120 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 34, NO.

1, JANUARY 2019

Mitigating the Risk of Voltage Collapse Using

Statistical Measures From PMU Data
Samuel C. Chevalier , Student Member, IEEE, and Paul D. H. Hines , Senior Member, IEEE

Abstract—With the continued deployment of synchronized pha- and control systems rely on voltage magnitude data for control
sor measurement units (PMUs), high sample rate data are rapidly decisions.
increasing the real time observability of power systems. Prior re- Across a variety of complex systems, there is increasing ev-
search has shown that the statistics of these data can provide useful
information regarding network stability, but it is not yet known idence that indicators of emerging critical transitions can be
how this statistical information can be actionably used to improve found in the statistics of state variable time series data [3].
power system stability. To address this issue, this paper presents Termed Critical Slowing Down (CSD) [4], this phenomenon
a method that gauges and improves the voltage stability of a sys- most clearly appears as elevated variance and autocorrelation in
tem using the statistics present in PMU data streams. Leveraging time-series data [5]. More recently, CSD has been successfully
an analytical solver to determine a range of “critical” bus voltage
variances, the presented methods monitor raw statistical data in investigated in the power systems literature, and strong con-
an observable load pocket to determine when control actions are nections have been drawn between bifurcation theory and the
needed to mitigate the risk of voltage collapse. A simple reactive elevation of certain statistics in voltage and current time series
power controller is then implemented, which acts dynamically to data [6]–[9]. In particular, [9] presents a method for analytically
maintain an acceptable voltage stability margin within the system.
calculating the time series statistics associated with a stochas-
Time domain simulations on 3-bus and 39-bus test cases demon-
strate that the resulting statistical controller can outperform more tically forced dynamic power system model. These results are
conventional feedback control systems by maintaining voltage sta- leveraged in this paper in order to predict key statistics of a
bility margins while loads simultaneously increase and fluctuate. power system that is approaching a critical transition. Others,
Index Terms—Synchronized phasor measurement units, voltage
such as [10], have developed control methods that use voltage
collapse, critical slowing down, holomorphic embedding, first magnitude declination rate measurements, but do not explicitly
passage processes. use statistical information as are presented in this paper.
As reviewed in [11], power systems are liable to experience
a variety of critical transitions, including Hopf, pitchfork, and
I. INTRODUCTION
limit-induced bifurcations. This paper is primarily concerned
N ORDER to optimize the use of limited infrastructure, with the slow load build up, reactive power shortages, and other
I transmission systems frequently operate close to their stabil-
ity or security limits. Although economically advantageous [1],
Long Term Voltage Stability (LTVS) processes that may con-
tribute to a Saddle-Node bifurcation of the algebraic power flow
this can lead to elevated blackout risk given the fluctuating na- equations. Classic voltage stability, which refers to a power sys-
ture of supply and demand [2]. Consequently, stability margin tem’s ability to uphold steady voltage magnitudes at all network
estimation is an essential tool for power system operators. Pre- buses after experiencing a disturbance, is lost after a network
dicting the onset of voltage instability, though, is often made experiences this sort of bifurcation [12]. The methods in this
difficult by reactive support systems and tap changing trans- paper aim to preserve such voltage stability and thus prevent a
formers that hold voltage magnitudes high as load increases. system from experiencing voltage collapse.
Although voltage support is essential for reliable operations, The goal of this paper is to describe and evaluate a control
these controls can sometimes hide the fact that a system is ap- system that uses the variance of bus voltages to reduce the prob-
proaching a voltage stability limit, particularly when operators ability of voltage collapse in a stochastic power system. This
control system leverages a number of innovative tools to per-
Manuscript received October 18, 2017; revised May 9, 2018; accepted August form this task. The first uses the load scaling factor from the
12, 2018. Date of publication August 23, 2018; date of current version December Holomorphic Embedding Load Flow Method (HELM) [13] to
19, 2018. The work of P. D. H. Hines was supported in part by U.S. National represent a slowly varying stochastic variable, such as changes
Science Foundation (NSF) Award ECCS-1254549. Paper no. TPWRS-01579-
2017. (Corresponding author: Samuel Chevalier.) in overall load levels over time. Second, a First Passage Process
S. C. Chevalier is with the Department of Mechanical Engineering, (FPP) [14] is used to identify critical loading thresholds given
Massachusetts Institute of Technology, Cambridge, MA 02139 USA (e-mail:, the statistics of the slower load changes. Finally, a full order
[email protected]).
P. D. H. Hines is with the Department of Electrical and Biomedical dynamical system model is used to analytically predict the ex-
Engineering, University of Vermont, Burlington, VT 05405 USA (e-mail:, pected algebraic variable covariance matrix of the system, given
[email protected]). stochastic load noise excitation, for the previously identified
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. critical loading level. The associated critical variances from this
Digital Object Identifier 10.1109/TPWRS.2018.2866484 matrix are then used as a reference signal to control a static VAR

0885-8950 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
CHEVALIER AND HINES: MITIGATING THE RISK OF VOLTAGE COLLAPSE USING STATISTICAL MEASURES FROM PMU DATA 121

Fig. 1. Two bus model with generator, load and shunt capacitor B s .

compensator (SVC). This reference signal is dynamically up-

dated as network configurations change and equilibriums shift.
In building this controller (see Fig. 8), this paper combines static Fig. 2. Load bus voltage as a function of P D for various shunt values. As
algebraic voltage collapse analysis through HELM, the first pas- more reactive power is injected into the system, the voltage magnitude at the
point of bifurcation drifts upward toward nominal operating voltage.
sage probability, statistical estimation (based on system model
and load noise assumptions) and network feedback in order to
leverage the statistical properties of PMU data to reduce the risk voltage of 1 p.u. to the point of bifurcation, decreases. This is
of voltage instability. clearly seen in Fig. 2 by the convergence of the arrows associated
The remainder of this paper is organized as follows. Section II with the ∗ symbols (bifurcation) and the • symbols (nominal
motivates the use of voltage variance as an indicator of stabil- operating voltage) as reactive support increase. On the other
ity and presents the methods employed in the statistical con- hand, as load increases, the magnitude of the derivative of the
troller which is developed in this paper. Section III describes the PV curve (with respect to load) increases, suggesting that this
new statistical controller along with two conventional reactive derivative is a useful indicator of proximity to the bifurcation.
power controllers that are used to benchmark the results against. If load varies stochastically with known variance σP2 D , (1)
Section IV demonstrates the statistical controller on a 3 bus describes the expected voltage variance via the delta method [16]
power system consisting of a generator, a load bus, and an SVC at the load bus:
bus. For further validation, Section V tests the controllers on the   2
IEEE 39 bus system. Finally, conclusions and ideas for future dVt 
σV t ≈
2
σP2 D , (1)
research are presented in Section VI. dPD E[P D ]

II. BACKGROUND where E[PD ] is the expected value of the load. Predicting the
distance to static voltage collapse with variance measurements
This section motivates the use of bus voltage variance as a can be accomplished by (i) drawing the PV curve for a system,
measure of stability and presents the methods and tools used to (ii) defining a loading margin (in terms of complex power P +
build our statistical controller. jQ) on the curve which should not be exceeded, and then (iii)
calculating the expected bus voltage variance at this threshold. If
A. Bus Voltage Variance in a 2 Bus Power System measured voltage variance exceeds this threshold value then the
system may be at risk of exceeding its stability limits. If reactive
A variety of systems, such as capacitor banks, tap changing
support is high, the voltage magnitude of the system may be
transformers, and various Flexible AC Transmission System
an unreliable real time measure of voltage stability. The bus
(FACTS) devices, are employed in power systems to ensure
voltage variance statistic, however, can potentially tell a more
that voltage magnitudes remain sufficiently high. As a result,
complete story about system stability. In the following sections,
voltage magnitudes are an imperfect indicator of the proximity
the stability information encoded in the variance is leveraged in
of a system to its voltage stability limits. To understand how an
order to make real-time, data-driven control decisions.
overloaded system with high voltage magnitudes may have a
compromised voltage stability margin, the definition of loading
margin in [15] (p. 262) is first given: “For a particular operating B. System Model Overview
point, the amount of additional load in a specific pattern of load A stochastically forced power system can be modeled with a
increase that would cause a voltage collapse is called the loading set of Differential-Algebraic Equations (DAEs) of the form
margin.”
Consider the system in Fig. 1, where capacitive shunt support ẋ = f (x, y) (2)
Bs and a constant power load P + jQ are placed at the “to”
0 = g(x, y, u(t)) (3)
bus and a generator with fixed voltage is located at the “from”
bus. Fig. 2 shows the power-voltage curves that result if the where f , g represent the differential and algebraic systems, x,
load’s power factor is held fixed with several different amounts y are the differential and algebraic state variables, and u(t)
of reactive power injection. represents the time-varying stochastic (net) load fluctuations [9],
As reactive support increases, the system can sustain larger [17], [18]. Neglecting for the moment the slow changes in load
increases in load before voltage collapse occurs. However, if level, the complex load at time t can be represented by (4):
reactive resources are used to maintain voltages at their nominal
levels (1 p.u.), the load margin, as measured from the operating S(t) = S0 (1 + u(t)) (4)
122 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 34, NO. 1, JANUARY 2019

with the dynamics of the fast load fluctuations given by the Prior work [22] provides an important foundation for using
Ornstein-Uhlenbeck process expressed in (5): HELM to solve for the static stability margins of a power sys-
tem. After generating the holomorphic voltage functions, the
u̇ = −Eu + Σξ (5)
largest, positive zero of the numerator of the Padé approximant
where E is a diagonal matrix of inverse time correlations, ξ is approximates the maximum power transfer point of the system.
a vector of zero-mean independent Gaussian random variables This method, though, scales all loads at uniform rates, and does
whose standard deviations are given on the diagonal of the n × n not account for more than one single generator bus in the system.
diagonal matrix Σ. This paper assumes that a grid operator In order to solve these problems, we derive a new method for
can estimate the statistics of load fluctuations (E and Σ) from scaling loads from a known base case solution. This approach
measurements. allows loads and generators to scale at different rates.
In the conventional CPF problem, generation participation
C. Computing the Algebraic Variable Covariance Matrix rates are assigned to generators to pick up excess load as it is
The process for deriving the approximate covariance matrix scaled. This is not the approach we took. For mathematical sim-
for all variables in a stochastically forced power system is de- plicity, we instead solve the base case power flow solution and
rived in [9]. This computation allows one to characterize the then fix the generator voltage phase angles. As load increases at
statistics of a system that is approaching a bifurcation. This the load buses, generation throughout a system increases quasi-
method is based on linearizing the equations encompassed by proportionally to the electrical distance between the generator
(2) and (3) and then algebraically solving for Δẋ and Δu̇ by and the load. Electrically proximal generators respond with the
eliminating the algebraic variable vector Δy: largest generation increases, while electrically distant genera-
       tors respond with smaller increases; we justify these simplifying
Δẋ As −fy g−1 y gu Δx 0 assumptions in [23]. Incoporating droop-coefficient-based gen-
= + ξ. (6) erator loading rates remains for future work.
Δu̇ 0 −E Δu Σ
We originally derived the full details of the method in [23].
where As is the standard state matrix. Using z = The mathematics are too lengthy to be shown in this paper, but
 
Δx Δu , (6) can be rewritten with compact matrices A and they are summarized in the remainder of this subsection. We
B via begin by defining a holomorphic voltage function for the ith
power system bus voltage via the following power series:
ż = Az + Bξ. (7)
∞

As introduced in [19], the Lyapunov equation (8) can be solved Vi (s) = Vi [n]sn , (9)
numerically1 to calculate the covariance matrix of z, where A n =0
and B are defined in (7):
where the variable s is a complex holomorphic function pa-
Aσz2 + σz2 A = −BB  . (8) rameter. The s = 0 condition yields the complex system wide
voltages for a given base case power flow solution (which may
 Δy is given by Δy = KΔz where
Since the linearized output
be solved for via HELM or Newton-Raphson). If this power flow
K ≡ −gy−1 gx −gy−1 gu , the state variable covariance matrix
solution is known, Vi [0] is known ∀i in the system. With this
can be transformed into the algebraic variable covariance matrix
definition, the holomorphically embedded power flow equation
via σy2 = Kσz2 K  . A subset of the diagonal entries of σy2 contain
at the ith PQ bus in an N bus power system may be stated:
the bus voltage variances.

N
Si∗ + ski Si∗
D. Adapting HELM to Solve CPF Yi,k Vk (s) = , i ∈ PQ. (10)
Vi∗ (s∗ )
k =1
The Continuation Power Flow (CPF) problem is a classic ap-
Equation (10) has the following attributes:
proach to understanding and predicting voltage instability. As r The holomorphic parameter s scales the load as it is in-
outlined in [20], CPF involves drawing PV curves given load and
generation increase rates using iterative Newton-Rapson meth- creased from s = 0. If s is real, the power factor of the
load is constant as apparent power scales.
ods. As introduced in [13], iterative techniques, such as Newton- r Si = Pi + jQi is the complex power injection at bus i.
Raphson, can encounter a numerical issues, such as divergence r The exponent ∗ denotes complex conjugation.
or finding undesirable low-voltage solutions, when solving the r The parameter ki , which can be positive, negative, or 0,
nonlinear power flow equations, particularly when a system ap-
proaches a Saddle-node bifurcation. An alternative is to use the corresponds to the rate at which bus i will be loaded as
Holomorphic Embedding Load-flow Method (HELM), which s increases from 0. If ki = 0, the load at bus i will not
uses complex analysis and recursive techniques to overcome change as s increases.
these numerical difficulties. If one exists, HELM is guaranteed The holomorphically embedded equations at voltage con-
to compute the high voltage power flow solution [21]. trolled buses (PV and reference) are given by
Vi (s) = Vi ej θ i , i ∈ {PV ∪ r} . (11)
1 Singularity of the state matrix is required for (8) to have a solution. As
the system approaches a singularity-induced bifurcation and A s approaches Generator voltages are independent of s since reactive power
singularity, the predicted variance will approach infinity. limits are not considered in this formulation. With the structure
CHEVALIER AND HINES: MITIGATING THE RISK OF VOLTAGE COLLAPSE USING STATISTICAL MEASURES FROM PMU DATA 123

given by (9), (10), and (11), the formulations given in [23, eqs.
(4.65)–(4.89)] may be used to recursively compute the power
series coefficients of Vi (s). Once done, the complex bus voltage
at PQ bus i for some arbitrary loading level s = sl may be
computed via

∞ 

Vi ej θ i = Vi [n]sn  . (12)

n =0 s=s l

If the voltage functions are evaluated from s = 0 to the voltage

collapse load s = sc , the CPF voltage magnitude curves may be
drawn analytically. In order to identify the critical load s = sc ,
the formulations in [23, eqs. (3.46)–(3.50)] may be used to
generate the Padé approximants A and B of the power series
Vi (s):
c −1
N
N c2−1
n A[n] (sn )
V [n] (s ) =
nN=0
c −1
. (13) Fig. 3. (a) The voltage magnitude for load buses of the 39 bus system as s is
n =0
2
B[n] (sn ) scaled, as computed by our adaptation of HELM; the curves in this panel are
n =0 drawn analytically. The thick black curve is the voltage magnitude of bus 20,
From this formulation, sc will be equal to the smallest, positive whose load is decreasing as s increases, according to equation (14). (b) The
voltage magnitude difference (error) for each PV curve between the NRPF and
real root of A [22]. HELM solutions. The error is numerically insignificant until s approaches the
In order to validate this continuation method, which we refer bifurcation point (s ≈ 2).
to as CPF via HELM, we test our method on the IEEE 39 bus
system. We first define the vector k which has length 39. The
elements of this vector contain the respective arbitrary loading
rates of the buses in the system.
⎧
⎪
⎨ 1 i ∈ {3, 4, 7, 8}
ki = −0.2 i = 20 (14)
⎪
⎩
Fig. 4. An illustration of the values that the holomorphic parameter s, which
0 otherwise starts at a base load s = 0, can attain during its random walk. sm is the load level
at which the probability of s hitting the Saddle-Node bifurcation sc exceeds a
The loads of the system scale in the following way, where S0 i probability limit.
is the base load at bus i in the system:
Si = S0 i + ski S0 i . (15) load buildup. This section builds on the First Passage Process
literature to systematically compute the probability that load
In this example, we scale s from 0 to 1.99, at which point the will not increase beyond a collapse threshold during a given
Saddle-Node Bifurcation occurs. Once the Padé approximants time period. To do so, we consider the holomorphic parameter
are known for each bus, we scale s and solve for the complex s from (10) which will scale the base complex power of the ith
voltages at each bus. The resulting PV curves are shown in Panel bus according to (15). To capture slow stochastic load changes,
(a) of Fig. 3. the evolution of parameter s is modeled as a Wiener Process
We also validated HELM against conventional Newton in which s begins at the origin and takes Gaussian-distributed
Raphson Power Flow (NRPF). As s was increased and the load steps with variance 2D:
was scaled via HELM, we solved for load bus voltages using
NRFP. We then plotted the difference between the HELM and s[0] = 0 (16)
the NRPF voltage magnitudes in panel (b). These results suggest √
that CPF via HELM computes load bus complex voltages with s[k + 1] = s[k] + 2D · N (1, 0). (17)
a very low degree of error for a given level of load increase. Ac-
The values which s may attain are shown in Fig. 4, where sc
cordingly, the critical loading levels computed by the numerator
corresponds to a Saddle-Node bifurcation of the algebraic power
of (13) are a sufficiently accurate approximation for the exact
flow equations. More explicitly, if the complex load at some bus
voltage collapse load values.
reaches S = S0 (1 + sc ), then the system’s voltage will collapse
(for notational simplicity, non-unity loading rates are not taken
E. Deriving a Probabilistic Loading Margin From First into account). If s is allowed to drift over a time period Δt,
Passage Processes
and an absorbing boundary condition (the point of collapse) sits
CPF (via HELM or NRPF) allows one to estimate how much at s = sc , the survival probability (SP) of the system may be
of a load margin exists between a current operating point and computed. The SP refers to the probability that that the system
voltage instability. However, it does not compute the probabil- will not experience voltage collapse due to the wandering of s
ity that a particular system will destabilize due to stochastic during Δt. As derived in [14], the SP for a system starting at
124 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 34, NO. 1, JANUARY 2019

Fig. 6. Reference Based Controller (RBC). Local terminal bus voltage V t is

the only feedback signal.

real time (PMU) load voltage observability and a controllable

Fig. 5. The interaction between base load, the fast noise u(t) of the Ornstein- reactive power resource, such as a Synchronous Condenser or a
Uhlenbeck process, and the slow noise from the randomly walking Holomorphic
scaling parameter s is portrayed. Static VAR Compensator (SVC) that can support load voltage.

s = 0 may be estimated as A. Reference Based Controller Overview

  The RBC does not rely on the Wide Area Measurement Sys-
sc
SP ≈ erf √ , (18) tem (WAMS); instead, it uses a local voltage terminal mea-
4DΔt surement Vt as a feedback signal to control the reactive power
where erf is the error function, and D is the diffusion coefficient injected by a “quasi-static SVC” device2 This relatively simple
of (17) which is based on load variability. Equation (18) gives approach to feedback SVC control is illustrated in Fig. 6. In this
the probability that the parameter s will not cross the voltage diagram, Δbsvc is the change in susceptance at the SVC and the
collapse threshold sc at any point during time Δt. regulator gain Kr is tuned to properly correlate voltage changes
Additionally, we introduce the value sm , where 0 < sm < sc . with reactive power injections. The reactive power changes are
This scalar value corresponds to the maximum allowable load limited by the size, in MVAr, of the SVC. The “BAF” block rep-
level S = S0 (1 + sm ) that can be reached before some operator resents a Buffered Average Filter that provides a rolling average
specified probability of voltage collapse grows too high. Said bus voltage magnitude over an operator-specified time window
differently, sm is chosen such that if the load starts diffusing Tw . After Tw seconds, the BAF computes average voltages and
from S = S0 (1 + sm ), the probability that it will not reach a the SVC adjusts its reactive power injection as needed. Finally,
voltage collapse load of S = S0 (1 + sc ) will be equal to opera- the “Network” block represents the physical feedback provided
tor specified survival probability SP∗ . For example, if an opera- by the natural evolution of bus voltages due to control input,
tor specifies that the system must have a survival probability of load fluctuations, and system dynamics.
at least SP∗ over future time interval Δt, then S0 (sc − sm ) will
be the load margin between the specified probabilistic threshold B. Mean Based Controller Overview
and true voltage collapse:
  Similar to the Automatic Voltage Control (AVC) system out-
∗ sc − sm lined in [24], the MBC relies not only on local terminal voltage,
SP ≈ erf √ , (19)
4DΔt but also on bus voltage magnitude data from a WAMS. These
data, sampled at 30 Hz, are also passed through a BAF with
where D is a diffusion coefficient that sets the standard deviation
time window Tw and then are each subtracted from some critical
of the stochastic load deviations per unit time. As a result, the
magnitude μcrit and summed together. The thresholds imposed
probability of voltage collapse VCP∗ is given by:
by μcrit are chosen based on minimum tolerable voltages (such
VCP∗ = 1 − SP∗ . (20) as 0.98 p.u., for example); probabilistic security margins are
not considered. As illustrated in Fig. 7, the voltage magnitudes
The connection between first passage processes and voltage V1 · · · Vn represent WAMS data from PMUs, and the gain Km
collapse is further detailed in [23]. We assume that system load is set based on how the operator wishes for the WAMS feedback
changes will incorporate both fast and slow load changes on and the local feedback to interact.
top of some base load condition. The interaction between these The “step” function in Fig. 7 (and Fig. 8) operates exactly as
fluctuations and the base load are illustrated in Fig. 5. the unit step function of equation (21). Its purpose is to ensure
that only WAMS bus voltages that are lower than the critical
III. CONTROLLER DESIGN voltage μcrit impact the feedback signal.
This section introduces a Variance Based Controller (VBC), 
x x≥0
which is subsequently shown to successfully mitigate the prob- x · u(x) = (21)
0 x<0
ability of voltage collapse. For benchmarking and illustration
purposes, we also introduce two other, more conventional, con-
trollers: a Mean Based Controller (MBC) and a Reference Based 2 In typical power system modeling, SVC devices can be dynamically modeled

Controller (RBC). For clarity, we introduce these controllers in with sets of ODEs. Since we are using the statistics of buffered time series data to
make control decisions, we have the SVC take discrete, rather than continuous,
reverse order of complexity (least to most). In actual imple- control action every T w (time window) seconds. We therefore refer to the device
mentation, both the MBC and VBC controller systems require as a “quasi-static SVC”.
CHEVALIER AND HINES: MITIGATING THE RISK OF VOLTAGE COLLAPSE USING STATISTICAL MEASURES FROM PMU DATA 125

Algorithm 1: Variance Based Controller (VBC).

START
1 Perform CPF via HELM on load pocket
2 Determine voltage collapse loading factor sc
3 Based on desired probabilistic security margin,
determine loading factor sm s.t. 0 < sm < sc
4 Computationally scale loads based on sm and then
analytically solve for load pocket critical variance
Fig. 7. Mean Based Controller (MBC). The WAMS acquire voltage magnitude 5 Use critical variances and magnitude constraints as
data V 1 · · · V n to use as feedback signals for the SVC. inputs to Fig. 8 controller
if New State Estimate Data Available then
Return to START
else
Return to 5
end

Fig. 9. 3 Bus test case. Aggregate generation (bus 1) feeds an aggregate load
pocket (bus 2) with voltage supported by a local SVC (bus 3).

system with identical simulation parameters between tests. In

order to provide a complete system description, the simulation,
Fig. 8. Variance Based Controller (VBC). Wide Area Measurement Systems control, and data files have been publicly posted online3 for
(WAMS) gather load pocket voltage magnitude data. Buffered Average Filters
(BAFs) and Buffered Variance Filters (BVFs) are used to quantify bus voltage open source access.
magnitude and variance. The step functions ensure that only critically low
magnitude and critically large variance measurements have effect on bS V C .
A. System Overview
In our three-bus test case (Fig. 9), aggregate generation is
In the context of Fig. 7, xi = μcrit − BAF(Vi ).
connected to a heavily loaded aggregate load pocket (such as a
city). The voltage magnitude of this load pocket is supported by
C. Variance Based Controller Overview a fully controllable quasi-static SVC, and a PMU feeds voltage
The VBC builds on the tools described in Section II in the fol- magnitude data back to the SVC in real time. In the case of the
lowing way. CPF via HELM is used to quickly determine how RBC, these PMU data are neglected.
much the load within a load pocket of concern may increase be- We outfit the generator with a 4th order Synchronous Ma-
fore the system undergoes static voltage collapse. Next, the First chine (SM), a 4th order Automatic Voltage Regulator (AVR),
Passage Process is used to determine the load level below which and a 3rd order Turbine Governor (TG). System parameters are
the probability of voltage collapse is sufficiently low. Using this approximately based on the WSCC 9-bus system, and com-
loading level, the critical bus voltage variances are found by ponent models are those described in [18]. At the SVC, we
leveraging the analytical covariance matrix solver along with choose a buffering time window of Tw = 3 s. The load of the
load noise estimation. Finally, these critical variances are used load pocket is constant power (PQ); the fast load fluctuations
as a feedback signal to control the reactive power injected by are described by the Ornstein-Uhlenbeck process of (5) and the
the quasi-static SVC device, as shown in Fig. 8. slow load variations are monotonically increased (see panel (b)
The VBC process is formally described in Algorithm 1. The of Fig. 11). Both fast and slow load fluctuations were applied to
BVF, or Buffered Variance Filter, is similar to the BAF in that the active and reactive power demands equally in order to hold
it computes the variance from a window of measurement data. power factor constant.
The constant Kv is a feedback gain parameter for the variance To compare the three controllers, we (i) initialized the heavily
measurements, and is tuned to allow the controller to use both loaded 3 bus system, (ii) performed a time domain simulation
the voltage magnitude and the variance feedback signals. with a stochastically increasing load, and (iii) measured the sur-
vival time achieved by each controller. Fast Ornstein-Uhlenbeck
IV. 3 BUS SYSTEM ILLUSTRATION noise is applied at each integration time step of Δt = 0.01. For
In order to illustrate and compare the effectiveness of the
controllers, we test each (RBC, MBC and VBC) on a three bus 3 https://github.com/SamChevalier/VoltageCollapse3Bus
126 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 34, NO. 1, JANUARY 2019

Fig. 10. (a) A noisy time domain signal V(t) with slowly varying equilibrium
changes (sped up for illustration purposes). A Savitzky-Golay Filter is applied
to V(t). (b) The filtered signal is subtracted from the noisy signal in order to Fig. 11. (a) The load bus voltage magnitude over the span of the simulations
generate the high frequency voltage fluctuations. This difference signal is used associated with all three controllers up to the point of voltage collapse. (b) The
to compute bus voltage variance. active power demand at the load bus (identical for all three simulations).

TABLE I
SIMULATION RESULTS SUMMARY

each simulation, we record the random fast and slow noise vec-
tors applied to the loads such that each controller experiences
identical simulation realizations.
In order to estimate the high frequency variance σ 2 of a volt- Fig. 12. The discretely measured (every T w = 3 seconds) load bus voltage
age signal whose underlying equilibrium point is constantly variance is plotted over the simulation lifespan for the VBC.
shifting due to the slow load fluctuations the real-time mea-
surements must first be detrended. To do so, we employ a 2nd
order FIR Savitzky-Golay Filter (SGF) to the voltage time se- At this point, the VBC begins to take control action due to the
ries data and then subtract the smoothed voltage signal from the extreme increases in the bus voltage variance. Since it relies
original data. This yields the zero-mean high frequency voltage only on bus voltage magnitude data, the MBC is unaware that
perturbations, as illustrated in Fig. 10. additional control action is needed and fails to maintain stability.
In Fig. 12, the bus voltage variance crosses the “critical”
threshold just before t = 400. The VBC simulation begins to
B. Simulation Results
call for increasing SVC support and thus prevents the system
With each controller, we simulated the system up until the from bifurcating at t = 490, when the MBC system fails. As can
point of voltage collapse. As previously indicated, the fast and be inferred from Fig. 2 and equation (1), the bus voltage variance
slow noise vectors for the simulation were computed and saved begins to show an exponential increase when the system load
before running each simulation, such that each controller expe- approaches the stability limit. As a result, the control signal
eas − σc rit ),
2 2
rienced an identical simulation case. Table I summarizes two associated with the bus voltage variance, Kv (σm
primary test results: the amount of time each controller kept also begins to increase exponentially. This explains the upward
the system “alive” (prevented bifurcation) and the amount of trend of the bus voltage magnitude for the VBC test during the
load increase that the system was able to sustain. Clearly, the last 100 seconds of simulation (seen in panel (a) of Fig. 11).
Variance Based Controller most effectively preserved voltage It is helpful to consider a critical point when the VBC and the
stability while load increased. To further illustrate these re- MBC take very different control actions. To do so, Fig. 13 zooms
sults, Fig. 11 shows the load bus voltage magnitude over time in on Fig. 11 to the window of time from t = 459 to t = 463.
(panel (a)) for all three controllers until the point of bifurcation In panel (b) of Fig. 13, the load fluctuations from t = 459 to
and the active power demand (panel (b)) at the load bus. t = 462 spike downwards despite a slow upward trend. Since
The results in Fig. 11 show that all three controllers take the system is operating close to the stability limit at this point,
identical action until roughly 200 seconds. At this point, the bus voltages spike high, above 1 per unit. Therefore, since the
PMU feedback signal of the voltage magnitude from the load mean voltage over the time window from t = 459 to t = 462
bus (bus 2 in Fig. 9) begins to drop low enough to warrant control appears relatively high, the MBC takes almost no control action.
action. The RBC simulation bifurcates at around 400 seconds, The VBC, on the other hand, measures an extremely high bus
but the MBC is able to maintain stability until about 490 seconds. voltage variance and thus takes strong control action, despite
CHEVALIER AND HINES: MITIGATING THE RISK OF VOLTAGE COLLAPSE USING STATISTICAL MEASURES FROM PMU DATA 127

Fig. 13. (a) The load bus voltage over a period of four seconds for two Fig. 15. Bus voltage magnitudes from simulations of the 39 bus test case with
controllers, where both controllers taken control action at t = 462 based on three different control systems, as load increases up until the point of voltage
measurements taken over the time window of t = 459 to t = 462. (b) The collapse. (a) Results from the VBC. (b) Results from MBC. (c) Includes results
associated active power demand at the load bus. for RBC. In each panel, the SVC bus voltage (bus 40) is noted.

Fig. 16. Bus voltage variance in the 39 bus test case, as load increases over
time.

Fig. 14. IEEE 39 bus system with added SVC Bus. 270 seconds longer than the MBC and 579 second longer than
the RBC. To better understand the success of the VBC, Fig. 16
the relatively high mean voltage magnitude (which is above shows the bus voltage variance and the average critical voltage
0.99 p.u.). This is but one of many examples of the VBC taking variance. Since each bus has a unique critical voltage variance,
control action when the MBC does not. As more and more SVC as computed by (II-C), for the sake of graphical clarity, only the
support is added to the system, the mean voltage magnitude average critical variance is shown.
becomes an unreliable signal for system voltage health as the Because the VBC measures the differences between the mea-
bifurcation voltage drifts closer to nominal system voltage. Bus sured and critical variances, and then scales these values by Kv
voltage variance, on the other hand, is a robust indicator of a and sums them across buses, large increases in variance (which
system’s proximity to voltage collapse. are expected as a system approaches its stability limit) lead to
very large reactive power injections, even when voltage mag-
V. 39 BUS SYSTEM TEST RESULTS nitude remains relatively “high”. These variance increases are
clearly seen in Fig. 16.
For further validation, we tested the controllers on a modified
version of the IEEE 39 bus system. As shown in Fig. 14, an
VI. CONCLUSION
SVC bus (bus “40”) was added to the system and connected
to 4 other buses to form an observable (via PMU) load pocket In this paper, we introduce and provide test results for a new
with reactive support. To test the controllers in this system, reactive power control system that uses bus voltage variance as
monotonically increasing slow load changes were applied to all a control signal to improve voltage stability. Tests of this system
load pocket buses (3, 4, 14, 15, 16, 17, and 18), in addition on a three-bus test case show that the Variance Based Controller
to fast mean reverting Ornstein-Uhlenbeck load noise. As with (VBC) can maintain voltage stability if load increases to 32.8%
the three-bus results in Fig. 11, the results clearly illustrate that above nominal, whereas a Mean Based Controller (MBC) al-
the Variance Based Controller improves voltage stability most lows for a load increase of only 27.0% above nominal, and the
effectively, relative to the reference controllers. Reference Based Control (RBC) allows for a load increase of
Fig. 15 shows the voltage evolution for the tests correspond- only 21.2%. Tests of the new control system on the 39 bus test
ing with all three controllers. The VBC deters voltage collapse case, in which load was steadily increasing, show that the VBC
128 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 34, NO. 1, JANUARY 2019

deterred voltage collapse 270 and 579 seconds longer than the [14] S. Redner, A Guide to First-Passage Processes. Cambridge, U.K.:
MBC and RBC, respectively. Both sets of results clearly show Cambridge Univ. Press, 2001.
[15] S. Greene, I. Dobson, and F. L. Alvarado, “Sensitivity of the loading
that statistical information can be valuable in reducing the risk margin to voltage collapse with respect to arbitrary parameters,” IEEE
of voltage collapse. Trans. Power Syst., vol. 12, no. 1, pp. 262–272, Feb. 1997.
Future work aims to extend the validation of the VBC to [16] G. W. Oehlert, “A note on the delta method,” Amer. Statist., vol. 46, no. 1,
pp. 27–29, 1992.
understand how it functions in the context of a larger system [17] M. Amini and M. Almassalkhi, “Investigating delays in frequency-
with more realistic load profiles. Similarly, the variance-based dependent load control,” in Proc. IEEE Innovative Smart Grid Technol.—
controller could be extended to include other types of statistical Asia, Nov. 2016, pp. 448–453.
[18] F. Milano, Power System Analysis Toolbox Reference Manual for PSAT
warning signs, such as autocorrelation. In order to provide for- Version 2.1.6., 2.1.6 ed., 2010.
mal performance guarantees for the proposed statistical control [19] C. W. Gardiner, Handbook of Stochastic Methods for Physics, Chemistry,
system, there is a need for additional studies to describe the and the Natural Sciences, 4th ed. Berlin, Germany: Springer, 2004, sec.
4.4.6, pp. 109–112.
conditions under which including voltage variance (and other [20] V. Ajjarapu and C. Christy, “The continuation power flow: A tool for
statistical) feedback in a reactive power control system leads to steady state voltage stability analysis,” IEEE Trans. Power Syst., vol. 7,
improved voltage control performance and stability. Addition- no. 1, pp. 416–423, Feb. 1992.
[21] S. Rao, Y. Feng, D. J. Tylavsky, and M. K. Subramanian, “The holomor-
ally, it would be useful to reformulate CPF via HELM to in- phic embedding method applied to the power-flow problem,” IEEE Trans.
corporate generation increase rates, derived from droop control Power Syst., vol. 31, no. 5, pp. 3816–3828, Sep. 2016.
settings, in the holomorphic voltage functions for PV buses from [22] M. Subramanian, “Application of holomorphic embedding to the power-
flow problem,” Master’s thesis, School Elect., Comput. Energy Eng.,
(11). The incorporation of these generation increase rates could Arizona State University, Tempe, AZ, USA, Jul. 2015.
allow for the computation of an even more realistic load margin [23] S. Chevalier, “Using real time statistical data to improve long term voltage
and thus better variance predictions. Finally, in moving towards stability in stochastic power systems,” Master’s thesis, Dept. Elect. Eng.,
University of Vermont, Burlington, VT, USA, Oct. 2016.
a more practical implementation of these methods, future work [24] Z. Liu and M. D. Ili, “Toward PMU-based robust automatic voltage con-
aims to understand the interaction between the controllers devel- trol (AVC) and automatic flow control (AFC),” in Proc. IEEE PES Gen.
oped in this paper and the other mechanisms which contribute Meeting, Jul. 2010, pp. 1–8.
to voltage collapse such as overexcitation limiters (OELs) and
on-load tap changing (OLTC) transformers.

REFERENCES
[1] I. Dobson, B. A. Carreras, V. E. Lynch, and D. E. Newman, “Complex
systems analysis of series of blackouts: Cascading failure, critical points,
and self-organization,” Chaos, vol. 17, 2007, Art. no. 026103.
[2] T. Ohno and S. Imai, “The 1987 tokyo blackout,” in Proc. IEEE PES Samuel C. Chevalier (S’13) received the B.S. and
Power Syst. Conf. Expo., Oct. 2006, pp. 314–318. M.S. degrees in electrical engineering from the Uni-
[3] M. Scheffer et al., “Early-warning signals for critical transitions,” Nature, versity of Vermont, Burlington, VT, USA, in 2015 and
vol. 461, pp. 53–59, 2009. 2016, respectively, and is currently working toward
[4] C. Wissel, “A universal law of the characteristic return time near thresh- the Ph.D. degree in mechanical engineering with the
olds,” Oecologia, vol. 65, no. 1, pp. 101–107, 1984. Massachusetts Institute of Technology, Cambridge,
[5] V. Dakos, E. H. Van Nes, P. D’Odorico, and M. Scheffer, “Robustness MA, USA. His research interests include stochastic
of variance and autocorrelation as indicators of critical slowing down,” power system stability, renewable energy penetration,
Ecology, vol. 93, pp. 264–271, 2012. and smart grid applications.
[6] G. Ghanavati, P. D. H. Hines, T. I. Lakoba, and E. Cotilla-Sanchez, “Under-
standing early indicators of critical transitions in power systems from au-
tocorrelation functions,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 61,
no. 9, pp. 2747–2760, Sep. 2014.
[7] D. Podolsky and K. Turitsyn, “Random load fluctuations and collapse
probability of a power system operating near codimension 1 saddle-node
bifurcation,” in Proc. IEEE Power Energy Soc. Gen. Meeting, Jul. 2013,
pp. 1–5.
[8] E. Cotilla-Sanchez, P. Hines, and C. Danforth, “Predicting critical tran-
sitions from time series synchrophasor data,” IEEE Trans. Smart Grid,
vol. 3, no. 4, pp. 1832–1840, Dec. 2012. Paul D. H. Hines (S’96–M’07–SM’14) received
[9] G. Ghanavati, P. D. H. Hines, and T. I. Lakoba, “Identifying useful sta- the B.S. degree from the University of Washington,
tistical indicators of proximity to instability in stochastic power systems,” Seattle, WA, USA, and the M.S. degree from Seattle
IEEE Trans. Power Syst., vol. 31, no. 2, pp. 1360–1368, Mar. 2016. Pacific University, Seattle, WA, USA, both in elec-
[10] R. Sodhi, S. C. Srivastava, and S. N. Singh, “A simple scheme for wide trical engineering, in 1997 and 2001, respectively,
area detection of impending voltage instability,” IEEE Trans. Smart Grid, and the Ph.D. degree in engineering and public pol-
vol. 3, no. 2, pp. 818–827, Jun. 2012. icy from Carnegie Mellon University, Pittsburgh, PA,
[11] A. A. P. Lerm, C. A. Canizares, and A. S. e Silva, “Multiparameter bifur- USA, in 2007. He is currently an Associate Professor
cation analysis of the south brazilian power system,” IEEE Trans. Power with the Department of Electrical and Biomedical
Syst., vol. 18, no. 2, pp. 737–746, May 2003. Engineering, on a secondary appointment with the
[12] P. Kundur et al., “Definition and classification of power system stability,” Department of Computer Science, the University of
IEEE Trans. Power Syst., vol. 21, no. 3, pp. 1387–1401, Aug. 2004. Vermont, Burlington, VT, USA. He is also the Vice-Chair of the IEEE PES
[13] A. Trias, “The holomorphic embedding load flow method,” in Proc. IEEE Working Group on Cascading Failures and a co-founder of Packetized Energy,
Power Energy Soc. Gen. Meeting, Jul. 2012, pp. 1–8. Burlington, VT, USA, a distributed energy software company.