Vol 63, No. 11;Nov 2013 377 Jokull Journal COMPARATIVE STUDY BETWEEN FUZZY LOGIC AND ARTIFICIAL NEURAL NETWORK (ANN) ALGORITHMS FOR SINGLE PHASE SHUNT ACTIVE POWER FILTERS (SAPFs) ALIYU SABO, NOOR IZZRI ABDULWAHAB, HAMISU USMAN, MOHD AMRAN MOHD RADZI, NASHIREN FARZILAH MAILAH Abstract-The excessive use of power electronics devices in industrial, commercial and residential purposes have lead to the deterioration of supply current and voltage wave forms, injecting harmonic pollution on to the supply system, which no doubt generates a major concern to power system engineers. Restricted standards set by IEEE-519 or IEC-61000-3-2 necessitates that, those unwanted harmonic injected current (into the utility networks) to fall below a specified range. Present day technological development, particularly in the field of power converters, introduces the application of active power filters as a modern weapon for harmonic current mitigation and reactive power compensation leading us to IEEE norms realization. This paper depicts comprehensively, comparison in terms of total harmonic distortion (THD) and simulation results, the most effectively fast response in harmonic mitigation between fuzzy logic based active power filter (APF) control and its artificial neural network (ANN) controller counterpart. In both cases, results were developed via simulation studies under MATLAB/SIMULINK environment. Keywords: Shunt active power filter, harmonics, total harmonic distortion, artificial neural network, fuzzy logic controller and power factor. 1. INTRODUCTION Present day technological development, particularly in the field of power electronic offers remarkable achievements in terms of its fast switching and high quality capability, less heavy in weight, operational flexibility, robustness, size minimization etc. On the other hand, they also dispense harmonics in networks due to their nonlinearity nature. Harmonics level minimization in the grid caused due to these development becomes a serious issue threatened power system engineers particularly, the utility companies. Classical passive filters techniques are adopted during the 20 th age, which are now seen as an armature technology as it adversely depends on networks impedance resulting in an unwanted resonance issues. Active power filters which is adopted as the latest matured development in handling effectively current harmonics mitigation, power factor optimization and reactive power compensation due to its flexibility and robustness in handling many types of harmonic pollution agent’s devices (power electronic devices) also providing quick response to load variations. This new trend finally abolishes the armature passive filters usage and its drawbacks. Various configurations of APF have been exploited together with its distinct control strategies in ensuring optimal operating state [1][2][3]. Several controller techniques for shunt APF’s have been developed in the past [4][5]. In [4][5][6], harmonic current extraction strategies can be realize via time-based, heterodyne, frequency-based, instantaneous power compensation and pattern recognition and learning (where mainly neural network technique). Similarly, previous work of the shunt APF presents an optimization algorithm in extracting current harmonics [4][7]. In general, shunt APF works on the principle of current harmonic injection. Its main job is to deliver compensated current to the supply network at the common coupling point, as such, to cancel the harmonics current and reactive power drawn by the nonlinear loads. SAPF configuration could be either voltage source inverter (VSI) or current source inverter (CSI) types, depending on applications to be used. But (VSI) type is the most widely employed in active filtering due to its simplicity and popular in recognition topology. Basically, shunt APF comprises the power circuit and control circuit. The power circuit incorporates the IGBT or MOSFET semiconductors switching devices with an interfacing inductor for taking care of the harmonics ripples after compensation by the active filter, DC
11
Embed
COMPARATIVE STUDY BETWEEN FUZZY LOGIC AND ARTIFICIAL NEURAL NETWORK (ANN) ALGORITHMS FOR SINGLE PHASE SHUNT ACTIVE POWER FILTERS (SAPFs
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Vol 63, No. 11;Nov 2013
377 Jokull Journal
COMPARATIVE STUDY BETWEEN FUZZY LOGIC AND ARTIFICIAL
NEURAL NETWORK (ANN) ALGORITHMS FOR SINGLE PHASE SHUNT
ACTIVE POWER FILTERS (SAPFs)
ALIYU SABO, NOOR IZZRI ABDULWAHAB, HAMISU USMAN, MOHD AMRAN MOHD RADZI,
NASHIREN FARZILAH MAILAH
Abstract-The excessive use of power electronics devices in industrial, commercial and residential
purposes have lead to the deterioration of supply current and voltage wave forms, injecting harmonic pollution
on to the supply system, which no doubt generates a major concern to power system engineers. Restricted
standards set by IEEE-519 or IEC-61000-3-2 necessitates that, those unwanted harmonic injected current (into
the utility networks) to fall below a specified range. Present day technological development, particularly in the
field of power converters, introduces the application of active power filters as a modern weapon for harmonic
current mitigation and reactive power compensation leading us to IEEE norms realization. This paper depicts
comprehensively, comparison in terms of total harmonic distortion (THD) and simulation results, the most
effectively fast response in harmonic mitigation between fuzzy logic based active power filter (APF) control and
its artificial neural network (ANN) controller counterpart. In both cases, results were developed via simulation
studies under MATLAB/SIMULINK environment.
Keywords: Shunt active power filter, harmonics, total harmonic distortion, artificial neural network,
fuzzy logic controller and power factor.
1. INTRODUCTION
Present day technological development, particularly in the field of power electronic offers remarkable achievements
in terms of its fast switching and high quality capability, less heavy in weight, operational flexibility, robustness,
size minimization etc. On the other hand, they also dispense harmonics in networks due to their nonlinearity nature.
Harmonics level minimization in the grid caused due to these development becomes a serious issue threatened
power system engineers particularly, the utility companies. Classical passive filters techniques are adopted during
the 20th
age, which are now seen as an armature technology as it adversely depends on networks impedance resulting
in an unwanted resonance issues. Active power filters which is adopted as the latest matured development in
handling effectively current harmonics mitigation, power factor optimization and reactive power compensation due
to its flexibility and robustness in handling many types of harmonic pollution agent’s devices (power electronic
devices) also providing quick response to load variations. This new trend finally abolishes the armature passive
filters usage and its drawbacks.
Various configurations of APF have been exploited together with its distinct control strategies in ensuring optimal
operating state [1][2][3]. Several controller techniques for shunt APF’s have been developed in the past [4][5]. In
[4][5][6], harmonic current extraction strategies can be realize via time-based, heterodyne, frequency-based,
instantaneous power compensation and pattern recognition and learning (where mainly neural network technique).
Similarly, previous work of the shunt APF presents an optimization algorithm in extracting current harmonics [4][7].
In general, shunt APF works on the principle of current harmonic injection. Its main job is to deliver compensated
current to the supply network at the common coupling point, as such, to cancel the harmonics current and reactive
power drawn by the nonlinear loads. SAPF configuration could be either voltage source inverter (VSI) or current
source inverter (CSI) types, depending on applications to be used. But (VSI) type is the most widely employed in
active filtering due to its simplicity and popular in recognition topology. Basically, shunt APF comprises the power
circuit and control circuit. The power circuit incorporates the IGBT or MOSFET semiconductors switching devices
with an interfacing inductor for taking care of the harmonics ripples after compensation by the active filter, DC
Vol 63, No. 11;Nov 2013
378 Jokull Journal
capacitor for maintaining DC voltage by the inverter. While the control circuit is the main brain of the filter, which
control the semiconductor switching gating signal for realization by the shunt APF. The depiction of the shunt APF
configuration is shown (fig 1).
Fig. 1 Single phase shunt APF configuration
In this paper, fuzzy logic and neural network algorithms in shunt APF control is been compared with results
obtained discussions to verify the most effective algorithm in shunt APF control for current harmonics mitigation.
Section 2 of this paper entails the control system development (2.1 for neural network and 2.2 for fuzzy logic
algorithms). Section 3 depicts the simulation outputs and lastly, section 4 dispenses the conclusion.
2. CONTROL TECHNIQUE
2.1 Neural network algorithm
This section explains the operation of the SAPF with adaptive/artificial neural network (ANN) control in harmonic
extraction. The algorithm is based on the concept of sum sine and cosine parts, attached each to an appropriate
coefficient which represents a signal periodically, as such estimates the polluted component. Load current
presented by fundamental and harmonic components as below (for each sample k in a digital operation) with △t as
the sampling time and w as the fundamental frequency.
With and as the amplitudes of the sine and cosine components of the measured nonlinear current, where n,
is the number of harmonics, up to N maximum number. In vectorial form,
(3)
With the weight matrix = and the sine and cosine vector
.
Vol 63, No. 11;Nov 2013
379 Jokull Journal
Fig. 2 conventional ANN extraction topology
Fig. 2 shows an ANN algorithm with a sole purpose of generating the right value of by training accurately .
The delta rule block is the brain of this extraction circuit (algorithm). The W-H minimizes the average square error
between the actual measured signal and the estimated signal written as
Where is the square of the vector and
The harmonic orders N, determines the weight matrix dimension presents a drawback of excess computation
and simulation time (to complete a single iteration of the computation). In the comparison aspect for the ANN, a
modified ANN algorithm is proposed with an advantage of mitigating the drawback presents by the above classical
algorithm.
Where, ,
, is the sampling time and alpha ( ) is the learning rate.
Fig. 3 A modified ANN extraction topology
Equation (5) above gives the algorithm (modified W-H algorithm) the liberty irrespective of N Present to alone
update the weights of the fundamental part. Mathematically, the elements are orthogonal for all, as such, the
modification is so much feasible and suitable, resulting in fast estimation and computation (enhancing the speed of
the iteration). A factor range 0 to 1 is incorporated. Fig. 3 above presents the configuration of the modified
proposed ANN algorithm which is far limited in complexity to its classical algorithm.
Vol 63, No. 11;Nov 2013
380 Jokull Journal
Challenge has been recorded in the implementation of this method as in a selection of a good alpha value ( ).
Previous works have been carried out which reports some suitable range of the factor ( ) [8] but, this paper skips the
learning rate value selection as appropriate and good simulation results with a preferred alpha value are recorded.
Recommendations for further study/work on a good range of alpha value would be beneficial. In the classical
version of the ANN algorithm (4), would follow harmonic order N if decided. However, in (5), only
contains the fundamental component as such, the alpha value is taken below 1/N which realizes an optimum
performance by the algorithm.
The nonlinear load currents fundamental sine component deducted from the nonlinear load current produces the
harmonic current .
This work further proposes a preferred value based on this two considerations: how quick the algorithm can
generates the current harmonics that will be treated as the reference signal by the pulse-width modulation (PWM)
in switching the gates of the VSI which produces the compensating current and how feasible the scheme can
generates and produce the harmonic current ) to minimizes total harmonic distortion (THD) . The complete
model of the simulation in MATLAB/SIMULINK environment using the modified ANN is shown in fig 4 below.
Fig. 4 Complete simulation model of the system
2.2 Fuzzy logic algorithm
2.2.1 Harmonic extraction technique
Basically, harmonic extractions can be achieved either in frequency domain or time domain approach. In this work,
a synchronous reference frame (SRF) a time domain was used for the extraction of the harmonic reference current to
the APF. The SRF is only applicable for three-phase system, which we modeled it as its three-phase under balanced
system condition and only one-phase was selected for the reference current harmonic extraction. In synchronous
reference frame, the distorted currents due to non-linear loads are transformed from a-b-c stationary reference frame
to its d-q reference frame. The SRF for the extraction of the harmonic signal is depicted in fig 5.
pwm
Discrete,Ts = 5e-005 s.
powergui
Vc
VS
v+-
VM3
v+-
VM2
v+-
VM1
VL
g
A
B
+
-
UB2
A
B
+
-
UB1
Subtract
Conn2Conn1
Subsystem1
In1
In2
Out1
Subsystem
RL R1
Uref Pulses
PWMG
0.003
LRL
ISIL
IH1
-1
Gain1
-1
Gaini
+-
CM2
i+
-
CM1
C2
C1
ACVS
Vol 63, No. 11;Nov 2013
381 Jokull Journal
Fig 5 SRF for extraction of harmonic
2.2.2 Fuzzy Logic Controller
Fuzzy logic controller (FLC) was first introduced by Professor Zadeh Lotfi of University California Berkeley in
early 1960s. He proposed a technique of how to process an imprecise data with complex input. The idea of Zadeh’s
was fully utilized after the introduction and availability of modern computers and controllers applications. Fuzzy
logic controllers gain interest by many researchers and system engineers in the application of control system analyst,
as well as in the control algorithm for shunt APF applications. FLC has advantages of simplicity in design
procedures that does not require any accurate mathematical modeling, can work with an imprecise input of the
system and it can also work with non-linearity. Fuzzy controller is very robust than classical controllers such as PI
and PID controllers [5]. In our study, a Mamdani fuzzy controller was chosen and designed with linguistic term “if
then” rules. The linguistic variables for the rule base was selected as, positive large (PL), positive medium (PM),
positive small (PS), negative large (NL), negative medium (NM), negative small (NS) and zero (Z). However, Fig. 2
depicts the general structure of FLC.
The design of the FLCis characterize as follows.
1. Five memberships function for each two inputs error (e) and its derivative (Δe) are used.
2. Seven memberships function for the output.
3. Mamdani implication in the design was also used.
4. Centre of area (COA) was used for the Deffuzzification process.
5. Implication using “min-max” approach
6. twenty five rules for both error and change of error are used.
Triangular membership functions are used due to its simplicity, completeness and easy to implement.
Fig. 6 Structure of FLC
Vol 63, No. 11;Nov 2013
382 Jokull Journal
As it was defined for the linguistic rule variables, table 1 below shows the “if then” rules for the five membership
functions selected for each of the input error (e) and the change of error (Δe). For the two inputs, only twenty five
possible rules are possible based on (5*5) =25 “if then” combinations rules.
Table 1 Linguistic variable rules for the FLC
Complete simulink model of the single-phase shunt APF is shown in fig 7. Parallel RC load with full bridge rectifier
are employed for the non-linear load which are the main source of harmonic producing load due to their non-
linearity in nature.
Fig. 7 Complete simulink model of single-phase shunt active power filter
3. SIMULATION RESULTS
3.1 ANN results and discussions
The performance and effectiveness of the modified ANN algorithm is discussed below, with the single phase shunt
APF system was tested and observed by simulation studies in MATLAB SIMULINK. The simulation’s sole aim
was to be accurately chosen at some ranges which will be implemented experimentally. The nonlinear load was
built by a series resistor (RS = 50Ω) and inductor (LS = 3mH) with a diode rectifier connected to a capacitor (CL =
470μF) and a resistive load (RL = 15Ω). The closed loop system has the following parameters (L = 1mH), r = 0.1Ω,
C = 1000μF, VS = 100V, 50HZ, Vdc = 150V, fswitching = 5 kHz. Discrete solver was used which performed the
simulation digitally at time sample Ts= 50μs.
Initially, the circuit was simulated without connecting the SAPF which records a THD of 16.59%. Chosen alpha
values 0.0009 to 0.001 at the SAPF presence is considered.THD measurements at distinct values of alpha are
reported as in table I above. The learning rate’s good chosen value was 0.004 (as shown in the table below) with a
THD measured within an acceptable range. Smaller value could reduce THD but, dispenses less fast increase in
for each k sample. These selection was made as to help the scheme produce a good and stable quicker and also
make it obeyed the IEEE-519-1992 THD% requirements.
Vol 63, No. 11;Nov 2013
383 Jokull Journal
TABLE I. THD SIMULATION RESULT FOR SPECIFIED ALPHA VALUES ( )
LEARNING RATE (α) THD MEASUREMENT (%)
0 (without SAPF) 33.59%
3.84%
3.87%
3.92%
3.83%
3.94%
4.06%
4.06%
4.06%0.0008
0.0009
0.001
0.002
0.003
0.004
0.005
0.006
The waveform of the current harmonics extracted by the ANN algorithm via a good chosen alpha value (0.004) is
presented below (fig. 8). Similarly, fig. 9 presents the simulated waveform for the supply voltage Vs (a), the supply
current IS (b) and the nonlinear load current IL (c) respectively all at 0.004. From the outputs (Table I and Fig. 5), the
THD was 3.92% with a stable , resulting in compensated to nearly in phase with VS and as such achieving a
nearly maximum PF.
Fig. 8 Harmonic extraction by ANN algorithm at 0.004 learning rate