Determining and analysing production losses due to ice on wind turbines using SCADA data

(1)

Determining and analysing production losses due to ice on wind turbines using

SCADA data

Oscar Felding

Sustainable Energy Engineering, master's level 2021

Luleå University of Technology

Department of Engineering Sciences and Mathematics

(2)

Abstract

Wind turbines are becoming a more common sight and a more important part in the power grid. The benefits are mainly that wind energy is a renewable energy source and a single wind turbine can produce enough electricity to cover several households’ annual electricity need and not producing carbon dioxide as a rest product. Drawbacks are fluctuation in wind speed, which makes it difficult to regulate. The turbines need to be placed far from cities which cause losses in transmission in the national power grid.

In cold areas with long winters there is a risk of high energy losses due to iced blades. If there is ice accretion on the wind turbine blades it can cause a production loss and in extension economical losses by not selling the electricity. Finding those events is of high interest and there are methods to prevent and remove ice. However, there are occasions when there is ice on the blades, but no sensors signal this, and the production loss is a fact.

There is a presumed production loss of 5-25 % annually due to icing on wind turbines in Sweden, depending on where the site is located. There is no general method for detecting ice in the industry but there are several methods available developed by different parties.

In this master’s thesis, a software has been developed in cooperation with Siemens Gamesa Renewable Energy to identify production losses on wind turbines due to icing using historical SCADA data. The software filters the raw data to construct a reference curve, to which data during cold weather is compared. It was found that low temperature causes ice losses, and the risk of an ice loss increases as temperature decreases. The annual losses at investigated wind farms were 4-10 % of the expected annual production.

Keywords: Wind turbines, ice accretion, ice losses, SCADA, cold climate

Sammanfattning

Vindkraftverk blir en allt vanligare syn och en viktigare del i kraftnätet. Fördelarna är framförallt att det är en förnybar energykälla, det blir inga koldioxidutsläpp när vindkraftverken har installerats och ett vindkraftverk kan täcka flera hushålls årliga elbehov. Nackdelar är att vinden inte går att kontrollera och elproduktionen inte är garanterad eller konstant. Vindkraftverk placeras långt ifrån tätorter, vilket leder till förluster under distribution.

I kalla regioner med långa vintrar uppstår en risk för energiförluster på grund av nedisade turbinblad. Om det finns ispåbyggnad på turbinbladen kan det orsaka produktionsförluster och följaktligen en ekonomisk förlust.

Det är av stort intresse i att upptäcka dessa och det finns flera metoder för att förbygga is och även avisning. Det antas vara produktionsförluster på 5-25 % årligen på grund av is i Sverige, beroende på vindparkens placering.

Det finns ingen generell metod för att upptäcka is inom industrin, men det finns flera metoder utvecklade av olika parter.

I det här examensarbetet har en mjukvara utvecklats i samarbete med Siemens Gamesa Renewable Energy för att upptäcka produktionsförluster hos vindkraftverk orsakade av nedisade turbinblad genom att använda SCADA-data. Mjukvaran filtrerar rådata för att beräkna en referenskurva, mot vilken data för kallt väder kan jämföras. Den visar att det finns korrelation mellan låg temperatur och produktionsförluster samt att risken för produktionsförlust ökar då temperaturen sjunker. De årliga produktionsförlusterna hos de undersökta vindparkerna var 4-10 % av den förväntade årliga produktionen.

Nyckelord: Vindturbin, ispåbyggnad, isförluster, SCADA-data, kallt klimat

(3)

Preface

It is hard to focus on ones work when there is an ongoing global pandemic and loved ones are at risk. Staying at home and barely meeting with friends is challenging for all of us.

Working on my thesis at Siemens Gamesa Renewable Energy is a huge opportunity for me and a good start in my career. I would like to thank my supervisor Emil Thalin for support and guidance throughout this thesis work. I also want to thank my colleague Zandra Lotthagen for help developing the software, finding raw data and thoughtful discussions.

(4)

List of Figures

1 Icing events, class A, B and C. Graphics by Timo Karlsson, Winter Wind 2020 . . . 2

2 Components of wind turbine. Graphics by Encyclopædia Britannica, Inc [3] . . . 4

3 Airfoil with the definition of lift, L, and drag, D, force. F is the resulting force, V∞ is incoming free flow wind speed and α is angle of attack. Graphics by Martin O. L. Hansen [10] . . . 4

4 Type of ice as function of air temperature and wind speed Graphics by S. Fikke [7]. . . 5

5 Different types of ice. . . 6

6 Illustration of reference power curve and potential energy of free flow wind. Graphics by PSU Aerospace Engineering [20]. . . 6

7 An iced (left) and normal (right) airfoil. Higher velocity gives a lower pressure. Interrupted airflow on an iced airfoil does not have as high wind speed as non-iced airfoil. Graphics by Narges Tabatabei, 2018 [25]. . . 7

8 Illustration of low power detection curve. The black curve is the sales power curve and the blue is when the turbine software detects low production due to ice. Graphics are SGRE internal work. 7 9 Power curves as function of wind speed, measured from nacelle anemometer, from raw data. . . . 8

10 Flowchart of the algorithm. The SCADA data is binned and filtered, one for reference curve and one for potential ice losses. The reference curve has a 10th and 90th percentile, which are used to compare to the potential ice points. Points from Filter 2 with production under the 10th percentile are considered iced points. . . 10

11 Anemometers measuring wind speed and temperature. . . 11

12 Reference power curves in blue for sites using temperature threshold. Green are all points accepted by Filter 1. 80 % of values are between the black lines. . . 14

13 Summer months reference power curves for Site 2, 3 and 6. . . 15

14 Actual power curves for 2016. Green is reference values, red is ice losses and blue is all points. Black lines are 10th and 90th percentile. . . 16

15 Actual power curves for 2017. Green is reference values, red is ice losses and blue is all points. Black lines are 10th and 90th percentile. . . 17

16 Actual power curves for 2018. Green is reference values, from temperature threshold, red is ice losses and blue is all points. Black lines are 10th and 90th percentile. . . 18

17 Power curves for 2019. Green is reference values, red is ice losses and blue is all points. Black lines are 10th and 90th percentile. . . 19

18 Production losses per site and year, using temperature threshold reference curve. . . 20

19 Production losses per site and year, using summer months reference curve . . . 21

20 Ice losses in temperature range 3 °C to −2 °C, Site 1. . . 22

21 Ice losses in temperature range −2 °C to −6 °C, Site 1. . . 23

22 Losses at temperatures below −6 °C, Site 1. . . 24

23 The risk of ice points, 2018. Green bars are magnitude of all measured points, blue are the ones where ice was detected. The red points are the probability of ice at a given temperature. . . 26

24 Risk of ice points, 2019. Green bars are magnitude of all measured points, blue are the ones where ice was detected. The red points are the probability of ice at a given temperature. . . 27

25 Relative losses per turbine and site, 2018 . . . 28

26 Relative losses per turbine and site, 2019 . . . 29

27 Production losses as function of temperature, 2016-2019 . Each point represents one year and one site. . . 30

28 Production losses as function of relative humidity, 2016-2019. Each point represents one year and site. . . 31

29 Production losses as function of precipitation, 2016-2019. Each points represents one year and site. 31 30 Monthly average temperature for February and December. Graphics from SMHI [21]. . . 31

31 Monthly average deviation temperature for February and December. Graphics from SMHI [21]. . 32

32 Monthly average precipitation for February and December. Graphics from SMHI [21]. . . 32

List of Tables

1 Manipulated ice losses, P_loss^∗ , at different sites and years, and number of detected ice points per turbine,Npoint^∗ , for both temperature threshold and summer months reference curve. "Quota

(6)

3 Amount of ice points with no ice-related alarm codes, in %. The top is using temperature threshold and the bottom is summer months. . . 25 4 Different sites, mean temperature, relative humidity and precipitation, averaged for January -

March and October-December. . . 30

Nomenclature

A Area of disc rotor m²

h Hub height m

m Mass kg

P Power W

P_loss Production losses due to ice % P_loss^∗ Manipulated production losses %

p Pressure Pa

R₀ Gas constant of dry air J/(kgK)

T Temperature °C

t Time s

V Wind speed m/s

ρ Air density kg/m³

IEA International Energy Agency SGRE Siemens Gamesa Reneable Energy

SCADA Supervisory Control and Data Acquisition std Standard conditions

site Site conditions eq Equivalent value

SMHI Swedish Meteorological and Hydrological Institute

(7)

1 Introduction

The potential of wind power is an old concept, from ships and windmills to wind turbines. In 2019, there was 19.8 TWh (10% of national production) of electricity produced from wind in Sweden and 1270 TWh (4.8 %) globally [4]. It is increasing and the market is expanding. The cold and sparsely populated areas in northern Europe are suitable for wind farms with cold, dense air having higher kinetic energy per unit of volume. However, with the cold climate there are challenges. Ice accretion on the blades affects the aerodynamics and lowers the power output. It can also lead to increased wear in the turbine and cause a risk of ice throw [13] [25].

The purpose of this thesis is to detect, analyse and understand production losses due to icing on wind turbine blades using computer software. Detection will be done using historical SCADA data and filtering out ice losses.

It will be investigated if there are correlation with precipitation, temperature and/or relative humidity and ice losses.

Knowing how much electricity production is lost due to icing is of interest because it is a loss of income when there is less production. There are sensors on the turbines to detect ice. However, the built-in sensors or software systems will not detect all icing events and therefore there is need for also investigating raw data after it has been produced.

1.1 Siemens Gamesa Renewable Energy

This thesis is done at Siemens Gamesa Renewable Energy (SGRE). German Siemens’s wind division merged with the Spanish company Gamesa in 2016 and has its onshore headquarters in Pamplona, Spain. Siemens AG owns 67 % of the shares and the remaining 33 % is owned by the Spanish shareholders. The areas of operation are onshore and offshore wind power and service of wind turbines. SGRE has 25 000 employees globally, whereof approximately 100 in Sweden [27].

SGRE have more than 1.5 GW wind power installed in Sweden and more than 100 GW globally (2019) [2]. The onshore turbines manufactured by SGRE have a nominal output from 2.1 MW with a rotor diameter 114 m to 6.6 MW and 170 m rotor. The largest offshore turbine in development will have 14MW nominal power with 220 mrotor radius [27] [28].

This thesis is a continuation and based on the thesis work done by Zandra Lotthagen at SGRE [14].

1.2 Literature review

Icing impact on Wind Turbine Production, 2014 is the doctoral thesis by Neil N. Davis where he investigated how power production is impacted by icing by analysing numerical models and observations. Davis used turbine power data and nacelle wind speed to see the power production spread. Several models were developed, both statistical, numerical and physical. Using power threshold curves for ice free data, ice could be detected during low production. The thesis uses 0.1 quantile of the non-iced data points to construct a curve to identify production losses. Davis acknowledges that the nacelle produces heat and temperature threshold for ice losses needed to be at least 2.5 °C above freezing, preferably 3 °C.

Numerical weather prediction model was used to investigate cloud parameters. Microphysical scheme needs to be chosen with care, since it can have a large impact on ice model results.

An ice blade model was developed to simulate ice accretion and ice mass on turbine blades. The amount of ice varied significantly if there were ice ablation present.

The statistical model simulates how the ice blade model and production losses correlate. The model also investigates how weather forecast and ice mass accumulation for six wind farms [5].

(8)

Task 19 was issued by the International Energy Agency (IEA), a task group called Task 19 "Wind Energy in Cold Climates" and was an international cooperation. The purpose was to investigate icing impact on wind turbines. Among others, it resulted in a code to calculate ice losses on wind turbines using Supervisory Control and Data Acquisition (SCADA). If the ice sensors do not recognise iced blade, an icing event is triggered when actual power is below the reference power and the temperature is below 3 °C. The three general steps were calculating an un-iced power curve, then start-stop timestamps to categorise different icing events and finally calculating the production losses from those icing events. The icing events are divided in three classes, A, B and C, see Fig- ure 1. Class A and B are underproduction due to icing and class C is overproduction due to iced anemometer [12] [26]. The reasoning behind using software and the anemometer installed on the nacelle was that using met masts would be significantly more expensive. This method is not ex- act, since the anemometer is placed behind the rotor. However, the method of using anemometers is cheap, simple and there is no need to install a met mast.

Figure 1: Icing events, class A, B and C. Graphics by Timo Karlsson, Winter Wind 2020

Research has been done on the effects of icing on airfoils, amongst others, there is one report by Yan Li et al. In the study, tests are preformed on a NACA7715 airfoil to find out how ice accumulates by using a water nozzle and a wind tunnel. By changing the wind speed and pitch angle of the airfoil, it was found that with low angle, ice accretion was on the front of the blade and the affected area was relatively low. At higher angles the icing rate was higher and larger icing area on the underside of the airfoil. Icing also gets worse at higher wind speeds [13].

Investigating the impact of icing on wind turbine aerodynamics is of importance in order to understand why and how icing affects the production power. Horn ice, rigid ice or streamwise ice were types of ice investigated to see how the shape of the ice caused separation in the flow of air around the turbine blades. To accomplice the investigation of air flow around the rotor and blades, two approaches were compared, computational fluid dynamics (CFD) and blade element method (BEM). The result of the study shows that the power coefficient is lower when the airfoil is iced [25].

1.3 Purpose and goals

The purpose is to identify and analyse production losses on wind turbine due to icing on the blades using computer software and historical data. The goals for this thesis are:

• Construction of automatic model that uses SCADA-data and calculates ice losses according to definition.

• Validate model by comparing with former internal work and thesis projects.

• Calculate ice losses for sites with SGRE turbines.

• Finding correlation between ice losses, temperature and precipitation, relative humidity.

• Determining risk of ice losses as a function of temperature.

1.4 Limitations

• Increased wear on turbine components due to icing and maintenance cost will not be investigated, if sites or turbines with higher production losses due to icing have higher maintenance cost.

• No consideration to the topography of sites is production losses.

• It is not investigated how many data point are required to give an accurate result when constructing the reference power curves.

• The time period of the data used is one calendar year.

(9)

• Sensitive data, see below

1.4.1 Sensitive data

Sensitive data have been censored to hide confidential information in this report. All presented values are manipulated with a confidential value to be able to compare results. Ploss represent calculated production losses due to ice, in %, from the expected annual production. It is then manipulated to Ploss^∗ to present, using a confidential value, mf actor. It is done similar with Npoint for all data points where ice is detected, to manipulated ice points Npoint^∗ when using a confidential value nf actor.

P_loss^∗ = P_loss mf actor

∗ 100 N_point^∗ = N_point

nf actor

(10)

2 Wind turbines and ice formation

It is of interest to know how much power production is lost due to ice, relative production losses, in order to calculate the financial losses. Being able to foresee icing events is beneficial to predict how much electricity that can be produced in such situations. It is also important to know which measures can be implemented to reduce the losses.

2.1 Wind turbines

The blades of the turbine catches the incoming wind, which makes the rotor rotate. The rotor is connected to a generator that converts kinetic energy to electrical energy. A gearbox is used in some wind turbines to speed up the rotational speed, but the generator can be designed as a direct drive for low rpm and thus there is no need for a gearbox. There are brakes on the drive train in order to stop the rotor and a yaw drive to rotate the nacelle and have the rotor face the incoming wind. An anemometer and other weather instruments are located at the back of the nacelle [16].

Figure 2: Components of wind turbine. Graphics by Encyclopædia Britannica, Inc [3]

The shape of an airfoil creates a pressure difference that generate a lift force. This force then applies torque on the rotor. The theoretical maximum efficiency can not be higher than 59.3% (Betz’s limit) since the wind still moves after generating lift [16] [24].

Figure 3: Airfoil with the definition of lift, L, and drag, D, force. F is the resulting force, V∞ is incoming free flow wind speed and α is angle of attack. Graphics by Martin O. L. Hansen [10]

The kinetic energy per unit of time, the wind power, in the free flow wind is defined by the equation:

P = 1 2

dm dt V²= 1

2ρAV³ (1)

(11)

The swept area by the blades, A, is πr² and r is the length of the blades. Standard conditions at sea level for air is Tstd = 15°C, pstd = 101 325 Pa, R0 = 287.058 Jkg⁻¹K⁻¹ and ρstd = 1.225 kg/m³ [16] Ch.2. Denser air has more power and can be seen as a function of temperature and altitude. Neither the air density nor pressure are measured, however it is possible to calculate the equivalent wind speed, Veq, of the free flow wind using equation 2. Then it is possible to compare the potential power in the free flow wind at different temperatures and altitudes [5].

Veq= Vsite

ρstd

ρ_site

1/3

(2)

Air is assumed an ideal gas and density, ρ, is a function of pressure, p, and temperature, T .

ρ = p

R0T (3)

Equation 3 in 2 to remove density, R0 is cancelled out.

Veq= Vsite

pstd/Tstd

psite/Tstd

^1/3

(4)

Air pressure can be seen as a function of altitude, h, and the pressure lowers at higher altitude.

p_site= p_std(1 − 2.25577 ∗ 10⁻⁵h)^5.25588 (5) Now putting equation 5 in 4, pstdis cancelled and it is possible to use temperature and altitude when calculating Veq.

Veq= Vsite

Tstd

Tsite

(1 − 2.25577 ∗ 10⁻⁵h)^5.25588^1/3

(6)

Where h is the hub height in meters above sea level [26] and Tsiteis measure temperature at the nacelle. Lower temperature increases equivalent wind speed, while higher altitude decreases equivalent wind speed.

2.2 Types of ice

Rime ice, wet snow and glaze ice have different densities and occur at different weather conditions, depending on temperature, wind speed and humidity. Super cooled droplets in clouds or fog freeze when in contact with a surface. Wet snow could also freeze on the blades. Glaze ice is densest and hardest, but smoother. Denser ice with low air content will require more heat to melt [7] [17].

Figure 4: Type of ice as function of air temperature and wind speed Graphics by S. Fikke [7].

(12)

(a) Glaze ice. Photo by A. Lukas [15] (b) Hard rime ice. Photo by M. Sochacki [23]

(c) Soft rime ice. Photo by Unknown [29]

Figure 5: Different types of ice.

2.3 Reference power curve

The reference power curve is how much power should be produced at a given wind speed. The electrical power output increases as a function of wind, from cut-in speed until the turbines rated power. Turbines have a cut-out speed as well, when the turbine is turned off for safety concerns at high winds [16].

Figure 6 to the right is an example of a reference power curve with cut-in and cut-out wind speed.

The rated power varies from different turbines. It also shows the potential kinetic energy in the wind.

Figure 6: Illustration of reference power curve and potential energy of free flow wind. Graphics by PSU Aerospace Engineering [20].

2.4 Wind turbine operation in ice conditions

Intuitively, ice should only appear below a temperature of 0 °C, however large ice accretion will take a long time to melt. Therefore, temperatures above zero also need to be considered. A temperature threshold of 3 °C is used in several studies, but there is no general consensus [11] [14] [26]. The nacelle itself is also producing heat, from the gearbox and the brakes if used heavily, which can cause a higher measured temperature value.

Figure 7 below shows how icing affects the aerodynamics of an airfoil. Since the lift force is created by a pressure difference, higher speed will create a lower pressure and higher lift force. The shape of an airfoil is designed to operate without ice and the shape changes along the radius of the blade. The ice disrupts the streamlines, causing flow separation, creating vertices and a lower speed, which leads to reduced lift force. A lower force will generate lower torque and less electricity produced.

(13)

Figure 7: An iced (left) and normal (right) airfoil. Higher velocity gives a lower pressure. Interrupted airflow on an iced airfoil does not have as high wind speed as non-iced airfoil. Graphics by Narges Tabatabei, 2018 [25].

2.4.1 Detecting ice

Figure 8: Illustration of low power detection curve. The black curve is the sales power curve and the blue is when the turbine software detects low production due to ice.

Graphics are SGRE internal work.

If there are high winds, low temperature and low production, there is risk of ice. The sensors on the nacelle should detect this, but an anemometer can be faulty due to ice as well [27]. The turbines have software programmed with the reference power curve, or sales power curve, and a Low Production Detection Curve (LPDC), see Figure 8, and if power production is under LPDC, de- icing or Operation with ice start, if available. If ice is detected, the turbine software will display an alarm code.

2.4.2 De-icing using blade heating

Blade heating is implemented in some of the wind farms. The heaters are installed in the blades, on the leading edge and beneath the blade surface. On the blade is a mat of carbon fibre which leads a current that melts the ice. The rotor has to be stopped and no power in generated, otherwise the cooling of the convection from the wind speed would neglect the heating from the mat. There is also a risk that melted water runs over a less heated area of the blade and freezes again [14] [17].

2.4.3 Operation with ice

Operation with ice (OWI) is a software that changes the pitch of the blades, to reduce lost production during icing event. The angle of attack changes and so does the coefficient of lift, hence losses are reduced, but not negligible. Neither does it remove the ice from the blades. An advantage compared to blade heating is that OWI does not require the rotor to stop, but the biggest advantage is that it requires no extra hardware to use and thus is cheap to implement. It is important that there is sufficient lubrication of gears and bearing when using OWI, because the blades need to rotate. Changing the pitch angle increases drag and the load on the blades [14] [27].

(14)

3 Theory

This section describes how raw data is processed, how the algorithm is set up and filters are defined.

Relative production losses can be calculated using measured wind speed and energy production and compare those to a theoretical reference power curve. The production loss is the difference of measured value to desired value.

3.1 SCADA data

Supervisory Control and Data Acquisition, SCADA, is data collected from several sensors on each turbine. It is used to control and follow-up the wind turbines’ performance. The SCADA data is sampled every 10 minutes as data points, and is an average of the time period of 600 seconds. The data for wind turbines consists of power production, wind speed, ambient temperature, alarm code, turbine status. Another parameter shows how many second the turbine has been active during the time period.

3.1.1 Grid limitations

If the power grid cannot handle the produced power, power output is de-rated in accordance with turbine supply agreement. This can occur during heavy winds at the same time as the grid operator needs to limit power output of the wind farm [1]. Figure 9b shows an example of raw data from a wind farm where production has been limited at times. This will be logged as an alarm code in the SCADA data.

(a) Unfiltered power curve. The turbines at this specific site have blade heating installed, hence negative production at some data points.

(b) Limited power curve, with de-rated power at 80 and 90%.

Figure 9: Power curves as function of wind speed, measured from nacelle anemometer, from raw data.

3.1.2 Percentiles

The percentile is a numeric divider and shows the percentage of how many numbers in a sequence is lower than said percentage. The median is the 50th percentile [9]. The 10th percentile is used to detect underproduction and 90th to find overproduction [26].

3.1.3 Outliers

Outliers are unwanted data points and consists of point outliers, contextual outliers or collective outliers. They can be caused by the human factor, sensor error (measurement), experiment setup errors, data processing errors or natural (not an error) [8]. They can cause calculation errors and reshaping power curves.

3.2 R & Rstudio

R is a programming language, similar to Python and is used to visualise data. Rstudio is the interface in which the code is written. It is free to use, open source, and has several packages for visualisation and filters that can be used for sort data. Filters work by looking at a column in a data set and removing unwanted values, by comparing strings or integers [18] [19].

3.3 Algorithm

The software has SCADA data as input to the algorithm, which filters out data point when the turbine is operating as desired. Another filter, which is a temperature threshold of 3 °C is used to remove potential ice

(15)

point. This filter is compared to using summer months instead of a temperature threshold. Summer starts when the daily mean temperature has been above 10 °C for five days in a row [22]. The data is binned into wind speed bins and a reference power curve is made from the mean value in each bin. This is Filter 1 in Figure 10.

The software also makes curves for the 10th and 90th percentile using the values from the same bins. The data used is on a yearly basis, to be consistent.

An ice point will be considered started if OWI is active or blade heating is activated. After two consecutive ice points in time, the following point will automatically be considered an ice point, even though it might not be underproduction. The reasoning is if it is cold, ice need more than 10 minutes to melt [26]. It will also remove fluctuation and measurement errors. This is Filter 2.

When an ice point is detected, the power output is compared to the reference curve for the same wind speed to calculate the relative production loss. 10 minutes is 1/6 hour.

Electrical power loss [W] ∗ 1

6hour = Electrical energy loss [Wh] (7) 3.3.1 Conditions reference curve

The data is sorted into bins of 0.5 m/s, then filtered for sufficient wind, over cut-in and below cut-out. Thereafter the power output is above zero, Psite > 0 W, and finally the turbine is functional and active. Temperature threshold of three degrees celcius, Tsite> 3 °C, or summer months will be investigated.

3.3.2 Conditions ice points

An ice point is when there is a production loss due to icing on the blades. It occurs when there is under production in a bin of 0.5 m/s, below the 10th percentile and a temperature below three degrees celcius, Tsite

< 3 °C. Two consecutive ice points in time will result in the following data point being considered an ice point.

If blade heating or OWI is actived, it is an ice point.

The most time consuming steps in the algorithm are comparing data points where ice could appear to the 10th percentile and when downloading data. Both time consuming steps are dependant on how much data is wanted or used.

3.4 Visualisation

All ice points are counted as well as summarised to compare the sites and turbines. The magnitude of a loss is the difference from the ice point to the reference power curve in the same bin. The ice points are then plotted in a power curve, ice loss curve and a histogram with temperature to see the probability of ice losses. It is also of interest to see if the individual turbines have different losses. They will be compared with temperature, relative humidity and precipitation. It is also possible to plot different temperature ranges.

(16)

Figure 10: Flowchart of the algorithm. The SCADA data is binned and filtered, one for reference curve and one for potential ice losses. The reference curve has a 10th and 90th percentile, which are used to compare to the potential ice points. Points from Filter 2 with production under the 10th percentile are considered iced points.

(17)

4 Method

In this section data gathering and handling is described, as well as the changes to the software.

4.1 Data gathering

SCADA data contains 52 560 data points for each turbine every year. The data point holds an average over the time period, to reduce data points and minimise fluctuation, that shows a variance in power output. Wind speed is measured with an anemometer. The anemometer is placed in the wake of the rotor, see Figure 2, and therefore the measurements come with a high degree of uncertainty. However, the uncertainty is considered constant and the measurements can therefore be used for relative calculations. Power output is measured at grid terminal at each turbine.

(a) KK Anemometer V010M (b) FT7 Ultra-

sonic

Figure 11: Anemometers measuring wind speed and temperature.

The code is developed to work on all sites, but in this report and for convenience, six sites in Sweden will be investigated. Site 1, 2, 4 and 5 are located in central Sweden with good wind conditions. In order to compare and control, site 6 is in the south of the country and site 3 more north. The distance between those sites is about 900 km. The sites have different number of turbines and nominal power. Sites 2, 3, 4 and 5 have blade heating. Only site 1, 2 and 3 have OWI installed.

Data from Swedish Meteorological and Hydrological Institute (SMHI) has been used. SMHI has measuring points located all over Sweden and data from the closest placed measuring point to each site has been used. The data is logged in csv-files and contain parameters for wind speed, temperature and precipitation. Temperature is sampled every hour, while precipitation is sampled daily. The distance between the SMHI measuring point and the site can cause uncertainties [21].

ERA5 is a dataset stretching back to 1979 founded by European Centre for Medium-Range Weather Forecast.

It has hourly data for temperature, relative humidity, wind speed, wind direction etc. It is open to public use.

ERA5 has fixed measuring points and in between those points the weather parameters are calculated [6].

When comparing mean temperature for correlation with production losses, only winter months have been considered to obtain a more relevant temperature value. A year with very warm summer and cold winter could have the same annual mean temperature as a damped summer and a mild winter.

4.2 Data handling

The data is filtered to remove unwanted values, i.e. zero production due to maintenance, below cut-in wind speed or too high winds. In order to decrease the risk of using ice point in the reference power curve, a temperature threshold of 3 °C is used. The temperature threshold will be compared to the summer months. Using summer months should reduce the risk of ice, but winters tend to have higher winds and therefor it is necessary to

(18)

Summer occurs at different dates at different places. At some places in the South of Sweden it occurs in early May, but there are areas in Sweden where it does not always become summer. To be consistent, May-September are chosen as summer months.

4.3 Software update

The software has been re-written to be more user friendly, dividing the code into chunks. Changing input parameters will then be quick to implement for specific uses.

Two new filters have been implemented in the software, which filters out point when the turbine is de-rated and if the turbine is not active for all 600 seconds during the 10 minutes in a measured point.

The reference curve has also been using a temperature threshold as well as a summer reference curve, to compare and see the differences and advantages of each reference curve.

The software can sort the ice losses in temperature ranges, to see where production losses due to ice occur and how significant the probability of ice is. Using a temperature threshold of −6 °C, glaze ice should not accumulate any more and at wind speeds over 7 m/s, there should be more hard rime ice. It is possible to see on which turbines said losses are.

(19)

5 Result

The main purpose of this thesis is to update the software detecting icing events on wind turbine blades. To do that, a reference power curve is constructed using non-iced measuring points. It is presented if using a temperature threshold of 3 °C or summer months May - September is the optimal way of doing the reference power curve. When detecting icing events, measured values are compared to the reference power curve. These points are then counted and the magnitude is summarised to compare the actual production to the expected annual production. The relative losses at the investigated sites are presented with the number of icing points at each site.

It is also presented in which temperature range production losses occur and the risk of an icing event. The production losses can also be seen at individual turbine and correlation with weather.

5.1 Software update

The filter function is heavily used in the software. It is possible to change from temperature threshold to summer months reference curve. A new layup of the code makes it easier to have an overview of the software with chunks of code. The calculated production losses are similar to former internal work.

5.2 Reference power curves, temperature threshold

The reference power curves are constructed with values defined in section 3.3 at each of the six investigated sites. To give a more stable curve, values from a full year are used. The green values in Figures 12a - 12f are all accepted points, the blue line is the mean value in each bin. The 10th (lower) and 90th (upper) percentiles are the black lines, which contains 80 % of the reference points. If a value is then below the 10th percentile line and less than 3 °C, then the loss is measured as the difference between the blue line and the actual measured production.

(20)

(a) Site 1. (b) Site 2.

(c) Site 3. (d) Site 4.

(e) Site 5. (f) Site 6.

Figure 12: Reference power curves in blue for sites using temperature threshold. Green are all points accepted by Filter 1. 80 % of values are between the black lines.

Site 1 has the most turbines, then site 4. Site 2 and 5 have the same number of turbines and site 6 and 3 have the fewest. The reference curve varies from year to year.

5.3 Summer month or temperature threshold for reference power curve

Summer occurs at different times of the year depending on latitude. Using the same months for Site 3 and 6 do not mean the same, since the mean temperature is different. Site 2 had a lowest temperature of −5 °C during May - September, Site 3 had −8 °C and suspected ice losses in reference curve. Site 6 has a lowest measured temperature of 1 °C.

(21)

(a) Site 2. (b) Site 3.

(c) Site 6.

Figure 13: Summer months reference power curves for Site 2, 3 and 6.

The summer reference curves have fewer values than the temperature threshold, which make the percentile lines tighter, but also a risk of outliers affecting the curve. Depending when turbines have been upgraded, using summer months can omit turbines upgrades during the autumn.

When looking at reference curves for Site 3, it is different from using summer months. In the temperature threshold there are no suspected ice losses. Site 6 reference curves are more similar. With narrower percentiles, more class A icing events could be identified.

5.4 Actual power curves

The actual power curves in Figures 14 - 17 show all points in blue, red are points defined as iced points and green points are points used for reference curve that are above 3 °C. The calculations have been done using summer months reference curve as well, but the plots are not presented. Values below 0 % nominal power are when blade heating is activated and rotor has stopped. Since the sites have a different number of turbines, ice points per turbine is presented in Table 1.

(22)

(a) Actual power curve, site 1, 2016. (b) Site 2, 2016.

(c) Site 3, 2016. (d) Site 4, 2016. Some turbines were upgraded to higher nominal power during the year.

(e) Site 5, 2016. (f) Site 6, 2016. Some turbines have been upgraded to higher nominal power during the year.

Figure 14: Actual power curves for 2016. Green is reference values, red is ice losses and blue is all points. Black lines are 10th and 90th percentile.

In Figure 14, it is observed that two sites have turbines of different nominal power. This can cause problems, since production losses at the turbines with higher nominal power can be overlooked, since the reference curve is lowered by the turbines with lower nominal power.

Site 3 is the smallest, but is de-rated on multiple occasions. If these values had been used, the reference curve would have been different and lower. Site 1-5 have both class A and B icing events.

(23)

(a) Actual power curve, site 1, 2017. (b) Actual power curve, site 2, 2017. Few accepted values at higher wind speeds causing the percentiles to drop.

(c) Site 3, 2017. (d) Site 4, 2017. Most turbines have been upgraded.

(e) Site 5, 2017. (f) Site 6, 2017.

Figure 15: Actual power curves for 2017. Green is reference values, red is ice losses and blue is all points. Black lines are 10th and 90th percentile.

In 2017, more turbine have been upgraded to higher nominal power, but there are still differences, seen in Figure 15. At Site 2, there are few accepted values at higher wind speeds, but some at low production. The 10th percentile have two drops that could cause calculation errors and ice losses omitted. The reason why these points are tolerated is unknown, but they are not in present when using the summer months reference curve.

It is possible that there have been maintenance and the turbine is in a start-up face.

Site 3 is de-rated less time, but still has limits at 80 % again. Site 1-5 all have icing event class B, but site 2 and 3 have fewer of class A.

(24)

(a) Actual power curve, site 1, 2018. (b) Actual power curve, site 2, 2018.

(c) Site 3, 2018. (d) Site 4, 2018. All turbines have been upgraded.

(e) Site 5, 2018. (f) Site 6, 2018.

Figure 16: Actual power curves for 2018. Green is reference values, from temperature threshold, red is ice losses and blue is all points. Black lines are 10th and 90th percentile.

Site 1 has a larger spread of icing points in 2018, but not in any of the other three years investigated. There are more ice points, but the production losses are the lowest. Site 6 now has all turbine upgraded, but also more values de-rated. Values close to 100 % nominal power are not included in reference curve, which makes the reference curve narrow. This can be due to that the temperature was low, but not low enough to cause ice accreation. The other sites all have class A and B icing events.

(25)

(a) Actual power curve, site 1, 2019. (b) Site 2, 2019.

(c) Site 3, 2019. (d) Site 4, 2019.

(e) Site 5, 2019. (f) Site 6, 2019.

Figure 17: Power curves for 2019. Green is reference values, red is ice losses and blue is all points. Black lines are 10th and 90th percentile.

Figure 17 show site 6 has a narrower reference curve in 2019, but still de-rated at many times. Site 5 has a wide spread of reference values, which is not seen in earlier years. Class A icing events are still visible.

(26)

Table 1: Manipulated ice losses, P_loss^∗ , at different sites and years, and number of detected ice points per turbine,Npoint^∗ , for both temperature threshold and summer months reference curve. "Quota losses" is production losses for temperature threshold reference curve divided by production losses for summer months reference curve.

Site Year Production losses Detected ice points Production losses Ice points Quota

%, manipulated per turbine, manipulated %, manipulated per turbine, manipulated losses Temperature ref. Temperature ref. Summer ref. Summer ref.

1 2016 104 1.84 102 1.71 1.02

1 2017 81.5 2.36 85.1 2.64 0.958

1 2018 65.4 2.39 63.4 2.31 1.03

1 2019 102 3.04 99.5 2.95 1.03

2 2016 87.9 2.56 85.2 2.48 1.03

2 2017 86.1 2.43 83.1 2.33 1.04

2 2018 133 2.66 131 2.61 1.02

2 2019 95.0 2.90 93.3 2.83 1.02

3 2016 157 2.88 157 2.83 1.00

3 2017 159 2.84 159 2.90 1.00

3 2018 202 3.18 203 3.19 0.995

3 2019 144 3.08 139 3.03 1.04

4 2016 86.3 1.70 83.3 1.60 1.04

4 2017 80.0 2.64 78.2 2.61 1.02

4 2018 90.8 2.36 88.5 2.32 1.02

4 2019 115 2.90 111 2.80 1.04

5 2016 77.0 1.97 74.7 1.87 1.03

5 2017 65.4 2.14 64.4 2.10 1.02

5 2018 82.7 2.26 81.5 2.25 1.01

5 2019 99.2 1.63 96.5 1.60 1.03

6 2016 9.37 0.550 7.95 0.467 1.18

6 2017 10.6 0.666 9.96 0.637 1.06

6 2018 7.83 0.411 7.18 0.390 1.09

6 2019 6.24 0.542 5.61 0.490 1.11

The temperature threshold reference curve gives higher production losses due to icing compared to the summer months reference curve, as seen in Table 1 "Quota losses". However, for Site 1-5 it is less than 4 % for every site. This can be caused by the differences when summer arrives and that ice losses have been found in May and September. Site 6 have the highest difference between temperature threshold and summer months, and this can be due to that there are few ice losses at the site.

Figure 18 and 19 show a visualisation of the values presented in Table 1. Figure 18 is the values using temperature threshold and Figure 19 is using summer months.

(a) Production losses per site and year. (b) Ice points per turbine and year.

Figure 18: Production losses per site and year, using temperature threshold reference curve.

(27)

Site 3 has the highest losses and most ice points. Site 1, 2, 4 and 5 have similar losses and ice points most years.

Site 2 has higher losses in 2018 and more ice points in 2016 than the sites in the same region. Site 6 has the least ice losses and fewest ice points. In Figure 18b and 19b, it is not shown if it is ice class A or B.

The large spread of values in the reference curve gives fewer losses, compared to when the percentiles are closer.

Many class A icing event are omitted at site 5 in 2019, since the production losses are similar, but fewer icing points.

(a) Losses per site and year (b) Ice points per turbine and year

Figure 19: Production losses per site and year, using summer months reference curve

The character of the figures displaying losses and ice points per year is the same, except for Site 1 in 2017, when losses and ice points are higher using the summer months for reference. This is due to the similar values using temperature threshold and summer months.

5.5 Temperature ranges and risk of ice

It was investigated in which temperature ranges ice appear and how large the risk of an ice point was. Looking at different ranges, with thresholds at −2 °C and −6 °C. Figure 20, 21 and 22 present these for Site 1 for 2016- 2019. The temperature threshold has been used for reference curve when calculating ice losses. The temperature threshold can be changed with user input, if specific temperature ranges are of interest.

(28)

(a) 2016 (b) 2017

Figure 20: Ice losses in temperature range 3 °C to −2 °C, Site 1.

More class B icing events can be seen in 2016 at > − 2 °C, Figure 20a, while the other years have more class A icing events. In 2019, there were the most points. Ice points higher than the reference curve can be seen and this is because of the definition. It two ice points are detected, the next following will also be defined as an ice point, even if it is not under production.

(29)

(a) 2016 (b) 2017

Figure 21: Ice losses in temperature range −2 °C to −6 °C, Site 1.

In the range −2 to −6 °C, the shape for all years have a similar structure. In 2017 there were more class A icing events. Figure 21c shows more losses at higher wind speeds and more spread of points. More class B losses appear, but most are class A losses.

(30)

(a) 2016 (b) 2017

Figure 22: Losses at temperatures below −6 °C, Site 1.

Figure 22 shows ice losses below −6 °C and here the large spread of values a site 1 in 2018 have disappeared.

In 2016, 2017 and 2018 it was the fewest ice point in this temperature range. In 2019, there were the most ice point in the range and many class B losses. There are not necessarily more icing points at colder temperature, but the ice accretion is more substantial. In this temperature range and at higher wind speed, it is perhaps hard rime ice, which will stick harder than soft rime ice to the turbine blades.

Table 2: Losses at different temperature ranges at Site 1. Results show ice points per turbine, Npoint^∗ . Year +3>T>-2 -2>T>-6 -6>T Total points per turbine

2016 0.977 0.497 0.368 1.84

2017 1.25 0.751 0.363 2.36

2018 0.794 0.834 0.765 2.39

2019 1.55 0.706 0.782 3.04

5.5.1 No alarm codes

The ice sensors do not always detect ice on the blades and most likely those not identified are class A. Figure 8 is used to find under production, but only find the more significant losses. In Table 3, it is shown how large portion of detected ice points using the software have an ice-related alarm code.

(31)

Table 3: Amount of ice points with no ice-related alarm codes, in %. The top is using temperature threshold and the bottom is summer months.

Year Site 1 Site 2 Site 3 Site 4 Site 5 Site 6

2016 77 84 91 87 96 86

2017 64 78 85 76 86 100

2018 56 60 69 55 68 98

2019 61 63 70 57 53 79

Mean, site 65 71 79 69 76 91

2016 78 83 91 86 96 85

2017 67 89 85 76 86 100

2018 55 60 70 54 68 98

2019 60 63 70 56 53 78

Mean, site 65 74 79 68 76 90

There are similar values on Site 1-5, but Site 6 have the highest average. This can be presumed that there are so few losses and Site 6 and few events with under production due to ice. The class A ice losses are close to the reference curve and Figure 8 shows that for the de-icing to be activated, there needs to be a large production loss. When de-icing using blade heating is activated, there will be no production, hence ice points are omitted by the turbine software. Similar using OWI, changing the pitch can improve production, but the blades are designed for a specific pitch and non-iced blades. The changed pitch will also increase drag and that leads to higher loads on the blades.

Little to no difference between temperature threshold and summer months can be found. Site 2 in 2017 is the only site with a significant difference. Few values at higher wind speeds, Figure 15b, can fail to detect icing points.

5.5.2 Risk of ice

Figure 23 and 24 show the risk of ice points at each site. It is plotted in a histogram and starts at 3 °C, because of the algorithm, and displays the lowest measured temperature at the specific site. All measured points are considered in the green values and light blue are the sum of all detected ice points. The overall trend is there are more points closer to 0 °C, but the risk of ice points is increasing at lower temperatures. Site 5 has peaks of measurements at −1 °C, in Figure 23f and 24f, which suggest problems with the temperature sensors.

(32)

(a) Site 1, 2018 (b) Site 2, 2018

(c) Site 3, 2018 (d) Site 4, 2018

(e) Site 5, 2018 (f) Site 6, 2018

Figure 23: The risk of ice points, 2018. Green bars are magnitude of all measured points, blue are the ones where ice was detected. The red points are the probability of ice at a given temperature.

Site 6 has the lowest risk of ice and fewer point at colder temperatures. The probability does not exceed 10 % and the highest is found at +1 and +2 °C.

The risk of ice is not zero at any site at 3 °C, which indicates there could be ice at higher temperatures. There are not many, but should be investigated further and the magnitude of these points, if they are worth considering.

(33)

(a) Risk of ice at different temperatures, site 1, 2019 (b) Risk of ice, site 2, 2019

(c) Site 3, 2019 (d) Site 4, 2019

(e) Site 5, 2019 (f) Site 6, 2019

Figure 24: Risk of ice points, 2019. Green bars are magnitude of all measured points, blue are the ones where ice was detected. The red points are the probability of ice at a given temperature.

Site 1-4 have a general trend from a probability of 40 % at the coldest, to 20 % when it gets warmer. The risk of ice at Site 6 is 15-20 %, but there are fewer measure points in the temperature range.

(34)

5.6 Individual turbines

Each site has a different number of turbines. With the script it is possible to see which turbine has the most ice points. Only 2018 and 2019 is presented, since turbines at the same sites have had different nominal power.

Every bar on the x-axis is one turbine.

(a) Relative losses per turbine, site 1, 2018 (b) Relative losses per turbine, site 2, 2018

(c) Site 3, 2018 (d) Site 4, 2018

(e) Site 5, 2018 (f) Site 6, 2018

Figure 25: Relative losses per turbine and site, 2018

There are differences in losses per turbine and site 6 has the lowest relative losses, while site 3 has the highest.

Some fluctuation occur at all sites, but the majority have similar losses.

(35)

(a) Relative losses per turbine 2019 (b) Relative losses per turbine, site 2 2019

(c) Site 3 (d) Site 4

(e) Site 5 (f) Site 6

Figure 26: Relative losses per turbine and site, 2019

Site 5 have large differences in the turbines in 2019, which should be investigated further. One reason for the outliers could be human factor, that maintenance has been preformed, but not reported as an alarm code. If so, it would be classified as an ice loss if it were cold at the same time. Similar can be seen at Site 4, with two turbines have much higher losses. It can also depend on the location of the turbines, when the topography is different and hence different weather conditions.

(36)

5.7 Comparing with weather data

In Figure 27 and 28, the mean temperature and relative humidity for January - March and October - December each year is presented. Then in Figure 29, it is the total precipitation for said months and each year. The weather data is presented in Table 4. Visualisation of temperature, deviating temperature and precipitation for Sweden are presented in Figure 30, 31 and 32.

Table 4: Different sites, mean temperature, relative humidity and precipitation, averaged for January - March and October-December.

Site Year Mean temperature Relative humidity Precipitation

°C, winter %, winter mm, winter

1 2016 -3.0 86.3 283

1 2017 -2.2 86.4 575

1 2018 -3.6 86.5 432

1 2019 -3.5 86.8 445

2 2016 -3.5 86.1 266

2 2017 -3.0 86.2 296

2 2018 -4.2 86.2 292

2 2019 -3.1 86.5 366

3 2016 -5.1 85.4 169

3 2017 -4.8 86.1 263

3 2018 -6.1 84.2 218

3 2019 -6.3 85.7 243

4 2016 -3.6 88.3 116

4 2017 -2.9 88.7 238

4 2018 -4.1 87.9 210

4 2019 -3.8 88.8 228

5 2016 -4.2 88.6 186

5 2017 -3.5 88.7 305

5 2018 -5.0 89.3 238

5 2019 -4.5 89.5 292

6 2016 1.3 88.8 427

6 2017 2.4 90.0 711

6 2018 1.0 89.5 412

6 2019 2.6 90.7 739

Figure 27: Production losses as function of temperature, 2016-2019 . Each point represents one year and one site.

Site 3 is the coldest site and site 6 the warmest and there is a correlation in temperature and production losses. Site 1, 2, 4 and 5 have similar temperatures and losses. Site 6 is the only site with an average temperature above freezing, then it is logical that the production losses are the smallest.

The lower mean temperature in 2018 (Figure 16f) removed more values from the temperature threshold compared to in 2019 (Figure 17f). The mean temperature was more than 0 °C and the ice losses were similar. R²of the trend line is 0.7 for all data points.

(37)

Figure 28: Production losses as function of relative humidity, 2016-2019. Each point represents one year and site.

Lower relative humidity indicates higher production losses. R² is 0.6. Important is that relative humidity is not absolute humidity.

Figure 29: Production losses as function of precipitation, 2016-2019. Each points represents one year and site.

In Figure 29, precipitation is presented with production losses, but there is not a clear trend as in Figure 27 and 28 and the R²is 0.4. This is not expected, since snow and supercooled rain were presumed to increase ice accretion on turbine blades.

(38)

(a) February 2017 (b) December 2017 (c) February 2018 (d) December 2018 (e) February 2019

Figure 31: Monthly average deviation temperature for February and December. Graphics from SMHI [21].

(a) February 2017 (b) December 2017 (c) February 2018 (d) December 2018 (e) February 2019

Figure 32: Monthly average precipitation for February and December. Graphics from SMHI [21].

February 2018 was colder than usual, but less precipitation than the years before and after. December 2017 had much rain, while February 2017 was dryer. February 2019 was warmer than the earlier years. Site 2 and 3 had the most production losses in 2018, while site 1 had the least that year. Site 4 and 5 had the most losses in 2019.

The risk of ice increases as temperatures drop. However, it is more common with temperatures around 0 °C than −15 °C.

(39)

6 Discussion & conclusion

The code in the software has been simplified and faster with an interactive structure and new sections, to give a new structure than the one made by Lotthagen. The main simplification is choosing data and which parameters to download instead of general master data, which decreases download time. New filters for the reference curve have been implemented where there must not be power reduction during the 10 minute sample time and the turbine must be active for all 600 seconds. It is possible to change from using temperature threshold to summer months. Not using values when the turbine is de-rated will make the percentiles narrower and more accurate.

The software updates on turbines are important when constructing the reference curve. There is a risk of omitting ice points if the percentiles are wider because of conflicting nominal power output. Several sites have had turbines upgraded in nominal power and that has given uncertainties in the code. Since there is nothing wrong with the turbine, the width of the percentiles are increased and the mean value is not at 100 % nominal production. The risk is then that class A losses are omitted. One improvement of the software is to use two or more reference curves for the same site, when more class A icing events would be detected.

Site 1, the largest site, has five times the number of turbines as the smallest site. The time required is significantly more to do calculations. Using another programming language could be beneficial, to run the code faster. The amount of data in each bin in not investigated here. The suggestion is to calculate the reference curve using several years and save it, using to for further calculations.

6.1 Using summer months or 3 °C for reference curve.

There is no guarantee having 3 °C as temperature threshold sorts out all ice points since there were losses found in May and September. The threshold can remove non-iced point when the turbine is operating as desired, since it does not consider the derivative of temperature changes. Using the same summer months for all sites is unreliable, since summer arrives at different times at different sites. Nevertheless, the definition of summer is a mean daily temperature over 10 °C, all ice should have disappeared by that point. The risk then becomes to few values in each bin and generates an unreliable reference curve and percentiles. Several sites in colder regions have more wind during winter and there is more energy per unit of volume in the colder air. The summer months reference curve also showed lower ice losses, which indicates that the mean value is lower. The ice points without an ice related alarm code were very similar to using the temperature threshold reference curve, meaning the proportion is the same. The temperature threshold reference curve also finds more ice points.

One option is to use a reference curve from a different site with the same nominal power and platform that is free of ice. The wind is equalised and hence can be valid for other parts of the country or regions. There would also be a benefit of not having to calculate a new reference curve for each cite. It can then be more efficient.

Given the number of turbines operated by SGRE, it seems possible. Different anemometers are being used and calibrated differently, which causes uncertainties.

6.2 Weather factor

There is a correlation with temperature and production losses. The magnitude of losses increase and the use of de-icing is implemented. For Site 1-5, the risk of ice losses decreased with higher temperature, but there are more occasions closer to 0 °C than −15 °C. Site 6 has the least amount of losses, which was also expected since it is located in southern Sweden and the winter mean temperature was above zero and warmest.

6.3 Ice detection

The turbine software does detect icing events during cold periods and when there is low production. It could be set to a higher detection curve, but OWI increases wear on the blades and blade heating stops the rotor all together. Stopping the rotor to use blade heating would cost much more than having a small production loss.

The OWI software still have the rotor running, but it does not detect where on the blade the ice is located.

The ice will remain on the blade and there will higher loads on the rotor, but less production losses.

It is of higher interest to detect class B icing events, since the production losses are more significant then.

6.4 Uncertainties

(40)

Data from SMHI is not from sites, but the closest measuring station. The distance is up to 30 km from the sites and the measurements could be localised. When using the temperature measurements from the turbines there is a risk of heat from the nacelle affecting the output, if the brakes of gearbox are warm.

Alarm codes are sorted and should show if there is maintenance, but if no such alarm code is found, the software does not filter those values. If it at the same time is cold, it will be classified as an ice loss. This could help explain why some turbines have higher losses than others. There is a possibility of human error, but the risk is considered low.

The risk of iced anemometers (values over 90th percentile, icing event class C ) have not been investigated due to time restrictions.

(41)

7 Future work

The next step in developing the software is to have an interactive application for SGRE employees, where turbine data can be uploaded. The application then lets the user choose what type of reference curve should be used and which temperature threshold, which summer months or both. Similar should be used for the ice filter where the user can choose a temperature threshold to find icing events.

It should be investigated why some turbines have higher losses than others at the same sites, if it is location, altitude or random. The code can be further developed to give more detailed information about the alarm codes with known icing event. This can then be used to calculate how expensive the losses will be. It is also interesting to know which software the turbine has, if it is an older model the codes can be different. Visiting a site during a known icing event to validate the code would be valuable, using the code in real time. Also comparing the code to photos of the iced turbine blade. Trying to watch weather forecasts and later see if there are losses.

Machine learning should be investigated if it can be implemented to simplify detection of production losses due to ice.

(42)

References

[1] Mattias Ahlstedt. Interview and ongoing talks with Mattias Ahlstedt. Siemens Gamesa Renewable Energy, 2020.

[2] Annual report 2019. Siemens Gamesa Renewable Energy. 2019. url: https://www.siemensgamesa.com/

en-int/investors-and-shareholders/financial-information/annual-reports.

[3] Christopher A. Badurek. Wind turbine technology. Retrieved 17-12-2020. 11-11-2013. url: https://www.

britannica.com/technology/wind-turbine.

[4] Data & Statistics. Explore energy data by category, indicator, country or region. Retrieved 27-01-2021.

International Energy Agency, IEA. 2021. url: https://www.iea.org/data-and-statistics?country=

SWEDEN&fuel=Electricity\%20and\%20heat&indicator=ElecGenByFuel.

[5] Neil N. Davis. “Icing Impacts on Wind Energy Production”. English. PhD thesis. Denmark, 2014. isbn:

978-87-92896-61-2. url: https://orbit.dtu.dk/en/publications/icing-impacts-on-wind-energy- production.

[6] ERA5. Retrieved 20-01-2021. European Centre for Medium-Range Weather Forecasts. 7-01-2021. url:

https://www.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era5.

[7] S. Fikke. “COST727: Atmospheric icing on structures”. In: Veröffentlichung MeteoSchweiz 75 (2006), p. 110.

[8] Jim Frost. Guidelines for removing and handling outliers in data. Retrieved 11-11-2020. 2020. url: https:

//statisticsbyjim.com/basics/remove-outliers/.

[9] Jim Frost. Percentiles: Interpretations and Calculations. Retrieved 15-01-2021. 2020. url: https : / / statisticsbyjim.com/basics/percentiles/.

[10] Martin O. L. Hansen. Aerodynamics of Wind Turbines. Second Edition. 2008. isbn: 978-1-84407-438-9.

[11] Johan Hansson, Johannes Lindvall, and Øyvind Byrkjedal. Quantification of icing losses in wind farms.

Assessment and optimzation of the energy production of operational wind farms: Part 3. 2016. isbn: 978- 91-7673-299-1.

[12] Timo Karlsson. About Task 19. Acquired 08-02-2021. url: https://community.ieawind.org/task19/

home.

[13] Yan Li et al. “A wind tunnel experimental study of icing on wind turbine blade airfoil”. In: Energy Conversion and Management 85 (September 2014), pp. 591–595. doi: https://doi.org/10.1016/j.

enconman.2014.05.026.

[14] Zandra Canovas Lotthagen. “Defining, analyzing and determining power losses - due to icing on wind turbine blades”. MA thesis. Mälardalen University, MDH, 6-06-2020.

[15] A. Lukas. Ice on grass. 7-03-2010. url: https://commons.wikimedia.org/wiki/File:Ice_on_grass.

jpg.

[16] J.F. Manwell, J.G. McGowan, and A.L. Rogers. Wind energy explained. Theory. design and application, second edition. Wiley, 2009. isbn: 978-0-470-01500-1.

[17] Jennifer Pettersson and Sofia Sollén. “Experimental investigation of a de-icing system for wind turbine blades based on infrared radiation”. MA thesis. Luleå University of Technology, 2019. url: http://urn.

kb.se/resolve?urn=urn:nbn:se:ltu:diva-75480.

[18] R version 4.0.2. The R Foundation. 2020. url: https://www.r-project.org/.

[19] Rstudio version 1.3. Rstudio, PBC. 2020. url: https://rstudio.com/.

[20] Sven Schmitz. Power Curve. Acquired 15-02-2021. url: https://www.e-education.psu.edu/aersp583/

node/470.

[21] SMHI. Data. 2020. url: https://www.smhi.se/data.

[22] SMHI. Sommar. Site updated 14-12-2020, retrieved 09-02-2021. 28-10-2011. url: https://www.smhi.se/

kunskapsbanken/meteorologi/sommar-1.1084.

[23] Marcin Sochacki. Hard rime on top of Szczeliniec Wielki, Poland. 1-01-2004. url: https://commons.

wikimedia.org/wiki/File:Rime_ice.jpg.

[24] Vaishali Sohoni, S.C Gupta, and R.K. Nema. “A Critical Review on Wind Turbine Power Curve Modelling Techniques and Their Applications in Wind Based Energy Systems”. In: Journal of Energy Volume 2016 (2016). Article Id 8519785. doi: http://dx.doi.org/10.1155/2016/8519785.

[25] Narges Tabatabei. “Impact of Icing on Wind Turbine Aerodynamics”. PhD thesis. Luleå University of Technology, December 2018. isbn: 978-91-7790-229-4. url: http://urn.kb.se/resolve?urn=urn:nbn:

se:ltu:diva-71186.

[26] Task 19 Ice Loss (Release 2.2). IEA Wind Task 19 and VTT. 7-11-2019.

[27] Emil Thalin. Interview and ongoing talks with Emil Thalin. Siemens Gamesa Renewable Energy, 2020.

[28] The leader in renewable energy. Retrieved 23-01-2021. Siemens Gamesa Renwable Energy, S.A., 2021.

url: https://www.siemensgamesa.com/en-int.

[29] Unknown. Soft rime in Möttlingen/Germany. 13-12-2004. url: https://commons.wikimedia.org/wiki/

File:Raureif2.JPG.

Determining and analysing production losses due to ice on wind turbines using SCADA data