45
CHAPTER-IV
NN-FL TECHNIQUE TO SELF TUNE THE
PARAMETERS
46
CHAPTER-IV
NN-FL TECHNIQUE TO SELF TUNE CONTROLLER
PARAMETERS
4.0. INTRODUCTION
HVDC finds a major application in long transmission system.
Current in the plant increases rapidly whenever an irregularity
occurs. Abrupt increase in current could both damage the system and
also affect its efficiency. To avoid this damage and to maintain the
current in normal values PI (Proportional – Integral) controllers are
used. Alteration of the controller constraints is the preliminary option
to retain systems consistency. PK and IK are the variable parameters.
An expert hybrid technique incorporating FL and NN is adapted to
control these parameters.
The deviation in current value and degree at which this error
changes are acquired from the system current values. By using these
error and rate values, fuzzy logic is used to calculate the fuzzy gain
and by giving fuzzy gain as input to the neural network, proportional
and integral gain are obtained as its output. For the operation of FL, a
fuzzy rule base is formulated. Then the fuzzy gain is given as input to
the neural network and corresponding proportional and integral gain
values are obtained as its output. By using this proportional and
integral gain the PI controller makes the system to remain stable. The
47
detailed process of self tuning the PI controller parameters using this
hybrid technique is described in the following section.
4.1. HVDC TEST SYSTEM
Figure 4.1 HVDC Test System Ref [27]
Figure 4.2 Details of ac System Representation on Either Side
Test system in Fig4.1 is adapted from Ref[27] with slight
modifications The sending end of the system consists of a constant
voltage and constant frequency source behind a reactance and ac
filters of fifth, seventh, eleventh and thirteenth harmonics are
included to produce a pure signal. SCR at the sending end is
maintained quite large. Then AC to DC converted power is exchanged
over the dc connector which consists of a large smoothing reactor and
a twelfth harmonic filter which reduces the ripple content. The
48
receiving end consists of a similar ac source as that at the input side
but with a very low SCR. Fault may occur at any point of the
transmission system during transfer of power from sending terminal
to the receiving station. Faults may affect terminal stations. Here, the
faults that occur both in the rectifier side as well as the inverter side
are considered.
4.2. FAULTS IN A HVDC SYSTEM
Faults arise as a result of unexpected variations in voltage and
current due to some interruptions along the dc line. Apart from
interrupting the system, these faults also influence functioning of the
system by and large. The following faults are more frequent with a
HVDC system.
(i).Single Line to Ground Fault
This is a recurrent fault in power transfer system. This fault is a
short circuit between line and ground.
(ii). Line to Line Fault
This fault arises between transmission lines. A short among any
two lines is treated as a fault.
4.3. ERROR AND RATE CALCULATION
On reducing the sending end network of the HVDC system in to a
thevenin‟s voltage behind a thevenin‟s impedance the current
supplied to the dc system under normal condition is equal to 1 KA. So
a reference current value ( rI ) of 1 KA is considered in the problem. If
the current is above 1 KA or below 1 KA then it can be inferred that a
49
fault has occurred in the system. Once fault initiates, the immediate
step is to make a note of the variation in current. The deviation of
fault current with respect to reference is termed as error. The error in
the value of Current and degree at which current varies are
calculated using these equations.
nr III ----------------------------(4.1)
T
II p
----------------------------(4.2)
).( IGE ----------------------------(4.3)
)( . 1
GR ----------------------------(4.4)
where, rI ,the reference current , nI - measured current,
pI - previous
value of error, T , sampling rate, G and 1G are the gains for
normalization, E is the error and R is the rate. If the system current
and the reference value are the same, then the E is zero and, E varies
as per the variation in current. After calculating the error and rate
values the next step is to apply these values to fuzzy logic and
generate fuzzy rules to obtain the fuzzy gain.
4.3.1. Generating Dataset to Train Fuzzy Logic system.
The fuzzy rules are generated using error in current and rate at
which it deviates from the reference. For deriving required set of data,
50
initially a current dataset maxminminmin ,,2,, IIIIIII qq is acquired
within the range [ maxI ,minI ]. Here, the Imax is 2.5 KA and Imin is 0 KA.
For generating training dataset different current values are taken
between 0 to 2.5 KA. After generating current dataset, the deviations
in current and rate of deviation are found out for each current value.
The error dataset is termed as nEEE ,,, 21 and the rate dataset is
termed as nRRR ,,, 21 . After calculating the error and rate values, by
applying fuzzy logic we get fuzzy gain values as the output. Then by
using the training dataset the fuzzy rules are devised.
4.3.2. Generation of Fuzzy Rules
After generating training dataset next step is to devise a fuzzy rule
base. Input variables are fuzzified into three sets namely, large,
medium and small and the output variables are fuzzified into five sets
namely, very large, large, medium, small and very small. By using
these sets fuzzy rule are engendered. The fuzzy rules developed are
put in Table-4.1.
In the formulated technique a triangular symmetrical MF is used.
Fuzzy rules are generated by considering both normal and abnormal
conditions. Under non contingencies no function is carried by FL.
Under abnormal conditions based on the error and rate values a
corresponding fuzzy gain is obtained as the output.
51
Table 4.1 FUZZY RULES
S.No Fuzzy rules
1
2
3 4
5
6 7
8
9
if, E=large and R=large, then G=very
large
if, E=large and R=medium, then G=large
if, E=large and R=small, then G=small
if, E=medium and R=large, then G=large
if, E=medium and R=medium, then
G=medium
if, E=medium and R=small, then G=large
if, E=small and R=large, then G=small
if, E=small and R=medium, then G=large
if, E=small and R=small, then G=very
small
Fuzzy Logic is trained by adapting the above devised rules.
During an abnormality the error may be either large or small. Based
on the error value, adjusting appropriate parameters by the PI
controller is essential to make the system remain stable.
4.3.3. Obtaining the Fuzzy Gain
As a first step, fuzzy logic is trained with the generated dataset, by
employing the devised fuzzy rules. After training, if any error and rate
52
value as given as input, the fuzzy gives the corresponding fuzzy gain
as output.
4.4. OBTAINING CONTROLLER PARAMETERS USING NN
The role of Neural Network is used to calculate the proportional and
integral gains. The input to NN is fuzzy gain and its outputs are
and . A Back propagation algorithm is applied to train the NN.
Training and testing procedures are elaborated in the next few
sections. For practical applications, neural network is trained once
and testing can be done any number of times.
4.5. Structure of Neural Network Used in NN-FL Technique
Back propagation algorithm is utilised to train the neural network.
In our method the role of this intelligent network is to calculate
and gains of the controller. A manipulation over these values
improves the system stability.
The network used in our method consists of one node in the input
layer, n nodes in the hidden layer and two nodes in the output layer.
The Fuzzy gain is transferred to the input layer and the outputs from
the Neural Network are the numerical values of proportional and the
integral gains. Configuration of Neural Network applied in the present
problem is shown below.
53
Figure 4.3: Single input - two output neural network employed for
obtaining Kp and KI
4.5.1. Training the Neural Network
For training, initially an input dataset is chosen and a common
weight is assigned for the hidden layers. A similarity check is made
between the obtained and target outputs. Then, by error applying
Back Propagation, weights from the output level to the hidden layers
are first modified. Then adjustment of weights from invisible level to
the input level is done to achieve the required output. This process of
comparison with the necessary output and modifications of the
weights are repeated until the network reaches the requisite output.
The Neural Network obtained by applying these initial weights is
shown in Figure 4.3.
w11n
y1
w2n2
w2n1
w212
w221
F2n
F21
F22
F31
F32
w211
w222
F1
y2
x
w111
w112
Input layer Hidden layer Output layer
w11n
y1
w2n2
w2n1
w212
w221
F2n
F21
F22
F31
F32
w211
w222
F1
y2
x
w111
w112
Input layer Hidden layer Output layer
54
Figure 4.4 Neural Network after the Application of Initial Weights
Various stages in training NN are:
Step 1: Initially assign weights to the neurons of the input layers.
Step 2: Apply the obtained training dataset to NN. Here x is the input
and 1y and 2y are outputs of NN.
)(1
1
121 ryWyn
r
r
------- (4.5)
)(2
1
222 ryWyn
r
r
------- (4.6)
)exp(1
1)(
11 xwry
r ------- ( 7.4 )
0.5
y1
0.9
1
0.8
0.6
F2n
F21
F22
F31
F32
0.2
0.1
F1
y2
x
0.3
0.7
Input layer Hidden layer Output layer
55
Equations 4.5 to 4.7 show the activation functions performed in the
input and output layers respectively.
Step 3: Determine the value of and using equations 4.5 and 4.6.
Step 5: Find out the error of the network after one pass and propagate
that error backwards.
Step 6: This process is carried out until the error reaches a minimum
value.
After training is completed, the network becomes fit for practical use.
The process involved in the above steps is drawn in the form of flow
chart as follows:
56
Figure 4.5 Flow Chart to train a neural network
57
4.6. FAULT CLEARANCE
During normal condition the system current equals the reference
value 1 KA. When a fault occurs in the system, the current increases
rapidly and reaches a value in between 2 and 2.5 KA and is calculated
using equation 4.3 and 4.4.To reduce this current PI controllers are
used.
4.6.1. PI Controller
Figure 4.6 PI Controller
The PI controller shown in Figure 4.6 substitutes the values
obtained from the proposed technique and outputs a current to the
HVDC test system of Figure 4.1. at every time instant considered.
Variations in the values of pK and iK results in a change of controller
output. Iout can be attained from eq‟n 4.8 as.
T
Ipout dtEKEKI0
(4.8)
Where , outI is the current,pK and IK ,the proportional and integral
gains and E is the error.
If the current calculated from equation 4.8 deviates from the
value under normal system conditions, the process is repeated and
another set of proportional and integral gains are found from the
_
+
+
+
Ir
In
E
EKpP
dtEKI I
∑ Process ∑
58
proposed technique obtained and substituted in eq‟n 4.8. This
procedure is iteratively done till current returns to normal value.
In the presented work, the faults mentioned in section [4.2] on
either ac sides are considered. In inverter side, the maximum fault
current that occurs is 2 KA and for section 4.2 maximum current that
occurs is 2.5 KA. In rectifier side, the same highest fault currents of
1.5 KA occur for both the type of faults mentioned in 4.2.
4.7. SIMULATION ANALYSIS
Up to the preceding sections implementation of the proposed
technique‟ s carried out. Now the controller with the proposed features
is ready for practical use. So, the developed model has been tested by
subjecting the system to the faults mentioned in section 4.1 and the
analysis of Voltage and current waveforms is presented. The following
cases have been studied.
4.7.1. DC Line to Line Fault at the Inverter
This kind of fault is the severest fault when the connected AC
system is weak .The nature of the fault is balanced but most critical
due to low SCR of the inverter bus as the total power injection
becomes zero. The dc power oscillations may give rise to uncontrolled
dv/dt and di/dt stresses on the converter thyristors. Also oscillations
in the inverter ac bus voltage may be detrimental to the loads at the
inverter end .As seen from the figures the conventional PI controller
makes the system oscillate even after the fault is recovered .These
59
oscillations were minimized with the adapted technique. A 5.2 cycle dc
line to line fault at the inverter is simulated at the inverter bus at 0.5
cycles. From the graph of the inverter dc voltage we can observe that
the number of commutation failures are reduced a lot compared to
that of the conventional and fuzzy logic controller. Figures 4.10(a, b,
&c), 4.11(a,b and c) and 4.12(a, b& c) represent the currents and
voltages for this case.
4.7.2. Single Line to Ground Fault at the Inverter
A single line to Ground fault has been simulated at the inverter
end ac system for about 2.5 cycles. The inverter end ac system being
weaker this kind of fault results in sudden voltage collapses on all the
phases leading to commutation failures and other difficulties in
converter operation. This also leads to unbalanced operation of the
converter even after the fault is cleared. The inverter dc voltage
showed the resultant commutation failures. The firing instants are
now uncertain and inverter extinction angle loses control over the
HVDC link current recovery. The converter current regulator mostly
influences the transient performance under these conditions.
Therefore the comparative study of various controllers shows a
substantial difference in their performance. The proposed technique
exhibited best performance in terms of transient recovery, as the
damping is much faster. Figures 4.7(a, b, &c), 4.8(a ,b and c) and
4.9(a, b& c) represent the currents and voltages for this case.
60
4.7.3. Single Line to Ground Fault at the Rectifier
A 2.5 cycle Single Line to Ground Fault is created at the rectifier ac
bus. Oscillations in current are reduced.AC side SCR is very high .So
the waveforms are least affected by the controller actions. There isn‟t
any reversal of power. Figures 4.13(a, b, &c), 4.14(a, b and c)
represent the currents and voltages for this case
4.7.4. Line to Line Fault at the Rectifier
Variation of the dc link current , rectifier side dc voltage are shown
in figures 4.15(a, b &c) and 4.16 (a, b &c) for a 2.5 cycle fault at the
rectifier bus after the inductor at o.5 sec. The dc bus Voltage
completely collapses and may lead to commutation failure. During
fault the dc link current drops to zero and this condition leads to
complete de-energisation of the dc link .and current wave forms are
not much affected by the controller. Response resulting from the
proposed technique result in fewer oscillations as compared to
conventional and fuzzy controllers.
61
Figure 4.7(a) Dc Link Current for Single Line to Ground Fault
at the Inverter for Conventional Controller
Figure 4.7(b) Dc Link Current for Single Line to Ground Fault at
the Inverter for Fuzzy Controller
62
Figure4.7 (c) Dc Link Current for Single Line to Ground
fault at the Inverter for FL-NN Technique.
Figure 4.8(a) Rectifier end Dc Voltage for Single Line to Ground
Fault at the Inverter for Conventional Controller
63
Figure 4.8(b) Rectifier end Dc Voltage for Single Line to
Ground Fault at the Inverter for Fuzzy Controller
Figure 4.8 (c) Rectifier end Dc Voltage for Single Line to
Ground Fault at the Inverter for FL-NN Technique .
64
Figure 4.9(a) Inverter end Dc Voltage for Single Line to Ground
Fault at the Inverter for Conventional Controller
Figure 4.9(b) Inverter end Dc Voltage for Single Line to Ground
Fault at the Inverter for Fuzzy Controller
65
Figure 4.9 ( c) Inverter end Dc Voltage for Single Line to Ground
Fault at the Inverter for FL-NN Technique
Figure 4.10(a) Dc Link Current for Line to Line Fault at the
Inverter for Conventional Controller
66
Figure 4.10 (b) Dc Link Current for Line to Line Fault at the
Inverter for Fuzzy Controller
Figure 4.10( c ) Dc Link Current for Line to Line Fault at the
Inverter for FL-NN Technique
67
Figure 4.11(a) Rectifier end Dc Voltage for Line to Line Fault
at the Inverter for Conventional Controller.
Figure 4.11(b) Rectifier end Dc Voltage for Line to Line Fault
at the Inverter for Fuzzy Controller.
68
Figure 4.11 (c) Rectifier end Dc Voltage for Line to Line Fault
at the Inverter for FL-NN Technique.
Figures 4.11 (a,b and c) represent the rectifier side dc voltage during
the fault clearing process when the test system is subjected to a line
to line fault at the inverter side.
69
Figure 4.12(a) Inverter end Dc Voltage for Line to Line Fault at
the Inverter for Conventional Controller.
Figure 4.12(b) Inverter end Dc Voltage for Line to Line Fault
at the Inverter for Fuzzy Controller.
70
Figure 4.12(c) Inverter end Dc Voltage for Line to Line Fault at
the Inverter for FL-NN Technique .
Figures 4.12 (a, b and c) shows the performance comparison between
(1) conventional, (2) the fuzzy-based and (3) the hybrid PI controller
self tuning technique in clearing line-to-line fault at inverter.
Figure 4.13(a) DC link Current for Single Line to Ground Fault
at the Rectifier for Conventional Controller.
71
Figure 4.13(b) DC link Current for Single Line to Ground Fault
At the Rectifier for Fuzzy Controller.
Figure 4.13(c DC link Current for Single Line to Ground Fault
at the Rectifier for FL-NN Technique .
72
Figure 4.14(a) Rectifier end Dc Voltage for Single Line to Ground
fault at the Rectifier for Conventional Controller
Figure 4.14(b) Rectifier end Dc Voltage for Single Line to
Ground fault at the Rectifier for fuzzy Controller.
73
Figure 4.14(c ) Rectifier end Dc Voltage for Single Line to
Ground fault at the Rectifier for FL-NN Technique
Figures 4.14(a,b and c)epresent rectifier side dc voltage when the test
system is subjected to a single line to ground fault at the rectifier end
for the tree controllers.
Figure 4.15(a) DC link Current for Line to Line fault at the
Rectifier for Conventional Controller
74
Figure 4.15 (b) DC link Current for Line to Line fault at the
Rectifier Controller
Figure 4.15 (c) DC link Current for Line to Line fault at the
Rectifier for FL-NN Technique
75
Figure 4.16(a) Rectifier end Dc Voltage for Line to Line fault
at the Rectifier for Conventional Controller
Figure 4.16(b) Rectifier end Dc Voltage for Line to Line fault at
the Rectifier for Fuzzy Controller
76
Figure 4.16 (c) Rectifier end Dc Voltage for Line to Line fault at
the Rectifier for FL-NN Technique.
Figures 4.16(a,b,c) is an assesment of (1) conventional, (2) the fuzzy-
based and (3) the hybrid PI controller self tuning technique in
overcoming dc line-to-line fault at rectifier.
77
Table 4.2 Comparison of Fault Clearance times for Conventional,
Fuzzy & Fuzzy-NN Controllers
4.8. Discussion and Conclusions
The discussed methodology was programmed in MATLAB 7.10 and
its operation, simulated. A comparison of the traditional self tuning
technique with that of a fuzzy based technique and a neural network-
fuzzy logic based technique is shown in the above graphs by making a
Type of the Fault
Parameters
Fault Clearance Time for
Adopted Controller in Sec
Conventional Fuzzy Fuzzy-NN
SLG Fault at the
Inverter
IDCr
VDCr
VDCi
o.4
0.54
0.54
0.35
0.44
0.42
0.09
0.1
0.11
LL Fault at Inverter IDCr
VDCr
VDCi
0.5
0.49
0.54
0.4
0.44
0.49
0.12
0.1
0.1
SLG Fault at the
Rectifer
IDCr
VDCr
0.49
0.49
0.44
0.44
0.06
0.11
LL Fault at Rectifer IDCr
VDCr
o.47
0.49
0.44
0.44
0.12
0.11
78
comparison of the currents and voltages at the inverter and rectifier
for the considered faults.
A NN-FL technique to self tune the variables of the PI Controller in
a HVDC system, is implemented. When a fault occurs in the system
the current and voltage increases and by using this hybrid technique,
the system voltage and current can be made to return to their stable
values within a fraction of a second. The functioning of the system can
be assessed from the obtained results. The implementation results
showed that the fault clearance time of the hybrid technique is very
low compared to conventional methods and fuzzy based self tuning
methods. Thus it has been proved that the implemented methodology
improved controlling of HVDC systems effectively than traditional self
tuning methods. The same conclusions are mathematically verified
from the fault clearance times tabulated in table 4.2 for the
considered faults at the inverter and rectifier sides for various
currents and voltages.