76
1, JANUARY 2019
Distributed Optimal Frequency Control
Considering a Nonlinear Network-Preserving Model
Zhaojian Wang , Feng Liu , John Z. F. Pang, Steven H. Low , Fellow, IEEE, and Shengwei Mei , Fellow, IEEE
Abstract—This paper addresses the distributed optimal fre-
quency control of power systems considering a network-preserving
model with nonlinear power flows and excitation voltage dynam-
ics. Salient features of the proposed distributed control strategy
are fourfold, first, nonlinearity is considered to cope with large
disturbances, second, only a part of generators are controllable,
third, no load measurement is required, fourth, communication
connectivity is required only for the controllable generators. To
this end, benefiting from the concept of “virtual load demand,” we
first design the distributed controller for the controllable genera-
tors by leveraging the primal-dual decomposition technique. We
then propose a method to estimate the virtual load demand of each
controllable generator based on local frequencies. We derive in-
cremental passivity conditions for the uncontrollable generators.
Finally, we prove that the closed-loop system is asymptotically sta-
ble and its equilibrium attains the optimal solution to the asso-
ciated economic dispatch problem. Simulations, including small
and large-disturbance scenarios, are carried on the New England
system, demonstrating the effectiveness of our design.
Index Terms—Frequency control, network-preserving model,
distributed control, incremental output passivity.
I. INTRODUCTION
REQUENCY restoration and economic dispatch (ED) are
F two important problems in power system operation. Con-
ventionally, they are implemented hierarchically in a centralized
fashion, where the former is addressed in a fast time scale while
the latter in a slow time scale [1], [2]. While this centralized
hierarchy works well for the traditional power system, it may
not be able to keep pace with the fast development of our power
system due to: 1) slow response, 2) insufficient flexibility, 3)
low privacy, 4) intense communication, and 5) single point of
Manuscript received September 7, 2017; revised February 13, 2018 and June
15, 2018; accepted July 28, 2018. Date of publication August 1, 2018; date of
current version December 19, 2018. This work was supported in part by the Na-
tional Natural Science Foundation of China under Grants 51677100, U1766206,
and 51621065, in part by the U.S. National Science Foundation through Awards
EPCN 1619352, CCF 1637598, CNS 1545096, in part by ARPA-E Award
DE-AR0000699, and in part by the Skoltech through Collaboration Agreement
1075-MRA. Paper no. TPWRS-01380-2017. (Corresponding author: Shengwei
Mei.)
Z. Wang, F. Liu, and S. Mei are with the China State Key Laboratory
of Power System, Department of Electrical Engineering, Tsinghua Univer-
sity, Beijing 100084, China (e-mail:, wangzhaojiantj@163.com; lfeng@mail.
tsinghua.edu.cn; meishengwei@mail.tsinghua.edu.cn).
J. Z. F. Pang and S. H. Low are with the Department of Electrical Engi-
neering, California Institute of Technology, Pasadena, CA 91105 USA (e-mail:,
slow@caltech.edu; jzpang@caltech.edu).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TPWRS.2018.2861941
0885-8950 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 34, NO.
failure issue. In this regard, the idea of breaking such a hierarchy
is proposed in [3], [4].
In [5], an intrinsic connection between the asymptotic sta-
bility of the dynamical frequency control system with the
ED problem is proposed. It leads to a so-called inverse-
engineering methodology for designing optimal frequency con-
trollers, where the (partial) primal-dual gradient algorithm plays
an essential role [1], [6]. When designing optimal frequency
controllers, in choice of power flow models, including the lin-
ear model (usually associated with DC power flow, e.g. [5],
[7]–[15]) and the nonlinear model (usually associated with
AC power flow, e.g. [16]–[21]), is crucial. While the closed-
loop system can be interpreted in a linear model as carrying
out a primal-dual algorithm for solving ED, this interpreta-
tion of frequency control may not hold in a nonlinear model.
In addition to nonlinear power flow, excitation voltage dy-
namics are considered in [16]–[18], making the model more
realistic.
The aforementioned idea is further developed to enable the
design of distributed optimal frequency controllers. Roughly
speaking, the works of distributed optimal frequency control
can be divided into two categories in terms of different
power system models: network-reduced models e.g. [5],
[12]–[16], [18] and network-preserving models e.g. [7]–[11],
[17]. In network-reduced models, generators and/or loads
are aggregated and treated as one bus or control area, which
are connected to each other through tie lines. In [5], [13],
aggregated generators in each area are driven by automatic
generation control (AGC) to restore system frequency. [12],
[14]–[16], [18] further consider both the aggregated generators
and load demands in frequency control. In network-preserving
models, generator and load buses are separately handled
with different dynamic models and coupled by power flows,
rendering a set of differential algebraic equations (DAEs). In
[7], an optimal load control (OLC) problem is formulated and
a primary load-side control is derived as a partial primal-dual
gradient algorithm for solving the OLC problem. The design
approach is extended to secondary frequency control (SFC)
that restores nominal frequency in [8]. It is generalized in [9],
where passivity condition guaranteeing stability is proposed for
each local bus. Then, a unified framework combining load and
generator control is advocated in [10]. A similar model is also
utilized in [11], where only limited control coverage is needed.
Similar to [18], the Hamiltonian method is used to analyze the
network-preserving model in [17]. Compared with the network-
reduced model, the network-preserving model describes power
WANG et al.: DISTRIBUTED OPTIMAL FREQUENCY CONTROL CONSIDERING A NONLINEAR NETWORK-PRESERVING MODEL 77

systems more precisely and appear more suitable for analyzing

interactions between different control areas. Therefore, we
specifically consider the network-preserving model in this
work.
Almost all of aforementioned works assume that all buses
are controllable and load demands at all buses are accurately
measurable, especially for those proposing secondary frequency
control. Moreover, it is usually assumed that the communica-
tion network has the same topology of the power grid. These Fig. 1. Summary of notations.
assumptions are strong and arguably unrealistic for practice.
First, only a part of generators and controllable load buses can We confirm the performance of controllers via simulations on
participate in frequency control in practice. Second, the com- a detailed power system in Section VII. Section VIII concludes
munication network may not be identical to the topology of the the paper.
power grid. Third, it is difficult to accurately measure the load
injection on every bus. In extreme cases, even the number of II. POWER SYSTEM MODELS
load buses is unknown in practice. These problems highlight
some reasons why theoretical work in this area is hardly applied A. Power Network
in practical systems. Power systems are composed of many generators and loads,
In this context, a novel distributed frequency recovery con- which are integrated in different buses and connected by power
troller is proposed that only needs to be implemented on control- lines, forming a power network. Buses are divided into three
lable generator buses. To this end, a network-preserving power types, controllable generator buses, uncontrollable generator
system model is adopted, which is a more complex and practical buses and pure load buses. Denote controllable generator buses
model compared to that in [14], [15]. This work is an extension as NC G = {1, 2, . . . , nC G }, uncontrollable generator buses
of our former work [11], [14], [15]. However, in this paper, as NU G = {nC G + 1, nC G + 2, . . . , nC G + nU G }, and pure
we design a totally different controller considering a third-order load buses as NL = {nC G + nU G + 1, . . . , nC G + nU G +
nonlinear generator model with excitation voltage dynamics and nL }. Then the set of generator buses is NG = NC G ∪ NU G
nonlinear power flow. It is also motivated partly by [17], which and set of all the buses is N = NG ∪ NL . It should be noted
adopts a similar model, although our results are significantly that load can be connected to any bus besides pure load buses.
different from those in [17]. Differing from [11], the controller Let E ⊆ N × N be the set of lines, where (i, j) ∈ E if buses i
avoids load measurement, which greatly facilitates implemen- and j are connected directly. Then the whole system is modeled
√ G = (N , E). The admittance of each line
tation. By using LaSalle’s invariance principle, it is proved that as a connected graph
the closed-loop system converges to an equilibrium point that is Yij := Gij + −1Bij with Gij = 0 for every line. Denote
solves the economic dispatch problem. The salient features of the bus voltage by Vi ∠θi , where Vi is the amplitude and θi is
the proposed distributed optimal frequency controller are: the voltage phase angle. The active and reactive power Pij , Qij
1) Model: The network-preserving model of power system is from bus i to bus j is
used, including excitation voltage dynamics and nonlinear
power flow. This, unlike work on the linear model, returns Pij = Vi Vj Bij sin (θi − θj ) (1a)
a valid controller even under large disturbances. Qij = Bij Vi2 − Vi Vj Bij cos (θi − θj ) (1b)
2) Controllability: We allow an arbitrary subset of generator
buses to be controllable. For convenience, most notations are summarized in Fig. 1.
3) Controller: The distributed controller achieves the optimal
solution to economic dispatch while restoring the nominal B. Synchronous Generators
frequency, provided that certain sufficient conditions on For i ∈ NG , we use the standard third-order generator model
active power dynamics of uncontrollable generators and (e.g. [18], [22], [23]) (2a)-(2c). Here (2d) is the simplified
excitation voltage dynamics of all generators are satisfied. governor-turbine model, and (2e) is the excitation voltage con-
4) Communication: Communication is required between trol model:1
neighboring controllable generators only, and the com-
munication network can be arbitrary as long as it remains δ̇i = ωi (2a)
connected. ω̇i = (Pig − Di ωi − Pei )/Mi (2b)
5) Measurement: No load measurement is needed, and the   
controller is adaptive to unknown load changes. Ėq i = −Eq i /Td0i + Ef i /Td0i (2c)
The rest of this paper is organized as follows. In Section II, we
Ṗig = −Pig /Ti + ugi (2d)
introduce the power system model. Section III formulates the
optimal economic dispatch problem. The distributed controller Ėf i = h(Ef i , Eq i ) (2e)
is proposed in Section IV, and we further prove the optimal-
ity and stability of the corresponding equilibrium point in 1 In fact, (2a) and (2b) is a linearized model with respect to ω . The detailed
i
Section V. The load estimation method is proposed in Sectin VI. model can be found in [22, Chapter 3.9].
78 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 34, NO. 1, JANUARY 2019

In this model, Mi is the moment of inertia; Di the damp- respectively. In the following analysis, we use ηii and ηij as new

ing constant; Td0i the d-axis transient time constant; Ti the time state variables instead of δi and θi .
constant of turbine; δi the power angle of generator i; ωi the gen- To summarize, (1)–(4), (6b)–(6e), (7a), (7b) constitute the
erator frequency deviation compared to an steady state value; nonlinear network preserving model of power systems, which
Pig the mechanical power input; pj the active load demand; Pei is in a form of differential-algebraic equations (DAE).

the active power injected to network; Eq i the q-axis transient in-
ternal voltage; Eq i the q-axis internal voltage; Ef i the excitation D. Communication Network
voltage. Eq i is given by
In this paper, we consider a communication graph among

xdi  xdi − xdi the buses of controllable generators only. Denote E ⊆ NC G ×
Eq i =  E −  Vi cos (δi − θi ) (3) NC G as the set of communication links. If generator i and j
xdi q i xdi

can communicate directly to each other, we denote (i, j) ∈ E.
where xdi is the d-axis synchronous reactance, and xdi is the Obviously, edges in the communication graph E can be different
d-axis transient reactance. from lines in the power network E. For the communication
The active and reactive power (denoted by Qei ) injection to network, we make the following assumption:
the network are A1: The communication graph E is undirected and

Eq i V i connected.
Pei =  sin (δi − θi ) (4a)
xdi
III. FORMULATION OF OPTIMAL FREQUENCY CONTROL

V2 Eq i V i
Qei = i −  cos (δi − θi ) (4b) A. Optimal Power-Sharing Problem in Frequency Control
xdi xdi
The purpose of optimal frequency control is to let all the con-
For controllable generators i ∈ NC G , the capacity limits are trollable generators share power mismatch economically when
g restoring frequency. Then we have the following optimization
P gi ≤ Pig ≤ P i (5)
formulation, denoted by SFC.
g
where P gi , P i are lower and upper limits of Pig . 
SFC: g min fi (Pig ) (8a)
P i ,i∈NC G i∈NC G
C. Dynamics of Voltage Phase Angles   
To build a network-preserving power system model, relation s.t. Pig = pi − Pig ∗ (8b)
i∈NC G i∈N i∈NU G
between generators and power network should be explicitly es-
g
tablished. In this paper, loads for bus i ∈ N are simply modeled P gi ≤ Pig ≤ P i , i ∈ NC G (8c)
as constant active and reactive power injections. Then the fol-
lowing equations are used to dictate the power balance and where Pig ∗is the mechanical power of uncontrollable generator
voltage phase-angle dynamics at each bus: in the steady state. In (8a), fi (Pig ) concerns the controllable
generation Pig , satisfying the following assumption:
θ̇i = ω̃i , i ∈ N (6a) A2: The objective fi (Pig ) is second-order continuously differ-
 
entiable, strongly convex and fi (Pig ) is Lipschitz contin-
0 = Pei − D̃i ω̃i − pi − Pij , i ∈ NG (6b)
j ∈N i uous with Lipschitz constant li > 0. i.e. ∃ αi > 0, αi ≤
 
fi (Pig ) ≤ li .
0 = −D̃i ω̃i − pi − Pij , i ∈ NL (6c) To ensure the feasibility of the optimization problem, we
j ∈N i
 make an additional assumption.
0 = Qei − qi − Qij , i ∈ NG (6d) A3: The system satisfies
j ∈N i
  g   g∗  g
0 = −qi − Qij , i ∈ NL (6e) Pi ≤ pi − Pi ≤ P i (9)
j ∈N i
i∈NC G i∈N i∈NU G i∈NC G
where, pi , qi are active and reactive load demands, respectively;
ω̃i the frequency deviation at bus i; Ni the set of buses connected Specifically, we say A3 is strictly satisfied if all the inequali-
directly to bus i; D̃i the damping constant at bus i; D̃i ω̃i the ties in (9) strictly hold.
change of frequency-sensitive load [7].
In power system, line power flows are mainly related to power B. Equivalent Optimization Model With Virtual Load
angle difference between two buses rather than the power angles Demands
independently. Then, we define new variables to denote angle In (8b), load demands are injected to every bus, which some-
differences as ηii := δi − θi , i ∈ NG and ηij := θi − θj , i, j ∈ times cannot be measured accurately if at all. As a consequence,
N . The time derivative of ηii and ηij are the values of pi may be unknown to both the controllable gener-
η̇ii = ωi − ω̃i , i ∈ NG (7a) ators i, i ∈ NC G and the uncontrollable generators i, i ∈ NU G .
To circumvent such an obstacle in design, we introduce a set of
η̇ij = ω̃i − ω̃j , i, j ∈ N , i = j (7b) new variables, p̂i , to re-formulate SFC as the following equiva-
WANG et al.: DISTRIBUTED OPTIMAL FREQUENCY CONTROL CONSIDERING A NONLINEAR NETWORK-PRESERVING MODEL 79

lent problem, i.e., equivalent SFC (ESFC): Specifically, for i ∈ NC G , the controller is revised to:
 
 

ESFC: min fi (Pig ) (10a) ugi = Pig /Ti − kP ig ωi + fi (Pig ) + μi − γi− + γi+ (13a)
P ig ,i∈NC G i∈NC G
 
   
s.t. Pig = p̂i (10b) μ̇i = kμ i Pig − p̂i − (μi − μj ) − zij (13b)
i∈NC G i∈NC G j ∈N c i j ∈N c i
g
P gi ≤ Pig ≤ Pi , i ∈ NC G (10c) żij = kz i (μi − μj ) (13c)
+
where p̂i is the virtual load demand supplied by  generator i in γ̇i− = kγ i [P gi − Pig ]γ − (13d)
i
the
 steady state,
 which is a constant, satisfying i∈NC G p̂i =  g +
g∗ γ̇i+ = kγ i Pig − P i (13e)
i∈N pi − i∈NU G Pi . Obviously, the number of virtual γ i+
loads should be equal to that of the controllable generators.
where, kμ i , kz i are positive constants; Nci the set of neighbors of
Note that the power balance constraint (8b) only requires that
bus i in the communication graph; μi the local estimation of μ.
all the generators supply all the loads while it is not necessary to
Here, (13b) and (13c) are used to estimate μ locally, where only
figure out which loads are supplied exactly by which generators.
neighboring information is needed. zij is an auxiliary variable
Hence we treat virtual load demands p̂i as the effective demands
to guarantee the consistency of all μi .
supplied by generator i for dealing with the issue that only a part
For the Larangian multiplier μ, −μ is often regarded as the
of generators are controllable.
   g∗ marginal cost of generation. Theoretically, −μi should reach
Simply letting i∈NC G p̂i = i∈N pi − i∈NU G Pi , we
immediately have the following Lemma:  Since μ̇i = 0
consensus for all the generators in the steady state.
holds in the steady state, we have Pig − p̂i − j ∈N c i zij = 0.
Lemma 1. The problems SFC (8) and ESFC (10) have the
Hence, zij can be regarded as the virtual line power flow of edge
same optimal solutions.
(i, j) in the communication graph.

IV. CONTROLLER DESIGN B. Active Power Dynamics of Uncontrollable Generators

A. Distributed Frequency Control of Controllable Generators To guarantee system stability, a sufficient condition is given
1) Controller Design Based on Primal-Dual Gradient Al- for the active power dynamics of uncontrollable generators.
gorithm: Invoking the primal-dual gradient algorithm, the La- C1: The active power dynamics of uncontrollable generators
grangian of the ESFC (10) is given by are strictly incrementally output passive in terms of the
input −ωi and output Pig , i.e., there exists a continuously
    differentiable, positive semidefinite function Sω i such that
L= fi (Pig ) + μ Pig − p̂i
Ṡω i ≤ (−ωi − (−ωi∗ )) Pig − Pig ∗ − φω i (Pig − Pig ∗ )
i∈NC G i∈NC G i∈NC G
g
+ γi− (P gi − Pig ) + γi+ (Pig − P i ), i ∈ NC G (11)
where φω i is a positive definite function, and φω i = 0 holds
only when Pig = Pig ∗ .
where μ, γi− , γi+
are Lagrangian multipliers. Based on primal-
The condition C1 on the active power dynamics of uncontrol-
dual update, the controller for i ∈ NC G is designed as
lable generators is easy to verify. As an example, the commonly-
 
 used primary frequency controller
ugi = Pig /Ti − kP ig ωi + (fi (Pig ) + μ − γi− + γi+ ) (12a)
  ugi = −ωi + ωi∗ − kω i (Pig − Pig ∗ ) + Pig /Ti (14)

μ̇ = kμ Pig − p̂i (12b) satisfies C1 whenever kω i > 0. In this case, we have Sω i =
i∈NC G i∈NC G
g g∗ 2 g g∗ 2
2 (Pi − Pi ) with k1 > 0 and φ = k2 (Pi − Pi ) with 0 <
k1
+
γ̇i− = kγ i [P gi − Pig ]γ − (12c) k2 ≤ kω i · k1 .
i
 g +
γ̇i+ = kγ i Pig − P i γ i+
(12d)
C. Excitation Voltage Dynamics of All Generators
Similar to the uncontrollable generators, the following suffi-
where kP ig , kμ , kγ i are positive constants; ugi the control input;

cient condition on excitation voltage dynamics of all generators
fi (Pig ) the marginal cost at Pig . For any xi , ai ∈ R, the operator is needed to guarantee system stability, since we do not design
is defined as: [xi ]+ +
a i = xi if ai > 0 or xi > 0; and [xi ]a i = 0 specific excitation voltage controllers here.
otherwise. C2: The excitation voltage dynamics are strictly incremen-
2) Estimating μ by Second-Order Consensus: In (12b), μ is tally output passive in terms of the input −Eq i and output
a global variable, which is a function of mechanical powers and Ef i , i.e., there exists continuously differentiable, positive
loads of the entire system. To circumvent the obstacle in imple- semidefinite function SE i such that
mentation, a second-order consensus based method is utilized
to estimate μ locally by using neighboring information only. ṠE i ≤ −Eq i − (−Eq∗i ) Ef i − Ef∗ i − φE i (Ef i − Ef∗ i )
80 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 34, NO. 1, JANUARY 2019

where φE i is a positive definite function, and φE i = 0 Theorem 2 says, at equilibrium, the nominal frequency is
holds only when Ef i = Ef∗ i . recovered and marginal costs are identical for all controllable
C2 is also easy to satisfy. As an example, it can be verified that generators, implying the optimality of equilibrium.
the controller given in [23]
B. Stability
h(Ef i , Eq i ) = −Ef i + Ef∗ i − kE i (Eq i − Eq∗i ) (15)
In this section, the stability of the closed-loop system (16) is
with kE i > 0 satisfies C2. In this case, SE i = k23 (Ef i − Ef∗ i )2 proved. First we define a function as
with k3 > 0 and φ = k4 (Ef i − Ef∗ i )2 with 0 < k4 ≤ kE i · k3 .  
L̂ := fi (Pig ∗ ) + μi (Pig − p̂i )
i∈NC G i∈NC G
V. OPTIMALITY AND STABILITY  1   
− μi zij − μi (μi − μj )
After implementing the controller on the physical power sys- i∈NC G 2 i∈NC G j ∈N i
tem, the closed-loop system reads   g
+ γi− (P gi − Pig ) + γi+ (Pig − P i ) (18)
(1)–(4), (6b)–(6e), (7a), (7b) i∈NC G i∈NC G
(16)
(13a)–(13e) Denote x1 := (P g ), x2 := (μ, z, γi− , γi+ ), x := (x1 , x2 ). Then
In this section, we prove the optimality and stability of the L̂(x1 , x2 ) is convex in x1 and concave in x2 .
closed-loop system (16). Before giving the main result, we construct a Lyapunov can-
didate function composed of four parts: the quadratic part, the
A. Optimality potential energy part, conditions C1 and C2 related parts, as we
now explain.
Denote the trajectory of closed-loop system as v(t) = For i ∈ NG , the quadratic part is given by

η(t), ω(t), ω̃(t), P g (t), μ(t), z(t), γ − (t), γ + (t), Eq (t), V (t) .  1 1
Define the equilibrium set of (16) as Wk (ω, x) = Mi (ωi − ωi∗ )2 + (x − x∗ )T K −1 (x − x∗ )
2 2
V := {v ∗ |v ∗ is an equilibrium of (16)} (17) i∈NG
(19)
We first present the following Theorem2 .
where K = diag(kP ig , kμ i , kz i , kγ i ) is a diagonal positive defi-
Theorem 2. Suppose assumptions A1, A2 and A3 hold. In
nite matrix.
equilibrium of (16), following assertions are true. 
Denoting xp = (Eq i , Vi , δi , θi ), the potential energy part is
1) The mechanical powers Pig satisfy P gi ≤ Pig ∗ ≤ P i , ∀i ∈
g

NC G . Wp (xp ) = W̃p (xp ) − (xp − x∗p )T ∇x p W̃p (x∗p ) − W̃p (x∗p )

2) System frequency recovers to the nominal value, i.e. ωi∗ = (20)
0, ∀i ∈ NC G ∪ NU G , and ω̃i∗ = 0, ∀i ∈ N . where,
3) The marginal generation costs satisfy fi (Pig ∗ ) − γi−∗ +


∗  1 
γi+∗ = fj (Pj ) − γj−∗ + γj+∗ , i, j ∈ NC G .
 g 
W̃p (Eq i , Ei , Vi , δi , θi ) = Bii Vi2 + pi θ i
i∈N 2
4) Pig ∗ is the unique optimal solution of SFC problem (8). i∈N
 1
5) μ∗i is unique if A3 is strictly satisfied. − qi ln Vi − Vi Vj Bij cos (θi − θj )
Due to the page limit, here, we only outline the proof of 2
i∈N i∈N j ∈N i
Theorem 2. The detailed proof is given in Appendix. A of [24].
Outline of proof. From γ̇i− = γ̇i+ = 0 in (13d), it follows  Eq i V i
 xdi   2
− cos (δi − θi ) + Eq i
that P gi ≤ Pig ∗ ≤ P i , which is the first assertion. Set μ̇i =
g   

i∈NG
xdi i∈NG
2xdi (xdi − xdi )
0, add(13b) for i ∈ NC G , we  can prove the power bal-
g∗ g∗ (21)
ance P − p + i∈NU G Pi = 0, implying
i∈NC G i i∈N i

∗ ∗ ∗
ωi = ω̃i = ω̃j = 0. Combine (13a), (2d) and Ṗig = 0. We have 
Conditions C1 and C2 related parts are i∈NU G Sω i and
1
fi (Pig ∗ ) − γi−∗ + γi+∗ + ωi∗ + μ∗i = 0. From żij = 0 in (13c),

i∈NG T  (x d i −x  ) SE i respectively.
d0i
we get μ∗i = μ∗j = μ0 , implying the third assertion. Moreover,
di
The Lyapunov function is defined as
we can prove that (Pig ∗ , μ0 , γi−∗ , γi+∗ ) satisfies the KKT con-   SE i
dition of SFC. Since all the constraints of SFC are linear, A3 W = Wk + Wp + Sω i +   (22)
Td0i (xdi − xdi )
implies that Slater’s condition holds [25, Chapter 5.2.3]. In ad- i∈N UG i∈N G

dition, the objective function is strictly convex. Thus Pig ∗ is the Then, we give the following assumption.
unique optimal solution of SFC. If A3 is strictly satisfied, we A4: The Hessian of Wp satisfies ∇2v Wp (v) > 0 at desired
know ∃ i ∈ NC G that γi−∗ = γi+∗ = 0. Then, μ∗i = −fi (Pig ∗ ) is


g∗ equilibrium.
uniquely determined by Pi , implying the last assertion.  Since the voltage phase deviation between two neighboring
buses is not large in practice, A4 is usually satisfied. Detailed
2 In the rest of this paper, we use without explanation the superscribe “∗” to explanations can be found in Appendix. A of this paper.
stand for equilibrium of the system (16). The following stability result can be obtained.
WANG et al.: DISTRIBUTED OPTIMAL FREQUENCY CONTROL CONSIDERING A NONLINEAR NETWORK-PRESERVING MODEL 81


Theorem 3. Suppose A1–A4 and C1, C2 hold. For every i∈NU G Pig ∗ . Noticing that the
power imbalance
 is very small
v ∗ , there exists a neighborhood S around v ∗ where all trajecto- in normal operation, we have i∈NC G Pei ≈ i∈N p̂i . In fact,
ries v(t) satisfying (16) starting in S converge to the set V. In they are identical in steady state. Hence, we specify p̂i =
addition, each trajectory converges to an equilibrium point. Pei , which implies Pig − p̂i = Pig − Pei = Mi ω̇i + Di ωi . That
Due to the page limit, here, we only outline the proof of leads to an estimation algorithm of μi
Theorem 3. The detailed proof is given in Appendix. B of [24].
Outline of proof. First, we prove  
μ̇i = kμ i − (μi − μj ) − zij + Mi ω̇i + Di ωi
 
Ẇk ≤ − Di (ωi − ωi∗ )2 − D̃i (ω̃i − ω̃i∗ )2 j ∈N i j ∈N i

i∈NG i∈N

 + τi (−μi − fi (Pig ) + γi− − γi+ ) (26)
− (ωi − ω̃i )(Pei − Pei∗ )
i∈NG
 where 0 < τi < 4/li . This way, we only need to measure fre-
− (ω̃i − ω̃j )(Pij − Pij∗ ) quencies ωi at each bus locally, other than the global load de-
(i,j )∈E mands. Since the controller only needs μi of neighboring buses
 in the communication graph, it is easy to implement.
+ (Pig − Pig ∗ )(ωi − ωi∗ ) (23) Now, we reconstruct the closed-loop system by replacing
i∈NU G
(13b) with (26) in (16), which is
Then, we have the derivative of Wp
(1)–(4), (6b)–(6e), (7a), (7b)
 (Eq i − Eq∗i )(Ef i − Ef∗ i )  (Eq i − Eq∗i )2 (27)
Ẇp =   −  
(13a), (13c)–(13e), (26)
i∈NG
Td0i (xdi − xdi ) i∈NG
Td0i (xdi − xdi )
 We have the following lemma.
+ (ωi − ω̃i )(Pei − Pei∗ ) Lemma 4. Assertions 1)–5) in Theorem 2 still hold for the
i∈NG equilibrium of (27).
 The proof of Lemma 4 is similar to that of Theorem 2, details
+ (ω̃i − ω̃j )(Pij − Pij∗ ) (24) are given in Appendix C of [24]. Although only local frequency
(i,j )∈E is measured in (26), the same equilibrium can also be attained.
Consequently, the derivative of W is
  B. Discussion on Stability
ṠE i
Ẇ = Ẇk + Ẇp + Ṡω i +   Recall (2b), then (26) is derived to
i∈NU G i∈NG Td0i (xdi − xdi )
   
≤− Di (ωi − ωi∗ )2 − D̃i (ω̃i − ω̃i∗ )2 μ̇i = kμ i Pig − p̂i + p̂i − Pei − (μi − μj ) − zij
j ∈N i j ∈N i
i∈NG i∈N

 (Eq i − Eq∗i )2   + τi (−μi − fi (Pig ) + γi− − γi+ ) (28)
−   − φω i − φE i
i∈NG
Td0i (xdi − xdi ) i∈NU G i∈NG
Denote ρi = p̂i − Pei = Pei∗ − Pei , which is the difference of
electric power and its value in the steady state. We have the
≤0 (25) following assumption
As ∇2 W > 0, there exists a neighborhood set {v : W (v) ≤ A5: The disturbance can be written as ρi = βi (t)ωi , where
} for all sufficiently small > 0 so that ∇2v W (v) > 0. Hence, |βi (t)| ≤ β̄i and β̄i is a positive constant. In addition, the
there is a compact set S around v ∗ contained in such neighbor- set { t < ∞ | ωi (t) = ωi∗ } has a measure zero.
hood, which is forward invariant. Let Z1 := { v : Ẇ (v) = 0 }. Whenever ωi = ωi∗ , there alway exists such βi (t). A5 argues that
By LaSalle’s invariance principle [26, Theorem 4.4], the each ωi (t) = ωi∗ only happens at isolated points except equilibrium.
of trajectories v(t) starting from S converges to the largest in- Generally, this is reasonable in power system.
variant set Z + contained in S ∩ Z1 . From above analysis, if Denote the state variables of (27) and its equilibrium set are
Ẇ (v) = 0, v is an equilibrium point of the closed-loop system ṽ and Ṽ respectively. We have following stability result.
(16). Hence, v converges to Z + ∈ V. Finally, it is proved that Theorem 5. Suppose A1–A5, C1, C2 hold and (9) is not
the convergence of each v(t) starting from V is to a point by binding. For every ṽ ∗ , there exists a neighborhood S around ṽ ∗
following the proof of Theorem 1 in [18].  where all trajectories ṽ(t) satisfying (27) starting in S converge
to the set Ṽ whenever

VI. IMPLEMENTATION BASED ON FREQUENCY MEASUREMENT β̄i < τi Di (4 − τi li ). (29)
A. Estimation and Optimality Moreover, the convergence of each such trajectory is to a point.
Note that virtual load demands p̂i used in the controller (13) Due to the page limit, here, we only give an outline the proof
are difficult to directly measure or estimate
 in practice.
Lemma 1 of Theorem 5. The detailed proof is given in Appendix. D of
implies that any p̂i are valid as long as i∈NC G p̂i = i∈N pi − [24].
82
Fig. 2. The New England 39-bus system.
Outline of proof. We still use the Lyapunov function (22)
to analyze the stability of the closed-loop system (27). In this
situation, it is proved that Wk satisfies (23) if
 
− τi Di (4 − τi li ) < βi < τi Di (4 − τi li ) (30)
To achieve this, (29) is needed. Then, analysis of Ẇp , Ẇ as
well as the convergence to a point are same as those in the proof
of Theorem 3. 
In fact, the range of β(t) can be very large as long as
li is small enough. For example, set τi = 3/li , then β̄i (t) <
 between neighboring controllable generators.
3Di /li . As we know, if we change the objective function to
k i∈NC G fi (Pig ), k > 0, the optimal solution will not change.
Thus, li can be very small as long as k is small enough. In this
regard, (29) is not conservative. Moreover, we will illustrate in
Section VII that even (9) is binding, our controller still works.
VII. SIMULATION RESULTS
A. Test System
To test the proposed controller, the New England 39-bus sys-
tem with 10 generators as shown in Fig. 2, is utilized. In the sim-
ulation, we apply (26) to estimate the virtual load demands. All
simulations are implemented in the commercial power system
simulation software PSCAD [27], and are carried on a notebook
with 8 GB memory and 2.39 GHz CPU.
We control only a subset of these generators, namely G32,
G36, G38, G39, while the remaining are equipped with the pri-
mary frequency control given in (14). In particular, we apply the
controller (13) derived based on a simple model to a much more
realistic and complicated model in PSCAD. The detailed elec-
tromagnetic transient model of three-phase synchronous ma-
chines (sixth-order model) is adopted to simulate generators
with governors and exciters. All the lines and transformers take
both resistance and reactance into account. The loads are mod-
eled as fixed loads in PSCAD. The communication graph is
undirected and set as G32 ↔ G36 ↔ G38 ↔ G39 ↔ G32.
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 34, NO. 1, JANUARY 2019
TABLE I
CAPACITY LIMITS OF GENERATORS
Fig. 3. Diagram of the closed-loop system.
TABLE II
EQUILIBRIUM POINTS
The objective function is set as fi = 12 ai (Pig )2 + bi Pig ,
which is the generation cost of generator i. Capacity limits of
Pig and parameters ai , bi are given in Table I.
The closed-loop system is shown in Fig. 3, where each gener-
ator only needs to measure local frequency, mechanical power,
voltage and phase angle to compute its control command. Note
that there is no load measurement and only μi are communicated
B. Results Under Small Disturbances
We consider the following scenario: 1) at t = 10 s, there is
a step change of 60 MW load demands at each of buses 3,
15, 23, 24, 25; 2) at t = 70 s, there is another step change of
120 MW load at bus 23. Neither the original load demands nor
their changes are known to the generators.
1) Equilibrium: In steady states, the nominal frequency is
well recovered. The optimal mechanical powers are given in
Table II, which are identical to the optimal solution of (8) com-
puted by centralized optimization. Stage 1 is for the period from
10 s to 70 s, and Stage 2 from 70 s to 130 s. The values in
Table II are generations at the end of each stage. In Stage 1, no
generation reaches its limit, while in Stage 2 both G38 and G39
reach their upper limits. At the end of each stage, the marginal
generation cost −μi of generator i, converges identically (see
Fig. 8), implying the optimality of the results. The test results
confirm the theoretical analyses and demonstrate that our con-
troller can automatically attain optimal operation points even in
the more realistic and sophisicated model.
2) Frequency Dynamics: In this subsection, we analyze the
dynamic performance of the closed-loop system. For com-
parison, automatic generation control (AGC) is tested in the
same scenario. In the AGC implementation, the signal of
area control error (ACE) is given by ACE = Kf ω + Pij [28,
WANG et al.: DISTRIBUTED OPTIMAL FREQUENCY CONTROL CONSIDERING A NONLINEAR NETWORK-PRESERVING MODEL
Fig. 4. Dynamics of frequencies.
Fig. 5. Dynamics of mechanical powers under AGC.
Fig. 6. Dynamics of mechanical powers under the proposed control.
Chapter 11.6], where, Kf is the frequency response coefficient;
ω the frequency deviation; Pij the deviation of tie line power.
In the case studies, we can treat the whole system as one con-
trol area, implying Pij = 0. Hence, the control center computes
ACE = Kf ω and allocates it to AGC generators, G32, G36,
G38 and G39. In this situation, the control command of each
generator is ri · ACE, where i ri = 1. In this paper, we set
ri = 0.25 for i = 1, 2, 3, 4.
The trajectories of frequencies are given in Fig. 4, where the
left one stands for the proposed controller and the right one for
the AGC.
It is shown in Fig. 4 that the frequencies are recovered to the
nominal value under both controls. The frequency drops under
two controls are very similar while the recovery time under the
proposed control is much less than that under the conventional
AGC.
Mechanical power dynamics under the AGC and the proposed
controller are shown in Figs. 5 and 6, respectively. The left parts
show mechanical powers of G32, G36, G38, G39, while the
right parts show mechanical powers of other generators adopt-
ing conventional controller (14). With both controls, mechanical
powers of the generators adopting (14) remain identical in the
83
Fig. 7. Dynamics of bus voltages.
Fig. 8. Dynamics of −μ and z.
steady state. However, there are two problems when adopting
the AGC: 1) mechanical powers are not optimal; 2) mechanical
power of G39 violates the capacity limit. In contrast, the pro-
posed control can avoid these problems. In Stage 1 of Fig. 6, no
generator reaches capacity limits. In Stage 2, both G38 and G39
reach their upper limits. Then, G38 and G39 stop increasing
their outputs while G32 and G36 raise their outputs to balance
the load demands. In addition, the steady states of in both stages
are optimal, which are the same as shown in Table II.
We also illustrate in Fig. 7 the dynamics of voltage under
the proposed controller at buses 3, 15, 23, 24, 25. The voltages
converge rapidly, and only experience small drops when loads
increase. This result validates the effectiveness of the voltage
control.
The marginal generation cost of generator i under the pro-
posed controller, −μi , are shown in the left part of Fig. 8. They
converge in both stages and the steady-state values in Stage 2
are slightly bigger than that in Stage 1, as the load changes
lead to an increase in the marginal generation cost. Dynamics
of zij , (i, j) ∈ E are illustrated in the right part of Fig. 8, which
demonstrate that the steady state values do not change in the
two stages. In addition, the variation of z in transient is very
small as the deviation of μi is very small.
3) Comparison With Other Controllers in [11]: Although a
number of studies have been devoted to distributed frequency
control of power systems, most of them assume that all the
nodes are controllable except [11]. To make a fair comparison,
the controller proposed in [11] is adopted as a rival in our tests.
As shown in (8) of [11], each controller needs to predict the load
it should supply in steady state. However, it is hard to acquire an
accurate prediction in practice, which could lead to steady-state
frequency error.
Now we compare the two controls in the same scenario as
that in Section VII.B.2). The dynamics of frequencies with the
controller given in [11] are shown in Fig. 9. It is observed that
84
Fig. 9. Dynamics of frequencies with the controller in [11].
Fig. 10. Dynamics of frequencies and mechanical powers with generator trip.
there is a frequency deviation in steady state, as the prediction
is inaccurate. Although the deviation is usually quite small,
it is difficult to completely eliminate. In contrast, when the
proposed method is adopted, there is no frequency deviation in
steady state, as is shown in Fig. 4. This result is perfectly in
coincidence with the indication given by Theorems 2, 3 and 5
and Lemma 4 in this paper.
C. Performance Under Large Disturbances
In this subsection, two scenarios of large disturbances are
considered. One is a generator tripping and the other is a short-
circuit fault followed by a line tripping.
1) Generator Tripping: At t = 10 s, G32 is tripped, fol-
lowed by occurrence of certain power imbalance. Note that the
communication graph still remains connected. The output of
G32 drops to zero. Frequency and mechanical powers change
accordingly. System dynamics are illustrated in Fig. 10. The
left part of Fig. 10 shows the frequency dynamics, and the right
shows the mechanical power dynamics of controllable genera-
tors. It is observed that the system frequency experiences a big
drop at first, and then recovers to the nominal value quickly
as other controllable generators increase outputs to balance the
power mismatch. These results confirm that our controller can
adapt to generator tripping autonomously even if the tripped
generator is contributing to frequency control.
2) Short-Circuit Fault: At t = 10 s, there occurs a three-
phase short-circuit fault on line (4, 14). At t = 10.05 s, this line
is tripped by breakers. At t = 60 s, the fault is removed and line
(4, 14) is re-closed. Frequency dynamics and voltage dynamics
of buses 4 and 14 are given in Fig. 11, where the left part shows
frequency dynamics and the right shows voltage dynamics. It
can be seen that the frequency experiences violent oscillations
after the fault happens. And then it is stabilized quickly once
the line is tripped. Small frequency oscillation occurs when the
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 34, NO. 1, JANUARY 2019
Fig. 11. Dynamics of frequencies and voltages with line trip.
Fig. 12. Dynamics of −μ and mechanical powers with line trip.
line is re-closed. At the same time, the voltages of buses 4 and
14 drop to nearly zero when the fault happens. The voltages are
then stabilized to a new steady-state value in around 10s after the
fault line is tripped. They are slightly different from their initial
values because the systems operating point has changed due to
line tripping. When the tripped line is re-closed, the voltages
recover to the initial values quickly.
When line (4, 14) is tripped, the power flow across the power
network varies accordingly. As a consequence, the line loss also
changes, causing variations of mechanical powers, as shown in
Fig. 12. In the left part of Fig. 12, the inset is the dynamics of
mechanical power of G39 from 30 s to 110 s. Similarly, the inset
in the right part is dynamics of −μ of all generators. Mechanical
powers and their marginal costs all increase when the line is
tripped. However, the proposed controller compensates the loss
change autonomously.
These simulation results demonstrate that the proposed dis-
tributed optimal frequency controller can cope with large dis-
turbances such as generator tripping and short-circuit fault.
VIII. CONCLUSION
In this paper, we have designed a distributed optimal fre-
quency control using a nonlinear network-preserving model,
where only a subset of generator buses is controlled. We have
also simplified the implementation by relaxing the requirements
of load measurements and communication topology. Since non-
linearity of power flow model and dynamics of excitation volt-
age has been taken into account, our controllers can cope with
large disturbances. We have proved that the closed-loop system
asymptotically converges to the optimal solution of the eco-
nomic dispatch problem. We have also carried out substantial
simulations to verify the good performance of our controller
under both small and large disturbances.
WANG et al.: DISTRIBUTED OPTIMAL FREQUENCY CONTROL CONSIDERING A NONLINEAR NETWORK-PRESERVING MODEL
In this work, we have not considered the constraints on line
flows, since the controllable generators are selected arbitrarily
and may not suffice for congestion management. An interest-
ing problem is how to find out the minimal set of controllable
generators to fulfill the requirement of congestion management,
which is our future work.
APPENDIX
DISCUSSION ON THE REASONABLENESS OF ASSUMPTION A4
In this part, we discuss the reasonableness of Assumption
A4 by referring to [29, Lemma 3]. Reference [29] investigates
the control of inverter-based microgrids based on a network-
preserving model, while we extend some results to more com-
plicated synchronous-generator based bulk power systems. For
simplicity, we first present some notations following [29]. Com-
paring (1) and (4), Pei , Qei have same structures with Pij , Qij ,
respectively. We can treat the reactance of generator as a line

with admittance as 1/xdi , which connects i ∈ NG and inner
node of the generator. We denote the inner nodes of genera-

tors as NG . Then, we obtain an augmented power network,
whose incidence matrix is denoted by Ĉ. The set of nodes in
 
the augmented power network is denoted by N = N ∪ NG .

Denote V̂ = (Eq , V ), θ̂ = (δ, θ). Let |Ĉ| denote the matrix ob-
tained from Ĉ by replacing all the elements cij with |cij |. Define

Γ(V̂ ) := diag(|Bij |Vi Vj ), i, j ∈ N . Define A as
⎧
⎨ −|Bij | cos (θi − θj ), i = j, i, j ∈ N ;
⎪ nomic dispatch in power networks via a saddle point design approach,” in
Aij = diag(|Bii |), i = j, i, j ∈ N ; (A.1)
⎪
⎩   [14] Z. Wang, F. Liu, S. H. Low, C. Zhao, and S. Mei, “Distributed frequency
−cos (δi − θj )/xdi , i ∈ NG , j ∈ NG .
For simplicity, we use the following notation. For an n-
dimensional vector r := {r1 , r2 , . . . , rn }, the diagonal ma-
trix diag[(r1 , r2 , . . . , rn ) is denoted in short by [r]D . And
cos(·), sin(·) are defined component-wise.
From the definition of W in (22), ∇2v W (v) > 0 if and only
if ∇2v Wp (v) > 0, i.e. the matrix
 
Γ(V̂ )[cos (Ĉ T θ̂)]D [sin (Ĉ T θ̂)]D Γ(V̂ )|Ĉ|T [V̂ ]−1
D [18] T. Stegink, C. D. Persis, and A. van der Schaft, “A unifying energy-based
[V̂ ]−1
D |Ĉ|Γ(V̂ )[sin (Ĉ θ̂)]D
T
A + H(V̂ )
(A.2)
is positive definite, where
  

xd i
 0
H(V̂ ) = 2x d i (x d i −x d i ) D (A.3)
0 [qi /Vi2 ]D
In any steady state of power system (i.e.), the phase-angle
difference between two neighboring nodes is usually small. In
addition, the difference between δi and θi is also small. This
implies that the matrix in (A.2) is diagonal dominant as well as
its positive definiteness. Therefore, Assumption A4 is usually
satisfied and makes sense.
REFERENCES
[1] D. Cai, E. Mallada, and A. Wierman, “Distributed optimization decompo-
sition for joint economic dispatch and frequency regulation,” IEEE Trans.
Power Syst., vol. 32, no. 6, pp. 4370–4385, Nov. 2017.
85
[2] D. J. Shiltz, S. Baros, M. Cvetkovi, and A. M. Annaswamy, “Integration
of automatic generation control and demand response via a dynamic reg-
ulation market mechanism,” IEEE Trans. Control Syst. Technol., to be
published, doi: 10.1109/TCST.2017.2776864.
[3] A. Jokić, M. Lazar, and P. P. van den Bosch, “Real-time control of power
systems using nodal prices,” Int. J. Elect. Power Energy Syst., vol. 31,
no. 9, pp. 522–530, 2009.
[4] F. Dorfler, J. W. Simpson-Porco, and F. Bullo, “Breaking the hierarchy:
Distributed control and economic optimality in microgrids,” IEEE Trans.
Control Netw. Syst., vol. 3, no. 3, pp. 241–253, Sep. 2016.
[5] N. Li, C. Zhao, and L. Chen, “Connecting automatic generation control
and economic dispatch from an optimization view,” IEEE Trans. Control
Netw. Syst., vol. 3, no. 3, pp. 254–264, Sep. 2016.
[6] D. Feijer and F. Paganini., “Stability of primal-dual gradient dynamics
and applications to network optimization,” Automatica, vol. 46, no. 12,
pp. 1974–1981, Dec. 2010.
[7] C. Zhao, U. Topcu, N. Li, and S. H. Low, “Design and stability of load-
side primary frequency control in power systems,” IEEE Trans. Autom.
Control, vol. 59, no. 5, pp. 1177–1189, Jan. 2014.
[8] E. Mallada, C. Zhao, and S. Low, “Optimal load-side control for frequency
regulation in smart grids,” IEEE Trans. Autom. Control, vol. 62, no. 12,
pp. 6294–6309, Dec. 2017.
[9] A. Kasis, E. Devane, C. Spanias, and I. Lestas, “Primary frequency regu-
lation with load-side participation, part i: Stability and optimality,” IEEE
Trans. Power Syst., vol. 32, no. 5, pp. 3505–3518, Sep. 2017.
[10] C. Zhao, E. Mallada, S. Low, and J. Bialek, “A unified framework for
frequency control and congestion management,” in Proc. IEEE Power
Syst. Comput. Conf., 2016, pp. 1–7.
[11] J. Z. Pang, L. Guo, and S. H. Low, “Load-side frequency regulation
with limited control coverage,” ACM SIGMETRICS Perform. Evalu. Rev.,
vol. 45, no. 2, pp. 94–96, 2017.
[12] Z. Wang, F. Liu, S. H. Low, C. Zhao, and S. Mei, “Decentralized op-
timal frequency control of interconnected power systems with transient
constraints,” in Proc. 55th IEEE Conf. Decis. Control, Dec. 2016, pp. 664–
671.
[13] X. Zhang, N. Li, and A. Papachristodoulou, “Achieving real-time eco-
Proc. IEEE PES Gen. Meeting, Jul. 2015, pp. 1–5.
control with operational constraints, Part I: Per-node power balance,”
IEEE Trans. Smart Grid, vol. 10, no. 1, pp. 40–52, Jan. 2019.
[15] Z. Wang, F. Liu, S. H. Low, C. Zhao, and S. Mei, “Distributed frequency
control with operational constraints, Part II: Network power balance,”
IEEE Trans. Smart Grid, vol. 10, no. 1, pp. 53–64, Jan. 2019.
[16] T. Stegink, C. D. Persis, and A. van der Schaft, “A port-hamiltonian
approach to optimal frequency regulation in power grids,” in Proc. 54th
IEEE Conf. Decis. Control, Dec. 2015, pp. 3224–3229.
[17] T. W. Stegink, C. De Persis, and A. J. van der Schaft, “Stabilization
of structure-preserving power networks with market dynamics,” IFAC-
PapersOnLine, vol. 50, no. 1, pp. 6737–6742, 2017.
approach to stability of power grids with market dynamics,” IEEE Trans.
Autom. Control, vol. 62, no. 6, pp. 2612–2622, Jun. 2017.
[19] X. Zhang and A. Papachristodoulou, “A real-time control framework for
smart power networks: Design methodology and stability,” Automatica,
vol. 58, pp. 43–50, Aug. 2015.
[20] C. Zhao and S. Low, “Optimal decentralized primary frequency control
in power networks,” in Proc. 53rd IEEE Conf. Decis. Control, Dec. 2014,
pp. 2467–2473.
[21] S. Trip and C. De Persis, “Optimal generation in structure-preserving
power networks with second-order turbine-governor dynamics,” in Proc.
IEEE Euro. Control Conf., 2016, pp. 916–921.
[22] P. Kundur, Power System Stability and Control, vol. 7. New York, NY,
USA: McGraw-Hill, 1994.
[23] Y. Liu, C. Li, and Y. Wang, “Decentralized excitation control of multi-
machine multi-load power systems using hamiltonian function method,”
Acta Automatica Sinica, vol. 35, no. 7, pp. 919–925, 2009.
[24] Z. Wang, F. Liu, J. Z. Pang, S. Low, and S. Mei, “Distributed optimal
frequency control considering a nonlinear network-preserving model,”
IEEE Trans. Power Syst., vol. 34, no. 1, pp. 76–86, Jan. 2019.
[25] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
Cambridge Univ. Press, 2004.
[26] H. K. Khalil, Nonlinear Systems, vol. 3. Englewood Cliffs, NJ, USA
Prentice-Hall, 1996.
[27] “Home of pscad,” 2018. [Online]. Available: https://hvdc.ca/pscad/
86
[28] V. V. Arthur and R. Bergen, Power System Analysis. Upper Saddle River,
NJ, USA: Pretice-Hall, 2000.
[29] C. D. Persis, N. Monshizadeh, J. Schiffer, and F. Drfler, “A lyapunov
approach to control of microgrids with a network-preserved differential-
algebraic model,” in Proc. IEEE 55th Conf. Decis. Control, Dec. 2016,
pp. 2595–2600.
Zhaojian Wang (S’15) received the B.S. degree from Tianjin University,
Tianjin, China, in 2013, and the Ph.D. degree from Tsinghua University, Beijing,
China, in 2018. From 2016 to 2017, he was a joint Ph.D. student at California
Institute of Technology, CA, USA. He is currently a Postdoctoral Scholar with
Tsinghua University, Beijing, China. His research interests include power sys-
tem distributed control and microgrid planning.
Feng Liu (M’12) received the B.S. and Ph.D. degrees in electrical engineering
from Tsinghua University, Beijing, China, in 1999 and 2004, respectively. He is
currently an Associate Professor with Tsinghua University. From 2015 to 2016,
he was a Visiting Associate with California Institute of Technology, Pasadena,
CA, USA. He is the author/coauthor of more than 100 peer-reviewed technical
papers and two books, and holds more than 20 issued/pending patents. His
research interests include stability analysis, optimal control and game theory
based decision making in energy, and power systems. He was a Guest Editor
for the IEEE TRANSACTIONS ON ENERGY CONVERSION.
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 34, NO. 1, JANUARY 2019
John Z. F. Pang (S’13) received the B.S. degree in mathematics from Nanyang
Technological University, Singapore, in 2013. He is currently working toward
the Ph.D. degree at California Institute of Technology, Pasadena, CA, USA. His
research interests include power system stability analysis, network economics,
optimization, and approximation algorithms.
Steven H. Low (F’08) received the B.S. degree from Cornell University, Ithaca,
NY, USA, in 1987, and the Ph.D. degree from the University of California,
Berkeley, Berkeley, CA, USA, in 1992, both in electrical engineering. He is
currently the Gilloon Professor with the Computing and Mathematical Sciences
and Electrical Engineering Departments, California Institute of Technology,
Pasadena, CA, USA. Before that, he was with AT&T Bell Laboratories, Murray
Hill, NJ, USA, and the University of Melbourne, Melbourne, VIC, Australia.
He is a Senior Editor for the IEEE TRANSACTIONS ON CONTROL OF NETWORK
SYSTEMS, the IEEE TRANSACTIONS ON NETWORK SCIENCE AND ENGINEERING,
and is currently on the editorial board of NOW Foundations and Trends in
Networking, and in Power Systems, as well as that of the Journal of Sustainable
Energy, Grids and Networks. He is the author of the book Analytical methods
for network congestion control, Morgan & Claypool, 2017.
Shengwei Mei (SM’05–F’15) received the B.S. degree in mathematics from
Xinjiang University, Urumuqi, China, the M.S. degree in operations research
from Tsinghua University, Beijing, China, and the Ph.D. degree in automatic
control from the Chinese Academy of Sciences, Beijing, China, in 1984, 1989,
and 1996, respectively. He is currently a Professor with the Department of Elec-
trical Engineering, Tsinghua University. His research interests include power
system analysis and control, robust control, and complex systems.