in section I was derived from the existing wind turbines in section
II. However, due to the design difference of the two de-icing systems, the de-
icing system of section I consumes less power than the de-icing system of section
II. Thus, the initial estimate for the de-icing system of section I derived from
section II is set as a percentage of the derived de-icing system data, 40%. This
percentage is to be further estimated.

The resistance of the cables and the operation temperature are also to be es-
timated, as the resistance of the cables change depending on the temperature.
Equation 5 shows the correlation between the resistance of the cable and the
temperature of the cable.

R = Rref (1 + α(T − Tref )) (5)

Where R is the conductor resistance at temperature T in ◦ C. Rref is the conduc-


tor resistance at the reference temperature Tref in ◦ C. The α is the temperature
coefficient of the resistance of the material of the conductor. The value of the
α for the aluminum cable used is 0.004308 [29].

The operation temperature estimate is an overall operation temperature set for


all the data points. The initial value of the resistance was set to +10 % of the
given dc-resistance. The initial operation temperature was set to 60 ◦ C. The
estimation process of the model parameters in the training session is completed
when the error estimates, the MSE, the RMSE and the weighted RMSE reach a
plateau where they stop decreasing. It is an iterative process as figure 2 shows.
In this thesis the weighted RMSE was to be minimized. The weight was set to
favor the active power output to the grid at the PCC above the voltages at the
wind turbine buses. This was to assure that the power losses in the model of
the wind farm were better represented.

33
The resistance Rref , and the operation temperature T , of the cable was es-
timated together, see equation 5. The two were tuned until the plateau was
reached, and the weighted RMSE did not decrease anymore. Thereafter, the
percentage of the de-icing system of section I was estimated, and until the
weighted RMSE did not decrease.

The parameters estimated in section I were then used to evaluate the power
losses in section II, the same estimated cable resistance and operation temper-
ature. Furthermore, the power losses in the additional transformer in section
II was estimated. The magnetic losses in the transformer were modelled and
estimated with a resistor in parallel connected to the ground. This was also
done iteratively, until the weighted RMSE reached the plateau and could not
be minimized further.

34
5 Results

In the following chapter the results from the case study are presented. As the
study is largely quantitative, the focus of the result lies on the output from the
load flow analysis, the quasi-dynamic simulations from the model.

5.1 Framework for evaluating results

The results produced by the load flow model are measured in terms of accuracy.
That is, the difference in the output from the model compared to the historical
real system measurements. The target variables for each forecast of 2019, for
section I, was the measured voltage at the wind turbines and the measured active
power output to the grid at the PCC for every hour of. The target variable for
each forecast of 2019, for section II, was the measured active power output to the
grid at the PCC. To clarify, as the data was not consistent and the quality was
somewhat questionable, the data was pre-processed. Additionally, the unviable
data points were filtered out. Thus, the number of hours of data, the number
of data points available decreased. The data points with missing voltage data
were filtered out, as well as the data points where the difference between the
active power production of the wind turbines and the active power output to
the grid was too great. Figure 12 displays the two datasets, the training dataset
and the verification dataset.

Figure 12: Number of data points in each dataset

The total number of data points available, the data points considered for the
model, was 7420 after the data was filtered. This was the number of data points
deemed viable that fulfilled the criteria. Thus, the training dataset consists of
75% of the data points available (5565) and the verification dataset consists of
the left 25% of the data points available (1855).

5.2 Parameter estimates

In this chapter, the performance of the evaluated parameters is displayed. The


cable resistance, operation temperature and the de-icing system data of sec-
tion I was estimated utilizing the training dataset for section I. The additional
transformer in section II was also evaluated utilizing the training dataset, but
for section II and with the estimated parameter values from section I.

35
Table 11 displays the initial and the final value of the cable resistance, the
operation temperature as well as the percentage estimate of the derived de-
icing system power consumption data. Table 12 displays the error estimates
corresponding to the initial and the final conditions, respectively, for section I.

The resistance of the cables was given as the dc-resistance in the data sheets
provided. As mentioned previously, the ac-resistance of the cable is estimated
to be higher than the dc-resistance due to the skin effect. Therefore, as the wind
farm operates in ac-conditions the resistance of the cable is estimated within the
interval of +0% to +50% of the dc-resistance given. The Operation temperature
was estimated within the interval of 40 and 90 ◦ C and the de-icing system was
estimated between 40% and 70% of the de-icing system of section II.

Table 11: Section I: Estimated parameters


Resistance Operation De-icing
Conditions
(Rref ) [Ω/km] temperature, (T ) [◦ C] I [%]
Initial 0,1375 60 40
Final 0,1531 50 60

Table 12: Section I: results estimated parameters 5565 data points


MSE MSE
Conditions RMSE wRMSE WU WP CC
Ug PP CC
Initial 8,18E-04 1,67E-04 8,35E-04 1,43E-03 2 25
Final 8,14E-04 1,56E-04 8,29E-04 1,39-03 2 25

The MSE Ug for the voltage at the wind turbines and the MSE PP CC for the
active power output to the grid is calculated. From the MSE values the RMSE
value is calculated. The weighted RMSE is calculated utilizing the MSE values
and the weighted values, WU and WP CC . WU is the weight for the voltage and
WP CC is the weight for the active power at the PCC. The weights were decided
to make sure that the active power output at the grid had more impact on the
model than the voltages at the wind turbines. It is of more importance that
the difference between the measured active power output to the grid and the
simulated active power output to the grid is minimized than the measured and
simulated voltages at the wind turbines, in terms of power losses in the wind
farm.

Furthermore, the weighted RMSE was minimized. The MSE values for the
voltage and the active power output at the PCC indicate the improvement of
the model in accuracy for respective quantities. The RMSE value displayed in
comparison to the weighted RMSE shows the impact of the weighted values. The
weighted values were tested in different combinations, without much difference
in the results. The weighted values displayed were decided as there was a distinct
difference in that the error estimates where to value the active power output to
the grid higher than the voltages. These values decided on are used throughout
the thesis.

36
Tables 13 and 14 displays the error estimates of section II, for the initial and the
final parameters evaluated previously in section I. Furthermore, the magnetic
losses in the additional transformer, T4, in section II are modelled and estimated
with a parallel resistor. As there is no voltage data for the wind turbines in
section II, the only error estimate is the MSE for the active power at the PCC.

Table 13: Section II: Estimated parameters


Resistance Operation T4
Conditions
(Rref ) [Ω/km] temperature, (T ) [◦ C] .[kΩ]
Initial 0,1375 60 100
Final 0,1531 50 142

Table 14: Section II: results operation parameters 5565 data points
MSE MSE
Conditions RMSE wRMSE WU WP CC
Ug PP CC
Initial — 1,98E-04 — — — —
Final — 1,94E-04 — — — —

5.3 Model performance

In this chapter the performance of the trained model is evaluated utilizing a new
set of data. This is to investigate if the parameters estimated in the training
session of the model are properly estimated, and if the estimated model performs
well against the independent set of data not used for the estimation process.
The verification dataset is utilized, which provides the model with new data
points, specifically, new scenarios to evaluate the error of the model with. The
verification dataset consists of the 25% left data points of the available data
points that were not used to train the model. By exposure to new scenarios,
flaws of the model could be uncovered, for instance if parameters were not
estimated correctly. It could also provide the information that the model does
not support other scenarios and cannot be used to evaluate or estimate the
power losses.

Table 15: Section I: results operation parameters 1855 data points


MSE MSE
Conditions RMSE wRMSE WU WP CC
Ug PP CC
Initial 9,58E-04 1,82E-03 2,06E-03 9,19E-03 2 25
Final 9,61E-04 1,81E-03 2,05E-03 9,16E-03 2 25

Table 15 and table 16 displays the error estimates of the initial and final con-
ditions of the model for the verification dataset of section I and section II. The
initial condition is the initially guessed model parameters with the correspond-
ing error estimate values. The final condition is the finally estimated parameters
of the model with the corresponding error estimate values. As explained earlier,
there is no voltage data for section II at the wind turbines, and therefore the

37
Table 16: Section II: results operation parameters 1855 data points
MSE MSE
Conditions RMSE wRMSE WU WP CC
Ug PP CC
Initial — 1,064E-03 — — — —
Final — 1,062E-03 — — — —

MSE of the active power output to the grid is the only error estimate for section
II.

Through comparison of the results presented in tables 12, 14, 15 and 16, it is
obvious that the model performed better when exposed to the training dataset
than when exposed to the verification dataset.

The tables 12 and 15, indicate that the accuracy of the voltage barely changed,
as the MSE Ug barely changed, when comparing the results of the training
dataset and the verification dataset. However, the results show that the accuracy
of the active power output to the grid at the PCC decreased as the value for
the MSE Ppcc increased by a factor of about 10 when comparing the results of
the training dataset and the verification dataset for section I. Furthermore, the
RMSE value increased by a factor of about 2.5, when comparing the results of
the two datasets for section I.

The tables 14 and 16 indicate that the accuracy of the active power output to
the grid at the PCC decreased for section II when comparing the results for
the training dataset and the verification dataset. The MSE Ppcc increased by
a factor of about 5 for section II comparing the two datasets.

5.4 Estimated power losses

The power losses are estimated and compared for the wind farm utilizing all
the available data points to the model, the 7420 data points. This gives an
indication to the magnitude of the power losses for the data points evaluated
and an indication to the main cause or source to the power losses during the
year of 2019.

Table 17: Total losses over the observed period of time


Total production [TWh] 525
Real measured losses [TWh] 20.2
Estimated losses [TWh] 17.2
Real measured losses [%] 3.85
Estimated losses [%] 3.28

Table 17 displays the total amount of energy produced by the wind turbines of
the wind farm during the observed time, the 7420 data points. The table also
displays the measured losses of the wind farm and the estimated losses of the
model. Additionally, the measured and the estimated losses are presented as a

38
percentage of the total energy produced. The de-icing system is excluded from
the measured and estimated losses.

5.4.1 The de-icing system

As previously mentioned, the de-icing system within wind farms can consume
rather much power during the year, it depends a lot on the climate and the
weather. The power consumed by the de-icing system of the wind turbines
could have been power supplied to the grid. Arguably, the two different de-
icing systems are expected to consume different amounts of power due to their
designs. Thus, the share of power losses within each section that the respective
de-icing systems constitute for is expected to somewhat differ. Figures 13 and
14 displays the estimated power losses as well as the power consumption of the
de-icing system in each section, respectively.

Figure 13: Comparison of the active power losses and the consumed active power
by the de-icing system in section I

To clarify, figure 13 shows the estimated active power losses in section I and the
active power consumption of the de-icing system in section I. Figure 14 shows
the estimated active power losses in section II and the active power consumption
of the de-icing system in section II. It should be noted that the de-icing system in
section I is derived from the de-icing system in section II, it is a rough estimate.
The estimated active power losses in each section includes the component power
losses. To be noted, this is the estimated power losses and the de-icing system
data for 7420 data points, the data points available, and not the whole year.
However, this gives an indication to how large the shares of the de-icing systems
constitute for the estimated power losses of respective sections in the wind farm.

39
Figure 14: Comparison of the active power losses and the consumed active power
by the de-icing system in section II

The figures 13 and 14 displays that the active power consumed by the de-icing
system of each section takes up a rather large share of the power losses. The de-
icing system of section II takes up a significantly larger share than the de-icing
system in section I, compared to the power losses in each section, respectively.
It is estimated that the de-icing system of section I and II constituted for about
16.2% and 32.0% of the power losses, respectively.

40
5.4.2 The additional transformer

The power losses due to the additional transformer, T4, in section II was esti-
mated. This was done by modelling and estimating the magnetic losses with a
parallel resistor connected to ground. For the available data points of the year
of 2019, 7420 data points, the power losses of section II are estimated. Thus,
the figure 15 shows the power losses with the additional transformer and the
scenario where the transformer is removed, without the additional transformer.

Figure 15: Comparison of the active power losses with and without the extra
transformer in section II

If the transformer were removed from the wind farm and the necessary modi-
fications of the wind farm was done, the power losses would decrease by about
3.4% in section II according to the data. This would increase the total active
power output to the grid by about 0.08% from section II.

41
6 Analysis and discussion

This chapter analyses and discusses the results and puts them into the theoretical
context of what was presented in the research background. Additionally, relevant
points regarding sustainability are brought up and explored.

6.1 Scrutiny of results

The results of the model are first and foremost based on the training session
and then the verification session of the model. The training session of the
model, where 75% of the available data points were used, proved that through
modification of the initial estimated parameters it was possible to increase the
accuracy of the model. The verification of the model, where 25% of the available
data points were used, indicates that the estimated model parameters of the
training of the model could be used for future scenarios.

The results of the MSE, the RMSE and the weighted RMSE display a measure
of accuracy of the model. The weighted RMSE depends on how the quantities
are weighted against one another, whereas the MSE and the RMSE does not.
The weight for the calculations of the weighted RMSE was set to favor the active
power output to the grid over the voltages at the wind turbines.

The tables 12 and 14 show that from the initial results of the MSE, RMSE
and the weighted RMSE it was possible to enhance the accuracy of the model
by modification of the parameters shown in Table 11. As the resistance of the
cables together with their operation temperature and the percentage of the de-
icing system of section I was modified, the error estimates values improved.
It was possible to improve the accuracy of the model for section I. It is also
possible that the accuracy of the model for section II was improved with the
modification of the parameters, as the MSE PP CC decreased from the initial
value to the final value. Although, this is arguably not enough to say that
the overall representation of section II is good enough as it does not take the
voltages at the wind turbines into consideration. As the MSE, the RMSE and
the weighted RMSE values improved by the modification of the parameters, the
power losses estimated became more accurate. Thus, the difference between
the simulated and the measurements of the active power output to the grid
decreased.

Tables 15 and 16 shows that the final parameter estimates represented the wind
farm for future scenarios better than the initial parameters estimates. Table
15 also shows that there is a slight trade-off between the voltage accuracy and
the active power output to the grid at the PCC accuracy with the estimated
parameters for section I for future scenarios. The error estimates show slight
improvement, for section II. It shows that the method of utilizing estimated
parameters of section I for section II was possible. Although, this is taking into
consideration the active power output to the grid and not taking into considera-
tion the voltages at the wind turbines. As explained earlier, this was due to the
lack of validation data for section II. Thus, it is unknown if the model accuracy
improved in terms of voltage for section II.

42
The verification of the model does not entirely prove that the model can be used
for future scenarios, as the results of the error estimates of the model shows
worse accuracy for the verification dataset than the training datasets in tables
12, 14, 15 and 16. As mentioned before, the results show that the accuracy of
the voltage for section I did not change much depending on whether the model
was exposed to the training or the verification dataset. However, the results of
the error estimates shows that the accuracy of the active power output to the
grid at the PCC decreased, as the MSE Ppcc increased by a factor of about 10
and 5 for section I and section II, respectively.

The underlying cause to the decreased accuracy for the model when exposed
to the verification dataset than when exposed to the training data set is that
the model parameters were estimated using the training dataset. The error
of the model was meant to be minimized for the training dataset. Hence, the
accuracy of the model for another dataset would probably be worse. However,
it is interesting that the error estimates show such a large difference between
the two datasets in accuracy. The reason for this could be that the data points
in the verification dataset are of worse quality than the data points in the
training dataset. There were more missing production data points that needed
to be filled in the verification dataset than in the training dataset. Hence, it is
difficult to determine how well the model performs for future scenarios by only
comparing the error estimates from the training dataset and the verification
dataset. Therefore, it could be useful to use an additional dataset to determine
the performance of the model for future scenarios.

Table 17 indicate that the model was fairly accurate. The difference of the
estimated energy losses and the measured energy losses for the observed period,
the 7420 data points, differ by about 0.57 %. This indicates that the model is
an overall good representation of the wind farm in terms of power losses, and
the model could be used as a tool to further identify the power losses in the
wind farm if more data is acquired.

The estimated parameters were used to estimate the power losses of the model
for 7420 data points, out of 8760 data points, for the year of 2019. The estimated
power losses are compared to the de-icing system power consumption for each
section of the wind farm. It shows that the share of power losses that the de-
icing systems constitute for is much larger for section II than for section I. This
may be because of how the de-icing system data for section I was derived and/or
that the estimated percentage was inaccurate. However, it seems that the most
likely reason for this is that the de-icing system for the wind turbines in section
II consumes more power in general compared to the de-icing system for the
wind turbines in section I, due to the designs. This could also indicate that the
de-icing system of section II is running an excessive amount as it constitutes for
a rather large portion of the power losses in section II.

The additional transformer exists due to the design of the wind farm. If the
transformer were removed, the power losses of section II would somewhat de-
crease. However, it constitutes such a small portion of the power losses within
section II, that it does not seem to be the reason for the power loss difference
between the two sections.

43
6.2 Data quality

The data measured was for every hour of the year of 2019. Out of the 8760
hours, 8760 data points for the model, 7420 data points were deemed viable to
use. The first 75% of the data points were used in the training session of the
model. The last 25% data points were used to verify the model. Arguably the
data quality of the measurements of the wind farm was not that great.

The combined actual active power production of the wind turbines compared
to the actual active power output at the PCC was to lie within the set interval.
This was to eliminate outliers within the data set, as well as to assure a certain
quality of the data. The elimination process takes the missing data points, the
missing production data into consideration, as well as outliers in the data, where
there was unreasonably high power losses in the wind farm. Thus, 7420 data
points of the 8760 data points available was deemed viable to use.

The data available was inconsistent. There were missing data points of produc-
tion for almost all wind turbines, even though the data was sorted. The load
data for the de-icing system of the section I, Type A wind turbines, did not exist
and therefore it was derived from the data of the de-icing system of section II,
Type B wind turbines. The power load of the de-icing system at section I was
an initial rough estimate from section II. One would need actual data of the
de-icing system of section I to be able to interpret the results of the wind farm
better, and to make a more accurate model to reality.

The production power data missing was filled with similar wind turbines pro-
duction data. The data was taken from nearby wind turbines, assuming that
the wind speed was similar. Hence, the data of power production of the wind
turbine missing would be like the power production data of the wind turbine
the data was taken from.

Arguably, the data quality of the input data of the wind farm is greater at
section II compared to section I. This is solely due to the fact that the de-icing
data of section I is missing, whereas it is not at section II. The data quality
of the validation data of the wind farm is far greater at section I compared to
section II. This is because of that the voltage data is missing at the section II,
Type B wind turbines. Thus, section II is missing an element of validation.
Hence, section II can only be evaluated through one measure, the active power
output at the PCC.

6.3 Source of error and model improvement

The model was built on somewhat questionable data. There are data points
where there was for instance missing power generation data of a couple of wind
turbines, which was filled using the data of nearby wind turbines. These data
points contain estimates of the different quantities. This means that some of
the data points used for the development of the model are not entirely true to
reality. Thus, the data is to some extent unreliable. Therefore, an improvement
to the model would be to improve the data used, make sure that the model is
based upon true values.

44
Another improvement option of the model would be to add more data or evaluate
and develop the model with more data. To make it more reliable, more than
one year of annual data for the model could be used.

It could be interesting to look at and estimate additional parameters of the


system components. This would suggest that a more overall accurate model of
the wind farm could be developed. It could also provide more information of
the components and the sources of power losses within the wind farm, and to
what extent these factors contribute to the power losses.

6.4 Sustainability

The benefits of decreasing the power losses to society in a wind farm is as


follows. Arguably, the benefits can be categorized as economic, ecological, and
social sustainability. Improving the efficiency of the wind farm, which means
improved power production from wind power. It becomes more important with
green sources of energy in the future, thus improving the wind power works
towards the United nations SDG 7th, clean and affordable energy production.

• Economic sustainability – Increasing the efficiency of the wind farm, pro-


ducing more power, or decreasing the power losses, means that potentially
the amount of power sold by the company increases. Increases the profit
of the wind farm.
• Ecological sustainability – Improving the power production of the wind
farm, producing more power, or decreasing the power losses, increases the
amount of green power produced to which in theory has the potential to
replace non-renewable power production.
• Social sustainability – Increasing the efficiency of the wind farm suggests
that more customers can invest in green power production or consumption
of green power. The customer has a positive impact on society by using
green power, contributes to a greener world and does not contribute to
the greenhouse effect.

45
7 Conclusion

In this chapter the conclusions drawn from this thesis are presented together
with proposals for future work.

7.1 Main findings

The selected approach to use load flow analysis to develop a model for power
loss source identification proved possible for the existing wind farm. A fairly
accurate model was developed in terms of power losses. The model was used to
estimate and compare the power losses of the wind farm and identify the source
to where there seem to be excessive amounts of power losses.

The additional transformer in section II exists and it is probably not of use


to remove it from the wind farm. The additional transformer in section II
constitutes for about 3.4% of the power losses in section II, which is less than
0.08% of the yearly produced power in section II. The internal power losses of
the additional transformer do not prove to be the main factor to the difference
in power losses between the two sections.

The de-icing system of section II seems to be the main reason for the difference
in power losses between the two sections. The de-icing system of section II
constitutes for about 30% of the power losses of section II, for the data points
available. In comparison, the de-icing system of section I constitutes for about
16% of the power losses in section I. If the de-icing systems were not running as
much during the year, potentially, an additional amount of power could be sold
instead. Hence, it would be of economic interest to decrease the power consumed
by the de-icing system during the year if it does not interfere with safe reliable
power production i.e. a significant amount of icing forms on the blades of the
turbines. Arguably, as the de-icing system data of section I was derived from
the de-icing system data of section II, it is not a true representation. Although,
if the percentage of the derived data of the de-icing system for section I was
increased further, it still would not constitute for as much share of power losses
as the de-icing system of section I. It adds to the point that the power losses
in the two sections compared to one another is significant due to the de-icing
systems.

Another point of interest is the control system of the wind farm. This is because
the reactive power data acquired at the wind turbines shows that the wind
turbines in section I produce a significant amount more reactive power than the
wind turbines in section II. It also shows that sometimes the wind turbines of
section I produce reactive power while section II consumes and vice versa, see
figure 10. Thus, the issue of the reactive power suggests that the control of the
wind farm is not ideal and that there is potential to decrease the power losses
in the wind farm through better control.

46
7.2 Future work

To continue the work with power loss identification and decreasing the power
losses in the existing wind farm, some proposals are presented below. This is to
address the problem and provide some sort of guideline for the future.

The de-icing system in section II should be studied further. According to the


results of the case study, the power consumption of the de-icing system in section
II seems excessive as it constitutes such a large portion of the power losses in
section II. This could mean that the de-icing system consumes an unnecessary
amount of power.

The control system of the existing wind farm should be studied further. This
is due to that share of produced reactive power between the two sections are so
different, and that during times one section consumes while the other produces
reactive power and vice versa. It could also be of interest to study whether it is
possible or not to change how the reactive power production of the wind farm
is controlled, as there is a potential loss of active power.

To further study the power losses of the wind farm it could be an option to
put measurement meters on the wind turbines to study the internal power con-
sumption of the wind turbines. Potentially, unnecessary power losses of the
wind turbines could be identified and improvement options for the wind tur-
bines could be of interest. It could also be of interest to put measurement
meters on other components, such as the main transformers and see if they are
operating correctly.

To improve the load flow model further, there is a need for more consistent data.
The de-icing system of the wind turbines in section I should be monitored and
recorded. To develop a more accurate model, the voltage of the wind turbines
in section II should also be monitored and recorded. This would suggest that
a model of the whole wind farm together could be developed, additionally this
could give a better overview of the power losses within the wind farm. Moreover,
other parameters of the components of the wind farm could be estimated further.
If there were more viable data points, it could be useful to develop the model
utilizing data from more than one year. This would probably lead to a more
reliable model and a more accurate representation of the wind farm, that could
be used to further identify the sources of the power losses within the wind farm.

47
References

[1] C. Philibert, “International energy technology collaboration and climate


change mitigation,” OECD/IEA, France, 2004.
[2] “Paris agreement,” United Nations, UNFCCC, December 2015. [Online].
Available: https://unfccc.int/sites/default/files/english paris agreement.
pdf
[3] (2015, November) The 17 goals, sustainable development goals. New York,
US. . [Online]. Available: https://sdgs.un.org/goals
[4] (2019, May) Falling renewable power costs open door to
greater climate ambition. Abu Dhabi, AE. . [Online]. Avail-
able: https://www.irena.org/newsroom/pressreleases/2019/May/
Falling-Renewable-Power-Costs-Open-Door-to-Greater-Climate-Ambition
[5] (2017, June) Vindkraftsstatistik 2016. Energimyndigheten. . [On-
line]. Available: http://www.energimyndigheten.se/globalassets/
statistik/officiell-statistik/statistikprodukter/vindkraftsstatistik/
vindkraftstatistik-2016.pdf
[6] J. Farrell, “Is bigger best report - part 1: Limits to scale in wind,” Is Bigger
Best Report, October, 2017.

[7] M. Froese, “Cracking the icing problem on turbine blades,” Windpower


Engineering Development, March 2018. [Online]. Available: https://www.
windpowerengineering.com/cracking-icing-problem-turbine-blades/
[8] B. Zhang, W. Hu, P. Hou, J. Tan, M. Soltani, and Z. Chen, “Review of
reactive power dispatch strategies for loss minimization in a dfig-based wind
farm,” Energies, vol. 10, no. 7, 2017.

[9] A. Larsson, “The power quality of wind turbines,” Doktorsavhandlingar vid


Chalmers tekniska högskola, 1654, pp. 1–49, January, 2000.
[10] B. K. Perera, P. Ciufo, and S. Perera, “Point of common coupling (pcc) volt-
age control of a grid-connected solar photovoltaic (pv) system,” in IECON
2013 - 39th Annual Conference of the IEEE Industrial Electronics Society,
Nov 2013, pp. 7475–7480.
[11] G. Andersson, “Modeling and analysis of electric power systems,” Lecture
227-0526-00, 2008.
[12] B. Shejal and S. Jamge, “Wind power stabilization to achieve proper grid
connection using power convertor dsp controller,” ICM2ST-10, AIP Con-
ference Proceedings, vol. 1324, pp. 147–150, nov 2010.
[13] L. F. Gaitán, J. D. Gómez, and E. Rivas-Trujillo, “Quasi-Dynamic
Analysis of a Local Distribution System with Distributed Generation.
Study Case: The IEEE 13 Nodes System,” TecnoL, vol. 22, pp. 140 – 157,
12 2019. [Online]. Available: http://www.scielo.org.co/scielo.php?script=
sci arttext&pid=S0123-77992019000300140&nrm=iso

48
[14] (2003, July) Explore the world of gray box testing? TASSQ.
. [Online]. Available: http://extremesoftwaretesting.com/Articles/
WorldofGrayBoxTesting.html

[15] M. Binieli, “Machine learning: an introduction to


mean squared error and regression lines,” October
2018. [Online]. Available: https://www.freecodecamp.org/news/
machine-learning-mean-squared-error-regression-line-c7dde9a26b93/

[16] A. Duun-Henriksen, S. Schmidt, R. Røge, J. Møller, K. Nørgaard,


J. Jørgensen, and H. Madsen, “Model identification using stochastic differ-
ential equation grey-box models in diabetes,” Journal of diabetes science
and technology, vol. 7, pp. 431–440, 03 2013.
[17] (2006, Nov) Sir isaac newton - mathematical laws black box theory. New-
science-theory. . [Online]. Available: https://www.new-science-theory.
com/isaac-newton.php
[18] (2020, Dec) White box testing. Software Testing Fundamentals. . [Online].
Available: https://softwaretestingfundamentals.com/white-box-testing/
[19] J. Hauth, “Grey-box modelling for nonlinear systems,” 2008.

[20] (2019, March) Quasi-dynamic harmonic load flow. DigSilent.


. [Online]. Available: http://www.digsilentiberica.es/noticia/69/
Quasi-Dynamic-Harmonic-Load-Flow/
[21] A. B. Hyndman, Rob J. Koehler, “Another look at measures of forecast
accuracy,” International Journal of Forecasting, no. vol22(4), p679-688,
2006.
[22] Z. Tu, W. Xie, W. Hürst, S. Xiong, and Q. Qin, “Weighted root mean
square approach to select the optimal smoothness parameter of the varia-
tional optical flow algorithms,” Optical Engineering, vol. 51, pp. 7202–, 03
2012.
[23] (2007, Jan) Nordic grid code 2007 (nordic col-
lection of rules). Nordel. . [Online]. Avail-
able: https://eepublicdownloads.entsoe.eu/clean-documents/pre2015/
publications/nordic/planning/070115 entsoe nordic NordicGridCode.pdf
[24] (2013, March) Entso-e network code for requirements for grid
connection applicable to all generators. entsoe. . [Online].
Available: https://eepublicdownloads.entsoe.eu/clean-documents/
pre2015/consultations/Network Code RfG/120626 final Network Code
on Requirements for Grid Connection applicable to all Generators.pdf
[25] (2007, Jan) Nordic grid code 2007 (nordic collection of rules).
Nordel. page 61, section: Definitions - appendix 1. [Online]. Avail-
able: https://eepublicdownloads.entsoe.eu/clean-documents/pre2015/
publications/nordic/planning/070115 entsoe nordic NordicGridCode.pdf
[26] I. Machado and I. Arias, “Grid codes comparison,” Gothenburg, 2006.
[Online]. Available: https://hdl.handle.net/20.500.12380/174164

49
[27] (2015, December) Kraftkabel handboken. nkt cables AB, Falun,
Sweden. . [Online]. Available: https://www.nkt.se/fileadmin/user upload/
NKT kraftkabelhandboken.pdf

[28] (2014, August) Heating effects for remote power de-


livery over bundle cables. Belden. . [Online]. Available:
https://www.newark.com/wcsstore/ExtendedSitesCatalogAssetStore/
cms/asset/pdf/americas/common/techarticles/belden/heating-effects.pdf

[29] CIRRIS, “Temperature coefficient of copper,” Aug 2014. [On-


line]. Available: https://www.cirris.com/learning-center/general-testing/
special-topics/177-temperature-coefficient-of-copper

50
TRITA-EECS-EX-2021:103

www.kth.se

You might also like