Adaptive Control of a Wind Turbine With Data Mining and Swarm Intelligence by n.rajbharath


									28                                                                                  IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 2, NO. 1, JANUARY 2011

        Adaptive Control of a Wind Turbine With Data
               Mining and Swarm Intelligence
                              Andrew Kusiak, Member, IEEE, and Zijun Zhang, Student Member, IEEE

    Abstract—The framework of adaptive control applied to a wind                    [13]. Munteanu et al. [14] designed two control loops to opti-
turbine is presented. The wind turbine is adaptively controlled to                  mize power at low frequency and high frequency scenarios [14].
achieve a balance between two objectives, power maximization and                    However, in the previous research, the investigation of another
minimization of the generator torque ramp rate. An optimization
model is developed and solved with a linear weighted objective. The                 beneficial avenue to reduce the cost of producing wind power,
objective weights are autonomously adjusted based on the demand                     reducing the cost of wind turbine operation and maintenance, is
data and the predicted power production. Two simulation models                      limited relative to power optimization. Senjyu et al. [15] studied
are established to generate demand information. The wind power                      the impact of limited activation of the blade pitch angle on the
is predicted by a data-driven time-series model utilizing historical                power output [15]. Kusiak et al. [16] developed a framework
wind speed and generated power data. The power generated from
the wind turbine is estimated by another model. Due to the intrinsic                of anticipatory control for optimization of wind turbine perfor-
properties of the data-driven model and changing weights of the                     mance expressed in four different metrics [16]. Although the
objective function, a particle swarm fuzzy algorithm is used to solve               published literature offers valuable insights into performance
it.                                                                                 optimization of wind turbines by considering multiple objec-
   Index Terms—Adaptive control, blade pitch angle, data mining,                    tives, determining the importance of each objective in response
electricity demand simulation, generator torque, neural networks,                   to changing wind conditions is an open issue.
optimization, particle swarm fuzzy algorithm, power prediction.                        In this paper, 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
                            I. INTRODUCTION                                         changes in wind speed and electricity demand. The smoothing
                                                                                    of the drive train is accomplished by minimizing the torque
       HE growing awareness of climate change, the environ-                         ramp rate rather than controlling the rotor speed presented in
T      ment, fuel supply uncertainty, and raising energy costs
have elevated interest in renewable energy. As one of the most
                                                                                    [16]. The former reduces extreme loads, which translates into a
                                                                                    lower maintenance and operation cost. To model the turbine, a
viable sources of renewable energy, wind power is undergoing                        data-driven approach is introduced. To realize the adaptive tur-
rapid expansion.                                                                    bine control, estimates of future electricity demand and wind
   Although wind energy research has expanded in scope to                           power to be produced at the same time are desired. A time-se-
cover domains such as, for example, wind energy conversion                          ries model extracted by a data-driven approach predicts the fu-
[1], [2], the design of wind turbines [3], [4], the condition mon-                  ture wind power. A simulation model is used to generate the
itoring of wind turbines [5], [6], reliability studies [7]–[9], and                 future demand due to the lack of demand data in this research.
the design of wind farms [10], [11], numerous challenges are                           Supervisory Control and Data Acquisition (SCADA) data
ahead as large-scale wind technology is new and is likely to be                     from turbines installed at a large wind farm (150 MW) have
further developed. The cost of producing electricity from the                       been used in this research. To develop and validate the models
wind is intricately related to the successful development and                       proposed in this research, 0.1-Hz (10-s) data from three ran-
commercialization of wind turbines. One meaningful way to re-                       domly selected wind turbines are used.
duce the cost is to maximize the power generated from a wind
turbine. Numerous approaches to increase power output by op-
                                                                                                       II. PROBLEM FORMULATION
timizing the power coefficient have been published in the lit-
erature. Boukhezzar et al. [12] designed a nonlinear controller
for optimizing the power of the DFIG generator [12]. Wang et                        A. Adaptive Control
al. [13] investigated an intelligent maximum power extraction                          The framework of adaptive wind turbine control is illustrated
algorithm to improve the performance of wind turbine systems                        in Fig. 1. In this paper, a wind turbine is optimized subject to the
                                                                                    following two objectives: power maximization and minimiza-
   Manuscript received January 09, 2010; revised August 05, 2010; accepted          tion of the torque ramp rate. A weighted linear combination
August 27, 2010. Date of publication September 02, 2010; date of current ver-
sion December 15, 2010. This work was supported by the Iowa Energy Center
                                                                                    of the two objectives is used in the bi-objective optimization
under Grant 07-01.                                                                  model. The values of two weights are impacted by the amount
   The authors are with the Intelligent Systems Laboratory, The University of       (excess or deficit) of the generated power and the projected
Iowa, Iowa City, IA 52242 USA (e-mail:                    power demand.
   Color versions of one or more of the figures in this paper are available online
at                                                         As illustrated in Fig. 1, three models, the wind turbine power
   Digital Object Identifier 10.1109/TSTE.2010.2072967                               generation model, the wind power prediction model, and the
                                                                 1949-3029/$26.00 © 2010 IEEE

                                                                                                 TABLE I
                                                                                             DATA DESCRIPTION

                                                                    the noncontrollable and controllable parameters, as shown in the


                                                                    where is the current time, represents the sampling time of
                                                                    data (10-s data), and       is the function describing the wind
                                                                    turbine energy conversion process learned by the data-mining
Fig. 1. Framework of adaptive wind turbine control.                    1) Algorithm Selection: To build the model (1), the following
                                                                    seven data-mining algorithms have been considered: neural
                                                                    network [19]–[21], neural network ensemble [22], -nearest
electricity demand model, are established to realize the adap-      neighbor [23], support vector machine [24], [25], boosting tree
tive control framework. The wind power prediction model and         (regression) [26], [27], classification and regression tree [28],
the electricity demand model are utilized as references in deter-   and random forest (regression) [29].
mining values of the weights of the objectives. The wind turbine       Table I describes the industrial data selected from three wind
power generation model aims to accurately estimate the power        turbines to conduct the research discussed in this paper. The data
generated from a wind turbine with the control settings. Estab-     are partitioned into three datasets: training dataset, test dataset,
lishing accurate models of power generation and power predic-       and validation dataset. Turbine 1 provides the training and test
tion is challenging, and in this paper, it is accomplished with     datasets, and the data from Turbines 2 and 3 are used to vali-
data-mining algorithms that have been successfully applied in       date the models derived by various data-mining algorithms. The
other domains, e.g., industry and service [17], [18].               training dataset is used to extract models by the data-mining
   The energy of the wind and the electricity demand are            algorithms. The test dataset is applied to test the accuracy of
variable, and therefore, the weights assigned to the two corre-     the models and selection of the best-performing data-mining al-
sponding objectives need to be adjusted accordingly. A novel        gorithm. The validation dataset examines the robustness of the
optimization algorithm, the Particle Swarm Fuzzy Algorithm,         data-driven model established by the selected algorithm.
is developed to determine the weights of the two objectives and        The four metrics (2)–(5), the mean absolute error (MAE), the
solve the optimization model that is composed of data-driven        standard deviation of mean absolute error (SD of MAE), the
models developed in this paper. Two controllable parameters,        mean absolute percentage error (MAPE), and the standard de-
blade pitch angle and generator torque, are the solution to         viation of mean absolute percentage error (SD of MAPE), are
the optimization problem, and they are considered the recom-        applied to evaluate the performance of the data-driven models
mended settings for the control system of a wind turbine to
manipulate wind turbines.                                                                                                           (2)
   To maintain good quality of data-driven models, a relearning
scheme is applied to the models using the continuously collected
data.                                                                                                                               (3)
B. Wind Turbine Power Generation Model
   The power conversion of a wind turbine is represented by the                                                                     (4)
4-tuplet           , where represents the wind speed (non-
controllable parameter), is the generator torque (controllable
parameter), is the blade pitch angle (controllable parameter),
and is the power generated by the wind turbine (response pa-
rameter). The value of the response parameter is impacted by                                                                        (5)
30                                                                               IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 2, NO. 1, JANUARY 2011

                            TABLE II                                                                          TABLE IV

                                                                                 accuracy reported in Table II, it is still impressive considering
                                                                                 that the model was built from the data generated at Turbine 1
                                                                                 and tested on Turbines 2 and 3.

                                                                                 C. Wind Power Prediction Model

                                                                                   To determine wind turbine power at time , the time-series
                                                                                 prediction model with the structure presented in (6) is utilized


                                                                                 where the notation is the same as in model (1).
Fig. 2. First 100 points of test results produced by the neural network model.
                                                                                    The time-series prediction model employs the past observed
                                                                                 events to determine its future values. In the candidate model (6),
                            TABLE III                                            the past states of the generated power itself and wind speed at
       PREDICTION ACCURACY RESULTS PRODUCED BY THE NEURAL                        time periods         to         are considered. A parameter selec-
                         NETWORK MODEL
                                                                                 tion strategy, the wrapper with a random search approach [30],
                                                                                 [31], is applied to select the most significant parameters. The pa-
                                                                                 rameters,        ,       ,       ,       , and      , are selected,
                                                                                 and model (6) is instantiated as model (7)


   Comparative analysis of the test results of the models built                  where the notation is the same as in model (1).
by the data-mining algorithms is shown in Table II. The model                       1) Algorithm Selection: The seven data-mining algorithms
derived by the neural network algorithm outperforms all models                   of Section II-B1 are applied to build power prediction models.
in estimating the power generated by a wind turbine. The cor-                    The training dataset of Table I is utilized to extract models, and
responding values of MAPE (0.02) and SD of MAPE (0.07) are                       the test dataset (Table I) is used to test their accuracy at time .
the lowest in Table II. Thus, the neural network algorithm will                     Table IV presents the test results [expressed with the metrics
be used to construct the wind turbine power generation model.                    (2)–(5)] of the power prediction models derived by the seven
   Fig. 2 illustrates the first 100 values of the observed (mea-                  data-mining algorithms. The neural network and the neural net-
sured) power and the power predicted by the neural network                       work ensemble provided results with an MAPE of 0.06. How-
model.                                                                           ever, since the values of MAE, SD of MAE, and SD of MAPE
   2) Model Validation: The validation dataset is used here to                   of the neural network ensemble are all smaller than those of the
assess the robustness of the neural network extracted model in                   neural network, the neural network ensemble is considered for
Section II-B1. Two validation datasets are used: Dataset 1 is col-               building the power prediction model.
lected from wind turbine 2, and validation dataset 2 is collected                   Fig. 3 illustrates the results for the first 100 points of power
from wind turbine 3. The details of the two validation datasets                  prediction by the neural network ensemble using validation
are presented in Table I. The four metrics (2)–(5) are used to as-               dataset 1.
sess the accuracy of the power model with the results presented                     2) Model Validation: The validation dataset used in
in Table III. The MAPE for validation dataset 1 is 0.05, and the                 Section II-B2 is used here to assess the robustness of the neural
MAPE for validation dataset 2 is 0.03. Although the accuracy                     network ensemble model. Table V presents the validation
of this model tested on two different turbines is lower than the                 results of this model. Compare the MAPE listed in Tables IV

Fig. 3. First 100 test points of the observed power and the power predicted by   Fig. 5. Simulated demand data from 7:00 A.M. to 9:00 A.M.
the neural network ensemble.

                             TABLE V                                             horizontal axis represents the time measured in 10-s intervals,
      TEST RESULTS OF POWER PREDICTION BY DATA-DRIVEN MODELS                     the same as the data frequency used to develop the models in
                                                                                 Section II-B. In this model, each demand value is a random
                                                                                 number generated from a normal distribution with the mean
                                                                                 following the pattern (solid line) in Fig. 4. The standard devi-
                                                                                 ation is arbitrarily fixed at 50.
                                                                                    As presented in Fig. 4, the mean demand reflects a scenario
                                                                                 where the mean electricity demand is low (200 kW) in the
                                                                                 evening after office hours (specifically, 6:00 P.M.) and in the
                                                                                 early morning before office hours (specifically, 8:00 A.M.).
                                                                                 After 8:00 A.M., the mean demand increases and reaches a
                                                                                 maximum at 9:00 A.M.. Between 9:00 A.M. and 5:00 P.M., the
                                                                                 mean demand remains constant, and it begins to decline after
                                                                                 5:00 P.M. because the staff leaves the office.
                                                                                    This demand model M1 is expressed as


Fig. 4. Electricity demand pattern of model M1.
                                                                                                                                    if        (9)
and V. It is obvious that this time-series data-driven model is
feasible, and its performance is consistent in power prediction.

D. Electricity Demand Simulation Model                                                                                              if
   Modeling electricity demand is outside the scope of this re-
search. In this paper, the electricity demand data has been sim-                 where       presents the demand data generated from the normal
ulated.                                                                          distribution, is the mean of the normal distribution,
   One basic demand simulation model is developed to generate                        ,                  ,       ,          (8:00 A.M.),
daily power demand at 10-s intervals based on the previous lit-                  (9:00 A.M.),                (5:00 P.M.),             (6:00 P.M.),
erature [32] and the daily electricity consumption by heating                                 (12:00 P.M.), and         .
ventilating and air conditioning (HVAC) systems. It is known                        In this model, the demand data is generated from a normal dis-
that modeling customer demand is a challenge, as it depends on                   tribution (8) and the mean demand computed. For the demand
factors such as the type of buildings, the region size, the time                 data generated from (8), two constraints (10) and (11) are used
of day, and so on. The demand considered is the net energy of                    to prevent the value of demand becoming negative or exceeding
other generation resources.                                                      the maximum capacity of the office space
   This demand model is established to describe a pattern that
simulates the usage of electricity in a micro-grid dominated by                                                                              (10)
business offices (see Fig. 4).                                                                                                                (11)
   The vertical axis of Fig. 4 represents the electricity demand
and its maximum of 1500 kW. This arbitrary value of demand                        Fig. 5 shows a portion of the demand data generated from
should match the generation of a single wind turbine. The                        model (8), (9) from 7:00 A.M. to 9:00 A.M..
32                                                                     IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 2, NO. 1, JANUARY 2011

E. Optimization Model                                                         III. PARTICLE SWARM FUZZY ALGORITHM (PSFA)
   The model considered in this paper maximizes the power gen-            In this research, a wind turbine is controlled adaptively
erated from a wind turbine and minimizes the generator torque          according to the predicted power generation and the power
ramp rate. Maximizing the generated power estimated by the             demand. The two control strategies impact the values of the
model of Section II-B is equivalent to                                 weights associated with the two objective functions of model
   . Here, the difference between the maximum power that could         (14). To produce the control strategy associated with the model
be potentially generated by a wind turbine and the power output        objectives, a particle swarm fuzzy algorithm is developed.
from the model is minimized. Two constraints are considered to            The particle swarm fuzzy algorithm involves two phases. In
construct this boundary, the maximum turbine capacity of 1500          the first phase, the weights of the two objectives are determined
kW, and the maximum energy that could be extracted from the            by a fuzzy algorithm [33] to ascertain the importance of two ob-
wind according to Betz’s law, expressed as           [16].             jectives (discussed in Section II-E) in the optimization problem
   The torque ramp rate         is expressed in                        based on the predicted power and demand information. This al-
                                                                       gorithm is expressed by the following:

where is the generator torque at time ,             is the generator                                                                  (17)
torque at time         .                                                                               if
   As the scales of the two objectives differ, the generated output                                    if                             (18)
is scaled as presented in the following:                                                               if
     The biobjective optimization model is expressed in

                                                                       where is the predicted power and demand from the power pre-
                                                                       diction model (6) and demand models (9) or (11), respectively;
                                                                               is the membership function;      is the weight factor to de-
                                                                       termine the final weight;        is the final weight used in the ob-
                                                                       jective of the optimization model (14).           is the maximum
                                                                       wind turbine capacity (here 1500 kW) and                    is arbi-
                                                                       trarily fixed at 200 kW and expressed in the [0, 1] interval. If the
                                                                       predicted power is less than 200 kW, then                 indicates
                                                                       that the wind speed is low and the smoothing control of the gen-
                                                                       erator torque is less significant than power maximization. Thus,
                                                               (14)    the weight assigned to the torque smoothing objective in (14)
where is the model’s overall objective;        and are the set-        will be close to 0. If customer demand is lower than 200 kW,
tings of the blade pitch angle and the generator torque at time        this implies low energy demand, and the weight assigned to the
 ; and      and     are the weights of the two objectives deter-       generated power objective in (14) should be small. The value of
mined by the particle swarm fuzzy algorithm (PSFA) (discussed                  indicates that the turbine’s generation capacity has been
in Section III). The remaining notation is the same as in (1) and      attained or the demand level is at maximum. If the predicted
(12).                                                                  power is greater or equal to 1500 kW, the control should focus
   In model (8), two controllable parameters, and , are uti-           on minimizing the torque ramp rate to prevent the fatigue of the
lized to realize the adaptive control of the wind turbine. As pre-     drive train and thus reduce the maintenance cost. The demand,
sented in model (14), to control the wind turbine conservatively,      which is greater or equal to 1500 kW, implies a high level energy
the ranges of the two controllable parameters are expressed in         consumption and power maximization needs to be emphasized.
the following:                                                            The second phase of the proposed algorithm involves the
                                                                       particle swarm optimization (PSO) algorithm [34]. In the PSO,
                                                               (15)    each particle represents the candidate optimal solutions of
                                                                       model (14) and can be expressed as                 , where      is a
                                                                       two-dimensional vector                     at the th iteration, and
   The lower bound ( 0.57) and the upper bound (90.61) for the             is another two-dimensional vector                  . The vector
blade pitch angle that cannot be exceeded are obtained from the             represents the position of each particle at iteration , and
industrial data used in this research. In addition, the increment      vector represents the velocity associated with each particle at
(or decrement) of blade pitch angle is set in the range                iteration . The initial value of each dimension of the particle’s
            . The value of the generator torque is expressed in        position is generated from a uniform distribution, where          is
percentage [0, 100], rather than N/m . The adjustment of the           generated from                                               and
generator torque is done at increments (or decrements) in the          is generated from                                                  .
range                      .                                           The initial values of the particle’s velocity are all set at 0 at

                            TABLE VI

the initialization step. The optimization procedure is presented

 Step 1) Determine the weights of fitness function           ,
          which is the objective function of model (14) that
          takes the weights of objectives produced by the            Fig. 6. Standardized fitness value.
          fuzzy algorithm.
 Step 2) Initialize the particle size , the position of each
          particle           , and its velocity          , where     pitch angle. The particle swarm fuzzy algorithm determines the
                          ,         , and       .                    weights of the objective function and the recommended control
 Step 3) Initialize the local best for each particle by              settings optimizing the two control objectives, the maximum
                    , and estimate the initial global best           power and the minimum torque ramp rate.
          by                           , for                and         2) Convergence of the Particle Swarm Fuzzy Algorithm:
                 .                                                   In this experiment described next, convergence of the particle
 Step 4) Repeat until the stopping criterion is satisfied             swarm fuzzy algorithm (PSFA) based on the data point of
                                                       .             Table VI is examined. Ten particles are created, and the stop-
                                                                     ping criterion is set to 1500 iterations. The convergence of the
For each particle                                                    PSFA is evaluated based on the fitness value. To simplify the
     Step 4.1) Create random vectors            and                  evaluation procedure, a fitness value every five iterations from
     by generating                 for           .                   5–100 and the fitness value at iteration 1500 are examined (see
     Step 4.2) Update the velocities of particles by                 Fig. 6). The horizontal axis in Fig. 6 represents the number of
                                                    and update       iterations, and the vertical axis shows the standardized fitness
     the particle positions by              .                        value expressed as
     Step 4.3) Update the local best by            , if
                     .                                                                                                             (21)
     Step 4.4) Update the global best by             , if
                     .                                               according to the notation of Section III.
                                                                        As shown in Fig. 6, the rapid drop of the standardized fit-
   The above algorithm requires parameter initialization. The        ness value for the global best indicates the quick convergence of
parameter is an inertial constant [35]. It controls the impact       the PSFA. In the 1500 iterations, the standardized fitness value
of the previous velocity on the current velocity, and usually it     drops to 1 at the 50th iteration (converges from 50th iterations)
has a value between 0 and 1. It is fixed arbitrarily at 0.5 in this   and constantly keeps this value in the following iteration.
research. Parameters and are two constants that reflect how              3) Optimization Results: The optimization results based on
much the movement of particles is impacted by the local and          this single point (Table VI) are illustrated in Table VII. The first
global best. Both values are arbitrarily fixed at 2 [36]. The PSO     column of Table VII presents the recommended control strategy
algorithm used in this research performs better than the SPEA        for the wind turbine. Here, the power generated is 1139.55 kW
reported in [16]. The computational time advantage can be ten-       (slightly higher than original power output of 1112 kW shown
fold and more. The computational efficiency gain comes from           in Table VI). The ramp rate of the generator torque is nearly
the flight function of the PSO. The performance of the PSO            zero. The two weights indicate the current control preference.
algorithm is presented in Section IV-A2.                             Weight-Torque Ramp is 0.97 and Weight-Power is 0.03. The
                                                                     values of the weights indicate that the control preference focuses
                 IV. INDUSTRIAL CASE STUDY                           on the smoothness of the drive train rather than the maximiza-
                                                                     tion of the generated power.
A. Single-Point Optimization
   1) Description of the Data Point: In this section, the particle   B. Multipoint Optimization
swarm fuzzy algorithm is demonstrated for a single data point           To simulate and demonstrate the adaptive control approach
that has been randomly selected from the test dataset of Table I.    of a wind turbine, the multipoint optimization based on selected
This data point is partially illustrated in Table VI.                data from the test dataset in Table I is introduced. This dataset
   The data in Table VI reflects a scenario where the current         reflects a scenario where the predicted power (supply) is higher
power generated from a wind turbine is high and the simulated        than the demand at the beginning of the period, and then the
demand is low. The current settings of two controllable param-       power demand gradually exceeds the predicted power. It con-
eters are 76.90 for the generator torque and 1.07 for the blade      tains 300 data points from 12/30/2008 8:17:20 A.M. to 12/30/
34                                                                          IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 2, NO. 1, JANUARY 2011

                                                                        TABLE VII
                                                 TEST RESULTS OF POWER PREDICTION BY DATA-DRIVEN MODELS

Fig. 7. Convergence speed for six data points.                               Fig. 9. Comparison of the optimized torque ramp rate and measured torque
                                                                             ramp rate.

Fig. 8. Estimated weights for two objectives.

                                                                             Fig. 10. Comparison of the optimized and original power.
2008 9:07:10 A.M.. The demand data is generated from the de-
mand model discussed in Section II-D.
   1) Stopping Criterion of the PSFA: Before running the mul-                   The optimized power generated from a wind turbine and the
tipoint optimization, an experiment is designed to evaluate the              original power it produced is illustrated in Fig. 10.
stopping criterion for the PSFA. Six data points, indexed as 20,                Figs. 9 and 10 clearly demonstrate the change in the wind
552, 1020, 1500, 1947, and 3155, are randomly selected from                  turbine control strategies. As initially predicted, the power gen-
the test dataset of Table I. The PSFA has run 1500 iterations for            erated is higher than the demand; therefore, the wind turbine is
each point. Fig. 7 illustrates the convergence of the PSFA for the           controlled for torque smoothness. In this case, a higher weight
six points. It is obvious that the PSFA converges quickly, most              is assigned to minimizing the generator torque ramp rate. How-
of the time within 60 iterations. However, to be more conserva-              ever, the direction of the weights gets changed over time. After
tive, the number of iterations that the PSFA needs to run is set             the initial period, the control strategy switches to maximizing
to 100.                                                                      power in response to the electricity demand gradually exceeds
   2) Optimization Results: In this section, the test dataset from           the predicted power. The weight assigned to the power maxi-
12/30/2008 8:17:20 A.M. to 12/30/2008 9:07:10 A.M. is merged                 mization overwhelms the weight of minimizing the torque ramp
with the demand data generated from demand model. Fig. 8                     rate. Due to the trade-off between the two objectives, increase
illustrates the weight values assigned to the two objectives of              in the torque ramps can be observed. Constraint (14) limits the
model (14). As Fig. 8 shows, Weight-Power is initially lower                 maximum change of torque ramps. It can be adjusted as needed.
than Weight-Torque Ramp; however, over time Weight-Power                        Fig. 11 demonstrates that the measured blade pitch angle re-
dominates Weight-Torque Ramp. Since, in the PSFA, Weight-                    mains essentially constant at 0.53 .
Power and Weight-Torque Ramp are associated with the demand                     To realize the adaptive control in practice, new set points need
and predicted power, Fig. 8 characterizes the scenario discussed             to be computed within the data sampling frequency (here 10 s).
before in Section IV-B. Fig. 9 shows the optimized generator                 The computational results with the 10-s data demonstrate that
and the measured torque ramp rate.                                           the set points of the torque and the blade pitch angle are usually
KUSIAK AND ZHANG: ADAPTIVE CONTROL OF A WIND TURBINE WITH DATA MINING AND SWARM INTELLIGENCE                                                                35

                                                                                   [8] R. Karki, P. Hu, and R. Billinton, “A simplified wind power generation
                                                                                       model for reliability evaluation,” IEEE Trans. Energy Convers., vol. 21,
                                                                                       no. 2, pp. 533–540, Jun. 2006.
                                                                                   [9] P. J. Tavner, J. Xiang, and F. Spinato, “Reliability analysis for wind
                                                                                       turbines,” Wind Energy, vol. 10, no. 1, pp. 1–18, 2006.
                                                                                  [10] R. Barthelmie, S. Frandsen, M. Nielsen, S. Pryor, P. Rethore, and H.
                                                                                       Jørgensen, “Modelling and measurements of power losses and turbu-
                                                                                       lence intensity in wind turbine wakes at Middelgrunden offshore wind
                                                                                       farm,” Wind Energy, vol. 10, no. 6, pp. 517–528, 2007.
                                                                                  [11] J. Mora, J. Barón, J. Santos, and M. Payán, “An evolutive algorithm
                                                                                       for wind farm optimal design,” Neurocomputing, vol. 70, no. 16, pp.
                                                                                       2651–2658, 2007.
                                                                                  [12] B. Boukhezzar and H. Siguerdidjane, “Nonlinear control with wind
                                                                                       estimation of a DFIG variable speed wind turbine for power capture
                                                                                       optimization,” Energy Convers. Manage., vol. 50, no. 4, pp. 885–892,
Fig. 11. Comparison of the optimized and measured blade pitch angle.
                                                                                  [13] Q. Wang and L. Chang, “An intelligent maximum power extraction
                                                                                       algorithm for inverter-based variable speed wind turbine systems,”
                                                                                       IEEE Trans. Power Electron., vol. 19, no. 5, pp. 1242–1249, Sep.
computed within 3 s, which is acceptable. The computational ef-                        2004.
ficiency could be further improved by applying different search                    [14] I. Munteanu, N. Cutululis, A. Bratcu, and E. Ceanga, “Optimization of
                                                                                       variable speed wind power systems based on a LQG approach,” Control
strategies and code development techniques.                                            Eng. Practice, vol. 13, no. 7, pp. 903–912, 2005.
                                                                                  [15] T. Senjyu, R. Sakamoto, N. Urasaki, T. Funabashi, H. Fujita, and H.
                            V. CONCLUSION                                              Sekine, “Output power leveling of wind turbine generator for all op-
                                                                                       erating regions by pitch angle control,” IEEE Trans. Energy Convers.,
   Adaptive control of a wind turbine maximizing power gen-                            vol. 21, no. 2, pp. 467–475, Jun. 2006.
eration and minimizing the torque ramp rate was presented.                        [16] A. Kusiak, Z. Song, and H. Zheng, “Anticipatory control of wind tur-
Data-mining algorithms were utilized to generate nonpara-                              bines with data-driven predictive models,” IEEE Trans. Energy Con-
                                                                                       vers., vol. 24, no. 3, pp. 766–774, Sep. 2009.
metric models of wind turbine power generation and the wind                       [17] M. J. A. Berry and G. Linoff, Data Mining Techniques: For Marketing,
power prediction model. An integrated model with a linear                              Sales, and Customer Relationship Management, 2nd ed. New York:
combination of weighted objectives was created. The weights                            Wiley, 2004.
                                                                                  [18] S. K. Shevade, S. S. Keerthi, C. Bhattacharyya, and K. R. K. Murthy,
associated with the objectives were estimated based on the                             “Improvements to the SMO algorithm for SVM regression,” IEEE
predicted power and demand. The demand was generated from                              Trans. Neural Netw., vol. 11, no. 5, pp. 1188–1193, Sep. 2000.
two simulation models. A novel optimization approach, the                         [19] H. Siegelmann and E. Sontag, “Analog computation via neural net-
                                                                                       works,” Theor. Comput. Sci., vol. 131, no. 2, pp. 331–360, 1994.
particle swarm fuzzy algorithm (PSFA), was developed to solve                     [20] G. P. Liu, Nonlinear Identification and Control: A Neural Network Ap-
the model developed in this paper.                                                     proach. London: Springer, 2001.
   Industrial data from three wind turbines of the same type                      [21] M. Smith, Neural Networks for Statistical Modeling. New York: Van
                                                                                       Nostrand, 1993.
was utilized in this study. The data sampling frequency was                       [22] L. K. Hansen and P. Salamon, “Neural network ensembles,” IEEE
0.1 Hz. The data from one turbine was used to develop and test                         Trans. Pattern Anal. Mach. Intell., vol. 12, no. 10, pp. 993–1001, Oct.
data-driven models, and the data of two other turbines was em-                         1990.
                                                                                  [23] Shakhnarovish, Darrell, and Indyk, Nearest-Neighbor Methods in
ployed to validate these models. Comparative analysis of var-                          Learning and Vision. Cambridge, MA: MIT Press, 2005.
ious data-mining algorithms was performed. The convergence                        [24] B. Schölkopf, C. J. C. Burges, and A. J. Smola, Advances in Kernel
and the stopping criterion of the PSFA were investigated.                              Methods: Support Vector Learning. Cambridge, MA: MIT Press,
   The feasibility of adaptive control of wind turbines was                       [25] I. Steinwart and A. Christmann, Support Vector Machines. New
demonstrated. In addition to the two objectives applied in this                        York: Springer-Verlag, 2008.
paper, additional objectives, for example, mitigation of wind                     [26] J. H. Friedman, “Stochastic gradient boosting,” Comput. Statist. Data
                                                                                       Anal., vol. 38, no. 4, pp. 367–378, 2002.
turbine vibrations, could be considered.                                          [27] J. H. Friedman, “Greedy function approximation: A gradient boosting
                                                                                       machine,” Ann. Statist., vol. 29, no. 5, pp. 1189–1232, 2001.
                               REFERENCES                                         [28] L. Breiman, J. H. Friedman, R. A. Olshen, and C. J. Stone, Classifica-
    [1] M. Arifujjaman, M. Iqbal, and J. Quaicoe, “Performance comparison              tion and Regression Trees. Monterey, CA: Wadsworth, Inc., 1984.
        of grid connected small wind energy conversion systems,” Wind Eng.,       [29] L. Breiman, “Random forests,” Mach. Learning, vol. 45, no. 1, pp.
        vol. 33, no. 1, pp. 1–18, 2009.                                                5–32, 2001.
    [2] Ö. Mutlu, E. Akpınar, and A. Balıkcı, “Power quality analysis of wind     [30] I. H. Witten and E. Frank, Data Mining: Practical Machine Learning
        farm connected to Alaçatı substation in Turkey,” Renewable Energy,             Tools and Techniques, 2nd ed. San Francisco, CA: Morgan Kauf-
        vol. 34, no. 5, pp. 1312–1318, 2009.                                           mann, 2005.
    [3] D. Laino, C. Butterfield, R. Thresher, and D. Dodge, “Evaluation of        [31] R. Kohavi and G. H. John, “Wrapper for feature subset selection,” Artif.
        select IEC standard wind turbine design cases on the combined experi-          Intell., vol. 97, no. 1–2, pp. 273–324, 1997.
        ment wind turbine model,” Wind Energy, Amer. Soc. Mech. Engineers,        [32] D. J. Swider, “Compressed air energy storage in an electricity system
        Solar Energy Division (Publication) SED, vol. 14, pp. 47–48, 1993.             with significant wind power generation,” IEEE Trans. Energy Convers.,
    [4] K. Saranyasoontorn and L. Manuel, “A comparison of wind turbine                vol. 22, no. 1, pp. 95–102, Mar. 2007.
        design loads in different environments using inverse reliability tech-    [33] B. Mirzaeian, M. Moallem, V. Tahani, and C. Lucas, “Multiobjective
        niques,” Wind Energy, vol. 126, no. 4, pp. 1060–1068, 2004.                    optimization method based on a genetic algorithm for switched reluc-
    [5] A. Kusiak, H. Zheng, and Z. Song, “On-line monitoring of power                 tance motor design,” IEEE Trans. Magn., vol. 38, no. 3, pp. 1524–1527,
        curves,” Renewable Energy, vol. 34, no. 6, pp. 1487–1493, 2009.                May 2002.
    [6] A. Kusiak, H. Zheng, and Z. Song, “Models for monitoring wind farm        [34] J. Kennedy and R. C. Eberhart, “Particle swarm optimization,” in IEEE
        power,” Renewable Energy, vol. 34, no. 3, pp. 583–590, 2009.                   Int. Conf. Neural Networks, 1995, pp. 1942–1948.
    [7] A. Leite, C. Borges, and D. Falcão, “Probabilistic wind farms gen-        [35] Y. Shi and R. Eberhart, “Parameter selection in particle swarm opti-
        eration model for reliability studies applied to Brazilian sites,” IEEE        mization,” in Proc. 7th Annu. Conf. Evolutionary Programming, 1998,
        Trans. Power Syst., vol. 21, no. 4, pp. 1493–1501, Nov. 2006.                  pp. 591–600.
36                                                                                  IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 2, NO. 1, JANUARY 2011

     [36] M. A. Abido, “Optimal design of power-system stabilizers using par-       Engineers, etc. His current research interests include applications of computa-
          ticle swarm optimization,” IEEE Trans. Energy Convers., vol. 17, no.      tional intelligence in automation, wind and combustion energy, manufacturing,
          3, pp. 406–413, Sep. 2002.                                                product development, and healthcare.
                                                                                       Prof. Kusiak is an Institute of Industrial Engineers Fellow and the Editor-in-
                                                                                    Chief of the Journal of Intelligent Manufacturing.
                           Andrew Kusiak (M’89) received the B.S. and M.S.
                           degrees in engineering from the Warsaw University
                           of Technology, Warsaw, Poland, in 1972 and 1974,
                           respectively, and the Ph.D. degree in operations                                  Zijun Zhang (S’10) received the B.S. degree (2008)
                           research from the Polish Academy of Sciences,                                     from the Chinese University of Hong Kong, Hong
                           Warsaw, in 1979.                                                                  Kong, China, in 2008, the M.S. degree from the Uni-
                              He is currently a Professor at the Intelligent Sys-                            versity of Iowa, Iowa City, in 2009, and is currently
                           tems Laboratory, Department of Mechanical and In-                                 working toward the Ph.D. degree in the Department
                           dustrial Engineering, The University of Iowa, Iowa                                of Mechanical and Industrial Engineering, The Uni-
                           City. He speaks frequently at international meetings,                             versity of Iowa, Iowa City.
                           conducts professional seminars, and does consulta-                                   His research concentrates on data mining and
tion for industrial corporations. He has served on the editorial boards of over                              computational intelligence applied to systems
40 journals. He is the author or coauthor of numerous books and technical pa-                                modeling, monitoring, and optimization in wind
pers in journals sponsored by professional societies, such as the Association for                            energy and HVAC domains. He is a member of the
the Advancement of Artificial Intelligence, the American Society of Mechanical       Intelligent Systems Laboratory at The University of Iowa.

To top