JOURNAL OF LATEX CLASS FILES 1

Complex-Frequency Synchronization of
Converter-Based Power Systems
Xiuqiang He, Member, IEEE, Verena Häberle, Student Member, IEEE,
and Florian Dörfler, Senior Member, IEEE

Abstract—In this paper, we study the phase-amplitude coupled velopment of renewable energy integration. The loss of syn-
dynamics in converter-based power systems from the complex- chronism under grid disturbances has occurred in renewable
frequency perspective. A complex-frequency quantity represents power plants, followed by a large-scale generation interruption
the rate of change of the voltage amplitude and the phase angle by
its real and imaginary parts, respectively. This emerging notion is [6]. Such synchronization stability issues become increasingly
arXiv:2208.13860v2 [eess.SY] 5 Sep 2022

of significance as it accommodates the multivariable characteris- challenging due to heterogeneous network characteristics and
tics of general power systems where active and reactive power are various converter control strategies. On the network side, P /θ
inherently coupled with both the voltage amplitude and phase. and Q/V dynamics become tightly coupled, especially in dis-
We propose the notion of complex-frequency synchronization to tribution networks (with low X/R ratios) [7], which are being
study the phase-amplitude coupled stability of a power system
with grid-forming virtual oscillator-controlled converters. To do increasingly penetrated by distributed energy resources. On
so, we formulate the system into a linear fast system and another the converter-control side, numerous types of control strategies
linear slow system. This linearity property makes it tractable have been developed, e.g., phase-locked loop based control [8],
to analyze fast complex-frequency synchronization and slower droop control [9], virtual synchronous machine control [10],
voltage stabilization. From the perspective of complex-frequency [11], matching control [12], or multivariable control [13]. For
synchronization, we provide novel insights into the equivalence of
virtual oscillator control to complex-power-frequency droop con- any particular type of the controls, the synchronization stability
trol, the stability analysis methodologies, and the stability criteria. with respect to a single converter connected to an infinite grid
Our study provides a practical solution to address challenging has been well-studied [8]–[14]. Among these, the main focus
stability issues in converter-dominated power systems. was on the single-converter P /θ dynamics (at most with some
Index Terms—Complex frequency, grid-forming control, syn- attention on the additional impact of Q/V control on the P /θ
chronization stability, virtual oscillator control, voltage stability. dynamics [10], [11]). Nonetheless, there are few studies on the
coupled stability between P /θ and Q/V dynamics, especially
I. I NTRODUCTION for multi-converter interconnected systems in general.
YNCHRONIZATION in ac power systems is a state when
S the rate of evolution of the phase angle in all generating
units is identical [1]. This concept is at the core of rotor-angle
Grid-forming (GFM) control provides a preferred solution
for the control of converter-dominated power systems. One of
the state-of-the-art GFM controls is virtual oscillator control
stability for conventional power systems. Synchronization is [15]. An advanced version, termed dispatchable virtual oscil-
also a prerequisite for frequency and voltage stability as well lator control (dVOC) [16], leads to almost globally asymptotic
as smooth and efficient power transmission. However, syn- stability in terms of synchronization (θ), voltage (V ) and meet-
chronization in power systems is a challenging issue because ing power setpoints in a dVOC-controlled networked system
the systems are inherently multivariable and nonlinear. Gen- [16]–[18]. The results were theoretically rigorous, experimen-
erally, active power or phase angle (P /θ) and reactive power tally validated [19], and often reproduced [20]. Moreover, the
or voltage (Q/V ) are nonlinearly coupled together [2]. dVOC reduces to a droop control under a small voltage ampli-
The synchronization in conventional power systems is man- tude deviation [19]. There is a growing consensus that dVOC
aged by the interaction of synchronous generators through is the most performant GFM control [20]. We emphasize,
high-voltage transmission networks. The dominantly inductive however, that some restrictive assumptions were applied in the
network characteristics lead to an approximate P /θ and Q/V stability analysis regarding dVOC, and thus, the theoretical
separation [2]. In analytical studies for synchronization/rotor- results in [16]–[18] apply only for a prespecified nominal
angle stability, a standard assumption is that voltage remains equilibrium and a network with a uniform r/` ratio. The
constant [3]. The synchronization dynamics can then be repre- stability problem with usually drooped equilibria in general
sented by classical swing equation models [3] or the celebrated networks remains still open.
Kuramoto model [4]. To further address nonlinearity, tran-
sient stability is investigated separately from small-disturbance An emerging notion, complex frequency, was recently for-
stability, where distinct analytical techniques are applied [5]. mulated in [21]. This new notion is foundational and practical.
These techniques work well for conventional power systems. A complex-frequency quantity, which can be considered as
Currently, power systems are witnessing an increasing uti- the derivative of a complex angle [22], represents the rate of
lization of power converters as a result of unprecedented de- change of both the voltage amplitude and the phase angle. It
thus fits well with the multivariable coupling characteristics
This work was supported by the European Union’s Horizon 2020 research between P /θ and Q/V dynamics. With the two-dimensional
and innovation program under Grant 883985.
The authors are with the Automatic Control Laboratory, ETH Zurich, 8092 “frequency” information, it has also provided practical insights
Zurich, Switzerland. Email:{xiuqhe,verenhae,dorfler}@ethz.ch. on power system state estimation and control [23], [24].
JOURNAL OF LATEX CLASS FILES 2

In this paper, we study the phase-amplitude coupled stabil- current io,k ∈ C and a terminal voltage v k ∈ C. We express
ity of a dVOC-controlled power system from the “complex- v k with the amplitude vk and the phase angle θk as
frequency” perspective. We define the notion of complex-
v k = vk (cos θk + j sin θk ) = vk ejθk . (1)
frequency synchronization, reveal the linear part of the dVOC
dynamics to be equivalent to a complex-power-frequency The network equation is given by
droop control, and show that complex-frequency synchroniza-
tion can be achieved with the linear part on the fast time io = Y v, (2)
scale. On the slower time scale, the system converges from T T
where io := [io,1 , · · · , io,N ] and v := [v 1 , · · · , v N ] . The
the synchronous state to a voltage steady state. The slow power-flow equations of the network are given by
system is also linear when viewed from a “complex angle” XN
y kl vk ejθk vl e−jθl ,



perspective. Consequently, the original phase-amplitude cou- sk = (3)
l=1
pled nonlinear stability problem is converted into two linear
subproblems: complex-frequency synchronization and voltage where sk := pk + jqk = v k io,k denotes the power injection at
stability. We apply linear system theory to solve these two node k, and y kl is the kth row and lth column entry of Y .
subproblems separately. By doing so, the challenge of directly The converters are controlled by a dVOC controller, whose
treating nonlinear stability is avoided, and we can handle the terminal voltage behavior is given in complex-voltage coordi-
cases of a drooped equilibrium and a nonuniform network. We nates as [16]
 p? − jq ? v ? − vk
derive quantitative stability conditions in the time domain and

v̇ k = jω0 v k +ηejϕ k ?2 k v k −io,k +ηα k ? v k , (4)
also provide admittance models and criteria in the frequency vk vk
domain. Our study provides a solution with high practicality where ω0 denotes the fundamental frequency, p?k , qk? , and vk?
to address the phase-amplitude coupled stability issue. denote the active power, reactive power, and voltage setpoints,
The rest of this section recalls some basic notation. The respectively, ejϕ with ϕ ∈ [0, π/2] denotes the rotation oper-
system to be studied is described in Section II. In Section III, ator to adapt to the network impedance characteristics, and
we reformulate the system from the complex-frequency/-angle η, α > 0 are control gains. In (4), we use complex-valued
perspective. In Sections IV and V, we present stability analysis ordinary differential equations (ODEs) as short-hands for two-
in the time and frequency domains, respectively. Case studies dimensional (real-valued) ODEs in αβ coordinates. Hence,
are shown in Section VI. Section VII concludes the paper. derivatives should also be understood as real-valued.
Notation: The set of real and complex numbers is denoted In the dVOC control law in (4), the first term induces
by R and C, respectively. A real or complex scalar is denoted harmonic oscillations of frequency ω0 , the second term syn-
by x or x, respectively. The complex conjugate of x is denoted chronizes the relative phase angles to the power setpoints
by x. A vector with real or complex entries is denoted by x via current feedback, and the third term regulates the voltage
or x, respectively. A matrix with real or complex entries is amplitude. We notice that the first two terms are linear whereas
denoted by A or A, respectively. Given a matrix A ∈ Cm×n , the third term is nonlinear.
AT and AH denote its transpose and Hermitian transpose, re- The overall system model is obtained by interconnecting the
spectively. For a real scalar x, a complex scalar x, a real vector network equation (2) and the node dynamics (4), i.e.,
x, and a complex vector x, we use |x|, |x|, kxk, kxk to denote h 
N
 i
the absolute value, the modulus, the Euclidean norm (xT x)1/2 , v̇ = jω0 v + ηejϕ diag (p?k − jqk? )/vk?2 k=1 − Y v
and the Euclidean norm (xH x)1/2 , respectively. The distance
 
N
+ ηαdiag (vk? − vk )/vk? k=1 v,
of a point x to a set C is denoted by kxkC := minz∈C kz − xk.
For a vector x ∈ Cn , diag (x) denotes the diagonal matrix where all shunt loads in the network admittance matrix Y can
formed from x. For a complex scalar, vector, or matrix, R(·) be absorbed into the corresponding setpoints (p?k − jqk? )/vk?2
and I (·) denote the real and imaginary parts, respectively. such that we consider Y ∈ CN ×N a complex-valued and
symmetric Laplacian matrix with zero row sums [27].
In this work, we are concerned with the stability (with
II. S YSTEM D ESCRIPTION AND DVOC
respect to both θ and V ) of this dVOC-based power system.
The system has been shown to be almost globally asymptoti-
We consider a connected and three-phase balanced power
cally stable under certain parametric conditions [16]–[18]. The
system. Among these, there are N converter nodes and M load
results were achieved by a set of nonlinear stability analysis
nodes. The converter nodes are modeled as voltage sources
approaches including Lyapunov analysis and singular pertur-
since we apply GFM controls. The load nodes are represented
bation analysis. The analysis was theoretically rigorous. How-
by linear RLC or constant-impedance branches, which is a
ever, two restrictive assumptions were required to make the
common assumption typically used for analysis [25]. All nodes
problem tractable: first, the resistance-inductance (r/`) ratios
in the network are interconnected by transmission lines. When
are assumed to be uniform [17, Assumption 1], and second,
ignoring the network dynamics, we obtain a phasor approxi-
the setpoints are considered to be consistent [17, Condition 1].
mation of the network. A reduced network is further obtained
As opposed to this, in this work, we take a complex-frequency
using Kron reduction, in which the load nodes are eliminated
perspective and provide a novel and more general analysis,
[26]. The reduced network is represented by an admittance
where the two assumptions are not required. To do so, we
matrix Y ∈ CN ×N . For each converter, we define an output
reformulate the system using the notion of complex frequency.
JOURNAL OF LATEX CLASS FILES 3

Im vk vk v v
III. C OMPLEX F REQUENCY AND S YSTEM
v
R EFORMULATION v k
v vk vk2
j v vk vk2
A. Complex Frequency vk
k

Consider the complex voltage expression in (1) with non- voltage trajectory vk (t )
Re
zero amplitude. We define a complex angle ϑk to lead to
v k = eϑk [22], which can be seen as a transformation between Fig. 1. The instantaneous complex frequency $k = εk + jωk is comprised
complex-voltage coordinates and complex-angle coordinates. by a radial component εk and an rotating component jωk .

Definition 1. (Complex angle [22]): Given the voltage ampli- to normalizing the power as ς k := ρk + jσk := sk /vk2 =
tude vk (in per unit) and the phase angle θk (in radian), we (pk + jqk )/vk2 , where ρk := pk /vk2 and σk := qk /vk2 are re-
define the complex angle ϑk := uk + jθk , where uk := ln vk is ferred to as normalized active and reactive power, respectively.
called voltage logarithm (dimensionless, or with units similar We then derive from (3) the normalized power-flow equations,
to radian). their conjugates, and the vector form as
XN
Definition 2. (Complex frequency [21], [22]): We define the ςk = y eϑl −ϑk , (5)
l=1 kl
complex frequency $k := ϑ̇k = v̇k /vk +j θ̇k = εk +jωk by the XN XN v
time-derivative of ϑk , where εk = v̇k /vk is the normalized rate ςk = y eϑl −ϑk = y l
, (6)
l=1 kl l=1 kl v k
of change of voltage (rocov), and ωk is the angular frequency.
ς = [ς 1 , · · · , ς N ]T = diag {1/v k }N


k=1 Y v. (7)
The notion of complex frequency has been recently devel-
oped in [21] and [22]. A geometrical interpretation of it is Proposition 1. Both the voltage phasor ratio v l /v k and the
provided as follows [28]. Consider $k = ϑ̇k = v̇ k /v k = normalized power ς k remain invariant in a synchronous state.
v̇ k v k /vk2 = (v̇ k + v̇ ⊥ )v k /vk2 , where v̇ k is decomposed into a d vl vl

v̇ l v̇ k

v

radial component v̇ k and a rotating component v̇ ⊥ , which are
Proof: From dt v = v k v − v k
= v l ϑ̇l − ϑ̇k , it
l k k
d lv
parallel and perpendicular to v k , respectively (see Fig. 1). We follows that dt v k = 0 in the synchronous state where ϑ̇l = ϑ̇k .
notice that v̇ k v k denotes a real number because v̇ k and v k It then follows from (6) that ς k also remains invariant in a
have opposite phases, and then v̇ ⊥ v k denotes an imaginary synchronous state.
number. Hence, the radial and rotating components correspond Since v l /v k = vl /vk ej(θl −θk ) , v l /v k remaining invariant
to the real and imaginary parts of the complex frequency, implies that the voltage amplitude ratio and the phase angle
respectively. In other words, the notion of complex frequency difference remain invariant. In terms of voltage amplitudes,
allows representing two-dimensional frequency information, they are allowed to change with time in a synchronous state.
i.e., the rate of change of voltage in the radial direction and This property is of significance because the constant-voltage
the angular speed in the rotational direction. We denote by assumption is dropped. It is of course not of relevance in
$0 := jω0 the nominal complex frequency. We also note that a steady state but during transients. We note that nonzero
the ln (v) coordinate in Definition 1 and its differential dv/v rocov has also been used as a global load-generation imbalance
in Definition 2 have been used before in conventional power signal in dc microgrids [32], and related logarithm voltage and
system static and dynamic analysis [29], [30]. normalized power coordinates have also been used in [33].

B. Complex-Frequency Synchronization D. DC Complex Power-Flow Equations

Definition 3. (Complex-frequency synchronization): Consider We define complex-angle differences as ϑlk := ϑl − ϑk .
N nodes in a connected network. The voltage trajectories Assume small ϑlk , which implies vk ≈ vl and θk ≈ θl . We can
achieve complex-frequency synchronization if all complex fre- then approximate eϑlk as eϑlk ≈ 1 + ϑlk by the leading terms
quencies ϑ̇k = v̇ k /v k converge to a common constant complex of the Maclaurin Series. Substituting
PN this approximation into
frequency $sync as t → ∞, i.e., ϑ̇k → $sync , t → ∞, ∀k. (6), we obtain that ς k ≈ l=1 y kl (1 + ϑlk ). We thus define
dc complex power flow as follows.
Complex-frequency synchronization implies both angular
frequency synchronization θ̇k → I ($sync ) and rocov syn- Definition 4. (DC complex power flow) Assume small complex
chronization v̇k /vk → R($sync ). Typically, one is interesting angle differences ϑlk ≈ 0, the dc complex power flow ς dc
k is
in v̇k = 0 [31]. We use (nonzero) rocov together with fre- XN
quency as a global variable to indicate power imbalance. ς dc
k
:= y kl (1 + ϑlk ), (8)
l=1
or formulated in the vector form as
C. Normalized Power-Flow Equations
ς dc = diag {1 − ϑk }Nk=1 Y 1N + Y ϑ = Y ϑ,


In a synchronous state under the classical notion, the power (9)
flow in (3) is invariant, i.e., balanced, where the phase angle
where ς dc := [ς dc dc T
1 , · · · , ςN ] , ϑ
:= [ϑ1 , · · · , ϑN ]T , 1N de-
differences remain invariant while the voltage amplitudes are
notes the column vector of all ones, and Y 1N = 0N .
assumed to be constant. For a complex-frequency synchronous
state, the power flow in (3) is not invariant due to the voltage Remark 1. The presented dc complex power-flow equations
amplitude variations. To find an invariant “power flow” to augment the classical dc power-flow equations from the com-
represent the load/generation balance in this case, we resort plex angle perspective. The classical ones are stated for active
JOURNAL OF LATEX CLASS FILES 4

power only [34]. For a branch with admittance y between Thus, a q-v droop control of the form v̇k /vk = η(σk? − σk ) +
node k and l, the dc complex power flowing from node k to l is ηα(u?k −uk ) can be derived from (12) for ϕ = π/2. In analogy,
ς dc = y (ϑk − ϑl ) as in (8). The classical dc power flow reads v̇k /vk = η(σk? − σk ) + ηα(vk? − vk )/vk? is obtained from the
pdc = (θk − θl )/x, which holds with two additional assump- classical dVOC [16]. These two q-v droop controls reveal to
tions: the network is dominantly inductive, i.e., y = 1/(jx), be highly similar around the operating region where vk ≈ vk? .
and vk ≈ vl ≈ 1. The augmented one degenerates into the This claim holds due to 1 − vk /vk? ≈ ln vk? − ln vk following
classical one when applying these two assumptions. from the Taylor Series ln x = x − 1 + o(x − 1).
Remark 2. The dc complex power flow is lossless across the
network, as observed by 1TNς
dc
= 1TN Y ϑ = 0 from (9). This
IV. S TABILITY A NALYSIS IN THE T IME D OMAIN
holds for any network impedance characteristics. Note that the The overall system model used for stability analysis is es-
original power flow (without being normalized) is lossy still. tablished by interconnecting the network equation (2) and the
E. Linear Virtual Oscillator Control node dynamics (12) as
We proceed with a linear virtual oscillator control (LVOC)
v̇ = $0 v + ηejϕ (diag (ς ? ) − Y ) v + ηαΨ(v)v, (14)
formulation for converters to achieve complex-frequency syn-
chronization. The LVOC is defined by the first two terms of where ς ? := [ς ?1 , · · · , ς ?N ]T , Ψ(v) := diag {u?k −ln |v k |}N


k=1 .
the dVOC control law in (4) as
v̇ k = $0 v k + ηejϕ ς ?k v k − io,k ,


(10) A. Separation into Fast and Slow Systems
where ς ?k = (p?k − jqk? )/vk?2 denotes the normalized power The system (14) is composed of a linear part (the first two
setpoint. The dVOC formulation was interpreted from a con- terms) and a nonlinear part (the third term). The linear part
sensus synchronization perspective in [16]. We provide another enables synchronization, while the nonlinear part establishes
insightful interpretation from the viewpoint of droop control. a voltage steady state. In the original dVOC articles [16]–
[18], these two terms were separated by choosing α small,
Proposition 2. The LVOC in (10) is equivalent to a complex- i.e., down-tuning the voltage regulation. Here, we consider
power-frequency droop control (complex droop control for the time-scale separation between the two parts such that they
short) can be treated separately. To do so, we incorporate a low-pass
ϑ̇k = $0 + ηejϕ ς ?k − ς k .


(11) filter into the amplitude feedback to slow down the voltage
Proof: The proof is straightforward by expanding (11) regulation. The filter is also necessary for harmonic filtering
as ϑ̇k = v̇ k /v k and ς k = (pk − jqk )/vk2 = (v k io,k )/vk2 = in actual practice. The system in (14) is then modified as
io,k /v k .
v̇ = $0 v + ηejϕ (diag (ς ? ) − Y ) v + ηαdiag u? − uf v


Proposition 2 suggests that LVOC (10) can possibly achieve  T
complex-frequency synchronization similar to a droop control, τ u̇f = ln |v 1 |, · · · , ln |v N | − uf ,
where the complex frequency ϑ̇k indicates the balance of (15)
the normalized power ς k . This equivalence also suggests the where u? := [u?1 , · · · , u?N ]T , uf is the filtered version of u :=
T
ln |v 1 |, · · · , ln |v N | = R(ϑ), and τ > 0 is the filter time

compatibility of LVOC with respect to most existing GFM
controllers, including p-f droop control, highly damped virtual constant. Using ϑ̇k = v̇ k /v k , the system (15) can then be
synchronous machine (VSM) [4], highly damped matching represented in complex-angle coordinates equivalently as
control, etc. This compatibility is not surprising since the
complex droop control in (11) augments the classical p-f droop ϑ̇ = $0 1N + ηejϕ (ς ? − ς) + ηα (u? − uf )
(16)
control θ̇k = ω0 +η (p?k − pk ), which is the core building block τ u̇f = u − uf = R(ϑ) − uf .
of all GFM controls.
Linear Fast System: On the time scale of fast synchro-
F. Voltage Regulation for the LVOC nization dynamics, uf in (15) is considered to be constant.
When applying ϕ = π/2 in (11), the real part of it results in Consequently, the synchronization dynamics are represented
a constant-reactive-power control with respect to the amplitude by the fast system as
part, i.e., v̇k /vk = η (σk? − σk ). To enable voltage regulation,
we augment the LVOC with a voltage amplitude regulation v̇ = $0 v + ηejϕ (diag (ς ∗ ) − Y ) v, (17)
term. One option is to add a proportional control term as in the where ς ∗ := ς ? + αe−jϕ (u? − uf ) is the equivalent power
dVOC in (4). Alternatively, one can resort to controlling the reference for this fast system. Note that the fast system (17) is
voltage logarithm ln vk using another dVOC controller similar linear when seen from the complex-voltage coordinate frame.
to the classical dVOC in (4) as Linearly Approximate Slow System: On the slower time
v̇ k = $0 v k + ηejϕ ς ?k v k − io,k + ηα (u?k − uk ) v k , (12)


scale of voltage amplitude regulation, we assume that the faster
complex-frequency synchronization is already achieved. For a
where uk = ln vk and u?k := ln vk? . The dVOC in (12) is equiv-
tractable analysis of voltage stability, we thus propose to apply
alent to a voltage regulation-enabled complex droop control
the dc complex power-flow equations derived in Section III-D.
as in (11), which reads in complex-angle coordinates as
This is reasonable because the system has already reached
ϑ̇k = $0 + ηejϕ ς ?k − ς k + ηα (u?k − uk ) .


(13) a synchronous state and the synchronized angle differences
JOURNAL OF LATEX CLASS FILES 5

ϑlk are quite small. Consequently, the dc complex power- synchronization. Hence, previous works [16]–[18] adopted the
flow equations hold well. When approximating ς in (16) with terminology almost global synchronization.
ς dc = Y ϑ, the slow system is obtained as Condition 1 relies on the calculation of eigenvalues of A.
While Condition 1 is mathematically intuitive, it does not
ϑ̇ = $0 1N + ηejϕ (ς ? − Y ϑ) + ηα (u? − uf )
(18) give any physical intuition. In the following, we provide a
τ u̇f = R(ϑ) − uf . parametric stability condition via a Lyapunov analysis.
The slow system is formulated in complex-angle coordinates Condition 2. (Parametric condition): There exists a maximal
to accommodate the dc complex power-flow equations. When angle difference δ̄ ∈ [0, π/2) and a maximal voltage-amplitude
seen from this coordinate frame, the slow system (18) is linear. ratio deviation γ̄ ∈ (0, 1) such that
Remark 3. When viewed from a different coordinate frame,
limt→∞ |θk − θl | ≤ δ̄, limt→∞ vk /vl − 1 ≤ γ̄ (19)
both systems in (17) and (18) become nonlinear. We highlight
that the adoption of coordinates matters. This is an interesting hold for any node pair (k, l) in any synchronous state. More-
and important observation such that we can apply linear sys- over, the power references ς ∗k , the rotation operator ejϕ , and
tem analysis to yield actionable results on the phase-amplitude the network admittance matrix Y satisfy
coupled stability issue. The difficulty of directly analyzing the
1 + cos δ̄ 2
nonlinear coupled dynamics of the system (14) is thus avoided. max R(ejϕ ς ∗k ) < (1 − γ̄) λ2 R(ejϕ Y ) ,


(20)
k 2
Remark 4. The system under disturbances is supposed to first
reach a synchronous state and then a voltage steady state. This where R(ejϕ Y ) is formed by the real part of the entries
does fit actual operational requirements, where the deviation of ejϕ Y , and it is therefore a Laplacian matrix, and λ2 (·)
range for voltage is much larger than that for frequency (e.g., denotes the second largest eigenvalue (a positive real number).
~±10% for voltage vs. ~±1.0% for frequency) [35]. In view Theorem 2. Under Condition 2, the system (17) achieves
of synchronization dynamics being faster than voltage conver- complex-frequency synchronization at λ1 for almost all initial
gence, the two subproblems, complex-frequency synchroniza- state v 0 except for the non-generic initial condition ψ T v =0
1 0
tion and voltage stability, can be analyzed separately. where the voltages v converge to 0N .

B. Complex-Frequency Synchronization Stability Analysis The proof is provided in Appendix. Condition 2 first places
a constraint on the angle differences and voltage ratios in the
We first analyze complex-frequency synchronization sta-
synchronous state. This constraint is intuitively reasonable for
bility by focusing on the fast system. The fast system (17)
power system operation since a healthy synchronous operating
is a linear autonomous system with system matrix A :=
point is generally characterized by rather small angle differ-
$0 I N + ηejϕ (diag (ς ∗ ) − Y ). The complex eigenvalues of
ences and voltage ratios. Condition 2 then bounds the power
A are denoted as λ1 , λ2 , · · · , λN , where R(λ1 ) ≥ R(λ2 ) ≥
setpoints in terms of the algebraic connectivity λ2 (·) of the
· · · ≥ R(λN ). We term λ1 dominant eigenvalue and assume
graph corresponding to the real part of the rotated network
that λ1 has algebraic multiplicity one and R(λ1 ) > R(λ2 ).
admittance matrix. This bound quantifies a few well-known
This assumption is generic and reflects the actual require-
engineering insights: the network should be sufficiently well
ment that power systems normally operate with only one
connected (as reflected by λ2 (·) > 0) and should not be
fundamental-frequency component. For the dominant eigen-
heavily loaded (cf. the left-hand side of (20)), and the rotation
value λ1 , we denote by ψ T and φ1 the left and right eigenvec-
1 angle ϕ should match the impedance angle as closely as pos-
tors, respectively (here, ψ 1 = φ1 since A is symmetric). We
sible [7]. In terms of application, Condition 2 is actionable.
provide a spectral stability condition for complex-frequency
The synchronization stability assessment can be performed
synchronization of the system (17).
in a decentralized fashion with the right-hand side network
Condition 1. (Spectral condition): The entries of φ1 are connectivity information in (20).
nonzero, and R(λ2 ) < 0. Previous dVOC stability studies [16]–[18] applied two re-
strictive assumptions, i.e., the network is assumed to have a
Theorem 1. Under Condition 1, the system (17) achieves
homogeneous r/` ratio, and the setpoints are considered to
complex-frequency synchronization at λ1 for almost all initial
be consistent with the power-flow equations. As opposed to
state v 0 except for the non-generic initial condition ψ T v =0
1 0 this, our analysis from the perspective of complex-frequency
where the voltages v converge to 0N .
synchronization drops both assumptions in light of Condition
The proof is provided in Appendix. Under Condition 1, the 2. Namely, we address a network with arbitrary impedance
response of the linear system (17) is dominated by the domi- parameters and consider arbitrary power and voltage setpoints.
nant modal response φ1 ψ T v eλ1 t . The other modes decay to
1 0
These two aspects are of significance. First, an actual power
zero, and consequently the system synchronizes at λ1 . In Con- network is generally composed of multiple-level sub-networks
dition 1, if one entry of φ1 is zero (this is not generic), then that are with different impedance characteristics. Second, an
the dominant modal response of the corresponding node will actual system almost always operates at a droop point as a
be zero. If the initial state v 0 satisfies ψ T v = 0 (this is also
1 0
result of load and generation variations as well as the power
not generic), then the overall dominant modal response will and voltage setpoints being often specified individually and
be zero. Both cases are not expected by complex-frequency thus inconsistently.
JOURNAL OF LATEX CLASS FILES 6

C. Voltage Stability Analysis are linear [18]. The converter equivalent admittance can then
We next show that for the slow system (18) voltage stability be derived from y equ,k (s) := −io,k (s)/v k (s). For example,
is guaranteed. We rewrite (18) in real-valued variables as y equ,k (s) = (s − jω0 )e−jϕ /η − ς ∗k for the case considering
only the LVOC dynamics for converters. Synthesizing all the
u̇ = ηρ0? − ηG0 u + ηB 0 θ + ηα (u? − uf )
converter equivalent admittances, the transmission line admit-
θ̇ = ω0 1N − ησ 0? − ηB 0 u − ηG0 θ (21) tances, and the load admittances, an admittance matrix Y 0 (s)
τ u̇f = u − uf , can be formed to represent the entire fast system in the same
way as forming a static admittance matrix. Seen from any
where u+jθ = ϑ, G0 +jB 0 = ejϕ Y , and ρ0? −jσ 0? = ejϕ ς ? . node k, the network-side aggregated admittance, y agg,k (s), can
In a steady state, u̇ = 0N but θ̇ 6= 0N . To denote the steady- be obtained from the Schur complement of the corresponding
state equilibrium, we define the center-of-angle coordinate as block of Y 0 (s) [26]. Consider k = 1, without loss of gener-
θ0 := 1T N θ/N . Then θ̇0 = ω0 − η 1N σ /N . We denote the
T 0?
ality (for the case of k 6= 1, swap the first and kth rows and
angle deviations from θ0 as δ := θ − 1N θ0 . The system (21) then the first and kth columns in Y 0 (s)). To apply the Schur
is transformed to the center-of-angle coordinate frame as complement, we partition Y 0 (s) into four blocks as
u̇ = ηρ0? − ηG0 u + ηB 0 δ + ηα (u? − uf ) 
a(s)1×1 bT (s)1×(M +N −1)

Y 0 (s) = .
δ̇ = η 1N 1T 0? 0?
− ηB 0 u − ηG0 δ


N σ /N − σ (22) c(s)(M +N −1)×1 D(s)(M +N −1)×(M +N −1)
τ u̇f = u − uf . The network-side aggregated admittance is thus obtained by
We show in Lemma 1 in Appendix that the system in (22) has y agg,k (s) := a(s) − bT (s)D −1 (s)c(s) − y equ,k (s), where the
T local converter equivalent admittance is excluded. Note that we
a unique equilibrium, which is denoted by [uT T T
s , δ s , us ] .
are able to represent the admittance model of the fast system
Theorem 3. The system (22) is globally asymptotically stable by a complex-coefficient transfer function because we consider
T
with respect to the equilibrium [uT T T
s , δ s , us ] . The voltages a three-phase balanced system and an admittance matrix in the
u and the frequencies θ̇ of the slow system (21) globally form a(s)
 −b(s) 
in αβ coordinates is equivalent to a(s) +
converge to us and ω0 + η 1T N I (e ς )/N , respectively.
jϕ ? b(s) a(s)
jb(s) in complex-voltage coordinates [37].
Theorem 3 guarantees voltage stability of the slow system,
the proof of which is provided in Appendix. We interpret B. Admittance Model of the Slow System
the steady-state frequency deviation η 1T
N I (e ς )/N as the
jϕ ?
The admittance model of the slow system (18) is shown in
droop gain multiplied by the average of all the individual Fig. 2(b), where we consider the Kron-reduced static network,
rotated active power setpoints. We relate this result to the fact since the slow system (18) was defined with static dc power
that the dc complex power flow is lossless, cf. Remark 2. flow. Hence, for each converter node k in (18), we consider
V. S TABILITY A NALYSIS IN THE F REQUENCY D OMAIN ϑ̇k = $0 + ηejϕ ς ?k − ηejϕ ς dc ?
k + ηα (uk − uf,k )
(23)
State-space analysis in the time domain typically requires u̇f,k = 1/τ (R(ϑk ) − uf,k ) .
the global information of the system. Impedance-/admittance- We define the rotated power setpoint as ρ0? 0? jϕ ?
k − jσk := e ς k ,
based analysis in the frequency domain is considered as a local 0? 0?
consistent with the vector formulation ρ − jσ = e ς in jϕ ?
approach [36], where the admittance model can be developed (21). Similarly, we define the rotated power flow as ρ0dc k −
either via local measurements and identification or via a state- jσk0dc := ejϕ ς dc
k . The node dynamics in the frequency domain
space model. We first provide methods to derive admittance are then obtained as
models from the state-space models shown before. We then
1 s + τ ηα
    0dc   0?
ρk + αu?k
 
present stability analysis by applying the Nyquist criterion. s+1 0 uk ρk
+ = , (24)
Note that the modeling and analysis are large-signal because η 0 s θk −σk0dc −σk0? + ω0 /η
| {z }
both systems are linear. Y dc
equ,k (s)

A. Admittance Model of the Fast System where Y dc equ,k (s) denotes the converter equivalent admittance

The admittance model of the fast system (17) is shown and [ρ0?k + αu?k , −σk0? + ω0 /η]T denotes the equivalent refer-
in Fig. 2(a), where each node is represented by a dynamic ence. To simplify aggregation, we remove the equivalent refer-
admittance. When considering the network dynamics, the net- ence such that a zero-input system can be considered. This is
work topology should be preserved (without performing Kron justified, as the systems with and without the reference input
reduction). The admittance model for transmission lines and share the same stability properties, cf. (34). Following the
RLC loads can be obtained readily, e.g., y l (s) = 1/(rl + s`l ) rotated power in (24) and the static network admittance Y ,
for a transmission line with resistance rl and inductance `l . we obtain the admittance matrix of the entire slow system as
When deriving the converter equivalent admittance, we can Y dc (s) := R(ejϕ Y ) ⊗ [ 10 01 ] + I (ejϕ Y ) ⊗ 01 −1
 
0
further incorporate the voltage- and current-loop dynamics and (25)
+ diag {Y dc N


the LC filter dynamics (see Fig. 3), as well as the LVOC equ,k (s)}k=1 ,

dynamics. This is feasible provided that the inner-loop con- where ⊗ denotes the Kronecker product. We can derive from
trollers in complex-voltage and complex-current coordinates Y dc (s) the network-side aggregated admittance, denoted by
JOURNAL OF LATEX CLASS FILES 7

dc dc T
io,k (s ) vk (s ) [ k
, k
] [uk , k ]T

k
uk Re(e j Y ) [ 10 10 ]
dc dc
yequ,k (s ) yagg,k (s ) 0
Yequ, k
(s ) Yagg,k
(s )
yl (s ) k Im(e Y )j
[ 10 0
1
]

N 1 dynamic M load N 1 reduced static and

converter network nodes converter rotated network
(a) nodes (b) nodes

Fig. 2. (a) The admittance model of the fast system, where a topology-preserved dynamic network is considered. (b) The admittance of the slow system,
where a Kron-reduced static network is considered because the system is defined with static dc complex power flow.

lf rf io,k vk (0.2, 0.7, 1.0) (0.3 j 0.6)/3 (0.4, 0.5, 1.0)

ek C2 C3
cf gc pk , qk , vk
if ,k vk 1 0.4 j 0.4 0.3 j 0.2 1
LVOC & 0.2 j 0.4 3 3 0.2 j 0.3
curr. loop volt. loop volt. regulator
ek if ,k vk
Fig. 3. The overall control of a grid-forming virtual oscillator-controlled C1 setpoints (pk , qk , vk )
converter in the network. The inner-loop controllers can be found in [18]. 1 base power 2 MVA
(0.6, 0.4, 1.0) base volt. 690 V
0.1 j 0.2
base freq. 50 Hz
dc
Y agg,k (s),by following a similar procedure as for the fast sys- Fig. 4. A three-bus system and its network impedance parameters and
tem. The admittance model is represented by a real-coefficient converter setpoints. Note that the network has non-uniform r/` ratios and
2 × 2 transfer function matrix, accounting for the existence of the converters have inconsistent setpoints.
the voltage regulation term.
Proposition 4. The slow system is stable if and only if N2 = 0.
C. Admittance-Based Criterion for Synchronization Stability The proof of Proposition 4 is immediate by applying the
We consider lk (s) := y agg,k (s)/y equ,k (s) as an admittance generalized Nyquist stability criterion [39, Theorem 4.7].
ratio to derive an admittance-based criterion for synchroniza-
tion stability [38]. Let N1 be the number of counterclockwise VI. C ASE S TUDIES
encirclements of the point (−1+j0) for lk (s) when s traverses We illustrate the theoretical results by case studies on a
the Nyquist contour, where lk (s) has P1 unstable poles. Let Z1 simplified three-bus system in Fig. 4. The topology represents
be the number of unstable poles of the closed-loop fast system. a Kron-reduced IEEE 9-bus system as in [17]. The line param-
It follows from the argument principle that Z1 = P1 − N1 eters (non-uniform) and converter setpoints (inconsistent) are
[39, Lemma 4.8]. Here, we seek Z1 ≤ 1, i.e., all closed-loop indicated in Fig. 4. We model the converters by an electromag-
poles are stable except one accounting for complex-frequency netic transient (EMT) average-value model, where both the
synchronization. inner-loop dynamics and the LC filter dynamics are retained
and the controllers are implemented in αβ coordinates [18].
Proposition 3. A necessary and sufficient condition for the
We choose the controller gains η = 0.04 and α = 5 in per-
fast system to be able to achieve complex-frequency synchro-
unit, which provide a frequency droop slope of η = 4% and
nization is Z1 ≤ 1.
a voltage droop slope of 1/α = 20% [40]. We choose the
The proof of Proposition 3 is straightforward by applying filter time constant τ = 0.005 s (a larger τ will deteriorate
Theorem 1, i.e., R(λ2 ) < 0 ⇒ Z1 ≤ 1. dynamic performance). Moreover, we choose ϕ = π/4 as a
compromise choice to adapt to the heterogeneous r/` ratios
Remark 5. Parameter tuning or power setpoint specifying
in the grid. The system is identified to be stable for complex-
can be performed quickly by Condition 2, where inner-loop
frequency synchronization by checking that Condition 1 holds.
dynamics and transmission line dynamics are ignored. On
We illustrate complex-frequency synchronization and volt-
this basis, Proposition 3 can be further leveraged for stability
age convergence with the simulation results in Fig. 5. The
checks to confirm the synchronization stability under the tuned
system black-starts from a point close to the origin. Before
parameters or specified power setpoints, where the dynamics
0.3 s, the voltage regulation is disabled such that we can ob-
are taken into account.
serve the phenomenon of complex-frequency synchronization
with the nonzero rocov (rate of change of voltage) wave and
D. Admittance-Based Criterion for Voltage Stability the frequency wave. In particular, the voltages increase at an
We consider Lk (s) := Y dc dc
agg,k (s)Y equ,k (s)
−1
as a return- exponential rate in the synchronous state. At 0.3 s, the voltage
ratio matrix to derive an admittance-based criterion for voltage regulation is enabled. The voltages are then boosted close to
stability [38]. Let N2 be the number of counterclockwise

en- the nominal point, during which complex-frequency synchro-
circlements of the origin by det I + Lk (s) when s traverses nization persists. At 0.5 s, we change the power setpoint ς 1
the Nyquist contour. Note that Lk (s) has no unstable poles but from 0.6 − j0.4 to −0.1 − j0.9. Afterwards, we observe a
has a pole at the origin, cf. (24), and thus, the Nyquist contour fast complex-frequency synchronization process, which is fol-
should exclude the origin. Moreover, the unity-feedback sys- lowed by the voltages setting down to a new steady state. The
tem of Lk (s) is assumed to have no hidden unstable modes. active and reactive power sharing follows when entering the
JOURNAL OF LATEX CLASS FILES 8

angle diff. (rad)

ified the admittance model and the criterion under more cases
voltage (pu)
0.04
0.5 by eigenvalue calculation with the state-space model (34).
0 0.03
Re v1 ) |12 |
-0.5 0.02 |13 | VII. C ONCLUSION
Re v2 )
Re v3 )
0.01 |23 |
-1 | v1| We investigated the complex-frequency synchronization and
0.6 voltage stability of dVOC-controlled converter-based power

reactive power (pu) active power (pu)

80
p1 systems. By applying complex frequency as a two-dimensional
0.4
rocov (1/s)

1
60 p2
2
p3
signal to indicate the imbalance of complex power, we define
40 3 0.2
the notion of complex-frequency synchronization to uniformly
20 0 treat the angle and amplitude dynamics driven by the com-
0
-0.2 plex power imbalance. We show that the dVOC can achieve
0.6
complex-frequency synchronization on the fast time scale and
frequency (pu)

1 then stabilize the voltage amplitudes on the slower time scale.

0.4 q1 An insightful understanding of dVOC is provided by revealing
1

0.99 2 q2 its equivalence to a complex-power-frequency droop control.

0.2 q3
3
From this equivalence, we can consider the dVOC as a direct
0
0.98 augmentation of the classical p-f droop control. We also theo-
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
time (s) time (s) retically show that the dVOC has superior stability properties
Fig. 5. Simulation verification of complex-frequency synchronization, voltage in the general case of droop equilibria and nonuniform net-
convergence, small complex-angle differences, and power sharing. works. Our study provides both time-domain stability analysis
4 4
methodologies and frequency-domain stability criteria. The
achieved results are practical to deal with stability issues in
imag. axis

imag. axis

0.12 2
2 converter-based power systems, such as standalone microgrids
0.04 0
(0, j 0) and offshore wind power plants connected to high-voltage dc
0
( 1, j 0) -2 transmission systems. For general power systems with hetero-
-2 -4 geneous devices, complex-frequency synchronization is also
-2 -1 0 1 2 0 2 4 6
real axis real axis an interesting topic. The nonlinear stability analysis for such
(a) (b)
Fig. 6. Nyquist plot (a) for the fast system and (b) for the slow system. The
systems still faces challenges, while it is feasible to perform
former is asymmetric about the real axis because l1 (s) is complex-coefficient. impedance-based analysis by linearization at an equilibrium.

A PPENDIX
steady state. During the voltage regulation, the complex angle
differences vary around zero, which supports the assumption Proof of Theorem 1: Since the dominant eigenvalue λ1
of dc complex power flow underlying the previous analysis. has algebraic multiplicity one, the dominant mode is denoted
Given that λ2 R(ejϕ Y ) ≈ 14.2, we emphasize that Con-

by eλ1 t . The response of v k is composed of a series of modes,
dition 2 is satisfied easily for arbitrary power setpoints within where we separate out the dominant modal response as v 1k . By
the capacity range. This indicates that complex-frequency syn- spectral decomposition, state equation decoupling, and modal
chronization can be readily achieved in the case where the response representation [5, Sec. 12.2.4], we obtain that v 1 :=
network dynamics can be neglected. When considering the [v 11 , · · · , v 1N ]T = φ1 ψ T v eλ1 t , where v 0 is an initial state.
1 0
network dynamics, the synchronization stability is adversely We first consider the case where ψ T v 6= 0. We obtain the
1 0
affected [17]. We test the viability of the admittance-based cri- following limit for any node pair (k, l)
terion in Proposition 3 to assess the destabilizing effect of the vk v 1 + (v k − v 1k )
network dynamics. The admittance model shown in Fig. 2 is lim = lim k1
t→∞ v l t→∞ v + (v l − v 1 )
established with the LVOC dynamics and the network dynam- l l
(26)
ics. We compare the synchronization stability under η = 0.04 φk1 z 0 + (v k − v 1k )e−λ1 t φ
= lim = k1 ,
and 0.12. Without considering the network dynamics, both t→∞ φ z + (v − v 1 )e−λ1 t φl1
l1 0 l l
cases are identified to be stable by Conditions 1 or 2. The
where φl1 6= 0 is the lth entry of φ1 , and the third equality
system with the presence of the network dynamics, however,
holds because R(λ1 ) is larger than the real part of the eigen-
becomes unstable under the larger gain η = 0.12. The Nyquist
values for the other non-dominant modes in (v k − v 1k ), i.e.,
plot of l1 (s) in Fig. 6(a) indicates N1 = −1 in this case, while
R(λ1 ) > R(λ2 ). We then obtain for the system (17) that
l1 (s) has two unstable poles so that Z1 = 3 > 1. However, we
allow at most one unstable closed-loop pole as synchronous v̇  XN v 
lim k = $0 + ηejϕ ς ∗k − lim y kl l
complex frequency. We validate by the Nyquist criterion and t→∞ v k t→∞ l=1 vk
eigenvalue calculation that we need η < 0.082 for stability.  XN φ  (27)
We also test the viability of the admittance-based criterion in = $0 + ηejϕ ς ∗k − y kl l1 = λ1 ,
l=1 φk1
Proposition 4 to assess
the voltage convergence. The Nyquist
plot of det I + L1 (s) is shown in Fig. 6(b), where N2 = 0 where the last equality holds due to Aφ1 = λ1 φ1 , namely,
PN
$0 φk1 + ηejϕ ς ∗k φk1 − m=1 y km φm1 = λ1 φk1 . It follows


confirms the voltage stability, cf. Proposition 4. We have ver-
JOURNAL OF LATEX CLASS FILES 9

from (27) that the complex frequencies converge to the dom- where the first inequality holds because the smallest Rayleigh
inant eigenvalue, i.e., $sync = λ1 . Power system operation quotient of ΠP v orthogonal to the eigenvector 1N of the
requires that only one fundamental-frequency component is smallest eigenvalue is the second smallest eigenvalue. The
present in the synchronous state. This is an additional require- second inequality holds due to the Cauchy-Schwarz inequality.
ment for Definition 3. It is apparent that R(λ2 ) < 0 fulfills When ψ T v 6= 0, it holds that limt→∞ (v k /v l ) = φk1 /φl1
1 0
this requirement since the non-dominant modes decay to zero. as in (26), where φl1 6= 0 holds for all l since (19) requires the
We finally consider the special case where ψ T v = 0. In
1 0
synchronous-state voltage amplitude ratio to be nonzero and
this case, the dominant modal response v 1 = φ1 ψ T v eλ1 t =
1 0
bounded. Then, the synchronous-state constraints in (19) can
0. The voltages converge to 0N due to R(λ2 ) < 0. be converted into ∠φk1 − ∠φl1 ≤ δ̄ and |φk1 |/|φk1 | − 1 ≤
Proof of Theorem 2: We first show that Condition 2 γ̄. Following an analogous argument as in the proof of [17,
guarantees that the voltages with initial condition satisfying Lemma 2], the term v H P R(ejϕ Y )P v can be bounded as
ψT v 6= 0 converge to the eigenspace spanned by φ1 , which is
1 0 v H P R(ejϕ Y )P v
denoted by S. A matrix P := I N − φ1 φH 1
/(φH φ ) is defined
1 1 2
as the projector onto the subspace orthogonal to S. Then, the min φk1 1 + cos δ̄
2
≥ λ2 R(e Y ) kvkS k
jϕ


distance of the point v to the set S is given by kvkS = kP vk. 2 2 (31)
max φk1
Notice that P H = P and P 2 = P . The square of the distance k
2
is kvkS = v H P v. We consider the Lyapunov function 2 2 1 + cos δ̄
≥ λ2 R(ejϕ Y ) kvkS (1 − γ̄)


.
1 2 1 2
V (v) := kvkS = v H P v. (28)
2 2 By substituting (30) and (31) into (29), we observe that the
The time derivation of V (v) along the trajectories of the parametric relationship in (20) is sufficient for
system (17) is given by V̇ (v) = 21 v H (AH P + P A)v. From V̇ (v) < 0, ∀v ∈
/ S. (32)
Aφ1 = λ1 φ1 and P = I N − φ1 φH 1
/(φH φ ), we ob-
1 1
tain that A − λ1 I N = (A − λ1 I N )P , and further that We conclude from (28) and (32) that the system (17) is
A = AP + λ1 (I N − P ). It follows that P A = P AP and asymptotically stable with respect to the eigenspace S for any
AH P = P AH P . Therefore, we obtain initial states v 0 satisfying ψ T v 6= 0. From limt→∞ (v k /v l ) =
1 0
1 φk1 /φl1 , we further obtain that lim (v̇ k /v k ) = λ1 as in (27).
t→∞
V̇ (v) = v H P A + AH P v


2 (29) This shows that the system (17) synchronizes at λ1 .
2
= ηv H P R(ejϕ diag (ς ∗ )) − R(ejϕ Y ) P v.
  Moreover, the asymptotic stability implies limt→∞ kvkS =
0, which suffices to show R(λk ) < 0 for all k > 1. It then
It follows that V̇ (v) ≤ 0 ⇐⇒ v H P R(ejϕ diag (ς ∗ ))P v ≤ follows that the voltages converge to 0N for ψ T v = 0.
1 0
v H P R(ejϕ Y )P v. In the following, we bound the left-hand
and right-hand sides of this inequality separately. Lemma 1. The system in (22) has a unique equilibrium
T T T
Since R(ejϕ diag (ς ∗ )) is a diagonal matrix, we obtain [uT T T T
s , δ s , us ] , where [us , δ s ] is the solution of
 0
G + αI N −B 0   ρ0? + αu?
  
2
v H P R(ejϕ diag (ς ∗ ))P v ≤ kvkS max R(ejϕ ς ∗k ). (30) u
k  B0 G0  s = 1N 1T 0?
N σ /N − σ
0? 
. (33)
δs
To bound the right-hand side term v H P R(ejϕ Y )P v, we 0NT
1NT
0
| {z } | {z }
define another projector onto the orthogonal subspace of the A b
span of 1N as Π := I N − 1N 1T T
N /N satisfying Π = Π and Proof: We assume that there are equilibria in the system
Π = Π. Since R(e Y ) is a Laplacian matrix, its smallest
2 jϕ
(22). Then, it holds that u̇ = 0N , δ̇ = 0N , and u = uf in the
eigenvalue is zero and the corresponding right and left eigen-
equilibria. Thus, (33) follows from (22). The set of equations
vectors are 1N . It follows that R(ejϕ Y ) = R(ejϕ Y )Π =
(33) has a unique solution if and only if the rank condition
ΠR(ejϕ Y ). We bound the right-hand side term as follows
rank (A) = rank ([A, b]) = 2N holds. The top N rows of A
v H P R(ejϕ Y )P v = v H P ΠR(ejϕ Y )ΠP v are linearly independent because of the presence of αI N . The
≥ λ2 R(ejϕ Y ) kΠP vk

2 middle N rows have a row rank of N − 1 because B 0 and G0
= λ2 R(ejϕ Y ) v H P v − v H P 1N 1T


are Laplacian. Also, the middle  0N rows of [A, b] have a row
N P v/N 1 0
1 1TN σ 0? /N − σ 0? =
T

rank of N −1 because of N B , G , N
= λ2 R(ejϕ Y ) v H P v − v H P (P 1N )(P 1N ) P v/N
H


 T T 
0N , 0N , 0 . Therefore, the rank condition holds, indicating
≥ λ2 R(ejϕ Y ) kvkS 1 − kP 1N k /N the system (22) has a unique equilibrium.

2 2

Proof of Theorem 3: We define an error coordinate as
= λ2 R(ejϕ Y ) kvkS 1 − 1T N P 1N /N

2

T T T
[ũT , δ̃ , ũT T T T T T T T
f ] := [u , δ , uf ] − [us , δ s , us ] to shift the
= λ2 R(ejϕ Y ) kvkS 1T N φ1 φ1 1N /(N φ1 φ1 )

2 H H


equilibrium to the origin. The resulting error dynamics are
2 1 then given as
= λ2 R(ejϕ Y ) kvkS PN


2×
N k=1 φk1 ˙ = −ηG0 ũ + ηB 0 δ̃ − ηαũf
ũ
˙
"N N Xl<k
#
X 2 X
δ̃ = −ηB 0 ũ − ηG0 δ̃ (34)
φk1 + 2 φk1 φl1 cos ∠φk1 − ∠φl1 ,
k=1 k=1 l=1
˙ f = ũ − ũf .
τ ũ
JOURNAL OF LATEX CLASS FILES 10

A positive-definite Lyapunov function is given by [16] M. Colombino, D. Groß, J.-S. Brouillon, and F. Dörfler, “Global phase
and magnitude synchronization of coupled oscillators with application
1 T 1 T 1 to the control of grid-forming power inverters,” IEEE Trans. Autom.
ũ ũ + δ̃ δ̃ + ηατ ũT
W (ũ, δ̃ f , ũ) = f ũf . Control, vol. 64, no. 11, pp. 4496–4511, 2019.
2 2 2
[17] D. Groß, M. Colombino, J.-S. Brouillon, and F. Dörfler, “The effect
The time derivative of it along (34) is of transmission-line dynamics on grid-forming dispatchable virtual os-
T cillator control,” IEEE Trans. Control Netw. Syst., vol. 6, no. 3, pp.
Ẇ (ũ, δ̃ f , ũ) = −η ũT G0 ũ − η δ̃ G0 δ̃ − ηαũT
f ũf , 1148–1160, 2019.
[18] I. Subotić, D. Groß, M. Colombino, and F. Dörfler, “A Lyapunov frame-
which is negative semi-definite. From Ẇ (ũ, δ̃ f , ũ) = 0, (34), work for nested dynamical systems on multiple time scales with appli-
T cation to converter-based power systems,” IEEE Trans. Autom. Control,
and 1T T T T
N δ̃ = 0, we obtain that [ũ , δ̃ , ũf ] = 03N . It follows vol. 66, no. 12, pp. 5909–5924, 2021.
from the LaSalle’s invariance principle [41, Corollary 4.2] that [19] G.-S. Seo, M. Colombino, I. Subotic, B. Johnson, D. Groß, and
F. Dörfler, “Dispatchable virtual oscillator control for decentralized
the system (34) is globally asymptotically stable with respect inverter-dominated power systems: Analysis and experiments,” in Proc.
to the origin. Equivalently, the system (22) is globally asymp- IEEE Appl. Power Electron. Conf. Expo., 2019, pp. 561–566.
T
totically stable with respect to the equilibrium [uT T T
s , δ s , us ] .
[20] M. Lu, S. Dhople, and B. Johnson, “Benchmarking nonlinear oscillators
For the slow system (21), u → us and θ̇ → δ̇ + 1N θ̇0 → for grid-forming inverter control,” IEEE Trans. Power Electron., vol. 37,
no. 9, pp. 10 250–10 266, 2022.
ω0 1N − η 1N 1T 0?
N σ /N as t → ∞. Namely, the voltages u [21] F. Milano, “Complex frequency,” IEEE Trans. Power Syst., vol. 37, no. 2,
and frequencies θ̇ of the slow system (21) globally converge pp. 1230–1240, 2022.
to us and ω0 + η 1T N I (e ς )/N , respectively.
jϕ ? [22] Y. Gu, Y. Li, and T. C. Green, “The nature of synchronization in
power systems: a revelation from communication theory,” arXiv preprint
arXiv:2103.16608, 2021.
R EFERENCES [23] W. Zhong, G. Tzounas, M. Liu, and F. Milano, “On-line inertia esti-
mation of virtual power plants,” Electr. Power Syst. Res., vol. 212, p.
[1] A. Sajadi, R. W. Kenyon, and B.-M. Hodge, “Synchronization in electric 108336, 2022.
power networks with inherent heterogeneity up to 100% inverter-based [24] F. Sanniti, G. Tzounas, R. Benato, and F. Milano, “Curvature-based
renewable generation,” Nat. Commun., vol. 13, no. 1, pp. 1–12, 2022. control for low-inertia systems,” IEEE Trans. Power Syst., pp. 1–4, 2022.
[2] P. Kundur, J. Paserba, V. Ajjarapu, G. Andersson, A. Bose, C. Canizares, [25] A. H. El-abiad and K. Nagappan, “Transient stability regions of multi-
N. Hatziargyriou, D. Hill, A. Stankovic, C. Taylor, T. Van Cutsem, machine power systems,” IEEE Trans. Power Appar. Syst., vol. PAS-85,
and V. Vittal, “Definition and classification of power system stability no. 2, pp. 169–179, 1966.
IEEE/CIGRE joint task force on stability terms and definitions,” IEEE [26] F. Dörfler and F. Bullo, “Kron reduction of graphs with applications to
Trans. Power Syst., vol. 19, no. 3, pp. 1387–1401, 2004. electrical networks,” IEEE Trans. Circuits Syst. I-Regul. Pap., vol. 60,
[3] A. Bergen and D. Hill, “A structure preserving model for power system no. 1, pp. 150–163, 2013.
stability analysis,” IEEE Trans. Power Appar. Syst., vol. PAS-100, no. 1, [27] S. E. Tuna, “Synchronization of linear oscillators coupled through a
pp. 25–35, 1981. dynamic network with interior nodes,” Automatica, vol. 117, p. 109008,
[4] F. Dörfler and F. Bullo, “Synchronization and transient stability in power 2020.
networks and nonuniform Kuramoto oscillators,” SIAM J. Control Op- [28] F. Milano, “A geometrical interpretation of frequency,” IEEE Trans.
tim., vol. 50, no. 3, pp. 1616–1642, 2012. Power Syst., vol. 37, no. 1, pp. 816–819, 2021.
[5] P. Kundur, N. J. Balu, and M. G. Lauby, Power System Stability and [29] M. Ilic, “Network theoretic conditions for existence and uniqueness of
Control, 3rd ed. New York, NY, USA: McGraw-Hill, 1994. steady state solutions to electric power circuits,” in Proc. IEEE Int. Symp.
[6] NERC/WECC Joint Task Force, “1200 MW fault induced solar photo- Circuits Syst., vol. 6, 1992, pp. 2821–2828 vol.6.
voltaic resource interruption disturbance report,” NERC, Atlanta, GA, [30] L. Chen, Y. Min, and W. Hu, “An energy-based method for location
USA, Tech. Rep., 2017. of power system oscillation source,” IEEE Trans. Power Syst., vol. 28,
[7] K. De Brabandere, B. Bolsens, J. Van den Keybus, A. Woyte, J. Driesen, no. 2, pp. 828–836, 2012.
and R. Belmans, “A voltage and frequency droop control method for [31] P. Yang, F. Liu, T. Liu, and D. J. Hill, “Augmented synchronization of
parallel inverters,” IEEE Trans. Power Electron., vol. 22, no. 4, pp. power systems,” arXiv preprint arXiv:2106.13166, 2021.
1107–1115, 2007. [32] J. Zhao and F. Dörfler, “Distributed control and optimization in dc
[8] X. He, H. Geng, R. Li, and B. C. Pal, “Transient stability analysis microgrids,” Automatica, vol. 61, pp. 18–26, 2015.
and enhancement of renewable energy conversion system during LVRT,” [33] R. Kogler, A. Plietzsch, P. Schultz, and F. Hellmann, “A normal
IEEE Trans. Sustain. Energy, vol. 11, no. 3, pp. 1612–1623, 2020. form for grid forming power grid actors,” 2021. [Online]. Available:
[9] L. Huang, H. Xin, Z. Wang, L. Zhang, K. Wu, and J. Hu, “Transient https://arxiv.org/abs/2106.00644
stability analysis and control design of droop-controlled voltage source [34] B. Stott, J. Jardim, and O. Alsac, “DC power flow revisited,” IEEE
converters considering current limitation,” IEEE Trans. Smart Grid, Trans. Power Syst., vol. 24, no. 3, pp. 1290–1300, 2009.
vol. 10, no. 1, pp. 578–591, 2019. [35] E. Commission, “Commission Regulation (EU) 2016/631 of 14 April
[10] Z. Shuai, C. Shen, X. Liu, Z. Li, and Z. J. Shen, “Transient angle stability 2016, establishing a network code on requirements for grid connection
of virtual synchronous generators using Lyapunov’s direct method,” of generators,” Off. J. Eur. Union, 2016.
IEEE Trans. Smart Grid, vol. 10, no. 4, pp. 4648–4661, 2019. [36] M. Amin and M. Molinas, “Small-signal stability assessment of power
[11] P. Ge, C. Tu, F. Xiao, Q. Guo, and J. Gao, “Design-oriented analysis electronics based power systems: A discussion of impedance- and
and transient stability enhancement control for a virtual synchronous eigenvalue-based methods,” IEEE Trans. Ind. Appl., vol. 53, no. 5, pp.
generator,” IEEE Trans. Ind. Electron., pp. 1–1, 2022. 5014–5030, 2017.
[12] C. Arghir, T. Jouini, and F. Dörfler, “Grid-forming control for power [37] L. Harnefors, “Modeling of three-phase dynamic systems using complex
converters based on matching of synchronous machines,” Automatica, transfer functions and transfer matrices,” IEEE Trans. Ind. Electron.,
vol. 95, pp. 273–282, 2018. vol. 54, no. 4, pp. 2239–2248, 2007.
[13] A. Tayyebi, A. Anta, and F. Dörfler, “Grid-forming hybrid angle control [38] B. Wen, D. Boroyevich, R. Burgos, P. Mattavelli, and Z. Shen, “Inverse
and almost global stability of the DC-AC power converter,” IEEE Trans. nyquist stability criterion for grid-tied inverters,” IEEE Trans. Power
Autom. Control, pp. 1–16, 2022. Electron., vol. 32, no. 2, pp. 1548–1556, 2016.
[14] M. Chen, D. Zhou, A. Tayyebi, E. Prieto-Araujo, F. Dörfler, and [39] S. Skogestad and I. Postlethwaite, Multivariable feedback control: anal-
F. Blaabjerg, “Generalized multivariable grid-forming control design for ysis and design, 2nd ed. Chichester, UK: John Wiley & Sons, 2001.
power converters,” IEEE Trans. Smart Grid, vol. 13, no. 4, pp. 2873– [40] M. Lu, S. Dutta, V. Purba, S. Dhople, and B. Johnson, “A grid-
2885, 2022. compatible virtual oscillator controller: Analysis and design,” in Proc.
[15] B. B. Johnson, M. Sinha, N. G. Ainsworth, F. Dörfler, and S. V. Dhople, IEEE Energy Convers. Congr. Expo., 2019, pp. 2643–2649.
“Synthesizing virtual oscillators to control islanded inverters,” IEEE [41] H. K. Khalil, Nonlinear Systems, 3rd ed. Englewood Cliffs, NJ, USA:
Trans. Power Electron., vol. 31, no. 8, pp. 6002–6015, 2016. Prentice-Hall, 2002.