Inverse Power Flow

1
Inverse Power Flow Problem

Ye Yuan, Member, IEEE, Steven H. Low, Fellow, IEEE,
Omid Ardakanian, Member, IEEE, and Claire J. Tomlin, Fellow, IEEE
arXiv:1610.06631v3 [cs.SY] 20 Feb 2020
Abstract
This study considers the inverse power flow problem which is to infer the nodal admittance matrix
(and therefore the operational network structure of the power system) from voltage and current phasors
measured at a number of buses. We show that the admittance matrix can be uniquely identified from
a sequence of measurements of different steady states when the every node in system is measured and
a reduced admittance matrix (from Kron reduction) can be determined when it contains some hidden
nodes. Furthermore, we propose efficient algorithms based on graph theory and convex relaxation to
uncover the actual admittance matrix of radial and mesh systems with hidden nodes. In particular,
we provide theoretical guarantees of the recovered admittance matrices for radial systems. Simulations
performed on standard test systems confirm that these algorithms provide accurate estimates of the
admittance matrix from noisy data.
Index Terms
Inverse Power Flow Problem, System Identification, Kron Reduction, Algebraic Graph Theory,
Phasor Measurement Units.
I. I NTRODUCTION
The power industry has witnessed profound changes in recent years. These changes are mostly
driven by the widespread adoption of distributed energy resources (DER), active participation of
customers in emerging energy markets, and rapid deployment of measurement, communication,
and control infrastructure resulting in an unprecedented level of visibility and controllability,
especially for distribution grids. Despite the increased amount of uncertainty, these changes
Ye Yuan is with School of Artificial Intelligence and Automation and State Key Lab of Digital Manufacturing Equipment
and Technology, Huazhong University of Science and Technology, Wuhan, China. Steven H. Low is with Department of
Computer and Mathematical Sciences and Electrical Engineering, California Institute of Technology, Pasadena, United States.
Omid Ardakanian is with Department of Computing Science, University of Alberta, Edmonton, Canada. Claire J. Tomlin is with
Department of Electrical Engineering and Computer Sciences, University of California, Berkeley, United States. Correspondence
to slow@caltech.edu.
February 21, 2020 DRAFT

2
offer ample opportunities to system operators to improve power system stability and efficiency
by leveraging advanced optimization and control techniques [1]. Most of these techniques require
the knowledge of the real-time network topology, which has been traditionally lacked in many
cases due to the limited visibility into the state of the power system.
The inverse power flow (IPF) problem we define in this paper concerns the estimation of
the nodal admittance matrix, which describes the network topology, from synchronized mea-
surements of voltage and current phasors (i.e., magnitudes and phase angles) which can be
obtained from phasor measurement units (PMUs) [2] or conventional supervisory control and
data acquisition (SCADA) technology. The IPF problem is the host of several crucial smart grid
applications, affecting real-time system operation and long-term planning, the most important of
which are:
1) State Estimation combines the knowledge of the admittance matrix with a set of known state-
variables to determine the unknown ones, e.g., voltage magnitude and phase angle of unobserved
buses, thereby building a real-time model of the network which enables the operators to justify
technical and economical decisions and to uncover potential operational problems. The state
estimation can be used to compensate for the transducer error, the time skewness of telemetry
data, and the absence of monitoring devices at particular nodes.
2) Optimization & Control techniques determine a sequence of operations that can transition
the power system from one steady state to another steady state that meets certain stability and
efficiency targets. These techniques typically require the knowledge of the network topology or
information about the state of the power system.
3) Event Detection concerns identifying faults, line outages, and other critical events, such
as transformer tap changes, capacitor and switching operations from changes in the real-time
network model. The idea of fault detection is to detect the occurrence of a fault when it happens
while fault location concerns pinpointing the fault to within a small section of a distribution
feeder.
4) Cybersecurity is the problem of identifying the potential vulnerabilities of a power system
and designing strategies to protect it from the potential cyber attacks using telemetry data along
with information about its topology.
In this paper, we lay out the theoretical foundation for the IPF problem and explore its
application in detecting faults and other critical events. Using the bus injection model (BIM)
[3], we propose efficient algorithms to identify the admittance matrix. In particular, we show that

3
when the system has no hidden state the admittance matrix can be uniquely identified from a
sequence of complex voltage and current measurements corresponding to different steady states.
Should there be some hidden states in the network, we show that a reduced admittance matrix
(from Kron reduction [4]) can be determined; we develop the following algorithms for identi-
fying the admittance matrix of the original system: a) for radial networks, a graph-theoretical
approach based on graph decomposition, maximal clique searching and composition, b) for
mesh networks, an algorithm based on low rank and sparse matrix decomposition. Furthermore,
a convex relaxation has been proposed to obtain a computationally efficient algorithm. Power
flow simulations are performed on the IEEE 14-bus system to back up the theoretical results
and evaluate their sensitivity to the measurement noise introduced by transducers.
The paper is outlined as follows: we formulate the IPF problem in Section II-A and propose
a solution for the case that the system is fully observable in Section II. We evaluate the
identification accuracy of the proposed algorithms in Section IV. When the system has hidden
nodes, we propose two efficient algorithms to solve the IPF problem for radial and mesh networks
in Section III. We survey related work in Section V and conclude the paper by presenting
directions for future work in Section VI.
II. IPF WITHOUT HIDDEN NODES
In this section we consider the IPF problem when voltage and current phasor measurements
are available at every bus in the system. We formulate the identification problem as a constrained
least square problem and convert it into an equivalent unconstrained least square problem. We
note that Y has a certain structure that can be exploited when solving the IPF problem— (a)
Y must be a symmetric complex matrix (i.e., Y ∈ SN ) and (b) Y encodes the topology of a
connected graph (or a connected tree for radial networks).
A. Problem formulation
Let C denote the set of complex numbers, R the set of real numbers, and N the set of
integers. For A ∈ Cn×n , Re(A) and Im(A) denote matrices with the real and imaginary parts of
A, respectively. Let Sn be the set of all n × n symmetric matrices. The transpose of a matrix A
is denoted AT and its Hermitian (complex conjugate) transpose is denoted AH . A[i, j] denotes
the element of A located at ith row and jth column. We define I as the identity matrix with an
appropriate dimension and define 1 as an all-1 column vector with an appropriate dimension.

4
A power system can be modeled by an undirected connected graph G = (N , E) where N :=

{1, 2, . . . , N } represents the set of buses, and E ⊆ N × N represents the set of lines, each
connecting two distinct buses. A bus j ∈ N can be a load bus, a generator bus, or a swing bus.
Let Vj be the complex voltage at bus j and sj be the net complex power injection (generation
minus load) at that bus. We use sj to denote both the complex number pj + iqj and the real pair
√
(pj , qj ) depending on the context. We use i both to denote −1 and an index depending on
the context. For each line (i, j) ∈ E, we denote its series admittance by yij . The bus admittance
matrix of this system is denoted Y , which is an N × N complex-valued matrix whose off-
P
diagonal elements are Yij = −yij and diagonal elements are Yii = − j6=i Yij , assuming that
there is no shunt element (this assumption can be relaxed). Hence, the current injection vector
can be expressed as I = Y V . The bus injection model (BIM)1 is defined by the following power
flow equations describing the Kirchhoff’s law for a given time index, k ∈ {1, . . . , K}:
X
yijH |Vi (k)|2 − Vi (k)VjH (k) , ∀i ∈ N ,

si (k) = (1)
j∈Ni
where Ni is the set of nodes directly connected to node i. Rewriting this formula in vector form
for all time indices yields the following equation for a bus i:
 H  
si (1)  
H
 ViH (1)   1 V (1) V 2 (1) . . . V N (1) Yi1
 si (2)    
 V H (2)   V1 (2) V2 (2) . . . VN (2)    Yi2 
 
 i =  . . (2)
 ..   .. .. .. ..   . 
 .   . . . .   . 
 H 
si (K)
ViH (K)
V1 (K) V2 (K) . . . VN (K) YiN
| {z } | {z } | {z }
IiK VK Yi
The IPF problem is: given voltage and power (or current) measurements of different steady-
states, i.e., Vi (k) and si (k) for k = 1, . . . , K, whether it is possible to recover the admittance
matrix Y . If so, how many steady-state measurements are required to uniquely identify Y ;
otherwise, what part of Y can be identified and where additional sensors should be installed so
that we can identify the entire admittance matrix.
Since given (Vi (k), si (k)) we can form the current vector IiK in (2), we henceforth assume
we have voltage and current measurements (Vi (k), Ii (k), k = 1, . . . , K).
1
We use the bus injection model in this paper; however, our approach can be generalized to other models, e.g., the DC
power flow model or the Distflow model [5].

5
B. Identification algorithm
The admittance matrix Y can be obtained from solving the optimization problem below:
min 0, (3)
X
s.t.: V K Y = I K , Y ∈ SN , Yii = − Yij , ∀i.
j6=i
h i
in which I K is a K × N matrix, i.e., I K = I1K I2K . . . INK . This feasibility problem can
be also written as a constrained least squares problem using matrix Frobenius norm:

Ŷ K,l2 , arg min V K Y − I K F (4)
Y
X
s.t.: Y ∈ SN , Yii = − Yij , ∀i.
j6=i
Define
h iT
vec(Y ) = Y11 Y21 . . . YN 1 Y12 Y22 . . . YN N ,
and apply the vec operator to the objective function and constrains of the above problem, we
obtain:

I ⊗ V K vec(Y ) − vec(I K )

min 2
(5)
vec(Y )∈CN 2 ×1
X
s.t.: Y ∈ SN , Yii = − Yij , ∀i,
j6=i
in here ⊗ is the Kronecker product. This holds because vec(ABC) = (C T ⊗ A)vec(B). Let
2 −N )/2×1
svec : CN ×N → C(N be a mapping from a symmetric complex matrix to a complex
vector defined as:
h iT
svec(Y ) = Y21 Y31 . . . YN 1 Y32 Y42 . . . YN N −1 .
It can be readily seen that svec is a bijection for any matrix Y satisfies a) Y ∈ SN and b)
2 ×(N 2 −N )/2
1T Y = 0. Based on this definition, we have vec(Y ) = Γsvec(Y ), where Γ ∈ RN
maps svec(Y ) to the vectorized admittance matrix as illustrated below.2
Example 1. For the network depicted in Figure 1, the Γ matrix has the following form
2
If Y includes shunt elements so that 1T Y 6= 0 then svec(Y ) will include diagonal elements (Y11 , . . . , YN N ).

6
Fig. 1: An illustrative example of a three-node power system with two lines connecting bus 1
to buses 2 and 3 respectively.
   
Y11 −1 −1 0
   
Y21   1 0 0
   
   
Y   0 1 0
 31   
    
Y   1 0 0
 12    Y21
    
Y22  = −1 0  Y31  .
−1  
  
   
Y32   0 0 1
    Y32
    | {z }
Y13   0 1 0
    svec(Y )
   
Y23   0 0 1
   
Y33 0 −1 −1
| {z } | {z }
vec(Y ) Γ
Based on the definition of Γ, the constrained l2 optimization problem can be converted to an

unconstrained l2 optimization:

K
K

min2 I ⊗V Γ svec(Y ) − vec(I )
svec(Y )∈C(N −N )/2×1 |
{z } , (6)

F 2
in which I denotes an identity matrix.

We define    
K
Re(F ) −Im(F ) Re(vec(I ))
F̃ =   , and b̃ =  .
K
Im(F ) Re(F ) Im(vec(I ))
The optimization problem (6) can be written as

˜
min2 F̃ f (Y ) − b̃ ,

(7)
f˜(Y )∈R(N −N )×1 2
in which f˜(Y ) , [svec(YR )T svec(YI )T ]T , YR = Re(Y ), and YI = Im(Y ). This least square
problem yields a solution provided that M̃ has full column rank:
−1
f˜(Y ) = F̃ T F̃ F̃ T b̃. (8)

7
We compute the solution of the original optimization problem (6) from the solution of the
optimization problem (7) by taking the inverse map of f˜. A sufficient condition to guarantee the
exactness of the solution is that V K has full column rank.
Proposition 1 (Exactness). If V K has full column rank, the optimization problem (7) has a
unique solution as in (8).
2 ×(N 2 −N )/2
Proof: Since Γ ∈ RN and has full column rank (this can be checked easily), there
2
exists a matrix Γ† such that Γ† Γ = I. For the Kronecker product I ⊗ V K ∈ CKN ×N , I ⊗ V K
has full column rank when V K has full column rank; therefore, F̃ and F have full column rank
given the fact that rank(F̃ ) = 2rank(F ).
Finally, we prove by contradiction that if F̃ has full column rank, the solution to the opti-
mization problem (7) is unique. Suppose there exists f˜(Y1 ) and f˜(Y2 ) (f˜(Y1 ) 6= f˜(Y2 )) such that
F̃ f˜(Y1 ) = b̃ and F̃ f˜(Y1 ) = b̃, then

F̃ f˜(Y1 ) − f˜(Y2 ) = 0,
which contradicts the full column rank assumption.
Remark 1. Our approach can be easily extended to the case of nonzero shunt elements where
1T Y 6= 0 by changing the definitions of svec(Y ), Γ and f (·).
Remark 2. When V K does not have full row rank, we can characterize the part of the admittance
matrix that is identifiable (see [6]).
We can add the element-wise positivity constraint to this problem if the conductance and
susceptance of each line are positive3 .

min F̃ f˜(Y ) − b̃ .

(9)
f˜(Y )≥0 2
The above problem is known as nonnegative least squares and can be solved using different
methods, such as the active set method [7].
3
The conductance of a line is always positive, the susceptance can be negative or positive depending on whether the line
is inductive or capacitive.

8
Fig. 2: Left: the true topology of the power system; Middle: the topology that we can infer
from sensory data if we neglect the effect of hidden nodes; Right: the topology that we would
like to obtain.
III. IPF WITH HIDDEN NODES
In the previous section we solve the IPF problem when voltage and current measurements
are available at all buses. In this section we consider the case when measurements are available
only at a subset of the buses. In a distribution system, for example, measurements are typically
available at the substation and customer meters but not throughout the grid.
In any case, solving the IPF problem with hidden nodes is more challenging due to the intrinsic
difficulties explained in Fig. 2. Specifically, the left figure shows a power system with a single
hidden node and a sparse topology. Even though there is only one hidden node, the inferred
network topology from the measurements will be a full mesh as shown in the middle figure. The
right figure shows the topology of the network interconnecting the measured nodes. A desired
identification algorithm should correctly recover/identify this topology, neglecting the additional
links that are introduced because of the hidden node.
In what follows, we first show that, in the presence of hidden nodes, the algorithm in Section
II can identify a Kron-reduced admittance matrix Ȳ defined in (13) below for the nodes where
measurements are available (Corollary 1) in Section III-A. Moreover Ȳ is the best that can be
identified for general networks. In Section III-B we show that when the network is a tree then it
is indeed possible to identify uniquely the original admittance matrix Y from its Kron reduction
under reasonable assumptions. In Section III-C, we propose a low rank and sparse decomposition
method for mesh networks.
A. Kron-reduced admittance matrix Ȳ
We call a bus/node an measured bus/node if measurements of its voltage and current injection
are available for identification and a hidden bus/node otherwise. Let M1 and M2 represent the

9
set of measured and hidden nodes respectively.
Assumption 1. We make the following assumptions:

1.1 The underlying graph G is connected.
1.2 Hidden nodes have zero injection Ii (k) = 0 for all i ∈ M2 .
We make an important assumption that the current injection is zero at every hidden node. This
is reasonable since, if generators or loads are connected to a node, their injections or withdrawals
will likely be measured.
Let H be the number of hidden nodes. Without loss of generality we label the buses so that the
first N −H buses are measured and the last H buses are hidden, i.e., M1 := {1, . . . , M , N −H}
and M2 := {N −H+1, . . . , N }. We partition the bus admittance matrix Y into four sub-matrices:
     
Y11 Y12 G G12 B B12
Y =   =  11  + i  11  = G + iB (10)
Y21 Y22 G21 G22 B21 B22
Here Y11 ∈ SN −H describes the connectivity among the measured nodes, Y12 = Y21T ∈ C(N −H)×H
the connectivity between the measured and the hidden nodes, and Y22 ∈ SH the connectivity
among the hidden nodes. For i ∈ M1 , (Ii (k), Vi (k), k = 1, . . . , K) denote the current and
voltage measurements at bus i at time k. To simplify notation, we index the entries of Y22 , not
by i, j = 1, . . . H, but by i, j = N −H+1, . . . , N . We index the entries of Y12 by i = 1, . . . , N −H
and j = N − H + 1, . . . , N and similarly for Y21 = Y12T , as well as submatrices Gij , Bij in (10).
For i ∈ M2 , we have Ii (k) = 0, ∀k, and Vi (k) is the voltage at bus i but is not available for
identification. To simplify notation, We abuse V1 (k) to denote both the voltage at bus 1 at time
k and the vector of all voltages at measured buses at time k, depending on the context; similarly
for V2 (k) and I1 (k). Then
    
I (k) Y Y V (k)
 1  =  11 12   1  , ∀k (11)
0 Y21 Y22 V2 (k)
If Y22 is invertible then eliminating V2 from (11) yields a relation
I1 (k) = Ȳ V1 (k), ∀k (12)
between currents and voltages at measured nodes through the Kron-reduced admittance matrix

10
Ȳ ∈ Sm defined as:
Ȳ , Y11 − Y12 Y22−1 Y12T (13)
for the set of measured nodes. In the rest of this subsection we first justify the invertibility of
Y22 and hence the definition of Ȳ . Proposition 1 then implies that Ȳ can be identified from
sufficient voltage and current measurements. Moreover Ȳ is the best we can identify for general
networks because multiple admittance matrices Y may reduce to the same Ȳ .
Assumption 2. The admittance matrix Y , G + iB defined in (10) satisfies:

2.1 The off-diagonal elements of G are non-positive and the off-diagonal elements of B are
non-negative.
P P
2.2 Diagonal dominance: G22 [i, i] ≥ − j6=i G22 [i, j] and −B22 [i, i] ≥ j6=i B22 [i, j] for any
i.
These assumptions mean that the series impedances of the lines are resistive and inductive.
Lemma 1. If H < N , then G22 0, −B22 0 and G22 − B22 0.
Proof: From the Gershgorin Theorem and Assumption 2.2, all eigenvalues of the submatrix
G22 lie in the right-half plane including the origin and all eigenvalues of B22 lie in the left-half
plane including the origin. Together with the fact that G22 and B22 are symmetric, we have
G22 0 and −B22 0.
This implies that G22 − B22 0. We now show that indeed G22 − B22 0. Suppose for the
sake of contradiction that there exists a nonzero x ∈ RH such that xT (G22 − B22 )x = 0. Denote
by A22 := G22 − B22 and A21 := G21 − B21 so that
X X
A22 [i, i] = (−A22 [i, j]) + (−A21 [i, j])
i,j∈M2 :j6=i i∈M2 ,j∈M1
Then
X X
xT (G22 − B22 )x = A22 [i, j]xi xj − A22 [i, j]x2i (−A21 [i, j])x2i

+
i,j∈M2 :i6=j i∈M2 ,j∈M1
X X
= (−A22 [i, j])(xi − xj )2 + (−A21 [i, j])x2i (14)
i,j∈M2 :i<j i∈M2 ,j∈M1
By definition A22 [i, j] = A21 [i, j] = 0 if i 6∼ j. For i, j ∈ M1 ∪ M2 , if i ∼ j, then Yij =

11
Gij + iBij 6= 0, i.e., at least one of Gij = −gij ≤ 0 or Bij = −bij ≥ 0 is nonzero. This implies
that −A22 [i, j] = −G22 [i, j] + B22 [i, j] > 0 and −A21 [i, j] = −G21 [i, j] + B21 [i, j] > 0 for all
i ∼ j. Therefore, for xT (G22 − B22 )x = 0 in (14), we must have:
1. xi = xj if i ∼ j is a connection between hidden nodes in M2 ;
2. xi = 0 for any hidden node i ∈ M2 connected to at least one observed node j ∈ M1 .
Since the network is connected, for every hidden node i ∈ M2 , there is a path that connects
the hidden node to an observed node k ∈ M1 . For all nodes j on this path from i to k, the
above properties implies that xj = 0. Since this is true for all hidden nodes, we have x = 0, a
contradiction. Hence G22 − B22 0.
Recall that the network is connected and has N nodes of which H are hidden.
Proposition 2. Under Assumptions 1 and 2, if H < N then Y22 is invertible.
Proof: We prove that 0 is not an eigenvalue of Y22 . Suppose for the sake of contradiction
that there exists a nonzero vector (v + iw) such that (G22 + iB22 )(v + iw) = 0, i.e.,
G22 v − B22 w = 0 and G22 w + B22 v = 0,
This implies
(G22 + B22 )v + (G22 − B22 )w = 0,
(15)
(G22 + B22 )w − (G22 − B22 )v = 0.
Since G22 − B22 0 by Lemma 1, we can eliminate v from (15) to get
(G22 − B22 ) + (G22 + B22 )(G22 − B22 )−1 (G22 + B22 ) w = 0

Multiplying wT on the left we have
wT (G22 − B22 )w + ŵT (G22 − B22 )−1 ŵ = 0
where ŵ = (G22 + B22 )w. This contradicts G22 − B22 0 unless w = 0. But w = 0 implies that
(G22 − B22 )v = 0 from (15), meaning v = 0. This is a contradiction and hence Y22 is invertible.
Proposition 2 guarantees that Y22 is invertible and hence the Kron-reduced admittance Ȳ in
(13) is well defined under Assumption 2. Moreover, because of (12), the algorithm in Section
II can identify the Kron-reduced admittance matrix Ȳ from sufficient number of voltage and

12
current measurements (Proposition 1).
Corollary 1. Suppose Assumptions 1 and 2 and the condition in Proposition 1 hold. If H < N ,
then the Kron-reduced admittance matrix Ȳ can be identified from voltage V1K and current I1K
measurements at the measured nodes.
An admittance matrix Y ∈ SN specifies a unique weighted undirected graph G = (N , E) with

N := {1, . . . , N } and E ⊆ N × N such that there is an edge (i, j) if and only if Y [i, j] 6= 0.
We sometimes refer to Y and its underlying graph jointly as the graph G = (N , E, Y ). Its Kron
reduction Ȳ similarly specifies a unique weighted graph Gc = (M1 , Ec , Ȳ ) that can be obtained
from G through Algorithm 1. Each iterative step in the algorithm removes a hidden node i ∈ M2
Algorithm 1 Graph Condensation Algorithm

1: Input: a graph G = (N , E, Y ) with N nodes and a set of measured nodes M1 =
{1, 2, . . . , M }
2: for v = M + 1 : N do
3: Remove hidden node v from N = N − {v} and all edges from E that are incident on v;
4: For all node pairs w and l that are neighbors of v, add an edge between w and l to E;
5: Update the admittance matrix Y = Y /Y [i, i] using eq. (16).
6: end for
7: return Gc = G and Ȳ = Y .
and connects all its neighbors to each other. This step can be represented algebraically as an
update on the admittance matrix to compute its Schur complement of Yii . Specifically we can
partition possibly after permutation an admittance matrix Y of an appropriate dimension into
the form:  
Y (i, i) Y (i, i]
Y = .
Y (i, i]T Y [i, i]
where Y [i, i] ∈ C is the ith diagonal element of Y and

   
Y [1, 1] ... Y [1, i − 1] Y [1, i + 1] ... Y [1, N ] Y [1, i]
 .
.. . .. .
.. .. .. ..   .. 

 . . . 


 . 

   
Y [i − 1, 1] . . . Y [i − 1, i − 1] Y [i − 1, i + 1] ... Y [i − 1, N ]
 Y [i − 1, i]
Y (i, i) =   , Y (i, i] =  .
  
Y [i + 1, 1] . . . Y [i + 1, i − 1] Y [i + 1, i + 1] ... Y [i + 1, N ] Y [i + 1, i]
   
 .. .. .. .. .. ..   .
..


 . . . . . . 





Y [N, 1] . . . Y [N, i − 1] Y [N, i + 1] ... Y [N, N ] Y [N, i]

13
Kron Reduction with respect to M1

<latexit sha1_base64="fmv6/xgIcGPx3JJdd4R1A+hBrrE=">AAACF3icbVDLSgMxFM3UV62vUZdugq3gqsxUQZcFN4IIVewD2qFk0ts2NJMZkoxShv6FG3/FjQtF3OrOvzHTzkJbDwQO59xL7jl+xJnSjvNt5ZaWV1bX8uuFjc2t7R17d6+hwlhSqNOQh7LlEwWcCahrpjm0Igkk8Dk0/dFF6jfvQSoWijs9jsALyECwPqNEG6lrl69kKPAt9GKaCviB6SGWoCKgGusQlzoB0UNKeHI96bqlrl10ys4UeJG4GSmiDLWu/dXphTQOQGjKiVJt14m0lxCpGeUwKXRiBRGhIzKAtqGCBKC8ZJprgo+M0sP9UJonNJ6qvzcSEig1DnwzmV6p5r1U/M9rx7p/7iVMRLEGQWcf9WOeBk5Lwj0mTX4+NoRQycytmA6JJFSbKgumBHc+8iJpVMruSblyUylWT7M68ugAHaJj5KIzVEWXqIbqiKJH9Ixe0Zv1ZL1Y79bHbDRnZTv76A+szx8pHJ86</latexit>
Y Ȳ
Y /Y [1, 1] ...
Graph Condensation Algorithm 1
<latexit sha1_base64="yrxkDnAy/DKXKALcSTkAB/6NbNA=">AAACB3icbVDLSsNAFJ3UV62vqEtBBovgqiR1YZeVLnRZwT6gDWUymTRDJzNhZiKU0J0bf8WNC0Xc+gvu/BunaRbaeuDC4Zx7uZzjJ4wq7TjfVmltfWNzq7xd2dnd2z+wD4+6SqQSkw4WTMi+jxRhlJOOppqRfiIJin1Gev6kNfd7D0QqKvi9nibEi9GY05BipI00sk9vJEoi2BI8IFzlIrxmYyGpjmLojuyqU3NywFXiFqQKCrRH9tcwEDiNCdeYIaUGrpNoL0NSU8zIrDJMFUkQnqAxGRjKUUyUl+U5ZvDcKAEMhTTDNczV3xcZipWaxr7ZjJGO1LI3F//zBqkOG15GeZJqwvHiUZgyqAWclwIDKgnWbGoIwiY6xRBHSCKsTXUVU4K7HHmVdOs197JWv6tXm42ijjI4AWfgArjgCjTBLWiDDsDgETyDV/BmPVkv1rv1sVgtWcXNMfgD6/MHPzGY3g==</latexit>
Fig. 3: Two equivalent schemes to compute the Kron reduced Ȳ .
Then each iterative step of Algorithm 1 updates the admittance matrix by:
Y /Y [i, i] = Y (i, i) − Y (i, i]Y −1 [i, i]Y (i, i]T . (16)
The algorithm repeats this step until all hidden nodes are removed from the original graph,
producing the Kron-reduced graph and its admittance matrix Gc = {M1 , Ec , Ȳ } [4] (shown in
Fig. 3).
Given an admittance matrix Y , each partition (Y11 , Y12 , Y22 ) in (10) defines uniquely a Kron-
reduced matrix Ȳ := Ȳ (Y11 , Y12 , Y22 ) given by (13). This mapping is clearly not injective in
general, i.e., given an M × M symmetric matrix Ȳ ∈ SM (possibly with Ȳ 1 = 0) there
are generally multiple N × N symmetric matrices Y that can be partitioned into (non-unique)
(Y11 , Y12 , Y22 ) whose Kron reductions are the given Ȳ , as long as N ≥ M .
Example 2. Consider a (Kron-reduced) admittance matrix for a two-node network:

 
θ −θ
Ȳ =  .
−θ θ
The following 3 × 3 admittance matrice Y with the given partition has Ȳ as its Kron reduction:
 
  θ + θ0 −θ −θ 0
Y11 Y12  
Y :=   =   −θ θ 0 .

Y12T Y22  
−θ0 0 θ0
for arbitrary θ0 ∈ R. The underlying network is shown in Fig. 1 with node 3 as the hidden node,
so that Ȳ corresponds to the Kron-reduced admittance matrix for nodes 1 and 2 in Fig. 1. In
this case the hidden node has degree 1. Another 3 × 3 admittance matrices Ỹ that also has Ȳ

14
as its Kron reduction is:

 
  θ1 0 −θ1
Ỹ11 Ỹ12  
Ỹ :=   = 
 0 θ2 −θ2 ,

Ỹ12T Ỹ22  
−θ1 −θ2 θ1 + θ2
as long as (θ1 , θ2 ) satisfies
θ1 θ2
= θ.
θ1 + θ2
The network underlying Ỹ is isomorphic to that in Fig. 1 with node 1 being the hidden node,
so that Ȳ is the Kron-reduced admittance matrix for nodes 2 and 3 in Fig. 1. In this case the
hidden node has degree 2.
Example 2 shows that in general only the Kron-reduced admittance matrix Ȳ is identifiable
from measurements at the measured nodes. For arbitrary networks it is impossible to identify
the original admittance matrix Y whose Kron reduction yields Ȳ . We next show the surprising
result that, when the network is a tree, then the original admittance matrix Y can indeed be
identified even in the presence of hidden nodes.
B. Radial networks: exact identification
Consider a radial network and suppose we have identified a Kron-reduced admittance matrix
Ȳ from partial voltage and current measurements. In this section we develop a novel algorithm
to compute the original admittance matrix Y from Ȳ under the following additional assumptions.
Assumption 3. The admittance matrix Y satisfies:

3.1 The underlying graph G(Y ) is a tree.
3.2 Every hidden node has a degree ≥ 3.
3.3 There is no shunt element in Y , i.e., Y 1 = 0.
Assumption 3.2 is a necessary condition for identification as shown in Example 2 where the
hidden node has a degrees 1 or 2.
We start with some definitions. Consider an undirected graph G = (N , E) where N :=
{1, . . . , N } is the set of nodes and E ⊆ N × N is the set of edges. A complete graph is one
in which all nodes are adjacent. A subgraph of G is a graph G 0 = (N 0 , E 0 ) with N 0 ⊆ N and

15
E 0 ⊆ E. A clique of G is a complete subgraph of G. A maximal clique of G is a clique that is

not a subgraph of another clique of G. We say G is a tree if there is exactly one path between
every two nodes. A forest is a disjoint union of trees.
For our purposes, an admittance matrix Y is a square complex symmetric matrix (we usually
assume Y satisfies Assumption 3.3). Given any N × N admittance matrix Y , let G(Y ) :=
(N (Y ), E(Y )) denote the graph where N (Y ) := {1, . . . , N } and there is an edge (i, j) ∈ E(Y )
if and only if Y [i, j] 6= 0. We sometimes refer to G(Y ) as the underlying graph of Y and write
G := (N , E) when Y is clear from the context. Consider two N × N admittance matrices Y1
and Y2 . We define two functions of Y1 , Y2 and their underlying graphs. First Y3 := Y1 + Y2 is
also an admittance matrix and its underlying graph G(Y3 ) = (N (Y3 ), E(Y3 )) is the graph with
the same set of nodes and edges in both graphs, i.e., E(Y3 ) := E(Y1 ) ∪ E(Y2 ). When the matrices
are clear from the context, we also denote the function Y3 = Y1 + Y2 by G3 = G1 ⊕ G2 . Note that
if Y1 and Y2 satisfy Assumption 3.3, so does Y3 . Second define the N × N matrix Y4 := Y1 /Y2
as a function of (Y1 , Y2 ) by:

 Y1 P


 [i, j] if i ∼ j and (i, j) 6∈ E2
Y4 [i, j] := − j Y4 [i, j] if i = j


 0

otherwise
The underlying graph G(Y4 ) is a subgraph of G(Y1 ) where edges in G(Y2 ) have been removed.
When the matrices are clear from the context, we also denote the function Y4 = Y1 /Y2 by
G4 = G1 /G2 . Note that Y4 satisfies Assumption 3.3 by definition.
1) Exact identification of Y : Fix an (unknown) admittance matrix Y and assume its underlying
graph G := G(Y ) is a tree. Suppose its Kron-reduced admittance matrix Ȳ and its underlying
graph Ḡ := G(Ȳ ) are given. For example Ȳ is obtained according to Corollary 1 from partial
voltage and current measurements at measured nodes in M1 .
Decomposition of Ḡ. Let E1 denote the subset of all edges of Ḡ that are between measured
nodes in the original graph G, and E2 denote the subset of all edges of Ḡ that have been added
by Step 4 of the graph condensation Algorithm 1 when hidden nodes are removed from G. By
definition of E1 , E2 , we have Ḡ = (M1 , E1 ∪ E2 ).
Lemma 2. E1 ∩ E2 = ∅.
Proof: If there exists an edge (i, j) ∈ E1 ∩ E2 , then (i, j) must be an edge in the original

16
graph G and nodes i and j must also be connected through a path consisting of only hidden
nodes. This creates a loop and contradicts that G is a tree. Hence E1 ∩ E2 = ∅.
Since Ḡ = (M1 , E1 ∪ E2 ), Lemma 2 motivates decomposing Ḡ into two subgraphs, G1 :=
(M1 , E1 ) and G2 := (M1 , E2 ), both defined on M1 of measured nodes but with disjoint edge
sets. While the graph Ḡ := G(Ȳ ) is defined by the Kron-reduced admittance matrix Ȳ , at this
point the graphs G1 , G2 are only defined in terms of the graph Ḡ (in fact in terms of G) and are
not associated with any admittance matrices. Define the matrices:
Ȳ1 := Y11 − diag{1T Y11 } (17a)
Ȳ2 := diag{1T Y11 } − Y12 Y22−1 Y12T . (17b)
The key observation, stated in the next result, is that G1 , G2 have simple structures, that the
matrices defined in (17) are indeed admittance matrices, and that G1 , G2 are the underlying
graphs of these admittance matrices.
Theorem 1 (Separability). Suppose the admittance matrix Y satisfies Assumptions 2 and 3. Then
1. G1 is a forest.
2. G2 = ⊕i Ci for some Ci are edge-disjoint maximal cliques each with more than 2 nodes.4
3. G1 = G(Ȳ1 ) and G2 = G(Ȳ2 ) so that Ḡ = G1 ⊕ G2 .
Proof: For the first assertion, G1 is a forest since it is a subgraph of the tree G. For the
second assertion G2 is a collection of maximal cliques Ci due to Step 4 of the graph condensation
Algorithm 1. To show that the maximal clique (in each) Ci is of size at least 3, suppose Ci consists
of mi (measured) nodes and, in the original graph G, these mi measured nodes “surround” hi
hidden nodes, i.e., the neighbors of each of these hidden nodes are either hidden nodes or
nodes in Ci . Let dj denote the degrees of hidden nodes j = 1, . . . , hi . These mi + hi nodes
form a (connected) subtree of G with exactly mi + hi − 1 edges. Since mi of these edges are
between measured and hidden nodes and hi − 1 edges are between hidden nodes, we must
have hj=1 dj = mi + 2(hi − 1) and hence mi = 2 + hi=1
P i P i
(di − 2). Since hi ≥ 1 and di ≥ 3
(Assumption 3.2), we have mi ≥ 3. To show that Ci and Cj are edge-disjoint, suppose for the
sake of contradiction that there is an edge (k, l) in both Ci and Cj . By the definition of G2 , (k, l)
4
Strictly speaking, each Ci is a subgraph of G2 with M1 as its node set. It consists of a single maximal clique and the
remaining isolated nodes in M1 . We will abuse notation and use Ci to both refer to this subgraph of G2 and to the maximal
clique in Ci .

17
is not an edge in the original graph G. Since nodes k, l are both in Ci , there is a path from k to
l in G that consists of only hidden nodes connected to measured nodes in the maximal clique
Ci . Since nodes k, l are both in Cj , there is disjoint path from k to l in G that consists of a set
of hidden nodes connected to nodes in Cj . These two paths form a loop in G, a contradiction.
Hence Ci and Cj do not share any edge in G2 .
For the third assertion, note that the matrix Ȳ1 defined in (17a) is symmetric and hence an
admittance matrix. The diagonal entry Y11 [i, i] of Y11 is the negative sum of the off-diagonal
entries of the ith row/column of the original admittance matrix Y (plus any shunt element at
bus i), so that the ith entry of 1T Y11 is equal to the ith row sum of Y12 (plus any shunt element
at bus i). Hence Ȳ1 satisfies Assumption 3.3 if Y does. Moreover, by the definition of G1 , the
edges in E1 correspond exactly to the off-diagonal entries of Y11 that are nonzero. This implies
that the graph G(Ȳ1 ) that underlies the admittance matrix in (17a) is indeed G1 .
The matrix Ȳ2 defined in (17b) is also symmetric and hence is an admittance matrix. If Y
satisfies Assumption 3.3, then
1T Y11 = −1T Y12T , 1T Y12 = −1T Y22 .
This implies
diag{1T Y11 } = diag{1T Y12 Y22−1 Y12T },
i.e., Ȳ2 defined in (17b) satisfies Assumption 3.3 when Y does.

Next we show that G2 = G(Ȳ2 ). From (17)
Ȳ = Y11 − Y12 Y22−1 Y12T = Ȳ1 + Ȳ2 ,
and hence Ḡ := G(Ȳ ) = G1 ⊕ G2 with G(Ȳ1 ) = G1 . Therefore we have G(Ȳ2 ) = G2 . This

concludes the proof.
Finding the maximum clique of a graph is termed the clique problem and is NP-complete in
general. There are many algorithms for solving the clique problem, such as the Bron-Kerbosch
algorithm, which we adopt in Algorithm 2.
Identification of Y . Next we propose an algorithm to obtain Y11 , Y22 and Y12 , and therefore the
original admittance matrix Y .
The decomposition of Ḡ into G1 and G2 guaranteed by Theorem 1 allows us to partition the set
M1 into a subset of internal measured nodes that are not connected to any hidden nodes and a

18
Algorithm 2 Graph Decoupling Algorithm

1: Input: a condensed graph Ḡ
2: Set G 0 = Ḡ, i = 1.
3: while G 0 has a clique with more than two nodes do
4: Use Bron-Kerbosch Algorithm to find a clique (≥ 3 nodes, together with other isolating
nodes) Ci in G 0 ,
5: Let G 0 = G 0 /Ci , i = i + 1,
6: end while
7: return G2 = ⊕i Ci and G1 = Ḡ/G2
disjoint subset of boundary measured nodes that connect to some hidden nodes. We can similarly
partition M2 into a subset of internal hidden nodes that are not connected to any measured nodes
and the disjoint subset of boundary hidden nodes that connect to some measured nodes. The
decomposition of Ḡ into G1 and G2 identifies only the types of measured nodes, but not those
of hidden nodes. We can hence arrange the original admittance matrix Y into the following
structure (only the upper triangular submatrix is shown as Y is symmetric):
 
Y11,11 Y11,12 0 0
   
Y11 Y12  Y 11,22 Y12,21 0 
Y =   =:  . (18a)


Y22 
 Y22,11 Y 
22,12 
Y22,22
Here, for Y11 , the submatrix Y11,11 corresponds to connectivity among the internal measured
nodes, Y11,22 corresponds to connectivity among the boundary measured nodes, and Y11,12 cor-
responds to connectivity between the internal and boundary measured nodes. Similarly, for Y22 ,
the submatrix Y22,11 corresponds to connectivity among the boundary hidden nodes, Y22,22 to that
among the internal hidden nodes, and Y22,12 to that between the internal and boundary hidden
nodes. The submatrix Y12,21 corresponds to connectivity between the set of boundary measured
nodes and the set of boundary hidden nodes. Denote the inverse Y22−1 by:
 
X22,11 X22,12
X22 := Y22−1 =:  .
T
X22,12 X22,22

19
−1
We have from (18a), X22,11 = (Y22,11 − Y22,12 Y22,22 T
Y22,12 )−1 from Woodbury formula
   
Y11,11 Y11,12 0 0
Ȳ = Y11 − Y12 Y22−1 Y12T =   −  . (18b)
T T
Y11,12 Y11,22 0 Y12,21 X22,11 Y12,21
Since we can compute Ȳ from partial voltage and current measurements, we can identify
submatrices Y11,11 and Y11,12 for internal measured nodes from Ȳ according to (18b). The edges in
T
E1 correspond to the off-diagonal entries of [Y11,11 Y11,12 ] as well as Y11,12 , and they form a forest
T
(Theorem 1). The edges in E2 correspond to the off-diagonal entries of Y11,22 −Y12,21 X22,11 Y12,21 ,
and they form a collection of cliques. Recall that both G1 and G2 have M1 as their node set;
see the example in Section IV or Figure 4.
G(Y )
<latexit sha1_base64="b8SZawwe/873TTo0tiqfWrVSUvE=">AAAB9XicbVDLSgMxFL1TX7W+qi7dBItQN8NMXeiy4EJXUsE+pB1LJs20oZnMkGSUMvQ/VHChiFv/xZ1/Y6btQlsPBA7n3Ms9OX7MmdKO823llpZXVtfy64WNza3tneLuXkNFiSS0TiIeyZaPFeVM0LpmmtNWLCkOfU6b/vA885v3VCoWiRs9iqkX4r5gASNYG+muE2I9IJinF+Py7XG3WHJsZwK0SNwZKVXtpwzPtW7xq9OLSBJSoQnHSrVdJ9ZeiqVmhNNxoZMoGmMyxH3aNlTgkCovnaQeoyOj9FAQSfOERhP190aKQ6VGoW8ms5Rq3svE/7x2ooMzL2UiTjQVZHooSDjSEcoqQD0mKdF8ZAgmkpmsiAywxESbogqmBHf+y4ukUbHdE7tybdq4ginycACHUAYXTqEKl1CDOhCQ8Aiv8GY9WC/Wu/UxHc1Zs519+APr8wdikZY2</latexit>
Y
<latexit sha1_base64="OtenIk1rubj2RKFoivKdetF2aYY=">AAAB6HicbZBNT8JAEIan+IX4hXr00khMPJEWDnqTxIsnA4l8GGjIdpnCynbb7G5NSMMv8OJBY7z6k/Tkv3EpHBR8k02evO9Mdmb8mDOlHefbyq2tb2xu5bcLO7t7+wfFw6OWihJJsUkjHsmOTxRyJrCpmebYiSWS0OfY9sfXs7z9iFKxSNzpSYxeSIaCBYwSbazGfb9YcspOJnsV3AWUrr6qmer94mdvENEkRKEpJ0p1XSfWXkqkZpTjtNBLFMaEjskQuwYFCVF5aTbo1D4zzsAOImme0Hbm/u5ISajUJPRNZUj0SC1nM/O/rJvo4NJLmYgTjYLOPwoSbuvInm1tD5hEqvnEAKGSmVltOiKSUG1uUzBHcJdXXoVWpexWy5WGU6rdwlx5OIFTOAcXLqAGN1CHJlBAeIIXeLUerGfrzXqfl+asRc8x/JH18QPJUI8t</latexit>
Ḡ1 := G(Ȳ1 ) Ȳ1 := Y11
<latexit sha1_base64="NXmAnuiqizpudZXX45skKNnvXho=">AAACFHicbVDLSsNAFJ3UV62vqEs3g0WoCCGpC0UQCi50JRXsQ5oQJtNJO3TyYGYilNCPcCPoj7hxoYhbF+78GydpF7X1wMCZc+7l3nu8mFEhTfNHKywsLi2vFFdLa+sbm1v69k5TRAnHpIEjFvG2hwRhNCQNSSUj7ZgTFHiMtLzBRea37gkXNApv5TAmToB6IfUpRlJJrn5ke4indoBkHyOWXo5GrnV2PvWv5AV3Sj509bJpmDngPLEmpFwznjI811392+5GOAlIKDFDQnQsM5ZOirikmJFRyU4EiREeoB7pKBqigAgnzY8awQOldKEfcfVCCXN1uiNFgRDDwFOV2bZi1svE/7xOIv1TJ6VhnEgS4vEgP2FQRjBLCHYpJ1iyoSIIc6p2hbiPOMJS5VhSIVizJ8+TZtWwjo3qjUrjGoxRBHtgH1SABU5ADVyBOmgADB7AC3gD79qj9qp9aJ/j0oI26dkFf6B9/QInv6Kn</latexit>
<latexit sha1_base64="cwUwn7TpVOgke1p8CvC5g+iZz1g=">AAACHXicbVDLSgMxFM34tr7qY+cmKIIby6QKiiAUXOhSwbZKpw6ZNFNDMw+SO2IN8yNu/AO/wY0LRVy4EZf+iWnrwteBCyfn3EvuPUEqhQbXfXeGhkdGx8YnJgtT0zOzc8X5hZpOMsV4lSUyUacB1VyKmFdBgOSnqeI0CiSvB539nl+/5EqLJD6BbsqbEW3HIhSMgpX84pYXUGXOcp/s7p35hpB8wwN+BaYlaDv3TP8RhIbk5ycDH3u5X1x1S24f+C8hX2S1sn79cXC3ZI784qvXSlgW8RiYpFo3iJtC01AFgkmeF7xM85SyDm3zhqUxjbhumv51OV6zSguHibIVA+6r3ycMjbTuRoHtjChc6N9eT/zPa2QQ7jSNiNMMeMwGH4WZxJDgXlS4JRRnILuWUKaE3RWzC6ooAxtowYZAfp/8l9TKJbJZKh/bNA7RABNoGa2gdUTQNqqgQ3SEqoihG3SPHtGTc+s8OM/Oy6B1yPmaWUQ/4Lx9AvT9pXw=</latexit>
diag{1T Y11 }
1 2 3 4 5 6 7
1 2 1 2
1 X X
Y11,11 Y11,12
<latexit sha1_base64="JGfW6F7NswQvodspDfxFWzq46k8=">AAAB8HicbVDLSgMxFL1TX7W+qi7dBIvgQsqkLnThoiCIG6GibS3tUDJp2oYmmSHJCGXoV7hxoYhbP8edX+EvmD4W2nrgwuGce7n3njAW3Fjf//IyS8srq2vZ9dzG5tb2Tn53r2aiRFNWpZGI9ENIDBNcsarlVrCHWDMiQ8Hq4eBy7NcfmTY8Uvd2GLNAkp7iXU6JdVKj0U4xPsF41M4X/KI/AVokeEYKZXRz9z0Irirt/GerE9FEMmWpIMY0sR/bICXacirYKNdKDIsJHZAeazqqiGQmSCcHj9CRUzqoG2lXyqKJ+nsiJdKYoQxdpyS2b+a9sfif10xs9zxIuYoTyxSdLuomAtkIjb9HHa4ZtWLoCKGau1sR7RNNqHUZ5VwIeP7lRVIrFfFpsXTr0riAKbJwAIdwDBjOoAzXUIEqUJDwBC/w6mnv2Xvz3qetGW82sw9/4H38AGBZkms=</latexit>
X X
<latexit sha1_base64="giFh1h8/iTywGBf4eA9WYvRkaak=">AAAB8HicbVDLSgMxFL1TX7W+qi7dBIvgQsqkLnThoiCIG6GibS3tUDJp2oYmmSHJCGXoV7hxoYhbP8edX+EvmD4W2nrgwuGce7n3njAW3Fjf//IyS8srq2vZ9dzG5tb2Tn53r2aiRFNWpZGI9ENIDBNcsarlVrCHWDMiQ8Hq4eBy7NcfmTY8Uvd2GLNAkp7iXU6JdVKj0U4xPsGlUTtf8Iv+BGiR4BkplNHN3fcguKq085+tTkQTyZSlghjTxH5sg5Roy6lgo1wrMSwmdEB6rOmoIpKZIJ0cPEJHTumgbqRdKYsm6u+JlEhjhjJ0nZLYvpn3xuJ/XjOx3fMg5SpOLFN0uqibCGQjNP4edbhm1IqhI4Rq7m5FtE80odZllHMh4PmXF0mtVMSnxdKtS+MCpsjCARzCMWA4gzJcQwWqQEHCE7zAq6e9Z+/Ne5+2ZrzZzD78gffxA2Hekmw=</latexit>
2
4 5 4 5 X
X
3 X X X
7 X
4 X X
5 X X X
6 6
6 X X X
3 7 X 3
Ḡ := G(Ȳ ) Ḡ2 := G(Ȳ2 )

Ȳ := Y11 Y12 Y22 1 Y12
T <latexit sha1_base64="UvvyXqKiucNFSkGIL+P1clwEsww=">AAACFHicbVDLSsNAFJ3UV62vqEs3wSJUhJDUhSIIBRe6kgr2IU0Ik+mkHTqZhJmJUEI/wo2gP+LGhSJuXbjzb5ykXdTWAwNnzrmXe+/xY0qEtKwfrbCwuLS8Ulwtra1vbG7p2ztNESUc4QaKaMTbPhSYEoYbkkiK2zHHMPQpbvmDi8xv3WMuSMRu5TDGbgh7jAQEQakkTz9yfMhTJ4SyjyBNL0cjr3p2PvWv5AV3Sj709LJlWjmMeWJPSLlmPmV4rnv6t9ONUBJiJhGFQnRsK5ZuCrkkiOJRyUkEjiEawB7uKMpgiIWb5keNjAOldI0g4uoxaeTqdEcKQyGGoa8qs23FrJeJ/3mdRAanbkpYnEjM0HhQkFBDRkaWkNElHCNJh4pAxIna1UB9yCGSKseSCsGePXmeNKumfWxWb1Qa12CMItgD+6ACbHACauAK1EEDIPAAXsAbeNcetVftQ/sclxa0Sc8u+APt6xcq4KKp</latexit>
Ȳ2 := diag{1T Y11 } Y12 Y22 1 Y12

T
<latexit sha1_base64="5772r7dpvkUZxuCrh0QChroLJYk=">AAACEHicbVDLSsNAFJ3UV62vqEs3g0Wsm5DUhSIIBRe6kgr2IU0ok+mkHTp5MDMRSsgnuHGnv+HGhSJuXbrzb5ykXdTWAwNnzrmXe+9xI0aFNM0frbCwuLS8Ulwtra1vbG7p2ztNEcYckwYOWcjbLhKE0YA0JJWMtCNOkO8y0nKHF5nfuidc0DC4laOIOD7qB9SjGEkldfVD20U8sX0kBxix5DJNz86nfpXcvkuPunrZNMwccJ5YE1KuGU8Znutd/dvuhTj2SSAxQ0J0LDOSToK4pJiRtGTHgkQID1GfdBQNkE+Ek+QHpfBAKT3ohVy9QMJcne5IkC/EyHdVZbarmPUy8T+vE0vv1EloEMWSBHg8yIsZlCHM0oE9ygmWbKQIwpyqXSEeII6wVBmWVAjW7MnzpFk1rGOjeqPSuAZjFMEe2AcVYIETUANXoA4aAIMH8ALewLv2qL1qH9rnuLSgTXp2wR9oX7+jGKFf</latexit>
<latexit sha1_base64="pBXXwACLK6H5fuDMnYTGhDxQh0U=">AAACEnicbVC7SgNBFJ2Nrxhfq5Y2i0HQImF3LRRBCNhYiETIk2SzzE4myZDZBzOzQlj2G2wsrPwGGxsLRWyt7EQ/xtlsCk08MHMP59zLzD1OQAkXuv6pZObmFxaXssu5ldW19Q11c6vG/ZAhXEU+9VnDgRxT4uGqIILiRsAwdB2K687wLPHr15hx4nsVMQqw5cK+R3oEQSElWz1oO5BFzfjktGlHhhEXkmLG8jbNuBMVjDgVOhVbzetFfQxtlhgTki/pFw/3d99fZVv9aHd9FLrYE4hCzluGHggrgkwQRHGca4ccBxANYR+3JPWgi7kVjVeKtT2pdLWez+TxhDZWf09E0OV85Dqy04ViwKe9RPzPa4Wid2xFxAtCgT2UPtQLqSZ8LclH6xKGkaAjSSBiRP5VQwPIIBIyxZwMwZheeZbUzKJxWDSvZBqXIEUW7IBdsA8McARK4ByUQRUgcAMewTN4UW6VJ+VVeUtbM8pkZhv8gfL+A/2ToOI=</latexit>
1 2 1 2 <latexit sha1_base64="4WAKB3Kmwvgmbs2+9dGwb6crGL8=">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</latexit>
1
2 Y11,11 Y11,12
0 0
-
3 4 5
<latexit sha1_base64="JGfW6F7NswQvodspDfxFWzq46k8=">AAAB8HicbVDLSgMxFL1TX7W+qi7dBIvgQsqkLnThoiCIG6GibS3tUDJp2oYmmSHJCGXoV7hxoYhbP8edX+EvmD4W2nrgwuGce7n3njAW3Fjf//IyS8srq2vZ9dzG5tb2Tn53r2aiRFNWpZGI9ENIDBNcsarlVrCHWDMiQ8Hq4eBy7NcfmTY8Uvd2GLNAkp7iXU6JdVKj0U4xPsF41M4X/KI/AVokeEYKZXRz9z0Irirt/GerE9FEMmWpIMY0sR/bICXacirYKNdKDIsJHZAeazqqiGQmSCcHj9CRUzqoG2lXyqKJ+nsiJdKYoQxdpyS2b+a9sfif10xs9zxIuYoTyxSdLuomAtkIjb9HHa4ZtWLoCKGau1sR7RNNqHUZ5VwIeP7lRVIrFfFpsXTr0riAKbJwAIdwDBjOoAzXUIEqUJDwBC/w6mnv2Xvz3qetGW82sw9/4H38AGBZkms=</latexit> <latexit sha1_base64="giFh1h8/iTywGBf4eA9WYvRkaak=">AAAB8HicbVDLSgMxFL1TX7W+qi7dBIvgQsqkLnThoiCIG6GibS3tUDJp2oYmmSHJCGXoV7hxoYhbP8edX+EvmD4W2nrgwuGce7n3njAW3Fjf//IyS8srq2vZ9dzG5tb2Tn53r2aiRFNWpZGI9ENIDBNcsarlVrCHWDMiQ8Hq4eBy7NcfmTY8Uvd2GLNAkp7iXU6JdVKj0U4xPsGlUTtf8Iv+BGiR4BkplNHN3fcguKq085+tTkQTyZSlghjTxH5sg5Roy6lgo1wrMSwmdEB6rOmoIpKZIJ0cPEJHTumgbqRdKYsm6u+JlEhjhjJ0nZLYvpn3xuJ/XjOx3fMg5SpOLFN0uqibCGQjNP4edbhm1IqhI4Rq7m5FtE80odZllHMh4PmXF0mtVMSnxdKtS+MCpsjCARzCMWA4gzJcQwWqQEHCE7zAq6e9Z+/Ne5+2ZrzZzD78gffxA2Hekmw=</latexit>
4 5 4
5 Y11,22 X 0
6
<latexit sha1_base64="9/Ohc09eMLyIKwhghzZIwtDktc0=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgQcLuetCDh4AgXoSIJjEkS5idzCZDZmaXmVkhLPkKLx4U8ernePMr/AUnj4MmFjQUVd10d4UJZ9q47peTW1peWV3Lrxc2Nre2d4q7e3Udp4rQGol5rB5CrClnktYMM5w+JIpiEXLaCAeXY7/xSJVmsbw3w4QGAvckixjBxkrNZifzvBPfH3WKJbfsToAWiTcjpQq6ufseBFfVTvGz3Y1JKqg0hGOtW56bmCDDyjDC6ajQTjVNMBngHm1ZKrGgOsgmB4/QkVW6KIqVLWnQRP09kWGh9VCEtlNg09fz3lj8z2ulJjoPMiaT1FBJpouilCMTo/H3qMsUJYYPLcFEMXsrIn2sMDE2o4INwZt/eZHU/bJ3WvZvbRoXMEUeDuAQjsGDM6jANVShBgQEPMELvDrKeXbenPdpa86ZzezDHzgfP2Nkkm0=</latexit>
6 Y11,11 Y11,12 6 Ȳ22 diag{1T Ȳ22 }

=
<latexit sha1_base64="giFh1h8/iTywGBf4eA9WYvRkaak=">AAAB8HicbVDLSgMxFL1TX7W+qi7dBIvgQsqkLnThoiCIG6GibS3tUDJp2oYmmSHJCGXoV7hxoYhbP8edX+EvmD4W2nrgwuGce7n3njAW3Fjf//IyS8srq2vZ9dzG5tb2Tn53r2aiRFNWpZGI9ENIDBNcsarlVrCHWDMiQ8Hq4eBy7NcfmTY8Uvd2GLNAkp7iXU6JdVKj0U4xPsGlUTtf8Iv+BGiR4BkplNHN3fcguKq085+tTkQTyZSlghjTxH5sg5Roy6lgo1wrMSwmdEB6rOmoIpKZIJ0cPEJHTumgbqRdKYsm6u+JlEhjhjJ0nZLYvpn3xuJ/XjOx3fMg5SpOLFN0uqibCGQjNP4edbhm1IqhI4Rq7m5FtE80odZllHMh4PmXF0mtVMSnxdKtS+MCpsjCARzCMWA4gzJcQwWqQEHCE7zAq6e9Z+/Ne5+2ZrzZzD78gffxA2Hekmw=</latexit>
<latexit sha1_base64="JGfW6F7NswQvodspDfxFWzq46k8=">AAAB8HicbVDLSgMxFL1TX7W+qi7dBIvgQsqkLnThoiCIG6GibS3tUDJp2oYmmSHJCGXoV7hxoYhbP8edX+EvmD4W2nrgwuGce7n3njAW3Fjf//IyS8srq2vZ9dzG5tb2Tn53r2aiRFNWpZGI9ENIDBNcsarlVrCHWDMiQ8Hq4eBy7NcfmTY8Uvd2GLNAkp7iXU6JdVKj0U4xPsF41M4X/KI/AVokeEYKZXRz9z0Irirt/GerE9FEMmWpIMY0sR/bICXacirYKNdKDIsJHZAeazqqiGQmSCcHj9CRUzqoG2lXyqKJ+nsiJdKYoQxdpyS2b+a9sfif10xs9zxIuYoTyxSdLuomAtkIjb9HHa4ZtWLoCKGau1sR7RNNqHUZ5VwIeP7lRVIrFfFpsXTr0riAKbJwAIdwDBjOoAzXUIEqUJDwBC/w6mnv2Xvz3qetGW82sw9/4H38AGBZkms=</latexit>
<latexit sha1_base64="auibt/nZRv5uLNjYnLC5Xx3jBfA=">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</latexit>
3 3
T
Ȳ22 := Y11,22
<latexit sha1_base64="CprelPrS8SjE+7s0Ya+i/OGCnPE=">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</latexit>
Y12,21 X22,11 Y12,21
Fig. 4: Illustration of admittance matrices and their underlying graphs: (Y, G(Y )), (Ȳ , Ḡ),
(Ȳ1 , Ḡ1 ), and (Ȳ2 , Ḡ2 ).
In the rest of this subsection we focus on identifying the remaining submatrices Y11,22 , Y12,21 as
well as Y22 = (Y22,11 , Y22,12 , Y22,22 ) of Y . For this purpose we assume without loss of generality
that all measured nodes are boundary measured nodes, i.e., the rows and columns corresponding
to submatrices Y11,11 and Y11,12 as well as their contributions to the diagonal entries of Y11,22
have been removed from Y . Then
 
  Y11,22 Y12,21 0
Y11 Y12  
Y =   =:  . (19)

 Y22,11 Y22,12
Y22  
Y22,22

20
Our goal is to identify Y in (19) given its Kron-reduction:
T
Ȳ2 = Y11,22 − Y12,21 X22,11 Y12,21 .
Theorem 1.2 implies that the underlying Kron-reduce graph G(Ȳ ) is a disjoint collection of
maximal cliques Ci among boundary measured nodes. By hidden nodes in a maximal clique Ci
of the Kron-reduced graph Ḡ, we mean the nonempty set of hidden nodes in the original graph
G that are connected either to the measured nodes in Ci or other hidden nodes in Ci . A measured
node can be in multiple cliques Ci though Ci are edge-disjoint (Theorem 1.2).
Lemma 3. Suppose the admittance matrix Y satisfies Assumptions 2 and 3. A measured node
can connect to only one hidden node in any cliques Ci of which it is a member.
Proof: If a measured node connects to more than one hidden node in a maximal cliques Ci ,
then there exists a loop since there is a path between any two hidden nodes in Ci , a contradiction.
We further assume without loss of generality that G(Ȳ ) consists of a single clique; other-
wise, we can repeatedly apply Algorithm 3 below to each clique separately to determine the
corresponding submatrices and then combine them to obtain Y22 and Y12 .
Remark 3. With this further assumption, Lemma 3 guarantees that Y12 has exactly one nonzero
element in each row.
Recall that there are M (boundary) measured nodes, indexed by 1, . . . , M , so that Y11,22 in
(19) is M × M . Suppose there are Hb boundary hidden nodes, indexed by M + 1, . . . , M + Hb ,
and Hi := H − Hb internal hidden nodes, indexed by M + Hb + 1, . . . , M + H. Then Y22,11 in
(19) is Hb × Hb and Y22,22 is Hi × Hi . Suppose each measured node i ∈ {1, . . . , M } is connected
to the hidden node h(i) ∈ {M + 1, . . . , M + Hb } by a line with series admittance yih(i) . From
Remark 3 we know there is a unique h(i) for each i, but voltage and current measurements only
identify the identity of each measured node i, but not the hidden node h(i) it is connected to
(nor the values of H, Hb , Hi ). The next result suggests a method to identify all measured nodes
that are connected to the same boundary hidden node.
Proposition 3. Suppose the admittance matrix Y satisfies Assumptions 2 and 3. Two measured
nodes i and j are connected to the same hidden node if and only if the off-diagonal entries of

21
rows i and j of Ȳ2 are proportional, i.e., there exists γ(i, j) 6= 0 such that
Ȳ2 [i, k]
= γ(i, j), k 6= i, j, k = 1, . . . , M.
Ȳ2 [j, k]
Proof: Application of Theorem 1 to the admittance matrix Y in (19) implies that
Ȳ2 := diag{1T Y11,22 } − Y12,21 X22,11 Y12,21

T
. (20)
Remark 3 implies row i of Y12,21 can be written as −yih(i) uTh(i) where uj is the Hb -dimensional
column vector with “1” in its jth position and “0” elsewhere. Hence
 
T
−y1h(1) uh(1)
..
 h ih i
T
Y12,21 X22,11 Y12,21 =  X −y1h(1) uh(1) . . . −yM h(M ) uh(M ) .
 
.
 22,11


−yM h(M ) uTh(M )
T
Denote by βij the (i, j) entry of X22,11 . Then row i of Y12,21 X22,11 Y12,21 is
h i
yih(i) · βh(i)h(1) y1h(1) βh(i)h(2) y2h(2) · · · βh(i)h(M ) yM h(M ) .
Consider two measured nodes i, j ∈ {1, . . . , M }. If they are connected to the same hidden node,
T
then h(i) = h(j) and hence row i and row j of Y12,21 X22,11 Y12,21 are proportional:
T

Y12,21 X22,11 Y12,21 [i, k] yih(i)
T
= =: γ(i, j), ∀k 6= i, j.
Y12,21 X22,11 Y12,21 [j, k] yjh(i)
The necessity of the proposition then follows from (20).

Conversely, suppose
T

Y12,21 X22,11 Y12,21 [i, k]
T
= γ(i, j), ∀k 6= i, j
Y12,21 X22,11 Y12,21 [j, k]
for some γ(i, j) 6= 0. Then
yih(i) βh(i)h(k) ykh(k) yih(i) βh(i)h(k)

= = γ(i, j), ∀k 6= i, j
yjh(j) βh(j)h(k) ykh(k) yjh(j) βh(j)h(k)
and therefore
βh(i)h(k)
=: γ̂(i, j), ∀k 6= i, j
βh(j)h(k)
Next we shall show that (h(k), ∀k 6= i, j) span all columns of X22,11 , which is equivalent to

22
show

βh(i)h(j) βh(i)h(i) βh(i)k
, ∈ , k = 1, . . . , Hb .
βh(j)h(j) βh(j)h(i) βh(j)k
We prove the sufficiency by contradiction and consider the following three scenarios:
• If every hidden boundary node connects to at least two measured nodes. Therefore, there
exist two indices i1 , j1 such that h(i1 ) = h(i) and h(j1 ) = h(j). We can let k = i1 and
k = j2 respectively and conclude the proof by combining the definition of hidden boundary
nodes. This contradicts the invertibility of X22,11 .
• Secondly, we consider the case that h(i) connects to exact one measured node i and h(j)
connects to at least two measured nodes. Without loss of generality, assume that h(j)
connects to two measured nodes j and j1 . Following the derivation above and noticing
h(j) = h(j1 ), we have
βh(i)h(k)
=: γ̂(i, j), ∀k 6= i.
βh(j)h(k)
−1 −1 T
Note that X22,11 = Y22,11 − Y22,12 Y22,22 Y22,12 . Consider the (h(i) − M )th row of matrix
−1
X22,11 , all the elements should be 0 excepting the edge between h(i) and h(j). This can be
shown by computing the adjugate of X22,11 . This leads to h(i) only connects h(j) directly
or through a path, implying h(i) can maximally connect to one hidden node. On the other
side, h(i) only connects to one measured node and therefore this violates Assumption 3.2.
• Finally, if both node h(i) and h(j) only connect to one measured node. We can rearrange the
matrix X22,11 by combining the h(i)th and h(j)th rows and the h(i)th and h(j)th columns
in a submatrix as
 
βh(i)h(i) βh(i)h(j) βh(i)1 ··· βh(i)Hb −2
 
 β
 h(j)h(i) βh(j)h(j) βh(j)1 ··· βh(j)Hb −2 
 
A A12
 11
 
X22,11 = 
 β1,h(i) β1,h(j) β11 ··· β1,Hb −2   =:
.
. . .. .. .. A21 A 22
.. ..
 

 . . . 

βHb −2,h(i) βHb −2,h(j) βHb −2,1 ··· βHb −2,Hb −2
A11 must be invertible, otherwise it will contradict the invertibility of X22,11 . Note that
−1
rank(A12 ) = 1. Consider the submatrix (X22,11 )12 = A−1 −1 −1
11 A12 (A22 − A21 A11 A12 ) , its
−1
rank is not more than 1. If rank((X22,11 )12 ) = 0, this means h(i) only connect to one

23
hidden node h(j). Also, h(i) only connects to one measured node and therefore this violates
−1
Assumption 3.2. If rank((X22,11 )12 ) = 1, then h(i) and h(j) connect at least two the same
hidden nodes. Yet it is not hard to see that this connected graph, in this case, must have a
loop, therefore it violates the tree Assumption 3.1.
Combining all above cases, we therefore show that, by contradiction, the sufficiency holds.
Note that if there are only M = 3 measured nodes then Assumptions 3.1 and 3.2 imply that
all of them must be connected to the same boundary hidden node.
Given the Kron-reduced admittance matrix Ȳ2 , Proposition 3 allows us to group together the
(boundary) measured nodes that are connected to the same (boundary) hidden node. This also
identifies the number of boundary hidden nodes, even though we do not know (yet) the number
or identity of internal hidden nodes nor the connectivity among the nodes. We can re-arrange
the submatrix matrix Y12,21 into a form easier for identification.
Specifically let measured nodes 1, . . . , k1 be connected to hidden node M +1, measured nodes
k1 +1, . . . , k2 to hidden node M +2, . . . , measured nodes kHb −1 +1, . . . , kHb := M to hidden node
M +Hb . Note that Proposition 3 yields the values for Hb and (k1 , k2 , . . . , kHb = M ) even though
it provides no information about the value of H, the total number of hidden nodes. To simplify
h i
notation, denote the series admittance yih(i) of line (i, h(i)) by yi . Then Y12 = Y12,21 0 where
Y12,21 is M × Hb and can be arranged into the following simple form:
 
y1 0 ··· 0
 . .. .. .. 
 .. . . . 
 
   
y
 k1 0 ··· 0   ŷ1 0 ··· 0
   
k1 +1 · · · ···
0 y 0   0 ŷ 0 
2
Y12,21 =  . =: .
   
 .. .
.. .
.. .
.. 
 . .
 .. .. .. ..
   . . 

 
0
 yk 2 · · · 0   0 0 ··· ŷHb
 . .. .. .. 
 ..

 . . . 
0 0 · · · ykHb
where for i = 1, . . . , Hb , ŷi is a (ki −ki−1 )-dimensional column vector corresponding to ki −ki−1
measured nodes that are connected to the hidden node M + i. Since Y has zero row sum by
Assumption 3.3, the diagonal matrix diag{1T Y11 } = diag{1T Y11,22 } = diag(yi = yih(i) , i =

24
1, . . . , M ). We have
Y12 Y22−1 Y12T = diag(ŷj , j = 1, . . . , Hb ) X22,11 diag(ŷjT , j = 1, . . . , Hb )

 
β11 ŷ1 ŷ1T β12 ŷ1 ŷ2T ··· β1Hb ŷ1 ŷH T
b
 
 β ŷ ŷ T β ŷ ŷ T
· · · β ŷ ŷ T 
 21 2 1 22 2 2 2Hb 2 Hb 
=  .. .. .. .

 . . · · · . 

βHb 1 ŷHb ŷ1T βHb 2 ŷHb ŷ2T · · · βHb Hb ŷHb ŷH T
b
Then the admittance matrix corresponding to the graph Ḡ2 in Theorem 1 is:
   
y1 0 · · · 0 β11 ŷ1 ŷ1T β12 ŷ1 ŷ2T · · · β1Hb ŷ1 ŷHT
b
   
T T
 y 2 · · · 0   β ŷ
22 2 2 ŷ · · · β ŷ ŷ
2Hb 2 Hb 

Ȳ2 =  − . (21)
  
.. ..   ... ..

 . . 


 . 

T
yM βHb Hb ŷHb ŷH b
Recall that we have already identified the Kron-reduced admittance matrix Ȳ2 , i.e., we know every
entry of Ȳ2 on the left-hand side of (21). We now explain how to identify (βij , i, j = 1, . . . , Hb )
and (yi = yih(i) , i = 1, . . . , M ) on the right-hand side of (21). In particular, (yi = yih(i) , i =
1, . . . , M ) yields Y12 of the original admittance matrix Y .
Identification of (yi = yih(i) , i = 1, . . . , M ). Let Ȳ2,k1 be the diagonal submatrix consisting of
the first k1 rows and columns of Ȳ2 corresponding to the first k1 measured nodes connected to
the first hidden node M + 1:
 
y1 · · · 0
.. 
 
.. T
Ȳ2,k1 :=  . .  − β11 ŷ1 ŷ1

 
yk1
   
y1 · · · 0 y
 1 h
..   . 
  i
..
=  . .  − β11 ·  ..  y1 · · · yk1 . (22)

   
yk1 yk1
We first explain how to identify (β11 , ŷ1 ) on the right-hand side of (22) from the knowledge
of Ȳ2,k1 on the left-hand side of (22). The identification of other bii , ŷi corresponding to ki −
ki−1 measured nodes connected to the hidden node M + i from the diagonal blocks Ȳ2,ki :=
diag(yki−1 −1 , . . . , yki ) − βii ŷi ŷiT can be done similarly. Then we will explain how to identify
non-diagonal blocks of Ȳ2 .

25
1) Case 1: kj ≥ 2. In this case, hidden node M + 1 is connected to two or more measured

nodes indexed by i = 1, . . . , k1 (without loss of generality, we consider a special case
when j = 1). Consider the first two measured nodes and the corresponding 2 × 2 principal
submatrix of Y2,k1 : for i, j = 1, 2

 y − β y2 if i=j
i 11 i
Ȳ2,k1 [i, j] = (23)
 −β11 yi yj if k 6= j
This leads to the following equations in (b11 , y1 , y2 ):
y1 − β11 y12 = Ȳ2,k1 [1, 1] =: a1
−β11 y1 y2 = Ȳ2,k1 [1, 2] =: a2
y2 − β11 y22 = Ȳ2,k1 [2, 2] =: a3
yielding:
a1 a3 − a22 a1 a3 − a22 a2 (a1 + a2 )(a2 + a3 )

y1 = , y2 = , β11 = − 2 (24)
a2 + a3 a1 + a2 (a1 a3 − a22 )
To identify other (yj , j > 2), note that
−β11 y1 yj = Y2,k1 [1, j], j = 3, . . . , k1
yielding
Y2,k1 [1, j]
yj = −
β11 y1
where b11 and y1 are given by (24).

2) Case 2: Once we have recover the coefficients for hidden boundary nodes with at least
two connections to measured nodes in case 1, next, we can treat these recovered hidden
nodes as measured nodes and repeat the above procedure until no hidden node is left. A
key step is to construct a new Kron reduced matrix once parts of the admittance matrix
has been found. Let the original Y have the following partition (note that this partition is

26
different from the partition in )

 
Y11,11 0 Y12,11 0
 
 0 Y 0 Y 
11,22 12,22 
Y = , (25)

Y T 0 Y 22,11 Y 
22,12 
 12,11
T T
0 Y12,22 Y22,12 Y22,22
where Y11,11 represents admittance between a subset of measured nodes that are connected
to hidden nodes with at least 2 connections to measured nodes, Y11,22 represents admittance
between the remaining measured nodes. The Kron reduced admittance matrix
    −1  
T
Y11,11 0 Y 0 Y Y Y 0
Ȳold =   −  12,11   22,11 22,12   12,11 . (26)
T T
0 Y11,22 0 Y12,22 Y22,12 Y22,22 0 Y12,22
Based on the results in Step 1, one can recover Y11,11 , Y12,11 and X22,11 . We aim to expand
the dimension by including these nodes as measured nodes.
   
Y
 11,11 0 Y 12,11
  0  h i
 −1
Ȳnew =  0 Y11,22 0  − Y12,22  Y22,22 0 Y12,22
T T
Y22,12 . (27)
  
   
T
Y12,11 0 Y22,11 Y22,12
T T T
Note that Y22,22 Y22,12 = −X22,12 X22,11 from matrix inversion lemma. The new blocks of
Ȳnew can be recovered as follows:
−1 T −1
Ȳnew,33 = Y22,11 − Y22,12 Y22,22 Y22,12 = X22,11 ,
−1 T T −1 † −1
Ȳnew,23 = Y12,22 Y22,22 Y12,22 = −Y24 X12 X22,11 = (Y12,11 Ȳold,12 )T X22,11 ,
(28)
−1 −1
Ȳnew,22 = Ȳold,22 + Y12,22 Y22,22 T
Y22,12 (Y22,11 − Y22,12 Y22,22 T
Y22,12 )−1 Y22,12 Y22,22
−1 T
Y12,22
−1 T
= Ȳnew,23 X22,11 Ȳnew,23 .
Therefore, a new Kron reduced admittance Ȳ can be constructed.

3) For any hidden node that connects to one or zero measured node, these hidden nodes will
eventually have more than one connections to measured nodes once the other hidden nodes
and therefore can be recovered. It is easy to show that there will never exist a scenario that
all the hidden nodes have at most 1 connection to measured nodes for a tree graph. To see
this, we have a first observation: for any clique, H ≥ M as every hidden node connects to
a different measured node. On one hand, the sum of all hidden nodes’ degrees is greater

27
than 3H under the Assumption 3. On the other hand, it is at most 2(H − 1) + M , which is
twice the sum of all edges between hidden nodes and the number of connections between
hidden nodes and measured nodes. However, 2(H − 1) + M < 3H, contradiction!
−1
4) If all hidden nodes are hidden boundary nodes, i.e., Y22 = Y22,11 , then Y22 = X22,11 and
hence the entire admittance matrix Y can be identified. If there are hidden nodes that are
not hidden boundary nodes, we can treat hidden boundary nodes as measured nodes now
and repeat the above procedure based on Woodbury formula.
Combine the above results, it guarantees the recovery of all hidden nodes. After repeating
this process and obtaining Y12 and Y22 , we can recover diag{1T Y11 } and consequently Y11 from
eq. (17a):
Y11 = Ȳ1 + diag{1T Y11 }. (29)
Algorithm 3 Recover Y from Ȳ

1: Input: Ȳ1 and Ȳ2
2: for any pair of nodes (j, k) do
3: Compute γ[j, k] from Ȳ2 .
4: end for
−1
5: Solve for Ŷ12 ,X22 from (23) and (28) and set Ŷ22 = X22
6: Solve fordiag{1T Y11 } = Ŷ12 X22 Ŷ12T + YG2 and Ŷ11 according to (29)
Ŷ Ŷ12
7: Set Ŷ = 11 T
Ŷ12 Ŷ22
8: if the graph corresponding to Ŷ22 contains a loop then
9: Repeat the above process and treat Ŷ22 as Ȳ
−1 T
10: Obtain Ŷ22,11 , Ŷ22,12 , Ŷ22,22 such that Ŷ22 = Ŷ22,11 − Ŷ22,12 Ŷ22,22 Ŷ22,12 .
 
Ŷ11 Ŷ12 0
11: Set Ŷ = Ŷ21 Ŷ22,11 Ŷ22,12 

T
0 Ŷ22,12 Ŷ22,22
12: Set Ŷ22 = Ŷ22,22
13: end if
14: return Y = Ŷ
2) Assumption relaxation: Nonzero shunt element: Next, we relax one of the assumption on
zero shunt element. Consider a graph satisfies only 3.1 and 3.2 but not 3.3 in Assumption 3
(i.e., there exist nodes with nonzero shunt elements), from Ȳ2 in eq. (21), we outline the key
differences below. The procedure is unchanged for using γ detecting hidden nodes as well as
recognizing how many measured nodes that this hidden node connects to. We consider the
following scenarios:

28
1. For any hidden node that connects to two or more measured nodes, given Ȳ2 , (dk , yk ),
k = 1, 2 can be identified using the following computation: for k = 1, 2 and l = 1, . . . , M

 −β y y if k 6= l
11 k l
Ȳ2 [k, l] = (30)
 dk − β11 y 2 if k = l
k
It is easy to see that d1 and d2 can be recovered. Yet, without the equality d1 = y1 and
d2 = y2 (due to zero shunt element assumption), we can not solve the following equations
as they are linearly dependent:
β11 y12 = a1 , β11 y1 y2 = a2 , β11 y22 = a3 .
Yet we can set y1 = 1 and derive expressions for β11 and y2 . The identified parameters
β11 y2
yield: β̂11 = y12
, ŷ2 = y1
.
2. For any hidden node that connects to one or zero measured node, these hidden nodes will
eventually have more than one connections to measured nodes once the other hidden nodes
become measured.
These scenarios in the zero shunt element case are still valid provided that we set an element to
1 (as a reference) in every ŷi . As a result, the identified parameters differs from the true parameters
by a diagonal scaling matrix: X̂ = ΓX Γ, Ŷ = Y Γ, Ŷ
11 11 13 =Y
13 Γ, Ȳˆ
new12,11 = Ȳ ,
new12,11 new22 new22
where
Γ , diag{yˆ1 [1], 1, . . . , 1, yˆ2 [1], 1, . . . , 1, . . . , yˆs [1], 1, . . . , 1}.
| {z } | {z } | {z }
k1 k2 ks
Next, we look into how these scaling affect the new blocks of Ȳˆnew : Ȳˆnew,11 = Ȳnew,11 , Ȳˆnew,12 =
Ȳnew,12 = 0, Ȳˆnew,33 = X̂11
−1 −1
= ΓX11 Γ. The new Ȳˆnew has the following form comparing it with
Ȳnew in the zero shunt element case:
 
 Ȳ11 0 Ȳ13 Γ 
ˆ
Ȳnew =  0

Ȳ22 Ȳ23 Γ  .

(31)
 
ΓȲ13T ΓȲ23T ΓȲ33 Γ
A new Kron reduced admittance Ȳˆ can be recovered. We can then repeat the procedure by
setting new reference elements to identify non-hidden boundary nodes (if there is any). As the
scaling does not change the topology, it is not hard to show that the identified topology should

29
be the same as the original graph, even though we can not get the right parameters and nor the
diagonal element of Y11 , i.e., diag{1T Y11 }.
C. Mesh Networks: Low-rank and Sparse Decomposition
For general (possibly mesh) networks, we propose algebraic solutions to such a problem. The
first step is similar to the case without hidden nodes: Ȳ is computed from successive current
and voltage measurements. When there exist hidden nodes, inferring Y from Ȳ is an ill-posed
problem since there often exist more variables than equality constraints. We therefore need to
resort to additional prior knowledge about the original system to constrain the solution space.
Interestingly, some prior knowledge about Ȳ is available: first, Y11 corresponds to a sparse
graph and is therefore sparse; second, Y12 Y22−1 Y12T has low rank (it has a maximal rank equal to
the number of hidden states h). We consider the following optimization problem in which Ȳ is
decomposed into a sparse matrix A and a low-rank matrix B:
h i
−1 T = arg min kAk + λrank (B)
Y11 Y12 Y22 Y12 0
A,B
(32)
s.t. A − B = Ȳ .
Here λ ∈ R+ is a weight that balances the sparsity of Y11 and the number of hidden states.
Without the knowledge of the true admittance matrix, it is impossible to have any identifiability
property as discussed in [8]. In what follows, we propose a computationally efficient algorithm
(convex relaxation) which relaxes (32). We then study how good such a relaxation is.
We can relax the optimization problem (32) to the following form
min kAk1 + λkBk∗

(33)
s.t. A − B = Ȳ .
We use `1 optimization as a convex relaxation of `0 optimization, and nuclear norm kBk∗ ,

P p
i λi (B H B) as a convex relaxation of the rank optimization.
The optimization problem in eq. (33) can be recast as a semidefinite program (SDP). To solve
it, the first step is to use the fact that the spectral norm k · k is the dual norm of the nuclear
norm k · k∗ [8]: kW0 k∗ = max{trace(W0T Y )| kY k ≤ 1}. Furthermore, the spectral norm admits

30
a simple semidefinite characterization:

 
tIn Y
kY k = min t s.t.   0.
t T
Y tIn
From duality, we can obtain the following SDP characterization of the nuclear norm and prob-
lem (33) can be rewritten as
λ
min kAk1 + (trace(W1 ) + trace(W2 ))
A,B,W1 ,W2 2
 
(34)
W B
s.t.  1  0, A − B = Ȳ .
B T W2
Lemma 4. Given a complex matrix W0 and consider a constructed W̃0

 
Re(W0 ) −Im(W0 )
W̃0 =  , (35)
Im(W0 ) Re(W0 )
we have the following properties: 1) rank(W̃0 ) = 2rank(W0 ) and 2) kW̃0 k∗ = 2kW0 k∗ .
Proof: Easy to show.

Based on the above lemma, the optimization problem (34) can be converted to the following
convex optimization:
λ
min 1Tn Z1n + (trace(W̃1 ) + trace(W̃2 ))
A,B,W̃1 ,W̃2 ,Z 2
 
W̃1 W̃0
s.t.  0
W̃0T W̃2
− Z[i, j] ≤ A[i, j] ≤ Z[i, j], ∀(i, j) (36)

 
Re(Ȳ ) −Im(Ȳ )
A−B = 
Im(Ȳ ) Re(Ȳ )
A, B has the form expressed in eq. (35).
Remark 4. The last constraint on A, B can be easily turned into linear constraints on their
coefficients.
Once we obtain A, B from this optimization, we might be able to reconstruct the complex

31
matrices Y11 and Y12 Y22−1 Y21 . However, there is no guarantee for general networks and admittance
matrices [8].
IV. S IMULATIONS
A. Simulated data
In this section we implement the proposed algorithms in MATLAB and evaluate their iden-
tification accuracy by performing simulations in MATPOWER [9]. The optimization problems
are solved using the CVX toolbox [10]. We run power flow analysis on the IEEE 14-bus test
system, representing a portion of a power system in the Midwestern U.S., which has 14 buses, 11
aggregated loads, and 5 generators, 3 of which are synchronous compensators used for reactive
power support [11]. To validate our identification algorithms in three-phase distribution systems,
we have also performed simulations on IEEE test distribution networks, which are not presented
here (see our previous work [12] for the identification results in distribution systems).
We assume that PMUs are installed at selected buses and that they can precisely measure the
voltage and current magnitudes and phase angles that we obtain from power flow calculations,
unless stated otherwise. For each scenario, we run 100 steady state simulations, each pertaining
to a time slot, to determine the voltage and current magnitude and phase angle of every bus, while
varying the real and reactive power demand of the loads across the time slots. Specifically, for
a given time slot, the real and reactive power consumption of a constant PQ load are computed
by multiplying a scaling factor drawn from a uniform distribution over the interval [0.8, 1.2]
by the real and reactive power consumption data provided in [11]. We obtain the admittance
matrix of this system using a built-in function of the power flow simulator. It turns out that the
absolute values of nonzero complex elements of the admittance matrix are between 1.86 and
40.06, reminding the readers that a complex number’s absolute value is its distance from zero
in the complex plane.
We first consider the scenario that every bus is equipped with a PMU. Assuming that the self
admittance of bus 7, i.e., the transformer bus, is known, Fig. 5 shows the identification error,
defined as |Y − Ŷ |, and the vertical color bar indicates the mapping of data values into colors.
Hence, the color of a cell located at row i and column j represents the value of |Y [i, j]− Ŷ [i, j]|.
It can be seen that the identification error using measurements of 15 time slots (K = 15) is
quite small compared to the absolute value of the elements of Y , and this error does not vary
much if we use more observations. We next analyze the sensitivity of the proposed algorithm

32
#10 -5
14 8
13
12 7
11
6
10
bus number 9 5
8
7 4
6
3
5
4 2
3
2 1
1
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14
bus number
Fig. 5: The identification error when there is no hidden state. The color of a cell located at row
i and column j represents the value of |Y [i, j] − Ŷ [i, j]|. It can be seen that the identification
error using measurements of 15 time slots (K = 15) is quite small compared to the absolute
value of the elements of Y .
14 0.015
13
12
11
10
0.01
9
bus number
5 0.005
1
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14
bus number
Fig. 6: The identification error when white Gaussian noise is added to both complex voltage and
current measurements. Similar to the previous case, the errors are sufficiently small.

33
to the measurement error which is typically introduced by the transducers. To this end, white
Gaussian noise with a signal-to-noise ratio of 125 is added to both complex voltage and current
measurements. The signal-to-noise ratio is chosen such that the measurement accuracy lies within
the reported range for existing PMU technology. Fig. 6 shows the absolute identification error for
this case. Similar to the previous case, the errors are sufficiently small. In general, we observe
that the identification error increases as we decrease the signal-to-noise ratio and it becomes
really large when the signal-to-noise ratio drops below 100; at this point we say that the Y
matrix cannot be identified from data.
B. Experimental data
Next, we test the proposed IPF algorithm on the IEEE 5-bus benchmark shown in the left panel
of Fig. 7. The network contains 4 dynamic loads, 2 generators and 7 transmission lines. The
system admittance matrix is obtained from [13] with the absolute values of nonzero complex
elements of the admittance matrix are between 3.80 and 31.32. Simulation is performed on
OPAL-RT real time simulator [14] in our lab shown in the right panel of Fig. 7. The generator is
represented by a sixth order model with turbine governors and excitation systems. Assume that the
sampling frequency of PMU is 20 Hz and the dynamic loads change over time. We collect PMU
data (nodal voltages and injected currents) of all buses for 40 seconds. Note that the frequency
fluctuates with changes of the load demands, which is consistent with field measurements; PMU
has 0.49% average total vector error in simulation. Fig. 8 shows the identification error, defined
as |Y − Ŷ | (element-wise absolute value), and the vertical color bar indicates the mapping of
data values into colors. One can see that the IPF is capable of identifying the topology correctly
and estimating system parameters with small error from noisy data with the maximum parameter
estimation error 0.2126.
C. Hidden node case
Given a graph shown in Fig. 9 (left), if sensors are deployed at nodes {1, 2, 6, 7, 8, 9, 12}, we
can use Kron reduction to obtain the graph shown in Fig. 9 (right).
In this example the hidden nodes are M2 = {3, 4, 5, 10, 11}, from which nodes {3, 5, 10, 11}
have degree less than 3, and therefore, cannot be identified from data. We now illustrate how
the proposed algorithm can identify the actual admittance matrix including Node 4.

34
Fig. 7: Left: The network topology of the 5-bus benchmark systems. Right: The OPAL-RT
real-time simulation platform that are used for data generation.
0.2
5
0.18
0.16
4
0.14
Bus Number
0.12
3
0.1
0.08
2 0.06
0.04
1 0.02
0
1 2 3 4 5
Bus Number
Fig. 8: The identified error of admittance matrix along buses for the 5-bus benchmark example
we tried.
The first step is to decompose the graph corresponding to the estimated Ȳ matrix to two graphs
using Algorithm 2, one is a collection of cliques and the other one is a tree and some isolated
nodes. The second step is to apply Algorithm 3 to identify the original admittance matrices for
all cliques. The final step is to take the union of the cliques and the graph obtained in the second
step. These steps are shown in Fig. 10.
Indeed, even the original graph does not satisfy the assumption on the degree of hidden nodes.

35
Fig. 9: An example of a 12-bus network. Left: The original network topology. Green circles
denote the observed nodes while yellow ones denote the hidden nodes. Right: The topology of
the condensed graph obtained from Kron reduction. The red lines denote the edges in G1 and
the grey ones denote the ones in G2 .
Decomposition
Step 1: +
Clique Tree
Estimate # of
hidden nodes
Step 2:
Composition
Step 3: +
Fig. 10: A step-by-step illustration of the proposed algorithm. Step 1 is described in Algorithm 2
that one separates the graph to a number of cliques and a forest. Step 2 is described in Algorithm 3
that one recovers the admittance submatrix from every clique. Step 3 combines all the recovered
graphs (together with admittance submatrices) to recover the whole admittance matrix.
We can recover a graph with smallest number of nodes that is a) consistent with Ȳ and b) satisfies
Assumption 3.

36
7 8
2 3 4 5
Fig. 11: An example of a 8-bus network, where 5 nodes are measured, and there exists three
hidden nodes.
Consider another graph with the following topology in Fig. 11, the admittance matrix has the
following form5
 
Y 0 0 0 0 −Y16 0 0
 11 

 Y22 0 0 0 0 −Y27 0  
 

 Y33 0 0 0 −Y37 0  
 
 Y 44 0 0 0 −Y 
48 
Y = .


 Y 55 0 0 −Y 
58 
 

 Y 66 −Y 67 −Y 68


 

 Y 77 0 

Y88
Through direct computation, we have

 
Y11 − X11 Y162 −X12 Y16 Y27 −X12 Y16 Y37 −X13 Y16 Y48 −X13 Y16 Y58
 

 Y22 − X22 Y272 −X22 Y27 Y37 −X23 Y27 Y48 −X23 Y27 Y58 
 
Ȳ = 
 Y33 − X22 Y372 −X22 Y37 Y48 −X23 Y37 Y58 .
 

 Y44 − X33 Y482 −X33 Y48 Y58 
2
Y55 − X33 Y58
For zero shunt element case, we follow the procedure in Algorithm 3. First of all, we compute
5
for notational simplicity, we used Yij to denote the Y [i, j] element in this section.

37
the γ  
∗ 0 0 0 0
 
 ∗ 1 0 0
 
 
γ=
 ∗ 0 0.
 

 ∗ 1
∗
It means that node 2 and 3 are connected to a same hidden node, and node 4 and 5 are connected
to a same hidden node. It is therefore straightforward to compute Y22 , Y33 , Y44 , Y55 , X22 , X23 , X33
from Ȳ and furthermore Y27 = −Y22 , Y37 = −Y33 , Y48 = −Y44 , Y58 = −Y55 . Next we obtain a
“new” Ȳ by including two new hidden boundary nodes as measured nodes.
  −1 
h i X22 , X23
Y 0 X12 Y16 X13 Y16   
 new,11 

 X23 , X33 

   

 −Y 27 0 

 
 −Y 0  
  37 
Ȳnew = 
 diag{Y22 , Y33 , Y44 , Y55 }   ,
0 −Y48 
  
  
 
0 −Y58
 
 
  −1 
 
 X22 , X23 
   
X23 , X33
h i
where X12 Y16 X13 Y16 can be computed from the corresponding row of Ȳ and
 −1  
h i X22 , X23 X Y
Ȳnew,11 = Y11 − X11 Y162 + X12 Y16 X13 Y16    12 16  .
X23 , X33 X13 Y16
It is noted that nodes 2, 3, 4, 5 have no connections to non hidden boundary nodes, then they
can be removed.
 h i 
Y11 − Y66−1 Y162 Y66−1 Y16 Y67 Y68 .
     
Ȳnew =
 
Y77 0 −1
Y67 h i 
   − Y66   Y67 Y68 .
0 Y88 Y68
The last equality due to matrix inverse lemma. We can repeat the procedure, by noting that the
size of Ȳnew is three, and furthermore solve for Y11 , Y77 , Y88 , Y66 , Y67 , Y68 , Y16 , therefore recover

38
the original admittance matrix Y .

For nonzero shunt element case, γ yields the same form. It means that node 2 and 3 are
connected to a same hidden node, and node 4 and 5 are connected to a same hidden node.
It is therefore straightforward to compute Y22 , Y33 , Y44 , Y55 from Ȳ . However, Y27 , Y37 , Y48 , Y58
can not be inferred. Indeed, we have to parameterize Ŷ27 = 1 and Ŷ48 = 1 and compute
Ŷ37 = Y37 /Y27 , Ŷ58 = Y58 /Y48 , X̂22 = X22 Y372 , X̂23 = X23 Y37 Y58 , X̂33 = X33 Y582 .
Next we obtain a “new” Ȳ by including two new hidden boundary nodes as measured nodes.
  −1 
h i X̂22 , X̂23
Y 0 X̂12 Y16 X̂13 Y16   
 new,11 

 X̂23 , X̂33 

   

 −1 0 

 
 −Ŷ 0   
  37 
Ȳnew = 
 diag{Y22 , Y33 , Y44 , Y55 }   ,
 0 −1 
  
 
 
0 −Ŷ58
 
 
  −1 
 
X̂ , X̂
 22 23 
 
 
X̂23 , X̂33
h i
where X̂12 Y16 X̂13 Y16 can be computed from the corresponding row of Ȳ and
 −1  
h i X̂22 , X̂23 X̂ Y
Ȳnew,11 = Y11 − X11 Y162 + X̂12 Y16 X̂13 Y16    12 16  .
X̂23 , X̂33 X̂13 Y16
It is noted that nodes 2, 3, 4, 5 have no connections to non hidden boundary nodes, then they
can be removed.
 h i 
Y11 − Y66−1 Y162 Y66−1 Y16 Y67 Y68 Γ−1
      
Ȳnew = ,
 
−1
Y77 0 −1
Y67 h i
−1
 Γ   − Y66   Y67 Y68  Γ 
0 Y88 Y68
where Γ , diag{Y27 , Y48 }. We compute
Y77 Y88 Y67 Y68

Ŷ77 = 2
, Ŷ88 = 2 , Ŷ67 = , Ŷ68 = , Ŷ11 = Y11 , Ŷ16 = Y16 , Ŷ66 = Y66 .
Y37 Y58 Y37 Y58
This indicates that we can recover the true topology.

39
V. R ELATED W ORK
The availability of high-precision, high-sample-rate measurements of transmission and distri-

bution networks in recent years has given impetus to research on topology identification, state
estimation, and fault detection.
The IPF problem that identifies both topology and admittance matrix has been studied ex-
tensively in transmission networks [15], [16], [17] as well as distribution grids [18], [19] using
single-phase a.c. and d.c. power flow models. For example, the topology identification problem
is cast as a sparse subspace learning problem in [15] and an efficient algorithm is proposed
to estimate the admittance matrix of the underlying power system from the measured power
injection of different buses. In [20], the topology of an urban (mesh) distribution network is
inferred from voltage magnitude, and real/reactive power measurements carried out by smart
meters at the end-nodes. A graphical model is built to describe the probabilistic relationships
between different voltage measurements using lasso. In [18] a graphical learning based approach
is developed to estimate the radial grid topology from nodal voltage measurements. The learning
algorithm is based on conditional independence tests for continuous variables over chordal
graphs. An efficient algorithm for topology identification of a power system is also proposed
in [17] drawing on ideas from compressive sensing and graph theory. The authors assume that
power and phase angle measurements are available for all nodes.
Different algorithms have been developed in the literature to make this inference, using
various techniques such as weighted least square, maximum-likelihood/maximum-a-posteriori
estimation, minimum spanning tree, sparse recovery, Lasso/Group Lasso, blind identification,
quickest change detection theory, as well as graphical model learning. There are three limitations
in the current literature that we propose overcome. First, most of the literature focuses on
topology identification or change detection, but there is not much work on joint topology and
parameter identification, with notable exceptions of [15], [21], [22]. Second, most papers require
measurements at every node in the network, with the exceptions of [23], [24], [22].
VI. C ONCLUSIONS
This paper lays out the foundation of the inverse power flow problem which concerns inferring
the admittance matrix of a power system from synchronized measurements of voltage and current
obtained from a subset of its buses. The algorithms proposed in this work are efficient, robust
to noise, and can jointly address state estimation and topology identification problems with

40
theoretical guarantees, if certain conditions are met. These findings are supported by high-fidelity
power flow simulations performed on standard test systems.
The plausibility of our results underlines that much value can be extracted from synchrophasor
data, especially in distribution systems which are not typically monitored beyond the substation.
In companion works [6], [25], we apply the results in this paper to the topology identification
and event detection in distribution networks. In future work, we plan to extend our framework to
three phase power flow models, which takes the mutual coupling between phases into account,
develop efficient algorithms for identifying the admittance matrix of radial distribution systems
with few measurement nodes, and analyze the sensitivity of the identification results to non-
stationary measurement errors.
VII. ACKNOWLEDGEMENT
We wish to thank Mr. Wei Zhou (Huazhong University of Science and Technology) for
discussion.
R EFERENCES
[1] O. Ardakanian, S. Keshav, and C. Rosenberg, Integration of Renewable Generation and Elastic Loads into Distribution
Grids, ser. SpringerBriefs in Electrical and Computer Engineering. Springer International Publishing, 2016.
[2] NASPI, “North American Synchrophasor Initiative,” https://www.naspi.org.
[3] S. H. Low, “Convex relaxation of optimal power flow?part i: Formulations and equivalence,” IEEE Transactions on Control
of Network Systems, vol. 1, no. 1, pp. 15–27, 2014.
[4] F. Dorfler and F. Bullo, “Kron Reduction of Graphs With Applications to Electrical Networks,” IEEE Transactions on
Circuits and Systems I: Regular Papers, vol. 60, no. 1, pp. 150–163, Jan 2013.
[5] M. Baran and F. F. Wu, “Optimal Sizing of Capacitors Placed on a Radial Distribution System,” IEEE Transactions on
Power Delivery, vol. 4, no. 1, pp. 735–743, Jan 1989.
[6] O. Ardakanian, Y. Yuan, V. Wong, R. Dobbe, S. Low, A. von Meier, and C. J. Tomlin, “On identification of distribution
grids,” IEEE Transactions on Control of Network Systems, 2019.
[7] S. Wright and J. Nocedal, “Numerical optimization,” Springer Science, vol. 35, pp. 67–68, 1999.
[8] V. Chandrasekaran, S. Sanghavi, P. A. Parrilo, and A. S. Willsky, “Rank-Sparsity Incoherence for Matrix Decomposition,”
SIAM Journal on Optimization, vol. 21, no. 2, pp. 572–596, 2011.
[9] R. D. Zimmerman, C. E. Murillo-Sanchez, and R. J. Thomas, “MATPOWER: Steady-State Operations, Planning, and
Analysis Tools for Power Systems Research and Education,” IEEE Transactions on Power Systems, vol. 26, no. 1, pp.
12–19, Feb 2011.
[10] M. Grant and S. Boyd, “CVX: Matlab software for disciplined convex programming, version 2.1,” http://cvxr.com/cvx,
Mar. 2014.
[11] IEEE, “14 Bus Test Case,” https://www.ee.washington.edu/research/pstca/pf14/pg tca14bus.htm.

41
[12] O. Ardakanian, Y. Yuan, R. Dobbe, A. von Meier, S. Low, and C. Tomlin, “Event detection and localization in distribution
grids with phasor measurement units,” in IEEE Power Energy Society General Meeting, July 2017, pp. 1–5.
[13] “IEEE 5 bus test system data,” http://shodhganga.inflibnet.ac.in/bitstream/10603/26549/14/14 appendix.pdf, [Online; ac-
cessed 20-Jan-2019].
[14] “OPAL-RT real-time digital simulator,” https://www.opal-rt.com/simulator-platform-op5600/, [Online; accessed 20-Jan-
2019].
[15] X. Li, H. V. Poor, and A. Scaglione, “Blind topology identification for power systems,” in Proc. IEEE SmartGridComm,
2013.
[16] V. Kekatos, G. B. Giannakis, and R. Baldick, “Grid topology identification using electricity prices,” in IEEE PES General
Meeting — Conference Exposition, July 2014, pp. 1–5.
[17] M. Babakmehr, M. G. Simoes, M. B. Wakin, and F. Harirchi, “Compressive Sensing-Based Topology Identification for
Smart Grids,” IEEE Transactions on Industrial Informatics, vol. 12, no. 2, pp. 532–543, April 2016.
[18] D. Deka, S. Backhaus, and M. Chertkovy, “Estimating distribution grid topologies: A graphical learning based approach,”
arXiv preprint arXiv:1602.08509, 2016.
[19] Y. Liao, Y. Weng, M. Wu, and R. Rajagopal, “Distribution grid topology reconstruction: An information theoretic approach,”
in North American Power Symposium, Oct 2015, pp. 1–6.
[20] Y. Liao, Y. Weng, and R. Rajagopal, “Urban Distribution Grid Topology Reconstruction via Lasso,” in IEEE PES General
Meeting, July 2016, pp. 1–6.
[21] J. Yu, Y. Weng, and R. Rajagopal, “PaToPa: A data-driven parameter and topology joint estimation framework in distribution
grids,” May 2017, arXiv preprint:1705.08870.
[22] S. Park, D. Deka, S. Backhaus, and M. Chertkov, “Topology and parameter estimation in distribution grids with terminal
node measurements,” 2019, preprint.
[23] H. Zhu and G. B. Giannakis, “Sparse overcomplete representations for efficient identification of power line outages,” IEEE
Transactions on Power Systems, vol. 27, no. 4, pp. 2215–2224, November 2012.
[24] D. Deka, S. Backhaus, and M. Chertkov, “Structure learning in power distribution networks,” IEEE Transactions on Control
of Network Systems, vol. 5, no. 3, pp. 1061–1074, September 2018.
[25] Y. Yuan, X. Tang, W. Zhou, W. Pan, X. Li, H.-T. Zhang, H. Ding, and J. Goncalves, “Data driven discovery of cyber
physical systems,” Nature communications, vol. 10, no. 1, pp. 1–9, 2019.

42
Ye Yuan (M’13) received the B.Eng. degree (Valedictorian) from the Department of Automation, Shanghai
Jiao Tong University, Shanghai, China, in September 2008, and the M.Phil. and Ph.D. degrees from the
PLACE
Department of Engineering, University of Cambridge, Cambridge, U.K., in October 2009 and February
PHOTO
2012, respectively. He has been a Full Professor at the Huazhong University of Science and Technology,
HERE
Wuhan, China since 2016. Prior to that, he was a Postdoctoral Researcher at UC Berkeley, a Junior
Research Fellow at Darwin College, University of Cambridge, and has been holding visiting researcher
positions at California Institute of Technology, Massachusetts Institute of Technology, and Imperial College London. His research
interests include system identification and control with applications to cyber-physical systems.
He is the recipient of China National Recruitment Program of 1000 Talented Young Scholars, Dorothy Hodgkin Postgraduate
Award, Microsoft Research PhD Scholarship, Cambridge Overseas Student Award, Chinese Government Award for Outstanding
Students Abroad, Henry Lester PhD Scholarship and a number of best paper awards in IEEE conferences.
Steven Low (Fellow, IEEE) received the B.S. degree from Cornell University, Ithaca, NY, USA, and the
Ph.D. degree from the University of California, Berkeley, CA, USA, both in electrical engineering.
PLACE
He is a Professor with the Department of Computing and Mathematical Sciences and the Department
PHOTO
of Electrical Engineering, California, Institute of Technology, Pasadena, CA, USA. Before that, he was
HERE
with AT&T Bell Laboratories, Murray Hill, NJ, USA, and the University of Melbourne, Australia. His
current research interest is in the control and optimization of power systems.
Dr. Low is a senior editor of the IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS and the IEEE
TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING, is on the editorial boards of NOW Foundations, Trends
in Networking, Electric Energy Systems, as well as Journal on Sustainable Energy, Grids and Networks.
Omid Ardakanian (M’15) received the B.S. degree in Computer Engineering from Sharif University
of Technology, Iran, and the MMath and PhD degrees both in Computer Science from the University of
PLACE
Waterloo, ON, Canada. He is an Assistant Professor with the Department of Computing Science, University
PHOTO
of Alberta, Edmonton, AB, Canada. Prior to that, he was a Postdoctoral Fellow in the Department
HERE
of Electrical Engineering and Computer Sciences, University of California, Berkeley, USA, and in the
Department of Electrical and Computer Engineering, University of British Columbia, Canada.
Dr. Ardakanian served as a Guest Editor of IEEE TRANSACTIONS ON SMART GRID. He was a recipient of the Postdoctoral
Fellowship from Natural Sciences and Engineering Research Council of Canada (2015), and a number of best paper awards in
ACM and IEEE conferences.

43
Claire Tomlin (Fellow, IEEE) received the M.Sc. degree from Imperial College London, London, U.K., in
1993, the B.A.Sc. degree from the University of Waterloo, Waterloo, ON, Canada, in 1992, and the Ph.D.
PLACE
degree from the University of California at Berkeley, Berkeley, in 1998, all in electrical engineering.
PHOTO
She is a Professor in the Department of Electrical Engineering and Computer Sciences, University of
HERE
California, Berkeley, and holds a part-time Research Professor position in the Department of Aeronautics
and Astronautics, Stanford University, Stanford, CA. She has held Visiting Research Positions with NASA
Ames, Honeywell Labs, and the University of British Columbia. Her research interests include control systems, specifically hybrid
control theory, and she works on air traffic control automation, flight management system analysis and design, and modeling
and analysis of biological cell networks.
Prof. Tomlin was a recipient of the MacArthur Fellowship (2006), the Eckman Award OF the American Automatic Control
Council (2003), the MIT Technology Review’s Top 100 Young Innovators Award (2003), the AIAA Outstanding TEACHER
Award (2001), an NSF Career Award (1999), a Terman Fellowship (1998), and the Bernard Friedman Memorial Prize in Applied
Mathematics (1998).

Inverse Power Flow

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Inverse Power Flow

Uploaded by

Copyright:

Available Formats

1

Inverse Power Flow Problem

February 21, 2020 DRAFT

February 21, 2020 DRAFT

II. IPF WITHOUT HIDDEN NODES

February 21, 2020 DRAFT

A power system can be modeled by an undirected connected graph G = (N , E) where N :=

February 21, 2020 DRAFT

February 21, 2020 DRAFT

Based on the definition of Γ, the constrained l2 optimization problem can be converted to an

in which I denotes an identity matrix.

The optimization problem (6) can be written as

February 21, 2020 DRAFT

which contradicts the full column rank assumption.

February 21, 2020 DRAFT

III. IPF WITH HIDDEN NODES

A. Kron-reduced admittance matrix Ȳ

February 21, 2020 DRAFT

set of measured and hidden nodes respectively.

Assumption 1. We make the following assumptions:

If Y22 is invertible then eliminating V2 from (11) yields a relation

I1 (k) = Ȳ V1 (k), ∀k (12)

February 21, 2020 DRAFT

Ȳ , Y11 − Y12 Y22−1 Y12T (13)

Assumption 2. The admittance matrix Y , G + iB defined in (10) satisfies:

Lemma 1. If H < N , then G22  0, −B22  0 and G22 − B22  0.

By definition A22 [i, j] = A21 [i, j] = 0 if i 6∼ j. For i, j ∈ M1 ∪ M2 , if i ∼ j, then Yij =

February 21, 2020 DRAFT

Proposition 2. Under Assumptions 1 and 2, if H < N then Y22 is invertible.

G22 v − B22 w = 0 and G22 w + B22 v = 0,

Since G22 − B22  0 by Lemma 1, we can eliminate v from (15) to get

(G22 − B22 ) + (G22 + B22 )(G22 − B22 )−1 (G22 + B22 ) w = 0

Multiplying wT on the left we have

wT (G22 − B22 )w + ŵT (G22 − B22 )−1 ŵ = 0

February 21, 2020 DRAFT

current measurements (Proposition 1).

An admittance matrix Y ∈ SN specifies a unique weighted undirected graph G = (N , E) with

Algorithm 1 Graph Condensation Algorithm

where Y [i, i] ∈ C is the ith diagonal element of Y and

February 21, 2020 DRAFT

Kron Reduction with respect to M1

Fig. 3: Two equivalent schemes to compute the Kron reduced Ȳ .

Y /Y [i, i] = Y (i, i) − Y (i, i]Y −1 [i, i]Y (i, i]T . (16)

Example 2. Consider a (Kron-reduced) admittance matrix for a two-node network:

February 21, 2020 DRAFT

as its Kron reduction is:

as long as (θ1 , θ2 ) satisfies

B. Radial networks: exact identification

Assumption 3. The admittance matrix Y satisfies:

February 21, 2020 DRAFT

E 0 ⊆ E. A clique of G is a complete subgraph of G. A maximal clique of G is a clique that is

February 21, 2020 DRAFT

Ȳ1 := Y11 − diag{1T Y11 } (17a)

Ȳ2 := diag{1T Y11 } − Y12 Y22−1 Y12T . (17b)

February 21, 2020 DRAFT

1T Y11 = −1T Y12T , 1T Y12 = −1T Y22 .

i.e., Ȳ2 defined in (17b) satisfies Assumption 3.3 when Y does.

Ȳ = Y11 − Y12 Y22−1 Y12T = Ȳ1 + Ȳ2 ,

and hence Ḡ := G(Ȳ ) = G1 ⊕ G2 with G(Ȳ1 ) = G1 . Therefore we have G(Ȳ2 ) = G2 . This

February 21, 2020 DRAFT

Algorithm 2 Graph Decoupling Algorithm

February 21, 2020 DRAFT

Ḡ := G(Ȳ ) Ḡ2 := G(Ȳ2 )

Lemma 1. If H < N , then G22 0, −B22 0 and G22 − B22 0.

Since G22 − B22 0 by Lemma 1, we can eliminate v from (15) to get