Engineering Intelligent Systems, Vol.8, No. 1, pp.

53-58, March 2000
53
REAL POWER OPTIMIZATION WITH LOAD FLOW USING
ADAPTIVE HOPFIELD NEURAL NETWORK

Kwang Y. Lee, Fatih M. Nuroglu, and Arthit Sode-Yome
Department of Electrical Engineering
The Pennsylvania State University
University Park, PA 16802


Abstract: This paper presents real power optimization with load flow using an adaptive Hopfield neural network. In
order to speed up the convergence of the Hopfield neural network system, the two adaptive methods, slope
adjustment and bias adjustment, were used with adaptive learning rates. Algorithms of economic load dispatch for
piecewise quadratic cost functions using the Hopfield neural network have been developed for the two approaches
In stead of using the typical B-coefficient method, this paper uses actual load flow to compute the transmission
loss accurately. These methods for optimization has been tested in the IEEE 30-bus system to demonstrate its
effectiveness. The performance of the proposed approaches is evaluated by comparing the results of the slope
adjustment and the bias adjustment methods with those of the conventional Hopfield network, and an additional
improvement was demonstrated by the use of momentum in the adaptive learning approaches.

Keywords: Optimal power flow, economic load dispatch, Hopfield neural networks, adaptive Hopfield
neural networks


1. INTRODUCTION

In power system, the operation cost at each time needs
to be minimized via economic load dispatch (ELD) [1].
Since the introduction by Hopfield [2,3] The Hopfield
neural networks have been used in many different
applications. An important property of the Hopfield
neural network is the decrease in energy by finite
amount whenever there is any change in inputs [4].
Thus, the Hopfield neural network can be used for
optimization. Tank and Hopfield [5] described how
several optimization problem can be rapidly solved by
highly interconnected networks of a simple analog
processor, which is an implementation of the Hopfield
neural network. Park and others [6] presented the
economic load dispatch for piecewise quadratic cost
functions using the Hopfield neural network. The
results of this method compared very well with those of
the numerical method in an hierarchical approach [7].
King and others [8] applied the Hopfield neural network
in the economic and environmental dispatching of
electric power systems.
These applications, however, involved a large
number of iterations and often showed oscillations
during transients. . Recently, in order to speed up the
convergence of the Hopfield neural networks, Lee and
others [9] developed adaptive Hopfield neural network
methods, slope adjustment and bias adjustment
methods, with adaptive learning rates. In these studies,
however, the transmission loss was either calculated by
the traditional B-coefficient method or ignored all
together. This paper incorporates both the adaptive
Hopfield network for faster convergence and the load
flow for accurate calculation of the transmission loss.


2. ECONOMIC LOAD DISPATCH

The cost function of ELD problem is defined as
follows:


( )
C C P
i i
i
=

, (1)

C P
a +b P +c P , , P P P
a +b P +c P , , P P P
a +b P +c P , k, P P P
i i
i i i i i i i i
i i i i i i i i
ik ik i ik i k- i i
( )
fuel 1
fuel 2
fuel
1 1 1
2
1
2 2 2
2
1 2
2
1
·
≤ ≤
≤ ≤
≤ ≤
¹
'
¹
¹
¹
¹
¹
M
,(2)

where
C
i
(P
i
) : cost of the i
th
generator
P
i
: the power output of generator i
a
ik
, b
ik
, c
i k
: cost coefficients of the i
th
generator
for fuel type k.


In minimizing the total cost, the constraints of
power balance and power limits should be satisfied:

Engineering Intelligent Systems, Vol.8, No. 1, pp. 53-58, March 2000
54
a) Power balance
The total generating power has to be equal to the
sum of load demand and transmission-line loss:

+ D L P
i
i
− ·

0 , (3)
where D is total load and L is transmission loss
The transmission loss is obtained by load flow.
Traditionally, the transmission loss was calculated by
the B-coefficient method [1] and, for simplification, it
was often assumed zero.

b) Maximum and minimum limits of power
The generation power of each generator has some
limits and it can be expressed as
P P P
i i i
≤ ≤ , (4)
where P and P are respectively the minimum and
maximum power generation limits.


3. HOPFIELD NETWORKS FOR ELD

The energy function of the continues-time Hopfield
network [3] is defined as
E T VV I V V
ij i j
j i
i i i i
i i
= − − +
∑ ∑ ∑ ∑
1
2
θ , (5)
where V
i
, I
i
, and θ
i
are respectively the output, external
input and threshold bias of neuron i.

The dynamics of the neurons is defined by
dU
dt
T V I
i
ij j i
j
= +

, (6)
where U
i
is the total input to neuron i and the sigmoidal
function can be defined as
V g U g
U
U
1
U +
U
i i i i
i
i i
· · ·
+ −
¸
¸

_
,

( ) ( ) λ
θ
0
1
exp
0
. (7)
Stability is known to be guaranteed since the energy
function is bounded and its increment is found to be
nonpositive. The time evolution of the system is a
motion in state-space that seeks out minima in E and
comes to a stop at such points.
In order to solve the ELD problem, the following
energy function is defined by augmenting the objective
function (1) with the constraint (2):

E = A D+ L P B a +b P+c P
i
i
i i i i i
i
1
2
( )
1
2
( )
2 2
− +
∑ ∑
,(8)

where a
ik
, b
ik
, c
ik
are the cost coefficients as discrete
functions of P
i
defined in (1).

By comparing (8) with (5) whose threshold is
assumed to be zero, the weight parameters and external
input of neuron i in the network [6] are given by
T A Bc
ii i
· − − ,
T A
ij
· − , (9)
I A D L
Bb
i
i
· + − ( )
2
.
From (6), the differential synchronous transition mode
used in computation of the Hopfield neural network is

U k U k TV k + I
i i ij j i
j
( ) ( - 1) ( ) − ·

. (10)

The sigmoidal function (7) can be modified [6] to
meet the power limit constraint as follows:

V k + = P - P
1
1 +
U k
U
P
i
i
i
i i
i
( 1) ( )
exp
( )
0

+
¸

1
]
1
+
θ
(11)


4. ADAPTIVE HOPFIELD NETWORKS

The traditional approach in solving the economic load
dispatch (ELD) problem using the Hopfield neural
network requires a large number of iterations and often
oscillates during the transient [6] and [9]. In order
to speed up convergence, two adaptive adjustment
methods are developed in this paper: slope adjustment
and bias adjustment methods [9].


- 600 - 400 - 200 0 200 400 600
100
150
200
250
Ui
Pi
U0=50100 150 200


Figure 1. Sigmoidal threshold function with different values of the
gain parameter.


4.1 Slope Adjustment Method

In transient state, the neuron input oscillates around the
threshold value, zero in Fig. 1. Some neuron inputs
Engineering Intelligent Systems, Vol.8, No. 1, pp. 53-58, March 2000
55
oscillate away from the threshold value. If the gain
parameter is set too high, the oscillation will occur at
the saturation region. If the slope in this region is too
low, the neurons can not go to the stable state and will
cause instability.

Since energy is to be minimized and its convergence
depends on the gain parameter U
0
, the gradient-descent
method can be applied to adjust the gain parameter as

U k U k
E
U
s 0 0
0
1 ( ) ( ) + · − η


, (12)
where η
s
is a learning rate.

From (8) and (11), the gradient of energy with respect
to the gain parameter can be computed as







E
U
E
P
P
U
i
i
i
n
0 0 1
·
·

. (13)
The update rule of (12) needs a suitable choice of
the learning rate η
s
. For speed and convergence, a
method to compute adaptive learning rates is developed
following the procedure in Ku and Lee [10]. It can be
shown [9,10] that the optimal learning rate is

η
s
s
g

·
1
2
, max
, (14)
where g g k g k E k U
s s s ,max
: max ( ) , ( ) ( ) / · · ∂ ∂
0
.


4.2 Bias Adjustment Method

There is a limitation in the slope adjustment method in
that slopes are small near the saturation region of the
sigmoidal function, Fig 1. If every input can use the
same maximum possible slope, convergence will be
much faster. This can be achieved by changing the bias
to shift the input near the center of the sigmoidal
function. The bias can be changed following the similar
gradient-descent method used in the slope adjustment
method:

θ θ η

∂θ
i i b
i
k k
E
( ) ( ) + · − 1 , (15)
where η
b
is a learning rate.
The bias can be applied to every neuron as in (7),
therefore, from (8) and (11), a derivative of energy with
respect to a bias can be individually computed as



∂θ



∂θ
E E
P
P
i i
i
i
· . (16)

The adaptive learning rate is also developed following
the similar procedure [10]. It can be shown that the
optimal learning rate is
η
b
b
g k

· −
1
( )
, (17)
where

g k T
V
V
b ij
i
j
( ) ·
∑ ∑

∂ θ

∂ θ
. (18)
Again, any other learning rate larger than η
b

does
not guarantee a faster convergence.


4.3 Momentum

The speed of convergence can be accelerated by adding
momentum in the update processes. The momentum
can be applied when updating the input in (10), the gain
parameter in (12), and the bias in (15):

U k U k TV k U k
i i ij
j
j u i
( ) ( ) ( ) ( ) − − · + −

1 1 α ∆ , (19)
U k U k
E
U
U k
s s 0 0
0
0
1 1 ( ) ( ) ( ) · − − + − η


α ∆ , (20)
θ θ η

∂θ
α θ ( ) ( ) ( ) k k
E
k
b b
· − − + − 1 1 ∆ , (21)
where α’s are momentum factors.


5. SIMULATION RESULTS

The results of slope adjustment and bias adjustment
methods are compared at different load demands. Then
the results of the slope and bias adjustment methods
with adaptive learning rates and momentum applied are
compared with each other. Three load levels are tested:
220 MW, 283.4 MW and 380 MW. The 220 MW and
380 MW load levels were used to force one or more
units to hit lower and upper limits. The IEEE 30-bus, 6
generator system is used for loadflow. The line data is
given in [13]. All the graphs are based on the nominal
load, 283.4 MW, to demonstrate the convergence. A
Compaq 90 MHz Pentium computer was used for
simulation.
For real power optimization with loadflow, first the
load flow is run to obtain initial generator power and
the system loss. Second, the adaptive Hopfield
networks update the slope and bias parameters and
generator power. Then, with the new generator power,
Engineering Intelligent Systems, Vol.8, No. 1, pp. 53-58, March 2000
56
load flow is run to obtain the new system loss and
generator power. Then, if the system converges, the
program will stop, otherwise the process will be
repeated. This procedure is shown in Fig. 2.


TABLE I
COST COEFFICIENTS FOR PIECEWISE QUADRATIC COST
FUNCTIONS

U Min P1 P2 Max F a b c
F1 F2 F3
1 50 100 190 200 1 0.000 1.900 .00355
1 2 3 2 0.000 2.000 .00376
3 0.000 2.200 .00415
2 20 35 50 80 1 0.000 1.700 .01700
1 2 3 2 0.000 1.750 .01750
3 0.000 2.050 .02350
3 10 25 35 1 0.000 1.000 .06250
1 2 2 0.000 1.200 .0825
4 10 20 30 1 0.000 3.250 .00834
1 2 2 0.000 3.650 .01234
5 15 30 50 1 0.000 3.000 .02500
1 2 2 0.000 3.300 .03500
6 12 25 40 1 0.000 3.000 .02500
1 2 2 0.000 3.300 .03500
GENERATIONS COST COEFFICIENTS


Slope adjustment method with load flow was run for
fixed and adaptive learning rates at three load demands.
The results are shown in Table II for comparison. For
283.4 MW, the generation cost was plotted for fixed
learning rate in Fig 3. In Fig 3, the effectiveness of the
adaptive Hopfield neural networks is clearly shown.



Initial Loadflow
Hopfield N. N.
Loadflow
Converged?
Optimal Cost
Stop
Yes
No


Figure 2 : Flowchart for the optimization with using load flow.


Bias adjustment method with loadflow was run for
fixed and adaptive learning rates at three load demands.
The results are shown in Table II for comparison. For
283.4 MW, the generation cost was plotted for fixed
,in Fig 3., and adaptive, in Fig 4, learning rates.

0 1 2 3 4 5
x 1 0
4
7 5 0
8 0 0
8 5 0
9 0 0
9 5 0
1 0 0 0
I t e r a t i o n s
C
o
s
t
S l o p e
b i a s
c o n v e n t i o n a l



Fig.3: Cost of the using Hopfield network (dashed-dotted line), slope
adjustment method with fixed learning rate (solid line) and bias
adjustment method with fixed learning rate (dashed line).



TABLE II
RESULTS FOR CONVENTIONAL HOPFIELD NETWORK (A), THE
SLOPE ADJUSTMENT METHOD WITH FIXED LEARNING RATE
η=1.0 (B) AND THE BIAS ADJUSTMENT METHOD WITH FIXED
LEARNING RATE η=1.0 (C)

U A B C A B C A B C
1 134.80 135.11 133.93 180.05 179.12 180.86 200 200 200
2 37.602 37.299 37.352 49.246 49.05 49.19 78.821 78.937 78.935
3 16.495 16.340 16.492 19.853 19.75 19.83 25.837 25.802 25.822
4 10.704 10.774 11.009 15.961 16.31 15.12 29.210 29.886 29.890
5 15.369 15.451 16.027 16.916 17.51 17.18 31.826 31.976 31.758
6 12.083 12.093 12.165 12.953 13.13 12.86 29.995 29.784 29.993
Tot. P. 227.05 227.06 226.97 294.99 294.88 295.45 396.39 396.385 396.39
Cost 589.63 589.73 590.06 810.06 810.08 810.24 1391 1390.89 1390.9
Lost 7.053 7.062 6.974 11.582 11.62 11.645 16.389 16.385 16.398
Iter. 48000 31000 25000 50000 30000 30000 49000 31700 31500
U0 100 100 100 100 100 100 100 100 100
Theta 0 0 50 0 0 50 0 0 50
Time NA 3m 45s 3m 4s NA 3m 37s 3m 37s NA 3m 51s 3m 53s
η
0 1 1 0 1.0 1.0 0 1 1
220 M 283.4 MW 380 MW



0 1 2 3 4 5
x 1 0
4
7 5 0
8 0 0
8 5 0
9 0 0
9 5 0
1 0 0 0
I t er at i ons
C
o
s
t
b i a s
s l o p e


Figure 4: Cost of the using slope adjustment method with load flow for
adaptive learning rate (solid line) and slope adjustment method with
load flow for adaptive learning rate (dashed-dotted line).
Engineering Intelligent Systems, Vol.8, No. 1, pp. 53-58, March 2000
57

Momentum is applied to the input of each method
and gain parameter for slope adjustment method to
speed up the convergence of the system. When the
momentum is applied to the system, the number of
iterations is significantly decreased. In Fig. 5 , the
momentum factor of 0.9 and 0.97 is applied to input
and gain parameter of slope adjustment methods, and
the number of iterations is reduced to about 15 percent
of that without momentum of slope adjustment method.
The system bus information (after convergence) is
shown in Table IV.

As seen in Table III, for 220 MW load, Units 4,5,and
6 have 10.606, 15.357, and 12.074 MW, respectively;
which are all very close to lower limits. Also for 380
MW load, Units 1,2, and 4 have 199.999, 79.024, and
29.809 MW, respectively, which are close to 200, 80,
and 30 MW upper limits , respectively.



TABLE III
RESULTS FOR ADAPTIVE SLOPE ADJUSTMENT METHOD WITH
MOMENTUM 0.9 APPLIED TO INPUT AND GAIN PARAMETER
WITH LOAD FLOW AND ADAPTIVE BIAS ADJUSTMENT
METHOD WITH MOMENTUM 0.9 APPLIED TO INPUT WITH LOAD
FLOW.


U Slope Bias Slope Bias Slope Bias
1 135.466 133.977 182.984 180.173 199.999 199.986
2 37.178 37.474 48.691 49.188 79.024 79.448
3 16.405 16.518 19.61 19.897 25.76 25.687
4 10.606 10.971 15.385 15.334 29.809 29.435
5 15.357 15.891 15.951 17.443 32.16 32.507
6 12.074 12.15 12.629 12.939 29.628 29.322
Total Power 227.086 226.981 295.252 294.974 396.324 389.385
Cost 589.647 589.965 810.278 810.203 1390.997 391.16
Lost 7.0861 9.981 11.852 11.574 16.324 9.385
Iterations 3200 2700 4000 3000 3000 2500
U0 100 100 100 100 100 100
Theta 0 50 0 50 0 50
Time 25s 21s 30s 23s 25s 20s
n 1.00E-04 1 1.00E-04 1 1.00E-04 1
220 MW 283.4 MW 380 MW



In Fig. 6, momentum 0.9 is applied to input of bias
adjustment method with adaptive learning rate using
load flow. Reduction of the iteration number for the
bias adjustment method is similar to the case for slope
adjustment method.

0 0 . 5 1 1 . 5 2 2 . 5 3
x 1 0
4
7 5 0
8 0 0
8 5 0
9 0 0
9 5 0
1 0 0 0
I t e r a t i o n s
C
o
s
t
W i t h M o m e n t u m
W i t h o u t M o m e n t u m



Figure 5: Cost of the using slope adjustment method with load flow for
adaptive learning rate without momentum (solid line) and with
momentum 0.9 applied at input and momentum 0.97 at gain parameter
(dashed-dotted line).



0 0 . 5 1 1 . 5 2
x 1 0
4
7 5 0
8 0 0
8 5 0
9 0 0
9 5 0
1 0 0 0
I t e r a t i o n s
C
o
s
t
W i t h M o m e n t u m
W i t h o u t M o m e n t u m


Figure 6: Cost of the using bias adjustment method with load flow for
adaptive learning rate without momentum (solid line) and with
momentum 0.9 applied at input (dashed-dotted line).



TABLE IV
BUS INFORMATION OF 283.4 MW LOAD WITH ADAPTIVE SLOPE
ADJUSTMENT METHOD WITH MOMENTUM


Bus # V (p.u) P (MW) Q(MVAr) Bus # V (p.u) P (MW) Q(MVAr)
1 1 182.984 -41.067 16 0.934 -3.495 -2.414
2 1 27 24.362 17 0.919 -8.999 -5.899
3 1 -10.381 28.257 18 0.912 -3.199 -0.965
4 1 15.387 28.366 19 0.906 -9.499 -3.482
5 1 -78.241 39.272 20 0.909 -2.198 -0.59
6 1 12.63 30.189 21 0.907 -17.496 -11.741
7 0.984 -22.795 -12.687 22 0.908 0.007 0.705
8 0.989 -2.399 -0.825 23 0.908 -3.197 -1.798
9 0.942 0 2.139 24 0.892 -8.693 -7.841
10 0.922 -5.792 1.698 25 0.893 0.01 0.485
11 0.987 -7.603 -1.413 26 0.87 -3.494 -2.842
12 0.958 -11.176 -2.238 27 0.904 0.004 2.26
13 0.99 0.007 -0.586 28 0.984 0.005 -2.409
14 0.938 -6.193 -0.331 29 0.882 -2.399 -0.952
15 0.927 -8.174 -5.318 30 0.87 -10.599 -1.296

Engineering Intelligent Systems, Vol.8, No. 1, pp. 53-58, March 2000
58

For both figures (Fig 5. and Fig. 6), 283.4 MW load
was used. The results of the cases for both methods are
shown in Table 4.


6. CONCLUSIONS

In this paper, the slope and bias adjustment are applied
to the real power optimization with loadflow. For all
three load levels, slope and bias adjustment methods
gave very good responses. Especially with momentum
applied, the iteration numbers are very low and the
computation time is around 20-25 seconds. Around
lower and upper limits of generator, the system
converged very smoothly.


ACKNOWLEDGEMENTS

This work was supported in part by the National Science
Foundation under grants INT-9605028 and ECS-
9705105.


REFERENCES

1. A. J. Wood and B. F. Wollenberg, Power
Generation, Operation, and Control, Second
Edition, John Wiley & Sons, New York, NY (1996)
2. J. J. Hopfield, ‘Neural networks and physical
systems with emergent collective computational
abilities’, Proceedings of National Academy of
Science, USA., Vol. 79, pp. 2554-2558, April
1982.
3. J. J. Hopfield, ‘Neurons with graded response have
collective computational properties like those of
two-state neurons’, Proceedings of National
Academy of Science, USA., Vol. 81, pp. 3088-
3092, May 1984.
4. J. A. Freeman and D. M. Skapura, Neural Networks
Algorithms, Applications, and Programming
Techniques, Addison-Wesley Pub. Co., Inc., July
1992.
5. W. D. Tank and J. J. Hopfield, ‘Simple neural
optimization networks: an A/D converter, signal
decision circuit, and a linear programming circuit’,
IEEE Transactions on Systems, Mans, and
Cybernetics, Vol. CAS-33, No. 5, pp. 553-541,
January/February 1992.
6. H. Park, Y. S. Kim, I. K. Eom, and K. Y. Lee,
‘Economic load dispatch for piecewise quadratic
cost function using hopfield neural networks’,
IEEE Transactions on Power system Apparatus
and Systems, Vol. 8, No. 3, pp. 1030-1038, August
1993.
7. C. E. Lin, and G. L. Viviani, ‘Hierachical economic
dispatch for piecewise quadratic cost functions’,
IEEE Transactions on Power Apparatus and
Systems., Vol. PAS-10, No. 6., June 1984.
8. T. D. King, M. E. El-Hawary, and F. El-Hawary,
“Optimal environmental dispatching of electric
power system via an improved hopfield neural
network model”, IEEE Transactions on Power
Systems, Vol. 1, No. 3, pp. 1559-1565, August
1995.
9. K. Y. Lee, A. Sode-Yome, and J. H. Park, ‘Adaptive
Hopfield Neural Networks for Economic Load
Dispatch’, IEEE Trans. on Power Systems, Vol 13
No 2 (May 1998) pp 519-526
10. C. C. Ku and K. Y. Lee, ‘Diagonal recurrent neural
networks for dynamic systems control,” IEEE,
Transactions on Neural Networks’, Vol. 6, No. 1,
pp. 144-156, January 1995.
11. M. A. El-Sharkawi and D. Niebur, Application of
Artificial Neural Networks to Power Systems,
IEEE Power System Society, 96 TP 112-0, IEEE,
1996.
12. J. J. Hopfield and W. D. Tank, ‘Neural computation
of decisions in optimization problems’, Biological
Cybernetics, Vol. 52, pp 141-152, 1985.
13. K. Y. Lee, Y. M. Park and J. L. Ortiz, ‘United
Approach to Optimal Real and Reactive Power
Dispatch’, IEEE Trans. on Power Appratus &
Systems, PAS-104 1985 pp 1147-1153

In order to speed up convergence. Traditionally. bik. zero in Fig. The dynamics of the neurons is defined by dU i dt = ∑ j T ij V j + Ii . the transmission loss was calculated by the B-coefficient method [1] and. external input and threshold bias of neuron i. Ii. Some neuron inputs . the neuron input oscillates around the threshold value. and θi are respectively the output. 1. it was often assumed zero. Sigmoidal threshold function with different values of the gain parameter.1) = ∑ TijV j ( k ) + I i . for simplification. ADAPTIVE HOPFIELD NETWORKS (5) where Vi. (4) (9) Bbi . j (10) where P and P are respectively the minimum and maximum power generation limits. 2 Ii = A( D + L ) − From (6).P i ) 1 + Pi (11)  Ui(k) + θi  1 + exp −  U0   3. the differential synchronous transition mode used in computation of the Hopfield neural network is U i ( k ) − U i ( k . (3) where D is total load and L is transmission loss The transmission loss is obtained by load flow. b) Maximum and minimum limits of power The generation power of each generator has some limits and it can be expressed as Pi ≤ Pi ≤ Pi . Vol. By comparing (8) with (5) whose threshold is assumed to be zero. 53-58.1 Slope Adjustment Method In transient state.(8) 2 i 2 i 100 00 -6 -400 -200 0 Ui 200 400 600 Figure 1. cik are the cost coefficients as discrete functions of Pi defined in (1). 4.8. i i 2 i j 4. i where aik. In order to solve the ELD problem. two adaptive adjustment methods are developed in this paper: slope adjustment and bias adjustment methods [9]. Tij = − A . 1. The time evolution of the system is a motion in state-space that seeks out minima in E and comes to a stop at such points. The sigmoidal function (7) can be modified [6] to meet the power limit constraint as follows: Vi ( k + 1) = ( P i .Engineering Intelligent Systems. HOPFIELD NETWORKS FOR ELD The energy function of the continues-time Hopfield network [3] is defined as E=− 1 ∑ ∑ Tij ViV j − ∑ I i Vi + ∑ θ i Vi . pp. March 2000 54 a) Power balance The total generating power has to be equal to the sum of load demand and transmission-line loss: D + L − ∑ Pi = 0 . No. (7) Pi 150 Stability is known to be guaranteed since the energy function is bounded and its increment is found to be nonpositive. 250 U0=50100 150 200 200 where Ui is the total input to neuron i and the sigmoidal function can be defined as U Vi = g i ( λ U i ) = g i ( i ) = U0 1  U +θi 1 + exp  − i  U0   . the following energy function is defined by augmenting the objective function (1) with the constraint (2): E= 1 1 A( D + L − ∑ Pi )2 + B ∑ (ai + bi Pi + ci Pi 2 ) . (6) The traditional approach in solving the economic load dispatch (ELD) problem using the Hopfield neural network requires a large number of iterations and often oscillates during the transient [6] and [9]. the weight parameters and external input of neuron i in the network [6] are given by Tii = − A − Bci .

Fig 1. If every input can use the same maximum possible slope. All the graphs are based on the nominal load. It can be shown that the optimal learning rate is η∗ = − b where gb ( k ) = ∑ ∑ Tij ∂ Vi ∂ V j . therefore. Second. ∂θ i ∂ Pi ∂θ i (16) The results of slope adjustment and bias adjustment methods are compared at different load demands. the oscillation will occur at the saturation region. March 2000 55 oscillate away from the threshold value. max The speed of convergence can be accelerated by adding momentum in the update processes. SIMULATION RESULTS where ηb is a learning rate. 283. a method to compute adaptive learning rates is developed following the procedure in Ku and Lee [10]. and the bias in (15): Ui (k ) − U i ( k − 1) = ∑ TijV j ( k ) + α u ∆U i ( k − 1) . ∂θ ∂θ (18) 1 . the gradient-descent method can be applied to adjust the gain parameter as U 0 ( k + 1) = U 0 ( k ) − η s ∂E . g b (k ) (17) (12) where ηs is a learning rate. with the new generator power. (14) U 0 (k ) = U 0 ( k − 1) − η s θ ( k ) = θ ( k − 1) − η b ∂E + α s ∆U 0 ( k − 1) . The line data is given in [13]. Then the results of the slope and bias adjustment methods with adaptive learning rates and momentum applied are compared with each other. If the slope in this region is too low. The IEEE 30-bus. ∂θ i (15) 5. This can be achieved by changing the bias to shift the input near the center of the sigmoidal function. For speed and convergence. Since energy is to be minimized and its convergence depends on the gain parameter U0. ∂θ 4. There is a limitation in the slope adjustment method in that slopes are small near the saturation region of the sigmoidal function. the adaptive Hopfield networks update the slope and bias parameters and generator power. the gain parameter in (12).2 Bias Adjustment Method where α’s are momentum factors. g s (k ) = ∂ E (k ) / ∂ U 0 .Engineering Intelligent Systems. i=1 i ∂U 0 n (13) 4. Three load levels are tested: 220 MW. ∂ U0 The adaptive learning rate is also developed following the similar procedure [10].4 MW and 380 MW. Then.3 Momentum The update rule of (12) needs a suitable choice of the learning rate ηs. For real power optimization with loadflow. any other learning rate larger than η ∗ does b not guarantee a faster convergence. ∂E + α b ∆θ ( k − 1) .4 MW. 6 generator system is used for loadflow. from (8) and (11). The 220 MW and 380 MW load levels were used to force one or more units to hit lower and upper limits. a derivative of energy with respect to a bias can be individually computed as ∂ E ∂ E ∂ Pi = . 283. The momentum can be applied when updating the input in (10). 53-58. The bias can be applied to every neuron as in (7). the gradient of energy with respect to the gain parameter can be computed as ∂E = ∂U 0 Again. ∂U0 (20) (21) where g s. No. (19) j . ∑∂P ∂ E ∂ Pi . A Compaq 90 MHz Pentium computer was used for simulation. 1. . If the gain parameter is set too high. first the load flow is run to obtain initial generator power and the system loss. pp. The bias can be changed following the similar gradient-descent method used in the slope adjustment method: θ i (k + 1) = θ i ( k ) − η b ∂E . It can be shown [9.max : = max g s ( k ) . convergence will be much faster. From (8) and (11). Vol. to demonstrate the convergence.10] that the optimal learning rate is η∗ s = 1 2 g s.8. the neurons can not go to the stable state and will cause instability.

0 (B) AND THE BIAS ADJUSTMENT METHOD WITH FIXED LEARNING RATE η=1. THE SLOPE ADJUSTMENT METHOD WITH FIXED LEARNING RATE η=1.009 16.995 29.19 19.299 16.73 7.000 0. U0 Theta Time η A 134.8.3: Cost of the using Hopfield network (dashed-dotted line).12 16.582 11.645 50000 30000 30000 100 100 100 0 0 50 NA 3m 37s 3m 37s 0 1.05 49.000 0.12 180.200 3.750 2.886 29.00355 .39 1391 1390.000 0.602 16.13 12.774 15. in Fig 4.700 1.4 MW.39 396.02500 .993 396.953 13.01700 .000 0.05 589.86 49.9 16.062 31000 100 0 3m 45s 1 283.352 16.837 25. This procedure is shown in Fig.784 29.45 810. pp.51 17. 53-58. and adaptive.027 12.000 3.961 16.935 25. Vol.822 29.853 19.250 3. learning rates.00415 .93 37.31 15.492 11. the generation cost was plotted for fixed learning rate in Fig 3.000 0.389 16. No.000 1.0 (C) 220 M B 135.000 0. the effectiveness of the adaptive Hopfield neural networks is clearly shown. Then.000 0.97 590.495 10.451 12. For 283. N.900 2.05 179.053 48000 100 0 NA 0 C 133. TABLE I COST COEFFICIENTS FOR PIECEWISE QUADRATIC COST FUNCTIONS GENERATIONS P1 P2 Max F F1 F2 F3 50 100 190 200 1 1 2 3 2 3 20 35 50 80 1 1 2 3 2 3 10 25 35 1 1 2 2 10 20 30 1 1 2 2 15 30 50 1 1 2 2 12 25 40 1 1 2 2 COST a 0.300 3.000 0.11 37.000 3.18 12.99 294. P. 1.02500 . 0 1 2 Iterations 3 4 x 10 5 4 Figure 4: Cost of the using slope adjustment method with load flow for adaptive learning rate (solid line) and slope adjustment method with load flow for adaptive learning rate (dashed-dotted line).083 227.802 25.650 3.210 29.000 0. Cost 850 bias 800 slope 750 Bias adjustment method with loadflow was run for fixed and adaptive learning rates at three load demands.000 COEFFICIENTS b c 1.00376 . U 1 2 3 4 5 6 Tot.369 12.80 37.826 31.200 1.050 1.0 1. In Fig 3.093 227.06 589.02350 . 3 4 5 6 TABLE II RESULTS FOR CONVENTIONAL HOPFIELD NETWORK (A). 2.75 19.385 396. March 2000 56 1 0 0 0 load flow is run to obtain the new system loss and generator power.Engineering Intelligent Systems.01750 . if the system converges.06 6.03500 .890 31.89 1390.03500 Cost 950 900 850 bias Slope conventional 800 750 0 1 2 Iterations 3 4 x 10 5 4 U Min 1 2 Fig.165 226.340 10.398 49000 31700 31500 100 100 100 0 0 50 NA 3m 51s 3m 53s 0 1 1 Slope adjustment method with load flow was run for fixed and adaptive learning rates at three load demands.in Fig 3.000 0. The results are shown in Table II for comparison.63 7.976 31.08 810.000 0. slope adjustment method with fixed learning rate (solid line) and bias adjustment method with fixed learning rate (dashed line).62 11.000 2.821 78. the generation cost was plotted for fixed .24 11.385 16.0 380 MW A B C 200 200 200 78.000 0. otherwise the process will be repeated.916 17.83 15. Initial Loadflow Hopfield N. the program will stop.06250 .0825 . For 283.000 0.4 MW. Cost Lost Iter.88 295. The results are shown in Table II for comparison.704 15.937 78.246 49.06 810.01234 .300 ..758 29.974 25000 100 50 3m 4s 1 Loadflow No Converged? 1000 Yes Optimal Cost 950 Stop 900 Figure 2 : Flowchart for the optimization with using load flow.00834 .86 294. .4 MW A B C 180.

841 0.809 MW. Units 1.628 29.482 -0.139 24 1.934 0.693 0.909 0.907 0. March 2000 57 1000 Momentum is applied to the input of each method and gain parameter for slope adjustment method to speed up the convergence of the system. 9 5 0 9 0 0 Cost Without Momentum 8 5 0 With Momentum 8 0 0 7 5 0 0 0.278 810.385 15. momentum 0. 80. respectively.174 Q(MVAr) Bus # -41.15 Total Power 227.977 2 37.606.413 26 -2. 1 0 0 0 9 5 0 TABLE III RESULTS FOR ADAPTIVE SLOPE ADJUSTMENT METHOD WITH MOMENTUM 0.16 32.9 is applied to input of bias adjustment method with adaptive learning rate using load flow.518 4 10.984 0. respectively. Also for 380 MW load.87 P (MW) -3.00E-04 1 283. 15.198 -17.074 MW.199 -9.067 16 24. 6.599 Q(MVAr) -2.005 -2.257 18 28. 5 .927 P (MW) 182.939 29.919 0.272 20 30.443 32.97 is applied to input and gain parameter of slope adjustment methods.9 applied at input (dashed-dotted line).908 0.474 3 16. .606 10.99 0.01 -3. The system bus information (after convergence) is shown in Table IV.366 19 39.629 12.647 589.825 23 2. pp.952 -1.435 15.495 -8.997 391.892 0.891 6 12.965 -3.024 79. respectively. for 220 MW load.9 APPLIED TO INPUT WITH LOAD FLOW.912 0.241 12. No.485 -2.984 27 -10.0861 9.007 -3.16 11.5 2 x 10 4 Figure 6: Cost of the using bias adjustment method with load flow for adaptive learning rate without momentum (solid line) and with momentum 0.603 -11.687 22 -0.938 0. and 12.882 0.296 In Fig. which are all very close to lower limits.024. Reduction of the iteration number for the bias adjustment method is similar to the case for slope adjustment method.999 -3.899 -0.63 -22.999 199.984 180.984 0.586 28 -0. and 29.986 48.324 389.189 21 -12.26 -2. 1.5 1 1. the number of iterations is significantly decreased.87 0.9 and 0.331 29 -5.893 0.u) 1 1 1 1 1 1 0.942 0.Engineering Intelligent Systems.252 294.322 295.004 0.188 79.965 Lost 7.5 x 10 3 4 Figure 5: Cost of the using slope adjustment method with load flow for adaptive learning rate without momentum (solid line) and with momentum 0. and 4 have 199.499 -2.324 9.5 Iterations 2 2.59 -11.8.809 29.318 30 V (p.238 27 -0.987 0.074 12.193 -8. In Fig.405 16. Units 4.9 applied at input and momentum 0.466 133.448 19.399 -10.792 -7.496 0.9 APPLIED TO INPUT AND GAIN PARAMETER WITH LOAD FLOW AND ADAPTIVE BIAS ADJUSTMENT METHOD WITH MOMENTUM 0.798 -7.385 810.197 -8.897 25.173 199.574 16.906 0.698 25 -1.922 0.334 29. As seen in Table III. the momentum factor of 0.086 226.795 -2.842 2.852 11.61 19.414 -5.687 15.5 1 Iterations 1.357 15. 220 MW U Slope Bias 1 135.691 49.989 0.999.007 -6.5.2.971 5 15.507 12.and 6 have 10.974 396.178 37.705 -1.u) 0.904 0. and the number of iterations is reduced to about 15 percent of that without momentum of slope adjustment method.981 Cost 589.981 Iterations 3200 2700 U0 100 100 Theta 0 50 Time 25s 21s n 1.951 17.381 15. 53-58. and 30 MW upper limits .741 0.908 0.357.399 0 -5.176 0.203 1390.00E-04 1 1. When the momentum is applied to the system.97 at gain parameter (dashed-dotted line). TABLE IV BUS INFORMATION OF 283.4 MW 380 MW Slope Bias Slope Bias 182.00E-04 1 Without Momentum 9 0 0 Cost With Momentum 8 5 0 8 0 0 7 5 0 0 0.385 4000 3000 3000 2500 100 100 100 100 0 50 0 50 30s 23s 25s 20s 1.958 0. Vol. which are close to 200. 79.4 MW LOAD WITH ADAPTIVE SLOPE ADJUSTMENT METHOD WITH MOMENTUM Bus # 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 V (p.362 17 28.409 -0.494 0.387 -78.76 25.

Mans. J. April 1982. M. and a linear programming circuit’. 553-541. El-Hawary. on Power Systems. PAS-104 1985 pp 1147-1153 For both figures (Fig 5. Vol. Proceedings of National Academy of Science. May 1984. I. the iteration numbers are very low and the computation time is around 20-25 seconds. King. 96 TP 112-0. Skapura. Sode-Yome. The results of the cases for both methods are shown in Table 4. Vol. Y. Lee. Wood and B. 144-156. 53-58. IEEE Transactions on Systems. 4. “Optimal environmental dispatching of electric power system via an improved hopfield neural network model”.Engineering Intelligent Systems. CAS-33. No. Especially with momentum applied. and Programming Techniques. the slope and bias adjustment are applied to the real power optimization with loadflow. J. El-Hawary. Transactions on Neural Networks’. January/February 1992.. IEEE Transactions on Power Apparatus and Systems. Neural Networks Algorithms. Lin. March 2000 58 and Systems. No. K. . 9. A. 5. Y. Wollenberg. 79. and Control. Co. 8. Niebur.. August 1995. 283. Inc. Hopfield. J. Tank. Addison-Wesley Pub. J. 3. 12.. ‘Neural computation of decisions in optimization problems’. El-Sharkawi and D. J. ‘United Approach to Optimal Real and Reactive Power Dispatch’. W. the system converged very smoothly. IEEE Transactions on Power system Apparatus 13. IEEE. D. For all three load levels. 1030-1038. 1559-1565. ‘Economic load dispatch for piecewise quadratic cost function using hopfield neural networks’. USA. and J. Second Edition. Vol. Hopfield and W.4 MW load was used. slope and bias adjustment methods gave very good responses. IEEE Trans. Kim. PAS-10. Vol. J. August 1993. ‘Simple neural optimization networks: an A/D converter. Lee. Applications. and Fig. 11. J. 3. Power Generation. D. June 1984. John Wiley & Sons. K. Y. T. pp. Vol. NY (1996) 2. pp 141-152. 2554-2558. H. Ku and K. CONCLUSIONS 8. J. July 1992. Application of Artificial Neural Networks to Power Systems. 1996. 6. and F. 6). Lee. J. IEEE Transactions on Power Systems. 30883092. No. and K. Tank and J. Operation. No. Park and J. No. signal decision circuit. E. 52. ‘Adaptive Hopfield Neural Networks for Economic Load Dispatch’.8. Hopfield. 10. 5. M. 6. 6. and Cybernetics. 1. A. In this paper. Park. and G. K. D. REFERENCES 1. Y. IEEE Trans. IEEE Power System Society. L. H. Vol 13 No 2 (May 1998) pp 519-526 C. M. Hopfield. pp. A.. 1. Freeman and D. pp. S. pp. C. New York. Proceedings of National Academy of Science. Ortiz. Vol... 3. Eom. L. ‘Hierachical economic dispatch for piecewise quadratic cost functions’. ACKNOWLEDGEMENTS This work was supported in part by the National Science Foundation under grants INT-9605028 and ECS9705105. 1. pp. 6. Y. USA. 7. Vol. ‘Neurons with graded response have collective computational properties like those of two-state neurons’. A. 81. January 1995. pp. Around lower and upper limits of generator. M. 1985. Vol. on Power Appratus & Systems. C. ‘Neural networks and physical systems with emergent collective computational abilities’. ‘Diagonal recurrent neural networks for dynamic systems control. Lee. E. pp. F. Biological Cybernetics. Park. Vol. No. Y. Viviani.” IEEE.

Sign up to vote on this title
UsefulNot useful