Real-Time Distribution Grid State Estimation with Limited

Sensors and Load Forecasting

Roel Dobbe Daniel Arnold Stephan Liu

Electrical Engineering & Grid Integration Group, Electrical Engineering &
Computer Sciences, Lawrence Berkeley National Computer Sciences,
UC Berkeley Laboratory UC Berkeley
Berkeley, United States Berkeley, United States Berkeley, United States
[email protected] [email protected] [email protected]
Duncan Callaway Claire Tomlin
Energy & Resources Group, Electrical Engineering &
UC Berkeley Computer Sciences,
Berkeley, United States UC Berkeley
[email protected] Berkeley, United States
[email protected]

ABSTRACT 1. INTRODUCTION
High penetration levels of distributed generation (DG)
and electric vehicles (EVs) diversify power flow and bring
uncertainty to distribution networks, making planning
1.1 Motivation
and control more involved for distribution system oper- Electric distribution grids deliver power to end cus-
ators (DSOs). The increased risk of constraint violation tomers, by connecting to the high voltage transmis-
triggers the need to augment forecasts with real-time sion grid at a substation and stepping down voltage to
state estimation. This is economically and technically medium and service levels. In distribution networks,
challenging since it requires investing in a large num- load forecasts and modeling were traditionally sufficient
ber of sensors and these have to communicate with of- to inform the planning of infrastructure updates and the
ten older and slower supervisory control and data ac- scheduling of control equipment. Rapid diversification
quisition (SCADA) systems. We address distribution of power flow now challenges this approach in two ways.
grid state estimation via combining only a limited set of First, the occurrence of bidirectional and intermittent
sensors with load forecast information. It revisits open power flow due to DG from wind and photovoltaics (PV)
problems in a recent paper that proposes a Bayesian es- causes rapid voltage fluctuations. These are increas-
timation scheme. We derive the estimator for balanced ingly harder to predict, and can result in protection is-
power networks via rigorous modeling. An off-line anal- sues, such as desensitization and unintended islanding
ysis of load aggregation, forecast accuracy and number or tripping [11]. This can lead to accelerated structural
of sensors provides concrete engineering trade-offs to de- damage and potentially cascading failures, and yields
termine the optimal number of sensors for a desired ac- economic burden due to accelerated wear [14]. Espe-
curacy. This estimation procedure can be used in real cially if DG is connected to more sensitive feeders with-
time as an observer for control problems or off-line for out proper monitoring functionality to assess its effect,
planning purposes to asses the effect of DG or EVs on cost of integration can easily multiply by a factor 3 to
specific network components. 4 [2,7]. The rapid adoption of EVs will further aggravate
this situation [10], especially if charging is optimized for
electricity prices [16]. In addition to challenges in oper-
ation, the inability to assess the impact of DG and EVs
on the physical network causes utilities to impose con-
servative caps on the allowable PV capacity and num-
ber of EVs, hindering the transition to renewable energy
sources. These concerns have mobilized many DSOs to
build a stronger information layer on top of their physi-
cal infrastructure that exploits recent advances in sens-

978-1-5090-1772-0/16/$31.00 ©2016 IEEE

ing and communication to enable both forecasting and tructure, which can lead to packet failure and highly
real-time state estimation. unreliable state estimates.
In transmission systems, the need for system reliabil- Another set of papers considers the incorporation of
ity and economies of scale have long motivated the de- load forecasts. In the conventional setting, we saw [6,12]
velopment of these methods [3,8]. Traditional state esti- augmenting the fully observed measurement vector with
mation relies on an overdetermined formulation for the load forecasts. A more recent work by Schenato et. al
unknown/unmeasured variables to be observable. This studies the use of PMUs in distribution system state es-
means that the number of available measurements must timation. It takes a Bayesian approach, in which load
be greater than or equal to the number of unknowns (to statistics are used to determine a forecast 𝑥ˆ of the state
be estimated). Most methods rely on weighted least- variables [13]. The paper shows its accuracy is compara-
squares or least absolute value approaches [3]. In recent ble to that of the conventional WLS estimators, and es-
years, the wide adoption of phasor measurement units timation error confidence intervals can be computed off-
(PMUs) has led to state estimators that can provide in- line, allowing for engineering trade-offs between number
formation at a faster time resolution allowing the real- of sensors and estimation accuracy.
time assessment of dynamics in these networks [17, 19].
Applying these methods to distribution systems is 1.3 Overview and Contributions
conceptually similar, but technically and economically In this work, we consider distribution grid state es-
a more challenging task. Firstly, forecasting smaller ag- timation using a limited number of sensors and load
gregations of loads is more challenging, leading to errors forecast information. Our objective is to estimate the
especially in areas where power flow can change quickly voltage magnitude and angle at all buses in the network.
due to intermittent DG. Secondly, ensuring observabil- We develop an estimator for balanced radial networks,
ity requires equipping most key buses in a network with based on the DistFlow equations [4] augmented with
sensors, which will likely remain an infeasible invest- an equation to incorporate voltage angle measurements,
ment for most DSOs. This leads to the question of how allowing for the use of PMU sensors. We show how
to increase observability in distribution systems given minimum mean square estimation (MMSE) can be used
more uncertain forecasts and a limited number of real- to neatly integrate real-time measurements and forecast
time measurements. information. We find that the estimator minimizing a
quadratic cost function is the linear least squares esti-
1.2 Previous Work mator (LLSE) [18]. Figure 1 depicts the approach we
A key challenge in distribution grids is to estimate follow to construct a distribution state estimator.
an 𝑛-dimensional state vector in scenarios where only a The algorithm starts with a forecast of the nodal net
limited set of 𝑀 sensors is available, i.e. 𝑀 < 𝑛, which load vector ŝ𝑛 , and the covariance matrix Σ𝑠 . Via the
does not satisfy the requirements for conventional state power flow model a forecast can be constructed for both
estimation. As a result, the standard estimation prob- the measured quantities 𝑧ˆ and the non-measured quan-
lem is underdetermined (𝑀 < 𝑛) and hence ill-posed tities 𝑥
ˆ. The LLSE now compares the actual measure-
from a computational point of view. In practice, this ment of 𝑧 with the forecasted value 𝑧ˆ. A deviation (𝑧−ˆ
𝑧)
means that the state vector 𝑥 is not observable. Some from the forecasted value is an innovation and triggers
recent works have addressed this issue by introducing the Bayesian estimator to proposes an update for the
the use of so-called pseudo-measurements. These are forecasted estimator 𝑥 ˆ [18].
typically calculated using short-term load forecasts or
Network Real-time
historical data, and tend to be less accurate than real- model measurements
time measurements. Forecast of Bayesian
quantities
In [9], Göl and Abur propose a hybrid estimator that Load forecast
Modeling Estimation
estimate

combines a limited number of phasor measurement units Second order

statistics
(PMUs) with a high refresh rate (at the order of 30 Hz),
and a fully observed (𝑀 ≥ 𝑛) set of SCADA measure-
ments at slower refresh rate (order of 5 to 15 min). The Computed a priori Computed in real-time

weighted least absolute value method used helps to ro-

bustify the the estimate between each SCADA update, Figure 1: Schematic overview of Distribution Grid State
but does not address a scenario where a limited set of Estimator.
measurements is available. Furthermore, the estimator
is designed for transmission systems that can rely on The authors of [13] already made an important step in
robust communication networks. SCADA in distribu- the direction of enabling underdetermined distribution
tion systems often lack a reliable communication infras- grid state estimation. In this work we take a different
modeling approach and address four of their open chal- are given by
lenges. Firstly, the linear approximation used to express ∑
the state variables as a function of the loads is derived 𝑃𝑖𝑗 = 𝑃𝑗𝑘 + 𝑟𝑖𝑗 ℓ𝑖𝑗 + 𝑝𝑐𝑗 − 𝑝𝑔𝑗
from stopping a numerical algorithm to solve power flow 𝑘∈𝒞𝑗
∑
equations. This approach falls short of addressing the 𝑄𝑖𝑗 = 𝑄𝑗𝑘 + 𝑥𝑖𝑗 ℓ𝑖𝑗 + 𝑞𝑗𝑐 − 𝑞𝑗𝑔 (1)
needs expressed by utilities to understand the origin and 𝑘∈𝒞𝑗
accuracy of approximations and complicates the gener- 𝑣𝑗2 = 𝑣𝑖2 − 2
2(𝑟𝑖𝑗 𝑃𝑖𝑗 + 𝑥𝑖𝑗 𝑄𝑖𝑗 ) + (𝑟𝑖𝑗 + 𝑥2𝑖𝑗 )ℓ𝑖𝑗
alizability to unbalanced networks. We therefore focus
on first principles, and derive an estimator based on where 𝑗 ∈ 𝒞𝑖 , 𝑟𝑖𝑗 and 𝑥𝑖𝑗 are the resistance and reac-
power flow modeling. In an extended work we plan to tance of the line between nodes 𝑖 and 𝑗, 𝑃𝑖𝑗 , 𝑄𝑖𝑗 ∈ ℝ
show that this allows us to extend the formulation to denote the real and reactive power flowing on a branch
3-phase unbalanced networks. Secondly, by design, the out of node 𝑖 (towards node 𝑗), 𝑝𝑐𝑗 , 𝑞𝑗𝑐 ∈ ℝ are the real
Bayesian prior on the voltage phasor is a function of the and reactive power consumption at node 𝑗, 𝑝𝑔𝑗 , 𝑞𝑗𝑔 ∈ ℝ
feeder head nominal voltage. The nominal voltage hence the real and reactive generation at node 𝑗, 𝑣𝑖 is the volt-
needs to be measured to scale the priors before feeding 𝑝2 +𝑞 2
age magnitude at node 𝑖, and ℓ𝑖𝑗 = 𝑖𝑗𝑣2 𝑖𝑗 is the squared
it to the Bayesian estimator. In this work, this scal- 𝑖
magnitude of the current on branch (𝑖, 𝑗). In our set-
ing dependency is eliminated by formulating a Bayesian
ting we might also have access to the voltage angle 𝛿𝑖 .
prior on voltage phasor differences rather than on the
The nonlinear equation for voltage angle difference over
absolute phasor. This allows for the algorithm to be im-
a branch is given by
plemented in a distributed fashion, i.e. you could build
separate estimators for different (connected) parts of a −𝑥𝑖𝑗 𝑃𝑖𝑗 + 𝑟𝑖𝑗 𝑄𝑖𝑗
network without having to consult the feeder head volt- sin(𝛿𝑗 − 𝛿𝑖 ) = (2)
𝑣𝑖 𝑣𝑗
age. Thirdly, the estimator in [13] models the voltage
and current phasor in rectangular form, i.e. the complex We propose the approximation that assumes that ℓ𝑖𝑗 ≈
√ 𝑖 is formulated as 𝕍𝑖 = Re(𝕍)+ 𝑗Im(𝕍),
voltage at node 0 and the voltage magnitudes 𝑣𝑖 ≈ 1 p.u.. If we also
where 𝑗 = −1. Since available synchrophasors typi- define 𝑦𝑖 ≜ 12 𝑣𝑖2 , we end up with the linear equations
cally report in polar form, i.e. 𝕍𝑖 = 𝑉𝑖 ∠𝛿𝑖 , a nonlinear
transformation is needed to adjust sensor readings be- ∑
𝑃𝑖𝑗 ≈ 𝑃𝑗𝑘 + 𝑝𝑐𝑗 − 𝑝𝑔𝑗
fore feeding these to the estimator: Re(𝕍) = 𝑉𝑖 cos 𝛿𝑖 𝑘∈𝒞𝑗
and Im(𝕍) = 𝑉𝑖 sin 𝛿𝑖 . This can lead to undesirable ∑
𝑄𝑖𝑗 ≈ 𝑄𝑗𝑘 + 𝑞𝑗𝑐 − 𝑞𝑗𝑔 (3)
magnification of measurement errors in the voltage an-
𝑘∈𝒞𝑗
gel 𝛿𝑖 . Our method formulates the problem in polar
𝑦𝑖 − 𝑦𝑗 ≈ 𝑟𝑖𝑗 𝑃𝑖𝑗 + 𝑥𝑖𝑗 𝑄𝑖𝑗
form, bypassing this difficulty. Lastly, in this paper we
𝛿𝑖 − 𝛿𝑗 ≈ 𝑥𝑖𝑗 𝑃𝑖𝑗 − 𝑟𝑖𝑗 𝑄𝑖𝑗
extend the analysis by considering the effect of aggre-
gation, number of sensors and forecast accuracy. We Notice that these are the linearized DistFlow equations
show that the estimation error is nonlinear in the level [4, 5], augmented with an update equation for the volt-
of aggregation. We provide a number of important engi- age angle 𝛿𝑘 . The latter equation was studied more
neering trade-offs based on how much improvement the recently motivated by the development of PMUs for dis-
estimator offers compared to forecasted values, giving tribution grids [17]. A radial network can be modeled
insight into the investment needed to achieve a desired as a set 𝒱 of 𝑁 buses, connected by a set ℰ of 𝑁 − 1
level of estimation accuracy. wires. The network is equipped with 𝑀 synchrophasors
at the subset of nodes ℳ ⊆ 𝒱. Each synchrophasor can
measure the voltage magnitude 𝑣𝑖 and phase angle 𝛿𝑖 at
𝑖 ∈ ℳ.
For system analytical purposes, we will consider two
2. POWER FLOW MODELING different network models. First, we will focus on a line
Consider a distribution feeder with a radial network feeder, i.e. network with buses lined up in one lat-
topology represented by a graph 𝒢 ≜ (𝒱, ℰ), defined by eral. Each node has one upstream and one downstream
a set of buses or nodes 𝒱 (a bus or node 𝑖 is in the set branch. This topology allows us to assess the accuracy
{0, 1, . . . , 𝑁 }, i.e. ∣𝒱∣ = 𝑁 + 1), and a set of lines or of the estimator for a growing number of sensors (or
edges ℰ. We consider radial (tree) topologies, so that sensor-to-sensor distance) and load aggregation. Sec-
we can define 𝒞𝑖 ≜ {𝑗 ∈ 𝒱 : 𝑖 > 𝑗, (𝑖, 𝑗) ∈ ℰ} to be the ond, we use the insights gained about sensor placement
set of child nodes connected downstream to node 𝑖. to set up an analysis on the IEEE radial feeder with
The well studied and popular DistFlow equations [4] 37 buses. For the line feeder, the extended DistFlow
equations presented in (3) simplify to it turns out that the estimator minimizing the mean
square error is also linear in the measurements, i.e. the
𝑃𝑘+1 = 𝑃𝑘 − 𝑝𝑐𝑘+1 + 𝑝𝑔𝑘+1
𝑐 𝑔 linear least squares estimator (LLSE) denoted 𝐿[𝑋∣𝑍]
𝑄𝑘+1 = 𝑄𝑘 − 𝑞𝑘+1 + 𝑞𝑘+1
(4) is the linear function 𝑎+𝑏𝑍 of 𝑍 that minimizes 𝐸((𝑋 −
𝑦𝑘+1 = 𝑦𝑘 − 𝑟𝑘 𝑃𝑘 − 𝑥𝑘 𝑄𝑘
𝑎−𝑏𝑍)2 ). In the multivariate case, let 𝑋, 𝑍 be vectors of
𝛿𝑘+1 = 𝛿𝑘 − 𝑥𝑘 𝑃𝑘 + 𝑟𝑘 𝑄𝑘
random variables on some probability space. The LLSE
with 𝑘 = 0, 1, . . . , 𝑁 − 1. In the sequel, we will use the of 𝑋 given 𝑍 is the linear function 𝐿[𝑋∣𝑍] = 𝑎 + 𝐵𝑍
fact that in this approximation, the power flow equa- that minimizes 𝐸(∥𝑋 − 𝑎 − 𝐵𝑍∥22 ). Following [18], it
tions can be rewritten in terms of the downstream con- can be shown that optimal linear estimator for 𝑋 given
sumption and generation 𝑍 is given by
𝑁
∑ 𝑁
∑ cov(𝑋, 𝑍)
𝑃𝑘 = (𝑝𝑐𝑖 − 𝑝𝑔𝑖 ) = 𝑝𝑛𝑖 𝐿[𝑋∣𝑍] = 𝐸(𝑋) + (𝑍 − 𝐸(𝑍)) (6)
var(𝑍)
𝑖=𝑘+1 𝑖=𝑘+1
𝑁 𝑁 (5)
∑ ∑ For the multivariate case, if Σ𝑍 is invertible, we have
𝑄𝑘 = (𝑞𝑖𝑐 − 𝑞𝑖𝑔 ) = 𝑞𝑖𝑛
𝑖=𝑘+1 𝑖=𝑘+1 𝐿[𝑋∣𝑍] = 𝐸(𝑋) + Σ𝑋,𝑍 Σ−1
𝑍 (𝑍 − 𝐸(𝑍))

where superscript 𝑛 denotes net load. As a result, by In the sequel, we derive the parametrization for 𝑋 (esti-
combining (4) and (5), we can express voltage magni- mation quantities) and 𝑍 (measured quantities) in terms
tude and phasor difference over a branch as a function of of a shared set of random variables (forecasted load vec-
the net loads downstream of that branch. In the sequel, tor). We then derive the first and second order statistics
we will see that this allows us to transform the forecast on 𝑋 and 𝑍 from the available distributions in order to
statistics on loads into forecast statistics on voltage pha- implement the LLSE.
sor differences.
3.2 Parametrization of measured and estima-
3. CONSTRUCTING THE DISTRIBUTION tion quantities
GRID STATE ESTIMATOR Consider the vector with all the differences in squared
In this section, we introduce and prepare the imple- voltage magnitude stacked with the differences in volt-
mentation of the estimation scheme depicted in Fig- age angles over all the branches (i.e. for every set of
ure 1. First, we introduce the concept of MMSE and adjacent nodes) in the network:
LLSE. Second, we parametrize the measured and non- ⎡ ⎤
𝑦1 − 𝑦0
measured quantities as a function of the nodal net load ⎢ 𝑦2 − 𝑦1 ⎥
vector. Thirdly, it turns out, that for this estimator, ⎢ ⎥
⎢ .. ⎥
we require statistics on the measured and estimation ⎢ . ⎥
⎢ ⎥
quantities. These can be derived by transforming the ⎢ 𝑦𝑁 − 𝑦𝑁 −1 ⎥
Δy ≜ ⎢ ⎢ ⎥ ∈ ℝ2𝑁 (7)
available forecast statistics on the net load vector into 𝛿 − 𝛿 ⎥
⎢ 1 0 ⎥
forecast statistics on both the measured and to be es- ⎢ 𝛿2 − 𝛿1 ⎥
⎢ ⎥
timated quantities, using the linear model. Lastly, we ⎢ .. ⎥
⎣ . ⎦
present the analytical formulations for the distribution 𝛿𝑁 − 𝛿𝑁 −1
grid state estimator in the context of voltage magnitude
and angle estimation. Let r ≜ [𝑟0 ⋅ ⋅ ⋅ 𝑟𝑁 −1 ]⊤ and x ≜ [𝑥0 ⋅ ⋅ ⋅ 𝑥𝑁 −1 ]⊤ be vec-
tors with respectively resistances and reactances ordered
3.1 Minimum Mean Square Estimation for all the lines. Using Equation (4), we can build a
Consider (𝑋, 𝑍) to be a pair of random variables on model for all the voltage differences over wires through-
some probability space. Here, we are interested in es- out the network
timating 𝑋, given measurements 𝑍, i.e. we consider a ⎡ ⎤
𝑃0
Bayesian estimation formulation. Assume we are given ⎢ .. ⎥
a joint distribution of (𝑋, 𝑍). The idea is now to find ⎢ . ⎥
[ ]⎢ ⎥
an estimator 𝑌 = 𝑔(𝑍) that minimizes the mean square −diag(r) −diag(x) ⎢ ⎢ 𝑃 𝑁 −1 ⎥
⎥
Δy = ⎢ 𝑄0 ⎥
error 𝐸(∥𝑋 − 𝑌 ∥2 ). One can show that the minimum −diag(x) diag(r)
mean square estimate (MMSE) of 𝑋 given 𝑍 is equiva- ⎢ ⎢ ..
⎥
⎥ (8)
≜𝑍𝑏 ∈ℝ2𝑁 ×2𝑁 ⎣ . ⎦
lent to the conditional expectation 𝐸[𝑋∣𝑍] [18].
𝑄𝑁 −1
We consider the case in which both the estimator and 
the measurements are linear in a shared set of vari- ≜S∈ℝ2𝑁
ables for which distributions are available. In this case, = 𝑍𝑏 S
where S is the vector with real and reactive branch flows refer to the placement of our sensors
stacked vertically. With Equation (5), we can expresses ⎡ ⎤
𝑦𝑚 2 − 𝑦𝑚 1
the branch flows S in terms of the nodal net loads, which ⎢ .. ⎥
yields ⎢ . ⎥
⎢ ⎥
⎢ 𝑦𝑚𝑀 − 𝑦𝑚𝑀 −1 ⎥
Δy𝑚 ≜ ⎢ ⎥
⎢ 𝛿𝑚 2 − 𝛿𝑚 1 ⎥ ∈ ℝ
2(𝑀 −1)
⎡ ⎤ ⎢ ⎥
1 1 ⋅⋅⋅ 1 0 ⋅⋅⋅ ⋅⋅⋅ 0 ⎡ ⎤ ⎢ .. ⎥
⎢ .. .. . . .. ⎥ 𝑝𝑛1 ⎣ ⎦
⎢ .
⎢ 0 1 1 . . . . ⎥⎥⎢
⎢ 𝑝𝑛2 ⎥
⎥
⎢ .. ⎥⎢ .. ⎥ 𝛿𝑚𝑀 − 𝛿𝑚𝑀 −1
⎢ .. .. ⎥⎢ ⎥
⎢ . . . ⎥⎢ . ⎥
⎢ with 𝑚1 , . . . , 𝑚𝑀 ∈ ℳ. We can now formulate the
0 0 ⋅⋅⋅ 1 0 ⋅⋅⋅ ⋅⋅⋅ 0 ⎥ ⎢ 𝑛 ⎥
𝑝𝑁 ⎥
S = ⎢
⎢
⎥⎢ equations, by adding up the differences of all individ-
⎢ 0 0 ⋅⋅⋅ 0 1 0 ⋅⋅⋅ 1 ⎥⎢⎥
⎢ 𝑞1𝑛 ⎥
⎥
⎢ ual lines in between the sensors. I.e., we can formulate
⎢ .. .. .. .. ⎥ ⎢ 𝑞2𝑛 ⎥
⎢ . . . 0 1 1 . ⎥⎢
⎥
⎢
⎥
.. ⎥
a permutation matrix such that Δy𝑚 = 𝒫𝑚 Δy, and
⎢ .. .. .. ⎥⎣ . ⎦ hence
⎣ . . . ⎦
𝑛
𝑞𝑁
0 ⋅⋅⋅ 0 0 0 ⋅⋅⋅ 1
≜ 𝒫𝑏 s𝑛 Δy𝑚 = 𝒫𝑚 𝑍𝑛 s𝑛 = 𝑍𝑚 s𝑛 (11)
(9) where 𝑍𝑚 ≜ 𝒫𝑚 𝑍𝑛 ∈ ℝ2(𝑀 −1)×2𝑁 . This gives us an
where s𝑛 ∈ ℝ2𝑁 is a vector with the nodal net loads, expression for the measured quantities as a function of
real and reactive stacked vertically. Note that for a line the nodal load vector.
feeder, 𝒫𝑏 has an upper triangular structure, which gen-
eralizes to any radial feeder with buses indexed from the 3.2.2 Non-measured quantities - Voltage Estimation
top of the network (feeder head) down to the leaf nodes, We are interested to estimate voltage magnitude and
be it that not all lower triangular entries will be equal angle at all the 𝑁 − 𝑀 buses in the network that are
to 1. not equipped with a sensor. We aim to do this given
We have now expressed the differences in voltage mag- a measurement of the voltage phasor y𝑚 at a limited
nitude and angle over all 𝑁 lines in terms of the nodal number of 𝑀 buses in the network, and forecast statis-
load vector, i.e. tics on the load vector s𝑛 . We consider the differences
in voltage between a location we want to estimate and a
Δy = 𝑍𝑏 𝒫𝑏 s𝑛 ≜ 𝑍𝑛 s𝑛 (10) nearby sensor location. These differences are collected
in a vector Δy𝑒 to be estimated as a function of the
load vector s𝑛 , similar to the construction of the mea-
where 𝑍𝑛 ∈ ℝ2𝑁 ×2𝑁 . surement equation in (11), i.e.
Δy𝑒 = 𝑍𝑒 s𝑛 ∈ ℝ2(𝑁 +1−𝑀 ) (12)
3.2.1 Measured quantities where 𝑍𝑒 ∈ ℝ 2(𝑁 +1−𝑀 )×2𝑁
is constructed in the same
In our actual setting, we do not directly measure volt- way as 𝑍𝑚 in (11). In order to retrieve an estimate
age differences over all individual wires. Instead, we of the absolute voltage value, we can simply take the
place the sensors over a distance, i.e. over a path going nearest sensor reading and add/subtract the estimated
through multiple nodes, with potentially branches going difference between the location and that sensor location.
off from the path. The voltage difference over the path
can be rewritten as the sum of the individual differences 3.3 Forecast Statistics
of the branches lying on the path. For example, for a We consider a setting in which we have access to a
path from bus 0 through bus 1 to bus 2, we have load forecast of the nodal net load vector s𝑛 . For now,
we assume that the forecast algorithm outputs a fore-
𝑦2 − 𝑦0 = (𝑦2 − 𝑦1 ) + (𝑦1 − 𝑦0 ) cast for nodal loads with a Gaussian distribution, i.e.
= (𝑟1 𝑃1 + 𝑥1 𝑄1 ) + (𝑟0 𝑃0 + 𝑥0 𝑄0 ) s𝑛 (𝑡) ∼ 𝒩 (𝜇𝑠 (𝑡), Σ𝑠 (𝑡)), with Σ𝑠 (𝑡) nonsingular. Note
= (𝑟1 (𝑝𝑛0 + 𝑝𝑛1 ) + 𝑥1 (𝑞0𝑛 + 𝑞1𝑛 )) + (𝑟0 𝑝𝑛0 + 𝑥0 𝑞0𝑛 ) that encode the changing nature of the forecast over
= (𝑟1 + 𝑟0 )𝑝𝑛0 + 𝑟1 𝑝𝑛1 + (𝑥1 + 𝑥0 )𝑞0𝑛 + 𝑥1 𝑞1𝑛 time 𝑡 in the statistics of s𝑛 .
We are interested in deriving a forecast statistic of the
variables that we want to monitor in the distribution
Notice that we can form this equation by doing a row grid. In this setting, these are voltage phasor differ-
operation on Equation (10), i.e. adding up rows 1 and ences as described in the former section. These forecast
2 of 𝑍𝑛 . Imagine that we have a set of differences that statistics are valuable by themselves as a prediction of
what the voltage phasor at a certain bus with or without topology and impedance information of the network.
a sensor might do at a certain point in time. Moreover,
we will use the forecast statistics to inform our LLSE 4. SIMULATION RESULTS
estimator. For the LLSE, we are interested in deriving In this paper, we apply the DGSE method on two
the first and second moment statistics of this forecast, simple single phase models. We first explain the ex-
i.e. the mean and covariance. perimental setup as done on a lateral line feeder to
In the previous subsection, we expressed different quan- assess the affect of sensor distance and aggregation of
tities that we can measure or would like to estimate as load/generation. We then apply the estimator on a ra-
linear expressions of the nodal net load vector. Since a dial IEEE test feeder model.
linear combination of Gaussian variables is a Gaussian
variable, we can derive the distributions for our esti- 4.1 Analysis on lateral line feeder
mation quantities. As a result, we have that 𝑥(𝑡) ∼ We assess the distribution grid state estimator through
𝒩 (𝜇𝑥 (𝑡), Σ𝑥 (𝑡)) and 𝑧(𝑡) ∼ 𝒩 (𝜇𝑧 (𝑡), Σ𝑧 (𝑡)), with Σ𝑧 (𝑡) a Monte Carlo simulation on a line feeder. Simula-
nonsingular. We derive these quantities in the setting tions are done with a full single-phase model, using
of voltage phasor estimation, the procedure can be fol- a forward-backward sweep algorithm. The structure
lowed for the other estimation procedures (power flow or of the line feeder allows for convenient analysis and
aggregate load). We now have that our measurements is a fair generalization for radial networks. We con-
are voltage phasor differences, i.e. 𝑧 = Δy𝑚 and the es- sider different length feeders, i.e. 𝑁 ∈ {20, 50}, to re-
timation quantities are other voltage phasor difference, semble the difference in geographic span found in real-
i.e. 𝑥 = Δy𝑒 . Given the linear relationships with the world circuits. For a given feeder we assign realistic
load vector s𝑛 , we can now derive the statistics on 𝑧. impedance values sampled from a uniform distribution
The mean of 𝑧 is at 𝑟𝑘 , 𝑥𝑘 ∼ 𝒰 (0.01, 0.11)Ω. The feeder is energized at
𝜇𝑧 (𝑡) = 𝐸(Δy𝑚 ) the substation, located at bus 𝑁 , at 𝑣0 = 12kV.
= 𝐸(𝑍𝑚 s𝑛 ) At each bus we consider the presence of an aggrega-
(13)
= 𝑍𝑚 𝐸(s𝑛 ) tion of typical residential houses. For each house 𝑖 we
= 𝑍𝑚 𝜇𝑠 (𝑡) will forecast the real power net load according to a nor-
mal distribution 𝑝𝑖 ∼ 𝒩 (𝜇𝑝𝑖 , 𝜎𝑖𝑝 ), 𝑞𝑖 ∼ 𝒩 (𝜇𝑞𝑖 , 𝜎𝑖𝑞 ). The
and similarly, we have that 𝜇𝑥 (𝑡) = 𝐸(Δy𝑒 ) = 𝑍𝑒 𝜇𝑠 (𝑡). mean values 𝜇𝑝𝑖 , 𝜇𝑞𝑖 for each house are drawn randomly
The covariance of 𝑧 is from a uniform distribution 𝜇𝑝𝑖 ∼ 𝒰 (−1, 1) kW. Recent
Σ𝑧 (𝑡) = 𝐸((𝑧 − 𝜇𝑧 )(𝑧 − 𝜇𝑧 )⊤ ) work on forecasting of individual homes reports a mean
= 𝐸((𝑍𝑚 s𝑛 − 𝑍𝑚 𝜇𝑠 (𝑡))(𝑍𝑚 s𝑛 − 𝑍𝑚 𝜇𝑠 (𝑡))⊤ ) absolute percentage error of 20 − 30% [15]. Hence, for
= 𝑍𝑚 𝐸((s𝑛 − 𝜇𝑠 (𝑡))(s𝑛 − 𝜇𝑠 )⊤ )𝑍𝑚⊤ our normally distributed forecast we introduce a con-
= 𝑍𝑚 Σ𝑠 (𝑡)𝑍𝑚⊤ servative variance, i.e. 𝜎𝑖𝑝 = 30% ⋅ 𝜇𝑝𝑖 (and similarly for
(14) reactive power). In this formulation, we can capture
⊤
denotes matrix transpose. Similarly, we have that the variance related to DG or EVs. Nevertheless, we
the cross-covariance of 𝑥 and 𝑧 is Σ𝑥,𝑧 (𝑡) = 𝑍𝑒 Σ𝑠 (𝑡)𝑍𝑚⊤
. realize that a normal distribution is not accurate in all
This yields all the statistics we need to construct the scenarios, hence we aim to generalize the estimator in
distribution grid state estimator. our future work.
In the simulation we vary the aggregation, by increas-
3.4 Distribution grid state estimator ing the number of houses 𝑐 connected at each bus, 𝑐 ∈
We can now analytically derive the LLSE of 𝑋 given {1, . . . , 50}. We do this uniformly across the feeder. We
𝑌 . For our voltage estimation setting this yields also vary the number of sensors placed, by populating
the feeder with a sensors placed at a fixed and increasing
𝐿[Δy𝑒 ∣Δy𝑚 ] = 𝐸(Δy𝑒 ) + . . . distance 𝑑 ∈ {1, . . . , 20}. Here, we consider two types
ΣΔy𝑒 ,Δy𝑚 Σ−1 Δy𝑚 (Δy𝑚 − 𝐸(Δy𝑚 )) of sensors, voltage magnitude sensors and PMUs (mea-
= 𝑍 𝑒 𝜇𝑠 + . . . suring voltage magnitude and angle). We perform 5000
⊤
( )
⊤ −1
𝑍𝑒 Σ𝑠 𝑍𝑚 𝑍𝑚 Σ𝑠 𝑍𝑚 (Δy𝑚 − 𝑍𝑚 𝜇𝑠 ) Monte Carlo simulations to achieve statistically relevant
(15) insights given the assumed uncertainty in load forecast.
We coin this method the Distribution Grid State Es-
timator (DGSE). Notice that it is written in the form 4.2 Results
Δŷ𝑒 = 𝑓 (Δy𝑚 ), as all of the other information needed The LLSE is used to estimate the voltage phasor dif-
to evaluate the estimator are forecast statistics, which ference vector Δy𝑒 , given a vector of measured volt-
are known a priori. As such, (15) uses the statistical in- age phasor differences Δy𝑚 . Figure 2 shows the re-
formation of the net loads s𝑛 , in combination with the sult of one sample simulation, with two different sen-
 Voltage magnitude estimation with sensor at every 5-th bus casted voltage phasor differences. It is important to
1.08
note that the ARMSE is highly dependent on the un-
1.06 certainty as defined in the load forecast. Hence, we will
also analyze how well the LLSE does at improving the
|V|

1.04
accuracy of the forecasted values. Figure 3 shows the
Real voltage
Measured z
1.02 Forecasted x
Updated x Monte Carlo ARMSE results for varying levels of ag-
1
0 10 20 30 40 50 60
gregation and sensor distance. We first inspect the ac-
k curacy of the forecasted values. For all the scenarios
1.08
Voltage magnitude estimation with sensor at every 15-th bus
simulated, the ARMSE is bounded above by 0.008 p.u..
1.06
The ARMSE increases roughly linearly with sensor dis-
tance. It increases nonlinearly for smaller levels of load
aggregation, and then linearly for 𝑐 ≥ 12 houses/bus.
|V|

1.04
Real voltage
Measured z
1.02 Forecasted x
Updated x
1
0 10 20 30 40 50 60
k

Figure 2: Voltage Estimation sample result, comparing

the actual voltage profile with the forecasted profile and
the estimated profile. Top plot shows result for a sensor
at every 5-th bus, bottom plot for a sensor at every 15-th (a) ARMSE vs sensor distance and aggregation
bus.

sor distances. The figure focuses on voltage magnitude

and plots the absolute voltage magnitude by adding the
forecasted and estimated differences to the voltage mag-
nitude measured at the sensor closest in the upstream
direction, i.e.
upstream
𝑣ˆ𝑒 = 𝑣𝑚 + Δˆ𝑣𝑒 , (b) Back of plot
upstream (16)
𝐿[𝑣𝑒 ∣Δy𝑚 ] = 𝑣𝑚 + 𝐿[Δ𝑣𝑒 ∣Δy𝑚 ]
Figure 3: ARMSE [p.u.] for 5000 Monte Carlo sim-
The figure shows clearly how the LLSE is able to up- ulations. Two horizontal axes denote increasing levels
date the forecasted values to approach the actual volt- of aggregation 𝑐 and increasing distance between every
age profile more closely. Comparing the top and bot- two sensors 𝑑. Green is the ARMS value for the fore-
tom plot shows the importance of using the LLSE if the cast, blue denotes the ARMS value for the estimator
sensor distance is larger. This yields a first important after incorporating the real-time measurements.
insight that if one has a higher number of sensors, the
forecast will do a good job at predicting the voltage dif-
We now analyze the performance of the LLSE. Figure
ference between a bus with sensor and buses without.
4 shows the relative improvement in ARMSE achieved
As the number of sensors decreases, the forecast error
by the LLSE as compared to the forecast. This plot
propagates as a voltage difference is estimated for larger
yields some relevant engineering implications. For small
differences from a sensor.
sensor distances (many sensors), the relative improve-
We now define the key metric used to analyze the
ment is on the order of 30 - 35%. Increasing the sensor
performance of the estimator - the Average Root Mean
distance (i.e. decreasing the number of sensors) on the
Square Error (ARMSE). The ARMSE gives an empiri-
circuit yields to two different insights. First, for small
cal measurement of the deviation from the actual value
levels of aggregation 𝑐 ≤ 10, the LLSE relative improve-
of the estimated variable. Given 𝑇 Monte Carlo ground
ment drops to below 30%. This might be associated to
truth samples 𝑥 ∈ ℝ𝑛𝑥 , 𝑧 ∈ ℝ𝑛𝑧 , the ARMSE of some
the nonlinear effect in the ARMSE of the forecast and
estimator or prediction 𝑥˜(𝑡) is defined as
 deserves further analysis. Second, for higher levels of
 aggregation 𝑐 ≥ 11 houses/bus, we see that the relative
( )  ⎷ 1 ∑
𝑇
𝐴𝑅𝑀 𝑆𝐸 {˜ 𝑇
𝑥(𝑡)}𝑡=1 = ˜(𝑡)]∥22
∥𝑥(𝑡) − 𝑥 performance shoots up towards 60% as 𝑑 approaches 6,
𝑛𝑥 𝑇 𝑡=1 and then rises slower for 𝑑 ≥ 6, oscillating and leveling
(17) off around 62 - 68% improvement. Inspecting this fur-
In the following, we will apply this metric to the fore- ther, Figure 3 neatly shows that while the ARMSE of
 Effect of sensor distance for aggregation of 5 houses per bus
0.0035

MC ARMSE [V - p.u.]
0.003
% improvement estimate vs. forecast

0.0025
70
0.002
60
0.0015
2 4 6 8 10 12 14 16 18 20
50 Sensor distance - d [# buses]

40
Figure 5: ARMSE for varying sensor distance 𝑑, and
30 fixed aggregation (5 houses/bus). The error increases
linearly with sensor distance 𝑑.
20
20
15 50 Effect of aggregation for a sensor at every 10-th bus
40 0.003

MC ARMSE [V - p.u.]
10 30
5 20 0.0025
10
Sensor distance [# buses] 0 0
Aggregation [houses/bus] 0.002

0.0015

Figure 4: Average percentage improvement of the esti- 0.001

0 10 20 30 40 50
mator versus the forecast for 5000 Monte Carlo samples. Aggregation - c [# houses/bus]
Improvement levels off at 68% for higher aggregation
and larger sensor distance. Figure 6: ARMSE for varying aggregation 𝑐, and fixed
sensor distance (10 buses). The error increases non-
linearly for lower aggregation, and linearly for 𝑐 ≥ 12
the forecast increases for larger sensor distance, the cor- houses/bus.
responding ARMSE for the LLSE seems to stay around
the same value of 0.0023 p.u..
This leads to several valuable engineering insights. is subject of further study.
First, for circuits with low levels of aggregation (𝑐 ≤ 10
houses/bus), the forecasted values already do a fairly 4.3 Application to IEEE 37 node test feeder
good job at predicting the voltage phasor with an up- To apply the estimator on a radial network, we used
perbound on the ARMSE of 0.0055, and the LLSE im- a modified version of the IEEE 37 node distribution
proves this by roughly 28%. An interpretation for this is feeder model [1], depicted in Figure 7. The feeder volt-
that power flow does not reach such high values, leading age and power ratings were left unchanged (4.8 kV and
to little propagation of errors throughout the network. 2.5 MVA), as were line segment configuration assign-
Second, and more importantly, there is a low cut-off ments. However, in order to create a balanced network,
point for increasing the number of sensors on a circuit, all line segments were assumed to be three phase. To do
after which extra sensors do not improve the perfor- so, we averaged the diagonal impedances of the 3-phase
mance of the estimator significantly. Or in other words, impedance matrices to create single phase impedances.
even a few measurements combined with the forecast We ignored the transformer at node 775 and the volt-
information lead to satisfactory ARMSE performance age regulator at the feeder head. We assumed all loads
with the LLSE, improving the forecasted values by more were constant power, and set their mean values 𝜇𝑝𝑖 , 𝜇𝑞𝑖 to
than 60%. This is good news, as it might help crafting a the sum over the three phases, as provided in [1]. Syn-
viable business case for using a limited set of potentially chrophasor sensors were placed at five different nodes,
expensive sensors to achieve state-of-the-art estimation indicated by red hexagons in Figure 7. We ran 10000
performance. Monte Carlo simulations.
Figures 5 and 6 show the trend in the ARMSE for Figure 8 shows the result for one Monte Carlo run.
respectively increasing sensor distance and higher lev- Figure 9 shows the ARMSE metric for all buses. It is
els of aggregation. As expected and also observed in bounded by 0.0032 p.u. for the forecasted values and
the forecast, the ARMSE grows roughly linearly with 0.0006 p.u. for the values calculated by the DGSE es-
having fewer sensors on the circuit. We found that for timator. As expected, the error in the forecast at a
certain levels of aggregation the linear trend was still ob- specific bus tends to increase with two factors - electri-
served but with slightly more oscillation. Furthermore, cal distance from the closest sensor, and the amount of
we also retrieve the nonlinear trend for increasing levels power flowing through the branches on the path to the
of aggregation as observed in the ARMSE for the fore- closest sensor. For instance, the ARMSE for bus 730,
cast. Interpreting the exact nature of this relationship far away from an upstream sensor, decreased by 97%,
 time using a limited set of measurements, in the form of
799
724 the linear least squares estimator. The algorithm was
722 analyzed in terms of its performance, by inspecting the
712 707 role of the number of sensors and load aggregation in the
701 network, and applied to general radial network topolo-
742 713 704
720 gies. The algorithm can potentially be implemented in
705 702 a decentralized manner, as it relies solely on local volt-
714 age phasor differences and load statistics which can be
706 communicated a priori.
729 744 727
703
718 Next steps entail the theoretical analysis of the LLSE
725 and generalization of load statistics to a broader set
of distributions. We will also apply the method to a
728 set of real feeder models, and use data driven load and
730 DG forecasts. In addition, our formulation allows for
a natural extension to the more realistic case of three-
phase unbalanced systems. Lastly, we will study the
732 708 role of uncertainty in the parameters of the feeder model
709
731
(impedances and topology).
736
733 775 Acknowledgments
710 734 The authors would like to thank Alexandra von Meier
740 for facilitating industry feedback, and Walid Krichene
for providing valuable insights early on in this project.
735
737 738 711 741
6. REFERENCES
Figure 7: IEEE 37 node test feeder model, PMU sensors [1] IEEE Distribution Test Feeders,
are indicated with hexagons. http://ewh.ieee.org/soc/pes/dsacom/testfeeders/.
[2] The Impact of Localized Energy Resources on
Southern California Edison’s Transmission and
while the ARMSE for buses closer to a sensor, such as Distribution System. Technical report, Southern
706 and 725, remained relatively the same. Interest- California Edison Distribution Engineering and
ingly, the DGSE estimator in the chosen setting tends Advanced Technology, May 2012.
to level off the uncertainty over all buses that we start [3] A. Abur and A. Gómez Expósito. Power System
off with in the forecast, thereby increasing the overall State Estimation: Theory and Implementation,
confidence. In the presence of a network model and volume 24 of Power Engineering (Willis). CRC
load and DG forecasts, this type of analysis can be used Press, Mar. 2004.
to determine optimal sensor placement and offline er- [4] M. Baran and F. Wu. Optimal capacitor
ror bound calculation for estimated variables at specific placement on radial distribution systems. IEEE
locations in the network. Future work will introduce Transactions on Power Delivery, 4(1):725–734,
specific scenarios relevant to higher levels of renewables Jan. 1989.
to see how the DGSE method can serve as an observer [5] M. E. Baran and F. F. Wu. Optimal sizing of
for purpose of controlling voltage regulation and protec- capacitors placed on a radial distribution system.
tion equipment. Power Delivery, IEEE Transactions on,
4(1):735–743, 1989.
5. CONCLUSIONS AND FUTURE WORK [6] E. Blood, B. Krogh, and M. Ilic. Electric power
This paper addressed the challenge of formulating a system static state estimation through Kalman
distribution grid state estimator, for scenarios where filtering and load forecasting. In 2008 IEEE
fully observed sensor arrangements are not yet feasible, Power and Energy Society General Meeting -
and load forecasts are subject to increasing uncertainty. Conversion and Delivery of Electrical Energy in
It derived an algorithm that exploits the information the 21st Century, pages 1–6, July 2008.
in load forecasting and feeder models to construct prior [7] Eugene Shlatz, Nathan Buch, and Melissa Chan.
statistics of the variables to be estimated. It then used a Distributed Generation Integration Cost Study -
Bayesian approach update these prior statistics in real- Analytical Framework. Technical report,
 Voltage magnitude estimation with sensors at 5 out of 37 buses
1
Real voltage
0.99 Measured z
Forecasted x
0.98
Updated x
0.97
k

0.96
V

0.95

0.94

0.93

0.92
799 701 702 705 742 712 713 704 720 706 725 707 724 722 703 730 709 708 732 731 733 734 710 735 736 737 738 711 741 740 714 718 727 744 728 729 775
bus number

Figure 8: Example voltage profile with forecast and estimation update at all the buses, numbered as in Figure 7.
-3 ARMSE for each bus for sensors at 5 out of 37 buses
×10
4
Forecast
3 Estimation
ARMSE

0
701 702 705 742 712 713 704 706 725 707 724 722 703 730 708 732 731 733 734 710 735 736 738 711 741 740 714 718 727 744 728 775
bus number

Figure 9: ARMSE in p.u. for each non-measured bus.

California Energy Commission, Nov. 2013. [14] N. Seltenrich. The New Grid - Plugging into
[8] G. Giannakis, V. Kekatos, N. Gatsis, S.-J. Kim, California’s clean-energy future, 2013.
H. Zhu, and B. Wollenberg. Monitoring and [15] R. Sevlian and R. Rajagopal. Value of aggregation
Optimization for Power Grids: A Signal in smart grids. In 2013 IEEE International
Processing Perspective. IEEE Signal Processing Conference on Smart Grid Communications
Magazine, 30(5):107–128, Sept. 2013. (SmartGridComm), pages 714–719, Oct. 2013.
[9] M. Gol and A. Abur. A Hybrid State Estimator [16] E. Veldman and R. Verzijlbergh. Distribution
For Systems With Limited Number of PMUs. Grid Impacts of Smart Electric Vehicle Charging
IEEE Transactions on Power Systems, From Different Perspectives. IEEE Transactions
30(3):1511–1517, May 2015. on Smart Grid, 6(1):333–342, Jan. 2015.
[10] R. C. Green, L. Wang, and M. Alam. The impact [17] A. von Meier, D. Culler, A. McEachern, and
of plug-in hybrid electric vehicles on distribution R. Arghandeh. Micro-synchrophasors for
networks: A review and outlook. Renewable and distribution systems. In IEEE 5th Innovative
Sustainable Energy Reviews, 15(1):544–553, 2011. Smart Grid Technologies Conference, Washington,
[11] G. Kaur and M. Vaziri. Effects of distributed DC, 2014.
generation (DG) interconnections on protection of [18] J. Walrand. Probability in Electrical Engineering
distribution feeders. In IEEE Power Engineering and Computer Science: An Application-Driven
Society General Meeting, 2006, 2006. Course, volume 1. Quoi?, 1st edition, Mar. 2014.
[12] P. Rousseaux, T. Van Cutsem, and T. E. [19] B. Xu and A. Abur. Observability analysis and
Dy Liacco. Whither dynamic state estimation? measurement placement for systems with PMUs.
International Journal of Electrical Power & In Proc. 2004 IEEE Power Systtem Conf,
Energy Systems, 12(2):104–116, Apr. 1990. volume 2, pages 943–946. Citeseer, 2004.
[13] L. Schenato, G. Barchi, D. Macii, R. Arghandeh,
K. Poolla, and A. Von Meier. Bayesian linear
state estimation using smart meters and PMUs
measurements in distribution grids. In 2014 IEEE
International Conference on Smart Grid
Communications (SmartGridComm), pages
572–577, Nov. 2014.