Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

Received: 25 July 2017 Revised: 3 November 2017 Accepted: 7 December 2017

DOI: 10.1002/etep.2526

RESEARCH ARTICLE

Control for grid‐connected DFIG‐based wind energy system


using adaptive neuro‐fuzzy technique

K.V. Shihabudheen1 | S. Krishnama Raju2 | G.N. Pillai1

1
Indian Institute of Technology Roorkee,
Summary
Roorkee, India
2
Sree Vidyanikethan Engineering College,
Smooth operation and control of power electronic converters are essential to
Tirupati, India ensure wind energy systems in compliance with modern grid codes. This paper
proposes an intelligent adaptive control strategy for doubly fed induction
Correspondence
KV Shihabudheen, Indian Institute of generator–based wind energy system using recently proposed extreme learning
Technology Roorkee, Roorkee 247667, adaptive neuro‐fuzzy inference system (ELANFIS). ELANFIS is a type of neuro‐
India.
fuzzy systems, which combines erudition capabilities of extreme learning
Email: shihab4806@gmail.com;
shih1dee@iit.ac.in machine and unambiguous knowledge of fuzzy systems. In ELANFIS, premise
parameters are generated randomly with restraints to house fuzziness and
consequent parameters are identified using Moore‐Penrose generalized inverse
method. The vector control with proposed ELANFIS control strategy is tested
under various contingencies and is able to handle the uncertainties in the wind
speed and grid disturbance. The performance of the proposed technique is ver-
ified through real‐time digital simulator with hardware in loop configuration.

KEYWORDS
doubly fed induction generator, extreme learning adaptive neuro‐fuzzy systems, hardware in loop
(HIL), real‐time digital simulator (RTDS)

1 | INTRODUCTION

Renewable energy harvesting and utilization are widely popularized around the globe due to its long‐time economic
aspects and, most importantly, concerns about the environment. Out of the different renewable energy sources, wind
energy and solar energy are preferred for supplying small electrical loads in remote locations. Use of doubly fed induc-
tion generator (DFIG) in wind generation has drawn increased attention due to reduced losses, the ability to control
power factor and power easily, and better performance in wide chaotic and random variation of wind speed.1,2 Other

List of symbols: vs, stator voltage; vr, rotor voltage; Rs, stator resistance; Rr, rotor resistance; ωr, rotor speed; ωs, synchronous angular speed; λs, stator
flux; λr, rotor flux; ls, stator inductance; lr, leakage inductance; Lm, mutual inductance; vrd, vrq, d‐axis and q‐axis rotor voltages; vsd, vsq, d‐axis and q‐axis
stator voltages; λrd, λrq, d‐axis and q‐axis rotor fluxes; λsd, λsq, d‐axis and q‐axis stator fluxes; ird, irq, d‐axis and q‐axis rotor currents; isd, isq, d‐axis and
q‐axis stator currents; P, stator reactive power; Q, rotor reactive power; J, moment of inertia; Tm, mechanical torque; Te, electromagnetic torque; Ta,
wind turbine power; pp, pole pair; γ, air density; R, radius of the turbine; vw, wind velocity of the turbine; Cp, power coefficient; L, number of rules;
xj, jth crisp input; Aij, linguistic variable for the jth input and ith rule.; n, dimension of input; m, dimensional output; N, number of training data;
pil(l = 0, 1, …, n), real valued consequent parameters; (aj*, bj*, cj*), parameters of uniformly distributed membership function; (aij, bij, cij), premise
parameters corresponding to jth input variable for ith rule; μAik , membership grade; βi, crisp value of ELANFIS output for ith rule; wi, firing
strength of ith rule; wi , normalized firing strength of ith rule; tj, target output for jth data; y, output of ELANFIS model; H, fuzzy layer output
matrix; P, consequent parameter matrix; T, target matrix; R, range; dcc, distance between 2 adjacent centers of uniformly distributed membership
function; h, number of membership functions; e, error; Δe, change in error

Int Trans Electr Energ Syst. 2018;e2526. wileyonlinelibrary.com/journal/etep Copyright © 2018 John Wiley & Sons, Ltd. 1 of 18
https://doi.org/10.1002/etep.2526
2 of 18 SHIHABUDHEEN ET AL.

advantages of the DFIG topology are (1) converters with 20% to 40% of the rated power of the generator that is only
required for the operation, (2) efficient power capture, and (3) reduced mechanical stress.3
However, the DFIG‐based wind energy system is highly sensitive to voltage fluctuation that is caused by faults or load
changes in the distribution system. The system fault causes voltage reduction in the DFIG terminal, which leads to over
currents in stator terminals. This may lead to damage of the stator winding; moreover, such increasing current can affect
the rotor side converter (RSC) and direct current (DC) link due to a magnetic link between stator and rotor.4 Hence, an
accurate power and voltage controller should be required to maintain the system in steady level during uncertain
conditions such as fault and wind speed variations. In modern transmission systems, to maintain the stability of the grid,
wind energy systems are expected to remain connected to the grid, even in the presence of fault.5,6 The power and voltage
oscillations in the network after clearing the fault are major concerns in designing a controller for grid‐connected DFIG.
The schematic diagram of wind turbine coupled with DFIG is shown in Figure 1. The turbine is coupled to a gear-
box to attain the rated speed of the generator. Through an isolation transformer, stator of the generator is directly con-
nected to the grid and it delivers the major portion of the electric power generated from wind. The rotor is connected
through frequency converters, which comprises two self‐commutated pulse‐width modulation converters, known as RSC
and grid side converter (GSC). The capacitor acts as a DC link source, decoupling the operation of the two converters. The
RSC controls the rotor voltage to track the reference power inputs, and grid side converter maintains the DC link voltage
at a constant level.
Vector control is a widely used technique for independent control of active and reactive powers.7 The control of active
and reactive powers can easily be achieved with rotor current controller, via adjustment of rotor excitation voltages. Most
of the conventional control schemes are based on voltage oriented or stator flux control, and the reference stator and
rotor control parameters tracked use conventional proportional‐integral (PI) controllers to previous studies.8-10 A new
PI‐based rotor current control strategy using positive and negative (dq) reference frames is developed to provide precise
control of the rotor positive and negative sequence currents.11 A PI control–based DC vector control scheme that
produces an integrated control strategy for wind energy extraction, grid voltage control, and reactive power control is
proposed.12 A combined vector control and direct power control are developed for RSC of DFIG.13 Combination of these
two techniques provides fast response and simple implementation. However, in all the above schemes, turning the PI gains
is a challenging task when the plant is nonlinear and parameters are uncertain. It requires the knowledge of dynamic
modeling and the behavior of the DFIG.14 Proportional‐integral controllers that are designed for a particular operating
point have poor performance in transient conditions.
Nonlinear control of DFIG wind energy systems is one of the emerging technique for accurate control performance.
Sliding mode control15,16 is developed to regulate the rotor speed of DFIG systems. Mechter et al17 developed a
backstepping controller for accurate control of active and reactive powers in DFIG systems. In Yang et al,18 a feedback
linearization technique is implemented for maximum power point tracking (MPPT) of wind energy in DFIG‐connected
wind energy distribution system. Guo et al19 designed a reactive power control for DFIG wind farm using dynamic
programming‐based optimal control. An adaptive control technique is designed to optimize the extracted power from
the wind turbine and also regulate the reactive power to meet grid requirements in Meng et al.20 But all the above
techniques require an accurate dynamic model of DFIG for designing the control strategies.

FIGURE 1 Schematic diagram of doubly fed induction generator (DFIG)–based wind turbine system. GSC, grid side converter; RSC, rotor
side converter
SHIHABUDHEEN ET AL. 3 of 18

Accurate model of the generator is not required for designing an intelligent control system based on computational
intelligence. Among the computational intelligence techniques, fuzzy logic and artificial neural network (ANN) appear
as powerful options for identification and control of nonlinear dynamic systems such as power control systems. Soares
et al21 propose neural networks–based control of DFIG system. The neural networks are used to generate the reference
parameters to the control system by voltage and power regulators. The proposed neural network controller produced
good flexibility and adaptability to generate active and reactive powers. In Medjber et al,22 an MPPT control scheme
using neural network is proposed for optimum power transfer between DFIG machine and the grid using the indirect
vector control. A multilayer perceptron‐based power control of DFIG allows the converter connected to the rotor termi-
nals to operate with constant switching frequency.23 Later, adaptive neural networks control24 and robust power control
with generalized neural networks25 are suggested to achieve better efficiency and robustness. However, most of the
neural network–based control schemes use gradient‐based techniques to tune the parameters, but the learning using
gradient descent consumes more time for the training and can easily converge to local minima due to improper learning
steps. A fuzzy logic control for controlling RSC of grid‐connected DFIG wind energy system was proposed in DeAlmeida
et al.26 The active and reactive power flow from the machine to the grid is controlled using manual tuned fuzzy control
and obtained a reasonable result. Many authors have used fuzzy logic for fault ride through control,27 self‐tuning
control,28 stability control,29 decentralized reactive power control,30 and frequency control31 of grid‐connected DFIG
system. Type‐2 fuzzy logic–based PI control is also implemented for considering uncertainty in grid connection.32,33
The fuzzy logic control suffers from limitations such as selection and tuning of membership functions and rules to
achieve robust performances. On the other hand, ANN has limitations because of their black box nature, difficulty in
selecting the best structure and size, and the considerable training time to solve a specific problem. In neuro‐fuzzy
systems such as adaptive network–based fuzzy inference system (ANFIS), learning capability of ANN and knowledge
representation of fuzzy systems are judiciously combined to form a more accurate intelligent system.34,35 ANFIS model-
ing produces smoothness in output behavior and adaptability compared to neural networks and fuzzy modeling and
hence can be used for modeling and control.36 In Jabr et al,14 a vector control scheme using ANFIS techniques is devel-
oped to control the rotor side voltage source converter that controls independently the generated active and reactive
power as well as the rotor speed to track the maximum wind power point. A neuro‐fuzzy–based power control with
wireless coded orthogonal frequency division multiplexing for a variable speed DFIG wind energy generation system
was developed to achieve robustness and reliability of the generation system.37 ANFIS uses hybrid learning algorithm,
which is a combination of steepest descent and least squares estimation to find premise parameters and consequent
parameters during the learning stage. But learning using gradient technique can easily converge to local minima. More-
over, hybrid learning algorithm has strong computational complexity, and hence, training of ANFIS becomes slow.
In the past few years, extreme learning machines (ELM) have been extensively used for regression and control
problems.38-40 In ELM, the hidden layer parameters are randomly selected and the output layer parameters are
determined by minimum norm least squares estimation methods, resulting in negligible computation time and better
generalization. In Mahmoud et al,41 an ELM‐based intelligent control technique is proposed to increase the extracting
energy as well as to improve the power system performance using signal decomposition and grouping. It develops an
optimal wind turbine control with ensemble prediction technique by combining the Complimentary ensemble empirical
mode decomposition with adaptive noise (CEEMDAN) wind speed decomposition and ELM. An intelligent islanding detec-
tion and classification are performed in grid‐connected DFIG wind farms using ELM in Patnaik et al.42 In ELM‐based neuro‐
fuzzy approaches, the advantages of ELM and Takagi‐Sugeno‐Kang (TSK) fuzzy inference system can be integrated to achieve
the best performance in neuro‐fuzzy systems. Recently, a novel neuro‐fuzzy algorithm called extreme learning adaptive
neuro‐fuzzy inference system (ELANFIS) was proposed by incorporating ELM techniques into ANFIS structure.43,44 The
premise parameters of the ELANFIS are randomly generated to achieve faster learning speed without sacrificing generaliza-
tion capability. Extreme learning machine theory is applied to tune the parameters of fuzzy rules. In ELANIS, the random-
ness in the premise parameter tuning is reduced by incorporating explicit knowledge representation embedded in the fuzzy if
then rules. Like ELM, ELANFIS can also avoid the difficulties faced by gradient descent methods like stopping to local
minima, stopping criteria, and overfitting, which are also encountered in regular ANFIS algorithm. The effectiveness and
performance of ELANFIS are well tested for regression,44-46 control problems,43 and prediction of landslide displacement.47
This paper proposes ELANFIS‐based novel adaptive control technique for controlling the RSC of grid‐connected
DFIG. The proposed ELANFIS controllers are trained based on the performance of different PI controllers tuned for dif-
ferent operating conditions. Unlike the PI controller, the proposed ELANFIS controller gives robust performance under
the wide range of operating conditions. Further, the controller performance is verified with DFIG model implemented in
RSCAD software. Real‐time simulations are performed and tested under various possible contingencies.
4 of 18 SHIHABUDHEEN ET AL.

Initially, the mathematical model of the DFIG is derived and implemented in RSCAD environment, which is
processed through real‐time digital power system simulator (RTDS). Then ELANFIS controller is designed for RSC
and implemented in a digital signal processor–based Dspace module. Feasibility of the controller for real‐time applica-
tion is analyzed by sending error signals of the rotor currents from RSCAD model (after converting it into analog signals)
as controller input. Then, it is processed through the proposed ELANFIS controller with hardware in loop (HIL)
configuration and fed back as actuating input to RTDS. The performance of the proposed controlled system is compared
to PI controller for a short‐circuit 3‐phase fault and single‐phase fault conditions and also under variable wind speed
conditions.
This paper is organized as follows: Section 2 describes the mathematical modeling of DFIG system. Section 3 explains
the basic theories and learning algorithms of ELANFIS. Section 4 describes the proposed controller architecture and
training procedure. Real‐time simulation analyses are discussed in Section 5, and conclusions are given in Section 6.

2 | D F I G N O N L I N E A R MO D E L I N G

The mathematical modeling equations of DFIG are expressed in the space vector form. The equivalent circuit of DFIG in
the arbitrary reference frame is shown in Figure 2
where vs and vr are the stator and rotor voltages, respectively, Rs and Rr are the stator and rotor resistances,
respectively, ωr is the rotor speed, and ωs the synchronous angular speed. ls and lr are the stator and leakage inductances,
and Lm is the mutual inductance. λs and λr are the stator and rotor fluxes, which can be represented as


λs ¼ Ls is þ Lm ir
; (1)
λr ¼ Lr ir þ Lm is

where Ls = ls + Lm and Lr = lr + Lm.


Now, stator and rotor voltages can be represented as


vs ¼ Rs is þ dλs =dt þ jωs λs
: (2)
vr ¼ Rr ir þ dλr =dt þ jðωs −ωr Þλr

Since the synchronous reference frame model is widely used for control analysis, the above model is transferred to
the synchronous reference frame and the dq axis components of flux are obtained as
(
λsd ¼ Ls isd þ Lm ird
; (3)
λsq ¼ Ls isq þ Lm irq

(
λrd ¼ Lr ird þ Lm isd
: (4)
λrq ¼ Lr irq þ Lm isq

FIGURE 2 Equivalent circuit of doubly fed induction generator in an arbitrary reference frame rotating at synchronous speed
SHIHABUDHEEN ET AL. 5 of 18

The dq axis components of stator and rotor voltages become


(
vsq ¼ Rs isq − ωs λsd þ d λsd =dt
; (6)
vsd ¼ Rs isd þ ωs λsq þ d λsq =dt

(
vrd ¼ Rr ird þ d λrd =dt − ðωs − ωr Þλrd
: (7)
vrq ¼ Rr irq þ d λrq =dt þ ðωs − ωr Þλrq

The stator active and reactive powers can be derived as


 
P ¼ 1:5 vsd isd þ vsq isq ; (8)

 
Q ¼ 1:5 vsq isd − vsd isq : (9)

Swing equation of the machine can be written as

dωr 1
¼ ðT m − T e Þ; (10)
dt J
where Tm and Te are the mechanical torque and electromagnetic torque, respectively, and expressed as9

Ta 1
γπR2 Cp v3w
Tm ¼ ; Ta ¼ 2
; (11)
pp ωr

 
T e ¼ pp:Lm isq ird − isd irq ; (12)

where Ta is the wind turbine power, pp is the pole pair, γ is the air density, and R and vw are the radius and wind velocity
of the turbine. Cp is the power coefficient, and ωr is the rotor angular speed.

3 | EXTREME LE AR N I N G A N FI S ( E L A NF I S)

ELANFIS is developed by integrating ELM techniques with conventional ANFIS system. The premise parameters
are generated randomly with restraints to house fuzziness and using Moore‐Penrose generalized inverse method
consequent parameters are generated. Consider a TSK‐type fuzzy system, with L rules; the rules are typically of
the form:

Rule Ri : IF ðx 1 is Ai1 Þ and ðx 2 is Ai2 Þ and … andðx n is Ain Þ;


THEN βi ¼ pi0 þ pi1 x 1 þ pi2 x 2 þ … þ pin x n ; (13)
where i ¼ 1; 2; :…; L;

where xj is the jth crisp input and Aij is the linguistic variable for the jth input, n represents the dimension of
input, and pil(l = 0, 1, …, n) are the real valued parameters.
The structure of ELANFIS can be explained with the help of Figure 3. It consists of five layers, which includes input
layer, membership function layer, normalization layer, consequent parameter layer, and output layer.

Input layer: Each node of this layer indicated input variables, and then it transferred to membership
function layer.
Membership function layer: In this layer, each node represents a fuzzy membership function. The bell
shaped membership functions are used in membership function layer because
membership grades vary smoothly and have nonzero membership values at all
points. The mathematical representation of bell shape membership function is
given as
6 of 18 SHIHABUDHEEN ET AL.

FIGURE 3 Extreme learning adaptive neuro‐fuzzy inference system architecture

  1
g x j ; a; b; c ¼   ; (14)
x − c 2bij
1 þ  j aij ij 

where aij, bij, and cij are the premise parameters. cij determines position, and aij and bij determine the shape corresponding to jth
input variable for ith rule. In ELANFIS, the parameters of the membership function (premise parameters) aij, bij, and cij are ran-
domly selected with some constraints and consequent layer parameters pil are learned by least square estimation method.
 
The membership grade of Aik for the input xj is represented μAik x j , and it can be given as

   
μAik x j ; ajk ; bjk ; cjk ¼ g x j ; ajk ; bjk ; cjk : (15)

Normalization layer: In this layer, each firing strength of fuzzy rule is normalized by its total sum.
Consequent parameter layer: In this layer, the neural network part is embedded with consequent parameters and its
output is represented by wi ðx Þβi .
Output layer: It calculates the defuzzified output of the total system. Hence, the final output can be
represented by (21).

The firing strength of ith rule can be calculated by

wi ðx Þ ¼ μAi1 ðx 1 Þ ⊗ μAi2 ðx 2 Þ ⊗ … ⊗ μAin ðx n Þ; (16)

where ⊗ indicates and operator of the fuzzy logic. Since T‐norm (triangular norm) product gives a smoothing effect, it is
used to obtain the firing strength of a rule by performing the and on the membership grades of premise parameters. The
normalized firing strength of the ith rule can be found by

w i ðx Þ
wi ðx Þ ¼ L
; (17)
∑ w i ðx Þ
i¼1

“THEN” part or consequent layer is a linear neural network with pi as weight parameters. The system output can be
obtained by calculating the weighted sum of the output of each normalized rule. The system output is calculated as
SHIHABUDHEEN ET AL. 7 of 18

L
∑ βi wi ðx Þ L
y ¼ i¼1L ¼ ∑ βi wi : (18)
i¼1
∑ wi ðx Þ
i¼1

In an ideal case when there is no training error, the output of ELANFIS can be represented as

  L  
y x j ¼ ∑ βi wi x j ¼ t j for j ¼ 1; 2; …; N; (19)
i¼1

where tj is the respective target output vector and N is the number of arbitrary training samples. In TSK model, conse-
quent parameter is the linear combination of input variables and is written as

βi ¼ x ja T pi ; (20)
h iT
where x ja ¼ 1; x Tj and pi is the parameter matrix for each ith rule and given by pi = [pi0, pi1, …, pin]T.
Hence, the output (18) can be written as

  L  
y x j ¼ ∑ x ja T pi wi x j ¼ t j for j ¼ 1; 2; …; N: (21)
i¼1

This can be written in matrix form

H N×hn ðnþ1Þ Phn ðnþ1Þ×1 ¼ T N×1 ; (22)

where h indicates the number of membership functions, n is the dimension of the input data, and hn is the maximum
number of rules used. The fuzzy layer output matrix H and the consequent parameter matrix P are given as
      
H ðx 1 ; x 2 ; …; x N Þ ¼ x ja T w1 x j ; x ja T w2 x j ; …; x ja T wL x j ; P ¼ ½p1 ; p2 …; pL T ; and T ¼ ½t 1 ; t2 …; t N T : (23)

3.1 | ELANFIS learning algorithm


In ELANFIS, the ELM strategy is applied to tune the parameters. The premise parameters (aj, bj, and cj) are generated
arbitrarily in a constrained range. Consider N training data, with n attribute inputs [X1 X2 . . Xn]N × n and m dimen-
sional outputs ½ T 1 T 2 : : T n N×m . Now, the range of input can be defined as

Ri ¼ MAX fX i g − MIN fX i g for i ¼ 1; 2; …:n; (24)

where MAX{.} and MIN{.} indicate the maximum and minimum values of the function.
Step 1. Randomly select the premise parameters (aj, bj, and cj) according to the constraint ranges. The idea is to
use the qualitative knowledge embedded in the linguistic variables of the premise part of the fuzzy rules to
determine the parameters and hence to reduce the randomness in parameter tuning. The parameter cj is the
center of bell membership function. The initial centers (cj*) are selected such that the range of input is divided
into equal intervals. The parameter cj is randomly taken in the range such that the adjacent centers do not cross
each other.

dcc dcc
c*j − < cj < c*j þ : (25)
2 2

dcc is the distance between two adjacent centers of uniformly distributed membership function.
Based on the range of universe disclosure and number of membership functions, the ranges of other parameters aj
and bj are selected as follows:
8 of 18 SHIHABUDHEEN ET AL.

The premise parameter aj determines the membership function width. For h uniformly distributed membership
functions, the initial parameter aj* of the jth membership function is obtained as

Rangei
aj * ¼ : (26)
2h − 2

Based on the calculated value for the above parameter aj*, the premise parameter aj is randomly taken in the range

a*j 3a*j
≤ aj ≤ : (27)
2 2

The parameters aj and bj give the slope of membership function as bj/2aj. The initial value of bj in uniformly
distributed membership function is taken as 2. Even a minute change in bj drastically changes the slope of the
membership function; hence, the variation from initial value is restricted to ±0.1. So the range of bj is given by

1:9 ≤ bj ≤ 2:1: (28)

Step 2. Once randomly generated membership functions are available, final output can be written as linear
combination of consequent parameters. Find out H matrix using (23).
Step 3. Calculate the consequent parameter matrix P, which can be designed by finding Moore‐Penrose generalized
inverse of H. Hence,

 
P ¼ H T H H T * T: (29)

4 | C O N T R O L L E R DE S IG N

The designed method aims to control DFIG output active and reactive powers by controlling rotor currents. This is done
by RSC by generating suitable pulses to converter switches. Generally, vector control is implemented in dq reference
frame due to high accuracy and short term overload capability. The control strategy with the stator voltage oriented
scheme using PI controller is shown in Figure 4. Initially, currents and voltages are converted from ABC to dq synchro-
nous reference frame. The MPPT scheme of wind turbine produces required reference active power from given wind
speed, which is then compared with the measured active power to generate required direct axis current reference.
Similarly, required quadrature axis reference current is generated from difference between reference reactive power
and measured reactive power. This current reference is then compared with actual current, and it produces control
inputs to PI controllers. The current controller produces required rotor voltages vrd and vrq to track the reference powers.
Proportional‐integral controllers are widely used to control the powers through RSC. However, PI controllers are not
robust enough to perform under different operating conditions. As PI controllers are designed based on a linearized
model, it cannot give satisfactory performance under complex highly nonlinear environment. Different PI gains should
be required with different operating conditions like fault and change in wind speed conditions to get better performance.
Moreover, design and selection of PI gains are difficult, especially under nonlinear conditions.
In Jabr and Kar,48 a fuzzy algorithm is developed to generate PI gains to generate excellent control performance. The
fuzzy rules are used to generate the proportional and integral gains of the controller. The rules are designed based on

FIGURE 4 The control scheme for rotor side converter with proportional‐integral (PI) controller. MPPT, maximum power point tracking
SHIHABUDHEEN ET AL. 9 of 18

quantitative knowledge and extensive simulations. The trade‐off between performance and robustness is not usually con-
sidered in fuzzy logic, because there is no effective method to tune the membership functions for minimizing output
error of the system. The proposed ELANFIS controller integrates learning capabilities of ELM networks and the explicit
knowledge of the fuzzy systems. Imprecision and uncertainty of the system being modeled are considered by fuzzy logic,
and adaptability is considered by ELM.
The structure of ELANFIS controller to tune PI gains is shown in Figure 5. The absolute value of the error signal
between reference rotor current and actual current is used to calculate the scheduled proportional and integral gains
using the ELANFIS model. To apply the vector control to DFIG system, four ELANFIS gain tuners, two for d‐axis and
two for q‐axis PI controllers, are trained off‐line. Each of these units performs proportional and integral gain tuning.
The absolute value of error signal and change in error are considered as an input of ELANFIS controller. The training
is performed using ELANFIS algorithm. The training data used are collected from extensive simulations of the vector
controlled DFIG system with various PI gains so that trained tuner can tune the PI gains online based on the knowledge
of the different PI controllers under different operating conditions such as 3‐phase fault, single‐phase fault, and different
wind speed conditions. The optimum value of membership function is selected by validation data.
A total of 5420 data samples are collected; of these, 4000 samples are used for training and the rest 1420 used for
validation. Then d‐axis and q‐axis ELANFIS controllers are trained with corresponding input‐output data pairs. Initially,
the ELANFIS networks are trained with two membership functions. The membership function obtained after the train-
ing is shown in Figure 6. It is shown that almost interpretable membership functions are obtained, which will cover
entire input range. Fuzzy rules are listed in (30) where the jth input variable is assigned two linguistic values SMALLj
and LARGEj. As there are two input variables, 22 = 4 fuzzy rules will form the knowledge base of the fuzzy inference
system. The total number of fitting parameters is 24 comprising 12 premise parameters and 12 consequent parameters.
The parameters of q‐axis ELANFIS controller for proportional gain tuning are shown in Table 1.

FIGURE 5 The control scheme with extreme learning adaptive neuro‐fuzzy inference system (ELANFIS) controller. PI, proportional‐
integral

FIGURE 6 The membership functions of ELANFIS controllers


10 of 18 SHIHABUDHEEN ET AL.

TABLE 1 Parameters of ELANFIS for q‐axis control

Premise Parameters Consequent Parameters


a b c pi qi ri

SMALL1 0.258 2.001 0.305 24.272 −42.785 −5.803


LARGE1 0.196 1.981 0.998 4.427 72.421 −40.679
SMALL2 0.217 2.100 0.235 −107.766 73.421 58.254
LARGE2 0.249 1.906 0.643 58.995 −54.585 −12.199

The number of membership functions is varied from 2 to 8, and the optimum value is obtained corresponding to
maximum validation accuracy. The best value of membership function is obtained as five. The training root mean square
errors of d‐axis ELANFIS controller with optimum membership function are obtained as 2.047e‐07 and 8.619e‐07. The
training root mean square errors of q‐axis ELANFIS controller with optimum membership function are obtained as
3.466e‐07 and 5.913‐07. The membership function obtained after the training with optimum membership function is
shown in Figure 7.

If e is SMALL1 and Δ e is SMALL2 ; then o ¼ p1 e þ q1 Δe þ r 1 ; (30)


If e is SMALL1 and Δe is LARGE2; then o ¼ p2 e þ q2 Δe þ r 2 ;
If e is LARGE1 and Δe is SMALL2; then o ¼ p3 e þ q3 Δe þ r 3 ;
If e is LARGE1 and Δe is LARGE2; then o ¼ p4 e þ q4 Δe þ r 4 :

The PI gains are inputs to the standard PI controller part of the vector controller to generate the control signals vrd
and vrq. The final output of the controller with compensation is defined as

v*rd ¼ vrd − ðωs − ωr Þ σLr irq ; (31)

 
v*rq ¼ vrq þ ðωs − ωr Þ Lm ims þ σLr irq ; (32)

where σ ¼ 1−L2m =Ls Lr , and the terms in brackets are the compensation terms required to completely decouple the d‐axis
and q‐axis voltages. Then these d‐axis and q‐axis voltages are converted to ABC frame, which is then sent to a pulse‐
width modulation (PWM) controller to generate switching pulses to the RSC.

FIGURE 7 Optimum membership functions of ELANFIS controller


SHIHABUDHEEN ET AL. 11 of 18

5 | R E A L‐ TIME SIMULATION

Real‐time digital simulator is a digital electromagnetic transient power system simulator that operates in the real time
with hardware interfacing facilities. It facilitates real‐time simulations of complex power systems modeled using parallel
processors that allow simulations with a time step up to 250 ns. To evaluate the performance of proposed ELANFIS
controller, real‐time simulations are performed on the grid‐connected DFIG system shown in Figure 8. The model
consists of 2 MW DFIG coupled with wind turbine connected to 230 kV bus through 0.69/22 and 22/230 kV
transformers. To simulate fault conditions, faults are applied at points A, B, and C.
The real‐time simulation model is designed in RSCAD based on Figure 8. Points A, B, and C are directly connected to
the fault initialization beaker, and rotor of DFIG is connected to crowbar circuit. To minimize the losses, converters with
40% of the machine rating has been used. To extract energy from rotor and stator, the rated DFIG speed is chosen to be
1.2 times synchronous speed.
Schematic diagram for proposed controller‐based real‐time simulation process using HIL arrangement is shown in
Figure 9. RSCAD model is implemented on the workstation computer, is compiled, and interacts with RTDS to execute
on the required number of processor modules. ELANFIS controller is trained using Matlab and build in Dspace module
for performing controller action. The difference in actual current and reference current generated in RSCAD platform is
converted in to ±10 V analog signal using gigabit transceiver analog output card, which consists of a 16‐bit digital to ana-
log converter. These analog signals are sent to ELANFIS controller. The output from ELANFIS controller is converted
into digital signal using gigabit transceiver analog input (GTAI) card. Twelve analog input channels included in GTAI
card with each channel configured with an input range of ±10 V. Compatibility between GTAI card output and RSCAD
is maintained by adjusting the scaling factors. The experimental setup is shown in Figure 10.

FIGURE 8 System model for real‐time simulations. DFIG, doubly fed induction generator

FIGURE 9 Schematic diagram for real‐time simulation process. ELANFIS, extreme learning adaptive neuro‐fuzzy inference system; GTAI,
gigabit transceiver analog input; GTAO, gigabit transceiver analog output; RTDS, Real time digital simulator
12 of 18 SHIHABUDHEEN ET AL.

FIGURE 10 Experimental setup for real‐time simulations

To simulate the system in different operating conditions, two different types of uncertainties are introduced in the
operating conditions. Firstly, 3‐phase short‐circuit and single‐phase faults are introduced to study the behavior of the
proposed method in different fault conditions. Secondly, the system is tested under variable wind speed conditions.

5.1 | Three‐phase fault


To analyze the effect of 3 phase short‐circuit fault on the proposed controller‐based DFIG system, a fault initialization
breaker is connected at point A as in Figure 6. The fault signal is applied for the duration of 0.15 seconds at time t = 0.2 sec-
onds. The response of the stator voltage in kilovolt and per unit, for PI controller and ELANFIS, is shown in Figure 11. In PI
case, during fault period, the stator voltage is dropped to 0.2 pu and distorted voltage is observed after clearing the fault. The
response of ELANFIS is better after clearing the fault. The response of active and reactive powers is shown in Figure 12;
damped oscillations are observed in both the cases. It is noted that the magnitude of the oscillations produced during fault
time is less in ELANFIS controller compared to PI case. In ELANFIS case, an improvement of 35.7% is noticed in terms of
reduction in the magnitude of oscillations in the active power. After clearing the fault, the power oscillations are settled in
0.92 seconds in case of PI controller and 0.8 seconds in case of ELANFIS controller.
Figure 13 shows the stator current, rotor current, and the DC link voltage of the system for both the controllers. The
stator and rotor currents change abruptly at the beginning of fault and decrease slowly during the fault period. Damping
of current oscillations after clearing the fault is an important factor for maintaining grid stability. In case of ELANFIS

FIGURE 11 Stator voltages of (A) proportional‐integral and (B) ELANFIS controller for 3‐phase fault
SHIHABUDHEEN ET AL. 13 of 18

FIGURE 12 Active and reactive powers of (A) proportional‐integral and (B) ELANFIS controller for 3‐phase fault

FIGURE 13 Currents and direct current (DC) link voltages of (A) proportional‐integral and (B) ELANFIS controller for 3‐phase fault

controller, stator and rotor currents are settled in 0.55 and 0.6 seconds, respectively, whereas in PI controllers, currents set-
tled after 0.7 and 0.9 seconds, respectively, with distorted magnitudes. After clearing the fault, oscillations are developed in
the DC link voltage; however, the ELANFIS controller was able to damp out these well before the settling time. It is observed
that ELANFIS controller was able to damp out the current and voltage oscillations without tripping the protection system.

5.2 | Single‐phase fault


To study the performance of the proposed controller for single‐phase fault condition, a line to ground fault is applied at
point C (in Figure 8) using a fault initialization breaker. Fault signal is applied for the duration of 0.08 seconds at time
14 of 18 SHIHABUDHEEN ET AL.

0.2 seconds. The performances of the PI controller and ELANFIS controller are shown in Figures 14 and 15. After
clearing the fault, in PI case, the power transients have reached to 6.9 MW, whereas the ELANFIS marked a 46.7%
improvement in the peak overshoot, making the response smooth.
In PI controller, after clearing the fault, the oscillations were observed in active and reactive powers for more than
0.6 seconds, whereas in ELANFIS controller, active and reactive power oscillations were damped with minimum time.
After clearing fault, the stator voltage settles in 0.6 seconds with distortion in PI controller, whereas stator voltage settles

FIGURE 14 Active and reactive powers


of (A) proportional‐integral and (B)
ELANFIS controller for single‐phase fault

FIGURE 15 Stator and direct current (DC) link voltages of (A) proportional‐integral and (B) ELANFIS controller for single‐phase fault
SHIHABUDHEEN ET AL. 15 of 18

in 0.42 seconds in ELANFIS controller. For PI controller, DC link voltage reaches a peak value of 1.65 kV during fault
condition and produces oscillations after clearing the fault. In case of ELANFIS, DC link voltage reaches a peak value
1.58 kV only during fault and oscillations damped out quickly after clearing the fault. This indicates the effectiveness
of the ELANFIS controller during single line to ground fault condition.

FIGURE 16 Stator and direct current (DC) link voltages of (A) proportional‐integral and (B) ELANFIS controller for wind speed variation

FIGURE 17 Active power, reactive power, and rotor current of ELANFIS controller for wind speed variation
16 of 18 SHIHABUDHEEN ET AL.

FIGURE 18 Rotor speed and torque of (A) proportional‐integral and (B) ELANFIS controller for wind speed variation

5.3 | Performances with change in wind speed


The optimal power algorithm decides the power output references for a given wind speed. The reference powers vary
according to changes in wind speed, and the controller has to track these changes. In this simulation, initial wind speed
is set as 12 m/s and then it is reduced to 10 m/s.
The performances of the PI controller and ELANFIS controller are shown in Figure 16. When the wind speed
changes to 10 m/s, the active power is reduced to 1.15 MW. For both controllers, a constant stator voltage is observed
irrespective of changes in wind speed. A small dip in DC link voltage is noticed during the transition period and recov-
ered without any oscillations. In this case, the performance of ELANFIS controller is similar to PI controller, which sig-
nifies that the degree of adaptability is very minimum for change in wind speed conditions. Unlike PI controllers,
retuning of parameters is not need in ELANFIS‐based controller.
Figure 17 shows the DFIG active power, reactive power, and rotor current of ELANFIS controlled system for wind speed
change. A change in active power from 1.95 to 1.15 MW is observed during wind speed change; likewise, reactive power also
reduces from −0.1 to −0.01 MVar. The change in rotor current magnitude from 1 to 0.73 pu with change in phase sequence
shows the integration of RSC for supersynchronous and subsynchronous rotor speeds. The rotor speed (in pu) and electrical
torque generated (in pu) are shown in Figure 18. The rotor speed is reduced to 1.02 pu when wind speed reduces from its
initial value 1.2 pu. Initially, torque is about 0.75 pu and it reduces to 0.53 pu when wind speed changes. In case of ELANFIS
controller, a steady performance of rotor speed and torque is observed after 8 seconds compared to PI controller.

6 | CONCLUSIONS

ELANFIS is one of the neuro‐fuzzy systems, which combines the learning capabilities of ELM and the explicit knowl-
edge of the fuzzy systems. In this paper, ELANFIS is proposed for adaptive control of grid‐connected DFIG‐based wind
energy system. ELANFIS controller is designed for power electronic converters of DFIG for an effective performance
under uncertainties in operating conditions. The performance of the proposed strategy is analyzed through real‐time
simulations with RTDS/RSCAD in HIL configuration. An improved performance of the proposed strategy is observed
when tested under various contingencies. Further, a robust performance and adaptive nature of the ELANFIS controller
make it an effective alternative to the conventional PI controller.
Based on the current simulation results, in case of wind speed variation, the improvement of ELANFIS controller
performance is found to be small in comparison with a regular PI controller. For this reason, a realistic stochastic wind
speed profile can be taken to further evaluate the performance of wind turbine using ELANFIS controller. Further, more
specific training data can be chosen for better performance. An integrated control strategy for MPPT wind energy
SHIHABUDHEEN ET AL. 17 of 18

extraction, grid voltage control, active and reactive power control, etc can be developed using ELANFIS technique to
achieve intelligent performance in DFIG system.

ORCID
K.V. Shihabudheen http://orcid.org/0000-0002-8122-5417

R EF E RE N C E S
1. Xie D, Xu Z, Yang L, Ostergaard J, Xue Y, Wong KP. A comprehensive LVRT control strategy for DFIG wind turbines with enhanced
reactive power support. IEEE Trans Power Syst. 2013;28(3):3302‐3310.
2. Jabr HM, Lu D, Kar NC. Design and implementation of neuro‐fuzzy vector control for wind‐driven doubly‐fed induction generator. IEEE
Trans Sustain Energy. 2011;2(4):404‐413.
3. Kamel RM, Chaouachi A, Nagasaka K. Wind power smoothing using fuzzy logic pitch controller and energy capacitor system for
improvement microgrid performance in islanding mode. Energy. 2010;35(5):2119‐2129.
4. Hansen AD, Michalke G. Fault ride‐through capability of DFIG wind turbines. Renew Energy. Jul. 2007;32(9):1594‐1610.
5. Yunus AMS, Abu‐Siada A, Masoum MAS. Improvement of LVRT capability of variable speed wind turbine generators using SMES unit. In
Proc. IEEE PES Innov. Smart Grid Technol. Asia (ISGT); 2011: 1–7.
6. Raju SK, Pillai GN. Type 2 fuzzy logic–based robust control strategy for power sharing in microgrids with uncertainties in operating
conditions. Int Trans Electr Energ Syst. 2017;27:e2294. https://doi.org/10.1002/etep.2294
7. Muller S, Deicke M, De Doncker RW. Doubly fed induction generator systems for wind turbines. IEEE Ind Appl Mag. 2002;8(3):26‐33.
8. Xie D, Xu Z, Yang L, Ostergaard J, Xue Y, Wong KP. A comprehensive LVRT control strategy for DFIG wind turbines with enhanced
reactive power support. IEEE Trans Power Syst. Aug. 2013;28(3):3302‐3310.
9. Chen M, Yu L, Wade NS, Liu X, Liu Q, Yang F. Investigation on the faulty state of DFIG in a microgrid. IEEE Trans Power Electron.
2011;26(7):1913‐1919.
10. Yao J, Li H, Chen Z, et al. Enhanced control of a DFIG based wind‐power generation system with series grid‐side converter under
unbalanced grid voltage conditions. IEEE Trans Power Electron. 2013;28(7).
11. Xu L, Wang Y. Dynamic modeling and control of DFIG‐based wind turbines under unbalanced network conditions. IEEE Trans Power
Syst. Feb. 2007;22(1):314‐323.
12. Li S, Haskew TA, Williams KA, Swatloski RP. Control of DFIG wind turbine with direct‐current vector control configuration. IEEE Trans
Sustainable Energy. Jan. 2012;3(1):1‐11.
13. Mohammadi J, Vaez‐Zadeh S, Afsharnia S, Daryabeigi E. A combined vector and direct power control for DFIG‐based wind turbines. IEEE
Trans Sustainable Energy. July 2014;5(3):767‐775.
14. Jabr HM, Lu D, Kar NC. Design and implementation of neuro‐fuzzy vector control for wind‐driven doubly‐fed induction generator. IEEE
Trans Sustainable Energy. Oct. 2011;2(4):404‐413.
15. Jafari A, Shahgholian G. Analysis and simulation of a sliding mode controller for mechanical part of a doubly‐fed induction generator‐
based wind turbine. IET Gener Transm Distrib. 2017;11(10):2677‐2688.
16. Liu X, Han Y, Wang C. Second‐order sliding mode control for power optimisation of DFIG‐based variable speed wind turbine. IET Renew
Power Gener. 2017;11(2):408‐418.
17. Mechter A, Kemih K, Ghanes M. Backstepping control of a wind turbine for low wind speeds. Nonlinear Dynam. 2016;84(4):2435‐2445.
18. Yang B, Lin J, Wang L, Yao W, Wu QH. Nonlinear maximum power point tracking control and modal analysis of DFIG based wind tur-
bine. Int J Electr Power Energy Syst. 2016;74:429‐436.
19. Guo W, Liu F, Si J, He D, Harley R, Mei S. Approximate dynamic programming based supplementary reactive power control for DFIG
wind farm to enhance power system stability. Neurocomputing. 2015;170:417‐427.
20. Meng W, Yang Q, Sun Y. Guaranteed performance control of DFIG variable‐speed wind turbines. IEEE Trans Control Syst Technol. Nov.
2016;24(6):2215‐2223.
21. Soares O, Gonçalves H, Martins A, Carvalho A. Nonlinear control of the doubly‐fed induction generator in wind power systems. Renew
Energy. August 2010;35(8):1662‐1670.
22. Medjber A, Guessoum A, Belmili H, Mellit A. New neural network and fuzzy logic controllers to monitor maximum power for wind energy
conversion system. Energy. 2016;106(1):137‐146.
23. de Marchi RA, Dainez PS, Von Zuben FJ, Bim E. A multilayer perceptron controller applied to the direct power control of a doubly fed
induction generator. IEEE Trans Sustainable Energy. April 2014;5(2):498‐506.
18 of 18 SHIHABUDHEEN ET AL.

24. Setiawan I, Priyadi A, Miyauchi H, Purnomo MH. Adaptive B‐spline neural network‐based vector control for a grid side converter in wind
turbine‐DFIG systems. IEEJ Trans Elec Electron Eng. 2015;10:674‐682.
25. Boufounas EM, Boumhidi J, Ouriagli M, Boumhidi I. A robust power control of the DFIG wind turbine based on general regression neural
network and APSO algorithm. Int J Power Energy Syst. 2015;35:64‐73.
26. DeAlmeida R, Lopes J, Barreiros J. Improving power system dynamic behavior through doubly fed induction machines controlled by static
converter using fuzzy control. IEEE Trans Power Syst. 2004;19(4):1942‐1950.
27. Rashid G, Ali MH. Fault ride through capability improvement of DFIG based wind farm by fuzzy logic controlled parallel resonance fault
current limiter. Electr Pow Syst Res. 2017;146:1‐8.
28. Jazaeri M, Samadi AA. Self‐tuning fuzzy PI‐based controller of DFIG wind turbine for transient conditions enhancement. Int Trans Electr
Energ Syst. 2015;25:2657‐2673.
29. Jafari SH, Raoofat M, Samet H. Improving transient stability of double fed induction generator using fuzzy controller. Int Trans Electr
Energ Syst. 2014;24:1065‐1075.
30. Rezaei F, Esmaeili S. Decentralized reactive power control of distributed PV and wind power generation units using an optimized fuzzy‐
based method. Int J Electr Power Energy Syst. May 2017;87:27‐42.
31. Zhang S, Mishra Y, Shahidehpour M. Fuzzy‐logic based frequency controller for wind farms augmented with energy storage systems. IEEE
Trans Power Syst. March 2016;31(2):1595‐1603.
32. Krishnama Raju S, Pillai GN. Design and implementation of type‐2 fuzzy logic controller for DFIG‐based wind energy systems in
distribution networks. IEEE Trans Sustainable Energy. Jan. 2016;7(1):345‐353.
33. Krishnama Raju S, Pillai GN. Design and real time implementation of type‐2 fuzzy vector control for DFIG based wind generators. Renew
Energy. April 2016;88:40‐50.
34. Mitra S, Hayashi Y. Neuro‐fuzzy rule generation: survey in soft computing framework. IEEE Trans Neural Netw. 2000;11(3):748‐767.
35. Jang J‐SR. ANFIS: adaptive‐network‐based fuzzy inference systems. IEEE Trans Svst, Man, Cybern. 1993;23(3):665‐685.
36. Jang J‐SR, Sun C‐T. Neuro‐fuzzy modeling and control. Proc IEEE. 1995;83(3):378‐406.
37. Casella IRS, Capovilla CE, Sguarezi Filho AJ, Jacomini RV, Azcue‐Puma JL. An ANFIS power control for wind energy generation in smart
grid scenario using wireless coded OFDM‐16‐QAM. J Control Autom Electr Syst. 2014;25:22‐31.
38. Huang G‐B, Zhu QY, Siew CK. Extreme learning machine: theory and applications. Neurocomputing. 2006;70(1):489‐501.
39. Huang G‐B, Chen L, Siew C‐K. Universal approximaion using incremental constructuve feedforward networks with random hidden nodes.
IEEE Trans Neural Netw. 2006;17(4):879‐892.
40. Nizar AH, Dong ZY, Wang Y. Power utility nontechnical loss analysis with extreme learning machine method. IEEE Trans Power Syst.
Aug. 2008;23(3):946‐955.
41. Mahmoud TK, Dong ZY, Ma J. A developed integrated scheme based approach for wind turbine intelligent control. IEEE Trans
Sustainable Energy. July 2017;8(3):927‐937.
42. Patnaik RK, Patnaik UC, Dharmapandit O, Naeem MH. A New Islanding Detection and Classification Technique for the Grid Connected
DFIG Wind Farms. Bhubaneswar: 2015 IEEE International Conference on Power, Communication and Information Technology (PCITC);
2015:782‐788.
43. Pillai GN, Pushpak J, Nisha MG. Extreme learning ANFIS for control applications. IEEE Symposium on Computational Intelligence in
Control and Automation (CICA); 2014: 1‐8.
44. Shihabudheen KV, Pillai GN. Regularized extreme learning adaptive neuro‐fuzzy algorithm for regression and classification. Knowl‐Based
Syst. 2017;127(1):100‐113.
45. Shihabudheen KV, Pillai GN, Peethambaran B. Prediction of landslide displacement with controlling factors using extreme learning adap-
tive neuro‐fuzzy inference system (ELANFIS). Appl Soft Comput. December 2017;61:892‐904.
46. Shihabudheen KV, Mahesh M, Pillai GN. Particle swarm optimization based extreme learning neuro‐fuzzy system for regression and
classification. Expert Syst Appl. 2018;92:474‐484.
47. Shihabudhen KV, Peethambaran B. Landslide displacement prediction technique using improved neuro‐fuzzy system. Arab J Geosci.
November 2017;10:502.
48. Jabr HM, Kar NC. Fuzzy gain tuner for vector control of doubly‐fed wind driven induction generator. In Proc. IEEE Canadian Conf. Elect.
and Computer Engineering; 2006: 2266–2269.

How to cite this article: Shihabudheen KV, Raju SK, Pillai GN. Control for grid‐connected DFIG‐based wind
energy system using adaptive neuro‐fuzzy technique. Int Trans Electr Energ Syst. 2018;e2526. https://doi.org/
10.1002/etep.2526

You might also like