Вы находитесь на странице: 1из 10


2, MARCH 2015


Multi-Objective Optimal Power Flow Considering

Transient Stability Based on Parallel NSGA-II
Cheng-Jin Ye and Min-Xiang Huang

AbstractStability is an important constraint in power system

operation and the transient stability constrained optimal power
flow (OPF) has always received considerable attention in recent
years. In this paper, the defects of the existing models and algorithms around this topic are firstly analyzed, on the basis of
which, a multi-objective optimization method is proposed. The
basic idea of the proposed method is to model transient stability
as an objective function rather than an inequality constraint and
consider classic transient stability constrained OPF (TSCOPF) as
a tradeoff procedure using Pareto ideology. Second, a master-slave
parallel elitist non-dominated sorting genetic algorithm II is used
to solve the proposed multi-objective optimization problem, the
parallel algorithm shows an excellent acceleration effect and
provides a set of Pareto optimal solutions for decision makers to
select. An innovative weight assigning technique based on fuzzy
membership variance is also introduced for a more scientific and
objective optimal solution decision. Case study results demonstrate the proposed multi-objective method has many advantages,
compared with traditional TSCOPF methods.

Fuel cost coefficients of the th generator.

Rotor angle vector.
Angular speed vector.
Rotor angle of th generator at integration step
after the th expected fault.
Center of inertia of system rotor angles at
integration step after the th expected fault.
Moment of inertia of the th generator.
Number of Pareto optimal solutions.
Number of objective functions.
The th objective function value of the th
Pareto optimal solution.
Maximum of the th objective function.

Index TermsMembership, multi-objective optimization,

NSGA-II, OPF, parallel computing, power system, transient

Minimum of the th objective function.

Total active load.
Total transmission loss.


Weight of the th objective function.

Fuzzy membership of the th objective
function of the th Pareto optimal solution.

Network variable/variable vector.

System state variable/variable vector.

Execution time of NSGA-II on

Control variable/variable vector.


Execution time of NSGA-II on 1 processor.

Difference form of system state variable or

state variable value at time or

Number of parallel processors used.

Step counter.

Population size of PNSGA-II.

Integration step length.

Current iteration number of PNSGA-II.

Simulation time length.

Maximum iteration number of PNSGA-II.

Active output of the th generator.

Number of decision variables.

Number of generators.
Number of buses.
Number of branches.
Number of expected faults.

Manuscript received January 20, 2014; revised April 18, 2014 and June 06,
2014; accepted July 10, 2014. Date of publication July 25, 2014; date of current
version February 17, 2015. Paper no. TPWRS-01624-2013.
The authors are with the College of Electrical Engineering, Hangzhou
310027, China (e-mail: yechenjing@zju.edu.cn; Huangmx@zju.edu.cn).
Digital Object Identifier 10.1109/TPWRS.2014.2339352


PTIMAL power flow (OPF) has been a classic issue in

power system since it was proposed in the early 1960s
[1], [2]. In traditional OPF equations, stability constraint is usually excluded; however, the system operating at the point suggested by OPF may not be able to maintain transient stable
when subject to credible contingencies. Due to the rapid increase of electricity demand and the electricity market deregulation, power systems tend to operate closer to stability boundaries. As a consequence, transient stability has been one of the
main concerns in power system operation and the transient stability constrained OPF (TSCOPF) was brought into being.

0885-8950 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.


Mathematically, TSCOPF is commonly modeled as a complicated nonlinear programming (NLP) problem including
massive constraints of differential algebraic equations. The two
main approaches to study TSCOPF are simulation based on
numerical discretization [3] and constraint transformation [4].
Recently, various methods have also been proposed, including
trajectory sensitivity method [5], semi-infinite programming
[6], transient energy function method [7], implicit enumeration method [8], and single-machine equivalent method [9].
Among these approaches, a combination of the numerical
discretization with the interior point method (IPM) has been
considered as a mainstream. One important extension of IPM
is the so-called reduced-space IPM (RIPM) [10]. RIPM has
shown a great improvement for solving numerical-discretization-based TSCOPF problem. However, dimensionality curse
still exists for data-intensive TSCOPF. And IPM naturally
relies on convexity to obtain the global optimum. However,
TSCOPF in realistic modern power system is generally indeed
non-convex, so researchers are forced to adopt a set of hypothesis and simplifications, otherwise they may suffer a lot
from unsolvability or non-convergence problems or get local
optimum. To overcome this problem, intelligence algorithms
(IA) such as differential evolution algorithm (DE) [11] and
particle swarm optimization (PSO) [12] were introduced to
enhance robustness. The competition between IA and IPM is
one of the most active factors in TSCOPF research and is also
one of the main lines of this paper.
Though, great progress around TSCOPF has been reported
in the existing literatures mentioned above, the mathematical
models are always limited to single-objective optimization
[3][12]. Specifically, fuel cost is modeled as the sole objective
function, transient stability and voltage or branch load flow are
formulated as constraints. However, it must be noted that certain differences do exist between transient stability and static
security constraint. Static security constraints are rigid and
must be always respected during the power system operation;
while, contingencies are not bound to happen, sometimes, to
obtain a lower fuel or operation cost, some dispatch centers
allow generators to operate under a certain degree of instability
in a certain period of time. Therefore, modeling stability as an
inequality constraint is not in accordance with the fact that OPF
is a tradeoff procedure among the three factors of cost, security
and stability; while, until recently, relevant multi-objective
optimization research is inadequate and not systematic, hence
multi-objective modification of traditional TSCOPF problem is
of significant theoretical and practical value. The first motivation of this paper is to adopt Pareto optimal ideology to redefine
TSCOPF problem and propose a multi-objective OPF (MOPF)
model considering transient stability. Inspired by the satisfactory application results of DE and PSO in solving TSCOPF
[11], [12], the Non-dominated Sorting Genetic Algorithm II
(NSGA-II) is introduced to solve the proposed multi-objective
optimization problem, which has been demonstrated to be effective in research areas such as optimal distributed generation
(DG) allocation [13] and transmission expansion planning [14].
NSGA-II provides a set of Pareto optimal solutions for decision
makes to select. This is the most important and significant
advantage of the proposed method. In order to select unbiased
optimal solution more scientifically, a weight method based on
fuzzy membership variance is also recommended.


Due to a large number of SBSI (step-by-step integration), the

proposed MOPF problem is extremely time- consuming [3], especially for large-scale power systems and multiple contingencies. The second motivation of this paper is to adopt parallel
computing technology to recompose NSGA-II program structure and shorten the execution time. Simulation results show
that the proposed algorithm has an excellent acceleration effect
to solve data-intensive problems.
The most common problem of transient stability is to study
the post-fault trajectory stability problem. With a given system
operating point, it can be expressed as a classic Cauchy problem:
represents the dynamic behavior of the
system. is the system state variable value at the fault clearing
time. Assumption is made that
is a stable equilibrium point
of the above problem, a stable region denoted as
[15]. The basic condition to maintain stability of the
above system is:
is subjected to the following two
variables: a) power flow control variable including generator
power output, generator terminal voltage and transformer taps,
etc. b) network variable such as nodal voltages and branch
. Network variflows. It can be expressed as:
ables respect the power flow equation
problem (1) is also associated with control variable , and it can
. With a given , there is a corresponding
be written as
stable equilibrium point and a stable region . To highlight
the correlation of stable region and control variables, the stable
region is denoted by
. Thus, TSCOPF can be expressed as
follows [16]:
In the above TSCOPF model, (2.1) is the objective function
(fuel cost or the modification amount of scheduled contract
power generation in deregulated power market); (2.3) is
the static security constraints such as thermal constraints of
branches and nodal voltage constraints; (2.4) is the constraints
of generator power outputs or voltages and the constraints of
transformer taps, etc.
All kinds of Lyapunov direct methods such as transient
energy function method and the extended equal area criterion
method are unable to give enough robustness and calculation
speed [17], [18]. On the other hand, engineers have long been
accustomed to use simulation method to judge the transient
stability. If the simulation time is long enough, simulation
method is able to guarantee sufficient accuracy. And if the


so-called classical model is used, the simulation method is also

fast enough [19].
According to the basic idea of the simulation method reported
in [3], by application of the trapezoidal integration principle
on (2.6), TSCOPF can be transformed into a general nonlinear
programming problem. Specifically, the infinite time constraint
(2.7) is transformed into the following finite time algebraic constraint:
If is long enough, as long as the rotor angle is limited (such
as the maximum relative rotor angle is less than 150 ), the transient stability of the system can be guaranteed. (2.1)(2.6) and
(3) actually is an optimal parameter selection under a fixed terminal in optimal control theory [20]. Equation (2.6) can be converted into the following difference equation:


is the objective function quantifying transient stability,

and can be of various modeling approaches. Here, we only consider the rotor angle stability and adopt the following penalty
is the sum of punitive amount for rotor
As can be seen,
angle deviations exceeding the threshold after the occurrence
of the expected faults in a given duration. Penalty coefficient
in this paper. The position of COI is calculated as a
weighted average as follows [3]:


. Now, constraint (3) of limited
time and infinite dimension is completely converted into finite
dimensional algebraic constraint:
. And TSCOPF is
modeled as a conventional nonlinear programming problem as

Penalty operator

in (8) is expressed as follows:


In (8), is the maximum allowable rotor angle deviation, called

a threshold. is decided based on operational experience. Commonly, it is set to a value between 100 and 150 to maintain
sufficient stability margin [3], [11], [12].
is an objective function constituted by static power flow
variables such as transmission loss and voltage qualified rate.
Though this kind of function is of various forms, its modeling
and calculating is relatively simple, so
is not considered in
the following section of this paper.


The dispatch of power system is associated with a variety of
factors. In many cases, engineers concern more about how to
coordinate cost, security and stability of power system rather
than getting a single rigid TSCOPF solution. In this way, transient stability is modeled as one of OPF optimization objectives, rather than a constraint. To meet practical application, additional power flow optimization objectives can also be incorporated into the model. The multi-objective OPF model can be
expressed as follows:
is always modeled as the total cost of fuel consumed by

A. Basic Procedure of NSGA-II

In 1989, Goldberg proposed a method to calculate the fitness
based on the concept of Pareto optimal, the level of non-inferior
solutions and the corresponding selection operator were used to
optimize populations towards the Pareto optimal direction [21].
Based on Goldbergs idea, a variety of multi-objective genetic
algorithms (GA) have been proposed, including Full Fusion GA
(FFGA), Niched Pareto GA (NPGA), Non-dominated Sorting
GA (NSGA), Simplex Parallel GA (SPGA) etc. Among these,
NSGA proposed by Srinivas was considered generally as a direct expression of Goldbergs idea and of the best performance
[22]. Based on NSGA, Deb further proposed NSGA-II with an
elitism strategy, which achieves a calculation complexity decrease and avoids the setting of sharing parameter [23].
NSGA-II has been demonstrated to be among the most efficient algorithms for multi-objective optimization on a number
of benchmark problems. Its detailed implementation procedure
can be found in [23], a brief description of NSGA-II procedure
is shown in Fig. 1.
The population is initialized as usual and then sorted based on
non-domination levels into fronts. The first front is a set of chromosomes being completely non-dominant or not dominated by
any other individuals in the current population, the second front
being dominated by the chromosomes in the first front only and



Fig. 2. Master-slave topology of PNSGA-II using

Fig. 1. Flow chart of NSGA-II procedure.

the front goes so on. In addition to front rank, a new parameter

called crowding distance is calculated for each chromosome.
The crowding distance is a measure of how close a chromosome
is to its neighbors. Large average crowding distance will result
in better diversity in the population. Parents are selected from
the population by using Binary Tournament Selection based on
is based on front rank and crowding distance
as follows [22], [23]:
a) The chromosome with the higher front rank value is
greater than the other regardless of crowding distance,
and is selected;
b) The chromosome with the larger crowding distance is
greater than the others located in the same front, and is
The selected parent population generates offsprings after the
operation of crossover and mutation operators. The population
with the current population and current offsprings is sorted
again based on non-domination and only the best individuals
are selected, so elitism is guaranteed, where is the population
size. The selection is also based on front rank and the crowding
distance on the last front.
B. Master-Slave Parallel NSGA-II on MPI
Due to a heavy task of power flow calculation and a large
number of numerical integrations, MOPF is inevitably a dataintensive and time-consuming problem. To solve this problem,
parallel computing is a worth trying method. There are a few
of libraries to apply parallel computation, such as the SharedMemory Processor (SMP) approach and the Message Passing
Interface (MPI) approach, etc. The MPI approach is used in this
paper as it has the following advantages [24]:
a) MPI is the most widely-supported communication library
for high-performance computing. Free implementations
of MPI are open access on the Internet [25][27].
b) With MPIs approximately 125 functions, engineers
need not to program common communication structures
themselves. A complete message passing program can be
written with only six basic functions: MPI_Init, MPI_Finalize, MPI_Comm_ rank, MPI_Comm_size, MPI_Send,
and MPI_Recv.
c) MPI has portability to almost all the major platforms.


In the following part, the master-slave parallel NSGA-II

processors is abbreviated to PNSGA-II
. Fig. 2
shows the procedure of PNSGA-II
, and the detailed steps
can be described as follows:
Step 1) Processor 0
initialize MPI environment
and set PNSGA-II parameters including: the size
of population ; crossover probability , muta; the maximum iteration number
tion probability
and the lower and upper limits of the corresponding decision variables, etc.
Step 2) Processor 0 reads network data and broadcasts public
data through message passing interface, such as load
data, bus data, generator data, branch data and transformer data, etc. Processor1
receive and
store public data.
Step 3) Processor 0 sets the value of iteration counter:
and randomly generates a swarm of chromosomes,
each chromosome stands for a candidate solution to
MOPF, to constitute the parent population . Let
be .
the size of
Step 4) Processor 0 executes
, here, is an
integer and the swarm
is divided into
evenly, then these parts are sent to the rest
processors by processor 0, each part has chromosomes. If cannot be divided exactly, let integer
be the quotient and remainder of
respectively. In this way, Processor1
receive chromosomes; Processor
Step 5) Processor 1
calculate the objective function values of their own chromosomes based on the
fast-decoupled load flow method [28] and trapezoidal integration method [29].
Step 6) Processor 0 gathers the objective function values of
all chromosomes from processor 1
. Processor 0 sorts the parent population
into fronts
according to non-domination levels and assigns
each chromosome a crowding distance.
Step 7) Processor 0 uses Binary Tournament Selection, Simulated Binary Crossover (SBX) [30] and Polynomial
Mutation (PM) [31] operators to create an offspring
with the size of . Then, processor
0 executes the elitism strategy:
is formed.
combined population with the size of



Step 8) Processor 0 sorts the population

into fronts as:
according to non-dominaare chosen from the
tion. Offspring population
chromosomes of
in the order of their front rank.
Thus, the best non-dominated front
is of the
is chosen next,
highest selection priority and
followed by
, and so on. Processor 0 continues
this procedure until no more fronts can be accommodated. Say that
is the last non- dominated front
beyond which no other front can be accommodated.
To choose exactly offspring chromosomes needed
, Processor 0 sorts the chromosomes of
to fill
the last front
using the crowded-comparison
in descending order. Until now, an offspring population with the size of is completely
Step 9) Processor 0 executes
. If the iteration
, go to step 4. Otherwise, go to
step 10.
Step 10) Processor 0 selects the best Pareto optimal solution
of the current iteration using a weighted method
based on membership variance from the first front of
. Then all processors exit MPI environment and
terminate the PNSGA -II program.

D. Optimal Solution Selection

Due to the complexity of MOPF optimization, as well as the
vagueness of human thoughts, the traditional optimal solution
selection strategy based on subjective preference is of significant limitation. This paper proposes an innovative selection
method based on fuzzy membership and variance weight.
Membership indicates the optimization degree of objective
functions. Here, some fuzzy theory techniques are applied. The
fuzzy membership of the th objective function of the th Pareto
optimal solution can be expressed as

Objective function values of the Pareto optimal solutions are
normalized to a real between 0 and 1 by (11).
In this paper, we quantify the quality of a Pareto optimal solution through its weighted sum of fuzzy memberships. A weight
assignment method based on the membership variance is introduced to reduce subjectivity:


In (12),
is the fuzzy membership
variance of the th objective function. The proposed weighting
method naturally respects the constraint:
and emphasizes the objective of a larger fluctuation of optimization results
more. Specifically, the larger the fuzzy membership variance of
, the larger the weight assigned to ; and vice versa.
The sum of weighted fuzzy memberships of each solution is
calculated and used as an unbiased optimal solution indicator:

C. Search Space Reduction Technique

Due to the randomness in initialization, crossover and mutation process, some invalid chromosomes will be generated
inevitably. In order to help improving efficiency, an improved
search space reduction technique is introduced.
The sum of the active generation power that has been dis,
patched to all generators excluding the slack generators is
the lower and upper active power limits of the slack generator
, respectively. Assumption is made that
is approximated as 10% of
and the following two conditions are premeditated [11]:

If a candidate chromosome respects condition 1 or 2, it should

be considered as a bad one and polynomial mutation operator
will be repeatedly performed on it, until the two conditions are
no longer respected. This can greatly reduce the time required
to deal with invalid solutions. Further, for some invalid candidates that are against static security or other constraints, we can
prompt them to be automatically abandoned by setting their objective functions to large values, for example:


is the weighted sum of fuzzy memberships of the th

is used as selection priority value of the th
optimal solution.
solution. The solution with the maximum value is regarded
as the best unbiased optimal solution.
A. Parameters Setting
The IEEE 39-bus system is used to show the feasibility of the
proposed method, the capacity benchmark is 100 MVA, basic
network parameters are available in [32] and [33], and the fuel
cost coefficients are available in [34].
The population size of PNSGA-II is 50. There are 18 control
variables including 9 generator active power outputs and 9 generator bus voltages. The upper and lower limits of control variables and some other variables are shown in Table I.
The following two three-phase to ground faults are considered: Fault A: a three-phase to ground fault took place at bus
29 and cleared by tripping line 2829 after 0.1 s; Fault B: a
three-phase to ground fault took place at bus 26 and cleared by
tripping line 2526 after 0.1 s.



Fig. 5. Proportion of Pareto solutions in PNSGA-II (16) population.

Fig. 3. Location of the PNSGA-II (16) population after 50 iterations of evolution.

Fig. 4. Location of the PNSGA-II (16) population after 100 iterations of evolution.

The proposed approach is implemented on a PC cluster containing 8 nodes all equipped with two Intel Xeon 2.33-GHz
quad-core CPUs and an 8-GB DDR3 memory and connected on
a Gigabit Ethernet network. So the maximum number of available parallel processors is 32, equal to the number of real CPU
cores of the clusters.
B. Pareto Optimal Solutions
Figs. 3 and 4 show the location of the population got by
PNSGA-II (16) after 50 and 100 generations of evolution, respectively. As can be seen, the population assembling effect has
been quite obvious with a non-dominated front appearing after
50 iterations of evolution. After 100 iterations, the locations of
the solutions tend to be identical; the entire population is located at a few aggregation points. To avoid randomness, we run
PNSGA-II (16) 10 times repeatedly. The locations of populations are roughly the same, indicating that the parallel algorithm
has a strong global optimization ability and significant convergence stability.
Fig. 5 shows the proportion of Pareto optimal solutions in the
population. With the increasing of iteration, the solutions tend
to be identical, but the proportion of non-dominated solutions in
the population increases gradually. After 100 iterations, inferior
solutions which are dominated by others have been eliminated
Table II shows the solution set consisted by 11 Pareto optimal
solutions after 100 generations of evolution. Determined by the
characteristics of the proposed model, the two objective functions of the problem are conflicting. To reduce the cost, the generators of lower cost coefficients are dispatched to output more


power, in this way, the overall stability of the system is weaken

inevitably, and vice versa. Better to meet one objective between
the cost and stability, inevitably at the expense of the other one.
Take solution 2 and 11 as examples, the fuel cost of solution 2 is
the lowest but the system is transient instable after fault B as the
divergent rotor angle curves indicated. While, solution 11 with
a higher fuel cost is transient stable after fault B, according to
Fig. 6.
C. Comparison With TSCOPF Methods
In Table III, solution 11 is the most attention-catching, whose
reaches 0, such solutions with
are considered to
be strictly transient stable; accordingly, the solutions with the
are called the most transient stable solutions.
Table III is mean Manhattan Distance calculated by
is the th decision variable value of the most tranwhere
sient stable solution got by IPM-based TSCOPF.
is the
th decision variable value of the most transient stable solution
got by other certain method.
In the following section, comparisons between the four
methods are carried out by the following four aspects:
a) Availability and robustness. As a mature IA method,
open- source implementations of NSGA-II are available
on the Internet, its programming and debugging is much
easier than IPM, which relies on convexity to obtain the
global optimum. Table IV gives the execution time of
PNSGA-II and IPM on IEEE-39, 300 and 678 test systems, symbol X represents failure to converge. As can


Fig. 6. Relative rotor angle curves to COI under solution 2 and 11 after fault.
(a) Relative rotor angle curves undersolution 11 after fault B. (b) Relative rotor
angle curves undersolution 2 after fault B.


be seen, the proposed method is always able to get a result, while IPM suffers a lot from problems of non-convergence in large scale power systems.
b) Execution time. If the maximum iteration is 100, average
time of the proposed method on 16 processors is 12.1% of
DE-based TSCOPF, 7.86% of PSO-based TSCOPF, and
1.47 times of RIPM-based TSCOPF. Given in Table IV,
when 32 processors are used, average time of PNSGA-II
drops to 23.18 s and is only 1.22 times of IPM-based
TSCOPF. So compared with serial IA-based TSCOPF, the
proposed parallel method has an obvious decrease in time
and lags slightly behind serial IPM-based TSCOPF when
enough number of parallel processors is used.
c) The optimization ability of fuel cost and transient stability.


Fuel cost: Shown in Table III, fuel cost of the

most transient stable solution got by the proposed
method is much higher than PSO-based TSCOPF,
and slightly higher than IPM and DE-based TSCOPF.
So the ability array of obtaining lower fuel cost in
descending order is: PSO, IPM, DE, PNSGA-II.
Transient stability: In every DE [11] iteration, transient stable individuals with lower cost are selected
as seeds to push the population to converge, so the
final solution got by DE-based TSCOPF is generally
transient stable. PSO [12] uses the weighted sum of
rotor angle deviations to COI and fuel cost as fitness
to evaluate the particles. So the PSO-based TSCOPF
solution is not inevitably strictly transient stable.
As Table III indicated, IPM only needs 11 iterations
to get a strictly transient stable solution, DE-based
TSCOPF gets strictly transient stable solution after
50-iter. PSO-based TSCOPF fails to get strictly transient stable solution even in 100-iter, the proposed
method gets such solutions after 100-iter. So the
ability array of obtaining strictly transient stable
solution in descending order is: IPM, DE, PNSGA-II,
PSO. Obviously, IPM is of the strongest ability to
obtain theoretical TSCOPF solution, so it is used as
a benchmark in Eq. 14 to evaluate the optimization
abilities of different methods.
Given in Table III, with the same iteration,
of the
proposed method is slightly higher than
of DE-based
TSCOPF and much lower than
of PSO-based
TSCOPF, which is consistent with the order mentioned
above. So the optimization ability of the proposed method
is far greater than PSO-based TSCOPF and similar to
DE-based TSCOPF.
Shown in Table IV.
of the proposed method on IEEE
300 test system is also below 5% when 100 iterations of
evolution are carried out.
on IEEE-39 system with different iterations of PNSGA- II are further shown in Fig. 7.
gradually decreases with the increasing number of iteration and reaches 0.75% at 200-iter, which is a negligible difference. So the proposed method is able to get a
theoretical TSCOPF solution if enough iterations of evolution are executed.
d) Diversity of the final results. The final result of NSGA-II
is not a single solution but a Pareto solution set including
the strictly transient stable solutions suggested by
TSCOPF (e.g., solution 11), this is considered as the most
significant characteristic and advantage of the proposed
method. Decision-makers can select their ideal schemes
from Table II according to their preferences. For example:
i) Decision maker A claims that the system must be
strictly transient stable, then solution 11 is the ideal
ii) Decision maker B prefers a low fuel cost, rather than
the transient stability, and then solution 2 with the
is the suggested one.
iii) Decision maker C has no significant preference;
he may choose the unbiased optimal solution with
the best balance between
based on the
weighted method. Fig. 8 shows the selection priority value of all the Pareto solutions calculated by




Fig. 9. Pareto fronts of different multi-objective algorithms.

Fig. 10. Speedup factors with different numbers of processors on three IEEE
test systems.

Fig. 7. Euclidean distance of the most transient stable solution of PNSGA-II

and IPM.

the origin of coordinates, indicating that the algorithms except

PNSGA-II are of obvious precocious phenomenon. The array of
optimization effect in descending order is: NSGA-II, SPEA-II,
GSA, and MPSO.
D. Analysis of Speedups, Efficiency, System Scale, and Faults
Number Considered
Three test systems are used to check the performance of the
proposed method in this section. Table IV summarizes these test
systems and gives the average time of 5 PNSGA-II runs when
different numbers of processors are used.
Obviously, the execution time decreases rapidly as the
number of processors grows. To further evaluate the parallel
performance of the master-slave PNSGA-II, the speedup factor
and the efficiency
are introduced. These two indices
are calculated as follows [24]:

Fig. 8. Selection priority of Pareto solutions.

Eq. 13. Apparently, solution 3 with the maximum
(0.83) is considered as the suggested solution.
This method succeeds in avoiding the blindness of
traditional decision method.
The proposed method provides decision-makers more options and a larger flexibility, in this way, one only need to do
re-selection instead of repeating the optimization procedure
when his preference changes.
To compare the optimization ability of different multi-objective optimization algorithms, we use PNSGA-II, MPSO (Multiobjective PSO) [34], SPEA-II (Improved Strength Pareto Evolutionary Algorithm) [35], and GSA (Gravitational Search Algorithm) [36] to optimize the above IEEE-39 case. The maximum iterations of the four algorithms are all 50. The results are
shown in Fig. 9. As can be seen, the four algorithms all get clear
Pareto fronts, but front got by PNSGA-II is relatively closer to

Figs. 10 and 11 present the satisfactory speedup factors and
efficiencies of the proposed method. The highest speedup factor
reaches 28.7 and the corresponding efficiency is 89.7%. So the
master-slave parallel modification of NSGA-II program structure is valid and necessary.
Shown in Fig. 10, with the growing number of processors,
the speedup factor increases but the increasing speed gradually
slows down and the efficiency always decreases. Parallelization
is conducive to take full advantage of computing ability and
benefits shortening execution time. However, the overhead for
communication system is unavoidable, when the number of processors increases, communication task accounts for a growing
share in the whole execution procedure, causing the dropping of
and increasing speed of



Fig. 11. Efficiencies with different numbers of processors on three IEEE test


Fig. 12. Speedup factors on IEEE-39 test systems with different numbers of
faults considered.

larger number of faults are considered. If PNSGA-II is run on

2 processors, the increasing multiple of execution time from 2
faults to 10 faults is
, while this multiple
drops to
when 32 processors are used. It
implies that if our cluster hardware meets condition and provides more parallel processors than 32, this algorithm can overcome the impact of the increasing faults number and handle
multi-contingency cases.

Shown in Table IV.

of the proposed method on IEEE
300 test system reaches 3.9% when 100 iterations of evolution
are carried out, indicating the proposed method is still effective in real-sized power systems, but its applicability drops due
to the rapid increase of execution time. It must be noted that a
larger systems speedup factor is bigger than a smaller systems
one when the number of processors are the same. And the relation between the speedup factors and the number of processors is almost linear in the IEEE 678 power system. It implies
that on large-scale power systems the increase of processors will
not affect the efficiency much, this is also confirmed as the decreasing speed of efficiency in IEEE 678 is the slower than the
other two. So it is not hard to predict that, a more satisfactory
speedup factor and a higher efficiency can be obtained when applying PNSGA-II in large-scale power systems with thousands
of buses.
SBSI is the most time-consuming part in PNSGA-II procedure, while the SBSI time depends on the number of faults considered. So the proposed method is very sensitive to the number
of faults considered. As shown in Table V, with the increasing
faults number, mean Manhattan distance with IPM gradually
increases, indicating a slight drop of optimization ability on
more faults cases. On the other hand, average execution time
of PNSGA-II grows rapidly with the increasing faults number,
but the growing trend is not entirely linear. Reported in [37], execution time per iteration of reduced-space IPM may not change
much when faults number varies, so it is one of the main drawbacks of the proposed method.
Fig. 12 indicates the fact that the relation between the speedup
factors and the number of processors is more linear when a

A multi-objective model is proposed in this paper as an effective quantitative analysis tool for OPF associated with cost, security and transient stability. A parallel NSGA-II is introduced
to search the Pareto optimal solutions, as well as a weighted
membership method for optimal solution selection. The proposed method has the following main features:
i) Providing a Pareto optimal solution set, rather than
a single strictly transient stable solution for decision-makers to select their ideal schemes according to
different preference. This is essentially different from
traditional TSCOPF and the proposed method is considered to be able to get a theoretically strictly transient
stable solution as well if enough iterations of evolution
are carried out.
ii) As an IA method, PNSGA-II is open-source, and much
easier to code and debug with a more significant robustness than IPM.
iii) Compared with serial IA-based TSCOPF, the proposed
parallel method has an obvious decrease in time and
lags slightly behind serial IPM-based TSCOPF when
enough number of parallel processors is used. Satisfactory speedup factors demonstrate the feasibility of
application the master-slave parallel algorithm to reduce
the optimization time and the acceleration is more effective in more data-intensive systems. It is a big push for
the proposed method to be applied in realistic large-scale
or multi-contingency power systems.
Following the idea of the proposed method in this paper,
many interesting topics are worth investigating in future research work:
i. Static power flow objective function
is reserved in
the proposed multi-objective OPF model, and NSGA-II
is naturally not sensitive to the number of objective functions, so more objective functions such as transmission



loss, in addition to fuel cost and transient stability can be

ii. To shorten the overhead for communication and further
improve optimization speed, the implementation of the
parallel algorithm on a supercomputer rather than PC
cluster is a worth trying method.
[1] M. Huneault and F. D. Galiana, A survey of the optimal power flow
literature, IEEE Trans. Power Syst., vol. 6, no. 2, pp. 762770, May
[2] G. Sergio, Optimal reactive dispatch through interior point methods,
IEEE Trans. Power Syst., vol. 9, no. 1, pp. 136146, Feb. 1994.
[3] D. Gan, R. J. Thomas, and R. D. Zimmerman, Stability-constrained
optimal power flow, IEEE Trans. Power Syst., vol. 15, no. 2, pp.
535540, May 2000.
[4] L. Chen, Y. Taka, H. Okamoto, R. Tanabe, and A. Ono, Optimal operation solutions of power systems with transient stability constraints,
IEEE Trans. Circuits Syst., vol. 48, no. 3, pp. 327339, 2001.
[5] T. Nguyen and M. A. Pai, Dynamic security-constrained rescheduling
of power systems using trajectory sensitivities, IEEE Trans. Power
Syst., vol. 18, no. 2, pp. 848854, May 2003.
[6] Y. Xia and W. Ka, Dynamic constrained optimal power flow using
semi-infinite programming, IEEE Trans. Power Syst., vol. 21, no. 3,
pp. 1455l457, Aug. 2006.
[7] Y. Xia, K. W. Chan, and M. Liu, Direct nonlinear primal-dual interiorpoint method for transient stability constrained optimal power flow,
Proc. Inst. Elect. Eng., Gen., Transm., Distrib., vol. 152, no. 1, pp.
1116, 2005.
[8] H. Xin, D. Gan, Z. Huang, K. Zhuang, and L. Cao, Applications of stability-constrained optimal power flow in the east China system, IEEE
Trans. Power Syst., vol. 25, no. 3, pp. 14231433, Aug. 2010.
[9] A. Pizano-Martinez, C. R. Fuerte-Esquivel, and D. Ruiz-Vega, Global
transient stability-constrained optimal power flow using an OMIB
reference trajectory, IEEE Trans. Power Syst., vol. 25, no. 1, pp.
392403, Feb. 2010.
[10] Q. Jiang and G. Geng, A reduced-space interior point method for transient stability constrained optimal power flow, IEEE Trans. Power
Syst., vol. 25, no. 3, pp. 12321240, Aug. 2010.
[11] H. R. Cai, C. Y. Chung, and K. P. Wong, Application of differential
evolution algorithm for transient stability constrained optimal power
flow, IEEE Trans. Power Syst., vol. 23, no. 2, pp. 719728, May 2008.
[12] N. Mo, Z. Y. Zou, K. W. Chan, and T. Y. G. Pong, Transient stability
constrained optimal power flow using particle swarm optimization,
IET Gen., Transm., Distrib., vol. 1, no. 3, pp. 476483, 2007.
[13] K. Nekooei, M. M. Farsangi, H. N. Pour, and K. Y. Lee, An improved
multi-objective harmony search for optimal placement of DGs in distribution systems, IEEE Trans. Smart Grid, vol. 4, no. 1, pp. 557567,
[14] P. Maghouli and S. H. Hosseini, Multi-objective framework for
transmission expansion planning in deregulated environments, IEEE
Trans. Power Syst., vol. 24, no. 2, pp. 10511061, May 2009.
[15] H. K. Khalil, Nonlinear Systems, second ed. Upper Saddle River, NJ,
USA: Prentice Hall, 1996.
[16] D. Gan, D. Chattopadhyay, and X. Luo, An improved method for optimal operation under stability constraints, in Proc. IEEE PES Transmission and Distribution Conf., Dallas, TX, USA, Sep. 2003.
[17] Y. Xue, T. Van Cutsem, and M. Ribbens-Pavella, A simple direct
method for fast transient stability assessment of power systems, IEEE
Trans. Power Syst., vol. 3, no. 2, pp. 400412, May 1988.
[18] J. Tong, H. D. Chiang, and T. P. Conneen, A Sensitivity-based BCU
method for fast derivation of stability limits in electric power systems,
IEEE Trans. Power Syst., vol. 8, no. 4, pp. 14181428, Nov. 1993.
[19] D. Xia and G. T. Heydt, On-line transient stability assessment using
higher-order Taylor expansion, IEEE Trans. Power App. Syst., vol.
PAS-102, no. 7, pp. 20382046, 1983.

[20] L. S. Pontryagin et al., The Mathematical Theory of Optimal Processes. New York, NY, USA: Wiley Interscience, 1962.
[21] D. E. Goldberg, Genetic Algorithm for Search, Optimization, Machine
Learning. Boston, MA, USA: Addison-Wesley Longman, 1989.
[22] N. Srinivas and K. Deb, Multi-objective optimization using non-dominated sorting in genetic algorithms, Evol. Computat., vol. 2, no. 3, pp.
221248, 1994.
[23] K. Deb, A. Pratap, and S. Agarwal et al., A fast and elitist multiobjective genetic algorithm: NSGA-II, IEEE Trans. Evol. Computat.,
vol. 6, no. 2, pp. 182197, 2002.
[24] Y. Li, Y. J. Cao, Z. Y. Liu, L. Yi, and Q. Y. Jiang, Dynamic optimal reactive power dispatch based on parallel particle swarm optimization algorithm, Comput. Math. Applicat., vol. 5, no. 7, pp. 18351842, 2009.
[25] [Online]. Available: http://www-unix.mcs.anl.gov/mpi/mpich/
[26] [Online]. Available: http://www.lam-mpi.org/
[27] [Online]. Available: http://www.open-mpi.org/
[28] B. Stott and O. Alsac, Fast Decoupled load flow, IEEE Trans. Power
App. Syst., vol. PAS-93, no. 3, pp. 859869, 1974.
[29] D. P. Kothari and I. J. Nagrath, Modern Power System Analysis.
Noida, India: Tata McGraw-Hill Education, 2003.
[30] K. Deb and R. B. Agarwal, Simulated binary crossover for continuous
search space, Complex Syst., vol. 9, pp. 115148, 1995.
[31] M. M. Raghuwanshi and O. G. Kakde, Survey on multi-objective evolutionary and real coded genetic algorithms, in Proc. 8th Asia Pacific
Symp. Intelligent and Evolutionary Systems, 2004, pp. 150161.
[32] R. Zimmerman and D. Gan, MATPOWER: A Matlab Power System
Simulation Package, 2005.
[33] J. H. Chow, Power system toolbox version 2.0, Cherry Tree Scientific
Software, 2000.
[34] F. R. Zaro and M. A. Abido, Multi-objective particle swarm optimization for optimal power flow in a deregulated environment of power
systems, in Proc. 11th Int. Conf. Intelligent Systems Design and Applications, 2011.
[35] M. T. Al-Hajri and M. A. Abido, Multi-objective optimal power flow
using improved strength Pareto evolutionary algorithm (SPEA2),
in Proc. 11th Int. Conf. Intelligent Systems Design and Applications,
2011, pp. 10971103.
[36] A. Bhattacharya and P. K. Roy, Solution of multi-objective optimal
power flow using gravitational search algorithm, IET Gen., Transm.,
Distrib., vol. 6, no. 8, pp. 751763, 2012.
[37] G. Geng and Q. Jiang, A two-level parallel decomposition approach
for transient stability constrained optimal power flow, IEEE Trans.
Power Syst., vol. 27, no. 4, pp. 20632073, Nov. 2012.
Cheng-Jin Ye was born in Hangzhou, China, in
1987. He received the B.E. degree from the College
of Electrical Engineering, Zhejiang University,
Hangzhou, China, in 2010, where he is currently
pursuing the Ph.D. degree.
His research interests include power grid planning,
renewable energy generation technology, and power
system optimization.

Min-Xiang Huang was born in Hangzhou, China,

in 1955. He received the B.E. degree from North
China Electric Power College, Baoding, China, in
1980, and the M.S. degree from Zhejiang University,
Hangzhou, China, in 1983.
He is currently a full Professor at the Department
of Electrical Engineering, Zhejiang University,
Hangzhou, China. His research interest includes
power grid planning, power system reliability
analysis, and electricity market issues.