This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON SUSTAINABLE ENERGY 1

An Approximate Wind Turbine Control System Model for Wind Farm Power Control
Yi Guo, Student Member, IEEE, S. Hossein Hosseini, Student Member, IEEE, Choon Yik Tang, Member, IEEE, John N. Jiang, Senior Member, IEEE, and Rama G. Ramakumar, Life Fellow, IEEE
Empirical Model Sampling period. Number of data points. Approximate Model Static nonlinear elements. Time constants. Filter outputs. Damping ratios. Natural frequencies. Scalar gains. I. INTRODUCTION EING able to control a wind farm so that its power output is cooperatively maximized, or smoothly regulated, is imperative to successful and reliable integration of large-scale wind generation into the power grid. The design of a sophisticated wind farm control system (WFCS) for such control, however, is challenging for a variety of reasons. First, a wind turbine, by itself, is already a fairly complex system with highly nonlinear dynamics, strong electromechanical coupling, inherently uncertain parameters, and multiple control variables. Second, when hundreds of such turbines are immersed in a wind eld across a geographical region, they produce turbulence and wake effects that affect downstream turbines, causing their overall behavior to be complicated. Third, the large number of control variables to simultaneously handle, and the rich set of approaches to possibly use, further compound the complexity. Thus, it is challenging to design a WFCS, which perhaps is a reason why there has been relatively little work to date on the topic [1][7], compared to, say, single-turbine control (e.g., [8][19]). One way to cope with the complexities, adopted in [1][7], is to introduce a hierarchical architecture, which, as shown in Fig. 1, divides a WFCS into two parts: a central wind individual wind turbine confarm controller (WFC) and trol systems (WTCSs), each comprising a wind turbine and its controller. With this architecture, we may rst design, , a WTCS that tries to regufor each late its millisecond-to-second-timescale active and reactive and at some desired and , power outputs . Upon comregardless of its incoming wind speed pletion, we may then design a WFC that tries to regulate the second-to-minute-timescale wind farm power outputs and at some desired
AbstractWind farm power control is key to reliable large-scale wind integration. The design of a sophisticated wind farm controller, however, is challenging partly because there is a lack of models that appropriately simplify the complex overall dynamics of a large number of wind turbine control systems (WTCSs). In this paper, using system identication approaches, we develop a simple approximate model that attempts to mimic the active and reactive power dynamics of two generic WTCS models under normal operating conditions: an analytical model described by nonlinear differential equations, and an empirical one by input-output measurement data. The approximate model contains two partsone for active power and one for reactiveeach of which is a third-order system that would have been linear if not for a static nonlinearity. For each generic model, we also provide an identication scheme that sequentially determines the approximate model parameters. Finally, we show via simulation that, despite its structural simplicity, the approximate model is accurate and versatile, capable of closely imitating several different analytical and empirical WTCS models from the literature and from real data. The results suggest that the approximate model may be used to facilitate research on wind farm power control. Index TermsApproximate model, wind turbine control system, wind farm, power control.

BASIC NOMENCLATURE General Time. Actual active and reactive powers. Desired active and reactive powers. Steady-state values of actual powers. Wind speed. Operating region. Analytical Model State variables. Internal control signals. Internal feedback signals.

Manuscript received July 23, 2011; revised August 03, 2012; accepted September 03, 2012. This work was supported by the National Science Foundation under Grant ECCS-0926038 and Grant ECCS-0955265. Y. Guo, S. H. Hosseini, C. Y. Tang, and J. N. Jiang are with the School of Electrical and Computer Engineering, University of Oklahoma, Norman, OK 73019 USA (e-mail: yi.guo@ou.edu; s.h.hosseini@ou.edu; cytang@ou.edu; jnjiang@ou.edu). R. G. Ramakumar is with the School of Electrical and Computer Engineering, Oklahoma State University, Stillwater, OK 74078 USA (e-mail: ramakum@okstate.edu). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TSTE.2012.2217992

1949-3029/$31.00 2012 IEEE

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY

Fig. 1. Hierarchical architecture of a wind farm control system.

and , presumably from a grid operator, by adjusting the s and s based on feedback of the s and s and possibly estimates of the s. Hence, the architecture simplies the design of a WFCS, allowing us to sequentially tackle two (seemingly) easier problems on different timescales, as opposed to tackling a harder one. The architecture also offers us the option of designing a new single-turbine controller for each WTCS , or applying an existing one (e.g., [8][14]) that accepts and as inputs.1 Furthermore, it allows us to view the WFC as a second-to-minute-timescale supervisor that tells every WTCS how much power to generate, and focus on its design without delving too much into millisecond-to-second, turbine-level details. Although the hierarchical architecture makes the problem more manageable, it does not remove the fact that each WTCSbeing a composition of an already-complex wind turbine and a possibly-complicated controllertypically has complex dynamics. As a result, the subsequent design and analysis of a supervisory WFC may prove to be difcult, depending on our goal: if we are content with a simple design (e.g., distribute and evenly among the s and s, or proportionally based on the s) and a basic analysis (e.g., simulation studies only), then how complex a WTCS is probably does not matter. However, if we aim for a nifty design (e.g., adjust the s and s so that the WTCSs can exploit their correlation, interaction, and/or diversity to cooperatively achieve faster transient responses and better steady-state smoothness in and ) and a deeper analysis and understanding (e.g., theoretical characterization of the resulting transient and steady-state behaviors), then an overly complex WTCS may render the process very difcult or even impossible. Therefore, to achieve the latter, it is necessary to build a suitably simplied WTCS model. To this end, suppose we have developed, or are given, a WTCS under normal operating conditionscall it WTCS and wish to design a WFC. Also suppose, at our disposal, is a mathematical model parameterized by a vector call it WTCS which, like WTCS (or each WTCS in Fig. 1), maps inputs to outputs , i.e., . Consider the following conditions on the model WTCS : (C1) There exists a such that whenever WTCS and WTCS are driven by the same inputs , they produce approximately the same outputs .
1Single-turbine controllers that do not accept and , such as those that always attempt maximum power tracking (e.g., [15][19]), may not t well with this architecture.

(C2) WTCS may be a nonlinear dynamical system but has a favorable structure conducive to control systems analysis and design. (C3) There is a set of WTCSs in the literature such that for each WTCS in the set, there exists a such that whenever WTCS and the WTCS are driven by the same inputs , they produce approximately the same outputs . Note that if (C1) holds, WTCS with the specic value of would be an accurate approximation of WTCS and, thus, may be used in place of WTCS in the WFC design and WFCS analysis. If, in addition, (C2) holds, the design and analysis would be more likely to succeed due to the favorable structure of WTCS . If (C3) holds as well, WTCS with different values of would be able to also approximate a number of different WTCSs in the literature (or by different manufacturers), making it a versatile model that brings WTCS and those WTCSs under the same umbrella, distinguished only by . It follows that the design and analysis outcomes (e.g., new control techniques, stability criteria, and performance formulas) are applicable not only to WTCS , but perhaps also to those WTCSs, increasing their impact. Hence, having an approximate model WTCS that satises conditions (C1)(C3) is extremely valuable. This paper is devoted to the development of such a model. We rst assume, in Section II, that two generic models of WTCS are given, namely, an analytical model described by a set of continuous-time nonlinear differential equations, and an empirical model described by a set of input-output measurement data. The latter is motivated by the fact that in practice, what is available may just be a set of data, rather than a mathematical model, due to legacy and proprietary reasons. Based on standard system identication approaches [20] and typical WTCS characteristics, we then develop, in Section III, an approximate model WTCS , which attempts to imitate both the analytical and empirical models. For each of these two models, we also provide a parameter identication scheme that sequentially determines the required in (C1), which is a vector of 10 parameters (two functions and eight scalars). The approximate model, depicted later in Fig. 3(c), may be regarded as satisfying (C2) because it is made up of two structurally identical partsone for active power and the other for reactiveeach of which is a third-order system that would have been linear if not for a static nonlinear component at its inputs (i.e., a modied Hammerstein model [20], [21]). Next, we validate, in Section IV, the approximate model via simulation, showing that it has enough ingredients to closely imitate several different analytical and empirical models from the literature [1], [4], [14], [16] and from real data taken from an Oklahoma wind farm. The encouraging results suggest that the approximate model satises (C3) and, hence, may be used to facilitate the design and analysis of a second-to-minute-timescale supervisory WFC that yields a sophisticated WFCS. Finally, Section V concludes the paper. We stress that this paper is not about wind turbine modeling (as in, say, [22][24]), controller design (as in, say, [8][19]), and controller comparison ([4]), nor is it about wind farm modeling ([25][27]) and controller design ([1][7]). Rather, the paper is on wind turbine control system modelingnot just the

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
GUO et al.: APPROXIMATE WIND TURBINE CONTROL SYSTEM MODEL FOR WIND FARM POWER CONTROL 3

turbine, but the turbine plus its controllerin the context of the architecture in Fig. 1, for which there seems to be no prior work. We note that there has been signicant efforts by the Western Electricity Coordinating Council (WECC), Utility Wind Integration Group (UWIG), and IEEE on developing generic WTCS models [28][30]. Such models, however, are intended for accurately simulating the impact of wind turbine generator dynamics on power system transient stability without compromising vendor proprietary information, rather than for facilitating research on wind farm power control in the aforementioned context, where model simplicity is at premium. Therefore, those models and our approximate model have notably different purposes. Indeed, they have very different structure and level of details, which can be seen by comparing, for instance, [28, Figs. 1, 3, and 2024] with Fig. 3(c) of this paper. II. MODELS OF WIND TURBINE CONTROL SYSTEMS In this section, we describe a generic analytical model and a generic empirical model of a WTCS under normal operating conditions. These two models set the stage for the development of an approximate model that mimics the dynamic performance of the WTCS, when it provides primary generation services to the grid. A. Analytical Model To control a variable-speed wind turbine, a standard approach is to rst model its dynamics based on rst principles, and then design a controller based on known techniques. Regardless of the model and design, the resulting wind turbine control system (WTCS) typically can be represented in a block diagram form as in Fig. 2, and described generically by a set of continuous-time, nonlinear differential equations in state-space form as follows: (1) (2) Here, denotes time; is the system states combining the wind turbine states (e.g., stator and rotor uxes or currents, rotor angular velocity) and controller states (if any); is the initial states; and are functions depending on the particular wind turbine model (e.g., fourth-order [22] or second-order [23] doubly fed induction generator (DFIG) dynamics or second-order permanent magnet synchronous generator (PMSG) dynamics [24], and rigid-shaft or exible-shaft [23] mechanical dynamics) and the particular controller design (e.g., one of the designs in [1][3] and [8][14]) including their parameters (e.g., resistances, inductances, rotor moment of inertia, rotor swept area, -surface, air density, friction coefcient, controller gains); , , , and are, respectively, the system inputs and outputs representing the desired and actual active and reactive powers, where positive values mean toward the grid; is another system input representing the wind speed; is the internal control signals (e.g., rotor voltages, blade pitch angle, electromagnetic torque); and is the internal feedback signals (e.g., various voltages and currents, rotor angular velocity, actual powers).

Fig. 2. Block diagram of a wind turbine control system.

In this paper, we assume that a generic analytical model of a WTCS, in the form of (1) and (2), is given as the rst of two models considered. In order to represent as many WTCSs in the literature as possible, we make only two assumptions about the analytical model (1) and (2): rst, the inputs are always in a bounded operating region . Second, the WTCS is reasonably well-designed, i.e., the functions and are such that for each constant , there exist steady-state values , depending possibly on , such that for every , . Finally, we allow and to be absent, since many existing WTCSs do not consider the reactive power (e.g., [10][12]), and to be absent as well, since some existing WTCSs do not require it to be specied (e.g., [15][19]). However, we require and to be present, since they are essential to WTCSs. B. Empirical Model Although it is common to work with a mathematical model in research on WTCSs, in practice we may not have access to the inner working of a WTCS, due perhaps to legacy and proprietary reasons. Instead, what may be available to us is a set of input-output measurement data, so that we have no choice but to treat the WTCS as a black box. The set of data can take various forms, but more often than not includes the following information:

(3)

where is the sampling period which is usually on the order of seconds or minutes, and is the number of data points which is usually large. In this paper, we assume that a generic empirical model of a WTCS, in the form of (3), is given as the second of the two models considered. Similar to the analytical model (1) and (2), in the empirical model (3) the columns , and are optional but the columns and are mandatory. However, unlike the analytical one where the inputs can be arbitrarily specied, with this empirical model we have no control over the inputs , as they are simply given, in the rst three columns. This difference will be accounted for shortly. Remark 1: The two models of WTCSs in this section may be thought of as the WTCS in Section I.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
4 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY

C. Discussion The analytical model (1) and (2) and the empirical model (3) are what our approximate model intends to imitate. As will be described in Section III, our approach is based on postulating a static nonlinear model that matches the steady-state input-output characteristics, followed by enriching the model with linear dynamics so that it also matches the transient input-output behaviors. A benet of this input-output approach is that it bypasses the need to consider the internal dynamics and specic details of the underlying WTCS, thereby allowing major types of generation technologies such as DFIG, PMSG, and IG to be approximately described using a simple, consistent model. More important, such a model enables one to approximately describe a large number of same or different types of WTCSs within a wind farm in a unied fashion, so that researchers may focus on other pressing issues when designing a comprehensive WFCS and understanding its attainable performance. III. PROPOSED APPROXIMATE MODEL In this section, we develop a simple mathematical model that approximates the analytical and empirical WTCS models in Section II, and a parameter identication scheme that determines the model parameters in each case. The development consists of three steps in both cases, as described below. A. Approximating the Analytical Model Step 1: Mimicking the steady-state responses to constant inputs In general, to create a system that mimics another system, it is reasonable to demand that the two systems exhibit the same steady-state responses to constant inputs. With this in mind, we note that whenever the analytical model (1) and (2) is subject to constant inputs , its outputs asymptotically converge to some steady-state values , which depend only on and not on the initial states . This dependency suggests that there exist functions and , such that and . It also suggests a static nonlinear model of the form (4) (5) which is capable of mimickingat the very leastthe steadystate outputs of the analytical model (1) and (2) whenever the inputs are constant, or slow-varying. To visually connect this Step 1 with subsequent steps of the development, a block diagram of the model (4) and (5) is shown in Fig. 3(a). Remark 2: Throughout the paper, the subscripts 1 and 2 are used to distinguish between similar parameters or variables associated with the active and reactive powers [e.g., is for active and is for reactive in Fig. 3(a)]. The functions and are (innite-dimensional) parameters of the model (4) and (5), which can be identied by simulating the analytical model (1) and (2) with various constant inputs sufciently covering the operating region , observing the steady-state outputs, and employing interpolation. The following procedure provides the details:
Fig. 3. Step-by-step development of the proposed approximate model. (a) Block diagram after Step 1 of 3. (b) Block diagram after Step 2 of 3. (c) Block diagram after Step 3 of 3.

Procedure for Step 1 1) 2) 3) 4) 5) 6) 7) 8) 9) 10) 11) Pick a large . for . Pick Pick a large . Loop over . Let . Pick any . Simulate the analytical model (1) and (2) from to . Record . Let and . End loop. Determine and via interpolation on the data points obtained.

Applying the above procedure to identify the functions and , we obtain a basic model (4) and (5) that exhibits the same steady-state behavior as that of (1) and (2). Step 2: Mimicking the transient responses to staircase inputs The basic model (4) and (5) in Fig. 3(a) is able to match the steady-state response of the analytical model (1) and (2). However, it fails to produce any kind of transient one normally would expect with WTCSs because (4) and (5) are merely static functions mapping the inputs to the outputs

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
GUO et al.: APPROXIMATE WIND TURBINE CONTROL SYSTEM MODEL FOR WIND FARM POWER CONTROL 5

. To alleviate this drawback, we insert into Fig. 3(a) rst-order linear dynamics between and and between and , to arrive at a Hammerstein model (see [21] and [20, Ch. 5.2]) shown in Fig. 3(b) and given by (6) (7) and are the time constants. Note that in where steady-state, (6) and (7) reduce to (4) and (5). Hence, (6) and (7) are able to capture not only the steady-state behavior of the analytical model (1) and (2), but also the dominant mode of its transient behavior with proper choices of and . The time constants and can be identied using a general approach in system identication sometimes known as the prediction-error methods (see [20, Ch. 7]). With this approach, we rst choose specic inputs and use them to simulate the analytical model (1) and (2) and the model (6) and (7), the latter with different values of and . We then compare the outputs of the two models and determine the best and , which minimize the output differences. The following procedure details this approach, in which we choose the inputs to be random staircase signals because they tend to bring out the dominant mode in systems, and allow any norm to be used for measuring the output differences: Procedure for Step 2 1) Pick a large . for 2) Pick randomly, independently, and equiprobably. 3) Use the in Step 1. 4) Let for . 5) Pick any . 6) Simulate the analytical model (1) and (2) from to . 7) Record as . 8) Pick a large . and for . 9) Pick 10) Use the and identied in Step 1. 11) Pick any . 12) Loop over . and . 13) Let 14) Simulate the model (6) and (7) from to 15) Record as . 16) Calculate and where . 17) End loop. 18) Let . , , and

Remark 3: In the above procedure, and for represent the search space for the best and ; and represent the desired norms; and and are introduced to reduce the impact of the initial states [i.e., of the analytical model (1) and (2) and of the model (6) and (7)] on the parameter estimation process. Using the preceding procedure to identify the time constants and , we obtain a rened model (6) and (7) that has more exibility to better match the behavior of (1) and (2). Step 3: Mimicking the responses to realistic inputs Although the rened model (6) and (7) in Fig. 3(b) is more sophisticated than the basic model (4) and (5) in Fig. 3(a), it can only produce rst-order-like responses. If such responses are indeed what the analytical model (1) and (2) produces, or if what we desire is just a crude approximation, then the rened model (6) and (7) may be satisfactory. Otherwise, its accuracy may be unacceptable. At rst glance, this issue can be overcome by replacing the rst-order linear dynamics in (6) and (7) with higher-order ones. This approach, however, has a fundamental limitation: recall from Step 1 that and . Thus, if a WTCS does power regulation and does it well over a wide range of , then and for any in that range. As a result, in (6) and in (7) would both be insensitive to , so that even large uctuations in the wind speed would be completely absorbed by and , producing no uctuations in the active and reactive powers and , which may be unrealistic. To bypass this limitation, we introduce two second-order linear lters and add two linear terms to (6) and (7), to get a modied Hammerstein model depicted in Fig. 3(c) and dened by

(8)

(9)

(10)

. where

(11) and are the lter outputs, and are the damping ratios, and are the natural frequencies, and and are scalar gains. To see the rationale behind (8)(11), notice that the second-order linear lters in (8) and (9) are low-pass lters with unity DC gains. Hence, and may be seen as short-term averages of , which catch up to if it ever approaches constant, and and may be viewed as deviations of from its short-term averages, which uctuate around zero. It follows that the linear terms

and

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY

and in (10) and (11) enable uctuations in to induce uctuations in and , bypassing the aforementioned limitation and yielding a feature not possessed by the rened model (6) and (7). Moreover, because the steady-state values of these terms are zero when is constant, (10) and (11) also preserve the role of and as constant-inputs-to-steady-state-outputs maps (see Step 1). Finally, due to the tuning knobs , , , , , and , (8)(11) possess considerable (but not excessive) freedom to mimic the way uctuations in affect and of the analytical model (1) and (2). All of these explain the rationale behind (8)(11), which we will refer to from now on as the approximate model. The parameters , , , , , and can be identied using the general approach adopted in Step 2, i.e., the so-called prediction-error methods [20]. Indeed, a procedure analogous to the one in Step 2 may be constructed as follows: Procedure for Step 3 1) 2) 3) 4) 5) 6) 7) 8) 9) 10) Pick a large . Pick some specic Pick any . Simulate the analytical model (1) and (2) from to . Record as Pick a large . for Pick . Use the and identied in Step 1. Use the and identied in Step 2. Pick any . 11) Loop over 12) Let .

B. Approximating the Empirical Model As was mentioned in Section II-B, in practice we may be given an empirical model of a WTCS, dened by input-output measurement data of the form (3), and asked to design a WFC. Thus, it is desirable that our approximate model (8)(11)with suitable choices of parameterscan also imitate the empirical model (3), producing outputs that closely resemble the last two columns of (3), when the inputs are from the rst three columns. To come up with such suitable choices, reconsider the parameter identication scheme from Steps 13. Observe that this scheme is not immediately applicable here because Steps 1 and 2 require constant and staircase inputs , but with the empirical model (3) the inputs are whatever that are given. To circumvent this issue, below we modify the scheme, allowing it to handle any given inputs, and label the steps involved Steps 13, to distinguish them from, and to stress their parallel with, Steps 13 above. The modication yields the second parameter identication scheme, intended just for the empirical case. Step 1: Identifying the functions and Parallel to Step 1, the goal of this Step 1 is to construct a procedure for identifying the functions and , so that the basic model (4) and (5) in Fig. 3(a) is able to roughly mimic the empirical model (3). To do so, observe that the identication of (and, similarly, ) can be treated as a curve-tting problem with domain containing the inputs and range containing the output . Also observe that if we partition the domain into and write as where are the parameters and are the set indicator basis functions (see [20, Ch. 5.4]), then the optimal in the least-squares sense can be easily computed: each is simply the average of those for which . These observations suggest the following procedure, in which we partition the domain into three-dimensional grids, for simplicity: Procedure for Step 1 1) Let

. 13) Simulate the model (8)(11) from as 14) Record 15) Calculate and , and 16) End loop. 17) Let minimizes be the , and that minimizes , where .

to

. .

that be the . . 2) Pick a large 3) Let 4) Loop over 5) Let . , and . .

Remark 4: In the above procedure, may be different from the in Steps 1 and 2; and may be, say, staircases, ramps, or from realistic proles; may be from real data; and for represent the search space for the best . Note that the three procedures in Steps 13 collectively form a parameter identication scheme, which enables sequential determination of all the parameters of the approximate model (8)(11) (i.e., and , then and , then the rest).

. 6) If , let

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
GUO et al.: APPROXIMATE WIND TURBINE CONTROL SYSTEM MODEL FOR WIND FARM POWER CONTROL 7

3) Pick . 4) Use the and 5) Use the and 6) Pick any 7) Loop over 8) Let

for identied in Step 1. identied in Step 2. . .

where denotes the cardinality of . 7) End loop. 8) Determine and via interpolation on the (at most) data points obtained. Step 2: Identifying the parameters and Unlike going from Step 1 to Step 1 where the procedure undergoes signicant changes, only minor modications are needed to make the procedures in Steps 2 and 3 applicable to the empirical model (3) in this Step 2 and the next Step 3. In particular, the inputs now come from the rst three columns of (3), and now represent the desired norms, and and now play the role of and in nullifying the impact of the initial states: Procedure for Step 2 1) Rename from the empirical model (3) as . 2) Pick a large . 3) Pick and for . 4) Use the and identied in Step 1. 5) Pick any . 6) Loop over . and . 7) Let 8) Simulate the rened model (6) and (7) from to . 9) Record as . 10) Calculate and , where , and 11) End loop. 12) Let . Step 3: Identifying the parameters and Procedure for Step 3 1) Rename as 2) Pick a large from the empirical model (3) . . , . , and

. 9) Simulate the approximate model (8)(11) from to . 10) Record as . 11) Calculate

where and . 12) End loop. 13) Let minimizes

be the , and that minimizes

that be the .

Remark 5: The approximate model (8)(11) in this section may be viewed as the WTCS in Section I, with . Also, it may be regarded as satisfying (C2) in Section I because it has isolated static nonlinearities and is relatively simple compared to full-blown WTCS models, such as those in Section IV. IV. VALIDATION OF THE APPROXIMATE MODEL In this section, we validate via simulation the approximate model developed in Section III, showing that it is capable of closely imitating several analytical and empirical WTCS models from the literature and from real data. To enable the validation, we rst describe a wind turbine model, followed by the analytical and empirical WTCS models considered. We then describe the validation settings and results. A. Wind Turbine Model Consider a variable-speed wind turbine equipped with a DFIG.2 Under normal operating conditions, the turbine may be modeled by the following differential and algebraic equations [16], [22]:

2Due to space limitation, only DFIG is considered in the model validation. We note, however, that other types of generators, such as IG, may be implemented in a similar manner.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
8 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY

denote the stator and rotor; the frame; the uxes; the voltages; the currents; the resistances; the inductances; the leakage coefcient; the constant angular velocity of the synchronously rotating reference frame; the rotor angular velocity; the rotor moment of inertia; the friction coefcient; the mechanical torque; the electromagnetic torque; the air density; the rotor swept area of radius the -surface; the tip speed ratio; and the blade pitch angle. In addition, to have some diversity in the validation, consider the following two distinct sets of values for the wind turbine parameters: the rst set of values is adopted from [31], [32] and corresponds to a GE 3.6 MW turbine, while the second is adopted from MATLAB/Simulink R2007a and corresponds to a GE 1.5 MW turbine. These two sets of values are listed in the Appendix.

where

B. WTCS Models Next, consider four analytical WTCS models from the literature and an empirical WTCS model from real data, labeled as WTCS1WTCS5 and dened as follows: WTCS1 is made up of the GE 3.6 MW turbine model and the controller in Rodriguez-Amenedo et al. [1], which regulates and by adjusting , , and using ve PI blocks and a power-speed lookup table, as depicted in Fig. 4(a). Note that this controller assumes that the -axis of the synchronously rotating reference frame is aligned with the stator ux vector, i.e., , and that the reactive power is solely coming from the stator, i.e., . For more information about this controller, see [1] and related work [33], [4] (in particular, [1, Fig. 3], [33, Fig. 10], and [4, Figs. 6 and 7]). WTCS2 is made up of the same GE 3.6 MW turbine model and the controller studied in Fernandez et al. [4] and displayed in Fig. 4(b). Observe that this controller is similar to the one in [1] except that it uses to determine in the outer loop and to determine in the inner loop, whereas the one in [1] does the opposite. For more details about this controller, see [4] and [2] (especially, [4, Figs. 6 and 8] and [2, Fig. 4]). WTCS3, unlike WTCS1 and WTCS2, is made up of the smaller GE 1.5 MW turbine model and the nonlinear dual-

Fig. 4. Block diagrams of the controllers that yield WTCS1WTCS4. (a) Block diagram of the controller in Rodriguez-Amenedo et al. [1]. (b) Block diagram of the controller in Fernandez et al. [4]. (c) Block diagram of the controller in Tang et al. [14]. (d) Block diagram of the controller in Johnson et al. [16].

mode controller in Tang et al. [14], which uses the feedback linearization technique to cancel nonlinearities in the DFIG dynamics, and the gradient descent method to maximize or regulate and including the power factor, as outlined in Fig. 4(c). Notice that this controller assumes instead that the -axis is aligned with the stator voltage vector, i.e., , and that it does not assume . WTCS4 is formed by the mechanical dynamics of the GE 1.5 MW turbine model and the controller in Johnson et al. [16], which is implemented on the Controls Advanced Research Turbine (CART) at the National Renewable Energy Laboratorys (NRELs) National Wind Technology Center (NWTC) and also discussed in [34]. Sketched in Fig. 4(d), this controller maximizes the power capture in Region 2 by varying and keeping at its optimum, and prevents the power capture from exceeding the rated

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
GUO et al.: APPROXIMATE WIND TURBINE CONTROL SYSTEM MODEL FOR WIND FARM POWER CONTROL 9

Fig. 5. Imitating WTCS1 dened by the GE 3.6 MW turbine and the controller in Rodriguez-Amenedo et al. [1].

value in Region 3 by varying accordingly. In contrast to WTCS1WTCS3, this WTCS assumes no electrical dynamics and, thus, does not involve and , nor . Finally, WTCS5the only empirical model considered in this paperis a black box dened by a set of input-output measurement data taken from an actual GE 1.5 MW turbine within a wind farm located in northwest Oklahoma. This set of data has data points and was collected over 238 days at a sampling period of min. Moreover, the set of data ts the mold of (3), containing the mandatory and for , but not the optional , , and . In order to use this data set for second-level simulation in the sequel, we redene as min, assuming that one-day worth of data were taken over an hour. C. Validation Settings Given WTCS1WTCS5, suppose now we want to construct, for each WTCS , an approximate model (8)(11) that resembles its behavior. To this end, for each WTCS , we execute the rst parameter identication scheme in Steps 13 (if WTCS is analytical), or the second one in Steps 13 (if it is empirical), to obtain a specic approximate model with specic values of , , , , as well as , , , , (if the optional is indeed an output of WTCS ).

To evaluate how well the ve approximate models imitate WTCS1WTCS5, we consider 30 different scenarios. For each scenario, we generate inputs from to s, choosing to be a staircase signal with three random staircase values each lasting 1200 s, to be such that the desired power factor is kept constant at to be an actual wind prole from the afore0.995, and mentioned wind farm. For each WTCS and each scenario, we simulate both WTCS and its corresponding approximate model for 3600 s using the same inputs associated with the scenario, record the outputs of the two models, and calculate the root-mean-square error (RMSE) in between the two models after some initial transient. (Obviously, the smaller the RMSE, the better the approximation.) D. Validation Results Figs. 59 depict, respectively, the ve approximate models and how well they resemble WTCS1WTCS5. Although the gures have different sizes and styles, they share the same format: the rst row of subplots is associated with the active power; the second row, if present, is associated with the reactive power; the rst column displays the identied values of the approximate model parameters , , , , , , , , , , showing and as contour plots in Figs. 57 and as graphs in Figs. 8 and 9; and the second and third columns each show, for a selected scenario, the outputs

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
10 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY

Fig. 6. Imitating WTCS2 dened by the GE 3.6 MW turbine and the controller in Fernandez et al. [4].

of WTCS and those of its corresponding approximate model over 3600 s and over 60 s, the latter in zoom-in windows. Notice that although, in general, and are functions of , for WTCS1WTCS3 they are functions of only or equivalently (since in order to obtain a desired power factor of 0.995), and for WTCS4 and WTCS5 they are functions of only (since is not required). This explains why and can be shown as contour plots and graphs. Also note that due to space limitations, for each WTCS , we could only show the outputs for two selected scenarios (as opposed to showing both the inputs and outputs for all 30 scenarios). Finally, each gray dot in subplot 1 of Fig. 9 represents an empirical data point for WTCS5 and is included just to provide additional insight. Complementing Figs. 59 is Table I, which shows the minimum, maximum, and average RMSE in between WTCS1WTCS5 and their corresponding approximate models, taken over all the 30 scenarios. To get a sense of what the numbers in the table mean, one may refer to Figs. 59, which also state the RMSEs of the curves for the few selected scenarios. Observe from Figs. 59 and Table I that while the proposed approximate model is not without error, the magnitude of which is generally very small, sometimes even negligible, across all WTCSs and all scenarios. In particular, it is able to produce the right peaks and valleys at the right moments in all the

3600-s subplots and the 60-s zoom-in windowsexcept for the rst 500 s in Fig. 7 and rst 200 s in Fig. 8, which may be attributed to the approximate and analytical models having different initial states and, hence, different initial transients. These encouraging observations validate the approximate model in Fig. 3(c), demonstrating its ability to closely replicate the behaviors of the ve fairly different analytical and empirical WTCS models considered. Remark 6: The validation in this section may be thought of as verifying (C1) and (C3) in Section I. V. CONCLUSION In this paper, we have presented a simple approximate model, which tries to mimic generic analytical and empirical WTCS models under normal operating conditions, along with two parameter identication schemes, which determine the approximate model parameters in both cases. We have also demonstrated through simulation the accuracy and versatility of the approximate model in resembling several different analytical and empirical WTCS models from the literature and from real data. From the results, we conclude that the approximate model is a compelling candidate, based on which one may design and analyze a second-to-minute-timescale supervisory wind farm controller using advanced control techniques (e.g., model predictive control [35], distributed cooperative control [36], and quasi-linear control [37]), in future research.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
GUO et al.: APPROXIMATE WIND TURBINE CONTROL SYSTEM MODEL FOR WIND FARM POWER CONTROL 11

Fig. 7. Imitating WTCS3 dened by the GE 1.5 MW turbine and the controller in Tang et al. [14].

Fig. 8. Imitating WTCS4 dened by the mechanical dynamics of the GE 1.5 MW turbine and the controller in Johnson et al. [16].

APPENDIX For the GE 3.6 MW turbine, the values of its parameters are: m, m/s, and the -surface , where

is given by

, and . For the GE 1.5 MW turbine, the values of its parameters are: m,

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
12 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY

Fig. 9. Imitating WTCS5 dened by real data from an Oklahoma wind farm. TABLE I AVERAGE ROOT-MEAN-SQUARE TAKEN OVER 30 SCENARIOS EACH WTCS [8] A. D. Hansen, P. Sorensen, F. Iov, and F. Blaabjerg, Control of variable speed wind turbines with doubly-fed induction generators, Wind Eng., vol. 28, no. 4, pp. 411434, Jun. 2004. [9] R. G. de Almeida and J. A. Pecas Lopes, Participation of doubly fed induction wind generators in system frequency regulation, IEEE Trans. Power Syst., vol. 22, no. 3, pp. 944950, Aug. 2007. [10] L.-R. Chang-Chien and Y.-C. Yin, Strategies for operating wind power in a similar manner of conventional power plant, IEEE Trans. Energy Convers., vol. 24, no. 4, pp. 926934, Dec. 2009. [11] L.-R. Chang-Chien, W.-T. Lin, and Y.-C. Yin, Enhancing frequency response control by DFIGs in the high wind penetrated power systems, IEEE Trans. Power Syst., vol. 26, no. 2, pp. 710718, May 2011. [12] H. T. Ma and B. H. Chowdhury, Working towards frequency regulation with wind plants: Combined control approaches, IET Renew. Power Generat., vol. 4, no. 4, pp. 308316, Jul. 2010. [13] Y. Guo, S. H. Hosseini, J. N. Jiang, C. Y. Tang, and R. G. Ramakumar, Voltage/pitch control for maximization and regulation of active/reactive powers in wind turbines with uncertainties, IET Renew. Power Generat., vol. 6, no. 2, pp. 99109, Mar. 2012. [14] C. Y. Tang, Y. Guo, and J. N. Jiang, Nonlinear dual-mode control of variable-speed wind turbines with doubly fed induction generators, IEEE Trans. Control Syst. Technol., vol. 19, no. 4, pp. 744756, Jul. 2011. [15] H. Li, K. L. Shi, and P. G. McLaren, Neural-network-based sensorless maximum wind energy capture with compensated power coefcient, IEEE Trans. Ind. Applicat., vol. 41, no. 6, pp. 15481556, Nov./Dec. 2005. [16] K. E. Johnson, L. Y. Pao, M. J. Balas, and L. J. Fingersh, Control of variable-speed wind turbines: Standard and adaptive techniques for maximizing energy capture, IEEE Control Syst. Mag., vol. 26, no. 3, pp. 7081, Jun. 2006. [17] V. Galdi, A. Piccolo, and P. Siano, Designing an adaptive fuzzy controller for maximum wind energy extraction, IEEE Trans. Energy Convers., vol. 23, no. 2, pp. 559569, Jun. 2008. [18] B. Beltran, T. Ahmed-Ali, and M. El Hachemi Benbouzid, Sliding mode power control of variable-speed wind energy conversion systems, IEEE Trans. Energy Convers., vol. 23, no. 2, pp. 551558, Jun. 2008. [19] J. Creaby, Y. Li, and J. E. Seem, Maximizing wind turbine energy capture using multivariable extremum seeking control, Wind Eng., vol. 33, no. 4, pp. 361388, Jun. 2009. [20] L. Ljung, System Identication: Theory for the User, 2nd ed. Upper Saddle River, NJ: Prentice-Hall, 1999. [21] K. Narendra and P. Gallman, An iterative method for the identication of nonlinear systems using a Hammerstein model, IEEE Trans. Automatic Control, vol. 11, no. 3, pp. 546550, Jul. 1966. [22] R. Fadaeinedjad, M. Moallem, and G. Moschopoulos, Simulation of a wind turbine with doubly-fed induction generator by FAST and Simulink, IEEE Trans. Energy Convers., vol. 23, no. 2, pp. 690700, Jun. 2008. [23] Y. Lei, A. Mullane, G. Lightbody, and R. Yacamini, Modeling of the wind turbine with a doubly fed induction generator for grid integration studies, IEEE Trans. Energy Convers., vol. 21, no. 1, pp. 257263, Mar. 2006.

MINIMUM, MAXIMUM, ERROR (RMSE) IN

AND

FOR

m/s, and the

are .

-surface is given by , where , and the coefcients , and

ACKNOWLEDGMENT The authors are grateful to the anonymous reviewers for their helpful comments and to Prof. T. Runolfsson and W. Wang for valuable discussions. REFERENCES
[1] J. L. Rodriguez-Amenedo, S. Arnalte, and J. C. Burgos, Automatic generation control of a wind farm with variable speed wind turbines, IEEE Trans. Energy Convers., vol. 17, no. 2, pp. 279284, Jun. 2002. [2] A. D. Hansen, P. Sorensen, F. Iov, and F. Blaabjerg, Centralised power control of wind farm with doubly fed induction generators, Renew. Energy, vol. 31, no. 7, pp. 935951, Jun. 2006. [3] R. G. de Almeida, E. D. Castronuovo, and J. A. Pecas Lopes, Optimum generation control in wind parks when carrying out system operator requests, IEEE Trans. Power Syst., vol. 21, no. 2, pp. 718725, May 2006. [4] L. M. Fernandez, C. A. Garcia, and F. Jurado, Comparative study on the performance of control systems for doubly fed induction generator (DFIG) wind turbines operating with power regulation, Energy, vol. 33, no. 9, pp. 14381452, Sep. 2008. [5] C. F. Moyano and J. A. Pecas Lopes, An optimization approach for wind turbine commitment and dispatch in a wind park, Electric Power Syst. Res., vol. 79, no. 1, pp. 7179, Jan. 2009. [6] G. Tapia, A. Tapia, and J. X. Ostolaza, Proportional-integral regulator-based approach to wind farm reactive power management for secondary voltage control, IEEE Trans. Energy Convers., vol. 22, no. 2, pp. 488498, Jun. 2007. [7] P. Li, P.-K. Keung, and B.-T. Ooi, Development and simulation of dynamic control strategies for wind farms, IET Renew. Power Generat., vol. 3, no. 2, pp. 180189, Jun. 2009.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
GUO et al.: APPROXIMATE WIND TURBINE CONTROL SYSTEM MODEL FOR WIND FARM POWER CONTROL 13

[24] P. C. Krause, O. Wasynczuk, and S. D. Sudhoff, Analysis of Electric Machinery and Drive Systems, 2nd ed. Piscataway, NJ: Wiley-IEEE Press, 2002. [25] S. Frandsen, R. Barthelmie, S. Pryor, O. Rathmann, S. Larsen, J. Hojstrup, and M. Thogersen, Analytical modelling of wind speed decit in large offshore wind farms, Wind Energy, vol. 9, no. 1, pp. 3953, Jan. 2006. [26] A. E. Feijoo and J. Cidras, Modeling of wind farms in the load ow analysis, IEEE Trans. Power Syst., vol. 15, no. 1, pp. 110115, Feb. 2000. [27] A. S. Dobakhshari and M. Fotuhi-Firuzabad, A reliability model of large wind farms for power system adequacy studies, IEEE Trans. Energy Convers., vol. 24, no. 3, pp. 792801, Sep. 2009. [28] M. Asmine, J. Brochu, J. Fortmann, R. Gagnon, Y. Kazachkov, C.-E. Langlois, C. Larose, E. Muljadi, J. MacDowell, P. Pourbeik, S. A. Seman, and K. Wiens, Model validation for wind turbine generator models, IEEE Trans. Power Syst., vol. 26, no. 3, pp. 17691782, Aug. 2011. [29] Standard Models for Variable Generation North American Electric Reliability Corporation, Princeton, NJ, Special Report, 2010. [30] M. Singh and S. Santoso, Dynamic Models for Wind Turbines and Wind Power Plants National Renewable Energy Laboratory, Golden, CO, Subcontract Report, 2011. [31] N. W. Miller, J. J. Sanchez-Gasca, W. W. Price, and R. W. Delmerico, Dynamic modeling of GE 1.5 and 3.6 MW wind turbine-generators for stability simulations, in Proc. PES General Meeting, Toronto, Canada, 2003, pp. 19771983. [32] W. Qiao, W. Zhou, J. M. Aller, and R. G. Harley, Wind speed estimation based sensorless output maximization control for a wind turbine driving a DFIG, IEEE Trans. Power Electron., vol. 23, no. 3, pp. 11561169, May 2008. [33] R. Pena, J. C. Clare, and G. M. Asher, Doubly fed induction generator using back-to-back PWM converters and its application to variablespeed wind-energy generation, IEE Proc. Electric Power Applicat., vol. 143, no. 3, pp. 231241, May 1996. [34] L. Y. Pao and K. E. Johnson, Control of wind turbines, IEEE Control Syst. Mag., vol. 31, no. 2, pp. 4462, Apr. 2011. [35] E. F. Camacho and C. B. Alba, Model Predictive Control, 2nd ed. London, England: Springer, 2007. [36] J. Shamma, Cooperative Control of Distributed Multi-Agent Systems. Chichester, England: Wiley, 2008. [37] S. Ching, Y. Eun, C. Gokcek, P. T. Kabamba, and S. M. Meerkov, Quasilinear Control: Performance Analysis and Design of Feedback Systems With Nonlinear Sensors and Actuators. New York: Cambridge Univ. Press, 2010. Yi Guo (S08) received the B.S. degree from Tianjin Polytechnic University and the M.S. degree from Tianjin University, Tianjin, China, in 2002 and 2005, respectively, and the Ph.D. degree in electrical engineering from the University of Oklahoma, Norman, in 2012. His current research interests include control theory and applications, power system control and stability, and wind energy generation and integration.

Choon Yik Tang (S97M04) received the B.S. and M.S. degrees in mechanical engineering from Oklahoma State University, Stillwater, in 1996 and 1997, respectively, and the Ph.D. degree in electrical engineering from the University of Michigan, Ann Arbor, in 2003. He was a Postdoctoral Research Fellow in the Department of Electrical Engineering and Computer Science at the University of Michigan from 2003 to 2004 and a Research Scientist at Honeywell Labs, Minneapolis, from 2004 to 2006. Since 2006, he has been an Assistant Professor in the School of Electrical and Computer Engineering, University of Oklahoma, Norman. His current research interests include systems and control theory, distributed algorithms for computation and optimization over networks, control and operation of wind farms, and computationally efcient digital lter design.

John N. Jiang (SM07) received the M.S. and Ph.D. degrees from the University of Texas at Austin. He is an Associate Professor in the Power System Group in the School of Electrical and Computer Engineering, University of Oklahoma, Norman. He has been involved in a number of wind energy related projects since 1989 in design, installation of standalone wind generation systems, the market impact of wind generation in Texas, and recent large-scale wind farms development in Oklahoma.

S. Hossein Hosseini (S10) received the B.S. degree from Sharif University of Technology, Tehran, Iran, in 2009. He is currently working toward the Ph.D. degree in the School of Electrical and Computer Engineering at the University of Oklahoma, Norman. His current research interests include power system economics and nance, and renewable energy.

Rama G. Ramakumar (M62SM75F94 LF02) received the B.E. degree from the University of Madras, Madras, India, the M.Tech. degree from the Indian Institute of Technology, Kharagpur, India, and the Ph.D. degree from Cornell University, Ithaca, NY, all in electrical engineering. After a decade (total) of service on the faculty of the Coimbatore Institute of Technology, Coimbatore, India, he joined Oklahoma State University, Stillwater, in 1967, where he has been a Professor since 1976. In addition, he has been the Director of the OSU Engineering Energy Laboratory since 1987. In 1991, he was named the PSO/Albrecht Naeter Professor of Electrical and Computer Engineering and was promoted to Regents Professor in 2008. His research interests are in the area of energy conversion, energy storage, power engineering, and renewable energy. He has been a consultant to several national and supranational organizations in the eld of energy and has organized and presented short courses on renewable energy topics and engineering reliability. His contributions are documented in over 150 publications, which include four U.S. patents, contributed chapters in four books and seven hand books, and technical papers in various journals, transactions, and national and international conference proceedings. He is the author of the text book Engineering Reliability Fundamentals and Applications (Englewood Cliffs, NY: Prentice-Hall, 1993). Dr. Ramakumars past and present leadership activities in the IEEE Power and Energy Society include chairing the Awards Committee of the Technical Council, the award Subcommittee of the Power Engineering Education Committee, the Energy Development Subcommittee and the Renewable Technologies Subcommittee of the Energy Development and Power Generation Committee, the Working Group on Renewable Technologies, and the Fellows Working Group of the Power Engineering Education Committee. He is a member of the American and International Solar Energy Societies, the American Society for Engineering Education, and the IEEE Industry Applications Society. He is a Registered Professional Engineer in the State of Oklahoma.