You are on page 1of 14

Electric Power Systems Research 60 (2002) 179– 192

www.elsevier.com/locate/epsr

Active power filter for three-phase four-wire electric systems using


neural networks
Chakphed Madtharad, Suttichai Premrudeepreechacharn *
Power Electronic Research Laboratory, Department of Electrical Engineering, Chiang Mai Uni6ersity, Chiang Mai, Thailand

Received 5 February 2001; received in revised form 12 September 2001; accepted 5 October 2001

Abstract

This paper presents the design of neural networks compared with the conventional technique, a hysteresis controller for active
power filter for three-phase four-wire electric system. A particular three-layer neural network structure is studied in some detail.
Simulation and experimental results of the active power filter with both controllers are also presented to verify the feasibility of
such controller. The simulation and experimental result show that both controller techniques can reduce harmonics in three-phase
four-wire electric systems drawn by nonlinear loads and can reduce neutral current. The advantage of the neural network
controller technique over hysteresis controller technique are less voltage ripple of d.c. bus, and less switching loss. Furthermore,
the neural networks controller has better fault tolerance than the hysteresis controller. © 2002 Elsevier Science B.V. All rights
reserved.

Keywords: Active power filter; Power quality; Neural networks

1. Introduction factor correction capacitors. The capacitors may over-


heat and fail when they are exposed to significant
There has been a continuous proliferation of non- harmonic currents.
linear type of loads due to the intensive use of elec- The active approaches have proven to be very ef-
tronics control in all branches of industry as well as fective [1–5]. Three single-phase active power filters
by general consumers. Conventional rectifiers are har- (APF) can be used for this propose. However, a con-
monic polluters of the power distribution system. ventional three-phase, three-wire APF cannot be used
Non-linear loads, especially power electronics loads, in a three-phase four-wire system to eliminate har-
create phase displacement and harmonic currents in monics on the neutral wire. The process of filtering is
the main three-phase power distribution system. Both done in the time domain which is based on the prin-
make the power factor of the system worse. The pres- ciple of holding the instantaneous source voltage or
ence of harmonic currents can also lead to some spe- current within some reasonable tolerance of a sine
cial problems in three-phase systems. In a three-phase wave. The harmonic components are compensated in-
four-wire systems, harmonic currents can lead to large stantaneously using current control technique.
currents in the neutral conductors, which may easily The past decade had seen a dramatic increase in
exceed the conductor rms current rating. Harmonic interest in neural network systems. The application of
currents tend to flow through shunt-connected power neural networks promises high computation rate pro-
vided by the massive parallelism, a great degree of
robustness or fault tolerance due to the distribution
representation, and an ability for adaptation, learn-
* Corresponding author. ing, and generalization to improve performance. To-
E-mail address: suttic@eng.cmu.ac.th (S. Premrudeepreechacharn). day neural networks are actively explored in artificial

0378-7796/02/$ - see front matter © 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 8 - 7 7 9 6 ( 0 1 ) 0 0 1 8 5 - 7
180 C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192

Fig. 1. An APF and nonlinear loads considered in this paper.

Fig. 2. A controller for APF.


C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192 181

Fig. 3. Neural network structure for APF.

intelligence, psychology, engineering, and physics. The Table 1


neural networks can be applied to power electronics Training pattern for neural network controller
areas such as power converter control, current regula-
Input pattern Desired pattern
tion control, motor speed regulation, etc. [6– 8]. Neural
networks controllers are proposed in this paper as a i*ca−ica i*cb−icb i*cc−icc Va Vb Vc
mean to solve the problems introduced by non-lineari-
ties in APF. 1 1 1 1 1 1
This paper presents the application of neural net- 1 1 −1 1 1 −1
1 −1 1 1 −1 1
works controllers for APF for three-phase four-wire 1 −1 −1 1 −1 −1
distribution systems. The system considered in this −1 1 1 −1 1 1
paper is shown in Fig. 1. Section 2 of this paper −1 1 −1 −1 1 −1
provides the fundamentals of APF. In Section 3, the −1 −1 1 −1 −1 1
structure of the controller is discussed. In this paper, −1 −1 −1 −1 −1 −1
back-propagation neural networks is considered. The
proposed neural network controllers are presented. Fi- Table 2
nally, simulation results verifying the concept are also The system parameters used in simulating the system shown in Fig. 1
presented.
Utility source Voltage source (Vs) 75 Vrms
Frequency (F) 50 Hz
Source resistance (Rs) 2.5 V
Source inductance (Ls) 0.4 mH
2. Active power filter system
Non-linear load Load capacitance (Cl) 1500 mF
The main objective of the APF is to compensate Load resistance (Rl) 27 V
harmonics, reactive power, neutral current and unbal-
Active power filter (APF) APF resistance (Rc) 1.6 V
ancing of non-linear loads locally such that a.c. mains APF inductance (Lc) 12.2 mH
supplies only unity power-factor sinusoidal balanced APF capacitance (Cc) 2200 mF
three-phase currents. The APF draws the required cur-
182 C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192

Fig. 4. The system performance of APF: (a) The nonlinear load current (I1a); (b) The source current using hysteresis controller (Isa); and (c) The
source current using neural networks controller (Isa).
C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192 183

rents from the a.c. mains to feed harmonics, reactive three single phase inverters. The switches of APF must
power, and neutral current for balancing of load cur- support unipolar voltage and bipolar current, an
rents locally and causes balanced sinusoidal unity IGBT with antiparallel diode is needed to implement
power-factor supply currents under all operating each switch. The current which must be supported by
conditions. each switch is the maximum inductor current. The
Fig. 1 shows the basic APF scheme including a set maximum voltage which must be supported by con-
of non-linear loads on a three-phase four-wire distri- trollable switches is the maximum d.c. bus voltage.
bution systems. In this paper, the APF consists of The load may be either single phase, two phase or

Fig. 5. The frequency spectra for the system: (a) The nonlinear load current (I1a); (b) The source current using hysteresis controller (Isa); and (c)
The source current using neural networks controller (Isa).
184 C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192

Fig. 5. (Continued)

three phase and non-linear in nature. In this paper, we I*cx = I*sx − Ilx (1)
consider three single phase uncontrolled diode bridge
rectifiers with resistive– capacitive loading as non-linear where I*cx is the APF current for phase x; I*sx, source
unbalanced loads. This load draws a non-sinusoidal current for phase x; Ilx, load current for phase x.
current from the utility. Thus, the current references consist of the harmonic
The controller for APF is comprised of an inner current components drawn by load. By forcing the filter
current loop which actively shapes the line currents and output current to follow the reference, only a fundamen-
an outer voltage control loop which regulates the mag- tal frequency sinusoidal current is drawn from the utility
nitude of the line currents as shown in Fig. 2. The inner or distribution transformer. Hence, undesirable har-
current loop can use a current regulator scheme such as monic current components are removed from the utility
hysteresis to shape the line current. Among the various side of the system.
current control techniques, hysteresis current control is Nominal d.c. bus voltage must be at least
2Vrms in
the simplest and most extensively used technique. How- order to assure control over the shape of the APF current
ever, a fixed-band hysteresis current control has draw- at all times [3]. Since the compensating currents should
backs on various switching frequencies and has large be cyclic with the line frequency, the capacitor voltage
ripple current. In this paper, we use the neural networks will contain ripple but its average will be stable provided
controllers as current regulator for APF. The outer that there is an appropriate balance of power. Therefore,
voltage loop regulates the average capacitor voltage by we are not concerned with the voltage ripple as long as
using a proportional-integral (PI) controller. In addition, the voltage is always at least
2Vrms. To charge and
the magnitude of the line current is dictated by the outer maintain adequate charge on the d.c. side capacitor, a PI
voltage control loop. regulator will be used to control the flow of real power
The current references for each of the three phase APF from the a.c. side towards the d.c. side of the converter.
are derived by sensing the load currents and removing Since the converter is designed only to compensate
from their source references current which is sinusoidally harmonics, which does not include the fundamental, this
waveshaped as shown in Eq. (1). real power transfer merely compensates the losses in the
C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192 185

Fig. 6. The system performance of APF: (a) The nonlinear load current (I1a); (b) The source current using hysteresis controller (Isa); and (c) The
source current using neural networks controller (Isa); (d) The neutral of nonlinear load current (Inl); (e) The neutral of source current using
hysteresis controller (Inl); (f) The neutral of source current using neural networks controller (Inl); (g) The ac component of DC bus voltage using
hysteresis controller (Vdc); and (h) The ac component of DC bus voltage using neural networks controller (Vdc).

various filter components, switches and intercon- The choice of capacitance value and capacitor type
nections. Hence, the outer voltage control loop uses the depends on both the minimum necessary d.c. bus
PI controller to sense the average voltage across the voltage that it must support and also the rms and
capacitor and adjust the gain of reference currents to harmonic components of the current that pass through
maintain the desired bus voltage. If the gain of it. The capacitor must be capable of handling the
reference current is too large, power flow dictates that required rms current without overheating and the
the average capacitor voltage must increase as the harmonic components should not cause excessive ripple
capacitor absorbs the excess power delivered by the on the d.c. bus voltage. This knowledge is utilized in
source or vice versa. rating the converter switches.
186 C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192

Fig. 6. (Continued)

3. Neural networks and parameter variations which may occur in a con-


trolled plant. In this paper, back-propagation neural
Neural networks have the potential to provide an networks are utilized as pattern classifier. Back-propa-
improved method of deriving non-linear models which gation neural networks are an example of nonlinear
is complementary to conventional techniques. Neural layered feed-forward networks. It is a universal approx-
networks have self-adapting capabilities which make imator [9,10].
them well suited to handle non-linearities, uncertainties The structure of the proposed neural networks used
C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192 187

Fig. 6. (Continued)

for APF is shown in Fig. 3. The neural networks were sufficient information to cover the essential characteris-
trained using MATLAB software. We trained the neu- tic of entire control mapping. Table 1 summarizes the
ral networks with a learning rate of 0.25 until the error eight training patterns for the neural network con-
function was less than 5×10 − 5. troller. The gating signal for controllable switches is the
The proposed neural network controllers are based output of the neural networks passed through the hard
on hysteresis current control. The neural network con- limiter given by
troller requires the training patterns to learn the desired If Vx \ 0, gating signal= 1, else gating signal=0
control mapping. The training patterns should have (2)
188 C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192

Fig. 6. (Continued)

4. Simulation results Fig. 4 documents the performance of the system for


one operating point. Fig. 4(a) show the nonlinear load
This section discusses the operation of the system current. The source current using hysteresis with
shown in Fig. 1. The system was simulated and hysteresis band is 0.01 A and neural network
evaluated to learn more about the operation of the controllers have been shown in Fig. 4(b,c), respectively
APF. The system components of Fig. 1 that are used at a maximum switching frequency of 5.2 kHz.
in the simulation are described in Table 2. Simulation Fig. 5 shows the spectral performance of the
studies are carried out to predict the performance of system. The total harmonic distortion (THD) is based
the proposed APF using MATLAB. on the 1st–35th odd harmonic of current waveform
C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192 189

and a pure sinewave for the voltage. That means the 5. Experimental results
distortion voltages present in the source voltage have
not been considered with regard to the calculation of This section discusses the operation of the system
THD. Fig. 5 shows the significant reduction in the shown in Fig. 1. The system was experimental and
harmonics drawn from the utility by using APF. The evaluated to learn more about the operation of the
APF is effectively used to reduce its THD of nonlinear APF. The system components of Fig. 1 that are used in
load current as shown in Fig. 5(a) from 69.76 to 11.73% the experiment are described in Table 2 at a maximum
and 9.60% using hysteresis and neural network con- switching frequency of 5.2 kHz.
trollers in Fig. 5(b,c), respectively at a maximum Fig. 6 documents the performance of the system for
switching frequency of 5.2 kHz. one operating point. Fig. 6(a) shows the nonlinear load

Fig. 7. The frequency spectra for the system: (a) The nonlinear load current (I1a); (b) The source current using hysteresis controller (Isa); and (c)
The source current using neural networks controller (Isa).
190 C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192

Fig. 7. (Continued)

current. The source current using hysteresis and neu- cannot, as shown in Fig. 8(a,b). That is, the neural
ral network controllers have been shown in Fig. 6(b) network controller has a better fault tolerance than
and (c), respectively. Fig. 6(d– f) show the neutral the hysteresis controller.
current of nonlinear load, neutral source current us- The THD of source in different conditions is
ing hysteresis and neural network controllers, respec- shown in Fig. 9. Fig. 9(a) shows that the higher the
tively. switching frequency, the lower the THD of the
Fig. 6(g) and (h) show the a.c. components of the source. The higher the THD of the load, the higher
d.c. bus capacitor of APF using hysteresis and neural the THD of the source as shown in Fig. 9(b). Fig.
network controllers, respectively. Since the compen- 9(a,b) shows that the hysteresis and neutral network
sating currents should be cyclic with the line fre- controllers can reduced harmonics in almost the same
quency, the capacitor voltage will contain ripple but capacity under different conditions.
its average will be stable provided that there is an
appropriate balance of power. As previously dis-
cussed, we are not con-cerned with the voltage ripple
6. Conclusion
as long as the voltage always at least
2Vrms. The
small amount of ripple on the d.c. bus voltage sug-
This paper has focused on applying back-propaga-
gests that a smaller bus capacitor may be used if we
tion neural networks to control the three single-phase
are not concerned with transient operation.
APF compared with hysteresis controller at a maxi-
Fig. 7 shows the spectral performance of the sys-
tem. The APF is effective in reducing its THD of mum switching frequency of 5.2 kHz. The simulation
nonlinear load current as shown in Fig. 7 (a) from and experimental results with the same parameters
69.33 to 11.47% and 12.79% by using hysteresis and have shown that the APF with both controllers is
neural networks controllers in Fig. 7(b) and (c), re- able to compensate the distortion and reactive power
spectively at a maximum switching frequency of 5.2 drawn by the nonlinear loads. In addition, the APF
kHz. also helps to reduce the undesirable current that flows
In the case of having some error in phase A of in the neutral line of a three-phase four-wire distribu-
controller, neural networks controller can work effec- tion system. By simulation, the APF is effectively
tively all of the three phase but hysteresis controller used to reduce its THD of nonlinear load current
C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192 191

Fig. 8. The system performance of APF when there is an error in phase A of the controller: (a) The source current using hysteresis controller (Isa);
and (b) The source current using neural networks controller (Isa).

from 69.76 to 11.73% and 9.60%, and by experiment, Acknowledgements


from 69.33 to 11.47% and 12.79% using hysteresis and
neural networks, respectively. The advantages of the The authors would like to thank The Shell Centenary
neural network controller technique over hysteresis Scholarship Fund and Graduated School Fund of Chi-
controller technique are less voltage ripple of d.c. bus, ang Mai University who supported this research. This
and less switching loss. Furthermore the neural net- paper was written when the first author studied in
works controller has better fault tolerance than the Chiang Mai University under the support of Provincial
hysteresis controller. Electricity Authority (PEA).
192 C. Madtharad, S. Premrudeepreechacharn / Electric Power Systems Research 60 (2002) 179–192

Fig. 9. The THD of source under different conditions: (a) The relationship between switching frequency and THD of source; and (b) The
relationship between THD of load and THD of source.

four-wire electric distribution systems supplying non-linear


References loads, Electr. Power Syst. Res. 44 (1998) 93 – 100.
[5] S. Saetieo, R. Devaraj, D.A. Torrey, The design and implemen-
[1] C.L. Chen, C.E. Lin, C.L. Huang, An active filter for unbal- tation of a three-phase active power filter based on sliding mode
anced three-phase system using synchronous detection method, control, IEEE Trans. Ind. Appl. 31 (5) (1995) 993 – 1000.
IEEE Power Electron. Specialists Conf. Record (1994) 1451 – [6] F. Harashima, Y. Demizu, S. Kondo, H. Hashimoto, Applica-
1455. tion of neural networks to power converter control, Conf.
[2] P. Enjeti, W. Shireen, P. Packebush, I. Pitel, Analysis and design Record of IEEE Ind. Appl. Society IAS Annual Meeting (1989)
of a new active power filter to cancel neutral current harmonics 1086 – 1091.
in three phase four wire electric distribution systems, Conf. [7] B.R. Lin, R.G. Hoft, Power electronics converter control based
Record of IEEE Ind. Appl. Society (IAS) Annual Meeting on neural network and fuzzy logic methods, IEEE Power Elec-
(1993) 939 – 946. tron. Specialists Conf. Record (1993) 900 – 906.
[3] C.A. Quinn, N. Mohan, Active filtering of harmonic currents in [8] H.M. Tai, J. Wang, K. Ashenayi, Motor Speed Regulation
three phase, four-wire systems with three-phase and single-phase Using Neural Networks, Proc. of the IECON (1990) 1215 –1220.
non-linear loads, Proceedings of Applied Power Electron. Conf. [9] S. Haykin, Neural Networks, Macmillan, New York, 1994.
(APEC) (1992) 829 –835. [10] F.V. Laurene, Fundamentals of Neural Networks: Architectures,
[4] B. Singh, K. Al-Haddad, A. Chandra, Harmonic elimination, Algorithms, and Applications, Prentice-Hall, Englewood Cliffs,
reactive power compensation and load balancing in three-phase, NJ, 1994.