Computers and Electrical Engineering 36 (2010) 553564

Contents lists available at ScienceDirect

Computers and Electrical Engineering

journal homepage: www.elsevier.com/locate/compeleceng

A grid computing based approach for the power system

dynamic security assessment
Wang Xingzhi *, Yan Zheng, Li Li
Department of Electrical Engineering, Shanghai Jiao Tong University, Shanghai 200240, PR China

a r t i c l e i n f o a b s t r a c t

Article history: This paper addresses the problem of parallel dynamic security assessment applications
Received 25 September 2008 from static homogeneous cluster environment to dynamic heterogeneous grid environ-
Received in revised form 14 October 2009 ment. Functional parallelism and data parallelism are supported by each of the message
Accepted 29 December 2009
passing interface model and TCP/IP model. To consider the differences in heterogeneous
Available online 25 February 2010
computing resources and complexity of large-scale power system communities, a kernel-
based multilevel algorithm is proposed for network partitioning. Since the bottleneck in
Keywords:
distributed computation is low speed network communication, a bi-level latency exploita-
Dynamic security assessment
Grid computing
tion technique is introduced for numerically solving system differential equations. The pro-
Modularity posed grid-based implementation includes the core simulation engine, grid computing
Latency exploitation technique middleware, a Python interface and Python front-end utilities. Tests for a 39-bus network,
Adaptive load balancing a 4000-bus network and a 10,000-bus network are reported, and the results of these exper-
iments demonstrate that the proposed scheme is able to execute the distributed simula-
tions on computational grid infrastructure and provide efcient parallelism.
2010 Elsevier Ltd. All rights reserved.

1. Introduction

The simulation of a complex power system for stability analysis is difcult due to the large number of components that
need to be considered such as governors, exciters, loads, electronic converters, etc., thus increasing the number of the state
variables and consequently the complexity of the dynamic model [1]. In order to develop a practicable approach utilizing
limited computational and technical resources, it is often necessary to use the parallel and distributed computing technol-
ogies [25].
During the last decades, a number of power system simulators implementing dynamic security assessment with parallel
computing have been developed. One of the most common schemes is the data parallelism [68]. While these low-level
infrastructures are extremely powerful, they are not compatible with each other, nor are they readily accessible to an aver-
age computational electrical engineer. On the other hand, higher-level parallelization systems with a Web-based user inter-
face may help computer neophytes, but these systems lack programming exibility to implement a users analysis algorithm
for various research purposes [911].
In this paper, a grid computing based architecture was developed to produce higher-level application program interfaces
(APIs) to provide users with a scripting environment and to distribute dynamic security assessment on grid environments.
Functional parallelism and data parallelism are supported by each of the message passing interface (MPI) model and TCP/IP
model.

* Corresponding author.
E-mail addresses: xzwang.sjtu@gmail.com (X. Wang), yanz@sjtu.edu.cn (Z. Yan), lili_sjtu@sjtu.edu.cn (L. Li).

0045-7906/$ - see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compeleceng.2009.12.010
554 X. Wang et al. / Computers and Electrical Engineering 36 (2010) 553564

The standard approach for data parallelism has been to divide the vertices of a power gird into approximately equal-
weight partitions and minimize the number of cut edges between partitions. It is proved that minimizing edge-cut may
not reduce the parallel execution time of a dynamic security assessment application [1214]. This paper presents a weighted
form of stochastic-based objective to directly optimize various weighted power grid cuts. This makes the proposed algorithm
scalable to irregular tasks and heterogeneous resources than standard partitioning methods.
During the dynamic security assessment, there exist both very fast and very slow dynamic devices. An important concern
arises here is how to consider the contribution of the slow-varying subsystem when solving for the fast one throughout a
complete interval. The initial attempts to exploit latency were made in very large-scale integration circuit simulation in con-
nection with the waveform relaxation method. In the early 1990s, the waveform relaxation method was used for the time
domain simulation of power systems in transient stability studies. In the latency technique proposed in this paper, boundary
subsystem uses the self and neighboring historical iterate parameters as guesses for the current state of subsystems.
The remainder of this paper is organized as follows. Section 2 briey introduces the power system dynamic security anal-
ysis algorithm. In Section 3, kernel-based complex power grid community nding algorithm, bordered block diagonal form
class data parallelism and latency exploitation technique to speed up simulations are described. Section 4 presents a grid
computing based architecture for distributed dynamic security assessment. In Section 5, three different scale power system
networks are used to test the effectiveness of the proposed approach. Finally, conclusions are drawn in Section 6.

2. Power system dynamic security analysis

2.1. Electromechanical transient stability problem

A power system can be modeled as a set of generators and a set of loads interconnected via a transmission network. The
electromechanical stability analysis of a power system is a simulation in the time domain, lasting several seconds or min-
utes. First the power system is simulated in its operating state, then large disturbances and protective actions are simulated
and nally the simulation is continued for a few more seconds or minutes. Different components of the power system have
their greatest inuence on the stability of the system at different points in time of the response simulation. The electrome-
chanical stability analysis of an electrical power system emphasizes the rapidly responding electrical components of the sys-
tem, for instance voltage and currents at the loads.
The complete system model of a power network can be described by a set of following nonlinear differential and algebraic
equations:

X_ g f Xg ; Vb ; 1

Ib Xg; Vb Yb Vb ; 2
where Xg are the generator variables describing the dynamics of the system and Vb are the bus voltages of the network.Using
the implicit trapezoidal integration rule for improved numerical stability, Eq. (1) can be discretized and rearranged as
follows:
Xgt  Xgt1 ;
3
 2h ff Xgt ; Vbt f Xgt1 ; Vbt1 g 0;
where t 1; 2; . . . ; T denotes time steps, and h represents time step duration. The linearized form of Eqs. (2) and (3) for iter-
ations at each time step is given by
    
Ag  Bg  Xg Rg
: 4
Cg  Yb Vb Rb
Eq. (4) may also be written as
Xg Ag 1 Bg Vb Ag 1 Rg 5
and, in a decoupled form
Yb Yg Vb Rb  Cg Ag 1 Rg ; 6
where Yg  is the generator equivalent admittance matrix dened as

Yg  Cg Ag 1 Bg : 7

The electromechanical transient stability problem can now be solved using a Newton method, obtaining Xg and Vb iter-
atively from Eqs. (5) and (6) at each time step.

2.2. An algorithm for performing dynamic security assessment

The dynamic security assessment can be divided into three sequential activities: (1) contingency screening; (2) dynamic
contingencies analysis; and (3) preventive and corrective control. In this paper, the second activity is mainly analyzed since it
 X. Wang et al. / Computers and Electrical Engineering 36 (2010) 553564 555

Fig. 1. Alternating scheme for electromechanical transient stability analysis.

is known to be the bottleneck in the fast computations and it represents an essential tool in supporting the optimal identi-
cation of the preventive and corrective measures aimed to mitigate the effects of critical contingencies.
Since the system dynamics can be highly nonlinear, especially for large disturbances, the corresponding dynamic solu-
tions are computed by time domain numerical integration. Two schemes are always used to solve the equations. The rst
approach is called simultaneous scheme where Eq. (1) is discretized and then solved together with Eq. (2) at each step.
The second one is called alternating scheme where Eq. (1) is solved by an integration method, and Eq. (2) is solved separately.
In Fig. 1 a brief owchart of alternating scheme for electromechanical transient analysis is depicted. The data parallelism
is started from dashed block where the subsystem boundary values are exchanged to solve set of linear equations extracted
from Eq. (2).
In particular, since there are independent contingency analysis tasks applying different disturbances to the given power
system, a functional parallelism can be used for fast dynamic contingencies analysis of large-scale power systems. For a gi-
ven power system with a given contingency, if parameters change, one would like to know as quickly as possible how these
changes will affect system stability indices. The transient stability indices at different fault clearing times can be obtained by
combining the transient energy function method with the time domain simulations [15].

3. Distributed dynamic security assessment

To exploit the potentialities of distributed computing for reducing the time needed for performing dynamic security anal-
ysis, the data parallelism is adopted.

3.1. Kernel-based algorithm for nding network communities

A power system network can be considered as comprising of sub-networks via tie-lines to a group of busbars known
as boundary nodes, and partitioning the network into sub-networks among processors is the common way to achieve
556 X. Wang et al. / Computers and Electrical Engineering 36 (2010) 553564

parallelism. To achieve the maximum efciency in parallel/distributed computation, the following criteria should be consid-
ered in network partition:

(1) The size of boundary nodes is as small as possible, which directly related to the data to be exchanged among
processors.
(2) The ratios of computation loads to computing powers for each processor are approximately equal.

Many algorithms have been presented for network partition, and most of them are based on heuristic graph partitioning
scheme. Although total network can be divided evenly with above partition methods, the uneven ratios of computing loads
to computing powers are still considerable because of the differences in heterogeneous computing resources and complexity
of large-scale power system communities. Thus, this paper presents a new algorithm to nding the communities of complex
power grids, and then the community results can be used for further data parallelism application.
Given a set of vectors /a1 ; /a2 ; . . . ; /an , the proposed kernel-based algorithm seeks to nd subsystem communities
p1 ; p2 ; . . . ; pk that satises the following mathematical programming equations,
P
k
ds 2
max M lLs  2L ;
s1
8
P
k P
s:t: min xi k/ai  mc k2 ;
c1 ai 2pc

where L is the number of total branches, ls and ds are the number of branches and bus degrees that belongs to subsystem
P
xi /ai
community s, xi is the non-negative weight, and mc Pi c
a 2p
. xi
ai 2pc

In the Hilbert space, the minimum distances of Eq. (8) can be calculated by the following equation,
P P
2
2 xj Kai ; aj
aj 2p c aj al xj xl Kaj ; al
k/ai  mc k Kai ; ai  P P 2 ; 9
aj 2pc xj xj
aj 2pc

where Kai ; aj /ai  /aj is the kernel matrix.

Starting from a random subsystem communities, Eqs. (8) and (9) are solved iteratively to nd optimal power system com-
munity structures. Notice that the calculation speed of the proposed procedure depends on the scale of power grids. In order
to improve the stability of subsystem partitioning precise, our improved multilevel algorithm starts from a multilevel graph
partitioning results and uses the weighted kernel-based algorithm during the renement phase to directly optimize various
weighted power grid cuts. This makes the proposed algorithm scaleable to the differences in heterogeneous computing re-
sources and complexity of large-scale power system communities.

3.2. A bordered block diagonal form of linear equations

Handling the solutions of distributed linear algebraic equations is one of the most difcult tasks in transient stability
problem. Many parallel algorithms tried to solve linear equations, such as those incorporate the multiple factoring methods,
the W matrix method, parallel factorization and substation method, conjugate gradient method, and bordered blocked diag-
onal form (BBDF) class methods. Among these methods, BBDF class is preferred for its light communication. BBDF linear
equation can be solved by parallel LU factorization and parallel forwardbackward substitution.
To solve a linear algebraic problem with the form,
Ax b; 10
where A is usually a sparse matrix, the most effective method up to date is the use of triangular factorization along with
forward/backward substitution. In the BBDF class algorithms, the matrix A of Eq. (10) can be arranged into the BBDF by putt-
ing the boundary nodes behind other nodes. The border blocks are also known as the connection blocks as they connect the
diagonal blocks to the boundary node block. The border blocks and boundary node block are distributed into each
subsystem.
In the LU factorization process, the BBDF structure is preserved and computation among the diagonal blocks and border
blocks are completely independent. Dependence exists only between the border blocks and their corresponding diagonal
blocks. In the forward substitution process, there is also no data dependence among the diagonal blocks. The operation
on the diagonal and bordered blocks can proceed in parallel. Before the boundary node blocks are solved, a communication
step is required to globally distribute the intermediate results corresponding to the boundary node block. After the commu-
nication, forward and backward substitution can proceed in parallel completely.

3.3. Latency exploitation technique to speedup simulations

In the grid computing environment, the uneven ratios of computing loads to computing powers are still considerable be-
cause of the differences in heterogeneous computing resources and multi-scale complexity of large-scale power systems.
 X. Wang et al. / Computers and Electrical Engineering 36 (2010) 553564 557

During each iteration step of the traditional dynamic stability simulations, every boundary subsystem iteration parameters
are computed and exchanged once using the previous one-step information, and its frequency is determined by the ratio of
the fastest subsystem computing load to the worst computing power. This quite often results in a waste of computa-
tional resources and increasing of communications since the frequency may be unnecessarily high for other subsystems. La-
tency exploitation can optimize computational resources by using multiple time step prediction techniques. In the latency
technique proposed in this paper, the large time step with high ratio of computing load to computing power should always
be an integer multiple of the small ones,
DT pDt; 11
where p is a positive integer that is larger than unity, DT and Dt are, respectively, the global and the local time steps. To con-
sider the contribution of the subsystem with lower ratio of computing load to computing power, a bi-level approximation
method is presented.
In the rst level approximation, the subsystem with high ratio of computing load to computing power use the self pre-
vious iterate boundary values for the current state of the other subsystems. According to the multi-step prediction scheme,
the update is done as follows:
X
p X
p
^ k1
V Ai V ki Bi gtki ; xk ; V ki ; 12
i0 i1

where A and B are time-varying parameters of different time steps, g 0 is a set of nonlinear algorithm equations of bound-
ary subsystems.
In the rst level approximation method which already described, the subsystem with high ratio of load to power uses the
self historical boundary values for the current state of the other subsystems. This may require more iteration to converge and
therefore degrade the efciency. It can be enhanced by using the historical boundary values of neighboring subsystems. The
idea of second level approximation is to estimate contribution of other subsystems in injection currents from its previous
quantity.
The second level approximation updates the high ratio subsystem contribution in injection current by using self and
neighboring historical information in local step number k. Based on Diakoptics technique, an unknown current exchange will
be appropriately injected back to the subsystems for modeling the other neighboring subsystems at any boundary buses, i.e.

Vkl Y1 k ^k
l Il Cl Ilex l 1; . . . ; s; 13
where l is the subsystem identier, k is iteration number and Cl is the incidence matrix of the lth subsystem. If bus i is a
boundary bus and currents outgoing, then cij is 1. While, if bus i is a boundary bus and currents incoming, then cij is 1.
Moreover, if bus i is not a boundary bus, then cij is 0.
To consider the contribution of the subsystems with lower ratio of computing load to computing power, Eq. (13) is pre-
served for every boundary branch between two adjacent subsystems i and j:
^ k  V^ k Pp1 V km  V km 
V i j m1 i j
1  F  yij  F  yij  V^ ki  V
^ k  ^Ik 0;
j ijex 14
p
where F is a scalar that represents the operating state of the cut branch. If F is 1, then the cut branch is closed. While, if F is 0,
then cut branch is open and either side of the boundary bus has different quantities. Eq. (14) can be rewritten in matrix form
as follows,
" #
X Xs  
s ^k
1 p  1F  Dl  V l 1  F Xp1 km ^k
D
m1 l
 V l  Ilex 0; 15
l1
p l1
p

where Dl is the matrix with element dij cij  yij . Combining Eqs. (12) and (15) into a common frame work, the general bi-
level approximation expression for the piecewise linear equation solution will be obtained.

4. The grid implementation of distributed dynamic security assessment

To execute the dynamic security analysis by adopting the distributed algorithm described in the previous section, a grid-
based implementation is proposed. A complex distributed dynamic security assessment implementation usually has a mod-
ular architecture and consists of a number of loosely or tightly coupled components. Our implementation includes the core
simulation engine for power ow simulation, contingency screening, electromechanical transient stability analysis, preven-
tive and corrective control; grid computing middleware; a Python language wrapper of core simulation engine and lower-
level middleware; Python front-end utilities (Fig. 2). The aim of our research is to virtualize separate modules of the appli-
cation to run them efciently as services and access them on the grid.
Developing high-performance problem solving environments that allow power system researchers to easily harness the
power of the grid infrastructure is currently a difcult task. Although the Globus toolkit provides a number of modules that
implement grid services for security, resource discovery, data transfer/management, etc., these grid services can still be very
558 X. Wang et al. / Computers and Electrical Engineering 36 (2010) 553564

Fig. 2. Distributed scheme of the dynamic security assessment application.

challenging to use. To address this difculty we have begun the development of the Python based toolkits that support rapid
power system stability analysis application development through the use of modern software engineering techniques.
In the proposed environment, there exist a wide variety of modules that provide the basic interface to the Globus toolkit.
The exception module makes extensive use of exceptions for error handling. The resource acquisition module provides the
main interface to the Globus GRAM protocol to provide resource acquisition and management functionality. The Gass copy
module provides a protocol independent interface to transferring les. And the secure IO module provides an easy to use
interface to high-performance secure synchronous and asynchronous remote IO using the Grid Security Infrastructure to
support PKI authentication [16].
The core components are modules simulating power ow, contingency screening, transient stability, preventive and cor-
rective control. The details on numerical methods and functional/data parallelism algorithms employed in the solvers have
been described in the previous sections. A data parallelism algorithm was developed with large-scale power system simu-
lation loads distributed among the processors. The algorithm was implemented in an SPMD model, using MPI message pass-
ing interface with MPI Barrier points for synchronization. Distributed data pool technique is used to improve the data
exchange efciency. To study the inuence of various contingencies on the simulated processes we run a number of simu-
lations in parallel (shown in Fig. 2 as Functional parallelism simulation1. . .Functional parallelism simulation M blocks).
To provide efcient execution of a distributed dynamic security assessment on the grid computing resources, it is neces-
sary to take into account the load balance scheme. Considering the computational and communication cost of the subsystem,
the cost of l can be described as follows:

xl N v l 2  N el Ndl N2cv l 2  Ncel Ncdl ; 16

where N v l is the number of internal nodes, N el is the number of internal weighted branches, N dl is the equivalent weight of
internal dynamic units, N cv l is the number of boundary nodes, N cel is the number of boundary weighted branches, and N cdl is
the equivalent weight of boundary dynamic units. Performance models are often parameterized by values that represent
 X. Wang et al. / Computers and Electrical Engineering 36 (2010) 553564 559

power system simulation and load characteristics. In dedicated settings it is often sufcient to represent these characteristics
by a single value, or point value. However, point values are often inaccurate or insufcient representations for characteristics
that change over time [3]. One way to represent this variable behavior is to use a stochastic value, or distribution. In the sto-
chastic setting, we choose the value for f from the range given by the stochastic prediction of performance. This allows us the
exibility to choose larger or smaller values of f in the range. One approach to determine which value to choose for f is to
consider the mean and standard deviation information of the computational and communication cost, that is

min f TF m  mx TF sd  sdx ; 17

where mx , the mean of the predicted computational and communication cost x for subsystems. sdx , the standard deviation of
the predicted computational and communication cost x for subsystems. TF m , the Tuning Factor of mean, used to determine
the number of means to add to (or subtract from) the standard deviation value. TF sd , the Tuning Factor of standard deviation,
used to determine the number of standard deviations to add to (or subtract from) the mean value.
Notice that the load balance algorithm of Eq. (17) is based on the complex power grid community identication results
solved by Eq. (8). Thus, a fast and adaptive load balance results can be achieved. This makes the proposed load balance algo-
rithm scalable to irregular tasks and heterogeneous resources than the traditional graph partitioning methods.

5. Numerical experiments

The architecture presented in Section 4 was coded using Python for front-end scripting, MPICH and Globus middleware,
and C++ programming language for core simulation engine. The grid-enabled data parallelism was tested with different scale
power grid congurations on an IBM p690. The models performance is analyzed in time domain simulation.

5.1. Test case #1

The rst test case is a 39-bus New England system, as shown in Fig. 3. It consists of 10 synchronous machines with IEEE
type-1 exciters, and standard turbine and governor models. There are 18 loads in the system totaling 5045.5 MW and
1158.4MVAR.
For validating the proposed kernel-based method, ve sub-networks comparison results between Metis and the proposed
method are shown in Figs. 3 and 4. The Metis is considered as the most accurate graph clustering toolkit, therefore the pro-
posed method is compared against Metis. The result of Fig. 3 show that nine branches should be cut to form ve sub-net-
works. As it is shown in Fig. 4, the proposed method produces eight cut branches. The cut branch power ow comparison
results are listed in Table 1. It shows that proposed method produces better partitions than Metis, which is not surprising
since Metis minimize the edge cut values instead of the weighted kernel-based clustering distances.
As an event scenario for electromechanical transient stability analysis, a three phase short circuit is considered on the 1%
of lines 2629 at 0.0s and removed at 0.13s. Fig. 5 shows the generator angle simulation results using the transient stability
analysis approach. As it is shown, the maximum absolute angle error at 2.04s reaches 184, the generators are divided into
three clusters: (1) G9; (2) G1and G10; (3) other generators. If there is no preventive and corrective control scheme, the G9
will become transient angle unstable.
Distributed simulation is carried out on a different error tolerance values with different step sizes. The distributed dy-
namic security assessment algorithm is compared with non-distributed algorithm. The results of all distributed cases are

Fig. 3. New England data parallelism results using original graph partitioning.
560 X. Wang et al. / Computers and Electrical Engineering 36 (2010) 553564

Fig. 4. New England data parallelism results using the proposed method.

Table 1
Cut branch power ow results of different power grid partitioning methods.

Number Onsinal method Proposed method

Cut branch Branch ow (MVA) Cut branch Branch ow (MVA)
No. l 12 60 1727 34
No. 2 1617 129 12 60
No. 3 34 141 1415 91
No. 4 45 192 1624 99
No. 5 1516 279 2526 101
No. 6 2627 290 1718 138
No. 7 611 315 34 141
No. 8 23 404 1621 240
No. 9 1619 438 / /

Fig. 5. Generator angle results of New England power system with line faults.

compared with a single simulation with the same time step and error tolerance to evaluate the performance of distributed
simulation. The mean absolute percentage error (MAPE) measured its performance in the bus voltage amplitude/rotor angle
of G5. Table 2 shows the test results. This comparison shows that almost same accuracy for the distributed algorithm. In the
same step size, the smaller error tolerance results in the smaller MAPE. However, smaller MAPE cannot be ensured by the
smaller step size with same error tolerance.
It is not always easy to predict the optimal parallel processor number with maximum parallel speedup. In this paper, the
subsystem number with maximum network modularity is used for purposes of nding the optimal processors with
maximum parallel speedup. For validating the proposed optimal parallel processor prediction scheme, the comparative
 X. Wang et al. / Computers and Electrical Engineering 36 (2010) 553564 561

Table 2
Mean absolute percentage error for different simulation methods.

Case Non-distributed G5-MAPE (%) Distributed G5-MAPE (%)

|Vm| dG |Vm| dG
1: Step size le2, Error tolerance le5 4.74e7 1.25e4 4.74e7 1.25e4
2: Step size le2, Error tolerance le4 2.64e6 4.55e5 2.64e6 4.55e5
3: Step size le2, Error tolerance le3 2.16e5 0.00303 2.16e5 0.00303
4: Step size 2e2, Error tolerance le5 l.08e7 1.77e4 1.08e7 1.77e4
5: Step size 2e2, Error tolerance le4 2.34e6 0.00155 2.34e6 0.00155
6: Step size 2e2, Error tolerance le3 1.88e5 0.02667 1.88e5 0.02667

Table 3
Comparative results of network modularity and parallelism simulation.

Number of processors Network modularity Simulation time (s) Parallel speedup

1 0.00000 0.10528 1.00000
2 0.43266 0.21689 0.48541
3 0.51205 0.08937 1.17802
4 0.48299 0.09911 1.06225
5 0.46621 0.11158 0.94354
6 0.51111 0.13064 0.80588
7 0.49055 0.15348 0.68595
S 0.50354 0.16980 0.62002
9 0.50071 0.19207 0.54813
10 0.50425 0.20922 0.50320

results for network modularity and parallel speedup are depicted in Table 3. As it is shown, the subsystem number with
maximum network modularity is three, while the optimal number of processors with maximum parallel speedup is same.
From the viewpoint of preventive and corrective control, branch 25 and 26 and branch 1727 will be switched out at 2.0s
and switched on at 2.36s, respectively, due to controlled separation protection system actions. Fig. 6 shows the generator
frequency simulation results using the proposed transient stability analysis approach. Note that all the generator frequencies
are within the operative frequency range 0:95; 1:03 p.u.

5.2. Test case # 2

Such a distributed simulation procedure has been successfully tested on real large power system with different sizes of
sub-networks. The mid-scale network was a part of North America grid with approximately 4000 buses.
Initially for validating the proposed rst level latency approximation approach, percentage error results of multi-step pre-
diction methods are listed in compare with those obtained from the traditional prediction method in Table 4. It is clear that
the parallel simulation times of modied rst order Adams estimation are smaller than that of others. An application to mid-
scale network shows that by the modied rst order Adams prediction method the simulation time is reduced by up to

Fig. 6. Generator frequency results of New England system after controlled separation scheme.
562 X. Wang et al. / Computers and Electrical Engineering 36 (2010) 553564

Table 4
Percentage error results of different multi-step simulation methods.

Number of processors Simulation time of original method (s) PE of rst- order Adams method (%) PE of modied rst-order Adams method (%)
1 13.362412 9.75 15.05
2 7.190145 8.42 15.20
3 5.846166 6.65 13.67
4 5.446230 9.56 15.01
5 5.176049 9.06 15.23
6 4.754111 8.43 16.14
7 4.606308 8.40 15.13
8 4.636290 10.07 17.89
9 4.671215 9.27 16.58
10 4.809937 8.80 16.39

Fig. 7. Simulation times of different neighboring historical parameter estimation schemes for NA4000.

17.89% compared to the traditional approach. This table can serve as validations of the efciency of the rst level latency
approximation algorithm.
For validating the proposed second level latency approximation approach, simulation time comparative results for pro-
posed method and the traditional method are depicted in Fig. 7. For algorithm stable convergence consideration, number p of
second level latency approximation is 2. As illustrated in Fig. 7, the use of the second level latency approximation approach is
able to reduce the solution time. Besides that, with the increasing of the processor number, the parallel simulation time (no
data I/O time) and total simulation time have some minimum values. It can be explained that there are not enough compu-
tations to parallel and too many processors may contribute to the loss of efciency by some additional communications. The
optimal number of parallel processors can also be estimated by the number of maximum network modularity.

5.3. Test case # 3

The largest network was a version of national China grid with approximately 10,000 buses.
The proposed load balance algorithm uses the adaptive stochastic-based algorithm to optimize weighted national China
grid cuts. If TF m and TF sd are set to 0 and 1 respectively, the clustering quality results of the proposed method and Metis can
be shown in Figs. 8 and 9. Fig. 8 shows that the proposed method produces a little higher ratio association values than Metis.
From Fig. 9, we see that the proposed method gives a lower normalized cut values than Metis. Note that all our experiments
show that the proposed method consistently produces better partitions than Metis, which is not surprising since Metis min-
imizes the edge cut value instead of the normalized cut or ratio association. This indicates that it is effective to employ the
proposed stochastic-based method to solve the large-scale power grid partition problem.
The speedup and efciency measure for the grid-enabled data parallelism was measured by using the national China grid.
The speedup and efciency of computations is presented in Fig. 10. Simulation results show that the near linear parallel
speedup could be obtained by the proposed approach. This can be explained that the use of data parallelism did reduce
the calculation time. Notice that the part of the gain in data parallel may be nullied by the increasing data I/O, communi-
cation step and programming overheads. The optimal number of processors can be predicted by the number of maximum
network modularity.
 X. Wang et al. / Computers and Electrical Engineering 36 (2010) 553564 563

Fig. 8. Quality comparison between the Metis and the proposed method in terms of ratio association values for national China grid.

Fig. 9. Quality comparison between the Metis and the proposed method in terms of normalized cut values for national China grid.

Fig. 10. Speedup and efciency of national China grid.

6. Conclusions

In this paper, the development and implementation of a grid-based distributed computing approach for electromechan-
ical transient stability analysis of large-scale power system were presented. Numerical analysis techniques of test case #1
564 X. Wang et al. / Computers and Electrical Engineering 36 (2010) 553564

demonstrate the convergence and precision of the dynamic security assessment algorithm. The network partition compar-
ative results show that the proposed kernel-based algorithm can be used to nd the optimal electrical network community
structure. Test case #2 is developed to demonstrate the capability of the latency exploitation technique. An application
shows that by the modied rst order Adams prediction method the simulation time can be reduced by up to 17.89% com-
pared to the traditional approach. As illustrated in the gures, the use of the second level latency approximation approach is
able to speedup simulations, and the optimal number of parallel processors can be estimated by the number of maximum
network modularity. For the test of 10,000-bus network, a near linear speedup results are achieved for the test case. This
demonstrates the proposed distributed approachs capability to work with on-line security analysis applications.
To further test the proposed algorithm, different static/dynamic homogeneous/heterogeneous resources can be used for
setting up the complex grid computing environment. Although most of the Globus toolkit and the core simulation engine for
dynamic security assessment have been wrapped, there are a number of areas still to be completed. In particular, in the fu-
ture service grids and a set of common GUI components will be considered for the integration of other power system
applications.

Acknowledgment

This work was supported in part by the National Natural Science Foundation of China with the No. 90612018.

References

[1] Machowski Jan, Bialek Janusz W, RIBumby James. Power System Dynamics and Stability. John Wiley & Sons Ltd.; 1997.
[2] Falco DM. High performance computing in power system applications. In: Proceedings of the 2nd international meeting on vector and parallel
proceeding; September 1996.
[3] Schopf J, Berman F. Stochastic scheduling. In: Proc supercomputing; 1999.
[4] Berman F, Wolski R, Casanova H, Cirne W, Dail H, Faerman M, et al. Adaptive computing on the grid using AppLeS. IEEE Trans Parallel Distributed Syst
2003;14:36982.
[5] Lee Liang-The, Tao Der-Fu, Tsao Chieh. An adaptive scheme for predicting the usage of grid resources. Comput Elect Eng 2007;33:111.
[6] Chan KW, Dunn RW, Daniels AR. Efcient heuristic partitioning algorithm for parallel processing of large power systems network equations. IEE Proc
Gener Transm Distrib 1995;142:62530.
[7] Xue W, Shu JW, Yan JF, Wang XF, Han YD. Cluster-based parallel simulation for power system transient stability analysis. Proc CSEE 2003;23:3843.
[8] Jian Wu, Schulz Noel N, Gao Wenzhong. Distributed simulation for power system analysis including shipboard systems. Elect Power Syst Res
2007;77:112431.
[9] Santo Michele Di, Vaccaro Alfredo, Villacci Domenico, Zimeo Eugenio. A distributed architecture for online power system security analysis. IEEE Trans
Ind Electron 2004;51:123848.
[10] Ali M, Dong ZY, Li X, Zhang P. RSA-grid: a grid computing based framework for power system reliability and security analysis. In: Proceedings of the
2006 IEEE power engineering society general meeting; June 2006.
[11] Shi LB, Chang NC. Lan Z, Zhao DP, Zhou HF, Peter TC, et al. Implementation of a power system dynamic security assessment and early warning system.
In: Proceedings of the 2007 IEEE power engineering society general meeting; June 2007.
[12] Albert R, Barabasi AL. Statistical mechanics of complex networks. Rev Mod Phys 2002;74:4797.
[13] Pothen A. Graph partitioning algorithms with applications to scientic computing. Department of Computer Science, Old Dominion University,
Norfolk, VA, Tech. rep. TR-97-03; 1997.
[14] Dhillon I, Guan Y, Kulis B. Kernel k-means, spectral clustering and normalized cuts. In: Proceedings of the 10th ACM knowledge discovery and data
mining conference; 2004.
[15] Wang Xingzhi,Yan Zheng, Li Li, Xie Dong. A new kernel-based clustering algorithm for the multi-machine equivalent of large power systems. In:
Proceedings of the 11th international conference on electrical machines and systems; 2008.
[16] Foster I, Kesselman C, Tsudik G, Tuecke S. A security architecture for computational grids. In: Proceedings of 5th ACM conference on computer and
communications security conference; 1998; p. 8392.