Automatica 61 (2015) 18–26

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Distributed control and optimization in DC microgrids✩

Jinxin Zhao a , Florian Dörfler b
a
Department of Mechanical and Aerospace Engineering, University of California, Los Angeles, CA, 90024, United States
b
Automatic Control Laboratory, Swiss Federal Institute of Technology (ETH) Zürich, 8092, Switzerland

article info abstract

Article history: Due to their compatibility with renewable and distributed generation, microgrids are a promising opera-
Received 9 October 2014 tional architecture for future power systems. Here we consider the operation of DC microgrids that arise
Received in revised form in many applications. We adopt a linear circuit model and propose a decentralized voltage droop control
26 March 2015
strategy that is inspired by frequency droop control in AC networks. We demonstrate that our primary
Accepted 29 June 2015
droop control strategy is able to achieve fair and stable load sharing (even in presence of actuation con-
straints) or an economic dispatch of the generation formulated as a quadratic and linearly-constrained
Keywords:
optimization problem on the source injections. Similar to frequency droop control, voltage droop control
DC microgrids induces a steady-state voltage drift depending on the imbalance of load and generation in the micro-
Load sharing grid. To compensate for this steady-state error, we consider two secondary control strategies. A purely
Distributed control decentralized secondary integral control strategy successfully compensates for the steady-state voltage
Optimal control drifts yet it fails to achieve the desired optimal steady-state injections. Next, we propose a consensus fil-
Stability analysis ter that requires communication among the controllers, that regulates the voltage drift, and that recovers
the desired optimal injections. The performance and robustness of our controllers are illustrated through
simulations.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction remote locations) while providing high quality power supply, e.g.,
in hospitals, research facilities, and school campuses. Third and fi-
Driven by environmental concerns, renewable energy sources nally, microgrids are naturally designed to integrate small-scale
are rapidly deployed, such as photovoltaic and wind generation. distributed generation, i.e., power is generated where it is needed
These sources will, for the most part, be deployed as small-scale without transmission losses.
generation units in low-voltage distribution networks. As a natu- Microgrids have been proposed based on either alternative cur-
ral consequence, the conventional centralized and hierarchical op- rent (AC) or direct current (DC) paradigms. AC power grids have
eration of power grids is advancing towards distributed and flat been in service for many decades, and their components and opera-
architectures, and so-called microgrids have been proposed as con- tion are well understood. The operational paradigms from conven-
ceptual solutions. Microgrids are low-voltage electrical distribu- tional AC power transmission networks have been inherited in AC
tion networks, composed of distributed generations, storages and microgrids (Guerrero, Vasquez, Matas, de Vicuna, & Castilla, 2011).
loads. The advantages of microgrids are as follows: first, micro- However, using DC microgrids has the following advantages: there
grids are capable of connecting to the power transmission grid, is an increasing number of DC sources and storages (e.g., solar cells
but they are also able to island themselves and operate indepen- and Li-ion batteries), end-user equipment (e.g., electric vehicles),
dently, e.g., in case of an outage. Second, microgrids can be de- and most of the contemporary electronic appliances. In Nilsson and
ployed as stand-alone small-footprint power systems (possibly in Sannino (2004) it is demonstrated that many daily loads supplied
by AC nowadays can operate also with a DC supply. In comparison
to AC microgrids with DC sources, the efficiency is raised since con-
version losses of DC-to-AC inverters are removed—though conver-
✩ A preliminary version of part of this paper’s results is Zhao and Dörfler (2015).
sion losses arise in DC-to-DC converters for sources with different
This work was supported by ETH Zürich startup funds. The material in this paper
voltage levels. Finally, DC microgrids are widely deployed in air-
was presented at the 2015 American Control Conference, July 1–3, 2015, Chicago,
IL, USA. This paper was recommended for publication in revised form by Associate crafts and spacecrafts (Justo, Mwasilu, Lee, & Jung, 2013). In sum-
Editor Jun-ichi Imura under the direction of Editor Toshiharu Sugie. mary, DC microgrids are a promising technology that has already
E-mail addresses: [email protected] (J. Zhao), [email protected] (F. Dörfler). attracted much research attention.
http://dx.doi.org/10.1016/j.automatica.2015.07.015
0005-1098/© 2015 Elsevier Ltd. All rights reserved.
 J. Zhao, F. Dörfler / Automatica 61 (2015) 18–26 19

Literature review: The articles (Guerrero et al., 2011; Ito, illustrate the performance and robustness of our primary and sec-
Zhongqing, & Akagi, 2004; Salomonsson & Sannino, 2007) fo- ondary controllers. Finally Section 7 concludes the paper.
cus on the hardware implementation of DC microgrids. A hierar- Aside from the importance of DC microgrids in their own right,
chical control layout for DC microgrids is proposed in Guerrero we sincerely believe that the considered DC scenario also serves
et al. (2011): a primary controller rapidly stabilizes the grid, and as valuable and accessible introduction to many power system
a secondary controller (on a slower time scale) corrects for the operational paradigms that have nonlinear and complex parallels
steady-state error induced by primary control. An experimental in AC networks. A preliminary version of part of this paper’s results
system involving solar-cell, wind turbine and power storage is is Zhao and Dörfler (2015).
designed and constructed in Ito et al. (2004). A low-voltage DC
distribution system for sensitive loads is described in Salomon- Preliminaries and notation
sson and Sannino (2007). In Shafiee, Dragicevic, Vasquez, and
Guerrero (2014), a modeling method of DC microgrid clusters is Vectors and matrices: Let 1n and On be the n-dimensional vectors
described. A scenario-based operation strategy for a DC micro- of unit and zero entries, respectively. Let In ∈ Rn×n be the
grid is developed in Xu and Chen (2011), emphasizing the de- n-dimensional identity. Let diag(v) represent a diagonal matrix
tailed model and control of wind turbine and battery. Feasibility with the elements of v on the diagonal. For a symmetric matrix
conditions for DC microgrids with constant power loads were pro- A = AT , the notation A > 0, A ≥ 0, A < 0, and A ≤ 0 means that
posed in Simpson-Porco, Dörfler, and Bullo (2015). A coopera- A is positive definite, positive semidefinite, and negative definite
tive control paradigm is proposed in Nasirian, Moayedi, Davoudi, and negative semidefinite, respectively.
and Lewis (2015) to establish a distributed primary/secondary Algebraic graph theory: Consider a connected, undirected, and
control framework for DC microgrids with communication capabil- weighted graph G = (V , E , W ), where V = {1, . . . , n} is the set of
ities. Distributed controllers have been studied to regulate multi- nodes, E ⊂ V × V is the set of undirected edges, and W = W T ∈
terminal DC transmission systems which share similar problem Rn×n is the adjacency matrix with entries wij > 0, if (i, j) ∈ E and
aspects with DC microgrids. The controller proposed in Andreas- wij = 0 otherwise. The degree
n matrix D ∈ R
n×n
is the diagonal
son, Dimarogonas, Sandberg, and Johansson (2014) achieves fair matrix with elements dii = j=1,j̸=i w ij . The Laplacian matrix L =
power sharing and asymptotically minimizes the cost of the power LT ∈ Rn×n is defined by L = D − W , and it satisfies L ≥ 0 and L1n =
injections. Distributed controllers focusing on voltage control are 0n . If the graph is connected, then the null space of L is spanned by
studied in Tucci, Riverso, Vasquez, Guerrero, and Ferrari-Trecate 1n , and all the other n − 1 eigenvalues of L are strictly positive.
(2015) and Morstyn, Hredzak, Demetriades, and Agelidis (2015). In
Zonetti, Ortega, and Benchaib (2014) a unified port-Hamiltonian 2. DC microgrid model
system model is proposed, and the performance of decentralized
PI control is discussed for a multi-terminal DC transmission sys- For our purposes, a microgrid is a linear connected circuit with
tem. For AC microgrids a flat and distributed operation architecture associated undirected graph G(V , E , W ), nodes V = {1, . . . , n},
has been proposed in Dörfler, Simpson-Porco, and Bullo (2014) and and edges E ⊂ V × V . We assume that all lines in the DC microgrid
Simpson-Porco, Dörfler, and Bullo (2013), consisting of simultane- are purely resistive, and refer to Remark 3.1 for an extension of our
ous (without time-scale separation) primary, secondary, and ter- results of more general line impedances. The adjacency matrix W
tiary controls. Inspired by these AC operation strategies we seek is defined with nonzero entries wij = wji = 1/Rij for (i, j) ∈ E ,
similar solutions for DC microgrids. i, j ∈ V . The
where Rij is the resistance of the line connecting nodes 
j=1,j̸=i wij .
n
Contribution and contents: In this article, we propose a compre- diagonal degree matrix D ∈ Rn×n has elements dii =
hensive operational control strategy for DC microgrids in order to The admittance matrix Y is defined as Y = D − W . Thus, Y = Y T ∈
achieve multiple objectives. Rn×n is a real-valued Laplacian matrix satisfying 1Tn Y = OTn .
In Section 2, we introduce the considered DC microgrid model. We partition the set of nodes into m sources and n − m loads:
Inspired by the shortcomings of conventional DC droop control and V = VS ∪ VL . Throughout this paper we denote sources and loads
the merits of frequency droop control in AC systems, we propose by the superscripts S and L, respectively. The sources are assumed
a novel primary voltage droop control strategy in Section 3. Our to be controllable current sources with positive current injections
proposed primary control strategy is fully decentralized, and we IiS ≥ 0 and are assembled in the vector I S . Each source is con-
demonstrate that it is capable of stabilizing the grid while achiev- strained by its output current capacity Īi , i.e., IiS ∈ [0, Īi ]. The loads
ing load sharing and avoiding actuator saturation. As base scenario are assumed to be constant-current loads with negative current in-
we consider a purely resistive network with constant current loads, jections IiL ≤ 0 and are assembled in the vector I L . Following Kirch-
but we also discuss extensions to other load and network mod-
hoff’s and Ohm’s laws, the network model is built as1
els. In Section 4, we consider the economic dispatch of multiple  S
VS
  
generating units and formulate it as a convex optimization prob- I YSS YSL
lem. We demonstrate that the set of minimizers of the economic = T (1)
IL YSL YLL VL
dispatch are in one-to-one correspondence with the steady states
achieved by our primary voltage droop control with appropriately where the admittance matrix Y is partitioned according to sources
chosen control gains. As a result, we propose a selection of control and loads, and V S and V L represent the nodal voltages (potentials)
gains (droop coefficients) to achieve economic optimality in a de- of sources and loads, respectively. Since Y is a Laplacian matrix,
centralized way and without a model of the network or the loads. In 1Tn Y = OTn and a necessary feasibility condition for Eq. (1) is
Section 5, we discuss the limitations of droop control causing
steady-state voltage drifts, and we study secondary control strate-
gies to compensate for it. First, we consider fully decentralized 1 Loads in DC power systems are conventionally modeled as constant-current,
integral controllers and illustrate their limitations. Next, we pro- constant-impedance, constant-voltage or constant-power loads (Nilsson & Sannino,
pose a distributed consensus filter that relies on communication 2004). Often loads do not belong to a single category but display a combination
between local controllers. We show that this distributed control of the above properties. We mainly focus on pure constant-current loads which
arise primarily in electronic loads and also in some conventional loads such as LED
strategy is capable of regulating the voltage drifts while simulta- lighting. We find that these loads are the mathematically most challenging linear
neously achieving tertiary-level objectives such as load sharing or loads. In Remarks 3.2 and 3.3 we show how all our results extend to constant-
economic dispatch. In Section 6, we present simulation results to impedance loads and constant-voltage buses.
20 J. Zhao, F. Dörfler / Automatica 61 (2015) 18–26

Fig. 1. A DC microgrid with three sources and five loads. The blue dashed lines
indicate the communication among the secondary controllers (29) that we design
Fig. 2. Realization of droop control (7) as analog circuit.
in Section 5.

where Ii∗ ∈ [0, I¯i ] is an injection setpoint and the gain Ci > 0
1Tm I S + 1Tn−m I L = 0. (2)
is referred to as droop coefficient. Unless non-local (distributed
Fig. 1 shows an example network of a DC microgrid. or decentralized) secondary controllers or carefully tuned vir-
In this DC microgrid setup, we assume that source buses are tual impedance controllers are added, the control (5) does gen-
controllable voltage sources and load buses are passive current erally not achieve load sharing (especially for non-negligible line
sinks.2 The control objectives are (i) balancing of generation and impedances); see Nasirian et al. (2015) for a review. From a math-
load (as in (2)) (ii) in a stable fashion and (iii) subject to fair re- ematical perspective this shortcoming is essentially due to the ab-
source allocation (e.g., a fair load sharing), (iv) subject to possi- sence of a global variable such as the AC frequency.
ble actuation constraints (e.g., within source capacity limits), and Here we start from the observation that the conventional con-
(v) subject to load voltages within pre-described bounds. In this ar- troller (5) can be interpreted as the steady-state of the following
ticle, we show that the control objectives (i)–(iv) can be achieved proportional–integral droop controller:
in a plug-and-play fashion, that is, without knowledge of the sys- IiS = Ii∗ − Ci V̇iS − pi , (6a)
tem model and data and in a distributed way without centralized
coordination. We return to the control objective (v) in Remark 5.1. ṗi = Ci V̇iS . (6b)
Observe that (6a) mimics the AC frequency droop (3) and (6b) is an
3. Primary droop control and load sharing integral controller compensating for steady-state drifts similar to a
decentralized secondary frequency integral controller often added
We briefly review frequency droop control in AC microgrids to droop in AC systems. Inspired by this observation, the success
(Guerrero et al., 2011) to motivate our proposed control strategy of frequency droop control (3) in AC systems, and the limitation of
for DC microgrids. In AC microgrids the active power injection Pi at conventional DC droop control (5), we propose the primary voltage
source i is controlled to be proportional to its frequency deviation droop controller
θ̇i (from a nominal frequency) as
IiS = Ii∗ − Ci V̇iS . (7)
Pi = Pi − Ci θ̇i ,
∗
(3) Fig. 2 shows an analog circuit realization of the droop controller
(7) via a constant current source Ii∗ and a shunt capacitor Ci
where the control gain Ci > 0 is referred to as the droop coefficient,
reminiscent of shunt compensation in DC power systems (Karlsson
Pi∗ ∈ [0, P̄i ] is a nominal injection setpoint, and P̄i is the capacity
& Svensson, 2003). The proposed primary droop control (7) is a
of source i. For a particular selection of droop coefficients, it can
fully decentralized local proportional control strategy. In a digital
be shown that frequency droop control stabilizes the AC microgrid
implementation, each generating unit is controlled as a voltage
to a synchronous solution and achieves proportional load sharing source with terminal voltage ViS , and a micro-controller realizes
at steady state (Simpson-Porco et al., 2013), that is, every source i
the droop (7) based on measurement of the output current IiS .
injects active power Pi according to its capacity P̄i : Pi /P̄i = Pj /P̄j for
Similar to AC droop control (3), our controller (7) induces a
all sources i, j ∈ VS . A key feature of AC frequency droop control is
global variable, namely a constant voltage drift, that depends on
that it synthesizes the synchronous frequency as a global variable the load/generation imbalance: i∈VS Ii∗ + j∈VL IjL . Of course, this
indicating the load/generation imbalance in the microgrid (Dörfler
drift has to be compensated by a secondary controller, which will
et al., 2014; Simpson-Porco et al., 2013).
be done in Section 5. Before that we analyze the primary droop
As for AC systems, a primary objective in DC microgrids is to
control loop (1) and (7) by itself and show, among others, that it
synthesize local decentralized droop controllers that achieve pro- achieves stable proportional load sharing:
portional load sharing in the sense that
Theorem 3.1 (Primary Control and Load Sharing). Consider the
IiS /I¯i = IjS /I¯j for all i, j ∈ VS , (4) closed-loop droop-controlled microgrid (1) and (7). Then the following
statements hold:
where IiS ∈ [0, I¯i ] is the current injection of source i ∈ VS and
(1) Voltage drifts: all voltages ViS , i ∈ VS converge exponentially to
I¯i > 0 is its capacity. The conventional DC voltage-vs-current droop
V (t ) = V ∗ + v̇drift t · 1m , where V ∗ ∈ Rm is a constant vector and
controller is given by (see Nasirian et al., 2015, Shafiee et al., 2014,
the common voltage drift is
Zonetti et al., 2014)
Ij∗ + IjL
 
IiS ∗
= Ii − Ci ViS , (5) j∈VS j∈VL
v̇drift =  . (8)
Cj
j∈VS

2 This setup includes the case when a current load is attached to a source bus (2) Proportional load sharing: if the droop coefficients and nomi-
whose terminal voltage is controllable. nal injection setpoints are selected proportionally, that is, for all
 J. Zhao, F. Dörfler / Automatica 61 (2015) 18–26 21

i, j ∈ V S

Ci /I¯i = Cj /I¯j and Ii∗ /I¯i = Ij∗ /I¯j , (9)

then at steady state the load is shared proportionally.
Proof. The closed-loop state space model (1), (7) is
 ∗
VS
    
C V̇ S IS Y YSL Fig. 3. Π -model with resistive series and capacitive shunt impedances.
= − SST , (10)
0 IL YSL YLL VL
For proportional coefficients (9), we have Ci /Ii∗ = Cj /Ij∗ and the
where C = diag(C1 , . . . , Cm ). Since YLL is invertible (Dörfler & previous inequality equivalently reads as
(I −
−1 L
Bullo, 2013, Lemma II.1), we eliminate the variable V L = YLL
 Ii∗

YSL V ) from the second block of (algebraic) equations in (10).
T S 
IkL ≤− ∗
Ij − Cj = 0.
This elimination process, termed Kron-reduction in circuit theory k∈VL j∈VS
Ci
(Dörfler & Bullo, 2013), gives the Kron-reduced system
A similar calculation, for IiS ≤ Īi , for i ∈ VS , yields
V̇ S = −C −1 L̃V S + C −1 Ĩ (11)
Ij∗ + IkL
 
∗ −1 L −1 T
where Ĩ = IS − YSL YLL I and L̃ = YSS − YSL YLL YSL is again j∈VS k∈VL
IiS = Ii∗ − Ci ≤ Īi .
a positive semidefinite Laplacian (Dörfler & Bullo, 2013, Lemma

Cj
II.1) with a unique zero eigenvalue. By Sylvester’s Law of Inertia j∈VS
(Ostrowski & Schneider, 1962, Corollary 3), because C −1 > 0
The coefficients satisfy Ci /( j∈Vj Cj ) = Īi /( j∈Vj Īj )
 
and L̃ is symmetric, C −1 L̃ has the same number of negative, zero
=
Ii∗ /( j∈Vj Ij∗ ), thus the previous inequality also reads as

and positive eigenvalues as L̃. Thus, C −1 L̃ has one zero-eigenvalue
and all other eigenvalues are positive. It follows that all modes of 
Cj
the Kron-reduced system (11) exponentially decay to zero with  j∈Vj  
exception of the zero mode with right eigenvector 1m and left Ik ≥ (Ii − Īi )
L ∗
− Ij∗ = − I¯j .
eigenvector C 1m . This zero mode is integrated and all components k∈VL
Ci j∈Vj j∈VS
of the vector V̇S will exponentially converge to the same value v̇drift .
We project the differential–algebraic equations (10) onto the zero These inequalities complete the proof of Theorem 3.2. 
mode by summing all Eqs. (10) as 1Tm C V̇S = 1Tm I S +1Tn−m I L . In steady We conclude that the primary droop controller (7) achieves
state for V̇S = v̇drift 1m , we recover the voltage drift (8). Since V̇S (t ) stable proportional load sharing in a fully decentralized way and
converges exponentially to the constant v̇drift 1m , we have that V (t ) while respecting actuation constraints. However, as in AC systems,
converges exponentially to V ∗ + v̇drift t · 1m , where V ∗ ∈ Rm . This the droop controller (7) induces a steady-state voltage drift (8)
proves statement (1). which is proportional to the total injection imbalance i∈VS Ii∗ +
At steady state, the closed-loop injections are L

j∈VL Ij . Notice that the total injection imbalance is zero (and
hence v̇ drift = 0) only if a precise forecast of the total load
IiS = −Ci v̇drift + Ii∗ . (12) L ∗

j∈VL Ij is known and the nominal injections Ii can be scheduled
accordingly. Such a precise forecast is generally not available, the
Thus, /I¯i = (−Ci v̇drift + Ii )/Īi . The proportional load sharing
IiS ∗
nominal injections are fixed (typically to 0 or Īi ), and the loads are
objective IiS /I¯i = IjS /I¯j (for all i, j ∈ VS ) can be achieved by choosing changing with time. Another way to reduce the voltage drift v̇drift
Ii∗ and Ci proportionally as in (9). This proves statement (2) of in (8) is to choose large droop coefficients Ci . On the other hand,
Theorem 3.1.  the latter choice results in slow response of the system.
Theorem 3.1 gives a criterion for stable load sharing of the We will explicitly address the regulation of the voltage drift in
closed-loop system (1), (7), namely the droop coefficients need Section 5. Before that we turn to the tertiary control (or energy
to be picked proportional to capacity Ci = γ Īi , where γ > 0 is management) problem (in Section 4) as well as extensions to other
load and line models.
constant. Observe that Theorem 3.1 does not guarantee that the
injections satisfy the actuation constraint Ii ∈ [0, Īi ]. If the control Remark 3.1 (Extension to Π -Model). Consider a microgrid with
gains are chosen as in (9),then the actuation constraint is met if resistive–capacitive lines (e.g., underground cables) described by
and only if the total load j∈VL IjL can be satisfied by the maximal the Π -model (Kundur, 1994) illustrated in Fig. 3. In this case, the

injections (at capacity) i∈VS Īi . DC microgrid model is given by
 S
VS V̇ S
    
I YSS YSL
Theorem 3.2 (Actuation Constraints). Consider a stationary solution = T L + Cd (13)
of the closed-loop system (1) and (7) with droop coefficients and IL YSL YLL V V̇ L
setpoints chosen proportionally as in (9). The following statements are where Cd is a diagonal matrix with the shunt capacities as diagonal
equivalent: elements, i.e., Cd = diag(Cd1 , . . . , Cdn ).
(1) Injection constraints: 0 ≤ IiS ≤ Īi for all i ∈ VS ; The primary droop controller analogously regulates the micro-
grid and results in voltage drifts. This can be seen by substituting
(2) Load satisfiability: i∈VS Īi ≥ − j∈VL IjL ≥ 0.
 
the droop controller I S = I ∗ − C V̇ S into (13), then the closed-loop
Proof. The steady-state injections are given by (8), (12). The state space equation is
condition IiS ≥ 0, for i ∈ VS , translates to  ∗
V̇ S VS
       
C 0 I Y YSL
Cd + = L − SST . (14)
Ij∗ + IkL
 
0 0 V̇ L I YSL YLL VL
j∈VS k∈VL
IiS = Ii∗ − Ci  ≥ 0. Since the model (14) has the same structure as the Kron-reduced
Cj
j∈VS system (11), analogous arguments as in the proof of Theorem 3.1
22 J. Zhao, F. Dörfler / Automatica 61 (2015) 18–26

Since YSS > 0, the system (17a) is Hurwitz, and V S converges to the
vector V S (∞) = YSS (I − YSL V L ). Thus, droop control stabilizes a
−1 ∗

microgrid with constant-voltage buses. We will not further pursue

this model. 

4. Optimal economic dispatch and droop control

The proportional choice of droop coefficients (9) leads to fair

load sharing (4) among the sources proportional to their capacity.
However, this objective may not be desirable when the sources rely
on different energy generation and conversion mechanisms. For
example, solar cells have lower capacities compared with diesel
generators, but they may be preferred due to economic and en-
vironmental reasons. In the following, we consider an alternative
Fig. 4. Microgrid with impedance loads.
generation dispatch criterion, namely the economic dispatch for-
malized as the optimization problem
show that all voltage drifts exponentially converge to the common
constant value m
 1

Ij∗ +

IjL minimize f (u) = αi u2i (18a)
{u,V S ,V L } i =1
2
j∈VS j∈VL
v̇drift =  .
IS∗ + u VS
    
Cj + Cd,j + Cd,j
subject to L =Y . (18b)
j∈VS j∈VL I VL
All subsequent developments are analogous for the more detailed The optimization problem (18) is convex with quadratic objective
model (14) and we focus on the model (1), (7). 
and linear constraints. The coefficients αi > 0 are chosen based on
the marginal cost of power source i reflecting its fuel and opera-
Remark 3.2 (Extension to Constant-Impedance Loads). Consider a
tion costs, capacity, or other preferences. In case that the nominal
DC microgrid with additional constant resistive loads (arising, e.g.,
injection setpoints Ii∗ are zero, the decision variable ui equals the
in electric lighting and heating devices Nilsson & Sannino, 2004),
total generation of source i. For nonzero setpoints Ii∗ > 0, ui is the
as shown in Fig. 4. We assume that there is at least one resistive
reserve generation to meet the real-time demand.
load, we let Rj > 0 ∈ R be the impedance of load j, and we define
the associated shunt admittance as Yshunt,j = 1/Rj (which is zero if
Theorem 4.1 (Economic Dispatch). Consider the optimization prob-
Rj = 0). The resulting current balance equations (1) are
lem (18). The optimal injections are
 S
VS
  
I YSS YSL
= T , (15) u∗i = −c /αi , i ∈ Vs , (19)
IL YSL YLL + Yshunt VL
1Tm IS∗ +1Tn−m I L
where Yshunt = diag([1/R1 , . . . , 1/Rn−m ]). After implementing where c = 
i 1/αi
is a constant.
voltage droop control (7) at the source nodes and applying Kron
Proof. The Lagrangian associated to (18) is
reduction, the closed-loop system is
m
IS∗ + u
  
C V̇ S = −[YSS − YSL (YLL + Yshunt )−1 YSL
T
]V S 1

L(u, V , λ) = αi u2i + λT − YV
2 IL
+ [I ∗ − YSL (YLL + ZLL )−1 I L ]. (16) i=1
 S  S
V λ
Since the admittance matrix in (15) is positive semidefinite and where V = L and λ = ∈ Rn .
V λL
irreducibly diagonal dominant, we conclude that it is nonsingular
(Horn & Johnson, 1985, Corollary 6.2.27) and thus also positive The KKT conditions ∂∂L
V
= 0, ∂∂λL = 0 and ∂∂Lu = 0 are necessary
definite. Since the class of positive definite matrices is closed under and sufficient for optimality due to the convexity of (18) (Boyd &
the Schur complement (Zhang, 2006, Chapter 4), it follows that the Vandenberghe, 2004). The first condition is ∂∂L V
= −λT Y = 0.
Schur complement (YSS − YSL (YLL + ZLL )−1 YSL T
) in (16) is positive Since Y is a Laplacian matrix, null(Y ) = span(1n ). Thus, we have
definite. Hence, the system (16) is Hurwitz, and the states converge that λ = c 1n , where c ∈ R is a constant. The second condition
to the constant steady state [YSS − YSL (YLL + ZLL )−1 YSL T −1 ∗
] [I − is ∂∂L
u
= uT diag(αi ) + λTS = 0. It follows that u∗i = −c /αi . The
YSL (YLL + ZLL ) I ] without voltage drift.
−1 L constraint (18b) implies
In conclusion, a DC microgrid with constant resistive loads is
IS∗ + u
 
stabilized by fully decentralized voltage droop control (7). Essen- 1Tn = 1Tm IS∗ + 1Tn−m I L + 1Tm u = 1Tn YV = 0.
IL
tially, the loads absorb any injections from the sources. We will not
further pursue this model.  Since u∗i = −c /αi , then c

1/αi = (1Tm IS∗ + 1Tm I L ). 
Remark 3.3 (Extension to Constant-Voltage Buses). Consider a DC Theorem 4.1 gives the optimal injections ui as a function of the
microgrid with constant-voltage buses arising, e.g., in controllable nominal injections IS∗ , the (possibly unknown) loads I L , and the cost
electronic loads (chips), points-of-common coupling (PCCs), and coefficients αi . Observe from (19) that at optimality all marginal
substations. For notational simplicity, assume that all load buses costs are identical:
have constant voltages V L in (1) (our reasoning easily extends to a
more general case). After implementing voltage droop control (7), αi u∗i = αj u∗j i ∈ Vs . (20)
we obtain Note the similarity between the optimal injections (19) and the
steady-state injections of droop-controlled microgrid (8) and (12).
C V̇ = −YSS V + (I − YSL V )
S S ∗ L
(17a)
Based on this observation, we present the following result: the
L
I = T S
YSL V + YLL V . L
(17b) optimal solution of the economic dispatch (19) can be achieved
 J. Zhao, F. Dörfler / Automatica 61 (2015) 18–26 23

by appropriately designed droop controllers (7). Conversely, any

steady state of the droop-controlled microgrid (1) and (7) is the
optimal solution of the economic dispatch (19) with appropriately
chosen parameters.

Corollary 4.1 (Droop Control and Economic Dispatch). Consider the

following two injections:
(1) The optimal injection Ii∗ + u∗i of the economic dispatch problem
(18) with cost coefficients αi ; and
(2) The steady-state injections Ii∗ − Ci v̇drift of the droop-controlled
microgrid (1) and (7) with droop coefficients Ci .
These two injections are identical if and only if
Fig. 5. Analog circuit realization of the integral controller (24) in the form (25) with
αi Ci = αj Cj for all i, j ∈ VS . (21) zero initial conditions.

5.1. Decentralized integral control

Proof. In the economic dispatch optimization problem (18), the
unique optimal injection is
To compensate the steady-state drift (8), we augment every
1 1Tm IS∗ + 1Tn−m I L droop controller (7) with a local integral controller penalizing volt-
Ii∗ + ui = Ii∗ − . (22) age drifts. The resulting PI controller is
αi 1

αj
j∈VS
IiS = Ii∗ − Ci V̇iS − pi (24a)
The stationary injection induced by droop-control is
Di ṗi = V̇iS (24b)
∗
IkL
 
Ij + where pi is an integral control variable, and Di > 0 is a gain.
j∈VS k∈VL
Ii∗ − Ci v̇drift = Ii∗ − Ci  . (23) Notice that (24) is a generalization of the PI droop controller (6),
Cj which reduces to the conventional DC droop control (5) in steady
j∈VS
state. The decentralized integral controller (24) (and equivalently
First, observe that (22) and (23) are equal when substituting Ci = conventional DC droop controller (5)) successfully corrects for the
β/αi into (23), where β ∈ R is constant. Conversely, if (22) and steady-state voltage drifts but fails to recover the desired injections
(23) are equal for all i ∈ VS , i.e., for load sharing and economic optimality. Fig. 5 shows an analog
circuit realization of the decentralized integral controller (24). In a
Ij∗ + IkL
 
digital implementation, a micro-controller can be programmed to
1 1Tm IS∗ + 1Tn−m I L j∈VS k∈VL
∗
Ii − ∗
= Ii − C i , realize (24) based on source-current measurements.
αi 1
 
αj
Cj
j∈VS j∈VS Theorem 5.1 (Performance of Decentralized Integral Control). Con-
sider the closed-loop secondary-controlled microgrid (1) with the de-
i∈VS Ci / i∈VS αi = αi Ci . Since
1
 
it follows for every i ∈ VS that
centralized integral controller (24). Then the following statements
the left-hand side of this equality is constant for every i ∈ VS , we hold:
have αi Ci = αj Cj for all i, j ∈ VS . 
(1) All source voltages ViS (t ) converge to stationary values without
Notice that the optimal injection (in the sense of the economic drift.
dispatch (18)) can be achieved in a fully decentralized manner, (2) The steady-state source injections do generally not achieve
without any communication among the sources or a knowledge proportional load sharing (4).
of the microgrid and the loads when the droop gains are chosen (3) The steady-state source injections are generally not optimal with
as Ci = β/αi for some constant β > 0. In general, the optimal respect to the economic dispatch (18).
droop gain Di = β/αi and the droop gain Di = γ Īi for proportional
load sharing (satisfying the conditions (9)) are not identical unless Proof. The secondary controller (24) is identical to,
αi = c /Īi for some constant c ∈ R. We conclude that the
i (Vi − Vi (0)) + pi (0),
IiS = −Ci V̇iS + Ii∗ − D− 1 S S
(25)
economic dispatch (18) is a more versatile objective that includes
load sharing (4) as a special case. where ViS (0) and pi (0) are initial values of ViS and pi , respectively.
Finally, we remark that the proposed voltage droop controllers The closed-loop state space model is then obtained by combining
are designed to be optimal with respect to a (terminal) steady-state Eqs. (1) and (25)
cost, but they are decentralized and independent of the system
VS
   ∗     
model and the load profile. Provided that detailed and accurate C V̇ S ÎS Y YSL D −1 V S
model data is available, controllers that are also transiently optimal = − SST − (26)
0 IL YSL YLL VL 0
can be designed, but they will likely not be fully decentralized.
In short, such optimal controllers violate the ‘‘plug-and-play’’ where ÎS∗ = IS∗ + D−1 (V S (0) − p(0)).
philosophy in microgrid operation (Dörfler et al., 2014). The associated Kron-reduced system is given by

5. Secondary integral control C V̇ S = −(L̃ + D−1 )V S + Î , (27)

where D = diag(D1 , . . . , Dm ) and Î = ÎS − YSL YLL I .

∗ −1 L
The primary droop control (7) results in a generally non-zero
stationary voltage drift given in (8) which has to be compensated Since L̃ ≥ 0 and D−1 > 0, we have that L̃ + D−1 > 0. Then
by means of a secondary controller. In what follows, we investigate according to Sylvester’s Law of Inertia (Ostrowski & Schneider,
two secondary control strategies: a fully decentralized one and a 1962), −C −1 (L̃ + D−1 ) has strictly negative eigenvalues. Thus, with
distributed one. a constant vector C −1 Î, the voltage sources V S converge to the
24 J. Zhao, F. Dörfler / Automatica 61 (2015) 18–26

asymptotically stable equilibrium V S (∞) = (L̃ + D−1 )−1 Î. This The proof of Theorem 5.2 is presented in the Appendix. We
concludes the proof of statement (1). conclude our analysis with a remark on the load voltages that were
The closed loop (1) and (24) is (after Kron reduction) so far left out of the picture.
    
C V̇ S I I −L̃V S + Ĩ . Remark 5.1 (Control of the Load Voltages). In our analysis we fo-
= (28)
CDṗ I I −p cused on controlling the sources, and the load voltages V L are de-
  termined as function of the source voltages (and thus the current
I I
The matrix I I
in (28) has nullspace [−x x]T for any x ∈ Rm . injection setpoints Ii∗ and corrections through primary/secondary
Hence, the equilibria of (28) are for any τ ∈ R control) by inverting the Kron reduction: V L = YLL (I − YSLT V S ).
−1 L

    Clearly, if the loads are attached to controllable voltage (source)

−L̃V S + Ĩ = τ −x . buses, then their voltages can be directly controlled. In the more
−p x general under-actuated case, we show in the following how a de-
sirable load voltage profile can be guaranteed by scheduling the
These equilibrium subspaces include (but are not equal to) the
nominal source injections Ii∗ accordingly. Consider the network
open-loop equilibria which are achieved for x = Om . Observe that
load sharing (4) and economic optimality (18) can be achieved only balance equations (1) evaluated in steady state for I S = I ∗ − p∗ .
Recall from Dörfler and Bullo (2013) that any principal minor of
for particular steady states satisfying xi /Ci = xj /Cj or xi αi = xj αj for
the Laplacian Y is nonsingular (provided that the microgrid is con-
all entries i, j of x, see (19). However, these particular steady states
nected). Thus, we can eliminate the source voltage V S and obtain
are strict subsets of the equilibrium subspaces and can generally
(for generic initial conditions) not be reached. This concludes the
I L + Ȳ (I ∗ − p∗ ) = Ỹ V L , (30)
proofs of statement (2) and (3). 
−1 T
where Ỹ = YLL − YSL YSS YSL is the reduced admittance matrix and
We conclude that the fully decentralized secondary integral T −1
controller (24) eliminates the voltage drift v̇drift , but it also Ȳ = −YSL YSS . Eq. (30) relates the load bus voltages V L and the
obliterates the desired properties of proportional load sharing and source injections I ∗ , and it can be used to schedule the source injec-
economic optimality. tions I ∗ to achieve a desirable nominal load voltage profile V L , e.g.,
by solving a feasibility problem subject to upper and lower bounds
on the load voltages. To do so the network admittance matrix Y
5.2. Distributed consensus filter needs to be known as well as a forecast for the nominal load I L and
its worst-case deviations which can be mapped into the required
In the following, we focus on distributed secondary integral con- secondary control action p∗ .
trol strategies that are able to achieve the desired optimal injec- In such setup, a nominal operating point is scheduled offline
tions at the requirement of communication. based on a system model and a load forecast so that nominal volt-
The previous decentralized controllers (24) result in the station- age bounds are guaranteed, and the primary and secondary con-
ary injection IiS = Ii∗ − pi (t → ∞) which depend on initial val- trollers compensate for deviations from the nominally forecasted
ues, exogenous disturbances, and unknown load parameters and load profile in real-time. 
do not necessarily satisfy the optimality condition (20) of identical
marginal costs. Hence, we propose the following distributed con-
6. Simulation results
sensus filter to force an alignment of the marginal injection costs
αi pi :
We illustrate the performance of our proposed controllers in
IiS ∗
= Ii − Ci V̇iS − pi (29a) a simulation scenario. We consider the microgrid displayed in
Fig. 1, where VS = {a, b, c } and VL = {d, e, f , g , h}. The
m
 microgrid is operating in islanded mode, and the dashed blue lines
Di ṗi = Ci V̇ S − Bij αi pi − αj pj
 
(29b) indicate the communication topology among the controllers in
j =1
the distributed consensus filter (29). To achieve proportional load
where Di > 0 and the terms Bij = Bji ≥ 0 induce an undirected sharing, the droop coefficients Ci are chosen to be proportional
and connected communication graph among the sources Vs . The to the source capacity Ī as in (9). At t = 10 s the initial load
consensus filter (29) resembles the distributed averaging PI (DAPI) demand I L = [−1, −2, −3, −3, −2]T changes instantaneously to
controller proposed in Simpson-Porco et al. (2013), and it com- I L = [−4, −0.5, −1.5, −5, −0.5]T . We investigate the transient
bines the integral action (24) together with a consensus flow (Bullo, and stationary behavior of the closed-loop system using different
Cortés, & Martínez, 2009). At the price of requiring communication control strategies; see Fig. 6.
between the sources, the distributed consensus filter (29) achieves The simulation results using primary droop control (7) are
regulation of the voltage drifts while recovering the desired injec- shown in Fig. 6(a). The constant voltage drifts (v̇drift ) are visible
tions for proportional load sharing or economic optimality, respec- as nonzero and identical (in steady state) slopes in Fig. 6(a), and
tively. the load sharing ratios (IiS /Īi ) converge to the same steady-state
value, i.e., the load is shared proportionally. Fig. 6(b) shows the
Theorem 5.2 (Performance of Distributed Consensus Filter). Con- simulation results using decentralized integral control (24). The
sider the closed-loop secondary-controlled microgrid (1) with the dis- source voltages converge to constant values without drifts, but
tributed consensus filter (29). Then the following statements hold: the load sharing ratios do not converge to the same steady-state
values. Additionally, the red injection in Fig. 6(b) exceeds the value
(1) All source voltages ViS (t ) converge exponentially to stationary 1, that is, the associated source injection exceeds its capacity.
values without drift. Moreover, the blue injection in Fig. 6(b) shows that, the current
(2) The steady-state source injections achieve proportional load injection is negative at steady state, that is, the associated source
sharing (4) if the controller gains are chosen as in (9). should absorb (instead of supply) current, which reveals another
(3) The steady-state source injections are optimal with respect to the disadvantage of decentralized integral control. The simulation
economic dispatch problem (18) if the controller gains are chosen results using the distributed consensus filter (29) are shown in
as in (21). Fig. 6(c). Observe that the source voltages converge to constant
 J. Zhao, F. Dörfler / Automatica 61 (2015) 18–26 25

a a a

b b b

(a) Primary droop control (7). (b) Decentralized integral control (24). (c) Distributed consensus filter (29).

Fig. 6. Closed-loop performance of the microgrid under different control strategies. (For interpretation of the references to color in this figure legend, the reader is referred
to the web version of this article.)

values without drifts, and the load sharing ratios converge to −L̃VS∗ − p∗ + Ĩ = Om . Let ṼS = V S − VS∗ and P̃ = p − p∗ , then
the same values. Hence, the voltage drifts are regulated and system (A.1) becomes
proportional load sharing is achieved. Finally, note the different  
scales in the plots which indicate a superior transient performance
˙
Ṽ S

C −1 L̃ C −1 Im
 
ṼS
=−
of the distributed consensus filter (29). ˙
P̃ D−1 L̃ D (Im + Lc C −1 )
−1
P̃
    
7. Conclusions I O C −1 I L̃ O ṼS
=− , (A.2)
O D −1 I C + Lc O C −1 P̃
We proposed decentralized and distributed primary droop
and secondary integral control strategies in DC microgrids. We where O and I denote zero and identity matrices of appropriate
analyzed the properties and limitations of these control strategies, dimension. The characteristic equation of the negative system
and investigated their consistencies with tertiary-level objectives matrix in (A.2) reads as
such as proportional load sharing and an economic dispatch     
among the generating units. This work is a first step towards I O C −1 I L̃ O
det λI −
establishing an operation architecture for DC microgrids. In our O D−1 I C + Lc O C −1
initial setup, we assumed constant current or constant impedance  
I O
loads, and we considered purely resistive networks or networks = det
with lines modeled by the resistive–capacitive Π -model. In O D −1
ongoing and future work, we plan to study the robust performance
    −1  
I O C I L̃ O
in presence of transient stochastic disturbances as well as × det λ −
O D I C + Lc O C −1
different network models including resistive–inductive–capacitive     −1  
lines and constant power load models using the approximation I O C I L̃ O
= det λ −
proposed in Gentile, Simpson-Porco, Dörfler, Zampieri, and Bullo O D I C + Lc O C −1
(2014).  
I O
× det
Acknowledgments O D−1
  −1  
C I L̃ O
The authors wish to thank A. Davoudi and M. Andreasson for = det λI − .
I C + Lc O C −1 D −1
their helpful comments and suggestions.
 
L̃ O
Appendix. Proof of Theorem 5.2 The matrix is positive semidefinite with one zero
O C −1 D−1
 T
eigenvalue and corresponding eigenvector 1Tm O , and
Proof. The closed-loop state space model (1), (29) is
C −1
 
   ∗   S
 I
is positive semidefinite. Let σ > 0, and consider the
C V̇ S IS YSS YSL Im V I C + Lc
C −1
 
0n−m  =  I L  − YSLT YLL 0n−m  V L  perturbed matrix I
, which is positive definite. Ac-
I C + Lc + σ I
Dṗ IS∗ YSS YSL Lc C −1 + Im p cording
 to Sylvester’s
 Law of Inertia
 (Weintraub, 2011), the ma-
C −1 L̃ O
where p = [p1 . . . pm ]T , and Lc is the Laplacian matrix induced by
I
trix C + Lc + σ I
has one zero eigenvalue, and all
I O C −1 D−1
the communication graph with weights Bij = Bji ≥ 0. After Kron- other eigenvalues are positive. Recall that the eigenvalues of a ma-
reduction, the reduced system is trix are continuous functions of the matrix elements (via the char-
      acteristic equation). Since there is only a single zero and m − 1
C V̇ S L̃ Im VS Ĩ
=− + . (A.1) positive eigenvalues for any σ > 0, the number of zero eigenval-
Dṗ L̃ (Im + Lc C −1 ) p Ĩ
ues can either increase or remain unchanged as σ ↓ 0. For σ = 0, 0
Let p∗ = C 1m v̇drift = C 1m 1Tm Ĩ / i=1 Ci then 1Tm (Ĩ − p∗ ) = 0. It
m and [1m O]T are an eigenvalue and eigenvectorpair. Also, the
 range
C −1

L̃ O I
follows that (Ĩ − p∗ ) is in the range of L̃ and there is VS∗ so that space of
O C −1 D−1
and the null space of I C + Lc
do not
26 J. Zhao, F. Dörfler / Automatica 61 (2015) 18–26
 T
coincide. Thus, 1Tm Om is the only eigenvector associated to the Kundur, P. (1994). Power system stability and control. McGraw-Hill.
zero eigenvalue of the system (A.2). Hence, there is only one zero Morstyn, T., Hredzak, B., Demetriades, G. D., & Agelidis, V. G. (2015). Unified
distributed control for dc microgrid operating modes. IEEE Transactions on
eigenvalue and all the other eigenvalues are positive. Power Systems, PP(99), 1–11.
We conclude that the solutions of the system (A.2) converge Nasirian, V., Moayedi, S., Davoudi, A., & Lewis, F. L. (2015). Distributed cooperative
T control of dc microgrids. IEEE Transactions on Power Electronics, 30(4),
exponentially to [κ 1m Om ]T , where κ ∈ R. Equivalently, V S p

2288–2303.
∗ T Nilsson, D., & Sannino, A. (2004). Load modelling for steady-state and transient
converges to the constant vector VS∗T + κ 1Tm p
 
. Therefore, at analysis of low-voltage dc systems. In Industry applications conference, 2004.
steady state, V̇S (∞) = 0. This proves statement (1). 39th IAS annual meeting. Conference record of the 2004 IEEE. Vol. 2. October
(pp. 774–780).
To prove statements (2) and (3), we write (A.1) as
Ostrowski, Alexander, & Schneider, Hans (1962). Some theorems on the inertia of
     general matrices. Journal of Mathematical Analysis and Applications, 4(1), 72–84.
C V̇ S I I −L̃V S + Ĩ . Salomonsson, D., & Sannino, A. (2007). Low-voltage dc distribution system for
=
Dṗ I I + Lc C −1 −p commercial power systems with sensitive electronic loads. IEEE Transactions
on Power Delivery, 22(3), 1620–1627.
Shafiee, Q., Dragicevic, T., Vasquez, J. C., & Guerrero, J. M. (2014). Modeling, stability
 
I I
The matrix I I + Lc C −1
in the above equation has nullspace analysis and active stabilization of multiple dc-microgrid clusters. In Energy
conference (ENERGYCON), 2014 IEEE international. May (pp. 1284–1290).
[C 1m − C 1m ]T , and the equilibria of (A.1) are Simpson-Porco, J. W., Dörfler, F., & Bullo, F. (2013). Synchronization and power
    sharing for droop-controlled inverters in islanded microgrids. Automatica,
−L̃V S + Ĩ = τ C 1m 49(9), 2603–2611.
−p −C 1 m Simpson-Porco, J. W., Dörfler, F., & Bullo, F. (2015). On resistive networks of constant
power devices. IEEE Transactions on Circuits and Systems II: Express Briefs, 62(8),
811–815.
for some τ ∈ R. We multiply the equation −L̃V S + Ĩ = τ C 1m by 1Tm Tucci, M., Riverso, S., Vasquez, J. C., Guerrero, J. M., & Ferrari-Trecate, G. (2015). A
decentralized scalable approach to voltage control of DC islanded microgrids.
on both sides: the first term of left-hand side equals −1Tm L̃V S = 0,
ArXiv e-prints, March.
the second term of the left-hand side equals 1Tm Ĩ = 1Tm (IS∗ − Weintraub, Steven H. (2011). A guide to advanced linear algebra. (44).
Xu, Lie, & Chen, Dong (2011). Control and operation of a dc microgrid with variable
YSL YLL I ) = 1Tm IS∗ + 1Tn−m I L and the right-hand side equals
−1 L
generation and energy storage. IEEE Transactions on Power Delivery, 26(4),
1m τ C 1m = τ
S
i∈VS Ci . Thus, τ = (1m IS + 1n−m I )/(
T ∗ T L
i∈VS Ci ) and
 
2513–2522.
Zhang, Fuzhen (2006). The schur complement and its applications. Vol. 4. Springer.
p = τ C 1m . We conclude that the steady-state injections of sources Zhao, J., & Dörfler, F. (2015). Distributed control, load sharing, and dispatch in DC
are determined by I S (∞) = I ∗ − p = I ∗ − C τ 1m , which equals microgrids. In American control conference (ACC), 1–3 July (pp. 3304–3309).
the steady-state injections of the primary control system (1) and Zonetti, D., Ortega, R., & Benchaib, A. (2014). Modeling and control of high-voltage
direct-current transmission systems: From theory to practice and back. ArXiv
(7). Therefore, statements (2) and (3) in Theorem 5.2 follow from e-prints, June.
statement (2) of Theorem 3.1 and Corollary 4.1. 

References Jinxin Zhao is a Ph.D. candidate in the Department

of Mechanical and Aerospace Engineering, University of
California, Los Angeles. He received his M.Sc. degree in
Andreasson, M., Dimarogonas, D. V., Sandberg, H., & Johansson, K. H. (2014). Control Aerospace Engineering from University of Michigan, Ann
of MTDC transmission systems under local information. ArXiv e-prints, June. Arbor in 2013, and B.Sc. Degree in Mechanical Engineering
Boyd, S., & Vandenberghe, L. (2004). Convex optimization. Cambridge University from Zhejiang University in 2011.
Press. His current research interests focus on distributed
Bullo, F., Cortés, J., & Martínez, S. (2009). Distributed control of robotic networks.
control, coupled oscillators and CPG control.
Princeton University Press.
Dörfler, F., & Bullo, F. (2013). Kron reduction of graphs with applications to electrical
networks. IEEE Transactions on Circuits and Systems I: Regular Papers, 60(1),
150–163.
Dörfler, F., Simpson-Porco, J., & Bullo, F. (2014). Breaking the hierarchy: Distributed
control & economic optimality in microgrids. ArXiv e-prints, January.
Gentile, B., Simpson-Porco, J. W., Dörfler, F., Zampieri, S., & Bullo, F. (2014). On
reactive power flow and voltage stability in microgrids. In American control Florian Dörfler is an Assistant Professor at the Automatic
conference. Portland, OR, June (pp. 759–764). Control Laboratory at ETH Zürich. He received his Ph.D.
Guerrero, J. M., Vasquez, J. C., Matas, J., de Vicuna, L. G., & Castilla, M. (2011). degree in Mechanical Engineering from the University of
Hierarchical control of droop-controlled AC and DC microgrids—a general California at Santa Barbara in 2013, and a Diploma degree
approach toward standardization. IEEE Transactions on Industrial Electronics, in Engineering Cybernetics from the University of Stuttgart
58(1), 158–172. in 2008. From 2013 to 2014 he was an Assistant Professor
Horn, R. A., & Johnson, C. R. (1985). Matrix analysis. Cambridge University Press. at the University of California Los Angeles. His primary
Ito, Youichi, Zhongqing, Yang, & Akagi, Hirofumi (2004). Dc microgrid based research interests are centered around distributed control,
distribution power generation system. In Power electronics and motion control complex networks, and cyber-physical systems currently
conference, 2004. IPEMC 2004. The 4th international. Vol. 3 (pp. 1740–1745). IEEE. with applications in energy systems and smart grids. He
Justo, Jackson John, Mwasilu, Francis, Lee, Ju, & Jung, Jin-Woo (2013). Ac-microgrids is a recipient of the 2009 Regents Special International
versus dc-microgrids with distributed energy resources: A review. Renewable Fellowship, the 2011 Peter J. Frenkel Foundation Fellowship, the 2010 ACC Student
and Sustainable Energy Reviews, 24, 387–405. Best Paper Award, the 2011 O. Hugo Schuck Best Paper Award, and the 2012–2014
Karlsson, P., & Svensson, J. (2003). Dc bus voltage control for a distributed power Automatica Best Paper Award. As a co-advisor and a co-author, he has been a finalist
system. IEEE Transactions on Power Electronics, 18(6), 1405–1412. for the ECC 2013 Best Student Paper Award.