my paper

IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 26, NO. 1, JANUARY 2011 165

Application of Adaptive Network-Based FuzzyInference System for Sensorless Control of

PMSG-Based Wind Turbine WithNonlinear-Load-Compensation Capabilities

Mukhtiar Singh, Student Member, IEEE, and Ambrish Chandra, Senior Member, IEEE

Abstract—The precise information of permanent-magnet syn-chronous generator (PMSG) rotor position and speed is essen-tially required to operate it on maximum power points. Thispaper presents an adaptive network-based fuzzy inference system(ANFIS) for speed and position estimation of PMSG, where anANFIS-based model reference adaptive system is continuouslytuned with actual PMSG to neutralize the effect of parameter vari-ations such as stator resistance, inductance, and torque constant.This ANFIS-tuned estimator is able to estimate the rotor positionand speed accurately over a wide speed range with a great im-munity against parameter variation. The proposed system consistsof two back-to-back connected inverters, where one controls thePMSG, while another is used for grid synchronization. Moreover,in the proposed study, the grid-side inverter is also utilized as har-monic, reactive power, and unbalanced load compensator for athree-phase, four-wire (3P4W) nonlinear load, if any, at point ofcommon coupling (PCC). This enables the grid to always supply/absorb a balanced set of fundamental currents at unity powerfactor. The proposed system is developed and simulated usingMATLAB/SimPowerSystem (SPS) toolbox. Besides this, a scaledlaboratory hardware prototype is developed and extensive ex-perimental study is carried out to validate the proposed controlapproach.

Index Terms—Active-power filter, adaptive neuro-fuzzy systems,distributed generation, grid interconnection, permanent-magnetsynchronous generator (PMSG), power quality, renewable energy,sensorless control, wind energy.

I. INTRODUCTION

THE GROWING concern about global warming and theharmful effects of fossil-fuel emissions has created new

demand for cleaner and sustainable renewable energy sources.Among renewable sources, wind energy is one of the fastest

Manuscript received October 16, 2009; revised March 25, 2010; acceptedJune 11, 2010. Date of current version December 27, 2010. This work was sup-ported by the Government of India and the High Commission of India, Ottawa,ON, Canada. Recommended for publication by Associate Editor M. G. Simoes.

M. Singh is with the Department of Electrical Engineering, Ecole de Tech-nologie Superieure, Montreal, QC H3C 1K3, Canada, on leave from the Depart-ment of Electrical Engineering, Deenbandhu Chhoturam University of Scienceand Technology, Sonepat 131 039, India (e-mail: [email protected]).

A. Chandra is with the Department of Electrical Engineering, Ecole deTechnologie Superieure, Montreal, QC H3C 1K3, Canada (e-mail: [email protected]).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TPEL.2010.2054113

growing and lowest-priced renewable energy technologies avail-able today, costing between 4 and 6 cents/kWh, depending uponthe size of a particular project. Modern wind turbines have lot ofcommercially available topologies based on induction generator(fixed speed) and doubly fed induction generator/synchronousgenerator/permanent-magnet synchronous generator (PMSG)(variable speed) [1]–[3]. The variable-speed wind turbines aremore attractive, as they can extract maximum power at dif-ferent wind velocities, and thus, reduce the mechanical stresson WECS by absorbing the wind-power fluctuations. Recently,PMSG-based directly driven variable-speed WECS are becom-ing more popular due to the elimination of gear box and excita-tion system.

Since, the maximum power is the cubic function of generatorspeed for a given tip speed ratio, the continuous information ofgenerator position and speed is essentially required. For this pur-pose, generally shaft-mounted speed sensors are used, resultingin additional cost and complexity of the system. To alleviate theneed of these sensors, several speed-estimating algorithms basedon motional electromotive force (EMF), flux-linkage variation,and Kalman filter have been introduced in the past. However,the precise estimation of rotor position and speed is very diffi-cult as most of these suffer because of simplified computationsbased on several assumptions, ignorance of parameter varia-tions, and inaccuracy involved with low-voltage signal mea-surement at lower–speed, especially in case of directly drivenPMSG [4]–[8]. To overcome these problems, several indirectmethods for maximum power extraction have also been intro-duced. Bhowmik and Spee [9] have proposed power-coefficientpolynomial to estimate wind velocity. In this method, an iterativealgorithm is employed to determine the roots of a polynomial.As the polynomial may be up to seventh order, therefore, the ex-act calculation of the roots is very complex and time-consumingtask. Tan and Islam [10] reported to apply a 2-D lookup tableof power-coefficient or power-mapping method to estimate thewind velocity. The mapping-power technique may occupy a lotof memory space. While saving the memory space by reduc-ing the size of table, the control accuracy will be affected. Liet al. [11] have also suggested a similar artificial neural network(ANN) based power-coefficient compensation method.

In this paper, a novel adaptive network-based fuzzy infer-ence system (ANFIS) based position and speed estimator ofPMSG has been proposed for wide range of speed opera-tion. The ANFIS architecture has well-known advantages of

0885-8993/$26.00 © 2010 IEEE

166 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 26, NO. 1, JANUARY 2011

modeling a highly nonlinear system, as it combines the capabil-ity of fuzzy reasoning in handling uncertainties and capabilityof ANN in learning from processes [12]. Thus, the ANFIS isused to develop an adaptive model of variable-speed PMSGunder highly uncertain operating conditions, which also auto-matically compensates the distortion in measuring signal andany variation in parameters such as inductance, resistance, etc.A detailed comparison of simulation results supported by ex-perimental results obtained using proposed ANFIS estimatoris also provided with EMF-based method under variable-speedPMSG operation. Both the simulation and experimental studiesare also carried out for 20% variation in PMSG resistance andinductance to validate the robustness of proposed ANFIS algo-rithm against generator parameter variation. An error-gradient-based dynamic back-propagation method has been used for theonline tuning of ANFIS architecture. This estimated rotor speedis further utilized to find out the maximum possible power usingpower–speed characteristic of PMSG. An IGBT-based rectifieris utilized to implement the proposed control strategy.

Moreover, in the proposed study, the grid-side inverter rat-ing is optimally utilized by incorporating the power-qualityimprovement features. Normally, the grid-interfacing inverterhas very low utilization factor of 20%–30% with a possiblepeak of 60% of rated output due to the intermittent nature ofwind [13], [14]. Therefore, if the same inverter is utilized forsolving power-quality problem at point of common coupling(PCC) in addition to its normal task, then the additional hard-ware cost for custom power devices like active power filter(APF), static compensator (STATCOM), or VAR compensatorcan be saved. Thus, the authors have proposed a very simpleand cost-effective solution by using the grid-side inverter as aload harmonics, load reactive power, and load unbalance com-pensator of a 3P4W nonlinear unbalanced load at PCC in a dis-tribution network, in addition to its normal task of wind-powerinjection into the grid.

The paper is organized as follows. The system under con-sideration is discussed in Section II. The proposed controlstrategy for generator and grid-side converters is explained inSection III. The MATLAB/SPS-based simulation results aregiven in Section IV and the experimental results are discussedin Section V. Finally, Section VI concludes the paper.

II. SYSTEM DESCRIPTION AND CONTROL

The proposed system consists of a PMSG-based variable-speed WECS consisting two back-to-back inverters with a com-mon dc link. The generator-side inverter controls its speed toextract maximum power at different speeds, while the grid-side inverter delivers the renewable power to grid with 3P4Wnonlinear-load compensation simultaneously. The block dia-gram of proposed variable-speed WECS is shown in Fig. 1.

A. Generator Position/Speed Estimation and Control

The information about generator position and speed is veryessential for optimal control of PMSG, in order to extractmaximum power at different wind velocities. However, dueto nonideal rotor-flux distribution, parameter variation, and

Fig. 1. Block diagram of proposed system.

Fig. 2. Control diagram of proposed system.

rapidly changing operating conditions, it is extremely difficult toreproduce the exact model of PMSG for control purpose. There-fore, in order to reject any such kind of external disturbances, anadaptive model is continuously tuned in parallel of PMSG withthe help ANFIS structure. The complete schematic diagram ofproposed system is shown in Fig. 2.

1) Modeling of PMSG: Since the back-EMF is the functionof rotor position in stationary reference frame, therefore, it isconvenient to model PMSG in this frame. The voltage equationsof PMSG are as follows:

Vα = −Rsiα − Lsdiαdt

+ Eα (1)

Vβ = −Rsiβ − Lsdiβdt

+ Eβ (2)

where Vαβ are stator terminal voltages, Rs is stator resistance,Ls is stator inductance, iαβ are output currents, and Eαβ areback EMFs, which can be given as

Eαβ =

[Eα

Eβ

]= ωrλm

[− sin(θr )

cos(θr )

]. (3)

Here, ωr , θr , and λm are rotor speed, rotor position, andmagnetic-flux linkage, respectively.

SINGH AND CHANDRA: APPLICATION OF ADAPTIVE NETWORK-BASED FUZZY INFERENCE SYSTEM 167

Fig. 3. State-space model of PMSG.

On rearranging (1) and (2), and rewriting them in matrix form,we have[

iα

iβ

]=

[−Rs/Ls 0

0 −Rs/Ls

] [iα

iβ

]

+

[−1/Ls 0

0 −1/Ls

] ([Vα

Vβ

]−

[Eα

Eβ

]). (4)

Using the same structure for adaptive model, the current es-timator can be designed, if the accurate value of back EMF isknown in the following equation:

˙iαβ = A · iαβ + B · (Vαβ − Eαβ ) (5)

where the cap ‘ ’ indicates the estimated variables and

A =

[−Rs/Ls 0

0 −Rs/Ls

], B =

[−1/Ls 0

0 −1/Ls

],

˙iαβ =

⎡⎣ ˙iα

˙iβ

⎤⎦ , iαβ =

[iα

iβ

], Vαβ =

[Vα

Vβ

], and

Eαβ =

[Eα

Eβ

].

The state-space-equivalent diagram of (5) is shown in Fig. 3.2) ANFIS-Based Rotor Position and Speed Estimation:

Since the system is highly nonlinear and working under dif-ferent operating conditions, it is extremely difficult to estimatethe exact value of back EMF, which also carries the informa-tion about the rotor position and speed. To solve this prob-lem, an ANFIS-based dynamic model of PMSG is developed.For ANFIS training, the explicit mathematical model of plantis not required. Only the order of the system with adequateinput/output data from actual plant is needed to develop itsadaptive model [15]. The target output from the actual plant iscompared with the output of ANFIS-based dynamic model inparallel and error is used to update the weights. In the proposedstudy, the dynamic model is continuously tuned by forcing itsestimated currents to track the actual currents of PMSG usingANFIS structure with dynamic back-propagation method. Oncethe estimated currents are equal to actual currents, the adaptivemodel exactly represents the real model irrespective of any kindof parameter variations and external disturbances. This resultsin the accurate estimation of back EMFs given as follows:

Eαβ =

[Eα

Eβ

]= ωrλm

[− sin(θr )

cos(θr )

]. (6)

Fig. 4. Block diagram of ANFIS-based speed and position estimation.

Fig. 5. Schematic of ANFIS architecture.

Now, the position and speed can be easily estimated using theestimated value of back EMFs as θr = − tan−1(Eα/Eβ ) andωr = (1/p)(dθr /dt), where p is the number of pole pairs. Theestimated back EMFs are passed through low-pass filter to filterthe noise. The schematic diagram of rotor position and speedestimation is shown in Fig. 4.

3) ANFIS Architecture: An ANFIS based on Takagi–Sugeno–Kang (TSK) method having 2:6:3:3:2 architecture with

two inputs (estimated currents ˙iα and ˙iβ ) and two outputs (esti-mated back EMFs Eα and Eβ ) are used to develop the dynamicmodel of PMSG. The errors between the actual and estimatedcurrents (ξαβ = iαβ − iαβ ) are used to tune the preconditionand consequent parameters. The diagram of ANFIS architec-ture is shown in Fig. 5 and the node functions of each layer aredescribed as follows.

Layer 1: This layer is also known as fuzzification layer, whereeach node is represented by a square. Here, three membershipfunctions are assigned to each input. The trapezoidal and tri-angular membership functions are used to reduce the computa-tion burden, as shown in Fig. 6, and their corresponding node


Fig. 6. Fuzzy membership functions.

equations are as given as follows:

μA 1 (Xα ) = μB1 (Xβ ) =

⎧⎪⎪⎪⎨⎪⎪⎪⎩

1, X ≤ b1

X − a1

b1 − a1, b1〈X〈

0, X ≥ a1

a1 (7)

μA 2 (Xα ) = μB2 (Xβ ) =

⎧⎨⎩

1 − X − a2

0.5b2, |X − a2 | ≤ 0.5b2

0, |X − a2 | ≥ 0.5b2

(8)

μA 3 (Xα ) = μB3 (Xβ ) =

⎧⎪⎪⎪⎨⎪⎪⎪⎩

0, X ≤ a3

X − a3

b3 − a3, a3〈X〈

1, X ≥ b3

b3 (9)

where the value of parameters {ai , bi} changes according tothe error, and accordingly generates the linguistic value of eachmembership function. Parameters in this layer are referred aspremise parameters or precondition parameters.

Layer 2: Every node in this layer is a circle labeled as Π,which multiplies the incoming signals and forwards it to nextlayer

μi = μAi(Xα ) · μBi

(Xβ ), i = 1, 2, 3. (10)

Here, the output of each node represents the firing strength ofa rule.

Layer 3: Every node in this layer is represented as a circle.This layer calculates the normalized firing strength of each rulegiven as follows:

μi =μi

μ1 + μ2 + μ3, i = 1, 2, 3. (11)

Layer 4: Every node in this layer is a square node with a nodefunction given by

Oi = μi · fji = μi(ai

0 + ai1 · Xj ), i = 1, 2, 3 and j = α, β.

(12)where the parameters {ai

0 , ai1} are tuned as the function of inputs

{iα , iβ}. Parameters in this layer are also referred as consequentparameters.

Layer 5: This layer is also called output layer, which computesthe output as follows:

Y1 = μ1 · fA1 + μ2 · fA

2 + μ3 · fA3 (13)

and

Y2 = μ1 · fB1 + μ2 · fB

2 + μ3 · fB3 .

The output from this layer are now multiplied with the nor-malizing factor and passed through the low-pass filter to findthe estimated value of back EMFs Eα and Eβ .

4) Online Training of ANFIS Architecture: For the proposedANFIS architecture, a gradient-descent technique is used to re-duce the error (usually a cost function given by the squarederror), where the weights are iterated by propagating the errorfrom output layer to input layer. This backward trip of sucha calculation is termed as “back propagation.” The trainingalgorithm is completed in two stages, known as precondition-parameter tuning and consequent-parameter tuning, where theobjective function to be minimized is defined as

ξ2αβ = (iαβ − iαβ )2 . (14)

Precondition-parameter tuning: The precondition parame-ters are required to update the fuzzy membership functionsas discussed in previous section for Layer 1. To minimize theerror function ξ2

αβ by gradient-descent method, the change ineach precondition parameter must be proportional to the rate ofchange of the error function w.r.t. that particular preconditionparameter, i.e.,

ΔaAi= −η

∂ξ2α

∂aAi

, i = 1, 2, 3 (15)

where η is the constant of proportionality defined as the learningrate. Therefore, the new value of the consequent parameter isgiven as

aAi(n + 1) = aAi

(n) + ΔaAi, i = 1, 2, 3 (16)

or

aAi(n + 1) = aAi

(n) − η∂ξ2

α

∂aAi

, i = 1, 2, 3. (17)

Similarly,

bAi(n + 1) = bAi

(n) − η∂ξ2

α

∂bAi

, i = 1, 2, 3. (18)

Now, the partial derivative of (17) can be found by the chainrule of differentiation as follows:

∂ξ2α

∂aA 1

=∂ξ2

α

∂iα· ∂iα

∂Eα

· ∂Eα

∂μ1· ∂μ1

∂μ1· ∂μ1

∂μA 1

· ∂μA 1

∂aA 1

. (19)

On computing all the terms of (19) and putting in (17), wecan find the updated value of parameter aA 1 as follows:

aA 1 (n + 1) = aA 1 (n) − 2η · ξα (n) · k · fA1 (n) · μB1 (n)

· μ2(n) + μ3(n)μ1(n) + μ2(n) + μ3(n)

· μA 1 (n) − 1bA 1 (n) − aA 1 (n)

.

(20)

Similarly,

bA 1 (n + 1) = bA 1 (n) + 2η · ξα (n) · k · fA1 (n) · μB1 (n)

· μ2(n) + μ3(n)μ1(n) + μ2(n) + μ3(n)

· μA 1 (n)bA 1 (n) − aA 1 (n)

.

(21)


In the same manner, the precondition parameters for the re-maining fuzzy membership functions can be derived as follows:

bA 2 (n + 1) = bA 2 (n) − 2η · ξα (n) · k · fA2 (n) · μB2 (n)

· μ1(n) + μ3(n)μ1(n) + μ2(n) + μ3(n)

· 1 − μA 2 (n)bA 2 (n)

(22)

aA 3 (n + 1) = aA 3 (n) − 2η · ξα (n) · k · fA3 (n) · μB3 (n)

· μ1(n) + μ2(n)μ1(n) + μ2(n) + μ3(n)

· μA 3 (n) − 1bA 3 (n) − aA 3 (n)

(23)

bA 3 (n + 1) = bA 3 (n) + 2η · ξα (n) · k · fA3 (n) · μB3 (n)

· μ1(n) + μ1(n)μ1(n) + μ2(n) + μ3(n)

· μA 3 (n)bA 3 (n) − aA 3 (n)

.

(24)

Similarly, the precondition parameters for the fuzzy member-ship functions of second input can be found just by interchangingthe subscripts α ↔ β and A ↔ B in (20)–(24).

Consequent-parameter tuning: To tune the consequent pa-rameters, as discussed in Layer 4, the following updated lawsare developed:

aα0i

(n + 1) = aα0i

(n) − ηc ·∂ξ2

α

∂aα0i

, i = 1, 2, 3 (25)

aα1i

(n + 1) = aα1i

(n) − ηc ·∂ξ2

α

∂aα1i

, i = 1, 2, 3 (26)

where ηc is the learning rate for consequent parameters. Thederivative terms in (25)–(26) can be found by the chain rule asalready discussed in case of precondition parameters as follows:

∂ξ2α

∂aα0i

=∂ξ2

α

∂iα· ∂iα

∂Eα

· ∂Eα

∂aα0i

, i = 1, 2, 3 (27)

∂ξ2α

∂aα1i

=∂ξ2

α

∂iα· ∂iα

∂Eα

· ∂Eα

∂aα1i

, i = 1, 2, 3. (28)

On substituting the terms derived in (27)–(28) into (25)–(26),the updated value of consequent parameters can be derived asfollows:

aα0i

(n + 1) = aα0i

(n) − 2ηc · ξα · k · μi

μ1 + μ2 + μ 3,

i = 1, 2, 3 (29)

aα1i

(n + 1) = aα1i

(n) − 2ηc · ξα · k · μi

μ1 + μ2 + μ 3· iα ,

i = 1, 2, 3. (30)

Similarly, the consequent parameter for second input can befound just by replacing α by β in above equations.

5) Control of PMSG: The main function of variable-speedWECS is to extract maximum power at different wind velocities.The wind power captured by wind turbine depends on its powercoefficient (Cp ) as per the relation given by

PTurb = PwindCp =12ρπr2V 3

ω Cp (31)

Fig. 7. Power–speed characteristic of PMSG.

where ρ is the air density, Vω is the wind velocity, and r is theradius of circular swept area of rotor blades. However, for agiven turbine, Cp is not always constant and heavily depends onthe tip speed ratio λ, which is given as

λ =Tip speed

Wind speed=

ωrr

Vω. (32)

Consequently, different wind speeds will require the optimalvalues of tip speed ratio (λoptimal) to achieve the optimum valueof power coefficient (Cp,optimal). The aforementioned aspectsmake it very clear that to extract maximum power out of thevarying wind, we need to have a WECS that allows the changein rotor speed to achieve optimal aerodynamic conditions. Underoptimal aerodynamic conditions, the maximum power output atdifferent wind velocities is almost the cubic function of gen-erator speed, as shown in Fig. 7. Thus, to extract maximumpower, the generator speed is needed to be controlled accordingto power–speed characteristic [16]. The reference speed ob-tained from power–speed characteristic is compared with speedestimated in previous section and the difference is applied toproportional–integral (PI) controller. The output of PI resultsinto a torque-controlling current component given as follows:

i∗q =(

KP +KI

S

)(ω∗

r − ωr ) (33)

where KP and KI are proportional and integral gains for gener-ator speed control. The direct-axis reference current componentcan be set to zero in order to obtain maximum torque per amper-age (MTPA), as the generator is always supposed to run belowthe base speed. The control action is implemented with the helpof IGBT-based rectifier.

B. Grid-Side Inverter Control

The proposed system employs a four-leg, grid-interfacinginverter, where the fourth leg of inverter is utilized to compen-sate the neutral current of 3P4W distribution network. Here,the inverter is a key element, since it delivers the power fromrenewable to grid, and also solves the power-quality problemarising due to unbalanced nonlinear load at PCC. The duty ra-tio of inverter switches is varied in a power cycle such thatthe combination of load- and inverter-injected power appears as


Fig. 8. Control diagram of grid-side converter.

balanced resistive load to the grid, resulting in the unity powerfactor (UPF) grid operation. The schematic of grid-interfacinginverter control with 3P4W nonlinear unbalanced load compen-sation is shown in Fig. 8.

Usually, the load currents, and either of the grid or invertercurrents are required to compensate the load unbalance, reac-tive power, and harmonics [17]–[21]. This makes it necessaryto have at least eight current sensors for a 3P4W distributionnetwork. However, in the proposed control strategy, only grid-current sensing is required to force the grid currents to be purelysinusoidal at UPF without having any knowledge of inverter-and load-current profile. The control of grid-side inverter com-prises of two loops. The outer dc voltage control loop is usedto set the current reference for active-power control. The outputof dc control loop (im ) is multiplied with in-phase unity vectortemplates (Ua , Ub , and Uc ) to generate the reference grid cur-rents (i∗a , i∗b , and i∗c ). The reference grid neutral current (i∗n ) isset to zero, being the instantaneous sum of balanced grid cur-rents. The in-phase unit vector templates Ua , Ub , and Uc can beeasily computed from three-phase grid voltages va , vb , and vc

using phase-locked loop (PLL) [22].

III. SIMULATION RESULTS AND DISCUSSION

The proposed system is simulated in MATLAB/Simulinkenvironment using SimPowerSystem (SPS) blockset. The per-formance of ANFIS-based position and speed estimation forsensorless control of PMSG is demonstrated through the exten-sive simulation results. Besides this, a detailed comparison ofproposed method is also provided w.r.t. the conventional EMFobserver-based results under variable-speed operation, with dif-ferent value of PMSG parameters. The variable torque is appliedto the PMSG, and accordingly its speed and position are esti-mated. The traces of applied torque, generated voltage, actualand estimated speed, actual and estimated position, speed error,and generated output power are mainly shown in Figs. 9 and 10.In Fig. 11, the traces of grid voltages, grid currents, load currents,and inverter currents are shown. The traces of phase-a currentsand neutral currents for grid, load, and inverter are shown inFig. 12 w.r.t. phase-a grid voltage, just to demonstrate the loadharmonics, and unbalanced- and neutral-current-compensationcapabilities.

Fig. 9. Simulation results with nominal PMSG parameters. (a) Torque.(b) PMSG voltage. (c) PMSG actual and estimated speed. (d) PMSG actualand estimated position. (e) Speed error. (f) PMSG output power.

Fig. 10. Simulation results with 20% variation in PMSG resistance and in-ductance. (a) Torque. (b) PMSG voltage. (c) PMSG actual and estimated speed.(d) PMSG actual and estimated position. (e) Speed error. (f) PMSG outputpower.

In Fig. 9, the simulation results with nominal value of PMSGparameters are shown. A variable torque available from windturbine is applied on the PMSG, as shown in Fig. 9(a). The corre-sponding change in amplitude and frequency of generator volt-age can be easily seen in Fig. 9(b). The Fig. 9(c) shows the actual


Fig. 11. Simulation results showing the performance of grid-side inverter.(a) Grid voltages. (b) Grid currents. (c) Unbalanced load currents. (d) Invertercurrents.

Fig. 12. Simulation results: grid, load, inverter currents of phase-a and neutralw.r.t. phase-a grid voltage.

and estimated speeds based on ANFIS and EMF algorithms, andtheir corresponding actual and estimated positions are shown inFig. 9(d). The ANFIS estimator estimates the PMSG positionand speed accurately, as evident from the trace of speed error inFig. 9(e), whereas the EMF method estimates the speed with lotof oscillations and deviation from actual speed. The correspond-ing PMSG output power is shown in Fig. 9(f). The simulationresults with 20% variation in PMSG resistance and inductanceare shown in Fig. 10. The performance of EMF-based algorithmdeteriorates with the variation in PMSG parameters, while theANFIS estimator shows promising immunity to any such kindof variation, as evident from Fig. 10(e).

Actually, the back EMF method, as in (1) and (2), requiresthe explicit mathematical model of PMSG. It can give satis-factory performance under ideal conditions, i.e., exact valueof parameter must be known and the input signal should bedistortion free. All these requirements make it unsuitable forwind-power application where the parameter of machine mayvary with the passage of time. Moreover, the directly drivenwind turbines are supposed to operate at lower–speeds, wherethe output voltage may be very small with lower signal-to-noiseratio. On the other hand, the proposed ANFIS algorithm requiresless-accurate mathematical model of PMSG, where it sees thereal PMSG as a black box and tries to estimate the back EMF

Fig. 13. Schematic of experimental setup.

adaptively by forcing the estimated currents equal to actual cur-rents. Once the estimated currents are equal to actual currentsfor the same set of input voltages, the ANFIS model exactlyreplicates the actual model of PMSG, and hence, does not re-quire the exact model of PMSG parameters. The proposed modelautomatically takes the parameter variation and distortion in sig-nal into account and modifies the estimated signal adaptively.The aforementioned facts are verified both in simulation andalso with experimental results, where the ANFIS algorithm out-performs the conventional back-EMF method under dynamicoperating conditions with nominal and 20% variation in PMSGparameter.

The performance of grid-side inverter is shown in Fig. 11,where the grid-side currents, as shown in Fig. 11(b), are alwayskept as a balanced set of sinusoidal currents at UPF, even inthe presence of 3P4W unbalanced nonlinear current shown inFig. 11(c). The inverter supplies the load harmonics, load un-balance, load-neutral current, and generated power from PMSG,as shown in Fig. 11(b). Initially, the inverter current is constantand it starts increasing at t = 0.2 s in response to increase inPMSG input torque. As the inverter-injected current is morethan the load-current demand, the difference of these is beinginjected into the grid with a sinusoidal and balanced profile. Att = 0.425 s, the input torque starts decreasing, which results incorresponding decrease in inverter and grid currents.

In Fig. 12, it is shown that inverter current consists of load-current harmonics component, load-current reactive component,and the active-current component in proportion to generatedwind power. Moreover, the inverter is fully able to supply theload-neutral current locally, resulting in the zero-grid neutralcurrent, and hence, balanced grid operation. On visualizing theprofile of all traces carefully in Fig. 12, it can be easily noticedthat the grid phase-a current is in phase opposition to grid phase-a voltage. This indicates that the power is being injected intothe grid at UPF.

IV. EXPERIMENTAL RESULTS AND DISCUSSION

A scaled hardware prototype is developed and implemented,as shown in Fig. 13. The PMSG is mechanically driven bya three-phase induction motor coupled with ACS800 drivefrom ABB. The variable torque caused by variable wind speed


Fig. 14. Experimental results for PMSG voltage and current.

Fig. 15. Experimental results with nominal PMSG parameters. (a) ANFISestimator-based results. (b) EMF-estimator-based results.

Fig. 16. Experimental results with 20% variation in PMSG resistance andinductance: (a) ANFIS estimator-based results and (b) EMF-estimator-basedresults.

command generated from MATLAB/Simulink file is applied tothe drive, which, in turn, rotates the PMSG at variable speed.The output of PMSG is connected to grid by two back-to-backconnected inverters with a common dc link.

The proposed ANFIS speed and position estimation and con-trol algorithm is implemented using dSPACE 1104. An exten-sive experimental study is carried out to validate the proposedstrategy. The experimental results are shown in Figs. 14–18.

A. Generator-Side Control Performance

The generator-side experimental results are shown inFigs. 14–16. The variable amplitude and frequency of PMSGvoltage, as shown in Fig. 14, indicate its variable-speed oper-ation, where the speed is varied from 12% to 22% of its ratedspeed keeping in mind the low-speed operation of real windturbine. The PMSG speed and position is estimated in real time.The experimental results based on ANFIS estimator and EMFestimator with nominal and varying parameters are shown inFigs. 15 and 16, respectively. The ANFIS estimator is able to


Fig. 17. (a) Traces of grid currents just before and after compensation.(b) Traces of inverter currents just before and after compensation. (c) Tracesof grid voltage, grid currents, load current, and inverter current just before andafter compensation.

estimate the PMSG position and speed accurately under nominaland varying parameter conditions, as shown in Figs. 15(a) and16(a), respectively, while the EMF method shows poor estima-tion performance, as shown in Fig. 15(b). The performance ofEMF estimator further deteriorates with the variation in PMSG

Fig. 18. (a) Traces of grid voltage, grid current, load current, and invertercurrent. (b) Traces of grid voltage, grid current, load current, and inverter current.

parameters and also loses control at regular interval under dy-namic operating conditions, as shown in Fig. 16(b).

B. Grid-Side Control Performance

The performance of grid-side inverter is shown in Figs. 17and 18, where the inverter is utilized as multifunction device. InFig. 17, the experimental results are shown only for active fil-tering application of grid-interfacing inverter. The grid currentprofile is shown in Fig. 17(a), just before and after compensation.Here, it can be noticed that the grid is supplying a highly non-linear 3P4W unbalanced current before compensation, whichbecomes a perfectly balanced set of sinusoidal currents once af-ter the inverter is connected in the circuit. The inverter-injectedcurrents are shown in Fig. 17(b), where the inverter not onlysupplies the nonlinear unbalanced component of load current,but also compensates the load-neutral current demand locally.This allows the grid to supply only the active balanced com-ponent of load current. Moreover, the traces of grid voltage,grid current, load current, and inverter currents are shown inFig. 17(c), where the grid currents becomes perfectly sinusoidal


TABLE ISUMMARY OF INVERTER COMPENSATION CAPABILITIES

and in phase with the grid voltage once after the inverter startsinjecting the nonlinear, unbalanced, and reactive component ofload current.

In Fig. 18, the performance of grid-side inverter is shown,both in active filtering mode and renewable power injectionmode simultaneously. The grid current is purely sinusoidal andstarts decreasing with the increase in inverter-injected current,as shown in Fig. 18(a). This indicates that the grid is partiallysupporting the active-power load demand, while the inverter issupplying the rest of active load demand, load harmonics, andload reactive-power demand simultaneously. This fact can beeasily verified from Fig. 18(b), where the grid current startsdecreasing and becomes purely sinusoidal and in phase withgrid voltage, once after the inverter starts injecting the currentin the network. A comparative table showing the total harmonicdistortions (THDs) and unbalance factor (UF) before and aftercompensation is given (see Table I), where the percentage UFis calculated separately for each phase as follows:

%UFabc =|iabc − iavg . |

iavg .× 100.

V. CONCLUSION

The paper has presented a novel ANFIS-based speed- andposition-estimation algorithm for the sensorless control ofPMSG. The simulation and experimental results have been pro-vided under variable-speed operation. The estimator is able toestimate the rotor speed and position accurately under bothsteady-state and dynamic conditions. A detailed comparison isalso provided in between conventional back EMF and proposedANFIS algorithm under nominal and 20% variation in PMSGparameters. The proposed ANFIS algorithm outperforms theback-EMF method, where ANFIS estimator shows great immu-nity against parameter variation. It has also been demonstratedthat the grid-side inverter can perform all the duties of shuntAPF while maintaining the smooth bidirectional power flowsimultaneously. The simulation results supported by experimen-tal results are provided to validate the fact that the grid-side

inverter can act as multioperation device in order to utilize itsmaximum rating. The current unbalance, current harmonics, andload-reactive power demand of an unbalanced 3P4W nonlinearload at PCC are compensated effectively such that the grid cur-rents are always maintained as a balanced set (0% UF) of sinu-soidal currents (2.7% THD) at UPF. Moreover, the load-neutralcurrent is compensated locally by the fourth leg of inverter tokeep the zero neutral current on grid-side, and hence, resultingin balanced grid operation. The grid-interfacing inverter injectsthe wind-generated power, load-reactive power, load unbalance,and load harmonics simultaneously, and hence, enables the gridto supply/absorb only fundamental current at UPF.

APPENDIX

SYSTEM PARAMETERS

REFERENCES

[1] F. Blaabjerg, R. Teodorescu, M. Liserre, and A. V. Timbus, “Overviewof control and grid synchronization for distributed power generation sys-tems,” IEEE Trans. Ind. Electron., vol. 53, no. 5, pp. 1398–1409, Oct.2006.

[2] J. Morren and S. W. de Hann, “Ride-through of wind Turbines with doubly-fed induction generator during a voltage dip,” IEEE Trans. Energy Con-vers., vol. 20, no. 2, pp. 435–441, Jun. 2005.

[3] G. Saccomando, J. Svensson, and A. Sannino, “Improving voltage distur-bance rejection for variable-speed wind turbines,” IEEE Trans. EnergyConvers., vol. 17, no. 3, pp. 422–428, Sep. 2002.

[4] T. Senjyua, S. Tamakia, E. Muhandoa, N. Urasakia, H. Kinjoa,T. Funabashib, H. Fujitac, and H. Sekinea, “Wind velocity and rotor po-sition sensorless maximum power point tracking control for wind genera-tion system,” in Proc. 30th Annu. Conf. IEEE Ind. Electron. Soc. (IECON2004), vol. 2, pp. 1957–1962.

[5] S. Taniguchi, S. Mochiduki, T. Yamakawa, S. Wakao, K. Kondo, andT. Yoneyama, “Starting procedure of rotational sensorless PMSM in therotating condition,” IEEE Trans. Ind. Appl., vol. 45, no. 1, pp. 194–202,Jan./Feb. 2009.

[6] S. Bolognani, L. Tubiana, and M. Zigliotto, “Extended Kalman filtertuning in sensorless PMSM drives,” IEEE Trans. Ind. Appl., vol. 39,no. 6, pp. 1741–1747, Nov./Dec. 2003.

[7] Z. Chen, M. Tomita, S. Doki, and S. Okuma, “An extended electromo-tive force model for sensorless control of interior permanent-magnet syn-chronous motors,” IEEE Trans. Ind. Electron., vol. 50, no. 2, pp. 288–295,Apr. 2003.

[8] P. P. Acarnley and J. F. Watson, “Review of position-sensorless operationof brushless permanent-magnet machines,” IEEE Trans. Ind. Electron.,vol. 53, no. 2, pp. 352–362, Apr. 2006.

[9] S. Bhowmik and R. Spee, “Wind speed estimation based variable speedwind power generation,” in Conf. Rec. IEEE IECON 1998, Aachen,Germany, pp. 596–601.


[10] K. Tan and S. Islam, “Optimum control strategies in energy conversion ofPMSG wind turbine system without mechanical sensors,” IEEE Trans.Energy Convers., vol. 19, no. 2, pp. 392–399, Jun. 2004.

[11] H. Li, K. L. Shi, and P. G. McLaren, “Neural-network-based sensor-lessmaximum wind energy capture with compensated power coefficient,”IEEE Trans. Ind. Appl., vol. 41, no. 6, pp. 1548–1556, Nov. 2005.

[12] J. Shing and R. Jang, “ANFIS: Adaptive-network-based fuzzy inferencesystem,” IEEE Trans. Syst., Man, Cybern., vol. 23, no. 3, pp. 665–685,May 1993.

[13] N. Boccard, “Capacity factor of wind power realized values vs. estimates,”Elsevier Trans. Energy Policy, vol. 37, pp. 2679–2688, 2009.

[14] Capacity Factors at Kansas Wind Farms Compared With Total State Elec-trical Demand, July 2007 to June 2008. (2009, Jun.). [Online]. Avail-able: http://www.kcc.state.ks.us/energy/charts/Wind Capacity Factors atKansas Wind Farms Compared with Total State Electrical Demand.pdf.

[15] B. K. Bose, Modern Power Electronics and AC Drives. Upper SaddleRiver, NJ: Prentice-Hall, 2002.

[16] M. Singh and A. Chandra, “Power maximization and voltage sag/swellride-through capability of PMSG based variable speed wind energy con-version system,” in Proc. 34th Annu. Conf. IEEE Ind. Electron. Soc.(IECON 2008), Orlando, FL, pp. 2206–2211.

[17] S. Rahmani, A. Hamadi, N. Mendalek, and K. Al-Haddad, “Implemen-tation and performance of cooperative control of shunt active filters forharmonic damping throughout a power distribution system,” IEEE Trans.Ind. Electron., vol. 56, no. 8, pp. 2904–2915, Aug. 2009.

[18] V. Khadkikar, A. Chandra, and B. N. Singh, “Generalised single-phase p-qtheory for active power filtering: Simulation and DSP-based experimentalinvestigation,” IET Trans. Power Electron., vol. 2, no. 1, pp. 67–78, Jan.2009.

[19] P. Rodrıguez, J. I. Candela, A. Luna, L. Asiminoaei, R. Teodorescu, andF. Blaabjerg, “Current harmonics cancellation in three-phase four-wiresystems by using a four-branch star filtering topology,” IEEE Trans.Power Electron., vol. 24, no. 8, pp. 1939–1950, Aug. 2009.

[20] J. Dai, D. Xu, and B. Wu, “Implementation and performance of cooperativecontrol of shunt active filters for harmonic damping throughout a powerdistribution system,” IEEE Trans. Power Electron., vol. 24, no. 4, pp. 963–972, Mar./Apr. 2009.

[21] J. H. R. Enslin and P. J. M. Heskes, “Harmonic interaction between alarge number of distributed power inverters and the distribution network,”IEEE Trans. Power Electron., vol. 19, no. 6, pp. 1586–1593, Nov. 2004.

[22] A. Timbus, M. Liserre, R. Teodorescu, P. Rodriguez, and F. Blaabjerg,“Evaluation of current controllers for distributed power generation sys-tems,” IEEE Trans. Power Electron., vol. 24, no. 3, pp. 654–664, Mar.2009.

Mukhtiar Singh (S’08) received the B.Tech. andM.Tech. degrees in electrical engineering from Na-tional Institute of Technology (formerly known asR.E.C. Kurukshetra), Kurukshetra, India, in 1999 and2001, respectively. He is currently working towardthe Ph.D. degree at Ecole de Technologie Superieure,Universite du Quebec, Montreal, QC, Canada, un-der the National Overseas Scholarship, funded by theGovernment of India.

He was a faculty member at B.M.I.E.T., Sonepat,India, and K.I.E.T, Ghaziabad, India, during 2000–

2002. Since 2002, he has been an Assistant Professor in the Department ofElectrical Engineering, Deenbandhu Chhoturam University of Science andTechnology, Sonepat, India. Currently, he is on study leave. His research inter-ests include renewable energy sources, power quality, energy storage systems,electric vehicles, and power electronics and drives.

Ambrish Chandra (SM’99) was born in India in1955. He received the B.E. degree from the Univer-sity of Roorkee [currently the Indian Institute of Tech-nology (IIT)], Roorkee, India, in 1977, the M.Tech.degree from the IIT, New Delhi, India, in 1980, andthe Ph.D. degree from the University of Calgary, AB,Canada, in 1987.

He was a Lecturer, and later, was a Reader at theUniversity of Roorkee. Since 1994, he has been aProfessor in the Department of Electrical Engineer-ing, Ecole de Technologie Superieure, Universie du

Quebec, Montreal, QC, Canada. His current research interests include powerquality, active filters, static reactive-power compensation, flexible ac transmis-sion systems (FACTS), and renewable energy resources.

Dr. Chandra is a member of the Ordre des Ingenieurs du Quebec, Canada.

Documents

my paper