Professional Documents
Culture Documents
i
lower and higher value of confidence limits of
PL
i
lower and higher value of confidence limits of real
line flows at all network lines
QL
i
lower and higher value of confidence limits of
reactive line flows at all network lines
Pslack
i
lower and higher value of confidence limits of real
power at reference bus
Qslack
i
lower and higher value of confidence limits of
reactive power at reference bus
B. Training of Artificial Neural Network
The selected artificial neural network used for training has
one input layer with 8 neurons, one hidden layer with 24
neurons and one output layer with 48 neurons.
Using the stochastic load flow algorithm, training patterns
were developed by varying the base case loading of 5 bus
system and then scaling the input and output quantities were
done as explained before.
The proposed method is implemented using Stuttgart
Neural Network Simulator (SNNS) [13]. For final training of
the artificial neural network, sum of square error (SSE) is
minimized.
In case of training with backpropagation algorithm, the
initial training characteristic of sum of square error (SSE) is
shown in Fig. 3. SSE reduced to about 0.017 in initial 100
iterations. After this convergence of training was slow and
value of SSE reduced to 0.00022 in 105000 iterations where
the network was found to be giving quite accurate results
when tested with novel patterns not included in the training
set.
0
1
2
3
4
5
6
7
8
9
0 500 1000 1500 2000 2500 3000 3500
No. of iterations
S
u
m
o
f
s
q
u
a
r
e
e
r
r
o
r
Fig. 3. Training of artificial neural network using backpropagation algorithm
for 5 bus system.
In training with the quickprop, the initial training
characteristic of sum of square error (SSE) is shown in Fig. 4.
SSE reduced to about 0.016 in initial 100 iterations. After this
convergence of training was slow and value of SSE reduced to
0.00100 in 7600 iterations. Convergence became slower and
SSE reduced to 0.00001 in 203000 iterations where network
was found to be giving quite accurate results when tested with
novel patterns not included in the training set.
0
10
20
30
40
50
60
70
80
90
100
110
0 500 1000 1500 2000 2500 3000 3500
No. of iterations
S
u
m
o
f
s
q
u
a
r
e
e
r
r
o
r
Fig. 4. Training of artificial neural network using quickprop algorithm for 5
bus system
V. RESULTS AND DISCUSSION
The trained artificial neural networks have been tested on
the novel patterns, which were not part of patterns use for
training. The artificial neural networks gave confidence limits
of output quantities that are almost same as found from the
classical stochastic load flow solutions when same input
quantities of these novel patterns, used for the trained artificial
neural network, are used as input data for the classical
stochastic load flow. To demonstrate the successful
implementation of the proposed approach, results for one such
novel pattern, which considered 105.5% loading of base case
and not used as training pattern, from trained artificial neural
networks for both cases, one trained using backpropagation
algorithm and another trained using quickprop algorithm, and
also from classical stochastic load flow approach are
presented in Table I - Table V.
TABLE I
COMPARISON OF CONFIDENCE LIMITS OF VOLTAGE MAGNITUDES FOR 5 BUS
SYSTEM FOR NOVEL PATTERN
Bus
no.
Confidence
Limits for
voltage
magnitude by
stochastic load
flow (p.u.)
Confidence
Limits for voltage
magnitude by
ANN using back-
propagation
(p.u.)
Confidence
Limits for voltage
magnitude by
ANN using
quickprop (p.u.)
1
2
3
4
1.0086 1.0196
1.0428 1.0487
1.0166 1.0250
1.0158 1.0244
1.0088 1.0197
1.0429 1.0488
1.0167 1.0252
1.0158 1.0245
1.0087 1.0197
1.0429 1.0487
1.0167 1.0251
1.0159 1.0245
A comparison of the output quantity results found from the
both artificial neural networks and classical stochastic load
flow approach shows that the confidence limits are same up to
the second digit of decimal for the phase angles that are given
in degree as their unit. For the confidence limits for bus
5
voltage magnitude, active and reactive power flows in lines
and active and reactive powers at slack bus, which are given
in p.u., the results from both artificial neural networks and
classical stochastic load flow approach are exactly same up to
third place of decimal and most of the result quantities are
similar up to the forth decimal place. This accuracy level of
matching of results in these confidence limits is more than
confidence limits of phase angles, which is quite justifiable
because these quantities are given in p.u. where as phase
angles are given in degrees.
TABLE II
COMPARISON OF CONFIDENCE LIMITS OF PHASE ANGLES FOR 5 BUS SYSTEM
FOR NOVEL PATTERN
Bus
no.
Confidence
Limits for phase
angle by
stochastic load
flow (deg.)
Confidence Limits
for phase angle by
ANN using back-
propagation (deg.)
Confidence Limits
for phase angle by
ANN using
quickprop (deg.)
1
2
3
4
-7.1098 -5.8880
-3.2140 -2.6973
-5.6969 -4.8524
-6.0732 -5.1798
-7.1080 -5.8859
-3.2129 -2.6960
-5.6948 -4.8510
-6.0714 -5.1765
-7.1082 -5.8860
-3.2130 -2.6963
-5.6951 -4.8510
-6.0714 -5.1767
TABLE III
COMPARISON OF CONFIDENCE LIMITS OF ACTIVE POWER FLOWS IN LINES FOR 5
BUS SYSTEM FOR NOVEL PATTERN
Line
no.
Confidence
Limits for active
power flow in
lines by
stochastic load
flow (p.u.)
Confidence
Limits for active
power flow in
lines by ANN
using back-
propagation
(p.u.)
Confidence
Limits for active
power flow in
lines by ANN
using quickprop
(p.u.)
1
2
3
4
5
6
7
0.8528 1.0264
0.3964 0.4644
0.2402 0.2812
0.2733 0.3167
0.5225 0.6356
0.1603 0.2379
0.0412 0.0924
0.8526 1.0261
0.3962 0.4644
0.2400 0.2812
0.2732 0.3166
0.5225 0.6354
0.1603 0.2379
0.0412 0.0924
0.8525 1.0260
0.3962 0.4643
0.2401 0.2812
0.2732 0.3165
0.5224 0.6354
0.1603 0.2379
0.0412 0.0923
TABLE IV
COMPARISON OF CONFIDENCE LIMITS OF REACTIVE POWER FLOWS IN LINES
FOR 5 BUS SYSTEM FOR NOVEL PATTERN
Line
no.
Confidence
Limits for
reactive power
flow in lines by
stochastic load
flow (p.u.)
Confidence
Limits for
reactive power
flow in lines by
ANN using back-
propagation
(p.u.)
Confidence
Limits for
reactive power
flow in lines by
ANN using
quickprop (p.u.)
1
2
3
4
5
6
7
-0.1051 -0.0373
0.0082 0.0324
0.0343 0.0473
0.0292 0.0409
0.0709 0.0953
-0.0606 -0.0435
-0.0272 -0.0173
-0.1050 -0.0373
0.0082 0.0323
0.0343 0.0473
0.0292 0.0409
0.0709 0.0952
-0.0606 -0.0435
-0.0272 -0.0173
-0.1051 -0.0373
0.0082 0.0324
0.0343 0.0473
0.0292 0.0409
0.0709 0.0952
-0.0606 -0.0435
-0.0272 -0.0173
TABLE V
COMPARISON OF CONFIDENCE LIMITS OF ACTIVE AND REACTIVE POWERS AT
SLACK BUS FOR 5 BUS SYSTEM FOR NOVEL PATTERN
Slack
bus
Confidence
Limits for
slack bus
power by
stochastic load
flow (p.u.)
Confidence
Limits for slack
bus power by
ANN using
back-
propagation
(p.u.)
Confidence
Limits for slack
bus power by
ANN using
quickprop (p.u.)
Active
power
1.2508 1.4894 1.2505 1.4891 1.2504 1.4891
Reactive
power
-0.0960 -0.0056 -0.0961 -0.0056 -0.0961 -0.0057
This comparison of results, which shows the similar
outputs from artificial neural networks and classical stochastic
load flow, validates the successful training of artificial neural
networks. It is clear from these results that the proposed
method using artificial neural network is quite suitable for
solving the stochastic load flow for power systems.
The proposed method can certainly be applied to larger
power systems. The inputs and output quantities for input and
output layers of artificial neural network are well defined in
present approach and these will be similar for any size of
power system, only their numbers and corresponding number
of neurons in the artificial neural network will increase. One
hidden layer is usually found appropriate for power
applications in literature and our study also found one hidden
layer sufficient for successful training of artificial neural
networks. Training patterns and testing patterns can also be
generated in similar way by choosing an appropriate range
around the base case stochastic load flow condition for that
system and generating the stochastic load flow results in that
range considering a small step size. As sigmoid activation
function for neurons has been successfully used in the present
study, same activation function will be quite appropriate for
the artificial neural networks for larger size power systems.
The use of sigmoid activation function needs outputs in the
range of 0 and 1 only, which requires scaling of output
quantities, so different scaling factors are needed for different
systems because scaling factor will depend on the range of
values of the output quantities of the training and testing
patterns. It is important to note that while defining scaling
factors large range of output values should be used so that it
can cover a large range of load level variation from base case
conditions.
We are attempting to apply the proposed method for larger
size power systems and will present the results in a future
transaction paper. The implementation of the proposed
artificial neural network based stochastic load flow method to
larger system will further establish the effectiveness of this
method to the real size power systems and will provide a very
useful technique to power engineers in effectively dealing
with demand side load uncertainties for power system
planning and operation.
6
VI. CONCLUSIONS
This paper has proposed a new method for stochastic load
flow analysis using artificial neural networks.
Backpropagation and quickprop algorithms have been applied
for the training of the artificial neural networks. It is clear
from the results that though quickprop algorithm is somewhat
faster in SSE reduction than backpropagation initially, but
later on both show similar characteristics in getting a low
value of SSE. Artificial neural networks trained by both
algorithms yield good results and directly give the confidence
limits of output quantities when tested on novel patterns not
included in the training set, which were validated by
comparing the results achieved from the classical stochastic
load flow for the same novel patterns. Therefore, both of these
artificial neural network training algorithms are equally good
for the stochastic load flow problem. Artificial neural
networks are extremely fast and predict the results
instantaneously. It is clear from this that the proposed method
using artificial neural network is very much suitable for
solving the stochastic load flow for power systems in
extremely fast way and provide a very useful technique for
power system planning and operation. The proposed method
can be applied for larger power systems. Future work will
present the application of the proposed method for the larger
power systems.
VII. REFERENCES
[1] B. Borkowska, "Probabilistic load flow," IEEE Trans. on Power
Apparatus and Systems, Vol. PAS-93, No. 3, pp. 752-759, May/June
1974.
[2] R. N. Allan, B. Borkowska, and C. H. Grigg, "Probabilistic analysis of
power flows," Proceedings of IEE, Vol. 121, No. 12, pp. 1551-1556,
Dec. 1974.
[3] J. F. Dopazo, O. A. Klitin, and A. M. Sasson, "Stochastic load flows,"
IEEE Trans. on Power Apparatus and Systems, Vol. PAS-94, pp. 299-
309, 1975.
[4] A. O. Ekwue, "Interactive stochastic security enhancement algorithm,"
International Journal of Electrical Power and Energy Systems, Vol. 6,
No. 1, pp. 51-54, January 1984
[5] A. O. Ekwue, and R. N. Adams, "optimal power rescheduling for system
security using stochastic second-order load flow," International Journal
Electrical Power and Energy Systems, Vol. 11, No. 4, pp. 277-282,
October 1989.
[6] M. Th. Schilling, A. M. Leite da Silva, R. Billinton, and M. A. El-Kady,
"Bibliography on power system probabilistic," IEEE Trans. on Power
Systems, Vol. 5, No. 1, pp. 1-11, February 1990.
[7] Henrique Steinherz Hippert, Carlos Eduardo Pedreira, and Reinaldo
Castro Souza, Neural Networks for Short-Term Load Forecasting: A
Review and Evaluation, IEEE Trans. On Power Systems, Vol. 16, No.
1, February 2001.
[8] Alexandre P. Alves da Silva, and Luciano S. Moulin, Confidence
Interval for Neural Network Based Short-Term Load Forecasting, IEEE
Trans. On Power Systems, Vol. 15, No. 4, November 2000.
[9] Li Zhang, Peter B. Luh, and Krishnan Kasiviswanathan, Energy
Clearing Price Prediction and Confidence Interval Estimation with
Cascaded Neural Networks, IEEE Trans. On Power Systems, Vol. 18,
No. 1, February 2003.
[10] T. T. Nguyen, "Neural network load flow," IEE Proceedings, Part C,
Vol. 142, No.1, pp. 51-58, January 1995.
[11] P. D. Wasserman, Neural computing theory and practice, New York:
Van Nostrand Reinhold, 1989.
[12] G. W. Stagg, and A. H. El-Abiad, Computer Methods in Power System
Analysis, McGraw-Hill, Inc., 1968.
[13] SNNS (Stuttgart Neural Network Simulator) Version 3.2 Manual,
developed at Institute for Parallel and Distributed High Performance
System (IPVR), University of Stuttgart, Postfach, Germany, 1993.
VIII. BIOGRAPHIES
Amit Jain (M2005) graduated from KNIT, India in
Electrical Engineering. He completed his masters
and Ph.D. from Indian Institute of Technology, New
Delhi, India.
He was working in Alstom on the power
SCADA systems. He has been in Korea in 2002
working as Post-doctoral researcher in the Brain
Korea 21 project team of Chungbuk National
University. He was Post Doctoral Fellow of the
Japan Society for the Promotion of Science (JSPS)
at Tohoku University, Sendai, Japan. Currently he is a Post Doctoral Research
Associate at Tohoku University, Sendai, Japan. His fields of research interest
are power system real time monitoring and control, artificial intelligence
applications, reliability analysis, parallel processing and nanotechnology.
S. C. Tripathy obtained his Ph.D. degree in
Electrical Engineering from University of
Minnesota, Minneapolis, USA in 1970. Then he
joined the faculty of the Indian Institute of
Technology, New Delhi, India. He was professor
there till he retired in 1998. He is currently a
professor at the Institute of Technology and
Management Gurgaon, India.
He had also been Head of the Center for Energy
Studies at Indian Institute of Technology, New
Delhi. He is Fellow of IEE (London) and IE (India). He has been visiting
professor to many reputed universities in Europe and Canada. His fields of
interest are Electric Power System Analysis and Control.
R. Balasubramanian obtained his Ph.D. degree from
IIT Kanpur. He is a professor at Indian Institute of
Technology, New Delhi, India. He is a senior member
of IEEE. He was chairman of IEEE Delhi section for
year 2001 and 2002.
His areas of research include Power System
Planning, Operation & Control, Energy System
Modeling & Management, and Power from Non-
conventional Energy Sources including Energy
Storage Devices. He has guided 10 doctoral scholars
and published about 80 research papers in the national and international
journals of repute.
Yoshiyuki Kawazoe was born in Japan on
December 16, 1947. He graduated from Tohoku
University in Japan in 1970. He earned his Ph.D.
from Tohoku University in 1975. Since 1975 he is a
faculty member in Tohoku University and a full
professor since 1990. He is the Director of the
Supercomputer center, Institute for Materials
Research, Tohoku University and also head of
Materials database committee, Japan Institute of
Metals. He also holds the Advisory Professor
positions at Fudan University, Shanghai, and Senan University, Chongching,
China.
He is a world renowned scientist in computational material science area.
He is editor-in-chief of many material science book series of Springer-Verlag,
Germany. He is also one of the leaders of the Japanese nanogrid project for
networking of supercomputer to realize an ultra-supercomputer for advance
nanotechnology research. He has more than 500 scientific paper publications.