IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 58, NO.

2, FEBRUARY 2011 421

A Shunt Active Power Filter With Enhanced

Performance Using ANN-Based Predictive
and Adaptive Controllers
Avik Bhattacharya, Student Member, IEEE, and Chandan Chakraborty, Senior Member, IEEE

Abstract—This paper attempts to improve the dynamic per- pensating current [8]. A critical evaluation of such techniques
formance of a shunt-type active power filter. The predictive and is recently reported by the authors [8].
adaptive properties of artificial neural networks (ANNs) are used Recently, ANNs have attracted much attention in different
for fast estimation of the compensating current. The dynamics of
the dc-link voltage is utilized in a predictive controller to generate applications, including the APF [9]–[20], [22]. Dash et al.
the first estimate followed by convergence of the algorithm by an [14] computed the Fourier coefficients of the signal by using
adaptive ANN (adaline) based network. Weights in adaline are adaline, and Chen and O’Connell [15] used an ANN that is
tuned to minimize the total harmonic distortion of the source trained with GA and back propagation. Lai et al. [16] used a
current. Extensive simulations and experimentations confirm the Hopfield neural network for real-time computation of frequency
validity of the proposed scheme for all kinds of load (balanced and
unbalanced) for a three-phase three-wire system. and harmonic content of the signal. Improved performance
has been observed compared to discrete Fourier transform,
Index Terms—Adaline, current control, nonlinear load, shunt fast Fourier transform, or Kalman-filtering-based approaches.
active power filter (APF), total harmonic distortion (THD), voltage
source inverter. Tey et al. [17] reported a modified version of [10]. An addi-
tional PI controller is used to regulate the dc-link voltage. A full
I. I NTRODUCTION “neuromimetic” strategy involving several adalines has been
reported by Abdeslam et al. [18]. The controller can adapt for

H ARMONIC compensations have become increasingly

important in power systems due to the widespread use
of adjustable-speed drives, arc furnace, switched-mode power
unbalance and change in working conditions. Lin [19] proposed
an intelligent neural-network-based harmonic detection, which
is first trained with enough data (1400 patterns). The working
supply, uninterruptible power supply, etc. Harmonics not only model could compute the harmonic components with only one-
increase the losses but also produce unwanted disturbance half of the distorted wave. An adaline-based harmonic compen-
to the communication network, more voltage and/or current sation is reported by Singh et al. [20]. Weights are computed
stress, etc. Different mitigation solutions, e.g., passive filter, online by the LMS algorithm. Luo et al. [21] demonstrated a
active power line conditioner, and also hybrid filter, have been 200-kVA laboratory prototype for a combined system for har-
proposed and used [1]–[8]. Recent technological advancement monic suppression and reactive power compensation using an
of switching devices and availability of cheaper controlling de- optimal nonlinear PI controller, whereas a two-stage recursive
vices, e.g., DSP-/field-programmable-gate-array-based system, least square based adaline is reported by Chang et al. [22] for
make active power line conditioner a natural choice to compen- harmonic measurement.
sate for harmonics. Shunt-type active power filter (APF) is used Note that parallel developments on predictive control tech-
to eliminate the current harmonics. niques are reported for power controllers. These are also ap-
The dynamic performance of an APF is mainly dependent plied to APF [23], [24]. The implementation of APF using
on how quickly and how accurately the harmonic components power balance at the dc link is reported by Singh et al. [25]. The
are extracted from the load current. Many harmonic extrac- dc-link voltage has been used to find the peak magnitude of the
tion techniques are available, and their responses have been supply current for self-supporting dc bus. However, no detail
explored. Proposed techniques include traditional d−q [2] and analysis of the dynamics of the dc-link voltage is available.
p−q theory [3]–[5] based approaches and application of adap- This paper is an integration of predictive and adaptive control
tive filters [6], wavelet [7], genetic algorithm (GA), artificial techniques for fast convergence and reduced computations. Two
neural network (ANN), etc., for quick estimation of the com- ANN-based controllers are used for such purpose. The predic-
tive controller generates the first estimate of the compensating
Manuscript received July 10, 2009; revised December 17, 2009 and April 10, current quickly after the change in load is detected. The change
2010; accepted June 12, 2010. Date of publication August 30, 2010; date of in voltage across the capacitor is used for this purpose. This
current version January 12, 2011.
The authors are with the Department of Electrical Engineering, Indian Insti- is followed by an adaline-based controller to fast converge to
tute of Technology Kharagpur, Kharagpur 721302, India (e-mail: avik.iitkgp@ the steady value. This paper is organized in seven sections.
gmail.com; [email protected]). Section II deals with the formulation of the problem. Section III
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. investigates the dynamics of the dc-link voltage. Predicting
Digital Object Identifier 10.1109/TIE.2010.2070770 algorithm is covered in Section IV. Adaline-based algorithm is

0278-0046/$26.00 © 2011 IEEE

Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on October 02,2023 at 15:32:36 UTC from IEEE Xplore. Restrictions apply.
 422 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 58, NO. 2, FEBRUARY 2011

Fig. 1. (a) APF to compensate for a nonlinear load. (b) Single phase of shunt APF.

presented in Section V. Section VI presents the simulation and value of Iα1 following which an adaline-based ANN is used
experimental results. Section VII concludes the work. for fast convergence. Note that the inverter also draws a small
current (isα (t)) to maintain the dc-link voltage.
II. E STIMATION OF C URRENT R EFERENCE
III. C ONTROL OF DC-L INK VOLTAGE IN APF
Fig. 1(a) shows the APF compensating a nonlinear load.
Fig. 1(b) shows the corresponding schematic diagram. A The dynamics of the dc-link voltage is an indirect measure of
general expression for the load current [corresponding to the performance of the APF. Whenever there is a change in the
Fig. 1(b)] is load, the voltage across the dc-link capacitor also undergoes a
corresponding change. A controller is used to keep the voltage
iL (t) = iα1 (t) + iβ1 (t) + ih (t). (1) regulated at a desired value. In this section, a simple analysis
of the dynamics of the dc-link voltage is first carried out.
The in-phase and quadrature components of the phase current at
Parameters that govern the dynamics are identified, following
fundamental frequency are iα1 and iβ1 , respectively. All other
which an algorithm is developed to estimate the compensating
harmonics are included in ih . The per-phase source voltage and
current of the APF.
the corresponding in-phase component of the load current may
To maintain the dc-bus voltage to a desired magnitude, the
be expressed as
capacitor draws in-phase (i.e., in phase with the source voltage)
vs (t) = Vm cos ωt (2) current isα . This is in addition to the compensating current ic .
From the power balance equation
iα1 (t) = Iα1 cos ωt. (3)
dvdc
Assuming that the APF will compensate for harmonic and pdc = Cdc vdc (5)
dt
reactive power, the compensating current becomes
where pdc is the power required to maintain the voltage vdc
ic (t) = iL (t) − iα1 (t) = iL (t) − Iα1 cos ωt (4) across the dc link.
From the power balance equation
where Iα1 is the peak magnitude of the in-phase current that the
   
mains should supply and hence needs to be estimated. Once Iα1 vsi (t)isαi (t) − Rf i2sαi (t) + i2ci (t)
estimation is over, the reference current for the APF may easily i=a,b,c i=a,b,c
be set as per (4). iL (t) may be measured using current sensors.
1  d 2 
In our proposed scheme, estimation of Iα1 is carried out by two − Lf i (t) + i2ci (t) = idc (t)vdc (t) = pdc (6)
stages. A single-layer ANN-based algorithm first predicts the 2 dt sαi
i=a,b,c
Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on October 02,2023 at 15:32:36 UTC from IEEE Xplore. Restrictions apply.
BHATTACHARYA AND CHAKRABORTY: SHUNT ACTIVE POWER FILTER WITH ENHANCED PERFORMANCE 423

where Rf and Lf are the resistance and inductance of the

inductor that is connected in between the point of common
coupling and the voltage source inverter. Note that isα supplies
the system loss at the steady state and charges/discharges
the capacitor during transient to maintain the dc-link voltage.
Considering that “power” is a scalar quantity, (6) for a balanced
three-phase system may be expressed as
 
3vs (t)isα (t) − 3Rf i2sα (t) + i2c (t)
3 d 2 
− Lf isα (t) + i2c (t) = idc (t)vdc (t) = pdc . (7)
2 dt
Fig. 2. Block diagram of ANN-based peak value predictor.
Applying small perturbations in ic , isα , vdc , and vs , around
an operating point, the following new set of variables may be ities [26]:
obtained: ∗
Cdc Vdc
Is < (16)
3KP Lf
ic (t) = Ic + Δic (8)
KP V s
isα (t) = Isα + Δisα (9) Is ≤ . (17)
2Rf KP + Lf KI
vdc (t) = Vdc + Δvdc (10)
All the ac quantities in (16) and (17) are expressed as rms value.
vs (t) = Vs + Δvs (11) ∗
Is , Vdc , and Vs are the source current, reference dc-link voltage,
and source voltage, respectively. Equations (16) and (17) are
where Ic , Isα , and Vs are rms and Vdc is the dc value of the
used to generate an initial guess of KP and KI and also to set
corresponding quantities at the operating point.
their limits.
Again, in steady state
 2 
3Vs Isα − 3Rf Isα + Ic2 = 0. (12) IV. ANN-BASED FAST E STIMATION OF
C OMPENSATING C URRENT
Substituting (8)–(12) in (7), the following equation is obtained:
An ANN-based PI controller plays a dual role. It ensures
3(Δvs Isα + Vs Δisα ) − 6Rf (Isα Δisα + Ic Δic ) faster reference generation and is also accountable for better
  regulation of dc-bus voltage. The structure of the system (i.e.,
dΔisα dΔic dΔvdc
−3Lf Isα + Ic = Cdc Vdc . (13) ANN-tuned adaptive PI controller) is shown in Fig. 2. To
dt dt dt reduce computational burden, a single-layer ANN structure is
Converting the variables to s-domain and after rearranging, (13) used. The input vector as expressed in (18) is fed to the state
may be expressed as exchanger. In our scheme, error voltage and its gradient are
chosen as the state of the system to ensure faster corrective
KGs (s)G1 (s)G2 (s) ∗ action
ΔVdc (s) = ΔVdc (s)
1 + KGs (s)G1 (s)G2 (s) ∗
u = [Vdc vdc ]T . (18)
G2 (s)G3 (s)
− ΔIc (s)
1 + KGs (s)G1 (s)G2 (s) The task of the state generator block is to generate states x1 and
x2 as follows:
G2 (s)G4 (s)
+ ΔVs (s) (14)
1 + KGs (s)G2 (s)G4 (s) δx1
x1 = ve (k) x2 = (19)
δk
where K is the small-signal gain. Detail derivation of (14) from
∗
(13) is available in the Appendix. where ve (k) = Vdc − vdc (k). The output error z(k) is repre-
Equation (14) confirms that the ripple in the dc-link voltage sented as
∗
depends on ΔVdc , ΔIc , and ΔVs . In our present problem, the
distortion in source voltage and reference dc-bus voltage is not z(k) = vo (k) − vo (k − 1). (20)
considered. Therefore, (14) further modifies to
The output vo (k) is fed to output state to estimate Iα1 .
G2 (s)G3 (s) Neuron cell generates controlling signal through interrelated
ΔVdc (s) = − ΔIc (s). (15) gathering [27], [28] as
1 + KGs (s)G1 (s)G2 (s)

This explores the possibility of extracting an estimate of the 

2
u(k) = u(k − 1) + wi (k)xi (k) (21)
compensating current from the change in vdc .
i=1
The gains of the PI controller used to regulate the dc-
link voltage are governed by the following two inequal- where wi is the weight of the system.
Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on October 02,2023 at 15:32:36 UTC from IEEE Xplore. Restrictions apply.
 424 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 58, NO. 2, FEBRUARY 2011

Here, a neuron is trained by Hebb’s rule [27], [28]. Therefore,

the change of weight of the neuron cell at kth instant may be
represented as

wi (k + 1) = (1 − c)wi (k) + ηri (k) (22)

ri (k) = z(k)u(k)xi (k) (23)

where ri is the progressive signal, η is Hebb’s studying ratio

(learning rate), and “c” is a constant. Substituting (22) and (23)
in (21), the following equation may be obtained:

Δwi (k) = wi (k + 1) − wi (k)

= − c [wi (k) − ηz(k)u(k)xi (k)/c] . (24)

Δwi (k) is the change of weight at kth step. Weights of

the neuron are tuned according to Hebb’s assumption. Hebb’s
assumption is popularly known as the covariance algorithm

Δwi (k) = Fi (y1 (k), xi (k)) (25)

where Fi (∗ ) is a function of both postsynaptic and presynaptic

signals and y1 is the output of the individual neuron. If Fi (∗ ) is
differentiable, then δFi /δwi may be represented as
δFi η
= wi (k) − z(k)u(k)xi (k). (26)
δwi c
From (26), the change of weight in kth sample may be ex-
pressed as Fig. 3. (a) Adaline-based harmonic extraction for a three-phase balanced
network. (b) Adaline-based harmonic extraction for a three-phase unbalanced
δFi (k) network.
Δwi (k) = −c . (27)
δwi (k) where ts is the step size in discrete domain. The square of error
Thus, by adjusting the values of wi (k), KP and KI are tuned. terms for kth sample may be expressed as

The weights wi (k) are searched according to the negative slope s2 (k) − 2s(k)a(k)
of function Fi (∗ ). Equations (26) and (27) are used to tune the ε (k) =
2
+1 (31)
a2 (k)
parameter used in (28) and (29). Finally, for the PI controller,
the weights are represented by
where
w1 (k + 1) = w1 (k) + ηI z(k)x1 (k) (28)
a2 (k) = Iα1
2
(k) cos2 (ωkts ) + Iβ1
2
(k) sin2 (ωkts ). (32)
w2 (k + 1) = w2 (k) + ηP z(k)x2 (k). (29)
Equation (31) may also be represented as
Whenever the ANN is initiated, it starts with a set of controller 
gains to generate the first estimate of the compensating current. s2 (k) − 2s(k) X T (k)α(k)αT (k)X(k)
ε (k) =
2
+1
These initial values of controller parameters are set by offline X T (k)α(k)αT (k)X(k)
training of the ANN. The controller parameters are then ad- (33)
justed following (28) and (29) to regulate the dc-link voltage.
where the vector

V. A DAPTIVE C URRENT D ETECTION T ECHNIQUE α(k) = [Iα1 (k), Iβ1 (k)] (34)

The ANN in Section IV provides an initial guess for any and the input vector
change in system dynamics. To generate more accurate refer-
ence for APF, load current samples are fed to the adaline-based X(k) = [cos ωkts , sin ωkts ]T . (35)
network shown in Fig. 3. Adaline is designed to minimize the
total harmonic distortion (THD) of source current. Uncompen- Equation (33) is further modified to fit in terms of a quadratic
sated source current sample s(k) may be represented as equation as
 T  T 
R
 A αT (k)x(k) α (k)x(k) − B αT (k)x(k) + C = 0
s(k) = Iα1 cos(ωkts )+Iβ1 sin(ωkts )+ In cos(nωkts +φn ) (36)
n=2
(30) where A = (1 − ε2 (k)), B = 2s(k), and C = s2 (k).
Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on October 02,2023 at 15:32:36 UTC from IEEE Xplore. Restrictions apply.
BHATTACHARYA AND CHAKRABORTY: SHUNT ACTIVE POWER FILTER WITH ENHANCED PERFORMANCE 425

Fig. 5. Performance of the APF with predictive ANN (experimental results).

Fig. 4. Performance of the APF with predictive ANN (simulation results). Top
Top waveform: Source current of phase A (scale: 20 A/div). Bottom waveform:
waveform: Source current of phase A (scale: 5 A/div). Middle waveform: Load
Load current of phase A (scale: 20 A/div). Middle waveform: Compensating
current of phase A (scale: 5 A/div). Bottom waveform: Compensating current
current of phase A (scale: 10 A/div). Time scale: 10 ms/div.
of phase A (scale: 5 A/div). Time scale: 20 ms/div.

Equation (36) is minimized by conjugate gradient (CG)

method [27]–[29].
Thus, the error function (i.e., THD) is minimized to calculate
the in-phase component of the fundamental load current. The
compensating current is then calculated according to (4). Con-
vergence of this method is faster than existing adaline-based
schemes due to the use of less number of tuning blocks. The
orthogonal relationship between the input vectors reduces the
computational burden of the system.
Fig. 3(a) shows the details of current detection for a three-
phase balanced system, while Fig. 3(b) shows the same for an
unbalanced (three-phase and three-wire) network. W1a , W1b , Fig. 6. Performance of the APF with adaline (simulation results). Top wave-
form: Source current of phase A (scale: 5 A/div). Middle waveform: Load
and W1c are the corresponding in-phase components of the current of phase A (scale: 5 A/div). Bottom waveform: Compensating current
current for phase-a, phase-b and phase-c, respectively. of phase A (scale: 5 A/div). Time Scale: 20 ms/div.

VI. S IMULATION AND E XPERIMENTAL R ESULTS

Simulations have been conducted for balanced and unbal-
anced loads using SIMULINK for different controller configu-
rations. Switching frequency of the inverter is set at 10 kHz, and
the dead time of the inverter is set at 2 μs. An experimental pro-
totype is also made in the laboratory to verify the performance
of the proposed predictive and adaptive ANN-based reference
generation technique for APF. A current-controlled inverter
using insulated-gate bipolar transistors, which is available in
the laboratory, is used for this purpose. A 110-V 50-Hz mains
supplying a load of 3 kVA is considered. A dSPACE-1104- Fig. 7. Performance of the APF with adaline (experimental results). (a) Load
current of phase A (scale: 20 A/div). (b) Source current of phase A (scale:
based software, as well as hardware environment, is created 20 A/div). (c) Compensating current of phase A (scale: 20 A/div). Time scale:
where the ANN routine is written in MATLAB/C/C++ code. 20 ms/div.
The whole system is built in SIMULINK where the ANN
routine is called whenever necessary. estimate helped, the waveform quality is poor due to the lack
First, simulation study is made for the case with only pre- of any corrective mechanism in the system. The experimental
dictive algorithm. Fig. 4 shows the waveform for balanced and simulation results matched.
and nonlinear load. A diode bridge feeding a highly inductive Next, the adaptive algorithm is tried. Simulation and ex-
network is treated as the nonlinear load. The figure shows the perimentation are conducted to check the performance of the
source, load, and compensating currents in top-to-bottom order. system for a step change in load. Balance three-phase non-
Load change has occurred at 37 ms. The initial estimate of the linear load is considered similar to the case with predictive
source current is extracted from the dip in capacitor voltage algorithm. Figs. 6 and 7 show the corresponding simulation
(according to the algorithm explained in Section IV). Fig. 5 and experimental results. In the figures, t1 shows the time when
shows the corresponding results from the experimentation. The the load is changed, and t2 shows the instant when the system
THD of the source current in simulation and experimental has reached steady state. The controller took 36 and 42 ms
results are 14% and 14.8%, respectively. Although a quick to converge in simulations and experimentation, respectively.
Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on October 02,2023 at 15:32:36 UTC from IEEE Xplore. Restrictions apply.
 426 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 58, NO. 2, FEBRUARY 2011

Fig. 8. Performance of the APF with predictive and adaptive controllers

(simulation results). (a) Load current of phase A (scale: 5 A/div). (b) Source
current of phase A (scale: 20 A/div). (c) Compensating current of phase A
(scale: 5 A/div). Time scale: 10 ms/div. Fig. 11. Performance of the APF in the case of single phasing (experimental
results). (a) Source current of phase A (scale: 10 A/div). (b) Compensating
current of phase A (scale: 10 A/div). (c) Load current of phase A (scale:
20 A/div). Time scale: 10 ms/div.

Fig. 9. Performance of the APF with predictive and adaptive controllers

(experimental results). (a) Load current of phase A (scale: 20 A/div). (b) Source
current of phase A (scale: 20 A/div). (c) Compensating current of phase A
Fig. 12. Performance of the APF with synchronous reference theory-based
(scale: 10 A/div). Time scale: 10 ms/div.
reference generation technique (simulation results). (a) Source current of
phase A (scale: 5 A/div). (b) Load current of phase A (scale: 5 A/div).
(c) Compensating current of phase A (scale: 5 A/div). Time scale: 20 ms/div.

harmonic contents (particularly the lower order harmonics) for

both compensated and uncompensated source currents. The
THD of the compensated source current is found to be 4.5%.
Some experiments are also conducted for unbalanced nonlinear
load. A severe unbalance such as single phasing is considered
as shown in Fig. 11.
The controller is found to operate satisfactorily. Finally,
the performance of the proposed controller is compared with
that of a traditional controller (such as synchronous reference
frame based). The synchronous reference frame controller is
simulated (using the cutoff frequency of the low-pass filter as
10 Hz). Fig. 12 shows corresponding dynamic performance.
The source current, load current, and compensating current
are shown in top-to-bottom order. The controller took 51 s to
converge, whereas the proposed controller with predictive and
Fig. 10. Harmonic spectra for the compensated source current (series 1) and adaptive algorithm converged within less than one-quarter of a
load current (series 2) of Fig. 9. cycle with acceptable current quality.

Note that the system took more time to converge as there is no

VII. C ONCLUSION
predictive mechanism; however, the waveform quality is much
better with the THDs of 5.1% and 5.3% for the simulation and An integration of predictive and adaptive ANN-based con-
experimental results, respectively. troller for a shunt-type APF has been presented in this paper
Now, to have the advantage of predictive and adaptive con- to improve the convergence and reduce the computational re-
trollers, the system is run with both the algorithms. Figs. 8 quirement. The predictive algorithm is derived from an ANN-
and 9 show the situation with both the predictive and adaptive based PI controller used to regulate the dc-link voltage in the
controllers in operation. The results have confirmed very satis- APF. This is followed by an adaline-based THD minimization
factory performance in terms of waveform quality and response technique. Adaline is trained by CG method to minimize THD.
time. The THDs for both the simulation and experimental Use of only two weights and two input vectors makes the
results remain less than 5%. Fig. 10 shows the individual convergence very fast. The system is extensively simulated
Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on October 02,2023 at 15:32:36 UTC from IEEE Xplore. Restrictions apply.
BHATTACHARYA AND CHAKRABORTY: SHUNT ACTIVE POWER FILTER WITH ENHANCED PERFORMANCE 427

in MATLAB/SIMULINK. An experimental prototype is pro- R EFERENCES

duced in the laboratory using dSPACE1104. The results from [1] B. Singh, K. Al-Haddad, and A. Chandra, “A review of active power filters
experiments match well with the simulation, confirming the for power quality improvements,” IEEE Trans. Ind. Electron., vol. 46,
no. 5, pp. 960–971, Oct. 1999.
usefulness of the proposed technique. [2] S. Bhattacharya, T. M. Frank, D. M. Divan, and B. Banerjee, “Active filter
system implementation,” IEEE Ind. Appl. Mag., vol. 4, no. 5, pp. 47–63,
Sep./Oct. 1998.
A PPENDIX [3] H. Akagi, Y. Kanazawa, and A. Nabae, “Instantaneous reactive power
compensators comprising switching devices without energy storage com-
Taking Laplace transformation of (13), the following equa- ponents,” IEEE Trans. Ind. Appl., vol. IA-20, no. 3, pp. 625–630,
tion may be obtained: May 1984.
[4] F. Z. Peng, G. W. Ott, Jr., and D. J. Adams, “Harmonic and reactive power
compensation based on the generalized instantaneous reactive power the-
3ΔVs (s)Isα + ΔIsα (s)(3Vs − 6Rf Isα − 3sLf Isα ) ory for three phase four wire system,” IEEE Trans. Power Electron.,
vol. 13, no. 6, pp. 1174–1181, Nov. 1998.
−ΔIc (s)(6Rf Ic + 3sLf Ic ) = sΔVdc (s)Cdc Vdc . (A1) [5] R. S. Herrera and P. Salmeron, “Instantaneous reactive power theory: A
reference in the nonlinear loads compensation,” IEEE Trans. Ind. Elec-
tron., vol. 56, no. 6, pp. 2015–2022, Jun. 2009.
Consider [6] H. Karimi, M. Karimi-Ghartemani, and M. R. Iravani, “An adaptive filter
for synchronous extraction of harmonics and distortions,” IEEE Trans.
G1 (s) = − 3(Vs − 2Rf Isα + sLf Isα ) (A2) Power Del., vol. 18, no. 4, pp. 1350–1356, Oct. 2003.
[7] M. Forghani and S. Afsharnia, “Online wavelet transform-based control
1 strategy for UPQC control system,” IEEE Trans. Power Del., vol. 22,
G2 (s) = (A3) no. 1, pp. 481–491, Jan. 2007.
Cdc Vdc s [8] A. Bhattacharya, C. Chakraborty, and S. Bhattacharya, “Current compen-
G3 (s) = 3(2Rf Ic + sLf Ic ) (A4) sation in shunt type active power filters,” IEEE Ind. Electron. Mag., vol. 3,
no. 3, pp. 38–49, Sep. 2009.
G4 (s) = 3Isα . (A5) [9] A. Hamadi, S. Rahmani, and K. Al-Haddad, “A hybrid passive filter
configuration for VAR control and harmonic compensation,” IEEE Trans.
Ind. Electron., vol. 57, no. 7, pp. 2419–2434, Jul. 2010.
Substituting (A2)–(A5) in (A1), the following equation may be [10] B. Widrow and M. A. Lehr, “30 years of adaptive neural networks:
derived: Perceptron, madaline and backpropagation,” Proc. IEEE, vol. 78, no. 9,
pp. 1415–1442, Sep. 1990.
[11] C.-C. Tsai, H.-C. Huang, and S.-C. Lin, “Adaptive neural network control
ΔVs (s)G4 (s) + ΔIsα (s)G1 (s) − ΔIc (s)G3 (s) of a self-balancing two-wheeled scooter,” IEEE Trans. Ind. Electron.,
vol. 57, no. 4, pp. 1420–1428, Apr. 2010.
1
= ΔVdc (s) . (A6) [12] A. Dinu, M. N. Cirstea, and S. E. Cirstea, “Direct neural-network
G2 (s) hardware-implementation algorithm,” IEEE Trans. Ind. Electron., vol. 57,
no. 5, pp. 1845–1848, May 2010.
The relation between isα and vdc may be expressed as [13] G. W. Chang, C.-I Chen, and Y.-F. Teng, “Radial-basis-function-based
neural network for harmonic detection,” IEEE Trans. Ind. Electron.,
vol. 57, no. 6, pp. 2171–2179, Jun. 2010.
dvdc
isα = K (A7) [14] P. K. Dash, D. P. Swain, A. C. Liew, and S. Rahman, “An adaptive linear
dt combiner for on-line tracking of power system harmonics,” IEEE Trans.
Power Syst., vol. 11, no. 4, pp. 1730–1735, Nov. 1996.
where K is the small-signal gain and isα is the part of the source [15] Y. M. Chen and R. M. O’Connell, “Active power line conditioner with a
neural network control,” IEEE Trans. Ind. Appl., vol. 33, no. 4, pp. 1131–
current used to stabilize the dc-bus voltage. Instantaneous dc- 1136, Jul./Aug. 1997.
link voltage is compared with the reference dc voltage, and the [16] L. L. Lai, C. T. Tse, W. L. Chan, and A. T. P. So, “Real time frequency and
error voltage in Laplace domain may be expressed as harmonic evaluation using artificial neural network,” IEEE Trans. Power
Del., vol. 14, no. 1, pp. 52–59, Jan. 1999.
∗ [17] L. H. Tey, P. L. So, and Y. C. Chu, “Improvement of power quality using
Vdc
Ve (s) = − Vdc (s). (A8) adaptive shunt filter,” IEEE Trans. Power Del., vol. 20, no. 2, pp. 1558–
s 1568, Apr. 2005.
[18] D. O. Abdeslam, P. Wira, J. Merckle, D. Flieller, and Y. A. Chapuis,
A PI controller may be used to maintain the capacitor voltage. “A unified artificial neural network architecture for active power filters,”
Therefore, the output of the PI controller may be expressed as IEEE Trans. Ind. Electron., vol. 54, no. 1, pp. 61–76, Feb. 2007.
[19] H. C. Lin, “Intelligent neural network-based fast power system har-
  monic detection,” IEEE Trans. Ind. Electron., vol. 54, no. 1, pp. 43–52,
KI
Vo (s) = KP + Ve (s). (A9) Feb. 2007.
s [20] B. Singh, V. Verma, and J. Solanki, “Neural network-based selective
compensation of current quality problems in distribution system,” IEEE
Thus, (A7) may be modified as Trans. Ind. Electron., vol. 54, no. 1, pp. 53–60, Feb. 2007.
[21] A. Luo, Z. Shuai, W. Zhu, and Z. J. Shen, “Combined system for har-
 ∗
 monic suppression and reactive power compensation,” IEEE Trans. Ind.
ΔVdc
ΔIsα (s) = KGs (s) − ΔVdc (s) (A10) Electron., vol. 56, no. 2, pp. 418–428, Feb. 2009.
s [22] G. W. Chang, C. I. Chen, and Q. W. Liang, “A two-stage adaline for
harmonics and interharmonics measurement,” IEEE Trans. Ind. Electron.,
vol. 56, no. 6, pp. 2220–2228, Jun. 2009.
where [23] J. H. Marks and T. C. Green, “Predictive transient-following control
  of shunt and series active power filters,” IEEE Trans. Power Electron.,
KI vol. 17, no. 4, pp. 574–584, Jul. 2002.
KP + = Gs (s). (A11)
s [24] M. Routimo, M. Salo, and H. Tuusa, “A novel simple prediction based
current reference generation method for an active power filter,” in Proc.
IEEE APEC, 2004, pp. 3215–3220.
Substituting values of ΔIsα (s) in (A1), (14) is obtained (in [25] B. N. Singh, B. Singh, A. Chandra, and K. Al-Haddad, “Design and
Section III). digital implementation of active filter with power balance theory,” Proc.
Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on October 02,2023 at 15:32:36 UTC from IEEE Xplore. Restrictions apply.
 428 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 58, NO. 2, FEBRUARY 2011

Inst. Elect. Eng.—Electr. Power Appl., vol. 152, no. 5, pp. 1149–1160, Chandan Chakraborty (S’92–M’97–SM’01) was
Sep. 2005. born in Kolkata, India, on August 17, 1965. He
[26] M. H. Rashid, Recent Development in Power Electronics: Selected Read- received the Ph.D. degrees in electrical engineer-
ing. Piscataway, NJ: IEEE Press, 1996. ing from the Indian Institute of Technology (IIT),
[27] S. Kumar, Neural Networks—A Class Room Approach. New Delhi, Kharagpur, India, and Mie University, Tsu, Japan, in
India: Tata McGraw-Hill, 2004. 1997 and 2000, respectively.
[28] F. M. Ham and I. Kostanic, Principles of Neurocomputing for Science & He is currently with the Department of Electrical
Engineering. New Delhi, India: Tata McGraw-Hill, 2002. Engineering, IIT Kharagpur, as an Associate Pro-
[29] J. Nocedal and S. J. Wright, Numerical Optimization. New York: fessor. His research interests include power convert-
Springer-Verlag, 1999. ers, motor drives, electric vehicles, and renewable
energy.
Dr. Chakraborty is a Fellow of the Indian National Academy of Engineering.
Avik Bhattacharya (S’09) received the B.E. de- He was the Technical Program Chair of the 2nd National Power Electronics
gree in electrical and electronics engineering from Conference held in Kharagpur in 2005 and also the Technical Program Cochair
Sikkim Manipal University, Sikkim, India, in 2001 of the ICIT2006, ICIT2008, and ICIT2010 held in Mumbai, Melbourne, and
and the M.Tech. degree in machine drive and power Chile, respectively. He is on the ADCOM of the IEEE Industrial Electronics
electronics from the Indian Institute of Technology, Society. He is one of the Associate Editors of the IEEE T RANSACTIONS ON
Kharagpur, India, in 2005, where he is currently I NDUSTRIAL E LECTRONICS and IEEE Industrial Electronics Magazine and an
working toward the Ph.D. degree in the Department Editor of the IEEE T RANSACTIONS ON S USTAINABLE E NERGY. He received
of Electrical Engineering. the Bimal Bose Award in Power Electronics in 2006 from the Institute of
His research interests include power converter Electronics and Telecommunication Engineers (India).
topologies and their control, power quality, and soft-
computing techniques.

Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on October 02,2023 at 15:32:36 UTC from IEEE Xplore. Restrictions apply.