You are on page 1of 14

Journal of Cleaner Production 275 (2020) 122776

Contents lists available at ScienceDirect

Journal of Cleaner Production


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

Optimal operation of water-energy microgrids; a mixed integer linear


programming formulation
Faegheh Moazeni a, *, Javad Khazaei b, c
a
Civil & Environmental Engineering Department, School of Science Engineering and Technology, Penn State Harrisburg, 777 W Harrisburg Pike,
Middletown, PA 17057, USA
b
Electrical Engineering Department at Penn State Harrisburg, 777 W Harrisburg Pike, Middletown, PA, 17057, USA
c
Architectural Engineering Department at Penn State University Park, State College, PA 16801, USA

a r t i c l e i n f o a b s t r a c t

Article history: Increasing energy efficiency in water distribution systems is one of the most important aspects of a
Received 8 March 2020 sustainable water infrastructure. Additionally, providing this energy from renewable sources is an
Received in revised form essential step towards a cleaner production in energy infrastructures. Therefore in this paper, an opti-
8 June 2020
mization model is developed to minimize the energy consumption of a water-energy microgrid system. A
Accepted 10 June 2020
Available online 19 July 2020
day-ahead economic dispatch model is developed to minimize the daily cost of energy in the water-
energy microgrid. The energy consumption of water system is minimized using tank’s and pump’s
Handling Editor: M.T. Moreira scheduling and operation, hydraulic factors, and daily demand. Particularly, the electricity consumption
of the pump is minimized by adjusting its head gain and flow rate via a variable speed. The energy unit is
Keywords: composed of an aggregated conventional power generation, solar photovoltaic, wind generation, and
Water-energy microgrid battery energy storage system. To offer a global optimum for the proposed non-linear programming
Cost optimization formulation, bivariate piecewise linear approximation is used to linearize the pump’s power consump-
Mixed integer non-linear programming tion and gain, and conservation of energy-mass. In addition, univariate piecewise linear approximation is
(MINLP)
used to linearize the energy consumption function of conventional power generation systems. The
Mixed integer linear programming (MILP)
optimization results of the mixed integer linear programming and mixed integer non-linear program-
Co-optimization
Piecewise linearization ming formulations for the studied water-energy microgrid system are compared and discussed. These
models are developed to concurrently minimize the electricity consumption of a micro water-energy
distribution network in two scenarios of (1) standalone water distribution system operation and (2)
an integrated islanded micro water-energy system with electrical loads.
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction et al., 2013), or energy efficiency of water-energy nexus. One


main aspect of energy efficiency in water-energy networks is to
The rapid pace of urbanization is constantly raising the demand allocate electricity generation resources optimally to manage the
for water and energy resources, challenging the urban economy as power consumption of pumps in a water distribution network and
well as the sustainability of the environment. By the year 2035, the meet the electricity demand of the water-energy network
global need for energy alone is estimated to be 40% higher than that (Scarazzato et al., 2017; Baleta et al., 2019). In addition, the demand
of 2010. Thus, it is imperative to explore the interlinks between response capability of water distribution networks can be studied
energy and water so that both resources are managed and by controlling the components of water distribution networks. The
conserved sustainably (Dai et al., 2018). There has been some re- main focus of this work is on optimization of water-energy systems
ports on development and optimization of water-energy systems in from the energy efficiency point of view and comparison between
a coordinated framework, which can be divided into modeling and linear and nonlinear formulation of the energy efficiency in water-
regulatory challenges (Scott et al., 2011; Bragalli et al., 2012; Lele energy systems.
A few research papers are available that studied the energy ef-
ficiency of both water and energy networks simultaneously. For
* Corresponding author. Penn State Harrisburg, 777 W Harrisburg Pike, Middle- instance, the cost of energy system with conventional energy
town, PA, 17057, USA. sources in a water-energy network was minimized using a mixed
E-mail addresses: moazeni@psu.edu (F. Moazeni), khazaei@psu.edu (J. Khazaei).

https://doi.org/10.1016/j.jclepro.2020.122776
0959-6526/© 2020 Elsevier Ltd. All rights reserved.
2 F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776

Nomenclature Pnom Nominal wind power in kW


k
Ppump Power of pump k in kW
aDG
i ; bi ; gi
DG DG
Cost function coefficients of generation unit i max
PPV;i Maximum PV output power in kW
ai;j Unique continuous variable for each segment in min
PPV;i Minimum PV output power in kW
piecewise linearization of bivariate functions PSTC PV output power in standard temperature
Dhk Head gain of pump k in meters max
PW;i Maximum wind turbine output power in kW
li Unique continuous value for each breakpoint in min
PW;i Minimum wind turbine output power in kW
piecewise linearization QD Demand flow in m3 =h
gi;j ; gi;j Unique binary variable for each upper/lower triangle QR Flow injected to the pipe from the reservoir in m3 =h
R
Qmax Maximum amount of water in the reservoir m3 =h
in piecewise linearization of bivariate functions
R
Qmin Minimum amount of water in the reservoir m3 =h
A Water flow index matrix
ak , bk , ck Pump’s characteristics constants QT Flow to the tank in m3 =h
C Factor for relative roughness of the pipe. Q0T Initial amount of water in the tank m3 =h
T
Qmax Maximum amount of water required in the tank
D Pipe diameter in meters
EiBatt Available energy in battery in kWh m3 =h
QminT Minimum amount of water required in the tank m3 =h
FiDG cost of ith DG plant in $
Fi ðPi Þ Cost function of generation units as a function of Pi Q m;n Vector of pipe flows between nodes m and n in m3 =h
Gac ðtÞ Solar irradiance a time t in kW=m2 rk Normalized speed of the pump
GSTC Solar irradiance at standard temperature in W= m2 Rmax
down
Round down maximum limit of DG unit i
hi Unique binary variable for each segment in piecewise Rmax
up Round up maximum limit of DG units
linearization Sk Status of pump k
Hm , Hn Pressure head at nodes m and n, respectively, in Sdn
i Shut down status of DG unit i
meters. Sup
i
Startup status of DG unit i
K Temperature coefficient of PV panel. t Time in hour.
KiPV ; Kiwind ; Kibatt Cost coefficients of PV, wind, and BESS unit Tc ; Ti Reference and current temperature of solar cells in + C
Lm;n Length of the pipe connecting nodes m and n in ux ; uy Binary variables introduced to linearize up/down
meters status of DG units
PiW Active power of ith wind generation unit in kW UiDG Status of DG unit i
PiBatt Active power of ith battery unit in kW UT; DT Minimum up/down time
PiDG Active power of ith DG in kW vi Cut-in wind speed in .
PiPV Active power of ith PV unit in kW vo Cut-out wind speed in m=s
max
Pi;chg Maximum battery charging power in kW vt measured wind speed in m=s
max
Pi;dis Maximum battery discharging power in kW vnom Nominal wind speed m=s
PiDG;max Maximum DG output power in kW Zm ,Zn Static head at node m and n in m.
PiDG;min Minimum DG output power in kW

integer non-linear programming (MINLP) formulation (Li et al., and Parvania, 2019). Nonetheless, most of the existing studies in
2018a). Another study proposed a co-optimization model of wa- energy efficiency and optimization of water-energy networks
ter and energy systems using non-linear mathematical program- ignored the fact that RE sources play an important role in
ming (Santhosh et al., 2014). The energy system, however, was only addressing the energy needs of future energy infrastructure and
composed of grid-connected power plants and renewable energy considered conventional fossil-fuel-based power plants in their
(RE) sources were excluded. In a different example, the impact of formulation.
various management strategies on energy consumption and cost of Microgrids have been introduced as promising solutions to
two water utilities was investigated via energy and water quality provide a flexible and cost-efficient energy in a decentralized
management systems (EWQMSs) (Cherchi et al., 2015). The results manner to isolated communities (Li et al., 2020; Nazari-Heris et al.,
offered insightful management and governing remarks about 2020) or when a natural disaster results in power shutoff from the
water-energy nexus, but did not discuss modeling and optimization utility grid (Rezvani et al., 2015; Yin et al., 2017). Water-energy
of energy sources (whether RE or conventional). Also, a graph microgrids are propitious solutions that can be applied to a wide
theory-based network was developed to represent the water and range of geographic or economic contexts such as islanded com-
energy flows as a directed graph (Tsolas et al., 2018). To optimize munities (Guo et al., 2014; Lu et al., 2017). There exists a few
the water and energy generation, an optimization model was research on modeling the operation and performance of water-
developed to remove the redundant sub-graphs that would result energy microgrids. For example, the techno-economic potential of
in maximum yield of water while minimized the energy con- a microgrid serving an industrial-scale drinking water plant was
sumption in conventional energy systems. Mixed integer linear studied in (Soshinskaya et al., 2014) and it was concluded that
programming models were used in (Oikonomou and Parvania, adding solar and wind generation to the microgrid increases the
2018; Oikonomou et al., 2018) to co-optimize the operation of benefits received from the demand response strategy. The study,
water distribution systems in a conventional power system. In a however, did not include details of water distribution network and
recent study, an optimization model was proposed to co-optimize a mainly focused on energy supply modeling. In (Palma-Behnke
water desalination facility with variable speed pumps, water tanks, et al., 2011), a neural-network-based energy management system
and a conventional thermal power plant-based energy system to was developed to minimize the cost of operation while supplying
provide day-ahead energy and frequency regulation (Oikonomou the water and electric load demands. However, the water
F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776 3

distribution system was treated as a constant load and detailed disadvantages of each model are discussed in terms of compu-
modeling of variable speed pumps and constraints of water dis- tational efficiency and optimal error,
tribution network were ignored. A detailed legacy infrastructural 3 The energy usage of a water distribution network with fixed-
component study was conducted to investigate the parallel con- and variable-speed pumps in an islanded water-energy micro-
cepts of electric and water distribution for resiliency of future grid system with renewable energy sources and storage units is
water-energy microgrids (Falco and Webb, 2015). Optimal sched- minimized on a daily basis,
uling of energy sources or water distribution network, however, 4 Distributed generation (DG)’s ramp rate limit, on/off status,
was not discussed in this work. A facility at University of Arizona startup and shutdown statuses, and minimum up/down times
was used by National Renewable Energy Laboratory (NREL) in (Daw are considered in the formulation of the optimization problem
et al., 2018) to investigate the feasibility of water-energy microgrids for islanded water-energy microgrid,
in providing demand response and flexibility in energy supply us- 5 A detailed mathematical model was considered for the pumps in
ing renewable energy sources. The study merely focused on the water systems relating the head gain and pump’s speed to
experimental validation and demonstration of capabilities of the consumed power, this realistic model makes the output
microgrids in providing demand response rather than optimizing power of pump as a bivariate function of pump’s speed and head
the schedule of water-energy network simultaneously. Optimal gain,
demand-side management of a micro water-energy network was 6 To realistically model the water level in the tank, which is
studied in (Li et al., 2018b). It was concluded that by controlling controlled by a pump, a constraint representing the relationship
water loads, the grid flexibility is significantly increased. The pro- between the status of pump and tank level is formulated,
posed nonlinear optimization model was converted into a mixed- 7 Effect of pump’s speed on resource allocation of water-energy
integer convex programming problem using quasi-convex hull microgrid is studied through case studies.
relaxation. The studied method considered pumps with constant
speed, which simplified the pump’s output power as a univariate The rest of the paper is organized in the following order. Section
function of head gain. In practice, pump’s output power is a 2 explains the modeling procedure for the optimization of the
bivariate function of head gain and pump’s speed. In a recent study, micro water system, section 3 discuses the electricity generation
optimized scheduling of microgrids with high-inertia thermal dy- modeling of the micro energy system supporting the micro water
namics and lossless water storage system was proposed (Nguyen system, section 4 presents the co-optimization model of an inte-
et al., 2020). It was concluded that energy saving was achieved if grated water-energy microgrid system, section 5 elaborates the
water and energy demands are preserved at the same time. linearization of water and energy models using piecewise linear
While the existing research has made some progress in functions, section 6 includes the case studies and results, and sec-
advancing water-energy microgrids research, some limitations still tion 7 concludes the paper.
exist that need to be addressed for cooperation of these two critical
infrastructures. First, existing studies ignored a detailed modeling 2. Micro water system model
in both water and energy networks. When the goal is to optimize
the operation of water-energy nexus, water and energy networks The water-energy microgrid system is illustrated in Fig. 1. The
should be optimized as a whole, and thus, all the properties, ele- energy network includes an aggregated wind farm with 500 kW
ments, and constraints of both systems must be built in the opti- capacity, a battery energy storage system (BESS), a conventional
mization model. Second, prior studies on water-energy distributed generation (DG) unit, a 500 kW solar PV farm, and
optimization (e.g. (Oikonomou and Parvania, 2018; Oikonomou aggregated electrical loads. The wind farm generates electricity
et al., 2018; Oikonomou and Parvania, 2019)) often neglected that depending on the available wind speed on a daily basis. The con-
future energy supply will come from isolated energy supplies with ventional DG will be scheduled by a unit commitment problem
the least need for the fossil-fuel-based power plants (Mariam et al., including the ramp up/down limit rates, on/off status, and mini-
2016). Consequently, islanded microgrids with renewable energy mum up/down times. Furthermore, the solar PV generates elec-
sources and storage units should be central to the energy produc- tricity depending on the available solar irradiance and outdoor
tion of the water-energy network optimization models. Resource temperature. The BESS is a bi-directional system that can provide a
allocation in islanded mode of operation significantly depends on demand response capability or store the excess energy for demand
the availability of renewable energy sources and simultaneous response purposes. The water distribution network includes seven
optimization of water demand and electric loads. Third, existing nodes, a reservoir, a tank, a pump and five residential nodes sup-
studies mainly simplified the problem by considering constant- plying water to houses. On/off status of pump is considered in the
speed pumps in water distribution networks and ignored the formulation to allow pumps to turn off when the power demand is
bivariate non-linearities in pump’s output power formulation. met. Summation of residential water usage at nodes 3e7 will be the
Furthermore, the relationship between the status of pumps close to water demand of the water-energy system. The pump’s speed is
the water tanks in controlling the tank’s level was not considered. considered variable so that it can be adjusted to save the electricity
As such, these detailed models should be added to the optimization cost of the entire system. While a pump’s characteristic curve is
problem to result in a more accurate resource allocation and energy obtained for a certain motor speed and impeller diameter, it can be
management. To address these limitations, a novel formulation is developed in a way that is adaptable for any speed and any diam-
developed in this study to optimize the operation of a water-energy eter. Such is accomplished by applying the affinity laws (Walski
microgrid such that. et al., 2003). The affinity laws present a direct relationship be-
tween the pump’s flow rate and its speed, and a direct correlation
1 Piecewise linear approximation of bivariate and univariate between the pump’s head and the square of its speed. This means
functions are used to convert the mixed integer non-linear that once the pump’s characteristic curve is obtained for a given
optimization model of the system to a globally optimal mixed speed, variable speeds can change the pump’s head and flow rate.
integer linear programming (MILP) model, Thus in this paper, a variable speed is contemplated for the pump so
2 A detailed comparison is carried out to compare the results of that the pump’s head and flow rate could be adjusted to utilize the
proposed MILP model with MINLP model, and advantages and lowest electrical power for a given water demand at any hour.
Pump’s electricity consumption is considered as the electricity
4 F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776

Fig. 1. Structure of the studied water-energy microgrid system.

demand of the energy network, which will be optimized in the power usage will be considered as one of the micro energy system’s
following section. The water sector is designed based on the tank’s input loads. Equation (1) represents the electricity usage of kth
and pump’s operation and scheduling, the hydraulic elements and pump varying with the pump’s flow rate and its head gain at hour t
physical properties such as flow rates, nodal heads, pipes’ lengths (Ulanicki et al., 2008)
and diameters, friction factor, and water demand diurnal pattern.
To validate the proposed models, the water-energy microgrid k
Ppump;t ¼ Skt BQtm;n Dhkt (1)
system shown in Fig. 1 with seven water distribution nodes and a
common AC bus with distributed generation is designed. All 
models are solved with the “OPTi” solver in MATLAB software. The 1 pump k is ON
Skt ¼ (2)
micro water model is designed to minimize the electricity usage, 0 otherwise
while the hydraulic demands are met at all the nodes in the dis-
tribution system. To do that, a supervisory control and data where B is a constant representing the specific weight of water at
acquisition (SCADA) system is implemented in the water distribu- 20  C and the efficiency of the pump (assuming it runs at 80% ef-
tion network to collect data related to flow rates and status of the ficiency) and Dhkt is the head gain, which is the dissipated pump’s
pump via smart meters. The developed optimization model will energy gained by the water. This energy will help the water rectify
result in a daily operation schedule for the pump which minimizes the head loss occurring due to the pipes’ friction or fittings. In (1), Skt
the cost of electricity consumption while meeting the daily water is a binary variable to indicate the on/off operating modes of the
demand. This schedule will be sent back to the SCADA in the form of
pump at hour t and Qtm;n is the flow of the pipe from node m to node
assignments. Following the given pattern or assignments, the
n. When a pump is ON, the flow of the associated pipe, Qtm;n , is
SCADA controls the physical elements of the system including
water levels in the tank and reservoir, opening/closing valves, or equal to the flow of the pump and when the pump is OFF, the pipe is
switching the pumps on and off, which can eventually conclude considered to be closed. Also, Qtm;n  0 to ensure the water will not
energy optimized conditions. Such cyber-physical elements in the flow backwards through the pump. The head gain, Dh, is obtained
water network can be integrated into the micro energy network, from equation (3)
also a cyber-physical unit, where the physical components of the
 2  2
two water and energy systems are operated together within one
Dhkt ¼ ak Qtm;n þ bk Qtm;n rtk þ ck rtk (3)
single system. Since pumps are the most energy intensive in a
water distribution network, the output of the micro water opti-
mization model, i.e. the active power usage of the hydraulic pump, where ak ; bk ; ck are the pump’s characteristics coefficients (Ulanicki
will represent the total power consumption of the micro-water et al., 2008). A pump is installed in a water distribution system to
system. Several papers have focused on managing power con- serve two main purposes: 1) to compensate elevation differences
sumption of pumps in residential applications (Al-Sumaiti et al., (also known as static head) and 2) to rectify head losses due to
2017) or scheduling of water pumps and their impact on elec- pipe’s friction and fittings. This is done by adding some amount of
tricity consumption (Al-Sumaiti et al., 2019c, b). The readers are head by the pump to the water. While the static head is indepen-
encouraged to refer to these papers for more information on pump dent of the pump’s flow rate, the head loss is highly dependent on it
scheduling and achieving energy efficiency by demand response. In (this is explained in details later in (7)). Therefore, the pump’s head
section 3 of this paper, i.e. the micro energy system modeling, this is a function of its flow rate and by extension, under affinity laws, is
a function of the speed (Walski et al., 2003). Due to the fact that the
F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776 5

flow rate changes hourly with demand, the pump’s speed is also
considered to be variable. As shown in (3), the head gain is a Lm;n
F m;n ¼  1:85 (8)
quadratic function of the water flow rate passing through the 0:278*C*D2:63
m;n
nkm
pump, as well as the normalized speed of the pump (rtk ¼ nknom
,
where nm is the measured speed and nnom is the nominal speed of where Lm;n and Dm;n are the length and diameter of the pipe con-
the pump, both in revolutions per minute (rpm)). The gain (short necting nodes m and n, respectively, and C is the pipe’s roughness in
for energy head gain) is the dissipated pump’s energy gained by the Hazen-Williams equation that captures the effect of friction in the
water. pipe, assumed to be 100 for this study, which is the pipe’s rough-
The constraint (4) states that at any hour, the amount of water ness for a 20-year old cast iron pipe.
supplied by the reservoir to a micro water network is equal to the
water accumulated in the water tank plus the amount consumed 3. Modeling of microgrid
(in the form of water demand) plus the amount flowing in the
pipes. This is known as conservation of mass theorem, which is To obtain the optimal power dispatch of a water-energy
represented as microgrid, the economic dispatch model is developed as a co-
optimization problem. The energy generation unit contains a
microgrid that operates with an aggregated DG unit, solar and
X
N wind energy sources, and energy storage devices. Moreover, the
A:Qtm;n  QtR þ QtT ¼ QtD (4) electrical loads consist of a constant 24-h daily load plus the
m;n¼1 active power consumption of the water system as illustrated in
Fig. 1. Parameters of the microgrid system can be found in
where Q R is the flow injected to the pipe from the reservoir in m3 = Table C1.
h, Q T is the flow to the tank in m3 =h, and Q D is the water demand in
m3 =h which is equal to the summation of water usage at each node 3.1. Conventional generation units
(Houghtalen et al., 2016). The left-hand side of equation (4) exhibits
the water flow in the network of pipes, where A is the index matrix The input to a conventional generation system is fuel and the
representing the water flow between nodes such that, output is the electrical power measured in kW. By drawing the
input-output curves of a conventional generation unit, also known
as heat-rate curve, and representing the fuel consumption in $/h,
8
< 1 water flows from node m to n the fuel-cost curve is calculated, which is represented as a
Am;n ¼ 1 water flows from node n to m quadratic function of real power generation (Saadat, 1999). The
: cost function of a distributed generation (DG) system (Fi;t DG ) is
0 no water flows between nodes m and n
(5) represented as

To keep the water level in a tank within the minimum amount  2


T ) and the
DG
Fi;t ¼ aDG
i þ bi Pi;t þ gi
DG DG DG DG
Pi;t (9)
required to supply the network for one day (Qmin
maximum amount allowed to avoid overflowing the tank (Qmax T ),
DG is the cost of ith DG plant in $ (Wood et al., 2013). In (9),
where Fi;t
the status of the pump connected to the tank can be controlled such
that ai , bi , and gi are cost coefficients of generation units, which nor-
mally are given. To represent the constraints of the DG units in the
economic dispatch, a unit commitment formulation is utilized,
8
< if which is formulated in (10)-(14). This problem is well-studied in
Q0T þ QtT < Qmin
T
Turn the pump ON
(6) economic dispatch of smart grid systems and the readers are
: if Q0T þ QtT > Qmax
T
Turn the pump OFF encouraged to refer to (Morales-Espan ~ a et al., 2013) for more in-
formation. Equation (10) is the power limit constraint of DG unit i
In this equation, Q0T is the initial water level in a tank in m3 = h. and Ui;t DG 2f0; 1g is a binary variable indicating on/off status of DG
Water tanks are always set up at a high elevation to provide suffi- unit i. Ramp up and ramp down limits of the DG units are expressed
cient nodal pressure for the network. Therefore, a pump will be in equations (11) and (12), respectively. Also, two binary variables,
installed to provide the extra energy required to overcome the up
Si;t and Sdn
i;t , are added to represent the startup and shut down
negative static head between the two points. This can be expressed
by Bernoulli’s equation in (7) (Houghtalen et al., 2016), status of DG units, respectively, which are expressed in (13) and
(14).

   1:85 DG DG;min DG DG DG;max


Ui;t Pi  Pi;t  Ui;t Pi (10)
Hm;t  Hn;t þ Zm;t  Zn;t þ Dhkt ¼ Fm;n Qtm;n (7)
DG DG
which represents the difference between the energy head of two Pi;t  Pi;t1  Rmax DG
up Ui;t1 (11)
nodes of m and n in the water network is equal to the head loss
between the two nodes. This head loss can be compensated by the DG DG
Pi;t þ Pi;t1  Rmax DG
down Ui;t (12)
head gain from the pump(s) existing on the pipe(s) connecting the
two nodes (Houghtalen et al., 2016). In equation (7), Hm and Hn  
present the head (short for pressure head) of nodes m and n, Sup
i;t
DG
¼ Max 0; Ui;t DG
 Ui;t1 ; ci 2 DG; ct2T (13)
respectively, in meters, Zm and Zn are the elevations at nodes m and
n, respectively, in meters. According to Hazen-Williams equation,  
the head loss of a pipe (F m;n ) can be computed using equation (8) Sdn DG DG
i;t ¼ Max 0; Ui;t1  Ui;t ; ci 2 DG; ct2T (14)
(Houghtalen et al., 2016),
6 F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776

Equations are then converted to a standard linear programming cells in  C, respectively. The temperature data for April 11, 2020 was
format using (15) and (16). also adopted from the Penn State University (PSU) SURFRAD sta-
tion’s website (Global Monitoring Laboratory, 2020), which is
up
min
up
Si;t located on the grounds of PSU’s agricultural research farm. The 24-
Si;t
h temperature, solar irradiance, and output power of PV unit are
s:t:Sup
i;t
0 illustrated in Fig. 2.
up DG DG (15) The cost function of solar photovoltaic (PV) generation system is
Si;t  Ui;t  Ui;t1
represented by
up
Si;t  Mux  
PV PV
up DG DG Fi;t Pi;t ¼ KiPV Pi;t
PV
(21)
Si;t  Ui;t  Ui;t1 þ Mð1  ux Þ

where KiPV is a fixed purchased power generation coefficient of PV


min Sdn
i;t unit i in $/kW, which was adopted from (Jin et al., 2017). Some more
Sdn
i;t
sample cost coefficients of solar system can be found in (Al-Sumaiti
s:t:Sdn
i;t  0 et al., 2019d). The parameters of the PV system are provided in the
Sdn DG DG (16) Appendix section. The PV power generation is restricted to the
i;t  Ui;t1  Ui;t
maximum and minimum available power, represented by
Sdn
i;t  Muy
  min
PPV;i;t PV
 Pi;t max
 PPV;i;t (22)
Sdn DG DG
i;t  Ui;t1  Ui;t þ M 1  uy

max is usually expressed using the daily weather forecast of


where PPV;i;t
where ux ; uy 2f0; 1g are two new binary variables and M is a large
number. It is noted that
up
Si;t and Sdn the output PV power. Several studies focused on modeling uncer-
i;t will be added to the objective
tainty of PV generated power (Al-Sumaiti et al., 2019a) or optimally
function of the co-optimization problem with a fixed startup and
tracking and modeling the PV power output (Obukhov et al., 2020),
shutdown cost (Cup and Cdn ) and the other constraints in (15) and
however, a sample 24-h data was used in this study based on
(16) will be added to the constraints of the co-optimization
sample solar irradiance data.
model. The relationship between the generator’s on/off status,
startup, and shutdown status is called the logical DG constraint,
which is formulated by 3.3. Wind generation cost
DG
Ui;t DG
 Ui;t1 ¼ Sup
i;t
 Sdn
i;t (17) The wind power plant’s output generated power is highly un-
certain and is characterized as a random function of wind speed.
Minimum up and down time limits for DG unit i are also added
The wind turbine’s output power is a function of wind’s speed,
to the optimization problem, which are expressed as
expressed by
X
t 8
Stup;i  Ui;t
DG
; t2fUT; …; Tg > if vt < vi and vt > vo
(18) >0
>
>
i¼tUTþ1 < vt  vi
W
Pi;t ¼ Pnom if vi  vt  vnom (23)
>
> vnom  vi
X >
>
t :P if vnom  vt  vo
Stdn;i  1  Ui;t
DG
; t2fUT; …; Tg (19) nom
i¼tDTþ1
W is the wind power at time t, v is the wind speed at time t
where Pi;t t
where UT and DT are minimum up and down times of DG unit i in m/sec, vnom is the nominal wind speed in m/sec, vi is the cut-in
(Rajan et al., 2005). Parameters of the unit commitment problem wind speed in m/sec, vo is the cut-out wind speed in m/sec
for the DG were adopted from (Wood et al., 2013; Rajan et al., (Hetzer et al., 2008). In this study, a sample 24-h wind data was
2005), which can be found in Table C1 in the Appendix section. adopted for April 11, 2020 from the Pennsylvania State Climatolo-
gist website that provides detailed climatic information (The
3.2. Solar generation cost
15
T (° C)

The proposed microgrid model has an aggregated 500 kW PV


10
unit. The 24-h PV power profile was calculated via converting a
c

sample solar irradiance data into output power. The daily irradiance 5
5 10 15 20
(W/m2 )

data was adopted from (Global Monitoring Laboratory, 2020) for


April 11, 2020 and the output power of the PV was calculated using 1000
(20), 500
ac,t
G

Gac;t 0
PV 5 10 15 20
Pi;t ¼ PSTC ð1 þ KðTc  Ti ÞÞ (20)
GSTC
(kW)

500

250
PV

wherePiPV is the active power of ith PV unit in kW,PSTC is the PV


t
P

output power in standard temperature condition (STC),Gac;t is the 0


solar irradiance at time t, K is the temperature coefficient of the PV 5 10 15 20
Time (hour)
panel, which is equal to  0:0047(Hijjo et al., 2017). In addi-
tion,Tc and Ti are the reference and current temperatures of solar Fig. 2. 24-h temperature, solar irradiance, and PV output power data.
F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776 7

Pennsylvania State Climatologist, 2020). Using the sample wind


speed data for 500 kW aggregated wind farm, cut-in, cut-out, and
Batt
Ei;t Batt
ðtÞ ¼ Ei;t1 Batt
þ Pi;t Dt (28)
nominal wind speed of 5, 45, and 15 m/s, respectively (Hetzer et al.,
2008), the wind farm’s output power data was calculated from (23) where Dt is the step size of the economic dispatch, which is
for a 24-h operating period. The results are illustrated in Fig. 3. considered as 1 h in this study. The energy level of the battery
The following equation presents the cost function of the wind should also be limited between the minimum and maximum
generation system (Augustine et al., 2012) available thresholds (Xiaoping et al., 2010)
  Min Batt Max
W
Fi;t W
Pi;t ¼ Kiwind Pi;t
W
(24) EBatt;i  Ei;t  EBatt;i (29)

The cost function of battery energy storage unit is expressed by


and Kiwind is the fixed purchased power coefficient of the wind  
Batt Batt
generation system, which was adopted from (Jin et al., 2017). Fi;t Pi;t ¼ Kibatt Pi;t
Batt
(30)
Similar to solar power, the wind energy generation is also bound to
the maximum and minimum available power, represented by
and Kibatt is a fixed purchased power coefficient for the battery
min W max storage, and is adopted from (Jin et al., 2017). In this study, the
PW;i;t  Pi;t  PW;i;t (25)
strategy for charging and discharging the battery is to ensure the
max is the maximum available wind power at time t, which is demand is always met. The storage will be charged to keep the
and PW;i;t
energy between the maximum and minimum limits as forced by
calculated by the sample wind speed data using (23) and is illus- the battery constraints. The charging happens with the cheapest
trated in Fig. 3. available energy in the micogrid during the hours that the demand
is completely met and more energy is available. More information
can be found in (Xiaoping et al., 2010; Pazheri et al., 2012;
3.4. Battery energy storage cost Mahmoodi et al., 2015; Jin et al., 2017).

In the economic dispatch model, the energy storage can operate


4. Co-optimization of micro water and microgrid
in both charging and discharging modes. Additionally, the model
incorporates the state of charge (SoC) of the storage unit to account
In the co-optimization model, the main objective is to minimize
for the availability of the battery. The storage operation is formu-
the total cost of power generation in an integrated water-energy
lated as the following (Xiaoping et al., 2010)
microgrid system shown in Fig. 1. The optimization problem de-
8 cides on choosing the cheapest energy resource depending on its
>
> Batt
>0 ðCharging ModeÞ
>
> if Pi;t availability (during the day), amount of hourly load, and cost of
<
if Batt
Pi;t <0 ðDischarging ModeÞ (26) generation. The proposed system is designed for an islanded water-
>
> energy microgrid, it resembles an islanded community that is
>
> Batt
: if Pi;t ¼ 0 ðStandby ModeÞ supported by a microgrid. The microgrid controller will manage all
the resources in this system at the point of interconnection. The
Batt is the storage power in kW at time t. The constraint
where Pi;t microgrid controller defines the operating points of the system
considering customers’ needs (Ton and Reilly, 2017). The co-
associated with charging or discharging mode of the battery can be
optimization model, presented by (31), aims to simultaneously
calculated by (Xiaoping et al., 2010; Eyisi et al., 2019)
minimize the overall cost of the combined micro water and energy
Max Batt Max systems, while meeting all the constraints of each individual sys-
Pi;chg  Pi;t  Pi;dis (27)
tem. In this model, the power requirement of the water network
becomes one of the constraints, in the form of an electrical load, to
Max and P Max are the maximum charging and discharging
where Pi;chg i;dis be met by the objective function.
powers available at time t. The dynamics of the energy generation
of the battery, at time t, can be represented by (Xiaoping et al., XTi X
NG;i

2010; Eyisi et al., 2019) Min DG


Fi;t PV
þ Fi;t W
þ Fi;t Batt
þ Fi;t þ Cup Sup
i;t
þ Cdn Sdn
i;t (31)
t¼1 i¼1

12 X
N G;i NX
pump
DG PV W Batt k
Pi;t þ Pi;t þ Pi;t þ Pi;t ¼ Pload þ Ppump;t (32)
v (m/sec)

10 i¼1 k¼1

8
t

6
5 10 15 20 5. Linearization of non-linear functions in the MINLP model
400
5.1. Linear approximation of cost functions in energy systems
(kW)

200 (univariate functions)


w,t
P

Cost functions of DGs expressed in (9) are univariate functions.


0
5 10 15 20 Piecewise linear approximation can be achieved by dividing the
Time (hour) overall cost function to n breakpoints (or sampling coordinates)
and approximating each segment with linear segments
Fig. 3. 24-h wind speed and output power data. ½ðPi ; Fi ðPi ÞÞ; ðPiþ1 ; Fi ðPiþ1 Þ; ði ¼ 1; …n 1Þ as illustrated in Fig. 4.
8 F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776

(39) ensure that only two consecutive l values take nonzero values
referring to (33). Equations 35e39 represent the linear approxi-
mation of generator cost functions for each DG unit.

5.2. Linear approximation of water system constraints (bivariate


functions)

In this section, the non-linear behavior of pump powers and


Bernoulli’s equation will be approximated using linear functions.
Let us define
 
DP Qtm;n ; rtk ¼ BQtm;n Dhkt (40)
Fig. 4. Piecewise linear approximation of a univariate cost function.

By replacing Dhkt from (3), the non-linear bivariate function of


For any given active power generation, e.g., Pi  P i  Piþ1 , the pump power is represented as

cost function value (Fi ðP i Þ) can be approximated by convex com-   h  2  2 i


bination of Fi ðPi Þ and Fi ðPiþ1 Þ via a unique continuous value, l2 DP Qtm;n ; rtk ¼ BQtm;n ak Qtm;n þ bk Qtm;n rtk þ ck rtk (41)
½0; 1 for each breakpoint i, such that
k
The pump’s output power is then represented by Ppump;t ¼
P i ¼ li Pi þ ð1  li ÞPiþ1 (33) Skt DPðQtm;n ; rtk Þ, which is a non-linear combination of pump’s
bivariate function (DPðQtm;n ; rtk Þ) and binary variable Sk , which can
Fi;t ðP i Þ ¼ li Fi ðP i Þ þ ð1  li ÞFi ðPiþ1 Þ (34) be linearized using Big M method (Copeland and Carter, 2014) as
the following
The above equations can be written for a standard linear pro-  
gramming solver by defining model variables and constraints that MSkt  DP Qtm;n ; rtk  MSkt (42)
force any Pi to be associated with an appropriate pair of consecutive
breakpoints. Each breakpoint will be associated with a continuous      
variable l2½0; 1, and each segment will be associated with a binary k
Ppump;t  M 1  Skt  DP Qtm;n ; rtk  Ppump;t
k
þ M 1  Skt
variable, h, (e.g., for segment i in interval ½Pi ; Piþ1 , li 2 ½0; 1 and a
(43)
binary variable hi are defined). Therefore, for n breakpoints and n
1 segments, n continuous values and n  1 binary variables are The bivariate function DPðQtm;n ; rtk Þ should then be approxi-
defined. The readers are encouraged to refer to (D’Ambrosio et al., mated to a linear function using piecewise linearization technique.
2010) for more information on piecewise linear functions and Next, referring to (7), let us consider DHðQtm;n ; rtk Þ as a bivariate
their MILP representation for univariate functions. The piecewise
function representing the Bernoulli’s equation.
linear approximation of cost function Fi ðPi Þ is expressed using the
following constraints
     1:85
DH Qtm;n ; rtk ¼ Hm;t  Hn;t þ Zm;t  Zn;t ¼ Fm;n Qtm;n  Dhkt
X
n
(44)
x
Pi;t ¼ li Pi ; x ¼ DG1 ; DG2 ; …; DGN (35)
i¼1 By replacing Dhkt from (3), one can find
   1:85  2  2
  X n
DH Qtm;n ; rtk ¼ Fm;n Qtm;n  ak Qtm;n  bk Qtm;n rtk  ck rtk
x
Fi;t Pi;t ¼ li Fi ðPi Þ (36)
i¼1 (45)

X
n1 which is another bivariate function that needs to be approximated
hi ¼ 1 (37) by piecewise linearization technique to represent the MINLP model
i¼1 by a linear programming model.
The approximation process can be explained by taking a look at
li  hi1 þ hi (38) bivariate functions plotted in Fig. 5 where x sampling coordinates
on Q-axis and t sampling coordinates on r axis can be considered.
X
n The functions DPðQtm;n ; rtk Þ and DHðQtm;n ; rtk Þ can then be evaluated
li ¼ 1 (39)
i¼1

In the above equations, (35) expresses the fact that any given
point Pi  P i  Piþ1 can be represented as a linear combination of
two consecutive breakpoints weighted by li as represented in (33).
In this equation, liþ1 ¼ 1  li . Equation (36) is a linear approxi-
mation of non-linear function Fi;t x as a convex combination of

function values at breakpoints weighted by li referring to (34) and


considering the fact that only two consecutive l values will be
nonzero only for any point. Furthermore, (37) ensures that only one
binary variable h out of n  1 takes a value of 1. Constraints (38) and Fig. 5. Piecewise linear approximation of bivariate functions (DH and DP).
F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776 9

for each breakpoint (Qi ; rj )(i ¼ 1; …; x, j ¼ 1; …; t). For any given intervals ½Qi ; Qiþ1 Þ and ½rj ; rjþ1 Þ in Fig. 5, binary variables gi;j and gij
point say (Q ; r) with Qi  Q  Qiþ1 and rj  r  rjþ1 , a rectangle of and will be associated to the upper and lower triangles in the right
vertices ðQi ;rj Þ, ðQiþ1 ;rj Þ, ðQi ;rjþ1 Þ, and ðQiþ1 ; rjþ1 Þ can be considered side rectangle. Furthermore, (54) imposes that only non-zero aij
(see Fig. 5). By drawing a diagonal for this rectangle, two triangles values are the ones associated with three vertices of the triangle
known as upper and lower triangles are formed (see the right-side chosen in (53). Therefore, Equations (48)-(54) are linear approxi-
plot on Fig. 5). The lower triangle is differentiated by a binary mation of (1), (3), and (7). Furthermore, equation (16) represents
variable g and the upper triangle is differentiated by g. The function the non-linear behavior of tank level, which can be linearized using
value at (Q ; r) is approximated by convex combination of the the big M method as the following (Copeland and Carter, 2014)
function values evaluated at the vertices of the triangle containing  
T
Qmin  MSkt < Q0T þ Q T < Qmax
T
þ M 1  Skt (55)
(Q ; r). For example, if the bivariate function is considered to be
DPðQtm;n ; rtk Þ, the function’s value at (Q ; r) is represented by
According to equation (55), when Skt ¼ 1, the left-hand
      inequality does not impose any constraint, and thus Q0T þ
DPðQ ; rÞ ¼ li DP Qi ; rj þ mj DP Qiþ1 ; rjþ1 þ 1  li  mj Pn
Q T < Qmax
T . Otherwise, the right-hand inequality does not impose
(46) T < Q T þ Q T . The tank flow is therefore
any constraint, meaning Qmin 0
8   bounded by the minimum and maximum amounts. Consequently,
> rjþ1  rj
< DP Qiþ1 ; rj if r  rj þ ðQ  Qi Þ the overall optimization model is converted to a linear program-
P¼ Qiþ1  Qi (47)
>   ming model, which has a global feasible solution.
:
DP Qi ; rjþ1 Otherwise
6. Case studies
where li 2½0; 1, mj 2½0; 1, and ð1 li mj Þ2½0; 1 are continuous
variables representing the weights for the convex combination of A water-energy microgrid presented in Fig. 1 is studied. It is
the vertices of the appropriate triangle which includes the point (Q ; noted that the system size in an islanded water-energy network is
r). The above convex combination of the bivariate function can be normally limited to a few nodes. However, for large-scale imple-
extended to DHðQtm;n ; rtk Þ similarly. In linear programming, (46) and mentation of the economic dispatch problem in a smart grid system
(47) and two equations developed for linear approximation of with hundreds of nodes, the readers are encouraged to refer to
(Hansen et al., 2015; Mokarram et al., 2019).
DHðQtm;n ; rtk Þ are formulated by introducing x  t continuous vari-
The electrical loads accounted in the micro-energy system
ables ai;j 2½0; 1 and extending the equations to a three-
indicate the total electricity consumption of micro-energy system
dimensional space such that
over a 24-h period of operation. A typical profile for the load of a
x X
t
community is adopted from (Mariam et al., 2016) and demon-
X
ai;j ¼ 1 (48) strated in Fig. 6.
i¼1 j¼1 The total water demand in the studied system is equal to the
summation of all the demands at the nodes where the end-
customers take out the water (i.e.
x X
X t
QD ¼ QD3 þ QD4 þ QD5 þ QD6 þ QD7 in Fig. 1). It is worth mentioning
Qtm;n ¼ ai;j Qi (49)
i¼1 j¼1
that the water demand is the water that is taken out of the system
at each node, and will not be recycled back into the water-energy
microgrid but instead will enter the wastewater system. The pa-
x X
X t
rameters of the system are provided in the Appendix section.
rtk ¼ ai;j rj (50)
Two optimization problems will be considered; 1) optimization
i¼1 j¼1
of water demand (“Opt W”), where the objective function is to
minimize the power consumption of the pump with constraints
  x X
X t   represented in (1)-(7) and 2) co-optimization of water-energy
DH Qtm;n ; rtk ¼ ai;j DH Qi ; rj (51)
microgrid (“Co-opt W-E00 ) represented in (31). For each model,
i¼1 j¼1
the non-linear functions in the optimization formulations were
approximated to MILP models using piecewise linear functions for
  x X
X t   n ¼ 3 (two segments) and n ¼ 11 (10 segments). The results ob-
DP Qtm;n ; rtk ¼ ai;j DP Qi ; rj (52) tained from linear formulations were then compared with the
i¼1 j¼1
original MINLP models. Table 1 provides the properties of each
model for micro water system optimization (Opt W) and water-
x1 X
X t1  energy microgrid system co-optimization (Co-opt W-E). In this
gi;j þ gij ¼ 1 (53)
i¼1 j¼1

P Load
aij  gi;j þ gi;j þ gi;j1 þ gi1;j1 þ gi1;j1 þ gi1;j (54) 600
Power (kW)

Constraint (48) denotes that there exists only two consecutive a 400
values with nonzero values. Constraints (49) (50) represent the
liner approximation of flows and pump speeds, respectively. Con- 200
straints (51) and (52) denote the linear approximation of DH and 0 5 10 15 20 25
DP, respectively. Constraint (53) represents the fact that any given Time (Hour)
solution will be located in only one triangle for the convex com-
bination. For example, consider the rectangle corresponding to Fig. 6. 24-h electrical load data.
10 F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776

Table 1
Problem properties of MINLP and MILP methods. MINLP MILP (n=3) MILP (n=11)

Power Consumption (kW)


300
Co-opt Opt

W-E W
# MINLP MILP MILP MINLP MILP MILP 200
of n¼3 n ¼ 11 n¼3 n ¼ 11
Vars 17 45 397 8 33 369
Bin vars 4 24 256 1 19 243
Const 56 155 1213 22 106 1114 100
Lin ineq 14 33 155 4 18 130
Lin eq 4 9 9 1 4 4
Nonlin equ 3 0 0 2 0 0 0
Bounds 31 89 793 14 19 737 0 2 4 6 8 10 12 14 16 18 20 22 24
Time (hour)

table, the number of decision variables (vars), the number of binary Fig. 8. 24-h power consumption of micro water via MINLP and triangle method MILP
variables (bin vars), the number of constraints (const), number of with n ¼ 3 and n ¼ 10.

linear inequality (lin ineq), linear equality (lin eq) and non-linear
equality (nonlin equ), and number of bounds are given.
water usage, following the peak hours of the water demand. The
As shown in Table 1, the original MINLP model includes 17
numerical results of these models are presented in Appendix B,
variables and 56 constraints for the co-optimization model and 8
Table B1. The total CPU time over 24-h period of operation associ-
variables and 22 constraints for the optimization of water demand
ated with MINLP model of the standalone water system is 1.043 s,
problem. Conversion of MINLP models to MILP has significantly
while it is 11.49 s in MILP model with n ¼ 3 and 3272 s in MILP with
expanded the size of optimization problems as it can be verified in
n ¼ 11. This is justified based on the high number of variables and
Table 1. For example, the MILP model with 11 breakpoints for the
constraints in MILP models, particularly with 10 segments. A
co-optimization of water-energy microgrid resulted in introducing
comparison between the results of the two MILP models clarifies
397 variables and 1213 constraints. Although the MILP models
that as the number of breakpoints increased, the results were closer
guarantee a global optimum, larger computing times are expected
to those from MINLP, thereby smaller error percentage (computed
for the central processing unit (CPU) for these models.
using (56)). As confirmed, in the MILP with n ¼ 11, the errors were
almost zero. This also agreed with the graphs in Fig. 8. However, as
6.1. Optimization of water and standalone energy networks expected from the size of the MILP with n ¼ 11 from Table 1, the
computing time of this model is very large, which is one of its
Water demand continuously fluctuates on an hourly and sea- drawbacks. None of the three models concluded any unsolved
sonal basis. During the day, the hours at which families are involved nodes within the time limit. The error was estimated using the
in more water-related activities, such as cooking and cleaning, following equation
show higher usage in the 24-h water demand profile. In the sum-  
mertime, the average daily usage of water is about 25% higher than Linear solution  Nonlinear solution
Error ð%Þ ¼ *100% (56)
the annual daily average (Viessman et al., 1998). Therefore, the Nonlinear solution
water demand calculations for this study was based upon an hourly
pattern of water usage throughout a summer day, with peaks
occurring at 7AM, 11AM, and 6PM (as shown in Fig. 7).
Solving the optimization model for the water demand (“Opt W00 )
6.1.1. Co-optimization of water-energy microgrid
resulted in power consumption of the studied water system (prior
Co-optimization of the studied water-energy microgrid using
to its integration with the energy system) for the water demand
both MINLP and MILP (n ¼ 3 and 11) models is done in this section.
shown in Fig. 7. Solution of the “Opt W00 problem for the daily water
Results of the daily operation of the system with each optimization
demand using MINLP and MILP models are presented in Fig. 8.
model are shown in Figs. 9e11, where the top subplots depict the
As can be seen in Fig. 8, the results of MINLP and MILP with 10
economic dispatch with a constant-speed pump set at 75% of its
segments (n ¼ 11) are perfectly identical, which is due to the larger
nominal speed, and the bottom subplots illustrate the effect of
breakpoints and thus tighter gap between the non-linear and linear
variable-speed pump on the daily dispatch.
lines within each segment. In addition, the hourly power con-
The network is developed to supply power to the electrical load
sumption in all three models mimicked the diurnal pattern of the
exhibited in Fig. 6 and the electrical consumption of the water
network (pump) shown in Fig. 8. The model would also result in
charging the BESS module if extra energy is available. It is evident
2500
from Figs. 9e11 that the daily contribution of DG is much less than
the generation of the renewable sources, which contributes to
Water Usage (m3 /h)

2000
reducing CO2 emissions, hence a cleaner power generation in all
1500 three case studies. In all three cases, the battery charging occurred
1000 during the hours of the day that the electrical demand was low and
extra renewable energy power was available, agreeing with the
500 pattern presented in Figs. 6 and 7. Comparing the results of power
generation profiles of MINLP and the two MILP models, one can
0
0 2 4 6 8 10 12 14 16 18 20 22 24 identify the almost identical patterns of MINLP and MILP with
Time (hour) n ¼ 11. It is also observed that minimum up/down limit constraints
of the DG unit have limited the scheduling of the DG throughout
Fig. 7. 24-h water demand for the studied network. the day. Therefore, the DG unit’s operation occurs during the initial
F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776 11

DG PV Wind Battery Load


DG PV Wind Battery Load

1000
Power (kW)

1000

Power (kW)
500
500

0
0

-500 -500
0 5 10 15 20 25 0 5 10 15 20 25

1000 1000
Power (kW)

Power (kW)
500 500

0 0

-500 -500
0 5 10 15 20 25 0 5 10 15 20 25
Time (Hour) Time (Hour)

Fig. 9. 24-h electrical load dispatch of water-energy microgrid for an MINLP co- Fig. 11. 24-h electrical load dispatch of water-energy microgrid for an MILP co-
optimization considering two conditions of constant pumps’ speed (top subplot) and optimization with n ¼ 11 considering two conditions of constant pumps’ speed (top
variable pump’s speed (bottom subplot). subplot) and variable pump’s speed (bottom subplot).

MINLP MILP (n=3) MILP (n=11)


DG PV Wind Battery Load 60

40
1000
20
Power (kW)

Cost ($/h)

500 0
0 5 10 15 20 25
60
0
40
-500
0 5 10 15 20 25 20

1000 0
0 5 10 15 20 25
Power (kW)

Time (hour)
500

Fig. 12. Total hourly cost of electricity generation via the studied methods.
0

-500
0 5 10 15 20 25 operation of the pumps in water distribution systems. Finally,
Time (Hour) compared to constant-speed cases, the variable-speed of pump
allowed more charging of energy storage during hours 12e16 for all
Fig. 10. 24-h electrical load dispatch of water-energy microgrid for an MILP co- three cases, which is justified by the fact that the power con-
optimization with n ¼ 3 considering two conditions of constant pumps’ speed (top sumption of pump is reduced at lower speeds.
subplot) and variable pump’s speed (bottom subplot).
The numerical results are presented in Appendix B, Table B2.
Similar to the standalone water system optimization, the MILP co-
hours of the day, when PV generation has not started yet and BESS optimization with n ¼ 11 has the largest computational times
does not have enough energy to supply the load, and during the (2647 s) among the three co-optimization models, which was ex-
night, when the PV generation becomes zero. It can also be pected based on the large number of variables and constraints
observed that batteries are charged during hours 1, 2, when the involved in computation. However, compared with the standalone
load is minimum and hours 12e15, when renewable sources have optimization, a considerable improvement in CPU time of the co-
excess power. The electrical load trend was almost the same for all optimization is observed, indicating a more efficient optimization
three models. Second row subplots in Figs. 9e11 show the effect of compared to standalone water optimization. The error percentages
variable-speed pumps on economic dispatch of the system. The in all three co-optimization models (computed from (56)) are zero
pump’s normalized speed (r) was allowed to change between 0.5 or close to zero.
and 1 during the day. Comparing the results with constant-speed Furthermore, the hourly cost of energy generation resulted from
pumps in Figs. 9e11, it is observed that the total demand has the three optimization models with variable and fixed pump speed
reduced significantly throughout the day. It is also noted that the are shown in Fig. 12. The upper plot demonstrates the hourly cost in
operation of DG has significantly reduced throughout the day, constant-speed scenario and the lower plot depicts the variable-
which contributes to reduction in CO2 emissions and cleaner pro- speed case study. As it can be seen, the hourly cost of operation is
duction. The peak demand reductions during hours 8, 12, and 18 enforced by the DG unit cost as it provides a more expensive cost
can also be observed in all three optimization models, indicating compared to PV, wind, or BESS. This can also be verified by looking
savings on daily peak demand charges with variable-speed at the economic dispatch of the DG unit in Figs. 9e11. In addition,
12 F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776

the results of the MILP with large number of breakpoints (n ¼ 11) which is an economic burden for these facilities to upgrade their
are very close to those of the MINLP model. There is a small dif- process to an autonomous control system. A combined water-
ference between the MILP with n ¼ 3 outcomes and the MINLP energy system’s operation offers significant savings that can
results, which agree with the error percentages provided in table justify the cooperation of interlinked water and energy sectors in
B2. Furthermore, the average hourly costs achieved from the future. Future work will focus on integration of water treatment
variable-speed pump are lower compared to the fixed-speed case facilities and grid-connection capabilities of the water-energy
study. For example, the maximum hourly cost of generation during microgrid systems in presence of uncertainties and load changes.
hours 7e10 and 17e20 is around $45, but the price is less than $40
for the same hours in variable-speed case (bottom subplot in Funding
Fig. 12). Therefore, significant savings can be achieved by incorpo-
rating variable-speed operation of pumps during the day in water This research did not receive any specific grant from funding
distribution networks. agencies in the public, commercial, or not-for-profit sectors.

6.2. Implications for theory and practice CRediT authorship contribution statement

The day-ahead economic dispatch proposed in the MINLP and Faegheh Moazeni: Conceptualization, Methodology, Software,
MILP co-optimization models result in significant cost reduction Validation, Formal analysis, Investigation, Writing - original draft,
and energy saving at the micro water-energy system due to the Writing - review & editing, Visualization. Javad Khazaei: Concep-
minimization of the power demand of the pumps, reduction of the tualization, Methodology, Software, Validation, Formal analysis,
peak demands, and allocation of the cheapest source for power Investigation, Writing - original draft, Writing - review & editing,
generation at any hour in the micro energy system. Additionally, Visualization.
incorporating renewable sources in the models helps moving the
micro water-energy system towards sustainability and clean power Declaration of competing interest
production, while contributes to carbon management and green-
house gas mitigation. This becomes a considerable source of energy The authors declare that they have no known competing
saving and cost offsetting parameter in metropolitan water-energy financial interests or personal relationships that could have
networks, where there are hundreds of pumps and several water appeared to influence the work reported in this paper.
tanks and reservoirs with high demands for electricity.
Appendix A
7. Conclusion
The base water demand, pressure head of the tank, pipe’s di-
In this paper, a mixed integer nonlinear programming (MINLP)
ameters, and pump’s flow are calculated from the following equa-
model was formulated to minimize the cost of operation of a water-
tions for Middletown, PA, where Penn State campus is located. In
energy microgrid. To add global optimality to the MINLP formula-
reality, these data are available as part of the SCADA recordings
tion, piecewise linear functions were used to approximate the
(Rasekh, Amin and Hassanzadeh, Amin and Mulchandani, Shaan
nonlinear behavior of cost function of conventional power plants,
and Modi, Shimon and Banks, M Katherine, 2016).
pump’s output power, and conservation of energy-mass as linear
The base water demand for each street is calculated as
functions. In addition, the MINLP model and MILP formulation with
11 breakpoints are in agreement with minimum error, however, the Water demand ¼ NB  NO  R  P
CPU computational time and number of variables have significantly
reduced the convergence speed of the MILP formulation with 11 where NB is the number of buildings obtained from Google Earth
breakpoints. The MILP formulation with 3 points provides a fast NO ¼ number of occupants per building assumed based on the type
convergence time and CPU time close to MINLP formulation, of the building, R ¼ average rate of water usage (m3 =d=capita)
however, there is a small percentage error on the final answer. The which is obtained from PA’s water system design manual for each
MILP model with 11 breakpoints results in global optimum with building’s application (Pennsylvania Department of Environmental
minimum error, however, the convergence time and number of Protection, 2014), and P ¼ peaking factor (%) obtained from
variables/constraints are significantly increased. The co- Goodrich formula (Brie re, 2014)
optimization of the water-energy microgrid provides the mini-
mum cost and maximum use of renewable energy sources to P ¼ 180t 0:1
reduce the power consumption of water distribution systems. The
proposed models can directly be applied in SCADA center of water- where t is the time in day. The base demand is then multiplied by
energy systems for daily economic dispatch problem that utilizes the 24-h multipliers from (Walski et al., 2003) to acquire the hourly
renewable energy sources and storage units to meet the water water demand profile presented in Fig. 7. According to the mass
demands of a community. As a result, the reduced power demand continuity equation, the pipe’s diameter can be calculated from the
of the water sector is satisfied at a lower cost, while contributing to in-pipe flow and velocity of the water inside the pipe, assuming a
a cleaner power production at the energy sector. It was concluded full-flow model, such that (Houghtalen et al., 2016)
that daily energy savings can be achieved by incorporating
rffiffiffiffiffiffiffi
variable-speed pumps in the water distribution network. Use of 4Q

variable-speed pumps allowed reduction in power consumption of pV
water distribution network, which reduced the daily power gen-
eration of conventional DG units. Also, reduction of the load where Q is the in-pipe flow calculated from mass balance at each
resulted in more charging of battery energy storage units during node (m3 =s), V is the velocity, assumed to be 0.9 m=s which is a
afternoon, when excess renewable energy generation was available. typical velocity according to PA’s design manual, and D is the pipe’s
The main challenge in applying the proposed models could be the diameter (m). The calculated diameter will be rounded up to the
lack of SCADA system in many water networks in the United States, next commercially available pipe size.
F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776 13

The elevations at each node, Z values, are from Google Earth.


Also, the pipes’ lengths are equal to the streets’ lengths, estimated Table C2
from Google Earth. All pipes are assumed to be 20-year old cast- Parameters of water distribution system

iron with the roughness of C ¼ 100 in the Hazen-Williams head- Parameter Values Parameter Values
loss equation (Houghtalen et al., 2016). The tank and reservoir B 2:75e3 Ti 78 $/MWh
pressure heads are exported from EPANET (a hydraulic simulator)
Q0T 0 m3 =h L12 300 m
design of the network. The pumps’ properties, i.e., a; b; c, are T
Qmin 763:4070 m3 =h D12 1m
adopted from (Ulanicki et al., 2008). T
Qmax 1:2723e3 m3 =h C 100
R
Qmin 1:272e3 m3 =h a  1:0941e4
R
Qmax 1:0e6 m3 =h b 3:9929e2
Appendix B Z1 ; Z2 20; 60 m c 223.32
Z3 ; Z4 44; 43 m Z5 ; Z6 ; Z7 40; 37; 34 m

Computational results of the standalone micro water optimi-


zation, as well as micro water-energy co-optimization are pre-
sented in Tables B1 and B2.

References

Al-Sumaiti, A.S., Salama, M.M., El-Moursi, M., 2017. Enabling electricity access in
Table B1
developing countries: a probabilistic weather driven house based approach.
Results for Micro Water Optimization
Appl. Energy 191, 531e548.
MINLP Al-Sumaiti, A.S., Ahmed, M.H., Rivera, S., El Moursi, M.S., Salama, M.M., Alsumaiti, T.,
2019a. Stochastic PV model for power system planning applications. IET Renew.
CPU time #of #of Ave. Error Unsolved Power Gener. 13 (16), 3168e3179.
Al-Sumaiti, A.S., Salama, M., El-Moursi, M., Alsumaiti, T.S., Marzband, M., 2019b.
(sec) nodes Iter. (%) nodes Enabling electricity access: a comprehensive energy efficient approach miti-
1.043 0 0 NA 0 gating climate/weather variabilityePart II. IET Generation. Transm. Distrib. 13
MILP (n ¼ 3) (12), 2572e2583.
11.489 25 483 0.36 0 Al-Sumaiti, A.S., Salama, M., El-Moursi, M., Alsumaiti, T.S., Marzband, M., 2019c.
MILP (n ¼ 11) Enabling electricity access: revisiting load models for AC-grid operation-part I.
3272 3147 117,033 0.0003 0 IET Generation. Transm. Distrib. 13 (12), 2563e2571.
Al-Sumaiti, A.S., Salama, M., Konda, S.R., Kavousi-Fard, A., 2019d. A guided pro-
cedure for governance institutions to regulate funding requirements of solar PV
projects. IEEE Access 7, 54203e54217.
Augustine, N., Suresh, S., Moghe, P., Sheikh, K., 2012. Economic dispatch for a
microgrid considering renewable energy cost functions. In: 2012 IEEE PES
Innovative Smart Grid Technologies (ISGT). IEEE, pp. 1e7.
Table B2 Baleta, J., Mikulci
c, H., Klemes, J.J., Urbaniec, K., Dui
c, N., 2019. Integration of energy,
Results for Water-Energy Microgrid Co-Optimization water and environmental systems for a sustainable development. J. Clean. Prod.
215, 1424e1436.
MINLP Bragalli, C., D’Ambrosio, C., Lee, J., Lodi, A., Toth, P., 2012. On the optimal design of
water distribution networks: a practical MINLP approach. Optim. Eng. 13 (2),
CPU time #of #of Ave. Error Unsolved 219e246.
re, François G., 2014. Drinking-water Distribution, Sewage, and Rainfall
Brie
(sec) nodes Iter. (%) nodes
Collection. Presses inter Polytechnique.
2.488 0 0 NA 0
Cherchi, C., Badruzzaman, M., Gordon, M., Bunn, S., Jacangelo, J.G., 2015. Investi-
MILP (n ¼ 3)
gation of cost and energy optimization of drinking water distribution systems.
14.80 18 0 0.078 0 Environ. Sci. Technol. 49 (22), 13724e13732.
MILP (n ¼ 11) Copeland, C., Carter, N.T., 2014. Energy-water Nexus: the Water Sector’s Energy Use,
2647 342 8934 0.000 0 vol. 43200. Congressional Research Service, Washington, DC, USA.
Dai, J., Wu, S., Han, G., Weinberg, J., Xie, X., Wu, X., Song, X., Jia, B., Xue, W., Yang, Q.,
2018. Water-energy nexus: a review of methods and tools for macro-assess-
ment. Appl. Energy 210, 393e408.
Daw, J.A., Kandt, A.J., Macknick, J.E., Giraldez Miner, J.I., Anderson, K.H.,
Appendix C Armstrong, N., Adams, J., 2018. Energy-water Microgrid Opportunity Analysis at
the University of arizona’s Biosphere 2 Facility. Tech. Rep. National Renewable
The parameters used in the water and energy systems are pro- Energy Laboratory (NREL), Golden, CO (United States).
D’Ambrosio, C., Lodi, A., Martello, S., 2010. Piecewise linear approximation of
vided in this section in Tables C1 and C2. functions of two variables in milp models. Oper. Res. Lett. 38 (1), 39e46.
Eyisi, C., Al-Sumaiti, A.S., Turitsyn, K., Li, Q., 2019. Mathematical models for opti-
mization of grid-integrated energy storage systems: a review. In: 2019 North
American Power Symposium (NAPS). IEEE, pp. 1e5.
Falco, G.J., Webb, W.R., 2015. Water microgrids: the future of water infrastructure
resilience. Procedia Eng. 118, 50e57.
Global Monitoring Laboratory, 2020. Pennsylvania state university SURFRAD
Table C1
network. https://www.esrl.noaa.gov/gmd/grad/surfrad/pennstat.html.
Parameters of microgrid system
(Accessed April 2020).
Parameter Values Guo, L., Liu, W., Li, X., Liu, Y., Jiao, B., Wang, W., Wang, C., Li, F., 2014. Energy
management system for stand-alone wind-powered-desalination microgrid.
DG aDG ¼ 1:82, bDG ¼ 0:065, gDG ¼ 0:000044 IEEE Trans. Smart Grid. 7 (2), 1079e1087.
P DG;min ¼ 0, P DG;max ¼ 500 kW, Rmax max
up ¼ Rdown ¼ 200 kW
Hansen, T.M., Roche, R., Suryanarayanan, S., Maciejewski, A.A., Siegel, H.J., 2015.
DT ¼ UT ¼ 3, Cup ¼ 0:01 $/kW, Cdn ¼ 0:01$/kW Heuristic optimization for an aggregator-based resource allocation in the smart
grid. IEEE Trans. Smart Grid. 6 (4), 1785e1794.
PV K PV ¼ 0:001 $/kW, PPV
min ¼ 0,
Hetzer, J., David, C.Y., Bhattarai, K., 2008. An economic dispatch model incorporating
PSTC ¼ 500 kW, GSTC ¼ 1000 W=m2 wind power. IEEE Trans. Energy Convers. 23 (2), 603e611.
Wind K wind ¼ 0:001 $/kW, PWmin ¼ 0, P
nom ¼ 500 kW Hijjo, M., Felgner, F., Frey, G., 2017. Pv-battery-diesel microgrid design for buildings
Batt K batt ¼ 0:65, Pchg
Max ¼ 500, P Max ¼ 500 kW
dis subject to severe power outages. In: 2017 IEEE PES PowerAfrica. IEEE,
0
EBatt Min ¼ 0, E Max ¼ 800 kWh
¼ 700, EBatt pp. 280e285.
Batt
Time Dt ¼ 1, Ti ¼ 24 H Houghtalen, R.J., Osman, A., Hwang, N.H., 2016. Fundamentals of Hydraulic Engi-
neering Systems. Prentice Hall, New York.
14 F. Moazeni, J. Khazaei / Journal of Cleaner Production 275 (2020) 122776

Jin, X., Mu, Y., Jia, H., Wu, J., Jiang, T., Yu, X., 2017. Dynamic economic dispatch of a Applications in Smart Grid (CIASG). IEEE, pp. 1e8.
hybrid energy microgrid considering building based virtual energy storage Pazheri, F., Othman, M., Malik, N., Safoora, O., 2012. Economic and environmental
system. Appl. Energy 194, 386e398. dispatch at highly potential renewable area with renewable storage. Int. J. En-
Lele, U., Klousia-Marquis, M., Goswami, S., 2013. Good governance for food, water viron. Sustain Dev. 3 (2), 177.
and energy security. Aquatic Procedia 1, 44e63. Pennsylvania Department of Environmental Protection, 2014. PA water system
Li, Q., Yu, S., Al-Sumaiti, A., Turitsyn, K., 2018a. Modeling and co-optimization of a design manual, Part II. http://www.depgreenport.state.pa.us/elibrary/.
micro water-energy nexus for smart communities. In: 2018 IEEE PES Innovative (Accessed April 2020).
Smart Grid Technologies Conference Europe (ISGT-Europe). IEEE, pp. 1e5. Rajan, D., Takriti, S., et al., 2005. Minimum up/down polytopes of the unit
Li, Q., Yu, S., Al-Sumaiti, A.S., Turitsyn, K., 2018b. Micro watereenergy nexus: commitment problem with start-up costs. IBM Res. Rep 23628, 1e14.
optimal demand-side management and quasi-convex hull relaxation. IEEE Rasekh, Amin, Hassanzadeh, Amin, Mulchandani, Shaan, Modi, Shimon, Banks, M
Trans. Control. Network Sys. 6 (4), 1313e1322. Katherine, 2016. Smart Water Networks and Cyber Security.
Li, Cuiping, et al., 2020. Economic dispatching strategy of distributed energy storage Rezvani, A., Gandomkar, M., Izadbakhsh, M., Ahmadi, A., 2015. Environmental/
for deferring substation expansion in the distribution network with distributed economic scheduling of a micro-grid with renewable energy resources. J. Clean.
generation and electric vehicle. J. Clean. Prod. 253, 119862. Prod. 87, 216e226.
Lu, X., Zhou, K., Yang, S., 2017. Multi-objective optimal dispatch of microgrid con- Saadat, H., 1999. Power System Analysis. WCB/McGraw-Hill.
taining electric vehicles. J. Clean. Prod. 165, 1572e1581. Santhosh, A., Farid, A.M., Youcef-Toumi, K., 2014. Real-time economic dispatch for
Mahmoodi, M., Shamsi, P., Fahimi, B., 2015. Economic dispatch of a hybrid microgrid the supply side of the energy-water nexus. Appl. Energy 122, 42e52.
with distributed energy storage. IEEE Trans. Smart Grid. 6 (6), 2607e2614. Scarazzato, T., Panossian, Z., Teno rio, J., Perez-Herranz, V., Espinosa, D., 2017.
Mariam, L., Basu, M., Conlon, M.F., 2016. Microgrid: architecture, policy and future A review of cleaner production in electroplating industries using electrodialysis.
trends. Renew. Sustain. Energy Rev. 64, 477e489. J. Clean. Prod. 168, 1590e1602.
Mokarram, M.J., Niknam, T., Aghaei, J., Shafie-khah, M., Catalao, J.P., 2019. Hybrid Scott, C.A., Pierce, S.A., Pasqualetti, M.J., Jones, A.L., Montz, B.E., Hoover, J.H., 2011.
optimization algorithm to solve the nonconvex multiarea economic dispatch Policy and institutional dimensions of the watereenergy nexus. Energy Pol. 39
problem. IEEE Sys. J. 13 (3), 3400e3409. (10), 6622e6630.
Morales-Espan ~ a, G., Latorre, J.M., Ramos, A., 2013. Tight and compact MILP Soshinskaya, M., Crijns-Graus, W.H., van der Meer, J., Guerrero, J.M., 2014. Appli-
formulation for the thermal unit commitment problem. IEEE Trans. Power Syst. cation of a microgrid with renewables for a water treatment plant. Appl. Energy
28 (4), 4897e4908. 134, 20e34.
Nazari-Heris, M., Mirzaei, M.A., Mohammadi-Ivatloo, B., Marzband, M., Asadi, S., The Pennsylvania State Climatologist, 2020. Hourly climatology data. http://climate.
2020. Economic-environmental effect of power to gas technology in coupled psu.edu/data/. (Accessed April 2020).
electricity and gas systems with price-responsive shiftable loads. J. Clean. Prod. Ton, D., Reilly, J., 2017. Microgrid controller initiatives: an overview of R&D by the
244, 118769. US Department of Energy. IEEE Power Energy Mag. 15 (4), 24e31.
Nguyen, H., Al-Sumaiti, A.S., Turitsyn, K., Li, Q., El Moursi, M.S., 2020. Further Tsolas, S.D., Karim, M.N., Hasan, M.F., 2018. Optimization of water-energy nexus: a
optimized scheduling of micro grids via dispatching virtual electricity storage network representation-based graphical approach. Appl. Energy 224, 230e250.
offered by deferrable power-driven demands. IEEE Trans. Power Syst. Ulanicki, B., Kahler, J., Coulbeck, B., 2008. Modeling the efficiency and power
Obukhov, S., Ibrahim, A., Diab, A.A.Z., Al-Sumaiti, A.S., Aboelsaud, R., 2020. Optimal characteristics of a pump group. J. Water Resour. Plann. Manag. 134 (1), 88e93.
performance of dynamic particle swarm optimization based maximum power Viessman, W., Hammer, M.J., Perez, E.M., Chadik, P.A., 1998. Water Supply and
trackers for stand-alone PV system under partial shading conditions. IEEE Ac- Pollution Control.
cess 8, 20770e20785. Walski, T.M., Chase, D.V., Savic, D.A., Grayman, W., Beckwith, S., Koelle, E., 2003.
Oikonomou, K., Parvania, M., 2018. Optimal coordination of water distribution en- Advanced Water Distribution Modeling and Management.
ergy flexibility with power systems operation. IEEE Trans. Smart Grid. 10 (1), Wood, A.J., Wollenberg, B.F., Sheble , G.B., 2013. Power Generation, Operation, and
1101e1110. Control. John Wiley & Sons.
Oikonomou, K., Parvania, M., 2019. Optimal participation of water desalination Xiaoping, L., Ming, D., Jianghong, H., Pingping, H., Yali, P., 2010. Dynamic economic
plants in electricity demand response and regulation markets. IEEE Sys. J. dispatch for microgrids including battery energy storage. In: The 2nd Interna-
https://doi.org/10.1109/JSYST.2019.2943451. tional Symposium on Power Electronics for Distributed Generation Systems.
Oikonomou, K., Parvania, M., Khatami, R., 2018. Optimal demand response sched- IEEE, pp. 914e917.
uling for water distribution systems. IEEE Trans. Ind. Inform. 14 (11), 5112e5122. Yin, S., Zhang, S., Andrews-Speed, P., Li, W., 2017. Economic and environmental
Palma-Behnke, R., Benavides, C., Aranda, E., Llanos, J., Saez, D., 2011. Energy man- effects of peak regulation using coal-fired power for the priority dispatch of
agement system for a renewable based microgrid with a demand side man- wind power in China. J. Clean. Prod. 162, 361e370.
agement mechanism. In: 2011 IEEE Symposium on Computational Intelligence

You might also like