You are on page 1of 5

ARTICLE IN PRESS

G Model
EPSR-3456; No. of Pages 5

Electric Power Systems Research xxx (2012) xxxxxx

Contents lists available at SciVerse ScienceDirect

Electric Power Systems Research


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

Short communication

On Krons diakoptics
Fabian M. Uriarte
The Center for Electromechanics of The University of Texas at Austin, 10100 Burnet Road, Austin, TX 78757, United States

a r t i c l e

i n f o

a b s t r a c t

Article history:
Received 27 September 2011
Received in revised form 25 January 2012
Accepted 25 January 2012
Available online xxx

Diakoptics is a well-known method of tearing electric networks into computationally smaller subsystems. This paper exposes two often-overlooked, important properties related to diakoptics. One is that
branches are not required to tear networks; the other is that the order of the boundary network is strongly
dependent on the power system formulation variablea choice commonly made too prematurely during software development. It is concluded that, rst, tearing zero-immittance branches (meshes and
nodes) offers more disconnection points than branch tearing; second, that tearing meshes can result
in less boundary variables than tearing nodes, and, hence, reduce the computation effort of solving the
boundary network.
2012 Elsevier B.V. All rights reserved.

Keywords:
Mesh
Node
Partitioning
Parallel
Power
Simulation
Diakoptics
Kron
Multicore
System

1. Introduction
Kron [1,2] derived from physical principles that electric networks formulated in the form of (1) could be partitioned into p
subsystems and reformulated as (2):
Aorig x = b

A1

A2

DT1

DT2

(1)

..

Ap
DTp

Expressing (2) in compound matrix form results in (3), where


each compound matrix is expanded in (4) and (5). In (5), the
subscript r represents the number of boundary variables that are
formed when partitioning the original network into p partitions.

Ablock
DT

D1
x1
b1
D2 x2 b2

..
. .
. .. = ..
Dp xp bp
u
0
Q

D
Q

Aorig = original (unpartitioned) network coefcient matrix;


x = network variables vector (node voltages or mesh currents);
b = excitation vector; Ai = network coefcient matrix for subsystem
i; xi = network variables vector for subsystem i; bi = excitation
vector of subsystem i; Di = connection matrix linking subsystem
is internal variables to its boundary variables; p = number of
partitions; Q = boundary immittance matrix; 0 = zero matrix or
vector; u = boundary variable vector (branch voltages or currents).


b
0

A2
..

u1
Q1
u2

u = . ;Q =
..

ur

D(i, j)

(3)

A1

Ablock =
(2)

x
u

x1
b1
D1

x2
b2
D2
; x= . ; b= . ; D= . (4)

..
..
..
Ap
xp
bp
Dp

Q2
..

(5)

Qr

= 1, if xi is positively coupled to uj
= 1, if xi is negatively coupled to uj
= 0, if xi is not coupled to uj .

Expanding the rows of (3) results in


Tel.: +1 512 232 8079; fax: +1 512 471 0781.
E-mail address: f.uriarte@cem.utexas.edu

x = A1
b A1
Du
block
block
T

Qu = D x.

0378-7796/$ see front matter 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.epsr.2012.01.016

Please cite this article in press as: F.M. Uriarte, On Krons diakoptics, Electr. Power Syst. Res. (2012), doi:10.1016/j.epsr.2012.01.016

(6)
(7)

ARTICLE IN PRESS

G Model
EPSR-3456; No. of Pages 5

F.M. Uriarte / Electric Power Systems Research xxx (2012) xxxxxx

Substituting (rst) x from (6) into (7) and (then) u from (7) back
into (6) results in (8). (Eq. (8) is also known in mathematics as
Woodburys method for inverting modied matrices [35].) Substituting the expanded matrices of (4) and (5) in (8) produces (9) and
(10), which is the diakoptical solution to (1) as originally proposed
by Kron.
u

x = A1
b A1
D(DT A1
D Q)
block
block
block

(DT A1
b)
block

(D Ablock b)

Fig. 1. Original oating unpartitioned electrical network.

(10)

In the context of electromagnetic transient simulations of large,


complex power systems [6,7], the sequential solution of (9) and
(10) at each time step is more efcient than the solution of (1)
due to the parallelization opportunities offered by (9) [3,8]. This
solution technique has seen application to power systems since
such problems were solved longhand.
This section summarized Krons diakoptics and its equations
of solution. The following sections expose two often-overlooked,
important properties related to diakoptics.
2. First property: zero-immittance tearing
In relation to (3), this section shows that it is not necessary for
the boundary immittance matrix Q to exist, i.e., Q = 0. (This property
was apparently not recognized by Kron.) The matrix Q represents
the immittance (if any) of all boundaries that form when partitioning a network, where branch tearing is only a special case.
Two situations where Q = 0 occurs are presented next. The
rst is when power systems are formulated in mesh currents as
variables. In this case, it can be said that its mesh currents circulate around open spaces of zero conductance. The second is
whenanalogouslypower systems are formulated in node voltages as variables, where it can be said that nodes appear as galvanic
junctions of zero impedance. If said meshes or nodes are bisected,
it constitutes a tearing of zero-immittance branches which makes
Q = 0. Henceforth, zero-immittance tearing may be referred to as
mesh or node tearing, or as topological tearing.
Substituting Q = 0 in (3) results in (11), where after following
the derivations from (6) and (7) through (9) and (10), results in the
solution form shown as (12) and (13) instead. Apparently the only
difference between solution sets (9) and (10) and (12) and (13) is
in the solution of ubut there are more.

Ablock
D

(11)

1 1
A1 b1
A1 D1
x1

x
A
b
A1

2
2

2
2 D2

=
=

x
.
.

. . .. u (12)

.
xp

A1
p bp

A1
p Dp

u = (DT A1
D)
block

(DT A1
b)
block

(13)

The advantages of zero-immittance tearing are:

ibc

(8)

1 1
A1 b1
A1 D 1
x1

x2 A1
b2 A1
D2

2
x = =
2

... .. .. u (9)
.
.

xp

A1
A1

p bp
p Dp

T 1
T 1
u = (D Ablock D Q)

iab

1) In diakoptics, Q exists only if boundary branches can be dened.


This is not always possible and is a limitation that does not exist
in zero-immittance tearing.
2) In zero-immittance tearing, a network can be partitioned anywhere: inside or outside power apparatus, or at single- or
three-phase buses. This exibility increments the number of
choices available to graph-theoretic algorithms [9], which are
normally used to search for the disconnection points of power
systems.
The disadvantages of zero-immittance tearing should also be
pointed out:
1) Zero-immittance tearing retains the non-zero structure of Aorig .
While this is not detrimental, it is preferable to decrease the
non-zero count when using sparse solvers [10]. Diakoptics does
this fairly wellbut only if (intentionally) tearing mutual inductances in nodal formulations [11], or shunt-impedances at buses
in mesh formulations [12]. The former is a rather obscure and
counter-intuitive approach to tearing; the latter, is self-evident
and intuitive [7].
2) The total system order increases for every zero-immittance tear.
This occurs in mesh formulations due to mesh bisections, and
in nodal formulations due to node bisections. The bisections
create new variables internal to each subsystem interfaced at
a boundary. This consideration is only important if using fullmatrix solution techniques, which are rarely preferable over
sparse ones.
It is not difcult to show that the advantages of zero-immittance
tearing outweigh its disadvantages as computational performance
is more sensitive to good network segregations (i.e., balanced subsystems with minimal tears) than it is to a few matrix ll-ins. Two
examples are provided next to illustrate zero-immittance tearing.
2.1. Mesh formulation example
Consider the ungrounded network in Fig. 1. Placing two voltage
sources of unknown value across the open spaces bisects iab and ibc
into iab1, iab2 and ibc1, ibc2, respectively. Fig. 2 shows the disconnection points after the bisection. From circuit theory, the voltage
sources can be torn as shown in Fig. 3, as the boundary variables
vab and vbc are common to both subsystems.
The matrices and vectors for the torn network of Fig. 3
are Ablock = diag(A1 , A2 ) = diag(Rmesh1 , Rmesh2 ), x = [ x1
[ imesh1

imesh2 ] , imesh1 = [ iab1


T

ibc1 ] , imesh2 = [ iab2


T

x2 ] =
T

ibc2 ] , u =

[ vab vbc ] , and b = [ b1 b2 ] = [ emesh1 emesh2 ] .


Matrices Rmesh1 , Rmesh2 , vectors imesh1 , imesh2 and emesh1 , emesh2
are the mesh resistance matrices, mesh current vectors, and EMF
vectors in subsystems 1 and 2, respectively. After forming D, which
depends on a networks topology, the prior matrices and vectors
can be substituted in (12) and (13) to obtain a partitioned-network
solution. As noted, boundary branches do not exist in Fig. 1: this
produces the condition Q = 0.

Please cite this article in press as: F.M. Uriarte, On Krons diakoptics, Electr. Power Syst. Res. (2012), doi:10.1016/j.epsr.2012.01.016

ARTICLE IN PRESS

G Model
EPSR-3456; No. of Pages 5

F.M. Uriarte / Electric Power Systems Research xxx (2012) xxxxxx

Add unknown
voltage sources

iab1

vab

iab2

ibc1

vbc

ibc2

Fig. 2. Bisection of open circuits from adding voltages sources forms a boundary or disconnection point.

Tear voltage sources

Subsystem 1

iab1

vab

iab2

ibc1

vbc

ibc2

Subsystem 2

Fig. 3. Tearing the voltage sources creates two subsystems having common boundary variables.

Fig. 4. Original grounded unpartitioned electrical network.

2.2. Nodal formulation example


Consider the grounded network in Fig. 4. Placing current sources
of unknown values in line with the galvanic junctions bisects va , vb ,
and vc , into va1 and va2 , vb1 and vb2 , and vc1 and vc2 , respectively.
Fig. 5 shows the current sources added to the disconnection point.
From circuit theory, current sources can be torn as shown in Fig. 6,
as boundary variables ia , ib , and ic are common to both subsystems
[1318].
The matrices and vectors for the torn network of
Fig. 6 are Ablock = diag(A1 , A2 ) = diag(Gnodal1 , Gnodal2 ), x =
[ x1

x2 ] = [ vnodal1

vnodal2 = [ va2

vnodal2 ] ,

vnodal1 = [ va1
T

vb1 vc1 ],
T

vb2 vc2 ], u = [ ia ib ic ] ,and b = [ b1 b2 ] =


T

[ inodal1 inodal1 ] . Matrices Gnodal1 , Gnodal2 , and vectors vnodal1 ,


vnodal2 and inodal1 , inodal2 are the nodal conductance matrices, node

voltage vectors, and current injection vectors for subsystems 1 and


2, respectively.
Similar to the mesh case, after forming D, the matrices and
vectors afore can be substituted in (12) and (13) to obtain a
partitioned-network solution. As noted, boundary branches do not
exist in Fig. 4. This also produces the condition Q = 0.

3. Second property: formulation choice


A boundary or disconnection point is where two or more
subsystems interface. Determining the number and location of
these boundaries to produce p partitions is a difcult problem to
solveyet a solution is required prior to partitioning a power system.

Fig. 5. Addition of unknown current sources between short circuits forms a boundary or disconnection point.

Please cite this article in press as: F.M. Uriarte, On Krons diakoptics, Electr. Power Syst. Res. (2012), doi:10.1016/j.epsr.2012.01.016

ARTICLE IN PRESS

G Model
EPSR-3456; No. of Pages 5

F.M. Uriarte / Electric Power Systems Research xxx (2012) xxxxxx

Fig. 6. Tearing the current sources creates two subsystems having common boundary variables.

Mesh Formulation

Nodal Formulation

[A2]

[A2]
[A1]

i2

[A1]

i1
[A5]

i5

i3
u1

p=5
r=1

u2

u1

[A3]

u3

i4

[A5]
[A4]

[A3]

u4

[A4]

p=5
r=4

Fig. 7. A partitioning situation that shows how the number of boundary variables r depends on the formulation choice.

The number of boundary variables r (i.e., the total number of


boundary voltage or current sources introduced) adversely affects
the computation time of u in (13); hence, it is important to choose
a formulation variable (nodes or meshes) such that r remains small
as p increases past two. It appears that the importance of this choice
is unknownor deliberately ignoredin much of the literature as
software manufacturers and books advocate nodal formulations
without showing how it performs in partitioned simulation scenarios.
Consider a grounded network where one (of many possible) disconnection points has been identied in Fig. 7. At this particular
disconnection point, there are p = 5 partitionsall interfaced at the
same boundary. This is not unusual in situations where p > 2. Shown
on the left is the disconnection point when the network is formulated in mesh currents. In this formulation, r = 1. Shown on the right
is the same disconnection point, but formulated in node voltages,
which results in r = 4. As noticed, the values of r in each case are
different.
The value of r has a negative impact to  in (13). Since  is dense,
the (1) frequent LU re-factorization (e.g., due to switching) and (2)
the forwardbackward solutions at each time step (two algorithms
of O(r2 ) each) suggests that it is prudent to keep r smallespecially
if p is to scale with the number of available parallel processing units.
As a supporting case to demonstrate that r depends on the
formulation choice, consider a hypothetical power system (rather
difcult to convey visually) with the disconnection points described
by Table 1. If such power system is grounded, the number of boundgnd
ary variables rmesh in a mesh formulation is given by (14); if the
power system is ungrounded, the number of boundary variables
gnd
rmesh is given by (15). For the same grounded power system, a nodal
gnd

formulation results in a number of boundary variables rnodal given


in (16).
Comparing the values of r in (14), (15), and (16), mesh formulations can demand less time to solve the boundary equations in

Table 1
List of disconnection points for an arbitrary partitioning case of p = 10 partitions.
Disconnection
point i (di )

Number of
conductors at di (Ni )

Number of subsystems
interfaced at di (si )

1
2
3
4
5

3
3
3
3
3

5
10
2
6
3

partitioned simulations. This nding is rarely recognized as important, and is commonly obfuscated by the ease in which nodal
matrices are formed instead [12].
gnd

rmesh =

d


Ni = (3 + 3 + 3 + 3 + 3) = 15

(14)

i=1

ungnd

rmesh =

d


(Ni 1) = (2 + 2 + 2 + 2 + 2) = 10

(15)

i=1

gnd

rnodal =

d


Ni (si 1) = 3(4 + 9 + 1 + 5 + 2) = 63

(16)

i=1

4. Conclusions
Tearing networks does not require boundary branches
as originally proposed by Kron. Tearing meshes and nodes
(zero-immittance tearing) increases the number of possible disconnection points (or boundaries) to choose from;
this is an important consideration for graph-theoretic
graph
segregations
algorithms,whichfurthermorepermits
as ne-grained as one power apparatus per subsystem.

Please cite this article in press as: F.M. Uriarte, On Krons diakoptics, Electr. Power Syst. Res. (2012), doi:10.1016/j.epsr.2012.01.016

G Model
EPSR-3456; No. of Pages 5

ARTICLE IN PRESS
F.M. Uriarte / Electric Power Systems Research xxx (2012) xxxxxx

There are several situations where mesh formulations results


in less computational effort (i.e., a smaller r) to solve the boundary
network: a well-known bottleneck in diakoptics-based partitioning. This is an important consideration that is often overlooked, and
merits attention early in the software design stage when deciding
between nodal and mesh formulations.
The results presented by this work are applicable to multicore
electromagnetic transient simulations of large, complex power systems [6]. Tearing meshes and nodes (instead of branches) results
in a better selection of disconnection points when using graph theory to determine where to partition a power system. Choice of the
correct (or best-suited) formulation method for a power system
simulation problem results in faster multicore (parallel) simulation results due to the reduced number of boundary variables (r).
While no one formulation method is best-suited for all problems,
the formulation choice should be assessed by computer programs
before running parallel simulations.
References
[1] G. Kron, Diakoptics, The Piecewise Solution of Large-Scale Systems, MacDonald
& Co., London, 1963.
[2] A. Brameller, M.N. John, M.R. Scott, Practical Diakoptics for Electrical Networks,
Chapman & Hall, London, 1969.
[3] I.S. Duff, A.M. Erisman, J.K. Reid, Direct Methods for Sparse Matrices, Oxford
University Press, Oxford, 1986.
[4] H.V. Henderson, S.R. Searle, On deriving the inverse of a sum of matrices, SIAM
Rev. 23 (1981) 5360.

[5] A. Klos, What is diakoptics? Int. J. Elec. Power 4 (1982) 192195.


[6] F.M. Uriarte, R.E. Hebner, A.L. Gattozzi, Accelerating the simulation of shipboard
power systems, in: Grand Challenges in Modeling & Simulation, The Hague,
Netherlands, 2011.
[7] F.M. Uriarte, Multicore simulation of an ungrounded power system, IET Elec.
Syst. Trans. 1 (2011) 3140.
[8] Z. Quming, S. Kai, K. Mohanram, D.C. Sorensen, Large power grid analysis using
domain decomposition, in: Design, Automation and Test in Europe, 2006. DATE
06. Proceedings, 2006, pp. 16.
[9] G. Karypis, V. Kumar, hMETIS: a Hypergraph Partitioning Package, Department
of Computer Science & Engineering, University of Minnesota, Minneapolis, MN,
1998.
[10] T.A. Davis, Direct Methods for Sparse Linear Systems, SIAM, Philadelphia, 2006.
[11] F.M. Uriarte, K.L. Butler-Purry, Diakoptics in shipboard power system simulation, in: North American Power Symposium (NAPS), Southern Illinois
University, Carbondale, 2006, pp. 201210.
[12] F.M. Uriarte, A tensor approach to the mesh resistance matrix, IEEE Trans. Power
Syst. 26 (2011) 19891997.
[13] P.W. Aitchison, A. Klos, Network tearing an alternative method, Int. J. Elec.
Power 13 (1991) 308320.
[14] A. Sangiovanni-Vincentelli, C. Li-Kuan, L.O. Chua, A new tearing approach
node tearing nodal analysis, in: IEEE International Symposium on Circuits and
Systems, 1975, pp. 143147.
[15] C. Yue, X. Zhou, R. Li, Node-splitting approach used for network partition and
parallel processing in electromagnetic transient smiulation, in: International
Conference on Power System Technology, Singapore, 2004.
[16] B.M. Zhang, N.A. Xiang, S.Y. Wang, Unied piecewise solution of power-system
networks combining both branch cutting and node tearing, Int. J. Elec. Power
11 (1989) 283288.
[17] A. Kalantari, S.M. Kouhsari, An exact piecewise method for fault studies in
interconnected networks, Int. J. Elec. Power 30 (2008) 216225.
[18] S. Esmaeili, S.M. Kouhsari, A distributed simulation based approach for detailed
and decentralized power system transient stability analysis, Elec. Power Syst.
Res. 77 (2007) 673684.

Please cite this article in press as: F.M. Uriarte, On Krons diakoptics, Electr. Power Syst. Res. (2012), doi:10.1016/j.epsr.2012.01.016

You might also like