Data-driven fatigue load monitoring in a wind farm

Data-driven fatigue load monitoring in a wind farm Luis Vera-Tudela

¨ Mathematik und Naturwissenschaften Von der Fakultät fur der Carl von Ossietzky Universität Oldenburg zur Erlangung des Grades und Titels eines

D OKTORS DER I NGENIEURWISSENSCHAFTEN D R .-I NG angenommene Dissertation von Herrn Luis Vera-Tudela geboren in Lima, Peru

Gutachter: Prof. Dr. Martin Kuhn ¨ Zweitgutachter: Prof. Dr. Gerard van Bussel Tag der Abgabe: 30.01.2018 Tag der Disputation: 29.06.2018

Since all models are wrong, the scientist must be alert to what is importantly wrong. George E. P. Box Acknowledging what you don’t know is the dawning of wisdom. Charlie Munger

iv

Abstract The expected lifetime of wind turbines is a factor directly proportional to their energy output. Thus, a better estimation should lead to a better valuation of the financial appealing of wind turbines. Moreover, their expected lifetime is a choice made by manufacturers, with typical values around 20 to 25 years, and as such affects their design. The number of years wind turbines should operate is used to calculate the accumulated fatigue loads they should withstand during their operational lifetime. To monitor accumulated fatigue loads with additional purpose-specific sensors would be a logical step to better understand the lifetime of wind turbines. However, this is an expensive and complex option and loads measurement are typically left to prototypes. But, since most commercial wind turbines record their operational conditions with Supervisory Control and Data Acquisition (SCADA) systems, estimating their accumulated fatigue loads from SCADA data became a field of research on its own more than a decade ago. This thesis contributes to this field by elaborating on how to define a methodology based on data mining techniques that describe the relation between 10-minute statistics of SCADA data and accumulated fatigue loads. Furthermore, efforts are centred on how to transform implementation challenges into scientific questions, where the contributions are conceived to improve the usability of the methodology. Some of the most important findings are: the research for more accurate models needs to be balanced with the needs of a minimum viable application; after there is a transparent transformation of application requirements to scientific problems, one can focus on a reproducible, simple and robust approach. Also with respect to more technical aspects, numerical regression problems can be simplified when potential input variables are methodologically selected and a problem with less variables can take advantage of simpler and more robust algorithms. The first two chapters serve to frame the problem and to scope the research. The first chapter defines what can be understood as fatigue load monitoring and provides a list of challenges. The second chapter clarifies concepts around modelling and monitoring, which serve to frame the research decisions made for the specific methodology. Then, in chapters three to five, specific problems are introduced next to original

vi

scientific contributions; each chapter targets key dilemmas: chapter three evaluates alternative methods to select sets of input variables; chapter four proposes metrics to evaluate quality of prediction and assess the impact that wake conditions have on prediction quality; chapter five quantifies the deterioration of prediction quality when fatigue loads are assessed on other turbines. Then, chapter six brings all previous steps together and presents how to implement the methodology in a commercial wind farm. Before last, chapter seven discusses other approaches not considered in the specific methodology. Finally, chapter eight closes the thesis with a brief summary and outlook.

Zusammenfassung Die erwartete Lebensdauer von Windenergieanlagen ist ein Faktor, der direkt proportional zu ihrer Energieabgabe ist. Daher sollte ihre genauere Einschätzung zu einer ¨ besseren Bewertung der finanziellen Attraktivität von Windenergieanlagen fuhren. ¨ Daruber hinaus wird die erwartete Lebensdauer vom Hersteller, mit typischen Werten zwischen 20 und 25 Jahren ausgewählt, ein Einfluss in ihre Auslegung. Dies wird ¨ verwendet, um die akkumulierten Ermudungslasten, die sie während ihrer Betriebslebensdauer aushalten sollten zu berechnen. ¨ ¨ Die Uberwachung akkumulierter Ermudungslasten durch die Installation zusätzlicher spezifischer Sensoren wäre ein logischer Schritt um die Lebensdauer von Windenergieanlagen besser zu verstehen. Diese Option ist teuer, komplex und daher sind Lastmessungen typischerweise auf Prototypen beschränkt. Da jedoch die meisten ¨ ¨ kommerziellen Windenergieanlagen uber ein Uberwachungsund Datenerfassungssys¨ ¨ tem (SCADA) verfugen, wurde die Ableitung akkumulierter Ermudungslasten aus SCADA-Daten vor mehr als einem Jahrzehnt zu einem eigenständigen Forschungsgebiet. Diese Arbeit liefert einen Beitrag zur Entwicklung einer Methodik, die auf Datamining Techniken basiert, die den Zusammenhang zwischen 10-Minuten-Statistiken von ¨ ¨ den Fall von TurSCADA-Daten und akkumulierter Ermudungsbelastung speziell fur binen in Windparks beschreiben. Besonderes Augenmerk wird darauf gelegt Implementierungsherausforderungen in wissenschaftliche Fragestellungen zu transformieren und sich dann auf wissenschaftliche Hypothesen zu konzentrieren, welche auf die Klärung oder Verbesserung der Anwendbarkeit der Methodik abzielen. Einige der wichtigsten Ergebnisse sind: die Suche nach genaueren Modellen muss ¨ ¨ mit den Anforderungen einer moglichst einfachen Anwendung in Ubereinstimmung gebracht werden; nach einer transparenten Transformation der Anwendungsanforderungen in wissenschaftliche Probleme kann man sich auf einen reproduzierbaren, einfachen und robusten Ansatz konzentrieren. Auch in Bezug auf eher technische Aspekte, ¨ numerische Regressionsprobleme konnen vereinfacht werden, wenn potenzielle Ein¨ ¨ gangsgrossen methodisch ausgewählt werden; daruber hinaus kann ein Problem mit weniger Variablen die Vorteile einfacherer und robusterer Algorithmen nutzen.

viii

Abstract / Zusammenfassung

Die ersten beiden Kapitel dienen dazu, das Problem zu umreissen und Vere¨ infachungen zu rechtfertigen, die unternommen wurden, um die durchgefuhrten ¨ Forschungsarbeiten zu erfassen. Im ersten Kapitel wird definiert, was als Uberwachung ¨ der Ermudungsbelastung verstanden werden kann und welche Herausforderungen zu bewältingen sind. Das zweite Kapitel erklärt Konzepte zur Modellierung und ¨ Uberwachung, die dazu dienen, die Forschungsentscheidungen zu treffen, die zur Bildung der spezifischen Methodik getroffen wurden. ¨ neben den wissenschaftlichen Anschließend wird in den Kapiteln drei bis funf ¨ Beiträgen eine kurze Einfuhrung in die spezifische Problematik gegeben. Jedes Kapi¨ tel beantwortet eine Schlusselfrage: Kapitel drei befasst sich mit der Auswahl von Eingangsvariablen, Kapitel vier schlägt Metriken zur Bewertung der Qualität von Vorhersagen vor und bewertet die Auswirkungen von Nachlaufbedingungen, und ¨ quantifiziert die Verschlechterung der Vorhersagequalität, wenn geschätzte Kapitel funf ¨ ¨ Ermudungslasten auf andere Windenergieanlagen ubertragen wird. Kapitel sechs fasst ¨ die Implementierung der alle vorherigen Schritte zusammen und zeigt ein Beispiel fur Methodik in einem kommerziellen Windpark. Kapitel sieben dokumentiert alternative Ansätze, die nicht in der Routine der Methodik enthalten sind. Abschliessend ¨ zukunftige ¨ fasst Kapitel acht die Ergebnisse zusammen und gibt Anregungen fur Forschungsaufgaben.

Contents Abstract / Zusammenfassung

v

List of Tables

xii

List of Figures

xv

List of Abbreviations 1

2

xxi

Introduction

1

1.1

State of the art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1

1.2

Problem scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2

1.2.1

In relation to monitoring fatigue loads . . . . . . . . . . . . . . . .

2

1.2.2

In relation to the use of SCADA data . . . . . . . . . . . . . . . . .

3

1.3

Research questions and objectives . . . . . . . . . . . . . . . . . . . . . .

6

1.4

Structure of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

7

Monitoring and modelling lifetime

9

2.1

Monitoring and modelling objectives . . . . . . . . . . . . . . . . . . . . .

9

2.2

Type of models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10

2.2.1

Physical models . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10

2.2.2

Data-driven models . . . . . . . . . . . . . . . . . . . . . . . . . .

11

2.2.3

Stochastic models . . . . . . . . . . . . . . . . . . . . . . . . . . . .

12

2.2.4

Hybrid models . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

12

Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

13

2.3.1

Monitoring strategies . . . . . . . . . . . . . . . . . . . . . . . . .

15

2.3.2

Relevance of monitoring . . . . . . . . . . . . . . . . . . . . . . . .

16

2.3.3

Monitoring as a process . . . . . . . . . . . . . . . . . . . . . . . .

19

2.3.4

From monitoring to modelling . . . . . . . . . . . . . . . . . . . .

20

2.3.5

Monitoring and statistics . . . . . . . . . . . . . . . . . . . . . . .

21

2.3

x

CONTENTS

2.4

3

2.4.1

Standards and guidelines . . . . . . . . . . . . . . . . . . . . . . .

23

2.4.2

Data, measurements and quantities . . . . . . . . . . . . . . . . .

24 27

3.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

28

3.1.1

Universe of potential inputs . . . . . . . . . . . . . . . . . . . . . .

28

3.1.2

Input variable selection problem . . . . . . . . . . . . . . . . . . .

29

On the selection of input variables for a wind turbine load monitoring system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

32

3.2.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

32

3.2.2

Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

34

3.2.3

Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

38

3.2.4

Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

38

3.2.5

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

43

Quality control - The influence of wind farm conditions

47

4.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

48

4.1.1

Definition of quality . . . . . . . . . . . . . . . . . . . . . . . . . .

48

4.1.2

The impact of wind farm flow conditions . . . . . . . . . . . . . .

49

4.2

5

23

Robustness - The selection of input variables

3.2

4

Other considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Analysing wind turbine fatigue load prediction: The impact of wind farm flow conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

50

4.2.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

50

4.2.2

Data description and methodology . . . . . . . . . . . . . . . . . .

53

4.2.3

Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

54

4.2.4

Fatigue loads calculation . . . . . . . . . . . . . . . . . . . . . . .

55

4.2.5

Input variable selection . . . . . . . . . . . . . . . . . . . . . . . .

58

4.2.6

Regression analysis . . . . . . . . . . . . . . . . . . . . . . . . . . .

60

4.2.7

Evaluation of predictions . . . . . . . . . . . . . . . . . . . . . . .

61

4.2.8

Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

66

4.2.9

Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

67

Model generalisation - The selection of regression models

71

5.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

71

5.1.1

Imbalance data sets . . . . . . . . . . . . . . . . . . . . . . . . . . .

72

5.1.2

Predicting loads for other turbines . . . . . . . . . . . . . . . . . .

73

CONTENTS

5.2

6

7

xi

Monitoring Fatigue Loads in Wind Farms from SCADA Data - Quantifying Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

75

5.2.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

76

5.2.2

Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

77

5.2.3

Fatigue loads estimation . . . . . . . . . . . . . . . . . . . . . . . .

79

5.2.4

Data description . . . . . . . . . . . . . . . . . . . . . . . . . . . .

79

5.2.5

Input variable selection . . . . . . . . . . . . . . . . . . . . . . . .

80

5.2.6

Regression models . . . . . . . . . . . . . . . . . . . . . . . . . . .

81

5.2.7

Performance metrics . . . . . . . . . . . . . . . . . . . . . . . . . .

83

5.2.8

Test on similar data . . . . . . . . . . . . . . . . . . . . . . . . . . .

85

5.2.9

Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

87

5.2.10 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

91

5.2.11 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

92

Synthesis - Data-driven fatigue load monitoring in a wind farm

95

6.1

Business understanding . . . . . . . . . . . . . . . . . . . . . . . . . . . .

97

6.2

Data understanding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

6.3

Data preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

6.4

Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

6.5

Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

6.6

Deployment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

Alternative approaches

107

7.1

Grey box modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

7.2

Time-series modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

7.3

Normal behaviour models for wind turbine vibrations: Comparison of neural networks and a stochastic approach . . . . . . . . . . . . . . . . . 112

8

7.3.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

7.3.2

Data description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

7.3.3

Methods and Results . . . . . . . . . . . . . . . . . . . . . . . . . . 117

7.3.4

Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

7.3.5

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

Summary and Outlook

133

8.1

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

8.2

Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

xii

References

139

Appendix A: Open source code

155

List of publications

157

Acknowledgements

159

Curriculum Vitae

161

Erklärung

163

List of Tables Table 2.1

Overview of scientific publications and models used according to

specific target values for wind turbine applications. . . . . . . . . . . . . Table 2.2

25

Measures (techniques) and some of their features (cost, online,

diagnosis, deployed and wind turbine components). Reproduced from Yang et al., 2014 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

26

Table 3.1

List of set of predictors and brief description of methods used . . .

37

Table 3.2

List of all potential input variables . . . . . . . . . . . . . . . . . . .

39

Table 3.3

Evaluation of neural networks with various set of predictors . . .

43

Table 3.4

List of set of predictors and brief description of methods used . . .

45

Table 4.1

Statistics of SCADA signals and loads measured . . . . . . . . . . .

54

Table 4.2

Wind farm flow scenarios and sectors per wind turbine (distance

-in rotor diameters- to closest neighbour) . . . . . . . . . . . . . . . . . . Table 4.3

Predictors selected from available statistics of SCADA signals . . .

Table 4.4

Average performance of predictions for ∆eq of blade root bending

moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

55 60 63

Table 4.5

Relative error of predictions for ∆eq of blade root bending moments 65

Table 4.6

Overall damage equivalent loads (∆eq ) normalised with measure-

ments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Table 5.1

Data available for the assessment. Statistics of SCADA signals and

mechanical loads. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Table 5.2

Input variables selected from 10-minute statistics of SCADA data.

Table 5.3

Average prediction performance for ∆eq of blade root bending

moments on similar data. . . . . . . . . . . . . . . . . . . . . . . . . . . . Table 5.4

66

80 81 87

Average prediction response for ∆eq of blade root bending moments

on similar data. Over-performing models are listed with gray background. 87

xiv

LIST OF TABLES

Table 5.5

Average predictions performance for ∆eq of blade root bending

moments in tests on dissimilar data. Best results are highlighted in gray background. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Table 5.6

90

Average prediction response for ∆eq of blade root bending moments

in tests on dissimilar data. Best results are highlighted in gray background. 90 Table 5.7

Prediction error of accumulated damage equivalent loads (∆eq nor-

malised with measurements) with Feed-Forward Neural Networks (FNN) and K-Nearest Neighbours (K-NN) for blade root bending moments in edgewise and flapwise directions, respectively. . . . . . . . . . . . . . . . Table 7.1

91

First four statistical moments of the value distributions shown

in Figure 7.9d for the normalised measurements and the reconstructed signals with each one of both models. . . . . . . . . . . . . . . . . . . . . 124 Table 7.2

Performance of both models, using the metrics defined in Equations

(7.2), (7.9)–(7.11), namely the mean of the absolute error, Equation (7.9), its standard deviation, Equation (7.10), the mean square error, Equation (7.2) and its standard deviation, Equation (7.11). . . . . . . . . . . . . . . 125

List of Figures Figure 1.1

Scheme representing different linear life consumptions over a

component time in operation (dotted line) including potential safety risk and additional use gained with or without monitoring (Hyers et al., 2006) Figure 2.1

Schematic of a bathtub-shaped failure rate during the operating

lifetime of machinery (Faulstich et al., 2011) . . . . . . . . . . . . . . . . . Figure 2.2

13

Example of main systems, sub-systems and components of mod-

ern wind turbines (Tchakoua et al., 2014) . . . . . . . . . . . . . . . . . . . Figure 2.3

4

14

Schematic representation of deterioration (P-F curve), identified

at time P, which leads to failure at time F based on the monitoring of two different attributes: temperature (left) and vibration (right) (Fischer and Coronado, 2015) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 2.4

The relation of (condition) monitoring systems and maintenance

strategies (Besnard et al., 2010) . . . . . . . . . . . . . . . . . . . . . . . . . Figure 2.5

18

Progression of failure over time (left) and along component-to-

system hierarchy (right) (Hyers et al., 2006) . . . . . . . . . . . . . . . . . . Figure 2.7

17

Reported reliability in wind turbine sub-assemblies from an on-

shore survey (Faulstich et al., 2011) . . . . . . . . . . . . . . . . . . . . . . Figure 2.6

16

20

Overview of condition monitoring (top) and its relation with

diagnostic (bottom left) and prognostics (bottom right) (Tchakoua et al., 2014) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 2.8

21

Probabilistic representation of deterioration prognosis. Proba-

bilistic degradation (a), failure detection (b) and quality of prediction (c) (Dragomir et al., 2009). pdf: probability density function, perf: performance, pred: prediction, RUL: remaining useful life, TT: time to failure, T: time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 2.9

22

Example of measurements carried out to investigate a gearbox

(Sheng, 2014) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

23

xvi

LIST OF FIGURES

Figure 3.1

Venn diagram displaying the relationship between different set of

potential input variables and target values . . . . . . . . . . . . . . . . . . Figure 3.2

28

Diagrams representing different approaches for input variable

selection algorithms (May et al., 2011): (a) wrapper, (b) embedded and (c) filter. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 3.3

Simplified fatigue load monitoring system . . . . . . . . . . . . .

Figure 3.4

Number of times that a predictor has been used in load monitoring

31 33

systems for wind turbines (from Cosack and Kuhn, ¨ 2009 to Vera-Tudela and Kuhn, ¨ 2013). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 3.5

Relation between potential input variables (simply variables),

features and predictors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 3.6

36

(a) Wind turbine 01 in offshore wind farm EnBW Baltic 1; (b)

Operating conditions in measured data set . . . . . . . . . . . . . . . . . Figure 3.7

34

38

(a) Regression for blade out of plane bending moment using a

feed-forward network trained with all potential input variables; (b) Error per value in the data set for the same network . . . . . . . . . . . . . . . Figure 3.8

Partial ranking of potential input variables based on their Spear-

man coefficient, shown in descendent order . . . . . . . . . . . . . . . . . Figure 3.9

39 40

Symbolic representation of cross-correlation between predictors,

shown using set Ranking . . . . . . . . . . . . . . . . . . . . . . . . . . . .

41

Figure 3.10 Dendrogram of predictors from set Ranking (variances are excluded) 42 Figure 3.11 (a) Scree plot of the 10 principal components; (b) Pareto of variance explained by the 10 principal components . . . . . . . . . . . . . . . . . . Figure 4.1

Probability density function of a process within lower and upper

specification limits, LSL and USL respectively (Montgomery, 2009) . . . . Figure 4.2

56

Normalised ∆eq for the blade root flapwise bending moment (blade

1 of wind turbine B01) over wind speeds . . . . . . . . . . . . . . . . . . . Figure 4.6

53

Relative number of 10-minute observations per scenario and wind

turbine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 4.5

52

Normalised Wind rose (left) and location of wind turbines B01

and B08 in EnBW Baltic 1 layout (right) . . . . . . . . . . . . . . . . . . . Figure 4.4

48

Probability density function of two predictor models with unequal

bias and variance, modified from Montgomery, 2009. . . . . . . . . . . . . Figure 4.3

42

Example ranking of correlation coefficients between statistics of

SCADA signals and ∆eq blade root edgewise bending moment (displaying

57

LIST OF FIGURES

only 10 highest coefficients) . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 4.7

59

First 1000 ∆eq of blade root bending moments in flapwise direction

in scenario all data (blade 1 of wind turbine B01) . . . . . . . . . . . . . . Figure 4.9

58

Example heatmap showing high correlation between the 10 statis-

tics of SCADA signals presented in Fig. 4.6 . . . . . . . . . . . . . . . . . Figure 4.8

xvii

62

Predicted and measured ∆eq of blade root bending moment in

flapwise direction in scenario all data (blade 1 of wind turbine B01) . . .

62

Figure 4.10 Prediction error for ∆eq of blade root bending moment in edgewise (top) and flapwise (bottom) directions (blade 1 of wind turbine B08) . . .

64

Figure 4.11 Probability density function of prediction error for ∆eq of blade root bending moment in edgewise (top) and flapwise (bottom) directions (blade 1 of wind turbine B08) . . . . . . . . . . . . . . . . . . . . . . . . . Figure 5.1

Sketch of a classification problem with imbalanced classes in a

data set (Ertekin et al., 2007). . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 5.2

65

73

Example of a Six Sigma process with the mean of its distribution

shifted by ±1.5σ with respect to the distribution centred within the specification levels (Montgomery, 2009) . . . . . . . . . . . . . . . . . . . . Figure 5.3

Procedure followed to estimate the performance of fatigue load

estimations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 5.4

74 77

Schema of probability density functions of prediction errors. One

centred on the target value and within tolerance limits (continuous line), another including a bias with low precision (discontinuous line) (VeraTudela and Kuhn, ¨ 2017). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 5.5

Normalised wind rose (left) and location of wind turbines B01

and B08 in EnBW Baltic 1 layout (right) (Vera-Tudela and Kuhn, ¨ 2017). . . Figure 5.6

78 80

Distribution of data sets, model development (train), tests on

similar and dissimilar data. Sets of data are named using turbine (B01, B08) and blade names (b1, b2). . . . . . . . . . . . . . . . . . . . . . . . . Figure 5.7

84

Test on similar data for ∆eq of blade root bending moment in flap-

wise direction. Serie of 100 data points. Measured (grey) and predicted (black) ∆eq with Linear Regression (LR, left) and Extra Tree Bagging Ensemble (ETBE, right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 5.8

85

Evaluation in fair test for DEL of blade root bending moment in

flapwise direction. Measured (grey) and estimated values (black) with Linear Regression (LR, left) and Extra Tree Bagging Ensemble (ETBE, right). 86

xviii

LIST OF FIGURES

Figure 5.9

Test on dissimilar data for ∆eq of blade root bending moment

in flapwise direction. Serie of 100 data points. Measured (grey) and estimated values (black) with Linear Regression (LR, left) and Extra Tree Bagging Ensemble (ETBE, right). . . . . . . . . . . . . . . . . . . . . . . .

88

Figure 5.10 Test on dissimilar data for ∆eq of blade root bending moment in flapwise direction. Measured (grey) and estimated values (black) with Linear Regression (LR, left) and Extra Tree Bagging Ensemble (ETBE, right). 88 Figure 5.11 Test on dissimilar data. Probability distribution of prediction errors for ∆eq of blade root bending moment in flapwise direction. Models developed with partial data of B1b1 were evaluated with similar data (B1b1) and dissimilar data (B1b2, B8b1 and B8b2). Results from Linear Regression (LR, left) and Extra Tree Bagging Ensemble (ETBE, right) are shown. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Figure 6.1

89

A schematic representation of the Cross Industry Standard Process

for data mining (CRISP-DM) (Provost and Fawcett, 2013). The process starts with the business understanding phase to then focus on the datamining problem itself during the next four phases that lead to the final deployment phase. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

96

Figure 6.2

Schematic description of the data warehouse for the research. . . 101

Figure 6.3

Example ranking (some) potential input variables to predict dam-

age equivalent loads of blade root bending moment on flapwise direction (left) and identification of redundant variables (right). . . . . . . . . . . . 102 Figure 6.4

Measured and estimated damage equivalent loads of blade root

bending moments in edgewise (left) and flapwise (right) directions, normalized with measurements values. Estimations with K-Nearest Neighbours (blue triangles) and Feed-Forward Neural Networks (green circles) in hold out data from B08 are displayed. . . . . . . . . . . . . . . . . . . . 103 Figure 6.5

Estimated accumulated fatigue loads of blade root bending mo-

ments in edgewise (left) and flapwise (right) directions, obtained with K-Nearest Neighbours (left) and Feed-Forward Neural Networks (right) in various turbines at the EnBW Baltic 1 wind farm . . . . . . . . . . . . 104 Figure 7.1

Alternative connection of white and black box models in grey box

modelling (Sohlberg and Jacobsen, 2008). Parallel (left) and serial (right). . 108 Figure 7.2

Alternative approaches considered during the investigation. . . . 109

Figure 7.3

Example timeseries of blade root out of plane bending moment,

LIST OF FIGURES

xix

estimated with a NARX network (Vera-Tudela and Kuhn, ¨ 2013). . . . . . . 110 Figure 7.4

Location of the Alpha Ventus wind farm at Borkum West in the

North Sea (54.3◦ N–6.5◦ W) together with a picture of turbine AV04. Examples of (a,b) wind speed and (c,d) tower acceleration signals in Alpha Ventus wind farm. In the left one illustrates the time-domain of both quantities while in the right the frequency domain is shown for the full datasets of October 2014. Scatter plot v × a is shown in the inset of (a). All data was normalized to fulfil all confidentiality protocols, namely normalization to maximal values (see text). Both wind velocity and tower acceleration are normalized to maximal observed values, vmax and amax respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 Figure 7.5

Schematic illustration of the neural network (NN) approach, in-

dicating the development and test of neural networks for signal reconstruction. Here the particular case of tower acceleration reconstruction is given. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 Figure 7.6

Schematic illustration of the Langevin (stochastic) approach. Here

the particular case of tower acceleration reconstruction at Alpha Ventus is given. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 Figure 7.7

Illustration of the stochastic approach, while integrating the evolu-

tion Equation (7.4), composed by its two contributions, the deterministic contribution D(1) which governs the tendency of the evolving observable (blue), and the stochastic fluctuations accounted by D(2) (red) added to the deterministic part. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 Figure 7.8

Illustration of (a) the first conditional moment for three different

bin values of the tower acceleration as defined in Equation (7.5). Here v∗ is given by the average velocity found during October. Numerical result for (b) the drift D(1) (a, v) and (c) the diffusion D(2) (a, v) in the Langevin equation given by Equation (7.6), plotted as function of a alone, i.e., they are projected at the a-axis to emphasize the linear and quadratic dependency of D(1) and D(2) respectively for the largest range of acceleration values. In red a polynomial fit of both functions is shown. We use bins for the velocity with a width of 0.017 in units of maximal velocity while acceleration was taken in bins of width 0.017 of maximal tower acceleration.122 Figure 7.9

Sample of data series of tower acceleration from measurements at

AV04 in Alpha Ventus and the acceleration reconstruction models using (a) NARX neural networks and (b) the Langevin model. In both cases,

xx

LIST OF FIGURES

one also plots (black lines) the corresponding series of measurements for better comparison. The corresponding value distribution of these series is displayed in (c) with symbols for the full month of November 2014. . . 123 Figure 7.10 Spectrum of original signal (top) and the corresponding reconstructed signals, one using the neural network (middle) and the other using our proposed stochastic approach (bottom). Notice that the high frequency domain is plotted in a linear frequency scale, while the low frequency domain is plotted in the logarithmic scale. . . . . . . . . . . . 126 Figure 7.11 Two-point statistics of the tower acceleration (lines) and the corresponding reconstructed signals, i.e., value distributions of ∆a(t) = a(t + τ) − a(t). From top to bottom one has τ = 1, 2, 4, 8 and 16 s. The vertical shift of the distribution is for better visualization. . . . . . . . . . 127

List of Abbreviations Abbreviation

Definition

ABE

Adaptive boosting ensemble

AI

Artificial intelligence

ANN

Artificial neural network

ARMA

Auto-regressive moving average

BMU

Federal Ministry for the Environment, Nature Conservation and Nuclear Safety

BMWi

Federal Ministry for Economic Affairs and Energy, Germany

BSH

Federal maritime and hydro-graphic agency, Germany

C

Clustering

CI

Computational intelligence

CM

Condition monitoring

CMS

Condition monitoring system

CRISP-DM

Cross industry standard process for data mining

DEL

Damage equivalent load

DIBt

German institute for the construction sector

DLC

Design load case

DM

Data mining

DNV-GL

Det Norske Veritas - Germanischer Lloyd

EA

Evolutionary algorithms

EC

Ensemble classifiers

EnBW

¨ Energie Baden-Wurttemberg AG

ES

Expert systems

xxii

LIST OF ABBREVIATIONS

ETBE

Extra tree bagging ensemble

FA

Frequency analysis

FL

Fuzzy logic

FNN

Feed-forward neural network

GWEC

Global wind energy council

IVS

Input variable selection

KM

K-means clustering

IEC

International Electrotechnical Commission

KNN

K-nearest neighbours

LIDAR

Light detection and ranging

LM

Load monitoring

LMS

Load monitoring system

LR

Linear regression

LSL

Lower specification limit

M

Malahanobis distance

MA

Modal analysis

MAD

Median absolute deviation

MAPE

Mean absolute percentage error

MAS

Multi-agent system

MEASNET

Network of European Measuring Institutes

MLP

Multilayer perceptron

MTBF

Mean time between failures

MTTR

Mean time to repair

NARX

Nonlinear autoregressive exogenous model

NDT

Non-destructive test

NN

Neural network

NREL

National Renewable Energy Laboratory, USA

NTL

Natural tolerance limits

OA

Optimization algorithms

O&M

Operation and maintenance

LIST OF ABBREVIATIONS

PCA

Principal component analysis

PF

Physics of failure

PPM

Parts per million

PR

Polynomial regression

QCC

Quality control charts

RAVE

Research at alpha ventus

RE

Relative error

RMSE

Relative mean square error

RNN

Recurrent neural network

RPM

Rotations per minute

RT

Regression tree

RUL

Remaining useful life

R2

Coefficient of regression

SA

Stochastic approach

SCADA

Supervisory control and data acquisition

SDOFAE

Standard deviation of absolute error

SDOFSE

Standard deviation of square error

SHM

Structural health monitoring

SI

Swarm intelligence

SVM

Support vector machines

TTF

Time to failure

USL

Upper specification limit

VDI

The association of German engineers

VM

Virtual models

WTM

Wait to maintain

xxiii

Chapter 1 Introduction Ageing wind turbines are of interest to understand and better predict the mechanisms that lead to their deterioration and lose of performance. Knowledge gained from better estimating their lifetime will benefit current and future installations. At the end of 2016, there were more than 486 GW of wind power capacity installed globally (GWEC, 2017). The spread and deepness of the wind energy market penetration (already present in more than 80 countries with 29 of them having more than 1 GW of installed wind capacity according to GWEC, 2017), indicates that improving performance and lifetime extension of wind turbines is becoming a very competitive business. Monitoring consumed lifetime with accumulated fatigue loads should help understand the connection between design criteria to the particular loading that wind turbines experience during their operational lifetime. In this thesis, a data-driven methodology that estimates fatigue loads from 10-minute statistics of SCADA data is presented.

1.1

State of the art

Technological development of wind turbines can be partially attributed to their design meant to facilitate machinery production in series. This reasonable approach improved their competitiveness with other energy sources as they benefited from economies of scale. Such progress was made at the cost of taking some conservative design assumptions. In order to ensure that wind turbines are also able to withstand various site-specific wind conditions, they are designed to comply with standarized site assumptions, also known as type classes according to IEC, 2005a. In general, wind turbines are installed in sites with less-severe atmospheric conditions than those assumed for the type class they were designed to. The difference

2

1. Introduction

between a conservative design criteria and a less-severe environment, that is the design reserve, implies the possibility of using machinery beyond its design lifetime. To maximize machinery usage, it is then of interest to verify and to quantify the margin left between conservative design assumptions and machinery deterioration at specific sites and under real operational conditions. However, the expected lifetime of wind turbines is commonly a value prescribed by designers early in the design process of new turbines. Although the number of years is not regulated, its value can be expected to be around 20 and 25 years, for onshore and offshore applications respectively. Such values are based on financial return and probability of failure (Veldkamp, 2008). Thus, design lifetime is used as an input to calculate accumulated fatigue loads, namely the fatigue loads envelope that mechanical components have to withstand during their operation. Therefore, if one investigates the inverse problem, it is reasonable to assume the possibility of monitoring consumed lifetime based on their accumulated fatigue loads. Furthermore, one could compare fatigue load estimations to design envelopes and then associate their difference to site-specific wind conditions (M. Hau, 2006). Some applications around this relationship are of current interest in the wind energy industry: advanced control, condition based O&M strategies, re-powering and lifetime extension. An interest that has been fueled by the almost ubiquitous record of operational data via Supervisory control and data acquisition (SCADA) systems.

1.2

Problem scope

The motivation that led to this thesis was to further develop the approach proposed by Cosack, 2010, which relates fatigue loads and SCADA data for single wind turbines, whereas the emphasis in this thesis was centred around turbines installed in wind farms. Thus, in a broad sense, its conception implied the idea of an improvement. It is then necessary to clarify its previous state-of-the-art before giving further details about how this improvement was achieved. While in this chapter the thesis scope and its contribution are stated, in the second chapter presents and reviews relevant concepts to comprehend decisions made during the formation of the approach.

1.2.1

In relation to monitoring fatigue loads

Monitoring accumulated fatigue loads, as presented in this thesis, is aimed to provide an indicator of lifetime consumed in wind turbines, which uses are envisioned to help

1.2 Problem scope

3

improve operation and maintenance (O&M) strategies as well as to maximize the use of the machinery. It is worth notice that the relationship between lifetime and fatigue loading is not investigated in this thesis, but only the possibility of estimating the last one by other sources rather than its direct measurement, i.e. nor structural damage models were studied neither the variations in material properties were taken into account when making estimations. Fig. 1.1 shows an idealized representation of life consumption (vertical axis) over its current time in operation (horizontal axis) as presented by Hyers et al., 2006. It presents how continuous monitoring can contribute to identify potential safety risks, e.g. in machinery exposed to severe usage or, on the contrary, to discover additional use gained when machinery is subjected to mild usage. Both options are defined with continuous lines, on top and below what is described as current service lifetime. The evaluations are made when the design life of component is drawn in the horizontal and when the baseline time without diagnosis is drawn in the vertical. It is worth notice that the schematic is meant to explain and discuss the motivation behind monitoring. Despite its obvious benefits, monitoring accumulated fatigue loads is limited by the complexity and extra costs associated to the installation of purpose-specific sensors at each of the positions relevant to safeguard the structural integrity of wind turbines (Cosack, 2010). Then, the challenge is to find a cost-effective alternative to the installation of extra sensors, which is expected to be penalized with an increase of uncertainty in the estimation of fatigue loads.

1.2.2

In relation to the use of SCADA data

Since most modern wind turbines include Supervisory Control and Data Acquisition (SCADA) systems, it is of interest to describe the relationship between 10-minute statistics recorded with SCADA systems and accumulated fatigue. Previous to this thesis other researchers explored and reported their relationship: Cosack, 2010; Cosack and Kuhn, ¨ 2009; Hofemann et al., 2011; T. S. Obdam and Braam, 2009; T. S. Obdam et al., 2009a,b; T. Obdam et al., 2010; Perisic et al., 2011; Smolka, Kaufer et al., 2012; Smolka and P. W. Cheng, 2013; Smolka, Quappen et al., 2011. In the context of this thesis, fatigue loads refer to represent the overall global or system loads in the turbine, i.e. bending moments at blade roots, bending moments and torque at main shaft and bending moments at the top and bottom of the tower. Specific sections and hot spots in wind turbine components are not investigated. But, despite their common goal in estimating loads as an alternative to their direct

4

1. Introduction

Figure 1.1: Scheme representing different linear life consumptions over a component time in operation (dotted line) including potential safety risk and additional use gained with or without monitoring (Hyers et al., 2006) measurement, investigations had different specific views when it came to the potential use of load monitoring systems, which invited to their careful review. Furthermore, they used difference data sources, which ranged from synthetic data out of aero-elastic simulations to measurements from selected onshore and offshore wind turbines. Despite these differences, a summary of some of the most relevant findings should shed more light into the specificities of the field: • There is a relationship between 10-minute statistics of SCADA data and 10-minute fatigue load indicators. • Damage equivalent load is the preferred indicator that has been used to represent fatigue loading. • Each 10-minute instance can be modelled independently of its predecessors, thus in this thesis preference is given to time-independent regression models when using 10-minute statistics. • Polynomial regressions are too complex to model the large number of 10-minute statistics recorded with SCADA systems (around 100).

1.2 Problem scope

5

• Regression models built with artificial neural networks are suited to model the relationship between 10-minute statistics of SCADA data and fatigue loads. • Synthetic data from simulations can be used for exploratory work, but free field measurements are preferred for model development and validation. • Coefficient of determinations (R2 ) are used to assess how models based on SCADA data explain the variability of fatigue loads. • Models are considered satisfactory when coefficients of determination are higher or equal to 0.9 (R2 ≥ 0.9). • Estimations made for free-stream conditions are deemed satisfactory, while those made for wake conditions are not. Although the major characteristics of the approach seemed to be defined, some questions arose after a careful description of the overall modelling process. In rather colloquial form, modelling fatigue loads starts with the record of what is considered to be a representative time of wind turbines lifetime. Then, some of the 10-minute statistics of SCADA data are chosen and fed to a regression model (an artificial neural network) for the optimization of its parameters on part of the data. Afterwards, the regression model is used to estimate 10-minute fatigue loads indicators on data hold out and results are reported. But, since it is no possible to investigate turbines from cradle to grave, the exercise is rather a proof of concept, which is expected to be extended by others to reproduce with larger data sets. At this point, and motivated by maintenance management problems (Sanz-Bobi, 2014), the following list of challenges were considered of interest: • When are records considered to be representative in order to define measurement campaigns? • How are data quality issues handled when applications are offline and online? • What is the smallest set of predictors needed to build a predictive model? • How is prediction quality assessed and what type of assessment is preferred towars its use in the industry? • How is prediction quality affected by wind farm flow conditions? • What alternative regression models are better suited for the task?

6

1. Introduction

• How and for which industrial application is the outcome of the process better suited?

1.3

Research questions and objectives

The objectives in this research are based on the idea that this thesis is a communication with the body of knowledge in the field, and that readers take the effort to read it interested in following on the findings here communicated. Thus, it has three objectives: first, to propose a methodology (process) to monitor fatigue loads on wind turbines installed in wind farms, which is structured to facilitate its improvement over time. This implies that the process steps have to be clearly defined in order to be replaced by better approaches as research is followed up by other investigators. Second, to clarify the process steps through various intermediate scientific questions, which should serve to structure future research field carried out by others. Since some of the process steps can be seen as a field of research of its own, e.g. selecting input variables, some interesting questions and hypotheses are left outside the scope of this thesis and were considered less relevant for the challenges that formed this thesis and were listed in the prvious section. Third, to focus on those questions that are expected to lead the methodology towards its implementation at the industry. This last objective forms the chore part that motivated this thesis and can be perceived by the list of challenges described in the previous section. Thus, the usability of the approach became the mantra that helped to scope the thesis with the following specific questions: 1. What is the smallest set of predictors needed from 10-min statistics of SCADA to estimate fatigue loads with an acceptable prediction quality? 2. What is a suitable metric that serve to assess prediction quality meant for their use in applications? 3. How fatigue loads prediction quality changes when they are made for other turbines in the same wind farm? 4. How should be load monitoring development tailored towards its implementation?

1.4 Structure of the thesis

1.4

7

Structure of the thesis

The first chapter introduced the reader to the state of the art in fatigue load monitoring of wind turbines and clarified the research questions and motivations for writing this thesis. Next, chapter 2, provides an explicit distinction between concepts that served to frame the problem. Later, chapters 3 to 7 discuss specific problems in the baseline approach and present original scientific contributions; chapter 3 describes a filter approach to select a minimum set of input variables from the ’universe’ of SCADA data and responds to the first objective of the thesis; chapter 4 proposes metrics to better assess prediction quality and evaluates the impact of wind farm flow conditions, which answers the second question, and; chapter 5 quantifies the deterioration of prediction quality when they are made for other turbines in a wind farm, which answers the third question of the thesis. Then, chapter 6 presents the implementation of the methodology in a wind farm, which answers the fourth question of the thesis; then, chapter 7 briefly documents other approaches to monitor wind turbines, which is meant to communicate what other ideas were considered during the construction of the approach and were deemed less relevant. Finally, chapter 8 closes the thesis with a summary and an outlook. Conscious of the diversity of interests that researchers have during their investigations (Sword, 2012) and motivated by Cortazar, 1963, this thesis is organised to offers various paths to its readers: 1. Normal path, followed by those who sequentially read the thesis from begin to end. 2. Curious path, for those who are only interested to gain an overview of the research field and its current challenges. They are suggested to start in chapter 2, then continue with the introduction parts of chapters 3 to 7, and might skip the rest. 3. Explorer path, conceived for those who are only interested in reading state-of-theart research in monitoring fatigue loads of wind turbines. They should explore the sections 3.2, 4.2, 5.2 and 7.3, the list of publications and might skip the rest. 4. Researcher path, followed by those who are looking for a knowledge gap and how to orient their own research on monitoring fatigue loads of wind turbines. They should start with the summary and outlook in chapter 8, continue with chapter 1 and plan their own path.

8

1. Introduction

Chapter 2 Monitoring and modelling lifetime At this stage, various ideas and concepts need clarification and this chapter will serve to explain them. First the main intention attained with the models investigated in this thesis is explained. Second, the type of models available to build a methodology that monitors fatigue loads on wind turbines are enumerated and compared, which serves to justify the specific approach selected in this thesis. Third, the specific methodology is discussed with respect to more established monitoring techniques, like structural health, condition and performance monitoring. Finally, the state of the art i monitoring wind turbines is framed with respect to current standards, guidelines and data recorded. These considerations are needed to justify decisions made to construct the specific methodology presented in this thesis.

2.1

Monitoring and modelling objectives

Monitoring means to observe regularly a quality (of something), thus it implies the quality is not constant over time, and the monitoring is meant to restore or avoid the change in quality; thus monitoring in wind turbines could refer to provide information for decision making aimed at either the increment of their energy performance or the reduction of their mechanical deterioration. Although both areas of research are sometimes combined, most of the investigations focus only on one of them, since the methods applied differ. Those investigations around increasing performance are referred to electrical energy gains and can be found in literature related to wind speed prediction, wind power forecast, wind turbine (or wind farm) control and wind farm optimization. The term monitoring in this thesis only deals with deterioration, failure rate and failure duration of mechanical components.

10

2. Monitoring and modelling lifetime

Models presented and discussed in this thesis are meant to have a high prediction capability as requirement. Although it might seem evident at first sight, this is explicitly stated since literature around fatigue loads shows a field preference for explanatory models, those where physical relationships have priority over predictive models. Although it is commonly assumed that explanatory models also have high predictive power, Shmueli, 2010 challenged this belief, since explanatory models are based on idealized processes, whose independent idealized variables are not directly recorded. Explanatory models used in wind turbines are faced with the challenge to measure wind direction in front of wind turbines and to characterise wake conditions over the whole rotor (Gaumond et al., 2014). When prediction is the main goal, rather than explanation, modelling and validation seem to have substantial differences and requirements. For predictive models observational rather than experimental data are preferred, as they include a more realistic context in which predictions are expected to be made. The opposite is experienced also as true, i.e. experimental data obtained under controlled conditions are preferred to test hypothesis made with explanatory models. Thus, in this thesis models that search for causality are not a goal during their construction, although they are preferred if available.

2.2

Type of models

The previous chapter explained that modelling fatigue loads of wind turbines in wind farms with SCADA data was meant to estimate their lifetime. Then, the next question to answer was how would it be better accomplished, since selecting a type of modelling approach means accepting its advantages and disadvantages. There is a large body of literature detailing many options as shown by the overview of those used in wind energy presented in Table 2.1, which lists references according to targets and models utilised. Fortunately, Welte and K. Wang, 2014 classified the different approaches to estimate lifetime of wind turbine components according to the field of knowledge used. They identified physical, stochastic, data-driven and hybrid models, each of them will be briefly discussed in the following sections.

2.2.1

Physical models

Physical models describe lifetime or failure in physical terms, which makes them highly dependent on specific parameters of the machinery and a precise representation of the

2.2 Type of models

11

environmental conditions, i.e. loading situations. Welte and K. Wang, 2014 referred them as white-box or transparent models because virtually all model information is explicitly declared, which should facilitate their understanding and improvement. Aero-hydroservo-elastic models (in general aero-elastic models, or simply physical models in this thesis) are widely used to estimate lifetime in terms of accumulated fatigue loads, either to design wind turbines (IEC, 2005a) or to decide upon their lifetime extension (DNV-GL, 2016b). They are as complex as the individual fields listed in their denomination and, as a consequence, there is a large body of literature describing the phenomena they model, which can be found describe in Burton et al., 2002; DNV/RISO, 2002; Gasch and Twele, 2011; M. O. L. Hansen, 2008; E. Hau, 2006; Manwell et al., 2002 and Jamieson, 2011. Moreover, to ensure the integrity of wind turbines and to safeguard them from uncertainties in material strength and aero-elastic calculations, safety margins are used to increase calculated loads and to reduce expected material strengths during their design. Since physical models require both machinery details and a deep understanding of the process governing the change in properties monitored, they are based on physical abstractions, which rely on specific attributes required as inputs for calculations (Rossi, 2014; Shmueli, 2010). These attributes, like for example: wind speed and wind direction, are better obtained in controlled conditions than in open field measurements (Gaumond et al., 2014). In the case of lifetime prediction of wind turbines, some assumptions made are the linear accumulation of fatigue loads and the determination of S-N curves independent of the median load (Castillo and Fernández-Canteli, 2009; Vassilopoulos and Keller, 2014), which are considered valid within the scope of this thesis and in accordance to current industry standards (IEC, 2001, 2005a).

2.2.2

Data-driven models

Data-driven models are known under various names: artificial or computational intelligence (AI, CI), machine learning, data-mining or statistical learning (Bishop, 2006; Hastie et al., 2013; Reis and Pati, 2000; Tran et al., 2010; Welte and K. Wang, 2014). They are built based on the relationships between input variables and target values, and are optimised to minimize a cost-function, which is related to their estimation error. They are also named black-box models because the relationship between their parameters and external criteria from the phenomena is not clearly revealed. It is worth note that Adams et al., 2011 described structural health monitoring, in general terms, to be a problem of pattern recognition for which data-driven models

12


are well-suited. Additionally, the extensive research of Kusiak and his colleagues, exemplified with results presented by Kusiak and Z. Zhang, 2011; Kusiak, Song et al., 2009; Kusiak and Z. Zhang, 2012; Kusiak and Zheng, 2010; Kusiak et al., 2010, expanded the number CI applications to wind energy problems. With respect to monitoring fatigue loads in wind turbines, previous research reports have indicated the suitability of artificial neural networks to model the relationship between 10-minute statistics of SCADA data and fatigue loads (Cosack, 2010; Noppe et al., 2016; T. Obdam et al., 2010; Smolka and P. W. Cheng, 2013; Smolka, Quappen et al., 2011). Also higher frequency (above 1 Hz) time-series have been explored using Langevin dynamics (Rinn, Lind, Wächter et al., 2016), for example to reconstruct series with mechanical loading that are used to estimate equivalent fatigue loads (Lind, Herráez et al., 2014).

2.2.3

Stochastic models

Stochastic models are based on probabilities of failure and serve to estimate uncertainties and confidence intervals around expected lifetime (Welte and K. Wang, 2014). They are based on the parametrization of probability distributions, which use data sets with failure rates and time between failures in wind turbines. Thus, they are better suited for those problemas associated to known and quasi constant failure rates, and when external conditions do not dominate inter-turbine variations in the data set, like in the flat region of the bathtub-shaped failure rate curve, which is shown in Fig. 2.1. A simple model, based on a constant failure rate, can be related to the concept of timebased maintenance. Since stochastic models are generic by nature, they can be used for general assessments. However, they do not include specific information related to the deterioration of materials in components.

2.2.4

Hybrid models

Hybrid models are those formed by the combination of two or more models previously described (Welte and K. Wang, 2014). A more specific type of hybrid models are grey models, which imply the combination of a physical (white) model and another datadriven (black) one. They are arguably the less developed type of models in wind energy. Schröder, 2014explored the use of an hybrid model to estimate fatigue loads in turbines of different rating, it was formed by a neural-network developed with a specific turbine and placed in series with a linear regression model to improve its generalization when used for turbines of different rating. More details about this type of models is given in section 7.1.

2.3 Monitoring

13

Figure 2.1: Schematic of a bathtub-shaped failure rate during the operating lifetime of machinery (Faulstich et al., 2011) Hybrid models are created with the intention to acquire or increase an advantage or to avoid or mitigate a certain disadvantage (Oussar and Dreyfus, 2001). According to Y. Liu and Nayak, 2012, the future of modelling (and monitoring) has a multi-disciplinary character, which should lead to the integration of various specific fields of knowledge, like vibration or signal processing, data mining, hardware optimization and intelligent material. In a review of the field, Kusiak, Z. Zhang and A. Verma, 2013 described hybrid models for wind speed prediction next to those based on statistical, physical and data-mining approaches; they reported that statistical and data-mining models were utilized for short- and medium-term prediction, while physical models were utilized for long-term prediction.

2.3

Monitoring

The concept of monitoring something (e.g. an attribute in a process) supposes the act of supervising it (that attribute) and implies the intention to control it. Therefore, one can refer to monitoring as the process that passively oversees the quality of something, and which is expected to return actionable information aimed to maintain or restore that quality to a given target value. The value of monitoring is deeply related to usability of its outcome, actionable information, and to the potential value gained when the quality

14


under surveillance is restored to its target value. Then, to design a complete monitoring system, one needs to include its two main characteristics: the first one consist of continuously assessing an attribute related to its quality and comparing it to a given target level, the second one refers to having or creating actionable information from its supervision (Hyers et al., 2006), which assumes that one knows exactly what to do after the monitored attribute has deviated from its expected quality. Since the definition of quality generally depends on more than one attribute, like temperature or vibration, monitoring is a multidimensional concept. Thus, the use of the term monitoring can lead to confusion as it may refer to supervising the whole wind farm, a single wind turbine, a system (pitch system), a sub-system (pitch mechanical transmission) or a specific component (pitch motor). In general, literature about monitoring wind turbines commonly refers to components and systems, whereas fatigue load monitoring is aimed at a system level.

Figure 2.2: Example of main systems, sub-systems and components of modern wind turbines (Tchakoua et al., 2014) Monitoring, as a research topic, is a complex endeavour due to the many components found in modern wind turbines as shown in Fig. 2.2, which is exacerbated by the various attributes required to define their quality state. For example, to monitor the quality

2.3 Monitoring

15

state of blades, one might be interested in supervising bending moments, vibration levels and crack initiation or propagation; to monitor drive-train bearings one might be rather interested in vibration levels and the metallic particulate found in oil. Thus, this multi-dimensionality and complexity needs to be kept in mind when the outcome of the fatigue load monitoring system is foreseen as integrated into other monitoring systems. Obviously, a simple load monitoring approach rather than a complex one is preferred.

2.3.1

Monitoring strategies

Many concepts in monitoring find their analogy with human-health related concepts (Crabtree, 2011). One can understand structural health monitoring (SHM), condition monitoring (CM) and performance monitoring (PM) as similar approaches meant to assess a quality or attribute related to the health of wind turbines, which slightly differs in their methods and data used. In general, they can be loosely defined as the examination of vital signals with the purpose to detect abnormal behaviour that can lead to system failure or underperformance. Fischer and Coronado, 2015, provide a good example, which is included in Fig. 2.3; it describes anomaly detection (P) via monitoring two attributes. On the left side, failure (F) is detected with temperature, while on the right side a better P-F interval is found when vibration is used to detect it (Besnard et al., 2010). The difference among these three monitoring terms (PM, CM and SHM) is subtle and their boundaries are not clearly defined, which could lead to misplace the requirements on load monitoring systems. One way to infer their differences is by looking at their measurements used and their complexity: PM is associated to data from Supervisory Control and Data Acquisition (SCADA) systems (K.-S. Wang et al., 2014), CM can be related to a mixture of SCADA and purpose-specific sensors, while SHM uses data from purpose-specific sensors. Additionally, PM and CM are normally referred to the investigation of operating wind turbines, while SHM can also include results from experiments in bench tests. Also, if one takes a closer look to some definitions proposed for SHM, CM and PM and their relation with wind turbine health problems, one might better understand them. Adams et al., 2011 refer to SHM as the discipline that assess health over the whole life-cycle of a system, including tests during the design phase and thus, it is capable of using passive measurement and actuators to perform analysis. According to Y. Liu and Nayak, 2012 SHM is the process to implement a strategy that serves to recognise damage, where diagnosis is the combination of damage detection, localization, type

16


Figure 2.3: Schematic representation of deterioration (P-F curve), identified at time P, which leads to failure at time F based on the monitoring of two different attributes: temperature (left) and vibration (right) (Fischer and Coronado, 2015) identification and severity estimation. Therefore, the main idea behind SHM strategies is that it is the process that identifies changes in qualities or attributes of the system. Now, if the definition of SHM indicates that it evaluates the whole life-cycle of the system, including non-destructive tests (NDT) carried out in laboratory conditions, like blade static loading, one can infer that CM and PM are limited to measurements in operational wind turbines. Additionally, due to their large use and numerous commercial applications for drive-train analysis, CM is typically associated to evaluations in rotating components and is mostly associated to performance evaluation and fault detection (Adams et al., 2011; Y. Liu and Nayak, 2012; J. Wang, Gao et al., 2014). Thus, PM is rather a cost-effective alternative to CM, which is based solely on the analysis of available SCADA data, where the state of wind turbine components is inferred from historical data (Kusiak, Z. Zhang and A. Verma, 2013). Both SHM and CM use similar sensing techniques: strain-gauges, piezo-electric, transducers, acoustic, optic fibre Bragg grating and accelerometers, whereas SHM offers a wider range of options for analysis and CM is limited to operational conditions (Y. Liu and Nayak, 2012). Therefore, a monitoring fatigue loads system, as it is conceived conceived in this thesis, fell into the performance monitoring (PM) category.

2.3.2

Relevance of monitoring

Since maintenance is the traditional term that describes how to maintain machinery operating according to their design conditions, there is a relation between monitoring and maintenance (Besnard et al., 2010; McMillan and Ault, 2014). This relation is shown in Fig. 2.4 under the name condition monitoring system. Monitoring is located under

2.3 Monitoring

17

the branch of condition based maintenance, because it supervises the current quality state of machinery to create actionable information. Thus, the system only intervenes when needed and not on a regular time-basis, as in the case of time-based maintenance. In Fig. 2.4, condition base maintenance is meant to be part of preventive maintenance and the actionable information is expected to be available before a failure occurs, otherwise the effort will be made to correct the failure, as in corrective maintenance (W. Li, 2009; McKone and Weiss, 2002; Walker and Stark, 2010).

Figure 2.4: The relation of (condition) monitoring systems and maintenance strategies (Besnard et al., 2010) Thus, the relevance of monitoring can be inferred from its alternatives. Starting from the bottom left of Fig. 2.4, inspection can be deemed sufficient if machinery is confined to a close area, where maintenance crews and spare parts are ready and available (such as in manufacturing plants), which is not the case for maintaining one or more wind farms. This is more relevant in the case of offshore wind farms, where accessibility to sites is another restriction to take into account when planning maintenance. Additionally, starting from the right side of Fig. 2.4, time-based maintenance could be considered a reasonable alternative if the mean time between failures (MTBF) and the mean time to repair (MTTR) indicators were fairly stable and similar for all components and systems in wind turbines. This is a good alternative for those maintaining a large number of equivalent machines with mature technology, which are expected to experience a fairly constant failure rate. However, it is not the case in wind turbines (see Fig. 2.5), where the result of a survey of operation and maintenance of onshore

18


wind turbines in Germany is displayed (Faulstich et al., 2011). Fig. 2.5 shows a large discrepancy between frequencies and downtimes per component. Thus, logistic operation becomes another criteria to consider as part of the problem. From the left of Fig. 2.4, corrective maintenance might be of interest for those conditions where malfunctioning machinery does not affect overall productivity or where the cost of other alternatives are too high. Unscheduled maintenance fells into corrective one, and is undesired if it means to stop generating energy; de-rating wind turbines, as a function of their deterioration of systems of components instead of stopping its production is an alternative action.

Figure 2.5: Reported reliability in wind turbine sub-assemblies from an onshore survey (Faulstich et al., 2011) An example that highlights the importance of monitoring components is the average lifetime of gearboxes in wind turbines, reported to be between 6 and 8 years (Besnard et al., 2010), which themselves are expected to withstand around 20 years of operational lifetime. Additionally, the difference between electrical and mechanical failures is another relevant information shown in Fig. 2.5. While the first ones are highly probable and have relatively low downtime, the downtimes caused by the second ones are high and have a low frequency of occurrence. Since this statistic was obtained from onshore

2.3 Monitoring

19

wind turbines, their impact on offshore wind farms can be assumed to be worst, when taking into consideration larger distances to machinery and lower accessibility due to weather conditions. In onshore cases minor failures (mostly electrical) cause only 5% of the downtime, and major ones (mostly mechanical) cause 95% of the downtime (Faulstich et al., 2011). Thus, since load monitoring systems depend on the reliability of electrical components, they should be designed robust and failure resistant, e.g. they should include alternative signals to allow for continuous predictions.

2.3.3

Monitoring as a process

Although a system is said to have failed after a failure event has occurred, it is more appropriate to consider failure as a process rather than as an event. Only if seen as a process, monitoring should be taking into account the different stages of failure modes leading to a failure event. But, to consider failure modes or how machinery tends to fail or to break, one needs to know the most probable course after an imminent failure has been detected. It is due to this complexity that measurement requirements to monitor a wind turbine can easily exceed 300 channels (A. P. Verma, 2012; Z.-Y. Zhang and K.-S. Wang, 2014). Then, it is not a surprise that most wind turbine monitoring systems use a mixture of standard operational data from SCADA systems and purpose-specific sensors. This requirement, when looking at monitoring as a process, can be recognised as a weakness in the isolated development of load monitoring systems (or other at system level) since it looks into a specific part of the whole problem. Hyers et al., 2006 depicted, in Fig. 2.6, four zones of mechanical failure versus time to failure (left) as well as how a failure moves up in the hierarchy, from material to component, and so on, until it reaches plant (turbine) level, which can cause a larger damage (right). The schematic in Fig. 2.6 highlights the idea that monitoring systems aim to detect and alert over possible failures early on during the operational time. Down the road, at the beginning of the critical detection horizon located on the horizontal axis (on the left-hand side) the system must be alerted, but only return a warning to repair or remove equipment when failure is imminent, at the critical prediction horizon. Once the system has passed the economical removal point, failure causes economical losses and maintenance strategy moves from preventive to corrective, as described in Fig. 2.4. Therefore, load monitoring systems can also be envisioned as an extra indicator, which could also serve to investigate the relationship between failures and accumulated fatigue loads.

20


Figure 2.6: Progression of failure over time (left) and along component-to-system hierarchy (right) (Hyers et al., 2006)

2.3.4

From monitoring to modelling

Since monitoring and maintenance are related, most of their activities should serve to understand their interconnection. Thus, the relationship between maintenance and monitoring, shown in Fig. 2.4 can also be inverted, as depicted in Fig. 2.7. In this case, (condition) monitoring starts with pre-processing data via data acquisition, signal processing and ends with feature extraction, where the relationships between measurements and failures are mapped. In this case, diagnosis is associated to the detection of failures that are later used for corrective maintenance actions. Prognostic or fault prediction is rather used in preventive maintenance for the allocation of future planned maintenance activities. In monitoring, one needs first to identify the deterioration phenomena with the aim of understanding it in a context (diagnosis). Later one has to project how this phenomena is expected to act on the machine (prognosis) (Dragomir et al., 2009; Hyers et al., 2006; Mahamad, 2010; Schwabacher and Goebel, 2006). Then, to appropriately model fatigue loads from 10-minute of SCADA data for monitoring, it is reasonable to map the steps in the process. Therefore, modelling from SCADA data to monitoring fatigue

2.3 Monitoring

21

Figure 2.7: Overview of condition monitoring (top) and its relation with diagnostic (bottom left) and prognostics (bottom right) (Tchakoua et al., 2014) loads can be considered as a three-steps process: data pre-processing, diagnosis and prognosis. Pre-processing consists of handling data before they are usable. Diagnosis describes the current loading state with respect to design loads or to other similar turbines. Prognosis goes a step forward and indicates the most probable scenario when current loading state continues in the future.

2.3.5

Monitoring and statistics

Statistics has a central role in monitoring, since most of the work is empirical and data are obtained from observations taken in real conditions. Fig. 2.8 presents a probabilistic representation of prognosis. Sub-figure (a), time to failure (TTF) is depicted together with prediction limits curves, which increases with prediction time as shortterm predictions are expected to have a higher degree of confidence than long-term ones. Then, in sub-figure (b), the quality of probabilistic predictions is related to their accuracy and variance with respect to their predicted time to failure. Finally, in subfigure (c), good predictions are described as centred around TTF (accurate) and with lower variance than acceptable bounds (precise). Since data-driven models transform high-dimensional data sets (measurements) into simpler problems via feature extraction, it is necessary to clearly state how data have been treated before its use, since the differences are expected to affect the results (Aggar-

22


Figure 2.8: Probabilistic representation of deterioration prognosis. Probabilistic degradation (a), failure detection (b) and quality of prediction (c) (Dragomir et al., 2009). pdf: probability density function, perf: performance, pred: prediction, RUL: remaining useful life, TT: time to failure, T: time

wal, 2013; Davies and Gather, 1973; Dragomir et al., 2009). Furthermore, the expansion of data-driven techniques, like artificial intelligence, machine learning, data mining, etc., has led to search for relationships in any data stored in wind turbines (Kusiak et al., 2009a), which creates another challenge: how to simplify multiple indicators to make decisions under uncertainty. Although yet far away to be a reality, these complexity and multi-dimensionality character of monitoring have motivated research that leans towards smart self-diagnosis CMS systems (Tchakoua et al., 2014). Thus, the development of load monitoring systems should include a clear instruction on how to use its outcome in a larger framework.

2.4 Other considerations

2.4

23

Other considerations

There are other two points worth of consideration: first, international standards and guidelines that build the frame of reference for modelling and monitoring wind turbines, and second, data, measurements and quantities typically recorded for those purposes.

2.4.1

Standards and guidelines

Although few in quantity, standards and guidelines for the design of wind turbines serve as a first reference to understand aero-elastic models used to represent wind turbine behaviour. Those from the International Electrotechnical Commission (IEC, 2001, 2005a,b) are the de-facto reference to develop and validate models for wind turbines. Design requirements for wind turbines, explained in IEC, 2005a, defines how to implement aero-elastic models to assess wind turbines loads; it enumerates those standard conditions, and design load cases (DLC), which are expected to occur during the lifetime of wind turbines. Other standards refer to the evaluation of wind power performance, IEC, 2005b, and to the measurements of mechanical loads, IEC, 2001.

Figure 2.9: Example of measurements carried out to investigate a gearbox (Sheng, 2014) Guidelines written by certification bodies clarify the spirit of the standards. Relevant

24


examples are those by Det Norske Veritas (DNV, 2010), by the Federal maritime and hydro-graphic agency (BSH, 2007), by the German institute for the construction sector (DIBT, 2004, in German), and by The association of German engineers (VDI, 2014). Furthermore, standards specifically relevant for monitoring wind turbines are those referred to condition monitoring systems (CMS) and lifetime extension assessment services, since these are matured products commercially available. The Germanischer guideline for the certification of condition monitoring systems for wind turbines (Germanischer Lloyd, 2003) specifies requirements to design and implement CMS in modern wind turbines. CMS are also discussed in literature (Giebhardt, 2007; Sanz-Bobi, 2014). Finally, Det Norske Veritas - Germanischer Lloyd released two documents regarding lifetime extension: a standard DNV-GL, 2016b and a certification process DNV-GL, 2016a. The last two documents are used during the evaluation of the load monitoring system.

2.4.2

Data, measurements and quantities

Even if this thesis is based on the analysis of SCADA data, there are several attributes that can be measured to monitor wind turbines health condition. The aggregated recommendations from literature (Fischer and Coronado, 2015; Giebhardt, 2007; Kusiak, Z. Zhang and A. Verma, 2013; Verbruggen, 2003) included: vibration, oil particles, conductivity and pH value, acceleration, thermography, strains, torque, bending, shear, oscillations, displacements, shaft RPM and position, acoustics, electrical power, generator phase, process parameters, visual inspection, performance monitoring, self diagnosis sensors and meteorology. Sheng, 2014 gives an example of this complex analysis in Fig. 2.9, which shows measurements available in a typical system that monitors the health of gearboxes in wind turbines. Furthermore, Yang et al., 2014 list measurement techniques associated to measured attributes, which is reproduced in Tab. 2.2. It also relates techniques to components and includes their associated cost, the possibility to obtain a degradation diagnosis and whether they have been deployed or not. It is worth noticing that all techniques already deployed to monitor blades, for which the foreseen low-cost load monitoring methodology in this thesis was validated, are found in the range of medium to veryhigh cost; specially the direct measurement of mechanical loads with fibre optic strain gauges, which is deemed to have a very high cost (Yang et al., 2014). Then, the main advantage of load monitoring systems solely from SCADA data appears to be their lower cost with respect to direct measurement in blades, which are expected to provide a higher accuracy.

2.4 Other considerations

Table 2.1: Overview of scientific publications and models used according to specific target values for wind turbine applications. Target values Models References C, MLP, NN, PR, Kusiak and W. Li, 2010a,b; Kusiak et al., Wind & power SVM, VM 2009a,b; S. Li et al., 2001; Zheng and Kusiak, prediction 2009 ARMA, DM, FA, Friedmann et al., 1993; Haddad et al., 2014; MA, NN, OA, Kusiak and Z. Zhang, 2010; Luo et al., 2014; Vibration PF, SI, QCC, Sawalhi et al., 2014; Sheldon et al., 2014; Sheng, WTM 2011; Swartz et al., 2008; Z. Zhang and Kusiak, 2012; Z. Zhang, A. Verma et al., 2012 C, EA, KM, M, Dempsey and Sheng, 2012; Kusiak and Z. NN, SI Zhang, 2011; Kusiak, Song et al., 2009; KuPower siak and A. Verma, 2013; Kusiak and Z. Zhang, performance 2012; Kusiak, Z. Zhang and Xu, 2013; Kusiak and Zheng, 2010; Kusiak et al., 2010; Z. Zhang, Zhou et al., 2014 DM, EC, FL, Gray and Watson, 2010; Kusiak and W. Li, NN, MAS, PF, 2011; Kusiak and A. Verma, 2011a, 2012; SanFailure PR tos, Villa et al., 2012; Schlechtingen, 2012; detection Schlechtingen and Ferreira Santos, 2011; Tian et al., 2011; Zaher et al., 2009 Maintenance DM, ES, FL, NN Cruz Garcia et al., 2006 ¨ 2009; M. Hau, NN, PF, PR, SA Cosack, 2010; Cosack and Kuhn, 2006; Lind, Herráez et al., 2014; Noppe et al., Load 2016; T. Obdam et al., 2010; Perisic et al., 2011; monitoring Smolka and P. W. Cheng, 2013; Smolka, Quappen et al., 2011; Veldkamp, 2008; Vera-Tudela and Kuhn, ¨ 2017 ARMA: auto-regressive moving average, C: clustering, DM: data mining, EA: evolutionary algorithms, EC: ensemble classifiers, ES: expert systems, FA: frequency analysis, FL: fuzzy logic, KM: k-means algorithm, M: Mahalanobis distance, MA: modal analysis, MAS: multi-agent system, MLP: multilayer perceptron, NN: neural networks, OA: optimization algorithm, PF: physics of failure, PR: polynomial regression, QCC: quality control charts, SA: stochastic approach, SI: swarm intelligence, SVM: support vector machines, VM: virtual models, WTM: Wait to maintain

25

26


Table 2.2: Measures (techniques) and some of their features (cost, online, diagnosis, deployed and wind turbine components). Reproduced from Yang et al., 2014 No. Technique Cost Online Diag. Inst. Components 1 Thermocouple Low Y N Y 1,3,6,9,11 2 Oil particle Low Y N Y 1,5 3 Vibration Low Y Y Y 4,5,6,7,8,9,10 4 Ultrasonic test LowY N Y 2,10 Medium 5 Electric test Low Y N Y 6 6 Vibro-acoustic Medium Y N N 2,5,6,7 7 Oil quality MediumN Y N 1,5 High 8 Acoustic High Y N N 2,5,6,7 Torsional vibration Low Y N Y 5,8 9 10 Fibre optic strain Very high Y N Y 2 gauges 11 Thermography Very high Y N N 2,3,5,6,7,8,9,11 12 Shaft torque Very high Y N Y 2,7,8 13 Shock pulse Low Y N N 1,5 Components: 1-Bearings, 2-Blades, 3-Converter, 4-Foundation, 5-Gearbox, 6Generator, 7-Main bearing, 8-Main shaft, 9-Nacelle, 10-Tower, 11-Transformer

Chapter 3 Robustness - The selection of input variables Before constructing and optimizing a prediction model that estimates fatigue loads of wind turbines in wind farms, the first question in this thesis refers to whether a small set of input variables can be identified to make such estimations arises. If the answer is positive, the task would be to find the appropriated method to determine its members. According to Che et al., 2017, a smaller set implies simple and more parsimonious models, ideally without redundant inputs, which should facilitate their understanding and simplify their construction. Searching for input variables, Cosack, 2010 focused on finding those input variables useful to explain the loading phenomena; he explored how disturbances in physical values were reflected (at the same time) in variations of 10-minute statistics of SCADA and fatigue loads, i.e. he looked for excitations that could be detected by correlation analysis. His exploration was carried out in form of experiments, where disturbances were controlled and used as independent variables in aero-elastic simulations. Examples of disturbances, were those affecting wind speed and turbulence intensity. In that search, fatigue loads and variables representing 10-minute statistics of SCADA were the dependent variables, whose variations were recorded to search for cause and effect relationships. Afterwards, T. Obdam et al., 2010 selected attributes based on a heuristic search on all SCADA data available. Later on, Smolka and P. W. Cheng, 2013 assessed the relationship between SCADA and fatigue loads with Spearman coefficients to select the most suitable ones. Noppe et al., 2016 selected those SCADA signals included in a physical equation that estimates thrust loading. In continuation of this previous work, in section 3.1, the input variable selection

28

3. Robustness - The selection of input variables

problem is discussed using set theory, to later continue with the original contribution in section ??, which assesses various filter algorithms that select input variables for the monitoring system.

3.1

Introduction

The introduction to select input variables is explained in two steps: in the first one (see section 3.1.1) the initial search field is limited with set theory; in the second one (see section 3.1.2) the approach used to form the set of predictors is explained.

3.1.1

Universe of potential inputs

The first goal was to limit the search of input variables required to estimate fatigue loads. Thus, let’s consider Y to be the sample set of all target values, that is the recorded set of 10-minute fatigue loads for a given sensor in wind turbines obtained from the specific measurement campaign, like e.g. blade root bending moment in flapwise direction. Also, let’s consider Y to be assumed representative of the entire population, that is of all fatigue loads that would have been measured at that given sensor position during the complete lifetime of the turbine.

(X) Selected input variables (S) 10-minute SCADA statistics

(Y) Fatigue loads

(P) Predictors (M) Measurement campaign (Z) Universe of all explanatory variables

Figure 3.1: Venn diagram displaying the relationship between different set of potential input variables and target values Then, let’s concede the adjective representative to sample Y , when no bias or better said, one that is as small as possible is brought into the inference of fatigue loads when

3.1 Introduction

29

using it (Borovicka et al., 2012). At this point, a measurement campaign (M) including one year of SCADA and load measurements would be deemed suitable to represent site-conditions (MEASNET, 2016). Once the set of targets Y is defined, let Z be the universe of all explanatory variables that have a relationship with the behaviour of Y . Let P be the set of predictors within the universe of all explanatory variables Z (P ⊆ Z) that is ideal to predict the target values Y . Fig. 3.1 displays the relationship between the set of all explanatory variables Z, the set of variables recorded in a measurement campaign M, the ideal set of predictors P and that of fatigue loads Y . It is clear that the whole set of predictors P is not necessarily recorded (or not desired to do so) within a measurement campaign M (P 6⊂ M) or is not available by default. But, the set of available 10-minute statistics of SCADA data S is assumed available (thus S ⊂ M), but it is not a proper subset of the predictors P (S 6⊂ P). Therefore, in order to define an appropriated set of input variables to search for, the focus in this thesis is constrained to a set X, which is a proper subset of the SCADA data S (X ⊂ S), and represent the intersection between S and P (P ∩ S, X ⊆ P). Therefore, it is assumed that the set of predictors within the SCADA data X should serve to estimate Y with higher uncertainty than P. This assumption is reviewed a posteriori and refers to the assessment of prediction quality in chapter 4. At this point it is worth to remark that uncertainty on fatigue loads estimations is also present when calculated with aeroelastic simulations, due to uncertainties in calculations themselves, material properties and environmental conditions (Veldkamp, 2006).

3.1.2

Input variable selection problem

The selection of input variables X is an important step to create data-driven models, as the reduction or removal of redundant variables is expected to improve prediction quality, either quantitatively in terms of accuracy and precision, or qualitatively in terms of simplicity, understandability and robustness (Galelli et al., 2014; Ghosh et al., 2014). Input variable selection (IVS) algorithms can be classified according to their relation with regression models that predicts target variables Y . Thus, there are three types: filter, wrapper and embedded algorithms (May et al., 2011; Mehmood et al., 2012). Fig. 3.2 illustrates these relationships and some of their characteristics. a) Wrapper algorithms also use a mathematical criteria to select predictors, but they take into account the optimization of regression models. Thus, they assure better model performance or at least they do not worsen prediction error. Since regression models are part of their process, they require more computing time, but allow

30


the introduction of optimization techniques like genetic algorithms (Eiben and Smith, 2015). b) Embedded algorithms are an intrinsic part of training regression models, where low weights linked to potential input variables are associated to them having lower impact than others. Thus, they could be removed from the model without a substantial decrease in prediction quality. Cosack, 2010 previously investigated pruning. c) Filter algorithms select relevant predictors based on a mathematical function that assess the relevance of each potential input variable. Since they completely separate the input variable selection process from the optimization of regression models, the selection is transparent and can be used to evaluate the relevance of each individual variable or group of variables (Wei et al., 2015). Smolka and P. W. Cheng, 2013 previously investigated a ranking approach. To select input variables for regression models, priority was given to investigate filter algorithms because: • The selection procedure is carried out regardless of regression models, which allowed to investigate them separately. • The transparency of the selection procedure is rated higher than obtaining better accuracy in predictions, as it served to discuss results with respect to concepts from aero-elastic simulation of wind turbines. • Understanding the physical relationship between predictors selected is relevant to trust the selection made as previous investigations used different input variables (Cosack, 2010; T. Obdam et al., 2010; Smolka and P. W. Cheng, 2013), but obtained equivalent results. Before finishing, it is important to clarify that in input variable selection (IVS) problems (Guyon and Elisseeff, 2003) it is of common use to refer as ’variables’ those attributes directly measured and as ’features’ their transformations into new variables. In this research one may get confused, as a typical SCADA system measure signals but store only their statistics. Since the focus of this investigation is limited to a condensed time domain (10-minute) statistics of SCADA signals are named ’variables’ and, thus, the use of the term ’features’ is reserved to their transformations, e.g. like following a principal component analysis (PCA).

3.1 Introduction

Figure 3.2: Diagrams representing different approaches for input variable selection algorithms (May et al., 2011): (a) wrapper, (b) embedded and (c) filter.

31

32

3.2


On the selection of input variables for a wind turbine load monitoring system

The content of this sub-chapter is identical to the journal article Vera-Tudela and Kuhn, ¨ ¨ 2014: L. Vera-Tudela, M. Kuhn, On the selection of input variables for a wind turbine c load monitoring system, Procedia Technology 15 (2014) 727–737. Elsevier (used with permission). https://doi.org/10.1016/j.protcy.2014.09.045

Abstract The use of statistics from standard SCADA signals as inputs to monitor fatigue loads in wind turbines is a promising alternative to its continuous measurement with extra sensors and is the topic of this contribution. One of the fundamental challenges is the selection of input variables, a problem reflected by the various set of predictors used in previous investigations; a situation that hinders the generalization of the methodology. This work attempts to establish a single optimum set. First, it evaluates the impact of using different methods to select input variables on the accuracy of their predictions and then performs statistical tests on their residuals to establish a selection criterion between sets. It discusses the findings and relates the similarities and differences to the high cross-correlation found in the set of candidates. Finally, it drafts recommendations for the development of a robust system.

Keywords Wind turbine; fatigue; load monitoring; artificial neural networks; machine learning; input variable selection; feature selection; dimensional reduction; stepwise regression; principal component analysis.

3.2.1

Introduction

The mechanical design loads of a wind turbine, calculated to dimensioning its structural components and to define its controller strategy, are commonly not monitored during its operation. They are usually calculated following international standards (IEC, 2005a) and could be considered a design assumption. The validation of design loads is normally limited to prototypes (IEC, 2001) due to the complexity and costs associated to operate and maintain an extra number of sensors over its life-time, which typically spans 20 years for current commercial wind turbines.

3.2 On the selection of input variables for a wind turbine load monitoring system

33

Figure 3.3: Simplified fatigue load monitoring system A load monitoring system should fill the existing gap between the estimated load envelopes during the design phase of the turbine and the costly continuous load measurements, by providing a fatigue load indicator during the operation of the turbine. Therefore, one can expect that optimizations requiring information regarding the structural integrity of the turbine may benefit from such a system, i.e. control strategies, operation and maintenance, upgrade of components, condition monitoring, etc. Previous investigations, see Cosack and Kuhn, ¨ 2009, to Vera-Tudela and Kuhn, ¨ 2013, have demonstrated the potential of using some statistics from standard SCADA signals as input variables to a load monitoring system, which uses artificial neural networks (ANN) as transfer function. A simplified diagram is shown in Fig. 3.3 with the intention to focus the discussion on the first block, the selection of input variables. Output of the block named transfer function are typically damage equivalent loads (DEL), which are used in wind turbine standards (IEC, 2001, 2005a) to describe fatigue loads (also detailed in Cosack, 2010; Cosack and Kuhn, ¨ 2009; Smolka, Quappen et al., 2011) and are estimated based on a rain-flow counting algorithm (H. Sutherland, 1999). While previous researches agree on ANN as a suitable transfer function, where most of them use a feed-forward (FNN) (Cosack and Kuhn, ¨ 2009 to Smolka and P. W. Cheng, 2013 and Vera-Tudela and Kuhn, ¨ 2013) and few of them a recurrent one (RNN) (Hofemann et al., 2011; Vera-Tudela and Kuhn, ¨ 2013), there is less agreement on the statistics of SCADA signals that should be used as predictors, as can be seen in Fig. 3.4, which lists the number of times a predictor has been used in twelve systems investigated (Cosack and Kuhn, ¨ 2009 to Vera-Tudela and Kuhn, ¨ 2013). It is clear that the mean of generator speed, electrical power and pitch angle are common to almost all sets; however, the importance of other variables is not clear. Since not all reports indicated the methods used to select predictors in detail, the author’s judgement is used to present an overview of the challenge. In general, all reports found good regression values for wind turbines in normal operation and free stream conditions; a situation that may lead to the disbelieve that a system based on ANN will work no matter what inputs are provided. This work evaluates the impact of five methods to select input variables for one fatigue load indicator. Due to limits of space, priority is given to list necessary parameters to reproduce each method, a brief explanation is given and further literature is indicated

34


Figure 3.4: Number of times that a predictor has been used in load monitoring systems for wind turbines (from Cosack and Kuhn, ¨ 2009 to Vera-Tudela and Kuhn, ¨ 2013). where necessary.

3.2.2

Methods

Artificial neural networks The transfer function of the load monitoring system, shown in Fig. 3.3, is based on artificial neural networks (ANN), which are mathematical non-linear functions of their inputs (see Dreyfus, 2003) and are commonly used in classification or regression problems, like monitoring fatigue loads in wind turbines (see Cosack and Kuhn, ¨ 2009, to Vera-Tudela and Kuhn, ¨ 2013). As described in Russo et al., 2013, and summarized in eq. 3.1, each neuron weights (ωi j ) and offsets (bi ) its inputs (Xi ) to pass the result (Y j ), transformed with a sigmoid function, to the next neuron in the network. Y j = Σi ωi j Xi + b j

(3.1)

The design of the network is done by presenting a set of correlated input variables and target values to the initial network in order to obtain its optimum coefficients (i, j, ωi j and bi ); this process (training) is repeated over several cycles (epochs) to reduce the


35

error (residuals) between predicted and target values; training is stopped when one of the defined goals is reached, minimizing prediction error over a reasonable amount of time. The feed-forward network (FNN), which is claimed to return an exact representation of a continuous function (Angus, 1991), is one of the most used types and is employed in this work as it has been the preferred one in previous investigations (Cosack and Kuhn, ¨ 2009, to Vera-Tudela and Kuhn, ¨ 2013). The architecture of the FNN used has been optimized using all potential input variables listed in Tab. 3.2; it includes a single hidden layer with 38 neurons and uses sigmoid activation function. To minimize prediction error, the Levenberg-Marquardt back-propagation optimization is used to train the network. To avoid over-fitting the data available is divided in three sub-sets as in Flexer, 1996, 70% of which is for training, 20% is for validation and 10% is hold to test the performance of the network after training, which is stopped if the performance gradient of the training set falls below 1.0*e-6, performance gradient of the validation set remains constant during 4 epochs, or a maximum number of 200 epochs is reached. The input variable selection problem The problem of selecting input variables for data-driven models is the core of this study, its importance is evident by the variety of inputs used in previous investigations (Cosack and Kuhn, ¨ 2009, to Vera-Tudela and Kuhn, ¨ 2013), shown in Fig. 3.4, and it is amply discussed in various publications (Guyon and Elisseeff, 2003, to Williams, 2011). In the following, definitions presented in Guyon and Elisseeff, 2003, are used to differentiate variables; thus, statistics obtained from SCADA signals are referred as potential input variables, from which new features could be generated (variables created out of potential input variables) or predictors can be selected (variables that are used as inputs for the transfer function). The relation between potential input variables, features and predictors is shown in Fig. 3.5. Finally, only predictors are used as input variables for the load monitoring system shown in Fig. 3.3. A total of 117 potential input variables (13 statistics of 9 SCADA signals) have been identified for load monitoring systems of wind turbines, as shown in Fig. 3.4. The numbers in the graph are found adding the quantity of times each has been used as predictor in previous investigations and it is a qualitative indicator of its relevance. Whereas it is clear that mean of generator speed, electrical power and pitch angle can be considered predictors, the selection of reminding ones can be discussed. Since each investigation represents a different research stage, the authors had dissimilar goals and

36


Figure 3.5: Relation between potential input variables (simply variables), features and predictors

used diverse methods, they are summarized here. The first analysis of input variables for a load monitoring system is found in Cosack, 2010, where it is demonstrated with the help of aero-elastic simulations that a set of 84 potential input variables, which correspond to 12 (10-min) statistics of 7 SCADA signals, could be used as predictors to estimate fatigue loads in a wind turbine. After the validation of the method, described also in Cosack and Kuhn, ¨ 2009, the number of predictors is reduced with an embedded pruning algorithm (May et al., 2011) that successively removes the weakest connection in the network until the root mean square error of prediction increases 5% with respect to the original one. After the first optimization the system is left with an average of 18 predictors. Later on, in T. S. Obdam et al., 2009a, to T. Obdam et al., 2010, the same approach is followed. In this case, the set of potential input variables corresponds to a smaller subset, using only 10 potential input variables for blade flapwise bending moment in onshore tests and 6 potential input variables for tower fore-aft bending in an offshore test. The selection of predictors is carried based on a heuristic trial and error approach. More details are provided in Smolka, Quappen et al., 2011, where in an offshore application 40 potential input variables are investigated (5 statistics from 8 SCADA signals); in this case, the selection of predictors follows a wrapper algorithm (May et al., 2011), where the evaluation of each potential input variable is based on its individual impact on the prediction accuracy. In Smolka, Kaufer et al., 2012, the investigation focuses on evaluating two set of potential input variables to represent sea conditions. Finally, in Smolka and P. W. Cheng, 2013, a set of 28 predictors for tower bending and 25 for blade bending are selected out of a set of 64 potential input variables with a filter algorithm (May et al., 2011), which ranks the candidates based on their Spearman coefficient. No feature has been created (see Fig. 3.5) in previous investigations to identify optimum predictors.


37

Set of predictors for analysis In this contribution the analysis focuses on filters and leaves the pruning method out of scope, since the last focuses on the evaluation of the transfer function itself. In order to find the most parsimonious system (Dreyfus, 2003) a filter could be performed before the transfer function is trained, after which a pruning algorithm could be used as verification. A list of the methods used is given in Tab. 3.1. Four sets of predictors are created with distinct filters apart from the first one, the baseline, which uses all variables; the last set is formed by features. The second set is based on the correlation of each candidate with the target value and ranks them with its Spearman coefficient, using as threshold a correlation of 0.5 to select predictors. The third set uses the previous one as initial model to perform a stepwise regression, which tests the significance of each potential input variable to be accepted or rejected in the model. Table 3.1: List of set of predictors and brief description of methods used Set Nr. Set of predictors Description 1 Complete All potential input variables, listed in Tab. 3.2 2 Ranking Simplification based on set nr. 1 & ranking of Spearman coefficients (Corr. > 0.5) 3 Stepwise Regres- Optimization based on set nr. 2, stepwise regression sion & t-test (Significance > 5%) 4 Ranking & Correla- Reduction based in set nr. 2 & crossed correlation tion of single variables (Corr < 0.95) 5 Ranking & Cluster Reduction based on set nr. 2 & clustering (1-Corr > 0.55) 6 Principal Compo- Relevant non-correlated features based on set nr. nents 1 (Explanation > 85%) Based on the third set, highly cross-correlated candidates (i.e. with a coefficient higher than 0.95) are discarded to form the fourth set of predictors; to create the fifth set, the same is performed but the selection uses hierarchical cluster of variables, where correlated cluster with values higher than 0.55 are eliminated. Additionally to the selection of a subset of potential input variables, a reduction of dimensions is investigated by searching new features based on the candidates. A set of principal components (Härdle and Hlavka, 2007), mutually orthogonal (uncorrelated), is created by transforming the original set of variables in to a new coordinate system that better explains the variance of the target values (May et al., 2011). More details about each method is included in section 3.2.4.

38


3.2.3

Data

The data used for this study is described in Vera-Tudela and Kuhn, ¨ 2013, and corresponds to measurements carried out in turbine 01 from offshore wind farm EnBW Baltic 1, which is formed by 21 Siemens 2.3-93 wind turbines. The wind farm is located in the German Baltic See, at 13 Km north of the Darß peninsula. Wind turbine 01 is located in the upper-western side of the layout of wind farm EnBW Baltic 1, as shown in a red circle Fig. 3.6(a), next to the electrical sub-station.

Figure 3.6: (a) Wind turbine 01 in offshore wind farm EnBW Baltic 1; (b) Operating conditions in measured data set To better evaluate the results found in previous investigations, shown in Fig. 3.4, data from westerly wind (free stream) is selected. Thus, 13450 values of 10-min statistics of SCADA signals and mechanical loads with a sample frequency of 50 Hz are available; recording dates are from March 2013 to Oct 2013. Only data from normal power production is used, as shown by the normalized mean power versus mean wind speed, shown in Fig. 3.6(b). For the discussion in this work, damage equivalent loads (DEL) of blade out of plane bending moment is estimated for each set of 10-min statistics of SCADA signals, also described in H. Sutherland, 1999, and used as target values. The maximum set of 56 potential input variables is shown in Tab. 3.2.

3.2.4

Results

To evaluate the impact of the different methodologies and set of predictors six sets are created: a baseline set, which uses all input variables listed in Tab. 3.2, four obtained with different filters and the last one based on an orthogonal transformation. To create


39

Table 3.2: List of all potential input variables Field Quantity Values SCADA signals 7 electrical power, generator speed, wind speed, pitch angle, nacelle fore-aft acceleration, nacelle side-side acceleration, yaw direction maximum, minimum, mean, range, variance, Statistics 8 standard deviation, skewness, kurtosis the first set of predictors, the FNN described in section 3.2.2 is trained using all 56 potential input variables. This network is optimized and its generalization is verified to assure the variance of its error does not affect the results. Fig. 3.7(a) shows the regression plot of the blade out of plane bending moment and the errors (residuals) for the data set described in section 3.2.3. The prediction during test returned a regression value of 0.9 and showed a normal distribution of errors with a mean root square of 0.11% and a standard deviation of 4.88% with respect to the target values. Results are similar to those found in Cosack and Kuhn, ¨ 2009, to Vera-Tudela and Kuhn, ¨ 2013, and serve as baseline.

Figure 3.7: (a) Regression for blade out of plane bending moment using a feedforward network trained with all potential input variables; (b) Error per value in the data set for the same network The second set of predictors is based on individual correlation and uses a ranking of Spearman coefficients, which are calculated for each potential input variable with respect to the blade out of plane bending moment. In Fig. 3.8, a partial ranking of the most correlated variables is shown in descendent order. In this case, an arbitrary threshold value of 0.5 is used and the first 23 candidates are selected as predictors; the analysis of the predictions, shown in Tab. 3.3, does not differ from the result found in

40


the baseline, shown in Fig. 3.7. Additionally, an evaluation of the residuals by means of a t-test, included also in Tab. 3.3, does not reject the null hypothesis that assumes both come from a normal distribution with the same mean error. The third set of predictors uses as initial model the outcome of the previous one and focuses on the evaluation of potential sets as a whole. A stepwise regression algorithm (backwards and forward) evaluates each potential input variable with respect to the initial model; the decision to accept or reject each variable is based on its impact on the regression of the entire set, its t-statistics and its probability of significance.

Figure 3.8: Partial ranking of potential input variables based on their Spearman coefficient, shown in descendent order The resulting set includes more variables and returns 28 predictors. Its results (regression, mean and standard deviation of error) do not significantly differ to previous cases, with exception of a maximum error of 92%. A plot of errors (not included) similar to Fig. 3.7(b) shows it is a single out of two outliers, apart from which the response resemble that of previous sets with maximum around 40%. However, when the residuals of baseline and stepwise regression cases are evaluated with a t-test, the null hypothesis, which assumes that the errors come from a normal distribution with same mean error, is rejected. The evaluation of cross-correlation coefficients between potential input variables show that they are highly correlated and some of them can be considered redundant, a


41

Figure 3.9: Symbolic representation of cross-correlation between predictors, shown using set Ranking

fact that is explained by the selection of several statistics as candidates. This correlation explains the variability of predictors used in previous investigations, shown in Fig. 3.4, and the unexpected acceptance of variables that were previously discarded in set ranking. In Fig. 3.9, a symbolic representation of the cross-correlation between predictors in set ranking is included. The fourth set of predictors is based on the previous one, eliminating those redundant candidates with higher cross-correlation factor than 0.95, the darker areas of Fig. 3.9. In the case of variance and standard deviation, only the last one is kept in all cases. To decide which candidates are kept in other cases, priority is given to those with a higher Spearman coefficient as ranked in Fig. 3.8. With this simplification, the set is further reduced to only 9 predictors, all of them being highly correlated to the target value and low cross-correlated with the rest. The evaluation of the network trained with this set, included in Tab. 3.3, returns reasonable results: a small reduction of the regression with similar values of mean, standard deviation and maximum error. The statistical evaluation of the residuals also rejects the null hypothesis. The fifth set of predictors is also based on the third one, but unlike the fourth one, it analyses the correlation of clusters instead of single variables. In Fig. 3.10, a dendrogram is used to represent the hierarchical connection of candidates in clusters, where variables shown correspond to set Ranking after variances are removed; values in the horizontal axis indicate the (1-Corr) distance, where ‘Corr’ is the correlation between clusters, thus highly correlated clusters are connected at smaller values. An arbitrary minimum distance of (1-Corr) > 0.55 is used to select predictors and define those redundant in the system. Using this method, the fifth set includes 17 predictors, which are highly

42


Figure 3.10: Dendrogram of predictors from set Ranking (variances are excluded) correlated to the target value and low correlated to the rest. The evaluation of the predictions show good results, being the null hypothesis for the residuals accepted in this case.

Figure 3.11: (a) Scree plot of the 10 principal components; (b) Pareto of variance explained by the 10 principal components Finally, the sixth set of predictors is formed by features created with a principal component analysis, described in Bishop, 1995; Dreyfus, 2003. The original coordinates (candidates) are transformed into a new one (principal components), where the new ones


43

are mutually orthogonal (i.e. uncorrelated) and expected to maximize the explanation of variance in the target value. These new features are ranked by the magnitude of the eigenvalues of the transformation matrix and are an indicator of the variance they can explain. A scree plot shows the eigenvalues of the first 10 principal components in Fig. 3.11(a), while a Pareto in Fig. 3.11(b) indicates the individual and cumulative explanation of variance by the first 10 principal components. There is no firm method to define how many principal components should be included as predictors; however, in this work a compromise between large variation in successive eigenvalues (shown in Fig. 3.11(a)) and a low increase in explanation of variance (shown in Fig. 3.11(b)) is found using 7 principal components, which explain 85% of the target value variance. Results, shown in Tab. 3.3, are comparable with the results of the previous sets, being the null hypothesis accepted in this case. A table with all predictors used in each set studied is included in Appendix A. Table 3.3: Evaluation of neural networks with various set of predictors Set Set of Nr. Perc. Reg. NumberMean Std. Max h p Nr. predic- of of R2 [-] of erdev. er(ttest) (ttest) [-] tors pred. total epochs ror [%] ror [-] [-] [-] [-] [%] [-] [%] [%] 1 Complete 56 100.0 0.900 19 0.11 4.88 30.19 0 1.0E+0 2 Ranking 23 41.1 0.893 18 0.02 5.06 49.70 0 2.2E-1 3 Stepwise 28 50.0 0.898 18 0.22 4.94 92.14 1 2.2E-6 Regression 16.1 0.885 13 0.09 5.24 57.34 1 4 Ranking 09 4.8E-3 & Corr. 5 Ranking 17 30.4 0.891 13 0.07 5.10 32.15 0 6.5E-1 & Clustering 6 Principal 07 12.5 0.902 15 0.02 7.29 46.01 0 5.8E-1 Components

3.2.5

Conclusions

Four filters and one dimensional reduction algorithm were defined to select an optimal subset of predictors to monitor the damage equivalent load for blade out of plane bending moment. Data from a wind turbine in normal operation and free stream

44


conditions were used to evaluate the prediction using a feed-forward neural network. The results reproduced the high variability of predictors used in previous publications and pointed at the high cross-correlation between potential input variables as main reason. Statistical test of residuals of predictions rejected the null hypothesis in those cases where cross-correlation was expected to affect the selection process. The most parsimonious system, i.e. with the smallest number of predictors, was achieved with the transformation of the candidates into mutually uncorrelated predictors (nr. 6), with the creation of features, but at the cost of losing expertise knowledge about the relation between predictors and target values. The method that optimizes the set of predictor by discarding those highly cross-correlated potential variables (nr. 4) resulted as a fair compromise between number of predictors and available expertise knowledge. To further develop the load monitoring system for its use in more complex scenarios (wake flow, transients, operation and faults, etc.) is recommended to evaluate the set of potential input variables with more robust coefficients, in special those from information theory (entropy, mutual information, etc.), which can provide more information about the relation between candidates and should allow the modeller to better optimize the selection of predictors. Acknowledgements This research is funded by the German Federal Ministry for the Environment, Nature Conservation and Nuclear Safety (BMU) as part of the research project “Control of offshore wind farms by local wind power prediction as well as by power and load monitoring” (0325215A). Appendix A. Set of predictors Tab. 3.4, describes the predictors selected in each set mentioned in Tab. 3.1.


45

Table 3.4: List of set of predictors and brief description of methods used Set Set of predictors Nr. of pre- Name of predictors Nr. dictors 1 Complete 56 See combination of signals and statistics in Tab. 3.2 Wind speed (mean, min, max, range, std, 2 Ranking 23 var), power (mean, min, max), generator rotation (mean, min, max), nacelle side-side acc. (max, range, std, var), nacelle foreaft acc. (max, range, std, var), pitch angle (mean, min, max) 3 Stepwise Re- 28 Wind speed (mean, skew, kur), power gression (mean, min, range, std, skew, kur), generator rotation (min, std, kur), nacelle side-side acc. (mean, std, skew, kur), nacelle fore-aft acc. (mean, range, std, skew), pitch angle (mean, range, std, kur, skew), yaw direction (max, std, kur) 4 Ranking & Corr. 09 Wind speed (mean, std), power (mean), generator rotation (mean), nacelle side-side acc. (std), nacelle fore-aft acc. (std), pitch angle (mean, min, max) 5 Ranking & Clus- 17 Wind speed (mean, min, max, range, std), ter power (mean, max), generator rotation (mean, min), nacelle side-side acc. (range, std), nacelle forea-aft acc. (max, range, std), pitch angle (mean, min, max) 6 Principal Com- 07 First seven principal components ponents

46


Chapter 4 Quality control - The influence of wind farm conditions After selecting the input variables for the prediction model from the 10-minute statistics of SCADA data, the next question to answer is how to measure the success of regression models for fatigue loads. This is of importance because it depends on the objective that needs to be accomplished with the output, in other words, it is subjective to the goal attained in monitoring wind turbines. Therefore, it is relevant to explicitly differentiate between model assessment and that of the usability of estimations. The expected deployment of the methodology needs to be taken into account during its development (Sanz-Bobi, 2014). In previous investigations, regression models were mostly evaluated using coefficient of regressions R2 , the means and standard deviations of prediction errors. Values of R2 were used to indicate how well a model explained the variance of the targets with that one of the predictors, while mean and variance indicated model accuracy and precision (Rawlings et al., 1998). Furthermore, according to Cosack, 2010; T. Obdam et al., 2010; Smolka and P. W. Cheng, 2013, in wake conditions the impact of physical disturbances were found higher and values of R2 were found lower than in free stream conditions, which led them to indicate that models were unsatisfactory. Therefore, in the following section 4.1, definitions and metrics from process quality are proposed to evaluate regression models for non-linear relationships taking into account their expected use. Then, section 4.2 includes an original scientific contribution, where the impact of wind farm flow conditions is assessed during the construction of models.

48

4. Quality control - The influence of wind farm conditions

4.1

Introduction

In section 4.1.1 concepts from process quality are proposed to assess prediction quality; later, in section 4.1.2, the impact of wind conditions on the quality of prediction is discussed and the previously introduced metrics are employed.

4.1.1

Definition of quality

Once monitoring lifetime based on fatigue load estimation is considered a process, it is necessary to select the measurable attribute to control process quality. Since the goal is to estimate fatigue loads, it follows that its quality should be associated to its error of prediction (calculated as: error = prediction − measurement), where a positive error means over-prediction and a negative one under-prediction. Therefore, the resulting set of prediction errors is represented as a distribution. It is the analysis of this distribution that leads to assess the quality of the prediction process. According to Montgomery (Montgomery, 2009), there are two different ways to examine the quality of a process when looking at the distribution of its governing attribute. There is an end-user perspective, where quality refers to its fitness for use, which is an intuitive but unhelpful definition as there is no end-user requirements to compare with at this point in the analysis. But, there is also an internal-view to quality, where quality is defined as inversely proportional to the variability of the attribute of interest. In other words, a methodology that return an average of prediction errors centred on zero with the lowest possible variance of prediction errors is deemed of having the best possible quality.

Figure 4.1: Probability density function of a process within lower and upper specification limits, LSL and USL respectively (Montgomery, 2009) Then, following the ideas from Tjahjono et al., 2010 and from Montgomery (Montgomery, 2009) to control quality, the process quality can be statistically characterised

4.1 Introduction

49

defining that 99.73% of its predictions should be within the boundaries of a lower and an upper specification level (LSL and USL respectively). In case LSL and USL are not given as in this case, then a multiple (three, 3) of the standard distribution σ (then 3σ ) is used to set those boundaries. Fig. 4.1 presents the probability density function of a continuous normally distributed attribute together with its limits. This self-defined limits are named natural tolerance limits (NT L). They can be used in the absence of product requirements as benchmark for its current state and to evaluate any future improvement. This preliminary definition of prediction quality can be used to assess the usability of the model, as the calculation of fatigue loads from aero-elastic simulations of wind turbines can also be described as a distribution around its true value (Veldkamp, 2006). Defining a more restrictive or permissive definition (1σ to 5σ ) would be a choice available for designers after the system is implemented.

4.1.2

The impact of wind farm flow conditions

Based on the results from Cosack, 2010; T. Obdam et al., 2010; Smolka and P. W. Cheng, 2013, fatigue load predictions were understood unsuitable for their deployment in wind farm flow conditions; however, after the previous clarifications and the separation of evaluating model predictions and model usability, this should be re-assessed in relation with its usability for end-of-life decisions, like e.g. re-powering or lifetime extension. In lights of the standard for lifetime extension of wind turbines DNVGL-ST-0262 (DNV-GL, 2016b), the use of a load monitoring system is foreseen as to pre-select those turbines better suited for consideration within a mode detailed assessment according to the standard DNV-GL, 2016b. Therefore, the problem of predicting fatigue loads in wind farm flow conditions can be re-phrased as the need to quantify the impact that wake flow has on the natural tolerance limits (NT L) for the process. A higher bias and variance is expected, so an assessment is done to quantify whether the difference is significant or not. The suitability of fatigue load predictions from 10-minute SCADA data depends on the usability of its outcome for a given application. Therefore, at this stage, the higher uncertainty in estimations can be managed by using its outcome as a filter to better allocate resources before commitments are made to invest in a more detailed lifetime assessment approach according to standards (DNV-GL, 2016b).

50

4.2


Analysing wind turbine fatigue load prediction: The impact of wind farm flow conditions

The content of this sub-chapter is identical to the journal article Vera-Tudela and Kuhn, ¨ ¨ 2017: L. Vera-Tudela, M. Kuhn, Analysing wind turbine fatigue load prediction: The impact of wind farm flow conditions, Renewable Energy 107 (2017) 352-360. Elsevier c

(used with permission). http://dx.doi.org/10.1016/j.renene.2017.01.065

Abstract Lifetime evaluation with fatigue loads is commonly used in the design phase of wind turbines, but rarely during operation due to the cost of extra measurements. Fatigue load prediction with neural networks, using existing SCADA signals, is a potential cost-effective alternative to continuously monitor lifetime consumption. However, although assessments for turbines in wind farm flow have been pointed out as deficient, the evaluations were limited to single cases and the implication for the design of a monitoring system was not discussed. Hence, we proposed metrics to evaluate prediction quality and, using one year of measurements at two wind turbines, we evaluated predictions in six different flow conditions. The quality of fatigue load predictions were evaluated for bending moments of two blades, in edgewise and flapwise directions. Results, based on 48 analyses, demonstrated that prediction quality varies marginally with varying flow conditions. Predictions were accurate in all cases and had an average error below 1.5 %, but their precision slightly deteriorated in wake flow conditions. In general, results demonstrated that a reasonable monitoring system can be based on a neural network model without the need to distinguish between inflow conditions.

Keywords wind turbine; wake effects; fatigue damage; lifetime extension; condition monitoring; neural networks.

4.2.1

Introduction

As the wind energy field matures, ageing wind turbines become an area of interest to reduce the cost of energy. In this context, models for lifetime estimation are relevant to assess design, operation and maintenance, lifetime extension and upgrade strategies.

4.2 Analysing wind turbine fatigue load prediction: The impact of wind farm flow conditions

51

Welte and K. Wang, 2014, classify the different approaches into four categories: physical, stochastic, data-driven and combined. Out of them, a physical model is usually applied during the design phase of wind turbines according to IEC 61400-1 (IEC, 2005a). In this paper we will present a data-driven model for its potential use during the operational phase. Veldkamp, 2008, shows how the uncertainty in fatigue lifetime calculation from physical models dominate the selection of safety factors, which are chosen to achieve a conservative design (Butterfield, 1997). Therefore, the continuous monitoring of fatigue loads, during the time wind turbines are in service, might provide an up-to-date review of lifetime consumption with respect to design specifications. Unfortunately, the physical models are restricted to manufacturers due to the high level of design details they contain, while the required continuous measurement of mechanical loads is a costly affair. Therefore, the development of a data-driven model for fatigue lifetime estimation during operation represents a logical step. Kusiak and his collaborators have extensively reported data-driven models as suitable for various monitoring tasks in wind turbines. Examples given are short-term wind power prediction (Kusiak and W. Li, 2010a), power optimization (Kusiak et al., 2010), vibration control (Kusiak and Z. Zhang, 2012) and the prediction, analysis and diagnosis of faults (Kusiak and W. Li, 2011; Kusiak and A. Verma, 2012). In this context, Cosack and Kuhn, ¨ 2009 proposed the prediction of wind turbine fatigue loads from 10-minute statistics of available operational SCADA signals. This approach is based on neural networks and does not require extra sensors. T. Obdam et al., 2010, evaluated its deployment in wind farms combined with measurements in a single turbine. Smolka and P. W. Cheng, 2013, investigated its application on offshore wind turbines and its implications on measurements. Nowadays, as wind energy is produced by larger wind farms, quantifying the effects that wakes cause in downstream turbines, i.e. a reduction in power production and an increase of mechanical loads, reveals to be an important step to improve their performance (Barthelmie et al., 2009). Unfortunately, fatigue loads predictions in wake conditions have been previously discussed and pointed out as unsatisfactory (Cosack, 2010; T. Obdam et al., 2010; Smolka, Quappen et al., 2011). However, there is neither a definition of prediction quality nor an assessment of how prediction quality varies in different wind farm flow conditions. We considered both of relevance to assess the current state and to measure improvements to this process. To do so, in our applied research we propose to use the definition of quality in a

52


process given by Montgomery, 2009, in which the quality of the fatigue load predictions process can be described as inversely proportional to its variability and can be related to a lower (LSL) and upper specification limit (USL), exemplified in Fig. 4.2. The challenge is then how to select such specification limits. Therefore, in order to help determine justifiable limits, we quantify the variability of the fatigue load predictions process.

7target , < low 7bias , < high

LSL

Target

USL

Figure 4.2: Probability density function of two predictor models with unequal bias and variance, modified from Montgomery, 2009.

In this paper we also present the impact that wind farm flow conditions have on the quality of fatigue loads predictions, which are made with neural network models and 10-minute statistics of operational SCADA signals. We estimate prediction capability and present its natural tolerance limits as reference for the future design of a load monitoring system. Our ideal envisioned model will have no bias (on target) and a low variance. To focus the analysis and report only around the influence of wind conditions in prediction quality, we followed the approach carried in previous publications by other authors (Cosack, 2010; Cosack and Kuhn, ¨ 2009; T. Obdam et al., 2010; Smolka and P. W. Cheng, 2013; Smolka, Quappen et al., 2011). Thus, we limited ourselves to the study of neural networks models and analysed data from each turbine independently. The study of alternative data-mining algorithms, the definition of minimum length of measurements required for system development, the evaluation in distinct wind turbines and their impact on the bias-variance of prediction, are topics of relevance under current investigation that will be reported in following articles.


4.2.2

53

Data description and methodology

Data from the research project Baltic I (Kuhn ¨ et al., 2014) were obtained through a one year measurement campaign (from March 2013 until March 2014), performed in wind turbines B01 and B08 at the offshore wind farm EnBW Baltic 1 (Fig. 4.3), located north of the peninsula of Darß, in the German Baltic Sea. The wind farm is formed by 21 Siemens wind turbines of 2.3 MW and 93 m of hub height. The measurement institute WIND-consult carried out the campaign according to the standard for measurement of mechanical loads in wind turbines (IEC, 2001). Afterwards, measurements were stored at servers of the wind farm operator EnBW. Finally, 10-minute statistics of equivalent operational SCADA signals and fatigue loads (see Tab. 4.1) were calculated from available 50 Hz sampling series (Cosack and Kuhn, ¨ 2009). NORTH

B01

WEST

5% 10% 15% 20% SOUTH

EAST

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

B08

0° 270°

90° 180°

Figure 4.3: Normalised Wind rose (left) and location of wind turbines B01 and B08 in EnBW Baltic 1 layout (right) Since typical SCADA systems can capture several hundreds of parameters, including continuous signals and discrete status codes (Kusiak and A. Verma, 2011b), for clarity, the hereafter mentioned as SCADA data refer only to 10-minute statistics of continuous operational signals included in Tab. 4.1. Only measurements of normal operational conditions (including all wind directions) were used for the analysis; transient events like system start or stop were not included. Furthermore, wind speeds were recorded with a cup anemometer located in the meteorological station on top of the nacelle, which was previously calibrated by Siemens to match the turbine power curve. Consequently, wind speed records were a good reference for the wind speed in front of the turbine

54


(Kuhn ¨ et al., 2014). Mechanical loads were recorded at two blades of each wind turbine, thus four sets of data were available. Table 4.1: Statistics of SCADA signals and loads measured Description Parameter SCADA signals Wind speed (cup anemometer), electrical power, generator speed, pitch angle, yaw direction and nacelle acceleration (fore-aft and side-to-side) 10-minute statis- Mean, maximum, minimum, range, varitics ance, standard deviation, skewness and kurtosis Mechanical Blade root edgewise bending moment and loads blade root flapwise bending moment As incorrect measurements may lead to erroneous statistics, we investigated the measurements with the Hampel identifier (see Davies and Gather, 1973). It consisted on a moving window filter with the median x˜k (k is the number of points in a window) and the median absolute deviation (MAD = mediani (|xi − median j (x j )|) as measures to determine the scatter of values. Values lying outside the range [x˜k − z · MAD, x˜k + z · MAD] were labelled as potential outliers (H. Liu et al., 2004). For this paper values of k = 15 and z = 4.5 were selected to define the moving range. Furthermore, only 10-minute data sets without transient operation were evaluated.

4.2.3

Scenarios

Since measurements included normal operational conditions in all wind directions, the positions of turbines B01 and B08 (see Fig. 4.3) allowed us to investigate the impact of wind farm flow scenarios. To achieve it, we classified the measurements into 12 sectors of 30 degrees according to incoming wind direction. Table 4.2 lists the scenarios and the data used from each wind turbine, including the distance to the closest wind turbine in each case (as a multiple of rotor diameters D). In Bustamante et al., 2015, we reported that also far away neighbour wind turbines (more than the 10 rotor diameters apart recommended in IEC, 2005a) had an impact on the measured mechanical loads, hence we included them to define our scenarios. In total six categories were studied: free stream refers to undisturbed flow, single wake where the flow was disturbed by one neighbouring wind turbine, mixing wake where it was disturbed by more than one, multiple wake where turbines were aligned in a single column (southern direction or 180 deg in Fig. 4.3), platform wake where the


55

incoming flow passed by the electrical substation (north of B01), and all data referred to all measurements available. Fig. 4.3 (right) indicates that south western wind was dominant during our recordings, which is seen in the different number of observations per scenario. Fig. 4.4 displays the relative number of observations collected per scenario investigated. In total 32062 and 29843 observations were available for wind turbine B01 and B08 respectively. Table 4.2: Wind farm flow scenarios and sectors per wind turbine (distance -in rotor diameters- to closest neighbour) Ref. Wind flow B01 B08 (a) Free stream 30◦ , 210◦ , 240◦ , 150◦ , 330◦ 270◦ , 300◦ , 330◦ (b) Single wake 0◦ (6D), 30◦ (13D), 120◦ (12D), 240◦ (11D), 270◦ (9D) (c) Mixing wake 60◦ (11D), 90◦ (9D), 60◦ (10D), 90◦ (9D), 120◦ (10D), 210◦ (16D) ◦ 150 (11D) (d) Multiple 180◦ (6D) 180◦ (6D) wake (e) Platform 0◦ (2D) 300◦ (9D) wake (f) All data all sectors all sectors

4.2.4

Fatigue loads calculation

After the classification, we calculated fatigue from mechanical load measurements ¨ relation, N f ∆Sm = following two rules (H. Sutherland, 1999). The first one was the Wholer K, which relates the number of cycles at failure (N f ) with a constant loading amplitude ¨ (∆S) to the power of the Wholer’s coefficient (m), where K is a variable proportional to material damage at failure. These empirical curves (also known as S − N curves), obtained through experiments, define the expected load cycles materials can withstand before a crack develops and leads to failure. In absence of blade design characteristics (geometry and material), we assumed the power law formulation to describe the S − N curve for the composites, thus we considered the slope of the S − N curve to remain constant and independent of the mean stress. Although this assumption is expected to overestimate loads, we considered it fair

56

4. Quality control - The influence of wind farm conditions 0.6 B01 B08

Relative number of observations [-]

0.5

0.4

0.3

0.2

0.1

0

a

b

c

d

e

Figure 4.4: Relative number of 10-minute observations per scenario and wind turbine

to illustrate that is possible to relate fatigue loads and SCADA data, which is presented in sections 4.2.5 and 4.2.6 (H. J. Sutherland and Mandell, 2004; Vassilopoulos and Keller, 2014). We used a value of m = 10 to estimate the fatigue damage caused by the loading history (Veldkamp, 2006). To monitor lifetime one is interested in the ratio of current damage to the prescribed failure level, which is defined as D = N∆Sm /K, and which is equal to one (D = 1) when the assumed failure level is reached. Also, as wind turbines are exposed to variable loading, we followed Miner’s rule, D = (∑ni=1 Ni ∆Sim )/K (Ragan and Manuel, 2007), which describes how to combine variable amplitudes ∆Si and n cycles with the help of a rainflow cycle-counting algorithm. However, for our analysis we did not have information about the failure level or the design envelope of the wind turbine. Then, in order to monitor fatigue loading, we applied the commonly used summation of variable loading into a single equivalent load range with eq. 4.1 (Cosack, 2010). This single one ∆eq represents a constant range that would cause the same damage as the variable amplitudes over the same number of cycles. In our evaluation we used Nre f = 1.0E + 07 as reference for the number of cycles (Veldkamp, 2006). As our calculations had the same reference number, we combined equivalent ranges from different (k) scenarios with eq. 4.2 to evaluate overall lifetime (∆eq ). In this case, the weighting factor ωi represents the probability of occurrence of each scenario (see Fig. 4.4).


57

n

∆eq = ( ∑ Ni ∆Sim /Nre f ))1/m

(4.1)

i=1 k

m ∆eq = [ ∑ (ωi · ∆Seq,i )]1/m

(4.2)

i=1

Fig. 4.5 shows the distribution (in boxplots) of the normalised ∆eq calculated from measurements of blade root flapwise bending moments, which are binned using wind speeds. Fig. 4.5 presents the direct relationship between wind speeds and loading, as well as the range of variation that other conditions might have on the ∆eq and that will be derived from the selected predictors, presented in Tab. 4.3. In Bustamante et al., 2015,

DEL blade root flapwise bending moment [-]

we provided a detailed description of the loading condition in the wind farm.

0.8

0.6

0.4

0.2

4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 Wind speed [m/s]

Figure 4.5: Normalised ∆eq for the blade root flapwise bending moment (blade 1 of wind turbine B01) over wind speeds

As we focused on the prediction of 10-minute damage equivalent load ranges (∆eq ) for the blade root bending moments, in edgewise and flapwise directions, we refer to them as fatigue load indicators for simplicity. In our investigation we assessed their prediction by means of neural networks fed with statistics of 10-minute SCADA signals; the limitations of lifetime estimation by means of fatigue loads are explained in Cosack, 2010; Veldkamp, 2006, 2008.

58


4.2.5

Input variable selection

In a previous investigation (Vera-Tudela and Kuhn, ¨ 2014) we determined that a filter (Guyon and Elisseeff, 2003) was the best approach to select input variables for a load monitoring system. The approach filters out SCADA data not highly correlated to the load to be monitored as well as redundant data, therefore its name. To assess the strength between each 10-minute statistic of SCADA signals and each fatigue load indicator, we used the Pearson coefficient (Corr) and built a ranking. The coefficient between one potential input (x) and one target (y) is calculated as Corr(x, y) = ∑M i=1 [(xi − p 2 2 x) · (yi − y)]/ (xi − x) (yi − y) , where M is the number of measurements available. To evaluate the statistical accuracy of this relationship we used a bootstrap algorithm (Hastie et al., 2013), which consisted on multiple randomized sampling. With this in mind we extracted 1000 samples from our measurements and calculated the Pearson coefficient. We used the average coefficient to rank potential inputs and a threshold of Corr ≥ 0.5 to determine an initial sub-set of potential inputs. To clarify the procedure we present the boxplot in Fig. 4.6 as example; it displays the sub-set ranking of potential

(10) Nacelle fore-aft acceleration (std)

(9) Wind speed (min)

(8) Electrical power (min)

(7) Wind speed (mean)

(6) Wind speed (max)

(5) Electrical power (mean)

(4) Generator speed (min)

(3) Generator speed (max)

(2) Generator speed (mean)

0.9 0.8 0.7 0.6 (1) Electrical power (max)

Correlation coeff. [-]

input variables to predict the ∆eq of blade root edgewise bending moment.

Figure 4.6: Example ranking of correlation coefficients between statistics of SCADA signals and ∆eq blade root edgewise bending moment (displaying only 10 highest coefficients) In Vera-Tudela and Kuhn, ¨ 2014, we demonstrated that some of the potential input variables were highly correlated, which made one of them irrelevant to improve to the system. For example, if two variables were highly correlated, one was sufficient to


59

achieve the same prediction capability than when both were used. Thus, they could be considered redundant. Therefore, we evaluated their pairwise correlation (xi , x j ) and removed those with a very high correlation using a threshold of Corr ≥ 0.95. The example heatmap of correlations between potential input variables, presented in Fig. 4.7, confirms that some statistics are highly correlated (for example variables (2), (3) and (4), all related to the generator speed (see Fig. 4.6)). To select what potential input variable to keep we used the ranking created in the previous step. This reduced the complexity of the system and facilitated its understanding. Table 4.3 lists predictors selected for both ∆eq of blade root bending moments. We reduced the number of inputs to better understand the relationship between inputs and target loads. But we were aware that our models were constructed to have a strong predictive power based solely on correlations found in observations. Therefore, they were different than those models meant to explain the relationship between variables and are based on experiments (Shmueli, 2010). Our understanding is that edgewise loading is mainly driven by gravitational loading, thus one can fairly assume its amplitude to be constant and relate its variation to the rotation of the rotor, which was represented by the generator speed and the electrical power, as shown in Fig. 4.6. In the case of the flapwise loading, which is dominated by the thrust load, the signals describing wind speed dominated the ranking. 1

(1) (2)

0.9

(3) (4)

0.8

(5) 0.7

(6) (7)

0.6

(8) (9)

0.5

(10)

(9)

(8)

(7)

(6)

(5)

(4)

(3)

(2)

(1)

(10)

Figure 4.7: Example heatmap showing high correlation between the 10 statistics of SCADA signals presented in Fig. 4.6

60


Table 4.3: Predictors selected from available statistics of SCADA signals Target DEL Predictors Blade root bending mo- Wind speed (max & range), electrical power (max), ment in edgewise direction generator speed (mean), nacelle fore-aft acceleration (std) & nacelle side-side acceleration (range) Blade root bending mo- Wind speed (max & range), electrical power (mean), ment in flapwise direction generator speed (min), pitch angle (max & range), nacelle fore-aft acceleration (std) & nacelle side-side acceleration (std)

4.2.6

Regression analysis

The final step consisted on mapping the relationship between statistics of SCADA signals (x) and fatigue loads (y). Previous research have indicated the suitability of neural networks to map the relationship between SCADA data and fatigue loads (Cosack, 2010; Cosack and Kuhn, ¨ 2009; T. Obdam et al., 2010; Smolka and P. W. Cheng, 2013; Smolka, Quappen et al., 2011; Vera-Tudela and Kuhn, ¨ 2014). Thus, we optimized feed-forward neural networks to estimate each ∆eq of blade root bending moments (y), ˆ in edgewise and flapwise directions, for each of the six wind farm flow conditions (Table 4.2) using the selected predictors (Table 4.3). nh

ni

y(x, ˆ w) = f ( ∑ ω j h( ∑ ω ji xi )) j=0

(4.3)

i=0

The models consisted of two-layer feed-forward neural networks (Bishop, 2006) as described by eq. 4.3, where ni represented the number of predictors (ni = 6 for the edgewise direction and ni = 8 for the flapwise direction) and ω ji , ω j the parameters to be optimized. They included a logistic sigmoid activation function h(·) for the nh hidden layer neurons (nh = 30) and a linear function f (·) for those in the output layer. To optimize the network parameters we used the Levenberg-Marquardt backpropagation algorithm (Bishop, 2006), which was set to minimise the mean square error (see eq. 4.6), where M was the number of values. We followed the hold out procedure (Bishop, 2006) to test the system with data it has not seen. Data were randomly divided in three sub-sets: training (70%), validation (20%) and test (10%). Also, to avoid over-fitting the system, a maximum of 200 epochs was set to stop optimizing coefficients with the training set. Results presented correspond to those of the test group. To evaluate networks performance we used the following metrics: coefficient of determination R2 (eq. 4.4), mean absolute percentage error MAPE (eq. 4.5) and the relative mean square error rMSE (eq. 4.6) (Kusiak and Z. Zhang, 2011; Kusiak and A. Verma,


61

2012; Z. Zhang, Zhou et al., 2014). 2 ∑M i=1 (yî − yi ) ´ ∑M i=1 (yi − y)

(4.4)

1 M yî − yi ∑ | yi | · 100% M i=1

(4.5)

1 M yî − yi 2 ∑ ( yi ) · 100% M i=1

(4.6)

R2 = 1 −

MAPE =

rMSE =

The quality of fatigue load predictions can be better described and evaluated if this is expressed in terms of prediction error, as this value is expected to be around the target value and should be minimized. Thus, we used the relative error rei (eq. 4.7), its mean re ¯ and its standard deviation σre , to evaluate the variability of prediction error in each scenario and its relation to quality. Furthermore, to assess the capability of our prediction process we took the statistics of relative errors (re ¯ and σre ) and estimated the natural tolerance limits (NT L) in eq. 4.8, with z = 3. Commonly employed in statistical quality control (Montgomery, 2009), the natural tolerance limits include 99.73% of the predictions within its three-sigma limits. Note that NT L evaluates only the variation of predicted fatigue loads (obtained with eq. 4.3) with respect to the target values, which were defined as those obtained by calculations (obtained with eq. 4.1 and eq. 4.2). It does not account for the variations due to fatigue load calculation theory. yî − yi · 100% yi

(4.7)

NT L = re ¯ ± z · σre %

(4.8)

rei =

4.2.7

Evaluation of predictions

To evaluate the impact of wind farm flow conditions on the prediction of fatigue loads and to ensure its reproducibility, we performed four replications (we used data from two blades at two wind turbines) to predict two fatigue load indicators in the six wind farm flow scenarios. Thus, in total we evaluated 48 neural networks (due to the combination of: two fatigue load indicators, two blades, two wind turbines, six wind flow scenarios) Fig. 4.8 and Fig. 4.9 depict the prediction for ∆eq of blade root bending moment in flapwise direction, when data from all wind flow conditions were used to optimize the

62


DEL blade root flapwise bending moment [-]

0.8 measured predicted

0.6

0.4

0.2

0

200

400

600

800

1000

Figure 4.8: First 1000 ∆eq of blade root bending moments in flapwise direction in scenario all data (blade 1 of wind turbine B01) neural network. Similar patterns were obtained in all cases evaluated for both fatigue load indicators. Fig. 4.8 and Fig. 4.9 illustrate how predictions followed measured loads. 1 0.9

Predicted value [-]

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.1

0.2

0.3

0.4 0.5 0.6 0.7 Measured value [-]

0.8

0.9

1

Figure 4.9: Predicted and measured ∆eq of blade root bending moment in flapwise direction in scenario all data (blade 1 of wind turbine B01) The performance of predictor models were measured using three metrics: coefficient of regression (R2 ), mean absolute percentage error (MAPE), and relative mean square error (rMSE). Table 4.4 shows their averages over four replications of each target fatigue load indicator.


63

Predictor models returned higher R2 for platform wake and free stream than for the other wind conditions (single, mixing and multiple wake), hence their regressions better explain the variance of fatigue load indicators. The similar variation of MAPE reveals a higher accuracy of predictions. The rMSE of predictions for ∆eq of blade root bending moment in edgewise direction are very close to zero, which demonstrates a high accuracy and small variance; the rMSE of predictions for ∆eq of blade root bending moment in flapwise direction are higher but still relatively low. Table 4.4: Average performance of predictions for ∆eq of blade root bending moments Reference scenario Edgewise Flapwise Wind farm flow R2 MAPE rMSE R2 MAPE rMSE (Reference) [-] [-] [-] [-] [-] [-] (a) Free stream 0.936 1.141 0.038 0.913 8.468 1.38 (b) Single wake 0.945 1.569 0.049 0.868 10.552 1.974 (c) Mixing wake 0.731 2.625 0.074 0.905 7.920 1.146 (d) Multiple wake 0.890 2.159 0.125 0.901 9.738 1.650 (e) Platform wake 0.986 0.863 0.016 0.957 6.770 0.939 (f) All data 0.817 1.911 0.042 0.899 8.963 1.504 Results in Tab. 4.4 show higher R2 , lower MAPE and rMSE of predictions for ∆eq of blade root bending moment in edgewise direction, which demonstrates a higher accuracy and precision than predictions made for ∆eq of blade root bending moment in flapwise direction. Box plots of prediction errors are displayed in Fig. 4.10. The spread of error is symmetric and is larger in wake conditions with the exception of platform wake. The spread is around six times larger in predictions for ∆eq of blade root bending moment in flapwise direction. To compare the impact of different wind flow conditions on fatigue load prediction, we also estimated the probability density function (pd f ) of errors. Fig. 4.11 presents the pd f for predictions of ∆eq of blade root bending moments in edgewise and flapwise direction respectively. The prediction errors for the edgewise direction have a lower order of magnitude than those for the flapwise direction. Fig. 4.11 demonstrates that predictions for platform wake (scenario e) and free stream (a) conditions have a higher probability to better match measured values than those for single wake (b) and multiple wake (d). The difference between free stream (a) and the remaining wake conditions is not so clear as that one between platform wake (e)

64

4. Quality control - The influence of wind farm conditions 0.10

Edgewise

0.05 0.00

Prediction error [-]

-0.05 -0.10

0.40

Flapwise

0.20 0.00 -0.20 -0.40

a

b

c

d

e

f

Figure 4.10: Prediction error for ∆eq of blade root bending moment in edgewise (top) and flapwise (bottom) directions (blade 1 of wind turbine B08) and the others. Mixing wake (c) is a special case, with lower probability than free stream (a) for the edgewise direction and higher for the flapwise direction. Results obtained when the predictor model was optimized using all data are always intermediate with respect to those from the other cases. To judge the quality of fatigue loads predictions we evaluated the statistics of relative errors. Results in Tab. 4.5 confirm that predictions errors are higher in wake conditions than in free stream. Prediction errors for a simple system, optimized to predict in all wind flow conditions, show intermediate quality. In all the cases, with exception of predictions in single wake for ∆eq of blade root bending moment in edgewise direction, neural networks slightly over-predicted fatigue loads. We evaluated also the prediction of the overall damage equivalent load for one year of measurements. The overall lifetime ∆eq were estimated with a single monitoring system, based on all data (scenario f), and also with the combination of individual predictions for each wind flow condition (scenarios a-e) considering the specific probability of occurrence (see Fig. 4.4). Table 4.6 presents the normalised overall ∆eq for the blade bending moments (in edgewise and flapwise directions) at each of the blades investigated. The overall ∆eq estimated with measurements were used as reference for the normalisation. It is clear that both systems predict fatigue loads with a similar accuracy.


65

40 a b c d e f

Probability density function [-] Flapwise Edgewise

30 20 10 0 -0.06

-0.04

-0.02

0

0.02

0.04

0.06

8 a b c d e f

6 4 2 0 -0.6

-0.4

-0.2

0

0.2

0.4

0.6

Figure 4.11: Probability density function of prediction error for ∆eq of blade root bending moment in edgewise (top) and flapwise (bottom) directions (blade 1 of wind turbine B08)

Finally, it is worth mentioning that target fatigue loads are also affected by the uncertainties incurred during their estimation, which is dominated by those from material fatigue strength, stress concentration, mean wind speed and turbulence intensity (Veldkamp, 2006). Uncertainties associated to mean wind speed and turbulence intensity are around 3% and 6%, respectively (Roeth, 2010; Veldkamp, 2006). Those related to material properties are around 5-10% (Toft et al., 2012). Therefore, the uncertainties in our estimation (Tab. 4.5 and 4.6) can be considered satisfactory.

Table 4.5: Relative error of predictions for ∆eq of blade root bending moments Reference scenario Edgewise Flapwise Wind farm flow re ¯ σre NT L re ¯ σre NT L (Reference) [%] [%] [%] ± [%] [%] [%] [%] ± [%] (a) Free stream 0.04 1.72 0.04 ± 5.17 0.95 11.22 0.95 ± 33.65 (b) Single wake -0.07 2.14 −0.07 ± 6.42 1.47 11.28 1.47 ± 39.85 (c) Mixing wake 0.04 2.14 0.04 ± 6.42 1.50 10.68 1.50 ± 32.04 (d) Multiple wake 0.03 2.22 0.03 ± 6.67 1.40 12.21 1.40 ± 36.65 (e) Platform wake 0.05 1.13 0.05 ± 3.40 0.76 9.13 0.76 ± 27.37 (f) All data 0.07 2.15 0.07 ± 6.45 1.27 10.94 1.27 ± 32.83

66


Table 4.6: Overall damage equivalent loads (∆eq ) normalised with measurements Reference scenario Turbine - blade Wind farm flow B01-b1 B01-b2 B08-b1 B08-b2 (Reference) [%] [%] [%] [%] Edgewise (a-e) Experts 99.97 93.67 99.95 99.98 (f) All data 99.96 93.66 99.88 100.19 Flapwise (a-e) Experts 98.29 98.53 97.54 96.97 (f) All data 98.45 98.74 97.32 97.07

4.2.8

Discussion

From the results we conclude that fatigue loads predicted with neural networks using SCADA data are accurate and on average centred around measured values. However, the precision of prediction varies when the regression analysis is carried out for different wind farm flow conditions. In most of the cases, precision decreases for wake conditions. Fatigue load predictions for the blade root bending moment in edgewise direction were more accurate than those for in flapwise direction. The relative mean square error (rMSE) was very close to zero (see Tab. 4.4). This can be due to its lower variance and its deterministic nature, since edgewise loading is dominated by alternating gravitational loading, which is caused by the weight of the rotating blades. Data driven models depend on data available and the results cannot be directly generalized. Although the networks were optimized using the hold out method to avoid over-fitting, evaluations on new data sets are expected to return higher prediction errors. However, its impact on our conclusions is reduced since we investigated four replications per each fatigue load indicator. A better description of the fatigue behaviour of the blade material, one that includes the effect of the mean stress level for example, should lead to the creation of a better predictor model based on SCADA data. Additionally, whereas it is considered appropriated, another indicator, like fatigue damage (∆m eq ) can be used as target values to train a different predictor model following the same procedure. The use of our current system for the same purpose would increase the errors presented in Tab. 4.6 by an exponential factor equal to m = 10. In the end, the best predictor model achievable will be as good as the data used for its development. Although the number of measurements in each scenario varied significantly when we subdivided the data to explore wind conditions (see Fig. 4.4), we considered the


67

smallest data set (scenario platform wake) sufficient to optimize its corresponding network, as this one was previously simplified following the results of our previous research (Vera-Tudela and Kuhn, ¨ 2014). In this case, the number of points was more than 10 times the number of coefficients. In all cases the Levenberg-Marquardt optimization stopped before we reached the maximum number of epochs defined. In Smolka and P. W. Cheng, 2013 the number of points and the design of measurement campaigns for load monitoring is further discussed. Our results agree with those previously reported by other authors (Cosack, 2010; Cosack and Kuhn, ¨ 2009; T. Obdam et al., 2010; Smolka and P. W. Cheng, 2013; Smolka, Quappen et al., 2011), who evaluated other wind turbine models and reported a lower coefficient of determination (R2 ) in wake flow. We also extended the research scope by subdividing wake flow into special cases: single, mixing, multiple and platform wake. Our results in different wind farm flow conditions were also reported with respect to the quality of fatigue loads predictions. This step is considered necessary to design a load monitoring system. The results indicate that precision of fatigue load predictions is the key parameter to improve the quality of short-term predictions. In our one year evaluation of damage equivalent load (∆eq ) predictions errors cancelled out, since predictions were highly accurate and centred on the true value. Finally, the natural tolerance limits (NTL) can be used as a reference to later specify more strict upper and lower specification limits (USL and LSL) for a wind turbine fatigue load monitoring system.

4.2.9

Conclusion

The quality of fatigue load predictions estimated with models based on neural networks were investigated for different wind farm flow conditions. To ensure reproducibility the analysis was replicated four times, using one year measurements in two blades at two offshore wind turbines. Statistics of SCADA signals highly correlated to damage equivalent load (∆eq ) of blade root bending moments were selected as predictors, eliminating those highly correlated among themselves. Neural networks were optimized to estimate fatigue loads in wind flow corresponding to: free stream, single wake, mixing wake, multiple wake and platform wake. A scenario including all measurements was also considered. In total 48 cases were evaluated. Overall ∆eq were calculated with a system based on all data (one expert) and with the combination of predictions from each wind flow condition (five experts). Results demonstrate that wind farm flow conditions have an impact on the qual-

68


ity of fatigue load predictions, especially for short-term evaluation. Also 10-minute predictions made with neural networks are more accurate for ∆eq blade root bending moment in edgewise direction than for flapwise direction. In this case, average relative errors were found below 0.1% and 1.5% respectively. Fatigue loads predictions for the edgewise direction had a standard deviation of 1.9%, while predictions for the flapwise direction had a standard deviation of 11.2%. The precision decreased for cases with higher spread of fatigue loads, which is the case in wake conditions. The calculation of an overall ∆eq for each wind turbine blade confirmed that longterm ∆eq estimation can be deemed as highly accurate. Despite its simplicity, systems based on a single neural network (one expert), based on data from all wind conditions, returned similar overall predictions than systems combining the results from models optimized for different wind conditions (five experts). For a more specific situation, like for example sectorial ∆eq analysis for performance optimization, results for specific wind conditions can be used as guideline for model selection. Results presented can also be used to define inherent fatigue load prediction capability. They can be regarded as a reference to design a load monitoring system around justifiable specification limits when the natural tolerance limits (NTL) are used as a baseline. For example, for the single model created with data from all wind conditions, we can specify that 99.73% of the predictions fall within the range of 0.07% ± 6.45% away from fatigue load values for the edgewise direction, for flapwise direction this range extends to 1.27% ± 32.83%. Although the results obtained in this applied research were positive and necessary to define a framework for its discussion, there are some design parameters that need to be specified before it can be implemented. In the future, we will report results from research dealing with the minimum quantity of measurements needed for development, how to limit the impact of missing records on prediction bias, prediction quality deterioration when the system is deployed in different wind turbines, and its test on a commercial wind farm. Finally, in its current form, the monitoring system can be envisioned to continuously update historical fatigue loading indicators of wind turbines in a wind farm based solely on their SCADA data. To better comprehend its usability, a prototype system can present results with various types of variables: an ordinal ranking of wind turbines according to their accumulated fatigue loading, an interval of current fatigue load per component, which is to be formed by its predicted value and its upper and lower specification limits, and finally; a categorical component-wise variable, which indicates if the loading is higher-than- or lower-than-expected.


69

Acknowledgement This work was funded by the German Federal Ministry of Economic Affairs and Energy as part of the research projects ”Probabilistic load description, monitoring, and reduction for the next generation of offshore wind turbines (OWEA Loads)”, grant number 0325577B and ”Control of offshore wind farms by local wind power prediction as well as by power and load monitoring (Baltic I)”, grant number 0325215A. We also acknowledge the cooperation with EnBW AG and Siemens Wind Power A/S, which allowed us to access measurement data for the analysis.

70


Chapter 5 Model generalisation - The selection of regression models After defining the quality of models that estimate fatigue loads in chapter 4, the next challenge is to quantify the deterioration of prediction quality, when the defined models are applied to estimate fatigue loads on other wind turbines. Regression models are built on the assumption that data available for their construction (training data) are representative of those used to validate them (test data) and to future data used after their deployment (Borovicka et al., 2012). Furthermore, they are optimized to minimize cost functions that are proportional to their prediction errors with the training data set (Hastie et al., 2013). Cosack, 2010; T. Obdam et al., 2010; Smolka and P. W. Cheng, 2013, indicated that fatigue load predictions were poor when they were made for wind turbines exposed to wake flow conditions. The challenge was then to quantify the deterioration of prediction quality, to map counter-measures to reduce this deterioration and to assess their impact on usability. In section 5.1, the problem of making predictions for other wind turbines in a homogeneous wind farm is discussed in light of the difference between data sets (free stream and wake conditions), to later continue with the original contribution in section 5.2, which investigates deterioration of prediction quality and assesses the impact of using other regression models.

5.1

Introduction

Section 5.1.1 describes the problem of making predictions for data sets with different characteristics using ideas developed for imbalance data sets in classification problems;

72

5. Model generalisation - The selection of regression models

afterwards, section 5.1.2 quantifies the deterioration of prediction quality when predictions are made on different turbines in a homogeneous wind farm (i.e. turbines of the same rating produced by the same manufacturer).

5.1.1

Imbalance data sets

It is assumed that models trained on data sets with given characteristics will be used to make predictions on data sets with equivalent ones (Borovicka et al., 2012). However, when making predictions for similar wind turbines in wind farms (same manufacturer, model and type), wind conditions differ because of different ambien site-conditions and due to the wakes created by wind farm layouts (K. S. Hansen and Larsen, 2005). Smolka and P. W. Cheng, 2013, identified this challenge and referred to it as the problem of making estimations in dissimilar (different) data sets. In the following paragraphs, the problem and approaches to mitigate it are described with ideas taken from the class imbalance problem. According to Haixiang et al., 2017, the problem of learning from class-imbalanced data has gained relevance during last years in data mining. Despite the fact that its theoretical constructions is vastly focused on classification problems (Ertekin et al., 2007; Guo et al., 2008; He and Garcia, 2010), the strategies that address the problem can be used as a frame of reference to improve the predictions on different turbines. Fig. 5.1 presents a typical binary classification problem, where there are two classes with a quality attribute that describes them, which is represented as a continuous distribution in the horizontal axis. It is not possibly to completely identify all elements of each class, since there are data within the margin, which can be attributed to both classes. When more elements of one class exist, a larger bias appears because the prediction error is lower when elements within the margin are associated to the majority class. In case of extreme imbalance, a classifier would identify all elements as one class only. Similarly, in an regression problem, the prediction of values using a model trained with dissimilar data is expected to be biased towards the center of the distribution of the population used to develop it. Following the review from Borovicka et al., 2012, there are three approaches to improve models in presence of imbalance data: • Algorithm level methods suggest the evaluation of alternative models or to perform modifications to the baseline model. In case of the second one, changes are made to independently learn only one class or to penalize models with a cost function associated to the risk of large errors.

5.1 Introduction

73

Figure 5.1: Sketch of a classification problem with imbalanced classes in a data set (Ertekin et al., 2007). • Ensemble methods propose the use of a group of models as forming a committee, which finally estimates the output. The learning process can be done either independently (bagging ensemble) or iteratively (boosting ensemble). It is expected that differences in models will average down the risk of large prediction errors. • Data-level methods recommend a change in the distributions of classes in data sets to balance them before models are developed. This is done either by undersampling the majority class, over-sampling the minority one or by a combination of both. Thus, models are trained in a less-efficient but more generic way. In section 5.2, alternative models as well as bagging and boosting ensembles are investigated to improve the generalisation of the model for making fatigue load predictions in another turbine of a homogeneous wind farm.

5.1.2

Predicting loads for other turbines

From the results in previous investigations (Cosack, 2010; T. Obdam et al., 2010; Smolka and P. W. Cheng, 2013), fatigue load predictions were understood unsuitable for their deployment in wind farm flow conditions. This was deemed after obtaining lower coefficients of regression R2 and higher error of predictions. Cosack, 2010, identified the deterioration from the change in disturbances when changing from free stream to wake flow conditions. T. Obdam et al., 2010, assessed a neural network based model, created with data from a wind turbine in free stream making predictions for another one in wake conditions, but used the latter data set to stop the improvement of the model, which improved prediction but prevented the analysis on independent data. Smolka and P. W. Cheng, 2013, inferred the nature of the problem using fair and unfair tests, which

74


refer to the evaluation with similar (balanced) and dissimilar (imbalanced) data sets respectively. Although this deterioration in quality of fatigue load predictions is comprehensible as wind conditions differ in free stream and wake conditions (K. S. Hansen and Larsen, 2005), its identification is not sufficient to determine whether models are still usable or not. Since metrics used are not connected to product requirements during this research, it is better to quantify the impact that wind farm flow has on the natural tolerance limits (NT L) of the prediction process, in order to assess the generalisation of fatigue loads estimations.

Figure 5.2: Example of a Six Sigma process with the mean of its distribution shifted by ±1.5σ with respect to the distribution centred within the specification levels (Montgomery, 2009) To clarify the idea of prediction bias, with respect to mature products, lets use the example in Fig. 5.2, which shows specification limits in a Six Sigma process. In this example Montgomery, 2009 presents an imaginary ±1.5σ shift in the distribution, which could be associated for example to our case of predicting loads for other turbines, but a shift that occurs within both boundaries lower and upper specification limits (LSL and USL). Although it is far from ideal, the quality of the process presented is within bounds and thus deemed suitable and under control, regardless of its deterioration. From the example one can infer that it is not the variation in prediction quality that matters the most, but its relation to the limits of the process using its outcome. Therefore, next section is limited to quantify the change in natural tolerance limits (NT L) as means to describe the deterioration in prediction quality.

5.2 Monitoring Fatigue Loads in Wind Farms from SCADA Data - Quantifying Quality

5.2

75

Monitoring Fatigue Loads in Wind Farms from SCADA Data - Quantifying Quality

The content of this sub-chapter quantifies the change in fatigue load predictions when they are made for different turbines (of the same type and manufacturer) located in the same wind farm. It also describe the variation in prediction quality obtained when different regression models are assessed.

Abstract The estimation of accumulated fatigue loads on wind turbines is of relevance to decide on their lifetime extension. Their assessment, according to the standard DNVGLST-0262, include fatigue loads calculation from aero-elastic simulations, a review to their monitoring records and a site inspection. It is a complex and expensive task to perform on large fleets for which a filter criteria to prioritize efforts is needed. Thus, fatigue load estimations from 10-minute statistics of SCADA data can help to close that gap. However, prediction quality has not been quantified for estimations made on different turbines due to the lack of measurement data. Therefore, in this paper we report its quantification using one-year measurements from two blades in two turbines at an offshore wind farm. We assessed fatigue loads prediction for blade root bending moments in edgewise and flapwise directions. Furthermore, we compared the impact that seven regression models would have on prediction quality. Results indicate that high prediction accuracy (below 1% error) for fatigue loads in edgewise and flapwise blade bending is achieved with Feed-Forward Neural Networks and K-Nearest Neighbours respectively, but that lower precision in estimations make the approach unsuited to be used as single-factor for decision making. Thus, in its current form, fatigue load estimations are better used as a filter to select turbines of interest from a fleet before resources are committed to a standard lifetime extension assessment.

Keywords wind turbine; fatigue damage; lifetime extension; condition monitoring; data mining; regression analysis.

76


5.2.1

Introduction

As the number of installed wind turbines increases over the years, so does the need to better monitor their deterioration and estimate their lifetime. The first need is observed in the development of optimised operation and condition based maintenance strategies, which are preferred over traditional approaches, typically associated to maintenance at fixed time intervals (Hyers et al., 2006). The second need is observed with the growth of re-powering and lifetime extension services for wind turbines in the last years of their expected operational lifetime (DNV-GL, 2016b). The condition of wind turbines is monitored with the large amount of data recorded either through their supervisory and control data acquisition (SCADA) systems or via purpose-specific sensors in condition monitoring systems (CMS) (K.-S. Wang et al., 2014; Z.-Y. Zhang and K.-S. Wang, 2014). The standard for life extension DNVGL-ST-0262 (DNV-GL, 2016b) includes two elements, its practical part refers to the inspection of monitoring records and to in-site inspections, which are preceded by its analytical part that includes the estimation of accumulated fatigue loads, which are calculated with aero-elastic models and are compared to design conditions (IEC, 2001, 2005a). It is expected that a continuous monitoring of fatigue loads in wind turbine would serve as liaison between continuous monitoring and fatigue load estimations. Gray and Watson, 2010, already demonstrated the relationship between fatigue damage accumulation and failure occurrence in wind turbine gearboxes. But the continuous measurement of mechanical loads is limited due to the cost of extra measurements (Cosack, 2010); thus, estimating fatigue loads with models based on 10-minute statistics of SCADA and neural networks has been previously investigated (Cosack, 2010; T. Obdam et al., 2010; Smolka and P. W. Cheng, 2013). Developing on those ideas, we described how to select a minimum set of predictors from SCADA data and determined the impact of wind farm conditions on prediction quality (Vera-Tudela and Kuhn, ¨ 2014, 2017). Despite progress made, all previous efforts were limited to single wind turbines, which can lead to over-optimistic results, since evaluations on other turbines are expected to lead to larger prediction errors (Borovicka et al., 2012; Guo et al., 2008; Kotsiantis et al., 2006; Smolka and P. W. Cheng, 2013). Therefore, in this paper we estimate fatigue loads for two equivalent turbines (same manufacturer, model and type), which are placed in the offshore wind farm EnBW Baltic 1 and are exposed to different wind farm flow conditions. Additionally, we compare the impact of seven regression models as counter-measure for the reduction in prediction quality (Buitinck et al., 2013; Pedregosa et al., 2011): five single algorithms (Feed-Forward

5.2 Monitoring Fatigue Loads in Wind Farms from SCADA Data - Quantifying Quality

77

Neural Networks (FNN), Linear Regression (LR), Support Vector Machines (SVM), K-Nearest Neighbours (K-NN) and Regression Tree (RT)) and two ensemble algorithms (Extra Tree Bagging Ensemble (ETBE) and Adaptive Boosting Ensemble (ABE)). We constrain our analysis to estimate only damage equivalent loads for blade root bending moments in flapwise and edgewise directions at four blades (two per turbine). We then quantify the reduction in prediction quality and discuss results to better assess the potential of this approach for its use in lifetime extension projects.

5.2.2

Methods

Figure 5.3 describes the approach followed. Data pre-processing refers to the calculation of damage equivalent loads, which is described in section 5.2.3 as well as the transformation of 50 Hz measurement into 10-minute statistics, which is detailed in section 5.2.4. Inputs selection refers to the collection of predictors for the regression analysis, documented in section 5.2.5. Regression analysis included the utilisation and evaluation of various models, explained in sections 5.2.6 and 5.2.7. Test on similar data dealt with the evaluation of regression models in the same turbine, as shown in section 5.2.8. Performance is evaluated when fatigue loads are estimated for other wind turbines in the wind farm, and thus it is presented in section 5.2.9. Data PreProcessing

Inputs Selection

Regression Analysis

Test on similar data

Performance (on dissimilar data)

Figure 5.3: Procedure followed to estimate the performance of fatigue load estimations. One of the limitations of data-driven models is found in the necessary assumption that data available for its construction is representative of future data, i.e. that would occur during wind turbine lifetime (Borovicka et al., 2012). As mentioned in section 5.2.1 (introduction), costly mechanical loads measurements are expected to be carried out in single wind turbines during a short period of time. Thus they do not necessarily include all relevant wind flow conditions (IEC, 2001, 2005a). Smolka and P. W. Cheng, 2013, investigated the requirements for measurement campaigns meant for developing fatigue load monitoring systems for offshore wind turbines. In Seifert et al., 2017, we investigated the minimum number consecutive of measurements needed. Furthermore, we divided the problem into two parts. The first part dealt with the need to quantify the expected reduction in performance (Vera-Tudela and Kuhn, ¨ 2017) when models, developed with values from one data set (i.e. one blade of a given wind

78


turbine), are tested with another data set (i.e another blade or another turbine). The second part consisted on the evaluation of alternative regression models to assess their performance on dissimilar data and thus complement previous investigations, which were limited to neural networks (Cosack, 2010; T. Obdam et al., 2010; Smolka and P. W. Cheng, 2013; Vera-Tudela and Kuhn, ¨ 2014, 2017). Thus in one hand we explored the reduction in quality, while in the other we assessed possible improvements attained when using different regression models. Figure 5.4 is a qualitatively representation of foreseen prediction errors when models are evaluated in similar data sets (continuous line) and dissimilar data sets (discontinuous line). A wellperforming model is expected to return a distribution of predictions centred on the target with a low standard deviation (closer to the continuous line than to the discontinuous line in Figure 5.4), at best within the boundaries of low and high specification limits (LSL and HSL) that define the minimum quality of the prediction process. Thus, we quantified the variation in performance when using other seven regression models, which is further described in section 5.2.6.

7 target ,