Upload
uiowa
View
1
Download
0
Embed Size (px)
Citation preview
lable at ScienceDirect
ARTICLE IN PRESS
Renewable Energy xxx (2009) 1–8
Contents lists avai
Renewable Energy
journal homepage: www.elsevier .com/locate/renene
Dynamic control of wind turbines
Andrew Kusiak*, Wenyan Li, Zhe SongDepartment of Mechanical and Industrial Engineering, 3131 Seamans Center, The University of Iowa, Iowa City, IA 52242-1527, USA
a r t i c l e i n f o
Article history:Received 8 April 2009Accepted 25 May 2009Available online xxx
Keywords:Wind turbineWind energyData miningModel predictive controlEvolutionary computation algorithmControl strategy optimization
* Corresponding author.E-mail address: [email protected] (A. Ku
0960-1481/$ – see front matter Published by Elsevierdoi:10.1016/j.renene.2009.05.022
Please cite this article in press as: Kusj.renene.2009.05.022
a b s t r a c t
The paper presents an intelligent wind turbine control system based on models integrating the followingthree approaches: data mining, model predictive control, and evolutionary computation. To enhance thecontrol strategy of the intelligent system, a multi-objective model is proposed. The model involves fivedifferent objectives with different weights controlling the wind turbine performance. These weights areadjusted in response to the variable wind conditions and operational requirements. Three control factors,wind speed, turbulence intensity, and electricity demand are considered in eight computationalscenarios. The performance of each scenario is illustrated with numerical results.
Published by Elsevier Ltd.
1. Introduction
Reduction of operations and maintenance costs [23,24] is a key tothe expansion of wind industry destined to grow in the next decades[28]. Development of control solutions [2,14,22,27] is a validapproach to reduce these costs. Awell designed wind turbine controlsystem should maximize not only the energy captured from thewind but also extend the lifetime of turbine components, e.g., thegearbox. It is known that smoothing the turbine power output isimportant in its integration with the electricity grid. All operationsand maintenance considerations have to be properly managed. Forexample, for a high wind speed and low electricity demand, a windturbine can be operated so that smoothing the rotor speed andgenerator torque becomes a priority. Wind turbine control tech-nology is relatively new, and opportunities exist to improve turbineperformance in the presence of operations and maintenanceconstraints. An intelligent wind turbine control system increasingcompetitiveness of wind energy is needed.
Most literature on wind turbine control have focused on maxi-mizing power [2,4,15,17,18,27] for wind speed in the range betweenthe cut-in and the cut-out wind speed. This goal is usually achievedby controlling the generator torque so that the rotor speedproducing the optimum power coefficient is attained. The windpower is maximized predominantly when the wind speed is belowits rated value and the blade pitch angle is fixed. Besides the
siak).
Ltd.
iak A, et al., Dynamic co
traditional control strategies (i.e., mainly feedback and adaptive-tracking based), predictive control [11,16,19] has been used tooptimize the capture of the wind power. Henriksen [11] investi-gated a model predictive control approach to pitch controllerdesign [3,21]. The model predictive control approach with bladepitch and generator torque as two control inputs was discussed inRef. [16]. The research reported in Refs. [11,16] was based onsimulated wind speeds in a reactive rather than predictive mode.Unlike traditional energy conversion systems, where the fuel inputcan be controlled, the speed of the wind cannot be controlled.However, knowing the wind speed ahead of time is useful incontrolling a wind turbine. Wind speed prediction was consideredin Ref. [19], where a linear wind speed time series model wasapplied to determine wind speed in a short-time horizon (i.e.,seconds). The wind turbine power output was assumed to bea linear function of the wind speed [19]. Thus, knowing wind speedahead of time would allow smooth power generation.
In this paper, an intelligent system for control of wind turbines isintroduced. Unlike the research reported in the literature[2,4,11,14–19,22,27], where wind turbine models are obtained fromthe first principles and aerodynamics, in the proposed intelligentcontrol system, data mining algorithms extract the turbine modelsfrom the process data, i.e., SCADA (Supervisory Control and DataAcquisition) system [26]. A time series model [6,13,20] is used topredict wind speed, and MPC (Model Predictive Control) optimizesthe process variables of wind turbines.
The model considered in this paper considers five weightedobjectives. The weights are adjusted according to eight typicalscenarios defined by wind conditions and operational requirements.
ntrol of wind turbines, Renewable Energy (2009), doi:10.1016/
0
200
400
600
800
1000
1200
1400
1600
2 4 6 8 10 12 14
Pow
er (
kW)
Wind speed (m/s)
10-second average
Fig. 1. Power curve of a 1.5 MW wind turbine for 10-s average data.
0
200
400
600
800
1000
1200
1400
1600
2 3 4 5 6 7 8 9 10 11 12
Pow
er (
kW)
Wind speed (m/s)
10-minute average
Fig. 3. Power curve of a 1.5 MW wind turbine for the 10-min average data.
A. Kusiak et al. / Renewable Energy xxx (2009) 1–82
ARTICLE IN PRESS
2. Data description
Modern wind turbines are equipped with sensors for controland monitoring purposes. The data sampling frequency can behigh (e.g., milliseconds), however, for specific applications, thehigh frequency data is usually aggregated (e.g., averaged) overa certain time period (e.g., 10 s, 10 min). At present, the 10-mindata standard is widely used in industry. Analysis of 10-min datadoes not allow observing important details. Figs. 1–3 show thepower curves of a 1.5 MW wind turbine when it is operatingbetween the cut-in and the rated wind speed. The power curveshown in Fig. 1 is scattered though the status of the wind turbineis considered to be normal. It can be observed that the spread ofthe points which create the power curve increases around therated wind speed. This increased variability translates into vari-able power output and results in loads that could be hazardous tothe drive train components.
Averaging the 10-s data over the 1-min data has resulted ina power curve with a reduced spread of the data points (see Fig. 2).For the 10-min average data (the industry standard), the datapoints follow a typically displayed pattern (see Fig. 3). Note that thescattered graph in Fig. 3 is due to insufficient number of data pointsincluded in the 10-s data set.
The results presented in Figs. 1–3 indicate that the wind turbinecontrol system considered in this research needs to be improved.An intelligent wind turbine control system is discussed and illus-trated with the data from a 1.5 MW wind turbine. Table 1 lists theprocess parameters used in the study.
Two major controllable parameters (i.e., the parameters opti-mizing the energy conversion process) are the blade pitch angleand the generator torque. In addition, three response parametersof interest to this research are the wind speed, rotor speed, and
0
200
400
600
800
1000
1200
1400
1600
2 4 6 8 10 12 14
Pow
er (
kW)
Wind speed (m/s)
1-minute average
Fig. 2. Power curve of a 1.5 MW wind turbine for 1-min average data.
Please cite this article in press as: Kusiak A, et al., Dynamic coj.renene.2009.05.022
wind turbine power output. They are important indicators of theenergy conversion process. The rotor speed and the generatorspeed are highly correlated, and in fact they are modeled bya linear function. In this paper, the rotor speed is selected asa response parameter to be included in an objective function.Rapid changes in the rotor speed accelerate the failure of itsmechanical components.
The data used for this case study was collected at a samplinginterval of 10 s for a day. This data set satisfactorily represents thedata generated at different turbines across many time horizons.The wind speed varied in the interval [2.97 m/s, 13.16 m/s]. Afterinitial denoising (e.g., removing turbine down time, wind speedbelow the cut-in speed, and the wind above the cut-out speed),2054 valid data points were considered. The data for the windspeed between the cut-in and the cut-out speed ranges wasconsidered in particular because this operational region presentsa major opportunity to optimize the wind turbine power gener-ation process. The distribution of data used in this paper is shownin Fig. 4.
3. Data driven intelligent wind turbine control system
In this paper, an intelligent wind turbine control system isdeveloped using data mining algorithms to optimize the windturbine energy conversion process when the wind speed isbetween the cut-in and the rated speed. Fig. 5 shows the basiccomponents (two modules) of the control system and its infor-mation flow.
The first module extracts the process and the time series modelsfrom the historical SCADA and wind data. These models areupdated once their performance degrades, which is accomplishedby comparing the model’s predicted values with the actualmeasurements recorded by SCADA and anemometers.
A time series model predicts wind speed at short-time intervals,which can be traced back to Ref. [19], where a linear wind speedtime series model was built. Once an accurate wind speed timeseries model is obtained, wind turbine is optimized with a modelpredictive control algorithm. Although several different algorithms
Table 1Wind turbine parameters used in the case study.
Parameter Name Unit
v Wind speed m/s�
x1 Blade pitch angle �
x2 Generator torque Nmy1 Wind turbine power output kWy2 Rotor speed rpm
ntrol of wind turbines, Renewable Energy (2009), doi:10.1016/
2.97 3.68 4.40 5.11 5.82 6.54 7.25 7.96 8.68 9.39 10.1010.8211.5312.2412.96
WindSpeed
0
10
20
30
40
50
60
70
80
No
of o
bs
Fig. 4. Wind speed distribution of 2054 data points.
0.000 0.069 0.137 0.206 0.274 0.343 0.412 0.480 0.549 0.617 0.6860
200
400
600
800
1000
1200
1400
No
of o
bs
Fig. 6. Turbulence intensity distribution (1-min time intervals).
A. Kusiak et al. / Renewable Energy xxx (2009) 1–8 3
ARTICLE IN PRESS
are compared, such as the boosting tree regression [8,9], linearregression, and neural network [1,10] the wind speed time seriesmodel (1) is built by a neural network ensemble, and the predictoris selected using the genetic wrapper approach [25,29].
vðtÞ ¼ gðvðt � 1Þ; vðt � 2Þ; vðt � 6ÞÞ (1)
The process models (2) and (3) are built by a neural networkalgorithm [7,12]. Since the sampling rate is 10 s, and the time delayof the wind turbine control system is not longer than 20 s, oneprevious state is sufficient in building models.
y1ðtÞ ¼ f1ðy1ðt � 1Þ; x1ðtÞ; x1ðt � 1Þ; x2ðtÞ; x2ðt � 1Þ;vðtÞ; vðt � 1ÞÞ (2)
y2ðtÞ ¼ f2ðy2ðt � 1Þ; x1ðtÞ; x1ðt � 1Þ; x2ðtÞ; x2ðt � 1Þ;vðtÞ; vðt � 1ÞÞ (3)
The second module includes the MPC (model predictive control)component and an EC (evolutionary computation) solver. The MPCmodel is formulated based on the extracted models and the current
Wind turbine Evolutionary
Data miningalgorithm
Identified windturbine model
SCA
DA
dat
a
Optimal settings ofpitch angle andgenerator torquecontrol
Update theprocess model
when notaccurate enough
MPC form
Current wind turbinestatus
Database
Historicaltraining data
Fig. 5. Intelligent control sy
Please cite this article in press as: Kusiak A, et al., Dynamic coj.renene.2009.05.022
wind turbine status as well as the wind speed. Then current windconditions and operational requirements are used to determine theimportance of the optimization objectives of the MPC model (hereweights associated with the MPC objective functions). Updatingprocess models is an important issue in MPC control; however, thispaper does not focus on this topic. Rather it emphasizes identifi-cation of various optimization objects according to different windconditions and operational requirements.
Control systems in the presently available wind turbines usuallyfollow one control strategy, i.e., when the wind speed is betweenthe cut-in and rated one, the power output is optimized byfollowing the maximum theoretic optimal power coefficient.Actually there are other factors to be considered besides maximi-zation of the power output. For example, smoothing the poweroutput or rotor speed variation is important depending on the windconditions and operational demands. An intelligent control systemshould provide more options to control a wind turbine based on thewind conditions and operational requirements. Five differentobjectives are considered in optimizing the energy conversionprocess as shown in (4).
Current wind conditionsand operationalrequirements
algorithm
Wind dataData mining
algorithm
Identified windtime series model
ulation
Update the timeseries modelwhen notaccurate enough
Historicaltraining data
stem of wind turbine.
ntrol of wind turbines, Renewable Energy (2009), doi:10.1016/
m
Table 2Classification of four wind speed scenarios.
ScenarioNumber
Wind/turbulence
Time stamp Wind speed characteristic Turbulence Intensity Characteristic
Min windspeed [m/s]
Ave windspeed [m/s]
Max windspeed [m/s]
Percentage ofwind speedhigher than7 m/s
Minturbulenceintensity
Aveturbulenceintensity
Maxturbulenceintensity
Percentage ofturbulenceintensityhigher than0.06
1 High/High ‘‘2:14:40 PM’’to ‘‘2:21:20PM’’
7.43 9.63 12.22 100.00% 0.03 0.06 0.16 54.84%
2 High /Low ‘‘1:05:40 PM’’to ‘‘1:10:30PM’’
7.02 9.58 11.66 100.00% 0.02 0.05 0.09 43.33%
3 Low/High 12:30:10 PM’’to ‘‘12:35:00PM’’
3.75 5.88 8.47 17.24% 0.04 0.16 0.32 89.66%
4 Low/Low 4:36:10 AM’’to ‘‘4:41:00AM’’
5.29 6.11 6.70 0.00% 0.01 0.03 0.06 4.17%
A. Kusiak et al. / Renewable Energy xxx (2009) 1–84
ARTICLE IN PRESS
J ¼w1JPower þ w2JRotor þ w3JP ramp þ w4JG ramp
þ w5JPitch ramp ð4Þ
where: JPower is a function to minimize the distance between thepower output to its upper limit and therefore maximizing thepower output, JRotor is a function to minimize rotor speed ramp,JP_ramp is a function to minimize power output ramp, JG_ramp isa function to minimize generation torque ramp, JPitch_ramp isa function to minimize pitch angle ramp.
The weights w1, w2, w3, w4, w5 in (4) are assigned differentvalues depending on the priority assigned to the correspondingobjective. Each weight is in the interval [0, 1], andw1þw2þw3þw4þw5¼1. The MPC model is defined in (5):
in J�w1; w2; w3; w4; w5; JPower; JRotor; JP ramp;
JG ramp; JPitch ramp�
(5)
It is worth mentioning that minimizing J(w1, w2, w3, w4, w5, JPower,JRotor, JP_ramp, JG_ramp, JPitch_ramp) is a challenge as the objective func-tion and the constraints are nonlinear. To solve this type of optimi-zation problem, an evolutionary strategy algorithm is proposed [5].However, this paper focuses on analysis of the weights for differentscenarios rather than the solution of the function.
4. Adjusting objectives based on wind conditionsand operational requirements
In this paper, the following three factors, wind speed, windturbulence, and electricity demand, are considered in deriving thecontrol priorities (i.e., the weights w1, w2, w3, w4, w5).
Table 3Scenario classification according to wind status and electricity demand.
Scenario number Wind speed Turbulence intensity
1 High (7w12.5 m/s) High (I> 0.06)23 Low (I< 0.06)4
5 Low (3.5w7 m/s) High (I> 0.06)67 Low (I< 0.06)8
Please cite this article in press as: Kusiak A, et al., Dynamic coj.renene.2009.05.022
Wind turbulence is an important metric describing the degreeof variability of the wind speed. If the current wind speed is highlyvariable, maximizing the power capture may significantly damagethe turbine‘s mechanical components. An intelligent wind turbinecontrol system should optimize the control priorities based onwind turbulence.
4.1. Wind turbulence intensity
Turbulence intensity is a measure of the overall level of windturbulence [30], and it is defined in (6).
It ¼st
vt(6)
where st is the standard deviation of the wind speed variationabout the mean wind speed vt at time stamp t, and vt is the meanwind speed over a certain interval at time t.
The mean wind speed vt over n consecutive sampling datapoints is defined in (7):
vt ¼1n
Xt
i¼ t�nþ1
vðiÞ (7)
where n is the number of consecutive data points.The standard deviation st is computed as follows (see (8)):
st ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
n� 1
Xt
i¼ t�nþ1
ðvari � var
vuut Þ2 (8)
Demand Weight
w1 w2 w3 w4 w5
High 0.4 0.15 0.2 0.2 0.05Low 0.2 0.25 0.25 0.25 0.05High 0.6 0.1 0.15 0.15 0Low 0.1 0.25 0.25 0.25 0.15
High 0.55 0.15 0.15 0.15 0Low 0.2 0.25 0.25 0.25 0.05High 0.6 0.1 0.1 0.1 0.1Low 0.2 0.2 0.2 0.2 0.2
ntrol of wind turbines, Renewable Energy (2009), doi:10.1016/
Table 4Summary of power output generation.
Scenarionumber
Original average poweroutput [kW]
Optimized average poweroutput [kW]
Original STD of poweroutput [kW]
Optimized STD of poweroutput [kW]
1 795.01 1278.36 171.99 153.802 1098.36 87.243 907.60 1427.70 228.90 125.614 554.90 55.695 200.40 372.62 139.82 226.936 271.72 153.237 275.00 535.55 36.97 64.588 498.89 62.51
A. Kusiak et al. / Renewable Energy xxx (2009) 1–8 5
ARTICLE IN PRESS
where vart is the variation of wind speed at time t, and is computedfrom (9):
vart ¼ vðtÞ � vt (9)
where vart is the mean wind speed variation at time t as shown in(10):
vart ¼1n
Xt
i¼ t�nþ1
vari (10)
The wind turbulence intensity (6) implies that current windturbulence intensity at time t is estimated based on historical datapoints. Alternatively, the wind turbulence intensity could becomputed based on the data obtained from the time seriesprediction model.
In this paper, the turbulence intensity is computed over a 1-mintime horizon (i.e., 6 consecutive 10 s data points) or a 5-min timehorizon (i.e., 30 consecutive 10 s data points).
Fig. 6 shows the distribution of turbulence intensity computedover a 1-min time interval. According to Fig. 6, It is generally smallerthan 0.274 most of the time, and the mode of It is around 0.069.
In comparing the distributions of turbulence intensity calcu-lated from one minute to 5 min time intervals, the common trend isobvious. For example, most turbulence intensities are between0.03w0.36, and the highest frequency occurs around 0.1, thoughthere are some small deviations.
For an intelligent wind turbine control system, 1-min datareflects resolution, and it is suitable to classify the current turbu-lence status. In this paper It¼ 0.06 is used as the threshold todistinguish between high turbulence intensity and low turbulenceintensity. Note that this threshold value was established based onthe specific data sets. More research could be conducted in thefuture to find a better classification rule.
4.2. Wind conditions
Wind speed is an important factor to be considered by theintelligent control system. For control purposes, three regions areusually considered [27]:
Table 5Summary of standard deviations for four parameters.
Scenarionumber
Original STD of Rotorspeed [rpm]
Optimized STD of Rotorspeed [rpm]
Original STD of bladepitch Angle [�]
1 0.25 0.24 4.392 0.103 0.37 0.27 3.574 0.095 1.79 1.69 4.096 1.397 0.82 0.30 0.008 0.40
Please cite this article in press as: Kusiak A, et al., Dynamic coj.renene.2009.05.022
(1) Wind speed smaller than the cut-in speed.(2) Wind speed larger the rated speed.(3) Wind speed that is between cut-in and the rated speed.
To optimize the wind turbine operations, this paper suggests furtherclassification of the wind speed region between the cut-in and the ratedwind speed. In this way, more precise control strategies can be imple-mented. The data used in this paper is collected from a 1.5 MW turbinewith cut-in speed of 3.5 m/s and rated speed of 12.5 m/s. The speed of7 m/s is used as an additional threshold category (see Fig. 4) to definethe following four wind speed scenarios shown in Table 2.
Table 2 illustrates the classification of wind conditions according towind speed and turbulence intensity. The threshold of wind speed is7 m/s and the threshold of turbulence intensity is 0.06. Considerscenario 1 for example, as all the wind speed is higher than 7 m/s, thisperiod of wind conditions is defined as high wind speed period. As54.84% turbulence intensity is higher than 0.06 and the maximum isaround 0.16, the scenario is also defined as high turbulence intensity.Similarity, scenario 2 is characterized as high wind speed and lowturbulence intensity. Scenario 3 is characterized as low wind speed andhigh turbulence intensity and Scenario 4 is low wind speed and lowturbulence intensity.
4.3. Electricity demand
Electricity demand is important in turbine control becausewhen it is low, there is no need to maximize the power output, andmore attention should be given to smoothing the power output. Onthe other hand, if the demand is high, more emphasis should begiven to maximizing the power output. In this paper electricitydemand (i.e., low or high) is also considered as a factor in classifyingthe operational scenarios.
4.4. Classification of operational scenarios
Eight operational scenarios are discussed in this paper accordingto the turbulence intensity, wind speed, and electricity demand.Table 2 shows the details of various scenarios and weights used inmodel (4). Each scenario represents a combination of weights,
Optimized STD of bladepitch angle [�]
Original STD of generatortorque [Nm]
Optimized STD of generatortorque [Nm]
1.91 1152.89 881.822.47 563.293.96 1470.40 747.912.15 400.584.54 934.38 1614.484.34 1142.182.36 207.19 438.800.72 446.31
ntrol of wind turbines, Renewable Energy (2009), doi:10.1016/
400
600
800
1000
1200
1400
1:05:40 PM 1:06:40 PM 1:07:40 PM 1:08:40 PM 1:09:40 PM
Pow
er O
utpu
t (k
W)
OptimizedOriginal
Fig. 7. Original and optimized power output in the interval ‘‘1:05:40 PM’’ to ‘‘1:10:30PM’’.
17.5
18.0
18.5
19.0
19.5
20.0
1:05:40 PM 1:06:40 PM 1:07:40 PM 1:08:40 PM 1:09:40 PM
Rot
or s
peed
(rp
m)
OptimizedOriginal
Fig. 8. Original and optimized rotor speed in the interval ‘‘1:05:40 PM’’ to ‘‘1:10:30PM’’.
-2
0
2
4
6
8
10
12
1:05:40 PM 1:06:40 PM 1:07:40 PM 1:08:40 PM 1:09:40 PM
Pit
ch a
ngle
(de
gree
)
OptimizedOriginal
Fig. 10. Original and optimized pitch angle in the interval ‘‘1:05:40 PM’’ to ‘‘1:10:30PM’’.
200
250
300
350
400
450
500
550
600
650
4:36:20 AM 4:37:10 AM 4:38:00 AM 4:38:50 AM 4:39:40 AM
Pow
er O
utpu
t (k
W)
Optimized
Original
Fig. 11. Original and optimized power output in the interval ‘‘4:36:20 AM’’ to ‘‘4:40:10AM’’.
A. Kusiak et al. / Renewable Energy xxx (2009) 1–86
ARTICLE IN PRESS
which differentiate the importance of the five objectives,wherew1þw2þw3þw4þw5¼1 and w1, w2, w3, w4, w5 arebetween 0 and 1 (see Table 3). The weight combinations are derivedbased on the heuristic domain knowledge. More research is neededto algorithmically generate these weights.
4.5. Computational results
The computational results reported in this section are based onthe dynamic equations extracted by the neural network algorithmand the wind speed time series model built from a neural networkensemble. The MPC model (5) is solved by an evolutionary strategyalgorithm for constrained optimization problems with certain fixedparameter settings (i.e., the population size, selection pressure). Foreach scenario, during a fixed time period, model (5) is solved to find
3000
4000
5000
6000
7000
8000
9000
1:05:40 PM 1:06:40 PM 1:07:40 PM 1:08:40 PM 1:09:40 PM
Gen
erat
orto
rque
(N
m)
OptimizedOriginal
Fig. 9. Original and optimized generator torque in the interval ‘‘1:05:40 PM’’ to‘‘1:10:30 PM’’.
Please cite this article in press as: Kusiak A, et al., Dynamic coj.renene.2009.05.022
the optimal pitch angle and generator torque settings for thestarting time stamp t. For the next sampling time tþ 1, model (5) issolved again based on the previously found optimal control settingsat time t (i.e., pitch angle and generator torque). This simulationcontinues until the fixed time period ends. Then the optimizedwind turbine status is compared with the original one for that fixedtime period.
4.6. Summary of optimization results
Table 4 illustrates the optimization results of power output forcomputational eight scenarios. The power output has increased ineach of the scenarios listed in Table 4, except scenario 4 (shown inbold), caused by the low electricity demand. STD (Standard
16
17
18
19
20
21
4:36:20 AM 4:37:10 AM 4:38:00 AM 4:38:50 AM 4:39:40 AM
Rot
or s
peed
(R
PM
)
OptimizedOriginal
Fig. 12. Original and optimized rotor speed in the interval ‘‘4:36:20 AM’’ to ‘‘4:40:10AM’’.
ntrol of wind turbines, Renewable Energy (2009), doi:10.1016/
1000
2000
3000
4000
5000
4:36:20 AM 4:37:10 AM 4:38:00 AM 4:38:50 AM 4:39:40 AM
Gen
erat
or t
orqu
e (N
m)
OptimizedOriginal
Fig. 13. Original and optimized generator torque in the interval ‘‘4:36:20 AM’’ to‘‘4:40:10 AM’’.
A. Kusiak et al. / Renewable Energy xxx (2009) 1–8 7
ARTICLE IN PRESS
Deviation) of power output implies the quality and smoothness ofpower output. Five optimized scenarios have smaller values of STDof the power output than the original ones. In scenarios 5, 7 and 8 ofTable 4 the power quality is diminished in order to satisfy theelectricity demand (shown in bold).
Table 5 illustrates the smoothness of rotor speed, blade pitchangle, and generator torque. Optimization has resulted in smoothervalues of rotor speed in eight scenarios. Comparing with the orig-inal blade pitch angle and generator torque, the optimized valuesare smoother in high wind speed scenarios (1–4). At low windspeed (scenarios 5 through 8 of Table 5), the smoothness of bladepitch angle and generator torque (shown in bold) has diminished tobenefit the power output.
The details of two illustrative operational scenarios 4 and 7 ofTable 5 are discussed next.
4.7. Operational scenario 4: high wind speed, low turbulenceintensity, and low electricity demand
Scenario 4 is concerned with high wind speed, low turbulenceintensity and low electricity demand. As electricity is demand low,w1is set as 0.1,w2, w3 and w4 are set as 0.25, w5is set as 0.15 (seeTable 3). In this case, the optimized power output is lower but muchsmoother than the original one. Due to low electricity demand, theincrease of generation is diminished to improve the power quality.In the same way, rotor speed, generator torque and blade pitchangle are smoothed after optimization. Figs. 7–10 show the resultsof optimization of the power output, rotor speed, generator torque,and pitch angle.
0
2
4
6
8
10
4:36:20 AM 4:37:10 AM 4:38:00 AM 4:38:50 AM 4:39:40 AM
Pit
ch a
ngle
(de
gree
)
Optimized
Original
Fig. 14. Original and optimized pitch angle in the interval ‘‘4:36:20 AM’’ to ‘‘4:40:10AM’’.
Please cite this article in press as: Kusiak A, et al., Dynamic coj.renene.2009.05.022
4.8. Operational scenario 7: Low wind speed, low turbulenceintensity, and high electricity demand
Here is the situation of low wind speed, low turbulence inten-sity and high electricity demand. As wind speed is low, the originallow power output cannot satisfy the high electricity demand. Sohigher w1is needed. In this case, w1 is set as 0.6,w2, w3, w4 and w5
are set to 0.1. Figs. 11–14 show the results of optimization of poweroutput, rotor speed, pitch angle, and generator torque.
As presented in Figs. 11–14, the generated power has increasedat the expense of its quality to satisfy the electricity demand. Therotor runs smoother after optimization. The original generatortorque does not vary much and the original blade pitch angle isconstant for the range of wind speeds considered in this scenario.The optimized blade pitch angle and generator torque have resul-ted in increased the power output.
5. Conclusion
In this paper, an intelligent system for control of wind turbineswas presented. The system integrated data mining, evolutionarycomputation, predictive control, and time series approaches. A timeseries model for prediction of wind speed was proposed. A dynamicfunction was built with five different weights determined byvarious operational scenarios. The system modifies the controlobjectives by observing the wind conditions and electricitydemand. Wind conditions were characterized by the wind speedand wind turbulence intensity. Eight scenarios were defined basedon the wind conditions and the electricity demand.
Each scenario was illustrated with computational results. Theoriginal wind turbine status and the optimized ones werecompared by simulation. The results produced by the intelligentcontrol system are better than those of the current wind turbinecontrol system. For turbulent wind, the intelligent control systemsmoothed the power output, generator torque, and rotor speedwithout compromising the electricity demand.
Further research should focus on automating the weightgeneration based on wind conditions and electricity demand. Otherobjectives could be considered by the predictive control model toexplore other aspects of the wind energy conversion process.Additional research is needed to improve the prediction accuracy ofthe wind speed, which is of importance in the proposed approach.Denoising techniques could be applied to enhance the data quality.
Acknowledgments
This research has been sponsored by the Iowa Energy Center,Grant No. IEC-07-01.
References
[1] Bishop C. Neural networks for pattern recognition. Oxford: University Press;1995.
[2] Boukhezzar B, Siguerdidjane H, Maureen Hand M. Nonlinear control ofvariable-speed wind turbines for generator torque limiting and poweroptimization. ASME Transactions: Journal of Solar Energy Engineering2006;128(4):516–30.
[3] Camacho EF, Bordons C. Model predictive control. London, UK: Springer; 1999.[4] Datta R, Ranganathan VT. A method of tracking the peak power points for
a variable speed wind energy conversion system. IEEE Transactions on EnergyConversion 2003;1(1):163–8.
[5] Eiben AE, Smith JE. Introduction to evolutionary computation. New York:Springer; 2003.
[6] Ernst B, Oakleaf B, Ahlstrom ML, Lange M, Moehrlen C, Lange B, et al. Pre-dicting the wind. IEEE Power & Energy Magazine 2007;5:78–89.
[7] Espinosa J, Vandewalle J, Wertz V. Fuzzy logic, identification and predictivecontrol. London, UK: Springer; 2005.
[8] Friedman JH. Stochastic gradient boosting. Computational Statistics & DataAnalysis 2002;38(4):367–78.
ntrol of wind turbines, Renewable Energy (2009), doi:10.1016/
A. Kusiak et al. / Renewable Energy xxx (2009) 1–88
ARTICLE IN PRESS
[9] Friedman JH. Greedy function approximation: a gradient boosting machine.Annals of Statistics 2001;29(5):1189–232.
[10] Haykin S. Neural networks: a comprehensive foundation. New York: Mac-millan Publishing; 1994.
[11] Henriksen LC. Model predictive control of a wind turbine; 2007.[12] Kelouwani S, Agbossou K. Nonlinear model identification of wind turbine with
a neural network. IEEE Transactions on Energy Conversion 2004;19(3):607–12.
[13] Ma L, Luan S, Jiang C, Liu H, Zhang Y. A review on the forecasting of windspeed and generated power. Renewable and Sustainable Energy Reviews2009;13(4):915–20.
[14] Manwell JF, McGowan JG, Rogers AL. Wind energy explained: theory, designand application. 1st ed. London, UK: John Wiley; 2002.
[15] Morimoto S, Nakayama H, Sanada M, Takeda Y. Sensorless output maximiza-tion control for variable-speed wind generation system using IPMSG. IEEETransactions on Industry Applications 2005;41(1):60–7.
[16] Muhando EB, Senjyu T, Urasaki N, Yona A, Funabashi T. Robust predictivecontrol of variable-speed wind turbine generator by self-tuning regulator. In:IEEE Power Engineering Society General Meeting; 2007. p. 1–8.
[17] Muljadi E, Butterfield CP. Pitch-controlled variable-speed wind turbinegeneration. IEEE Transactions on Industry Applications 2001;37(1):240–6.
[18] Munteanu I, Cutululis NA, Bratcu AI, Ceanga E. Optimization of variable speedwind power systems based on a LQG approach. Control Engineering Practice2005;13:903–12.
[19] Nanayakkara N, Nakamura M, Hatazaki H. Predictive control of wind turbinesin small power systems at high turbulent wind speeds. Control EngineeringPractice 1997;5(8):1063–9.
Please cite this article in press as: Kusiak A, et al., Dynamic coj.renene.2009.05.022
[20] Riahy GH, Abedi M. Short term wind speed forecasting for wind turbineapplications using linear prediction method. Renewable Energy 2008;33(1):35–41.
[21] Rossiter JA. Model-based predictive control: a practical approach. New York:CRC Press; 2003.
[22] Song YD, Dhinakaran B, Bao XY. Variable speed control of wind turbines usingnonlinear and adaptive algorithms. Journal of Wind Engineering and Indus-trial Aerodynamics 2000;5(3):293–308.
[23] Walford CA. Wind turbine reliability: understanding and minimizing windturbine operation and maintenance costs. Available:. Albuquerque, NM: San-dia National Laboratory www.prod.sandia.gov/cgi-bin/techlib/access-control.pl/2006/061100.pdf; 2006.
[24] Wiser R, Bolinger M. Annual report on U.S. wind power installation, cost, andperformance trends: 2006. Available:. Golden, CO: NREL, US Department ofEnergy http://www.nrel.gov/wind/pdfs/41435.pdf; 2007.
[25] Witten IH, Frank E. Data mining: practical machine learning tools and tech-niques. 2nd Ed. San Francisco: Morgan Kaufmann; 2005.
[26] Zhu YC. Multivariable system identification for process control. New York:Pergamon Press; 2001.
[27] Johnson KE, Pao LY, Balas MJ, Fingersh LJ. Control of variable-speed windturbines: standard and adaptive techniques for maximizing energy capture.IEEE Control Systems Magazine 2006;26(3):70–81.
[28] http://www.awea.org.[29] Kohavi R, John GH. Wrapper for feature subset selection. Artificial Intelligence
1997;97(1–2):273–324.[30] Burton T, Sharpe D, Jenkins N, Bossanyi E. Wind energy handbook. New York:
Wiley; 2001.
ntrol of wind turbines, Renewable Energy (2009), doi:10.1016/