Distributed robust secondary frequency control of inverter-based
microgrids under time-varying communication delays
Milad Gholami, Gianni Bianchini, Antonio Vicino
Abstract— This paper presents a robust secondary control
strategy for frequency synchronization and active power shar-
ing for inverter-based microgrids. The problem is addressed in a
multi-agent fashion where the local controllers of the distributed
generators play the role of agents, and communication is af-
fected by time-varying delays. The approach is fully distributed
and based on a synergic combination of linear consensus and
integral sliding-mode control. Lyapunov analysis is presented
to assess the stability properties of the closed loop. Delay-
dependent stability conditions are expressed as a set of linear
matrix inequalities whose solution yields appropriate control
gains such that frequency restoration is achieved despite delays
and active power sharing constraints. Simulations confirm the
effectiveness of the proposed control strategy.
I. I NTRODUCTION
Microgrids (MGs) are small-scale power grids consist-
ing of localized grouping of heterogeneous renewable Dis-
tributed Generators (DGs), storage systems, and loads. MGs
operate either in islanded, autonomous mode or connected
to the main power grid. The control of an AC MG has been
recently standardized into a three-layered, nested, control
architecture [1]. The inner Primary Control (PC) is generally
implemented in a decentralized way and provides commands
to each DG’s power converter. It guarantees voltage and
current stability while establishing a desired sharing of power
among DGs through droop power control terms.
Since DGs have no inertia, a Secondary Control (SC) is
needed to compensate for frequency and voltage devia-
tions introduced by the droop terms. Lastly, a Tertiary
Control (TC) may adapt the droop coefficients online to
meet additional power-flow constraints. Among the three,
the frequency SC task is of more practical interest to make
the integration of renewable generation compliant in the
existing fixed-frequency power transmission paradigm. It is
also worth mentioning that temporarily modifying the MG
frequency while preserving the power sharing among DGs is
also useful to perform seamless transition of the MG from
islanded to grid-connected mode [2].
Early SCs were centralized [1]. These solutions are now
discouraged in favour of distributed approaches because
they scale better with the MG size, are more robust to
failures, and dispense from costly central computing and
communication units [3]. Pioneering work that solves the
frequency SC problem in a distributed fashion is reported
The authors are with Dipartimento di Ingegneria dell’Informazione e
Scienze Matematiche, Università di Siena, Siena 53100, Italy (emails:
{gholami, giannibi, vicino}@diism.unisi.it.)
This work was supported in part by the Italian Ministry for Research
in the framework of the 2017 Program for Research Projects of National
Interest (PRIN) under Grant 2017YKXYXJ.
in [4]. There, a distributed averaging proportional integral
scheme is proposed, and power compensation terms are
aimed to meet the desired power sharing among DGs. In
contrast with [4], where the SC set-points are assumed known
to all the DGs, later leader-follower strategies have been
proposed. These allow for arbitrary adjusting of the MG
frequency in a distributed way by acting on the leader DG
only. As an example, [5] proposes a leader-based interaction
rule and proves its stability features in the local sense.
Although the mentioned strategies require a communica-
tion infrastructure, network-induced communication delays,
which may degrade the MG performance and even desta-
bilize it, have seldom been considered. The impact on SC
of a constant identical delay for all communication links is
analyzed in [6] and [7]. An SC that accounts for time-varying
delay is considered in [8]. A distributed finite-time control
protocol for frequency and voltage restoration under a unique
constant delay affecting the communication network is pro-
posed in [9], whereas [10] only addresses the finite-time volt-
age control problem. Along this line, recent works [11]–[13]
suggest event-triggered SC approaches with time-varying
communication delays. Although all works mentioned above
are promising, they all rely on perfect knowledge of the DG
mathematical models and power measurements.
Thus motivated, this paper proposes a leader-oriented fre-
quency SC strategy capable of restoring DG frequencies
to the desired value and establishing active power sharing
accuracy in spite of delays in the communication between
DGs and unknown load variations. We design a distributed
SC consensus protocol consisting of two distinct terms: a
nonlinear sliding-mode-based discontinuous term and a lin-
ear time-delay consensus protocol. The nonlinear part of the
proposed scheme is an Integral Sliding Mode (ISM)-based
discontinuous term using only local measurements, which
is employed to suppress the effect of disturbances on the
MG emerging behavior. The second term aims to compensate
for the unavoidable deviations of the DG output frequencies
from the set-point value while preserving the power sharing
accuracy condition. In comparison with [6]–[9], [11]–[13],
the proposed strategy requires neither power measurements
nor the use of globally known frequency set-points, thus
allowing for SC design independently from the PC loop (i.e.,
the frequency restoration approach presented in this work
is inherently robust against load variations). As opposed
to [10], the frequency restoration problem and achieving
active power-sharing are the focus of this paper. The main
contributions of the present work are as follows:
 (i) We demonstrate that in the absence of disturbance terms
affecting the MG, the linear time-delay consensus protocol
forces the output frequency of each DG to synchronize
with the set-point value despite time-varying communication
delays. (ii) We prove that considering the presence of the dis-
turbance terms yields after a finite transient a sliding motion
along which the MG exhibits the same trajectories of the
disturbance-free linear dynamics under the linear time delay
consensus protocol, for which the achievement of frequency
restoration and active power sharing accuracy was previously
demonstrated. (iii) By employing a Lyapunov-Krasovskii
function, we derive a delay-dependent control gain tuning
procedure, represented as a set of Linear Matrix Inequalities
(LMIs) whose solution allows for finding the appropriate
control gains. (iv) Lastly, it is worth mentioning that although
the proposed SC has some discontinuous control components
since they appear only in their time derivatives, the actual
control inputs are smooth. Therefore, they can safely be used
to feed the inner PCs [1]. Preliminary results of this research,
limited to the voltage SC design, have been presented in [14].
The article is structured as follows: some preliminary
notions and notations are recalled in Section II. Section III
provides the nonlinear inverter-based MG model for SC pur-
poses. The frequency SC problem, along with the proposed
strategy and the main results are illustrated in Section V.
Computer simulations are discussed in Section VI. Finally,
Section VII provides concluding remarks.
II. M ATHEMATICAL P RELIMINARIES AND N OTATIONS
Notation: The complex, real, and positive
√ numbers sets
are denoted by C, R, and R+ , and ı = −1. Let A =
[Aij ] ∈ Rn×n , its transpose is A⊺ . Assume A symmetric,
A ≻ 0 (A ⪰ 0) denotes A positive (semi-) definite. In is
the n-dimensional identity matrix and 1N is a vector with
all N elements equal to 1. card(X ) is the cardinality of set
X . Finally, the SIGN(·) operator denotes the discontinuous
multi-valued function
 the injected real power Pi (t) and reactive power Qi (t) are
1
 if Si (t) > 0
SIGN(Si (t)) ∈ [1, −1] if Si (t) = 0
−1

if Si (t) < 0
Graph Theory: A graph GN (V, E) is a topological tool used
to describe pairwise interactions between nodes, or agents
in the multi-agent field. The node set is V = {1, 2, . . . , N },
whereas the edge set E ⊆ {V × V} describes the pairwise
interactions between nodes. The adjacency matrix of GN
is A = [αij ] ∈ CN ×N , and αij denotes the edge’s weight.
If (i, j) ∈
/ E then αij = 0, otherwise αij ̸= 0. A directed
path is an alternating sequence of nodes and edges with
both endpoints of an edge appearing adjacent to it in the
sequence. Node i is a root node if it can be reached from
any other vertex by traversing a directed path.
III. M ICROGRID MODELING FOR SC DESIGN
A MG is a cyber-physical system with deeply intertwined
components. Interactions among DGs at the cyber and elec-
trical layers can thus be encoded using Graph Theory (see
Fig. 1).
MG model: As shown in Fig. 2, a DG includes a DC source,
a DC/AC inverter, a smoothing filter, a bus connector, and
the PC, which embeds current, voltage, and droop control.
Its aim is to enforce MG stability under the desired power
sharing among DGs. By applying the separation principle,
the inner current and voltage PC loops can be neglected [15],
thus yielding the following simplified DG model [5]
δ̇i (t) = ωi (t) = ω̄i (t) − kPi Pim (t), ∀i ∈ V (1)
kvi v̇i (t) =−vi (t) + v̄i (t) − kQi Qm
i (t), (2)
where δi is the i-th DG voltage’s phase angle, ωi its
frequency in rad/sec, and vi its magnitude in Vrms . The
frequency and voltage SC inputs are ω̄i and v̄i , respectively,
kvi > 0 is the voltage gain, and kPi and kQi > 0 are the so-
called droop coefficients selected to meet the given real and
reactive power sharing specifications. Pim and Qm i denote
the measured real and reactive power through the low-pass
sensor dynamics
τPi Ṗim (t) =−Pim (t) + Pi (t), τPi > 0, (3)
τQi Q̇m
i (t) =−Qm
i (t) + Qi (t), τQi > 0, (4)
where Pi and Qi are the actual active and reactive powers
at the DG output ports. The electrical interaction among
DGs over the MG can be encoded by a complex-weighted
connected graph G e (V, E e ) being V the DG set, E e the
power lines, and Ae = [Gij + ıBij ] ∈ CN ×N the complex
adjacency matrix, where Gij and Bij denote the conductance
and the inductive dominant susceptance between two DGs,
respectively. If no connection between the i-th DG and the
k-th DG exists, then Gik = Bik = 0. Let Nie = {k :
k ∈ V, k ̸= i, Bik ̸= 0}. Under the reasonable assumption
that the power transmission lines of the MG network are
lossless, i.e., Gij = 0 and Ae = [ıBij ] (see [5, Remark 1]),
obtained as
Pi (t) = PiL (t) + k∈N e vi (t)vk (t)|Bik | sin(δi (t) − δk (t)),
P
i
Qi (t) = QL 2
P
Pi (t) + vi (t) k∈Nie |Bik |
+ k∈N e vi (t)vk (t)|Bik | cos(δi (t) − δk (t))
i
(5)
L L
where Pi and Qi are the absorbed powers by load i,
whereas the remaining terms account for the injected power
flows due to voltage drops and current flows. Lastly, to
incorporate various types of loads, the so-called ZIP model
is applied [5]
PiL (t) =Piℓ1 vi2 (t) + Piℓ2 vi (t) + Piℓ3 ,
ℓ1 2 ℓ2 ℓ3
QL
i (t) =Qi vi (t) + Qi vi (t) + Qi , (6)
where the pairs (Piℓ1 , Qℓi 1 ), and (Piℓ2 , Qℓi 1 ), and (Piℓ3 , Qℓi 3 )
denote the loads at constant impedance (Z), current (I), and
power (P), respectively. Finally, it is worth mentioning that
if the SC is inactive, then ω̄i and v̄i in (1)-(2) correspond
to the nominal MG utility frequency and rated voltage, e.g.,
2π50Hz ≈ 314rad/sec and 220Vrms ≡ 310Vph−0 .
In the sequel, the following reasonable assumption is made: distributed way, DGs are assumed to be provided with
Assumption 1 The actual real and reactive power flows computing and communication facilities embedded within
their controllers, which play the role of agents. Agents
DG 1 DG 4 communicate over a network whose topology is encoded by
a real-weighted graph G c (V, E c ), where V is the same node
Reference set used to model electrical network G e (V, E e ), whereas E c
Load 1
Load 4 is the set of the available communication links. The set of
Secondary
Control
Secondary
Control
communication neighbors of the i-th DG is Nic = {j ∈
Transmission
Line 1
DG 1 DG 4
Transmission V : (i, j) ∈ E c }. According to the leader-follower paradigm,
DG 2 DG 3 Line 3
DG 2 Secondary Secondary
DG 3 we further assume the existence of a root node, labeled
Control Control
by 0 in the augmented graph GcN ({0} ∪ V, EcN ), aimed to
Communication Network Model
provide the SC set-points to a non-empty set of DGs. Clearly,
Load 2 Transmission Load 3
card(EcN ) > card(E c ). Lastly, the set of communication
Line 2
neighbors of the i-th DG with respect to graph GcN is denoted
Electrical Network Model
as Nci = {j ∈ V : (i, j) ∈ EcN }.
Fig. 1: Cyber and physical representation of a secondary
controlled inslanded MG composed of four DGs and loads. IV. F REQUENCY RESTORATION CONTROL PROBLEM
Output
Prime
Inverter
LCL Filter
Connector Let the desired frequency set-point provided by the virtual
Source

RL
node 0 be denoted by ω0 > 0. In the absence of SCs, and
LCL
by letting ω̄i = ω0 , it turns out from (1) that the resulting
PWM Power
steady-state (ss) equilibrium is such that ωi (∞) = ωss < ω0 ,
Calculation thus frequency restoration is required. Indeed, following [4]
and [5], the synchronization condition depends on the droop
Current Voltage Droop
Control Loop Control Loop Control coefficients and the balancing of real power as
PN m
Fig. 2: Block diagram of the PC. k=1 Pk (∞)
ωss = ωi (∞) = ω0 − P N 1
∀ i ∈ V (9)
k=1 kPk
at the DG’s output-ports in (5) are assumed bounded-
in-magnitude by a-priori known constants as |Pi (t)| ≤ and preserves the so-called power-sharing PC objective
Q
ΠP i , |Qi (t)| ≤ Πi , ∀ i ∈ V. kPi Pi (∞) = kPk Pk (∞) ∀ i, k ∈ V, (10)
Remark 1 Assumption 1 is justified because the power
flowing in the lines and/or absorbed by the load is bounded whose aim is to enforce a well-defined ratio between the
everywhere due to: a) the passive behaviour of loads and DG injected real power flows. Consider now (1), clearly the
lines; b) the bounded operating range of the DC-AC power- achievement of frequency restoration
converters due to their physical limits; c) the presence of ωss = ωi (∞) = ω0 ∀ i ∈ V, (11)
protection apparatus and inner voltage and current PC.
This keeps the power flows within pre-specified ranges as under the constraint (10) implies that the frequency SC inputs
discussed, e.g., in [15] and [16]. should be equal to each other at the steady state
Let us now define di (t) to be equal to the right-hand side ω̄i (∞) = ω̄ss ∀i∈V (12)
of (1) except the control term, i.e., di (t) = −kPi Pim (t). By
differentiating it along the trajectories of (1)-(6), we have being ω̄ss the frequency SC input at the steady state.
d˙i (t) = −kPi Ṗim (t), with i ∈ V. From Assumption 1 and
V. F REQUENCY SECONDARY CONTROL DESIGN
thanks to the stability imposed by the primary control, it
holds that To solve the SC problem in a distributed setting accounting
for network-induced communication delays and power flow
∃ Π ∈ R+ : |d˙i (t)| ≤ Π, ∀ t ≥ 0. (7) constraints, we consider a two-component local interaction
Thus Eq. (1) can be expressed in the next augmented form protocol as follows

ω̇i (t) =ω̄˙ i (t) + d˙i (t) with d˙i (t) = −kPi Ṗim (t) (8) ω̄˙ i (t) = u̇ci (t) + u̇di (t). (13)

where ω̄˙ i is the actual control that needs to be designed, and The first component u̇ci (t) satisfies the linear consensus
d˙i plays the rule of a time-varying disturbance term. protocol dynamics
u̇ci (t) = − j∈Nc kij (ωi (t − τ (t)) − ωj (t − τ (t)))
P
Remark 2 It is worth remarking that the upper bound (7) on
signals d˙i (t) can easily be estimated by means of a terminal P i

− j∈N c k ij uci (t − τ (t)) − ucj (t − τ (t))
i
sliding-mode disturbance observer implemented within each (14)
local controller, see, e.g., [17] and [18]. where kij and k ij are scalar control gains to be designed,
Cyber MG model: To achieve the SC objectives in a such that if (i, j) ∈ EcN then kij > 0 and k ij > 0,
otherwise kij = 0 and k ij = 0. τ (t) is a uniform time- (25) yields that
varying delay associated with communication channels. The
Vi (t) = 0, ∀t ≥ 0 → Si (t) = 0, ∀t ≥ 0, (25)
common assumption of rate-bounded time-varying delays is
made [7], [10], and [19]. i.e. a sliding motion along the manifold Si (t) = 0 takes place
Assumption 2 Let known bounds τ ⋆ , τg , exist, such that from the initial time on. According to the equivalent control
method for analyzing the sliding-mode dynamics [21], the
0 ≤ τ (t) ≤ τ ⋆ , |τ̇ (t)| ≤ τg . (15) trajectories of the discontinuous closed-loop system can be
To correlate the data received from a DG for feedback achieved by solving the equivalent closed-loop system where
purposes, we further assume that: the discontinuous control term u̇di (t) is substituted by the
associated equivalent control u̇di,eq (t). The latter is computed
Assumption 3 τ (t) is measurable for all t ≥ 0.
by solving the equation Ṡi = 0. It follows from (22) that
Remark 3 Communication protocols usually include a data-
packet time-stamp. Thus, the requirement for detectable u̇di,eq (t) = −d˙i (t), (26)
delays is costless. Once the delay τ (t) is detected, by means
Replacing (26) for u̇di (t) in the discontinuous closed-loop
of local buffers, each controller is enabled to retrieve its own
dynamics (8), (13)-(18), we can get the disturbance-free MG
state at that time t−τ (t), and performs (14). This assumption
dynamics (20). □
is common in networked applications, see, e.g., [20] and [7].
The point now is to design u̇ci (t) in (20) to achieve the control
The second control component u̇di (t) in (13) is the discon- objectives (11) and (12) asymptotically. Up to now, we found
tinuous nonlinear control that (13) degenerates into (20) for all t ≥ 0 according to
Theorem 1. It follows that the achievement (11) and (12)
u̇di (t) = −mi SIGN(Si (t)), (16) along the trajectories of (13), could be met if and only if the
where mi are scalar control gains and the sliding variable synchronization error vector

⊺
Si (t) is defined as e(t) = e1 (t), . . . , eN (t) with ei (t) = ωi (t) − ω0
Si (t) = ωi (t) − zi (t) (17) (27)
and the disagreement vector associated with uci (t), ∀i ∈ V
where zi (t) satisfies
⊺
⊺
ε(t) = ε1 (t), . . . , εN (t) = Ω uc1 (t), . . . , ucN (t)
żi (t) = u̇ci (t), (28)
(18)
zi (0) = ωi (0). go to zero, where Ω = IN −1N 1⊺N /N is the so-called average
disagreement matrix, which satisfies
Theorem 1 Consider the MG dynamics (1)-(6) under As- Ω = Ω⊺ , 1⊺N Ω = 0⊺N .
sumptions 1-3 along with the local interaction rule (13)-(18).
Let the gain parameters mi be chosen such that Indeed, one may observe that εi = 0 ∀ i implies uci (t) =
ucj (t) ∀ i, j ∈ V.
mi > Π. (19)
Let us now compute the compact form of the error dynamics
Then, the trajectories of the closed-loop system converge to on the sliding manifold by substituting (14) into (8) as
those of the linear disturbance-free MG dynamics P
ω̇i (t) = − j∈Nc kij (ωi (t − τ (t)) − ωj (t − τ (t)))
P i
ω̇i (t) =u̇ci (t) (20)


− j∈N c k ij uci (t − τ (t)) − ucj (t − τ (t)
i

Proof of Theorem 1. Consider the Lyapunov function (29)

Then by differentiating ei (t) in (27) along the trajectories
1 2
Vi (t) = 2 Si (t) . (21) of (29), one has
By straightforward manipulations, we obtain that along the ėi (t) = −kP i0 ei (t − τ (t))
trajectories of (8), (13)-(18) the time derivative of the sliding − j∈N c kij (ei (t − τ (t)) − ej (t − τ (t)))
i
variable Si (t) is given by − j∈N c k ij (uci (t − τ (t)) − uci (t − τ (t))).
P
i
(30)
Ṡi (t) = ω̇i (t) − żi (t) = d˙i (t) + u̇di (t)
(22) Moreover, by simple manipulation, the error dynamics asso-
= d˙i (t) − mi SIGN(Si (t)).
ciated with the vectors e(t) and ε(t) can be written as
Hence,    c c  
ė(t) K K e(t − τ (t))
= c (31)
V̇i (t) = Si (t) · Ṡi (t) = Si (t) · d˙i (t) − mi · |Si (t)|. ε̇(t) ΩK c ΩK ε(t − τ (t))
(23) | {z } | {z }| {z }
χ̇(t) A χ(t−τ (t))
Therefore, by (7) and (19)
V̇i (t) ≤ −|Si (t)| · (mi − Π) < 0. (24)
Note that, by (17) and (18), Vi (0) = 0, which along with
 c c
where K c = [kij
c
] ∈ RN ×N and K = [k ij ] ∈ RN ×N have manipulations taking advantage of (15), it holds
entries as
 P V̇2 (t) ≤ χ(t)⊺ Qχ(t) − χ(t − τ (t))⊺ Q(1 − τg )χ(t − τ (t)),
 j∈Nci kij if i = j (37)
c c t
kij = −kij if (i, j) ∈ EN ̸ j ,
Z
and i =

0 otherwise V̇3 (t) ≤ τ ⋆ χ̇(t)W χ̇(t) − χ̇(s)⊺ W χ̇(s)ds. (38)
t−τ (t)
 P Now, by applying the Jensen inequality to the integral term
j∈Nic k ij if i = j
in (38), it yields that

c
k ij = c
−k ij if (i, j) ∈ EN and i ̸= j , Rt
− t−τ (t) χ̇(s)⊺ W χ̇(s)ds ≤

0 otherwise
Rt
⊺ Rt

We are now in a position to present sufficient conditions − t−τ (t) χ̇(s)ds W t−τ (t) χ̇(s)ds
guaranteeing the global asymptotic stability of the closed- Hence, (38) can be recast as follows:
loop collective error dynamics on the sliding manifold (31). Rt
⊺ Rt
V̇3 (t) ≤ τ ⋆ χ̇(t)W χ̇(t) −


χ̇(s)ds W t−τ (t) χ̇(s)ds
Theorem 2 Consider a MG of V = {1, 2, . . . , N } DGs as in t−τ (t)
(39)
(1)-(6), under the SC (14). Let the communication topology
c and
GN be connected and undirected. Let node 0 be a root node
V̇3 (t) ≤ τ ⋆ χ̇(t)W χ̇(t) (40)
over GcN (0 ∪ V, EcN ). Assume the communications subjected
to time-varying delays τ , with given upper-bounds τ ⋆ and τg Finally, summing up (36), (37) and (40), and adding the
as in (15). Let Assumptions 1, 2, and 3 be in force. Let there identically zero terms
exist 2N × 2N symmetric positive definite matrices Q, R,

2 χ(t)⊺ M ⊺ + χ(t − τ (t))⊺ T ⊺ ×
P , 2N ×2N free matrices M , T , X, and positive scalars kij Rt
(41)
∀(i, j) ∈ EcN and k ij ∀(i, j) ∈ ENc
, such that the following χ(t) − χ(t − τ (t)) − t−τ (t) χ̇(t)ds
LMIs hold: Rt Rt

2 t−τ (t) χ̇⊺ (t)dsR × χ(t) − χ(t − τ (t)) − t−τ (t) χ̇(t)ds
Ξ11 −M ⊺ + T −A − M ⊺ + R⊺ 02N ×2N
 
(42)
−T ⊺ − R⊺ −A⊺ 
 
 ⋆ Ξ22
Ξ= ≺0 and 2χ̇(t)⊺ F ⊺ χ̇(t) − Aχ(t − τ (t)) where F is a nonsin-
 ⋆ ⋆ −2R 02N ×2N  gular matrix, we get
⋆ ⋆ ⋆ Ξ44
(32) V̇ (t) ≤ ρ(t)⊺ Σρ(t) (43)
with with
Ξ11 =A + A⊺ + Q + M ⊺ + M Ξ11

−M ⊺ + T −A − M ⊺ + R⊺ 02N ×2N

Ξ22 = − Q(1 − τg ) − T ⊺ − T  ⋆ Ξ22 −T ⊺ − R⊺ ⊺
−A F
Σ=

Ξ44 =τ ⋆ P + X ⊺ + X ⋆ ⋆ −2R 02N ×2N 
⋆ ⊺
Then the error dynamics on the sliding manifold (31) are ⋆ ⋆ ⋆ τ W +F +F
(44)
globally asymptotically stable.  ⊺
Rt
Proof of Theorem 2 Pick the Lyapunov function ρ(t) = χ(t)⊺ , χ(t − τ (t))⊺ , t−τ (t) χ̇(s)⊺ ds, χ̇(t)⊺
V (t) = V1 (t) + V2 (t) + V3 (t) (33) (45)
Thus if Σ ≺ 0 is satisfied, then V̇ (t) is negative definite.
where Clearly, Σ ≺ 0 if and only if Θ = Γ−1 ΣΓ ≺ 0 where
V1 (t) = χ(t)⊺ χ(t) Γ = diag{I, I, I, F −1 }. It is easily seen that if (32) holds,
Rt
V2 (t) = t−τ (t) χ(s)⊺ Qχ(t)ds then Θ ≺ 0 by taking F = X −1 and W = F P F −1 . Once a
(34)
R0 Rt solution to the LMI problem (32) is found, the control gains
V3 (t) = −τ ⋆ t+θ χ̇(s)⊺ W χ̇(s)dsdθ kij and k ij to be used in the local interaction rule (14) are
where W is some symmetric positive definite matrix. Differ- derived. □
entiating V1 (t) along the solution of (31), one has The straightforward consequence of Theorems 1 and 2 is that
provided that the gains mi are selected according to (19),
V̇1 (t) = 2χ(t)⊺ Aχ(t − τ (t)) (35)
and suitable values for the gains kij and k ij are obtained by
By employing
Rt the Newton-Leibnitz formula χ(t − τ (t)) = solving the LMI problem (32) then the MG dynamics (1)-(6)
χ(t) − t−τ (t) χ̇(s)ds, Eq. (35) can be rewritten as under Assumptions 1-3 along with the local interaction rule
Rt (13)-(18) guarantees the frequency restoration condition (11)
V̇1 (t) = χ(t)⊺ (A + A⊺ )χ(t) − 2χ(t)⊺ A t−τ (t) χ̇(t)ds and preserves the active power sharing accuracy condition
(36) (10) despite the presence of disturbance terms and the time-
Taking the time derivative of V2 (t) and V3 (t) along the delayed measurements.
solution of (31), and performing lengthy but straightforward
 VI. N UMERICAL SIMULATION
The proposed frequency SC is tested on a MG composed
of N = 4 DGs. Its rated voltage of 220Vrms per phase at
50Hz. The MG is implemented according to (1)-(6) under the
MATLAB/Simulink environment. The ode1 solver is used
with a sample time of 50µs. The MG parameters are listed
e
in Table I. The electrical and communication topologies, GN
c
and GN , are as in Fig. 1. Only DG 1 can access the SC
set-points from virtual node 0. Communication delays are

TABLE I: Parameters of the MG Test System.

DG 1 DG 2 DG 3 DG 4
τP1 0.016 τP2 0.016 τP3 0.016 τP4 0.016
τQ1 0.016 τQ2 0.016 τQ3 0.016 τQ4 0.016
kP1 6e−5 kP2 3e−5 kP3 2e−5 kP4 1.5e−5 Fig. 3: (a) DG’s local frequencies ωi (t); (b) Comparison
−4 −4
kQ1 4.2e kQ2 4.2e kQ3 4.2e−4 kQ4 4.2e−4 between the actual (Pi /Pj ) and the expected kPi /kPj , i =
kv1 1e−2 k v2 1e−2 k v3 1e−2 kv4 1e−2 1, 2, 3, 4 active power sharing ratio; (c) DG’s local frequency
Real and reactive local loads - (W, VAR) SC signals ω̄i (t); (c) DG’s local frequency SC signals ω̄i (t);
P1ℓ1 0.01 P2ℓ1 0.01 P3ℓ1 0.01 P4ℓ1 0.01 (d) DG’s local voltages vi (t).
ℓ2 ℓ2
P1 1 P2 2 P2ℓ3 3 P4ℓ3 4
P1ℓ3 1e4 P2ℓ3 1e4 P3ℓ3 1e4 P4ℓ3 1e4
Qℓ11 0.01 Qℓ21 0.01 Qℓ31 0.01 Qℓ41 0.01
of communication time-varying delays, the MG operating
Qℓ12 1 Qℓ22 2 Qℓ32 3 Qℓ42 4
working points changes due to either the activation of the
Qℓ13 1e4 Qℓ23 1e4 Qℓ33 1e4 Qℓ43 1e4
voltage SC at t = 40s, or the disconnection/connection of
Lines - [Ω−1 ] B12 = 10 B23 = 10.67 B34 = 9.82
the loads, resp., at t = 20s, and t = 35s. More specifically,
when our frequency SC is inactive during Event 1, all the
time-varying and such that, for each oriented link (i, j) ∈ DGs frequencies deviate from their reference values, and the
EcN , τ̇ (t) is randomly uniformly distributed within (−1, 1). PC cannot succeed in restoring them (see Fig. 3 (a)). Once
The delay bounds in (15) are τ = 0.5s, and τg = 1000. activated at t = 5s, the proposed distributed SC not only
We selected an upper bound (7) on signals d˙i (t) as Π = robustly restores the DG frequencies to the expected values
0.05 by employing the terminal sliding-mode disturbance (see Fig. 3 (a)) but also preserves the active power sharing
observer proposed in [17]. Thus the parameters mi = 0.1 (see Fig. 3 (b)). Let us further note that, once at t = 75s
set according to (19). To account for a more realistic use ω0 is changed from 50Hz to 50.1Hz, the MG correctly
case, the presence of the time-delay voltage SC proposed modifies its synchronous frequency while preserving the
in [14], is also considered. The appropriate control gains in desired power sharing, cfr. Figs. 3 (a) and (b). Lastly, it
(14), achieved by verifying the feasibility problem of LMI is clear from Fig. 3 (c) that the frequency SC shows a
in (32) through Yalmip Toolbox via SeDuMi solver [22], are satisfactory performance and smooth control signals.
k10 = 2.18, k12 = 1.58, k21 = 1.65, k23 = 1.7, k32 = 1.69,
k34 = 1.65, k43 = 1.83, k 12 = 1.91, k 21 = 1.65, k 23 = 1.7, VII. C ONCLUSIONS
k 32 = 1.7, k 34 = 1.65, k 43 = 1.83. To show the robustness This paper has proposed a frequency SC protocol for
features of the proposed SC to different variations on the inverter-based microgrids accounting for disturbance terms
MG operating working-points, during the simulation, the and time-varying delays in the communication links. The
following events are scheduled: method improves the current State of the Art because it
• Event 1 (t = 0s): Only the PC is active, with ω̄i (t) = is fully distributed, model-free, and robust against network-
2π50Hz ≈ 314rad/s and v̄i (t) = 220Vrms ; induced time-varying delays. Future research activities could
• Event 2 (t = 5s): The frequency SC loop with ω̄i (t) as in be devoted to considering more general, possibly multi-time
(14) and ω0 = 2π50Hz ≈ 314rad/s is closed; varying communication delays and switching topologies.
• Event 3 (t = 20s): Load 4, i.e. (P4L , QL4 ) is disconnected; R EFERENCES
• Event 4 (t = 35s): Load 4, i.e. (P4L , QL4 ) is connected; [1] J. M. Guerrero, J. C. Vasquez, J. Matas, L. G. De Vicuña, and
• Event 5 (t = 40s): The voltage SC loop with v̄i (t) as in M. Castilla, “Hierarchical control of droop-controlled ac and
[14] is closed, with v0 = 220Vrms ; dc microgrids—a general approach toward standardization,”
• Event 6 (t = 70s): The leader reference for the frequency IEEE Transactions on industrial electronics, vol. 58, no. 1,
SC is changed from 50Hz to 50.1Hz. 2010.
[2] S. D’Silva, M. Shadmand, S. Bayhan, and H. Abu-Rub,
The simulation results in Fig. 3 confirm that the proposed “Towards grid of microgrids: Seamless transition between
frequency SC scheme effectively and robustly accomplishes grid-connected and islanded modes of operation. ieee open
the control objectives (10) and (11) despite the presence j ind electron soc 2020; 1 (1): 66–81,” 2020.
 [3] F. Dörfler, J. W. Simpson-Porco, and F. Bullo, “Breaking [20] G. Chen and F. L. Lewis, “Leader-following control for
the hierarchy: Distributed control and economic optimality multiple inertial agents,” International Journal of Robust and
in microgrids,” IEEE Transactions on Control of Network Nonlinear Control, vol. 21, no. 8, 2011.
Systems, vol. 3, no. 3, 2016. [21] V. Utkin and J. Shi, “Integral sliding mode in systems
[4] J. W. Simpson-Porco, F. Dörfler, and F. Bullo, “Synchro- operating under uncertainty conditions,” in Proceedings of
nization and power sharing for droop-controlled inverters in 35th IEEE conference on decision and control, IEEE, vol. 4,
islanded microgrids,” Automatica, vol. 49, no. 9, 2013. 1996.
[5] F. Guo, C. Wen, J. Mao, and Y.-D. Song, “Distributed sec- [22] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear
ondary voltage and frequency restoration control of droop- matrix inequalities in system and control theory. SIAM,
controlled inverter-based microgrids,” IEEE Transactions on 1994.
industrial Electronics, vol. 62, no. 7, 2014.
[6] S. Liu, X. Wang, and P. X. Liu, “Impact of communication
delays on secondary frequency control in an islanded micro-
grid,” IEEE Transactions on Industrial Electronics, vol. 62,
no. 4, 2014.
[7] J. Lai, H. Zhou, X. Lu, X. Yu, and W. Hu, “Droop-based
distributed cooperative control for microgrids with time-
varying delays,” IEEE Transactions on Smart Grid, vol. 7,
no. 4, 2016.
[8] J. Lai, X. Lu, and X. Yu, “Stochastic distributed frequency
and load sharing control for microgrids with communication
delays,” IEEE Systems Journal, vol. 13, no. 4, 2019.
[9] B. Ning, Q.-L. Han, and L. Ding, “Distributed finite-time
secondary frequency and voltage control for islanded micro-
grids with communication delays and switching topologies,”
IEEE Transactions on Cybernetics, vol. 51, no. 8, 2020.
[10] A. Andreotti, B. Caiazzo, A. Petrillo, and S. Santini, “Dis-
tributed robust finite-time secondary control for stand-alone
microgrids with time-varying communication delays,” IEEE
Access, vol. 9, 2021.
[11] G. Lou, Y. Hong, W. Gu, and J. Xie, “Event-triggered time
synchronization strategy for distributed secondary control
with consideration of time-varying delays,” CSEE Journal
of Power and Energy Systems, 2022.
[12] K. Zhao, F. Xiao, and B. Wei, “Distributed asynchronous
event-triggered secondary control of microgrids with time-
varying communication delays,” in 2022 41st Chinese Con-
trol Conference (CCC), 2022.
[13] G. Zhao, L. Jin, H. Cui, and Y. Wang, “Distributed dynamic
event-triggered secondary control for islanded microgrids
with disturbances and communication delays: A hybrid sys-
tems approach,” IEEE Transactions on Industrial Informat-
ics, 2022.
[14] M. Gholami, A. Pilloni, A. Pisano, Z. A. S. Dashti, and E.
Usai, “Robust consensus-based secondary voltage restoration
of inverter-based islanded microgrids with delayed communi-
cations,” in 2018 IEEE Conference on Decision and Control
(CDC), IEEE, 2018.
[15] A. Pilloni, A. Pisano, and E. Usai, “Robust finite-time fre-
quency and voltage restoration of inverter-based microgrids
via sliding-mode cooperative control,” IEEE Transactions on
Industrial Electronics, vol. 65, no. 1, 2017.
[16] A. Bidram, A. Davoudi, F. L. Lewis, and J. M. Guerrero,
“Distributed cooperative secondary control of microgrids
using feedback linearization,” IEEE transactions on power
systems, vol. 28, no. 3, 2013.
[17] K. Shao, J. Zheng, H. Wang, X. Wang, and B. Liang,
“Leakage-type adaptive state and disturbance observers for
uncertain nonlinear systems,” Nonlinear Dynamics, vol. 105,
2021.
[18] S. Mobayen and F. Tchier, “Nonsingular fast terminal
sliding-mode stabilizer for a class of uncertain nonlinear
systems based on disturbance observer,” Scientia Iranica.
Transaction D, Computer Science & Engineering, Electrical,
vol. 24, no. 3, 2017.
[19] E. Fridman and Y. Orlov, “Exponential stability of linear
distributed parameter systems with time-varying delays,”
Automatica, vol. 45, no. 1, 2009.