A Two-Step Distribution System State Estimator

with Grid Constraints and Mixed Measurements

Miguel Picallo∗ , Adolfo Anta† , Bart de Schutter∗ and Ara Panosyan‡
∗ Delft Center for Systems and Control, Delft University of Technology
Delft, The Netherlands
m.picallocruz,[email protected]
† Austrian Institute of Technology
Vienna, Austria
‡ General Electric Global Research
Munich, Germany

Abstract—In this paper we consider the problem of State Therefore, the power flow equations need to be solved for
Estimation (SE) in large-scale, 3-phase coupled, unbalanced the 3 phases simultaneously, and approximations like the
distribution systems. More specifically, we address the problem fast decoupled power flow [6], [7] cannot be used. Another
of including mixed real-time measurements, synchronized and
unsynchronized, from phasor measurement units and smart limitation is the lack of sufficient real-time measurement units
meters, into existing SE solutions. We propose a computationally in distribution networks: although Phasor Measurement Units
efficient two-step method to update a prior solution using the (PMUs) and smart meters are being deployed, their high cost
measurements, while taking into account physical constraints [8] prevents installing the required number of sensors to make
caused by zero-injection buses. We test the method on a bench- the system fully observable.
mark test feeder and show the effectiveness of the approach.
Some methods have been proposed to perform SE by
exploiting the network structure and the nature of the mea-
Index Terms—Distribution System State Estimation, Phasor
Measurement Units, Power Flow Computation, Power System surements, such as branch-current-based SE [9]–[11]. Other
Modeling, Recursive Filtering methods use the few measurements available to improve the
SE accuracy of previous methods based on load assump-
I. I NTRODUCTION tions only [12]–[14]. Yet, most of this work is limited to
The operation of a power network requires accurate moni- transmission networks, small-scale distribution networks, or
toring of its state: bus voltages, line currents, consumption, and single-phase networks. Some recent work extends it to more
generation. This is specially relevant in transmission networks general networks, but assumes availability of measurements
where volatile and distributed generation and consumption at every bus [15], which at the moment is not the case
cause bidirectional power flow and voltage drops. Normally, for most distribution networks. Additionally, when traditional
State Estimation (SE) consists in estimating the bus voltage methods [1], [5] deal with network constraints, such as zero-
phasors using the power flow equations derived from the injection buses, become computationally expensive as the size
structure of the network, represented by the admittance matrix. of the network increases. This may be a limitation in large
SE is typically performed by taking several measurements and distribution systems when fast SE results may be required
solving a weighted least-squares problem using an iterative every few seconds. In the current paper, we will address theses
approach like Newton-Raphson [1]–[5]. issues.
Typically distribution networks used to have a simple radial In particular, our main contributions can be summarized
structure with a single source bus, the point of common as follows: First, we include network equality constraints
coupling (PCC), connected to the grid and injecting power, and more efficiently by a dimension reduction similar to [16].
therefore SE was not as necessary as in transmission networks. Then, we extend some recently proposed methods that split
This is changing given the increasing number of sources the problem [12], to large-scale, 3-phase coupled, unbalanced
of distributed generation like PV-panels, electrical vehicles, and constrained distribution systems, with synchronized and
batteries, etc. injecting power into the network and causing unsynchronized real-time measurements affected by both mag-
bidirectional power flow. Consequently, SE becomes necessary nitude and angle noises.
in distribution networks. One of the major difficulties for The rest of the paper is structured as follows: Section II
its implementation lies in the structural difference between presents some background about power networks. Section III
transmission and distribution networks: in the latter, there discusses the different types of measurements available for SE.
are coupled phases, unbalanced loads, and lower X/R ratios. Section IV presents the standard methodology for SE. Section
V presents our contributions: it proposes the linear dimension
This project has received funding from the European Union’s Horizon 2020
research and innovation programme under the Marie Skłodowska-Curie grant reduction method and extends recent SE methodologies to split
agreement No 675318 (INCITE). the problem. Section VI defines the test case simulation and
shows its results. Finally, Section VII presents the conclusions physical constraints for the voltage states by defining the
and proposes topics for future work. set of indices of zero-injection buses ε = {i, · · · , j}:
II. D ISTRIBUTION S YSTEM M ODEL (S)ε = 0, (I)ε = 0 (4)
A distribution system consists of buses, where power is where (·)ε denotes the elements at indices in ε.
injected or consumed, and branches, each connecting two 3) Real-time measurements: They can be a range of voltages
buses. This system can be modeled as a graph G = (V, E, W) and currents measurements from PMUs, smart meters, or
with nodes V = {1, ..., Nbus } representing the buses, edges conventional remote terminal units. There are two possi-
E = {(vi , vj ) | vi , vj ∈ V} representing the branches, ble kinds: GPS-synchronized sensors measuring magnitude
and edge weights W = {wi,j | (vi , vj ) ∈ E, wi,j ∈ C} and phase angle, and unsynchronized sensors providing
representing the admittance of a branch, which is determined only magnitude. We model the noises with a low standard
by the length and type of the line cables. deviation for magnitude and angle, σmag ≈ 1% and σang ≈
In 3-phase networks buses may have up to 3 phases, so that 0.01 rad respectively, according to the IEEE standard for
the voltage Vi at bus i lives in C≤3 (and the edge weights PMUs [17].
wi,j ∈ C≤3×≤3 ). The state of the network is then typically Synchronized measurements can be expressed using a lin-
T
represented by the vector bus voltages Vbus = [Vsrc V T ]T ∈ ear approximation (see appendix) with magnitude and an-
N +3
C , where Vsrc ∈ C denotes the known voltage at the
3
gle noise caused by the measurements and by an imperfect
source bus, and V ∈ CN the voltages in the non-source buses, synchronization. For a number NmeasL of measurements
where N depends on the number of buses and phases per bus. zmeasL ∈ CNmeasL we have:
Using the Laplacian matrix Y ∈ C(N +3)×(N +3) of the
weighted graph G, called admittance matrix [1], the power zmeasL ≈ CmeasL V + diag(CmeasL V )(ωmag + jωang ) (5)
flow equations to compute the currents I and the power loads where CmeasL is the matrix mapping state voltages to
S are: measurements, ωmag ∼ N (0, σmag Id,NmeasL ), ωang ∼
      
Isrc Vsrc Y Y Vsrc ¯ N (0, σang Id,NmeasL ). Since σmag and σang are similar [17],
=Y = a b , S = diag(I)V (1)
I V Yc Yd V for simplicity we will use the same value for both: σmeas =
¯ denotes the complex conjugate, diag(·) represents σang = σmag . Then, for measurement i at phase l of bus n
where (·) we have:
the diagonal operator, converting a vector into a diagonal
matrix. Separating Y in blocks according to the indices of (C V ) = (CmeasL )i,• V =
⎧ measL i
the source bus Vsrc , see (1), the voltage V for the non-source ⎪
⎪ V for a voltage measurement
⎨ nl
buses can be rewritten as: (Y )nl ,• V for a current measurement (6)
V = Yd−1 I − Yd−1 Yc Vsrc ⎪
⎪ (Y )nl ,ml (Vnl − Vml ) for a branch-current i → m
⎩ measurement
(2)
V0 = V |I=0 = −Yd−1 Yc Vsrc
where (·)i,• denotes row i.
with V0 denoting the voltage without loads Unsynchronized measurements have a nonlinear relation
III. M EASUREMENTS CmeasNL (V ) with the voltages V . For NmeasNL measure-
ments zmeasNL ∈ RNmeasNL we have:
Several sources of information can be available to solve the
SE problem: zmeasNL = CmeasNL (V ) + diag(CmeasNL (V ))ωmeasNL (7)
1) Pseudo-measurements: They can be load estimations for
with ωmeasNL ∼ N (0, σmeas Id,NmeasNL ). For measurement i at
every hour, based on predictions and known installed load
phase l of bus i we have:
capacity at every bus, and represented by Spsd . Since they
are estimations rather than actual measurements, their noise (CmeasNL (V ))i =
⎧ 
is modeled with a relative large standard deviation (a ⎪ V  for a voltage measurement
⎪
⎨ nl 
typical value can be σpsd ≈ 50% [12]): (Y ) V  for a current measurement
 nl ,• 
⎪ (Y )n ,m (Vn − Vm ) for a branch-current i → m
S = Spsd + Spsd ωpsd = (Ppsd + jQpsd )(1 + ωpsd ) ⎪
⎩ l l l l
measurement
where j is the imaginary unit, ωpsd ∼ N (0, σpsd Id,N ), Id,N (8)
is the identity matrix with size N ×N , and P = {S}, Q = Since the measurement noises in (5) and (7) are small
{S} denote the real and imaginary parts of the power. The according to the PMUs standard [17], their covariance
noise covariance of these measurements is then: matrices can be approximated using the measurements
 2  2 instead of the actual value:
Σpsd = E[ S − Spsd  ] = σpsd
2
diag( Spsd  ) (3) 2
ΣmeasL = E[ |zmeasL − CmeasL V | ]
where E[·] denotes expectation,|·| element-wise magnitudes 2 2
2
= 2σmeas diag(|CmeasL V | ) ≈ 2σmeas
2
diag(|zmeasL | )
of a complex vector, and (·)2 element-wise square.  2
2) Virtual measurements: They are buses with no loads con- ΣmeasNL = E[ zmeasNL − CmeasNL (V ) ]
nected, zero injection buses. They can be modeled as = σmeas
2
diag(CmeasNL (V )2 ) ≈ σmeas
2 2
diag(zmeasNL )
 IV. S TANDARD M ETHODOLOGY FOR S TATE E STIMATION V. P ROPOSED M ETHODS FOR S TATE E STIMATION
In this section we propose new SE methods that significantly
The standard methodology for SE computes the maxi-
improve the standard methodology for (9). First, we present a
mum likelihood estimation by solving a constrained nonlinear
method that simplifies the problem and reduces its complexity,
weighted least-squares problem with all measurements and
while obtaining the same solution; then, we propose another
estimations: (Spsd )εc , zmeasL , zmeasNL , their noise covariance,
method that significantly reduces the computation cost while
and the system constraints (Spsd )ε = 0, where εc denotes
achieving approximately the same accuracy.
the complement of ε within the set {1, . . . , N }. Typically,
this problem is solved using the Newthon-Raphson method A. Weighted least-squares in subspace
[1]. However, since this requires the use of gradients and the Here we propose an alternative to Lagrangian multipliers in
power flow formulas in (1) are not holomorphic, i.e. complex order to include the constraints using a dimension reduction
differentiable, the problem is typically solved in real variables similar to [16]. Therefore, we use a linear transformation that
T
using a polar representation Vpolar = [|V | ∠V T ]T : produces feasible solutions. For convenience, we first define
the space of feasible solutions of (9) as:
min z − h(Vpolar ) W −1
s.t. g(Vpolar ) = 0 (9)
Vpolar
{V | Iε = (Yd )ε V + (Yc )ε,• Vsrc = 0}
(10)
where xA = xT Ax for some A
2
0, z is the vector of = {V | ∃x ∈ CN −|ε| s.t. V = F x + Vp }
measurements in rectangular coordinates:
where (Yd )ε,• ∈ C|ε|×N denotes the rows of Yd corresponding
to indeces in ε; Vp is a solution satisfying Iε = 0, and F is a
z = [(Ppsd )Tεc , (Qpsd )Tεc , {zmeasL }T , {zmeasL }T , zmeasNL
T
]T
basis of the nullspace of (Yd )ε,• :
and h(·) is a nonlinear function mapping Vpolar to measure- (Yd )ε,• F = 0 (11)
ments. The function g(·) indicates the zero loads: g(Vpolar ) =
[(Ppsd )Tε , (Qpsd )Tε ]T , and the weight matrix W −1 is the in- Therefore, V0 is chosen for Vp , and F is built using an
verse of the measurement noises covariances in rectangular orthonormal basis of the kernel of (Yd )ε,• e.g. by computing
variables: the singular value decomposition of (Yd )ε,• = U SV ∗ , where
(·)∗ denotes complex conjugate transpose, and taking the
W = diag(Σrect,psd , Σrect,measL , ΣmeasNL ) columns of V with zero singular value. Then, F = ker((Yd )ε,• )
and F ∗ F = Id , where ker(·) denotes the kernel subspace.
where here we use the function diag(·) to denote the oper- If we consider the rectangular representation of voltages
ator converting several matrices into a single block-diagonal Vrect = [{V }T {V }T ]T instead of the polar representation
matrix, and the suffix (·)rect stands for rectangular coordinates: Vpolar , the subspace equations can still be represented linearly
in real variables:
2 2
Σrect,measL ≈ σmeas
2
diag(|zmeasL | ,|zmeasL | )) {Vrect | {Iε } = 0, {Iε } = 0}
   
(Ppsd )2εc (Ppsd Qpsd )εc 
Σrect,psd = σpsd
2  {(Yd )ε,• } −{(Yd )ε,• }
diag
(Ppsd Qpsd )εc (Qpsd )2εc = Vrect  V
 {(Yd )ε,• } {(Yd )ε,• } rect
  (12)
{(Yc )ε,• Vsrc }
+ =0
In the literature [1], [3], the constraints in (9) are typically {(Yc )ε,• Vsrc }
included in two ways:
= {Vrect | ∃x̃ ∈ C2N −2|ε| s.t. Vrect = F̃ x̃ + Ṽp }
1) either using Lagrangian multipliers λ, which increases
T
the size of the optimization variables to [Vpolar λT ] T ∈ where now
 
R2N +2|ε| ; {V0 }
Ṽp = Vrect,0 =
2) or including the equality constraints as measurements with {V0 }
a noise with a very small standard deviation, so that their  
{(Yd )ε,• } −{(Yd )ε,• }
corresponding weights become very large with respect F̃ = ker
{(Yd )ε,• } {(Yd )ε,• }
to other measurements. However, this produces an ill-
conditioned weights matrix, which may cause numerical Then problem (9) becomes:
problems.  
In order to efficiently solve either 1) or 2), different methods F̃ arg min z − h̃(F̃ x̃ + Ṽp ) + Vrect,0 (13)
x̃ W −1
like the orthogonal, hybrid, or Hachtel methods are typically
used [1], [3], [5], [18], [19]. However, these iterative ap- where h̃(·) is like h(·) but in rectangular coordinates, so
proaches increase the computation time, which may compro- that h̃(Vrect ) = h(Vpolar ). This again can be solved using the
mise the use of these SE techniques for real-time operation Newton-Raphson method with iterations:
of large distribution networks. Therefore, we propose a more
x̃k+1 = x̃k + Δx̃k (14)
computationally efficient method appropriate to the problem.
where Remark 1. The estimated current at any iteration k is
(HkT W −1 Hk )Δx̃k = HkT W −1 (z − h̃(F̃ x̃k + Ṽp )) Ik = (diag(V̄k−1 ))−1 S̄ (19)
(15)
Hk = ∇x̃ h̃(F̃ x̃ + Ṽp ) |x̃k = ∇Vrect h̃(Vrect ) |Vrect,k F̃
If the loads satisfy (S)ε = 0, then (Ik )ε = 0. This means
With the method in (13), we have eliminated the equality that there exists a vector x ∈ CN−|ε| s.t. Vk = F x + V0
constraints by embedding the solution into a smaller subspace and thus Vk is feasible for all k with respect to (10).
of feasible solutions. This simplifies the method and reduces Remark 2. Method b) can be extended to real variables
its computational cost with respect to using Lagrangian mul- to get Σprior in rectangular coordinates:
tipliers or large weights [1], [5], since we reduce the size of  
T
the optimization variables from [Vpolar λT ]T ∈ R2N +2|ε| or {Spsd }
Vrect,k+1 = Bk + Vrect,0 (20)
Vpolar ∈ R 2N
respectively, to x̃ ∈ R 2N −2|ε|
. Additionally, the {Spsd }
gain matrix (HkT W −1 Hk ) will not be ill-conditioned, since with
all weights have comparable magnitudes; and F ,F̃ can be     −1
{Yd−1 } −{Yd−1 }
2
computed offline. Moreover, with method (13) the constraints |Vk |
Bk = diag
will be satisfied by construction, which in not guaranteed if {Yd−1 } {Yd−1 } |Vk |
2

using large weights. Oppositely to [16], we keep the rectangu-  

diag({Vk }) diag({Vk })
lar representation, which keeps the transformation linear, and ·
choose a different representation of the subspace using the diag({Vk }) −diag({Vk })
orthonormal bases F , F̃ , in order to compare the method with and estimation error covariance:
the later ones in (22), (24) and (29), when transforming the  
{Spsd }2
covariance matrices Σprior and Σpost . Σrect,prior = B0 σpsd diag
2
B0T (21)
{Spsd }2
B. Power flow solution plus optimal linear update
Using Newthon-Raphson for (9) and (13) requires to solve The advantage of approach b) with respect to a) is that b) can
in each iteration the linear system (15) with approximately be solved in complex variables. In both cases, since the load
same size N as the network. Consequently, this can become estimates Spsd are known beforehand, this part of the problem
computationally expensive when applied to large distribution can be computed offline, taking this iterative computational
systems. Therefore, we propose an alternative method that cost outside the real-time operation of the system.
splits the problem into two parts as in [12]: first, we use 2) Updating with only synchronized measurements:
only the constraints and the load estimates Spsd available If there are not any unsynchronized measurements, the map-
beforehand to solve the power flow problem offline and to ping from the state variables to the real-time measurements (6)
obtain a prior solution defined as Vprior ; then, we update in real- is linear and thus holomorphic in x ∈ CN −|ε| . Consequently,
time the solution using the measurements zmeasL , zmeasNL to the update can be solved using complex variables to obtain a
compute a better estimate, represented by Vpost . We also prove linear expression. Different alternative methods are possible;
that two alternatives for the update, the minimum-variance here we present two of them and later show their equivalence:
and the maximum-likelihood estimator, are equal under certain a) If the prior solution Vprior is feasible, then there exists a
conditions. vector xprior s.t. Vprior = F xprior + V0 . It can be computed
1) Obtaining a prior solution: as xprior = F ∗ (Vprior −V0 ), with estimation error covariance
a) The prior estimate can be computed using the pre- Σxprior = F ∗ Σprior F , or xrect,prior = F̃ T (Vrect,prior −Vrect,0 ) for
vious method in (13) without real-time measure- real variables. Then the maximum likelihood update can be
ments: z = [(Ppsd )Tεc (Qpsd )Tεc ]T and h̃(Vrect ) = computed using weighted least-squares:
[Pεc (Vrect )T Qεc (Vrect )T ]T , where P (Vrect ) denotes P as   2
a function of Vrect . This method will estimate real and x − xprior
Vpost = V0 +F arg min
imaginary parts Vrect,prior , with estimation error covariance: x CmeasL (F x + V0 ) − zmeasL −1
Wpost
Σrect,prior ≈ F̃ (HkT W −1 Hk )−1 F̃ T (16) (22)
−1
where Wpost = (diag(F ∗ Σprior F, ΣmeasL ))−1 is the weight
b) An alternative approach using complex variables would be matrix. Taking derivatives at (22) we obtain the solution:
the method proposed in [12], after extending it to a 3-phase
coupled, unbalanced system. Using the voltage under no Vpost = F ((F ∗ Σprior F )−1 + F ∗ CmeasL
∗
Σ−1
measL CmeasL F )
−1

loads V0 as initial value in an iterative method inspired by ·((F ∗ Σprior F )−1 xprior + F ∗ CmeasL
∗
Σ−1
measL
(2), for the k-th iteration we have: ·(zmeasL − CmeasL V0 )) + V0
Vk+1 = Yd−1 (diag(V̄k ))−1 S̄ + V0 (17) (23)
Remark 3. Considering the first-order (and thus linear)
The estimation error covariance can be approximated by
approximation of the measurement function h̃(·), in [13] it
the covariance after the first iteration, which corresponds
is proven that solving the weighted least-squares problem
to the linear approximation in [20]:
in one step (13) and in two steps (i.e. weighted least-
Σprior ≈ Yd−1 (diag(V̄0 ))−1 Σpsd (diag(V0 ))−1 (Yd−1 )∗ (18) squares for power flow plus update (22)), produce exactly
 the same solution. Consequently, we can expect a similar 3) Updating with unsynchronized measurements:
accuracy between the one-step and the two-steps methods If there are nonlinear measurements, since the magnitude of a
when applied to the original nonlinear problem. complex number as in (8) is not a holomorphic function, the
b) The alternative approach proposed in [12] and extended update step 24 needs to be solved using real variables:
here, is to improve the solution using a minimum-variance ⎡ ⎤
{zmeasL − CmeasL V (Vrect,prior )}
linear update: Vrect,post = Vrect,prior + K̃ ⎣ {zmeasL − CmeasL V (Vrect,prior )} ⎦
  zmeasNL − CmeasNL (V (Vrect,prior ))
Vpost = Vprior + K zmeasL − CmeasL Vprior (24)
(29)
Then the error covariance of estimation Vpost is: Given that
   
{CmeasL V (Vrect,prior )} {CmeasL } −{CmeasL }
∗
Σpost = Σprior + K(ΣmeasL + CmeasL Σprior CmeasL )K ∗ = V
(25) {CmeasL V (Vrect,prior )} {CmeasL } {CmeasL } rect,prior
∗ ∗
−KCmeasL Σprior − Σprior CmeasL K
and defining
and the optimal gain K can be computed minimizing the ⎡   ⎤
expected error E[(Vpost − V )∗ (Vpost − V )] = tr(Σpost ): {CmeasL } −{CmeasL }
CmeasLNL = ⎣ {CmeasL } {CmeasL } ⎦
∗
K = Σprior CmeasL ∗
(CmeasL Σprior CmeasL + ΣmeasL )−1 (26) ∇Vrect,prior CmeasNL (V (Vrect )) |Vrect,prior
ΣmeasLNL = diag(Σrect,measL , ΣmeasNL )
Proposition 1. If Vprior is a feasible solution to (10), as the
solutions in Section V-B1, then the solution (24) satisfies (10). the first-order approximation of the estimation error covariance
of this nonlinear filter is:
Proof. For Vprior : (Yd )ε,• Vprior +(Yc )ε,• Vsrc = 0. Also, since the T
first term in K is F , because K starts with Σprior = F Σxprior F ∗ , Σrect,post ≈ K̃(ΣmeasLNL + CmeasLNL Σrect,prior CmeasLNL )K̃ T
T
we have (Yd )ε,• K = 0 due to (11). Therefore, Vpost from (24) +Σrect,prior − K̃CmeasLNL Σrect,prior − Σrect,prior CmeasLNL K̃ T
satisfies (10).  (30)
where now the optimal gain K̃ is:
Proposition 2. If Vprior is a feasible solution to (10), then both T T
alternatives (23) and (24) are equal. K̃ = Σrect,prior CmeasLNL (CmeasLNL Σrect,prior CmeasLNL +ΣmeasLNL )−1

Proof. We start from expression (23), using Woodbury’s iden- Remark 4. As in Proposition 2, it can be proven that the
tity [21] on the inverse term and (26) for K we get: first-order approximation of the estimation error covariance
(30) of (29) is equal to the one approximating the estimation
Vpost = F ((F ∗ Σprior F ) − (F ∗ Σprior F )F ∗ CmeasL
∗
error covariance of (22) including nonlinear measurements,
·(CmeasL Σprior CmeasL + ΣmeasL ) CmeasL F (F ∗ Σprior F ))
∗ −1 computed similarly as in (16).
·((F ∗ Σprior F )−1 xprior + F ∗ CmeasL
∗
Σ−1
measL VI. C ASE S TUDY
·(zmeasL − CmeasL V0 )) + V0 A simulation of 24 hours with 15 min intervals is run on
= (F (F ∗ Σprior F ) − KCmeasL F (F ∗ Σprior F )) a test case to compare the different methods for SE. Here we
·((F ∗ Σprior F )−1 xprior + F ∗ CmeasL
∗
Σ−1
measL describe the settings of the test case and analyze its results.
·(zmeasL − CmeasL V0 )) + V0 These algorithms are coded in Python and run on an Intel Core
i7-5600U CPU at 2.60GHz with 16GB of RAM.
= V0 + (Id − KCmeasL )F xprior
+(Id − KCmeasL )F (F ∗ Σprior F )F ∗ CmeasL
∗
Σ−1
measL
A. Settings
·(zmeasL − CmeasL V0 )) • System: We use the 123-bus test feeder [22] available online
(27) [23], see Fig. 1. This is a challenging example, since it is 3-
Doing some manipulation to the formula of K in (26) we get: phase coupled, unbalanced, and larger than other examples
in the literature [10], [12].
∗
(Id − KCmeasL )Σprior CmeasL Σ−1
measL = K (28) • Measurements (see Fig. 1): Voltage measurements (red cir-
cle for phasor, red square for magnitude only) are placed at
So finally we can convert the expression in (23) to (24):
buses 79, 95, 83 and 300, current measurements (blue dashed
Vpost = V0 + (Id − KCmeasL )F xprior + K(zmeasL − CmeasL V0 ) circle for phasor, blue dashed square for magnitude only) at
= Vprior + K(zmeasL − CmeasL Vprior )  buses 65 and 48, and branch current phasor measurements
(blue dashed arrow) at branch 150 (after the regulator)
The maximum-likelihood estimator of the weighted least- → 149. The standard deviation used to simulate noise
squares solution coincides with the minimum-variance esti- in the measurements is σmeas = 0.01 according to the
mator of the linear update, since we are considering Gaussian standards in [17]. This sensors locations are chosen since
distributions for the noise. Therefore, it is the same to use they correspond to nodes with big loads and/or towards the
one method or the other, when considering only synchronized end of a feeder, or at a branch transporting a large current.
measurements. A more extensive method to obtain optimal locations is not
 Figure 2: Load profiles are built aggregating over 50 house-
holds profiles from the dataset in [24].

Moreover, Fig. 3b shows that the two-step solution Vpost

performs statistically as well as all weighted least-squares
Figure 1: 123-bus test feeder with measurements location. The solutions VWLSHa , VWLSOr , VWLSsub , with a minimal decrease in
network image has been taken from [23]. accuracy, especially when compare to that of Vprior . However,
the methods using weighted least-squares are computationally
much more expensive than the two step method Vpost , see Fig.
within the scope of this paper and will be addressed in future
4.
work.
To summarize, the solutions including real-time measure-
• Load Profiles: They are built aggregating over 50 households
ments clearly improve the accuracy of the estimation. How-
profiles with maximum loads above 1 kW from the dataset
ever, the two-step approach has a much smaller computational
available in the DiSC simulation framework [24]. The values
cost and is more scalable to bigger networks for real-time
of Spsd are an average over all households multiplied by 50.
operation.
This way the loads are similar to the base loads provided
in the 123-bus test feeder, and the relative load standard VII. C ONCLUSIONS
deviation σpsd is approximately 50%. The load profiles can We have proposed a methodology for state estimation in
be seen in Fig. 2. large-scale, 3-phase coupled, unbalanced distribution systems,
• Nomenclature: Vprior denotes the prior solution (17); VWLSHa with mixed measurements: phasor-synchronized (linear) and
and VWLSOr are the traditional Hachtel and Orthogonal solu- magnitude-unsynchronized (nonlinear), zero load injection
tions of (9) based on weighted least squares [1]; VWLSsub is buses in the form of constraints, and different sources of
the weighted least-squares solution using the subspace (13); noise: magnitude and angle. We have shown that the proposed
and Vpost is the posterior estimate (29). method is as accurate as standard methods, but computation-
• Accuracy metric: The root mean square error (RMSE) and ally more efficient in real-time operation. The computational
the maximum absolute error (MaxAE) are determined for cost reduction is achieved by two factors: first using a di-
the voltages estimated by every method: mensionality reduction to the subspace of feasible solution,

N  2 and second splitting the problem into a two-step offline and
RMSEt,methodm = N1 i=1 Vmethodm ,t,i − Vt,i 
  online problem. The iterative offline method estimates the prior
maxAEt,methodm = maxi Vmethodm ,t,i − Vt,i  solutions, taking most of the computational cost, while the
where Vt,i , Vmethodm ,t,i denote the value of the i-element, at online problem updates this solution in a single step, without
time step t of the simulation, of the actual voltage and the requiring further iterations.
voltage estimated by method m respectively. Future work could include adding different sources of
distributed generation and investigating how they affect the
B. Results solutions; developing methodologies to compute the mini-
Fig. 3a shows how a solution with real-time measurements mum number and optimal allocation of measurement units to
like Vpost clearly outperforms the prior solution Vprior using guarantee a given accuracy level; and adding dynamic state
only load estimations. This is relevant in this case, because equations to exploit the historical sensor information.
the Vprior may produce an error of almost 0.05 pu and thus A PPENDIX
is unable to monitor network constraint violations, while the
A PMU measurement of a complex number u = |u| ejθu
estimation Vpost produces errors of at most 0.01 pu. This
with magnitude and angle noise can be expressed as:
shows that even when few measurements are available (in
this case only 7 measurements in a 123-bus network), using ũ = (|u| + ωmag )ej(θu +ωθ )
them can dramatically increase the accuracy of the prediction. = (|u| + ωmag )(cos(θu + ωθ ) + j sin(θu + ωθ ))
 [8] V. Madani, M. Parashar, J. Giri, S. Durbha, F. Rahmatian, D. Day,
M. Adamiak, and G. Sheble, “PMU placement considerations - A
roadmap for optimal PMU placement,” in Power Systems Conference
and Exposition (PSCE). IEEE, 2011, pp. 1–7.
[9] M. E. Baran and A. W. Kelley, “State estimation for real-time monitoring
of distribution systems,” IEEE Transactions on Power Systems, vol. 9,
no. 3, pp. 1601–1609, 1994.
[10] M. B. A. Kelley, “A branch current based state estimation method for
distribution systems,” IEEE Transactions on Power Systems, vol. 10, pp.
483–491, 1995.
[11] W.-M. Lin, J.-H. Teng, and S.-J. Chen, “A highly efficient algorithm in
treating current measurements for the branch-current-based distribution
state estimation,” IEEE Transactions on Power Delivery, vol. 16, no. 3,
pp. 433–439, 2001.
[12] L. Schenato, G. Barchi, D. Macii, R. Arghandeh, K. Poolla, and
A. V. Meier, “Bayesian linear state estimation using smart meters and
pmus measurements in distribution grids,” in 2014 IEEE International
Conference on Smart Grid Communications (SmartGridComm), Nov
(a) (b) 2014, pp. 572–577.
Figure 3: Box-plots of RMSE and MaxAE for the solutions. [13] M. Zhou, V. A. Centeno, J. S. Thorp, and A. G. Phadke, “An alternative
for including phasor measurements in state estimators,” IEEE Transac-
The red line indicates the median, the red square the mean. tions on Power Systems, vol. 21, no. 4, pp. 1930–1937, 2006.
[14] Y. Hu, A. Kuh, T. Yang, and A. Kavcic, “A belief propagation based
power distribution system state estimator,” IEEE Computational Intelli-
gence Magazine, vol. 6, no. 3, pp. 36–46, 2011.
[15] S. Bolognani, R. Carli, and M. Todescato, “State estimation in power
distribution networks with poorly synchronized measurements,” in 2014
IEEE 53rd Annual Conference on Decision and Control. IEEE, 2014,
pp. 2579–2584.
[16] Y. Guo, W. Wu, B. Zhang, and H. Sun, “An efficient state estimation
algorithm considering zero injection constraints,” IEEE Transactions on
Power Systems, vol. 28, no. 3, pp. 2651–2659, 2013.
[17] K. Martin, D. Hamai, M. Adamiak, S. Anderson, M. Begovic, G. Ben-
mouyal, G. Brunello, J. Burger, J. Cai, B. Dickerson et al., “Exploring
the IEEE standard C37. 118–2005 synchrophasors for power systems,”
IEEE Transactions on Power Delivery, vol. 23, no. 4, pp. 1805–1811,
2008.
[18] G. N. Korres, “A robust algorithm for power system state estimation
with equality constraints,” IEEE Transactions on Power Systems, vol. 25,
Figure 4: Box-plots of execution times for the solutions. The no. 3, pp. 1531–1541, 2010.
red line indicates the median, the red square the mean. [19] A. Simoes-Costa and V. Quintana, “A robust numerical technique for
power system state estimation,” IEEE Transactions on Power Apparatus
and Systems, no. 2, pp. 691–698, 1981.
ω [20] S. Bolognani and S. Zampieri, “On the existence and linear approxima-
where ωmag ∼ N (0, 1%|u|) (or ω̃mag = |u| mag
∼ N (0, 0.01)), tion of the power flow solution in power distribution networks,” IEEE
ωθ ∼ N (0, 0.01 rad), according to the standards in [17]. Using Transactions on Power Systems, vol. 31, no. 1, pp. 163–172, 2016.
[21] M. A. Woodbury, “Inverting modified matrices,” Statistical Research
trigonometric identities, considering that |ωθ | 1 so that Group Memorandum Reports. Princeton, NJ: Princeton University,
sin(ωθ ) ≈ ωθ , cos(ωθ ) ≈ 1, and neglecting second-order noise no. 42, 1950.
terms, we obtain: [22] W. Kersting, “Radial distribution test feeders,” IEEE Transactions on
Power Systems, vol. 6, no. 3, pp. 975–985, 1991.
ũ ≈|u| (cos(θu ) + j sin(θu )) + ωmag (cos(θu ) + j sin(θu )) [23] D. T. Feeders, “IEEE PES Distribution Systems
+ωθ |u| (− sin(θu ) + j cos(θu )) 
Analysis Subcommittee’s Radial Test Feeders,”
ewh.ieee.org/soc/pes/dsacom/testfeeders.html, 1991, [Online]. Accessed:
≈ u + ω̃mag u + jωθ u = u + u ω̃mag + jωθ 2016-10-20.
[24] R. Pedersen, C. Sloth, G. B. Andresen, and R. Wisniewski, “DiSC:
R EFERENCES A simulation framework for distribution system voltage control,” in
[1] A. Abur and A. G. Exposito, Power System State Estimation: Theory European Control Conference (ECC), 2015, pp. 1056–1063.
and Implementation. CRC Press, 2004.
[2] Y.-F. Huang, S. Werner, J. Huang, N. Kashyap, and V. Gupta, “State
estimation in electric power grids: Meeting new challenges presented by
the requirements of the future grid,” IEEE Signal Processing Magazine,
vol. 29, no. 5, pp. 33–43, 2012.
[3] A. Monticelli, “Electric power system state estimation,” Proceedings of
the IEEE, vol. 88, no. 2, pp. 262–282, 2000.
[4] B. Hayes and M. Prodanovic, “State estimation techniques for electric
power distribution systems,” in 2014 IEEE European Modelling Sympo-
sium (EMS), 2014, pp. 303–308.
[5] L. Holten, A. Gjelsvik, S. Aam, F. F. Wu, and W.-H. Liu, “Comparison
of different methods for state estimation,” IEEE Transactions on Power
Systems, vol. 3, no. 4, pp. 1798–1806, 1988.
[6] B. Stott and O. Alsaç, “Fast decoupled load flow,” IEEE Transactions
on Power Apparatus and Systems, no. 3, pp. 859–869, 1974.
[7] A. Garcia, A. Monticelli, and P. Abreu, “Fast decoupled state estimation
and bad data processing,” IEEE Transactions on Power Apparatus and
Systems, no. 5, pp. 1645–1652, 1979.