PERFORMANCE OPTIMIZATION OF WIND TURBINES

by

Mani Shekhar
An Abstract

Of a thesis submitted in partial fulfillment of the requirements for the M.Tech degree in Thermal Engineering under the esteemed guidance of

Dr. Y.K Mohanty

October 2013

Gandhi Institute of Engineering & Technology

(Affiliated under Biju Patnaik University of Technology, Rourkela) Gunupur-765022 , Dist: rayagada (Odisha)

CERTIFICATE

This is to certify that Mr. MANI SHEKHAR, M.tech student of 4th semester Thermal Engineering of Gandhi Institute of Engineering & Technology, Gunupur has successfully completed the thesis on Performance Optimization of wind Turbine in October 2013.

He has taken keen and genuine interest in research and acquiring material. We found him sincere and hardworking throughout the research & thesis preparation.

Project Guide Dr. Y. K. Mohanty

HOD, Thermal Engg. Dr. P.K. Talapatra

ACKNOWLEDGEMENT

I would like to express my sincere gratitude to my guide Dr. Y.K. Mohanty, for his devotion to this research. He has been the most instrumental person for my academic and research achievements. He provided the motivation, encouragement, guidance and advice.

I am extremely grateful to Mr. Vishal, NIT, Jaipur for providing sampling data, continuous support, valuable suggestions and feedback on my research.

I gratefully acknowledge the co-operation of all the members who have worked with me and provided advice, reviews, and suggestions.

And finally, and most importantly, I would like to express my sincere gratitude to my parents, who solidly supported me in pursuing my M.Tech degree.

ABSTRACT
Improving performance of wind turbines through effective control strategies to reduce the power generation cost is highly desired by the wind industry. The majority of the literature on performance of wind turbines has focused on models derived from principles versed in physics. Physics-based models are usually complex and not accurate due to the fact that wind turbines involve mechanical, electrical, and software components. These components interact with each other and are subjected to variable loads introduced by the wind as well as the rotating elements of the wind turbine. Recent advances in data acquisition systems allow collection of large volumes of wind energy data. Although the prime purpose of data collection is monitoring conditions of wind turbines, the collected data offers a golden opportunity to address most challenging issues of wind turbine systems. In this dissertation, data mining is applied to construct accurate models based on the turbine collected data. To solve the data-driven models, evolutionary computation algorithms are applied. As data-driven based models are nonparametric, the evolutionary computation approach makes an ideal solution tool. Optimizing wind turbines with different objectives is studied to accomplish different research goals. Two research directions of wind turbines performance are pursued, optimizing a wind turbine performance and optimizing a wind farm performance. The goal of single wind turbine optimization is to improve wind turbine efficiency and its life-cycle. The performance optimization of a wind farm is to minimize the total cost of operating a wind farm based on the computed turbine scheduling strategies. The methodology presented in the dissertation is applicable to processes besides wind industry.

LIST OF FIGURES
Figure 1.1 Figure 1.2 Figure 1.3 Figure 1.4 Figure 1.5 Figure 1.6 Figure 3.1 Figure 3.2 Figure 3.3 Figure 3.4 Figure 3.5 Figure 3.6 Figure 3.7 Figure 3.8 Figure 3.9 Figure 3.10 Figure 3.11 Figure 3.12 Figure 3.13 Figure 3.14 Figure 3.15 Figure 3.16 Figure 4.1 Figure 4.2 Figure 4.3 Figure 4.4 Figure 4.5 Figure 4.6 Figure 4.7 Figure 4.8 Figure 4.9 Figure 4.10 Wind turbine details Wind turbine size increases over time Indias wind power density map Indias different state wind power capacity in MW Indias cumulative wind power installation year by year World wind power capacity year by year The power corresponding to the first 300 observed and power model predicted points The predicted and observed value of the first 300 test points Results for the first 300 points used to test the drive-train acceleration model The predicted and observed value of the first 300 test points Test results of the tower acceleration model for first 300 points The predicted and observed value of the tower acceleration for first 300 test points Prediction results from the wind speed prediction model for the first 300 points First 300 points of one-step ahead prediction of the generated power The results of one-step ahead prediction of the drive-train acceleration for first 300 points The results of one-step ahead prediction of the tower acceleration for the first 300 points Convergence of the PSO algorithm for three instances of Case 1 Convergence of the PSO algorithm for three instances of Case 2 Optimized Vs observed blade pitch angle for Case 1 Optimized Vs observed generator torque for Case 1 Optimized Vs observed blade pitch angle for Case 2 Optimized Vs observed generator torque for Case 2 Wind farm boundary and the denition of the wind speed direction Wind turbine wake model An imaginary half cone of wind turbine Turbines affected by the wake of other turbines A wind turbine inside the cone of another turbine Layout solutions for different number of turbines given the wind characteristics in Table 4.1 Initial population and the population at the 100th generation Six layouts of turbines generated by using another initial population size Layout solutions for different numbers of turbines under the wind characteristics in Table 4.2 Wake loss due to the turbines added to the wind farm. 2 3 6 7 9 11 28 28 31 31 33 33 35 36 36 37 41 42 43 43 44 44 47 50
53

55
56 62 64 65 67 69

LIST OF TABLES
Table 1.1 Table 1.2 Table 1.3 Table 3.1 Table 3.2 Table 3.3 Table 3.4 Table 3.5 Table 3.6 Table 3.7 Table 3.8 Table 3.9 Table 3.10 Table 3.11 Table 3.12 Table 3.13 Table 4.1 Table 4.2 Table 4.3 Table 4.4 Indias largest wind power plants Top 10 wind power capacity countries by nameplate Top 10 wind power capacity countries by production Sample dataset Statistics of de-noising drive-train acceleration with two threshold settings Statistics of de-noising tower acceleration with two threshold settings Accuracy of power generation models extracted with six algorithms Parameter list for modeling wind turbine vibration List of parameters before and after parameter selection Accuracy of drive-train acceleration models extracted with six algorithms Selected parameters for modeling tower acceleration Accuracy of the tower acceleration models extracted with six algorithms Accuracy of wind speed prediction models extracted with six algorithms Results of one-step ahead prediction Summary of multi-point optimization of Case 1 Summary of multi-point optimization in Case 2 Wind scenario 1 Wind scenario 2 Optimized vs. Ideal energy production and wake loss for different numbers of turbines under wind scenario 1 Optimized vs. Ideal energy production and wake loss for different numbers of turbines under wind scenario 2 8-9 12 12 24 25 25 27 29 30 31 32 33 35
36

42
44 62 66 68 68

CONTENT
Chapter-1 Introduction
.. 1-3 .. 3-5 .. 5-10 .. 10-13 ... 13 ... 13-14 1.0 Introduction 1.1 Objective 1.2 Wind power in India 1.3 wind power industry 1.4 Objective of thesis 1.5 Scope of thesis

Chapter-2

Review of literature
15-19 20 . 21-22

2.0 Introduction 2.1 Review of literature

2.1.1 Wind Turbine Vibration and Power Generation 2.1.2 Wind farm layout problem 2.2 Conclusion

22

22

Chapter-3 Control of Wind Turbine Power and Vibration with a Data Driven Approach
3.0 Introduction 23 23-26 . 26

3.1 Data description and processing

3.2 Dynamic modeling and prediction for anticipatory control 3.2.1 Modeling power generation 3.2.2 Modeling wind turbine vibration

26-28 28-33 34

3.2.3 Prediction based on data-driven models

3.2.3.1 Modeling wind speed based on the time-series method 34-35 3.2.3.2 One-step ahead prediction 3.3 Multi-objective optimization 3.3.1 Anticipatory control process 3.3.2 Multi-objective optimization model 35-37

. 37 . 37-38 . 38-40

3.4 Case study and computational results 3.4.1 Stopping criteria of the PSO 3.4.2 Case 1 (w1 = 1 and w2 = w3 = 0) 3.4.3 Case 2 (w1 = 0 and w2 = w3 = 0.5) 3.5 Conclusion

41 41-42 42-43 43-44

45

Chapter-4

Design of Wind Farm layout

46-47 47-49

4.0 Introduction

4.1 Problem formulation and methodology 4.2 The wake loss model 4.3 Power curve

50-53 54-56 56-58

4.4 Discretization of wind speed and wind direction 4.5 Evolutionary strategy algorithm 4.5.1 The algorithm 4.5.2 Mutation

... 58-59

.... 59-60 60 60-61

4.5.3 Selection and recombination of parents 4.5.4 Tournament selection 4.6 Computational study 4.6.1 Wind scenario 1 4.6.2 Wind scenario 2 4.6.3 Wake loss 4.7 Conclusion

61 61-62 63-64 65-66 66-69 69-70

Chapter-5

Conclusion
. 71 . 71 . 72

5.0 Introduction 5.1 Scope of future work 5.2 Conclusion

References

.. 73-75

Chapter 1
1.0 Introduction

Introduction

Much of the electrical energy utilized worldwide comes from conventional sources such as fossil fuels (coal, oil and gas), nuclear energy, as well as hydro. In such a traditional power system, the electrical energy is generated by a large power plant and transported over long distances to the user through the transmission and distribution lines. However, electricity generation from conventional sources has some consequences. 1. Reserves for fossil fuels are fast being depleted leading to limited fuel supply. It is estimated that if available reserves continue to be consumed at the current rate, the world may run out of fossil fuel in a couple of decades. Nuclear energy also has limited fuel supply. 2. Fossil fuel can cause serious problems to the environment. In the process of burning fossil fuels in power plants to generate electricity, greenhouse gases such as carbon dioxide (CO2) and nitrous oxide are released to the environment which gives rise to smog, acid rain and other environmental problems. One big consequence of this is the buildup of CO2in the Earths atmosphere which causes global warming. 3. Nuclear and radiation accident with damaging effects on the environment and human life. One recent reminder is the Fukushima Nuclear Accident following the Japan earthquake and tsunami on 11 March 2011. Accidents from nuclear power plants can lead to the release of high amounts of radioactivity into the atmosphere. Radioactive materials may also be released into the ground thereby contaminating agricultural land and domestic water. Deaths from direct exposure as well as future cancer deaths in the surrounding population may also occur. The growing demand for electrical energy for industrial and domestic use, coupled with the limited amount of available fossil fuel reserves and its negative effects on the environment, have made it necessary to seek alternative and renewable energy sources. As a result, Renewable energy sources such as wind, solar, small scale hydro, and ocean energy are becoming increasingly integrated to traditional electric power systems. Amongst the renewable sources, wind technology has emerged as the clear leader in terms of total available installed capacity.

Renewable energy has been vigorously debated and pursued in the past decade. The growing public awareness of the environmental concerns, limited energy supply, and uncertain energy prices has spearheaded this debate.

Figure 1.1: Wind turbine details

Wind energy has experienced a remarkable expansion in the past years. The global cumulative capacity of wind power generation has increased 20 times in a ten year period and is expected to grow even faster in the future. In 2008, the U.S. Department of Energy announced an ambitious goal for wind energy to produce 20% of electricity by 2030. However, the challenge for wind energy generation is to be cost competitive with the energy produced from the conventional sources. The cost of wind generated power is largely derived from the turbine installation cost and the operation and maintenance (O&M) cost. Therefore, reducing the cost of installation and O&M makes wind power more cost-attractive. In this dissertation, performance of wind turbines is studied with the aim of reducing the O&M cost and thus lowering the cost of wind power generation.

Figure 1.2: Wind turbine size increases over time To improve wind turbine performance, models of wind turbines are investigated. Although various approaches have been studied to model conventional power systems, most of them are not applicable to modeling wind turbine systems. The main reason is that the static and dynamic characteristics of large-scale wind turbines differ from the conventional power plant systems. Therefore, novel theories for modeling wind turbine systems are needed.

1.1 Objective
Three research goals are set for this dissertation. The first goal is to develop nonlinear and non-parametric models accurately capturing performance of wind turbines. The wind turbine performance is expressed by the amount of the generated power and the level of turbine vibration. Since wind turbine vibration impact performance and life-cycle of wind turbine components, such as a gearbox, it is considered as an indicator of mechanical component usage. The second goal is to optimize wind turbine performance based on the developed models. Unlike the

previous two goals that concentrate on a single turbine, the third goal is to improve performance of a wind farm.
1) Control of Wind Turbine Power and Vibration with a Data Driven Approach:

In this the optimization of wind turbine performance will be based on an anticipatory control model. Both maximizing wind power generation and minimizing the drive-train and tower vibration are considered in the optimization. The drive-train and tower vibration is represented by the vibratory acceleration of drive-train and tower measured with accelerometers. 2) Adaptive Control of a Wind Turbine with Data Mining and Swarm Intelligence: An adaptive approach to wind turbine control is presented. It is designed to achieve a balance between power optimization and smooth drive train control in response to the changes in wind speed and electricity demand. The smoothing of the drive train is accomplished by minimizing the torque ramp rate rather than controlling the rotor speed presented in (Kusiak et al. 2009a). The former reduces extreme loads, which translates into a lower maintenance and operation cost. To model the turbine, a data-driven approach is introduced. To realize the adaptive turbine control, estimates of future electricity demand and wind power to be produced at the same time are desired. A time-series model extracted by a datadriven approach predicts the future wind power. A simulation model is used to generate the future demand due to the lack of demand data in this research. 3) Optimization of Wind Turbine Power and Its Variability with an Artificial Immune Network Algorithm: A new power optimization constraint constructed based on control chart theory is introduced to the optimization of wind turbine power generation process. Due to variations in wind speed, wind farms exhibit different characteristics from conventional power plants. Two important goals are to maximize the generated power and to minimize its variability in order to meet the grid requirements. The variability of the generated wind power has become an essential issue in the management of wind-driven energy production and other is power ramp rate (PRR) expresses the change of power for a given time interval, e.g., a minute or an hour. The prediction of power ramp rate results at 10-min to 60-min intervals was demonstrated in their research. Control of power ramp rates is beneficial to the power dispatching and grid

4) Designing a Wind Farm layout: It is a model to wind turbine place in order to capture maximum wind based upon wind distribution. It considers wake loss which can be calculated based of turbine location and wind distribution. 5) Scheduling a Wind Farm layout: In is a model for scheduling wind turbines of a wind farm. A particle swarm optimization algorithm with a small world network structure is introduced for solving this model. The solution generated by the algorithm defines operational status of wind turbines for a scheduling horizon selected by a decision maker. Different operational scenarios are constructed based on time series data of electricity price, grid demand, and wind speed. The computational results provide insights into management of wind farm. 6) Scheduling A Wind Farm With Data Driven Stochastic Optimization: It is a model to schedule wind turbines in a wind farm. A scheduling model determines activation/deactivation status of individual wind turbines to cope with the varying electricity prices, wind speed, and grid demand. Due to the complexity of the scheduling model, an enhanced particle swarm algorithm is introduced to compute the optimal solution. More advanced models can accurately depict the power generation process are needed to develop a more applicable wind farm scheduling model. Therefore, data mining algorithms are employed to develop wind power generation models for the wind farm scheduling problem.

1.2 Wind energy in India

The development of wind power in India began in the 1999s, and has significantly increased in the last few years. Although a relative newcomer to the wind industry compared with Denmark or the United States, India has the fifth largest installed wind power capacity in the world. In 2009-10 India's growth rate was highest among the other top four countries. As of 31 Jan 2013 the installed capacity of wind power in India was 19564.95 MW. It is estimated that 6,000 MW of additional wind power capacity will be installed in India by 2014. Wind power accounts for 8.5% of India's total installed power capacity, and it generates 1.6% of the country's power. The worldwide installed capacity of wind power reached 197 GW by the end of 2010. China (44,733 MW), US (40,180 MW), Germany (27,215 MW) and Spain (20,676 MW) are ahead of India in fifth position.[11] The short gestation periods for installing wind turbines, and the increasing reliability and

performance of wind energy machines has made wind power a favoured choice for capacity addition in India.

Figure 1.3: Indias wind power density map

Suzlon, an Indian-owned company, emerged on the global scene in the past decade, and by 2006 had captured almost 7.7 percent of market share in global wind turbine sales. Suzlon is currently the leading manufacturer of wind turbines for the Indian market, holding some 43 percent of market share in India. Suzlons success has made India the developing country leader in advanced wind turbine technology.

Figure 1.4: Indias different state wind power capacity in MW

Power Plant
Vankusawade Wind Park Cape Comorin Kayathar Subhash Ramakkalmedu Muppandal Wind Gudimangalam

Producer

Location

State

Total Capacity (MWe)

259

Suzlon Energy Ltd. Aban Loyd Chiles Offshore Ltd. Subhash Ltd. Subhash Ltd. Muppandal Wind Farm Gudimangalam Wind Farm Wescare (India) Ltd. Danida India Ltd. Mohan Breweries & Distilleries Ltd. MP Windfarms Ltd.

Satara District.

Maharashtra

Kanyakumari Kayathar Ramakkalmedu Muppandal Gudimangalam

Tamil Nadu Tamil Nadu Kerala Tamil Nadu Tamil Nadu Andhra Pradesh Gujarat Tamil Nadu Madhya Pradesh Karnataka

33 30 25 22 21

Puthlur RCI Lamda Danida Chennai Mohan

Puthlur Lamba Chennai

20 15 15

Jamgudrani MP

Dewas Chitradurga District Perungudi Kethanur

14

Jogmatti BSES

BSES Ltd. Newam Power Company Ltd. Kethanur Wind Farm Andhra Pradesh State Road Transport Corporation. Madras Cements Ltd. MMTCL

14

Perungudi Newam Kethanur Wind Farm

Tamil Nadu Tamil Nadu Andhra Pradesh Tamil Nadu Karnataka

12 11

Hyderabad APSRTC

Hyderabad

10

Muppandal Madras Shah Gajendragarh

Muppandal Gadag

10 15

Power Plant

Producer

Location

State

Total Capacity (MWe)

10.8

Shah Gajendragarh

Sanjay D. Ghodawat Tuppadahalli Energy India Private Limited Chettinad Cement Corp. Ltd. Shalivahana Green Energy. Ltd. Oil India Ltd.

Gadag Chitradurga District

Karnataka

Acciona Tuppadahalli

Karnataka

56.1

Poolavadi Chettinad

Poolavadi

Tamil Nadu

10

Shalivahana Wind Dangiri Wind Farm

Tirupur Jaiselmer

Tamil Nadu Rajasthan

20.4 54

Table 1.1: Indias largest wind power plants The growth of Renewable Energy in India is enormous and Wind Energy proves to be the most effective solution to the problem of depleting fossil fuels, importing of coal, greenhouse gas emission, environmental pollution etc. Wind energy as a renewable, non-polluting and affordable source directly avoids dependency of fuel and transport, can lead to green and clean electricity.

Figure 1.5: Indias cumulative wind power installation year by year

With an installed capacity of 19 GW (Mar 2013) of wind energy, Renewable Energy Sources (excluding large Hydro) currently accounts for 12.5% i.e. 27.5 GW of Indias overall installed power capacity. Wind Energy holds the major portion of 70% among renewable and continued as the largest supplier of clean energy. In its 12th Five Year Plan (2012-2017), the Indian Government has set a target of adding 18.5 GW of renewable energy sources to the generation mix out of which 11 GW is the Wind estimation and rest from renewable sources like Solar 4 GW and others 3.5 GW. The Wind Potential in India was first estimated by Centre for Wind Energy Technology (C-WET) at 50m hub-height i.e. 49 GW but according to the new survey at 80m hub height, the potential grows as much as 102 GW. This figure was adopted by the government as the official estimate. However, Lawrence Berkley National Laboratory (LBNL) has stated the potential is over 300 to 400 GW. One of the major advantages of wind energy is its inherent strength to support rural employment and uplift of rural economy. Further, unlike all other sources of power, wind energy does not consume any water-which in itself will become a scarce commodity. Overall the future of Wind Energy in India is bright as energy security and self-sufficiency is identified as the major driver. The biggest advantage with wind energy is that the fuel is free, and also it doesnt produce CO2 emission. Wind farm can be built reasonably fast, the wind farm land can be used for farming as well thus serving dual purpose, and it is cost-effective as compare to other forms of renewable energy.

1.3 Wind power industry

In 2010 wind energy production was over 2.5% of total worldwide electricity usage, and growing rapidly at more than 25% per annum. The monetary cost per unit of energy produced is similar to the cost for new coal and natural gas installations. Worldwide there are now over two hundred thousand wind turbines operating, with a total nameplate capacity of 282,482 MW as of end 2012.The European Union alone passed some 100,000 MW nameplate capacity in September 2012, while the United States surpassed 50,000 MW in August 2012 and China passed 50,000 MW the same month.

10

Figure 1.6: World wind power capacity year by year

World wind generation capacity more than quadrupled between 2000 and 2006, doubling about every three years. The United States pioneered wind farms and led the world in installed capacity in the 1980s and into the 1990s. In 1997 German installed capacity surpassed the U.S. and led until once again overtaken by the U.S. in 2008. China has been rapidly expanding its wind installations in the late 2000s and passed the U.S. in 2010 to become the world leader.

11

Top 10 countries by nameplate windpower capacity (2012 year-end)

Country China United States Germany Spain India UK Italy France Canada Portugal (rest of world) World total New 2012 capacity (MW) 12,960 13,124 2,145 1,122 2,336 1,897 1,273 757 935 145 6,737 44,799 MW Windpower total capacity (MW) 75,324 60,007 31,308 22,796 18,421 8,845 8,144 7,564 6,200 4,525 39,853 282,587 MW % world total 26.7 21.2 11.1 8.1 6.5 3 2.9 2.7 2.2 1.6 14.1 100%

Table 1.2: Top 10 wind power capacity countries by nameplate

Top 10 countries by windpower electricity production (2011 year-end)

Country United States China Germany Spain India Canada UK France Italy Denmark (rest of world) Windpower production (TWh) 120.5 88.6 48.9 42.4 24.9 19.7 15.5 12.2 9.9 9.8 67.7 % world total 26.2 19.3 10.6 9.2 5.4 4.3 3.4 2.7 2.1 2.1 14.7

Table 1.3: Top 10 wind power capacity countries by production

12

From the above data, it is very clear that wind industry growing very fast in the world. But many of the wind farms are running under capacity due the various problem of wind availability throughout year. Many of the research are going for the optimization of wind power and extract maximum of it. Many of the industries are using different design of turbine blades, keeping auxiliary generators, scheduling of farm by using different algorithm & data. It is very important for any booming industry. The future of wind industry is luxurious.

1.4 Objective of thesis

Out of 5 objectives, in this thesis I will discuss only about 2 objectives. These two are having a wide possibility to optimize the performance of turbines. 1) Control of Wind Turbine Power and Vibration with a Data Driven Approach: In this the optimization of wind turbine performance will be based on an anticipatory control model. Both maximizing wind power generation and minimizing the drive-train and tower vibration are considered in the optimization. The drive-train and tower vibration is represented by the vibratory acceleration of drive-train and tower measured with accelerometers.
2) Designing of wind farm layout: It is a model to wind turbine place in order to capture maximum wind based upon wind distribution. It considers wake loss which can be calculated based of turbine location and wind distribution.

1.5 Scope of thesis

This thesis includes the strategies and methodologies used in the analysis and design, optimization problem formulation, and implementation portions of this research. In addition, it includes worked example problems, a discussion of numerical results and conclusions. In Chapter 3, the optimization of wind turbine performance based on an anticipatory control model is discussed. Both maximizing wind power generation and minimizing the drive-train and tower vibration are considered in the optimization. The drive-train and tower vibration is represented by the vibratory acceleration of drive-train and tower measured with accelerometers. Chapter 4 presents a model for wind turbine placement based on the wind distribution in order to maximize the wind energy capture. The model considers wake loss, which can be calculated based on wind turbine locations, and wind direction. Since the turbine layout design is a constrained optimization problem, for ease of solving it, the constraints are transformed into a second objective function. Then a multi-objective evolutionary strategy algorithm is developed to

13

solve the transformed bi-criteria optimization problem, which maximizes the expected energy output, as well as minimizes the constraint violations. The presented model is illustrated with examples as well as an industrial application. Chapter 5 contains a discussion of the optimum solutions and the numerical results obtained during solution. This chapter also presents conclusions from the study and areas for further research.

14

Chapter 2

Review of Literature
2.0 Introduction

To solve emerging problems in the rapidly expanding wind energy industry, wind energy research has intensified in recent years. The published literature with wind energy can be categorized into, design of wind turbines (Laino et al. 1993; Saranyasoontorn and Manuel 2004), the design and reliability of wind farms (Barthelmie et al. 2007; Mora et al. 2007; Leite et al. 2007), control of wind turbines (Senjyu et al. 2006; Ko et al. 2008; Johnson et al. 2006; Munteanu et al. 2005), the prediction of wind power (Ko et al. 2008; Mutlu et al. 2009), wind energy conversion (Kusiak et al. 2009b; Kusiak et al. 2009c) and condition monitoring of wind turbines (Kusiak et al. 2009d; Kusiak et al. 2009e). Although the wind turbine efficiency topic has been studied in the literature, several shortcomings can be noted.

15

16

17

18

The wind power generation cost can be considered from two perspectives, the efficiency of power generation and the O&M cost of the wind turbine. Therefore, simply improving the efficiency of wind power generation cannot reduce the power generation cost as the cost of turbine consumed mechanical components may increase. Another shortcoming is that improving the performance model of a wind turbine does not easily scale up to that of a wind farm. Numerous factors, such as electricity demand, wind speed, and electricity price might impact the performance of a wind farm. Scheduling wind farms needs to be investigated. Moreover, the majority of the published studies about control of a wind turbine system involve Physics-based models. Since such models usually involve stringent assumptions, they do not adequately model working systems.

19

2.1 Review of Literature

In this thesis I am concern about two goals to optimize the performance. Which will be detailed describe in chapter 3 & chapter 4.

2.1.1 Wind Turbine Vibration and Power Generation In the published literature, wind turbine vibration and power generation have been studied as two independent topics. The published research on wind turbine vibration has primarily focused on the models derived from the first principles. Leithead and Connor (2000) studied dynamics of variable speed wind turbines and design of models to control wind turbines. Fadaeinedjad et al. (2008) investigated the impact of voltage sag on vibration of the wind turbine tower. They used three simulation programs, TrubSim, FAST and Simulink, to model wind turbines. Murtagh et al. (2008) investigated control wind turbine vibration by incorporating a passive control device. A passive control method using a tuned mass damper to mitigate vibrations of the blades and tower of a wind turbine was introduced. Hansen et al. (2006) discussed the estimation of aeroelastic damping of operational wind turbine modes based on experiments. Hansen et al. (2003) also presented analysis of vibrations in a three-blade wind turbine. Molinas et al. (2010) addressed extending the life of a gearbox by smoothing the transient generator torque to control vibration. Although first principles based models provided solid foundation of understanding the nature of wind turbine vibration, the real wind turbine vibration cannot be fully reflected due to the assumptions and the diverse source of wind turbine vibration (Wowk 1991). Predicting wind power is an important direction of the wind energy research since wind power prediction is critical to power system reliability and handling variability of the generated power (Bathurst et al. 2002). Long-term wind speed and power prediction is of interest to management of energy distribution (Barbounis et al. 2006). Short-term prediction of wind turbine parameters is key to anticipatory control of wind turbines (Senjyu et al. 2006) and on-line monitoring (Kusiak et al. 2009d, Kusiak et al. 2009e). Numerous approaches for predicting wind speed and power generated by wind turbines have been developed. Potter and Negnevitsky (2006) presented an adaptive neuro-fuzzy system to forecast wind speed considering both wind speed and direction. Boukhezzar et al. (2006) presented the estimation of wind speed based on the aerodynamic torque. Louka et al. (2008) applied Kalman filters as a post-processing method to enhance wind speed prediction. Flores et al. (2005) employed neural networks for wind speed prediction in time-scales can vary from 1 minute to an hour and designed a control system for active power generation. ElFouly et al. (2008) developed a linear time-series based models to predict both wind speed and direction. Damousis et al. (2004) used a fuzzy logic model trained by a genetic algorithm to predict wind speed and power over 0.5 to 2 hour horizons.

20

Based on the predicted wind speed, wind power models are conducted for wind power prediction. Sideratos and Hatziargyriou (2007) presented an advanced statistical method for wind power forecasting 48 hours ahead based on artificial intelligent techniques. Bessa et al. (2009) reported the adopting entropy concepts to the training of neural network to build hourly based power prediction model. Kariniotakis et al. (1996) developed a recurrent high order neural network for building power prediction models based on 10-minute average data. Besides the power prediction, wind power models are also discussed in the power optimization research. Boukhezzar and Siguerdidjane (2009) presented a non-linear controller for optimizing the power of the DFIG (Doubly Fed Induction Generator). Wang and Chang (2004) investigated an intelligent power extraction algorithm for improving the performance of wind turbine systems. Morimoto et al. (2005) researched the maximization of wind-turbine-generated power by controlling the current vector of the interior, permanent-magnet, synchronous generator. Muljadi and Butterfield (2001) addressed operating a variable-speed wind turbine with pitch control to maximize power while minimizing the loads. Moyano et al. (2009) adopted an operation strategy to optimize wind power on a wind park control level. In the published literature, statistical, Physics-based and control theory based wind power models are widely discussed. However, the drawback of such models is also obvious. Assumptions, delay of feedback loop and delay of measuring wind speed impair the results of power estimation. 2.1.2 Wind Farm layout Problem Although most current wind farm layout optimization studies focus on the layout of turbines (also the focus of this thesis), we also want to explore additional optimization topics that are closely related: layout of infrastructure, and sizing of energy storage system. Our ability to model and solve these optimization problems influences the performance of wind farms on multiple dimensions: profitability, environmental footprint and social responsibility. Each problem is explored individually with emphasis on modeling and optimization techniques. Directions of current research and suggestions for future research are discussed at the end of the review. Overall, we found that the current scope of the turbine and infrastructure layout literature is limited to simple problem set-up: at terrain, simple cost functions, and little integration of the turbine and infrastructure problems. The energy storage system sizing problem is closely related to the dynamic problem of wind energy dispatch. For all three problems, the main research contributions have been from the engineering literature. Based on these observations, we propose new research directions.

21

Bryony L. Du Pont et al.(2012) used a extended pattern for optimization of turbine placement layout in wind farms. Andrew Kusiak et al. (2009) presented a model for wind turbine placement based on the wind distribution. The model considers wake loss, which can be calculated based on wind turbine locations, and wind direction.

2.2 Conclusion
Much literature published which describes the optimization in case of variable wind speed, vibrations in a wind turbine and inability to capture maximum wind. Off shore wind turbine layout design is also a widely discussed literature. Many of the articles are publishing in day to day life. Many models, algorithm is presented in order to solve the various problems and optimize the performance of wind turbines. New improved turbine blade design to get less noise, less vibration and maximum speed in low wind area also is a widely discussed literature. Still lot of improvement to be done as wind industries are growing rapidly due to renewable source of energy.

22

CONTROL OF WIND TURBINE POWER AND VIBRATION WITH Chapter 3 A DATA DRIVEN APPROACH
3.0 Introduction
Improving performance of wind turbines is of interest to the wind energy industry as it provides an opportunity to reduce the cost of generating electricity. This cost is largely determined by the efficiency of wind turbines and maintenance of the equipment (U.S. Department of Energy 2008). This chapter investigates the optimization of wind turbine performance in based on an anticipatory control model (Kusiak et al. 2009a; Camacho and Bordons 1999; Rossiter 2003). Both maximizing wind power generation and minimizing the drive-train and tower vibration are considered in the optimization. The drive-train and tower vibration is measured with accelerometers. The task of modeling the power generation process and wind turbine vibration is challenging. In conventional analysis, the first principles and aerodynamics are frequently applied. However, assumptions made by the classical analysis approach limit its practical applications. In this chapter, a data-driven approach is utilized to model power generation and turbine vibration. The sources of wind turbine vibration are diverse (Wowk 1991). The focus here is on vibrations attributed to the control of wind turbines, e.g., control of the blade pitch and the generator torque. The performance optimization model is derived from data, and it includes constraints. As data-driven models are usually non-linear and non-parametric, conventional optimizers cannot solve them. A particle swarm optimization algorithm (Kennedy and Eberhart 1995; Shi and Eberhart 1998; Abido 2002) is applied here do solve this challenging problem. Solutions of this optimization problem are the optimal settings of the controllable parameters, generator torque and blade pitch angle.

3.1 Data Description and Processing

In this research, data collected from supervisory control and data acquisition (SCADA) systems installed at a large wind farm is utilized. The sampling frequency of data collected over a period of one month is 0.1 Hz (this is the highest frequency permitted for research). Though the SCADA system collects data on over 120 parameters, a subset of parameters is used in this research. The

23

domain knowledge and the results from the previous studies are used to select most relevant parameters. Table 3.1 demonstrates the sample data used in this research.

Table 3.1: Sample data set

To enhance accuracy of the data-driven models, the data is pre-processed to exclude data caused by sensor failures, transmission errors, and failures of various subsystems. These errors usually appear as the data exceeding physical constraints or missing values. The data quality also impacts accuracy of the models extracted by data-driven approaches. As the accelerometers are sensitive to noise, the Daubechies wavelet (Daubechies 1992) is applied to de-noise and smooth the measured drive-train and tower acceleration data. In particular, two threshold schemes of Daub 5 wavelets with 5 levels, Fix_Soft and Fix_Hard (Kobayashi 1998; Tang et al. 2000), are considered to obtain an acceptable threshold setting. Fix thresholding uses a fixed threshold multiplied by a small factor. Hard and soft thresholding could potentially produce different de-noising results. Hard thresholding is the usual process of setting to zero for the elements with absolute values lower than the threshold. Soft thresholding is an extension of hard thresholding, first setting to zero the elements with absolute values lower than the threshold, and then reducing the nonzero coefficients towards zero. The entropy type of the wavelet analysis can be accomplished with SURE (Stein's Unbiased Risk Estimate) (Stein 1981) which uses function (3.1). Entropy Threshold = -3.1

24

To increase de-noising efficiency, the best tree algorithm (Coifman and Wickerhauser 1992) is applied. The dataset from 11/1/2009 12:00:40 AM to 11/2/2009 4:09:40 AM is selected to perform the comparative analysis of two threshold settings. The following four metrics, the Mean Square Error (MSE), the Standard Deviation of Square Error (SD of SE), the Maximum of Square Error (Max SE), and the Minimum of Square Error (Min SE), are used to evaluate the two threshold settings (see (3.2) (3.5)).

-- 3.2

-- 3.3 -- 3.4 -- 3.5 Table 3.2 summarizes the comparative analysis of two threshold settings for de-noising the drive-train acceleration. Table 3.3 illustrates the same results for denoising the tower acceleration. As illustrated in both tables, the threshold setting, Fix_Hard, is more suitable for de-noising the two accelerations, and therefore it is selected for further study. Threshold Fix_Hard Fix_Soft MSE 16.1052 31.6875 SD of SE 28.0168 64.2476 Max SE 316.0088 786.8944 Min SE 0 0

Table 3.2: Statistics of de-noising drive-train acceleration with two threshold settings

Threshold Fix_Hard Fix_Soft

MSE 99.1767 141.5024

SD of MSE 161.2681 262.0488

Max SE 1840.8441 5098.8578

Min SE 0 0

Table 3.3: Statistics of de-noising tower acceleration with two threshold settings

25

The dataset after pre-processing is split into two parts, the training and the test dataset. The training dataset includes data from 11/1/2009 12:00:30 AM to 11/7/2009 5:58:30 PM, around 10000 points. The test dataset contains about 5000 data points, from 11/7/2009 5:59:10 PM to 11/8/2009 7:52:30 AM.

3.2 Dynamic Modeling and Prediction for Anticipatory Control

The anticipatory control of wind turbine vibration and power generation discussed in this chapter includes three data-driven models, a power generation model, a drive-train vibration model, and a tower vibration model. The power generation model describes the dynamic process of extracting power from the wind, while the other two models represent turbine vibration. Two controllable parameters, the generator torque and the blade pitch angle, are utilized to control the performance of the wind turbine. The concept of anticipatory control falls into the category of model predictive control (MPC) (Camacho and Bordons 1999; Rossiter 2003). It extends the MPC scheme by including controllable and noncontrollable parameters. It aims at generation of control settings ahead of the current process. 3.2.1 Modeling Power Generation The dynamic process of power generation is represented by the four-tuplet (v, , , P), where v represents the wind speed (non-controllable parameter), is the generator torque (controllable parameter), is the blade pitch angle (controllable parameter), and P is the power generated by the wind turbine (response parameter). The value of the response parameter, P, is impacted by the non-controllable and controllable parameters shown in (3.6). -- 3.6 Where t is the current time, T represents the sampling time of data (10-s data), and f (.) is the function describing the wind turbine energy conversion learned by the data-mining algorithms. To extract the data-driven model, six data-mining algorithms, neural network (Siegelmann and Sontag 1994; Liu 2001; Smith 1993), neural network ensemble (Hansen and Salamon 1990), support vector machine (Schlkopf et al. 1999; Steinwart and Christmann 2008), k nearest neighbor (Shakhnarovish et al. 2005), boosting regression tree (Friedman 2001; Friedman 2002) and random forest regression (Breiman 2001), are considered the most suitable algorithm is selected. Four metrics (3.7) (3.10), the mean absolute error (MAE), standard deviation of absolute error (SD of AE), mean absolute percentage error (MAPE) and standard deviation of absolute percentage error (SD of APE), are employed to

26

evaluate the performance of candidate algorithms.

The training and test data introduced in Section 3.1 is used in the comparative analysis. The test dataset is used to determine the performance and robustness of the data-driven models. Table 3.4 presents the test results derived by the data-driven models. As shown in Table 3.4, models extracted by three algorithms, neural network ensemble, neural network and k nearest neighbor, exhibit similar performance. However, the MAPE (0.0096) and SD of APE (0.0155) offered by the neural network model are slightly better than the MAPE and SD of APE of the other two. Hence, the neural network is selected as a more suitable algorithm to build the power generation model.

Table 3.4: Accuracy of power generation models extracted with six algorithms

Figure 3.1 and Figure 3.2 illustrate the prediction accuracy of the model developed by the neural network algorithm for the first 300 points from the test set. Figure 3.1 is the run chart, and Figure 3.2 presents the scatter plot of the observed and the predicted values.

27

Figure 3.1: The power corresponding to the first 300 observed and power model predicted points

Figure 3.2: The predicted and observed value of the first 300 test points

3.2.2 Modeling Wind Turbine Vibration In this research, vibrations of the drive-train are measured by the accelerometer installed at the bottom of a nacelle. The tower acceleration is measured by the accelerometer located at the connection of the nacelle and the tower. The wind turbine vibrations are largely due to the air passing through the wind turbine and the forces originated by the control system impacting the generator torque and the blade pitch angle. Based on the domain knowledge and

28

the past research, six parameters, the generator torque, blade pitch angle, wind speed, wind deviation, drive-train acceleration, and tower acceleration, are selected to model wind turbine vibration (see Table 3.5). In Table 3.5, the drivetrain acceleration and the tower acceleration are the response parameters. The generator torque and the blade pitch angle are controllable parameters, while the wind speed and wind deviation are non-controllable parameters. Parameter Drive-train acceleration Tower acceleration Generator torque Blade pitch angle Wind speed Wind deviation v wd Symbol Ad At

Table 3.5: Parameter list for modeling wind turbine vibration

Wind turbine vibration is a dynamic process. To model the drive-train acceleration that portrays the vibration in the drive-train system, the current and the past states of parameters need to be considered. In this section, the past three states of controllable and non-controllable parameters are included in the training and testing dataset mentioned in Section 3.1 to build the models of drive-train acceleration. Parameter selection is performed to reduce the dimensionality of the dataset and simplify the structure of the model. The wrapper genetic algorithm (Witten and Frank 2005; Kohavi and John 1997) is applied for parameter selection based on the training dataset of Section 3.1. Table 3.6 illustrates the parameters used in the selection process and the selected parameters, where, t is the current time, and T represents the time unit (10-s in this chapter). The model of the drive-train acceleration is expressed in (3.11) using the selected parameters and the past states of the drive-train acceleration itself.

The notation of Table 3.5 and Table 3.6 is used in model (3.11). Data-mining algorithms of Section 3.2.1 are used to model drive-train acceleration. The training and test datasets used here are the same as in Section 3.1.

29

Table 3.7 illustrates the test results of the drive-train acceleration models derived by the six data-mining algorithms. The neural-network ensemble algorithm provides lower values of MAPE and SD of APE than the other five algorithms, and it is selected to model the drive-train acceleration.

Table 3.6: List of parameters before and after parameter selection

30

Algorithm MAE Neural Network Ensemble 2.8376 Neural Network 2.8929 Support Vector Machine 11.4550 Boosting Regression Tree 3.3508 Random Forest (Regression) 14.5029 k Nearest Neighbor 4.3069

SD of AE 4.4902 4.4559 8.0720 4.9792 12.6100 5.6329

MAPE 0.0517 0.0569 0.3742 0.0686 0.3148 0.0920

SD of APE 0.0608 0.0621 0.4128 0.0674 0.1947 0.0850

Table 3.7: Accuracy of drive-train acceleration models extracted with six algorithms

Figure 3.3 presents results produced for the first 300 points by the drive-train acceleration model extracted by the neural network ensemble algorithm. Figure 3.4 illustrates the observed value and predicted value of these 300 points.

Figure 3.3: Results for the first 300 points used to test the drive-train acceleration model

Figure 3.4: The predicted and observed value of the first 300 test points

31

To establish the tower acceleration model, same parameter selection procedure and data mining algorithm for developing the drive train acceleration model are applied. Table 3.8 lists the parameters selected for modeling tower acceleration.

Table 3.8 Selected parameters for modeling tower acceleration

The model of the tower acceleration is expressed as (3.12).

Where the notation is the same as in model (3.11). Table 3.9 presents the testing results of the models extracted by different data- mining algorithms. The neural network ensemble and the neural network offer the best performance in estimating tower acceleration. Since the MAPE for the neural network ensemble is slightly better than that of the neural network, the neural network ensemble is chosen to build the tower acceleration model. Figure 3.5 illustrates the observed and predicted values of the tower acceleration for the first 300 test points using the neural- network ensemble model. The scatter plot in Figure 3.6 demonstrates the relationship between the predicted and observed values for the same 300 points.

32

Algorithm Neural Network Ensemble Neural Network Support Vector Machine Boosting Regression Tree Random Forest (Regression) k Nearest Neighbor

MAE SD of AE MAP SD of 4.2667 8.7580 E 0.0637 APE 0.1021 4.3324 8.7523 0.0649 0.1000 54.105 19.8097 1.3150 1.0944 9 7.0456 10.2048 0.1089 0.1285 12.605 14.3985 0.2652 0.4624 2 12.824 12.2709 0.2407 0.2074

9 Table 3.9: Accuracy of the tower acceleration models extracted with six algorithms

Figure 3.6: The predicted and observed value of the tower acceleration for first 300 test points

33

3.2.3 Prediction Based on Data-driven Models

To realize anticipatory control, the values of the generated power and the wind turbine vibration need to be predicted a step ahead (here 10-s). Based on the dynamic models (3.6), (3.11), and (3.12) of Section 3.2.1 and 3.2.2, the models for one-step ahead predictions are formulated in (3.13) (3.15).

Where the notation is the same as in Section 3.2.2. 3.2.3.1 Modeling Wind Speed Based on the Time-series Method
The values of all parameters in Eqs (3.13) (3.15) can be measured or computed except the wind speed at time t + T. A model for wind speed prediction at time t + T needs to be developed. In this research, the time-series model (see (3.16)) for wind speed prediction is developed based on its past states.

v(t ) = f (v(t - T), v(t - 2T), v(t - 3T), ..., v(t - nT))

-- 3.16

Eight past states of the wind speed are considered as candidates for inclusion in the model. The wrapper approach of Section 3.2.2 is applied to select the most suitable parameters. Among the eight past states, only the 7th state, v(t 7T), is considered as an unnecessary parameter. Thus, Eq (3.16) is reformulated as (3.17). v (t ) = f (v(t - T), v(t - 2T), v(t - 3T), v(t - 4T), v(t - 5T), v(t - 6T), v(t - 8T)) -- 3.17

To build model (3.17) from the data, six data-mining algorithms are utilized. The performance of these algorithms is summarized in Table 3.10. As The neural network algorithm provides the lowest value of all the four metrics, MAE, SD of AE, MAPE and SD of APE. Hence, it is selected for modeling wind speed.

34

Algorithm Neural Network Ensemble Neural Network Support Vector Machine Boosting Tree Random Forest k-NN

MAE 0.3624 0.3620 0.3931 0.4367 1.0870 0.4240

SD of AE 0.3518 0.3518 0.3516 0.3846 0.8108 0.3963

MAPE SD of APE 0.0448 0.0400 0.0447 0.0399 0.0512 0.0450 0.0551 0.0464 0.1297 0.0806 0.0529 0.0464

Table 3.10 Accuracy of wind speed prediction models extracted with six algorithms

Figure 3.7 illustrates wind speed prediction results for the first 300 test points

3.2.3.2 One-step Ahead Prediction A model for one-step ahead prediction of the generated power, the drivetrain acceleration, and the tower acceleration is discussed (see Eqs (3.13) (3.15)). The model is tested with the test dataset introduced in Section 3.1. Table 3.11 presents the values of four metrics (defined in Section 3.2.1) for one-step ahead wind speed prediction. The value of MAPE for predicting generated power at t + T = 10-s is 0.0096 (see Table 3.11), which is almost identical to the accuracy of the power prediction at time t (see Table 3.4). The accuracy of predicting drive-train acceleration (MAPE = 0.0785) at t + T = 10-s is also close to that at the current time (MAPE = 0.0517). The MAPE for predicting tower acceleration at t + T = 10-s is higher (MAPE = 0.1054) relative to its predicted value at time t (MAPE = 0.0637). In general, the prediction accuracy of the three models in 10-s ahead prediction is better than 90%, which presents an acceptable result. Figures 3.8 3.10 illustrate

35

the first 300 results for of the three parameters, generated power, drive- train acceleration, and tower acceleration.

Parameter MAE SD of AE Generated power 4.5592 3.5510 Drive-train acceleration 4.1247 5.7525 Tower acceleration 6.5365 9.9910

MAPE SD of APE 0.0096 0.0142 0.0785 0.0815 0.1054 0.1425

Table 3.11 Results of one-step ahead prediction

Figure 3.8: First 300 points of one-step ahead prediction of the generated power

Figure 3.9: The results of one-step ahead prediction of the drive-train acceleration for first 300 points

36

Figure 3.10: The results of one-step ahead prediction of the tower acceleration for the first 300 points

3.3 Multi-objective Optimization

3.3.1 Anticipatory Control Process The process of anticipatory control is implemented as an iterative, finite horizon optimization of wind turbine performance. The anticipatory control model generates control settings for the wind turbine at the current time t and t + T. For example, assume the optimal settings of controllable parameters, the generator torque and the blade pitch angle for time t and t + T are (t)*, (t + T)*, (t)* and (t + T)*. The wind speed at time t + T can be obtained from model (3.17). The power generated at time t and t + T is computed from (3.18) and (3.19). The optimal drive-train acceleration is expressed in (3.20) (3.21) and the optimal tower acceleration is formulated in (3.22) (3.23).

37

Where the notation is the same as in Section 3.2.2. To implement the anticipatory control approach, the settings at time t + T are computed based on the current status parameters. Once the wind turbine is operated with the computed control strategies, the status of the wind turbine will be sampled, and the sampled data will be stored in a database as the inputs for the next optimization step. This procedure is repeatedly implemented to realize the anticipatory control. 3.3.2 Multi-objective Optimization Model To optimize the wind turbine performance, the power generated in the interval t to t + T can be maximized as an expression of a linear combination of P(t)* and P(t + T)* (Kusiak et al. 2009a). This can be transformed as minimization of the linear combination of |P(t)* min{1500, 2.625v(t)3}| and |P(t + T)* min{1500, 2.625v(t)3}| shown in (3.24).

Where Tp is the weighted sum, the wp is the weight coefficient (between 0 and 1)

38

and the min{1500, 2.625v(t)3} is the boundary of theoretical maximum power that can be obtained from the wind kinetic energy (Kusiak et al. 2009a). Like the maximization of the generated power, the minimization of drive-train acceleration and the tower acceleration is formulated as (2.25) and (2.26).

Where w is the weight coefficient in [0, 1]. To define the feasible ranges of the parameters, constraints (3.28) (3.31), are defined.

In this study, the value of generator torque is normalized to [min = 0, max =100%]. In addition, the maximum rate of change of the generator torque is limited to 40%. Therefore, the upper and lower bound for generator torque can be determined as min{100, CurrentSetting + 40} and max{0, CurrentSetting 40}. To control the blade pitch angle, the maximum value of blade pitch angle is set to 90.61 degrees and the minimum value is fixed at 0.57 degree. The two values are obtained from the maximum and minimum values of the blade pitch angle included in the dataset. The maximum rate of the blade pitch angle change is 15 degrees. Then the upper bound and lower bound of the blade pitch angle control can be expressed as max{ 0.57 , CurrentSetting 15} and min{90.61, CurrentSetting + 15}. The optimization model is formulated in (3.32).

39

Where C(.) is the objective function that transforms the multi-objective model to a single- objective model (see (3.24) (3.27)). The remaining notation is the same as in Section 3.2.2. The four controllable parameters are the solutions to the optimization model. Solving the optimization model (3.32) is a challenge for classical optimization algorithms. A particle swarm optimization (PSO) algorithm (Kennedy and Eberhart 1995; Shi and Eberhart 1998; Abido 2002) is applied to solve this problem. To implement the PSO algorithm, some parameters need to be initialized, such as the number of particles. In this study, the number of particles is arbitrarily fixed at 10. The inertial parameter, , which controls the impact of previous velocity on the current velocity in the velocity function, should also be determined. Usually it is a value between 0 and 1. In this chapter, it is set to 0.5. Besides this parameter, in the velocity function, there are another two constants c1 and c2, which require an assignment of values. Here, c1 and c2, are both arbitrarily set to 2 in this study.

40

3.4 Case Study and Computational Results

In this section, the computational results of optimizing the wind turbine performance with the PSO algorithm are presented. Before computing the optimal solutions for the wind turbine control, the value of the weights used in the objective of model (3.32) should be determined. In Eqs (3.24) (3.26), the weights, P, D and T, are all assigned the value 0.5. This value indicates the equal importance of the three objectives, generated power, drive-train acceleration and tower acceleration at time t + T. Two cases of weight assignments are studied to show different wind turbine control preference. In Case 1, the w1 = 1 and w2 = w3 = 0. In this case, the wind turbine is supposed to be controlled in maximizing the generated power. In Case 2, the w1 = 0 and w2 = w3 = 0.5. This case translates to a circumstance whereby the preference of controlling the wind turbine reduces the turbine vibration. In this part, a multi-point optimization is introduced. It is based on data from 11/8/2009 7:47:10 AM to 11/8/2009 8:03:50 AM (about 16 minutes) in the dataset. 3.4.1 Stopping Criteria of the PSO The number of iterations of the PSO to converge needs to be determined. Three instances of the dataset used in Section 3.5 are randomly selected to show the convergence speed of the PSO algorithm. The standardized best global fitness based on function (3.33) is employed to evaluate the convergence of the PSO

Figure 3.11: Convergence of the PSO algorithm for three instances of Case 1

41

algorithm. Figure 3.12 illustrates the number of iterations the PSO needs to converge in Case 1 for all three instances. Figure 3.13 shows the similar concept for Case 2.

Figure 3.12 Convergence of the PSO algorithm for three instances of Case 2

Figures 3.11 and 3.12 illustrate that the PSO converges within 50 iterations. However, to set a more conservative setting criterion the iteration limit of 100 is applied. 3.4.2 Case 1 (w1 = 0 and w2 = w3 = 0) This section addresses the multi-point optimization result for Case 1 based on the data points of Section 3.4. Table 3.12 summarizes the optimization results of anticipatory control for these data points. Since in this case the wind turbine is controlled with the preference being to maximize only the generated power, the weight coefficient for maximizing the power generation model is set to 1, and for minimizing wind turbine vibration to 0. As shown in Table 3.12, the gain of maximizing power at time t + T can approach 37.49%. However, the drive-train acceleration and tower acceleration both increase. The power advantage comes from the increased generator torque. Figures 3.13 and 3.14 illustrate the optimized control strategies at time t + T versus the observed control strategies. Objective Generated power Drive-train acceleration Tower acceleration Original Value (Mean) 1089.0740 53.1174 54.5172 Optimized Value (Mean) Gain (Mean) 1497.3160 37.49% 65.4381 23.20% 85.3627 56.58%

Table 3.12 Summary of multi-point optimization of Case 1

42

Figure 3.14: Optimized vs observed generator torque for Case 1

2.4.3 Case 2 (w1 = 0 and w2 = w3 = 0.5) Case 2 represents a control preference emphasizing reduction of wind turbine vibration rather than increasing the power output. In this case, the weight assigned to maximizing power is w1= 0, and the weights assigned to mitigating drive-train and tower acceleration are w2 = w3 = 0.5. Table 3.13 summarizes the optimization results of anticipatory control for Case 2. The vibration is reduced by 2.07% and 4.63% for the drive-train and the tower, respectively, at time t + T. The generated power is reduced by 9.42%. Figures 3.15 and 3.16 demonstrate the

43

optimized value and observed value of the blade pitch angle and generator torque at time t + T.

Objective Generated power Drive-train acceleration Tower acceleration

Original Value (Mean) Optimized Value (Mean) 1089.074 986.4936 53.11743 52.01851 54.51721 51.99516

Gain (Mean) -9.42% -2.07% -4.63%

Table 3.13: Summary of multi-point optimization in Case 2

Figure 3.15: Optimized Vs Observed blade pitch angle for Case 2

Figure 3.16: Optimized Vs Observed generator torque for Case 2

44

3.5 Conclusion
An anticipatory control model for optimizing performance of wind turbines was discussed in this chapter. It was set to accomplish three goals, increasing the power extracted from the wind, and mitigating vibration of the drive-train system and the tower. By considering these goals, the power generation model, drivetrain acceleration model, tower acceleration model and the time-series wind speed prediction model were derived to establish this optimization model. Data-driven approaches were applied to constrict these models from wind farm data. A comparative analysis was performed to assess the performance of data-driven models derived by different data-mining algorithms and select the best performing model. To solve the constructed optimization model, a particle swarm optimization algorithm was applied. Two strategies of anticipatory control were implemented, control of a wind turbine by emphasizing maximization of power generation and reduction of wind turbine vibration. Single point and multi-point optimization results were presented. As the generator torque and power are correlated, the increased generator torque improves the power generated. However, as shown in the optimization results, increasing torque led to a higher acceleration of the drivetrain and the tower. To reduce vibration of the drive- train and tower, an optimization strategy resulting in reduced generator torque was recommended.

45

Chapter 4

Design of Wind Farm layout

4.0 Introduction

The wind energy market is rapidly expanding worldwide. This rapid growth of the wind energy industry has led to cost reduction challenges. There are various ways of reducing the cost of producing wind power: for example, the site selection, site layout design, predictive maintenance, and optimal control system design. The wind farm layout design is an important component of ensuring the protabilit y of a wind farm project. An inadequate wind farm layout design would lead to lower than expected wind power capture, increased maintenance costs, and so on. Equation (4.1) captures the cost of energy (COE). -- 4.1 Where CI is the initial capital cost ($) of the wind farm; FCR is the xed charge rate (%/year); CR is the levelized replacement cost ($/year); CO&M is the cost of maintenance and operations ($/kWh); AEP is the annual energy production (kWh/year). Note that at the design stage, AEP is the expected (planned) wind energy production. The annual energy production AEP is affected by the turbine availability, i.e., the number of operational hours in a year. Maximizing the AEP is an effective approach for reducing the cost of energy production. In this paper, AEP is improved by optimizing the wind farm layout design, specically minimizing the wake loss. Wind farm layout has been addressed in the literature Grady et al. and Mosetti et al. used a genetic algorithm to minimize a weighted sum of wind energy and turbine costs. The wind farm is divided into a square grid to facilitate the encoding of a 01 type solution. Lackner and Elkinton presented a general framework to optimize the offshore wind turbine layout. Details of how to solve the optimization problem are neither discussed nor any other wind farm layout design tools, such as WASP. Castro Mora et al. also used a genetic algorithm to maximize an economic function, which is related to turbine parameters and locations. The wind farm is represented with a square grid. The wind energy calculation is not based on the power curve function, and wind direction was not fully discussed in their optimization models. This paper extends the approach by developing specic

46

mathematical models to calculate the wake loss based on turbine locations. Solution of the constrained optimization problem is fully discussed with a double-objective evolutionary strategy algorithm, which can be easily extended by considering additional constraints.

4.1 Problem formulation and methodology

Modeling the wind farm layout design problem calls for assumptions. However, the assumptions made in this paper are acceptable in industrial applications and they could be modied or even removed, if necessary. Assumption 1. For a wind farm project, the number of wind turbines N is xed and known before the farm is constructed. A typical wind farm project has its total capacity goal dictated by various factors, e.g., nances and turbine availability. For example, to achieve 150 MW (Mega Watt) capacity, a hundred 1.5 MW turbines are needed.

Figure 4.1: Wind farm boundary and the denition of the wind speed direction

Assumption 2. Wind turbine location is characterized by its two-dimensional Cartesian coordinates (x,y), represented as a vector x of the length relatively small variations of surface roughness. The optimal solution computed in this paper is represented by N Cartesian coordinates (xi,yi), i = 1,,N Assumption 3. The wind turbines considered for a wind farm are homogenous, i.e., they all have the same power curve function

47

P = f(v) Where v is the wind speed at the wind turbine rotor with a xed height, P is the turbine power output. The wind turbine nacelle is usually controlled so that the rotor is oriented perpendicularly to the wind direction. For ease of turbine operations and maintenance, wind farm developers usually deal with one supplier of wind turbines delivering one turbine type. Assumption 4. Wind speed v (at a given location, height, and direction) follows a Weibull distribution pv (v, k, c) = k/c (v/c) k-1 e (v/c)k Where k is the shape parameter, c is the scale parameter, and pv(.) is the probability density function. Note that Assumption 4 maynot hold for short-time horizons; however, a great number of sites have shown Weibull distribution of the wind speed. Assumption 5. At a given height, wind speed (a parameter of the Weibull distribution function) v is a continuous function of the wind direction , i.e., k = k(), c = c(), 0 < < 360. In other words, wind speeds at different locations with the same direction share the same Weibull distribution across a wind farm. The parameter is also assumed to have a continuous probability distribution p(). In a relatively at terrain, Assumption 5 is a reasonable one. Moreover, if the wind farm does not cover a wide range of terrain, the wind speeds at a xed direction should share a similar distribution. Future research could consider the wind speed distributions changing with directions and locations. Wind direction is an important parameter in this paper. Fig. 4.1 illustrates wind direction for a wind farm, where North is dened as 900 and East is dened as 00. Though wind turbines may follow different layout patterns, here it is assumed that all turbines are placed within a circular boundary of a wind farm. Assumption 6. Any two turbines in a wind farm are separated from each other by at least 4 rotor diameters. Ensuring sufficient spacing between adjacent turbines reduces interactions, e.g., wind turbulence, thus diminishing the hazardous loads on the turbine. Given the rotor radius R, any two wind turbines located at (xi,yi) and (xj,yj) should satisfy the inequality (xi xj) 2 + (yi yj) 2 < 64 R2 Assumption 6 is based on certain domain heuristics. However, it can be replaced by more complex constraints. For example, the minimum distance between two turbines can be a function of the wind direction and its probability. For a single wind turbine located at (x,y) and wind direction , the expected

48

energy production is

-- 4.2 Integrating the expression in (4.2) for in the range 0360 provides the expected energy production of a single wind turbine

-- 4.3
In this paper, optimizing E(P) is considered equivalent to optimizing AEP as the number of operational hours is xed. Note that in Equation (4.3), the wind speed is integrated from 0 to innity, which is equivalent to integrating the wind speed from 0 to the cut-out speed. Once the wind speed is greater than the cutout speed, a wind turbine is shut down for safety and it produces zero energy.

Model (4) represents an ideal scenario for maximizing the total wind energy captured by turbines, where the interactions among turbines are neglected, with r being the radius of the wind farm. Note that the circular boundary of a wind farm can be easily modied to a rectangular boundary.

-- 4.4 The major interactions among the turbines result in the energy loss caused by the wind turbine wakes. A wind farm design for the maximum of wind energy capture must minimize the wake effects among turbines.

49

4.2 The wake loss model

Wake loss is an important factor in considering wind park layout design. When a uniform incoming wind encounters a wind turbine, a linearly expanding wake behind the turbine occurs. A portion of the free stream winds speed will be reduced from its original speed vup to vdown. Fig. 4.2 illustrates the basic concept of wake behind a wind turbine.

Fig. 4.2: Wind turbine wake model

The velocity decit is dened as the fractional reduction of free stream wind speed in the wake of the turbine, and it is calculated from equation (4.5).

Where CT is the thrust coefcient of the turbine; k is the wake spreading constant, and d is the distance behind the upstream turbine following the wind direction. Equation (4.5) can be simplied if CT and k are assumed to be constant for all turbines.

For any two turbines located at (xi,yi) and (xj,yj) in a wind farm, Vel_defi,j is dened as the velocity decit at turbine i in the wake of turbine j. The wind direction is from j to i, di,j is the distance between turbine j and i projected on the wind direction .

50

When a turbine is affected by the wakes of multiple turbines, the total velocity decit is computed as (4.8)

Where Vel_defi is the total wind speed decit at turbine i. However, for a given wind direction, not all turbines generate the wake effects at turbine site i. Given wind direction q, all wind turbines rotors are normally positioned perpendicular to the wind direction. The wake behind a turbine could be seen as a part of an imaginary cone. Fig. 4.3 illustrates a half cone formed by a turbine located at (x,y) and /2) is calculated as arctan (k), and the distance between A and the rotor center point is R/k. Fig. 4.4 shows 4.5 turbines in a wind farm. If wind blows from the West, turbine 2 is in the wake of turbine 1. Turbine 3 is in the wake of turbine 1 and turbine 4. Turbine 5 is not affected by the wake of any other turbine. Lemma 1. Given the wind direction q, any two wind turbines, i and j located within the wind farm, the angle bi,j, 0 , between the vector, originating at turbine js cone vertex A to turbine i and the vector, originating at A to turbine j, bi,j is

Proof: Consider wind direction in Fig. 5, origin O, and any two turbines i and j (without loss of generality, assume i =1, j =2) represented by vectors x1 and x2, located at T1 and T2, respectively. The vector originating at A to T1 (Turbine 1) should also have angle , and its length is Vector can be expressed as

51

Thus vector

is and

If the angle (0 ) between the vectors inside the cone;

Substituting, ,

into equation (4.10), can be written as the function of R, k, and the turbines positions x1 and x2. Let,

be the angle used to determine whether turbine i is in the cone of turbine j given the wind direction . Lemma 2. If wind turbine i is inside the wake of turbine j, di,j is the distance between turbine i and j projected on the wind direction ,

Proof: The vector from T1 and T2 is calculated as x2 - x1 on is

Thus the projection length of Thus di,j can be written as

52

Figure 4.3: An imaginary half cone of wind turbine

It is easy to observe that Vel_defi is a function of and all turbine locations. It is shown that only scaling parameter c of a Weibull distribution will be affected by the wake loss, i.e.,

Although the wake loss model considered in this paper is rather simple, other complex wake loss models can be incorporated into the proposed optimization approach. Once the wake loss can be calculated based on wind turbine locations, the power curve function needs be determined for estimation of the expected power production.

53

4.3. Power curve

Though a power curve P = f(v) usually resembles a sigmoid function, it could be described as a linear function with a tolerable error. For example, model (17) illustrates a linear power curve function, where vcut-in is the cut-in wind speed. If the wind speed is smaller than the cut-in speed, there is no power output because the torque is not sufcient enough to turn the generator. Similarly, if the wind speed is greater than the rated speed and smaller than the cut-out speed vcut-out, the wind turbine control system will keep the power output stable (xed at Prated) to protect the system from hazardous loads. When the wind speed is between the cut-in and rated speed, the power output follows a linear equation, where is the slope parameter, and is intercept parameter. In future research, the linear power curve function could be replaced by a more accurate nonlinear approximation.

Using the linear power curve function of equation (4.17) in equation (4.16)

54

Figure 4.4: Turbines affected by the wake of other turbines

55

Figure 4.5: A wind turbine inside the cone of another turbine

However, if the wind speed is discretized into small bins, the integration can be approximated with the Riemann sum. Later in the paper, the wind speed and its direction are discretized so that numerical integration can be performed. It needs to be noted that continuous wind characteristics are not available in the design of wind farms.

4.4 Discretization of wind speed and wind direction

Assume the wind direction is discretized into N +1 bins of equal width; Let 1; 2;...; N be the dividing points of wind direction with 0 1 2 . N 360 ; 0 = 00 ; N + 1 = 360 Each bin is associated with a relative frequency 0 i 1, i= 0,., N For example 0 is the frequency of bin [0 ,1], N is the frequency of bin [N ,360 ]. The frequency i can be easily estimated from the wind data. Wind speed is also discretized into Nv + 1 bins Let, v1, v2, v3,.., vNv be the dividing points of wind speed with vcut-in v1 v2 v3 . vNv vrated
0

56

Can be calculated as

Can be calculated as

57

Substituting equation (4.18) into model (4.4) makes it is easy to solve the resultant optimization model. Once wind speed and wind direction are discretized into small bins, ci(.) and k(.) can estimated from the historical wind data. Discretization makes the calculation of the expected power tractable. Though model (4.4) is a constrained nonlinear optimization problem, considering nonlinear power curves or more complex wake loss models, and additional constraints would increase its complexity. Thus selecting a flexible optimization algorithm is important. This model complexity warrants the use of a population based search algorithm, such as the evolutionary strategy algorithm.

4.5 Evolutionary strategy algorithm

The solution of model (4.4) can be encoded as a vector used by an Evolutionary strategy algorithm. The general form of the jth individual in the evolutionary strategy algorithm is dened as (zj , j), where zj and j are two vectors with 2N enteries. i.e.,

58

The pair

is the position of the rst turbine, while is used to mutate the Nth turbines position.

is used to mutate

the rst turbines position. The pair

is the position of the Nth turbine and

Each element of j is used as a standard deviation of a normal distribution with zero mean. j is used to mutate the solution zj. The initial population is generated by uniformly sampling low up low up from the range [ , ], where and are the lower and upper bounds for the standard deviation vector. In this paper, a multi-objective evolutionary algorithm is used as a basic algorithm to solve a constrained optimization problem. All the constraints are transformed to formulate the second objective function. To standardize the optimization problem, model (4.4) is transformed into a minimization problem with two objectives.

Minimizing Obj1 equals maximizing the expected energy output. The minimum of Obj2 is zero, which means all constraints are satised. To solve the double-objective optimization problem, an evolutionary strategy algorithm is used. This algorithm matches the complexity of the underlying problem. 4.5.1. The algorithm 1) Initialize three empty sets Parent, Offspring and Elite. Randomly generate child individuals (solutions) to form the initial children population and place them in Offspring. 2) Repeat until the stopping criterion is satised. 2.1 Find non-dominated solutions in Offspring and copy them into Elite. Remove dominated solutions in Elite. Reduce the size of Elite by clustering, if necessary.

59

2.2 Fitness assignment: Assign tness to individuals in Offspring and Elite. 2.3 Selection: use tournament selection to select Parent Individuals from Offspring U Elite and store them in Parent. 2.4 Recombination: Generate a new population Offspring by selecting two parents in Parent. 2.5 Mutation: Mutate the individuals in Offspring. 2.6 Assign tness to the individuals in Offspring. The stopping criterion in this thesis is the number of generations. 4.5.2. Mutation An individual (zj,j) can be mutated by following the equations(4.19) and (4.20), with j mutated first, zj mutated next

Where N(0, )is a random number drawn from normal distribution with 0 mean and standard deviations ;Ni,x(0, )and Ni,y(0,( )i=1...N) are two random numbers drawn from normal distribution with a mean of 0 and standard deviations,
;Ni,x(0, )is generated specifically for

is generated specifically for while N(0, )is for all entries; and . denotes the Hadamard matrix product. The new solution is generated from equation (4.20).

Where N(0,j) is a vector of the same size as zj. Each element of N(0,j) is generated from a normal distribution with a mean of 0 and the corresponding standard deviation in vector j. 4.5.3. Selection and recombination of parents To generate Child children, two parents are selected from the parent population and recombined Child times. Assume each time two parents are selected randomly to produce one child by using equation (4.21):

60

Selected Parents is a set consisting of the two indices of the randomly selected parents. 4.5.4. Tournament selection Tournament selection with replacement is used in this thesis to select out promising individuals going into the next generation based on their fitness values. The tournament size is a predefined parameter to control the selection pressure.

4.6 Computational study

To illustrate the concepts presented in this paper, numerical examples and an industrial case study are presented. The wind turbines used in this paper have the following parameters: rotor radius is 38.5 (m); cut-in speed is 3.5 (m/s); rated speed is 14 (m/s); rated power is 1500 (kW). For the linear power curve function, = 140.86, = -500. Hub height is 80 (m). The thrust coefficient CT is assumed to be 0.8, the spreading constant k is assumed to be 0.075 for land cases. Knowing the cut-in wind speed and the rated wind speed, wind speed is divided at Nv = 20 intervals of 0.5 (m) each. Similarly, the wind direction is divided at N = 23 intervals of 150 each. Some of the parameters of the evolutionary algorithm used in this paper are fixed throughout all experiments. The tournament size is fixed at 4. Initial population zj is generated by uniformly sampling from [-r,r]. During the optimization process, all solutions are checked for non-violation of the interval constraint. Similarly j is limited to [low,up] set at [1,r/20]. In other words, the minimum standard deviation for mutating the solution is set at 1 (m); the maximum standard deviation is heuristically set at 1/20 of the farm radius, and it increases with the farm radius. Based on the heuristic knowledge,

61

 is

calculated as

1/ 1/

is calculated as

The number of generations in this thesis is set to 100.

Table 4.1: Wind scenario 1

Figure 4.6: Layout solutions for different number of turbines given the wind characteristics in Table 4.1

62

4.6.1 Wind scenario 1 The illustrative example used in this thesis assumes that The farm radius r = 500 (m); The wind characteristic of this example is described in Table 4.1, which was obtained from analyzing historical wind data. The data in Table 4.1 can be read as follows: when wind direction is between 00 and 150, the wind speed follows a Weibull distribution with k = 2, c =13; the probability for wind blowing between 00 and 150 is zero. Similarly, when wind direction is between 900 and 1050, the wind speed follows a Weibull distribution with k = 2, c = 13; the probability for wind blowing between 900 and 1050 is 0.6. Table 4.1 shows that wind blows predominantly from 750 to 1050 with a probability of 0.8. To illustrate the algorithms ability to optimize the wind farm layout, different numbers of turbines (26 turbines) are considered. Fig. 4.6 shows the optimal layout solutions for different numbers of turbines. The optimal solutions are obtained through Parent/Child = 20/120, for 100 generations.

During the optimization process, the Elite set size are limited to 50. When there are more than 50 individuals in the Elite set, a clustering procedure is activated to delete individuals with the same objective function values or similar geometric positions. Based on Fig. 4.6, for two turbines to be placed within the farm radius, the algorithm places the two turbines on the boundary of the circle. The line connecting the two turbines appears to be perpendicular to the major wind direction (900 1050). When there are three turbines to be placed, the algorithm positions the two turbines (lower part of Fig. 4.6) across the major wind direction, and attempts to locate the third turbine as far as possible from the other two turbines, and to keep it from the wakes of the other two turbines given the major wind direction. As shown in Fig. 4.6, when there are 4, 5 or 6 turbines to be located, the algorithm selects the locations, avoiding the wake losses under the major wind directions. Fig. 4.7 shows the initial population and the population at the 100th generation for the 6 turbines. It is easy to see that the algorithm converges as the number of generations increases. However, 100 generations are enough to converge for the data used in this example.

63

Figure 4.7: Initial population and the population at the 100th generation

For a large number of turbines, the search space expands due the increased number of individuals in the ES algorithm. Thus increasing the initial population size seems to be a viable way to search for optimal solutions. Fig. 4.8 shows two optimal solutions generated with different population sizes. The triangle shapes indicate the locations generated when Parent/Child = 20/120. The square shapes show the locations generated when Parent/Child = 200/1200 Using Parent/Child = 20/120, The algorithm can nd Obj1 = 1.1939 x 10-5. Using Parent/Child = 200/1200, the algorithm nds Obj1 = 1.1937 x 10-5, which is a better value than in the case of a smaller initial population. Note that increasing the population size leads to better quality solutions. It is worth mentioning that the tournament size should also be modied when the population size changes.

64

4.6.2. Wind scenario 2 Table 4.2 shows the wind data collected from an industrial wind farm. For this wind farm, the wind blows predominantly from 1200 to 2250, this is a wider direction spectrum than that of wind scenario 1. Wind scenario 1 is easier to see whether the optimization results make sense in terms of minimizing wake loss. Fig. 4.9 shows the optimal layout solutions for different numbers of turbines (26). In this wind scenario 2, wind farm radius is still assumed to be 500 (m). All the solutions are generated with Parent/Child = 20/120, for 100 generations

Figure 4.8: Six layouts of turbines generated by using another initial population size

65

The line connecting the two turbines pretends to be perpendicular to the major wind direction (12002250). When there are three turbines to be placed, the algorithm positions the two turbines across the major wind direction, and attempts to locate the third turbine as far as possible from the other two turbines, while minimizing the wake loss generated by the third turbine from the major wind directions. As in Fig. 4.9, when there are 4, 5 or 6 turbines to be located, the algorithm selects the locations, avoiding the wake losses under the major wind directions. Note that when the wind direction displays a wide range, it is almost impossible for human heuristics to derive an optimal solution. In this case, a wind farm layout optimization tool is necessary for designing turbine locations. 4.6.3. Wake loss If it is assumed that there is no wake loss, the expected power can be calculated from equation (4.22) by substituting ci(.) with c(.) in Equation (4.18). In this scenario, optimal turbine locations only need to satisfy the distance constraints.

Table 4.2: Wind scenario 2

66

Figure 4. 9: Layout solutions for different numbers of turbines under the wind characteristics in Table 4.2

Thus equation (4.22) can be regarded as the ideal scenario where the maximum expected energy production from a wind turbine can be achieved. For a given number of turbines, equation (4.22) can be used to calculate the ideal energy production for a wind farm without considering wake loss. Once the wake loss is considered, adding additional turbines to the wind farm may cause energy loss. Tables 4.3 and 4.4 are generated based on the two different wind scenarios, 1 and 2.

67

Table 4.3: Optimized vs. Ideal energy production and wake loss for different numbers of turbines under wind scenario 1

The data in Table 4.3 is read as: when there are two turbines to be placed in a wind farm, the optimized solution generated by the evolutionary algorithm provides 28083.42 kW of expected energy production (note that the yearly production is not computed). If there is no wake loss, the ideal energy production is 28091.47 kW; the potential wake loss under this layout solution is calculated by using 28091.47 - 28083.42 = 8.05 Tables 4.3 and 4.4 show that as the number of turbines increases, the wake loss increases. Fig. 4.10 shows that for two different wind scenarios, the wake loss of wind scenario 1 is similar to the wake loss of wind scenario 2 when the number of turbines is 25. When the number of turbines is 6, the wake loss of wind scenario 2 is significantly higher than that of wind scenario 1. The reason is obvious, as wind scenario 2 includes a wider range of wind directions with a certain probability. However, in wind scenario 1, the wind direction is limited to 750 1050; other wind directions have almost zero probability, which makes it easier to place turbines and avoid wake loss.

Table 4.4: Optimized vs. Ideal energy production and wake loss for different numbers of turbines under wind scenario 2

Based on Fig. 4.10, for a given wind farm radius, it is obvious that there should be some limit on the number of turbines to be placed on a farm. Locating too many turbines will significantly increase wake loss. Based on the constraints listed in model (4.4), placing 7 turbines within a wind farm of a 500-m radius is not possible.

68

The evolutionary algorithm does not generate a feasible solution when the number of turbines is 7. The minimum distance between two turbines is violated.

Figure 4.10: Wake loss due to the turbines added to the wind farm.

4.7 Conclusion
A generic model for optimizing in-land wind farm layout was presented. The optimization model considered wind farm radius and turbine distance constraints. However, other constraints can be easily incorporated in this model. The model maximizes the energy production by placing wind turbines in such a way that the wake loss is minimized. As the optimization model is nonlinear, and it is hard to derive an analytical solution from the integration part, wind speed and wind direction are discretized into small bins so that the integration can be approximated with discrete summations. A bi-objective evolutionary strategy algorithm was adopted to solve the constrained nonlinear optimization problem. Both the expected energy production and the constraint violation were optimized. The optimal solution maximized energy production while satisfying all constraints.

69

Although the wake loss model considered in this thesis is a simple one, more complex wake loss models can be considered in the optimization approach discussed in the thesis. The complex nonlinear model was solved by the evolutionary strategy algorithm. Although a global optimal solution is not guaranteed, the quality of the generated solutions is acceptable for industrial applications.

70

Chapter 5
5.0 Introduction

Conclusion

A data-mining framework for optimization of wind turbines performance was introduced. The framework included construction and optimization of data-driven models. Data-mining algorithms were applied to identify accurate wind turbine models from the SCADA collected data. Once the accuracy of data-driven models was validated, optimization algorithms were utilized to compute the optimal solutions. Since data-driven models are not explicit, evolutionary computation algorithms, such as evolutionary strategy, particle swarm optimization and artificial immune systems, were considered. Although the global optimal solution was not always guaranteed by evolutionary computation algorithms, the computed solutions improved performance of wind turbines and wind farms. A generic model of turbines in-land placement is discussed in order to use maximum of the wind, by placing turbines at particular distance and particular angle, and getting maximum energy. By placing the turbines in such a manner also reduces the wake losses. The optimization model is non-linear equation so evolutionary strategy algorithm is used to solve it.

5.1 Scope of future work

In the single wind turbine study, one challenge of the future research was the wind speed signal sensed by anemometer mounted at the back of wind turbine nacelle. Due to the location of the wind speed sensor, the measured data could not describe the wind speed conditions at the front of wind turbine rotor. Control of wind turbines based on the current wind speed data impaired the performance of wind turbine. Therefore, a more advanced wind speed sensors, such as laser sensors, needed to be mounted on wind turbines to improve the wind turbine control. As far as designing of wind farm layout is concern, now our future research can focus on considering different terrains, e.g., the terrain height information could be incorporated into the model. Furthermore, wind turbine parameter selection could be considered based on the wind farm wind characteristics. For example, for a low wind speed area, there is no need to select large turbines with a high cut-in speed. In other words, it is desirable to maximize the wind turbine capacity utilization factor subject to investment and operational constraints.

71

5.2 Conclusion
An anticipatory control model for optimizing performance of wind turbines was developed in Chapter 3. It was set to accomplish three goals, increasing the power extracted from the wind, and mitigating vibration of the drive-train system and the tower. By considering these goals, the power generation model, drive-train acceleration model, tower acceleration model and the time-series wind speed prediction model were derived to establish this optimization model. To solve the constructed optimization model, a particle swarm optimization algorithm was applied. An evolutionary strategy algorithm is used to solve the non-linear optimization problem in chapter 4. The main goal of this chapter is to use maximum of the wind by placing turbines at appropriate position. Two wind scenarios is considered and according to this scenarios different layout of placement of turbines are discussed for optimizing the energy from that end scenario. With wake loss & without wake loss analysis has been done. In this thesis two objectives has been discussed and they are concentrated on: Increase the power extracted from the wind. Reduce the vibration of the drive train system. Reduce the vibration of the tower. Use the maximum of the wind by turbine placement layouts. Using different layouts for different wind scenarios as per locations.

72

References
1. A. Kusiak and Z. Zhang, Analysis of Wind Turbine Vibrations Based on SCADA Data, ASME Journal of Solar Engineering, Vol. 132, No. 3, pp. 031008-1 031008-12, 2010a. 2. A. Fraser, Simulation of genetic systems by automatic digital computers. I. Introduction, Australian Journal of Biological Sciences, Vol. 10, pp. 484491, 1957. 3. AWAE, http://www.awea.org/; 2008. 4. Cai Z, Wang Y. A multiobjective optimization-based evolutionary algorithm for constrained optimization. IEEE Transactions on Evolutionary Computation 2006;10:65875. 5. C. Fang, J. Lee and M. A. Schilling, Balancing Exploration and Exploitation through Structural Design: The Isolation of Subgroups and Organizational Learning, Organization Science, Vol. 21, No. 3, pp. 625 642, 2010. 6. Crespo A, Hernandez J, Frandsen S. Survey of modeling methods for wind turbine wakes and wind farms. Wind Energy 1999;2:124. 7. gg[14] B. Boukhezzar and H. Siguerdidjane, Nonlinear control with wind estimation of a DFIG variable speed wind turbine for power capture optimization, Energy Conversion and Management, Vol. 50, No. 4, pp. 885-892, 2009. 8. B. Boukhezzar, H. Siguerdidjane and M. Hand, Nonlinear Control of VariableSpeed Wind Turbines for Generator Torque Limiting and Power Optimization, ASME Transactions: Journal of Solar Energy Engineering, vol. 128, pp. 516-530, 2006. 9. D. Bche, P. Stoll, R. Dornberger and P. Koumoutsakos, Multiobjective evolutionary algorithm for the optimization of noisy combustion processes, IEEE Transactions on Man, Control, and Cybernetics Part C, vol. 32, pp. 460-473, 2002. 10. Grady SA, Hussaini MY, Abdullah MM. Placement of wind turbines using genetic algorithms. Renewable Energy 20 05;30:25970. 11. http:// www.simplydecoded.com. 12. http://en.wikipedia.org/wiki/Wind_power_in_India 13. http://www.thehindu.com/todays-paper/wind-energy-gives-some-relief-frompower-cut/article4632307.ece?css=print

14. http://www.indiainfoline.com/Markets/News/Hindustan-Zinc-announces-150-MWexpansion-in-wind-power-generation-capacity/3473752590

15. http://www.renewableenergyworld.com 16. http://www.indianwindpower.com 17. http://www.cs.cmu.edu/~jgc/Presentations/WindPower5-2009.ppt

73

18. http://www.moneycontrol.com/news/business/hindustan-zinc-announces-150-mwexpansionswind-power-generation-capacity_511217.html

19. http://www.hindustantimes.com/business-news/CorporateNews/CLP-power-Indiato-invest-Rs-1-800-cr-in-wind-energy/Article1-837514.aspx

20. http://www.hindustantimes.com/business-news/WorldEconomy/Wind-powerbase-to-rise-50-by-2015/Article1-1033842.aspx

21. http://www.sciencedaily.com/releases/2010/11/101123174322.htm 22. http://www.sciencedirect.com/science/article/pii/S0924013605005856 23. http://europepmc.org/abstract/MED/9268969/reload=0;jsessionid=6k0kjUjDp3Qim

uVKdkXu.28

24. Jensen NO. A note on wind generator interaction. Roskilde, Denmark: Riso National Laboratory; 1983. Technical Report Riso-M-2411. 25. L. Breiman, "Random Forests," Machine Learning, Vol. 45, No. 1, pp. 5-32, 2001. 26. Lackner MA, Elkinton CN. An analytical framework for offshore wind farm layout optimization. Wind Engineering 2007;31:1731. 27. Lange B, Waldl H, Guerrero AG, Heinemann D. Modeling of offshore wind turbine wakes with the wind farm program FLaP. Wind Energy 2003;6:87 104. 28. Manwell JF, McGowan JG, Rogers AL. Wind energy explained: theory, design and application. 1st ed. London: John Wiley & Sons; 2002. 29. M. H. Hansen, Improved modal dynamics of wind turbines to avoid stallinduced vibrations, Wind Energy, Vol. 6, No. 2, pp. 179-195, 2003. 30. Mosetti G, Poloni C, Diviacco B. Optimization of wind turbine positioning in large wind farms by means of a genetic algorithm. Journal of Wind Engineering and Industrial Aerodynamics 1994;51:10516. 31. Neustadter HE. Method for evaluating wind turbine wake effects on wind farm performance. Journal of Solar Energy Engineering 1985;107:2403. 32. P. Flores, A. Tapia and G. Tapia, Application of a control algorithm for wind speed prediction and active power generation, Renewable Energy, vol. 30, pp. 523-536,2005. 33. S. Forrest, B. Javornik, R. E. Smith and A. S. Perelson, Using Genetic Algorithms to Explore Pattern Recognition in the Immune System, Evolutionary Computation, Vol. 1, No. 3, pp. 191-211, 1993. 34. Stroock DW. A concise introduction to the theory of integration. 3rd ed. Birkhauser; 1998. 35. Walford CA. Wind turbine reliability: understanding and minimizing wind turbine operation and maintenance costs. Available from:. Albuquerque, New Mexico: Sandia National Laboratories www.prod.sandia.gov/cgibin/techlib/access-control.pl/20 06/06110 0.pdf; 2006.

74

36. Wiser R, Bolinger M. Annual report on U.S. wind power installation, cost, and performance trends: 2006. Available from. Golden, CO: NREL, US Department of Energy, http://www.nrel.gov/wind/pdfs/41435.pdf; 2007. 37. Zitzler E, Thiele L. Multiobjective evolutionary algorithms: a comparative case study and the strength Pareto approach. IEEE Transactions on Evolutionary Computation 1999;3:25771. 38. Eiben AE, Smith JE. Introduction to evolutionary computation. New York: Springer-Verlag; 2003. 39. Roger H, Charles J. Topics in Matrix Analysis. Cambridge; 1994. 40. Castro Mora J, Calero Baron JM, Riquelme Santos JM, Burgos Payan M. An evolutive algorithm for wind farm optimal design. Neurocomputing 2007;70:26518. 41. WASP, http://www.wasp.dk/; 2009

75