-
IEEE TRANSACTIONS ON SMART GRID, VOL. 6, NO. 3, MAY 2015
1137
Distributed Optimal Energy Management inMicrogrids
Wenbo Shi, Student Member, IEEE, Xiaorong Xie, Member, IEEE,
Chi-Cheng Chu, and Rajit Gadh
AbstractEnergy management in microgrids is typically for-mulated
as a nonlinear optimization problem. Solving it in acentralized
manner does not only require high computationalcapabilities at the
microgrid central controller (MGCC), but mayalso infringe customer
privacy. Existing distributed approaches,on the other hand, assume
that all generations and loads areconnected to one bus, and ignore
the underlying power dis-tribution network and the associated power
flow and systemoperational constraints. Consequently, the schedules
producedby those algorithms may violate those constraints and thus
arenot feasible in practice. Therefore, the focus of this paper is
onthe design of a distributed energy management strategy (EMS)for
the optimal operation of microgrids with consideration of
thedistribution network and the associated constraints.
Specifically,we formulate microgrid energy management as an optimal
powerflow problem, and propose a distributed EMS where the MGCCand
the local controllers jointly compute an optimal schedule.We also
provide an implementation of the proposed distributedEMS based on
IEC 61850. As one demonstration, we apply theproposed distributed
EMS to a real microgrid in GuangdongProvince, China, consisting of
photovoltaics, wind turbines, dieselgenerators, and a battery
energy storage system. The simulationresults demonstrate the
effectiveness and fast convergence of theproposed distributed
EMS.
Index TermsDistributed algorithms, distribution networks,energy
management, microgrids, optimal power flow (OPF),optimization.
I. INTRODUCTION
M ICROGRIDS are low-voltage distribution systems con-sisting of
distributed energy resources (DERs) andcontrollable loads, which
can be operated in either islandedor grid-connected mode [2]. DERs
include a variety of dis-tributed generation (DG) units such as
wind turbines (WTs),photovoltaics (PVs), and distributed storage
(DS) units such asbatteries. Sound operation of a microgrid
requires an energymanagement strategy (EMS) which controls the
power flowsin the microgrid by adjusting the power
imported/exported
Manuscript received July 18, 2014; revised October 2, 2014;
acceptedNovember 16, 2014. Date of publication December 9, 2014;
date of cur-rent version April 17, 2015. This work was supported by
the Researchand Development Program of the Korea Institute of
Energy Research underGrant B4-2411-01. This paper has been
presented in part at the IEEEInternational Conference on Smart Grid
Communications, Venice, Italy,Nov. 2014 [1]. Paper no.
TSG-00738-2014.
W. Shi, C.-C. Chu, and R. Gadh are with the Smart Grid Energy
ResearchCenter, University of California, Los Angeles, CA 90095 USA
(e-mail:[email protected]).
X. Xie is with the State Key Laboratory of Power Systems,
Department ofElectrical Engineering, Tsinghua University, Beijing
100084, China.
Color versions of one or more of the figures in this paper are
availableonline at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSG.2014.2373150
from/to the main grid, the dispatchable DERs, and the
con-trollable loads based on the present and forecasted
informationof the market, the generations, and the loads in order
to meetcertain operational objectives (e.g., minimizing costs)
[2].
Energy management in microgrids is typically formulatedas a
nonlinear optimization problem. Various centralized meth-ods have
been proposed to solve it in the literature, includingmixed integer
programming [3], sequential quadratic program-ming [4], particle
swarm optimization [5], neural networks [6],etc. The centralized
approaches [3][6] require high computa-tional capabilities at the
microgrid central controller (MGCC),which is neither efficient nor
scalable. Moreover, a central-ized EMS requires the MGCC to gather
information of theDERs (e.g., production costs, constraints, etc.)
and the loads(e.g., customer preferences, constraints, etc.) as the
inputs foroptimization. However, different DERs may belong to
differ-ent entities and they may keep their information private
[7].Customers may also be unwilling to expose their informationdue
to privacy [8]. Therefore, in this paper, we are interestedin
developing a distributed EMS which is efficient, scalable,and
privacy preserving.
Several distributed algorithms have been proposed for
theoperation of microgrids in the literature. In [7], a
distributedalgorithm based on the classical symmetrical
assignmentproblem is proposed. Energy management is formulated asa
resource allocation problem in [9] and distributed algo-rithms are
proposed for distributed allocation. A convexproblem formulation
can be found in [10] and dual decom-position is used to develop a
distributed EMS to maintain thesupplydemand balance in microgrids.
A privacy-preservingenergy scheduling algorithm in microgrids is
proposed in [8],where the privacy constraints are integrated with
the linearprogramming model and distributed algorithms are
developed.In [11], the additive-increase/multiplicative-decrease
algorithmis adopted to optimize DER operations in a
distributedfashion.
The problem with the existing distributed approaches
formicrogrid energy management [7][11] is that they considerthe
supplydemand matching in an abstract way where theaggregate demand
is simply equal to the supply. They assumethat all generations and
loads are connected to one bus andignore the underlying power
distribution network and theassociated power flow (e.g., Kirchhoffs
law) and system oper-ational constraints (e.g., voltage
tolerances). Consequently, theschedules produced by those
algorithms may violate thoseconstraints and thus are not feasible
in practice. It is worthnoting that distribution networks have been
taken into account
1949-3053 c 2014 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.
-
1138 IEEE TRANSACTIONS ON SMART GRID, VOL. 6, NO. 3, MAY
2015
in a few recent demand response studies [12]. However,the idea
of integrating distribution networks with distributedenergy
management in microgrids, where both supply-side anddemand-side
management (DSM) are considered has not beenexplored.
The focus of this paper is on the design of a distributed EMSfor
the optimal operation of microgrids with consideration ofthe
underlying power distribution network and the
associatedconstraints. More specifically, we consider a microgrid
con-sisting of multiple DERs and controllable loads. The
objectiveof the EMS is to control the power flows in the microgrid
inorder to: 1) minimize the cost of generation, the cost of
energystorage, and the cost of energy purchase from the main
grid;2) minimize the dissatisfactions of the customers in the
DSM;and 3) minimize the power losses subject to the DER
con-straints, the load constraints, the power-flow constraints,
andthe system operational constraints.
Specifically, we formulate energy management in micro-grids as
an optimal power flow (OPF) problem. The OPFproblem is difficult to
solve due to the nonconvex power-flowconstraints. We convexify the
OPF problem by relaxing thepower-flow constraints (see [13], [14]
for a tutorial on con-vex relaxation of OPF). Sufficient conditions
for the exactnessof the relaxation have been derived in [15][18],
whichhold for a variety of IEEE test systems and real-worldsystems.
Therefore, we focus on solving the relaxed OPFproblem (OPF-r) in
this paper. Note that most of the convexrelaxations of OPF [15][18]
assume a single-phase or bal-anced three-phase distribution
network. See [19] for convexrelaxation of OPF in multiphase
networks. The OPF-r prob-lem is a centralized convex optimization
problem. To solve itin a distributed manner, we propose a
distributed EMS wherethe MGCC and the local controllers (LCs)
jointly compute anoptimal schedule.
In order to illustrate the feasibility of the proposed EMSin
real systems, we provide an implementation based on theIEC 61850
standard which provides standardized communi-cation and control
interfaces for all DER devices. As onedemonstration, we apply the
proposed EMS to a real micro-grid in Guangdong Province, China,
consisting of PVs, WTs,diesel generators, and a battery energy
storage system (BESS).The simulation results demonstrate the
effectiveness and fastconvergence of the proposed EMS. We show that
the pro-posed EMS is able to manage the operations in the
microgridto achieve the objective while maintaining the bus
voltageswithin the allowed range. Furthermore, we discuss the
loca-tion effect, the load shedding/shifting, the ramping
constraint,and the trade-offs in the objective.
Compared with a preliminary version of this paper [1],this paper
provides a more realistic and detailed systemmodel by incorporating
the generation ramping constraint, theinverter capacity, the
battery efficiency, and the load shift-ing. We also provide an IEC
61850 implementation of theproposed distributed EMS to demonstrate
its feasibility inreal systems. A case study using the real pricing
data isprovided to give better insights into microgrid energy
manage-ment. A more comprehensive analysis of the proposed EMS
isgiven, including the discussions on the location effect, the
load
Fig. 1. System architecture.
shedding/shifting, the ramping constraint, and the trade-offs
inthe objective.
The rest of this paper is organized as follows. We intro-duce
the system model in Section II and propose the EMS inSection III.
Simulation results are provided in Section V andthe conclusion is
given in Section VI.
II. SYSTEM MODELIn this section, we describe the system model
for devel-
oping the proposed distributed EMS. We first give anoverview of
the system followed by the DG model, theDS model, and the load
model. We then model thepower distribution network using a branch
flow modeland formulate microgrid energy management as an
OPFproblem.
A. System OverviewA low-voltage power distribution network
generally has a
radial structure [12]. Thus, we consider a radial microgrid
con-sisting of a set of DG units denoted by G {g1, g2, . . . ,
gG},DS units denoted by B {b1, b2, . . . , bB}, and
controllableloads denoted by L {l1, l2, . . . , lL}. In the
microgrid, thereis a MGCC which coordinates the operations of the
DERs andthe controllable loads. At each of the DERs and the
loads,there is a LC which is able to coordinate with the MGCC
tocompute its schedule locally via a two-way
communicationinfrastructure. Fig. 1 shows the system
architecture.
In this paper, we use a discrete-time model with a
finitehorizon. We consider a time period or namely a
schedulinghorizon, which is divided into T equal intervals t,
denotedby T {0, 1, . . . , T 1}.
B. DG ModelWe consider both renewable and conventional DGs
in
the microgrid. Renewable DGs such as PVs and WTs
arenondispatchable and conventional DGs such as diesel are
dis-patchable. For each DG g G, we denote its complex outputpower
by sg(t) pg(t) + iqg(t), where pg(t) is the activepower and qg(t)
is the reactive power.
1) Renewable DG: A renewable DG unit such as PV orWT is not
dispatchable and its output power is dependenton the availability
of the primary sources (i.e., sun irradianceor wind). Therefore,
forecast is required in order to considerthem in the energy
management optimization. Methods for PVforecasting [20] and WT
forecasting [21] can be utilized. In
-
SHI et al.: DISTRIBUTED OPTIMAL ENERGY MANAGEMENT IN MICROGRIDS
1139
our model, we assume that the complex power of a renewableDG
unit over the scheduling horizon is given and there is nogeneration
cost for renewable DGs.
2) Conventional DG: A conventional DG unit such asdiesel is a
dispatchable source, its output power is a variablewith the
following constraints:
0 pg(t) pg, t T , (1)|pg(t) pg(t 1)| rgpg, t T , (2)
where pg is the maximum output power and rg (0, 1] is theramping
parameter.
For a given conventional DG unit, its reactive powergenerated at
the inverter is bounded by
pg(t)2 + qg(t)2 s2g, t T , (3)where sg is the capacity of the
inverter.
We model the conventional DG generation cost at each timet T
using a quadratic model [10]
Cg( pg(t)) g(
pg(t)t)2 + gpg(t)t + cg, (4)
where g, g, and cg are constants.
C. DS ModelWe consider batteries as the DS units in the
microgrid.
For a given battery b B, we denote its complex power bysb(t)
pb(t) + iqb(t), where pb(t) is the active power (posi-tive when
charging and negative when discharging) and qb(t)is the reactive
power. Let Eb(t) denote the energy stored inthe battery at time t.
A given battery can be modeled by thefollowing constraints:
pb pb(t) pb, t T , (5)
pb(t)2 + qb(t)2 s2b, t T , (6)Eb(t + 1) = bEb(t) + pb(t)t, t T ,
(7)
Eb Eb(t) Eb, t T , (8)Eb(T) Eeb, (9)
where pb is the maximum charging rate, pb is the maximum
discharging rate, sb is the capacity of the inverter, b (0,
1]captures the battery efficiency, Eb and Eb are the minimum
andmaximum allowed energy stored in the battery, respectively,and
Eeb is the minimum energy that the battery should maintainat the
end of the scheduling horizon.
We use a cost function to capture the damages to the batteryby
the charging and discharging operations. Three types ofdamages are
considered: 1) fast charging; 2) frequent switchesbetween charging
and discharging; and 3) deep discharging.We model the battery cost
as [22]
Cb(pb) b
tTpb(t)2 b
T2
t=0pb(t + 1)pb(t)
+ b
tT
(min(Eb(t) bEb, 0)
)2, (10)
where pb (pb(t), t T ) is the charging/discharging vec-tor, b,
b, b, and b are positive constants. b, b, and b
trade off among fast charging, switches between charging
anddischarging, and deep discharging. The above function is con-vex
when b > b. We choose b = 0.2, meaning thatthe cost function
penalizes deep charging when the energystored in the battery Eb(t)
is less than 20% of the batterycapacity Eb.
D. Load ModelWe consider two types of controllable loads in the
micro-
grid: interruptible and deferrable loads. An interruptible
loadcan be shedded and a deferrable load can be shifted in timebut
need to consume a certain amount of energy before adeadline.
For each load l L, we denote its complex power bysl(t) pl(t) +
iql(t) and it is bounded by
pl(t) pl(t) pl(t), t T , (11)
ql(t) ql(t) ql(t), t T , (12)
wherepl(t) and pl(t) are the minimum and maximum active
power, respectively, andql(t) and ql(t) are the minimum and
maximum reactive power, respectively.For a deferrable load, we
assume that its consumed energy
is bounded by
El
tTpl(t)t El, (13)
where El and El are the energy lower and upper
bound,respectively.For each load, we define a demand vector denoted
by
pl ( pl(t), t T ) and a cost function Cl(pl) which mea-sures the
dissatisfaction of the customer using the demandschedule pl.
The cost function of an interruptible load is dependent onthe
shedded load and can be defined as
Cl(pl)
tTl
(min
(pl(t) p fl (t), 0
))2, (14)
where p fl (t) is the forecasted load and l is a positive
constant.The above cost function is nonzero only when there is
loadshedding, i.e., pl(t) < p fl (t).
The cost function of a deferrable load is dependent on
theunfulfilled energy and can be defined as
Cl(pl) l
(
El
tTpl(t)t
)
, (15)
where l is a positive constant.
E. Distribution Network ModelA distribution network can be
modeled as a connected graph
G = (N , E), where each node i N represents a bus and eachlink
in E represents a branch (line or transformer). We denotea link by
(i, j) E . Power distribution networks are typicallyradial and the
graph G becomes a tree for radial distributionsystems. We index the
buses in N by i = 0, 1, . . . , n, andbus 0 denotes the feeder
which has a fixed voltage and flexiblepower injection.
-
1140 IEEE TRANSACTIONS ON SMART GRID, VOL. 6, NO. 3, MAY
2015
For each link (i, j) E , let zij rij + ixi,j be the
compleximpedance of the branch, Iij(t) be complex current from
busesi to j, and Sij(t) Pij(t) + iQij(t) be the complex
powerflowing from buses i to j.
For each bus i N , let Vi(t) be the complex voltage at busi and
si(t) pi(t) + iqi(t) be the net load which is the loadminus the
generation at bus i. Each bus i N \{0} is connectedto a subset of
DG units Gi, DS units Bi, and loads Li. The netload at each bus i
satisfies
si(t) = sli(t) + sbi(t) sgi(t), i N \ {0},t T , (16)where
sli(t)
lLi sl(t), sbi(t)
bBi sb(t), and
sgi
gGi sg(t).Given the radial distribution network G, the feeder
volt-
age V0, and the impedances {zij}(i,j)E , then the other
variablesincluding the power flows, the voltages, the currents, and
thebus loads satisfy the following physical laws for all
branches(i, j) E and all t T .
1) Ohms law:Vi(t) Vj(t) = zijIij(t). (17)
2) Power-flow definition:Sij(t) = Vi(t)Iij(t). (18)
3) Power balance:Sij(t) zij|Iij(t)|2
k:( j,k)ESjk(t) = sj(t). (19)
Using (17)(19) and in terms of real variables, we can modelthe
steady-state power flows in a given distribution network Gas [13]:
(i, j) E,t T
pj(t) = Pij(t) rijij(t)
k:( j,k)EPjk(t), (20)
qj(t) = Qij(t) xijij(t)
k:( j,k)EQjk(t), (21)
vj(t) = vi(t) 2(rijPij(t) + xijQij(t)
) +(
r2ij + x2ij)
ij(t),(22)
ij(t) = Pij(t)2 + Qij(t)2vi(t)
, (23)
where ij(t) |Iij(t)|2 and vi(t) |Vi(t)|2.Equations (20)(23)
define a system of equations
in the variables (P(t), Q(t), v(t), l(t), s(t)), whereP(t) (
Pij(t), (i, j) E), Q(t) (Qij(t), (i, j) E),v(t) (vi(t), i N \ {0}),
l(t) (ij(t), (i, j) E), ands(t) (si(t), i N \ {0}). The phase
angles of the voltagesand the currents are not included. But they
can be uniquelydetermined for radial systems [14].
F. Energy ManagementThe objective of the microgrid operator is
to minimize its
operational cost, while delivering reliable and
high-qualitypower to the loads. However, the introduction of DERs
makesit challenging to balance the supply and demand in a
micro-grid, especially in islanded mode where only DERs are usedas
the supply. The integration of DERs may also increase the
voltage significantly, which reduces the quality of the volt-age
received by other users in the distribution network. Thus,in this
paper, we study microgrid energy management aimingat achieving the
operational objective of the microgrid opera-tor, while meeting the
supplydemand balance and the voltagetolerance constraints.
We consider the following voltage tolerance constraints inthe
microgrid:
Vi |Vi(t)| Vi, i N \ {0},t T , (24)where Vi and Vi correspond to
the minimum and maximumallowed voltage according to the
specification, respectively.
The net power injected to the microgrid from the main gridis
given by
s0(t) =
j:(0,j)Es0j(t), t T . (25)
If the microgrid is operated in islanded mode, thens0(t) = 0. If
the microgrid is operated in grid-connected mode,then s0(t) is the
net complex power traded between themicrogrid and the main
grid.
We model the cost of energy purchase from the main gridat each
time t T as
C0(t, p0(t)) (t)p0(t)t, (26)where (t) is the market energy
price. Note that p0(t) can benegative, meaning that the microgrid
can sell its surplus powerto the main grid.
The objective of the energy management in the microgridis to: 1)
minimize the cost of generation, the cost of energystorage, and the
cost of energy purchase from the main grid;2) minimize the
dissatisfactions of the customers in the DSM;and 3) minimize the
power losses subject to the DER con-straints, the load constraints,
the power-flow constraints, andthe system operational constraints
(voltage tolerances).
We define P (P(t), t T ), Q (Q(t), t T ), v (v(t), t T ), l
(l(t), t T ), sg (sg(t), t T ), sb (sb(t), t T ), sl (sl(t), t T ),
s (sg, sb, sl, g G, b B, l L), and Cg(pg) tT Cg( pg(t)). Microgrid
energymanagement can be formulated as an OPF problem
OPFmin
P,Q,v,l,sg
gGCg(pg) + b
bBCb(pb) + l
lLCl(pl)
+ 0
tTC0(t, p0(t)) + p
tT
(i,j)Erijij(t)
s.t. (1)(3), (5)(9), (11)(13), (16),(20)(25),
where g, b, l, 0, and p are the parameters to tradeoff among the
cost minimizations and the power lossminimization.
III. DISTRIBUTED EMSThe previous OPF problem is nonconvex due to
the
quadratic equality constraint in (23) and is NP-hard to
solve
-
SHI et al.: DISTRIBUTED OPTIMAL ENERGY MANAGEMENT IN MICROGRIDS
1141
in general [13]. We, therefore, relax them to inequalities
ij(t) Pij(t)2 + Qij(t)2vi(t)
, (i, j) E,t T . (27)We then consider the following convex
relaxation of OPF:
OPF rmin
P,Q,v,l,sg
gGCg(pg) + b
bBCb(pb) + l
lLCl(pl)
+ 0
tTC0(t, p0(t)) + p
tT
(i,j)Erijij(t)
s.t. (1)(3), (5)(9), (11)(13), (16),(20)(22), (24), (25),
(27).
If the equality in (27) is attained in the solution to
OPF-r,then it is also an optimal solution to OPF. The sufficient
con-ditions under which the relaxation is exact have been
exploitedin [15][18]. Roughly speaking, if the bus voltage is
keptaround the nominal value and the power injection at each busis
not too large, then the relaxation is exact. Detailed condi-tions
when the voltage upper bound is not important can befound in [15]
and [16]. See [17], [18] on how to deal withthe voltage upper
bound. In this paper, we assume that thesufficient conditions
specified in [18] hold for the microgridand thus we focus on
solving the OPF-r problem.
The above OPF-r problem is a centralized optimizationproblem. In
order to design an efficient, scalable, and privacy-preserving EMS,
we propose a distributed algorithm to solvethe OPF-r problem using
the predictor corrector proximal mul-tiplier (PCPM) algorithm
(refer to [23] or the Appendix fordetails). Note that we use PCPM
to develop the proposeddistributed algorithm here. Other
distributed algorithms suchas the alternating direction method of
multipliers (ADMM)algorithm [24] may also apply to solve the OPF-r
problem ina distributed manner.
Initially set k 0. The LCs of the DERs and the loadsset their
initial schedules randomly and communicate them tothe MGCC. In the
meantime, the MGCC randomly choosesthe initial ski (t) pki (t)+
iqki (t) and two virtual control signals{ki (t)}tT , {ki (t)}tT for
each bus i N \{0}. i(t) and i(t)are the Lagrangian multipliers
associated with the active andreactive power at bus i,
respectively.
At the beginning of the kth step, the MGCC sends twocontrol
signals ki (t) ki (t)+ (pkli(t)+pkbi(t)pkgi(t)pki (t))and ki (t) ki
(t) + (qkli(t) + qkbi(t) qkgi(t) qki (t)) to theLCs of the DERs and
the loads connected to bus i, where is a positive constant. Then
the following occurs.
1) The LC of each conventional DG unit solves the follow-ing
problem:
EMS LC(DG)min
sggCg(pg) +
(ki
)Tpg +
(ki
)Tqg
+ 12
pg pkg
2 + 1
2
qg qkg
2
s.t. (1) (3),where ki (ki (t), t T ) and ki (ki (t), t T ).
Theoptimal sg is set as sk+1g .
Algorithm 1 Proposed Distributed EMS1: initialization k 0. The
LCs set the initial schedules randomly
and return them to the MGCC. The MGCC sets the initial ki (t),ki
(t) and the initial s
ki (t) randomly.
2: repeat3: The MGCC updates ki (t) and
ki (t) and sends two control
signals ki and ki to the LCs connected to bus i.
4: The LC at each DER and each load calculates a new scheduleby
solving the corresponding EMS-LC problem.
5: The MGCC computes a new sk+1(t) for each time t T bysolving
the EMS-MGCC problem.
6: The LC communicates the new schedule to the MGCC.7: The MGCC
updates k+1i (t) and
k+1i (t).
8: k k + 1.9: until convergence
2) The LC of each DS unit solves the following problem:EMS
LC(DS)
minsb
bCb(pb) +(ki
)Tpb +
(ki
)Tqb
+ 12
pb pkb
2 + 1
2
qb qkb
2
s.t. (5) (9).The optimal sb is set as s
k+1b .
3) The LC of each load solves the following problem:EMS
LC(Load)
minsl
lCl(pl) +(ki
)Tpl +
(ki
)Tql
+ 12
pl pkl
2 + 1
2
ql qkl
2
s.t. (11)(13).
The optimal sl is set as sk+1l .
4) The MGCC solves the following problem for eachtime t T :EMS
MGCC
minP(t),Q(t),
v(t),l(t),s(t)
0C0(t, p0(t)) + p
(i,j)Erijij(t)
(k(t)
)Tp(t)
(k(t)
)Tq(t)
+ 12
p(t) pk(t)
2 + 1
2
q(t) qk(t)
2
s.t. (20)(22), (24), (25), (27),
where k(t) (ki (t), i N \ {0}) and k(t) (ki (t), i N \ {0}). The
optimal s(t) is set as sk+1(t).
At the end of the kth step, the LCs communicate their
newschedules sk+1l , sk+1g , and s
k+1b to the MGCC and the MGCC
updates k+1i (t) ki (t) + ( pk+1li (t) + pk+1bi (t) pk+1gi (t)
pk+1i (t)) and
k+1i (t) ki (t)+ (qk+1li (t)+qk+1bi (t)qk+1gi (t)
qk+1i (t)) for all i N \ {0} and all t T . Set k k + 1,
andrepeat the process until convergence.
A complete description of the proposed distributed EMS canbe
found in Algorithm 1. When is small enough, the abovealgorithm will
converge to the optimal solution to OPF-r which
-
1142 IEEE TRANSACTIONS ON SMART GRID, VOL. 6, NO. 3, MAY
2015
Fig. 2. IEC 61850 implementation.
is also the optimal solution to OPF if the relaxation is exact
and( pkli(t)+pkbi(t)pkgi(t)pki (t)) and (qkli(t)+qkbi(t)qkgi(t)qki
(t))will converge to zero [23]. As we can see, the MGCC and theLCs
jointly compute the optimal schedule.
In the proposed distributed EMS, the private information ofthe
DERs and the loads is stored at the LC where the EMS-LCproblem is
solved locally. The MGCC solves the EMS-MGCCproblem using the
system information, including the topology,the power losses, the
market energy price, etc. The informationexchanged between the MGCC
and the LCs include only thecontrol signals and the schedules.
Therefore, the privacy ofthe DERs (i.e., production costs and
constraints) and the loads(i.e., customer preferences and
constraints) are both preservedby the proposed EMS.
IV. IEC 61850 IMPLEMENTATIONIEC 61850 is a set of international
standards designed
originally for communications within substation
automationsystems [25]. More recently, IEC 61850 draws attention
toresearchers in the area of microgrids as it provides
standard-ized communication and control interfaces for all DER
devicesto achieve interoperability in microgrids. Due to its
essentialrole in microgrid systems, it is important that the
microgridEMS is compatible with IEC 61850. To this end, we
providean implementation of the proposed distributed EMS based
onthe IEC 61850 standard.
In IEC 61850, each DER unit is modeled as a logicaldevice (LD).
The LD is composed of the relevant logicalnodes (LNs) which are
predefined groupings of data objectsthat serve specific functions.
Refer to [26] for detailed defini-tions of LNs for DER devices. The
LN named DER energy andancillary services schedule (DSCH) can be
used to implementthe proposed EMS. In a DSCH LN, an array of the
times-tamps and values can be read or written using the IEC
61850ACSI services (e.g., GetDataValues and SetDataValues)
andmultiple schedules can be defined in parallel.
In order to implement the proposed EMS, we use fourDSCH LNs:
active power, reactive power, price for activepower, and price for
reactive power. The active power andreactive schedules are the
results from solving the EMS-LCproblem. The control signals sent by
the MGCC ki and kican be viewed as the price for active power and
reactivepower, respectively. Fig. 2 illustrates the proposed IEC
61850implementation.
Fig. 3. Topology of the microgrid.
Algorithm 1 needs a few updates to support the proposedIEC 61850
implementation. In step 3, the MGCC sends thecontrol signals by
writing to the DSCH LNs of the pricefor active and reactive power.
In step 4, after the LC solvesthe EMS-LC problem using the control
signals read from theDSCH LNs of the prices, the new schedules are
written tothe corresponding DSCH LNs. In step 6, the LC sends the
newschedules to the MGCC using the IEC 61850 report controlblock
(RCB) mechanism. At convergence, the optimal sched-ules will be
stored in the DSCH LNs and used to control theDER device.
V. PERFORMANCE EVALUATIONIn this section, we demonstrate the
proposed distributed
EMS by applying it to a real microgrid. We first describethe
microgrid system and the setup used in the simulation.We then
present the simulation results in islanded and grid-connected mode
followed by the discussions on the locationeffect, the load
shedding/shifting, the ramping constraint, andthe trade-offs in the
objective.
A. Simulation SetupFig. 3 shows the configuration of a real
microgrid in
Guangdong Province, China. The numbers under the DERsand the
loads in the figure correspond to the maximum power.We use this
microgrid to demonstrate the proposed distributedEMS. In the
simulation, a day starts from 12 A.M. The timeinterval t in the
model is 1 h and we denote a day byT {0, 1, . . . , 23}, where each
t T denotes the hour of[t, t + 1].
We set the cost function of the diesel generation asCg( pg(t))
10( pg(t)t)2 + 70( pg(t)t). The generationramping parameter rg is
chosen to be 0.3. The capacity ofthe BESS Eb is 3 MWh and Eb is
chosen to be 0.1 MWh.We set Eb(0) = 1.5 MWh, Eeb = 1.0 MWh, and b =
0.95.The parameters in the cost function of the battery are
chosenas b = 1, b = 0.75, and b = 0.5. We choose l = 103 and
-
SHI et al.: DISTRIBUTED OPTIMAL ENERGY MANAGEMENT IN MICROGRIDS
1143
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
2430
40
50
60
70
80
90
100
110
120
130
140
150
Time
Pric
e ($/
MWh)
Fig. 4. Day-ahead price from CAISO.
Fig. 5. Output schedules in islanded mode.
l = 102 in the cost function of the interruptible and
deferrableloads, respectively. We assume that bus 5 is a deferrable
loadand the rest of the loads are all interruptible. For the
interrupt-ible loads, the maximum load shedding percentage is
chosenrandomly from the range [20%, 50%]. For the deferable
load,the start time is chosen randomly from [10, 13] and the
dead-line is chosen randomly from [20, 23]. The minimum
energyrequirement El is chosen randomly from [3pl, 5pl] and
theenergy upper bound is set to be El = El + 2pl, where plis the
maximum power of load l. Perfect forecasting of thePV, the WT, and
the loads is assumed. For the forecastingof the deferrable load, we
assume that the load consumespower at the maximum rate from the
start time until the con-sumed energy reaches the energy upper
bound. We use thereal day-ahead energy price from California ISO
(CAISO) inthe simulation as shown in Fig. 4.
B. Case StudyWe apply the proposed EMS to the microgrid using
the
setup described above. The voltage tolerances are set to be[0.95
p.u., 1.05 p.u.]. The parameters in the algorithm arechosen as g =
1, b = 1, l = 1, 0 = 1, p = 1, and = 0.75.
The day-ahead schedules produced by the proposed EMSin islanded
mode is shown in Fig. 5. In islanded mode, themicrogrid can utilize
only the DERs to serve the loads. Sincethe nondispatchable
renewable generations (i.e., PV and WT)in the microgrid only serve
for a small proportion of the loads,diesel is the main power
source. From the figure, we can seethat the total diesel generation
changes in the same trend as
Fig. 6. Output schedules in grid-connected mode.
Fig. 7. Net power injected to the microgrid.
the total load. We can also observe the
charging/dischargingcycles of the battery from the figure. The
battery is chargedwhen the renewable generation is high and
discharged whenit is low, serving as the storage for the renewables
in themicrogrid.
Fig. 6 shows the results in grid-connected mode. Comparedwith
Fig. 5, the total diesel generation is decreased sig-nificantly in
grid-connected mode. Diesel generation is nolonger the main power
source in grid-connected mode as themicrogrid is able to import
power from the main grid. Thebattery in grid-connected mode is the
storage for not onlythe renewables, but also the cheap power from
the main grid.It is also charged when the energy price is low and
dis-charged when the energy price is high, making profits for
themicrogrid.
The net power injected to the microgrid in grid-connectedmode is
shown in Fig. 7. It can be seen from the figure thatthe microgrid
imports power when the energy price is lowand exports power when
the price is high (t [16, 18]). Ifwe compare the total operational
cost of the microgrid (i.e.,the value of the objective function) in
the two modes, we canfind that grid-connected mode decreases the
operational costby ($6491 $5484)/$6491 = 15.5%.
Fig. 8 shows the dynamics of EMS-MGCC and EMS-LC
ingrid-connected mode. We can see that our proposed
distributedalgorithm converges fast. For the simulations, we also
verifythat the solution to the centralized OPF-r problem is the
sameas the solution to the distributed algorithm. We further
verify
-
1144 IEEE TRANSACTIONS ON SMART GRID, VOL. 6, NO. 3, MAY
2015
Fig. 8. Dynamics of EMS-MGCC and EMS-LC.
Fig. 9. Maximum and minimum bus voltage in the microgrid.
that the equality in (27) is attained in the optimal solution
toOPF-r, i.e., OPF-r is an exact relaxation of OPF.
C. Discussion1) Location Effect: In order to exemplify the
effect on volt-
age tolerances, we increase the line lengths by five times.Fig.
9 shows the maximum and the minimum bus voltagein the microgrid
over time. It can be seen from the figurethat the bus voltages in
the microgrid are well maintainedwithin the allowed range. The
maximum voltage reaches theupper bound when the nondispatchable
renewable generationis high and the minimum voltage reaches the
lower boundwhen the load is high. This is because the generation
injectspower to the distribution network and hence increases the
volt-age, while the load consumes power and thus decreases
thevoltage.
In order to understand how the voltage tolerances aremaintained
by the proposed EMS, we look into the demandreduction of each load
as shown in Fig. 10. From the figure, itcan be easily seen that the
demand reduction of bus 2 and 3is significantly more than the other
buses. This is becauseboth bus 2 and 3 are far from the generation
or the feeder andthus the voltage drop along the line is
significant. Therefore,more demand reduction is required at bus 2
and 3 in order tomaintain the bus voltage above the minimum allowed
voltage.Similarly, those loads close to the nondispatchable
renewablegeneration (bus 11 and 12) need to consume more in orderto
reduce the high voltage due to the generation, leading toless
demand reduction. The result shows the location effectthat a DSM
scheme may discriminate the loads based on theirlocations [12].
Fig. 10. Demand reduction of the loads.
Fig. 11. Load shifting.
2) Load Shedding/Shifting: Fig. 10 also illustrates how theloads
are shedded in grid-connected mode. As can be expected,the loads
are shedded in response to the price: more loadsare shedded when
the price is high in order to save cost.In islanded mode, load
shedding is mainly used to balancethe local supply. If we compare
the total amount of shed-ded loads in the two modes, we can find
that islanded mode(8.91 WM) makes more load shedding than
grid-connectedmode (6.21 MW).
Fig. 11 shows the load shifting in grid-connected mode. Ascan be
seen from the figure, the load is shifted from the timewhen the
energy price is high (t [14, 17]) to when the priceis low (t [19,
21]). The total consumed energy with loadshifting is the same as
without load shifting.
3) Ramping Constraint: Fig. 12 illustrates the effect of
theramping constraint on the objective. As can be seen fromthe
figure, the total cost is nonincreasing with the ramp-ing parameter
rg in both islanded and grid-connected mode.This is because the
stricter the ramping constraint is, the lessavailable power the
diesel generation can provide. It can bealso seen from the figure
that the marginal cost decreasesin both modes as rg increases. In
particular, the cost doesnot decrease much when rg 0.3, showing
that the dieselgeneration supply is relatively sufficient in that
region of rg.Furthermore, the marginal cost in islanded mode
decreasesmuch faster than in grid-connected mode when rg is
small.This is easy to understand as diesel is the main power
sup-ply in islanded mode and a strict ramping constraint onit would
cause more load shedding/shifting that leads to ahigher cost.
-
SHI et al.: DISTRIBUTED OPTIMAL ENERGY MANAGEMENT IN MICROGRIDS
1145
Fig. 12. Effect of rg on the objective.
TABLE ICOST COMPARISON UNDER DIFFERENT
4) Trad-Offs in the Objective: In the objective func-tion of the
optimization, there are several parameters(g, b, l, 0, p). Each is
associated with one cost mini-mization in the optimization. To
evaluate their effects on theproposed EMS, we conduct simulations
using different . Wechoose the baseline as the set of parameters
used in our previ-ous simulation. We then change the parameter one
at a timeand compare the individual costs as shown in Table I.
As can be seen from the table, the parameter affectsthe
trade-offs among different cost minimizations. A large would
decrease the corresponding cost in the optimization. Thechoice of
depends on the importance of the correspondingcost minimization in
energy management. For example, if themicrogrid operator is more
interested in minimizing the gen-eration cost, g can be increased
and the resulting generationcost would be decreased.
VI. CONCLUSIONIn this paper, we propose a distributed EMS for
the optimal
operation of microgrids. Compared with the existing dis-tributed
approaches, our proposed EMS considers the underly-ing power
distribution network and the associated constraints.Specifically,
we formulate microgrid energy management asan OPF problem and
propose a distributed EMS where theMGCC and the LCs jointly compute
an optimal schedule. Wealso provide an implementation of the
proposed EMS basedon the IEC 61850 standard. As one demonstration,
we applythe proposed EMS to a real microgrid in Guangdong
Province,China. The simulation results demonstrate that the
proposedEMS is effective in both islanded and grid-connected
mode.It is shown that the proposed distributed algorithm
convergesfast. A comprehensive analysis of its performance is
given.Future work includes implementing the proposed EMS in areal
system and analyzing its performance.
APPENDIXIntroduction to PCPM
In this paper, we develop a distributed EMS using the
PCPMalgorithm [23]. PCPM is a decomposition method for
solvingconvex optimization problem. At each iteration, it
computestwo proximal steps in the dual variables and one proximal
stepin the primal variables. We give a very brief description of
thePCPM algorithm below.
Consider a convex optimization problem with separablestructure
of the form
minxX ,yY
f (x) + g(y) (28)s.t. Ax + By = c. (29)
Let z be the Lagrangian variable for the constraint (29).The
steps of the PCPM algorithm to solve the problem are
given as follows.1) Initially set k 0 and choose the initial
(x0, y0, z0)
randomly.2) For each k 0, update a virtual variable zk := zk
+
(Axk + Byk c) where > 0 is a constant step size.3) Solve
xk+1 = arg minxX
{f (x) +
(zk
)TAx + 1
2
x xk
2},
yk+1 = arg minyY
{g(y) +
(zk
)TBy + 1
2
y yk
2}.
4) Update zk+1 := zk + (Axk+1 + Byk+1 c).5) k k + 1, and go to
step 2 until convergence.Steps 2 and 4 can be seen as a predictor
step and a corrector
step to the Lagrange multiplier, respectively. It has been
shownin [23] that the above algorithm will converge to a
primal-dual optimal solution (x, y, z) for a sufficient small
positivestep size as long as strong duality holds for the
convexproblem (28).
ACKNOWLEDGMENTThe authors would like to thank Y. Dong from
the
Department of Electrical Engineering, Tsinghua University,
forproviding the microgrid data, and N. Li from the Departmentof
Electrical Engineering, Harvard University, for
insightfuldiscussions and comments.
REFERENCES[1] W. Shi, X. Xie, C.-C. Chu, and R. Gadh, A
distributed optimal
energy management strategy for microgrids, in Proc. IEEE Int.
Conf.Smart Grid Commun. (SmartGridComm), Venice, Italy, Nov.
2014,pp. 206211.
[2] F. Katiraei, R. Iravani, N. Hatziargyriou, and A. Dimeas,
Microgridsmanagement, IEEE Power Energy Mag., vol. 6, no. 3, pp.
5465,May 2008.
[3] S. Choi, S. Park, D.-J. Kang, S.-J. Han, and H.-M. Kim, A
micro-grid energy management system for inducing optimal demand
response,in Proc. IEEE Int. Conf. Smart Grid Commun.
(SmartGridComm),Brussels, Belgium, Oct. 2011, pp. 1924.
[4] C. Cecati, C. Citro, and P. Siano, Combined operations of
renewableenergy systems and responsive demand in a smart grid, IEEE
Trans.Sustain. Energy, vol. 2, no. 4, pp. 468476, Oct. 2011.
-
1146 IEEE TRANSACTIONS ON SMART GRID, VOL. 6, NO. 3, MAY
2015
[5] S. Pourmousavi, M. Nehrir, C. Colson, and C. Wang, Real-time
energymanagement of a stand-alone hybrid wind-microturbine energy
systemusing particle swarm optimization, IEEE Trans. Sustain.
Energy, vol. 1,no. 3, pp. 193201, Oct. 2010.
[6] P. Siano, C. Cecati, H. Yu, and J. Kolbusz, Real time
operation ofsmart grids via FCN networks and optimal power flow,
IEEE Trans.Ind. Informat., vol. 8, no. 4, pp. 944952, Nov.
2012.
[7] A. Dimeas and N. Hatziargyriou, Operation of a multiagent
sys-tem for microgrid control, IEEE Trans. Power Syst., vol. 20,
no. 3,pp. 14471455, Aug. 2005.
[8] Z. Wang, K. Yang, and X. Wang, Privacy-preserving energy
schedul-ing in microgrid systems, IEEE Trans. Smart Grid, vol. 4,
no. 4,pp. 18101820, Dec. 2013.
[9] A. Dominguez-Garcia and C. Hadjicostis, Distributed
algorithms forcontrol of demand response and distributed energy
resources, in Proc.IEEE Conf. Decis. Control Eur. Control (CDC),
Orlando, FL, USA,Dec. 2011, pp. 2732.
[10] Y. Zhang, N. Gatsis, and G. Giannakis, Robust energy
managementfor microgrids with high-penetration renewables, IEEE
Trans. Sustain.Energy, vol. 4, no. 4, pp. 944953, Oct. 2013.
[11] E. Crisostomi, M. Liu, M. Raugi, and R. Shorten,
Plug-and-play dis-tributed algorithms for optimized power
generation in a microgrid,IEEE Trans. Smart Grid, vol. 5, no. 4,
pp. 21452154, Jul. 2014.
[12] W. Shi, N. Li, X. Xie, C.-C. Chu, and R. Gadh, Optimal
residentialdemand response in distribution networks, IEEE J. Sel.
Areas Commun.,vol. 32, no. 7, pp. 14411450, Jul. 2014.
[13] S. H. Low, Convex relaxation of optimal power flowPart
I:Formulations and equivalence, IEEE Trans. Control Netw. Syst.,
vol. 1,no. 1, pp. 1527, Mar. 2014.
[14] S. H. Low, Convex relaxation of optimal power flowPart
II:Exactness, IEEE Trans. Control Netw. Syst., vol. 1, no. 2, pp.
177189,Jun. 2014.
[15] N. Li, L. Chen, and S. H. Low, Exact convex relaxation of
OPF forradial networks using branch flow model, in Proc. IEEE Int.
Conf.Smart Grid Commun. (SmartGridComm), Tainan, Taiwan, Nov.
2012,pp. 712.
[16] L. Gan, N. Li, U. Topcu, and S. H. Low, On the exactness of
con-vex relaxation for optimal power flow in tree networks, in
Proc. IEEEConf. Decis. Control Eur. Control (CDC), Maui, HI, USA,
Dec. 2012,pp. 465471.
[17] L. Gan, N. Li, U. Topcu, and S. H. Low, Optimal power flow
in treenetworks, in Proc. IEEE Conf. Decis. Control Eur. Control
(CDC),Florence, Italy, Dec. 2013, pp. 23132318.
[18] L. Gan, N. Li, U. Topcu, and S. H. Low, Exact con-vex
relaxation of optimal power flow in radial networks, IEEETrans.
Autom. Control. [Online]. Available:
http://ieeexplore.ieee.org/xpl/articleDetails.jsp?tp=&arnumber=6843918
[19] L. Gan and S. H. Low, Convex relaxations and linear
approximationfor optimal power flow in multiphase radial networks,
arxiv:1406.3054,2014. [Online]. Available:
http://arxiv.org/abs/1406.3054
[20] R. Huang, T. Huang, R. Gadh, and N. Li, Solar generation
predictionusing the ARMA model in a laboratory-level micro-grid, in
Proc. IEEEInt. Conf. Smart Grid Commun. (SmartGridComm), Tainan,
Taiwan,Nov. 2012, pp. 528533.
[21] G. Sideratos and N. Hatziargyriou, An advanced statistical
methodfor wind power forecasting, IEEE Trans. Power Syst., vol. 22,
no. 1,pp. 258265, Feb. 2007.
[22] N. Li, L. Chen, and S. H. Low, Optimal demand response
based onutility maximization in power networks, in Proc. IEEE Power
EnergySoc. Gen. Meeting (PES GM), San Diego, CA, USA, Jul. 2011,
pp. 18.
[23] G. Chen and M. Teboulle, A proximal-based decomposition
methodfor convex minimization problems, Math. Program., vol. 64,
no. 1,pp. 81101, Mar. 1994.
[24] Q. Peng and S. H. Low, Distributed algorithm for
optimalpower flow on a radial network, 2014. [Online].
Available:http://arxiv.org/abs/1404.0700
[25] R. Huang, E.-K. Lee, C.-C. Chu, and R. Gadh, Integration of
IEC 61850into a distributed energy resources system in a smart
green building,in Proc. IEEE Power Energy Soc. Gen. Meeting (PES
GM), NationalHarbor, MD, USA, Jul. 2014, pp. 15.
[26] Communication Networks and Systems for Power
UtilityAutomationPart 7-420: Basic Communication
StructureDistributedEnergy Resources Logical Nodes, IEC Standard
61850-7-420, 2009.
Wenbo Shi (S08) received the B.S. degree fromXian Jiaotong
University, Xian, China, and theM.A.Sc. degree from the University
of BritishColumbia, Vancouver, BC, Canada, in 2009 and2011,
respectively, both in electrical engineering.He is currently
pursuing the Ph.D. degree from theSmart Grid Energy Research
Center, University ofCalifornia, Los Angeles, CA, USA.
His current research interests include demandresponse,
microgrids, and energy management sys-tems.
Xiaorong Xie (M02) received the B.Sc. degreefrom Shanghai Jiao
Tong University, Shanghai,China, and the Ph.D. degree from
TsinghuaUniversity, Beijing, China, in 1996 and 2001,
respec-tively.
He is an Associate Professor with the Departmentof Electrical
Engineering, Tsinghua University. Hiscurrent research interests
include analysis and con-trol of microgrids, and flexible ac
transmissionsystems.
Chi-Cheng Chu received the B.S. degree fromNational Taiwan
University, Taipei, Taiwan,and the Ph.D. degree from the University
ofWisconsinMadison, Madison, WI, USA, in 1990and 2001,
respectively.
He is currently a Project Lead with the SmartGrid Energy
Research Center, University ofCalifornia, Los Angeles, CA, USA. He
is aseasoned Research Manager who supervised andsteered multiple
industry and academia researchprojects in the field of smart grid,
radio frequency
identification technologies, mobile communication, media
entertainment,3-D/2-D visualization of scientific data, and
computer aided design.
Rajit Gadh received the Bachelors degree from theIndian
Institute of Technology, Kanpur, India; theMasters degree from
Cornell University, Ithaca, NY,USA; and the Ph.D. degree from
Carnegie MellonUniversity, Pittsburgh, PA, USA, in 1984, 1986
and1991, respectively.
He is a Professor with the Henry Samueli Schoolof Engineering
and Applied Science, University ofCalifornia, Los Angeles (UCLA),
CA, USA, andthe Founding Director of the UCLA Smart GridEnergy
Research Center, Los Angeles. His current
research interests include smart grid architectures, smart
wireless communica-tions, sense and control for demand response,
microgrids and electric vehicleintegration into the grid, and
mobile multimedia.
/ColorImageDict > /JPEG2000ColorACSImageDict >
/JPEG2000ColorImageDict > /AntiAliasGrayImages false
/CropGrayImages true /GrayImageMinResolution 200
/GrayImageMinResolutionPolicy /OK /DownsampleGrayImages true
/GrayImageDownsampleType /Bicubic /GrayImageResolution 300
/GrayImageDepth -1 /GrayImageMinDownsampleDepth 2
/GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages true
/GrayImageFilter /DCTEncode /AutoFilterGrayImages false
/GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict >
/GrayImageDict > /JPEG2000GrayACSImageDict >
/JPEG2000GrayImageDict > /AntiAliasMonoImages false
/CropMonoImages true /MonoImageMinResolution 400
/MonoImageMinResolutionPolicy /OK /DownsampleMonoImages true
/MonoImageDownsampleType /Bicubic /MonoImageResolution 600
/MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000
/EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode
/MonoImageDict > /AllowPSXObjects false /CheckCompliance [ /None
] /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false
/PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000
0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true
/PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ]
/PDFXOutputIntentProfile (None) /PDFXOutputConditionIdentifier ()
/PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped
/False
/CreateJDFFile false /Description >>>
setdistillerparams> setpagedevice