IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 64, NO.

10, OCTOBER 2019 3967

Robust Decentralized Secondary Frequency

Control in Power Systems: Merits and Tradeoffs
Erik Weitenberg , Yan Jiang , Changhong Zhao , Member, IEEE, Enrique Mallada , Member, IEEE,
Claudio De Persis , and Florian Dörfler

Abstract—Frequency restoration in power systems is I. INTRODUCTION

conventionally performed by broadcasting a centralized sig-
HE core operation principle of an ac power system is to
nal to local controllers. As a result of energy transition, tech-
nological advances, and scientific interest in distributed
control and optimization methods, a plethora of distributed
T balance supply and demand in approximately real time.
Any instantaneous imbalance results in a deviation of global
frequency control strategies have been proposed recently, system frequency from its nominal value. Thus, a central con-
which rely on communication amongst local controllers. In trol task is to regulate the frequency in an economically ef-
this paper, we propose a fully decentralized leaky integral ficient way and despite fluctuating loads, variable generation,
controller for frequency restoration, which is derived from a and possibly faults. Frequency control is conventionally per-
classic lag element. We study steady-state, asymptotic opti- formed in a hierarchical architecture: The foundation is made
mality, nominal stability, input-to-state stability, noise rejec- of the generators’ rotational inertia providing an instantaneous
tion, transient performance, and robustness properties of
this controller in closed loop with a nonlinear and multivari-
frequency response, and three control layers—primary (droop),
able power system model. We demonstrate that the leaky in- secondary automatic generation (AGC), and tertiary (economic
tegral controller can strike an acceptable tradeoff between dispatch)—operate at different time scales on top of it [1], [2].
performance and robustness as well as between asymp- Conventionally, droop controllers are installed at synchronous
totic disturbance rejection and transient convergence rate machines and operate fully decentralized, but they cannot by
by tuning its dc gain and time constant. We compare our themselves restore the system frequency to its nominal value.
findings to conventional decentralized integral control and To ensure a correct steady-state frequency and a fair power shar-
distributed-averaging-based integral control in theory and ing among generators, centralized AGC and economic dispatch
simulations. schemes are employed on longer time scales.
Index Terms—Decentralized control, power generation This conventional operational strategy is currently challenged
control, power system stability. by increasing volatility on all time scales (due to variable re-
newable generation and increasing penetration of low-inertia
sources) as well as the ever-growing complexity of power sys-
Manuscript received April 16, 2018; revised September 17, 2018; ac- tems integrating distributed generation, demand response, mi-
cepted November 20, 2018. Date of publication December 3, 2018; date crogrids, HVdc systems, etc. Motivated by these paradigm shifts
of current version September 25, 2019. The work of E. Weitenberg and
C. De Persis was supported in part by the NWO-URSES project EN-
and recent advances in distributed control and optimization, an
BARK, the DST-NWO IndoDutch Cooperation on “Smart Grids—Energy active research area has emerged developing more flexible dis-
management strategies for interconnected smart microgrids” and the tributed schemes to replace or complement the traditional fre-
STW Perspectief program “Robust Design of Cyber-physical Systems” – quency control layers.
“Energy Autonomous Smart Microgrids”. The work of Y. Jiang and E. Mal- In this paper, we focus on secondary control. We refer to [3,
lada was supported in part by the NSF under Grant CNS 1544771, Grant
EPCN 1711188, Grant AMPS 1736448, and Grant CAREER 1752362, Section IV-C] for a survey covering recent approaches amongst
and in part by the U.S. DoE Award DE-EE0008006. The work of C. Zhao which we highlight semicentralized broadcast-based schemes
was supported in part by ARPA-E under the NODES program under similar to AGC [4]–[6] and also highlight distributed schemes
Grant DE-AC36-08GO28308, and in part by the DOE under the EN- based on consensus-based averaging [7]–[12] or primal dual
ERGISE program Award DE-EE0007998. The work of F. Dörfler was
supported by ETH funds and the SNF Assistant Professor Energy under
methods [13]–[16], which all rely on communication amongst
Grant 160573. Recommended by Associate Editor Prof. Yann Le Gorrec. controllers. However, because of security, robustness, and eco-
(Corresponding author: Florian Dörfler.) nomic concerns it is desirable to regulate the frequency without
E. Weitenberg and C. De Persis are with the University of Gronin- relying on communication. A seemingly obvious and often ad-
gen, Groningen 9712 CP, The Netherlands (e-mail:, e.r.a.weitenberg@ vocated solution is to complement local proportional droop con-
gmail.com; [email protected]).
Y. Jiang and E. Mallada are with the Johns Hopkins University, Balti- trol with decentralized integral control [5], [7], [17]. In theory,
more MD 21218 USA (e-mail:, [email protected]; [email protected]). such schemes ensure nominal and global closed-loop stability at
C. Zhao is with the National Renewable Energy Laboratory, Golden a correct steady-state frequency, though in practice they suffer
CO 80401 USA (e-mail:, [email protected]). from poor robustness to measurement bias and clock drifts [4],
F. Dörfler is with the Automatic Control Laboratory, Swiss Federal
Institute of Technology (ETH) Zürich, Zürich 8092, Switzerland (e-mail:,
[5], [11], [18]. Furthermore, the power injections resulting from
[email protected]). decentralized integral control generally do not lead to an effi-
Color versions of one or more of the figures in this paper are available cient allocation of generation resources. A conventional remedy
online at http://ieeexplore.ieee.org. to overcome performance and robustness issues of integral con-
Digital Object Identifier 10.1109/TAC.2018.2884650 trollers is to implement them as lag elements with finite dc gain

0018-9286 © 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.

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
3968 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 64, NO. 10, OCTOBER 2019

[19]. Indeed, such decentralized lag element approaches have generation minus local load in the reduced model), and u ∈ Rn
been investigated by practitioners: Ainsworth and Grijalva [17] is a control input to be designed later. Finally, the magnetic en-
provides insights on the closed-loop steady-states and transient ergy stored in the purely inductive (lossless) power transmission
dynamics based on numerical analysis and asymptotic argu- lines is (up to a constant) given by
ments, Heidari et al. [20] provides a numerical certificate for
1 
n
ultimate boundedness, and Han et al. [21] analyzes lead-lag
U (θ) = − Bij Vi Vj cos (θi − θj )
filters based on a numerical small-signal analysis. 2 i,j =1
Here, we follow the latter approach and propose a fully de-
centralized leaky integral controller derived from a standard lag where Bij ≥ 0 is the susceptance of the line connecting genera-
element. We consider this controller in feedback with a nonlin- tors i and j with terminal voltage magnitudes Vi , Vj > 0, which
ear and multivariable multimachine power system model and are assumed to be constant.
provide a formal analysis of the closed-loop system concerning Observe that the vector of power injections
the following:
1) steady-state frequency regulation, power sharing, and dis- 
n

patch properties; (∇U (θ))i = Bij Vi Vj sin (θi − θj ) (2)

2) the transient dynamics in terms of nominal exponential j =1
stability and input-to-state stability with respect to disturbances
affecting the dynamics and controller; satisfies a zero net power flow balance: 1Tn ∇U (θ) = 0, where
3) the dynamic performance as measured by the H2 norm. 1n ∈ Rn is the vector of unit entries. In what follows, we also
All of these properties are characterized by precisely quan- write these quantities in compact notation as
tifiable tradeoffs—dynamic versus steady-state performance as
U (θ) = −1T Γ cos (B T θ), ∇U (θ) = BΓ sin (B T θ)
well as nominal versus robust performance—that can be set by
tuning the dc gain and time constant of our proposed controller. where B ∈ Rn ×m is the incidence matrix [22] of the power
We compare our findings with the corresponding properties of transmission grid connecting the n generators with m trans-
decentralized integral control, and we illustrate our analytical mission lines, and Γ ∈ Rm ×m is the diagonal matrix with its
findings with a detailed simulation study based on the IEEE diagonal entries being all the nonzero Vi Vj Bij s corresponding
39-bus power system. We find that our proposed fully decen- to the susceptance and voltage.
tralized leaky integral controller is able to strike an acceptable We note that all our subsequent developments can also be ex-
tradeoff between dynamic and steady-state performance and tended to more detailed structure-preserving models with first-
can compete with other communication-based distributed con- order dynamics (e.g., due to power converters), algebraic load
trollers. flow equations, and variable voltages by using the techniques
The remainder of this paper is organized as follows. developed in [7] and [9]. In the interest of clarity, we present
Section II lays out the problem setup in power system frequency our ideas for the concise albeit stylized model (1).
control. Section III discusses the pros and cons of decentral-
ized integral control and proposes the leaky integral controller.
Section IV analyzes the steady-state, stability, robustness, and B. Secondary Frequency Control
optimality properties of this leaky integral controller. Section V In what follows, we refer to a solution [θ(t), ω(t)] of (1) as a
illustrates our results in a numerical case study. Finally, synchronous solution if it is of the form θ̇(t) = ω(t) = ωsync 1n ,
Section VI summarizes and discusses our findings. where ωsync is the synchronous frequency.
Key to the analysis of part of the results in this paper (see Lemma 1 (Synchronization frequency): If there is a syn-
Section IV-B) is a strict Lyapunov function. A first attempt to chronous solution to the power system model (1), then the syn-
arrive at one was made in preliminary work [7]. The current chronous frequency is given by
paper is substantially different from [7], as it establishes several n 
novel and stronger results; provides additional context, motiva- P ∗ + ni=1 u∗i
tion, and possible implications; and discusses the tradeoffs that ωsync = i=1i n (3)
i=1 Di
arise from the tunable controller parameters.
where u∗i denotes the steady-state control action.
II. POWER SYSTEM FREQUENCY CONTROL Proof: In the synchronized case, (1b) reduces to Dωsync 1n +
∇U (θ) = P ∗ + u. After multiplying this equation by 1Tn and
A. System Model using 1Tn ∇U (θ) = 0, we arrive at the claim (3). 
Consider a lossless, connected, and network-reduced power Observe from (3) that ωsync = 0 if and only if all injec-
n ∗ ∗
system with n generators modeled by the swing equations [1] tions are balanced: i=1 Pi + ui = 0. In this case, a syn-
chronous solution coincides with equilibrium (θ∗ , ω ∗ , u∗ ) ∈
θ̇ = ω (1a) T n × {0n } × Rn of (1). Our first objective is frequency regu-
∗
M ω̇ = − Dω + P − ∇U (θ) + u (1b) lation, also referred to as secondary frequency control.
Problem 1 (Frequency restoration): Given an unknown con-
where θ ∈ T n and ω ∈ Rn are the generator rotor angles and stant vector P ∗ , we design a control strategy u = u(ω) to stabi-
frequencies relative to the utility frequency given by 2π50 or lize the power system model (1) to an equilibrium (θ∗ , ω ∗ , u∗ ) ∈
2π60 Hz. The diagonal matrices M, D ∈ Rn ×n collect the in- T × {0n } × R so that i=1 Pi∗ + u∗i = 0.
n n n

ertia and damping coefficients Mi , Di > 0, respectively. The Observe that there are manifold choices of u∗ to achieve this
generator primary (droop) control is integrated in the damping task. Thus, a further objective is the most economic allocation of
coefficient Di , P ∗ ∈ Rn is vector of net power injections (local steady-state control inputs u∗ given by a solution to the following

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
WEITENBERG et al.: ROBUST DECENTRALIZED SECONDARY FREQUENCY CONTROL IN POWER SYSTEMS: MERITS AND TRADEOFFS 3969

optimal dispatch problem: here. First, note that (7) can be explicitly integrated as

n
u = −T −1 (θ − θ0 ) − p0 = −T −1 (θ − θ0 ) (8)
minimizeu ∈Rn ai u2i (4a)
i=1 where we used θ0 = θ0 − T p0 as a shorthand. In what follows,
we study only the state [θ(t), ω(t)] without p(t) since p(t) is a

n 
n
function of θ(t) and initial conditions as defined in (8).
subject to Pi∗ + ui = 0 . (4b)
Next, consider the LaSalle function
i=1 i=1
1 T
The term ai u2i with ai > 0 is the quadratic generation cost V(θ, ω) = ω M ω + U (θ) − θT P ∗
for generator i. Observe that the unique minimizer u of the 2
linearly constrained quadratic program (4) guarantees identical 1
+ (θ − θ0 )T T −1 (θ − θ0 ). (9)
marginal costs at optimality [8], [10] as 2
ai ui = aj uj ∀i, j ∈ {1, . . . , n} . (5) The derivative of V along any trajectory of (1), (7) is
We remark that a special case of the identical marginal cost V̇(θ, ω) = −ω T Dω . (10)
criterion (5) is fair proportional power sharing [23] when the
coefficients ai are chosen inversely to a reference power P̄i > 0 Note that for any initial condition (θ0 , ω0 ) ∈ T × R , the sub-
n n

(normally the power rating) for every generator i given by level set Ω := {(θ, ω) | V(θ, ω) ≤ V(θ0 , ω0 )} is compact. In-
deed, Ω is closed because of continuity of V and bounded since
ui /P̄i = uj /P̄j ∀i, j ∈ {1, . . . , n} . (6) V is radially unbounded because of quadratic terms in ω and θ.
The set Ω is also forward invariant since V̇ ≤ 0 by (10).
The optimal dispatch problem (4) also captures the core ob-
In order to proceed, we define the zero-dissipation set as
jective of the so-called economic dispatch problem [24], and  
it is also known as the base point and participation factors E = (θ, ω) | V̇(θ, ω) = 0 = {(θ, ω) | ω = 0n } (11)
method [24, Ch. 3.8].
Problem 2 (Optimal frequency restoration): Given an un- and EΩ := E ∩ Ω. By LaSalle’s theorem [25, Th. 4.4], as t →
known constant vector P ∗ , we design a control strategy u = +∞, [θ(t), ω(t)] converges to a nonempty, compact, invariant
u(ω) to stabilize the power system model (1) to an equilib- set LΩ , which is a subset of EΩ . In the following, we show that
rium (θ∗ , ω ∗ , u∗ ) ∈ T n × {0n } × Rn where u∗ minimizes the any point (θ , ω ) ∈ LΩ is an equilibrium of (1), (7). Owing to
optimal dispatch problem (4). the invariance of LΩ , the trajectory [θ(t), ω(t)] starting from
Besides steady-state optimal frequency regulation, we will (θ , ω ) stays identically in LΩ and thus in EΩ . Therefore, by
also pursue certain robustness and transient performance char- (11) we have ω(t) ≡ 0 and hence ω̇(t) ≡ 0. Thus, every point
acteristics of the closed loop that we specify later. on this trajectory, in particular the starting point (θ , ω ), is an
equilibrium of (1), (7). 
III. FULLY DECENTRALIZED FREQUENCY CONTROL The astonishing global convergence merit of decentralized
The frequency-regulation Problems 1 and 2 have seen many integral control comes at a cost though. First, note that the
centralized and distributed control approaches. Since P ∗ is gen- steady-state injections from decentralized integral control (7)
erally unknown, all approaches explicitly or implicitly rely on u∗ = −T −1 (θ∗ − θ0 ) − p0
integral control of the frequency error. In the following, we fo-
cus on fully decentralized integral control approaches making depend on initial conditions and the unknown values of P ∗ .
use only of local frequency measurements: ui = ui (ωi ). Thus, in general, u∗ does not meet the optimality criterion (5).
Second and more importantly, internal instability due to decen-
A. Decentralized Pure Integral Control tralized integrators is a known phenomenon in control systems
[26], [27]. In this particular scenario, as shown in [11, Th. 1]
One possible control action is decentralized pure integral and [4, Proposition 1], the decentralized integral controller (7)
control of the locally measured frequency, i.e., is not robust to arbitrarily small biased measurement errors that
u= −p (7a) may arise, e.g., due to clock drifts [18]. More precisely, the
closed-loop system consisting of (1) and the integral controller
T ṗ = ω (7b) subject to measurement bias η ∈ Rn
where p ∈ R is an auxiliary local control variable, and T ∈
n
u= −p (12a)
Rn ×n is a diagonal matrix of positive time constants Ti > 0. The
closed-loop system (1), (7) enjoys many favorable properties, T ṗ = ω + η , (12b)
such as solving the frequency-regulation Problem 1 with global does not admit any synchronous solution unless η ∈ span(1n ),
convergence guarantees regardless of the system or controller that is, all biases ηi , for all i ∈ {1, . . . , n}, are perfectly iden-
initial conditions or the unknown vector P ∗ . tical [4, Proposition 1]. Thus, while theoretically favorable, the
Theorem 2 (Convergence under decentralized pure integral decentralized integral controller (7) is not practical.
control): The closed-loop system (1), (7) has a nonempty
set X ∗ ⊆ T n × {0n } × Rn of equilibria, and all trajectories
B. Decentralized Lag and Leaky Integral Control
[θ(t), ω(t), p(t)] globally converge to X ∗ as t → +∞.
Proof: This proof is based on an idea initially proposed in [7] In standard frequency-domain control design [19] a stable
while we make some arguments and derivations more rigorous and finite dc-gain implementation of a proportional-integral (PI)

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
3970 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 64, NO. 10, OCTOBER 2019

controller is given by a lag element parameterized as (θ∗ , ω ∗ , p∗ ) of the form

Ts + 1 α−1 θ̇∗ = ω ∗ (16a)
α = 1
 +
αT s + 1
proportional control
αT s + 1
  0n = − Dω ∗ + P ∗ − ∇U (θ∗ ) − p∗ (16b)
leaky integral control ∗ ∗
0n = ω − K p (16c)
where T > 0 and α  1. The lag element consists of a propor- ∗
where ω = ωsync 1n for some ωsync ∈ R. 
tional channel as well as a first-order lag often referred to as a By eliminating the variable p∗ from (16), we arrive at
leaky integrator. In our context, a state-space realization of a
decentralized lag element for frequency control is given by P ∗ − (D + K −1 ) ωsync 1n = ∇U (θ∗ ) . (17)
Equation (17) takes the form of lossless active power flow equa-
u = − ω − (α − 1)p tions [1] with injections P ∗ − (D + K −1 ) ωsync 1n . Thus, As-
αT ṗ = ω − p sumption 1 is equivalent to assuming feasibility of the power
flow (17), which is always true for sufficiently small P ∗ .
where T is a diagonal matrix of time constants, and α  1 is Under this assumption, we now show various properties of
scalar. In what follows, we disregard the proportional channel the closed-loop system (15) under leaky integral control (13).
(that would add further droop) and focus on the leaky integrator
to remedy the shortcomings of pure integral control (7). A. Steady-State Analysis
Consider the leaky integral controller
We begin our analysis by studying the steady-state character-
u= −p (13a) istics. At a steady state, the control input u∗ takes the value
T ṗ = ω − K p (13b) u∗ = −p∗ = −K −1 ω ∗ = −K −1 ωsync 1n (18)
−1
that is, it has a finite dc gain K similar to a primary droop
where K, T ∈ Rn ×n are diagonal matrices of positive control
control. The following result is analogous to Lemma 1.
gains Ki , Ti > 0. The transfer function of the leaky integral
Lemma 3 (Steady-state frequency): Consider the closed-
controller (13) at a node i (from ωi to −ui ) is given by
loop system (15) and its equilibria (16). The explicit synchro-
1 Ki−1 nization frequency is given by
Ki (s) = = (14) n
T i s + Ki (Ti /Ki ) · s + 1 P∗
ωsync = n i=1 i −1 . (19)
i=1 Di + Ki
i.e., the leaky integrator is a first-order lag with dc gain Ki−1
and bandwidth Ki /Ti . It is instructive to consider the following Unsurprisingly, the leaky integral controller (13) does not
limiting values for the gains.  regulate the synchronous frequency ωsync to zero un-
generally
1) For Ti  0, leaky integral control (13) reduces to pro- less i Pi∗ = 0. However, it can achieve approximate frequency
regulation within a prespecified tolerance band.
portional (droop) control with gain Ki−1 .
Corollary 4 (Banded frequency restoration): Consider the
2) For Ki  0, we recover the pure integral control (7). closed-loop system (15). The synchronous frequency ωsync takes
3) For Ki  ∞ or Ti  ∞, we obtain an open-loop system a value in a band around zero, which can be made arbitrarily
without control action. small by choosing the gains Ki > 0 sufficiently small. In par-
Thus, from a loop-shaping perspective for open-loop sta- ticular, for any ε > 0, if
ble single-input-single-output (SISO) systems, we expect good 
steady-state frequency regulation for a large dc gain Ki−1 , and a n
| ni=1 Pi∗ | 
n
Ki−1 ≥ − Di (20)
large (respectively, small) cut-off frequency Ki /Ti likely results ε
i=1 i=1
in good nominal transient performance (respectively, good noise
rejection). We will confirm these intuitions in the next section, then |ωsync | ≤ ε.
where we analyze the leaky integrator (13) in closed loop with While regulating the frequencies to a narrow band is suf-
the nonlinear and multivariable power system (1) and highlight ficient in practical applications, the closed-loop performance
its merits and tradeoffs as function of the gains K and T . may suffer since the control input (13) may become ineffective
because of a small bandwidth Ki /Ti . Similar observations have
also been made in [17] and [20]. We will repeatedly encounter
IV. PROPERTIES OF THE LEAKY INTEGRAL CONTROLLER
this tradeoff for the decentralized leaky integral controller (13)
The power system model (1) controlled by the leaky integrator between choosing a small gain K (for desirable steady-state
(13) gives rise to the closed-loop system properties) and large gain (for transient performance).
The closed-loop steady-state injections are given by (18), and
θ̇ = ω (15a) we conclude that the leaky integral controller achieves propor-
∗ tional power-sharing by tuning its gains appropriately.
M ω̇ = − Dω + P − ∇U (θ) − p (15b)
Corollary 5 (Steady-state power-sharing): Consider the
T ṗ = ω − K p . (15c) closed-loop system (15). The steady-state injections u∗ of the
leaky integral controller achieve fair proportional power-sharing
We make the following standing assumption on this system. as follows:
Assumption 1 (Existence of a synchronous solution): As-
sume that the closed-loop (15) admits a synchronous solution Ki u∗i = Kj u∗j ∀i, j ∈ {1, . . . , n} . (21)

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
WEITENBERG et al.: ROBUST DECENTRALIZED SECONDARY FREQUENCY CONTROL IN POWER SYSTEMS: MERITS AND TRADEOFFS 3971

Hence, arbitrary power-sharing ratios, as in (6), can be pre- The existence of (δ ∗ , ω ∗ , p∗ ) is guaranteed by Assumption 1.
scribed by choosing the control gains as Ki ∼ 1/P̄i . Similarly, Additionally, we make the following standard assumption con-
we have the following result on steady-state optimality. straining steady-state angle differences.
Corollary 6 (Steady-state optimality): Consider the closed- Assumption 2 (Security constraint): The synchronous solu-
loop system (15). The steady-state injections u∗ of the leaky tion (24) is such that B T δ ∗ ∈ Θ := (− π2 + ρ, π2 − ρ)m for a
integral controller minimize the optimal dispatch problem as constant scalar ρ ∈ (0, π2 ).
follows: Remark 1: Compared with the conventional security con-
n straint assumption [8], we introduce an extra margin ρ on the
minimizeu ∈Rn Ki u2i (22a) constraint to be able to explicitly quantify the decay of the Lya-
i=1 punov function we use in proofs of Theorems 7 and 8. 
By using Lyapunov techniques following [12], it is possi-

n 
n
ble to show that the leaky integral controller (13) guarantees
subject to Pi∗ + (1 + Di Ki )ui = 0 . (22b)
exponential stability of the synchronous solution (24).
i=1 i=1
Theorem 7 (Exponential stability under leaky integral con-
Proof: Observe from (21) that the steady-state injections (18) trol): Consider the closed-loop system (13) and (23). Let As-
meet the identical marginal cost requirement (5) with ai = Ki . sumptions 1 and 2 hold true. The equilibrium (δ ∗ , ω ∗ , p∗ ) is
Additionally, the steady-state equations (16b), (16c), and (18) locally exponentially stable. In particular, given the incremental
can be merged to the expression state
0n = DK u∗ + P ∗ − ∇U (θ∗ ) + u∗ . x = x(δ, ω, p) = col(δ − δ ∗ , ω − ω ∗ , p − p∗ ) (25)
By multiplying the left-hand side of this equation by we ar- 1Tn , ∗ ∗
the solutions x(t) = col(δ(t) − δ , ω(t) − ω , p(t) − p ), with ∗
rive at condition (22b). Hence, the injections u∗ are also feasible [δ(t), ω(t), p(t)] a solution to (13) and (23) that starts sufficiently
for (22) and thus optimal for program (22).  close to the origin satisfy for all t ≥ 0,
The steady-state injections of the leaky integrator are opti-
mal for the modified dispatch problem (22) with appropriately x(t)2 ≤ λe−α t x0 2 (26)
chosen cost functions. From (22b),the leaky integrator does not where λ and α are positive constants. In particular, when mul-
achieve perfect power balancing ni=1 Pi∗ + u∗i = 0 and under- tiplying the gains K and T by the positive scalars κ and τ ,
estimates the net load, but it can satisfy the power balance (4b) respectively, α is monotonically nondecreasing as a function of
arbitrarily well for K chosen sufficiently small. Note that in the gain κ and nonincreasing as a function of τ .
practice, the control gain K cannot be chosen arbitrarily small Proof: Consider the incremental Lyapunov function from
to avoid ineffective control and the shortcomings of the decen- [12], including a cross term between potential and kinetic ener-
tralized integrator (7) (lack of robustness and power sharing). gies, as
The following sections will make these ideas precise from the
1
perspectives of stability, robustness, and optimality. V (x) = (ω − ω ∗ )T M (ω − ω ∗ )
2
B. Stability Analysis + U (δ) − U (δ ∗ ) − ∇U (δ ∗ )T (δ − δ ∗ )
For ease of analysis, in this section, we introduce a change 1
of coordinates for the voltage phase angle θ. Let δ = θ − + (p − p∗ )T T (p − p∗ )
2
n 1n 1n θ = Πθ be the center-of-inertia coordinates (see e.g.,
1 T

[28], [9]), where Π = I − n1 1n 1Tn . In these coordinates, the + (∇U (δ) − ∇U (δ ∗ ))T M ω (27)
open-loop system (1) becomes where ∈ R is a small positive parameter.
δ̇ = Πω (23a) First, we will show that this is indeed a valid Lyapunov func-
tion, by proving positivity outside of the origin and strict nega-
∗
M ω̇ = −Dω + P − ∇U (δ) + u (23b) tivity of its time derivative along the solutions of (23).
For sufficiently small values of and if Assumption 2 holds,
where by an abuse of notation we use the same symbol U for
V (x) satisfies condition
the potential function expressed in terms of δ
U (δ) = −1T Γ cos (B T δ), ∇U (δ) = BΓ sin (B T δ). β1 x2 ≤ V (x) ≤ β2 x2 (28)

Note that B T Π = B T since B T 1n = 0n [22]. The synchronous for some β1 , β2 > 0 and for all x with B T δ ∈ Θ, by Lemma 14
solution (θ∗ , ω ∗ , p∗ )1 defined in (16) is mapped into the point in Appendix A. The derivative of V (x) can be expressed as
(δ ∗ , ω ∗ , p∗ ), with δ ∗ = Πθ∗ , satisfying conditions V̇ (x) = −χT H(δ)χ
δ̇ ∗ = 0n (24a) where χ(δ, ω, p) := col(∇U (δ) − ∇U (δ ∗ ), ω − ω ∗ , p − p∗ )
0n = −Dω ∗ + P ∗ − ∇U (δ ∗ ) − p∗ (24b) ⎡ ⎤
I 1
2 D − 12 I
∗
0n = ω − K p . ∗ ⎢ ⎥
(24c) H(δ) = ⎣ 12 D D − E(δ) 0n ×n ⎦ (29)
− 12 I 0n ×n K
1 Of course, care must be taken when interpreting the results in this section
since the steady-state itself depends on the controller gain K (see Section IV-A). and we defined the shorthand E(δ) = symm(M ∇2 U (δ)) with
Here we are merely interested in the stability relative to the equilibrium. symm(A) = 12 (A + AT ).

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
3972 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 64, NO. 10, OCTOBER 2019

We claim that for all δ, H(δ) > 0. To see this, apply ing to λm in (H(δ)). Then, for any vector z satisfying z = 1,
Lemma 12 from Appendix A to obtain H(δ) ≥ H (δ) with λm in (H(δ)) = eTm in H(δ)em in ≤ z T H(δ)z. Hence, we have
⎡ ⎤
2I 0n ×n 0n ×n β4 = min λm in (H(δ)) = minBT δ ∈Θ , z :z =1 z T H(δ)z
⎢ ⎥ B T δ ∈Θ
H (δ) := ⎣ 0n ×n D − (E(δ) + D2 ) 0n ×n ⎦.
where the last equality holds by noting that em in is one of the
0n ×n 0n ×n K− I
vectors z at which the minimum is attained.
Given that D and K are positive-definite matrices, one can select Now, suppose we multiply K by a factor κ > 1. Let H (δ) =
to be positive yet sufficiently small so that H (δ) > 0. H(δ) + block diag(0, 0, (κ − 1)K). The new value of β4 is
To show exponential decline of the Lyapunov function V (x), given by
which is necessary for proving (26), we must find some positive  n 
constant α such that V̇ (x) ≤ −αV (x). β4 = min z T H(δ)z + (κ − 1)Ki z2n
2
+i .
B T δ ∈Θ , z :z =1  i=1

We claim that a positive constant β3 , dependent on ρ from
=z T H (δ )z
Assumption 2, exists such that χ2 ≥ β3 x2 . To see this, we
note from Lemma 13 in Appendix A that a constant β3 exists so The argument of the minimization is not smaller than z T H(δ)
that z for any z. It follows that β4 ≥ minBT δ ∈Θ,z :z =1 z T H(δ)z =
∇U (δ) − ∇U (δ̄)2 ≤ β3 δ − δ ∗ 2 . (30) β4 . Similarly, if 0 < κ < 1, then β4 ≤ minBT δ ∈Θ,z :z =1 z T H
(δ)z = β4 . Hence, β4 is a monotonically nondecreasing func-
The claim then follows with β3 = min(1, β3 −1 ).
tion of gain κ. Likewise, α is a monotonically decreasing func-
In order to proceed, we set β4 := minBT δ ∈Θ λm in (H(δ)).
tion of β2 , which itself is a nondecreasing function of τ . 
Then, using (28), it follows that as far as B T δ ∈ Θ
Theorem 7 is in line with the loop-shaping insight that the
β3 β4 bandwidth Ki /Ti determines nominal performance: The decay
V̇ (x) ≤ −β4 χ2 ≤ −β3 β4 x2 ≤ − V =: −αV (x) . rate α is monotonically nondecreasing in Ki /Ti .
β2
For this inequality to lead to the claimed exponential stability,
we must guarantee that the solutions do not leave Θ. To do so, C. Robustness Analysis
we study the sublevel sets of V (x) and find one that is con- We now depart from nominal performance and focus on ro-
tained in Θ. Recall that the sublevel sets of V (x) are invariant bustness. Recall a key disadvantage of pure integral control: It
and thus solutions x(t) are bounded for all t ≥ 0 in sublevel is not robust to biased measurement errors of the form (12).
sets {x : V (x) ≤ V (x0 )} for which B T δ ∈ Θ. Hence, we re- We now show that leaky integral control (13) is robust to such
quire the initial conditions x0 of solutions x(t) to be within a measurement errors. In what follows, instead of (13), consider
suitable sublevel set {x : V (x) ≤ V (x0 )} where B T δ ∈ Θ. We leaky integral control subjected to the following measurement
now construct such a sublevel set. Let errors
ξ2 u = −p (32a)
c := β1 (31)
λm ax (BB T )
T ṗ = ω − K p + η (32b)
where ξ > 0 is a parameter with the property that any δ sat-
isfying B T δ − B T δ ∗  ≤ ξ also satisfies B T δ ∈ Θ. The param- where the measurement noise η = η(t) ∈ R is assumed to be
n

eter ξ exists because B T δ ∗ ∈ Θ, and Θ is an open set. Ac- an ∞-norm bounded disturbance. In this case, the bias-induced
cordingly, define the sublevel set Ωc := {x : V (x) ≤ c}, with c instability (see Section III-A) does not occur.
defined above, and note that any point in Ωc satisfies B T δ ∈ Θ. Let us first offer a qualitative steady-state analysis. For a
2 constant vector η, the equilibrium equation (16c) becomes
As a matter of fact V (x) ≤ c implies x2 ≤ λm a xξ(BBT ) and
2
therefore δ − δ ∗ 2 ≤ λm a xξ(BBT ) . This, in turn, implies that 0n = ω ∗ − K p∗ + η
B T (δ − δ ∗ )2 ≤ ξ 2 , and hence B T δ ∈ Θ by the choice of ξ. so that the closed loop (1), (32) will admit synchronous equi-
We conclude that any solution issuing from the sublevel set libria. Indeed, the governing equations (17) determining the
Ωc will remain inside of it. Hence, along these solutions, the synchronous frequency ωsync rewritten as
inequality V̇ (x) ≤ −αV (x) always holds true.
By the comparison lemma [25, Lemma B.2], this inequality (D + K −1 ) ωsync 1 = P ∗ − ∇U (θ∗ ) − K −1 η .
yields V (x(t)) ≤ e−α t V (x(0)), which we combine again with Observe that the noise terms η now take the same role as
(28) to arrive at (26) with λ = β2 /β1 . the constant injections P ∗ , and their effect can be made ar-
Finally, we address the effect of K and T on α by introducing bitrarily small by increasing K. We now make this qualitative
the scalar factors κ and τ multiplying K and T , and by studying steady-state reasoning more precise and derive a robustness cri-
the effect of manipulations of κ and τ on the exponential decline terion by means of the same Lyapunov approach used to prove
of V (x) and, therefore, of x(t). Note that α is a monotonically Theorem 7. We take the measurement error η as disturbance
increasing function of β4 = minBT δ ∈Θ λm in (H(δ)). Recall that input and quantify its effect on the convergence behavior along
for any vector z, we have the lines of input-to-state stability. First, we define the specific
λm in (H(δ))z2 ≤ z T H(δ)z robust stability criterion that we will use, adapted from [29].
Definition 1 (Input-to-state-stability with restrictions): A
with equality if z is the eigenvector corresponding to λm in system ẋ = f (x, η) is said to be input-to-state stable (ISS) with
(H(δ)). Let em in denote the normalized eigenvector correspond- restriction X on x(0) = x0 and restriction η ∈ R> 0 on η(·) if

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
WEITENBERG et al.: ROBUST DECENTRALIZED SECONDARY FREQUENCY CONTROL IN POWER SYSTEMS: MERITS AND TRADEOFFS 3973

there exist a class KL-function β and a class K∞ -function γ states, X . In the remainder of the proof, we fix η̄ such that
such that
η̄ 2 = α̂cμ
x(t) ≤ β(x0 , t) + γ(η(·)∞ )
with c defined as in (31) in the proof of Theorem 7.
for all t ∈ R≥0 , x0 ∈ X , and inputs η(·) ∈ Ln∞ satisfying Define the sublevel set Ωc , again as in the proof of Theorem 7.
We now claim that the solutions of the closed-loop system
η(·)∞ := ess sup η(t) ≤ η. cannot leave Ωc . In fact, on the boundary ∂Ωc of the sublevel
t∈R≥0
set Ωc , the right-hand side of (34) equals −α̂c + μ1 η2 , which
Theorem 8 (ISS under biased leaky integral control): is a nonpositive constant by the choice of η̄. Hence, a solution
Consider system (23) in a closed loop with the biased leaky in- leaving Ωc would contradict the property that V̇ (x) ≤ 0 for all
tegral controller (32). Let Assumptions 1 and 2 hold true. Given x ∈ ∂Ωc . We conclude that all solutions must satisfy (34) for
a diagonal matrix K > 0, there exist a positive constant η and all t ∈ R≥0 . Hence, we choose X = Ωc .
a set X such that the closed-loop system is ISS from the noise Having validated (34), we now derive the exponential bound
η to the state x = col(δ − δ ∗ , ω − ω ∗ , p − p∗ ) with restrictions (33). By the Comparison Lemma, the use of convolution integral
X on x0 and η on η(·), where (δ ∗ , ω ∗ , p∗ ) is the equilibrium and bounding η(t)2 by η(·)2∞ , we arrive at
of the nominal system, i.e., with η = 0. In particular, the
solutions x(t) = col(δ(t) − δ ∗ , ω(t) − ω ∗ , p(t) − p∗ ), with 1
V (x(t)) ≤ e−α̂ t V (x0 ) + η(·)2∞ .
(δ(t), ω(t), p(t)) a solution to (23), (32) for which x(0) ∈ X α̂μ
and η(·)∞ ≤ η are satisfied for all t ∈ R≥0 , are given by
We combine this inequality with (28) and (30) to arrive at (33)
x(t)2 ≤ λe−α̂ t x(0)2 + γη(·)2∞ (33) with λ = β2 /β1 and γ = (α̂β1 μ)−1 .
Finally, we address the effects of K and T on α̂ and γ by
where α̂, λ, and γ are positive constants. Furthermore, when introducing the scalar factors κ and τ multiplying K and T .
multiplying the gains K and T by the positive scalars κ and τ , As κ increases, there is no need to increase , while it is
respectively, then γ is monotonically decreasing (respectively, possible to increase μ. Analogously to the reasoning in the
nonincreasing) as a function of κ (respectively, τ ), and α̂ is proof of Theorem 7, increasing the value of κ for constant and
monotonically nondecreasing as a function of κ and nonin-
increasing μ cannot lower the value of β̂4 and α̂, but decreases
creasing as a function of τ .
the value of γ. If one decreases κ, but multiplies μ by the same
Proof: We start by extending the Lyapunov arguments from
the proof of Theorem 7 to take the noise η(t) into account, factor so as to keep β̂4 constant, μ will also decrease. This
guarantees α̂ remains constant in this case, preserving its status
obtaining again an upper bound of V̇ (x) in terms of V (x).
as a nondecreasing function of κ. However, a decrease in μ
From the proof of Theorem 7 recall the Lyapunov func-
results in an increase in γ, retaining its status as a decreasing
tion derivative V̇ (x) = −χT H(δ)χ − (p − p∗ )T η. Since for any function of κ. Therefore, α̂ is nondecreasing as a function of κ
positive parameter μ we have and γ is decreasing.
1 As in Theorem 7, τ affects only β1 and β2 , and the same
−(p − p∗ )T η ≤ μp − p∗ 2 + η2 result holds true: α̂ is a monotonically nonincreasing function
μ
of τ . Analogously, γ is monotonically nonincreasing in τ . 
one further obtains Theorem 8 shows that larger gains K (and T ) reduce
   (respectively, do not amplify) the effect of the noise η on the
0 0 0 1
V̇ (x) ≤ −χ H(δ) − 0 0
T
0 χ+ η2 . state x. This further emphasizes the tradeoff between frequency
0 0 μI μ banding and controller performance already touched on in
  Section IV-A. We further extend and formalize this tradeoff in
= Ĥ (δ ) Section V-D by means of an H2 performance analysis.
Remark 2 (Exponential ISS with restrictions): The KL–
Following the reasoning in the proof of Theorem 7, we note that function from the ISS inequality (33) is an exponential
Ĥ(δ) ≥ Ĥ (δ), where function, so the stability property is in fact exponential ISS with
⎡ ⎤ restrictions. The need to include restrictions X on the initial
2I 0n ×n 0n ×n conditions and η̄ on the noise is due to the requirement of
Ĥ (δ) := ⎣ 0n ×n D − (E(δ) + D2 ) 0n ×n ⎦.
maintaining the state response within the safety region Θ. 
0n ×n 0n ×n K − I − μI

It follows that for sufficiently small values of and μ, Ĥ(δ) ≥ D. H2 Performance Analysis
Ĥ (δ) > 0. To continue, let β̂4 := minBT δ ∈Θ λm in (Ĥ(δ)). As a All findings thus far show that the closed-loop performance
result, we find that for a positive constant α̂ = ββ3 β̂2 4 crucially depends on the choice of Ki and Ti . Small gains Ki
are advantageous for steady-state properties, large gains Ki and
1 Ti are advantageous for noise rejection, and the nominal perfor-
V̇ (x) ≤ −α̂V (x) + η2 (34) mance does not deteriorate when increasing Ki /Ti . To further
μ
understand this tradeoff we now study the transient performance
for all x such that B T δ ∈ Θ. in the presence of stochastic disturbances by means of the H2
We now again make sure that no solutions can leave the set norm. The use of the H2 norm for evaluating power network
Θ. To make this possible, it is necessary to impose a restriction performance was first introduced in [30]. This versatile frame-
on the magnitude of the noise η̄ and the set of possible initial work allows us to characterize various network properties such

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
3974 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 64, NO. 10, OCTOBER 2019

as resistive power losses [30], voltage deviations [31], the role Although a closed-form solution of (37) is generally hard to cal-
of inertia [32], and phase coherence [33], in the presence of culate, it is possible to provide a qualitative analysis by assuming
stochastic disturbances, and network-wide frequency transients homogeneous parameters as in the following result.
induced by step changes [34], [35]. Theorem 9 (H2 norm of leaky integrator): Consider the LTI
Here, in a stochastic setting, we investigate the effect of the power system model Gleaky in (35). Assume homogeneous pa-
gains K and T on the steady-state frequency variance in the pres- rameters, i.e., Mi = m, Di = d, Ti = τ , Ki = k, σζ ,i = σζ ,
ence of power disturbances and noisy frequency measurements and ση ,i = ση , ∀i ∈ {1, . . . , n}. Then, the squared H2 norm of
modeled as white noise inputs. More precisely, we compute the Gleaky is given by
H2 norm of the system (15) with output ω(t) and inputs in
(15b) and (15c). With this aim, we first linearize (15) around a
steady state (θ∗ , ω ∗ , p∗ ).2 Using ∇2 U (θ∗ ) = LB , where LB is Gleaky 2H2
a weighted Laplacian matrix [22], and redefining (θ, ω, p) as a nσζ2 
n
− k σ2 + σ2
deviation from steady state, the closed-loop model (15) becomes = +   md ζ  η  . (39)
2md i=1 2d mk 2 + d + dτ k + τ + λi τ 2
θ̇ = ω
M ω̇ = − Dω − LB θ − p In particular, setting k = 0 in (39) gives
T ṗ = ω − Kp .
We use Sζ ζ to denote the disturbances on the net power nσζ2 
n
ση2
Gintegrator 2H2 = + (40)
injection and Sη η to model the noise incurred in the frequency 2md i=1 2d (τ + λi τ 2 )
measurement required to implement the controller (13). Then,
by defining the system output as y = ω, we obtain the linear
time-invariant (LTI) system as where Gintegrator denotes the linearized power system model con-
⎡ ⎤ ⎡ ⎤  trolled by the pure integral controller (7).
θ̇ 0 I 0 θ Proof: Consider the orthonormal change of input, state,
⎣ ω̇ ⎦ = ⎣ −M −1 LB −M −1 D −M −1 ⎦ ω and output variables θ = U θ , ω = U ω , p = U p , y = U y ,
ṗ 0 T −1 −T −1 K p ζ = U ζ , and η = U η , where U is the orthonormal trans-
  formation that diagonalizes LB : U T LB U = diag{λ1 , . . . , λn }
=A
⎡ ⎤ with λi being the ith eigenvalue of LB in increasing order
    (λ1 = 0 < λ2 ≤ · · · ≤ λn ). The H2 norm is invariant under this
0 0 θ
ζ
+ ⎣M −1 Sζ 0 ⎦ , y = [0 I 0] ω . transformation and (35) decouples into n subsystems as
0 T −1 Sη
η   p
  =C ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
=B θ̇i 0 1 0 θi 0 0  
⎢ ⎥ ⎢ λi 1 ⎥⎣ ⎦ ⎢ σζ ⎥ ηp,i
(35) ⎣ ω̇i ⎦ = ⎣ − m − m − m ⎦ ωi + ⎣ m 0 ⎦
d
σ ηω ,i
The signals ζ ∈ Rn and η ∈ Rn represent white noise with unit ṗi 0 1
− τk pi 0 τη
variance, i.e., E[ζ(t)T ζ(τ )] = δ(t − τ )In and E[η(t)T η(τ )]  
τ
 
=A i =B i
= δ(t − τ )In , and Sζ = diag{σζ ,i , i ∈ {1, . . . , n}}, Sη =
diag{ση ,i , i ∈ {1, . . . , n}}. ⎡ ⎤
θi
We are interested in understanding the effects of Ki and
Ti on the system performance. To this aim, we will compute yi = [ 0 1 0 ] ⎣ ωi ⎦ . (41)
 
the H2 norm of (35) and compare it with that of the pure =C i pi
integrator, as well as the open loop system. From (14), we see
that for Ki  0 (respectively, for Ki  ∞) for i ∈ {1, . . . , n}
we recover the closed-loop system controlled by pure integral Then, based on (37) and (38), Gleaky 2H2 can be calculated by
control (7) (respectively, the open-loop system). Thus, in computing the norm of the n subsystems (41) (see, e.g., [30],
what follows, we denote the LTI system (35) by Gleaky , for [32], [36]–[38]). The key step is to solve n Lyapunov equations
K = 0n ×n by Gintegrator , and for Ki  ∞ by Gopen-loop .
The squared H2 norm of the LTI system (35) is given by ATi Q + QAi = −CiT Ci (42)
G2H2 T
= lim E[y (t)y(t)]. (36)
t→∞
where Q must be symmetric and can thus be parameterized as
Via the observability Gramian X, G2H2 can be computed as
 
G2H2 = tr(B T XB) (37) q11 q12 q13
Q= q12 q22 q23 . (43)
where X solves the Lyapunov equation q13 q23 q33
AT X + XA = −C T C. (38)
Whenever λi = 0 (42) has a unique solution Q. For λ1 = 0, the
2 Of course, care must be taken when interpreting the results in this section
system (41) has a zero pole that could render infinite H2 norm
since the steady-state itself depends on the controller gain K (see Section IV-A), and nonunique solutions to (42). We will later see that this mode
but here we are merely interested in the transient performance. is unobservable and thus the H2 norm is finite.

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
WEITENBERG et al.: ROBUST DECENTRALIZED SECONDARY FREQUENCY CONTROL IN POWER SYSTEMS: MERITS AND TRADEOFFS 3975

We now focus on the case λi = 0. Direct calculations show under leaky integral control is strictly smaller than that under
that pure integral control: Gleaky 2H2 < Gintegrator 2H2 . Moreover,
 
λi km 1 λi in the absence of power disturbances, σζ = 0, Gleaky 2H2 is a
q11 = − 2 q33 + − q33 (44a) strictly decreasing function of k ≥ 0 and τ ≥ 0.
d τ 2 τ
Remark 3 (Optimal H2 performance at open loop): Observe
q12 = 0 (44b) from (39) that in the absence of power disturbances (σζ = 0)
and in the presence of measurement noise (ση = 0), the opti-
q13 = λi q33 (44c)
mal gains are k  ∞ or τ  ∞, which from (14) reduces to
 
m km 1 the open-loop case. This insight is consistent with the noise
q22 = − 2 q33 + (44d) rejection bounds (33) in Theorem 8. Of course, the steady-
d τ 2
state characteristics in Section IV-A all demand a sufficiently
km small value of k, and power disturbances will typically be
q23 = − q33 (44e)
τ present as well. Nevertheless, these considerations pose the
where all solutions are parameterized in question of whether leaky integral control can ever improve
the open-loop performance Gopen-loop 2H2 := nσζ2 /(2md) ob-
1 tained for k, τ  ∞. We explicitly address this question
q33 = m   . (45)
2d 2
τ2 k +
m
dτ 2 + τd k + 1
τ + λi below. 
The next corollary (for proof, see Appendix B2) will use
Therefore, we obtain the characterization of the effect of τ on the performance as a
 σ 2
ση2 mechanism to derive an optimal choice for both k and τ that
ζ
Gleaky,i 2H2 = tr(BiT QBi ) = q22 +
q33 . (46) can ensure the improvement of the leaky integrator performance
m τ2
Gleaky H2 not only with respect to the pure integrator perfor-
By substituting (44d) and (45) into (46), we arrive at mance Gintegrator H2 but also with respect to the open-loop
Gleaky,i 2H2 = performance Gopen-loop H2 .
 σ2  Corollary 11 (H2 optimal tuning): Under the assumption of
σ η2 Theorem 9 and for any τ > 0, and k such that
τ2 − d + k
k ζ
σζ2
   + . (47)  2
2d τm2 k 2 + dτm2 + τd k + 1
τ + λi 2md k ση
> (48)
We now consider the case λi = 0, i.e., i = 1. Since λ1 = 0, d σζ
neither ω̇1 , nor ṗ1 , nor y1 depends on θ1 in (41). Thus, θi is not the closed-loop performance under the leaky integral control
observable, and we can simplify the system (41) to outperforms the open-loop system performance, i.e.,
   d    σ  
ω̇i − m − m1 ωi ζ
0 ηp,i Gleaky 2H2 < Gopen-loop 2H2 .
= + m ση
ṗi 1
− τk pi 0 ηω ,i Moreover, the global minimum of the H2 norm under leaky

τ
   τ
=A i =B i integral control is obtained by setting τ → τ ∗ = 0 and k to
  2    σ 2

ωi ∗ ση ζ
yi = [ 1 0 ] . k =d 1+ 1+ . (49)
  pi σζ d
Ci
Remark 4 (Necessity of condition (48)): We highlight that
Again, we solve the Lyapunov equation (42), but here Q = condition (48) is, in fact, necessary for improving performance
QT is a 2-by-2 matrix. A similar calculation as before yields beyond Gopen-loop H2 . When (48) is violated, ∂∂τ Gleaky 2H2 <
that Gleaky,1 2H2 is also given by (47) with λ1 = 0. Therefore, 0; see Appendix B2. In this case, if (48) does not hold, it is easy

Gleaky 2H2 = ni=1 Gleaky,i 2H2 , which is equal to (39). to see from (39) that Gleaky H2  Gopen-loop H2 as τ  ∞,
Finally, note from (7) and (13) that the leaky integra- which implies Gleaky H2 > Gopen-loop H2 . 
tor reduces to an integrator when K = 0n ×n . It follows that Corollary 11 suggests that the optimal controller tuning re-
Gintegrator 2H2 can be obtained by setting k = 0 in (39).  quires τ ∗ = 0, which reduces the leaky integrator to a propor-
Theorem 9 provides an explicit expression for the closed- tional droop controller with gain 1/k ∗ . However, setting τ to
loop H2 performance under leaky integral control (13) as well small values reduces the response time Ti /Ki = τ /k of the
as under pure integral control (7). Observe from (37), (39), and leaky integrator, which in an actual implementation will be lim-
(40) that power disturbances and measurement noise have an ited by the actuator’s response time (not modeled here). We
independent additive effect on the H2 norm. Thus, either of the point out, however, that Corollary 11 also shows that the leaky
two effects can be obtained by setting ση = 0 or σζ = 0. integrator provides performance improvements for any τ > 0,
The following corollary, whose proof is in Appendix B1, and thus this limitation will only affect the extent to which the
shows the supremacy of leaky integral control over pure integral H2 performance is improved.
control for any positive gain k. Furthermore, in the presence of The optimal value k ∗ in (49) also unveils interesting trade-
only measurement noise, increasing k or τ always improves offs between performance and robustness. More precisely, on
Gleaky 2H2 which is consistent with the ISS insights obtained the one hand, in the high-power disturbance regime σζ  ∞,
from Theorem 8. the optimal gain is k ∗  0. The latter choice of course weakens
Corollary 10 (Monotonicity of the H2 norm): Under the as- the robustness properties described in Section IV-B. On the other
sumptions of Theorem 9, for any k > 0 the closed-loop H2 norm hand, in the presence of large measurement errors ση  ∞, one

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
3976 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 64, NO. 10, OCTOBER 2019

all simulations below, a 300-MW step increase in active-power

load occurs at each of buses 15, 23, and 39 at time t = 5 s.

A. Comparison Between Controllers Without Noise

We implement each of the following controllers across the 10
generators to stabilize the system after the increase in load.
1) Distributed-averaging based integral control (DAI):
u= −p (50a)
T ṗ = A−1 ω − LAp . (50b)
T
Here, L = L is the Laplacian matrix of a communication
graph among the controllers, which we choose as a ring
graph with uniform weights 0.1. The matrix A is diagonal
with entries Aii = ai being the cost coefficients in (4a)
chosen as 1.0 for generators G3, G5, G6, G9, and G10,
and 2.0 for the rest. We choose the time constant Ti =
0.05 s for every generator i. The DAI control (50) is
known to achieve stable and optimal frequency regulation
as in Problem 2; see [7]–[12]. Even though DAI control is
based on a reliable and fast communication environment,
Fig. 1. 39-bus New England system used in simulations. we include it here as a baseline for comparison purposes.
2) Decentralized pure integral control (7) with time constant
loses the ability to properly regulate the frequency as k ∗  ∞, Ti = 0.05 s for every generator i.
i.e., the open-loop case. 3) Decentralized leaky integral control (13) with time con-
Remark 5 (Joint banded frequency restoration and optimal stant Ti = 0.05 s for every generator i. The gain Ki equals
H2 performance): This last discussion also unveils a critical 0.005 for generators G3, G5, G6, G9, and G10, and 0.01
tradeoff of leaky integral control: It may be infeasible to jointly for the rest. The values of Ki are proportional to the val-
satisfy (20) and (48) when the measurement noise ση is large.
ues of ai in DAI (50) so that the dispatch objectives (4a)
For a specified level ε of frequency restoration, the parameter k
that satisfies (20), or equivalently and (22a) are identical.
Fig. 2 (dashed plots) shows the frequency at G1 (all other
  ∗ −1
| i Pi | generators display similar frequency trends), and Fig. 3 shows
k≤ −d the active-power outputs of all generators, under the different
nε
controllers above and without noisy measurements. First, note
may not satisfy (48) and thus leads to worse performance than that all closed-loop systems reach stable steady-states; see Theo-
that of open loop. Of course, one can still take large values of rems 2 and 8. Second, observe from Fig. 2 that both pure integral
τ to mitigate this degradation, as in Remark 3. However, this and DAI control can perfectly restore the frequencies to the nom-
comes at the cost of lower convergence rate: Large τ leads to inal value, whereas leaky integral control leads to a steady-state
slow feedback. We refer to Section VI for further discussion of frequency error as predicted in Lemma 3. Third, as observed
these tradeoffs.  from Fig. 3, both DAI and leaky integral control achieve the
desired asymptotic power sharing (2:1 ratio between G3, G5,
V. CASE STUDY: IEEE 39 NEW ENGLAND SYSTEM G6, G9, G10 and other generators) as predicted in Corollary 5.
However, leaky integral control solves the dispatch problem (22)
In this section, we perform a case study with the 39-bus New thereby underestimating the net load compared to DAI, which
England system (see Fig. 1), which is modeled as in (1)–(2) solves (4); see Corollary 6. We conclude that fully decentralized
with parameters Mi (for the 10 generator buses), Vi , and Bij leaky integral controller can achieve a performance similar to
taken from [39]. The inertia coefficients Mi are set to zero for the communication-based DAI controller—though at the cost of
the 29 (load) buses without generators. Note that Mi ’s in our steady-state offsets in both frequency and power adjustment.
simulations are heterogeneous, which relaxes our simplifying
assumption in Section IV-D that Mi ’s are homogeneous and
allows for testing the proposed scheme under a more realistic B. Comparison Between Controllers With Noise
setting. For every generator bus i, the damping coefficient Di Next, a noise term ηi (t) is added to the frequency measure-
is chosen as 20 per unit (p.u.) so that a 0.05 p.u. (3 Hz) change ments ω in (50b), (7b), and (13b) for DAI, pure integral, and
in frequency will cause a 1 p.u. (1000 MW) change in the leaky integral control, respectively. The noise ηi (t) is sampled
generator output power. For every load bus i, Di is chosen as from a uniform distribution on [0, η i ], with η i selected such
1/200 of that of a generator. Note that the generator turbine- that the ratios of η i between generators are 1 : 2 : 3 : · · · : 10
governor dynamics are ignored in the model (1)–(2) leading to and [η 1 , η 2 , . . . ] = η = 0.01 Hz. The meaning of η here is
a simulated frequency response that is faster than in practice, consistent with that in Definition 1 and Theorem 8. At each
but the fundamental dynamics of the system are retained for generator i, the noise has nonzero mean η i /2 (inducing a con-
a proof-of-concept illustration of the proposed controller. For stant measurement bias) and variance ση2 ,i = η 2i /12.

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
WEITENBERG et al.: ROBUST DECENTRALIZED SECONDARY FREQUENCY CONTROL IN POWER SYSTEMS: MERITS AND TRADEOFFS 3977

Fig. 2. Frequency at generator G1 under different control methods. (a) DAI control. (b) Decentralized pure integral control. (c) Leaky integral
control.

Fig. 3. Changes in active-power outputs of all the generators without noise. (a) DAI control. (b) Decentralized pure integral control. (c) Leaky
integral control.

Fig. 4. Changes in active-power outputs of all the generators, under a frequency measurement noise bounded by η = 0.01 Hz. (a) DAI control.
(b) Decentralized pure integral control. (c) Leaky integral control.

Fig. 2 (solid plots) shows the frequency at generator G1, of controller performance are calculated for the frequency at
and Fig. 4 shows the changes in active-power outputs of all generator G1:
the generators under such a measurement noise. Observe from 1) the steady-state frequency error without noise;
Fig. 2(b)–(c) and Fig. 4(b)–(c) that leaky integral control is 2) the convergence time without noise, which is defined as the
more robust to measurement noise than pure integral control. time when frequency error enters and stays within [0.95, 1.05]
Fig. 4(a) and (c) show that DAI control is even more robust than times its steady-state;
the leaky integral control in terms of generator power outputs, 3) the frequency root-mean-square-error (RMSE) from its
which is not surprising since the averaging process between nominal steady-state, calculated over 60–80 s (the average
neighboring DAI controllers can effectively mitigate the effect RMSE over 100 random realizations is taken).
of noise—thanks to communication. The RMSE results from measurement noise ηi (t) generated
every second at every generator i from a uniform distribution
on [−η i , η i ], where the meaning of η i is the same as that in
C. Impacts of Leaky Integral Control Parameters Section V-B; ηi (t) has zero mean so that the performance in
mitigating steady-state bias and noise-induced variance can be
Next, we investigate the impacts of inverse dc gains Ki and observed separately. Fig. 5 shows these metrics as functions of
time constants Ti on the performance of leaky integral control. k. It can be observed that the steady-state error increases with k,
First, we set the integral time constant Ti = τ = 0.05 s for as predicted by Lemma 3; convergence is faster as k increases,
every generator i, and tune the gains Ki = k for generators G3, in agreement with Theorem 7; and robustness to measurement
G5, G6, G9, and G10; Ki = 2k for other generators to ensure the noise is improved as k increases, as predicted by Theorem 8 and
same asymptotic power sharing as above. The following metrics Corollary 10.

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
3978 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 64, NO. 10, OCTOBER 2019

τ increases, which is in line with Theorem 8 and predicted by

Corollary 10.
Finally, we discuss performance degradation if the response
time of leaky integral controller is smaller than the actuation
response time. The generator turbine-governor dynamics can
be modeled as first- or second-order transfer functions, with
dominant time constants in the range of [0.25 s, 2.5 s] for hy-
draulic turbines and [4 s, 7 s] for steam turbines [40, Ch. 9]. The
analogous time constant for our controller corresponds to the
parameter ratio Ti /Ki . For the simulations in Figs. 2–4, this
ratio was chosen as 10 s for generators G3, G5, G6, G9, G10
and of 5 s for others. Thus, they are compatible with actuation
through steam and hydraulic turbines. If this was not the case,
the controllers have to be slowed down and their performance
can be inferred using Figs. 5 and 6. Finally, we stress that the
proven robustness guarantees, i.e., input-to-state-stability of the
nonlinear model, will not be at stake, provided that the initial
conditions and the maximum noise magnitude are those charac-
terized in the proof of Theorem 8.

D. Tuning Recommendations
Our results quantifying the effects of the gains K and T on
the system behavior lead to a number of insights about tuning
Fig. 5. Steady-state error (upper), convergence time (middle), and
the gains in a practical setting. Specifically, a possible approach
RMSE (lower) of frequency at generator G1, as functions of the gain is as follows. First, the ratios between the values Ki−1 can be
k for leaky integral control. The time constants are T i = τ = 0.05 s for all determined using Corollary 5 and knowledge about the genera-
generators. tor operation cost. Second, a lower bound on the sum of these

values ni=1 Ki−1 can be obtained from Corollary 4 according
to the required steady-state performance. Since by Theorem 7
larger gains Ki are beneficial to faster convergence, it is prefer-
able to set the values of Ki−1 equal to the lower bound from
Corollary 4. Note that in Corollary 4, the value of ε is normally
specified in the grid code and is thus assumed to be known.
 The
grid code also specifies a worst-case power imbalance ni=1 Pi ∗
that frequency controllers have to counteract before the system
is redispatched. Specifically, in our simulations, we assumed
an admissible frequency deviation  ε = 0.3 Hz = 0.005 p.u., a
worst-case power imbalance ni=1 Pi ∗ = 1800 MW = 18 p.u.
(approximately
 the simultaneous loss of the two largest gener-
ators), and ni=1 Di = 2100 p.u. based on practical generator
droop settings and  load damping values. As a result of Corol-
lary 4, we obtained ni=1 Ki−1 = 1500 p.u., which together with
Corollary 5 leads to our choice of Ki = 0.005 for generators
G3, G5, G6, G9, G10 and 0.01 for the others. Third, with the in-
verse gains Ki−1 fixed, the time constants Ti can be determined
to strike a desired tradeoff between frequency convergence rate
Fig. 6. Convergence time (upper) and RMSE (lower) of frequency at and noise rejection. We outline two possible approaches in the
generator G1, as functions of the time constant T i = τ for leaky integral
control. The gains K i are 0.005 for G3, G5, G6, G9, G10 and 0.01 for following based on Theorem 8 or simulation data.
other generators. A possible approach to determine Ti is foreshadowed by the
proof of Theorem 8. The maximum noise magnitude η̄ (for
which input-to-state stability can be established in Theorem 8)
Next, we tune the integral time constants Ti = τ for all gen- is linear in β1 /β2 , which are both defined as functions of T in
erators and fix k = 0.005, i.e., Ki = 0.005 for G3, G5, G6, G9, the proof of Lemma 14. From their definitions, one learns that
G10 and Ki = 0.01 for other generators, for a balance between η̄ is a convex function of each of the values of T . By requiring
steady-state and transient performance. Since the steady-state that the value of η̄ exceeds the sensor noise estimate, one can
is independent of τ , only the convergence time (measured then find bounds on the values of Ti . Within these bounds one
for the case without noise) and RMSE (taken as the average should select the lowest values of Ti , as this is both beneficial
of 100 runs with different realizations of noise) of frequency for a faster convergence rate α̂ and a smaller deviation due to
at generator G1 are shown in Fig. 6. It can be observed that the disturbance γ η̄ 2 , as seen in the proof of Theorem 8.
convergence is faster as τ decreases, which is in line with If the system under investigation makes the above consid-
Theorem 7. Robustness to measurement noise is improved as erations for T infeasible, an alternative tuning approach for T

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
WEITENBERG et al.: ROBUST DECENTRALIZED SECONDARY FREQUENCY CONTROL IN POWER SYSTEMS: MERITS AND TRADEOFFS 3979

relies on simulation data. For example, consider the simplified sufficiently large droop gain Ki−1 . The answers to this question
case presented in Fig. 6, where there is a single time constant can be found in practical advantages: 1) leaky integral control
τ = Ti for all the generators i to be tuned. By means of regres- obviously low-pass filters measurement noise; 2) has a finite
sion methods, one can approximate the relationships between bandwidth thus resulting in a less aggressive control action
the frequency convergence time Tconv , the frequency RMSE more suitable for slowly ramping generators; and 3) is not
fRM SE , and the gain τ via the functions susceptible to windup (indeed, a PI control action with anti-
windup reduces to a lag element [19]). 4) Other benefits that we
Tconv (τ ) = aτ + b
did not touch upon in our analysis are related to classical loop
fRM SE (τ ) = ce−α τ + d shaping, e.g., the frequency for the phase shift can be specified
for leaky integral control (13) to give a desired phase margin
where a > 0, b ∈ R, c > 0, d ∈ R, α > 0 are constants. The (and thus also practically relevant delay margin) where needed
time constant τ can then be chosen according to the criterion for robustness or overshoot.
min γ Tconv (τ ) + fRM SE (τ ) In summary, our lag-element-inspired leaky integral control
τ ≥0 is fully decentralized, stabilizing, and can be tuned to achieve
robust noise rejection, satisfactory steady-state regulation, and a
where γ > 0 is a tradeoff parameter selected according to the
desirable transient performance with exponential convergence.
relative importance of convergence time and noise robustness.
We showed that these objectives are not always aligned, and
The unique optimal solution to this tradeoff criterion is
tradeoffs have to be found. Our tuning recommendations are
  
1 αc summarized in Section V-D. From a practical perspective, we
τ ∗ = max log ,0 . recommend to tune the leaky integral controller toward robust
α γa
steady-state regulation and to address transient performance
with related lead-element-inspired controllers [38].
VI. SUMMARY AND DISCUSSION
We believe that the aforementioned extension of the leaky
In the following, we summarize our findings and the various integrator with lead compensators is a fruitful direction for fu-
tradeoffs that need to be taken into account for the tuning of the ture research. Another relevant direction is a rigorous analysis
proposed leaky integral controller (13). of decentralized integrators with dead-zones that are often used
From the discussion following the Laplace-domain represen- by practitioners (in power systems and beyond) as alternatives
tation (14), the gains Ki and Ti of the leaky integral controller to finite-dc-gain implementations, such as the leaky integrator.
(13) can be understood as interpolation parameters for which the Finally, all the presented results can and should be extended to
leaky integral controller reduces to a pure integrator (Ki  0) more detailed higher-order power system models.
with gain Ti , a proportional (droop) controller (Ti  0) with
gain Ki−1 , or no control action (Ki , Ti  ∞). Within these ex- APPENDIX
treme parameterizations, we found the following tradeoffs: The
steady-state analysis in Section IV-A showed that proportional A. Technical Lemmas
power sharing and banded frequency regulation is achieved for
any choice of gains Ki > 0: Their sum gives a desired steady- We recall several technical lemmas used in the main text.
state frequency performance (see Corollary 4), and their ratios Lemma 12 (Matrix cross terms): [12, Lemma 15] Given any
give rise to the desired proportional power sharing [see Corol- four matrices A, B, C, and D of appropriate dimensions, we
lary (5)]. However, a vanishingly small gain Ki is required for have
asymptotically exact frequency regulation (see Corollary 6), i.e.,    
A BTC A − BTB 0
the case of integral control. Otherwise, the net load is always un- M := ≥ =: M .
C TB D 0 D − C TC
derestimated. On the one hand, in regard to stability, we inferred
global stability for vanishing Ki  0 (see Theorem 2) but also Lemma 13 (Bounding the potential function): [12, Lemma
an absence of robustness to measurement errors as in (12). On 5] Consider the Bregman distance Vδ := U (δ) − U (δ̄) − ∇U
the other hand, for positive gains Ki > 0 we obtained nominal
(δ̄)T (δ − δ ∗ ). The following properties hold for all δ, δ̄ that sat-
local exponential stability (see Theorem 7) with exponential rate
as a function of Ki /Ti and robustness (in the form of exponen- isfy B T δ, B T δ̄ ∈ Θ.
tial ISS with restrictions) to bounded measurement errors (see 1) There exist positive scalars α1 and α2 such that
Theorem 8) with increasing (respectively, nondecreasing) ro-
α1 δ − δ ∗  ≤ ∇U (δ) − ∇U (δ ∗ ) ≤ α2 δ − δ ∗ .
bustness margins to measurement noise as Ki (or Ti ) becomes
larger. From a H2 -performance perspective, we could quali- 2) There exist positive scalars α3 and α4 such that
tatively (under homogeneous parameter assumptions) confirm
these results for the linearized system. In particular, we showed α3 δ − δ ∗ 2 ≤ Vδ ≤ α4 δ − δ ∗ 2 .
that measurement disturbances are increasingly suppressed for
larger gains Ki and Ti (see Corollary 10), but for sufficiently Lemma 14 (Positivity of V ): Suppose that Assumption 2
large power disturbances a particular choice of gains Ki to- holds true and B T δ ∈ Θ. The Lyapunov function V in (27)
gether with sufficiently small time constants Ti optimizes the satisfies
transient performance (see Corollary 11), i.e., the case of droop
control. β1 x2 ≤ V (x) ≤ β2 x2
Our findings, especially the last one, pose the question
whether the leaky integral controller (13) actually improves for some positive constants β1 and β2 , with x given in (25),
upon proportional (droop) control (the case Ti = 0) with provided that is sufficiently small.

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
3980 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 64, NO. 10, OCTOBER 2019

Proof: This proof follows the same line of arguments as the If only power disturbances are applied, i.e., when ση = 0 in
proof of [12, Lemma 8], but accounts for our slightly different (39) and (40), then f (k) reduces to
Lyapunov function. We will bound V (x) in (27) term-by-term.

n
α1 k
The quadratic terms in ω − ω ∗ and p − p∗ are easily bounded fζ (k) = nα6 − . (54)
in terms of the eigenvalues of matrices M and T , respectively. i=1
α3 k 2 + α4 k + α5 (λi )
The terms in δ and δ ∗ is addressed in the second statement of
Clearly, for all k > 0, Gleaky 2H2 = fζ (k) < fζ (0) =
Lemma 13. These three terms lead to the early bound
Gintegrator 2H2 . Therefore, since Gleaky 2H2 = f (k) = fζ (k) +
min(λm in (M ), λm in (T ), α3 )x2 ≤ V (x)| =0 fη (k), it follows for all k > 0 that Gleaky 2H2 = fη (k) + fζ
(k) < fη (0) + fζ (0) = Gintegrator 2H2 . 
≤ max(λm ax (M ), λm ax (T ), α4 )x2 .
2) Proof of Corollary 11:
The cross term (∇U (δ) − ∇U (δ ∗ ))T M ω can be written as Proof: First notice that for ση2 − σζ2 k/d > 0, the first term
 T    of (39) is always positive and thus Gleaky H2 > Gopen loop H2
∇U (δ) − ∇U (δ ∗ ) 0 2M ∇U (δ) − ∇U (δ ∗ ) for all τ . As a result, one can only improve the performance
.
ω 2M 0 ω beyond open loop when ση2 − σζ2 k/d < 0, which is equivalent
This allows us to apply Lemma 12, which yields the to (48). The derivative of (39) with respect to τ equals

−∇U (δ) − ∇U (δ ∗ )2 − λm ax (M )2 ω2 ∂ n

−(ση2 − kd σζ2 )2d(2τ λi + 1)
Gleaky 2H2 = .
∂τ (2d[mk 2 + ( md + dτ )k + τ + λi τ 2 ])2
≤ (∇U (δ) − ∇U (δ ∗ ))T M ω i=1

≤ ∇U (δ) − ∇U (δ ∗ )2 + λm ax (M )2 ω2 . Therefore, ∂∂τ Gleaky 2H2 > 0 whenever (48) holds true. It fol-
lows that the minimal norm is the limit when τ = 0.
By applying the first statement of Lemma 13, we can bound the We now compute the derivative of fζ (k) as
entire Lyapunov function using

n
α1 (α3 k 2 − α5 (λi ))
β1 = min(λm in (M ) − λm ax (M )2 , λm in (T ), α3 − α22 ) fζ (k) = . (55)
i=1
(α3 k 2 + α4 k + α5 (λi ))2
β2 = max(λm ax (M ) + λm ax (M )2 , λm ax (T ), α4 + α22 ).
Notice that τ = 0 implies α5 (λi ) = τ (1 + λi τ ) = 0 so that
Finally, we select sufficiently small so that β1 > 0. 
! 
n
α1 (α3 k 2 )
fζ (k)!τ =0 = .
B. Proof of Corollaries i=1
(α3 k 2 + α4 k)2
We provide here the proof of Corollaries 10 and 11. Thus, when considering fη and fζ for τ = 0, we get
1) Proof of Corollary 10: ! ! !
Proof: For a given value of τ , consider the function f (k)!τ =0 = fη (k)!τ =0 + fζ (k)!τ =0


n
−α1 k + α2 α1 α3 k 2 − 2α2 α3 k − α2 α4
f (k) = nα6 + (51) =n .
α3 k 2 + α4 k + α5 (λi ) (α3 k 2 + α4 k)2
i=1 !
By setting f (k)!τ =0 = 0, the optimal value of k is obtained as
where α1 = σζ2 /d, α2 = ση2 , α3 = 2dm, α4 = 2d (m/d + dτ ), the unique positive root of the second-order polynomial
α5 (λi ) = 2d(τ + λi τ 2 ), and α6 = σζ2 /2md are all positive pa-
rameters. The function f (k) interpolates between Gleaky 2H2 = p(k) = α1 α3 k 2 − 2α2 α3 k − α2 α4
f (k) and Gintegrator 2H2 = f (0).  
= 2m σζ2 k 2 − 2dση2 k − ση2
We prove that if either power disturbances σζ or measurement
noise ση equals zero, then Gleaky 2H2 < Gintegrator 2H2 holds which is explicitly given by (49). 
true for all k > 0. In presence of only measurement noise, i.e.,
when σζ = 0 the function f (k) reduces to ACKNOWLEDGMENT

n
α2 The authors would like to thank D. Groß for various helpful
fη (k) = (52)
α3 k 2 + α4 k + α5 (λi ) discussions that improved the presentation of this paper.
i=1

whose derivative with respect to k is REFERENCES


n
α2 (2α3 k + α4 ) [1] J. Machowski, J. W. Bialek, and J. R. Bumby, Power System Dynamics.
fη (k) = − . (53) 2nd ed., Hoboken, NJ, USA: Wiley, 2008.
i=1
(α3 k 2 + α4 k + α5 (λi ))2 [2] H. Bevrani, Robust Power System Frequency Control, vol. 85. New York,
NY, USA: Springer-Verlag, 2009.
Clearly, for all k > 0, fη (k) < 0. An analogous reasoning holds [3] D. K. Molzahn et al., “A survey of distributed optimization and control
true when analyzing Gleaky 2H2 as a function of τ , which shows algorithms for electric power systems,” IEEE Trans. Smart Grid, vol. 8,
no. 6, pp. 2941–2962, Nov. 2017.
the second claimed statement. Furthermore, fη (k) < 0 also im- [4] F. Dörfler and S. Grammatico, “Gather-and-broadcast frequency control
plies that Gleaky 2H2 = fη (k) < fη (0) = Gintegrator 2H2 in power systems,” Automatica, vol. 79, pp. 296–305, 2017.

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
WEITENBERG et al.: ROBUST DECENTRALIZED SECONDARY FREQUENCY CONTROL IN POWER SYSTEMS: MERITS AND TRADEOFFS 3981

[5] M. Andreasson, D. V. Dimarogonas, H. Sandberg, and K. H. Johansson, [29] A. R. Teel, “A nonlinear small gain theorem for the analysis of control
“Distributed control of networked dynamical systems: Static feedback, systems with saturation,” IEEE Trans. Autom. Control, vol. 41, no. 9,
integral action and consensus,” IEEE Trans. Autom. Control, vol. 59, pp. 1256–1270, Sep. 1996.
no. 7, pp. 1750–1764, Jul. 2014. [30] E. Tegling, B. Bamieh, and D. Gayme, “The price of synchrony: Evaluating
[6] Q. Shafiee, J. M. Guerrero, and J. Vasquez, “Distributed Secondary Con- the resistive losses in synchronizing power networks,” IEEE Trans. Control
trol for Islanded MicroGrids – A Novel Approach,” IEEE Trans. Power Netw. Syst., vol. 2, no. 3, pp. 254–266, Sep. 2015.
Electron., vol. 29, no. 2, pp. 1018–1031, 2014. [31] M. Andreasson, E. Tegling, H. Sandberg, and K. H. Johansson, “Per-
[7] C. Zhao, E. Mallada, and F. Dörfler, “Distributed frequency control for sta- formance and scalability of voltage controllers in multi-terminal HVdc
bility and economic dispatch in power networks,” in Proc. Amer. Control networks,” in Proc. Amer. Control Conf., May 2017, pp. 3029–3034.
Conf., Chicago, IL, USA, Jul. 2015, pp. 2359–2364. [32] B. K. Poolla, S. Bolognani, and F. Dorfler, “Optimal placement of virtual
[8] F. Dörfler, J. W. Simpson-Porco, and F. Bullo, “Breaking the hierarchy: inertia in power grids,” IEEE Trans. Autom. Control, vol. 62, no. 12,
Distributed control & economic optimality in microgrids,” IEEE Trans. pp. 6209–6220, Dec. 2017.
Control Netw. Syst., vol. 3, no. 3, pp. 241–253, Sep. 2016. [33] E. Tegling, M. Andreasson, J. W. Simpson-Porco, and H. Sandberg, “Im-
[9] C. De Persis, N. Monshizadeh, J. Schiffer, and F. Dörfler, “A Lyapunov proving performance of droop-controlled microgrids through distributed
approach to control of microgrids with a network-preserved differential- pi-control,” in Proc. Amer. Control Conf., Jul. 2016, pp. 2321–2327.
algebraic model,” in Proc. IEEE Conf. Decis. Control, Las Vegas, NV, [34] M. Andreasson, E. Tegling, H. Sandberg, and K. H. Johansson, “Coherence
USA, Dec. 2016, pp. 2595–2600. in synchronizing power networks with distributed integral control,” in
[10] S. Trip, M. Bürger, and C. De Persis, “An internal model approach to Proc. IEEE 56th Annu. Conf. Decis. Control, Dec. 2017, pp. 6327–6333.
(optimal) frequency regulation in power grids with time-varying voltages,” [35] F. Paganini and E. Mallada, “Global performance metrics for synchroniza-
Automatica, vol. 64, pp. 240–253, 2016. tion of heterogeneously rated power systems: The role of machine models
[11] M. Andreasson, D. V. Dimarogonas, H. Sandberg, and K. H. Johansson, and inertia,” in Proc. 55th Annu. Allerton Conf. Commun., Control, Com-
“Distributed PI-control with applications to power systems frequency put., Oct. 2017, pp. 324–331.
control,” in Proc. Amer. Control Conf., Portland, OR, USA, Jun. 2014, [36] A. Mešanović, U. Münz, and C. Heyde, “Comparison of H∞ , H2 , and pole
pp. 3183–3188. optimization for power system oscillation damping with remote renewable
[12] E. Weitenberg, C. De Persis, and N. Monshizadeh, “Exponential conver- generation,” IFAC-PapersOnLine, vol. 49, no. 27, pp. 103–108, 2016.
gence under distributed averaging integral frequency control,” Automatica, [37] M. Pirani, J. W. Simpson-Porco, and B. Fidan, “System-theoretic perfor-
vol. 98, pp. 103–113, 2018. mance metrics for low-inertia stability of power networks,” Mar. 2017,
[13] N. Li, C. Zhao, and L. Chen, “Connecting automatic generation control arXiv:1703.02646.
and economic dispatch from an optimization view,” IEEE Trans. Control [38] Y. Jiang, R. Pates, and E. Mallada, “Performance tradeoffs of dynamically
Netw. Syst., vol. 3, no. 3, pp. 254–264, Sep. 2016. controlled grid-connected inverters in low inertia power systems,” in Proc.
[14] X. Zhang and A. Papachristodoulou, “A real-time control framework for IEEE 56th Conf. Decis. Control, 2017, pp. 5098–5105.
smart power networks: Design methodology and stability,” Automatica, [39] K. W. Cheung, J. Chow, and G. Rogers, “Power system toolbox, v 3.0.,”
vol. 58, pp. 43–50, 2015. Rensselaer Polytechnic Inst. Cherry Tree Sci. Softw., vol. 48, p. 53, 2009.
[15] C. Zhao, E. Mallada, S. H. Low, and J. W. Bialek, “A unified framework [40] P. Kundur, Power System Stability and Control. New York, NY, USA:
for frequency control and congestion management,” in Proc. Power Syst. McGraw-Hill, 1994.
Comput. Conf., 2016, pp. 1–7.
[16] 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.
[17] N. Ainsworth and S. Grijalva, “Design and quasi-equilibrium analysis
of a distributed frequency-restoration controller for inverter-based micro-
grids,” in Proc. North Amer. Power Symp., Manhattan, KS, USA, Sep.
2013, pp. 1–6. Erik Weitenberg received the B.Sc. degree in
[18] J. Schiffer, R. Ortega, C. A. Hans, and J. Raisch, “Droop-controlled mathematics and the M.Sc. degree in math-
inverter-based microgrids are robust to clock drifts,” in Proc. Amer. Con- ematics with a specialization in algebra and
trol Conf., Chicago, IL, USA, Jul. 2015, pp. 2341–2346. cryptography, in 2010 and 2012, respectively,
[19] G. F. Franklin, J. D. Powell, and A. Emami-Naeini, Feedback Control of from the University of Groningen, Groningen,
Dynamic Systems, vol. 2. Reading, MA, USA: Addison-Wesley, 1994. The Netherlands, where he is currently work-
[20] R. Heidari, M. M. Seron, and J. H. Braslavsky, “Ultimate boundedness ing toward the Ph.D. degree in control of cyber-
and regions of attraction of frequency droop controlled microgrids with physical systems.
secondary control loops,” Automatica, vol. 81, pp. 416–428, 2017. His current research interests include stability
[21] Y. Han et al., “Analysis of washout filter-based power sharing strategy– and robustness of networked and cyberphysical
An equivalent secondary controller for islanded microgrid without LBC systems, with applications to power systems.
lines,” IEEE Trans. Smart Grid, vol. 9, no. 5, pp. 4061–4076. 2018.
[22] F. Bullo, Lectures on Network Systems, 1 Ed., with contributions
by J. Cortes, F. Dörfler, and S. Martinez, 2018. [Online]. Available:
http://motion.me.ucsb.edu/book-lns
[23] J. M. Guerrero, J. C. Vasquez, J. Matas, L. G. de Vicuna, and M. Castilla,
“Hierarchical control of droop-controlled AC and DC microgrids–A gen-
eral approach toward standardization,” IEEE Trans. Ind. Electron., vol. 58,
no. 1, pp. 158–172, Jan. 2011. Yan Jiang received the B.Eng. degree in elec-
[24] A. J. Wood and B. F. Wollenberg, Power Generation, Operation, and trical engineering and automation from Harbin
Control, 2nd ed. Hoboken, NJ, USA: Wiley, 1996. Institute of Technology, Harbin, China, in 2013,
[25] H. K. Khalil, Nonlinear Control. M. J. Horton, Ed. London, U.K.: Pearson, and the M.S. degree in electrical engineering
2014. from Huazhong University of Science and Tech-
[26] P. J. Campo and M. Morari, “Achievable closed-loop properties of systems nology, Wuhan, China, in 2016. She is currently
under decentralized control: Conditions involving the steady-state gain,” working toward the Ph.D. degree at the Depart-
IEEE Trans. Autom. Control, vol. 39, no. 5, pp. 932–943, May 1994. ment of Electrical and Computer Engineering
and the M.S.E. degree at the Department of Ap-
[27] K. J. Åström and T. Hägglund, Advanced PID Control. Pittsburgh, PA,
plied Mathematics and Statistics, Johns Hopkins
USA: ISA-The Instrumentation, Systems and Automation Society, 2006.
University, Baltimore, MD, USA.
[28] P. W. Sauer and M. A. Pai, Power System Dynamics and Stability.
Her research interests include the area of control of power systems.
Hoboken, NJ, USA: Wiley, 1998.

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.
3982 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 64, NO. 10, OCTOBER 2019

Changhong Zhao (S’12–M’15) received the Claudio De Persis received the Laurea degree
B.Eng. degree in automation from Tsinghua Uni- in electronic engineering in 1996 and the Ph.D.
versity, Beijing, China, in 2010, and the Ph.D. degree in system engineering in 2000, both from
degree in electrical engineering from California the University of Rome “La Sapienza,” Rome,
Institute of Technology, Pasadena, CA, USA, in Italy.
2016. He is a Professor with the Engineering and
He is a Researcher with the National Re- Technology Institute, Faculty of Science and En-
newable Energy Laboratory, Golden, CO, USA. gineering, University of Groningen, the Nether-
His research interests include distributed control lands. Previously, he held faculty positions with
of networked systems, power system dynamics the Department of Mechanical Automation and
and stability, and optimization of power and mul- Mechatronics, University of Twente, and the De-
tienergy systems. partment of Computer, Control, and Management Engineering, Univer-
Dr. Zhao was a recipient of the Caltech Demetriades-Tsafka-Kokkalis sity of Rome “La Sapienza”. He was a Research Associate with the De-
Ph.D. Thesis Prize, the Caltech Charles Wilts Ph.D. Thesis Prize, and partment of Systems Science and Mathematics, Washington University,
the 2015 Qualcomm Innovation Fellowship Finalist Award. St. Louis, MO, USA, in 2000–2001, and with the Department of Electrical
Engineering, Yale University, New Haven, CT, USA, in 2001–2002. His
main research interest is in control theory, and his recent research in-
terests include dynamical networks, cyberphysical systems, smart grids,
and resilient control.
Prof. Persis was an Editor for the International Journal of Robust and
Nonlinear Control (2006–2013), an Associate Editor for the IEEE TRANS-
ACTIONS ON CONTROL SYSTEMS TECHNOLOGY (2010–2015), and for the
IEEE TRANSACTIONS ON AUTOMATIC CONTROL (2012–2015). He is cur-
rently an Associate Editor for the Automatica (2013-present) and for the
IEEE CONTROL SYSTEMS LETTERS (2017-present).

Florian Dörfler received his Ph.D. degree in

Enrique Mallada (S’09–M’13) received the In- Mechanical Engineering from the University of
geniero en Telecomunicaciones degree from California at Santa Barbara in 2013, and a
Universidad ORT, Montevideo, Uruguay, in 2005, Diplom degree in Engineering Cybernetics from
and the Ph.D. degree in electrical and computer the University of Stuttgart in 2008.
engineering with a minor in applied mathemat- He is an Assistant Professor with the Auto-
ics from Cornell University, Ithaca, NY, USA, in matic Control Laboratory, ETH Zürich, Zurich,
2014. Switzerland. From 2013 to 2014, he was an As-
He is an Assistant Professor of electrical sistant Professor with the University of California
and computer engineering with Johns Hopkins Los Angeles. His primary research interests in-
University, Baltimore, MD, USA. Prior to join- clude distributed control, complex networks, and
ing Johns Hopkins University in 2016, he was cyberphysical systems currently with applications in energy systems.
a Postdoctoral Fellow with the Center for the Mathematics of Informa- Prof. Dörfler is a recipient of the 2009 Regents Special International
tion, Caltech from 2014 to 2016. His research interests include the areas Fellowship, the 2011 Peter J. Frenkel Foundation Fellowship, and the
of control, dynamical systems and optimization, with applications to en- 2015 UCSB ME Best Ph.D. Award. His students were recipients or fi-
gineering networks such as power systems and the Internet. nalists for Best Student Paper Awards at the 2013 European Control
Prof. Mallada was recipient of the CAREER Award from the National Conference, the 2016 American Control Conference, and the 2017 PES
Science Foundation (NSF) in 2018, the ECE Director’s Ph.D. Thesis Re- PowerTech Conference. His articles received the 2010 ACC Student
search Award for his dissertation in 2014, the Center for the Mathematics Best Paper Award, the 2011 O. Hugo Schuck Best Paper Award, the
of Information (CMI) Fellowship from Caltech in 2014, and the Cornell 2012–2014 Automatica Best Paper Award, and the 2016 IEEE Circuits
University Jacobs Fellowship in 2011. and Systems Guillemin-Cauer Best Paper Award.

Authorized licensed use limited to: Johns Hopkins University. Downloaded on July 10,2020 at 16:04:53 UTC from IEEE Xplore. Restrictions apply.