Zhang Et Al-2017-Asian Journal of Control

Asian Journal of Control, Vol. 19, No. 3, pp.
996–1008, May 2017

Published online 12 December 2016 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/asjc.1424
OPERATIONAL OPTIMIZATION FOR MICROGRID OF BUILDINGS WITH

DISTRIBUTED SOLAR POWER AND BATTERY
Yuanming Zhang and Qing-Shan Jia
ABSTRACT
Improving building energy efficiency is significant for energy conservation and environmental protection. When
there are multiple buildings with solar power generation and batteries connected in a microgrid, coordinating the dis-
tributed energy supply and consumption may substantially improve the energy efficiency. We consider this important
problem in this paper and make the following three major contributions. First, we formulate the operation optimization
of a microgrid of buildings as a two-stage stochastic programming problem. Second, the problem is transformed into a
stochastic mixed-integer linear programming. Then the scenario method is used to solve the problem. Third, case stud-
ies of a university campus is presented. The numerical results show that coordinating the distributed solar power and
battery can reduce the operational cost of the microgrid.
Key Words: Microgrid of buildings, energy efficiency, solar power, stochastic programming, scenario method.
Nonenclature ptg,l,i Power purchased in node i for building load in

stage t.
Cit Electricity cost of building i in stage t. ptg,b,i Power purchased in node i for battery charging
ptc,i ∕ptd,i Charging/discharging power in stage t. in stage t.
t
Pc,i /Pc,i Minimum/maximum charging power of battery p b,l,i,j
Discharging battery power in node i delivered to
i. building load in node i in stage t.
Pd,i /Pd,i Minimum/maximum discharging power of bat- ptl,i Building load in node i in stage t.
tery i. Qn,i Capacity of battery i.
t t t
ppv,i Solar power generation of node i in stage t. x c,i
∕x d,i
Charging/discharging decision in stage t.
Ppv Rated power of BIPV. Bi , PVi , Li , G Battery, BIPV, load, and distribution grid
ptpv,l,i,j Solar power delivered from node i to satisfy 𝜇t , 𝜈 Time-of-use(TOU) electricity price and selling
building load in node j. price of solar power.
t
ppv,b,i,j Solar power delivered from node i to charge 𝜌 c,j ∕𝜌 d,j Charging/discharging coefficients.
battery in node j in stage t. 𝛾 t
i
SOC of battery i in stage t.
t
ppv,g,i Solar power fed into the grid in node i in stage t. Γi ∕Γi Minimum/maximum level of of SOC.
t
pg,i Total power purchased from the grid in node i in I pv , I 0 , Ipv,n Current generated by the incident light, reverse
stage t. saturation current, light-generated current in
nominal condition.
V , I Output voltage and current.
Manuscript received February 21, 2016; revised July 26, 2016; accepted
Rs , Rp Series resistance and parallel resistance.
September 19, 2016. N s , Np Number of pv cells connected in series and in
The authors are with Center for Intelligent and Networked Systems (CFINS), parallel.
Department of Automation, Tsinghua National Laboratory for Information
T , T ◦
Science and Technology (TNList), Tsinghua University, Beijing 100084, China. a n Actual and nominal (25 C) P-N junction temper-
Qing-Shan Jia is the corresponding author, (e-mail: jiaqs@tsinghua.edu.cn). ature.
This work is supported in part by the Key R&D Project of China Ga , Gn Actual and nominal (1000W∕m2 ) solar irradi-
(No. 2016YFB09010905), the National Natural Science Foundation of China
under grants (Nos. 61021063, 61074034, 61174072, 61222302, 91224008, and ance.
U1301254), the Tsinghua National Laboratory for Information Science and q, k Electron charge (1.60217646×10−19 C) and Boltz-
Technology (TNLIST) Funding for Excellent Young Scholar, the Program for
New Star in Science and Technology in Beijing (No. xx2014B056), and the 111
mann constant (1.30806503 × 10−23 J∕K).
International Collaboration Program of China (No. B06002). a, a ,
1 2 Diode ideality constants.
a
This is an open access article under the terms of the Creative Commons Voc,n , Isc,n Open-circuit voltage and short-circuit current.
Attribution-NonCommercial-NoDerivs License, which permits use and distri-
bution in any medium, provided the original work is properly cited, the use is
IX Indicator function, with IX = 1 if the statement
non-commercial and no modifications or adaptations are made. ’X ’ is true, and 0 otherwise.
© 2016 The Authors. Asian Journal of Control published by Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd.
Y. Zhang And Q. S. Jia: Operational Optimization for Microgrid of Buildings 997
Wgrid Power purchased in one day. considered the optimal planning of a community micro-
Wdemand Power demand in one day. grid and used a probabilistic minimal cut-set-based itera-
𝛽 Degree of self-sufficiency in one day. tive methodology in the planning with various renewable
energy sources. Zhang et al. [12] studied the optimal
scheduling of smart homes’ energy using a mixed-integer
I. INTRODUCTION linear programming considering different kinds of energy
systems. For the demand side, Amin Khodaei and
Buildings have become the largest energy consumer
Mohammad Shahidehpour [13] investigated the opti-
since 2011 in many countries all over the world [1].
mal energy management in a community-based micro-
In China, energy consumed by buildings is expected to
grid with mixed-integer programming and Lagrangian
be responsible for more than 35% of its energy con-
relaxation considering various kinds of residential loads.
sumption until 2020 [2]. Taking into account the energy
Roy et al. [14] evaluated different charging strategies for
crisis, environmental protection, and for the sustainable
hybrid electric vehicles in an office building microgrid
development of human society, much attention has been
with a PV system and evaluated their performance in
focused on improving building energy efficiency. Except
different scenarios. Liu et al. [15] proposed a heuristic
for developing advanced building material, designing
operation strategy for electric vehicle charging activi-
energy-efficient structure and retrofitting old buildings,
ties based on dynamical event triggering for real-time
there are generally two ways for the energy-efficiency
power allocation.
improvement. One way is to exploit renewable energy
In practice, a building microgrid usually consists of
sources, such as solar power, wind power, and biomass
various kinds of buildings located in different areas, e.g.,
energy. The other way is to optimize the operation of
in a university campus, where the power generation and
building equipment in order to economically and effi-
consumption are not centralized. Meanwhile, consider-
ciently use the energy.
ing the different functions of buildings, their demand
For electricity, distributed renewable power brings
profiles are different and sometimes complementary. For
challenges to the power grid considering their uncer-
example, power demand of the office buildings and teach-
tainty, variability and intermittency while the micro-
ing buildings often appears in daytime and that of the
grid technology provides an effective approach to
dormitory buildings often appears in nighttime. Then
integrate the renewable power. A microgrid usually
there arises the chance to improve the energy efficiency
consists of distributed power sources as power generat-
of the building microgrid by cooperatively managing
ing units and decentralized load as power consumers.
their power supply and demand. Power resources (like
The semi-autonomous operation and high self-governed
BIPV and storage battery) affiliated to buildings in valley
degree of microgrid can not only bring stability to
demand can supply power to buildings in peak demand,
the main power grid, but also improve regional energy
which can reduce the total power purchase from the util-
efficiency [3]. With development of building integrated
ity grid especially for areas with time-of-use electricity
power (such as building integrated photovoltaic, BIPV
price and lower the energy cost of the system. This joint
for short) and electricity storage technology, buildings
energy management of the building microgrid benefits
can generate and store power for self consumption and
the power consumers while at the same time brings the
can even feed surplus power into the grid. These charac-
teristics make buildings operate as a building microgrid. following challenges to the operation of the system. First,
Previous building microgrid research generally falls into the scheduling of power in different locations in some
two groups. In the first group, attention is payed to the periods is a multi-time-stage decision-making process
management and operation of various distributed energy and involves spatio-temporal constraints. This makes the
sources [4–6], where the energy consumption is usually optimization of the power scheduling intractable com-
assumed to be centralized. In the other group, research pared to energy management in a single-building micro-
focuses on the demand-side load management [7–10] grid. Second, the distributed renewable power generation
with the energy supply assumed to be centralized. and energy consumption are by nature random. This
A brief literature review of the energy management requires the power schedule to be not only efficient but
in the building microgrid with renewable energy is made also robust accordingly.
as follows. For the supply side, Guan et al. [4] investi- In this paper the problem of the joint operation
gated the energy management of various kinds of energy of a building microgrid is analyzed and discussed. The
systems in a building microgrid and modeled the prob- objective is to minimize the expected electricity cost
lem as a mixed-integer programming (MIP) with sce- in one day and the following major contributions are
nario method addressing the randomness. Che et al. [11] made. First, we formulate the operational optimization
998 Asian Journal of Control, Vol. 19, No. 3, pp. 996–1008, May 2017
of the building microgrid with distributed solar power control policy to optimize the operation of batteries and
and decentralized battery as a two-stage stochastic pro- power scheduling such that the electricity cost is mini-
gramming. In the first stage the day-ahead decisions mized and the system energy efficiency can be improved.
on battery charging/discharging are made. In the sec- Details of the description are as follows.
ond stage the real-time decisions for the power distribu-
tion are made. Second, this problem is converted into
2.1 BIPV system
a stochastic mixed-integer linear programming (SMILP)
with scenario method handling the uncertainties of solar The BIPV system has been used to harvest solar
power and building load. Third, numerical examples power for buildings for its advantageous application to
based on the campus of Tsinghua University in Beijing, both new and old buildings (Here we do not differen-
China are considered. The results show that the proposed tiate BIPV from BAPV, short for building-applied pho-
approach can find control policies that can make the tovoltaics [16]). The PV panel is the basic power gen-
building microgrid operate in an energy-efficient man- erating device of BIPV, which consists of PV modules
ner. The robustness of the schedule with respect to the first connected in series and then in parallel [17]. Taking
uncertainty in solar power and load as well as the com- advantage of the simulation accuracy of the two-diode
putational time of the approach are also discussed. PV model, the BIPV system is described following the
The rest of the paper is organized as follows. In formulation in [18]. Based on the physics of P-N junction
Section II, the operational optimization of the microgrid and the Thevenin’s theorem, the solar power generation
of buildings is formulated as a two-stage stochastic pro- process is as in Fig. 1,
gramming. In Section III, the problem is converted into The corresponding mathematical descriptions are
an SMILP. A scenario-based method is used to handle from (1) to (5).
the uncertainties in solar power and building load. In ( V + Rs I )
Section IV, two numerical examples based on university ppv = V Ipv − Id1 − Id2 − (1)
campus are presented to test the effectiveness of the pro- Rp
posed formulation. At last, a brief conclusion is made in ( ( ) ) Ga
Section V. Ipv = Np Ipv,n + KI Ta − Tn (2)
Gn
( )
( {q V + R I } )
s
II. PROBLEM FORMULATION Id1 = I0 exp −1 (3)
a1 kNs Ta
( )
Consider a microgrid of buildings with M nodes. In ( {q V + R I } )
s
each node i, with i = 1, 2, ...M, there are a building, a Id2 = I0 exp −1 (4)
BIPV system, and a storage battery system. All nodes can a2 kNs Ta
buy electricity from the utility grid as well as selling back ( )
surplus solar power. The topology of this system can be Isc,n + KI Ta − Tn
described with a complete graph in which all the nodes
I0 = N p {( ( )) }
exp Voc,n + KV Ta − Tn ∕(aVt,n ) − 1
are fully connected. Each edge ⟨i, j⟩ represents the power
(5)
line connecting nodes i and j, through which power can
be delivered from one to the other. BIPV in one node can Interested readers can refer to [18] and [19] for more
supply power to the local building (for electricity con- details of the mathematical description. The most impor-
sumption) and the local battery (for battery charging) tant factors that influence solar power generation are
prior to buildings and batteries in other nodes. Surplus
solar power can be sold back to the utility grid. Storage
battery can also be charged by the utility grid when the
electricity price is low. Discharged power from the bat-
tery can be used by local building and scheduled by other
buildings. Power shortage of the whole system is satisfied
by the utility grid.
In a T-stage operation, the system coordinator
needs to make decisions on scheduling of power and the
decisions in one stage influence the decisions afterwards,
which means that the operation of this system is spa-
tiotemporally coupled. Our objective is to make a T-stage Fig. 1. Circuit diagram of a two-diode PV model.
ground solar irradiance Ga and P-N junction temperature Taking into account the lifespan, it is required
Ta . Variations of these two factors lead to the fluctua- that battery should not be charged and discharged
tion of solar power. Also the stochastic nature of Ga and at the same time. Meanwhile, in order to avoid
Ta brings challenges to the prediction of solar power. over-charging/-discharging, the SOC is restricted
Therefore, Ga and Ta are our main considerations in the between the minimum level Γi and the maximum level
description of solar power. Γi . For emergency response, the SOC is further required
to be 𝛾it ≥ Γi + 𝛼 (for example 𝛼 = 0.2), instead of being
2.2 Dynamics of the storage battery 𝛾it ≥ Γi .
In our building microgrid, the battery is mainly

used to store power when there is a surplus and to 2.3 Power scheduling
supply power when there is a deficit. Considering the In the joint-operation, solar power and battery
time-of-use (TOU) electricity price, the battery can also power in one node can be scheduled by other buildings
store power when the rate is low while supply power when besides local consumption. As an example, the power
the rate is high. An important status indicator of battery delivery between two building nodes i and j is presented
is the state-of-charge (SOC), which indicates its remain- in Fig. 2. As aforementioned, PVi can supply solar power
ing energy. The SOC in one stage is influenced by both the to load Li and Lj , as well as to battery Bi and Bj . Surplus
SOC in the previous stage and the charging/discharging solar power can be fed to the utility grid G. The dis-
decision in the current stage. In a discrete-time form, the charged power from Bi can be consumed by Li and also be
dynamics of SOC and the charging/discharging limita- scheduled by Lj . The power shortage of Li is satisfied by
tion for battery in node i are formulated in (6)–(8), with purchasing power from G. Purchased power can also be
t = 1, 2, ..., T − 1. used to charge Bi . Node j has the same operation mode.
The mathematical formulation is shown as (9)–(11).
1. Dynamics of the SOC:
(𝜌c,i ⋅ ptc,i ⋅ xtc,i + ptd,i ⋅ xtd,i )Δt 1. Power from BIPV:
𝛾it+1 = 𝛾it + (6)
Qn,i M (
∑ )
ptpv,i = ptpv,l,i,j + ptpv,b,i,j ⋅ xtc,j + ptpv,g,i (9)
with 𝛾it ∈ [Γi , Γi ] ⊆ (0, 1).
j=1
2. Charging/discharging decision:
2. Charging/discharging power of storage battery:
xtc,i − xtd,i ≤ 1 (7)
⎧
⎪ pt = ( ∑ pt
M
with xtc,i ∈ {0, 1} and xtd,i ∈ {0, −1}.
+ ptg,b,i ) ⋅ xtc,i
3. Charging/discharging limitation: ⎪ c,i j=1
pv,b,j,i
{ ⎨ (10)
⎪ pt = ( ∑ pt ) ⋅ xt
M
Pc,i ⋅ xtc,i ≤ ptc,i ≤ Pc,i ⋅ xtc,i
(8) ⎪ d,i b,l,i,j d,i
Pd,i ⋅ (−xtd,i ) ≤ ptd,i ≤ Pd,i ⋅ (−xtd,i ) ⎩ j=1
Fig. 2. Power delivery between node i and node j . [Color figure can be viewed at wileyonlinelibrary.com]
load. As time goes by, more information about the solar

power and the building load can be acquired. Then a
more appropriate power scheduling decision can be made
inducing the actual electricity cost C(X), where there
̄
exists a recourse to the previous cost C(X). Note also
that although the first stage and second stage share the
same battery charing/discharing decision X, their power
Fig. 3. Two-stage decision process. [Color figure can be
schedulings are different. The expectation "E" of the
viewed at wileyonlinelibrary.com]
recourse cost is over the stochastic solar power and build-
ing load. The objective function (15) is further expressed
3. Power from distribution grid: as (16).
ptg,i = ptg,l,i + ptg,b,i ⋅ xtc,i (11) min E[C(X)] (16)

X
Note also that in order to make decisions on power

2.4 Power balance scheduling, the coordinator needs to collect information
Power balance is the coordination between power of all building nodes through the communication sys-
supply and demand. In the joint-operation, the power tem. As the sensing system is usually integrated into
balance for building load Li as (12). the energy system with no additional cost, in practice
the communication cost can be neglected for transmit-
∑
M
ting the information in one day’s operation. As for the
(ptpv,l,j,i + 𝜌d,j ⋅ ptb,l,j,i ⋅ xtd,j ) + ptg,l,i = ptl,i (12) uncertainty in communication, it can be considered as
j=1
another reason inducing solar power and building load
variations in the coordinator’s view. This can also be
handled with the scenario method which we will not go
2.5 Objective function into particulars.
Assume 𝜇t is the TOU price at stage t and 𝜈 is the
price for selling solar power. (In present Beijing, China, III. SOLUTION METHODOLOGY
we have 𝜈 < 𝜇t and 𝜈 is constant). At stage t, each node
can either make a decision to buy electricity and pay the The previous formulation includes the nonlinear
bill or to sell solar power and receive a revenue. The elec- cost function and the integer battery charg-
tricity cost for node i with the battery charing/discharing ing/discharging decisions, which is a stochastic
decision being xtc,i ∕xtd,i is mixed-integer nonlinear programming (SMINLP). In
this section, we first transform the SMINLP into a
Cit (xtc,i ∕xtd,i ) = 𝜇t ptg,i I{pt >0} − 𝜈ptpv,g,i I{pt >0} (13) stochastic mixed-integer linear programming (SMILP)
g,i pv,g,i
by linearizing the objective and constraints. Then the

Denote vector Xt = (xtc,1 ∕xtd,1 , ..., xtc,M ∕xtd,M ) and X =
scenario method is used to solve the problem with sce-
(X1 , ..., XT ), the total T-stage electricity cost is narios approximating the stochastic solar power and
building load.
∑ ∑
T M
C(X) = Cit (xtc,i ∕xtd,i ). (14)
t=1 i=1
3.1 Linearization of the cost and constraints
Considering the stochastic elements and following For the nonlinear constraints (6), (9), (10), (11) and
the two-stage stochastic programming formulation [20], (12), the following generic denotation is used to describe
our objective is to minimize the electricity cost with a the linearization process. Assume a nonlinear expression
second-stage expected recourse cost as (15). of y as (17), where x is continuous with x ∈ [0, Ymax ] and
̄ ̄ 𝛼 ∈ {0, 1}.
min C(X) + E[C(X) − C(X)] (15)
X ⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟
recourse y=x⋅𝛼 (17)
The two-stage decision process is shown in Fig. 3. Note Introduce another nonnegative variable 𝜃 and an equiv-
̄
that C(X) is the first-stage electricity cost based on alent linear expression of equation (17) is as (18)
the day-ahead prediction of solar power and building with which all the nonlinear constraints concern with
charing/discharging decisions can be linearized. Algorithm 1 Scenario Generation Algorithm

{ 1. Given the prediction of solar irradiance, temper-
y=𝜃 ̄ t , T̄ t , and p̄ t , for i =
(18) ature, and building load as G
0 ≤ 𝜃 ≤ 𝛼 ⋅ Ymax . a a l,i
1, 2, ...M and t = 1, 2, ...T;
For the objective function (14), whether cost is paid 2. According to the corresponding distributions, ran-
or revenue is received is determined by buying or selling domly generate the prediction errors as 𝜀̂ te1 , 𝜀̂ te2 , and 𝜀̂ ti ;
electricity, which is influenced by the charging/discharing
̄ t + 𝜀̂ t and T̄ t + 𝜀̂ t into (1)–(5) to cal-
3. Substitute G
decision of battery and the power scheduling. For each a e1 a e2
time stage t, introduce another two variables zt and culate the solar power p̂ tpv,i for one scenario and the
𝜙t , with zt being the total cost (revenue) at t and 𝜙t corresponding building load is p̂ tl,i = p̄ tl,i + 𝜀̂ ti .
being (19). 4. Then a scenario s for(solar power and building load )
can be expressed as s ∶ p̂ tpv,1 , ..., p̂ tpv,M , p̂ tl,1 , ..., p̂ tl,M .
∑
M
𝜙t = (ptl,i − ptpv,i − 𝜌d,i ptd,i xtd,i + ptc,i xtc,i ). (19)
i=1
∑
T
Then we have ̃S(
min E ẑ t )
{ t t t=1
𝜇 .𝜙 if 𝜙t > 0;
zt = (20) s.t.
𝜈.𝜙t if 𝜙t ≤ 0.
First-stage constraints:
For the T time stages, an equivalent expression of
the objective function is as the following optimization ∑
M
problem with linear cost (revenue) and constraints in p̄ tl,i = (̄ptpv,l,j,i + 𝜌d,j ⋅ p̄ tb,l,j,i ) + p̄ tg,l,i
each stage. j=1
M ( )
∑
∑
T
p̄ tpv,i = p̄ tpv,l,i,j + p̄ tpv,b,i,j + p̄ tpv,g,i
min E( zt ) j=1
t=1
(21) p̄ tg,i = p̄ tg,l,i + p̄ tg,b,i
s.t. zt ≥ 𝜇t ⋅ 𝜙t
zt ≥ 𝜈 ⋅ 𝜙t ∑
M
p̄ tc,i = p̄ tpv,b,i + p̄ tg,b,i , p̄ td,i = p̄ tb,l,i,j
t = 1, 2, ...T. j=1
where ’s.t.’ is short for ’subject to’. 𝛾̄it = 𝛾̄it−1 + [(𝜌c,i ⋅ p̄ t−1
c,i
+ p̄ t−1
d,i
)Δt]∕Qn,i
Second-stage constraints:
3.2 Stochastic programming with the scenario method
ẑ t ≥ 𝜇t ⋅ 𝜙̂ t , ẑ t ≥ 𝜈 ⋅ 𝜙̂ t
The scenario method has been widely used in a
variety of mathematical optimization problems with ran- ∑
M
domness, such as portfolio selection [21], risk manage- 𝜙̂ t = (̂ptl,i − p̂ tpv,i − 𝜌d,i ⋅ p̂ td,i + p̂ tc,i )
i=1
ment [22], and power management [23], to deal with
uncertainty. In our problem, the solar power and the ∑
M
building load are uncertain. Denote the prediction error p̂ tl,i = (̂ptpv,l,j,i + 𝜌d,j ⋅ p̂ tb,l,j,i ) + p̂ tg,l,i
j=1
of the solar irradiance, the temperature, and the load of
the i-th building by 𝜀e1 , 𝜀e2 , and 𝜀i , which we assume to M ( )
∑
be subject to Gaussian distribution as 𝜀e1 ∼ N(𝜇e1 , 𝜎e1 2
), p̂ tpv,i = p̂ tpv,l,i,j + p̂ tpv,b,i,j + p̂ tpv,g,i
𝜀e2 ∼ N(𝜇e2 , 𝜎e2
2
), and 𝜀i ∼ N(𝜇i , 𝜎i2 ). Then a scenario for j=1
solar power and building load can be generated through p̂ tg,i = p̂ tg,l,i + p̂ tg,b,i
the Algorithm 1.
∑
M
With S scenarios of solar power and building p̂ tc,i = p̂ tpv,b,i + p̂ tg,b,i , p̂ td,i = p̂ tb,l,i,j
load for the future T stages, the microgrid operation j=1
problem can be further approximated by the following
mixed-integer linear programming (MILP). 𝛾̂it = 𝛾̂it−1 + [(𝜌c,i ⋅ p̂ t−1
c,i
+ p̂ t−1
d,i
)Δt]∕Qn,i
Common constraints: the web site http://climate.dest.com.cn/. It is assumed

xtc,i − xtd,i ≤ 1 that all the building energy systems in the campus share
the same weather condition (the solar radiation and the
Pc,i ⋅ xtc,i ≤ p̂ tc,i (̄ptc,i ) ≤ Pc,i ⋅ xtc,i temperature). The TOU electricity price in Beijing area
is shown in Fig. 4. Price for selling electricity is 0.3
Pd,i ⋅ xtd,i ≤ p̂ td,i (̄ptd,i ) ≤ Pd,i ⋅ xtd,i
RMB/kWh, which is lower than the valley TOU price.
Γi ≤ 𝛾̂it (̄𝛾it ) ≤ Γi The rated power of the BIPV system is determined by
the available roof area of each building. The capacity of
Pc,i ⋅ xtc,i ≤ p̂ tpv,b,i (̄ptpv,b,i ) ≤ Pc,i ⋅ xtc,i
each battery is calculated according to that it can support
Pc,i ⋅ xtc,i ≤ p̂ tg,b,i (̄ptg,b,i ) ≤ Pc,i ⋅ xtc,i the operation of local building for at least 5 hours at the
peak demand. The minimum/maximum SOC of the bat-
Pd,i ⋅ xtd,i ≤ p̂ tb,l,i,j (̄ptb,l,i,j ) ≤ Pd,i ⋅ xtd,i tery is 0.3/0.9. The initial SOC is 0.3. The charging and
where t = 1, 2, ...T, i = 1, 2, ...M, variables with a bar the discharging coefficients are 0.9.
mean their decisions based on prediction, variables with
a hat mean their decisions in a future scenario s, and the
constraints include all S scenarios.
The first-stage constraints are based on the predic-
tion and the second-stage constraints are based on the
future realizations. The common constraints influence
decisions in both the two stages. Note that E ̃ S (∑T ẑ t )
t=1
is an approximation to the expectation with S scenar-
ios. Though the approximation error goes to zero when
the scenario number S goes to infinity, a large num-
ber of scenarios usually increases the computational
time. Fortunately, a lot of researches have been car-
ried out for scenario reduction [23–25], where we will
not go into particulars. Also note that in our two-stage
stochastic programming, all S scenarios share the same
charging/discharging decisions made day-ahead based on
both the prediction and the realizations. But the optimal
charging/discharging power and the power scheduling
decisions are different for each realization. Our objective Fig. 4. The TOU price in Beijing. [Color figure can be viewed
is to find the optimal battery charging/discharging deci- at wileyonlinelibrary.com]
sions which can minimize the total expected electricity
cost while not only accommodate all the randomness, but
also meet the power demand for each building node.
IV. NUMERICAL TESTS AND ANALYSIS
Two numerical examples based on the campus of

Tsinghua University in Beijing are tested in this section.
All tests are performed using the IBM CPLEX solver
(version 12.4) on a laptop with the CPU main frequency 2
GHz. The first example with three buildings is presented
in subsection 4.1 and a larger example with 18 buildings is
presented in subsection 4.2. The day-ahead building elec-
trical load is estimated using EneryPlus (version 7.2) with
the comfort indoor temperature between 22◦ C and 26◦ C.
The weather information acquired from the weather
station located in Tsinghua University for a typical sum-
mer working day (August 13, 2015) is used in the test. The
historical and realtime weather data can be acquired from Fig. 5. One-line sketch diagram of the 3-node system.
4.1 Case 1: A 3-node system both calculated independently. Strategy B and Strategy
D are calculated using 100 solar power and building load
The 3-node system consists of a dormitory, a dining
scenarios with their standard deviations (StD) being 0.1
hall and a teaching building, indexed with 1, 2, and 3 as
and 0.1. The average cost and the average computation
described with the one-line sketch diagram in Fig. 5. The
time of these four strategies with another 100 scenarios
storage battery can be seen as both an energy supplier
are shown in Table II and Table III.
and a load. Parameters of the 3-node system are listed
Note that in Table II, the cost of each node using A
in Table I. Estimated building load profiles using Ener-
is close to that using B. The cost of each node using C is
gyPlus in a 24-hour time horizon are presented in Fig. 6
close to that using D. The total cost using C or D is less
with different demand curves for three buildings.
than the total cost using A or B. The cost of node 2 using
4.1.1 Comparison of different strategies Table I. Parameters of the 3-node sys-

tem.
Battery control strategy from the proposed
two-stage stochastic programming (named D) is com- No. 1 2 3
pared to other three control strategies A, B and C.
PPV (kW) 100 150 300
Strategies A and B are from literature [4] and strategy Qn (kWH) 1000 1500 3000
C is from literature [26]. A is obtained based on the Pc∕d (kW) 0 0 0
prediction of solar power and building load without
Pc∕d (kW) 100 150 300
considering the randomness and the power scheduling
among different buildings (∀i, j ∈ M and i ≠ j, ptpv,l,i,j =
0, ptpv,b,i,j = 0 and ptb,l,i,j = 0,). B is obtained considering Table II. Average cost (RMB).
the randomness of solar and load but not considering A B C D
the power scheduling among different nodes either. C
is obtained based on the prediction of solar power and No. 1 363.4 359.2 343.4 396.4
building load without considering their randomness but No. 2 136.4 135.3 338.1 289.6
No. 3 921.1 914.0 626.4 620.1
considering the power scheduling among different build-
Total 1420.9 1408.5 1307.9 1306.1
ings. For each building, strategy A and strategy B can be
Fig. 6. Building load profiles of the 3-node system. [Color figure can be viewed at wileyonlinelibrary.com]
Table III. Average computation time(s). achieves a higher degree of self-sufficiency. This also
demonstrates the advantage for sharing the solar power
A B C D
and the batteries among buildings in the microgrid.
Time 0.20 3.46 0.35 46.36
Wgrid
𝛽 =1− (22)
Wdemand
A and B are much less than that using C and D. On the
contrary, the cost of node 3 using A and B are much more The relationship between the average cost and the
than that using C and D. Analysis are made as follows. scenario numbers for strategy D is shown in Fig. 8. As we
On the one hand, both A and B are local optimal control can see, when there are more than 40 scenarios, the aver-
strategies and their costs are close [4]. C and D are global age cost variation is less than 0.5%. This demonstrates
optimal control strategies which also have close costs. On that the proposed method can provide a good control
the other hand, in strategy A and B the decisions of node strategy with an appropriate number of scenarios.
2 are made by optimizing the local operation. Using A
and B the solar power can be sold to get a revenue and the 4.1.2 Sensitivity to load and solar variation
power bought from the grid only needs to meet its own Strategies A and C are obtained based on the pre-
demand curve. However, in strategies C and D the deci- dicted solar power and building load. Strategies B and
sions of node 2 need to consider the operation of all three
nodes. The surplus solar power is supplied to other nodes.
More traditional power is used to charge the battery so
that the battery may be used by the other buildings when
there is a high demand on power and when the TOU
price is high. The decrease of the revenue for selling solar
power and the increase of the cost for purchasing the tra-
ditional power induce a higher cost in node 2. For node
3, with strategies C and D it is possible for it to utilize
the solar power and the batteries in the other buildings
when its power demand is high. This reduces the power
purchased from the grid especially when the TOU price
is high, and therefore reduces its total cost. Furthermore,
the power scheduling among different buildings improves
the energy efficiency of the solar power and the power
storage system in strategy C or D, which brings a more
economic operation of the whole system.
Fig. 7. Comparison of 𝛽 . [Color figure can be viewed at
Note also that although the costs of strategy C and wileyonlinelibrary.com]
strategy D are close, the computation time for D is larger
than that for C. This is caused by the consideration of
randomness in D. In practice, if the prediction accuracy
of solar power and building load is high, strategy C is
a good enough control strategy. While if the prediction
accuracy is lower or the variations of solar and load are
large, strategy D is a better choice.
Considering the capability of autonomous oper-
ation, self-sufficiency (or self-adequacy) has been dis-
cussed for the energy management in different kinds of
smart microgrids [27–31]. A high self-sufficiency can not
only lower the electricity cost but also lower the emission
for generating traditional power. What’s more, for pol-
icy and technical reasons in the mainland China, local
consumption of renewable energy is encouraged by the
government. Considering these reasons, define 𝛽 as the Fig. 8. Relationship between average cost and
one-day degree of self-sufficiency as in (22). From Fig. scenario number. [Color figure can be viewed at
7 we can see that with strategies C and D the system wileyonlinelibrary.com]
D are obtained based on scenarios that are generated

from the random distribution of solar power and build-
ing load. We investigate the robustness of these strategies
with respect to the uncertain solar power and load in
this subsection. In order to test their sensitivity to solar
and load variations, we conduct two groups of numer-
ical experiments. In the first group, we fix the standard
deviation (StD) of the solar power at 0.10 and set the
standard deviation of the load to 0.10, 0.15, ..., 0.5. The
total costs using strategies A, B, C, and D are shown in
Fig. 9. In the second group, we fix the standard devia-
tion of the load at 0.10 and set the standard deviation
of the solar to 0.10, 0.15, ..., 0.5. The total costs using
strategies A, B, C, and D are shown in Fig. 10. In both
Fig. 9 and Fig. 10, we can see that strategies C and D
Fig. 9. Cost comparison (solar StD=0.10). [Color figure can
achieve lower costs than strategies A and B. As shown be viewed at wileyonlinelibrary.com]
in Fig. 11 and Fig. 12, when the standard deviation of
the solar is 0.50, strategy D achieves 15%, 13%, and 5%
lower cost comparing with A, B, and C, respectively.
When the standard deviation of the load is 0.50, strat-
egy D achieves 8%, 7%, and 1.5% lower cost comparing
with strategies A, B, and C, respectively. The comparison
between strategies A and C shows that scheduling power
among different buildings is more robust to address the
uncertainty of solar power and building load. The com-
parison between strategy C and strategy D shows that
the two-stage stochastic formulation has better perfor-
mance in addressing uncertainty. Here we have no risk
reference in the proposed formulation and all the com-
parisons are made with neutral risk. Conejo et al. [32]
has made discussions about various kinds of risk control
in electricity markets with uncertainty. The conditional
value-at-risk (CVaR) is implemented as a risk measure
incorporated into the the risk-neutral problems for deci- Fig. 10. Cost comparison (load StD=0.10). [Color figure can
sion making with different confidence levels of risk. The be viewed at wileyonlinelibrary.com]
purpose of comparison here is to show the advantage
of joint-operation of the building microgrids. If differ-
ent risk references need to be considered, the formu-
lations with CVaR in [32] can also be applied in the
joint-operation for a robust power schedule.
4.2 Case 2: An 18-node system

As discussed in Section III, the number of con-
straints increases when the number of scenarios increases.
In this subsection, a microgrid of 18 buildings located in
Tsinghua campus in Beijing is used to test the influence of
the scenarios number on the cost and computation time
of the proposed method. This 18-node system consists of
5 research buildings (No. 13–17), 1 office building (No.
18), 4 teaching buildings (No. 3–6) , 2 dining halls (No. Fig. 11. Relative cost saving (solar StD=0.10). [Color figure
1–2) , 4 dormitories (No. 7–10) , 1 cinema (No. 11), and can be viewed at wileyonlinelibrary.com]
Fig. 13. Influence of scenario number. [Color figure can be

Fig. 12. Relative cost saving (load StD=0.10). [Color figure viewed at wileyonlinelibrary.com]
can be viewed at wileyonlinelibrary.com]
Table IV. Parameter settings of the 18-node system. computational time is about 300 seconds. This demon-
No. PPV (kW) Qn (kWh) Pc∕d (kW) Pc∕d (kW) strates that the previously proposed two-stage stochastic
programming can provide a good enough control strat-
1 300 2000 0 200 egy for the operation of the building microgrid with
2 300 2000 0 200 appropriate computational time and scenarios.
3 400 4000 0 400
4 250 4000 0 400
5 250 4500 0 450
6 600 5000 0 500 V. CONCLUSION
7 1500 3500 0 350
8 2000 3500 0 350 Improving building energy efficiency is important
9 2500 8000 0 800 for energy saving. Taking advantage of the microgrid
10 1500 9000 0 900 technology, the joint operation of a building microgrid
11 500 8500 0 850 is considered in this paper. We formulate the joint oper-
12 800 8000 0 800 ation of the building microgrid with multiple buildings
13 800 9500 0 950 as a two-stage stochastic programming. The scenario
14 600 8500 0 850 method is used to address the uncertainties of solar
15 700 8500 0 850
power and building load. After linearization, CPLEX
16 900 8000 0 800
is applied to solve the problem. Numerical results show
17 800 8000 0 800
18 2000 10000 0 1000 that the energy efficiency of the building microgrid is
improved using the proposed method which consid-
ers power scheduling among different buildings. It also
1 concert hall (No. 12). These buildings are connected in shows that joint operation of the building microgrid
a same distribution grid. Parameter settings are listed in can better accommodate solar power and building load
Table IV. Note that when the building nodes are powered uncertainties compared to the operation of a single build-
by different distribution grids, the proposed formulation ing. Moreover, the tests show that a good enough control
can also apply for buildings connected in the same grid. strategy can be obtained with appropriate number of
Different numbers of scenarios are generated to test scenarios for the proposed method.
the proposed control method when the standard devia-
tion of the solar power is 0.10 and when the standard REFERENCES
deviation of the load is 0.10. For each number of scenar-
ios 10 tests are performed and the results are presented 1. Lu, N., T. Taylor, W. Jiang, J. Correia, L. R. Leung,
in Fig. 13. As we can see the average computational time and P. C. Wong, “The temperature sensitivity of
increases when the number of scenarios increases. How- the residential load and commercial building load,”
ever, when there are more than 35 scenarios, the relative 2009 IEEE Power & Energy Society General Meeting,
variation of the cost is less than 0.4%. The overall average Calgary, Alberta, Canada., pp. 1–7 (2009).
2. Cai, W., Y. Wu, Y. Zhong, and H. Ren, “China build- 14. Van Roy, J., N. Leemput, F. Geth, J. Buscher,
ing energy consumption: situation, challenges and R. Salenbien, and J. Driesen, “Electric vehicle charg-
corresponding measures,” Energy Policy, Vol. 37, ing in an office building microgrid with distributed
pp. 2054–2059 (2009). energy resources,” IEEE Trans. Sustain. Energy,
3. Katiraei, F. and M. R. Iravani, “Power management Vol. 5, pp. 1389–1397 (2014).
strategies for a microgrid with multiple distributed 15. Liu, N., Q. Chen, J. Liu, X. Lu, P. Li, J. Lei,
generation units,” IEEE Trans. Power Syst., Vol. 21, and J. Zhang, “A heuristic operation strategy for
pp. 1821–1831 (2006). commercial building microgrids containing EVs and
4. Guan, X., Z. Xu, and Q. S. Jia, “Energy-efficient PV system,” IEEE Trans. Sustain. Energy, Vol. 62,
buildings facilitated by microgrid,” IEEE Trans. pp. 2560–2570 (2015).
Smart Grid, Vol. 1, pp. 243–252 (2010). 16. Santos, I. P. D. and R. Ruther, “The potential
of building-integrated (BIPV) and building-applied
5. Stadler, M., Effect of heat and electricity storage
photovoltaics (BAPV) in single-family, urban res-
and reliability on microgrid viability: a study of com-
idences at low latitudes in Brazil,” Energy Build.,
mercial buildings in California and New York states,
Vol. 50, pp. 290–297 (2012).
Lawrence Berkeley National Laboratory, Berkeley,
17. Villalva, M. G., J. R. Gazoli, and E. R. Filho, “Com-
California, America (2009). prehensive approach to modeling and simulation of
6. Wang, Z., R. Yang, and L Wang, “Intelligent photovoltaic arrays,” IEEE Trans. Power Electron.,
multi-agent control for integrated building and Vol. 24, pp. 1198–1208 (2009).
micro-grid systems,” First Conf. Innov. Smart Grid 18. Ishaque, K., Z Salam, and Syafaruddin, “A com-
Technol., CA, USA, pp. 1–7 (2011). prehensive matlab simulink pv system simulator
7. Zong, Y., D. Kullmann, A. Thavlov, O. Gehrke, and with partial shading capability based on two-diode
H. W. Bindner, “Application of model predictive con- model,” Sol. Energy, Vol. 85, pp. 2217–2227 (2011).
trol for active load management in a distributed 19. Ishaque, K., Z. Salam, H Taheri, and Syafaruddin,
power system with high wind penetration,” IEEE “Modeling and simulation of photovoltaic (PV) sys-
Trans. Smart Grid, Vol. 3, pp. 1055–1062 (2012). tem during partial shading based on a two-diode
8. Zhao, P., S. Suryanarayanan, and M. G. Simoes, “An model,” Simul. Model. Pract. Theory, Vol. 19,
energy management system for building structures pp. 1613–1626 (2011).
using a multi-agent decision-making control method- 20. Birge, J. R. and F. Louveaux, Introduction to Stochas-
ology,” IEEE Trans. Ind. Appl., Vol. 49, pp. 322–330 tic Programming, Springer, New York (2011).
(2013). 21. Barro, D. and E. Canestrelli, “Dynamic portfolio
9. Zong, Y., D. Kullmann, A. Thavlov, O. Gehrke, and optimization: Time decomposition using the maxi-
H. W Bindner, “Active load management in an intel- mum principle with a scenario approach,” Euro. J.
ligent building using model predictive control strat- Oper. Res., Vol. 163, pp. 217–229 (2005).
egy,” 2011 IEEE Trondheim Power Tech, Jun. 19–23, 22. Chapelle, A., Y. Crama, G. Hubner, and J. P. Peters,
Trondheim, Norway, pp. 1-–6 (2011). “Practical methods for measuring and managing
10. Singh, A., “Multifunctional capabilities of grid operational risk in the financial sector: A clinical
study,” J. Banking Finance, Vol. 32, pp. 1049–1061
connected distributed generation system with
(2008).
non-linear loads,” Asian J. Control, Vol. 18, No. 4,
23. Growe-Kuska, N., H. Heitsch, and W. Romisch,
pp. 1537–1545 (2016).
“Scenario reduction and scenario tree construction
11. Che, L., X. Zhang, M. Shahidehpour, A. Alabdul-
for power management problems,” Power Tech Conf.
wahab, and A. Abusorrah, “Optimal interconnection Proc. IEEE Bologna, Vol. 3, pp. 7–18 (2003).
planning of community microgrids with renewable 24. Dupacova, J., N. Growe-Kuska, and W. Romisch,
energy sources,” IEEE Trans. Smart Grid, Vol. PP, “Scenario reduction in stochastic programming,”
pp. 99 (2015). DOI: 10.1109/TSG.2015.2456834. Math. Program., Vol. 95, pp. 493–511 (2003).
12. Zhang, D., N. Shah, and L. G. Papageorgiou, “Ef- 25. Heitsch, H. and W. Romisch, “Scenario reduction
ficient energy consumption and operation manage- algorithms in stochastic programming,” Comput.
ment in a smart building with microgrid,” Energy Optim. Appl., Vol. 24, pp. 187–206 (2003).
Conv. Manag., Vol. 74, pp. 209–222 (2013). 26. Zhang, Y. and Q. S. Jia, “Optimal storage bat-
13. Khodaei, A. and M. Shahidehpour, “Optimal oper- tery scheduling for energy-efficient buildings in a
ation of a community-based microgrid,” 2011 IEEE micro-grid,” Proc. 27th Chinese Control Decis. Conf.,
PES, Medellin, Colombia, pp. 1–3 (2011). pp. 5540–5545 (2015).
27. Ali Arefifar, S., Y. Abdel-Rady, I. Mohamed, Tarek, Yuanming Zhang received the B.S.
and H. M. El-Fouly, “Supply-Adequacy-Based Opti- degree in detection, guidance, and con-
mal Construction of Microgrids in Smart Distri- trol techniques from Harbin Institute of
bution Systems,” IEEE Trans. Smart Grid, Vol. 3, Technology and M.S. degree in control
pp. 1491–1502 (2012). theory and engineering in Automation
28. Zhao, B., Y. Shi, X. Dong, W. Luan, and Research and Design Institute of Metal-
J. Bornemann, “Short-Term Operation Scheduling lurgical Industry, in 2005 and 2012. He
in Renewable-Powered Microgrids: A Duality-Based is currently pursuing the Ph.D. degree at the Tsinghua
Approach,” IEEE Trans. Sustain. Energy, Vol. 5, University, Beijing, China. His research interests include
power system optimization, optimization theory and
pp. 209–218 (2014). solar power system.
29. Jiang, Q., M. Xue, and G. Geng, “Energy man-
agement of microgrid in grid-connected and
stand-alone modes,” IEEE Trans. Power Syst., Qing-Shan Jia (S’02-M’06-SM’11)
Vol. 28, pp. 3380–3390 (2013). received the B.E. degree in automation
in July 2002 and the Ph.D. degree in
30. Abu-Sharkha, S., R. J. Arnolde, J. Kohlerd, R. Lia, control science and engineering in July
T. Markvarta, J. N. Rossb, K. Steemersc, P. Wilsonb, 2006, both from Tsinghua University,
and R. Yao, “Can microgrids make a major contribu- Beijing, China. He is an associate pro-
tion to UK energy supply?” Renew. Sust. Energ. Rev., fessor at the Center for Intelligent and
Vol. 10, pp. 78–127 (2006). Networked Systems (CFINS), Department of Automa-
31. Huang, Q., Q. S. Jia, Z. Qiu, X. Guan, and tion, TNLIST, Tsinghua University, Beijing, China.
G. Deconinck, “Matching EV charging load with He was a postdoc at Harvard University in 2006, a
uncertain wind power: A simulation-based policy visiting assistant professor at the Hong Kong Univer-
improvement approach,” IEEE Trans. Smart Grid, sity of Science and Technology in 2010, and a visiting
Vol. 6, pp. 1425–1433 (2015). associate professor at Massachusetts Institute of Tech-
nology in 2013. His research interests include theories
32. Conejo, A. J., M. Carri’on, and J. M. Morales, Deci- and applications of discrete event dynamic systems
sion Making under Uncertainty in Electricity Mar- (DEDSs) and simulation-based performance evaluation
kets, Springer, New York (2010). and optimization of complex systems.

Zhang Et Al-2017-Asian Journal of Control

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Zhang Et Al-2017-Asian Journal of Control

Uploaded by

Copyright:

Available Formats

Asian Journal of Control, Vol. 19, No. 3, pp.

996–1008, May 2017

OPERATIONAL OPTIMIZATION FOR MICROGRID OF BUILDINGS WITH

Nonenclature ptg,l,i Power purchased in node i for building load in

In our building microgrid, the battery is mainly

load. As time goes by, more information about the solar

ptg,i = ptg,l,i + ptg,b,i ⋅ xtc,i (11) min E[C(X)] (16)

Note also that in order to make decisions on power

by linearizing the objective and constraints. Then the

charing/discharging decisions can be linearized. Algorithm 1 Scenario Generation Algorithm

Common constraints: the web site http://climate.dest.com.cn/. It is assumed

IV. NUMERICAL TESTS AND ANALYSIS

Two numerical examples based on the campus of

4.1.1 Comparison of different strategies Table I. Parameters of the 3-node sys-

D are obtained based on scenarios that are generated

4.2 Case 2: An 18-node system

Fig. 13. Influence of scenario number. [Color figure can be

You might also like