JOURNAL OF MODERN POWER SYSTEMS AND CLEAN ENERGY, VOL. 8, NO.

4, July 2020 627

Fully Distributed State Estimation for Power

System with Information Propagation Algorithm
Qiao Li, Lin Cheng, Wei Gao, and David Wenzhong Gao

Abstract—
—In this paper, a new fully distributed state estima‐ and more distributed. As a result, in these distributed power
tion (DSE) based on weighted least square (WLS) method and systems, the conventional centralized methods may not be ef‐
graph theory is proposed for power system. The proposed meth‐ ficient enough to work with the distributed devices. Also,
od is fully distributed so that the centralized facilities, e.g., su‐
pervisory control and data acquisition (SCADA) and central‐ distributed methods are usually more robust than centralized
ized estimators, are not required. Also, different from the exist‐ ones. Therefore, distributed state estimation (DSE) method is
ing DSE methods, the proposed method is a bus-level DSE needed for future power systems.
method, in which the power system is not required to be parti‐ Some efforts have been made in recent years to develop
tioned into several areas. In order to realize the proposed fully the DSE method. The most commonly used method is the
distributed DSE method, a novel information propagation algo‐
rithm is developed in this paper. This algorithm has great po‐
distributed multi-area state estimation as in [11]-[16]. In this
tential in future applications since it is useful to broadcast the type of methods, the entire power system is partitioned into
local information of the nodes to the entire system in a fully dis‐ several small areas. Each area has a local state estimator to
tributed network. The proposed DSE method is compared with estimate the states of its own buses, and communicates with
the conventional centralized state estimation method and exist‐ the nearby estimators to ensure the observability of the esti‐
ing multi-area DSE method in different models in this paper. mation. The state estimation in this method is realized by de‐
The results show that the proposed method has better perfor‐
mance than the traditional methods. composing the centralized state estimation problem into sev‐
eral smaller problems. However, since there are some re‐
Index Terms—
—Consensus protocol, state estimation, smart quirements for the area partitioning such as the size of the
grid communication, graph theory, distributed network.
areas, the way to divide the system and the overlapping be‐
tween the areas, this method may not be compatible with
I. INTRODUCTION many power systems. Also, this method is not fully distribut‐
ed since it has centralized structure (all sensors in an area

S TATE estimation is an important technique for power

system operation. It is usually performed by the energy
management system (EMS) [1]-[3] in control centers to ob‐
are connected to the local estimator) inside each area and
the state estimation algorithm is partially centralized (all da‐
ta in an area are assembled at the estimator for calculation).
tain the states of the power system. The major objectives of In addition, other methods are also proposed in [17], [18] to
state estimation are monitoring the states of power systems, address the DSE problem. However, the centralized facilities
improving measurement accuracy, and detecting bad mea‐ such as global positioning system (GPS) and supervisory
surements. Currently, some state estimation methods have control and data acquisition (SCADA) are required in these
been applied in power systems such as weighted least square methods. Therefore, these methods are not totally distributed
(WLS) method [4] - [7] and Kalman filter [8] - [10]. These as well.
methods have good performance in conventional centralized The latest works on the DSE are [19]-[21]. Reference [19]
power systems. However, due to the development of renew‐ proposes a fully distributed robust bilinear state-estimation
able energy, microgrid, and smart grid, the number of distrib‐ (D-RBSE) method for multi-area power systems. The meth‐
uted generators (DGs) and sensors in power systems grows od is an extension of bi-linear state estimation [22] into
dramatically, thus the future power systems tend to be more multi-area power systems. In this method, the state estima‐
tion problem in each area is solved locally and a small
Manuscript received: March 14, 2019; accepted: September 4, 2019. Date of amount of data is transmitted between different areas to
CrossCheck: September 4, 2019. Date of online publication: April 23, 2020. achieve the globally accurate estimation. In [20], a three-
This work was supported by the U. S. National Science Foundation (No.
1711951).
stage DSE method for AC-DC hybrid distribution network is
This article is distributed under the terms of the Creative Commons Attribu‐ proposed. In the first stage, the alternating direction multipli‐
tion 4.0 International License (http://creativecommons.org/licenses/by/4.0/). er method (ADMM) is used to perform the preliminary state
Q. Li, W. Gao, and D. W. Gao (corresponding author) are with the Depart‐
ment of Electrical and Computer Engineering, University of Denver, 2155 E estimation over the distributed multi-area system. The sec‐
Wesley Ave, Denver, CO 80210, USA (e-mail: lq0607@gmail. com; gao18222 ond stage includes a centralized nonlinear conversion system
[email protected]; [email protected]). which calculates the intermediate variables from the results
L. Cheng is with the Department of Electrical Engineering, Tsinghua Universi‐
ty, Beijing, China (e-mail: [email protected]). of the first stage considering the nonlinearity of the system.
DOI: 10.35833/MPCE.2019.000159 The third stage includes a centralized state estimation system
628 JOURNAL OF MODERN POWER SYSTEMS AND CLEAN ENERGY, VOL. 8, NO. 4, July 2020

which computes the final estimated states using the interme‐ z = Hθ + η (1)
diate variables from the second stage. Finally, a DSE meth‐ where z is the measurement vector of readings on all meters;
od based on parallelized stream computing on the real-time H is the observation matrix that represents the characteristic
cloud platform is proposed in [21]. In the computation, the of the power system; θ is the state of the power system such
interconnected power grids with tie lines are decoupled into as the phase angle and voltage; and η is the noise or error in
sub-regions to realize the state estimation. In sum, the latest the measurement.
works on DSE are based on the multi-area method, in which For example, the power system shown in Fig. 1 has four
they are still not totally decentralized. buses, in which bus 4 is set as the reference. There are four
According to the above discussion, the existing DSE meth‐ meters, i.e., M12, M24, M31 and M43, measuring the real power
ods are not totally distributed. This drawback makes the dis‐ on the transmission lines. The meters are connected by the
tributed power system less robust since the centralized struc‐ communication lines as shown by the dashed lines in Fig. 1.
ture still exists in the system. Also, due to the same reason,
these methods are not compatible with some other distribut‐ Centralized state estimator
ed techniques in power systems, e. g. distributed economic
dispatch methods [23], so that the development of more ad‐ Load 1 Load 2
vanced distributed techniques in power systems is limited. Bus 1 Bus 2
X32 X12 X24
To overcome this drawback of the current DSE methods, G1 G2
a true DSE method without any centralized structure or facil‐ ~ M31 M12 M24
~
ity, e.g. SCADA, local control center or GPS, is proposed in X43
this paper so that the state estimation can be realized with M43 Bus 4 (reference
only the distributed smart meters. Also, a novel information phase angle is 0)
propagation algorithm is proposed as the basis of the pro‐ (a)
posed DSE method. The information propagation algorithm Load 1 Load 2
can help the smart meters broadcast their local data to the Bus 1 Bus 2
entire system with a distributed communication network. In M31 X32 M12 X12 M24 X24
G1 G2
addition, the distributed WLS method for normal sensor net‐ Local Local Local
~ ~
work [24], DC power flow model and AC power flow mod‐ state state state
estimator estimator estimator
el [25] are adopted to develop the DSE method. X43 M43
The major contributions of this paper are summarized as
follows. Bus 3 Local Bus 4
1) The fully distributed, meter-level DC and AC state esti‐ state
estimator (reference
mation methods for power systems are developed without phase angle is 0)
(b)
the need for centralized structures or facilities. Unlike the ex‐
Measurement unit; Communication line
isting DSE method, the system is not required to be parti‐
Fig. 1. An example of state estimation for power system with four buses.
tioned into small areas.
(a) Centralized state estimation. (b) Proposed DSE.
2) A novel information propagation algorithm is proposed
to broadcast the local data of each node to others in a distrib‐
uted system. According to the DC power flow model and measurement
3) The proposed method has been tested in different pow‐ model (1), the system can be modeled as:
er system simulation models. The results show that the esti‐ é 1 1 ù
mation of the proposed distributed method is more accurate ê X - 0 ú
ê 12 X 12 ú
than the conventional centralized method. ê 1 ú
é 12 ù ê-
M 1 ú
This paper is organized as follows. In Section II, the ob‐ 0
stacle to apply the WLS-based DSE for sensor network [24] êM ú ê X 31 X 31 ú éθ 1 ù
êê 31 úú = ê ú êθ ú + η (2)
in power systems is explained at first. The information prop‐ êêM 24 úú ê 1 ú ê 2ú
ê ú ê 0 0 ú ëθ 3 û
agation algorithm is introduced as a useful tool to make the ëM 43 û ê X 24 ú
system observable. Based on the proposed information propa‐ ê ú
gation method, the proposed DSE is developed. In Section ê 1 ú
ê 0 0 - ú
III, the proposed algorithms are tested in several case studies ë X 43 û
in MATLAB/Simulink software. Finally, Section IV con‐ where X ij is the reactance of the line between the bus i and
cludes this paper. bus j; and θ i is the phase angle of the bus i.
In this system, the measurement is the measured real pow‐
II. PROPOSED APPROACH er on the transmission lines. The states to be estimated are
the phase angles on the buses except for the slack bus. The
A. Problem Statement observation matrix H is created according to the transmis‐
This paper will focus on the DC state estimation first. sion line parameters.
Based on the DC power flow model, the measurement pro‐ Typically, a centralized state estimation method as shown
cess in the power system can be modeled as: in Fig. 1(a) using WLS method [25] is adopted to estimate
LI et al.: FULLY DISTRIBUTED STATE ESTIMATION FOR POWER SYSTEM WITH INFORMATION PROPAGATION ALGORITHM 629

the states in power system. In this method, the readings w ij is a weight coefficient on the edge between node i and
from all meters are collected by the centralized state estima‐ node j; and Ni is the set of neighbor nodes of the node j.
tor. Then, the centralized state estimator performs the WLS For the entire graph, the information states of all nodes con‐
to estimate the states with the noisy measurement z by: verge to their average value ∑x i n if the weight coeffi‐
θ̂ = ( H T R-1 H ) H T R-1 z
-1
(3) cients w ij are chosen based on certain rules in [24].
where θ̂ is the estimation of the state θ; and R is the covari‐
Now, suppose that each node in the graph has its local in‐
formation, e.g., z i is the local information of the node i. The
ance matrix of measurement errors, which are dependent on
goal of the information propagation algorithm is to allow
the accuracy of meters.
each node in the graph accurately estimate the local informa‐
In a power system using DSE as shown in Fig. 1(b), the
tion of all nodes, i. e., the vector of all local information
DSE is performed in the nodes of the system. A node is the
[z 1 z 2 ...z n ] can be estimated by each node with the informa‐
combination of a meter and a local state estimator as shown
tion propagation algorithm.
in Fig. 1(b). The local state estimator performs the proposed
In order to achieve this goal, the information state x i in
information propagation and the proposed state estimation
(5) is considered as the estimated local information vector
from this paper. Also, each local state estimator connects to
x i =[ẑ i1 ẑ i2 ...ẑ in ], where ẑ ij is the estimation of the jth node’s
its neighbors by the communication lines (the dash lines in
local information by node i. Then, if the consensus protocol
the figure). In this system, since there is no centralized state
(5) can be modified so that the information state x i does not
estimator, the WLS method and other centralized state esti‐
converge to the average value but the local information vec‐
mation methods cannot be directly applied. For example, the
tor, i. e., klim x (k) =[z 1 z 2 ...z n ], the information propagation
node M 31 is only connected to the node M 12, so the measured ®¥ i

values M 24 and M 43 are unknown for M 31. Then, in order to is achieved. Based on this idea, the following modified con‐
perform centralized state estimation, the measurement model sensus protocol serves as the proposed information propaga‐
for the local state estimator in node M 31 is built as: tion algorithm:
é 1 1 ù x i (k + 1) = x i (k) + τI i0 ∑ w ij (x j (k) - x i (k) ) (6)
ê X - 0 ú éθ ù j Î Ni
éê 12 ùú = êê
M 12 X 12 úú ê ú
1

θ +η (4) The above equation can be re-written in vector form as:

ëM 31 û êê 1 1 úú ê 2 ú
-
ê X 0
X 31 úû 3
ë θ û éẑ i1 (k + 1)ù éẑ i1 (k)ù æ éẑ j1 (k)ù éẑ i1 (k)ù ö
ë 31
ê ú ê ú çê ú ê ú÷
êêẑ i2 (k + 1)úú êêẑ i2 (k)úú ç êẑ j2 (k)ú êẑ i2 (k)ú ÷
It is easy to verify that the term H T R-1 H from (3) is sin‐ úú = êê ú + τI i ∑ w ij ç êêê úú - êê úú ÷ (7)
0
êê
gular with the H matrix from (4), thus (3) cannot be solved   ú jÎN ç  ú ê  ú÷
ê ú ê ú ç êẑ (k)ú êẑ (k)ú ÷
i

accurately. This is to say that the system is unobservable for ëẑ in (k + 1)û ëẑ in (k)û è ë jn û ë in û ø
the node M 31 [25] and the state estimation cannot simply pro‐ where the initial values ẑ ij (0) can be arbitrary if i ¹ j, but
ceed in a distributed fashion. To solve this problem, an infor‐
ẑ ii (0) = z i since the local information z i is available at the ith
mation propagation algorithm is proposed to broadcast the lo‐
node itself; and I i0 is an n ´ n diagonal matrix which is the
cal measurement data (local information) of each node to the
same as the identity matrix but has a zero at the ith diagonal
entire system, so that the system will become observable for
entry:
all nodes.
I i0  diag(11...101...1) (8)
B. Information Propagation Algorithm
For the node i, the i row of (7) is:
th

ẑ ii (k + 1) = ẑ ii (k) + τ ´ 0 ´ ∑ w ij ( ẑ j (k) - ẑ i (k)) =

The function of the proposed information propagation al‐
gorithm is to broadcast the local information, i. e., the mea‐
j Î Ni
surements from the meter in power system, of each node to
ẑ ii (k) + 0 = ẑ ii (k) = ẑ ii (0) = z i(9)
the entire system without a centralized communication coor‐
dinator. Since the i row of the matrix I is 0, the second term in
th
i
0

The information propagation algorithm is based on the (9) is always 0. Thus, the ith row of the estimated local infor‐
consensus protocol technique [24]. In this method, the com‐ mation vector remains unchanged for all the times, i. e.,
munication network of a power system can be modeled as a ẑ ii (k) = ẑ ii (0) = z i for k = 01...¥.
graph, in which the meters with local state estimators, e. g., For the jth row of (9) in which j ¹ i, since ẑ jj (k) on the jth
M 32, M 12, M 24, and M 43 in Fig. 1(b), are treated as nodes, node is fixed at z j as discussed above, the consensus proto‐
while the communication lines between the nodes are the col works in the leader-follower mode [26]. In this mode,
edges of the graph. Suppose that x i is the information state the node j is the leader and other nodes including node i are
of the ith node in the graph. The vector X =[x 1 x 2 ...x n ] in‐ the followers. Thus, the jth row ẑ ij (k) of the information state
cludes the information states of all n nodes. Then, the con‐ x i on the node i will converge to z j, i.e., klim ẑ (k) = z j.
® ¥ ij
sensus protocol for the ith node is: According to the discussion above, the ith row of the infor‐
x i (k + 1) = x i (k) + τ ∑ w ij ( x j (k) - x i (k)) (5)
mation state x i is fixed at the local information z i of the
i jÎN node i, and the other rows ẑ j (k) converge to the local infor‐
where x i (k) is the information state x i at the time step k; τ is mation z j of other nodes. Therefore, the goal klim ®¥ i
x (k) =
the time interval between two consecutive update time steps; [z 1 z 2 ...z n ] is achieved.
630 JOURNAL OF MODERN POWER SYSTEMS AND CLEAN ENERGY, VOL. 8, NO. 4, July 2020

identity matrix; C = ( H T R-1 H ) H T R-1; and D = 0. The

-1
C. Proposed DSE
With the information propagation algorithm, the DSE for transfer function of this model is:
Y (s) θ̂ i (s) ( H T R-1 H ) H T R-1
power systems can be established. By implementing the in‐ -1

formation propagation algorithm in power systems, the local = = (16)

information z i is the measurement from the meter of the ith U(s) η̇ (s) sI + I i0 L w
node. The notation Z i is used to represent the estimated local Since the eigenvalues of L w are non-negative due to the
information vector Z i = x i =[ẑ i1 ẑ i2 ...ẑ in ] of node i. Equation convergence condition of the consensus protocol, the poles
(6) becomes: of the system (16) are on the left-half plane. Therefore, the
Z i (k + 1) = Z i (k) + τI i0 ∑ w ij (Z j (k) - Z i (k) ) (10)
model (16) of the DSE algorithm can be considered as a
j Î Ni low-pass filter. This is helpful to mitigate the noise in the
measurement.
By integrating the WLS method with the information prop‐
agation algorithm, the new DSE algorithm at node i is de‐ D. DSE with AC Power Flow
scribed as: Based on the proposed information propagation algorithm,
ìZ i'i (k) = z i (k) an AC DSE can be developed as well. The idea of this AC
ï DSE is to use the information propagation algorithm to
ïïZ i'j ( k ) = Z ij ( k ) j ¹ i
ï share the measured data with other nodes in the system, then
íZ i (k + 1) = Z i' (k) + τI i0 ∑ w Zij ( Z j' (k) - Z i' (k)) (11) the AC state estimation algorithm [25] is performed when
ï jÎN
ïï the information is well shared. To determine if the informa‐
i

ïθ̂ (k + 1) = ( H T R-1 H ) -1 H T R-1 Z (k + 1) tion is well shared, the information mismatch on the node i
î i i
can be calculated as:
where Z i' (k) is a modification of Z i (k) that the ith entry is re‐
 
placed by the actual measurement z i (k); w Zij is the a weight δ info
i = ∑ ( Z j' (k) - Z i' (k))  (17)
coefficient for the communication of information vector on jÎN ¥
the edge between node i and node j; and θ̂ i is the estimated
i

where  ×  ¥ is the infinity norm which gives the maximal val‐

states, i. e., phase angles, of all buses in the power system
calculated by node i. ue in the vector. When the information mismatch δ info
i is low‐
According to the DSE equation (11), the algorithm can er than a threshold value, the AC state estimation will be ex‐
work as a low-pass filter which will reduce the impact of ecuted.
the measurement errors. The reason is presented as follows. Different from the DC DSE discussed above, the AC DSE
Firstly, the consensus protocol (5) can be represented into measures not only the real power on the transmission line,
matrix form [22] as: but also the reactive power on the transmission line. Thus,
the information propagation algorithm will be applied to
X(k + 1) = X(k) - τL w X(k) (12)
share both the measured real power and reactive power. Let
where X(k) =[x 1 (k); x 2 (k); x 3 (k); ...; x n (k)] contains all infor‐ z iP and z iQ be the real power and reactive power measured on
mation states in the system; and L w is the weighted Lapla‐ the meter of node i, respectively. The estimated states are
cian matrix [24]. Similarly, the matrix form of the ith row of the phase angle and voltage magnitude on the buses in the
the DSE algorithm (11) can be written as: system. The algorithm of the AC DSE is listed in Algo‐
Z i (k + 1) = Z i' (k) - τI i0 L w Z i' (k) (13) rithm I.
Let Z i' (k) = Z i (k) + Dη, where Dη =[0...0dη i 0...0] and Algorithm I AC DSE
dη i is the difference of the measurement error on the meter Initialize all variables
of node i between two time steps. The equation becomes: while δ info
i is larger than the threshold do
Update the measurements z iP and z iQ
Z i (k + 1) = Z i (k) + Dη - τI i0 L w (Z i (k) + Dη)Þ Run (11) to share z iP
Z i (k + 1) - Z i (k) Dη Run (11) to share z iQ
= - I i0 L w Z i (k) (14) Calculate information mismatch δ info by (17)
τ τ end while
i

where I i0 L w Dη = 0 since Dη only has non-zero value at its ith Perform AC state estimation [23]
row and the elements at ith row of I i0 are all zeros. Then,
(14) is rewritten into continuous form by letting τ ® ¥, and
III. SIMULATION RESULTS
the DSE algorithm can be represented in the state space
form by adding the last sub-equation of (11):

{
A. Information Propagation Algorithm in Example Graphs
Ż i (t) = -I i0 L w Z i (t) + η̇ In order to verify that the proposed information propaga‐
(15)
θ̂ i (t) = ( H T R-1 H ) H T R-1 Z i (t)
-1
tion algorithm can share the local value of each vertex to the
others, the simulations are conducted in the communication
This is a state space model where the input u = η̇ = networks with two different graphs shown in Fig. 2, which
[0...0η̇ i 0...0] and η̇ i is the derivative of the measurement are the cycle graph and the tree graph, respectively [27],
error; the output is y = θ̂ i (t); A = -I i0 L w; B = I where I is an where V1 to V6 are the vertices of the graph.
LI et al.: FULLY DISTRIBUTED STATE ESTIMATION FOR POWER SYSTEM WITH INFORMATION PROPAGATION ALGORITHM 631

V2 V4 adopted to test the proposed DSE method. WSCC 9-bus sys‐

V1
tem is an approximated model of the WSCC power system
V2 V5 V6 in 60 Hz, as shown in Fig. 5, where the generators G1 and
V1 V5 G3 in the power system are hydro plant and thermal plant,
V3 respectively.

V4 Bus 2 Bus 7 Bus 8 Bus 9 Bus 3

V3 V6 Load 3 230 kV 13.8 kV
G2 18 kV 230 kV 230 kV
M4 G3
(a) (b) Line 5 M5 Line 6 M6
~ ~
Fig. 2. Two different graphs. (a) Cycle graph. (b) Tree graph. 128 MVA T2 T3 192 MVA
Line 2 Line 4
In the two case studies, the vertices V1, V2, V3, V4, V5 Bus 5
M3
Bus 6
and V6 have initial local values 8, 2, 25, 12, 13 and 3, re‐ 230 kV M2 230 kV
spectively. Then, at t = 0.5 s, the local values are turned into Load 1 Line 1 Line 3 Load 2
4, 6, 14, 9, 21 and 11. By applying the information propaga‐ Bus 4: 230 kV
tion algorithm, the results are shown in Figs. 3 and 4. (Reference) M1

30 T1
Measurement unit Bus 1
25 16.5 kV
Communication line
~ G1
Estimation

20
247.5 MVA
15
10 Fig. 5. WSCC 9-bus system.
5

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Considering the rapid development of renewable energy,
Time (s) the generator G2 is modeled as an inverter-based generator.
Node 1; Node 2; Node 3; Node 4; Node 5; Node 6 Three transformers T1, T2 and T3 step up the voltages of
Fig. 3. Information propagation algorithm in cycle graph. the generators to 230 kV. The impedances (in p. u.) of the
transmission lines are: 0.01+ j0.085 for line 1, 0.032 + j0.161
30
for line 2, 0.017 + j0.092 for line 3, 0.039 + j0.17 for line 4,
25 0.0085+ j0.072 for line 5, and 0.0119 + j0.1008 for line 6.
Estimation

20 There are three loads, load 1 (187.5 MVA), load 2 (153

15 MVA) and load 3 (170 MVA), on bus 5, bus 6 and bus 8, re‐
10 spectively. In addition, six meters, M 1, M 2, ..., M 6, are in‐
5 stalled to measure the real power on the transmission lines
in the system. In real systems, the meters can be built by the
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Time (s)
embedded systems such as advanced reduced instruction set
Node 1; Node 2; Node 3; Node 4; Node 5; Node 6 computing (RISC) machine (ARM), field-programmable gate
array (FPGA) or digital signal processor (DSP). Each node
Fig. 4. Information propagation algorithm in tree graph.
is connected with its neighbor nodes by the communication
lines. The communication lines should be bi-directional with
The results show that local data can be successfully propa‐ low latency. In practice, TCP/IP protocol based communica‐
gated in the graphs by the information propagation algo‐ tion methods such as Ethernet network and Wifi can be used
rithm. According to Figs. 3 and 4, the estimations can con‐ to build the communication lines. Also, in the simulation, in
verge to the correct values in about 0.15 s. Also, the informa‐
order to simulate the delay in communication, a delay com‐
tion propagation algorithm converges faster in the cycle
ponent is added on every communication line.
graph than in the tree graph. This is because the Laplacian
In a 25 s simulation, load 1 is increased by 30 MW at 5
matrix for a cycle graph usually has larger eigenvalues,
s, while the load 3 is decreased by 20 MW at 15 s. The
which means that the algorithm can converge more quickly.
Gaussian noise with zero mean and standard derivation σ =
The average value of the eigenvalues of the modified Lapla‐
0.01 is added on the meters. The proposed DSE algorithm
cian matrix of the cycle graph is 1.6667, and it is 1.3889 for
the tree graph, where the modified Laplacian matrix is the (11) is used to estimate the phase angle on each bus. Figure
Laplacian matrix of the information propagation algorithm. 6 shows the phase angle estimation results by the distributed
Therefore, the larger eigenvalues for cycle graph allow it to algorithm executed at node M1. The result is compared with
converge faster. the centralized state estimation and the actual value. It
shows the DSE has less noise than the centralized method.
B. State Estimation in Western System Coordinating Council This is because the consensus protocol is a recursive algo‐
(WSCC) 9-bus Power System rithm that has the inherent filtering effect, which reduces the
In this work, the WSCC 3-generator 9-bus system [28] is noise.
632 JOURNAL OF MODERN POWER SYSTEMS AND CLEAN ENERGY, VOL. 8, NO. 4, July 2020

0.04
Bus 7
tion propagation algorithm as discussed above. Therefore,
0.02 -0.075 the estimation of the state of bus 5 on meter M 2 is less accu‐
0 rate than on the other meters. But this problem can be easily
Phase angle (rad)

Bus 9 -0.080
-0.02 addressed by simply adding a low-pass filter to the meters.
-0.04 Bus 8 -0.085 C. State Estimation in IEEE 39-bus System
-0.06
Bus 6 -0.090 The IEEE 39-bus system is a power system model with
-0.08
Bus 5
13 14 15 16 17 18 10 generators and 39 buses, as shown in Fig. 8. The parame‐
-0.10
ters can be found in [29]. In the system, 37 meters running
-0.12 the proposed DSE are installed on the power lines in Fig. 8,
0 5 10 15 20 25
Time (s) which are connected by the communication network. The
Centralized state estimation; Proposed DSE; Actual value communication network in the system has multiple cycles,
Fig. 6. Comparison among actual value, estimation by centralized method e. g., the meters M1, M2, M25, M3, M4, M5, M6, M32, M7, M8,
and estimation by proposed DSE method at node M1. M9, M30 form a cycle. The existence of the cycles in commu‐
nication network makes this DSE system more robust. This
In order to evaluate the performance of the proposed meth‐ is because that there are more redundant paths in the net‐
od, the mean squared error (MSE) is adopted to compute the work to link any two meters. For example, if the communi‐
accuracy of the estimation. The MSE is defined by MSE = cation line between the meters M 1 and M 2 is broken, the in‐

T∑
1 T formation can still be transmitted between M 1 and M 2 by the
(a(k) - â (k)) , where a(k) is the actual value; â (k) is the
2

k=0
path M 2 ® M 25 ® M 3 ® ...® M 30 ® M 1. The results of the
estimated value; and T is the total time steps. The results are simulation are shown in Fig. 9. This figure only shows the
shown in Fig. 7. MSE of the estimation for the states of all buses in the sys‐
tem by the meter M 1 since all other meters have similar re‐
0.0000050 Estimation from M1 sults. According to the results, it is clear that the proposed
0.0000045 Estimation from M2
0.0000040 Estimation from M3 DSE is more accurate than the traditional centralized meth‐
0.0000035 Estimation from M4 od. This improvement of the estimation is also caused by the
Estimation from M5
0.0000030 low-pass filter characteristic of the information propagation
MSE

0.0000025 Estimation from M6

0.0000020
Centralized estimation algorithm.
0.0000015 G10 G8
0.0000010
0.0000005
~ ~
30 37
0 M29
5 6 7 8 9 26 M28 M37
Bus No. 29
2 25 M26 28
Fig. 7. Comparison of MSE between different methods. M27 38
M2 M25 3 27
M1
18 M35
1 M3
M24 ~ G9
M18 16 24
M17
In the bar chart, the bars are categorized into five groups, M16 M34
4 17 G6
in which each group represents the results for each bus in M30 M31
15 M21
39 M4 M15 21 ~
the power system. Meanwhile, inside each group, the bars 12 M36
~ M5 14 35
show the estimation errors of different nodes and methods. G1 M9 M6
5 M12 M33 M14 M19 19
Since the proposed DSE algorithm can be performed on ev‐ 9 M32 6 M22
22
ery node, each node has its own state estimation results. M8 7
M11
13 M20
M 11 M10 M23
Therefore, the left six bars in each group are the results of 8
7
10
M13
20 23
the proposed DSE executed on the six different nodes, and
31 32 33 34 36
the last bar represents the result of the centralized method
performed by the control center. According to the results, the G2 ~ G3 ~ G4 ~ ~ G5 G7 ~
error of our proposed method is much smaller than the cen‐ Measurement unit; Communication line

tralized method. On the other hand, the estimations of some Fig. 8. IEEE 39-bus system.
buses have higher errors in some meters, e.g., the estimation
of bus 5 by the M 2 (the second bar in the first group in Fig. 6
Proposed DSE
7) has higher error. This is because the meter M 2 can direct‐ 5 Centralized state estimation
4
ly measure the data from bus 5 without going through the in‐
MSE

3
formation propagation algorithm, which means that the mea‐ 2
surement noise is directly input into the state estimation algo‐ 1
rithm in M 2. However, for other meters, the data of bus 5 is 0
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
transmitted by the information propagation algorithm so that 3 5 7 9 11 13 15 17 19 21 23 25 27 29
the measurement errors are reduced during the communica‐ Bus No.
tion due to the low-pass filter characteristic of the informa‐ Fig. 9. MSE of different methods for IEEE 39-bus system.
LI et al.: FULLY DISTRIBUTED STATE ESTIMATION FOR POWER SYSTEM WITH INFORMATION PROPAGATION ALGORITHM 633

D. Comparison of Different Methods in IEEE 14-bus System mation methods. However, like most other DSE methods,
In order to verify the performance of the proposed DSE the multi-area DSE requires local estimator in each parti‐
method, the commonly used multi-area DSE method [11] is tioned area and centralized coordinator for the whole system.
adopted for comparison. The comparison is done in the Therefore, the method is not fully distributed. On the other
IEEE 14-bus system as shown in Fig. 10. All methods in hand, The AC DSE method in our paper has a much lower
this comparison use the same meters as shown in the figure. error than other methods. This is because the AC power
Also, for the multi-area DSE method, the system is parti‐ flow model is more accurate than the DC power flow mod‐
tioned [11] as the figure shows. el. In conclusion, the proposed DSE method achieves similar
or slightly better performance as the commonly used multi-
Area 4 area DSE, but it is a truly distributed method compared to
12 13 14
the existing methods.
TABLE I
ERRORS OF PHASE ANGLE ESTIMATION
6 11 10 9
Area 3
Method Average error of 1000 tests (p.u.)
7
Centralized state estimation 0.0163
8
Proposed DC DSE 0.0158
Area 1 Area 2 Proposed AC DSE 0.0021
1 5 4
Multi-area DSE 0.0161

2 3
E. Improving Measured Data
Power line; Communication line; t Bus; Meter On the other hand, the estimated states can be used to im‐
Fig. 10. IEEE 14-bus system. prove the accuracy of the measurements by the equation ẑ =
Hθ̂ , where θ̂ is the estimated phase angle obtained from the
proposed DSE method; and ẑ is the estimation of the power
The results of the simulations are listed in Table I. The re‐
flow. Using this method, the noise in the raw measurements
sults show the errors of the estimated phase angle from each
of the power flow from the nodes can be reduced, thus the
method. The errors are calculated by averaging the errors in
measurements are more precise. The case study is done on
1000 runs, in which the measurement noise follows normal
the IEEE 39-bus system. The estimation results from the me‐
distribution with 0 mean and 0.01 standard deviation. The re‐
ter M1 in IEEE 39-bus system are presented in Fig. 11 as an
sults show that the proposed DC DSE method has similar ac‐
example.
curacy as the multi-area DSE and the centralized state esti‐
0.0016
0.0014 Proposed DSE
0.0012 Centralized state estimation
0.0010
MSE

0.0008
0.0006
0.0004
0.0002
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37
Measurement unit No.
Fig. 11. MSE of power flow estimation for M1 in IEEE 39-bus system.

The results of the IEEE 39-bus system in Fig. 11 compare ments. In this case study, the system runs for 8 s and two
the MSEs of the power flow estimation between the pro‐ bad measurements (by increasing the raw measurement read‐
posed DSE method and the traditional centralized state esti‐ ing by 1.0) are injected between 2-3 s and 5-6 s on M1 and
mation method. According to the results, the proposed meth‐ M14, respectively. Based on the proposed DSE method, the
od has lower MSE for the estimation of each bus, which bad measurement can be detected by the ith meter by calcu‐
means that it is much more accurate than the centralized lating the measurement residual J i (θ̂ i ) [25]:
method, which is true for other nodes. n (z - ẑ )2
J i (θ̂ i ) = ∑
ij ij
(18)
F. Bad Measurement Detection j=1
2
σ ij
Bad measurement detection is one of the most important where z ij is the jth raw measurements; ẑ i = Hθ̂ i is the estima‐
topics in power system estimation [30], [31]. In a large sen‐ tion of the measurements; and σ ij is the standard deviation of
sor network, the hardware malfunction or cyber-attack may the measurement which is 0.01 in this system. In IEEE 39-
cause the sensors to provide incorrect readings. Therefore, it bus system, the number of estimated states θ̂ i is 29 and the
is very important to identify and eliminate bad measure‐ number of measurements is 37, hence the degrees of free‐
634 JOURNAL OF MODERN POWER SYSTEMS AND CLEAN ENERGY, VOL. 8, NO. 4, July 2020

dom for the state estimation is 37 - 29 = 8. The measurement tional Intelligence and Computing Research (ICCIC), Tamilnadu, In‐
residual J i (θ̂ i ) satisfies the χ 2 distribution. Then, according to
dia, Dec. 2014, pp. 1-5.
[7] H. Chen, L. Zhang, J. Mo et al., “Synchrophasor-based real-time state
the probability distribution function (PDF) of χ 2 distribution, estimation and situational awareness system for power system opera‐
by choosing 3σ confidence, the threshold value of J i (θ̂ i ) for tion,” Journal of Modern Power Systems and Clean Energy, vol. 4,
no. 3, pp. 370-382, Jul. 2016.
the bad measurement detection is 23.3. [8] J. Zhang, G. Welch, G. Bishop et al., “A two-stage Kalman filter ap‐
In the proposed method, bad data detection can be per‐ proach for robust and real-time power system state estimation,” IEEE
Transactions on Sustainable Energy, vol. 5, no. 2, pp. 629-636, Apr.
formed by any meter in the system. In Fig. 12, the residuals
J i (θ̂ i ) calculated by all nodes (the lines in different colors)
2014.
[9] E. Ghahremani and I. Kamwa, “Dynamic state estimation in power
are presented. The results show that the residuals on all system by applying the extended Kalman filter with unknown inputs
to phasor measurements,” IEEE Transactions on Power Systems, vol.
nodes exceed the threshold during 2-3 s and 5-6 s. It means 26, no. 4, pp. 2556-2566, Nov. 2011.
that the bad data injection on M1 and M14 are detected. [10] S. Wang, W. Gao, and A. S. Meliopoulos, “An alternative method for
power system dynamic state estimation based on unscented trans‐
40 form,” IEEE Transactions on Power Systems, vol. 27, no. 2, pp. 942-
500 23.3 23.3
20 950, May 2012.
400 [11] G. N. Korres, “A distributed multiarea state estimation,” IEEE Trans‐
0 actions on Power Systems, vol. 26, no. 1, pp. 73-84, Feb. 2011.
4.5 5.0 5.5 6.0 6.5 [12] G. N. Korres, A. Tzavellas, and E. Galinas, “A distributed implementa‐
300
Ji (θi)
^

tion of multi-area power system state estimation on a cluster of com‐

200 puters,” Electric Power Systems Research, vol. 102, pp. 20-32, Sept.
2013.
100
[13] A. Gomez-Exposito, A. de la Villa Jaen, C. Gomez-Quiles et al., “A
taxonomy of multi-area state estimation methods,” Electric Power Sys‐
0 1 2 3 4 5 6 7 8 tems Research, vol. 81, no. 4, pp. 1060-1069, Apr. 2011.
Time (s) [14] V. Kekatos and G. B. Giannakis, “Distributed robust power system
Fig. 12. Values of objective function J i (θ̂ i ) for bad data detection. state estimation,” IEEE Transactions on Power Systems, vol. 28, no.
2, pp. 1617-1626, May 2013.
[15] X. Li and A. Scaglione, “Robust decentralized state estimation and
tracking for power systems via network gossiping,” IEEE Journal on
IV. CONCLUSION Selected Areas in Communications, vol. 31, no. 7, pp. 1184-1194, Jul.
2013.
This paper proposes a new DSE method for power sys‐ [16] S. Chen, Z. Wei, G. Sun et al., “Multi-area distributed three-phase
tems. The proposed algorithm can dynamically estimate the state estimation for unbalanced active distribution networks,” Journal
states without any centralized facility. An information propa‐ of Modern Power Systems and Clean Energy, vol. 5, no. 5, pp. 767-
776, Nov. 2016.
gation algorithm is developed to broadcast any information [17] A. K. Singh and B. C. Pal, “Decentralized dynamic state estimation in
of vertex to the whole system. The information propagation power systems using unscented transformation,” IEEE Transactions on
algorithm can be useful in a distributed system since it does Power Systems, vol. 29, no. 2, pp. 794-804, Mar. 2014.
[18] J. Du, S. Ma, Y. C. Wu et al., “Distributed hybrid power state estima‐
not require the centralized unit to gather data. The proposed tion under PMU sampling phase errors,” IEEE Transactions on Signal
method is simulated in the WSCC 9-bus power system, Processing, vol. 62, no. 16, pp. 4052-4063, Aug. 2014.
IEEE 14-bus system, and IEEE 39-bus system in MATLAB/ [19] W. Zheng, W. Wu, A. Gomez-Exposito et al., “Distributed robust bilin‐
Simulink. The results show that the proposed DSE algorithm ear state estimation for power systems with nonlinear measurements,”
IEEE Transactions on Power Systems, vol. 32, no. 1, pp. 499-509,
can achieve better performance than the traditional central‐ Jan. 2016.
ized state estimation method. In future work, for a large [20] X. Kong, Z. Yan, R. Guo et al., “Three-stage distributed state estima‐
power system with thousands of meters and buses, a new tion for AC-DC hybrid distribution network under mixed measurement
environment,” IEEE Access, vol. 6, pp. 39027-39036, Jul. 2018.
method to reduce the computation will be explored. [21] L. Shan, J. Yu, J. Zhang et al., “Distributed state estimation based on
the realtime dispatch and control cloud platform,” in Proceedings of
REFERENCES 2018 2nd IEEE Conference on Energy Internet and Energy System In‐
tegration (EI2), Beijing, China, Oct. 2018, pp. 1-6.
[1] F. F. Wu, “Power system state estimation: a survey,” International [22] C. Gomez-Quiles, H. A. Gil, A. de la Villa Jaen et al., “Equality-con‐
Journal of Electrical Power & Energy Systems, vol. 12, no. 2, pp. 80- strained bilinear state estimation,” IEEE Transactions on Power Sys‐
87, Apr. 1990. tems, vol. 28, no. 2, pp. 902-910, May 2012.
[2] F. F. Wu, K. Moslehi, and A. Bose, “Power system control centers: [23] Q. Li, D. W. Gao, H. Zhang et al., “Consensus-based distributed eco‐
past, present, and future,” Proceedings of the IEEE, vol. 93, no. 11, nomic dispatch control method in power systems,” IEEE Transactions
pp. 1890-1908, Nov. 2005. on Smart Grid, vol. 10, no. 1, pp. 941-954, Jan. 2019.
[3] H. Li, A. T. Eseye, J. Zhang et al., “Optimal energy management for [24] M. Mesbahi and M. Egerstedt, Graph Theoretic Methods in Multia‐
industrial microgrids with high penetration renewables,” Protection gent Networks. Princeton: Princeton University Press, 2010.
and Control of Modern Power Systems, vol. 2, no. 1, pp. 1-12, Apr. [25] A. J. Wood and B. Wollenberg, Power Generation Operation and Con‐
2017. trol, 2nd ed. Hoboken: John Wiley & Sons, 1996, p. 195.
[4] A. Rahman and G. K. Venayagamoorthy, “Scalable cellular computa‐ [26] T. Gustavi, D. V. Dimarogonas, M. Egerstedt et al., “Sufficient condi‐
tional network based WLS state estimator for power systems,” in Pro‐ tions for connectivity maintenance and rendezvous in leader-follower
ceedings of 2015 Clemson University Power Systems Conference networks,” Automatica, vol. 46, no. 1, pp. 133-139, Jan. 2010.
(PSC), Clemson, USA, Mar. 2015, pp. 1-6. [27] R. Balakrishnan and K. Ranganathan, A Textbook of Graph Theory.
[5] X. Yang, Z. Hu, R. Fan et al., “Weighted least squares state estimation Berlin: Springer, 2012.
based on the optimal weight,” in Proceedings of 2015 Third Interna‐ [28] P. M. Anderson and A. A. Fouad, Power System Control and Stability.
tional Conference on Technological Advances in Electrical, Electronics Hoboken: John Wiley & Sons, 2008.
and Computer Engineering (TAEECE), Beirut, Lebanon, Apr. 2015, [29] A. Pai, Energy Function Analysis for Power System Stability. Berlin:
pp. 12-16. Springer Science & Business Media, 2012.
[6] P. Tripathi, J. Rahul, and N. A. Radhamohan, “Analysis of state esti‐ [30] N. Zivkovic and A. T. Saric, “Detection of false data injection attacks
mation for power system using weighted least square method,” in Pro‐ using unscented Kalman filter,” Journal of Modern Power Systems
ceedings of 2014 IEEE International Conference on IEEE Computa‐ and Clean Energy, vol. 6, no. 5, pp. 847-859, Sept. 2018.
LI et al.: FULLY DISTRIBUTED STATE ESTIMATION FOR POWER SYSTEM WITH INFORMATION PROPAGATION ALGORITHM 635

[31] J. Chen, G. Liang, Z. Cai et al., “Impact analysis of false data injec‐ Wei Gao received his bachelor’s degree in automation from Hebei Universi‐
tion attacks on power system static security assessment,” Journal of ty of Technology, Tianjin, China, in 2017. He is pursuing his Ph.D. degree
Modern Power Systems and Clean Energy, vol. 4, no. 3, pp. 496-505, in the Department of Electrical and Computer Engineering, University of
Jul. 2016. Denver, Denver, USA. His research interests are microgrid control, renew‐
able energy, and power system stability.

Qiao Li received the B. Eng. degree in automation engineering from Gui‐ David Wenzhong Gao received his M. S. and Ph. D. degrees in electrical
zhou University, Guiyang, China, in 2011 and the M.Eng. degree in control and computer engineering, specializing in electric power engineering, from
theory and control engineering from the Nanjing University of Aeronautics Georgia Institute of Technology, Atlanta, USA, in 1999 and 2002, respec‐
and Astronautics, Nanjing, China, in 2014. He is currently pursuing the tively. He is now with the Department of Electrical and Computer Engineer‐
Ph.D. degree with the Department of Electrical and Computer Engineering, ing, University of Denver, Denver, USA. He is an Associate Editor for
University of Denver, Denver, USA. His current research interests include IEEE Journal of Emerging and Selected Topics in Power Electronics. He
distributed power system, networked control, microgrid, energy storage sys‐ was an editor of IEEE Transactions on Sustainable Energy. He was the Gen‐
tem, and renewable energy. eral Chair for the 48th North American Power Symposium (NAPS 2016)
and for the IEEE Symposium on Power Electronics and Machines in Wind
Lin Cheng received the Ph. D. degree from Tsinghua University, Beijing, Applications (PEMWA 2012). His current teaching and research interests in‐
China, in 2001, where he is currently a Tenured Professor with the Depart‐ clude renewable energy and distributed generation, microgrid, smart grid,
ment of Electrical Engineering, Tsinghua University. His research interests power system protection, power electronics applications in power systems,
are power system reliability evaluation, power system dynamics and the power system modeling and simulation, and hybrid electric propulsion sys‐
multi-energy system planning. tems.