Sparse and Numerically Stable Implementation of a

Distribution System State Estimation based on

Multifrontal QR Factorization
Gustavo M. Hebling
Julio A. D. Massignan
João A. B. London Junior
Marcos H. M. Camillo
University of São Paulo
São Carlos, SP, Brasil
gustavo.hebling@usp.br, julio.massignan@usp.br, jbalj@sc.usp.br, marcos.camillo@copel.com

Abstract—Enhancing situational awareness of distribution net- computational performance are two requirements of the state
works is a requirement of Smart Grids. In order to fulfill estimation process. Even more so in recent years with the
this requirement, specialized algorithms have been developed increase of real-time applications and also data volumes at
to perform Distribution System State Estimation (DSSE). Due
to the particularities of such networks, those algorithms often operation centers.
rely on simplifications and approximations of the measurement Despite the consolidated position of the Weighted Least
model which make difficult to generalize their results. This paper Squares (WLS) state estimator for transmission systems, spe-
presents a sparse and numerically stable implementation of an cialized algorithms have been developed to perform Distri-
algorithm for DSSE, which does not require any additional bution System State Estimation (DSSE) [1]. The motivation
assumption from the traditional state estimation formulation.
The numerical stability is guaranteed by using Multifrontal QR was creating algorithms capable of dealing with three-phase
factorization and an optimal ordering technique is evaluated to unbalanced, asymmetrical, untransposed, large scale radial
reduce fill-in. Simulation results are carried out with IEEE three- networks, with a few real-time measurements along with
phase unbalanced test feeders to evaluate the algorithm. a high number of pseudo-measurements and also virtual-
Index Terms—Distribution System, Multifrontal QR, Numeri- measurements to characterize the loads [1]. Instead of using
cal Methods, State Estimation, Sparsity. bus voltages, those algorithms use currents flowing in the
network branches as state variables [2]. Another approach is to
use a formulation based on the admittance matrix of the system
I. I NTRODUCTION [3]. These alternative approaches do not completely solve the
central issue of DSSE, the ill-condition of the Gain Matrix,
Distribution Systems (DSs) are currently facing a new oper-
because the primary source of ill-conditioning is related to
ating environment due to the penetration of renewables, energy
different weights of measurements [4], as well as connections
storage, electric vehicles, and demand response. The utilities
between long and short lines, common to DS [5].
are given the task to manage a large amount of data from
Virtual measurements are used to ensure network observ-
new sensors, such as smart meters, for real-time applications
ability and usually high weights are assigned to this type of
to guarantee a reliable, safe and optimized operation of their
measurement in contrast to low weights assigned to pseudo-
networks in such a complex scenario. In this context, the state
measurements. This may cause ill-conditioning and some
estimator is the fundamental computational tool for real-time
proposals such as [6], [7] suggest modelling virtual measure-
monitoring of power systems, gathering the available measured
ments as equality constrains using Lagrangian Relaxation to
data and the network model to evaluate the operating state [1].
include them in the WLS estimator. However, this produces
State estimation is, therefore, the first step to perform a
some numerical instability and sparse techniques as well as
wide range of automated applications in energy management
row pivoting must be employed. A more stable approach to
systems, such as power flow, contingency analysis, load fore-
factorization is proposed in [8], [9], [10] where orthogonal
casting, stability margin calculations, short-circuit calculations
transformations are used to solve the least-square problem
and others. Since many of these applications provide the
via backwards substitution. This approach tends to produce
situational awareness of the grid in real time, accuracy and
excessive fill-ins whereas in [11] Givens rotations are used to
improve robustness and an ordering technique is applied to
This work was supported by CNPq, CAPES, FAPESP (grants 2018/00214-4 maintain sparsity.
and 2016/19646-6) and COPEL S/A (project PD PD-2866-0504/2018). Within this context, this paper presents an efficient and

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
numerically stable algorithm to perform DSSE using the k is the iteration index. The linear approximation at the point
traditional WLS nonlinear formulation with nodal voltages xk is:
as state variables, i.e. without any additional assumption or h(xk+1 ) ≈ h(xk ) + H(xk )4xk (4)
simplification on the measurement model. The contribution
resides in solving very ill-conditioned DSSE problems with Rewriting (1) considering the linear approximations pre-
the classical WLS state estimator formulation. The numerical sented beforehand we obtain:
stability of the proposed algorithm is enhanced through the z = h(xk ) + H(xk )4xk + e (5)
application of QR-based formulation in the WLS algorithm,
while computational efficiency is achieved through sparse Which is equivalent to:
routines based on the Multifrontal approach. An efficient 4z(xk ) = z − h(xk ) = H(xk )4xk + e (6)
ordering technique reduces the number of fill-ins exploiting
even further numerical stability and computational efficiency. From the linear approximations, it is possible to write a new
Simulations are performed on the three-phase unbalanced test objective function J(4x):
feeders: IEEE34, IEEE123 and IEEE US Low Voltage (342
J(4x) = [4z(xk ) − H(xk )4xk ]T W [4z(xk ) − H(xk )4xk ]
nodes).
(7)
This paper is structured as follows: in section II we present
The minimum of J(4x) is obtained when:
the traditional mathematical formulation for the state estima-
tion process based on the Normal Equation, some considera- ∂J(4x)
= H(xk )T W [4z(xk ) − H(xk )4xk ] = 0 (8)
tions about the condition number of the matrices involved in ∂x
this process as well as the different types of measurements The correction 4xk is then given by the solution of:
available; section III presents the proposed algorithm that
uses a QR factorization and an optimal ordering for fill-in 4xk = [H(xk )T W H(xk )]−1 H(xk )T W 4z(xk ) (9)
reduction; Section IV is reserved to simulation results and
Equation (9) is known as the Normal Equation and to
Section V presents the conclusions and final remarks.
obtain its solution it is necessary to invert or factorize
II. D ISTRIBUTION S YSTEM S TATE E STIMATION H(xk )T W H(xk ), which is called the Gain matrix G. Since it
is computationally inefficient to invert the Gain matrix, factor-
A. Conventional State Estimation based on the Normal Equa-
ization methods are used to obtain the solution such as Gauss
tion
Elimination, Cholesky Factorization, LU Decomposition and
The distribution system state estimation is based on the others. The numerical robustness of the solution of (9) depends
nonlinear model of measurements: not only on the chosen method for factorization but also on
z = h(x) + e, (1) the condition number of the Gain matrix [12].
It is possible to show that the condition number of the Gain
where z is the measurement vector (mx1), x is the state matrix K(G) (defined as the ratio between the largest and
variables vector (nx1), h(.) is the nonlinear state estimation smallest of its eigenvalues λ(G)) is close to the magnitude of
function (mx1) that relates the measurements to the states vari- the square of the condition number of the weighted Jacobian
ables, and e is the vector of measurement errors (mx1) usually matrix, as follows:
considered as independent random Gaussian variables with λmax (G) 1
zero mean and diagonal covariance matrix Rz (Rz = diag K(G) = ≈ [K(W 2 H)]2 (10)
2 2 2 λmin (G)
{σz1 , σz2 , ..., σzm }, where σzi is the standard deviation of
measurement zi ).
DSs have a set of particularities that make the resulting
Through the classical Weighted Least-Squares (WLS) esti-
system of the WLS even more prone to ill-conditioning. These
mator, the state estimate vector x̂ is obtained by minimizing
particularities are related to the topology of the network, the
the index J(x) given by:
number of buses and availability of measurements. A DS is
J(x) = [z − h(x)]T W [z − h(x)], (2) usually radial, with single-phase as well as two and three-
phase unbalanced and short branches, containing different
where W = Rz−1 . transformers connections and has a large number of nodes.
The minimum value for J(x) is obtained when its derivative, The number of virtual and pseudo-measurements with different
∂J(x)
∂x = 0, is equal to zero which can be written as: precision values from the real-time measurements also makes
H T (x̂)W [z − h(x̂)] = 0 (3) the weighting process a source of ill-conditioning.
These aforementioned characteristics have prevented the
The matrix H(x̂) is known as the Jacobian matrix and it is the use of the WLS state estimator based on the Normal Equa-
matrix of first derivatives of h(x) evaluated at the x = x̂ point. tion method for DSSE [13], [14]. Instead, some alternative
Given the nonlinear nature of the index J(x), an iterative formulations have been proposed in the literature such as
algorithm is used to obtain a solution to a linear equation, the Branch Current approach in [15] and the Admittance-
which calculates a correction, 4xk , to the state vector where matrix approach in [16]. These alternative formulations are

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
based on simplifications either in the state variables or in the for using a stable approach such as the method proposed in
measurement set that may not be as accurate when applied in this work. It is noteworthy that other approaches may also
modern DSs. be applied for dealing with virtual measurements, such as
using equality constraint formulation by Lagrangian methods
B. Real-Time Measurements
for DSSE [6] but will not be addressed in this work.
Real-time measurements represent the information gathered
by the SCADA system about the monitored electrical quanti- III. T HE PROPOSED S PARSE AND N UMERICALLY S TABLE
ties in the system typically composed by power flows, power A LGORITHM FOR DSSE
injections, voltage and current magnitudes across the network. In order to counter the associated ill-conditioning of DSSE
Another important kind of measurements are the synchronized using the traditional WLS, a new strategy is proposed in this
phasor measurements obtained by Phasor Measurement Units paper based on the application of an orthogonal method for
(PMUs) that monitor the voltage and current phasor (magni- solving the state estimation problem. Using the QR factoriza-
tude and angle). Another recent source of information in DSs tion as follows:
is the smart meter, which has recently been started to be widely
deployed in the low voltage networks. These smart meters W 1/2 H(xk ) = QT R (12)
typically provide power injections and voltage magnitudes at It is possible to rewrite the normal equation as:
a consumer level. These information build the measurement
vector z and the respective nonlinear model h(x) in (1). [RT QQT R]4xk = RT QW 1/2 [z − h(xk )] (13)
To obtain the diagonal elements of the weighting matrix Since QQT = I, where I is the identity matrix, we obtain:
2
W of the real-time measurements (Wi,i = 1/σzi , for i =
1, . . . , mrt , where mrt is the number of available real-time [RT R]4xk = RT QW 1/2 [z − h(xk )] (14)
measurements) the approach proposed in [17], [18] is used
Multiplying both sides by (RT )−1 :
according to (11).
 
 lf  R4xk = QW 1/2 [z − h(xk )] (15)
zi  pri
k
σz i = , (11) From (15), it is possible to obtain the vector 4x using back-
3
wards substitution since R was obtained via QR factorization
where, zilf is the value obtained from an exact load flow and it is, therefore, upper triangular. The resulting system is
solution (reference value) and pri is the associated precision better conditioned in terms of numerical robustness than that
of the measurement device. obtained with the normal equation. Consider the condition
number of the upper triangular matrix R:
C. Load Pseudo-Measurements
In the context of nowadays distribution networks, pseudo K(R) = kRkkRk−1 (16)
measurements consist of the majority of information available
about the system and their qualities directly impact on the K(R) = k[(QT )−1 W 1/2 H]kk[(QT )−1 W 1/2 H]−1 k (17)
DSSE accuracy [19]. Typically, these pseudo-measurements
T −1 1/2 1/2 −1 T
represent the load of distribution transformers that connect K(R) = k[(Q ) W H]kk[(W H) Q ]k (18)
medium-voltage to low-voltage circuits. Usually, these trans-
formers do not have any associated real-time measurement.
K(R) ≈ k(QT )−1 kk(W 1/2 H)kk(W 1/2 H)−1 kkQT k (19)
Therefore, pseudo-measurements are necessary for observ-
ability requirements. The approach of this work consists of Since Q is orthogonal, the additional following properties are
representing the power injections (loads) in the network as load true: QT = Q−1 and kQk = k(QT )−1 k = 1. Therefore we
pseudo-measurements. The information used to generate these can approximate the condition number of R as:
pseudo measurements is typically obtained by stratification of
the consumers into classes and using typical load patterns for K(R) ≈ k(W 1/2 H)kk(W 1/2 H)−1 k = K(W 1/2 H) (20)
each consumer. Thus, less accurate informations than the real- Equation (20) shows that the resulting system, when applying
time measurements are introduced in the measurement vector the QR factorization to the WLS formulation, has a condition
and represented with different precision levels. number of the same order of magnitude as the Jacobian matrix
D. Virtual Measurements whereas the system obtained with the Normal Equation has
a condition number approximately the square of that of the
Virtual measurements represent known electrical quantities
Jacobian matrix. In addition to being a better conditioned
in the system such as zero-injection nodes (nodes without
system, when using the QR factorization another benefit lies
loads and generation) and are incorporated as measurements
in the stability of the method when dealing with floating point
with low standard deviations (in this work 1e-6). The differ-
arithmetic.
ences among the variance values of the real-time, virtual and
In order to improve readability, we define the matrix A as:
pseudo-measurements are another well-known source of ill-
conditioning in DSSE [6], [7] and also an essential motivation A = W 1/2 H (21)

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
The corresponding QR factorization satisfies: form of heuristics which will be discussed further in the next
section. Then a symbolic analysis is performed in order to
Q̃R̃ = A + δA (22)
obtain the non-zero pattern of the matrix by performing a
Where kδAk
= O(machine ). The forward step is stable [20] symbolic Cholesky factorization of AT A. Next, there are the
kAk
and proving the backwards stability we obtain: numerical factorization phase and the proper solving phase
that uses backward substitution.
(A + 4A)x̃ = b (23) Figure 1 represents the underlying structure of the algorithm
k4Ak
in which the matrix is reordered. Its elimination tree is
where kAk= O(machine ). The first step in the proof is to obtained via the symbolic factorization, and the supernodes
write A + 4A as: are found in order to form the dense frontal matrices that will
(Q̃ + δQ)(R̃ + δR)x̃ = b (24) be subject to the numerical factorization using the Householder
QR method , whereas other proposals use Givens Rotations to
[Q̃R̃ + Q̃(δR) + R̃(δQ) + (δQ)(δR)]x̃ = b (25) obtain a orthogonal factorization [26]. The figure presents the
elimination tree with supernodes as well as the corresponding
Since Q̃R̃ = A + δA, and δA is small when compared to
R factor.This algorithm is available as an open-source package
A, we can write:
named SuiteSparse.

kR̃k kA + δAk
≤ kQ̃∗ k = O(1) (26)
kAk kAk

Using equation (26) and adding the terms of equation (25)

we can show that:

k4Ak kδAk k(δQ)R̃k

≤ + + (27)
kAk kAk kAk
k(δR)Q̃k k(δQ)(δR)k
+ + = O(machine )
kAk kAk
One consequence of the backwards stability of the QR
algorithm is that the obtained solution x̃ follows:
kx̃ − xk
= O(K(A)machine ), (28)
kxk Fig. 1. Matrix A, its QR factor R and the corresponding elimination tree.
Source: [22]
where K(A) is the condition number of the matrix A [20].
Overall, the proposed algorithm that uses the QR factorization
of the matrix W 1/2 H results in a smaller condition number B. Efficient Ordering for Fill-In Reduction
than that of the correspondent Gain matrix. This fact also has It is common to reorder the rows and columns of the
influence on the backwards stability of the method because of matrix being factorized in order to reduce the number of fill-
equation (28). Therefore, the proposed algorithm in addition ins it produces, that is, elements that are initially zero and,
to being better conditioned is also stable. after the factorization, become non-zero. The optimal ordering
problem consists of finding a permutation matrix P in order
A. Sparse QR Multifrontal Implementation to reduce the number of non-zeros created and the amount of
Householder Reflections [21] and Givens Rotations [22] are computation work in the factorization of P AP T [27], [28].
the methods usually applied to obtain a QR factorization. This optimization problem is NP-complete [29] and heuristics
In order to extract peak computational performance when are used in an attempt to solve it.
factorizing large sparse matrices other methods have been pro- The strategies to reduce the number of fill-ins are usually
posed such as the Multifrontal QR [23]. Traditional orthogonal related to the undirected graph of the matrix being factorized
methods for QR factorization of sparse matrices operate on because the factorization process is equivalent to a node
one row or column at a time [24], [25] and are not able elimination in the graph. Fill-in means, therefore, new edges
to achieve said maximum performance because of irregular created in the graph after the elimination process of a node
access of memory. The Multifrontal QR performs on groups and one strategy to reduce this number is to select nodes with
of rows and columns in a sequence of dense matrices called the minimum degree, that is, the number of connections to
frontal matrices that can be handled in parallel. be eliminated. Note that this number is updated after each
There are three phases in solving the resulting linear system elimination step because new edges are created. The resulting
in (15) when using a direct method such as the Multifrontal permutation matrix is that which orders the matrix being
QR. First, a fill-in reducing permutation is obtained using some factorized in the sequence that the minimum degree nodes will

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
be eliminated first. Consider the undirected graph of a 10x10 vector composed by the complex
p
exponential part of the angle
matrix presented in figure 2. A minimum degree ordering of the phasor (θip = ej.θi ). Working algebraically, obtain the
would result in a permutation such as that the column indexes following expression for the power flow equations:
of the original matrix would be 5,10,4,7,2,1, 3,6,8,9 in the
permutated matrix.
S p = [V p ].[θp ].Y pp∗ .[V p ].θp∗ + [V p ].[θp ].Y ps∗ .[V s ].θs∗ .
(32)
The above equation is used to build active and reactive
power flow models for steady state analysis. The derivatives
of such equation regarding the state variables (complex nodal
voltages in polar coordinates), in the same matrix representa-
tion, are:

∂S p
=J ii .[θp ].Y pp∗ .[V p ].θp∗ (33)
∂Vip
+[V p ].[θp ].Y pp∗ .J ii .θp∗ + J ii .[θp ].Y ps∗ .[V s ].θs∗
Fig. 2. Undirected graph of a 10x10 matrix. Source: [30]
∂S p
= [V p ].[θp ].Y ps∗ .J ii .θs∗ (34)
C. Three-Phase Models for DSSE ∂Vis
The three-phase network model to perform DSSE in this
work consists on the representation of each element of the DS ∂S p p
=j.eθi .[V p ].J ii .Y pp∗ .[V p ].θp∗ (35)
by its two-port admittance matrix. Different types of equip- ∂θip
ment and connections can be incorporated in the measurement −j.[V p ].[θp ].Y pp∗ .[V p ].J ii .θp∗
model by simply changing the respective admittance matrix, p
+j.eθi .[V p ].J ii .Y ps∗ .[V s ].θs∗
which is an implementation advantage when representing com-
plex distribution networks with several different components
[31]. Each element is modelled by the following two-port ∂S p
= −j.[V p ].[θp ].Y ps∗ .[V s ].J ii .θs∗ , (36)
model: ∂θis

I˙
 p   pp
Y Y ps
  p
V̇ where, J ii is a matrix with the same size as the number
= . , (29) of phases and only the ii element equal one and the others
I˙s Y sp Y ss V̇ s
equal zero. It is noteworthy that the above equations can be
where, p and s denote the terminals of the branch element, used for different types of equipment, connections and number
I˙p and I˙s are the current phasor injected in each terminal of of phases, a general model for steady state analysis such as
the branch element, V̇ p and V̇ s the voltage phasor of each DSSE. Details of three-phase models and respective admit-
terminal, and Y pp , Y ps , Y sp and Y ss are the admittance tance matrices to build each two-port model for distribution
sub-matrices that represent each equipment and the respective circuits, transformers, voltage regulators, capacitor banks and
parameters. other equipment can be found in [32], [33].
The power flow equations can be written for the respective
branch as the following matrix expression for the terminal p: IV. S IMULATION R ESULTS
This section presents the numerical results of the application
of the proposed algorithm to the IEEE 34, 123, and 342 dis-
S p = P p + j.Qp = V̇ p  (I˙p )∗ = V̇ p  (Y pp .V̇ p + Y ps .V̇ s )∗ ,
tribution test feeders. For all simulations, based on a reference
(30)
load flow solution, 100 test samples were generated via Monte
where, S p is a vector with the per-phase complex power at
Carlo methodology through the inclusion of random errors in
terminal p (active and reactive power flows, P p and Qp ), 
the reference values of the measurements (obtained from an
denotes the Hadamard product (element-wise), and ∗ denotes
exact load flow solution). Therefore, the measurement values
the complex conjugate. Separating the voltage phasor in its
used in each simulation were obtained according to:
exponential representation according to the following expres-
sion: zi = zilf + ui .σzi , (37)
where zi is the value of the i − th measurement that will
V̇ p = V p  θp = diag(V p ).θp = [V p ].θp , (31)
constitute the input data for a given simulation of the proposed
where [.] denotes in this paper, for the sake of notation algorithm, zilf is the corresponding reference value for i − th
simplification, a diagonal matrix formed by the elements of measurement obtained from an exact load flow solution, ui is a
a vector, V p is a vector composed by the voltage magnitude random variable that follows ui ∼ N (0, 1), that is, zero mean
in the i − th phase of the system (Vip = |Vip |) , and θp is a and unitary standard deviation and finally σzi is the standard

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
deviation of the i − th measurement given by equation (11). number of G is approximately the square of the condition
To define the standard deviation of each type of measurement, number of W 1/2 H.
the following meter precision were used in this paper: 30%
for pseudo-measurements, 5% for smart-meter measurements, TABLE I
2% for active and reactive power injection measurements, P ERFORMANCE METRICS FOR EACH SIMULATION SCENARIO .
1% for voltage magnitude measurements, 0.1% for PMU Scenario MAE Kema K(W 1/2 H) K(Gain)
measurements and 0.0001% for virtual measurements. 1 0.00058512 3.5981e-05 6.6843e+09 5.5326e+18
2 0.00040791 2.0207e-05 1.3536e+10 1.4447e+18
A. Accuracy Analysis
3 0.00041309 1.7887e-05 8.7496e+11 6.4480e+21
The IEEE US Low Voltage distribution test system with 342
nodes [34] was used to perform accuracy analysis. Figure 3
shows the system, in red the medium voltage network and in The MAE and Kema metrics were also graphed in order to
blue the low voltage network, note that this system contains visualize the simulation results for each Monte Carlo simula-
meshed networks as well. tion and also for each state variable. In figure 4 we present
the KEMA metric obtained in each Monte Carlo repetition.
In figure 5, we present the average value of the MAE metric
for each state variable, this metric is also condensed on table
II where the average MAE is presented for each scenario.
Also in figure 5, the state variables in the high voltage (HV)
system are highlighted as well as medium voltage (MV) and
low voltage (LV). The proposed algorithm returned the worst
performance when applied for Scenario 1, mainly because of
the reduced number of measurements (the global redundancy
(GR) of Scenario 1 is lower than the others).

TABLE II
D ETAILED ERRORS FOR ALL SIMULATION SCENARIOS .
Scenario Magnitude Error (p.u.) Angle Error (p.u.) GR
1 6.0290e-04 5.6715e-04 1.12
Fig. 3. 342 buses system used in simulations
2 4.2891e-04 2.9507e-04 1.28
3 3.9800e-04 2.4537e-04 1.50
Three different scenarios of measurement sets were created
to perform the simulations. The first one contains mostly
pseudo-measurements and no smart-meter measurements. In
the second one, a high number of smart-meter measurements
were used and, in the third scenario, SCADA measurements
of power flow in transformers were added to the measurement
set of the second scenario. Each simulation result for a given
Monte Carlo repetition is stored in order to evaluate the
accuracy of the proposed algorithm using metrics such as the
MAE (Mean Absolute Error):
nX
trials
1
M AEi = |x̂ki − xref
i |, (38) Fig. 4. KEMA metric for each Monte Carlo repetition
ntrials
k=1

where k represents each repetition of ntrials simulations of

the Monte Carlo method. Another useful metric evaluated is
the Kema, based on [35]:
v
u nbus nphases
1 u X X ~ ~
Kemak = t (|Vjφ − Vjφ,ref |) (39)
nbus
nb=1 φ=1

We evaluated the 342 nodes system with 100 Monte

Carlo repetitions for each measurement scenario previously
presented. The condensed results are presented in table I,
which also contains the condition number of the corresponding Fig. 5. MAE metric for all state variables
matrices W 1/2 H and G. As it is expected, the condition

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
B. Computational Aspects

Other distribution networks were simulated to understand

the computational work improvement associated with the QR
factorization method for state estimation. The aim was to
evaluate execution time and the number of iterations for DSs of
different sizes. In Table III, we present the obtained execution
times for convergence as well as the GR of the measurement
set used for three different-sized systems which contain 34,
123, and 342 nodes. The condition number of the matrix
W 1/2 H is presented as well as the condition number of the
corresponding Gain matrix.These simulations were performed
on a machine that has a 2.7 Ghz Core i5 with 8gb of RAM and
used a Unix OS and the algorithm was implemented using the
C programming language. The current implementation of the
proposed algorithm presents satisfactory results in terms of
execution time since convergence is obtained in 4 iterations
or less and total time necessary is quite low even for the
larger 342 buses system. Further improvements on the code Fig. 6. Iteration comparison between proposed QR and the Cholesky and
may decrease memory usage. QR factorization of the Gain matrix, IEEE 34 nodes system and 123 nodes
system, respectively.
In order to better illustrate the convergence of the proposed
QR factorization a comparison is shown in figure 6 for the
34 and 123 nodes systems. The y-axis is the log of the
maximum absolute variation of the state variables between
two consecutive iterations. Curves are shown for not only
the proposed method but also the traditional WLS approach
which uses the Gain matrix and two different factorization
methods are compared: Cholesky and QR applied in the Gain
matrix. The Cholesky factorization is the most commonly used
given the symmetric and positive definite characteristics of the
Gain matrix. The implementation of the Cholesky factorization
is also sparse-oriented and we used an open-source package
named CHOLMOD [36], which is part of SuiteSparse. None
of three tested systems resulted in convergence when using
the Gain matrix approach.

TABLE III
Fig. 7. Sparsity plot of the matrices obtained for the 342 nodes system
C OMPUTATIONAL ASPECTS CONSIDERING DIFFERENT- SIZED SYSTEMS .
System GR Time (s) Iterations K(W 1/2 H) K(Gain)
IEEE34 3.49 0.1282 4 1.093e+10 3.434e+18 V. C ONCLUSIONS
IEEE123 1.11 0.2159 4 1.081e+09 1.714e+17
In this paper, we presented a sparse and numerically stable
IEEE342 1.12 1.1025 4 8.752e+11 1.646e+21
implementation of an algorithm for DSSE, which does not
require any additional assumption from the traditional state
Besides the computation gain in execution time, because of estimation formulation. This was achieved by using an or-
the fill-in reducing ordering step, the R factor of the matrix thogonal method to reduce the ill-conditioning of the system
W 1/2 H is expected to have a reduced number of fill-ins as well as an efficient ordering technique to preserve sparsity,
when compared to the execution without the ordering step. In that is, to reduce the number of fill-ins in the factorization
figure 7, we present the structure of the matrix W 1/2 H of the process.
reference case for the 342 nodes system and its corresponding Through a proper formulation of the QR method into the
R factor with the permutation matrix applied. We also present numerical solution of the three-phase state estimation problem,
the R factor without the permutation applied, which contain the algorithm does not require the explicit calculation of the
many more fill-ins. The fill-ins were reduced by 85.54%. We state estimator Gain Matrix. This translates into a large im-
also present the R factor of the corresponding G matrix to provement on the condition number of the nonlinear systems,
show that the traditional approach would yield the highest and sensibly reduces the computational burden. By choosing
number of fill-ins. a Multifrontal QR factorization specialized in sparse matrices

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
it is possible to improve even further the computational effi- [18] M. R. M. Castillo et al., ”Offline Detection, Identification, and Cor-
ciency in order to deal with real time operation requirements rection of Branch Parameter Errors Based on Several Measurement
Snapshots”, IEEE Trans. Power Systems, Vol. 26, No. 2, May 2011.
and high volumes of data. [19] J. A. D. Massignan et al., ”In-Field Validation of a Real-Time Monitor-
A diverse set of measurements were used to evaluate ing Tool for Distribution Feeders” IEEE Trans. Power Delivery, Vol. 33,
No. 4, Aug. 2018.
the proposed algorithm in IEEE three-phase unbalanced test [20] L. N. Trefethen, and D. Bau III, , ”Numerical Linear Algebra”. SIAM,
feeders in different conditions of operation and measurement Philadelphia, PA. 1997
availability. The results show that the proposed algorithm [21] A. S. Householder, ”Unitary triangularization of a nonsymmetric ma-
trix”, ACM 5, 339–342 1958.
retains accuracy and also decreases the condition number [22] W. Givens, ”Computation of plane unitary rotations transforming a
of the system. Fill-in is greatly reduced and the current general matrix to triangular form”, SIAM J. Appl. Math. 6, 26–50 1958.
implementation requires a low execution time in a variety of [23] T. A. Davis, , ”Algorithm 915, SuiteSparseQR: multifrontal multi-
threaded rank-revealing sparse QR factorization”. ACM Transactions on
systems. Further improvements may improve these results by Mathematical Software, Vol. 38, No. 1, 2011.
using parallel computing techniques. [24] A. George, J. W. H. Liu, and E. G. Ng, ”A data structure for sparse
The proposed algorithm will be applied to solve DSSE for QR and LU factorizations”, SIAM J. Sci. Statist. Comput. 9, 1, 100–121
1988.
larger systems as well as real DS in future works. Even higher [25] A George, and M. T. Heath, ”Solution of sparse linear least squares
computational gains are expected in these cases where the ill- problems using Givens rotations”, Linear Algebra Appl. 34, 69–83 1980.
conditioning is more severe. [26] A. S. Costa, R. Salgado and P. Haas, ”Globally convergent state
estimation based on Givens rotations” 2007 iREP Symposium - Bulk
Power System Dynamics and Control - VII. Revitalizing Operational
R EFERENCES Reliability, 2007.
[27] T. A. Davis, 2006. Direct Methods for Sparse Linear Systems. SIAM,
[1] A. Primadianto and C. Lu, “A Review on Distribution System State Philadelphia, PA.
Estimation”, IEEE Trans. Power Systems, vol. 32, no.5, Sept. 2017. [28] P. R. Amestoy, , T. A. Davis, , and I. S. Duff, ”Algorithm 837: AMD,
[2] M. Pau, et. al., ”Efficient branch-current-based distribution system state an approximate minimum degree ordering algorithm”. ACM Trans. Math.
estimation including synchronized measurements,” IEEE Trans. Instr. Softw. 30, 3, 381–388, 2004.
Meas., vol. 62, no. 9, 2013. [29] M. Yannakakis, Computing the minimum fill-in is NP-complete, SIAM
[3] M.C. Almeida, L.F. Ochoa, ”An Improved Three-Phase AMB Distribution J. Alg. Disc. Meth., 2 , pp. 77–79, 1981.
System State Estimator”, IEEE Trans. Power Systems, vol. 32, no. 2, Mar [30] A. Gomez, and L. G. Franquelo, , ”An Efficient Ordering Algorithm to
2017. Improve Sparse Vector Methods”. IEEE Transactions on Power Systems,
[4] D. Atanackovic and V. Dabic, ”Deployment of real-time state estimator Vol 3. No. 4, 1988.
and load flow in BC Hydro DMS - challenges and opportunities,” in 2013 [31] F. Ahmad, A. Rasool, E. Ozsoy, S. Rajasekar, A. Sabanovic ad M.
IEEE PES General Meeting, Vancouver, BC, 2013 Elitaş, “Distribution system state estimation-A step towards smart grid”,
[5] A. Monticelli, C. A. F. Murari, and F. F. Wu, “A hybrid state estimator: Renewable and Sustainable Energy Reviews, Vol. 81, No.2, Jan. 2018
Solving normal equations by orthogonal transformations,” IEEE Trans. [32] W. H. Kersting, “Distribution system modeling and analysis”, CRC
Power Appar. Syst., vol. PAS-104, no. 12, pp. 3460–3468, 1985 Press, Boca Raton, Florida, 2001
[6] W. Lin and J. Teng, ”State Estimation for Distribution Systems with Zero- [33] P. Xiao, D. C. Yu and W. Yan, “A unified three-phase transformer model
Injection Constraints”, IEEE Transactions on Power Systems, Vol. 11, No. for distribution load flow calculations,” IEEE Transactions on Power
1, 1996 Systems, Vol. 21, No. 1, 2006
[7] V. Thornley, N. Jenkins, and S. White, “State estimation applied to active [34] IEEE, “IEEE Test Feeders.” [Online]. Available at:
distribution networks with minimal measurements,” 15th Power Syst. http://sites.ieee.org/pes-testfeeders/. [24-Sep-2019].
Comput. Conf., no. 5, 2005. [35] “Metrics for determining the impact of phasor measurements on power
[8] A. Monticelli, C. A. F. Murari, and F. F. Wu, “A hybrid state estimator: system state estimation, DRAFT,” 2006.
Solving normal equations by orthogonal transformations,” IEEE Trans. [36] Y. Chen et al., ”Algorithm 887: CHOLMOD, Supernodal Sparse
Power Appar. Syst., vol. PAS-104, no. 12, pp. 3460–3468, 1985. Cholesky Factorization and Update/Downdate.”. ACM Transactions on
[9] A. Simoes-Costa and V. Quintana, “A robust numerical technique for Mathematical Software, Vol. 35, No. 3, 2008.
power system state estimation,” IEEE Trans. Power Appar. Syst., vol.
PAS-100, no. 2, pp. 691–698, 1981.
[10] A. Simoes-Costa and V. H. Quintana, “An orthogonal row processing
algorithm for power system sequential state estimation,” IEEE Trans.
Power Appar. Syst., vol. PAS-100, no. 8, pp. 3791–3800, 1981.
[11] N. Vempati, I. W. Slutsker, adn W. F. Tinney, ”Enhancements to Givens
Rotations for Power System State Estimation”, IEEE Trans. on Power
Systems, 6 (2):842-849, 1991.
[12] R. Ebrahimian, and R. Baldick, ”State Estimator Condition Number
Analysis”. IEEE Power Engineering Review 21(5):64-64,
[13] S. Lefebvre, ; J. Prévost, ; L. Lenoir, ”Distribution State Estimation: A
necessary requirement for the smart grid.” IEEE PES General Meeting
— Conference Exposition. National Harbor: [s.n.], 2014. p. 1–5, 2014.
[14] M. Baran, ”Branch current based state estimation for distribution system
monitoring.” IEEE Power and Energy Society General Meeting, p. 12–15,
2012.
[15] M. Baran, and A. Kelley, ”State estimation for real-time monitoring of
distribution systems. IEEE Transactions on Power Systems”, v. 9, n. 3,
p. 1601–1609, 1994.
[16] C. N. Lu, ; J. H. Teng,; W. H. E. Liu, ”Distribution system state
estimation.” IEEE Transactions on Power Systems, v. 10, n. I, p. 229–240,
1995.
[17] R. Singh, B.C. Pal and R.B. Vinter, “Measurement Placement in Distri-
bution System State Estimation”, IEEE Transactions on Power Systems,
Vol. 24, No.2, May 2009.

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020