Professional Documents
Culture Documents
I. INTRODUCTION
2
associated limits at each bus.
II. THE PROPOSED ANN ARCHITECTURE AND FUNCTIONS
Fig. 1 shows the proposed feed forward neural network. The
architecture consists of an input layer, a hidden layer and an
output layer. Input vector is fed to the input layer for all the
buses of the selected system. The network weights and biases
are adjusted using the dotprod function in the toolbox and
adaptation is done using adaptwb function, which changes
both weights and biases. First to obtain the best combination of
number of neurons, training and transfer function a test system
is developed and the toolbox parameters are applied. The
following Toolbox functions are analyzed i)-Neural network
architecture and types, ii) Training functions, iii) Activation
functions, iv) Learning function, v) Initialization functions, and
vi) Performance functions
Lj = 1
i=g
ji
i =1
Vi
Vj
(1)
V L Z LL
G = GL
I K
where, F
Initially from the selected input data set, the minimum and
maximum values are found out. Then the FFNN architecture is
developed defining the number of neurons, training and
transfer functions. Further the input data is normalized and
filtered for redundancy. Then the network training parameters
such as learning ratio (0.08), learning increment (2.00),
number of epochs (400), Parameter Goal (1e-5), Minimum and
Maximum Gradient (1e-10 and 1e-10 respectively) have been
set based on the extensive analysis done. After training the
network for the set of samples, network performance is
evaluated with a new set of non-simulated data and the output
with the L-index values obtained from the conventional
algorithm is compared. The trained network nomenclature is
assigned and further a new set of NN is trained for
contingencies.
III. THE VOLTAGE STABILITY ASSESSMENT ALGORITHM
This paper uses the enhancement of the voltage stability of
LG
F LG
Y GG
IL
G
V
(3)
i=g
F
i =1
ji
Vi
ji + i j
Vj
(4)
For the stability, the index L j , must not be violated for any
of the nodes ' j ' . Hence the global indicator L describing the
stability of the complete subsystem is given by L = maximum
of L j for all j (load buses). Thus the indicator ' L ' is the
quantitative measure of the estimation of the distance of the
actual state of the system to the stability limit. The maximum
of the L-index gives the proximity of the system to voltage
collapse. An L-index away from 1 and close to 0 gives the
improved stability margin. The stability margin can also be
obtained as the distance of L from a unit value i.e. (1-L). Thus
the L-index gives a scalar number to each load bus and
performs consistent results. The stability margin and the
weakest bus can also be identified after obtaining the indices of
3
all the buses. Further the reactive power support measures and
the improved power flow can be performed in the system.
The algorithm developed is based on the operational load
flow, which will overcome the deficiencies of the conventional
load flow model. The model developed is suitable for
incorporating such various control strategies as Generator
prime mover response, Automatic Generation Control (AGC),
Generator tripping, Load shedding, and Load characteristics.
IV. THE ANN ALGORITHM AND FUNCTIONS
The proposed network architecture is a three layer feed
forward structure including input, output and hidden layer
using a back propagation algorithm. Following algorithmic
steps describes in detail the approach used for investigating the
different parameters and functions in the MATLAB toolbox.
Step 1: A conventional voltage stability algorithm is run with
the test system for simulated loading conditions. Using this
first the base case and the maximum loading conditions of the
test system are determined using the conventional software [5].
Then the load conditions are varied from base case till full load
and training samples are generated.
Step 2: Create a database for the input vector in the
following
form
where,
G1
119
115
T12
123
T13
54
52
55
1
T1
T11
111
G10
G2
G3
G4
50
2
T10
57
3
T2
139
T151
140
T4
T3
143
141
258
136
G6
G9
56
T14
11
T9
129
G5
7
T6
58
155
G7
G8
154
T5
149
157
147
T17
9
T7
10
T8
59
62
145
T16
61
63
267
Fig. 2. Functional Flowchart of the Voltage Stability Monitoring and Assessment
Tool
92
65
66
266
153
265
T18
166
5
TABLE 2
TOTAL GENERATION AND LOAD SCHEDULE FOR 367 NODE PRACTICAL SYSTEM
Description
MW
MVAR
1712.76
1621.06
--
646.31& -198.61
970.77
435
1712.23
1621.06
91.28
--
613.29
970.77
113.25
470.85
Sr.No
Voltage profile
(Equivalent Practical System)
MW
0.95
0.85
609.00&189.50
584.76
255
398.24
584.76
68.53
255.00
10
15
20
25
30
35
40
45
MVAR
Bus Numbers
L-Indices Values
(Equivalent Practical System)
0.4
Description
1.05
TABLE 3
TOTAL GENERATION AND LOAD SCHEDULE FOR REDUCED PRACTICAL
SYSTEM
Sr.No
1.15
0.3
0.2
0.1
0
0
10
15
20
25
30
35
40
45
Bus Numbers
V o lta g e P r o file fo r 3 6 7 N o d e p r a c tic a l S y s te m
L-Index values
(Lightly Loaded Conditions)
1 .0 5
0 .9 5
0 .9
358
337
316
295
274
253
232
211
190
169
148
127
85
106
64
43
22
0 .8 5
B u s N um bers
Case 1:
The developed equivalent 45-bus system from 367-bus
system represents ten generators, eighteen transformers and
forty-four transmission lines other than the loads. The loads
are combined into thirteen equivalent forms.
The details of the system grid total and the individual bus Lindex values used for training and testing are shown in the
table 3 and figures 6 respectively. The voltage profile shows a
variation in the voltage per-unit voltage magnitude which is as
high as 1.116 at bus number 28 and 0.864 at bus number 18.
Similarly the L-index values show the same effect as these
scalar values moves away from zero in the event of voltage
deviation. Thus these values are still very good indicators of
the bus voltage profile.
L-Index Values
Voltage in p.u
1 .1
0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1
10
15
20
25
30
35
40
45
Bus Numbers
Fig. 7. L-Index profile of the reduced practical system for a lightly loaded
condition
The load is varied for a light loaded condition and the input
vector for training the L-index is obtained for ten percent to
hundred and forty five percent of the base loading. Once the
proposed ANN architecture is well trained, (confirming it with
the post regression analysis) the intelligent stability indicator is
compared with the L-index value from the conventional
software. The results obtained for the non-simulated conditions
are as in the figure 7 and 8. Please note that the system loading
has been performed in a uniform manner getting the data set
for each five percent increase on all the bus. Then the load
changes are done in a random manner at each buses and the
data set for training has been extracted which has been referred
as abnormal way of loading. The L-index value and the error
for the load variation for an abnormal case is as in figure 9 and
10. Figures 8 and 10 shows the trained ANN structure is
capable of generating appropriate L index values during the
6
normal loading conditions as well as abnormal loading within
the maximum error of 0.05. This once again shows that once
trained effectively the proposed ANN structure can be used in
the practical system online irrespective of the system size.
very high. Around bus number 158 the voltages are low again.
Thus these are the areas, which are violating the voltage limits.
For the ANN training it is observed that for the reduced system
the L-index value is learned considerably accurate with a single
ANN.
Voltage profile
(Equivalent Practical System with
contingencies)
0.05
0
-0.05
0
10
15
20
25
30
35
40
45
Absolute Error
0.1
1.15
1.05
0.95
-0.1
0.85
0
Bus Numbers
10
15
20
25
30
35
40
45
Bus Numbers
Fig. 8. Absolute Error L-Index profile of the reduced practical system for a
lightly loaded condition (Trained ANN and Conventional software)
L-Index values
(Abnormal Conditions)
L-Index values
(Contingency Conditions)
0.2
L-Index Values
L-Index Values
0.25
0.2
0.15
0.1
0.05
0
0.15
0.1
0.05
0
-0.05
-0.05
10
15
20
25
30
35
40
10
45
15
20
25
30
35
40
45
Bus Numbers
Bus Numbers
Absolute Error
Absolute Error
0.02
0
-0.02
-0.04
-0.06 0
10
15
20
25
30
35
40
45
-0.08
Bus Numbers
10
15
20
25
30
35
40
45
Bus Numbers
Fig. 10. Absolute Error L-Index profile of the reduced practical system for an
abnormal loading condition (Trained ANN and Con. software)
Case 2:
In this case line 17-20 (equivalent to 59-63 in figure 3) is
deleted and simulated as line outage. Initially this line was
carrying 222.57 MW and 77 MVAR under base loading
condition. This was supplying to the load of 270 MW in bus
number 40 (157). The voltage profile is as in figure 11. The
voltage profile shows that the voltages at bus numbers from 58
to 66 is below one per unit and especially on bus number 61,
63 and 66 the voltages are very low. Like wise the voltage
magnitude from the bus numbers 115 to 136 the voltages are
Fig. 13. Absolute Error L-Index profile of the reduced practical system for a
contingency condition (Trained ANN and Conventional software)
This confirms that the ANN can be used for voltage stability
assessment quite efficiently when training is done properly.
When the load is light the trained value found to be having
more error. For moderate loading and for heavily loaded
condition the absolute error is less. The comparison of the Lindex values for a non-simulated condition as in figures 12 and
13 shows the effectiveness of the proposed ANN structure. As
the trained ANN does not need any further computation, the
proposed method is suitable especially on an energy
management center to perform online analysis. Further the
computation time for the offline training was around 120
seconds.
7
VIII. CONCLUSION
An artificial neural network technique for on line assessment
of power system voltage stability using a developed training
algorithm for all system buses has been presented with detail
steps involved with MATLAB neural network toolbox. Unlike
other reported techniques, the main advantage of the proposed
method is that the voltage stability indices for all the buses in
the power system can be calculated using the trained artificial
neural network at every monitoring period. The stability
margin and voltage profile for individual buses, the global
stability margin, as well as possible improvement measures of
the power system can be assessed at the same time during both
normal and contingent situations using the trained ANN.
Training and testing results form all cases, including
contingencies on a practical power systems network shows that
the proposed ANN algorithm is capable to learn and perform
as a tool for online voltage stability analysis by measuring the
L-indices for all the vulnerable buses.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]