Electrical Power and Energy Systems xxx (2011) xxx–xxx

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

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

Tie-line constrained distributed state estimation

Kalyan Dasgupta, K.S. Swarup ⇑
Department of Electrical Engineering, Indian Institute of Technology, Madras, Chennai 600 036, India

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

Article history: This paper presents an implementable distributed state estimation method for online analysis of power
Received 22 February 2008 systems having multiple, geographically separated areas. Distributed state estimation apart from giving a
Received in revised form 4 September 2010 faster solution, also improves the condition number of the resultant gain matrices. The method proposed
Accepted 3 December 2010
here uses the conventional WLS estimator (Gauss–Newton method) with equality constraints forcing the
Available online xxxx
tie line flows as calculated by the adjacent areas to be equal. Based on the topology of the network, the
system is partitioned into multiple areas and a processor is assigned to each of these areas for solving the
Keywords:
local state estimation problem. Simulations carried out on the IEEE 14, 30 and 118 bus systems show
State estimation
Distributed computing
good convergence properties and improvement on the condition number of the gain matrices when com-
Speedup pared with centralized algorithms.
Condition number Ó 2011 Elsevier Ltd. All rights reserved.

1. Introduction of sending it to a central control centre located hundred kilometers

State estimation is used in all Energy Management Systems
(EMS) to identify the present operating state of a system [1–4].
Once the estimates of the state variables are known, proper ac-
tions, if required (during emergency, normal insecure states), can
be taken to bring the system back to its normal secure state. For
proper monitoring of the system, the intervals at which the telem-
etry data are collected and the estimates of the state variables
made must be very less. At present state estimation in power
system control centre is done every ten to fifteen minutes. As the
size of a power system increases, collecting data and solving the
state estimation problem in a very short time in one control centre,
not only becomes very difficult, but also requires extra investment
in setting up long telemetry and communication lines. Besides,
policy and privacy considerations among companies operating in
a transmission system also make centralized state estimation
unsuitable.
Distributed state estimation (DSE) is one way of solving the
above mentioned problems. Shahidehpour and Wang in [4],
Tylavsky and Bose in [6] and Ramesh in [7] discuss the uses and is-
sues concerning parallel and distributed computing in online
power system analysis. In a distributed system, instead of having
one control centre, the system has multiple control centres located
in geographically separated areas. Each of these centres is
equipped with tracking instruments and software to collect data
and estimate the states of the buses local to it. Under such a dis-
tributed environment, the Remote Terminal Units (RTUs) collect
data and sends it to the control centre that is local to it; instead
⇑ Corresponding author. Tel.: +91 44 2257 4440; fax: +91 44 2257 4402.
E-mail address: swarup@ee.iitm.ac.in (K.S. Swarup).
away. The availability of advanced distributed computing technol-
ogies, both hardware and software makes this kind of real time
analysis possible. Since the size of the local state estimation prob-
lem is much smaller than the very large master problem, the time
taken to solve it and the number of iterations for it to converge is
also less when compared to the master problem. Apart from saving
in time and investment, distributed state estimation has certain
numerical advantages as well. The gain matrix of a large state esti-
mation problem is a severely ill-conditioned matrix. Gu et al. in [5]
discusses about the solution of large ill-conditioned power system
state estimation problems. Once the problem becomes smaller, the
ill-conditioning of the gain matrix decreases and the state variables
become less sensitive to errors in the measured values.
Distributed state estimation has been an area of research for
quite sometime. Some of the proposed methods of solving a
distributed state estimation problem are given in [8–15]. Wu and
Falcao in [8], Carvalho and Barbosa in [9], and Ebrahimian and
Baldick in [10] propose methods based on two adjacent areas shar-
ing a boundary bus. Lin and Lin in [11] proposed a hierarchical, tier
based state estimator [12]. In [13], Lin proposes a method based on
the recursive quadratic programming on the dual (RQPD) method.
In [14], Zhao and Abur propose a two level state estimator using
synchronized phasor measurements. The method proposed here
uses the conventional decoupled (P–h and Q–V), Gauss–Newton
method with equality constraints forcing the tie line flows as cal-
culated by the adjacent areas to be equal. It follows [8] very closely
because of its simplicity and effectiveness. However, in [8], two
areas are made to share a bus and the equality constraints forces
the state variables of the shared bus as calculated by the two areas
to be equal. An important advantage of the proposed method is
that the buses in one area are not referenced to a slack bus placed

0142-0615/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijepes.2010.12.010

Please cite this article in press as: Dasgupta K, Swarup KS. Tie-line constrained distributed state estimation. Int J Electr Power Energ Syst (2011),
doi:10.1016/j.ijepes.2010.12.010
2 K. Dasgupta, K.S. Swarup / Electrical Power and Energy Systems xxx (2011) xxx–xxx

in some other area. As a result, the observability of one area is

independent of the observability of other areas. Another advantage
lies in the fact that, the injection measurements of the boundary
buses are made use of, which is not so in the case of algorithms
where a bus is shared by two areas. A study of the speedup of solu-
tion time and the improvement in the condition number of the
gain matrices are also presented.
This paper is divided into four sections. The second section
deals with the mathematical formulation of the problem. The third
section deals with the issues relating to the implementation of the
distributed problem and also presents the results. Important con-
clusions are presented in the fourth section.
Fig. 1. Boundary bus connections.
2. Mathematical formulations of state estimation algorithms

2.1. Centralized state estimation model

The standard measurement model of a state estimation prob-

lem is given by
z ¼ hðxÞ þ e ð1Þ
where z = (m  1) measurement vector, x = (n  1) vector of state
variables, h = (m  1) vector of non-linear functions relating the
vectors z and the state vector x, e = (m  1) vector of measurement
errors, m = number of measurements and n = number of state
variables.
The measurements include the voltage magnitudes, real and
reactive power injections and flows. The state variables are the
bus voltage magnitudes V and their relative phase angles h.
The decoupled state estimator model, as given by Monticelli
Fig. 2. Extended area m and area n.
and Garcia in [16], is given by
Z p ¼ hp ðxÞ þ ep ð2Þ
The state estimate of the above system can be obtained by solv-
Z q ¼ hq ðxÞ þ eq ð3Þ ing for the Weighted Least Squares (WLS) problem for the individ-
ual areas (m and n) with linear equality constraints forcing the
where subscripts p and q are used to indicate the real and reactive flows Pij and Qij as calculated by areas m and n to be equal.
part of the measurement model. The objective function for the above mentioned two area sys-
The decoupled WLS estimator works by minimizing the follow- tem can be written as
ing objective function
X 1h iT h i1 h i
1 1 JðxÞ ¼
k
zkp  hp ðxÞ Rkp
k
zkp  hp ðxÞ
JðxÞ ¼ ½zp  hp ðxÞT  R1 T 1
p  ½zp  hp ðxÞ þ ½zq  hq ðxÞ  Rq 2
2 2 k¼m;n

 ½zq  hq ðxÞ ð4Þ 1 h iT h i1 h i

k k
þ zkq  hq ðxÞ Rkq zkq  hq ðxÞ ð7Þ
2
Rp and Rq are diagonal matrices consisting of the variances of
real and reactive power measurements respectively. The solution such that
is reached by iteratively solving for the following equations
m n
T 1 Pm n
ij  P ij ¼ 0 and Q ij  Q ij ¼ 0 ð8Þ
Dh ¼ G1
p  H p Rp  ½zp  hp ðh; VÞ ð5Þ
m
T 1
where P m ij and Q ij are the power flows calculated by area m and
DV ¼ G1
q  H q Rq  ½zq  hq ðh; VÞ ð6Þ n n
Pij and Q ij ; the flows calculated by area n.
where Gp = HTp R1 T 1 All other symbols have their usual meaning as defined in Eqs.
p Hp and, Gq ¼ Hq Rq H q are the gain matrices of the
@hp @h
decoupled equations, Hp ¼ @h and Hq ¼ @Vq . (1)–(4), with the superscript indicating the area where the variable
is calculated.
2.2. Distributed state estimation formulation Now assuming Gij << Bij and hi  hj, Pij and Qij can be written as:

Pij ¼ V i V j Bij ðhi  hj Þ ð9Þ

The proposed method is explained here by taking an example of
two areas sharing a tie-line. Fig. 1 show areas m and n sharing the Q ij ¼ V i ðBij ðV i  V j Þ  V i ysh Þ ð10Þ
tie line connecting buses i and j. Bus i is the internal boundary bus
and bus j is the external boundary bus of area m and vice versa for where Gij and Bij are the real and imaginary part of the (i, j) th ele-
area n. Area m can be extended to include bus j and area n can be ment of the admittance matrix, ysh is the half line charging admit-
extended to include bus i as shown in Fig. 2. When estimating the tance of the line i–j.
state variables of bus j the control centre in area m makes use of If V m m n n
i V j ¼ V i V j , ViVj in Eq. (9) would no longer play a part in the
the injection measurements of bus i and the flow measurements final derivation. Similarly, if V m n
i ¼ V i , the outermost Vi in Eq. (10)
Pij and Qij (active and reactive power flow) in the line connecting would no longer play a part in the final derivation.
bus i and j. Hence, ignoring ViVj in Eqs. (9) and (8) can be written as

Please cite this article in press as: Dasgupta K, Swarup KS. Tie-line constrained distributed state estimation. Int J Electr Power Energ Syst (2011),
doi:10.1016/j.ijepes.2010.12.010
 K. Dasgupta, K.S. Swarup / Electrical Power and Energy Systems xxx (2011) xxx–xxx 3

Pm n m m n n m m n n
ij  P ij ¼ ½Bij ðhi  hj Þ  ½Bij ðhi  hj Þ ¼ Bp h þ Bp h ized updating process. The second stage consisting of Eqs. (19) and
X k k (20), makes sure that the constraints in (8) are satisfied.
¼ Bp h ð11Þ e be the values of state variables after the first stage
Let ~
h and V
k¼m;n
update and let ‘i’ indicate the iteration number. ‘i’ should not be
where hm m n n
i ; hj ; hi and hj are phase angles of buses i and j as calcu- confused with bus i. Therefore
lated by area m and n respectively, Bm n
p and Bp are matrices of size ~hm ði þ 1Þ ¼ hm ðiÞ þ ½Gm 1  ½Hm T ½Rm 1 ½zm  hm ðhm ðiÞ; V m ðiÞÞ ð18Þ
(l  nm) and (l  nn) respectively, h (nm  1) and hn (nn  1) are
m p p p p p

the phase angle vectors of area m and area n respectively, nm and X X

nn are the number of buses in area m and area n respectively and kp ði þ 1Þ ¼ ð Bkp ½Gkp 1 ½Bkp T Þ1 Bkp ~hk ði þ 1Þ ð19Þ
l is the number of tie-lines connecting area m and area n. k¼m;n k¼m;n

In our case l = 1. The non-zero elements of Bkp are either Bij or

1 m T
Bij. hm ði þ 1Þ ¼ ~hm ði þ 1Þ  ½Gm
p  ½Bp  kp ði þ 1Þ ð20Þ
Similarly ignoring Vi in Eq. (10), m T m 1 m
where Gm p ¼ ½Hp  ½Rp  Hp is the gain matrix for the P-h equations of
Qm
ij  Q nij ¼ ½ðBij  ysh ÞV m
i  Bij V m
j   ½ðBij  ysh ÞV ni  Bij V nj  area m. If the off-diagonal elements of Gm p are neglected, then
X only the internal and external boundary buses i.e. buses i and
¼ Bm
q V
m
þ Bnq V n ¼ Bkq V k ð12Þ
j are affected by Eq. (20). Then (20) can be re-written for hm m
i and hj as
k¼m;n

hm ~m
i ði þ 1Þ ¼ hi ði þ 1Þ
Where Bm n
q and Bq are matrices of size (l  nm) and (l  nn) respec-
tively and V (nm  1) and Vn (nn  1) are the voltage magnitude
m
gm ~m ~m ~n ~n
ii ½ðhi ði þ 1Þ  hj ði þ 1ÞÞ  ðhi ði þ 1Þ  hj ði þ 1ÞÞ
vectors of area m and area n respectively. The non-zero elements 
ðg m m n n
ii þ g jj Þ þ ðg ii þ g jj Þ
of Bkq are (Bij  ysh) – (Bij  ysh), Bij and Bij.
ð21Þ
It is to be noted that Pij and Qij in (8) are the values calculated
using the initial estimates of the state variables of the boundary
buses. They are not the measured values. hm ~m
j ði þ 1Þ ¼ hj ði þ 1Þ

The Lagrangian functions Lp and Lq for the decoupled state esti- gm ~m ~m ~n ~n

jj ½ðhj ði þ 1Þ  hi ði þ 1ÞÞ  ðhj ði þ 1Þ  hi ði þ 1ÞÞ
mation problem are given by 
ðg m m n n
ii þ g jj Þ þ ðg ii þ g jj Þ
X 1
Lp ¼
k k
½zkp  hp ðhk ; V k ÞT ½Rkp 1 ½zkp  hp ðhk ; V k Þ þ kTp ð22Þ
k¼m;n
2
X k k where g m m
ii and g jj are the diagonal elements of the inverse gain ma-
 Bp h ð13Þ trix (½Gm
p 1
) of area m corresponding to buses i and j respectively.
k¼m;n Similarly g nii and g njj are the diagonal elements of ½Gnp 1 corre-
sponding to buses i and j respectively.
and
The voltage magnitudes V also get updated in the same fashion.
X 1 k k However for updating V, the recently updated values of h
Lq ¼ ½zkq  hq ðhk ; V k ÞT ½Rkq 1 ½zkq  hq ðhk ; V k Þ þ kTq (hm ði þ 1Þ) are used.
k¼m;n
2
X k k
 Bq V ð14Þ ~ m ði þ 1Þ ¼ V m ðiÞ þ ½Gm 1  ½Hm T ½Rm 1 ½zm  hm ðhm ði
V q q q q q
k¼m;n
þ 1Þ; V m ðiÞÞ ð23Þ
Now the necessary conditions for the solution of (13) and (14)
as seen by area m are, X X
kq ði þ 1Þ ¼ ð Bkq ½Gkq 1 ½Bkq T Þ1 Bkq V~ k ði þ 1Þ ð24Þ
@Lp T m 1 m m m m T k¼m;n k¼m;n
¼ ½Hm m
p  ½Rp  ½zp  hp ðh ; V Þ þ ½Bp  kp ¼ 0 ð15Þ
@hm
~ m ði þ 1Þ  ½Gm 1 ½Bm T kq ði þ 1Þ
V m ði þ 1Þ ¼ V ð25Þ
q q
@Lq T m 1 m m m m T
¼ ½Hm m
q  ½Rq  ½zq  hq ðh ; V Þ þ ½Bq  kq ¼ 0 ð16Þ
@V m where Gm m T m 1 m
q ¼ ½Hq  ½Rq  Hq is the gain matrix for the Q–V equations
of area m.
@Lp X k k @Lq X k k
¼ B h ¼ 0 and ¼ B V ¼ 0: ð17Þ Again in Eq. (21), if the off-diagonal elements of Gm q are ne-
@kp k¼m;n p @kq k¼m;n q glected, only V m m
i and V j get affected by Eq. (25). Then (25) can
be re-written for V ni and V nj as
where kp and kq are vectors of Langrangian multipliers of size
(l  1), Vm ~m
i ði þ 1Þ ¼ V i ði þ 1Þ

gm ~m ~n ~m ~n
m m ii ðBij  ysh Þ½ðBij  ysh ÞðV i ði þ 1Þ  V i ði þ 1ÞÞ  Bij ðV j ði þ 1Þ  V j ði þ 1ÞÞ
@hp @hq 
Hm
m ðBij  ysh Þ2 ðg m n 2 m n
ii þ g ii Þ þ Bij ðg jj þ g jj Þ
¼ m andHq ¼ :
p
@h @V m ð26Þ
For a fast decoupled state estimation analysis Hkp and Hkq are cal-
Vm ~m
culated at the nominal point before the analysis is started and kept j ði þ 1Þ ¼ V j ði þ 1Þ
constant throughout the iteration process. gm ~m ~n ~m ~n
jj Bij ½Bij ðV j ði þ 1Þ  V j ði þ 1ÞÞ  ðBij  ysh ÞðV i ði þ 1Þ  V i ði þ 1ÞÞ

ðBij  ysh Þ2 ðg m n 2 m n
ii þ g ii Þ þ Bij ðg jj þ g jj Þ
2.3. Proposed distributed state estimation algorithm
ð27Þ

The solution of Eqs. (15)–(17) can be obtained by updating the where g m m

ii and g jj are the diagonal elements of the inverse gain ma-
m 1
state variables in a two stage process by using the Gauss–Newton trix (½Gq  ) of area m corresponding to buses i and j respectively,
method combined with the fast decoupled load flow assumptions. and similarly with elements g nii and g njj for area n.
The first stage, given by Eq. (18), is similar to the standard central- If Bij >> ysh; Eq. (26) can be written as:

Please cite this article in press as: Dasgupta K, Swarup KS. Tie-line constrained distributed state estimation. Int J Electr Power Energ Syst (2011),
doi:10.1016/j.ijepes.2010.12.010
4 K. Dasgupta, K.S. Swarup / Electrical Power and Energy Systems xxx (2011) xxx–xxx

Vm ~m
i ði þ 1Þ ¼ V i ði þ 1Þ
The solution given above is specific to area m. The solution for
~m area n is exactly the same as given by Eqs. (18)–(29), with only
gm
ii ½ðV i ði þ 1Þ  V~ m ~n ~n
j ði þ 1ÞÞ  ðV i ði þ 1Þ  V j ði þ 1ÞÞ
 the superscripts m and n interchanged. It is to be noted that the
ðg m m n n
ii þ g jj Þ þ ðg ii þ g jj Þ constraints in (8) are unique in the fact that the flows Pij and Qij
ð28Þ and not Pji and Qji, were chosen. Eqs. (26) and (27) will have a direct
bearing on this. However, if Bij >> ysh, (8) no longer remains unique
and (27) as: i.e. even if Pji and Qji are used in (8), instead of Pij and Qij, the equa-
Vm ~m tions would still remain unchanged. The whole process can be rep-
j ði þ 1Þ ¼ V j ði þ 1Þ
resented by a flowchart shown in Fig. 3.
gm ~m ~m ~n ~n
jj ½ðV j ði þ 1Þ  V i ði þ 1ÞÞ  ðV j ði þ 1Þ  V i ði þ 1ÞÞ As shown in the flowchart, the state estimation calculations in

ðg m m n n
ii þ g jj Þ þ ðg ii þ g jj Þ all the control centres are initiated by a Central Control Centre.
ð29Þ Once an area converges, a flag is set to 1 and the other areas no
longer communicate with the area that has converged. The state

Initiate state estimation calculation in all the control centres

Start calculation Start calculation

for Area m for Area n

Read Measurement Data Read Measurement Data

Initialize the values of state Initialize the values of state

variables. Set flag (m) = 0 variables. Set flag (n) = 0

Update θm using (18) Communication Update θn using (18)

b/w processors
B Exchange values of flag Exchange values of flag A

No Flag (m) = 0 No
Flag (n) = 0
? ?
Yes Yes

Pass on values of θ m
i - θ m
j , g iim + g mjj Acquire values of θ im - θ jm , g iim + g mjj
Acquire values of θ in - θ jn , g iin + g njj Pass on values of θ in - θ jn , g iin + g njj

Update θ i m and θjm using (21) and (22) Update θ in and θjn using (21) and (22)

Update Vm using (23) Update Vn using (23)

No Flag (n) = 0 Flag (m) = 0 No

? ?
Yes

Pass on values of Vi - V jm , g iim + g mjj

m
Acquire values of Vi m - V jm , g iim + g mjj

Acquire values of Vi n - V jn , g iin + g njj Pass on values of Vi n - V jn , g iin + g njj

Update Vi m and V jm using (26) and (27) Update Vi n and V jn using (26) and (27)

No Max (θ m & V m) < ε Max (θn & V n) > ε No

? ?
Yes Yes

Set flag (m) = 1 A B Set flag (n) = 1

Stop Stop

Fig. 3. Flowchart for distributed state estimation.

Please cite this article in press as: Dasgupta K, Swarup KS. Tie-line constrained distributed state estimation. Int J Electr Power Energ Syst (2011),
doi:10.1016/j.ijepes.2010.12.010
 K. Dasgupta, K.S. Swarup / Electrical Power and Energy Systems xxx (2011) xxx–xxx 5

variables of the boundary buses that belong to other areas, but

shared with the area that has converged, are no longer updated
by any of the equations. In Fig. 3, a two area example has been ta-
ken and the assumption Bij >> ysh has been considered. When there
are multiple areas, the same protocol would be followed by each of
the areas.
Fig. 3 only shows the way the state variables are estimated. Bad
data detection however needs to be done before the estimation is
initiated. The process of bad data detection is independent in each
of the areas and requires no intervention (information) from other
areas. Bad data in the measurements related to the boundary buses
may lead to erroneous estimates of the state variables of the
boundary buses. To reduce the impact of bad data on the estimates
of the boundary buses, the redundancy of the local area adjoining
the tie lines has to be high. In case the boundary buses become
unobservable due to bad data elimination, adjacent areas (control
centres) may exchange tie line flow measurements to make them Buses Area 1 Area 2 Area 3 Area 4
observable.
Internal 1, 2, 3, 11, 20, 13, 14, 25, 26, 27,
Buses 5, 7 21 15 29, 30
Internal 9, 10, 17, 12, 16,
Boundary 4, 6 8, 24, 28
3. Implementation and results 19, 22 18, 23
Buses
External 8, 9, 10, 6, 16, 18, 4, 17,
3.1. Implementation issues Boundary 6, 22, 23
12, 28 24 19, 24
Buses
Experiments were carried out using Linux workstations con-
nected via 100Mbps Ethernet. The simulation program is coded Fig. 4. Partitioning the IEEE 30 bus system.
in C. Inter-processor communication was done through LAM/MPI
[17,18]. Each system had its own memory from which it accessed
data (distributed memory) [19,20]. The tests were done on IEEE 14, For communication with processors belonging to other areas,
IEEE 30 and IEEE 118 bus systems [21]. every computer keeps a schedule. The schedule gives information
Table 1 gives the details of the partitioning of the different test about which processor to communicate with and when and also
systems. It also gives details on the number of boundary buses and the number of tie lines for which to exchange data. The schedule
tie lines present in each system. Partitioning of the systems into should be optimized in a way such that queuing effect is mini-
different areas was done by visual inspection, keeping in mind that mized. Fig. 5 shows the way the processors communicate in the
the inter-area tie lines should be as less as possible. Columns 7 and IEEE 30 bus system. The schedule can be broken down into three
8 give the internal and external boundary buses of each area. There states. In state 1, Proc (Processor) 1 communicates with Proc 2
are cases where one internal boundary bus has tie lines with multi- for two tie lines and Proc 3 communicates with Proc 4 for one tie
ple external boundary buses and vice versa. Fig. 4 shows partition- line. Similarly, in state 2, Proc 1 communicates with Proc 3 for
ing of the IEEE 30 bus system. one tie line and Proc 2 communicates with Proc 4 for one tie line
In the distributed implementation, unlike in the centralized and so on. Note that when Proc 1 communicates with Proc 2, Proc
case (where there is a single system wide reference slack bus), 3 does it with Proc 4. Proc 3 is not made to wait for Proc 1 to finish
every area has its own reference slack bus. However, in order to communication with Proc 2; instead it utilizes this time to commu-
compare the distributed results with the centralized ones, a refer- nicate with Proc 4. When large numbers of processors are involved,
ence angle was assumed in each area and all the bus angles (of that the schedule has to be made in such a way that this waiting time is
area) were shifted by an amount such that the angle in one chosen minimized.
bus matches the assumed reference angle (for that area). Since the
bus voltage angles of an area are not dependent on a reference bus
angle based in another area, the observability of the gain matrix of 3.2. Convergence characteristics
an area is independent of the observability of other areas. Hence
for the distributed implementation to work, each area has to be The results for all the test cases carried out for distributed state
independently observable. The measurement data was taken from estimation analysis show good convergence properties. Table 2
steady state load flow analysis results, with some normally distrib- gives the average number of iterations taken by each of the distrib-
uted errors introduced. uted systems and their comparison with the centralized case. The
The amount of data communicated between two areas sharing 4th column in Table 2 gives the average number of iterations taken
tie lines is quite less and is as much as or lesser than as given in by the distributed implementation for all the tolerance limits. The
other algorithms. The number of data communicated per iteration, number of iterations in the distributed implementation of the IEEE
as shown in Fig. 3, is four times the number of tie lines shared. 118 bus system is less than the centralized state estimation.

Table 1
Details of test cases and their partitioning.

Test cases No. of buses No. of Areas Buses in each area Branches in each area No. of tie-lines Internal boundary buses External boundary buses
IEEE 14 14 2 7, 7 8, 9 3, 3 3, 2 2, 3
IEEE 30 30 4 7, 8, 7, 8 9, 8, 7, 8 5, 5, 4, 4 2, 5, 4, 3 5, 4, 4, 3
IEEE 118 118 6 22, 21, 17, 18, 25, 15 29, 25, 24, 27, 33, 16 7, 8, 14, 8, 13, 14 5, 6, 6, 5, 6, 6 7, 6, 12, 3, 7, 7

Please cite this article in press as: Dasgupta K, Swarup KS. Tie-line constrained distributed state estimation. Int J Electr Power Energ Syst (2011),
doi:10.1016/j.ijepes.2010.12.010
6 K. Dasgupta, K.S. Swarup / Electrical Power and Energy Systems xxx (2011) xxx–xxx

2 1
1.
Proc 1 Proc 2 Proc 3
1 1
2. Proc 1 Proc 3 Proc 2
2 2
3. Proc 1 Proc 4 Proc 2
Fig. 5. Schedule of communication between processors.
3.3. Speedup and improvement in condition number
Speedup can be categorized into (i) iterative speedup and (ii)
overall speedup. Iterative speedup is the speedup achieved in the
iterative part of the analysis. Overall speedup is the speedup
achieved when both the iterative and the non-iterative times are
considered. The non-iterative part of the analysis involves reading
measurement data, calculation of the gain matrix and its factoriza-
tion etc. Fig. 6 shows the overall time taken for analysis versus the
number of buses in both the centralized and distributed cases. The
dashed line in Fig. 6 is the plot of the time taken for the distributed
implementation versus the number of buses. Fig. 7 gives the bar
graph of overall speedup and iterative speedup versus number of
buses.
As can be seen from Figs. 6 and 7, the speedup (overall and iter-
ative) for the IEEE 14 bus system is less than 1 i.e. the time taken
Proc 4
for the distributed implementation of the IEEE 14 bus system is
more than the time taken for its centralized counterpart. This can
Proc 4 be attributed to the fact that the fraction of time taken for inter-
area communications in small systems make up the major part
of the total time taken for analysis. In large systems the time taken
Proc 3 for calculation of the state variables makes up the major part. Table
3 gives the speedup results as well as the results for the condition
number.
The condition number of the gain matrix [1,5] in the state esti-
mation analysis gives us an idea about the sensitivity of the solu-
tion vector to the measurement residual vector and also to errors
in the coefficient matrix (gain matrix). The condition number of
the gain matrix is given by
jðGÞ ¼ kGk  kG1 k: ð30Þ
where ||G|| is the norm of G given by
!
X
kGk ¼ max jGij j ð31Þ
i
j
Where Gij is the (i, j) th element of the gain matrix G.

Table 2
Performance of proposed algorithm.

Test Tolerance Average number of Average

cases limit iterations redundancy
Centralized Distributed
IEEE 14 0.01 2 3 3.05
0.001 3 5
0.0001 5 6.5
IEEE 30 0.01 2 3 2.65
0.001 3 4.5
0.0001 5 6
IEEE 118 0.01 3 3 3.02
0.001 5 4
0.0001 7 5.2
Fig. 7. Speedup versus No. of buses.

Fig. 6. Overall analysis time versus No. of buses.

Please cite this article in press as: Dasgupta K, Swarup KS. Tie-line constrained distributed state estimation. Int J Electr Power Energ Syst (2011),
doi:10.1016/j.ijepes.2010.12.010
 K. Dasgupta, K.S. Swarup / Electrical Power and Energy Systems xxx (2011) xxx–xxx 7

Table 3 as their centralized counterparts. This in turn leads to reduction in

Speedup and condition number of gain matrices. the condition number.
Test cases Iterative speedup Overall speedup Condition number
Distributed Centralized
IEEE 14 0.289 0.565 1012 1171
3.4. Accuracy measurement
IEEE 30 0.6397 2.316 2725 12,600
IEEE 118 6.340 7.651 70,741 614,731
The accuracy of the distributed state estimator (DSE) was mea-
sured by calculating the maximum deviations (error) in the esti-
mated phase angles and voltage magnitudes of the boundary
Very high condition numbers can lead to erroneous results,
buses from their true values. The l1 norm of the error vector
especially when there are errors in the measurement residual vec-
(kx  ^
xk1), i.e. the summation of the deviations of the estimated
tor or when there are presence of bad data. As can be seen from Ta-
values (^x) from the true values (x), was also calculated. These re-
ble 3, the difference in the condition numbers of the distributed
sults were compared with the results of a completely independent
and the centralized cases increase when the system size increases.
area state estimator (ASE). Table 4 gives the results of the accuracy
This phenomenon can be attributed to the fact that the state vari-
measurements. The unit for the voltage values is per unit (pu)
ables in an area in the distributed case are not influenced by the
while the unit for the angle values is degrees. The errors in the esti-
changes in the state variables of other areas, which is not the case
mated values of the internal buses are negligible and hence not
with the centralized implementation. In the centralized case the
shown. The maximum error in the estimated values for the phase
changes in the state variables of a bus is sensitive to changes in
angles and voltage magnitudes are represented by dh and dv
the state variables of all the other buses. All these interactions
respectively, while the l1 norm for the error vectors are represented
though weak tend to increase the condition number of the gain
by lh and lv.
matrix, making the solution vector more sensitive in nature. Hence
From Table 4 it can be observed that there is a large improve-
the values of ||G|| and ||G1|| in the distributed case are not as large
ment in the accuracy of the results in the external boundary buses
for the distributed state estimator (DSE), when compared with the
area state estimator (ASE) as the system size increases. The accu-
Table 4
Accuracy measurements. racy of the results for the internal boundary buses are almost the
same for both the cases. When compared with the results of the
Test cases Errors Internal boundary buses External boundary buses
centralized algorithm the distributed results for the boundary
DSE ASE DSE ASE buses are fairly accurate. The results of the centralized and distrib-
IEEE 14 dh 0.0060 0.0105 0.0176 0.0160 uted algorithm for the IEEE 30 bus system is given in Figs. 8 and 9.
dv 0.0004 0.0004 0.0010 0.0014 The coloured cells in Figs. 8 and 9 indicate that the considered bus
lh 0.0143 0.0203 0.0335 0.0401
is an external boundary bus for that particular area. Column 1 of
lv 0.0008 0.0008 0.0021 0.0026
IEEE 30 dh 0.0492 0.0506 0.0726 0.0797
the figures gives the bus numbers for all the boundary buses. Col-
dv 0.0036 0.0034 0.0068 0.0067 umn 2 of the figures gives the true values of the state variables
lh 0.2036 0.2129 0.2134 0.2296 used for these boundary buses. Columns 8, 9 and 10 give the differ-
lv 0.0247 0.0252 0.0368 0.0389 ence of the estimates from the true values for the distributed and
IEEE 118 dh 0.5854 0.5854 1.0648 1.6658
centralized cases.
dv 0.0341 0.0296 0.0493 0.1540
lh 5.3666 5.6193 9.5364 12.1112 The distributed results for the internal boundary buses are
lv 0.2026 0.1775 0.6014 1.0010 found to be better than the centralized results. The distributed re-
sults for the external boundary buses show a little more deviation
DSE – Distributed state estimator. ASE – Area state estimator. dh – Maximum error
in phase angles. dv – Maximum error in voltage magnitudes. lh and lv – l1 norm of the from their true values. However, as the system size increases (as
error vector in phase angles and voltage magnitudes. was observed for the IEEE 118 bus system) the results for the

Fig. 8. Phase angle results for the IEEE 30 bus system.

Please cite this article in press as: Dasgupta K, Swarup KS. Tie-line constrained distributed state estimation. Int J Electr Power Energ Syst (2011),
doi:10.1016/j.ijepes.2010.12.010
8 K. Dasgupta, K.S. Swarup / Electrical Power and Energy Systems xxx (2011) xxx–xxx
Fig. 9. Voltage magnitude results for the IEEE 30 bus system.
external boundary buses also tend to be better than the centralized
results.
4. Conclusions
An algorithm by which a large centralized state estimation
problem can be decomposed into multiple distributed state esti-
mation problems is presented. The proposed algorithm provided
good speedups, as well as good results on the estimates of the
boundary bus variables. The improvement in the condition number
of the gain matrices also gives the distributed implementation a
numerical advantage over centralized state estimators. The imple-
mentation was done in a synchronous environment.
This kind of an algorithm fits well in a deregulated environment
where transmission companies have their own control centres.
With improvements in phasor measurement technologies, phasor
data obtained from phasor measurement units (PMUs) also opens
up new areas, whereby, faster and more accurate estimates can
be obtained in a distributed environment. Control centres can ex-
change data amongst each other without compromising on their
privacy policy. However under such a deregulated environment
the protocols for communication between different control centres
need standardization.
References
[1] Abur A, Exposito AG. Power system state estimation. Marcel Dekker, Inc.; 2004.
[2] Wood AllenJ, Wollenberg BruceF. Power generation operation and
control. John Wiley & Sons; 1996.
[3] Grainger JohnJ, Stevenson WD. Power system analysis. McGraw Hill
International; 1994.
Please cite this article in press as: Dasgupta K, Swarup KS. Tie-line constrained distributed state estimation. Int J Electr Power Energ Syst (2011),
doi:10.1016/j.ijepes.2010.12.010
[4] Shahidehpour M, Wang Y. Communication and control in electric power
systems. Wiley Interscience; 2003.
[5] Gu JW, Clements K, Krumpholz GR, Davis PW. The solution of III- conditioned
power system state estimation problems via the method of Peters and
Wilkinson. IEEE Trans Power App Syst 1983;PAS-102(10):3473–80.
[6] Tylavsky Daniel J, Bose Anjan. Parallel processing in power systems. IEEE Trans
Power Syst 1992;7(2):629–37.
[7] Ramesh VC. On distributed computing for on-line power system applications.
Electr Power Energy Syst 1996;18(8):527–33.
[8] Wu FF, Falcao DM. Parallel and distributed state estimation. IEEE Trans Power
Syst 1995;10(2):724–30.
[9] Jose Beleza Carvalho, Maciel Barbosa F. Distributed processing in power
system state estimation. In: 10th mediterr electrotech conf, melecon, vol. III;
2000. p. 1128–31.
[10] Ebrahimian Reza, Baldick Ross. State estimation distributed processing. IEEE
Trans Power Syst 2000;15(4):1240–6.
[11] Lin SY, Lin CH. An implementable distributed state estimator. IEEE Trans
Power Syst 1994;9(3):1277–84.
[12] Van Cutsem T, Ribbens-Pavella M. Critical survey of hierarchical methods for
state estimation of electric power systems. IEEE Trans Power App Syst
1983;102(10):3415–24.
[13] Lin SY. A distributed state estimator for electric power systems. IEEE Trans
Power Syst 1992;7(2):551–7.
[14] Zhao Liang, Abur Ali. Multiarea state estimation using synchronized phasor
measurements. IEEE Trans Power Syst 2005;20(2):611–7.
[15] Conejo AJ, Torre S, Canas M. An optimization approach to multiarea state
estimation. IEEE Trans Power Syst 2007;22(1):213–21.
[16] Monticelli A, Garcia A. Fast decoupled state estimation and bad data detection.
IEEE Trans Power App Syst 1979;PAS-98(5):1645–52.
[17] Pacheco PeterS. Parallel programming with MPI. Morgan Kaufmann
Publishers, Inc.; 1997.
[18] LAM/MPI user’s guide version 7.1.2. The LAM/MPI team open systems lab.
Available: <http://www.lam-mpi.org/>; March 10, 2006.
[19] Bertsekas Dimitri P, Tsitsiklis John N. Parallel and distributed
computation. Printice-Hall; 1989.
[20] Wallach Y. Alternating sequential/parallel processing (Lecture notes in
computer science). Springer-Verlag; 1982.
[21] Power system test archive. Available: <http://www.ee.washington.edu/
research/pstca/>; February, 2007.