Stability and Control-Annurev 2022 PDF

Annual Review of Control, Robotics, and
Autonomous Systems
Stability and Control of
Power Grids
Annu. Rev. Control Robot. Auton. Syst. 2022.5:689-716. Downloaded from www.annualreviews.org
Tao Liu,1,∗ Yue Song,1,∗ Lipeng Zhu,1,2,∗

and David J. Hill1,3
1
Department of Electrical and Electronic Engineering, University of Hong Kong, Hong Kong,
Access provided by 190.153.84.69 on 02/22/23. For personal use only.
China; email: [email protected], [email protected], [email protected]

2
College of Electrical and Information Engineering, Hunan University, Changsha, China;
email: [email protected]
3
School of Electrical Engineering and Telecommunications, University of New South Wales,
Kensington, New South Wales, Australia
Annu. Rev. Control Robot. Auton. Syst. 2022. Keywords

5:689–716
power grids, stability, network systems, distributed control, machine
First published as a Review in Advance
on November 4, 2021 learning, deep learning
The Annual Review of Control, Robotics, and Abstract
Autonomous Systems is online at
control.annualreviews.org Power grids are critical infrastructure in modern society, and there are well-
https://doi.org/10.1146/annurev-control-042820- established theories for the stability and control of traditional power grids
011148 under a centralized paradigm. Driven by environmental and sustainability
Copyright © 2022 by Annual Reviews. concerns, power grids are undergoing an unprecedented transition, with
All rights reserved much more flexibility as well as uncertainty brought by the growing pen-
∗
These authors contributed equally to this article etration of renewable energy and power electronic devices. A new paradigm
for stability and control is under development that uses graph-based, data-
based, and distributed analysis tools. This article surveys classic and novel
results on the stability and control of power grids to provide a perspective
on this both old and new subject.
689
1. INTRODUCTION
Power grids are giant, complex systems where the loads (power consumers) receive an electricity
Bus: a conductor in a supply from generators (power sources) via power transfer over the underlying power networks.
grid maintained at a Instability events in power grids are unacceptable to society. The consequent blackouts can cause
specific voltage and huge economic losses and can even threaten the security of a society. Therefore, stability and
capable of carrying a
control have been issues of fundamental importance since the birth of power grids.
high current, making a
common connection Stability assessment is challenging due to the nonlinearity and complexity of the dynamical
between several behavior of power grids. Although a time domain simulation can provide details on system tra-
components (e.g., jectories before, during, and after a disturbance, it is unable to explain why the system is stable
generators and loads); or unstable under that disturbance. The stability theory of power grids was established by using
it is also referred to as
mathematical tools from dynamical systems, such as Lyapunov’s first and second methods for local
a node in some
literature stability and estimation of stability regions, respectively. With the decentralization of power gen-
eration driven by renewable energy, the focus of stability analysis is shifting to the role of power
Line: a series
network structures and the development of distributed stability certificates with the help of more
component (e.g., a
advanced tools from network systems theory. Section 2 surveys these classic and recent results on
transmission line or
transformer) that links power system stability.
two buses in a grid; it Frequency control and voltage control are two core control issues to maintain the secure and
is also referred to as a stable operation of power grids. Traditional power system control attempts to use both central-
branch in some
ized and decentralized control and focuses on transmission systems, partly due to the properties
literature
of traditional power systems, such as centralized power generation, unidirectional power flow,
and passive electricity distribution. However, the integration of renewables changes power sys-
tems from the traditional paradigm to a completely new one with distributed power generation,
bidirectional power flow, and active electricity distribution (1). Furthermore, as renewable energy
sources and other types of dynamical devices are connected to different voltage levels, new dynam-
ical behaviors will appear all over the grid. Power system control is therefore facing new challenges
and opportunities, which we briefly review in Section 3 from a control engineering point of view
and which include frequency control, voltage control, and control of microgrids (MGs).
Essentially, conventional studies on power grid stability and control use solutions that rely
heavily on mathematical models. Despite significant progress in the past few decades, these solu-
tions may not be useful in practical grids with extremely high complexity and salient variability.
Fortunately, thanks to the wide deployment of advanced information and communication tech-
nologies in today’s grids, massive amounts of operational data recording system-wide dynamics
in real time are available for stability analysis and control. This has helped to unlock the poten-
tial of data-based ideas and approaches, especially machine learning (ML) techniques, in tackling
conventionally challenging issues encountered by model-based methods. Hence, Section 4 selects
some representative research efforts to illustrate how data-based solutions effectively work on grid
stability analysis and control.
2. STABILITY THEORY OF POWER GRIDS

2.1. Stability Categories and Associated Modeling
A power grid is a complex system consisting of buses and lines, where the buses may connect
generators or loads and the lines link the buses and form a complex network structure. Each bus
i in the grid is featured by the active power injection Pi ∈ R, reactive power injection Qi ∈ R,
and voltage phasor Vi ∠θi ∈ C, with Vi , θi ∈ R being the voltage magnitude and phase angle, re-
spectively. A line connecting bus i and bus j is represented by an admittance yi j ∈ C. The power
network structure is represented by the so-called admittance matrix Y = [Yi j ] ∈ Cn×n , where
690 Liu et al.

Yij = Yji = −yij if there is a line between bus i and bus j, Yij = Yji = 0 if there is no line be-

tween bus i and bus j, and Yii = nj=1, j=i −Yi j . Generally, power system dynamics can be described
by the differential-algebraic equations
ẋ = f (x, y), 1a.
0 = g(x, y), 1b.
where x collects the state variables, such as the rotor angles and flux decays of generators and rotor
slips of induction motors, and y collects the algebraic variables, such as the voltage magnitudes
and phase angles of nodes in the grid. Equation 1a captures the dynamical behaviors of generators
and loads, and Equation 1b refers mainly to the circuit equation regarding the conservation of
power at buses with only static components.
Power system stability is essentially a single problem, i.e., the system-wide stability; however,
treating the stability problem as such does not help to figure out the physical mechanism of the
various forms of instability that a power grid may undergo. According to the main system variable
in which the instability event is observed, power system stability is generally classified into rotor
angle stability, voltage stability, and frequency stability (2). This classification corresponds to the
concept of partial stability in stability theory (3). We can also make specific assumptions about
each of the stability subcategories and obtain simplified models that capture the key factors in the
respective stability issues.
2.1.1. Rotor angle stability. Rotor angle stability, or simply angle stability, refers to the abil-
ity of synchronous machines—synchronous generators (SGs) and motors—in a power system to
remain in synchronism after being subjected to a disturbance (2). Angle stability is usually seen
as a generator-oriented issue and depends on whether each SG in the system can maintain or re-
store its equilibrium between electromagnetic torque and mechanical torque. Instability occurs
in the form of increasing angular swings of some generators, leading to their loss of synchronism
with other generators. Angle stability is a purely short-term problem, as the time frame of interest
is several seconds. According to the nature of the disturbances, it can be further classified into
small-disturbance angle stability and transient stability (i.e., large-disturbance angle stability).
Angle stability problems occur mainly in high-voltage transmission systems, where there is a
strong decoupling between P–θ and Q–V relations (4). Therefore, it is common in angle stability
analysis to assume a constant voltage magnitude Vi for each bus and focus only on the P–θ relation.
There are two major types of models for the study of angle stability: the network-reduced model
and the structure-preserving model. Let VG and VL denote the set of generator buses and set
of buses that are without generators and may connect loads. In the network-reduced model, all
loads are assumed to be constant impedances and incorporated into the admittance matrix. Based
on the bus partitioning
VG and VL , the admittance matrix can be accordingly partitioned into
YGG YGL −1
Y= YLG YLL
r
, and we obtain YGG = YGG − YGL YLL r
YLG by Kron reduction. The matrix YGG
represents the admittance matrix of the reduced network, where VL is eliminated and only VG
remains. This then leads to the following network-reduced model:

g
Mi θ̈i + Di θ˙i = Pi − ViV j Gri j cos(θi − θ j ) − ViV j Bri j sin(θi − θ j ), i ∈ VG , 2.

j=1
where Mi > 0 and Di > 0 denote the moment of inertia and damping coefficient of the generator,
respectively, and Gri j and Bri j denote the real and imaginary parts of [YGG
r
]i j , respectively. The
network-reduced model greatly reduces the problem dimension and has been commonly adopted,
especially in early works, which were done when powerful computers were unavailable. However,
www.annualreviews.org • Stability and Control of the Power Grid 691

the presence of the transfer conductance Gri j in Equation 2 causes path-dependent terms that
make the Lyapunov function not well defined (5). Although the conductance of physical lines
Gij = Re{Yij } is nearly zero in high-voltage transmission systems, Gri j is never negligible because it
comes from both the lines and the loads.
To bypass the difficulty caused by the network-reduced model, Bergen & Hill (6) first proposed
the structure-preserving model in 1981. This model takes the following form:

n
Mi θ̈i + Di θ˙i = Pi − ViV j Bi j sin(θi − θ j ), i ∈ VG ,
j=1

n
Di θ˙i = Pi − ViV j Bi j sin(θi − θ j ), i ∈ VL ,
3.
j=1
where Bij = Re{Yij } denotes the line susceptance. In Equation 3, the second-order equation cap-
tures the swing dynamics of the generators, while the first-order equation captures the load fre-
quency behavior under angle oscillations. In this model, the physical network structure of the
power system remains intact, and hence the physical line conductance Gij is reasonably assumed
to be zero. More importantly, there exist well-defined Lyapunov functions for the structure-
preserving model. This model has become popular in recent network science studies of power sys-
tem stability (7), and many variant models have been developed that include more detailed descrip-
tions of system components, such as generator flux decays and excitation controllers. We direct
readers to References 8 and 9 for an extensive survey of the family of structure-preserving models.
2.1.2. Frequency stability. Frequency stability refers to the ability of a power system to main-
tain a steady frequency following a severe system upset that results in a significant imbalance
between generation and load (2). It depends mainly on the ability to maintain or restore the bal-
ance between system generation and load, with minimum unintentional loss of load. Instability
occurs in the form of sustained frequency swings that lead to tripping of generating units and/or
loads. Frequency stability has both short- and long-term issues. The time frame of interest spans
from several seconds to several minutes, depending on the nature of the different controllers that
kick in after the occurrence of a power imbalance.
Since the frequency and angle deviations will not be very large during the time frame of fre-
quency stability analysis, linear models are commonly adopted. A typical example is the linear
structure-preserving model, which is obtained by substituting the approximations Vi = 1 and
sin (θ i − θ j ) = θ i − θ j into Equation 3. Unlike Equation 3, where the sinusoid functions induce
a substantial amount of nonlinearity and could cause instability, such a linear model is asymptot-
ically stable in general. Actually, a major topic in frequency stability is the design of controllers
that steer the angular frequency of each bus—say, θ˙i —back to the rated value (50 or 60 Hz) as
quickly as possible after the power imbalance. Note that the term frequency control is used more
often than the term frequency stability; since the study of frequency stability is oriented to control
techniques, we do not discuss it further in this section and instead elaborate the relevant results
in Section 3.
2.1.3. Voltage stability. Voltage stability refers to the ability of a power system to maintain
steady voltages close to their nominal values at all buses in the system after being subjected to a
disturbance (2). It is largely a load-oriented issue and depends mainly on the ability to maintain
or restore equilibrium between the load demand and the power supply that is transferred from
692 Liu et al.

generators via the network. The driving force for voltage instability is usually that the loads tend
to restore their power consumption after the disturbance to levels that may be beyond the transfer
capability of the power network. Instability occurs in the form of a progressive fall or rise of the
voltages of some buses. Voltage stability has both short- and long-term issues that involve fast-
acting load components (e.g., induction motors and electronically controlled loads) and slow-
acting load components (e.g., tap-changing transformers and thermostatically controlled loads),
respectively. The time frame of interest varies from a few seconds to tens of minutes.
In the time frame of short-term voltage stability problems, the behaviors of active powers, reac-
tive powers, voltage magnitudes, and phase angles are all intertwined, which implies that the model
cannot be simplified into the form of Equation 3. We have to resort to the general differential-
algebraic equation model to study short-term voltage stability. Such a complex model inevitably
makes it very hard to figure out the mechanism of short-term voltage stability, and many questions
in this field remain open.
For long-term voltage stability, a topic of major concern is the condition that can guarantee the
existence and proper sensitivity behavior of a long-term equilibrium. Here, the proper sensitivity
behavior is indicated mainly by the voltage regularity, i.e., ∂V i

∂Pi
∂Vi
≥ 0 and ∂Q i
≥ 0 (10). The study
of long-term equilibria can be carried out using a steady-state model obtained by letting ẋ = 0
in Equation 1. Furthermore, with some simplifications and substitutions, the steady-state model
reduces to the usual power flow equation:

n
Pi = ViV j Gi j cos(θi − θ j ) + ViV j Bi j sin(θi − θ j ),
j=1

n
Qi = ViV j Gi j sin(θi − θ j ) − ViV j Bi j cos(θi − θ j ). 4.
j=1
The tuple (Pi , Qi , Vi , θ i ) satisfying the power flow equation represents a long-term operating equi-
librium, which is also known as a power flow solution. Given the Pi and Qi of each bus, power flow
analysis aims to check whether there exists a voltage solution (Vi , θ i ) and how the voltage solution
changes when a perturbation is applied to Pi or Qi . Power flow analysis can provide mechanism-
based explanations for some short- and long-term voltage instability phenomena; for example, the
voltage stability limit is indicated by a singularity point of the power flow Jacobian (11).
2.1.4. Converter-driven stability. The above stability concepts work in conventional power
grids, where the power sources are mainly fossil fuels and operate in the form of SGs that bring a
large moment of inertia. Consequently, stability problems regarding angle, voltage, and frequency
involve mainly electromechanical phenomena, the timescale of which is no less than 10−1 s. By
contrast, future grids will increasingly rely on renewable generation based on converter-interfaced
generators (CIGs), which have significantly different features. The dynamics of CIGs have a much
wider timescale that leads to cross-couplings with both the electromechanical dynamics of SGs
and the electromagnetic transients of the network (12). The CIGs may cause instability phenom-
ena over a wide spectrum, which can be classified into fast- and slow-interaction issues. Fast-
interaction instability is driven by the fast dynamic interactions of CIGs and other fast-response
components (e.g., electromagnetic dynamics of lines and SG stators). The frequency of unstable
oscillations can be up to several kilohertz (13). The models for fast-interaction stability prob-
lems need to include the inner (current) control loops of CIGs and network electromagnetics.
Slow-interaction instability is driven by the slow dynamic interactions of CIGs and other slow-
response components (e.g., the electromechanical dynamics of SGs). The models for this type of

stability problems can be obtained by incorporating the outer (power and voltage) control loops or
phase-locked loops of CIGs into the conventional differential-algebraic equation model (e.g., see
14, 15).
Single-machine
infinite-bus (SMIB) Since CIG modeling involving electromagnetics is an emerging field, and standard models
system: a system have not yet been developed, this article focuses mainly on stability and control problems related
consisting of a single to electromechanical dynamics, which have sophisticated models.
generator and an
infinite bus that are
connected via a single 2.2. Traditional Methods
line; studies of this There are two general classes of results for stability assessment: time domain simulation and sta-
type of system focus on
bility theory. Time domain simulation refers to the numerical computation of the trajectory of a
how the generator is
dynamical system. It applies to much-higher-order models than Equation 3 in order to include
synchronized with the

bus, which represents a detailed descriptions of generators and loads, and it provides rich response curves during the pre-
giant system fault, fault-on, and postfault periods. As time domain simulation is not a theme of this article, we
do not discuss it further here. In the following, we briefly review the classic results on power sys-
tem stability theory, a discussion that also serves as background for the review of recent progress
in stability theory in Section 2.3.
2.2.1. Angle stability. The most classic theory of angle stability is the equal-area criterion for
a single-machine infinite-bus (SMIB) system. For generic systems with multiple generators, the
extensions have been made mainly based on the direct method in terms of Lyapunov functions,
otherwise called (transient) energy functions. The first Lyapunov function for angle stability anal-
ysis was proposed in the 1940s (16) and derived from the network-reduced model. The transfer
conductances need to be dropped in order to obtain a well-defined Lyapunov function; however,
as mentioned above, this operation may cause significant error in the stability assessment because
it effectively ignores the real power loads.
As stated above, the structure-preserving model allowed for the construction of rigorous Lya-
punov functions. A Lyapunov function is generally a sum of energies from system components;
for example, the Lyapunov function for Equation 3 consists of the kinetic energy of generators
and the potential energy of lines. Lyapunov functions have been established for many different
power system models that contain more detailed descriptions of generator and load dynamics (9).
The direct method estimates the stability region of the stable equilibrium point as a set of initial
points at which the Lyapunov function values are less than a critical energy. The critical energy is
usually given by the energy value at a certain unstable equilibrium point (UEP), while a different
choice of UEP leads to different features of the estimated stability region. For instance, the stabil-
ity region estimated by the closest UEP method is valid for checking the system stability after any
kind of disturbance, but this estimation is also highly conservative. By comparison, the stability
region estimated by the potential energy boundary surface method, the controlling UEP method,
and the boundary of the stability region–based controlling UEP method is much less conservative
but specific to the stability check after a particular disturbance. We direct readers to Reference 17
for a comprehensive study of these methods.
The extended equal-area criterion provides another viewpoint for critical energy. This method
projects the dynamics of a generic system into an equivalent SMIB system and then determines
the critical energy by applying the equal-area criterion to the SMIB system (18).
2.2.2. Voltage stability. Voltage stability theory has been established largely based on bifurca-
tion analysis. For instance, short-term voltage collapse is closely linked to the singularity-induced
bifurcation of the differential-algebraic equation model, which is also known as the impasse sur-
face (19). In addition, saddle-node bifurcation and limit-induced bifurcation are two common
694 Liu et al.

indicators of a long-term voltage stability limit. A saddle-node bifurcation occurs when the max-
imum loadability is reached, which mathematically corresponds to a singularity point of a power
flow Jacobian (11). A limit-induced bifurcation is caused by a generator hitting its reactive power
Single-load
limit and leads to a voltage irregular power flow solution (20). As with angle stability, there is a infinite-bus (SLIB)
simple and representative system for studying voltage stability: the single-load infinite-bus (SLIB) system: a system
system. The stability limit points of the SLIB system have analytical expressions. In general cases, consisting of a single
the stability limit points are computed mainly by the continuation power flow method (21) or load and an infinite
bus that are connected
point of collapse method (22).
via a single line;
In general, the traditional methods fall into the category of node-based analysis, i.e., the role studies of this type of
of node parameters in stability problems. Research efforts have focused mainly on including more system focus on how
detailed generator dynamics in energy functions (9) or quantifying the impact of generator and an equilibrium can be
load parameters using sensitivity methods (e.g., see 23, 24). This appears to be mainly because the maintained between
the load demand
concepts of angle stability and voltage stability are connected to the behaviors of generators and
and the power supply
loads, respectively. By contrast, the influence of the power network structure on stability has been from the bus via the
paid much less attention. Power network parameters are simply taken as some coefficients in the line
respective models. The next subsection reviews some of the recent progress on stability theory,
Algebraic
focusing in particular on the role of network structure. connectivity: the
second smallest
eigenvalue of LG
2.3. Graph-Based Methods
Laplacian matrix:
The motivation of graph-based stability analysis and the obtained results on different stability LG = [Li j ] ∈ Rn×n ,
problems are discussed below. where Lij = Lji = −w ij
for (i, j) ∈ E, Lij =
Lji = 0 for (i,j) ∈/ E,
2.3.1. Power grids as dynamical networks. Let us first introduce some basic graph notations.
and Lii = − nj=1 Li j
A weighted undirected graph is denoted by G(V, E, W), where V denotes the set of nodes and
E ⊆ V × V denotes the set of edges, with |V| = n and |E| = l, and W = diag{wi j } ∈ Rl×l is the Cut set: a set of all
edges with one
diagonal matrix, with w ij being the weight of edge (i, j) ∈ E. Signed graphs are considered here,
endpoint in V1 and the
where the edge weights can be either positive or negative. other endpoint in V2 ,
According to these graph theory preliminaries, a power network induces a graph by treating denoted by E cut , where
buses as nodes and lines as edges. Furthermore, power system dynamics such as those described by V1 , V2 = φ is an
Equation 3 are essentially a nonuniform multirate Kuramoto model, where the network parame- arbitrary partition of V
such that V1 ∪ V2 = V
ters are embedded into the system dynamics. This is similar to another relevant subject, dynamical
and V1 ∩ V2 = φ
networks, which focuses on how collective behaviors are influenced by the structural properties
of the underlying network. For instance, a well-known result says that network synchronization is Effective resistance:
the effective resistance
achieved by a sufficiently large algebraic connectivity of the underlying network (25). Therefore,
between two nodes
it is natural to push the stability theory of power grids further in the direction of exploiting the i, j ∈ V, i = j, is
role of power network structure (26). In fact, some early works attempted to explore the impact of rGeff (i, j) =
network structural factors on stability problems (e.g., see 27, 28), but the lack of advanced tools at (ei − e j )T L†G (ei − e j ),
that time prevented further development. Thanks to the progress on graph theory and network where ei ∈ Rn is a
science (e.g., see 29–31), the past decade has witnessed the rise of graph-based stability analysis of vector where the ith
entry is one and all
power grids.
other entries are zero
2.3.2. Angle stability. Table 1 gives representative references related to the graph-based theory
of angle stability. It turns out that the graph Laplacian matrices and cut sets and their associated
concepts (e.g., effective resistance and algebraic connectivity) have essential importance in charac-
terizing angle stability under small and large disturbances. Here, the graph has the same topology
of the power network. Two kinds of edge weight assignments are adopted in the literature. One
takes the line susceptances as edge weights, i.e., w ij = Bij ; the corresponding Laplacian matrix is

Table 1 Representative references for the graph-based theory of angle stability
Stability category Graph concepts used in the stability condition
Small-disturbance angle stability Laplacian matrix (32, 33, 36, 38, 39)
Effective resistance (35, 36, 39)
Algebraic connectivity (34)
Cut set (33, 38)
Transient stability Laplacian matrix (40, 42)
Effective resistance (35, 36, 43)
Algebraic connectivity (40–43)
Cut set (36, 37)
simply denoted as LG in the following. The other includes the angle information at an equilib-
rium θ e ∈ Rn , i.e., wi j = ViV j Bi j cos(θie − θ ej ), which may be negative; the corresponding Laplacian
matrix is denoted as LG (θ e ) in the following.
The existence and local stability of equilibria, which are two fundamental issues in small-
disturbance angle stability, are linked to graph concepts. Dörfler et al. (32) proposed a con-
dition in terms of LG for the existence of a unique and stable equilibrium within the region
{θi | |θi − θ j | < π2 , ∀(i, j) ∈ E}. Jafarpour & Bullo (33) further established an improved version of
this condition using cut set projection. A large algebraic connectivity defined over LG also guar-
antees the existence of stable equilibria (34). For the stability of equilibria, it has been proved that
an equilibrium θ e is hyperbolic and locally asymptotically stable if LG (θ e ) is positive semidefinite
and has only one zero eigenvalue (35, 36). An equilibrium is unstable if a cut set is formed by
lines with π2 < |θie − θ ej | < 3π 2
that induce negative weighted edges (37). The number of unstable
manifolds of a UEP can be described by linear matrix inequalities with respect to LG (θ e ) (38, 39).
Moreover, the local stability of an equilibrium is indicated by positive effective resistances defined
over LG (θ e ) (35, 36, 39).
Transient stability is also closely linked to these graph concepts. For a power grid with a negli-
gible moment of inertia (i.e., Mi /Di = ε), the stability region can be estimated in terms of LG and
algebraic connectivity (40). These results were later extended to a general case with nonnegligi-
ble inertia (41, 42). It turns out that a greater algebraic connectivity leads to a greater estimated
stability region. Song et al. (35) and Dörfler & Bullo (43) explored the relationship between tran-
sient stability and effective resistance and found that a better transient stability is indicated by a
smaller positive effective resistance. In addition, the properties of graph cut sets provide a theoret-
ical explanation for the common phenomenon that transient instability is initiated from the angle
separation across a cut set in the power network (36, 37). This phenomenon is usually supposed
to be related to generator behaviors, but this new finding indicates that it is attributable to the
structural properties of the underlying network.
2.3.3. Voltage stability. So far, the progress on graph-based voltage stability theory has been
made mainly in the existence of a power flow solution and the singularity of a power flow Jacobian,
which falls into the category of long-term voltage stability. Based on fixed-point theorems such as
Banach’s fixed-point theorem and Brouwer’s fixed-point theorem, some novel conditions for the
existence of power flow solution have been established (e.g., see 44–47). The admittance matrix,
which carries the structural information about the underlying power network, appears explicitly
in these conditions and indicates how the parameters of the power network, generators, and loads
accommodate a power flow solution as a long-term equilibrium. In addition, as a common type
of voltage stability limit, the power flow Jacobian singularity is closely linked to the properties of
the admittance matrix. A classic network-related result for SLIB systems says that a SLIB system
696 Liu et al.

reaches a singularity point of the power flow Jacobian when the equivalent admittance of the
load has the same modulus as the admittance of the single line (11). Through the use of more
advanced matrix analysis, this result has been extended to power grids without PV buses (e.g.,
distribution systems), relying on the development of a general admittance ratio of power network
to loads; a generic system reaches a singularity point of the power flow Jacobian when the general
admittance ratio is unity (48, 49). This extension figures out how the parameters of the power
network, generators, and loads stress the system to a stability limit.
By contrast, the graph-based study of short-term voltage stability is rather preliminary because
the associated differential-algebraic equation model is highly complex and hard to analyze. Alter-
natively, data-based methods are becoming popular in short-term voltage stability assessment,
which we elaborate on in Section 4.
2.3.4. Stability in microgrids. An MG is a low- or medium-voltage power grid consisting of

distributed energy resources (DERs) (e.g., renewable CIGs and energy storage systems) and loads.
It can be connected to a utility grid (i.e., the grid-connected mode) or operate autonomously to
provide electricity to local consumers (i.e., the islanded mode) (50). When working in the grid-
connected mode, it can manage its own power balance, either delivering power to or absorbing
power from the utility grid as a generator or load. When working in the islanded mode, it can
continue to supply power to local loads in case of utility grid failures and hence can increase the
system resiliency.
Apart from the specific issues related to CIGs, MGs also inherit common stability issues from
conventional power grids regarding angle, frequency, and voltage dynamics (51). Hence, it is nat-
ural that MG stability has also been revealed to be closely connected to graph concepts such as the
algebraic connectivity of the conductance matrix (i.e., the real part of the admittance matrix). A
greater algebraic connectivity indicates a better stability in MGs (52, 53). In addition, Song et al.
(54) proved that an MG loses stability in the case of zero algebraic connectivity, and continuing
to add CIGs into the MG using a tree-like topology makes the algebraic connectivity approach
zero. This implies that the tree-like connection, which is the simplest and cheapest way to inte-
grate CIGs, is harmful to MG stability. By contrast, adding more lines to form loops in the MG
network, which helps to maintain the algebraic connectivity, is beneficial to stability.
Overall, the graph-based theory of angle stability has been established in a systematic manner,
and graph-based ideas are receiving increasing attention in voltage stability and MG stability.
Future results in this direction are expected not only to add to stability theory but also to facilitate
new control strategies by means of grid flexibility.
2.4. Further Discussion

Apart from the aforementioned methods, some new trends are also emerging in the stability anal-
ysis of power grids, as shown below.
2.4.1. Distributed stability evaluation. The idea that a stability evaluation is supposed to be
done by centralized computation is almost never questioned. This is reasonable for a high-voltage
transmission grid because the stability condition usually requires system-wide data, and a control
center exists to collect all the required data and execute efficient computations. However, the
proliferation of distributed generation is decentralizing power grids into a huge number of small
autonomous subsystems. In this case, centralized computation is no longer adequate because it
is neither effective in communicating with the numerous subsystems nor friendly to the privacy
preservation of the subsystems.

In this context, distributed stability analytics become a more promising solution. So far, the dis-
tributed stability conditions have been established mainly by exploiting the structural information
carried by the system Jacobian matrix or Lyapunov function. For instance, using the observability
property, one can reconstruct the system Jacobian based on the data that an agent collects from
the distributed communication (55). The passivity of interconnected systems (56) and separable
Lyapunov functions (57–59) are applied to decompose system-wide conditions into a series of
conditions with respect to each subsystem. In general, the distributed stability criteria are derived
at the cost of increasing conservativeness. Sometimes special controllers are required in order to
meet the criteria (56, 59). More research efforts need to be put into reducing the conservativeness
of distributed stability criteria.
2.4.2. Beyond equilibrium-based analysis. Conventional stability analysis is oriented to

equilibria—that is, it requires the existence of a stable equilibrium point and focuses on the system
behavior after applying a disturbance to the equilibrium. Such an equilibrium-based paradigm may
become less adequate for future grids using renewable energy sources (60). The uncertain nature
of renewable energy means that it keeps injecting continuous disturbances into power grids, so
that the system may never reach a state of equilibrium. In this case, the stability analysis needs
to go beyond the existing equilibrium-based paradigm to capture this feature. Some recent works
have started rethinking the paradigm, replacing equilibria with consensus and related concepts in
the analysis. For instance, Zhu & Hill (41, 61) characterized the impact of continuous disturbances
on stability using a positively invariant set in terms of node and network parameters. To move to-
ward environmentally friendly power grids that use renewable energy sources, a new paradigm for
stability analysis deserves more attention.
3. POWER GRID CONTROL

To enhance stability or to stabilize the system after disturbances, various types of control methods
have been developed. Since angle stability is a fast, dynamic phenomenon with a timescale from
milliseconds to a few seconds, only limited local fast control actions (e.g., fast valving or the use of
braking resistors) have been available to prevent instability. Some more advanced exceptions are
excitation control using feedback linearization, preventive control, emergency control, generation
shedding, and controlled islanding against cascading events (for some examples, see 62–66), which
are becoming more common in practical power system control. Thus, this section focuses on
frequency control and voltage control. Of course, as more fast control devices and methods have
been used and developed (e.g., energy storage system and demand response), new control methods
to deal with angle instability are expected.
In addition, traditional power system control is usually developed for high-voltage networks,
because dynamic devices (such as SGs) are connected mainly to these networks in a traditional
power system. As more dynamic devices are connected to the grid at different voltage levels, dy-
namic issues will appear all over the grid. The development of the concept of MGs has brought
many advantages, including the integration of renewables and improved system resilience. In fact,
an MG can provide a local control platform to solve problems encountered within the MG locally.
However, due to the special properties of MGs, frequency and voltage control issues in MGs are
quite different from those in high-voltage networks. Therefore, this section also discusses con-
trol issues in MGs. We briefly review some aspects of frequency control and voltage control for
transmission and subtransmission networks and control of MGs from a control engineering point
of view. As we cannot cover everything in this field, we discuss only some aspects that we find
interesting.
698 Liu et al.

3.1. Frequency Control
The power system frequency, whose deviation indicates the power imbalance between generation
and demand, must be maintained almost constantly at a nominal value (50 or 60 Hz) during normal
conditions and must be kept within a prespecified threshold under abnormal conditions (e.g., the
trip of a large SG or the loss of a transmission line). Otherwise, underfrequency protection will
start load shedding to restore the frequency, and customers will be affected. If this process cannot
stop further frequency drops, the protection system disconnects large generation units to protect
SGs from physical damage; this in turn enlarges the power deficiency and may lead to cascading
failures and blackouts (67).
3.1.1. Traditional frequency control. To maintain the system frequency, traditional power
systems focus on generator-side control and make generation follow demand by including a so-
phisticated control mechanism with a long-term spinning reserve plan and a three-layer hierar-
chical control structure comprising primary, secondary, and tertiary frequency control (68). The
state-of-the-art load forecast techniques provide day-ahead (in hours) and near-real-time (in min-
utes) demand predictions. The former help the control center to decide which units should be
turned on at which generation level (i.e., the unit commitment). In close to near real time, the up-
dated near-real-time load forecast is used in the tertiary frequency control, or economic dispatch,
to periodically update set points of generators. The primary frequency control, or droop control,
stabilizes the frequency to a new equilibrium point after a contingency. The secondary frequency
control, or automatic generation control, then gradually drives the frequency back to the nominal
value by automatically adjusting the set points of some selected generators.
3.1.2. New issues. With the high penetration of renewables, the traditional generation-
following-demand control paradigm may be inadequate to regulate frequency due to the high
intertemporal variation and limited predictability of renewables (69). Under this circumstance,
the system uncertainties come from both demand and generation, neither of which can be pre-
dicted without errors. Moreover, because renewable generators are usually connected to the grid
via power converters, which usually do not provide inertia to the system, high use of renewables
may also reduce the total system inertia and make a system more sensitive to contingencies. As
more SGs are replaced by CIGs, the system inertia will be reduced significantly and may even
vary with respect to time. Ulbig et al. (70) showed that the inertia of a system with high use of
renewables decreases significantly and that the frequency with traditional generator-side control
will exceed the acceptable region during the transient response after a contingency, which can
be handled by the same control method for the system with no renewables. We refer readers to
References 60 and 71 for surveys of new issues caused by renewables.
3.1.3. New methods. To address these issues, new control methods and devices have been de-
veloped. First, for the generator side, new control methods have been developed for SGs, such
as event-triggered control to replace the traditional periodic sampling mechanism of automatic
generation control (72) and consensus-based control to replace automatic generation control (73).
Recently, wind turbines have also been used to help frequency regulation by optimally controlling
their rotor speeds and pitch angles (e.g., 74). Different control methods have also been designed
for CIGs, such as droop control, the virtual synchronous machine strategy, and virtual oscillator
control. More details on CIG control are provided in References 60 and 71 and references therein.
Energy storage devices have a high ramping capability and thus can help to eliminate fast fre-
quency deviations caused by renewables. For primary frequency control, energy storage devices

can provide droop control service and synthetic inertia (75). They can also participate in secondary
frequency control, where the key problem is how to fully consider different characteristics of SGs
and storage devices. Usually, the load signal is decomposed into a slow-moving but possibly bi-
ased component and a fast-moving but energy-limited one, and SGs and storage devices are used
to track these two components, respectively. In addition, how to consider different characteris-
tics of different storage devices (e.g., energy-to-power ratios) is an issue that deserves attention.
Mégel et al. (76) obtained some preliminary results in this direction, formulating the problem as
an optimization problem and then solving it in a distributed way.
Due to properties such as instantaneous response, economic efficiency, and distributed avail-
ability throughout the grid, load-side control (demand response) has received an increasing
amount of attention (77). Smart loads, such as loads with electric springs (78), have the ability
to adjust their active power continuously to follow the change in generation, leading to a new
control paradigm in which demand follows generation. Various types of load-shedding methods
have been successfully used in practice to restore the frequency in extreme situations (79); the dif-
ference between load shedding and load-side control is that the latter aims to secure the system’s
operation while minimizing the impact on end users. Two goals must therefore be considered
simultaneously: One is to be fully responsive, with high-resolution system-level control across
multiple timescales, and the other is to be nondisruptive, with an imperceptible effect on end-use
performance (77). Liu et al. (80) proposed a novel distributed switching control method to balance
the two conflicting targets, where the designed load-side controller can work in either a frequency
restoration mode or a load restoration mode determined by a well-designed switching signal.
Power systems traditionally have a hierarchical structure consisting of transmission, sub-
transmission, and distribution systems. Frequency control problems are usually considered at
the transmission level, with loads that can be the aggregation of loads in a subtransmission
network. Similarly, a load in a subtransmission network can be the aggregation of all loads
in the related distribution systems. How the control signal obtained in transmission systems
can be split down to the subtransmission systems and then to the distribution systems needs to
be carefully investigated. Addressing this issue may require using the idea of granular control,
where all controllable devices need to work together and behave exactly as required by the
higher-level controllers. However, due to the higher resistance/reactance ratio in these networks,
active power changes may also cause voltage problems, and thus both frequency and voltage must
be considered simultaneously. Zhang et al. (81) have reported some preliminary results along
these lines, using a distributed leader–follower consensus algorithm for frequency control and
decentralized voltage control to guarantee bus voltages.
A related issue is the dynamical model for each aggregate load. Loads are dynamical systems.
With CIGs, storage devices, and a load-side controller involved, the aggregate load dynamics will
be enhanced and may change quickly (82). Therefore, they will play an increasingly important
role in frequency control, and how to build an accurate load model deserves attention.
3.2. Voltage Control

The primary target of voltage control is to maintain bus voltages in a power system within ex-
pected ranges under diverse operating conditions. Various control actions can achieve this result,
such as generation rescheduling, excitation adjustment, reactive power compensation, on-load tap
changer regulation, and emergency load shedding (11).
3.2.1. Traditional voltage control. Traditional voltage control schemes can be categorized into
two classes based on the system status: voltage stability control, which is used in stressed or risky
700 Liu et al.

situations (11, 83), and automatic voltage control, which is used in normal conditions (84, 85). Volt-
age stability control aims to address the aforementioned voltage stability problems. During online
operation, if a system is predicted to be insecure with respect to potential contingencies, preven-
tive control strategies such as generation rescheduling can be used to regain sufficient security
margins. Since it may be impossible to preventively stabilize the system against all contingencies,
corrective control is used as a complementary countermeasure and is expected to be automati-
cally executed quickly after an emergency situation occurs. For example, although load shedding
can significantly affect the power supply, this type of action can effectively save the system from
voltage collapse induced by either small or large disturbances (86).
Automatic voltage control is devoted to optimally regulating system-wide bus voltages during
normal operation. Similar to traditional frequency control, automatic voltage control can be im-
plemented in a three-layer hierarchical control structure that includes primary control, secondary
control, and tertiary control (84). When automatic voltage control is used, the system is far from
an emergency situation and thus can be gradually governed toward optimized voltage profiles
via mild actions such as generation rescheduling, on-load tap changer adjustment, and voltage or
reactive power control (87).
3.2.2. New issues. Renewables also cause new voltage issues for power systems. In particular,
variable renewable generation may cause fast and dramatic voltage changes. In subtransmission
and distribution systems, such changes may be larger than those in transmission systems due to
their higher resistance/reactance ratios (88). Moreover, the reverse power caused by renewables
may lead to voltage rise issues and cause incorrect actions of protective devices and disconnec-
tion of equipment. This subsection therefore focuses on voltage control in subtransmission and
distribution systems.
Traditional voltage control in these networks uses on-load tap changers and capacitors that
have discrete steps (89) but may be inadequate for new voltage issues. First, on-load tap changers
and capacitors may not be fast enough for voltage regulation under the new circumstances (90)
due to their discrete control nature. Second, they may undergo excessive numbers of operations,
which may reduce their lifetimes significantly (91). Thus, new control methods are needed to deal
with these new issues.
3.2.3. New methods. As in frequency control, CIGs, storage devices, and flexible loads can all
participate in voltage control. In fact, these three types of devices, known as DERs, can provide fast
active and reactive power support for voltage regulation through the control of electronics that
interface them with the grid. Problems in this area include how to coordinate DERs scattered in a
network and how to cooperate with the existing voltage control devices (e.g., on-load tap changers
and capacitors). Load-side control provides an example. The two goals of load-side control dis-
cussed above in relation to frequency issues also apply here. On the one hand, traditional voltage
control devices have slow control actions; on the other hand, load-side control needs to minimize
the impact on end users. To deal with these issues, Tang et al. (92) proposed a novel coordinated
control framework to handle negative voltage impacts in a weak subtransmission system, where
traditional voltage controllers are used to regulate voltage deviations in the system all the time and
the load-side controller is activated only when needed—for example, when the traditional voltage
controllers cannot keep voltages within the expected ranges.
Another issue with using DERs for voltage control relates to the hierarchical structure of power
systems, where a subtransmission system may have many different DERs scattered in different dis-
tribution systems. It may be hard to control these DERs directly using the subtransmission volt-
age controller. Again, the idea of granular control may help. For example, Tang et al. (93) treated

distribution systems with DERs as clusters that serve as intermediaries between small DERs and
the subtransmission voltage controller. Due to the high resistance/reactance ratio, the active power
changes of DERs can also cause serious voltage issues in distribution systems. To address this is-
sue, the authors developed a novel two-module distributed control framework where a command-
following module is used to track the active power change command from the subtransmission
voltage controller by coordinating the active power of the DERs, and a power quality manage-
ment module is used to improve the power quality by coordinating both the active and reactive
power of the DERs.
Voltage issues in distribution systems can also be handled by network reconfiguration. For
example, Song et al. (94) formulated a new reconfiguration model that minimizes the network loss
and restricts the voltage volatility indices with the coordination of switched capacitor banks, which
provides some new insights regarding voltage issues in distribution systems with high renewable
penetration.
3.3. Control of Microgrids and Networked Microgrids

The concepts of MGs and networked MGs refer to key modules for future power grids where
distribution systems with distributed generators can be clustered into connected MGs. By con-
trolling a huge number of connected MGs properly, keeping a local power balance within each
MG, and making them help each other, we may solve all problems locally. The key problem then
becomes how to operate a stable MG and a networked MG.
3.3.1. Control of microgrids. To safely operate an MG, we need instantaneous control to stabi-
lize the MG when it is subjected to disturbances (i.e., controller design). To economically operate
an MG, we need to optimally manage its controllable devices within a certain time period (i.e., en-
ergy management system design). We must also integrate these ideas to form an automatic control
system that includes an upper-layer energy management system and a lower-layer instantaneous
control system. When operating an MG, a hierarchical control structure is usually adopted that
combines centralized and decentralized methods—that is, controlling the MG requires a control
center (95). This control method was originally developed for high-voltage power systems and
may reduce control performance or even jeopardize stability if the distinctive characteristics of
MGs are not properly considered.
Frequency and voltage are the two core interacting issues in ensuring stable MG operation.
Many works have focused on the controller design of CIGs. If used alone, this generation-type
control may have some limitations. First, CIGs such as wind and solar plants are usually connected
to the system via electronic devices that do not provide any system inertia. To overcome this issue,
droop control [e.g., P–f and Q–V droop control (96, 97)] is designed to maintain an instantaneous
power balance. Achieving better transient performance requires high droop gains, which may
jeopardize stability (98). Second, distributed renewable sources have a limited control capacity,
which limits the MG’s ability to survive large disturbances (99).
Load-side control and ancillary devices (e.g., energy storage and reactive power compensators)
offer possible ways to address these issues. In particular, controllable loads [e.g., voltage-dependent
loads (100) and loads with electric springs (78)] are used to regulate frequency by adjusting load
bus voltages within certain ranges. These ideas create a new way to regulate frequency and voltage
simultaneously by using load-side control. However, Lee et al. (78) showed that a cooperative
control method using a group of smart loads may have positive effects on system performance,
whereas a decentralized control method may lead to negative impacts and even instability. This
issue may also apply to the control of the other types of controllable devices in MGs. Therefore, a
702 Liu et al.

cooperative control method is needed to make controllable devices work properly in a coordinated
way.
In contrast to large-scale power systems, it is difficult to implement a supervisory control center
in MGs due to issues such as the high cost of communication lines (101). The plug-and-play nature
of MGs (102) also limits applications of centralized algorithms that are sensitive to changes. To
meet these new requirements, distributed algorithms have been extensively used to address issues
in MGs (for examples, see 55, 103, 104). Combining the hierarchical control idea with distributed
algorithms should enable the development of an automatic control system under which MGs can
operate stably and economically.
3.3.2. Control of networked microgrids. A single MG usually has limited operational flexi-
bility (105), which may limit its ability to tolerate large disturbances. A feasible way to solve this
problem is to connect nearby MGs and form a larger system (99); the key problem is then how to
make these MGs help each other without introducing new problems.
Different aspects of the concept of networked MG systems have begun to attract attention,
including self-healing (106), energy trading (107), and load sharing (108). However, most works
in these areas have focused on energy management issues. Liu et al. (109) studied the frequency
control problem in a two-MG system using voltage-dependent loads. The results showed that the
control performance of the networked MGs is significantly better than the performance of each
individual MG when they are subjected to the same disturbance, and the networked MGs could
survive a larger disturbance by helping each other maintain frequency and bus voltages within the
allowable ranges. However, important control issues remain open—for example, how to optimally
manage power trading between MGs and regulate the tie-line power between connected MGs for
stable operations. Again, the integration of hierarchical control and distributed algorithms may
provide a way to solve these problems.

Most of the traditional control methods are model based. Strictly speaking, every model is wrong
to some extent because of the complexity and many types of uncertainties involved, includ-
ing both structured and unstructured uncertainties, which may influence model-based control
performance.
Economic mechanisms have also led to an increase in stresses on the grid. For example, the
growing use of renewables increases the complexity of power balance control, and the use of
rooftop solar power and batteries in the distribution systems changes customers from pure power
consumers to prosumers who can also supply power to the grid. Solving these problems requires
new concepts, such as transactive energy—a system of economic and control mechanisms that use
value as the key operational factor and allow a dynamic balance between generation and demand
across the whole power system (110). All of these developments call for new ideas and methods,
which may lead to approaches such as combining big data with traditional control and game theory.
4. DATA-BASED STABILITY ANALYSIS AND CONTROL

In today’s power grids, with the availability of huge volumes of operational data acquired by ad-
vanced sensors such as phasor measurement units (PMUs) (111), data-driven solutions for power
grid stability analysis and control hold great promise. In particular, the advancement of ML tech-
nologies, especially the emerging deep learning (DL) approaches (112), has made such data-driven
solutions more tractable and powerful than ever before. In the following, we focus on how ML

approaches are leveraged to address various problems of power grid stability analysis and control
from a data-driven perspective. While other data-based methods that do not include explicit learn-
ing (e.g., time series analysis, signal processing, and data-enabled model predictive control) have
Environment: for
power grids, the actual also been developed in recent years, they fall outside the scope of this section. Unlike the sections
operational context of above that reviewed existing studies on model-based power grid stability and control based on
the grid or the virtual specific stability issues and the mathematical models adopted therein, here we group data-based
context specified by solutions into different classes for detailed review according to the inherent nature of the different
the grid’s numerical
ML methods. As shown below, ML methods of the same class can be exploited to address different
simulation model
stability issues once they are appropriately designed.
Dynamic stability/
security assessment
(DSA): a method of 4.1. Overview of Machine Learning–Based Solutions in Power Grids
quickly and reliably

The ML field comprises two basic classes of methods: supervised learning and unsupervised learn-
performing power grid
stability or security ing (113). While the former tries to learn a mapping from input–output pairs (e.g., classification
analysis, especially for and regression), the latter aims to uncover the inherent data distribution and characteristics of
online monitoring the inputs in the absence of outputs (e.g., clustering). If some learning instances are labeled with
contexts, to enhance explicit outputs, semisupervised learning can be performed to learn knowledge from both labeled
the grid’s awareness of
and unlabeled data. In addition to these ML approaches, there exists another special class of ML
risky situations
methods called reinforcement learning (RL) (114). In general, RL strives to train an agent by
iteratively exploring and interacting with a given environment.
Here, we further classify all of these methods into two major categories according to their
learning data flows: feedforward learning (which includes supervised, unsupervised, and semisu-
pervised learning) and interactive learning (which includes RL). Figure 1 illustrates how these
categories work in the context of power grids. As can be seen in the figure, both of them involve
two basic phases: offline learning and online application. Because most of the computational bur-
dens are concentrated on offline learning, the online application phases incur almost negligible
computational costs, which often enables ML-based solutions in power grids to achieve extremely
high online efficiency. This is a significant advantage over conventional model-based methods,
especially when the latter involve complicated optimization. In addition, with no limitations on
system scale or complexity, ML-based solutions are more applicable in practical contexts and can
therefore be extensively deployed in various grids. For detailed applications, although feedforward
ML can be widely applied to both grid dynamic stability/security assessment (DSA) and control,
interactive RL is generally more suitable for control issues, as shown below.
4.2. Data-Driven Power Grid Dynamic Stability/Security Assessment

Given a specific system (grid) with n major buses, data-driven DSA generally aims to quickly pre-
dict its stability/security status and/or margin after encountering a certain fault. Specifically, the
prediction is carried out by inferring the input–output mapping F : x → y for x = {x1 , x2 , . . . , xn }.
Here, x denotes the initial system states observed in prefault or early postfault stages, and y repre-
sents the eventual stability/security status or margin of the dynamic process induced by the fault.
For the state observed at bus i (1 ≤ i ≤ n), i.e., xi , it can be either measured/featured data points
obtained from specific snapshots of the system dynamics or time series trajectories sequentially
acquired from a given time window. To comprehensively capture the system dynamics, multiple
variables at bus i, such as the bus voltage magnitude, phase angle, frequency deviation, active power,
and reactive power (V, θ, f, P, and Q, respectively), can be gathered into xi . The output y can be
either a binary label indicating the stability/security status or a continuous value quantifying the
stability/security margin. Based on the means of feature extraction or learning from x, we divide
the reviewed data-driven DSA studies into three groups, as presented in Table 2.
704 Liu et al.

a Feedforward machine learning–based solutions in power grids
1 Offline learning 2 Online application
Data Model Model

Historical integration training Data-driven application Online decision-
Learning data set
operation records inference model making
Data Real-time
generation data acquisition
Numerical Practical
simulation model power grid
b Interactive reinforcement learning–based solutions in power grids

1 Offline learning 2 Online application
Data Data
Historical System dynamic feeding Data-driven feeding System dynamic
operation records states intelligent agent states
Offline data Interactive Agent Real-time

collection learning application data acquisition
Scenario Action Action

reproduction Numerical execution Adaptive execution Practical
simulation model action scheme power grid
Figure 1
(a) Feedforward machine learning–based and (b) interactive reinforcement learning–based solutions in power grids.
Table 2 Representative references for data-driven power grid dynamic stability/security assessment
Means Stability issues Learning methods
Snapshotted feature engineering Transient stability (115, 116) Decision tree (115)
Short-term voltage stability (117, 118) Random forest (119)
Long-term voltage stability (119, 120) Support vector machine (116, 119)
Frequency stability (121) Multiple linear regression (120)
Artificial neural network (119)
Extreme learning machine (117, 121)
Random vector functional link (118)
Temporal dependency learning Transient stability (122, 123) Convolutional neural network (122)
Short-term voltage stability (124, 125) Recurrent neural network (123)
Long-term voltage stability (126) Shapelet learning (124–126)
Spatial–temporal correlation Transient stability (128) Graph convolutional network (128, 129)
learning Short-term voltage stability (129–131) Recurrent neural network (130)
Shapelet learning (131)

4.2.1. Snapshotted feature engineering for dynamic stability/security assessment. This
line of work generally takes system states acquired from single snapshots or handcrafted feature
values calculated from specific snapshots as inputs to learn system stability patterns. Following this
basic idea, data-driven DSA has been carried out for various stability issues, including transient
stability (115, 116), short-term voltage stability (117, 118), long-term voltage stability (119, 120),
and frequency stability (121). With individual dimensions of inputs presented in the form of single
values, a wide variety of classic ML algorithms can be employed to derive DSA models, including
decision trees, random forests, support vector machines, multiple linear regression, multilayer
perceptron–based artificial neural networks, and randomized artificial neural networks such as
extreme learning machines and random vector functional links. These methods generally have
shallow learning structures, which fit well with snapshotted inputs involving flat data structures.
As these efforts merely take data points or featured values obtained from individual snapshots
of system dynamics as independent inputs for learning, they could fail to capture the inherent
evolution features hidden behind system dynamic trajectories. Without sufficient feature learning,
it may be difficult for the obtained DSA models to generalize to online application contexts, where
various unforeseen operation scenarios with complicated system dynamics constantly occur.
4.2.2. Temporal dependency learning for dynamic stability/security assessment. In con-

trast to the above data-driven DSA efforts, which learn information from certain snapshots in-
dependently, approaches based on temporal dependency learning take temporal dependencies or
correlations within the system dynamics into sufficient account for stability feature learning. In
particular, assuming that the intrinsic temporal dependencies are well preserved in system time se-
ries trajectories, they take multidimensional time series data acquired from individual buses as the
raw inputs to derive DSA models. To learn from such consecutive time series inputs, they generally
employ emerging DL techniques with a stronger capability in feature learning and representation,
such as convolutional neural networks (122) and recurrent neural networks (123). Due to their ad-
vantage in uncovering stability features from large-disturbance-induced system dynamics, these
DL-based methods have been widely applied to address stability issues on transient timescales.
However, we would like to note that, as the DSA models derived from DL are essentially black
boxes with no interpretability, it would be difficult for system operators to comprehend how they
work effectively in practical DSA.
Given the need for both temporal dependency learning and interpretable model derivation, a
new explainable learning approach called shapelet learning has been introduced in power system
online DSA in recent years (124–126). Similar to the well-known wavelet concept, shapelets are
local subshapes (subsequences) of time series that are discriminative enough to clearly distinguish
different classes of time series instances (127). The shapelets are first extensively searched from
all the time series instances and then taken as critical features to measure the differences between
individual time series instances via distance values. Doing this transforms the raw time series data
into flat distance data, which can be compatibly handled by various standard ML algorithms for
DSA model derivation. As the shapelets are essentially representative time series subsequences,
they are quite interpretable and can help provide intuitive insights into the critical evolution pat-
terns of system stability.
4.2.3. Spatial–temporal correlation learning for dynamic stability/security assessment. In

addition to paying special attention to temporal dependencies, efforts on spatial–temporal corre-
lation learning further take the inherent correlations between buses in the grid into explicit ac-
count. Specifically, considering the irregular graph-like structure of the power grid, some of them
apply a promising DL method called graph convolutional networks to tackle the online DSA issue
706 Liu et al.

(128, 129). With the help of such a network, the irregular power grid structure is embedded into
the learning procedure to account for networked spatial correlations between individual buses;
conventional convolutional neural network or recurrent neural network modules are further added
to enable temporal dependency learning. Recently, Zhu et al. (130) made a different attempt by
formulating spatial correlations as attention factors to govern recurrent neural network–based
temporal dependency learning.
On the basis of the shapelet learning approach, Zhu et al. (131) developed an interpretable
spatial–temporal feature learning scheme. By quantifying spatial correlations with geospatial in-
formation on practical substations (buses), this scheme constructed wide-area voltage animations
to incorporate system spatial information with temporal dynamics. The obtained DSA model can
interpret system voltage stability characteristics from a wide-area spatial–temporal perspective.
However, its reliability depends greatly on the precision of the geospatial information, which may
not always be guaranteed in practice.
4.3. Data-Driven Power Grid Control

When a data-driven DSA model identifies system instability or insecurity, corrective or preventive
control actions should be taken as soon as possible to restore stable or secure operation. In addi-
tion, in normal operating conditions without large disturbances, proper control measures should
be used to maintain variables such as bus voltages and frequencies within desired ranges. Similar
to the above data-driven DSA task, these control problems can be addressed from a data-driven
perspective. Table 3 summarizes representative studies involving data-driven power grid control
in three classes.
4.3.1. Feedforward machine learning–based solutions. Feedforward ML-based solutions re-

semble the data-driven DSA efforts reviewed above, with the potential of ML methods in mapping
inference sufficiently explored to figure out adaptive online control strategies. These solutions
have been implemented to tackle diverse control issues, including transient stability control (132,
133), voltage stability control (134, 135), and frequency control (121). Among them, various ML
algorithms have been employed to help derive data-driven control schemes, including decision
trees, support vector machines, artificial neural networks, extreme learning machines, and recur-
rent neural networks. The most straightforward way to implement these schemes is to directly
learn the mapping between system dynamic states and the optimal control actions (121, 134).
Table 3 Representative references for data-driven power grid control

Category Control problems Learning methods
Feedforward machine Transient stability control (132, 133) Decision tree (132)
learning–based approaches Voltage stability control (134, 135) Support vector machine (133)
Frequency control (121) Artificial neural network (134)
Extreme learning machine (121)
Recurrent neural network (135)
Interactive reinforcement Transient stability control (136) Q-learning (139)
learning–based approaches Voltage stability control (136) Deep Q-network (136)
Automatic voltage control (137) Soft actor–critic (137)
Frequency control (138) Deep deterministic policy gradient (138)
Wide-area damping control (139)
Other learning approaches Automatic voltage control (141, 142) Approximate dynamic programming (143)
Wide-area damping control (143) Knowledge search (141, 142)

Alternatively, by constructing a stability/security space with decision variables, a risky case can be
moved to the stable/secure region by properly adjusting the decision variables, which implicitly
yields the control strategy (132, 133). Zhu & Luo (135) recently developed a promising recurrent
neural network learning approach that incorporates the basic idea of model predictive control
into ML. By predicting system voltage changes after taking potential control actions, it can ef-
ficiently minimize the cost of control actions while satisfying voltage recovery requirements for
voltage stability control. However, we should mention that these solutions either rely heavily on
the preparation of a high-quality case base with optimal control experiences (121, 134) or depend
largely on the precision of system dynamics prediction for control strategy derivation (132, 133,
135).
4.3.2. Interactive reinforcement learning–based solutions. As illustrated in Figure 1b, in-

teractive RL is well suited to adaptive control in power grids. Let a system trajectory obtained
from the interactive procedure shown in Figure 1b be described as τ = {s0 , a0 , s1 , r1 , a1 , . . . , sT − 1 ,
rT − 1 , aT − 1 , sT , rT }, where si and ai are the observed system state and executed control action at
the ith (0 ≤ i ≤ T) time step, and ri + 1 (0 ≤ i ≤ T − 1) is the reward obtained by executing ac-
tion ai after observing system state si (with the system state changed to si + 1 ). The aim of RL is
to derive an optimized decision strategy π ∗ (a|s) that can maximize the accumulated rewards—i.e.,
−1 i
R(τ ) = Ti=0 γ ri+1 , where γ ࢠ [0, 1] is a discount factor that determines how important future
rewards are with respect to the current state. In power grids, relating the rewards R(τ ) to differ-
ent electrical variables for control performance optimization enables various control problems to
be handled, including transient stability control (136), voltage stability control (136), automatic
voltage control (137), frequency control (138), and wide-area low-frequency oscillation damping
control (139).
Classical RL algorithms such as Q-learning (114) generally define system states and actions in
discrete spaces, which would limit their applications in practical grids with diverse continuous vari-
ables. In recent years, by systematically combining DL with RL, an advanced RL paradigm called
deep reinforcement learning (DRL) with a stronger capability in learning representation and gen-
eralization has exhibited great potential in many complex control and decision-making problems,
including the game Go (140). Given the significant advantages of DRL, some typical algorithms
(e.g., deep Q-networks, the soft actor–critic algorithm, and the deep deterministic policy gradient
method) have been recently introduced into power grids and provided excellent performance on
various control tasks (136–138). However, RL/DRL approaches generally must learn the opti-
mized control strategies via numerous trial-and-error interactions, which makes them computa-
tionally costly. For practical large-scale grids involving tens of thousands of buses and transmission
lines, high-performance computing platforms would be needed.
4.3.3. Other learning approaches. Following the basic learning idea in ML-based solutions,
some related learning schemes have been developed for power grid control in recent decades
(141–143). Among them, a popular approach called approximate dynamic programming has been
employed to address typical control tasks such as low-frequency oscillation suppression (143). In
fact, approximate dynamic programming works in the same manner of interactive learning as RL
does but from a stochastic decision and programming perspective. Ma & Hill (141, 142) reported a
class of automatic coordinated voltage control schemes based on knowledge search. In this scheme,
a knowledge base consisting of representative optimally controlled cases is prepared and then
applied to online voltage control via knowledge search. Concretely, the case that best matches the
current online operation scenario is picked out from the knowledge base for application. Although
708 Liu et al.

no ML algorithm is explicitly involved, this method adopts the basic idea of nearest-neighbor-
based learning (113).

In the above subsections, instead of exhaustively introducing all the related ML applications, we
have selected some representative studies to categorize and review. For more extensive summaries
of ML/DL applications to power grid stability and control, we refer interested readers to several
recent surveys (144–146). Although ML-based solutions have shown great potential in power grid
stability and control, they still face many significant challenges in practice, and they still have a long
way to go before they will be widely applicable in practical grids. In the following, we briefly discuss
relevant challenging yet crucial issues that deserve more systematic research efforts in the future:
Incorporating model-based and data-based methods: Although the use of data-based solu-
tions assisted by advanced ML techniques to address various stability and control issues in
power grids is highly promising, their performance could be further boosted if the conven-
tional model-based methods were systematically incorporated with them. Doing so would
take advantage of the strengths of both model-based and data-based methods, helping create
more powerful solutions.
Reducing the reality gap between simulation models and practical contexts: Existing ML-
based solutions generally assume that offline numerical simulation models can effectively
mimic practical grids. However, due to unavoidable system modeling and simulation errors,
there are nonnegligible reality gaps between them, meaning that the obtained data-driven
solutions deviate from practical contexts. In this respect, it is necessary to improve the fidelity
of offline learning data to mitigate the drift of offline ML procedures.
Improving the interpretability of data-driven solutions: Unlike model-based methods, which
usually involve explicit physical significance, data-driven solutions—especially DL-based
alternatives—often make decisions in black-box form. In practice, it would be difficult for
system practitioners on the industrial side to accept such nontransparent decisions and be-
lieve that they can reliably work all the time. Hence, there is an imperative need to develop
more interpretable ML schemes without sacrificing learning performance. In fact, this can
not only promote the acceptance of ML solutions in the electric power industry but also pro-
vide interpretable knowledge uncovered from data to help practitioners better understand
the complicated characteristics of practical grids.
Addressing the rarity of unstable or unsafe scenarios: In spite of the high risk of encounter-
ing various faults, practical power grids can maintain stability in most cases and are unstable
or unsafe in only a few situations, which tend to be rare in practice. It can therefore be hard
to rigorously validate the reliability of a data-driven solution in practical contexts, especially
when simulation models are not precise enough to mimic practical conditions. To address
this issue, realistic data augmentation should be carefully carried out. In addition, the rar-
ity of unstable or unsafe scenarios may distort learning cases, but this can be mitigated by
specific tricks, such as over- or subsampling and cost-sensitive learning.
Enhancing the robustness to practical defective environments: Because data-driven solu-
tions are built on system operational data, the data quality acts as a cornerstone for their
reliable implementation in practice. In practical contexts, however, due to the inevitability
of defective measurement and communication conditions, PMU data errors or dropouts
and wide-area communication delays or failures reduce the quality of online PMU data.
Moreover, considering possible cyberattacks in modern power grids, the PMU data quality
issue may be more severe. In this regard, more robust ML schemes need to be designed

against various defective conditions. In recent years, some efforts have been made to tackle
specific types of defective data issues (118), yet more powerful solutions are still needed to
comprehensively address diverse low-quality data conditions.
Adapting to online changing conditions: Due to the uncertainty and variability introduced
by various renewable energy sources and loads, the operating conditions of practical grids
generally change over time, meaning that data-driven models that were initially learned from
limited offline scenarios are constantly encountering unforeseen scenarios. To enhance a
model’s adaptability to online changing conditions, it must be periodically updated via online
continuous learning. To this end, lifelong or incremental learning strategies (112) can be
carefully designed to help it closely follow changing conditions and enable new knowledge
discovery.
Transferring knowledge between related power grids: For existing ML-based solutions in
power grids, data-driven models are often devoted to specific systems. For new systems, or
even expanded existing systems, they are often learned from scratch, which incurs heavy of-
fline computational burdens, especially in practical large-scale systems. Although this does
not influence online efficiency, it could impede the efficient implementation of the online
model updates described above. In fact, if a practical large system shares some similarities
with a small one, a new data-driven model can be efficiently derived for the former by strate-
gically transferring knowledge first learned from the small one. In addition, instead of learn-
ing from scratch, this knowledge transfer may provide a better starting point to accelerate
offline learning in a large system.
Implementing ML-based schemes in a distributed manner: The majority of existing ML-
based solutions in power grids are assumed to be deployed in a centralized way, where
system-wide data such as PMU measurements are uploaded to the control center for ML
implementation. This could create a high risk of data leaks as well as heavy communication
burdens. If these problems become major concerns in practice (e.g., networked MGs being
sensitive to data privacy and communication flexibility), then distributed ML implementa-
tions such as federated learning (147) would be a necessity.
5. CONCLUDING REMARKS
This survey has provided a brief overview of power grid stability and control. In addition to
summarizing classical model-based analysis and control approaches, we have highlighted some
new trends, including stability issues induced by renewable energy sources, graph-based methods,
distributed stability analysis and control, and networked MG control. Given the potential of
data-based methods in tackling some conventionally challenging issues in practical grids, we have
also discussed recent progress in data-driven power grid DSA and control as well as some unsolved
problems. We hope this survey can inspire more innovative ideas, methods, and technologies to
further develop new solutions for secure yet sustainable operation and control of future power
grids.
DISCLOSURE STATEMENT
The authors are not aware of any affiliations, memberships, funding, or financial holdings that
might be perceived as affecting the objectivity of this review.
ACKNOWLEDGMENTS
This work was supported by grants from the Research Grants Council of the Hong Kong Spe-
cial Administrative Region under the Theme-Based Research Scheme (through project number
710 Liu et al.

T23-701/14-N) and the General Research Fund (through project numbers 17202414, 17208817,
17207918, 17256516, and 17209219).
LITERATURE CITED
1. Rocabert J, Luna A, Blaabjerg F, Rodriguez P. 2012. Control of power converters in AC microgrids.
IEEE Trans. Power Electron. 27:4734–49
2. Kundur P, Paserba J, Ajjarapu V, Andersson G, Bose A, et al. 2004. Definition and classification of power
system stability. IEEE Trans. Power Syst. 19:1387–401
3. Vorotnikov VI. 2012. Partial Stability and Control. Basel, Switz.: Birkhäuser
4. Sastry S, Varaiya PP. 1980. Hierarchical stability and alert state steering control of interconnected power
systems. IEEE Trans. Circuits Syst. 27:1102–12
5. Chiang HD. 1989. Study of the existence of energy functions for power systems with losses. IEEE Trans.
Circuits Syst. 36:1423–29
6. Bergen AR, Hill DJ. 1981. A structure preserving model for power system stability analysis. IEEE Trans.
Power App. Syst. PAS-100:25–35
7. Motter AE, Myers SA, Anghel M, Nishikawa T. 2013. Spontaneous synchrony in power-grid networks.
Nat. Phys. 9:191–97

8. Pai MA. 1989. Energy Function Analysis for Power System Stability. Boston: Kluwer Acad.
9. Padiyar K. 2013. Structure Preserving Energy Functions in Power Systems: Theory and Applications. Boca
Raton, FL: CRC
10. Hill DJ, Hiskens IA, Popovic DH. 1994. Stability analysis of power system loads with recovery dynamics.
Int. J. Electr. Power Energy Syst. 16:277–86
11. Van Cutsem T, Vournas C. 1998. Voltage Stability of Electric Power Systems. Boston: Kluwer Acad.
12. Hatziargyriou N, Milanovic JV, Rahmann C, Ajjarapu V, Canizares C, et al. 2020. Definition and clas-
sification of power system stability – revisited & extended. IEEE Trans. Power Syst. 36:3271–81
13. Wang X, Blaabjerg F. 2019. Harmonic stability in power electronic-based power systems: concept, mod-
eling, and analysis. IEEE Trans. Smart Grid 10:2858–70
14. Schiffer J, Zonetti D, Ortega R, Stanković A, Sezi T, Raisch J. 2016. A survey on modeling of
microgrids—from fundamental physics to phasors and voltage sources. Automatica 74:135–50
15. Vorobev P, Huang PH, Al Hosani M, Kirtley JL, Turitsyn K. 2018. High-fidelity model order reduction
for microgrids stability assessment. IEEE Trans. Power Syst. 33:874–87
16. Magnusson PC. 1947. The transient-energy method of calculating stability. AIEE Trans. 66:747–55
17. Chiang HD. 2011. Direct Methods for Stability Analysis of Electric Power Systems: Theoretical Foundation,
BCU Methodologies, and Applications. Hoboken, NJ: Wiley & Sons
18. Xue Y, Van Custem T, Ribbens-Pavella M. 1989. Extended equal area criterion justifications, general-
izations, applications. IEEE Trans. Power Syst. 4:44–52
19. Venkatasubramanian V, Schättler H, Zaborszky J. 1995. Local bifurcations and feasibility regions in
differential-algebraic systems. IEEE Trans. Autom. Control 40:1992–2013
20. Avalos RJ, Cañizares C, Milano F, Conejo AJ. 2009. Equivalency of continuation and optimization meth-
ods to determine saddle-node and limit-induced bifurcations in power systems. IEEE Trans. Circuits Syst.
I 56:210–23
21. Ajjarapu V, Christy C. 1992. The continuation power flow: a tool for steady state voltage stability anal-
ysis. IEEE Trans. Power Syst. 7:416–23
22. Cañizares C, Alvarado FL. 1993. Point of collapse and continuation methods for large AC/DC systems.
IEEE Trans. Power Syst. 8:1–8
23. Nguyen TB, Pai MA. 2003. Dynamic security-constrained rescheduling of power systems using trajec-
tory sensitivities. IEEE Trans. Power Syst. 18:848–54
24. Hou G, Vittal V. 2013. Determination of transient stability constrained interface real power flow limit
using trajectory sensitivity approach. IEEE Trans. Power Syst. 28:2156–63
25. Wang XF, Chen G. 2003. Complex networks: small-world, scale-free and beyond. IEEE Circuits Syst.
Mag. 3(1):6–20
26. Hill DJ, Chen G. 2006. Power systems as dynamic networks. In 2006 IEEE International Symposium on
Circuits and Systems, pp. 722–25. Piscataway, NJ: IEEE

27. Araposthatis A, Sastry S, Varaiya PP. 1981. Analysis of power-flow equation. Int. J. Electr. Power Energy
Syst. 3:115–26
28. Hill DJ, Bergen AR. 1982. Stability analysis of multimachine power networks with linear frequency
dependent loads. IEEE Trans. Circuits Syst. 29:840–48
29. Bapat RB. 2011. Graphs and Matrices. London: Springer
30. Olfati-Saber R, Fax JA, Murray RM. 2007. Consensus and cooperation in networked multi-agent sys-
tems. Proc. IEEE 95:215–33
31. Zelazo D, Bürger M. 2017. On the robustness of uncertain consensus networks. IEEE Trans. Control
Netw. Syst. 4:170–78
32. Dörfler F, Chertkov M, Bullo F. 2013. Synchronization in complex oscillator networks and smart grids.
PNAS 110:2005–10
33. Jafarpour S, Bullo F. 2019. Synchronization of Kuramoto oscillators via cutset projections. IEEE Trans.
Autom. Control 64:2830–44

34. Dörfler F, Bullo F. 2011. Topological equivalence of a structure-preserving power network model and
a non-uniform Kuramoto model of coupled oscillators. In 2011 50th IEEE Conference on Decision and
Control and European Control Conference, pp. 7099–104. Piscataway, NJ: IEEE
35. Song Y, Hill DJ, Liu T. 2019. On extension of effective resistance with application to graph Laplacian
definiteness and power network stability. IEEE Trans. Circuits Syst. I 66:4415–28
36. Song Y, Hill DJ, Liu T. 2020. Network-based analysis of rotor angle stability of power systems. Found.
Trends Electr. Energy Syst. 4:222–345
37. Song Y, Hill DJ, Liu T. 2018. Characterization of cutsets in networks with application to transient
stability analysis of power systems. IEEE Trans. Control Netw. Syst. 5:1261–74
38. Song Y, Hill DJ, Liu T. 2018. Network-based analysis of small-disturbance angle stability of power
systems. IEEE Trans. Control Netw. Syst. 5:901–12
39. Chen W, Wang D, Liu J, Chen Y, Khong SZ, et al. 2021. On spectral properties of signed Laplacians
with connections to eventual positivity. IEEE Trans. Autom. Control 66:2177–90
40. Dörfler F, Bullo F. 2012. Synchronization and transient stability in power networks and nonuniform
Kuramoto oscillators. SIAM J. Control Optim. 50:1616–42
41. Zhu L, Hill DJ. 2018. Stability analysis of power systems: a network synchronization perspective. SIAM
J. Control Optim. 56:1640–64
42. Wu L, Pota HR, Petersen IR. 2019. Synchronization conditions for a multirate Kuramoto network with
an arbitrary topology and nonidentical oscillators. IEEE Trans. Cybern. 49:2242–54
43. Dörfler F, Bullo F. 2010. Synchronization of power networks: network reduction and effective resistance.
IFAC Proc. Vol. 43(19):197–202
44. Simpson-Porco JW. 2018. A theory of solvability for lossless power flow equations—part I: fixed-point
power flow. IEEE Trans. Control Netw. Syst. 5:1361–72
45. Wang C, Bernstein A, Le Boudec JY, Paolone M. 2018. Explicit conditions on existence and uniqueness
of load-flow solutions in distribution networks. IEEE Trans. Smart Grid 9:953–62
46. Bernstein A, Wang C, Dall’Anese E, Le Boudec JY, Zhao C. 2018. Load flow in multiphase distribution
networks: existence, uniqueness, non-singularity and linear models. IEEE Trans. Power Syst. 33:5832–43
47. Nguyen HD, Dvijotham K, Yu S, Turitsyn K. 2019. A framework for robust long-term voltage stability
of distribution systems. IEEE Trans. Smart Grid 10:4827–37
48. Wang Z, Cui B, Wang J. 2017. A necessary condition for power flow insolvability in power distribution
systems with distributed generators. IEEE Trans. Power Syst. 32:1440–50
49. Song Y, Hill DJ, Liu T. 2019. Static voltage stability analysis of distribution systems based on network-
load admittance ratio. IEEE Trans. Power Syst. 34:2270–80
50. Lasseter R, Akhil A, Marnay C, Stephens J, Dagle J, et al. 2002. The CERTS microgrid concept. White
Pap., Transm. Reliab. Program, Off. Power Technol., US Dep. Energy, Washington, DC
51. Farrokhabadi M, Cañizares CA, Simpson-Porco JW, Nasr E, Fan L, et al. 2020. Microgrid stability
definitions, analysis, and examples. IEEE Trans. Power Syst. 35:13–29
52. Simpson-Porco JW, Dörfler F, Bullo F. 2013. Synchronization and power sharing for droop-controlled
inverters in islanded microgrids. Automatica 49:2603–11
712 Liu et al.

53. Colombino M, Groß D, Brouillon JS, Dörfler F. 2019. Global phase and magnitude synchronization of
coupled oscillators with application to the control of grid-forming power inverters. IEEE Trans. Autom.
Control 64:4496–511
54. Song Y, Hill DJ, Liu T. 2019. Impact of DG connection topology on the stability of inverter-based
microgrids. IEEE Trans. Power Syst. 34:3970–72
55. Song Y, Hill DJ, Liu T, Zheng Y. 2017. A distributed framework for stability evaluation and enhancement
of inverter-based microgrids. IEEE Trans. Smart Grid 8:3020–34
56. Yang P, Liu F, Wang Z, Shen C. 2020. Distributed stability conditions for power systems with hetero-
geneous nonlinear bus dynamics. IEEE Trans. Power Syst. 35:2313–24
57. Zhang Y, Xie L. 2015. Online dynamic security assessment of microgrid interconnections in smart dis-
tribution systems. IEEE Trans. Power Syst. 30:3246–54
58. Nandanoori SP, Kundu S, Du W, Tuffner FK, Schneider KP. 2020. Distributed small-signal stability
conditions for inverter-based unbalanced microgrids. IEEE Trans. Power Syst. 35:3981–90
59. Baros S, Bernstein A, Hatziargyriou ND. 2021. Distributed conditions for small-signal stability of power
grids and local control design. IEEE Trans. Power Syst. 36:2058–67
60. Milano F, Dörfler F, Hug G, Hill DJ, Verbič G. 2018. In 2018 Power Systems Computation Conference.
Piscataway, NJ: IEEE. https://doi.org/10.23919/PSCC.2018.8450880
61. Zhu L, Hill DJ. 2020. Synchronization of Kuramoto oscillators: a regional stability framework. IEEE
Trans. Autom. Control 65:5070–82
62. Lu Q, Sun Y, Mei S. 2001. Nonlinear Control Systems and Power System Dynamics. New York: Springer
63. Ruiz-Vega D, Pavella M. 2003. A comprehensive approach to transient stability control. I. Near optimal
preventive control. IEEE Trans. Power Syst. 18:1446–53
64. Ruiz-Vega D, Pavella M. 2003. A comprehensive approach to transient stability control. II. Open loop
emergency control. IEEE Trans. Power Syst. 18:1454–60
65. Bhui P, Senroy N. 2017. Real-time prediction and control of transient stability using transient energy
function. IEEE Trans. Power Syst. 32:923–34
66. Kamali S, Amraee T, Fotuhi-Firuzabad M. 2021. Controlled islanding for enhancing grid resilience
against power system blackout. IEEE Trans. Power Deliv. 36:2386–96
67. Kundur P. 1994. Power Systems Stability and Control. New York: McGraw-Hill
68. Wood AJ, Wollenberg BF, Sheblé GB. 2013. Power Generation, Operation, and Control. Hoboken, NJ:
Wiley. 3rd ed.
69. Hill DJ, Liu T, Verbic G. 2012. Smart grids as distributed learning control. In 2012 IEEE Power and En-
ergy Society General Meeting. Piscataway, NJ: IEEE. https://doi.org/10.1109/PESGM.2012.6344726
70. Ulbig A, Borsche TS, Andersson G. 2014. Impact of low rotational inertia on power system stability and
operation. IFAC Proc. Vol. 47(3):7290–97
71. Paolone M, Gaunt T, Guillaud X, Liserre M, Meliopoulos S, Monti A, et al. 2020. Fundamentals of
power systems modelling in the presence of converter-interfaced generation. Electr. Power Syst. Res.
189:106811
72. Yang L, Liu T, Hill D. 2021. Decentralized event-triggered frequency regulation for multi-area power
systems. Automatica 126:109479
73. Andreasson M, Dimarogonas DV, Sandberg H, Johansson KH. 2014. Distributed PI-control with applica-
tions to power systems frequency control. In 2014 American Control Conference, pp. 3183–88. Piscataway, NJ:
IEEE
74. Lyu X, Jia Y, Liu T, Chai S. 2021. System-oriented power regulation scheme for wind farms: the quest
for uncertainty management. IEEE Trans. Power Syst. 36:4259–69
75. Serban I, Marinescu C. 2014. Control strategy of three-phase battery energy storage systems for fre-
quency support in microgrids and with uninterrupted supply of local loads. IEEE Trans. Power Electron.
29:5010–20
76. Mégel O, Liu T, Hill DJ, Andersson G. 2018. Distributed secondary frequency control algorithm con-
sidering storage efficiency. IEEE Trans. Smart Grid 9:6214–28
77. Callaway DS, Hiskens IA. 2011. Achieving controllability of electric loads. Proc. IEEE 99:184–99
78. Lee CK, Chaudhuri NR, Chaudhuri B, Hui SR. 2013. Droop control of distributed electric springs for
stabilizing future power grid. IEEE Trans. Smart Grid 4:1558–66

79. Halevi Y, Kottick D. 1993. Optimization of load shedding system. IEEE Trans. Energy Convers. 8:207–13
80. Liu T, Hill DJ, Zhang C. 2016. Non-disruptive load-side control for frequency regulation in power
systems. IEEE Trans. Smart Grid 7:2142–53
81. Zhang C, Liu T, Hill DJ. 2018. Granular optimal load-side control of power systems with electric spring
aggregators. arXiv:1806.03679 [math.OC]
82. Zhang X, Hill DJ, Lu C. 2020. Identification of composite demand side model with distributed photo-
voltaic generation and energy storage. IEEE Trans. Sustain. Energy 11:326–36
83. Savulescu SC. 2014. Real-Time Stability in Power Systems: Techniques for Early Detection of the Risk of
Blackout. Cham, Switz.: Springer
84. Fusco G, Russo M. 2006. Adaptive Voltage Control in Power Systems: Modeling, Design and Applications.
London: Springer
85. Corsi S. 2015. Voltage Control and Protection in Electrical Power Systems: From System Components to Wide-
Area Control. London: Springer

86. Van Cutsem T, Vournas C. 2007. Emergency voltage stability controls: an overview. In 2007 IEEE Power
Engineering Society General Meeting. Piscataway, NJ: IEEE. https://doi.org/10.1109/PES.2007.386089
87. Sun H, Guo Q, Qi J, Ajjarapu V, Bravo R, et al. 2019. Review of challenges and research opportunities
for voltage control in smart grids. IEEE Trans. Power Syst. 34:2790–801
88. Baghsorkhi SS, Hiskens IA. 2012. Impact of wind power variability on sub-transmission networks.
In 2012 IEEE Power and Energy Society General Meeting. Piscataway, NJ: IEEE. https://doi.org/
10.1109/PESGM.2012.6345683
89. Sarimuthu CR, Ramachandaramurthy VK, Agileswari K, Mokhlis H. 2016. A review on voltage control
methods using on-load tap changer transformers for networks with renewable energy sources. Renew.
Sust. Energy Rev. 62:1154–61
90. Mahmud N, Zahedi A. 2016. Review of control strategies for voltage regulation of the smart distribution
network with high penetration of renewable distributed generation. Renew. Sust. Energy Rev. 64:582–95
91. Senjyu T, Miyazato Y, Yona A, Urasaki N, Funabashi T. 2008. Optimal distribution voltage control and
coordination with distributed generation. IEEE Trans. Power Deliv. 23:1236–42
92. Tang Z, Hill DJ, Liu T, Ma H. 2018. Hierarchical voltage control of weak subtransmission networks
with high penetration of wind power. IEEE Trans. Power Syst. 33:187–97
93. Tang Z, Hill DJ, Liu T. 2020. Distributed control of active distribution networks to support voltage
control in subtransmission networks. Int. J. Electr. Power Energy Syst. 117:105715
94. Song Y, Zheng Y, Liu T, Lei S, Hill DJ. 2020. A new formulation of distribution network reconfiguration
for reducing the voltage volatility induced by distributed generation. IEEE Trans. Power Syst. 35:496–507
95. Bidram A, Davoudi A. 2012. Hierarchical structure of microgrids control system. IEEE Trans. Smart Grid
3:1963–76
96. Chen F, Chen M, Li Q, Meng K, Zheng Y, et al. 2017. Cost-based droop schemes for economic dispatch
in islanded microgrids. IEEE Trans. Smart Grid 8:63–74
97. Schiffer J, Ortega R, Astolfi A, Raisch J, Sezi T. 2014. Conditions for stability of droop-controlled
inverter-based microgrids. Automatica 50:2457–69
98. Barklund E, Pogaku N, Prodanovic M, Hernandez-Aramburo C, Green TC. 2008. Energy manage-
ment in autonomous microgrid using stability-constrained droop control of inverters. IEEE Trans. Power
Electron. 23:2346–52
99. Hatziargyriou N. 2013. Microgrids: Architectures and Control. Chichester, UK: Wiley & Sons
100. Farrokhabadi M, Cañizares CA, Bhattacharya K. 2015. Frequency control in isolated/islanded micro-
grids through voltage regulation. IEEE Trans. Smart Grid 8:1185–94
101. Vandoorn T, De Kooning J, Meersman B, Vandevelde L. 2013. Review of primary control strategies for
islanded microgrids with power-electronic interfaces. Renew. Sust. Energy Rev. 19:613–28
102. Dörfler F, Simpson-Porco JW, Bullo F. 2014. Plug-and-play control and optimization in microgrids. In
53rd IEEE Conference on Decision and Control, pp. 211–16. Piscataway, NJ: IEEE
103. Hug G, Kar S, Wu C. 2015. Consensus plus innovations approach for distributed multiagent coordina-
tion in a microgrid. IEEE Trans. Smart Grid 6:1893–903
104. Dörfler F, Simpson-Porco JW, Bullo F. 2016. Breaking the hierarchy: distributed control and economic
optimality in microgrids. IEEE Trans. Control Netw. Syst. 3:241–53
714 Liu et al.

105. Ulbig A, Andersson G. 2015. Analyzing operational flexibility of electric power systems. Int. J. Electr.
Power Energy Syst. 72:155–64
106. Wang Z, Chen B, Wang J, Chen C. 2016. Networked microgrids for self-healing power systems. IEEE
Trans. Smart Grid 7:310–19
107. Wang H, Huang J. 2018. Incentivizing energy trading for interconnected microgrids. IEEE Trans. Smart
Grid 9:2647–57
108. Zhang Y, Xie L, Ding Q. 2016. Interactive control of coupled microgrids for guaranteed system-wide
small signal stability. IEEE Trans. Smart Grid 7:1088–96
109. Liu K, Liu T, Hill D. 2017. Frequency control in networked microgrids with voltage-sensitive loads. Pa-
per presented at the 10th Bulk Power Systems Dynamics and Control Symposium, Espinho, Port.,
Aug. 27–Sept. 1
110. GridWise Archit. Counc. 2015. GridWise transactive energy framework version 1.0. Tech. Rep., GridWise
Archit. Counc., Richland, WA

111. Phadke AG, Thorp JS. 2017. Synchronized Phasor Measurements and Their Applications. Cham, Switz.:
Springer. 2nd ed.
112. Goodfellow I, Bengio Y, Courville A. 2016. Deep Learning. Cambridge, MA: MIT Press
113. Bishop CM. 2006. Pattern Recognition and Machine Learning. New York: Springer
114. Sutton RS, Barto AG. 2018. Reinforcement Learning: An Introduction. Cambridge, MA: MIT Press
115. He M, Zhang J, Vittal V. 2013. Robust online dynamic security assessment using adaptive ensemble
decision-tree learning. IEEE Trans. Power Syst. 28:4089–98
116. Wang B, Fang B, Wang Y, Liu H, Liu Y. 2016. Power system transient stability assessment based on big
data and the core vector machine. IEEE Trans. Smart Grid 7:2561–70
117. Zhang Y, Xu Y, Dong Z, Zhang R. 2019. A hierarchical self-adaptive data-analytics method for power
system short-term voltage stability assessment. IEEE Trans. Ind. Inform. 15:74–84
118. Zhang Y, Xu Y, Zhang R, Dong ZY. 2019. A missing-data tolerant method for data-driven short-term
voltage stability assessment of power systems. IEEE Trans. Smart Grid 10:5663–74
119. Malbasa V, Zheng C, Chen PC, Popovic T, Kezunovic M. 2017. Voltage stability prediction using active
machine learning. IEEE Trans. Smart Grid 8:3117–24
120. Li S, Ajjarapu V, Djukanovic M. 2018. Adaptive online monitoring of voltage stability margin via local
regression. IEEE Trans. Power Syst. 33:701–13
121. Wang Q, Li F, Tang Y, Xu Y. 2019. Integrating model-driven and data-driven methods for power system
frequency stability assessment and control. IEEE Trans. Power Syst. 34:4557–68
122. Zhu L, Hill DJ, Lu C. 2020. Hierarchical deep learning machine for power system online transient
stability prediction. IEEE Trans. Power Syst. 35:2399–411
123. Yu JQ, Hill DJ, Lam AYS, Gu J, Li VOK. 2018. Intelligent time-adaptive transient stability assessment
system. IEEE Trans. Power Syst. 33:1049–58
124. Zhu L, Lu C, Sun Y. 2016. Time series shapelet classification based online short-term voltage stability
assessment. IEEE Trans. Power Syst. 31:1430–39
125. Zhu L, Lu C, Dong ZY, Hong C. 2017. Imbalance learning machine-based power system short-term
voltage stability assessment. IEEE Trans. Ind. Inform. 13:2533–43
126. Xie S, Lu C, Zhu L, Zhang J. 2018. Online long-term voltage stability assessment based on time series
shapelet extraction. In 2018 IEEE Innovative Smart Grid Technologies - Asia, pp. 1153–58. Piscataway, NJ:
IEEE
127. Ye L, Keogh E. 2011. Time series shapelets: a novel technique that allows accurate, interpretable and
fast classification. Data Min. Knowl. Discov. 22:149–82
128. Huang J, Guan L, Su Y, Yao H, Guo M, Zhong Z. 2020. Recurrent graph convolutional network-based
multi-task transient stability assessment framework in power system. IEEE Access 8:93283–96
129. Luo Y, Lu C, Zhu L, Song J. 2021. Data-driven short-term voltage stability assessment based on spatial-
temporal graph convolutional network. Int. J. Electr. Power Energy Syst. 130:106753
130. Zhu L, Hill DJ, Lu C. 2021. Intelligent short-term voltage stability assessment via spatial attention
rectified RNN learning. IEEE Trans. Ind. Inform. 17:7005–16
131. Zhu L, Lu C, Kamwa I, Zeng H. 2020. Spatial-temporal feature learning in smart grids: a case study on
short-term voltage stability assessment. IEEE Trans. Ind. Inform. 16:1470–82

132. Genc I, Diao R, Vittal V, Kolluri S, Mandal S. 2010. Decision tree-based preventive and corrective con-
trol applications for dynamic security enhancement in power systems. IEEE Trans. Power Syst. 25:1611–19
133. Tian F, Zhou X, Yu Z, Shi D, Chen Y, Huang Y. 2019. A preventive transient stability control method
based on support vector machine. Electr. Power Syst. Res. 170:286–93
134. Cai H, Ma H, Hill DJ. 2020. A data-based learning and control method for long-term voltage stability.
135. Zhu L, Luo Y. 2021. Deep feedback learning based predictive control for power system undervoltage
load shedding. IEEE Trans. Power Syst. 36:3349–61
136. Huang Q, Huang R, Hao W, Tan J, Fan R, Huang Z. 2020. Adaptive power system emergency control
using deep reinforcement learning. IEEE Trans. Smart Grid 11:1171–82
137. Wang W, Yu N, Gao Y, Shi J. 2020. Safe off-policy deep reinforcement learning algorithm for Volt-VAR
control in power distribution systems. IEEE Trans. Smart Grid 11:3008–18
138. Yan Z, Xu Y. 2020. A multi-agent deep reinforcement learning method for cooperative load frequency
control of a multi-area power system. IEEE Trans. Power Syst. 35:4599–608
139. Duan J, Xu H, Liu W. 2018. Q-learning-based damping control of wide-area power systems under cyber
uncertainties. IEEE Trans. Smart Grid 9:6408–18
140. Schrittwieser J, Antonoglou I, Hubert T, Simonyan K, Sifre L, et al. 2020. Mastering Atari, Go, chess
and shogi by planning with a learned model. Nature 588:604–9

141. Ma H, Hill DJ. 2014. Adaptive coordinated voltage control—part I: basic scheme. IEEE Trans. Power
Syst. 29:1546–53
142. Ma H, Hill DJ. 2014. Adaptive coordinated voltage control—part II: use of learning for rapid response.
143. Shen Y, Yao W, Wen J, He H, Jiang L. 2019. Resilient wide-area damping control using GrHDP to
tolerate communication failures. IEEE Trans. Smart Grid 10:2547–57
144. Duchesne L, Karangelos E, Wehenkel L. 2020. Recent developments in machine learning for energy
systems reliability management. Proc. IEEE 108:1656–76
145. Glavic M. 2019. (Deep) reinforcement learning for electric power system control and related problems:
a short review and perspectives. Annu. Rev. Control 48:22–35
146. Shi Z, Yao W, Li Z, Zeng L, Zhao Y, et al. 2020. Artificial intelligence techniques for stability analysis
and control in smart grids: methodologies, applications, challenges and future directions. Appl. Energy
278:115733
147. Yang Q, Liu Y, Chen T, Tong Y. 2019. Federated machine learning: concept and applications. ACM
Trans. Intell. Syst. Technol. 10:12
716 Liu et al.

AS05_TOC ARjats.cls January 27, 2022 15:37
Annual Review of
Control, Robotics,
and Autonomous
Contents Systems
Volume 5, 2022
An Historical Perspective on the Control of Robotic Manipulators

Mark W. Spong p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 1
Cognitive Science as a Source of Forward and Inverse Models of

Human Decisions for Robotics and Control
Mark K. Ho and Thomas L. Griffiths p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p33
Internal Models in Control, Bioengineering, and Neuroscience

Michelangelo Bin, Jie Huang, Alberto Isidori, Lorenzo Marconi, Matteo Mischiati,
and Eduardo Sontag p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p55
Behavior Trees in Robot Control Systems
Petter Ögren and Christopher I. Sprague p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p81
Methods for Robot Behavior Adaptation for Cognitive
Neurorehabilitation
Alyssa Kubota and Laurel D. Riek p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 109
Grappling Spacecraft
Carl Glen Henshaw, Samantha Glassner, Bo Naasz, and Brian Roberts p p p p p p p p p p p p p p p p p 137
Design and Control of Drones
Mark W. Mueller, Seung Jae Lee, and Raffaello D’Andrea p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 161
Contact and Physical Interaction
Neville Hogan p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 179
Multirobot Control Strategies for Collective Transport
Hamed Farivarnejad and Spring Berman p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 205
Observer Design for Nonlinear Systems with Equivariance
Robert Mahony, Pieter van Goor, and Tarek Hamel p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 221
Partially Observable Markov Decision Processes and Robotics
Hanna Kurniawati p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 253
Increasingly Intelligent Micromachines
Tian-Yun Huang, Hongri Gu, and Bradley J. Nelson p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 279
Magnetic Micro- and Nanoagents for Monitoring Enzymatic Activity
In Vivo
Michael G. Christiansen, Matej Vizovišek, and Simone Schuerle p p p p p p p p p p p p p p p p p p p p p p p p p 311
AS05_TOC ARjats.cls January 27, 2022 15:37
From Theoretical Work to Clinical Translation: Progress in

Concentric Tube Robots
Zisos Mitros, S.M. Hadi Sadati, Ross Henry, Lyndon Da Cruz,
and Christos Bergeles p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 335
Medical Robotics: Opportunities in China
Yao Guo, Weidong Chen, Jie Zhao, and Guang-Zhong Yang p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 361
Probabilistic Model Checking and Autonomy
Marta Kwiatkowska, Gethin Norman, and David Parker p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 385
Safe Learning in Robotics: From Learning-Based Control to Safe
Reinforcement Learning
Lukas Brunke, Melissa Greeff, Adam W. Hall, Zhaocong Yuan, Siqi Zhou,
Jacopo Panerati, and Angela P. Schoellig p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 411
Secure Networked Control Systems

Henrik Sandberg, Vijay Gupta, and Karl H. Johansson p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 445
Energy-Aware Controllability of Complex Networks
Giacomo Baggio, Fabio Pasqualetti, and Sandro Zampieri p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 465
Control of Microparticle Assembly
Xun Tang and Martha A. Grover p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 491
Stimuli-Responsive Polymers for Soft Robotics
Yusen Zhao, Mutian Hua, Yichen Yan, Shuwang Wu, Yousif Alsaid,
and Ximin He p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 515
Control of the Stefan System and Applications: A Tutorial
Shumon Koga and Miroslav Krstic p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 547
Turbulence and Control of Wind Farms
Carl R. Shapiro, Genevieve M. Starke, and Dennice F. Gayme p p p p p p p p p p p p p p p p p p p p p p p p p p p 579
Autonomous Airborne Wind Energy Systems: Accomplishments
and Challenges
Lorenzo Fagiano, Manfred Quack, Florian Bauer, Lode Carnel, and Espen Oland p p p p p p 603
Analysis and Control of Autonomous Mobility-on-Demand Systems
Gioele Zardini, Nicolas Lanzetti, Marco Pavone, and Emilio Frazzoli p p p p p p p p p p p p p p p p p p p 633
Control as an Enabler for Electrified Mobility
Andrew G. Alleyne and Christopher T. Aksland p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 659
Stability and Control of Power Grids
Tao Liu, Yue Song, Lipeng Zhu, and David J. Hill p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 689
Errata
An online log of corrections to Annual Review of Control, Robotics, and Autonomous
Systems articles may be found at http://www.annualreviews.org/errata/control

Stability and Control-Annurev 2022 PDF

Uploaded by

Copyright:

Available Formats

Stability and Control-Annurev 2022 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Stability and Control-Annurev 2022 PDF

Uploaded by

Copyright:

Available Formats

Annual Review of Control, Robotics, and

Tao Liu,1,∗ Yue Song,1,∗ Lipeng Zhu,1,2,∗

China; email: [email protected], [email protected], [email protected]

Annu. Rev. Control Robot. Auton. Syst. 2022. Keywords

2. STABILITY THEORY OF POWER GRIDS

690 Liu et al.

Mi θ̈i + Di θ˙i = Pi − ViV j Gri j cos(θi − θ j ) − ViV j Bri j sin(θi − θ j ), i ∈ VG , 2.

www.annualreviews.org • Stability and Control of the Power Grid 691

692 Liu et al.

behavior is indicated mainly by the voltage regularity, i.e., ∂V i

www.annualreviews.org • Stability and Control of the Power Grid 693

synchronized with the

in stability theory in Section 2.3.

694 Liu et al.

www.annualreviews.org • Stability and Control of the Power Grid 695

696 Liu et al.

2.3.4. Stability in microgrids. An MG is a low- or medium-voltage power grid consisting of

2.4. Further Discussion

www.annualreviews.org • Stability and Control of the Power Grid 697

2.4.2. Beyond equilibrium-based analysis. Conventional stability analysis is oriented to

3. POWER GRID CONTROL

698 Liu et al.

www.annualreviews.org • Stability and Control of the Power Grid 699

3.2. Voltage Control

700 Liu et al.

reactive power control (87).

www.annualreviews.org • Stability and Control of the Power Grid 701

3.3. Control of Microgrids and Networked Microgrids

702 Liu et al.

3.4. Further Discussion

4. DATA-BASED STABILITY ANALYSIS AND CONTROL

www.annualreviews.org • Stability and Control of the Power Grid 703

quickly and reliably

4.2. Data-Driven Power Grid Dynamic Stability/Security Assessment

704 Liu et al.

Data Model Model

b Interactive reinforcement learning–based solutions in power grids

Offline data Interactive Agent Real-time

Scenario Action Action

www.annualreviews.org • Stability and Control of the Power Grid 705

4.2.2. Temporal dependency learning for dynamic stability/security assessment. In con-

4.2.3. Spatial–temporal correlation learning for dynamic stability/security assessment. In

706 Liu et al.

4.3. Data-Driven Power Grid Control

4.3.1. Feedforward machine learning–based solutions. Feedforward ML-based solutions re-

Table 3 Representative references for data-driven power grid control

www.annualreviews.org • Stability and Control of the Power Grid 707

4.3.2. Interactive reinforcement learning–based solutions. As illustrated in Figure 1b, in-

708 Liu et al.

4.4. Further Discussion

www.annualreviews.org • Stability and Control of the Power Grid 709

710 Liu et al.

Nat. Phys. 9:191–97

www.annualreviews.org • Stability and Control of the Power Grid 711

Autom. Control 64:2830–44

712 Liu et al.

www.annualreviews.org • Stability and Control of the Power Grid 713

Area Control. London: Springer

714 Liu et al.