Multi-objective Optimization of Reactive

Power Flow Using Demand Profile

Classification
Rui He, G.A. Taylor, and Y.H. Song
Abstract--For

decades,

research

on

reactive

power

management as a snapshot problem employing a single objective

function has been widely performed. However, this cannot satisfy
the ISOs aim of optimization on several points simultaneously

g : Generator bus
i : Node number
L : a set of load points in a continuous period

for a successive period. Therefore, in this paper, a multi-objective

N g , N d : Number of the generator buses and the demand

optimal reactive power flow (ORPF) formulation with respect to

buses

the time-domain is proposed.

The proposed formulation

minimizes both losses and payment for the provision of the

reactive power service in the framework of the UK daily
Balancing Market. Prior to the optimization procedure, the
related control parameters can be ordered with the aid of a load
classification method, in order to simplify the control actions.
During the optimization, compromise programming is applied to
achieve the multi-objective. A case study on a 60-bus test system

P, Q : Active and Reactive power injection

S ijf , S ijt : Apparent power flow of the from and to bus

T : Tap-changer
V : Voltage of the bus
: Phase angle of the bus

is presented to illustrate the application of the proposed modeling

framework.

II.

INTRODUCTION

Index Terms--control action, demand profile classification,

multi-objective, partition, reactive power control, self-adaptive,
sub-interval, time-domain, transition-optimization

I.

NOMENCLATURE

d : Demand bus

s a consequence of the global restructuring of electricity

industries over the last decade, a wide range of reactive
power markets and pricing mechanisms have been proposed
and adopted [1]. With regard to the UK power system,
National Grid Transco (NGT) provides a balancing market in
which the reactive power service is provided [2]. In the
market, successful generators sign bilateral contracts with
NGT for the provision of reactive power, while those

f1 : Functional form of the objective to minimize loss

unselected generators that can be required to provide reactive

power service will receive a default payment which is less

f 2 : Functional form of the objective to minimize total

than the payment made by the market agreement [2, 3].

NGT, as a buyer of reactive power service from those
generators, has an incentive to keep such expenditure at a
minimum level. Furthermore, as the operator and owner of
transmission network, NGT also needs to consider final
operating decisions, such as the minimization of system real
losses, and lower cost of the operation and maintenance of the
whole network, etc. To achieve those targets, a compromise
programming method based on the daily load curve as

payment

f 1* , f

*
2

: Ideal solutions of the objectives in the

independent optimisation
The authors are with Brunel Institute of Power Systems, Brunel University,
Uxbridge, London (e-mails: [email protected], [email protected]).

presented in detail in [3] in order to minimize both losses and

payments.
However, the computational effort required in [3] for the
24-hour optimisation of real losses and payments requires
excessive calculation. This is because at the beginning of
every Settlement Period Duration (SPD), the control
parameters such as tap-ratios should be selected again.
Moreover, the frequent changing of the control equipment in
the system will increase the operational cost and accelerate the
depreciation of equipment, all of which are not in the
economic interest of the ISO.

III.

MASTER LEVEL LOAD CLASSIFICATION

To illustrate the classification process, an example of a

daily load demand curve imitating the UK demand curve [7] is
presented in Fig.1. During the 24 hours, we can see that the
peak demand appears at 19 oclock while the lowest value
appears at 4 oclock in the morning. Moreover, for a certain
period, the load varies slowly, such as from 9 oclock to 17
oclock. And for the other period, the load varies fast. So next,
the proposed method will partition the curve into several
segments in the time domain according to these
characteristics.

So, to speed up the calculation process and to simplify the

control actions, it is necessary to divide the load curve into
several sub-intervals. In each sub-interval control action takes
place only once. In [4-6] this problem is formulated in a
transition-optimised fashion. In [4], six typical operating
points are chosen to represent the daily load pattern. However,
such six points may be inadequate to represent the whole daily
load if the load varies more significantly in one day. In [6], an
adaptive limit method was proposed that could perform
effectively by using two statistic measurements when
partitioning the load profile. The number of groups of the
sub-intervals is calculated at the end of classification.
However, being a commercial organisation, NGT are likely to
want to decide the times of the control actions a day in
advance. To meet this goal, this paper proposes a method to
partition continuous load points into a certain number of
groups according to the characteristics of the load profile.
Afterwards, at the beginning of each sub-interval,
corresponding tap-ratios will be given and stay stable for the
whole sub-interval period. Thus, unnecessary change of the
control equipment can be avoided, and the required number of
control actions taking place in a day could be realised. So, in
this paper, a master-slave-level method is proposed to
optimise a multi-objective daily problem. The master level is
used to divide the daily load curve into several sub-intervals
and to reduce the depreciation cost via the fixing of
corresponding tap-ratios; while the slave level is used to solve
the multi-objective problem with fixed tap-ratios at each SPD.
The rest of the paper will initially introduce the master
level and the slave level respectively, then demonstrate the
technique with a 60-bus test case, and finally draw
conclusions based on the results and analysis.

Fig. 1. The 24-Hour Power Demand

A.

Two Measurements

The classification is based on the scattering and variability

analysis of the daily demand curve, and such analysis will
determine a tolerance for the partitioning. Starting from 0
oclock, so long as the consequent points are in the tolerance,
these points can be grouped into one segment. To judge
whether those points are in the tolerance or not, there are two
measurements in use, the standard deviation and the range.
The standard deviation represents the degree of the
spreading of a group of points. It can be used to classify
scattering points. However, if a skew point occurs, the
standard deviation method may not reflect the characteristics
of those points precisely. To solve this problem, the range
representing the difference between the highest and the lowest
values works as a supplementary measurement. Formulations
of these two measurements are given out in following content.
1) The Standard Deviation
n

Where

( L

i = 1

L )

(1)

L =

i =1

If we choose tolerance as S equal to 0.026, the actions will

happen at 7:00, 9:00 and 22:00 shown in Fig.2. If S is 0.024,
the actions will take place at 7:00, 9:00, 22:00 and 23:00
shown in Fig. 3.

S=0.026, R=0.048, the control actions will happen at 6:00,

7:00, 9:00, 19:00, 20:00, 22:00 and 23:00 shown in Fig. 5.
From Fig. 4 and Fig. 5, we can see that by introducing the
range as the supplementary measurement, the partitioning of
those load points could be more reasonable.

From Fig. 2 and Fig. 3, we can see that choosing a proper S

value will help to reduce control actions greatly. However,
after a long stable situation, at the coming skew point, even
the power demand increases suddenly, no control action is
made. This may cause the voltage collapse in the real system.
So, to amend this partitioning method, the range measurement
is needed.
Fig. 4. Classification Result 3 (S=0.026 R=0.06)

Fig. 2. Classification Result 1 (S=0.026)

Fig.5. Classification Result 4 (S=0.024 R=0.046)

B.

Expected Action Times

Using the method above, not until the end of the

classification, could the number of groups of those subintervals be explicitly predicted. So, for an ISO who wants to
decide the times of control actions before optimisation, this
method needs to be improved. A self-adaptive concept is
employed. Such self-adaptive process is based on the
comparison of number of the sub-intervals obtained in
calculation to the expected number. If the former is larger, it
means the tolerances of R and S are smaller. In this case, those

Fig. 3. Classification Result 2 (S=0.024)

2) The Range

R = Max ( L ) Min ( L )

(2)

Still, we choose S=0.026, and for the tolerance R is set to

be 0.06. Then the control actions will take place at 7:00, 9:00,
19:00,21:00 and 22:00 shown in Fig. 4. Moreover, if we set

tolerances should be decreased. And vice versa.

In Fig. 6, the flow chart shows the process of classification
using self-adaptive concept.
After the cycle of partitioning, adapting tolerances and
comparing, the serial points can be grouped into a certain subintervals.

Because in the application of the classification method,

each sub-interval is constrained by both standard deviation
and range measurements, the loads in a sub-interval can be
regarded as staying comparatively close. Consequently, it is
reasonable to use the mean of the loads in a sub-interval as a
parameter to determine the tap-ratios for further optimisation
of reactive power flow in the slave-level.

Min Distance =

f 1* f 1 2
f 2* f 2 2
)
+
(
)
f 1*
f 2*

(3)

Subject to:

Input expected n

Pgi Pdi P (Vi , i , Ti ) = 0

(4)

Q gi Qdi Q(Vi , i , Ti ) = 0

(5)

Vi min Vi Vi max

(6)

Pig min Pi Pig max

R0, S0

(7)

Update S, R

Qig min Qi Qig max

(8)

Partition L, get n

n=n?
N

(9)

Sijt S ijmax

(10)

Ti min Ti Ti max

n>n?
Y

R = R R R = R + R
S = S S S = S + S

output
Fig.6. Flow Chart of Self-Adaptive Load Classification Process.

IV.

S ijf S ijmax

SLAVE LEVEL MULTI-OBJECTIVE

OPTIMISATION

In the slave level, the multi-objective to minimize real

losses and payments for provision of reactive power services
simultaneously is carried out. Because the action times are
greatly reduced by the load classification method, and
corresponding tap-ratios are fixed at the beginning of each
sub-interval, the calculation efficiency is improved a lot.
For the multi-objective optimisation, the processes to
minimize loss and payment are not in the same direction. Due
to the inconsistency of the two independent objectives, a
compromise programming is used [3, 8, 9]. The formulation
of such multi-objective function is,

(11)

where Distance is the performance index measuring the

optimisation of the multi-objective function,.
V.

CASE STUDIES

A 60-bus system is constructed to test the proposed

method. This system is modelled on the basis of the southern
part of the UK network. The voltage ranges from 400kv to
33kv, and in total we assume there are 85 branches and 45 tapchangers in this system. And based on the data provided on
NGTs website, the prices for provision of the reactive power
services are created.
Because tolerances S and R will affect the partitioning
result, the choice of their initial value should be very crucial.
Due to the characteristics of the power system itself, the
control action should take place when the load demand meets
the peak point or when the change rate of the load demand is
very high. Based on this principle, the initial value of R is set
equal to the highest load difference in one hour, R0 =0.11, S0
=0.024.
A.

Results of Load Classification

Due to the 8-hour work policy, it is assumed that the

control action needs to be done exact 8 times a day. With the
expected number of groups equal to 8, after application of the
partitioning method, the daily power demand load curve can
be divided. And the tolerances of R and S are updated to 0.015
and 0.1091 separately. In Table I, considering these

tolerances, we have an explicit list of a series of standard

deviations and ranges indicating how the two measurements
function in the partitioning process.

greatly reduced. Besides, the losses dont change too much.

And the saving of the losses compared with the losses got in
the power flow is still great. Moreover, for those sub-intervals
that contain only one SPD, the loss doesnt change.
Table II

Table I
STANDARD DEVIATIONS AND RANGES

MULTI-OPTIMISATION WITHOUT LOAD CLASSIFICATION VS.

MULTI-OPTIMISATION WITH LOAD CLASSIFICATION

Time
00:0006:00
06:0007:00
07:0008:00
08:0019:00
19:0020:00
20:0021:00
21:0022:00
22:0023:00

Standard deviation > S=

0.015?
Standard
Y/N
deviation
0.0155
Y
0.0550
0.0264
0.0255

Y
Y
Y

Range > R= 0.1091?

Range

Y/N

0.0468

0.1100

0.0527
0.1105

Ti
me

N
Y

0.0248

0.0496

0.0248

0.0496

0.0419

0.0837

0.0330

0.0660

Fig. 7. Partitioned Load Demand

B.

Results of Multi-objective Optimisation

The results of the optimised losses and payments for 24

hours are shown in Table II. And also, Fig. 8 and Fig. 9 show
the comparison of losses and payments in detail.
From both Table II and Fig. 8, we can see that with the
proposed classification method, the control action times are

Action
4
Action
5
Action
6
Action
7
Action
8

With the
partition method
f-pay f-loss
f-pay

Without Partition
f-loss

0:00 46.876

190.693

47.041

442.833

46.443

600.984

1:00 46.455

216.325

47.557

762.735

51.083

225.192

2:00 46.437

184.811

48.665

235.982

47.566

316.583

3:00 45.894

209.305

50.145

304.978

49.712

251.667

4:00 45.763

172.636

49.299

222.94

48.665

292.846

5:00 45.985

181.633

47.390

218.438

49.763

305.058

199.588

50.807

380.500

50.807

380.500

315.137

65.391

662.593

65.391

662.593

Action
6:00 48.049
1
Action
7:00 61.055
2
Action
8:00 72.179
3

From Table I, we can see that period between 8:00 and

18:00 is grouped into the same sub-interval due to both effects
of standard deviation and range, as well as period 6:00. For
the rest of the points, standard deviation functions as the
primary measurement.
Based on Table I, Fig. 7 shows the time when the control
actions will happen during a day lively.

Independent
solution
f1-loss f2-pay

420.227

80.162

741.14

80.250

703.876

9:00 78.648

430.947

86.251

850.524

86.507

831.923

10:00 78.746

493.612

86.102

852.423

86.705

851.179

11:00 79.238

556.42

90.262

798.054

87.188

877.653

12:00 79.738

447.133

88.01

982.463

87.665

990.192

13:00 79.226

459.321

86.426

890.823

89.657

762.164

14:00 79.484

449.104

83.268

1054.571 86.525

907.994

15:00 76.394

516.687

79.182

1027.972 77.317

916.499

16:00 79.560

475.907

86.983

809.933

86.748

916.195

17:00 82.904

576.288

87.521

1207.699 86.001

1181.55

18:00 88.969

627.593

89.715

1115.343 89.191

1262.05

19:00 111.45

1211.93

123.60

1331.35

123.60

1331.35

20:00 85.838

993.833

92.620

1059.51

92.620

1059.51

21:00 75.191

414.757

81.557

734.458

81.557

734.458

22:00 57.782

261.204

64.640

511.862

64.640

511.862

23:00 49.613

237.922

52.654

427.587

52.654

427.587

In total

641.48

10243.02 1765.246. 17626.71 1768.17 17338.72

Saving

350.62

77408.19 226.85

70024.5

223.93

70312.48

Fig.8. Losses Without and With Load Classification

Fig.10. Variation of Tap-Ratio before and after the load classification

From both Table II and Fig. 9, more optimistic situation is

shown. The payment in total is reduced while the control
action times are fixed to 8 times a day. From this point, we can
see that the financial benefits are more obvious.
In short, from Table II, we can see that with the multiobjective optimisation, the real losses and total payments are
reduced. Furthermore, if the partitioning method is used
before the optimisation process, the reduction is obvious as
well, and to some degree, the reduction, for example, the
payment reduction, may increase more. These results indicate
that this partitioning method helps not only reduce the control
actions, but also the real losses and the payment.

VI.

CONCLUSIONS AND FUTURE WORK

The application of the proposed partitioning method has

been demonstrated in the newly proposed multi-objective
optimisation procedure. A 60-bus system representation of
part of the UK network has been employed to test the method.
Based on the above results, the following conclusions can be
obtained:
A. The proposed measurements can help partition the
demand curve into several groups according to the
load profile. Furthermore, with the self-adapting
process, the expected number of control actions can
be realized.
B. By fixing tap-ratios at the beginning of each subinterval, the optimisation process is still achieved.
Moreover, the results are more satisfying in an
economic sense.
C. The two-stage multi-objective optimisation process
can converge to a compromised optimal solution, in
which both losses and payments are close to the
minimum values.

Fig. 9. Payments Without and With Load Classification

C.

Variation of Tap-Ratio

In order to illustrate the variation of the tap-changers, a

typical point is chosen. Such tap-changer is installed between
bus 6(400kV) and bus 45(132kV), and in the daily
optimisation process, it has the highest change rate among all
those tap-changers. From Fig. 10, we can see that with the
proposed load classification method, the change of those tapratios can be reduced a lot. Besides, before the load
classification, the variation of the tap-changer is unpredictable.
However, after the load classification, the change is simpler
and more predictable.

To simplify the calculation process, those discrete

parameters such as tap-ratios are dealt with as continuous
variables. This may cause some inaccuracy in the calculation
procedure. However, it does not adversely affect the
application of the proposed load classification method. In
future research the author will apply an improved successive
linear programming method in order to solve this discrete
parameter problem.
VII.

REFERENCES

[1] K. Bhattacharya, J. Zhong, Reactive Power as an Ancillary Service, I4

Transactions on Power Systems, Vol. 16, No. 2, May 2001, pp.294-300.
[2] CUSU SCHEDULE 3, National Grid, version 1.2, 31st December
2002. Available: http:// www.nationalgrid.com/uk/indinfo/index.html

[3] R. He, Z.H.Bie, Y. H. Song, Yosuke Nakanishi and Chikashi Nakazawa,

focusing on reactive power market. The work on Multi-objective ORPF in

Multi-Optimisation of Reactive Power Flow In Balancing Market, in Proc.

Balancing Market was presented on 39 INTERNATIONAL UNIVERSITIES

2004 International Universities Power Engineering Conf., pp.1125-1129.

POWER ENGINEERING CONFERENCE (UPEC) in 2004.

[4] Y.Y. Hong, C. M. Liao, Short-term scheduling of reactive power

controllers, IEEE Trans. on Power Systems, vol.10, pp860-868, May 1995.
[5] G. A.Taylor, Y. H. Song, M.R.Irving, M.E.Bradley, AND T.G.Williams,
A Review of Algorithms and Heuristic Based Methods for Voltage/VAr
Control in Proc. IEEE-PES/CSEE International Conference on Power System
th

th

Technology, Kunming, China, Vol. 3, pp1660-1664, 13 17 October 2002.

[6] Z.H.Bie, Y. H. Song, X. F. Wang, G.A.Taylor, and M.R.Irving, A
Transition-Optimized Approach to Reactive Power and Voltage Control,

G.A. Taylor received his BSc from the University of London, UK in 1987
and MSc and PhD from the University of Greenwich, UK in 1992 and 1997,
respectively. He was the National Grid UK post-doctoral scholar at Brunel
University, UK from 2000-2003. He is currently a lecturer within the Brunel
Institute of Power Systems, School of Engineering and Design, Brunel
University, UK. His research interests include power system and network
optimisation.

th

presented at IEEE General Meeting 2004, Denver, Colorado, US, 6-10 June
2004.

Y.H. Song is a Professor of Electrical Energy Systems at Brunel University,

[7] Seven Year Statement, National Grid, May, 2002. Available: http://

UK, where he is Pro-Vice-Chancellor of the University. He is a Fellow of

www.nationalgrid.com/uk/library/index.html

Royal Academy of Engineering (UK) and a fellow of the IEE (UK). He was

[8] D.Chattopadhyay, James Momoh, A multiobjective operations planning

model with unit commitment and transmission constraints, I4 Transactions
on Power Systems, Vol. 14, No. 3, August 1999, pp.1078-1084.
[9] Jin Zhong, Kankar Bhattacharya, Towards a Competitive Market for
Reactive Power, I4 Power Engineering Society General Meeting, 2003. Vol.
2, July 2003, pp.1109-1109.

VIII.

BIOGRAPHIES

Rui He was born in Chengdu, China. After obtaining her Bachelors Degree
in Electrical Engineering from Zhejiang University, China in October 2002,
she started her doctorial research in power marketing under the supervision of
Prof. Y.H. Song in Brunel University, United Kingdom. She has been

awarded a DSc by Brunel University in 2002 for his significant research

contributions.