Implementation of a Neuro-Fuzzy Direct Torque

and Reactive Power Control for DFIM

R. V. Jacomini† , C. M. Rocha∗ , J. A. T. Altuna∗ J. L. Azcue∗ , C. E. Capovilla∗ , A. J. Sguarezi Filho∗
∗ Universidade Federal do ABC - UFABC, Santo André, SP - Brazil
† IFSP Hortolândia, SP - Brazil

Abstract—This paper proposes a Takagi-Sugeno neurofuzzy I. INTRODUCTION

inference system for direct torque and stator reactive power
control applied to a doubly fed induction motor (DFIM). The con- Doubly Fed Induction Machine (DFIM) is an attractive
trol variables (d-axis and q-axis rotor voltages) are determined alternative to cage rotor induction and synchronous machines
through a control system composed by a neuro-fuzzy inference
in high power applications. The interest occurs mainly in
system and a first order Takagi- Sugeno fuzzy logic controller.
Experimental results are presented to validate the controller applications such as: generator for renewable energy [1],
operation for variable speed under no-load and load conditions [2] and motor for ventilation systems and pumps. For such
and stator reactive power variation under load condition. For applications, the stator windings are directly connected to
this last validation, a PI controller is used to control the rotor the three phase grid with constant frequency and the rotor
speed, thereby its output is used to manipulate the torque in
windings are supplied by a bidirectional power converter.
order to follow the demanded speed value.
Index Terms—ANFIS, direct torque control, Doubly fed induc- This one is designed to handle about ±25% of the machines
tion motor, neuro-fuzzy control, reactive power control. rating [1], becoming attractive the utilization of DFIM in large
power systems.
Among all methods of electromagnetic torque and reactive
N OMENCLATURE
power control for DFIM, the direct control techniques have be-
rs , rr stator and rotor resistances; come the preferred strategies for high dynamic performances.
L s , Lr , Lm stator, rotor, and mutual The techniques consist of an algorithm that generates directly
inductances; the required rotor control voltage and their errors.
σ total leakage factor; Based on the technique of Direct Torque Control (DTC)
ω 1 , ω2 , ωr synchronous, slip, and rotor applied to induction machines [3], the Direct Power Control
angular frequency; (DPC) using hysteresis controllers associated with a switching
θ1 , θ2 , θr stator flux, slip, and rotor table was the first control strategy applied to the DFIM [4],
angles; [5]. These controllers have the disadvantage of operating with
P, Q active and reactive power; variable switching frequency. This fact is overcome in [6],
Tem electromagnetic torque; where the constant switching frequency is obtained by cal-
V , I, ψ voltage, current, and flux culating the required rotor control voltage for each switching
space vectors; period from the estimated stator flux, stator active and reactive
v, i, ψ voltage, current, and flux power, and their errors.
components; In this context, this paper proposes a neuro-fuzzy controller
p number of poles. based on direct torque and reactive power control for DFIM.
Differing from [7], in this paper the control strategy is applied
Superscripts to DFIM as a motor instead of generator.The controller pro-
∗ reference value; posed consisting of a Neuro-Fuzzy Inference System (NFIS)
 estimated value. combined to Takagi-Sugeno Fuzzy Logic Controller (T-S
FLC). From electromagnetic torque, stator reactive power, and
Subscripts rotor speed, the NFIS generates the direct and quadrature
s, r stator and rotor; axes reference values of rotor voltage components in the
α, β direct and quadrature axis expre- stator-flux-oriented reference frame. The T-S FLC determines
ssed at stationary reference frame; the increments in the voltage rotor components necessary to
d, q direct and quadrature axis expre- minimize the electromagnetic torque and stator reactive power
ssed at synchronous reference frame; errors.
m, n direct and quadrature axis expre- This paper is organized as follows: Besides the introductory
ssed at rotor reference frame; section, Vector control of DFIM is shown in section II;
1, 2 synchronous and slip. The proposed control scheme is presented in section III;

k,((( 
Main simulated and experimental results are considered in
3 Lm
section IV in order to validate the proposed control scheme; Ps = − ω1 ψsd irq (12)
and Finally, section V concludes the work. 2 Ls
From equations (10) and (9), the stator reactive power can
II. BASIC CONCEPTS OF DFIM VECTOR CONTROL
be given by:
A. Stator-flux-oriented control
3 Lm ψ 
sd
The modeling of DFIM is obtained at Stator-Flux-Oriented Qs = ω1 ψsd − ird (13)
(SFO) reference frame. This approach implies the alignment 2 Ls Lm
of the direct axis with the stator flux vector, i.e., ψ s,dq = ψsd , As a result, the model according to the orientation of
and therefore, ψsq =0. In this way, the stator and rotor voltages, the stator flux allows independent control of electromagnetic
stator and rotor flux linkages equations can be written as: torque/stator active and reactive power through the rotor
current components. In summary, Tem or Ps is controlled by
dψsd
V s,dq = rs I s,dq + + jω1 ψsd (1) irq and Qs by ird . Therefore, the rotor current components
dt can be controlled to regulate the electromagnetic torque/stator
dψ r,dq active and reactive power.
V r,dq = rr I r,dq + + jω2 ψ r,dq (2)
dt B. Direct torque and reactive power control
In contrast to Stator-Flux-Oriented Control (SFOC), which
ψ s,dq = ψsd = Ls I s,dq + Lm I r,dq (3) the rotor current components are used as control variables, in
DTC, the rotor flux components are used as control variables.
ψ r,dq = Lr I r,dq + Lm I s,dq (4) From equations (5), (10), (7), and (9), the electromagnetic
torque and stator reactive power can be expressed by:
According to the equations (3) and (4), the stator and rotor
currents can be calculated as: p
Tem = − kσ ψsd ψrq (14)
2
ψsd kr
I s,dq = − ψ (5)  
σLs σLs r,dq Qs = kσ ω1 ψsd ψrd −
Lr
ψsd (15)
Lm
ψ r,dq ks
I r,dq = − ψsd (6) where: kσ = 1.5 Lm /σLs Lr .
σLr σLr It is observed that the electromagnetic torque and stator
where ks =Lm /Ls and kr =Lm /Lr are the coupling factors of reactive power are controlled by the rotor flux components ψrq
the stator and rotor, respectively, and σ = (Ls Lr −L2m )/Ls Lr . and ψrd , respectively, due to ψsd to be constant. For a clear
The electromagnetic torque, stator active and reactive power understanding, it is used a vector diagram of DFIM, where the
can be expressed by: δψ is the angle between rotor and flux vectors.
3p  ∗

Tem = m I s,dq · ψ s,dq (7) III. PROPOSED DIRECT TORQUE AND STATOR
22 REACTIVE POWER CONTROL BASED ON
NEURO-FUZZY STRATEGY
3 ∗
Ps = e(V s,dq · I s,dq ) (8) From the electromagnetic torque and stator reactive power
2
references, the control strategy based on neural-fuzzy infer-
3 ∗ ence system calculates the required rotor voltage components
Qs = m(V s,dq · I s,dq ) (9)
2 ∗
vrd ∗
and vrq ∗
. The input variables of control strategy are Tem ,
∗
Under balanced grid voltage condition, taking into consid- Qs and rotor angular speed ωr . To minimize the error of
eration equation (1), and if the voltage drop across the stator steady-state and to have a fast dynamic response at torque and
resistance is neglected, the equation for the stator voltage in stator reactive power is added a Takagi-Sugeno fuzzy logic
steady-state conditions is given by: controller, which determines the increment necessary in the
components of direct axis and quadrature rotor voltage Δvrq
V s,dq = jω1 ψsd (10) and Δvrd .
It represents that the stator flux ψsd is a constant determined A. Adaptive neural-fuzzy inference system and training
by the supply voltage magnitude and synchronous frequency The rules base and membership functions are obtained in
ω1 . Therefore, from equations (3), (7), (8), and (10), the elec- an automated process by an Adaptive Neural-Fuzzy Inference
tromagnetic torque and stator active power can be expressed, System (ANFIS) by using input-output data set. The parame-
in terms of rotor current component irq , as: ters associated with the output and input membership functions
3 p Lm are adjusted by the combination of backpropagation learning
Tem = − ψsd irq (11) algorithm with a least squares type algorithm [8].
2 2 Ls


 According to the property of ANFIS, it is only suitable to for j=1,2,...,8.
the multiple input and single output system. Therefore, as the – Layer 3: The firing strength of each rule are normalized via
proposed system requires three inputs and two outputs, it is N neuron, resulting in:
∗
necessary to use two ANFIS: one to provide vrd (ANFIS 1) wj
∗ O3,j = wj = , (20)
and another to provide vrq (ANFIS 2).
w1 + w2 + w3 + ... + w8
∗
for j=1,2,...,8.
Tem , Q∗
s , ωr
– Layer 4: All nodes in this layer are adaptive. The outputs
ANFIS 1  vrd,1 of the nodes are calculated by the product of the normalized
N
 vrd,2 firing strengths and value of membership function for a first-
A1 N
∗
Tem order Sugeno fuzzy model.
A2

N
vrd,3
∗
 vrd,4 vrd ∗
B1 N Σ O4,j = wj (pj · Tem + qj · Q∗s + rj · ωr + sj ) (21)
Q∗s  vrd,5
B2 N for j=1,2,...,8. Where: O4,j is vrd,j or vrq,j , and pj , qj , rj , rj
 vrd,6
C1
N are the parameters of the consequent parameter set.
ωr  vrd,7 – Layer 5: In this last layer, the output O5 of the neuro-fuzzy
C2 N
 vrd,8 system is given by:
N

ANFIS 2 8

 vrq,1
N O5 = w(j) O4,j (22)
A1

N
vrq,2 j=1
∗
Tem  vrd,1 ∗ ∗
A2 N where: O5 is vrd or vrq .
∗
 vrq,4 vrq
B1 N Σ B. Training data
Q∗s  vrq,5
B2 N The required input-output training data are obtained from a

N
vrq,6 DFIM employing PI, using for this a SFOC simulation. From
C1
ωr 
N
vrq,7 this simulation is sampled 135 training data for 27 Operation
C2

N
vrq,8 Points (OP), as shown in Table I.
∗
As presented in Fig. 3, the membership functions generated
Tem , Q∗
s , ωr
by ANFIS 1 and ANFIS 2, after training, are equal. Therefore,
it simplifies the system shown in Fig. 1 for the compact one
shown in Fig. 2. The parameter set of the ith node (pi , qi , ri ,
layer 1 layer 2 layer 3 layer 4 layer 5 and si ) in layer 4 generated by ANFIS 1 and ANFIS 2 are
Fig. 1. Implemented NFIS architecture. given by:

Fig. 1 shows this structure, when can be observed that it

vrd,1 = [0.94568 − 0.00472 0.02345 1.81163]
contains eight fuzzy Takagi-Sugeno if-then rules [9] and two
membership functions for each input variable. The functions vrd,2 = [−0.55573 − 0.00472 0.02356 − 1.08688]
of the nodes for each layer are described as follows: vrd,3 = [0.94556 0.00471 0.02353 1.81163]
– Layer 1: All nodes in this layer are adaptive, with the vrd,4 = [−0.55521 0.00473 0.02348 − 1.08688]
function defined by:
vrd,5 = [1.01839 − 0.00513 0.00849 1.96272]
vrd,6 = [−0.62355 − 0.00513 0.00846 − 1.17759]
∗
O1,i = μAi (Tem ), i = 1, 2 (16) vrd,7 = [1.01851 0.00513 0.00847 1.96272]
O1,(i+2) = μ Bi
(Q∗s ), i = 1, 2 (17) vrd,8 = [−0.62407 0.00512 0.00848 − 1.17759]
O1,(i+4) = μCi (ωr ), i = 1, 2 (18)
∗
where: Tem , Q∗s and ωr are the inputs in the nodes O1 and vrq,1 = [−0.40813 − 0.00428 − 0.00097 56.39471]
Ai . The Bi and Ci are linguistic variables associated with vrq,2 = [0.30424 0.00427 − 0.00108 − 33.83683]
these nodes. At the implementation the linguistic variables are
vrq,3 = [−0.40804 0.00429 − 0.00105 56.39471]
denoted triangular function.
– Layer 2: The outputs of each node in this layer are the firing vrq,4 = [0.30335 − 0.00428 − 0.00101 − 33.83683]
strength wj , that are determined by T-norm operators, defined vrq,5 = [−0.49505 − 0.00465 − 0.05879 56.32524]
by algebraic product (Π Neuron). vrq,6 = [0.23448 0.00466 − 0.05877 − 33.79539]
vrq,7 = [−0.49514 0.00465 − 0.05877 56.32524]
 ∗

O2,j = wj = μA (Tem ) × μB (Q∗s ) × μC (ωr ), (19) vrq,8 = [0.23537 − 0.00466 − 0.05878 − 33.79539]


 TABLE I 1
A1 A2

μ(Tem)
T RAINING DATA 0.5

OP ∗
Tem Q∗s ωr ∗
vrq ∗
vrd 0
−12 −10 −8 −6 −4 −2 0 2 4 6 8 10 12
1 -13 -2200 471.25 27.611 -47.695 Tem(Nm)
2 0 -2200 471.25 17.232 -57.736 1
B1 B2

μ(Qs)
3 13 -2200 471.25 66.59 4.325 0.5
4 -13 0 471.25 32.048 15.225
5 0 0 471.25 44.308 5.242 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
6 13 0 471.25 56.253 -4.598 Qs(kVAR)
7 -13 2200 471.25 21.806 4.230 1
C1 C2
8 0 2200 471.25 34.578 -5.17

μ(ωr )
0.5
9 13 2200 471.25 46.978 -14.498
10 -13 -2200 377 -8.138 15.569 0
282.75 329.875 377 424.125 471.25
11 0 -2200 377 1.843 15.087 ωr (rad/s)
12 13 -2200 377 11.550 14.645
13 -13 0 377 -8.941 5.149 Fig. 3. Membership function of three inputs.
14 0 0 377 1.095 5.206
15 13 0 377 10.850 5.238
16 -13 2200 377 -8.635 -5.281
17 0 2200 377 1.371 -4.684
(1)
18
19
13
-13
2200
-2200
377
282.75
11.101
-59.797
-4.176
6.057
N R1 μN Δvrd
20 0 -2200 282.75 -51.493 15.57 ΔQs (k) Δvrd (k)
21 13 -2200 282.75 -43.488 24.964 Gqs ZE R2 Σ /Σμ Gvrd
22 -13 0 282.75 -49.931 -4.925
23 0 0 282.75 -42.118 5.17 (3)
24 13 0 282.75 -34.552 15.076 P R3 μP Δvrd
25 -13 2200 282.75 -39.078 -14.794
26 0 2200 282.75 -31.834 -4.199 (1)
27 13 2200 282.75 -24.774 6.145 N R1 μN Δvrd
ΔTem (k) Δvrq (k)
Gte ZE R2 Σ /Σμ Gvrq
C. The proposed T-S FLC
(3)
P R3 μP Δvrq
The proposed T-S FLC determines the increments Δvrq and
Δvrd from the error of electromagnetic torque ΔTem and Fig. 4. Proposed T-S FLC architecture.
stator reactive power ΔQs , respectively. In Fig. 4 is shown
the implemented architecture of the T-S FLC.
The architecture uses two gains for the inputs with fixed
values (Gte and Gqs ) and for the outputs are introduced
the gains Gvrd and Gvrq to match the values generated by
weighting with the actual values of controller output variables.
∗
Tem , Q∗
s , ωr
1) Membership functions: Fig. 5(a) uses a triangular mem-
vrd,1 bership functions for both inputs, where linguistic values are
vrd,2 negative (N ), positive (P ), and zero (ZE).
vrd,3
∗
vrd,4 vrd
Σ
 vrd,5
N
μ (Δvrd), μ (Δvrq )

N ZE P

N vrd,6 1 1
Δvrd, Δvrq

∗ A1
Tem 
A2 N vrd,7 0.5 0

 vrd,8 0 −1

B1 N −1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1

ΔQs, ΔTem
Q∗s 
ΔQs, ΔTem
B2 N vrq,1
 (a) Membership function for inputs. (b) Behavior of Δvrd and Δvrq in
C1 N vrq,2 terms of ΔQs and ΔTem .
ωr 
C2 N vrd,1 Fig. 5. Membership function for inputs ΔQs and ΔTem , and behavior of
 ∗
N vrq,4 vrq respective outputs Δvrd , Δvrq in function of ΔQs and ΔTem .
Σ
vrq,5
vrq,6
vrq,7
2) The fuzzy rule base: The T-S FLC has three rules (R1,
vrq,8
R2, and R3). The set of Δvrd and Δvrq increments appears
Fig. 2. Proposed NFIS architecture. as follows:


 Texas Instruments (TI) board (DSP TMS320F2812). In Dig-
ital Signal Processor (DSP) programming, two interruptions
R1 : IF (ΔQs or ΔTem ) IS (N ) THEN are used: 1) Interruption of 50μs for voltages and currents
(1) (1) (1) (1) (1) (1)
(Δvrd = a0 + a1 ΔQs or Δvrq = a0 + a1 ΔTem ) measurements, abc to αβ conversion of these variables, and
angular position estimation; 2) Interruption of 200μs for calcu-
lation of rotor parameters (speed and position) through reading
R2 : IF (ΔQs or ΔTem ) IS (ZE) THEN
(2) (2) (2) (2) (2)
pulses generated by encoder, execution of the algorithms T-
(2)
(Δvrd = a0 + a1 ΔQs or Δvrq = a0 + a1 ΔTem ) S FLC and NFIS that generates the values of vrd ∗
and vrq∗
,
transformation of V r,dq to V r,mn , and finally the execution of
R3 : IF (ΔQs or ΔTem ) IS (P ) THEN SVM algorithm.
(3) (3) (3) (3) (3) (3)
(Δvrd = a0 + a1 ΔQs or Δvrq = a0 + a1 ΔTem ) (23)
TABLE II
(i) (i) DFIM NOMINAL VALUES AND PARAMETERS
For each consequent rule, Δvrd and Δvrq are specified by
a first-order T-S model: Ps , Vs , f s 2.25 kW, 220 V, 60 Hz
rs , rr 1.2Ω, 1.24Ω
Ls , Lr , Lm 98.14 mH, 98.14 mH, 91.96 mH
Δvrd
(1)
= 11.22 + 12.25ΔQs p 4
(2)
Δvrd = 10.28 + 0.0048ΔQs
(3) A. Performance test for the control strategy
Δvrd = 11.22 − 12.25ΔQs
(1)
In this first test, simulated and experimental results are
Δvrq = 11.22 + 12.25ΔTem verified to evaluated the steady-state performance and dynamic
(2)
Δvrq = 10.28 + 0.0048ΔTem response of the control strategy. Initially, simulations are
(3)
Δvrq = 11.22 − 12.25ΔTem (24) performed as tuning tests for the gain values Gqs , Gps , Gvrd ,
and Gvrq of the T-S-FLC. The results are presented side by
Fig. 5(b) illustrates the behavior of the output variables side for a better comparison. Fig. 7 shows results for torque
Δvrd and Δvrq , which depends on the input variables ΔQs reference alternately varying from 10 Nm to -10 Nm, inside
and ΔTe , respectively. a period of 100m seconds, while the stator reactive power
3) Aggregation process: The output of T-S FLC acts to reference is maintained at zero level. For this test the DFIM
reduce the error. If the error of the electromagnetic torque and is driven by a DC motor. Fig. 7(a), (b), (g), and (h) show the
stator reactive power is positive, then, their outputs will have torque and stator reactive power responses. It can be noted that
a decrement. If the error is zero, the outputs will be zero. both, experimental and simulated responses perfectly follow
The overall output of the T-S FLC implemented is the their respective references.
weighted average of the three individual outputs, given by: The three-phase instantaneous stator and rotor currents are
shown in Fig. 7(c), (d), (i), and (j). Fig. 7(e) and (l) show
(1) (2) (3) the phase change between stator voltage and current, due to
μN Δvrd + μZE Δvrd + μP Δvrd
Δvrd (ΔQs ) = alteration of operating mode, from motor (positive torque)
μN + μZE + μP to generator (negative torque). Fig. 7(f) and (m) show the
(1) (2) (3)
μN Δvrq + μZE Δvrq + μP Δvrq speed change, caused by torque manipulation. Depending on
Δvrq (ΔTem ) = (25)
μN + μZE + μP the torque balance between the DC motor and the DFIM, the
machine sets accelerates or decelerates.
In T-S FLC controller, it is not necessary the defuzzification
interface [10]. This occurs due to in Takagi-Sugeno fuzzy B. Load test
controllers each rule is already crisp and the total result is In this third test, the load used is a DC machine operating
determined by the weighted sum of each rule, as is shown in as generator. The response for a 9 Nm load torque is shown
equations (25). in Fig. 8, with the motor operating at 192 rad/s and the
IV. SIMULATED AND EXPERIMENTAL RESULTS stator reactive power reference equal to 0 VAR. The speed
response, Fig. 8(a), shows a null steady-state error. Fig. 8(b)
This section presents simulated and experimental results in and (c) shows stator reactive power and torque, respectively.
order to validate the proposed DTC control strategy. With It can be verified that the torque Tem increases to account
this objective, it is chosen and performed four tests. The for load torque. In Fig. 8(d) and (e) is shown the three-phase
experimental setup comprises a 2.25 kW DFIM mechanically instantaneous currents of stator and rotor, respectively. The
coupled to a DC machine as load. The DFIM parameters are load insertion causes a current increase as expected.
given in Table II.
A power electronic converter comprised of Insulated-Gate V. CONCLUSION
Bipolar Transistor (IGBT) switches is used to supply the In this paper was proposed a control scheme that combines
machine rotor. The DTC strategy is implemented with a NFIS with Takagi-Sugeno fuzzy logic controller. This strategy


 C
Q∗s +
T-S FLC ∗
+ vrd
-
∗
+ V r,mn
ωr NFIS ejθ2 V LB SV M
∗
+ vrq
ωr∗ + ∗
Tem +
PI T-S FLC +
- -
θ2 ωr
- θr 
ωr +
θ1
DFIM
θ1
estimation

Tem V s,αβ V s,abc

Tem and Qs αβ

Qs calculation I s,αβ I s,abc

abc

GRID

Fig. 6. Neuro-fuzzy controller for direct torque control.

has shown that is possible decoupling torque and reactive

power control of a DFIM by direct manipulation of rotor
voltages, without the need of current controllers. The simu- ωr ωr∗

lated and experimental results confirm the effectiveness of the 2200

ωr (rpm)

control during several operating conditions. The performance 1834

1400
is good, inclusively attending the requirements of fast dynamic 0 1.2 2.4 3.6 4.8 6

response and minimal oscillation around the steady-state. Time (s)

(a) Rotor speed
Thus, this control strategy is an interesting tool for DFIM
Qs Q∗s
operational applications as motor.
Qs (kVAR)

0.5

−0.5
R EFERENCES
0 1.2 2.4 3.6 4.8 6
Time (s)
(b) Stator reactive power
[1] F. Blaabjerg and F. Iov. Wind power - a power soucer now enabled by
∗
Tem
power electronics. 9th Brasilian Power Electronics Conference, month Tem

2007. 10
Tem (Nm)

[2] R. Datta and V. T. Ranganathan. Variable-speed wind power gener- 5

ation using doubly fed wound rotor induction machine-a comparison 0

with alternative schemes. Energy Conversion, IEEE Transactions on, −5

0 1.2 2.4 3.6 4.8 6
Time (s)
17(3):414–421, Sep 2002. (c) Electromagnetic torque
[3] I. Takahashi and T. Noguchi. A new quick-response and high-efficiency
control strategy of an induction motor. Industry Applications, IEEE 10

Transactions on, IA-22(5):820–827, Sept 1986.

Is (A)

0
[4] R. Datta and V. T. Ranganathan. Direct power control of grid-connected
wound rotor induction machine without rotor position sensors. Power −10

Electronics, IEEE Transactions on, 16(3):390–399, May 2001. 0 1.2 2.4

Time (s)
3.6 4.8 6

[5] Lie Xu and P. Cartwright. Direct active and reactive power control of (d) Three-phase stator current
dfig for wind energy generation. Energy Conversion, IEEE Transactions
on, 21(3):750–758, Sept 2006. 10

[6] Dawei Zhi and Lie Xu. Direct power control of dfig with constant
Ir (A)

0
switching frequency and improved transient performance. Energy
Conversion, IEEE Transactions on, 22(1):110–118, March 2007. −10

0 1.2 2.4 3.6 4.8 6

[7] R. Vani Jacomini and E. Bim. Implementation of the direct power Time (s)

control of a doubly fed induction generator by using a takagi-sugeno (e) Three-phase rotor current

neuro-fuzzy inference system. In Industrial Electronics Society, IECON

2013 - 39th Annual Conference of the IEEE, pages 852–857, Nov 2013. Fig. 8. Experimental result to load test of 9 Nm.
[8] J. S. R. Jang. Adaptative-network-based fuzzy inference system.
IEEE Transactions on Systems, Man, and Cybernetics, 23(3):665–684,
May/June 1993.


 Te Te∗ Te Te∗
10 10

Te (Nm)

Te (Nm)
0 0

−10 −10

1 1.12 1.24 1.36 1.48 1.6 0 0.12 0.24 0.36 0.48 0.6
Time (s) Time (s)
(a) Electromagnetic torque (g) Electromagnetic torque

Qs Q∗s Qs Q∗s
500 500
Qs (VAR)

Qs (VAR)
0 0

−500 −500

1 1.12 1.24 1.36 1.48 1.6 0 0.12 0.24 0.36 0.48 0.6
Time (s) Time (s)
(b) Stator reactive power (h) Stator reactive power

10 10
Is,abc (A)

Is,abc (A)
0 0

−10 −10

1 1.12 1.24 1.36 1.48 1.6 0 0.12 0.24 0.36 0.48 0.6
Time (s) Time (s)
(c) Three-phase stator current (i) Three-phase stator current

10 10
Ir,abc (A)

Ir,abc (A)

0 0

−10 −10

1 1.12 1.24 1.36 1.48 1.6 0 0.12 0.24 0.36 0.48 0.6
Time (s) Time (s)
(d) Three-phase rotor current (j) Three-phase rotor current

200 vs is 200 vs is
10 10
100 100
vs (V)

vs (V)

is (A)
is (A)

0 0 0 0
−100 −100
−10 −10
−200 −200
1 1.12 1.24 1.36 1.48 1.6 0 0.12 0.24 0.36 0.48 0.6
Time (s) Time (s)
(e) Stator phase voltage and current (l) Stator phase voltage and current

1900 1900
ωr (rpm)

ωr (rpm)

1800 1800

1700 1700

1 1.1 1.2 1.3 1.4 1.5 1.6 0 0.12 0.24 0.36 0.48 0.6
Time (s) Time (s)
(f) Rotor speed (m) Rotor speed

Fig. 7. Simulated (a, b, c, d, e, and f) and experimental (g, h, i, j, l, and m) results under various torque steps with stator reactive power reference maintained
at zero level.

[9] T. Takagi and M. Sugeno. Fuzzy identification of systems and its

applications to modeling and control. Systems, Man and Cybernetics,
IEEE Transactions on, SMC-15(1):116–132, Jan 1985.
[10] H. Hellendoorn D. Driankov and M. Reinfrank. An introduction to fuzzy
control. S 1996.