You are on page 1of 11

2136

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 28, NO. 3, AUGUST 2013

A Multi-Objective PMU Placement Method


Considering Measurement Redundancy and
Observability Value Under Contingencies
Seyed Mahdi Mazhari, Student Member, IEEE, Hassan Monsef, Hamid Lesani, and
Alireza Fereidunian, Member, IEEE

AbstractThis paper proposes a multi-objective phasor


measurement units (PMUs) placement method in electric transmission grids. Further consideration is devoted to the early PMU
placement formulations, to simultaneously determine minimum
number of PMUs, as well as maximum measurement redundancy. Moreover, a new methodology is presented for valuation
of observability under contingencies, including line outages and
loss of PMUs. Furthermore, a generalized observability function
is introduced to allocate the PMUs in presence of conventional
non-synchronous measurements. The resultant optimization
problem is solved using Cellular Learning Automata (CLA),
introducing new CLA local rules to improve the optimization
process. The developed method is conducted on IEEE standard
test systems as well as the Iranian 230- and 400-kV transmission
grids, followed by a discussion on results.
Index TermsCellular learning automata (CLA), measurement
redundancy, observability, optimal placement, phasor measurement unit (PMU), power system contingencies.

Set of neighboring lines of bus .


Set of union of bus and neighboring buses of
the th bus.
Set of lines incident to bus .
Set of actions of the th learning automaton.
Set of inputs of the th learning automaton.
Set of actions probability vector of the th learning
automaton.
Constants:
Total costs imposed for placing PMU at bus ($).
Maximum project budget ($).
Minimum number of redundancy requirements
of bus .

NOTATION
The notation used throughout this paper is reproduced below
for quick reference.
Sets:
Set of network buses.
Set of network lines.
Set of lines which would be unobservable when
line is out.

Failure rate of PMU (fr/year).


Failure rate of line (fr/year).
Disturbance rate of line including events which
caused a failure in line or ones which are cleared
by the protective devices (fr/year).
Total number of network buses
Average outage time of PMU (hr).

Set of lines which would be unobservable when


PMU is lost.

Average outage time of line (hr).

Set of buses which have PMU.

Penalty parameter of the CLA algorithm.

Set of costs for placing PMUs at

Reward parameter of the CLA algorithm.


Monetary factor of observability value of line .

Very large number.


Manuscript received October 12, 2011; revised February 08, 2012, May 24,
2012, and September 06, 2012; accepted October 17, 2012. Date of publication
February 01, 2013; date of current version July 18, 2013. Paper no. TPWRS00907-2011.
S. M. Mazhari, H. Monsef, and H. Lesani are with the School of Electrical
and Computer Engineering, University of Tehran, Tehran, Iran (e-mail:
mazhari@ut.ac.ir; hmonsef@ut.ac.ir; lesani@ut.ac.ir).
A. Fereidunian is with the Electrical Engineering Faculty, K. N. Toosi
University of Technology and also with the School of Electrical and
Computer Engineering, University of Tehran, Tehran, Iran (e-mail: fereidunian@eetd.kntu.ac.ir; arf@ece.ut.ac.ir).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TPWRS.2012.2234147

Small number.
Functions:
Objective function of the PMU placement
problem.
Observability function of bus .
Observability function of bus due to
non-synchronous measurements of bus
between buses
).

0885-8950/$31.00 2013 IEEE

(or

MAZHARI et al.: MULTI-OBJECTIVE PMU PLACEMENT METHOD CONSIDERING MEASUREMENT REDUNDANCY

Penalty function of PMUs unreliability.


Penalty function of lines unreliability.
Penalty function of buses redundancies.
Penalty function of buses unobservability.
Probability of selecting action at the th iteration.
Observability value of each disturbance.
Variables:
Binary decision variable that is equal to 1 if PMU
is installed at bus and 0 otherwise.
Binary decision variable that is equal to 1 if bus
is observed and 0 otherwise.
Binary decision variable that is equal to 1 if bus
is zero-injection bus or there is a power injection
measurement at bus and 0 otherwise.
Binary decision variable that is equal to 1 if there
is a voltage measurement at bus and 0 otherwise.
Binary decision variable that is equal to 1 if there
is a flow injection measurement between buses
and 0 otherwise.
Binary decision variable that is equal to 0 if the
th power injection measurement has observed a
network bus and 1 otherwise.
I. INTRODUCTION

HASOR measurement units (PMUs) are crucial elements


of wide-area state estimation systems in transmission
grids, as they maintain a high quality observability on electrical
quantities of power system [1].
Power system operators should ensure that electrical quantities lie within their normal limits. They utilize state estimation
to observe the system conditions. Since full observability of the
system is a prerequisite to the state estimation, measurements
are distributed throughout the network and joined to the state
estimators [2]. Based on these measurements, state estimators
obtain an estimate of the state of the power system, while ensuring consistency of the estimates with the measurement set
[3]. It is said that a system is topologically observable, if all of
its bus voltage phasors can be estimated [2].
The PMUs are intelligent electronic devices (IEDs), which
provide synchronized real-time measurements of voltage phasors, as well as predetermined number of incident current phasors at the buses where they are located [4]. PMUs are synchronized in a wide-area interconnected network using global positioning systems (GPS). According to Ohms Law, when a PMU
is placed at a bus, neighboring buses also become observable.
Furthermore, if voltage phasors at both ends of a branch are observed, the branch current can be calculated using the Kirchhoffs Voltage Law (KVL) [5]. Accordingly, it is neither economical (due to PMU cost) nor possible (due to nonexistence of
communication facilities in some substations) and even nor necessary (due to Ohms and Kirchhoffs Laws) to install a PMU

2137

at every node of a wide-area interconnected network. Hence,


optimal PMU placement is formulated as a power system optimization problem and is investigated in various studies [5][18].
Considerable research has been devoted to PMU placement
problem in literature. The performance of genetic algorithm
and simulated annealing is investigated for solving the PMU
placement problem in [5] and [6], respectively. In [7], immunity
genetic algorithm is applied to a standard test system. While
satisfactory results are presented, obtained results may hardly
meet the topological observability constraints [15]. Although
these methods can find minimum number of PMUs, measurement redundancy, as a conflicting and progressive objective,
is rarely considered. Non-dominated sorting genetic algorithm
(NSGA) is proposed in [8] for simultaneous minimization of
number of PMUs and maximization of measurement redundancy; however, this method may not lead to the minimum
number of PMUs that can make the system observable [9]. A
binary search algorithm is proposed in [9] for system observability, which is capable of finding the global optimum, yet it is
computationally expensive.
Moreover, integer programming has been widely used for
PMU placement [10][13]. In [10], integer quadratic programming is proposed, considering single line outages and loss of
PMUs. However, effects of zero-injection buses are not considered in the developed algorithm. A generalized integer linear
programming is also used in [11]. The method considers power
network with and without conventional measurements; nevertheless, it needs some modifications to be used in multistage
planning. Moreover, it is not studied the contingencies. A new
integer linear programming based formulation is proposed for
multi-stage scheduling of PMU placement in [12]. Although the
method considers PMU losses, line outage contingencies are not
incorporated. Reference [13] conducts a binary integer linear
programming, considering system observability and minimum
metering economy as objectives. It proposes a new method to
model the conventional measurements and studies the effects
of unequal PMU costs and PMU outages, yet line outage contingencies are not considered there. In [14], a contingency-constrained PMU placement model is proposed, presenting a new
formulation for the problem, considering the effects of PMU
losses and line outages. Although, solution time is substantially
decreased, their proposed formulation may not meet topological observability constraints, while two or more zero-injection
buses are interconnected in the system [15]. Hence, infeasible
results are reported in some cases (see Section V-A).
Despite aforementioned literature, some researchers have
studied the PMU placement for purposes other than full observability [16], [17]. In [16], a gradual PMU placement is
proposed, considering depth of unobservability. Moreover, [17]
developed an algorithm in which PMU placement is studied for
bad data processing in state estimation.
To the best of our knowledge, worth of observability under
contingencies is not considered in previous studies yet, since it
is assumed that the network remains full observable in contingencies. However, meeting the mentioned constraint in presence
of PMU losses and line outages needs excessive PMUs, leading
to proposing PMUs for almost half of network buses [10], [12]
and [14], [15].

2138

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 28, NO. 3, AUGUST 2013

This paper reports a modified multi-objective PMU placement method. First, worth of observability is calculated for a
set of PMUs in contingency, by ranking the prospective contingencies. As the most note-worthy contribution of this paper, the
prevalent PMU placement problem formulation is extended and
new concepts are considered, as for maximizing measurement
redundancy and observability valuation under contingencies, all
in presence of conventional measurements. The contingencies
considered in this paper include line outages and PMU losses.
Moreover, as the second contribution of this paper, the optimization problem is tackled by Cellular Learning Automata (CLA),
introducing new CLA local rules to enhance the optimization
process. Finally, the developed method is applied to the IEEE
standard test systems as well as the Iranian 230- and 400-kV
transmission network, followed by presenting the results and
comparing them to those of previous research.
II. PROBLEM FORMULATION
A. Mathematical Modeling
In this paper, the multi-objective PMU placement problem is
formulated as in (1)(12). Equation (1) represents the objective
function of the PMU placement problem, in which, the first term
represents total cost of PMUs installed throughout the network.
The second term represents a penalty factor for observability redundancies in normal operation; and the third and fourth terms
provide penalty functions for buses unobservability under contingencies. Finally, the fifth term is a penalty function which accounts for the network unobservability in normal operation. The
PMU placement problem constraints are represented in (2)(3).
While (2) assures that whole buses possess their least predefined redundancies, and (3) guarantees that total project cost is
less than the project budget:

(1)
Subject to:
(2)
(3)
where:
(4)
(5)

(6)

(7)

(8)
(9)
(10)

(11)
(12)

B. Description on the Employed Terms


In the first term of (1), the total costs imposed for placing a
PMU at a bus are affected by its geographical position, type of
PMU, and status of the existing telecommunication system at
that bus [13].
The second term of (1), guarantees the achievement of maximum measurement redundancy while minimizing the number
of PMUs during the optimization process. In this term,
shows penalty function of buses redundancy and is calculated
through (4). In this equation,
is the maximum number of
network redundancies, which appears when PMUs are installed
at all of the network buses. Equation (5) shows how
is
calculated for a network. Accordingly, minimizing a penalty
function as
, can guarantee the achievement
of maximum measurement redundancy. Since, the mentioned
penalty parameter and the first term of (1) are not as the same
type, simultaneous optimization of these terms by a single
objective function cannot lead to the optimal solution [8]. To
overcome this deficiency, the penalty function is divided by
maximum network redundancy and the obtained result is multiplied by minimum PMU cost as shown in (1) and (4). In (4),
is a small quantity, which ensures that
in all situations.
It is empirically seen in simulations that if
,
the optimization algorithm always converges to the optimal
solution. In (1), the redundancy penalty function,
, is multiplied by PMU cost to increase the worth of redundancy during
optimization process. Moreover,
is selected, in order
to ensure that
is always less than a PMU cost
Hence, although an increase in number of PMUs decreases
the
, increases in the first term of (1) would be more than
decrease of the second term, thus leading to an increase in the
objective function. Therefore, the possibility of entrapping in
local minima is reduced, while the first two terms of (1) are
minimized.
Third and fourth terms of (1) show effects of PMU losses and
line outages, respectively; considering worth of observability
in contingencies. The aim of PMU placement is to make the
system topologically observable. Therefore, the system operators expect to observe all network disturbances (including faults,
blackouts, or ones which are cleared by the protective devices
and etc.) by efficient utilization of a number of PMUs at proper
buses. However, PMU losses and line outages may cause a set of

MAZHARI et al.: MULTI-OBJECTIVE PMU PLACEMENT METHOD CONSIDERING MEASUREMENT REDUNDANCY

disturbances not to be observed in contingency situations. To resolve this issue, the researchers have solved the PMU placement
problem with respect to
contingency criterion [10][15].
This criterion guarantees that the system is observable under
single contingencies. Although the
criterion substantially improves the system reliability, it may be economically
inefficient as it proposes excessive PMUs, comparing to the situation without this criterion [14], [15]. Similarly, observability of
some lines may have little significance [19]. On the other hand,
from the technical point of view, the system operator may tend
to observe the current flows through a corridor in any situation,
thus need to consider higher orders of reliability like
and
so on. Therefore, it is not reasonable to limit the observability
level of such lines by
criterion [20], [21]. Addressing
these concerns is the motivation behind proposing the observability value in this work, i.e., the third and fourth terms of the
objective function.
Therefore, worth of observability in contingencies is evaluated and the PMU placement is conducted based upon an economic objective function. Third and fourth terms of (1) show
effects of PMU losses and line outages, respectively. It is assumed that a PMU is the main element for sensing a set of line
disturbances which are observed directly or by the aid of this
PMU. Therefore, it can be said that the PMU cost is depreciated to sense a set of disturbance during its lifecycle. Hence,
each disturbance has a monetary value like (8). Since some of
network disturbance may be missed without PMUs within contingencies, it would be economical to place a PMU at a bus, if
that helps a considerable number of disturbances to be sensed
in such situation. Thus the financial benefit of installing a PMU
is significant comparing to its cost.
Worth of un-sensed disturbances due to PMU losses and line
outages are formulated through (6) and (7), respectively. In (6),
the annual PMU outage time
is multiplied by the
set of disturbances which are not sensed during its outage time.
It should be noted that the set of disturbance that are sensed
directly or by the aid of PMU is calculated as
.
Since the cost of PMU is depreciated to observe these disturbances, (6) is able to calculate worth of disturbances which are
not sensed because of the PMU outages. In this equation, the
union
operator is used to represent the simultaneous occurrence of disturbances in contingencies.
A monetary factor for observability value of line is considered in the penalty function, to represent the criticality of the
line. A line may have strategic role regarding its usage, the path
which it construct, and so on. For instance, transmission lines
which export electrical energy between two countries or two regions might have more importance for the system operator. In
order to meet these constraints in the proposed penalty function,
, which shows the monetary factor for observability value of
line , is added to (6). The factor
can be set equal to 1 for
a simple network or when the actual value of
is not determined. However, if
is set to , the optimization algorithm
tries to find the PMU layout in which line shall be observed
under contingencies.
In (7), similar to (6), the annual line outage time
is multiplied by the set of disturbances which
are not sensed during its outage time. Worth of each distur-

2139

is calculated by dividing the total PMU costs by


bance
total number of network disturbances as formulated in (8).
It must be noted that, (6) and (7) are formed based on single
contingencies [12][14]. However, they can easily be extended
to consider multiple simultaneous contingencies. Nevertheless,
unlike [12][14] in which power network remains observable
during contingencies, that might not be necessarily economical,
our proposed method may not lead to complete observability
in contingencies unless it is economically efficient. Since the
third and fourth terms of (1) represents the monetary worth of
observability in contingencies, the first four terms of (1) have
the same unit ($); hence, they can be simultaneously optimized
through a single objective function as reported in (1).
The fifth term of (1) shows penalty parameter of network unobservability in which
is the penalty function of bus unobservability and is calculated through (9). In this equation,
is a binary decision variable that is equal to 1 if bus i is observable and is set to 0 otherwise. Moreover, as illustrated in
the fifth term of (1), the unobservability penalty function
is multiplied by a very large number
. If all of the network
buses are observed, the
will return zero and no penalty is
added to the objective function. Hence, the objective function
will return the least value for the feasible solutions, comparing
to those of infeasible ones. It is empirically seen in simulations
that if
), the optimization algorithm always
converges to the optimal solution.

C. Proposed Observability Function


A new observability function is proposed in (10). This
equation remains unchanged in presence of conventional
non-synchronous measurements and zero-injection buses.
While the first term represents direct observability from neighboring PMUs, the effects of conventional measurements are
considered within the second and third terms. In the third term,
shows observability due to installation of a conventional
voltage phasor measurement at bus . If a voltage measurement
is installed on a bus, it measures voltage phasor of that bus, and
the bus becomes observable [15], [22].
In addition,
which is calculated as (11) represents the
times by which the th bus is observed due to the effects of zeroinjection buses or conventional measurements that are placed
at bus (or between buses and ). Since a branch flow measurement can easily measure the branch current, if the voltage
phasor at one end of the branch is observed, the voltage phasor
at the other end becomes observable according to Ohms Law
[15], [16], [22]. This effect is considered in the first term of (11).
For a set of buses that coincide to a zero-injection bus, plus that
zero-injection bus, if the voltage phasors are all observable except one, the unobservable bus also will be observed via KCL
[11][16]. This situation is modeled through the second term
of (11). In this term,
is a floor square bracket function and
guarantees the second term of (11) to be binary. Assume bus
as a zero-injection bus. If members of the set
are all observable except one,
would equal to
1 and the remaining bus also became observable, otherwise the
second term of (11) would equal to 0.

2140

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 28, NO. 3, AUGUST 2013

Fig. 1. Interaction between learning automata and environment [24][26].

Since (10)(12) are dependent, in order to calculate observability function for the entire buses of a network, a pseudo-repetitive procedure is conducted as follows:
using (13):
Step 1) Initialize for

Variable structure learning automata (VSLA) are represented


by a quartet
, where
is the set of inputs,
is the set of actions,
is the set of actions probability vector and is learning algorithm. The learning algorithm is a recurrence relation which is used to modify the action probability vector. At each instant , the action probability
vector is updated by the linear learning algorithm given in (14),
if the selected action is rewarded by the random environment,
and it is updated as given in (15) if the taken action is penalized
[24][26]:
(14)

(13)

(15)

other
Step 2) Calculate (12) for the entire network buses.
Step 3) Update
using (10)(11).
Step 4) Repeat steps (2)(3) until
all buses.

for

III. CELLULAR LEARNING AUTOMATA


Cellular automata are dynamically discrete systems which
their behavior relies on their local relevance [23], [24]. Cellular
automata are composed of a regular array of cells and each follows a simple rule. While each cell can assume a state from
a finite set of states, they act together to improve themselves
and to express complicated behavioral patterns [25]. The time
goes discrete and the cells update their states synchronously according to a local rule. The present state of each cell depends on
the previous states of a set of cells, including the cell itself, and
constitutes its neighborhood [26]. A cell is called a neighboring
of another cell if they can affect each other in a period of time
[25].
Cellular automata can be reinforced by learning automata to
be used as a powerful mathematical model for various problems
[26]. Learning automata are adaptive decision-making units
which select their current action based on past experiences
earned from the environment. Their performances are improved by learning how to choose the best action from a finite
set of allowed actions through repeated interactions with their
environment [24].
Automata are guided to the optimal actions considering interactions between automata and the adaptive unit, which is
called random environment. Fig. 1 illustrates how a stochastic
automaton works in feedback connected with a random environment. Accordingly, an action is chosen at random based on a
probability distribution kept over the action-set. Then, the given
action is sent as the input to the random environment. The environment responds the listed actions proportional to reinforcement signals, and the action probability vector is updated considering feedback from the environment [24][26]. Prevalent
learning automata try to find the optimal actions, so that average
penalty received from the environment is minimized. The objective of a learning automaton can be changed by maximizing average reward received from the environment or other multi-objective approaches [25].

represents total number of


In the above equation,
actions for each automaton. If
, the recurrence (14)
and (15) are called linear reward-penalty
algorithm,
if
, the given equations are called linear rewardpenalty
, and finally if
they are called linear
reward-inaction
.
The overall procedure of the CLA can be described as follows: each cell contains a learning automaton which initially selects an action in random, from a finite set of its allowed actions.
Then, the selected actions are rewarded or penalized according
to the actions selected by their neighbors (local rules) and also
the new state of the CLA (an environmental rule). Afterwards,
the entire learning automata update their probability function
based on the amount of received rewards or penalties. In the next
step, the same process is repeated except that the probability of
choosing a particular action is increased more than the others for
each automaton. By repeating the algorithm for several times,
convergence to the optimal solution is achieved [24][26].
IV. SOLUTION APPROACH
CLA is a robust algorithm, established on a network structure, where its cells cooperate to achieve a unique aim [25],
[28]. Hence, it can easily be adapted to the transmission network and PMU placement problem. In CLA, state of each cell
depends on the state of neighboring cells. Similarly, in the PMU
placement problem, installing a PMU at a bus depends on the
state of neighboring buses. Therefore, a CLA based algorithm
is completely consistent with the PMU placement problem, as it
structurally emulates the PMU placement problem. Moreover,
implementing the proposed reliability model needs to create a
set of sub-graphs
which are similar to the neighboring
cells of each automaton. Therefore, the CLA based algorithm
reduces the computational burden while solving the PMU placement problem. Moreover, the CLA algorithm has been used for
solving various problems so far and satisfactory results were
reported [26][31]. According to the above discussion, in this
paper the PMU placement problem is modeled through CLA;
though it could also be solved using other evolutionary algorithms.
Each network bus is considered as a cell which reinforced
by an
automaton. The motivations behind employing

MAZHARI et al.: MULTI-OBJECTIVE PMU PLACEMENT METHOD CONSIDERING MEASUREMENT REDUNDANCY

2141

Fig. 2. Neighboring lines of a CLA cell and CLA local rules. (a) First CLA local rule. (b) Second CLA local rule. (c) Third CLA local rule.

the
automata are as follows: As it is mathematically
proven that the
automata are -optimal [24], [25],
enlisting proper learning rate guarantees the achievement of the
optimal solutions [32]. Moreover, the significant performance
of the learning algorithms illustrated in (14)(15) for the
automata is reported in several studies [26][33].
Each automaton contains two allowed actions: expressing a
PMU to either be installed in the cell or not. Environmental
rule is defined based on (1) and all members of the set
are assumed as neighbors of the bus (CLA cell) . Local rules
which show local improvement of neighboring cells is achieved
by applying (1) to the set of neighboring buses and neighboring
lines which are defined as follows:
(16)
Neighboring lines for a simple case is shown in Fig. 2. Accordingly, in any iteration, CLA cells select proper actions based
on their actions probability vector. Then, the observability function is determined for all buses using (10)(12) in a pseudorepetitive procedure. Afterwards, objective function of the PMU
placement problem is calculated using (1). If the current CLA
status improved the objective function, the response sent from
the environment is positive and all the selected actions are primarily rewarded, otherwise penalized. Then, CLA local rules
are investigated for each cell using (1) and (16). According to
the local rules, the CLA cell is again rewarded or penalized. By
repeating the mentioned process for several times, the algorithm
converges to the optimal solution. In this paper, the CLA local
rules are defined as follows:
If the present action of a cell differs to that of the previous
one:
If the cell action differs to that of the previous actions of
some of its neighbors, and if at least one action among its
neighbors equals to that of the cell previous action, the
reward or penalty is based on local rule. Thus, if local
rule is improved, the cell action is rewarded, otherwise
penalized. This is schematically shown in Fig. 2(a).
If the cell action equals to that of all previous actions
of its neighbors, and if the number of neighbors which
have changed their action is more than those that have
not changed, the reward or penalty is based on local rule.
This is schematically shown in Fig. 2(b).
If the present action of a cell equals to the previous one:

If the cell action equals to the previous actions of some


of its neighbors, and if none have changed their actions,
the reward or penalty is based on local rule. This is
schematically shown in Fig. 2(c).
If a cell action has not received any reward or penalty by applying the local rules, it does not receive any secondarily reward
or penalty.
It must be noted that CLA local rules change in a wide range
and could be changed according to the problem statement. Nevertheless, the aim of these rules is to prepare a suitable way to
give reward or penalty to an action in which the cell be able
to find the optimal action through repeated interactions with its
neighbors and the environment. Accordingly, these rules are not
limited to the above mentioned ones, and new rules can be offered and tested. A set of CLA local rules and their applications
are investigated in [25][31]. Moreover, [33] presents a set of
publications regarding cellular automata, learning automata and
the CLA.
V. NUMERICAL STUDIES AND RESULTS
Aimed at solving the problem of PMU placement by the proposed method, software is developed within a user-friendly environment. The user can see the results by providing the GIS
ready map (or presenting complete network information) and
entering technical and economic data. To evaluate the performance of the proposed method, the problem is solved for IEEE
standard test systems [34] as well as for the Iranian 230- and
400-kV transmission networks [35] and compared to that of
previous research. The technical specifications of the computer
used for simulations are Centrino 1.8-GHz CPU with a 1 GB of
RAM.
A. First Scenario
In this scenario the solution is conducted for the IEEE standard test systems, and the obtained results are compared to those
of previous research.
As it can be seen in Table I, the proposed method could
achieve the best existing solution from the minimization of
PMU numbers viewpoint. Moreover, as shown in Table II, in
comparison with the previous methods, the proposed solution
has improved the measurement redundancy for the 30-bus test
system (19%). Accordingly, the second term of (1) which is
added to the objective function to find the maximum measurement redundancy has improved the obtained solutions. The

2142

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 28, NO. 3, AUGUST 2013

TABLE I
OBTAINED RESULTS OF THE PMU PLACEMENT FOR IEEE STANDARD TEST SYSTEMS

TABLE II
COMPARISON OF OBTAINED RESULTS BY SEVERAL METHODS FOR IEEE STANDARD TEST SYSTEMS

TABLE III
PMU PLACEMENT IN PRESENCE OF NON-SYNCHRONOUS CONVENTIONAL MEASUREMENTS FOR IEEE 57-BUS TEST SYSTEM

buses. Besides, the results of the IEEE 118-bus system reported


in [15] are also infeasible due to unobservability of buses 33 and
35. It should be noted that [7] has presented two results for the
118-bus test system including and excluding a new observability
rule proposed in [36]. Since this paper deals with the prevalent
observability rules, the corresponding results are reported from
[7].
Fig. 3. Part of the IEEE 118-bus test system.

execution times for test systems are 1.61, 19.19, 43.45, 337.09
and 1691.27 seconds, respectively.
The obtained results of [14] for 118-bus test system are better
than this paper proposed solutions in terms of minimization of
number of PMUs. Moreover, the results obtained in [14] are infeasible, as they cannot meet the topological observability of
the system due to buses 63 and 64 [15]. This part of the network
is shown in Fig. 3. In this figure, according to the results presented by [14], no PMU is installed in buses 59, 61, 63, 64 and
65; yet buses 59, 61 and 65 are observable due to the adjacent
PMUs or effects of zero-injection buses. Moreover, according to
[34], buses 63 and 64 are zero-injection buses. Since buses coinciding to an unobservable zero-injection bus (buses 63 or 64) are
not all observable, the zero-injection bus cannot be identified
as observable by applying the KCL at zero-injection bus [14].
Hence, the IEEE 118-bus test system cannot be completely observed using these PMUs formation. These authors have found
that the modeling proposed by [14] for zero-injection buses may
achieve infeasible solutions when two or more zero-injection
buses are connected together [15]. Accordingly, buses 13, 14
and 32 of the 39-bus test system reported in [14] are not topologically observable, while PMU can be installed in zero-injection

B. Second Scenario
In this scenario, effects of conventional measurements are investigated in various situations for the IEEE 57-bus test system
and obtained results are presented in Table III. As shown in
this table, while power injection measurements could reduce
the number of PMUs, average measurements redundancy is increased (3%). Besides, combination of power injection and current flow measurements has produced better solutions (10%).
Moreover, while effects of voltage measurements in a number
of PMUs are negligible for this case, they could properly improve measurement redundancies (9.5%). According to these
results, effects of non-synchronous conventional measurements
in PMU placement are non-negligible, but change according to
conventional measurement locations and situation of the zeroinjection buses.
The binary search algorithm that is proposed in [9] is implemented to solve the PMU placement problem in presence
of non-synchronous measurements. Obtained results of the
implemented direct search algorithm for the IEEE 57-bus test
systemwhich is achieved after approximately two hours of
computationare completely consistent with results presented
in Table III. Accordingly, the proposed observability function results appropriate solutions in presence of conventional
non-synchronous measurements.

MAZHARI et al.: MULTI-OBJECTIVE PMU PLACEMENT METHOD CONSIDERING MEASUREMENT REDUNDANCY

2143

TABLE IV
PMU PLACEMENT UNDER SINGLE CONTINGENCY FOR IEEE 57-BUS TEST SYSTEM

TABLE V
PMU PLACEMENT CONSIDERING WORTH OF STATE ESTIMATION UNDER CONTINGENCIES FOR IEEE 57-BUS TEST SYSTEM

C. Third Scenario
In this scenario, PMU placement is conducted under single
contingencies including line outages and loss of PMUs [14]. In
order to meet the full observability under single line outages,
no line must be unobservable when a line is out. Hence, the set
of
in (7) should be null for the entire line outages, which
leads to
. To do so, the monetary factor of observability
value of each line is set equal to a very large number
.
By minimizing the objective function (1), the optimization algorithm tends to choose the layouts in which
to decrease the fitness level. In addition, since the algorithm does not
tackle with the PMU outages, the
is set equal to zero. Similarly, to meet the full observability under single loss of PMUs,
is set to zero and
. The mentioned model is conducted on the IEEE 57-bus test system and obtained results are
compared to those of [14]. As it can be seen in Table IV, the
total number of required PMUs under single contingencies is
reported, beside average measurement redundancies. As shown
in the first row of this table that the CLA solution requires 19
PMUs under single line outages which is the same as the result reported in [14] from the number of PMUs point of view.
However, the proposed algorithm has improved the measurement redundancy by 8% which demonstrates the performance
of the second term of (1). Moreover, the second row of Table IV
compares the CLA results with those of [14] under single loss
of PMUs. According to these results, while the CLA proposed a
solution with 25 PMUs, the associated result in [14] requires 26
PMUs. In addition, the CLA solution has better measurement redundancy (5%). Hence, the proposed model has achieved better
solution for single loss of PMUs which shows the quality of the
CLA algorithm form the optimization viewpoint. Moreover, the
comparisons presented in Tables II and IV illustrate that the proposed worth of state estimation formulated in (6) and (7) results
proper solutions under boundary conditions.
In order to investigate the PMU placement under the proposed contingency model, the problem is conducted for the
same test case and obtained results are presented in Table V.
For this reason, rate of line outages and rate of PMU losses are
set equal to 0.08 (fr/year/km) and 20 (fr/year), respectively.
In addition, average fault clearance time for lines and PMUs

plus PMU costs are set equal to 5 , 2 and 5000 (US$),


respectively. Disturbance rate of transmission lines is assumed
twice of the failure rate.
As it can be seen in Tables IV and V, while total number of
PMUs in normal condition is 11, 19 PMUs are needed to cover
the single line outage condition [14]. However, the proposed
contingency approach needs 12 PMUs for line outage condition. According to this table, almost similar results are achieved
for other contingences including single PMU losses and simultaneous line and PMU outages. In addition, while disturbance
worth is doubled, total required PMUs are increased almost
29%. Based on this Table, since the PMUs are located in almost
half of network buses, it seems that excessive PMUs are needed
through
contingency PMU placement. On the other hand,
the proposed contingency approach has proposed more reasonable solutions in which worth of network disturbances is participated in PMU placement during contingencies.
D. Fourth Scenario
In this scenario, the PMU placement problem is investigated
for real Iranian 230- and 400-kV transmission network. The Iranian transmission grids are highly interconnected and managed
by 16 regional utilities owned by a holder state owned company
(Tavanir).
The 230- and 400-kV transmission lines have roughly lengths
of 24 000 and 15 000 km, respectively, all of them monitored
by Tavanir. The under-study network which is shown in Fig. 4,
contains 242 buses and 364 corridors. While contingency data
are set the same as third scenario, complete data listing can be
found in [35].
Obtained results are shown in Tables VI and VII. As it can be
seen, while the under-study network needs 71 PMUs in normal
condition, required PMUs are increased almost 35% to cover
the contingencies. In addition, sensitivity analysis presented in
Table VIII shows that in the most critical condition in which disturbances worth are quadrupled comparing to normal condition,
almost 58% of network buses need PMU. On contrary, while
disturbances worth are set as normal condition, required PMUs
are increased almost 35% comparing to normal operation. This
shows that exact values of disturbance worth are required for

2144

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 28, NO. 3, AUGUST 2013

TABLE VI
OBTAINED RESULTS OF PMU PLACEMENT FOR THE IRANIAN 230- AND 400- KV TRANSMISSION NETWORK

TABLE VII
OBTAINED RESULTS OF PMU PLACEMENT FOR THE IRANIAN 230- AND 400-KV TRANSMISSION NETWORK UNDER VARIOUS WORTH OF CONTINGENCIES

Fig. 5. Trajectory of best solution for the BPSO, CLA, and GA.

Fig. 4. Iranian 230- and 400-kV transmission network.

TABLE VIII
RESULTS OF THE PROPOSED PMU PLACEMENT MODEL FOR THE IRANIAN
TRANSMISSION NETWORK USING EVOLUTIONARY ALGORITHMS

(BPSO) algorithm proposed in [32] as well as the GA introduced


in [5]. Moreover, as an initial solution, PMU is selected for the
entire network buses. Each algorithm is conducted for 10 independent runs and obtained results are shown in Table VIII. As it
can be seen in this table, average results of the CLA algorithm
are better than those of the BPSO and GA. In addition, trajectory of the best solution for each algorithm is shown in Fig. 5.
According to these comparisons, optimal solution of the CLA is
better than the mentioned evolutionary algorithms and has been
achieved in less execution time.
E. Fifth Scenario

optimal planning. Hence, the authors propose that PMU placement be studied in two phases for practical networks: In the first
phase, PMU placement by the proposed approach be investigated using
. Then, PMUs be installed throughout the
network considering results of the first phase. Afterwards, network disturbances are stored during state estimation for a period
of time. In the second phase, the problem would be conducted by
the proposed approach using actual worth of disturbances based
on obtained information of real operational periods. Final output
of the mentioned phases leads to a PMUs placement which has
optimal performance in both normal and contingency situations.
In order to investigate optimality of the CLA solutions for the
Iranian transmission grid, the proposed objective function of (1)
is minimized by means of a binary particle swarm optimization

In learning automata-based algorithms, choosing the proper


learning rate is the most challenging issue. However, from
the learning automata theory, it is concluded that solution
optimality is inversely proportional to the learning rate [24],
[25]. Although, it is common to assume
in
this sort of problems [24][28], for confidence, IEEE standard
test systems which are investigated in [9] are assumed. Then,
the percentage of the converged runs (PCR) to the expected
results (or better than them) for different values of learning
rate is calculated for 100 independent runs. Table IX shows
the obtained results. In this table, (LR) is the learning rate,
(PRT) is running time of the algorithm which is divided by the
maximum running time of the algorithm when
and
(PSR) is the algorithm sampling rate which is divided by the
maximum sampling rate of the algorithm when
. It
can be observed that the proposed algorithm always converges
to the minimal solution if the learning rate set to
.

MAZHARI et al.: MULTI-OBJECTIVE PMU PLACEMENT METHOD CONSIDERING MEASUREMENT REDUNDANCY

2145

TABLE IX
PERFORMANCE EVALUATION OF THE PROPOSED ALGORITHM FOR DIFFERENT LEARNING RATES

TABLE X
SENSITIVITY ANALYSIS OF THE PROPOSED CLA LOCAL RULES FOR THE IRANIAN 230- AND 400-KV TRANSMISSION NETWORK

Since the proposed algorithm aims at finding the minimum


solution, in this paper the learning rate is set to
.
In addition, effects of the proposed CLA local rules are investigated in this scenario and obtained results are shown in
Table X. For this reason, the fourth scenario is repeated for
each Test of Table X for 10 independent runs. As it can be
seen in this table, while probability of achieving optimal solutions is increased, computational burden is almost doubled by
applying these rules to the optimization algorithm. Although it
is hard to highlight exact effects of each CLA local rule based
on these slight tests, it can be concluded that local rules have
properly improved rewarding strategy so that better solutions
are achieved in Test 2-Test 8 rather than Test 1. It should
be noted that, although in Test 8 computer executive operations are increased, the solution time is decreased since learning
automata could find their proper actions sooner.

VI. CONCLUSION
The multi-objective PMU placement problem was investigated as an optimization problem using CLA. The prevalent
PMU placement problem formulation is extended and new concepts of measurement redundancy and observability value in
contingencies were defined and added to the model. A new
observability function was proposed in which effects of conventional measurements and zero-injection buses were considered during a pseudo-repetitive process. Furthermore, effects of
conventional measurements on PMU placement were studied
in various situations. Finally, the problem was conducted for
the IEEE standard test systems as well as the Iranian 230- and
400-kV transmission network and compared to those of the previous research. The obtained results revealed that the proposed
approach can be used as an effective tool for optimal PMU
placement of a practical network under contingencies.
Further research may be conducted on preparing an integer
linear programming based formulation for the proposed objective function, and investigating the effects of installing redundant PMUs at a bus on the worth of observability in contingencies.

ACKNOWLEDGMENT
The authors would like to thank Prof. M. R. Meybodi and
Dr. J. Akbari-Torkestani for their valuable help, comments, and
suggestions.
The insightful and constructive comments of the editors and
the anonymous refereesthat greatly improved the content and
presentation of this manuscriptare thankfully acknowledged.
REFERENCES
[1] A. G. Phadke, Synchronized phasor measurements in power systems,
IEEE Comput. Appl. Power, vol. 6, no. 2, pp. 1015, Apr. 1993.
[2] A. Abur and A. G. Esposito, Power System State Estimation: Theory
and Implementation. New York: Marcel Dekker, 2004.
[3] A. S. Debs, Modern Power Systems Control and Operation. London,
U.K.: Kluwer, 1998.
[4] R. E. Wilson, Satellite synchronized measurements confirm power
equation, IEEE Potentials, vol. 13, no. 2, pp. 2628, Apr. 1994.
[5] F. J. Marn, F. Garca-Lagos, G. Joya, and F. Sandoval, Genetic algorithms for optimal placement of phasor measurement units in electric
networks, Electron. Lett., vol. 39, no. 19, pp. 14031405, Sep. 2003.
[6] T. L. Baldwin, L. Mili, M. B. Boisen, and R. Adapa, Power system observability with minimal phasor measurement placement, IEEE Trans.
Power Syst., vol. 8, no. 2, pp. 707715, May 1993.
[7] F. Aminifar, C. Lucas, A. Khodaei, and M. Fotuhi-Firuzabad, Optimal
placement of phasor measurement units using immunity genetic algorithm, IEEE Trans. Power Del., vol. 24, no. 3, pp. 10141020, Jul.
2009.
[8] B. Milosevic and M. Begovic, Nondominated sorting genetic algorithm for optimal phasor measurement placement, IEEE Trans. Power
Syst., vol. 18, no. 1, pp. 6975, Feb. 2003.
[9] S. Chakrabarti and E. Kyriakides, Optimal placement of phasor measurement units for power system observability, IEEE Trans. Power
Syst., vol. 23, no. 3, pp. 14331440, Aug. 2008.
[10] S. Chakrabarti, E. Kyriakides, and D. G. Eliades, Placement of synchronized measurements for power system observability, IEEE Trans.
Power Del., vol. 24, no. 1, pp. 1219, Jan. 2009.
[11] J. B. Gou, Generalized integer linear programming formulation for
optimal PMU placement, IEEE Trans. Power Syst., vol. 23, no. 3, pp.
10991104, Aug. 2008.
[12] D. Dua, S. Dambhare, R. K. Gajbhiye, and S. A. Soman, Optimal
multistage scheduling of PMU placement: An ILP approach, IEEE
Trans. Power Del., vol. 23, no. 4, pp. 18121820, Oct. 2006.
[13] N. H. Abbasy and H. M. Ismail, A unified approach for the optimal
PMU location for power system state estimation, IEEE Trans. Power
Syst., vol. 24, no. 2, pp. 806813, May 2009.
[14] F. Aminifar, A. Khodaei, M. Fotuhi-Firuzabad, and M. Shahideh-pour,
Contingency-constrained PMU placement in power networks, IEEE
Trans. Power Syst., vol. 25, no. 1, pp. 516523, Feb. 2010.
[15] S. M. Mahaei and M. Tarafdar-Hagh, Minimizing the number of
PMUs and their optimal placement in power systems, Elect. Power
Syst. Res., vol. 83, pp. 6672, 2012.

2146

[16] R. F. Nuqui and A. G. Phadke, Phasor measurement unit placement


techniques for complete and incomplete observability, IEEE Trans.
Power Del., vol. 20, no. 4, pp. 23812388, Oct. 2005.
[17] J. Chen and A. Abur, Placement of PMUs to enable bad data detection in state estimation, IEEE Trans. Power Syst., vol. 21, no. 4, pp.
16081615, Nov. 2006.
[18] B. Xu and A. Abur, Observability analysis and measurement placement for system with PMUs, in Proc. IEEE Power Syst. Conf. Expo.,
Oct. 2004, vol. 2, pp. 943946.
[19] S. Almeida, R. Pestana, and F. M. Barbosa, Severe contingencies
analysis in Portuguese transmission system, in Proc. 41st Universities Power Engineering Int. Conf., 2006, pp. 462466.
[20] Y. Sun and T. J. Overbye, Visualizations for power system contingency analysis data, IEEE Trans. Power Syst., vol. 19, no. 4, pp.
18591866, Nov. 2004.
[21] B. L. Silverstein and D. M. Porter, Contingency ranking for bulk
system reliability criterion, IEEE Trans. Power Syst., vol. 7, no. 3,
pp. 956964, Aug. 1992.
[22] F. Aminifar, M. Fotuhi-Firuzabad, M. Shahidehpour, and A. Khodaei,
Observability enhancement by optimal PMU placement considering
random power system outages, Energy Syst., no. 2, pp. 4565, 2011.
[23] E. M. Carreno, R. M. Rochaand, and A. Padilha-Feltrin, A cellular
automaton approach to spatial electric load forecasting, IEEE Trans.
Power Syst., vol. 26, no. 2, pp. 532540, May 2011.
[24] K. S. Narendra and K. S. Thathachar, Learning Automata: An Introduction. New York: Printice-Hall, 1989.
[25] M. R. Meybodi and H. Beigy, A mathematical framework for cellular
learning automata, J. Adv. Complex Syst., vol. 7, no. 3, pp. 295320,
2004.
[26] J. A. Torkestani and M. R. Meybodi, A cellular learning automatabased algorithm for solving the vertex coloring problem, Expert Syst.
Appl., vol. 38, pp. 92379247, 2011.
[27] A. Enami-Eraghi, J. Akbari-Torkestani, M. R. Meybodi, and A. H.
Fathy-Navid, Cellular learning automata-based channel assignment
algorithms for wireless mobile ad hoc networks, in Proc. Int. Conf.
Machine Learning and Computing, 2011, pp. 173177.
[28] M. Saheb-Zamani, F. Mehdipour, and M. R. Meybodi, Implementation of cellular learning automata on reconfigurable computing systems, in Proc. Canadian Conf. Electrical and Computer Engineering,
2003, vol. 2, pp. 11391142.
[29] H. Beigy and M. R. Meybodi, Cellular learning automata with multiple learning automata in each cell and its applications, IEEE Trans.
Syst., Man, Cybern., vol. 40, no. 1, pp. 5465, 2010.
[30] H. Beigy and M. R. Meybodi, A self-organizing channel assignment
algorithm: A cellular learning automata approach, Intell. Data Eng.
Autom. Learn., vol. 26, no. 90, pp. 119126, 2003.
[31] R. Rastegar and M. R. Meybodi, A new evolutionary computing
model based on cellular learning automata, in Proc. IEEE Conf.
Cybernetics and Intelligent Systems, 2004, vol. 1, pp. 433438.
[32] C. nsal, Intelligent navigation of autonomous vehicles in an automated highway system: Learning methods and interacting vehicles approach, Ph.D. dissertation, Virginia Polytechnic Institute and State
University, Blacksburg, Virginia, 1996.
[33] [Online]. Available: http://ce.aut.ac.ir/~meybodi/index.htm#publication.
[34] R. Christie, Power System Test Archive, Aug. 1999. [Online]. Available: http://www.ee.washington.edu/research/pstca.
[35] [Online]. Available: http://mazhariac.persiangig.com/.
[36] M. Hajian, A. M. Ranjbar, T. Amraee, and B. Mozafari, Optimal
placement of PMUs to maintain network observability using a modified BPSO algorithm, Int. J. Elect. Power Energy Syst., vol. 33, no.
1, pp. 2834, 2011.

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 28, NO. 3, AUGUST 2013

Seyed Mahdi Mazhari (S12) received the B.S.


(Hon.) degree from the University of Birjand, Birjand, Iran, in 2010 and the M.S. (Hon.) degree from
the University of Tehran, Tehran, Iran, 2012, both in
electrical engineering.
Currently, he is working as a research associate at
the research Center of power system operation and
planning studies, University of Tehran. His research
interest includes planning of the electric power distribution and transmission systems, power system operation, and artificial intelligence applications to power
system optimization problems.

Hassan Monsef received the B.Sc. degree from


Sharif University of Technology, Tehran, Iran,
in 1986, the M.Sc. degree with honor from the
University of Tehran in 1989, and the Ph.D. degree
from Sharif University of Technology in 1996, all in
power engineering.
He has been with the University College of Engineering, School of Electrical and Computer Engineering, University of Tehran, since 1996, where he
currently is an Associate Professor. His research interests are power system operation under deregulation, reliability of power system, power systems economics, renewable energy
systems, and its integration in smart grid.

Hamid Lesani received the M.S. degree in power


engineering from the University of Tehran, Tehran,
Iran, in 1975 and the Ph.D. degree in electrical engineering from the University of Dundee, Dundee,
U.K., in 1987.
Then, he joined the Department of Electrical and
Computer Engineering, University of Tehran, where
he currently serves as a Professor at the Center of Excellence for Control and Intelligent Processing. His
teaching and research interests is focused on design
and modeling of electrical machines and power systems.
Prof. Lesani is a member of the IEEE PES and the IEEE Iran Section.

Alireza Fereidunian (M08) received the M.Sc. and


Ph.D. degrees from the University of Tehran, Tehran,
Iran, in 1997 and 2009, respectively.
He is an Assistant Professor at the K. N. Toosi
University of Technology, Tehran. He is a Post-Doctoral Research Associate at the University of Tehran.
His research interests include Smart Grid, energy
distribution systems, and application of IT and AI
in power systems. Moreover, he works in complex
systems, systems reliability and human-automation
interactions areas, where he has invented the Adaptive Autonomy Expert System (AAES).
Dr. Fereidunian is an IEEE-SMC-HCI TC member and a member of INCOSE
(as INCOSE Iran point of contact).

You might also like