Electrical Power and Energy Systems 58 (2014) 140149

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

Supervision control for optimal energy cost management

in DC microgrid: Design and simulation
Manuela Sechilariu , Bao Chao Wang, Fabrice Locment
Universit de Technologie de Compigne, AVENUES-GSU EA 7284, BP 60203, rue du Docteur Schweitzer, 60203 Compigne, France

a r t i c l e

i n f o

Article history:
Received 30 January 2013
Received in revised form 7 January 2014
Accepted 18 January 2014

Keywords:
DC microgrid
Energy management
Prediction
Smart grid
Simulation
Supervision

a b s t r a c t
The development of microgrids could facilitate the smart grid feasibility which is conceived to improve
instantaneous grid power balancing as well as demand response. It requires microgrid control functions
as power balancing, optimization, prediction, and smart grid and end-user interaction. In literature, these
aspects have been studied mostly separately. However, combining them together, especially implementing optimization in real-time operation has not been reported. The difculty is to offer resistance to optimization uncertainties in real-time power balancing. To cover the research gap, this paper presents the
supervision design with predicted powers ow optimization for DC microgrid based on photovoltaic
sources, storage, grid connection and DC load. The supervision control, designed as four-layer structure,
takes into account forecast of power production and load power demand, storage capability, grid power
limitations, grid time-of-use tariffs, optimizes energy cost, and handles instantaneous power balancing in
the microgrid. Optimization aims to reduce the microgrid energy cost while meeting all constraints and is
carried out by mixed integer linear programming. Simulation results, show that the proposed control is
able to implement optimization in real-time power balancing with resistance to uncertainties. The
designed supervision can be a solution concerning the communication between loads and smart grid.
2014 Elsevier Ltd. All rights reserved.

1. Introduction
Aiming to avoid grid voltage uctuations [1,2], or even blackout,
at any time instant, the electric grid must balance power between
the production and the consumption with a small margin of error.
The grid capacity is built to satisfy the peak consumption. If the
peak consumption can be shifted during a day, referred to as peak
shaving, the power adjustment, often ensured by excess capacities working in stand-by mode, could be largely reduced. To build
a more robust utility grid, strategies and means of power management are being developed, as well as information on grid needs
and availability [3], which could assist in power balancing by
avoiding undesired injection and performing load shaving during
peak hours. For this, the smart grid is being created to facilitate
information exchange. Smart grid is electric networks that employ
innovative and intelligent monitoring, control communication, and
self-healing technologies to deliver better services for power producers and distributors, exible choices for end-users, reliability
and security of power supply [4,5]. Smart grid is expected mainly
for the following aspects: bidirectional power distribution; bidirec-

Corresponding author. Tel.: +33 344234964; fax: +33 344235262.

E-mail address: manuela.sechilariu@utc.fr (M. Sechilariu).
0142-0615/$ - see front matter 2014 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ijepes.2014.01.018

tional communication, and reduction mismatching between

supply and demand.
The concept of microgrid is proposed for better renewable energy penetration into the utility grid and helps energy management to respond to some grid issues, such as peak shaving, and
reduces energy cost [610]. Microgrids are considered as one of
the possible approaches helping to develop the smart grid [11].
By aggregating loads and multi-source, renewable and traditional,
microgrid can operate in both off-grid and grid-connected conguration. It is generally considered that microgrid controls on-site
generation and power demand to meet the objectives of providing
local power, ancillary services, and injecting power into the utility
grid if required [8]. Concerning microgrid approach, several main
advantages can be given: improving renewable energy penetration
level, facilitating the smart grid implementation, better energy
supply for remote areas, power balancing at local level with selfsupplying possibility, and maintaining load supply during islanding operation or off-grid mode [12]. Thus, the microgrid controller
becomes essential for balancing power and energy management,
and facilitates the sources pooling during islanding.
Depending on the usage of AC or DC bus for coupling different
elements within microgrid, AC microgrid, DC microgrid and hybrid
AC/DC microgrid structures exist [13]. At present, the DC grid is not
ubiquitous [14,15], but more HVDC transmission lines are being
built in MW level, while low voltage DC grid is being adopted,

M. Sechilariu et al. / Electrical Power and Energy Systems 58 (2014) 140149

141

Nomenclature
CG
grid energy cost ()
CLS
load shedding cost ()
CPVL
PV production limitation cost ()
CS
storage energy cost ()
Ctotal
microgrid energy cost ()
cG
grid energy tariff (/kW h)
cNH
grid energy tariff for normal hours (/kW h)
cPH
grid energy tariff for peak hours (/kW h)
cLS
load shedding tariff (/kW h)
cPVL
PV production limitation tariff (/kW h)
cS
storage energy tariff (/kW h)
CP
proportional gain
CREF
storage nominal capacity (Ah)
iPV
PV current (A)
iPV
PV current reference (A)
KD
distribution coefcient
KL
load shedding coefcient
KL_lim
load shedding limit coefcient
p*
power reference (W)
pG
grid power (W)
pG
grid power reference (W)
pG_I
grid injection power (W)
pG_S
grid supply power (W)
pG_I_lim
grid injection power limit (W)
pG_S_lim
grid supply power limit (W)
pG_I_prediction grid injection power prediction (W)
pG_S_prediction grid supply power prediction (W)
pL
load power (W)
pL_D
load power demand (W)
pL_lim
load power limit (W)
pL_max
load maximum power (W)
pL_prediction load power prediction (W)
pPV
PV power (W)

starting with data centers, for the reason of more efciency, less
cost, less occupied space, lower lifetime cost and more reliability
[1618].
Paper [13] presents a three-levels hierarchical control according
to ISA-95 and applied to AC or DC microgrids. This general approach
of hierarchical control for microgrids is conceived for a large-scale
power system, upstream in the utility grid hierarchy. Imitating the
behavior of a grid synchronous generator control, the proposed hierarchical control strategy aims at balancing power between multi
inverters coupled on the same bus without communication, while
controlling the power at the point of common coupling (PCC) at
the same time. The proposed hierarchical control is considered as
a part of the central control and does not take into account the prediction of the power generation and the energy optimization.
In [18], support for autonomous DC microgrid applications is
proposed by integrating the device-level service oriented architecture paradigm into the international standard IEC 61850 applications. In order to create self-manageable microgrid with
semantic-enabled plug-and-play process for distributed energy resources, this solution provides generic middleware platform required for vertical communication. However, the proposed
solution applied to the real microgrid power systems requires
additional control and regulation policy.
A high-level energy management supervision, by means of multi-agent systems, is presented in [19]. In this work, the authors focus on two-level architecture for multiple interconnected
microgrids aiming to manage distributed energy resources in order
to match the buyers and sellers in the energy market.

pPV_lim
PV limited power (W)
pPV lim
PV limited power reference (W)
pPV_MPPT PV MPPT power (W)
pPV_prediction PV power prediction (W)
pS
storage power (W)
pS
storage power reference (W)
pS_C
storage charging power (W)
pS_D
storage discharging power (W)
soc
storage state of charge (%)
SOCmax
SOC upper limit (%)
SOCmin
SOC lower limit (%)
SOC0
initial soc (%)
v
DC bus voltage (V)
v*
DC bus voltage reference (V)
vPV
PV voltage (V)
v PV
PV voltage reference (V)
v PV lim PV limited voltage reference (V)
v PV MPPT PV MPPT voltage reference (V)
vS
storage voltage (V)
Abbreviation
ACR
automatic current regulator
AVR
automatic voltage regulator
HMI
humanmachine interface
MPPT
maximum power point tracking
NH
normal hours
PH
peak hours
PI
proportionalintegral
PV
photovoltaic
P&O
Perturb & Observe
PWM
Pulse Width Modulation

A generalized formulation for intelligent energy management of

a microgrid is proposed in [20] using multiobjective optimization to
minimize the operation cost and the environmental impact. An articial neural network ensemble is developed to predict renewable
energy generation and load demand. In addition, a battery scheduling is proposed as a part of an optimal online energy management,
seen as a decision-making process. However, smart grid data exchanges online or dynamic energy pricing are not considered.
To increase penetration of small PV production into the grid, a
local hierarchical control with energy management is proposed
in [22]. The system is presented as multi-layer control structure,
each layer with a different function, and is based on an optimal
power ow management with predictions, which considers batteries ageing and day-ahead approach into the optimization process.
However, the exchange data with the smart grid, such as limitations of the grid capacity, is not taken into account. Moreover,
due to uncertainty of prediction and lack of grid information, the
grid power could be out of control.
Concerning the energy management two main approaches are
considered: rule-based and optimization-based approaches. Rulebased approach manages the system according to prexed rules,
such as simple rule base, multi-agent system [19] and fuzzy logic
approaches [20,21]. Optimization based approach manages the
system by mathematical optimization, carried out with objective
function and constraints. The optimization methods include the
articial intelligence joint with linear programming [20], linear
programming [21] or dynamic programming [22,23], and genetic
algorithms [24].

142

M. Sechilariu et al. / Electrical Power and Energy Systems 58 (2014) 140149

To sum up the results presented in these works, the rule based

system is simple and robust, but not guarantee the optimal performance with given operating conditions. Moreover, rules become
complex when facing different scenarios. The optimization gives
an optimal solution within given constraints and operation condition. However, optimization is usually treated as separate problem
from the power balancing strategy, and optimization is not implemented in real-time operation. Since optimization are preformed
based on prediction, errors between prediction and real condition
could result in degraded real-time operation by violating certain
constraints, or even result in failure. Thus, optimization and power
balancing should be designed together and coupled with adequate
interface.
Hence, in contrast of the above cited works, in this study the
goal is to design and develop a microgrid local supervision control
which can respond to all following requirements:
to design optimization and power balancing together and
couple them through adequate interface;
to implement optimization in real-time operation with
resistance to uncertainties;
to process prediction data used for optimization;
to exchange data with the smart grid and with the end-user;
to adapt various data in microgrid operation, such as real-time
market pricing and user demand.
Thus, this paper focuses on supervisory control of DC microgrid
which is supposed to manage the power ow in microgrid and
power ow exchanged with the utility grid, with the objective of
making full use of each source while respecting their constraints
on capacity and power. The microgrid based on PV sources, storage,
grid connection and DC load is presented in Section 2. The supervision control, designed as four-layer structure, which is supposed to
exchange data with the smart grid, deal with end-user demand,
forecast of PV production and load consumption is described in
Section 3. Taking into account forecasting data, storage capability,
grid power limitations, grid time-of-use tariffs, the predicted powers ow is optimized by mixed integer linear programming and
solved by CPLEX solver. The DC microgrid control is simulated
for two different cases, based on optimized powers ow and without optimization; the results are presented in Section 4. Concluding remarks are given in Section 5.

2. Microgrid overview
The microgrid presented in Fig. 1 is suggested for local PV
power generation combined with storage and grid connection,

which feed directly a DC load through their dedicated converters.

Concerning the DC bus, following considerations are assumed:
for the demand side, 90% of a tertiary buildings electrical load is
possible to be DC fed in efciency manner; for the grid side, the
power factor can be controlled at 1 [5].
Aiming to interact with the load environment (user, metadata,
smart grid), a supervision system is added, whose main role is to
balance instantaneous power in power system following an energy
management algorithm. The system operation must keep power
balance while respecting constraints of certain elements. Fig. 2
shows the powers ow in the microgrid, using unidirectional powers whose sign convention is always positive.
The system operates respecting the available storage level and
taking into account the grid connection. In case of insufcient PV
energy toward the load, the system security is ensured thanks to
the grid connection and by storage, as well as the load shedding
program. If any excess PV power, the storage could be charged
and the grid connection gives the possibility to trade it back.
2.1. PV sources control
PV system is controlled either by a maximum power point
tracking (MPPT) method or by an algorithm to output a limited
power pPV lim to protect the storage from overcharging or to maintain grid injection power within imposed constraint. In this study,
the chosen MPPT method is the very well known Perturb & Observe (P&O) [25,26], according to which PV sources produce a
MPPT power pPV MPPT . The PV production could be limited thanks
to the limited power reference pPV lim calculated by the supervision
with the algorithm described in Section 3.4.
The PV power pPV control strategy is shown in Fig. 3. The P&O
algorithm and the PV production limiting algorithm give at the
same time corresponding voltage reference v PV MPPT and v PV lim to
operate PV system. The maximum of these two references is taken
as the PV voltage control reference v PV , which represents the minimum power. Following v PV , the PV system is operated by voltage
and current double closed-loop control via automatic voltage regulator (AVR) and automatic current regulator (ACR).
The output of control, duty cycle of Pulse Width Modulation
(PWM) signal, is given to the power electronic devices to control
the PV system. During the MPPT operation, if a limited power within the MPPT ability is given, the proportionalintegral (PI) controller would increase the v PV lim . For v PV lim > v PV MPPT the v PV lim
reference is taken and the MPPT algorithm is stopped. By constrained power closed loop control, the PI controller controls the
PV power at the limited level. In case of low solar irradiance, the
PV output power ability is less than the limited power reference,
so the PI controller will decrease v PV lim until the lower limit, and
v PV MPPT is taken to control the PV system. So, the limited power

SUPERVISION SYSTEM
System states

pS _ C

Storage

pS

_D

POWER SYSTEM

v
PV power through load
DC Load
PV

Control

DC
*
vPV

DC

pG*

DC

DC
AC

pS*

DC

KL

PV
Sources

pG _ S

pG _ I
PV
Sources

Grid
connection

Storage

Fig. 1. DC microgrid overview.

Grid
connection
Fig. 2. Powers ow representation.

DC Load

M. Sechilariu et al. / Electrical Power and Energy Systems 58 (2014) 140149

143

Supervision algorithm

power pL, to keep power balance. The load limited power pL_lim is
controlled by the load coefcient KL dened by Eq. (6):

*
pPV
_ lim

K L pL

vPV

STOP
P&O

iPV

*
pPV
_ lim

pPV
+-

*
vPV

PI

*
vPV

+-

max

*
vPV
_ lim

AVR

*
iPV

+-

ACR

where pL_max is a constant as contractual subscribed maximum

power. It is supposed that the load could not be supplied with a
power that exceeds this limit. Given the constant denominator
pL_max, therefore KL e [0, 1] changes according to the available
power for supplying the load. If pL_D > pL_lim, the load would be shed
within the limit level. These operations should be controlled by the
supervision system. Temporarily load partial shedding could be a
solution to reduce utility grid mismatching, or to obtain less energy
consumption, within the limit agreed by end-user.

*
vPV
_ MPPT

lim =pL max

PWM

2.5. Power balancing principle

According to the powers sign convention (powers always
positive), the physical law of power balancing is described by:

Fig. 3. PV system control strategy.

control does not affect MPPT algorithm and MPPT power is

produced. Experiment validation of the control is provided in [27].
2.2. Storage control
Lead-acid batteries are selected as storage for the DC microgrid
applied to building, because of relatively low cost and mature technology [28]. The storage is operated by current closed-loop control,
and the storage power is controlled by supervision system which
calculates the corresponding power reference. The storage state
of charge soc must be respected to its upper and lower limitations,
SOCmax and SOCmin respectively, to protect the storage from overcharging and over discharging, as described by Eq. (1). The soc is
calculated by Eq. (2), with SOC0 as initial soc at t0, CREF as storage
nominal capacity (Ah) and vS as storage voltage. When the soc limit
is not reached, the PV production should not be limited, as in Eq.
(3).

SOC min 6 soct 6 SOC max

soct SOC 0
pPV

lim t

pPV

1
3600  v S  C REF

MPPT t

1
t

pS C t  pS D tdt

t0

if soct < SOC max

2.3. Grid connection control

The grid connection is controlled by current closed-loop control.
The grid power is controlled by supervision system which calculates the corresponding power reference. Furthermore, as messages transmitted by the smart grid via supervision system,
limits for grid supply power pG_S_lim and grid injection power
pG_I_lim could be imposed. By these two limitations, grid problems,
such as performing peak shaving, avoiding undesired injection or
downscaling injection uctuations caused by intermittent PV productions, can be improved. During microgrid operation, the grid
power should be controlled to satisfy Eqs. (4) and (5).

0 6 pG I t 6 pG

I lim

0 6 pG S t 6 pG

S lim

2.4. DC load control

The load power demand pL_D should be satised; nevertheless,
in case of insufcient storage and grid access limits, pL_D cannot
be fully met, and the load must be partially shed, forming load

pL t pG I t pS C t pG S t pS D t pPV t

with pL(t) = min (pL_D(t), pL_lim(t)) and pPV(t) = min (pPV_MPPT(t),

pPV_lim(t)).
Fluctuations in the DC bus voltage, which is noted v, are caused
by the difference between load consumption and PV generation.
The required power reference p* for power balancing is calculated
by regulating v with a proportional controller as in Eq. (8):

p t pPV t  pL t  C P v  t  v t

where v is the DC bus voltage control reference and CP is the proportional gain. For stabilizing the DC bus voltage, power balance
in the system is performed by adjusting storage and grid power.
Thus, p* is shared by the storage and the grid as in Eq. (9)

p t pG t pS t

with pG(t) = pG_I(t)  pG_S(t) and pS(t) = pS_C(t)  pS_D(t).

After calculating p*, the grid power reference pG and storage
power reference pS are calculated according to a distribution coefcient KD(t), as described in Eqs. (10) and (11).

pS t K D tp t

10

pG t p t  pS t

11

The power balancing of studied DC microgrid operation can work

with any KD value. To improve efciency and reduce energy cost,
optimization calculations using specically metadata (power predictions, power limits, real-time grid energy tariffs) are done by
the supervision which outputs the distribution coefcient, KD(t),
whose time depending values represent the predicted optimized
powers ow.
3. Supervision control design
The supervision system, proposed in Fig. 4, is designed in fourlayer structure, which consists of humanmachine interface (HMI),
prediction layer that predicts load consumption and PV production, energy management layer that optimizes the predicted powers ow, and operation layer that balance instantaneous power,
based on unique interface parameter KD(t), in power system.
The supervision control takes into consideration prediction data
and various constraints, such as grid power limits and storage
capacity, which can be fully respected for both optimization and
real operation. Each layer provides an independent function and
thus, the structure is exible and can be implemented in several
microcontrollers or computers so that real-time power balancing
control and complicated optimization can be executed at the same
time without affecting each other. The multi-layer structure

144

M. Sechilariu et al. / Electrical Power and Energy Systems 58 (2014) 140149

Supervision
User demand

K L _ lim

Human-machine interface

Metadata

Wheather
Load scheduling
Weather forecast

Prediction layer

Smart grid messages

Load management data

Energy management layer

Operation layer

p L _ prediction
p PV _ prediction

Fig. 6. Prediction layer design.

Power system states

3.3. Energy management layer

Power system
Fig. 4. Design principle of supervision.

simplies implementation of such complex control strategy. Thus,

the control of the power balancing is separated from energy management layer, yet they are linked through one interface parameter
KD(t). On one hand, the energy management layer is able to optimize the microgrid energy cost through predictive data and thus
obtain the predicted optimized power ow which is then translated into KD(t) sequence. On the other hand, the power balancing
control in operation layer is an independent function that can work
with any KD(t) value. The distribution coefcient KD(t) is a single
interface parameter yet represents the power ow from different
sources. Hence, the communication of KD(t) does not need high
speed communication between layers.

3.1. Humanmachine interface layer

The HMI design is presented in Fig. 5. User can specify the lowest limit of the load coefcient, KL_lim. Aiming to maintain power
balance, if the system requires KL < KL_lim, the operation is assumed
with the necessary KL value, but the user will be notied.

3.2. Prediction layer

Energy management layer (Fig. 7) interacts with the prediction

layer and controls the operation layer by calculating the distribution coefcient KD following an optimization method. The optimization goal is to obtain the best power distribution between the
grid and the storage, so to reduce energy cost, grid power peak consumption, load shedding and limiting PV production at the same
time. In this study, the smart grid message is supposed to provide
real-time grid energy tariffs and grid power limits, which assist in
reducing peak supply and avoid undesired injection. Furthermore,
thanks to the smart grid connection, this layer is able to inform the
grid operator of the grid supply power prediction pG_S_prediction and
grid injection power prediction pG_I_prediction.
The energy cost of system Ctotal consists of grid energy cost CG,
storage energy cost CS, PV production limitation cost CPVL and load
shedding cost CLS, as in Eq. (12).

C total C G C S C PVL C LS

By calculating the energy cost for each time duration Dt, CG is

dened by Eq. (13). According to this denition, the grid power
could be bought or sold.

CG

tF
X

1
6

3:6  10

cG ti  Dt  pG I ti pG S t i 

t i t 0

This study takes into account the same price for energy purchased or sold, and the grid energy tariff is dened by Eq. (14).

cNH 0:1=kW h for t 2 normal hoursNH

cPH 0:7=kW h for t 2 peak hoursPH

CS

tF
X

1
6

3:6  10

cS t i  Dt  pS C t i pS D t i 

ti t 0

with t i ft0 ; t 0 Dt; t 0 2Dt; . . . ; tF g and cs t 0:05=kW h

15

{
nvo

Load shedding limit

K L _lim
Fig. 5. HMI layer design.

14

Storage aging should be considered to give an energy tariff of

storage using. For this study, the storage energy cost CS and an arbitrary storage energy tariff are dened in Eq. (15).

Smart grid messages

User demand

13

with t i ft0 ; t 0 Dt; t 0 2Dt; . . . ; tF g

cG t

The prediction layer design is presented in Fig. 6. It calculates

possible PV power and load power evolutions for the next day.
Based on solar irradiance and temperature forecasting, and on PV
model, built with parameters identication or PV solar irradiance
mapping [29,30], the PV predictable power pPV_prediction could be
calculated with as less error as possible.
The load power supply can be predicted by statistical data and/
or by information of load scheduling from building management
system and with respect of user demand [31,32]. The load scheduling, that represents the operating program of building facilities,
building energy needs linked to weather, and the partial and full
load shedding program are supposed known, and they are implemented and updated continuously by the building management
system. According to the value of KL_lim chosen by the user, the load
prediction power pL_prediction can be calculated by this layer.

12

Grid time-of-use tariffs

pG _ S _ lim , pG _ I _ lim

nvi
nvi: network variable input
nvo: network variable output

p L _ prediction
p PV _ prediction

pG _ I _ prediction
pG _ S _ prediction

Fig. 7. Energy management layer design.

pG _ S _ lim
pG _ I _ lim
K D

145

M. Sechilariu et al. / Electrical Power and Energy Systems 58 (2014) 140149

The cost of PV system shedding and an arbitrary PV shedding

tariff are dened in Eq. (16).

C PVL

tF
X

1
6

3:6  10

cPVL t i  Dt  pPV

MPPT t i

 pPV

pG _ S _ lim , pG _ I _ lim
KD

Power system

SOCmin , SOCmax

lim t i 

p PV , p L , soc, pG , v

t i t 0

with t i ft 0 ; t0 Dt; t 0 2Dt; . . . ; t F g and cPVL t 1=kW h

16

*
pG* , pS* , K L , pPV
_ lim

The cost of load shedding and an arbitrary load shedding tariff

are dened in Eq. (17).

C LS

tF
X

3:6  106 ti t0

cLS ti  Dt  pL D t i  pL

lim t i 

with t i ft 0 ; t0 Dt; t 0 2Dt; . . . ; t F g and cLS t 1=kW h

17
Aiming to limit the power grid uctuations, grid power changing rate limits are introduced as:

pG ti  pG t i1 6 Limit
pG ti  pG t i1 P Limit

18

with t i ft 0 ; t0 Dt; t 0 2Dt; . . . ; t F g

As PV energy grid injection benets incentive tariffs, energy grid
injection by power grid charged storage is forbidden. Thus, the following limits are imposed in order to ensure storage energy charge
and grid injection only from PV production.

K L < K L _ lim alert user

Fig. 8. Operation layer design.

ent types of optimization problems. However, any other mixed

integer linear programming solver also can be used (LP_SOLVE,
GUROBI, . . . ). In addition, in order to express our problem in the
syntax of the solver and to call the solving algorithm of the solver,
a procedure written in C++ is used. This procedure output to a le
the optimal powers ow, which is the evolution of pS_D(t), pS_C(t),
pG_S(t), pG_I(t). The estimated optimum powers ow is then translated into a control parameter that is the optimum distribution
coefcient KD(t). Taking into account Eqs. (9) and (10), the optimum KD values are calculated by Eq. (21):

K D t

pS C t  pS D t
pS C t  pS D t pG I t  pG S t

21

3.4. Operation layer

pG ti P 0; pS t i P 0 if pPV t i  pL D t i P 0
pG ti 6 0; pS t i 6 0 if pPV t i  pL D ti < 0
with t i ft 0 ; t0 Dt; t 0 2Dt; . . . ; t F g

19

Finally, by considering the discrete time instant ti, from t0 to tF,

with the time interval Dt, the optimization problem can be completely mathematically expressed by (20):

Minimize C total C G C S C PVL C LS

with respect to :
8
pL ti pG I t i pS C ti pG S ti pS D t i pPV t i
>
>
>
>
>
SOC min 6 soct i 6 SOC max
>
>
>
>
tF
>
X
>
>
>
pS C ti  pS D t i Dt
> soct i SOC 0 3600v1S C REF
>
>
>
ti t0
>
>
>
>
>
< 0 6 pG I t i 6 pG I lim
0 6 pG S t i 6 pG S lim
>
>
>
>
p t i  pG t i1 6 Limit
>
>
> G
>
>
pG t i  pG t i1 P Limit
>
>
>
>
>
pG t i P 0; pS t i P 0 if pPV t i  pL D ti P 0
>
>
>
>
>
> pG t i < 0; pS t i < 0 if pPV ti  pL D t i < 0
>
>
:
pPV MPPT t  pPV lim t 0 if soct < SOC max

According to the energy management layer output (KD(t) and

grid power limits), the operation layer presented in Fig. 8 aims at
balancing power in the power system while meeting all
constraints.
If there is not sufcient power to supply the load (insufcient
PV production when the grid power is limited and the storage is
empty), the load shedding is performed. In such cases, the load
power has to be limited to pL_lim, which is calculated as in Eq. (22).

pL

20

ti ft 0 ; t0 Dt; t 0 2Dt; . . . ; t F g
Energy optimization are usually solved by linear programming [21]
or dynamic programming [22,33] technique. Dynamic programming can solve non-linear problem, while linear programming
solves problems satisfying linear forms. Linear programming can
be more efciently solved with less time and memories. It can be
veried that, mathematical formulation given by Eq. (20) follows
standard linear programming form, except for the last constraint
which does not take a linear form. However, this constraint can
be easily linearized by introducing for each time point ti, a variable
array which is integer; this is why the optimization formulation is a
mixed integer linear program [34].
In this study, the optimization problem is solved using the IBM
ILOG CPLEX solver [35], which is a powerful tool for solving differ-

lim

pPV pG

S lim

22

The system compares the actual load power pL with the load
power limit pL_lim; if pL > pL_lim the load would be shed within the
limit level. When the storage is full and the grid injection limit
does not permit absorbing all excess of PV production, the PV limited production is performed by calculating pPV lim :

pPV

lim

pL pG

I lim

23

To sum up the above described power balancing, and taking

into consideration the limits of storage and grid given in Eqs. (1),
(3), (4), and (5), the overall algorithm of the operation layer is
shown in Fig. 9.
Experiment validation of the technical feasibility of the operation layer for a constant value of KD is provided in [15]. As the prediction layer requires available and reliable forecast data, for the
PV sources geolocation and for few hours ahead, in this study,
the optimized DC microgrid is tested by simulation.
4. Simulation results
The microgrid control is simulated for an operation on 23rd of
April 2011 in Compiegne, France. The objective of this study is
more to validate a comprehensive approach rather than purely
numerical results. For this reason we do not give the numerical values of various system components studied, yet powers values are
based on our experimental multisource system platform
[12,15,36]. The simulation results of the DC microgrid were obtained under the MATLAB-Simulink. All the implemented auto-

146

M. Sechilariu et al. / Electrical Power and Energy Systems 58 (2014) 140149

Read from the energy management layer: K D , pG _ S _ lim , pG _ I _ lim

Read from the power system: v, pG , pPV , pL , soc, SOCmin , SOCmax
p* = pPV pL CP (v* v)
Yes

Yes

No

soc SOCmin

pS* = 0

No

p* 0

Yes

pS* = 0

pS* = K D p*

pG* = p* pS* with pG* _ S = pG*

Yes
*
G_S

pG* _ S pG _ S _ lim

No

soc SOCmax

pS* = K D p*

pG* = p* pS* with pG* _ I = pG*

No

Yes

No

pG* _ I pG _ I _ lim

p = pG _ I _ lim

= pG _ S _ lim

*
G

Update load shedding

control

Yes

soc SOCmin
No

pL _ lim = pPV + pG _ S _ lim

Yes

soc SOCmax
K L = pL _ lim pL _ max

Update PV limited power

reference

No
*
pPV
_ lim = pPV _ MPPT

*
pPV
_ lim = pL + pG _ I _ lim

*
pPV
_ lim = pPV _ MPPT

KL = 1

KL = 1

Fig. 9. Flowchart of operation layer control.

1500

(a)

P PV_MPPT_measurement

P PV_hourly average_

(W)

1000

0
9:00

10:00

11:00

12:00

13:00

14:00

15:00

16:00

17:00

18:00

Time
1500

(b)

PPV_measurement
MeasurePPV_prediction
Prediction

(W)

1000
PV

Nowadays services can be found providing worldwide PV power

forecast according to location and weather information. However,
public data of solar irradiance (W/m2) forecasting are not yet precise enough for a specied location. This is why is very difcult to
calculate the PV power prediction following real solar irradiance
forecasting. The measured PV MPPT power, whose values have
been recorded by our experimental device, is shown in Fig. 10a:
the green curve shows the real time PV power evolution, while
the gray bars are the hourly average PV power.
As this paper aims at demonstrating feasibility of the designed
supervision, to overcome the lack of real solar irradiance forecasting, the PV power prediction data are calculated from the real measurement data. PV power hourly prediction data given in Fig. 10b
are assumed having random 10% error with the hourly average
of the measurement data.
Load prediction data are supposed to be given by load management system, which implies additional uncertainties. In this study,
a simple arbitrary load power evolution is considered. The difference of load power and load power prediction is shown in
Fig. 11. In order to perform an optimized operation for the next
day, prediction layer is supposed to give to the energy management layer the forecasts of the PV power and the load power
hourly evolutions.
Peak hours during the day are assumed 11:0013:00 and
16:0018:00. Grid and storage constraints, as arbitrary values,

500

4.1. Optimization results

PV

matic controls, described earlier, are working satisfactorily. The

proportional controller, whose proportional gain CP is used in Eq.
(8), provides a wide control bandwidth adapting to simulation
step. However, synthesis for parameter tuning is not realized in
this work.

500

0
9:00

10:00

11:00

12:00

13:00

14:00

15:00

16:00

17:00

18:00

Time
Fig. 10. PV MPPT power evolution (a), hourly PV power measurement and
prediction (b).

are imposed for the system operation. To mitigate grid power

strong uctuations, grid power changing limit is imposed as
20 W/s. Arbitrary soc limits are considered as 45% and 55%, while
SOC0 = 50% and CREF = 130 Ah. Considering the storage capacity of

147

M. Sechilariu et al. / Electrical Power and Energy Systems 58 (2014) 140149

1500

age has to be proposed to supply the load. During the peak hours
the soc decreases continually, as shown in Fig. 12b. Hence, the storage is used to supply the load as much as possible with respect to
its soc lower limit. However, as the storage energy is not enough,
the grid power is also used to supply the load. This is why, during
peak hours, storage power and grid power are proposed to share
the necessary power to supply the load in an optimized manner
while respecting all constraints. This sharing power is proposed
in an intermittent manner by the used solver.

PL_measurement
PL_prediction
Measure
Prediction

(W)

1000

500

0
9:00

10:00

11:00

12:00

13:00

14:00

15:00

16:00

17:00

4.2. Powers ow simulation controlled by KD(t) optimum evolution

18:00

Time
Fig. 11. Load power measurement and day-ahead prediction.

our experimental platform, whose parameters values are used in

the simulation study, these soc limits are selected to show the system behavior with relevant storage events (full, empty) in a day
run. Grid power injection limit is imposed as 700 W, and supply
limit as 600 W. Based on the prediction information, the energy
management layer calculates the optimization problem by CPLEX
and gives the optimum powers ow evolution, as presented in
Fig. 12a. Corresponding KD(t) sequence is calculated by Eq. (21)
from the optimum powers ow evolution, as presented in
Fig. 12b. For performing a day optimization during 9 h, the data
resolution is chosen at 10 s/point, i.e. 3240 points each power
curve. The optimization program execution time is within 10 s
for a computer with CORE i5 processor. The optimum cost estimated by CPLEX is 0.517 .
Fig. 12 shows the optimization calculation results. It can be observed that for peak hours the PV production is not sufcient to
supply the load; so, the storage and the grid have to supply the
load for the remain part of power. The considered optimization
problem is formulated to minimize the global energy cost for the
whole period from 9:00 to 18:00, while respecting all constraints.
So, as for the peak hours the grid energy tariff is very high and largely superior to storage energy tariff, it seems normal that the stor-

The calculated KD(t) optimum evolution is given to the operation layer to run the power system following the conditions given
by 23rd of April 2011 (meteorological and load). The operation
powers ow, as real situation, simulated by MATLAB, is shown in
Fig. 13a. During this day operation, grid and storage share power
for supplying energy or for receiving energy at the same time. In
the rst off-peak hours (9:0011:00), grid mainly supplies the load
for reserving storage for peak hour supply.
During the rst peak hours (11:0013:00), the load is supplied
by storage and grid, the sharing proportion is determined by optimization calculation which aims also to reserve storage for supplying during second period of peak hours.
Just before 13:00 the surplus of the PV production is injected
into the grid in order to make the maximum prot. Aiming to reduce the energy cost by avoiding grid to supply during peak hours,
in the second peak hour period (16:0018:00), the storage is
mainly used for supplying the load. During 13:0015:00 with the
excess PV production, the storage is charged for supplying in the
second peak hour. Grid power injection limit and supply limit
are respected. Short time load shedding can be seen in the operation after 17:00, when the battery is empty. The load shedding is
performed based on instantaneous power information. To avoid
load shedding uctuations in PV power uctuating circumstances,
it is also possible to impose duration for load shedding in CPLEX
optimization, and optimized load shedding information could be

2000

2000

(a)

pPV

pG_I

pG_S

pS_C

pS_D

pPV_MPPT

pL

(a)

1500

Peak hours

Power (W)

1500

1000

pL_D

pPV_MPPT

pG_I

pG_S

pS_C

pS_D

pPV

pL

Peak hours

1000

500

500
0
9:00

10:00

11:00

12:00

13:00

14:00

15:00

16:00

17:00

16:00

56
soc 55
54
53
52
51
50
49
48
47
46
45
44
17:00 18:00

(b)

10:00

12:00

13:00

14:00

15:00

12:00

13:00

14:00

15:00

430
425
420
415
410
405
400
395
390
385
380
375
370
9:00

(b)

(V)

KD

11:00

11:00

16:00

17:00

18:00

Time

Time

1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
9:00

10:00

18:00

10:00

DC bus voltage

11:00

12:00

13:00

14:00

15:00

16:00

17:00

56
55
54
53
52
51
50
49
48
47
46
45
44
18:00

soc

9:00

Power (W)

pL_D

Time

Time
Fig. 12. Optimized powers ow (a), optimized KD(t) and soc evolution (b).

Fig. 13. Simulated powers ow (a), DC bus voltage and soc evolution (b) for
optimum KD(t).

148

M. Sechilariu et al. / Electrical Power and Energy Systems 58 (2014) 140149

given to operation layer to override the operation layer load shedding control. The energy cost is 0.512 , which is close to the optimization cost. The soc evolution with the optimum KD(t) and the
DC bus voltage are illustrated in Fig. 13b. The DC bus voltage uctuations are negligible compared to the value of 400 V, signifying
the power is well balanced.
By comparing the results presented in Figs. 12 and 13, it can be
seen that the simulated powers ow is slightly different from the
optimization due to the uncertainties of solar irradiance prediction
and load power prediction. During the solar uctuations between
16:00 and 17:00, the storage provided more powers. However, in
this case, the storage is still able to be the main load supply during
the second period of peak hours, but a slight load shedding occurs
when the soc reaches its low limit.
4.3. Powers ow simulation controlled by constant KD
In order to further analyze, a simulation case for a constant KD is
presented in Fig. 14. It is chosen as constant value KD = 0.5885
which is the average value of optimum KD(t) evolution shown in
Fig. 12b. In this case, the obtained energy cost, 0.652 , the difference with optimization is larger compared with using optimal
KD(t), and longer load shedding can be seen during this operation.
The soc evolution, illustrated in Fig. 14b, is very different from the
optimum soc evolution shown in Fig. 13b. Even optimization effect
is affected, the power balancing is robust. Regarding the DC bus
voltage illustrated in Fig. 14b, it can be seen that the DC bus voltage
remains stable with very slight uctuations, signifying the power is
well balanced.
4.4. Simulation results: comparison and discussion
Table 1 shows the energy cost of the microgrid Ctotal given by Eq.
(12) and occurrences of load shedding for these three cases: optimized operation by energy management layer with 10% uncertainties prediction data, simulated operation in case of a real PV
production with the calculated optimum KD(t), and simulated

2000

(a)

Power (W)

1500

pL_D

pPV_MPPT

pG_I

pG_S

pS_C

pS_D

pPV

11:00

12:00

13:00

14:00

15:00

16:00

17:00

18:00

Time

(V)

(b)

10:00

DC bus voltage

11:00

12:00

13:00

14:00

15:00

16:00

17:00

56
55
54
53
52
51
50
49
48
47
46
45
44
18:00

soc

430
425
420
415
410
405
400
395
390
385
380
375
370
9:00

KD

Ctotal cost
()

Load shedding
duration

Optimization

Optimum KD(t) as given in

Fig. 12b
Optimum KD(t) as given in
Fig. 12b
KD = 0.5885

0.517

No shedding

0.512

7 min

0.652

37 min

Simulation
Simulation

operation with constant KD = 0.5885. It can be seen that simulated

cost for the optimum KD(t) is close to the prediction cost, and the
error is within 1%, which is from the uncertainties of prediction.
For the constant KD case, the cost is 26% more than the prediction case; moreover, longer load shedding can be seen in this case.
Even with some uncertainties, the optimized KD(t) operates the
microgrid with respect of the utility grid requirements and storage
capacity. This comparison validates the presented simulation case
for the proposed supervision of the proposed DC microgrid.
As previously mentioned, the communication of KD(t) does not
need high speed communication between layers. So, the supervision control provides the possibility of re-performing optimization
and updating the KD(t) sequence during the actual operation without interrupting the power balancing. Thus, hourly or more
frequent optimization that updates KD(t) sequence, with latest prediction and power system status, is expected to give better energy
performance of the supervision control.
However, the limit of the supervision control is that optimizing
effectiveness is affected on the prediction precision. Predictions
uncertainties do not inuence power balancing but the optimal energy cost is affected. Future research should focus on enhancing
optimization performance, especially facing low prediction precision. An optimization technique that able to optimize power ow
with consideration on uncertainties of the prediction combined
with a rule based algorithm in operation layer that corrects KD(t)
in real time with respect of power system status can be developed
as one solution. Besides, a second storage can be installed as backup for correcting the errors between optimized power and real
operation.
5. Conclusions

Peak hours

500

10:00

Case
operation

pL

1000

0
9:00

Table 1
Comparison of different cases.

Time
Fig. 14. Simulated powers ow (a), DC bus voltage and soc evolution (b) for
constant KD = 0.5885.

Facing the advent of smart grid context, a microgrid control

combining power balancing, optimization and smart grid interaction is proposed through a multi-layer supervision control structure. The research issue of implementing optimization in realtime operation is particularly addressed.
Based on PV sources, storage, power grid connection and DC
load, the microgrid aims at self supply with limited access to grid.
Taking into account forecast of PV production and load power demand, the four-layer supervision system performs optimization
and implements optimization in instantaneous power balancing
through simple interface. It handles also constraints such as storage capability, grid power limitations, energy grid tariffs, grid peak
hours. The optimization is based on mixed integer linear programming, solved by CPLEX. Simulation results, even with uncertainties
of prediction and arbitrary energy tariffs, taken as assumptions in
this study, show that the proposed supervision design is able to
perform efciency and cost effective powers ow in real-time
operation with respect to constraints such as grid power limits
and storage capacity. Load shedding and PV power limiting ensures
power balancing in any case. The simulation shows that the optimization gives better energy performance while minimizing load
shedding and PV production limitation and the operation layer respects all constrains of power system elements.

M. Sechilariu et al. / Electrical Power and Energy Systems 58 (2014) 140149

On the other hand, the optimization efciency is based on prediction precision, which may limit the nal performance. The
designed operation layer can work with any KD value, so the
prediction errors and non-optimum KD does not affect the power
balance. However, if uncertainties are higher than 10%, the energy
cost and the load shedding duration could be severely affected. The
developed simulation will permit in further work to design an
additional supervision layer aiming to mitigate the differences
between the optimized powers ow and the real one.
To sum up, the feasibility of the proposed DC microgrid supervision control structure, that combines grid interaction and energy
management with power balancing, is proved by simulation results. Experimental test is going be carried out in real conditions
once the PV power prediction service is ready. Although the microgrid only refers to a building scale and involves only a few sources,
the idea of parameterize power balancing and interfacing with
optimization, as well as smart grid interaction, can be generalized
and thus can be used as solution for advanced energy management
for other microgrids to optimize local power ow and improve
future PV penetration.
References
[1] Lee TL, Hu SH, Chan YH. D-STATCOM with positive-sequence admittance and
negative-sequence conductance to mitigate voltage uctuations in high-level
penetration of distributed-generation systems. IEEE Trans Ind Electron
2013;60(4):141728.
[2] Wei-Lin H, Chia-Hung L, Chao-Shun C, Hsu CT, Te-Tien K, Cheng-Ta T, et al.
Impact of PV generation to voltage variation and power losses of distribution
systems. In: Proc 4th international conference on electric utility deregulation
and restructuring and power technologies; 2011. p. 14748.
[3] Liserre M, Sauter T, Hung JY. Future energy systems, integrating renewable
energy sources into the smart power grid through industrial electronics. IEEE
Ind Electron Mag 2010;4(1):1837.
[4] Peeters E, Belhomme R, Batlle C, Bouffard F, Karkkainen S, Six D, et al.
ADDRESS: scenarios and architecture for active demand development in the
smart grid of the future. In: Proc of CIRED 20th international conference on
electricity distribution; 2009. p. 14.
[5] Sechilariu M, Wang BC, Locment F. Building-integrated microgrid: advanced
local energy management for forthcoming smart power grid communication.
Energy Build 2013;59(1):23643.
[6] Lasseter RH, Eto JH, Schenkman B, Stevens J, Vollkommer H, Klapp D, et al.
CERTS microgrid laboratory test bed. IEEE Trans Power Delivery
2010;26(1):253140.
[7] Hatziargyriou N, Asano H, Iravani R, Marnay C. Microgrids. IEEE Power Energy
Mag 2007:7894.
[8] Guerrero JM, Chandorkar M, Lee TL, Loh PC. Advanced control architectures for
intelligent microgridsPart I: decentralized and hierarchical control. IEEE
Trans Ind Electron 2013;60(4):160718.
[9] Georgilakis PS. Integration of Distributed Generation in the Power System, M.
Bollen, F. Hassan. WileyIEEE Press, New Jersey (2011). Int J Electr Power
Energy Syst 2013;48:6970.
[10] Alvarez E, Campos AM, Arboleya P, Gutirrez AJ. Microgrid management with a
quick response optimization algorithm for active power. Int J Electr Power
Energy Syst 2012;43(1):46573.
[11] Lasseter RH. Smart distribution: coupled microgrids. Proc IEEE
2011;99:107482.
[12] Sechilariu M, Wang BC, Locment F. Building integrated photovoltaic system
with energy storage and smart grid communication. IEEE Trans Ind Electron
2013;60(4):160718.

149

[13] Guerrero JM, Vasquez JC, Matas J, de Vicuna LG, Castilla M. Hierarchical control
of droop-controlled AC and DC microgridsa general approach toward
standardization. IEEE Trans Ind Electron 2011;58(1):15872.
[14] Shenai K, Shah K. Smart DC micro-grid for efcient utilization of distributed
renewable energy. In: Proc of IEEE Energytech; 2011. p. 16.
[15] Wang BC, Sechilariu M, Locment F. Intelligent DC microgrid with smart grid
communications: control strategy consideration and design. IEEE Trans Smart
Grid 2012;3(4):214856.
[16] AlLee G, Tschudi W. Edison Redux: 380 Vdc brings reliability and efciency to
sustainable data centers. IEEE Power Energy Mag 2012;10:509.
[17] Patterson BT. DC, come home: DC microgrids and the birth of the Enernet.
IEEE Power Energy Mag 2012;10:609.
[18] Sucic S, Havelka JG, Dragicevic T. A device-level service-oriented middleware
platform for self-manageable DC microgrid applications utilizing semanticenabled distributed energy resources. Int J Electr Power Energy Syst
2014;54:57688.
[19] Kumar Nunna HSVS, Doolla S. Multiagent-based distributed-energy-resource
management for intelligent microgrids. IEEE Trans Ind Electron
2013;60(1):167887.
[20] Chaouachi A, Kamel RM, Andoulsi R, Nagasaka K. Multiobjective intelligent
energy management for a microgrid. IEEE Trans Ind Electron
2013;60(1):168899.
[21] Chakraborty S, Weis MD, Simoes MG. Distributed intelligent energy
management system for a single-phase high-frequency AC microgrid. IEEE
Trans Ind Electron 2007;54:97109.
[22] Riffonneau Y, Bacha S, Barruel F, Ploix S. Optimal power ow management for
grid connected PV systems with batteries. IEEE Trans Sustain Energy
2011;2(3):32532.
[23] Bo G, Mills JK, Dong S. Energy management control of microturbine-powered
plug-in hybrid electric vehicles using the telemetry equivalent consumption
minimization strategy. IEEE Trans Vehicular Technol 2011;60:423848.
[24] Bustos C, Watts D, Ren H. MicroGrid operation and design optimization with
synthetic wind and solar resources. IEEE Trans Latin America
2012;10:155062.
[25] Houssamo I, Locment F, Sechilariu M. Maximum power tracking for
photovoltaic power system: development and experimental comparison of
two algorithms. Renew Energy 2010;35(10):23817.
[26] Houssamo I, Locment F, Sechilariu M. Experimental analysis of impact of MPPT
methods on energy efciency for photovoltaic power systems. Int J Electr
Power Energy Syst 2013;46:98107.
[27] Wang BC, Houssamo I, Sechilariu M, Locment F. A simple PV constrained
production control strategy. In: Proc of IEEE international symposium on
industrial electronics; 2012. p. 96974.
[28] Tan X, Li Q, Wang H. Advances and trends of energy storage technology in
microgrid. Int J Electr Power Energy Syst 2013;44(1):17991.
[29] Houssamo I, Wang BC, Sechilariu M, Locment F, Friedrich G. A simple
experimental prediction model of photovoltaic power for DC microgrid. In:
Proc of IEEE international symposium on industrial electronics; 2012. p. 9638.
[30] Lorenz E, Hurka J, Heinemann D, Beyer HG. Irradiance forecasting for the
power prediction of grid-connected photovoltaic systems. IEEE J Select Topics
Appl Earth Observ Remote Sens 2009;2:210.
[31] Amjady N, Keynia F, Zareipour H. Short-term load forecast of microgrids by a
new bilevel prediction strategy. IEEE Trans Smart Grid 2010;1:28694.
[32] Ren P, Xiang Z, Qiu Z. Intelligent domestic electricity management system
based on analog-distributed hierarchy. Int J Electr Power Energy Syst
2013;46:4004.
[33] Fuselli D, De Angelis F, Boaro M, Squartini S, Wei Q, Liu D, et al. Action
dependent heuristic dynamic programming for home energy resource
scheduling. Int J Electr Power Energy Syst 2013;48:14860.
[34] Franco JF, Rider MJ, Lavorato M, Romero R. A mixed-integer LP model for the
optimal allocation of voltage regulators and capacitors in radial distribution
systems. Int J Electr Power Energy Syst 2013;48:12330.
[35] IBM.com. IBM ILOG CPLEX Optimizer. <http://www.ibm.com>.
[36] Locment F, Sechilariu M, Houssamo I. DC load and batteries control limitations
for photovoltaic systems. Experimental validation. IEEE Trans Power Electron
2012;27(9):40308.