Optimization and Engineering manuscript No.

(will be inserted by the editor)

Optimal Design and Dispatch of a System of Diesel

Generators, Photovoltaics and Batteries for Remote
Locations

Michael S. Scioletti · Alexandra M.

Newman · Johanna K. Goodman ·
Alexander J. Zolan · Sven Ley↵er

DRAFT as of November 27, 2016

Abstract Renewable energy technologies, specifically, solar photovoltaic cells,

combined with battery storage and diesel generators, form a hybrid system
capable of independently powering remote locations, i.e., those isolated from
larger grids. If sized correctly, hybrid systems reduce fuel consumption com-
pared to diesel generator-only alternatives. We present an optimization model
for establishing a hybrid power design and dispatch strategy for remote loca-
tions, such as a military forward operating base, that models the acquisition
of di↵erent power technologies as integer variables and their operation using
nonlinear expressions. Our cost-minimizing, nonconvex, mixed-integer, non-
linear program contains a detailed battery model. Due to its complexities, we
present linearizations, which include exact and convex under-estimation tech-
niques, and a heuristic, which determines an initial feasible solution to serve as
a “warm start” for the solver. We determine, in a few hours at most, solutions
within 5% of optimality for a candidate set of technologies; these solutions
closely resemble those from the nonlinear model. Our instances contain real
data spanning a yearly horizon at hour fidelity and demonstrate that a hy-
brid system could reduce fuel consumption by as much as 50% compared to a
generator-only solution.

M. Scioletti · A. Newman
Department of Mechanical Engineering, Colorado School of Mines, Golden, CO 80401
E-mail: [email protected], E-mail: [email protected]
J. Goodman
School of Chemical and Biomolecular Engineering, Georgia Institute of Technology, Atlanta,
GA 30332
E-mail: [email protected]
A. Zolan
Department of Mechanical Engineering, University of Texas at Austin, Austin, TX 78712
E-mail: [email protected]
S. Ley↵er
Mathematics and Computer Science Division, Argonne National Laboratory, Lemont, IL
60439 E-mail: ley↵[email protected]
2 Scioletti et al.

1 Introduction and Literature Review

Hybrid power systems integrate renewable energy technologies, such as solar

photovoltaic devices (PV), with energy storage systems (batteries) and diesel
generators to provide grid-quality electrical power to remote locations. If sized
and operated efficiently, hybrid systems are a cost-beneficial alternative to
grid extension and spot generation. To this end, we construct a mixed-integer,
nonlinear optimization model that determines the number and type of PV,
battery, and generator technologies to procure, and how to operate them on
an hourly basis so as to minimize costs, while adhering to constraints that
govern the operation of the system. Because nonlinearities lead to tractability
issues, we present linearizations that yield good approximations to the non-
linear model. We also assume that our inputs are deterministic. While even
higher fidelity models may be formulated, our goal is to make design deci-
sions, and to use the dispatch as a guide for these decisions [29]. Real-time
dispatch would require a more detailed model, taking the design as fixed. Our
load profiles demonstrate significant variability over the course of one year,
and this serves as a proxy for making our design decisions robust to stochastic
loads. Furthermore, as we discuss throughout the paper, even our one-stage
deterministic model is associated with instances that are difficult to solve; our
proposed techniques enhance that solvability, and, to our knowledge, demon-
strate current state of the art. Accounting for stochasticity with a two- or
multi-stage approach would result in an intractable model.
Optimally determining design and dispatch is an NP-hard problem that
involves modeling nonlinearities and integer restrictions. It is common to sep-
arate the problem into one of (i) design or (ii) dispatch and then solve; how-
ever, this does not guarantee global optimality of the solution, because it is
a restriction of the problem. Our computational tests show that, for our in-
stances, basing an entire design and dispatch solution simply on a design that
examines maximum load can result in solutions that use 50% more fuel. More
tailored heuristics show promise in providing good, but not optimal, results to
the design problem [37, 15, 14, 38, 25], but often rely on dispatch strategies
set a priori ([13] and [17]) to satisfy demand constraints. Some authors use
multiple objectives such as cost, reliability, and emissions reduction or Pareto
optimality assessment techniques as part of a heuristic strategy [19], but have
difficulty establishing solution quality. Shortening the time horizon [43] and/or
reducing the variability in daily demand [19] increases tractability of the prob-
lem; however, seasonal changes in demand could significantly impact design
decisions.
HOMER (Hybrid Optimization Model for Electric Renewables), at the
time of this writing, is the most widely used design and dispatch program
[36, 48, 53, 31, 41] and represents a simulation model that, for a year-long de-
mand profile, uses fixed dispatch strategies and ranks resulting solutions based
on life-cycle cost [8]. Few modeling e↵orts include deterministic methods such
as linear programming (LP) and/or mixed integer programming (MIP) to solve
the design and dispatch problem, especially as a monolith for a year-long hori-
Hybrid Power System Design and Dispatch 3

zon with hourly time fidelity. A MIP with wind power, batteries, and genera-
tors [16] produces results comparable to HOMER’s; however, it fails to address
the complexities associated with battery modeling and solves the problem in
two steps: (i) design solutions result from running the MIP for a curtailed time
horizon and then (ii) dispatch solutions follow for the entire year given design
from (i). A linear program solves a year-long problem at hourly time fidelity
to understand the operating relationships between the technologies within a
hybrid system over a 20-year ownership timeline [33]; but, the model considers
identical 24-hour demand periods for the entire year and lacks battery mod-
eling detail. Given the unpredictability of renewable energy, [10] presents a
stochastic model to account for the variations in windspeed by solving a year-
long problem decomposed into day-long, i.e., 24-hour, sub-problems; however,
the authors fix technologies in their hybrid system with the goal of optimiz-
ing the sizing of an energy storage system, which could consist of batteries,
flywheels, and supercapacitators [9]. Alternately, [46] develops a nonconvex
mixed-integer, nonlinear program (MINLP) to describe the design and dis-
patch of a distributed combined heat and power generation system using Solid
Oxide Fuel Cells (SOFC), PV, and batteries for commercial buildings for a
time horizon of one year (8,760 hours). Due to the complexities of modeling
SOFCs, the authors do not attempt to model batteries or PV in detail. By
developing a convex under-estimation of the MINLP through a linearization
technique for bi- and tri-linear terms, the authors present a MIP that, with
the help of a bounding algorithm which takes days to run, solves year-long
instances to a gap less than 8%.

Through the use of a heuristic, which serves to provide the solver with
an initial feasible solution and linearizations, which include exact and con-
vex under-estimation techniques, our research contributes to the literature by
solving the design and dispatch problem to within 5% of optimality given a
candidate set of technologies in a matter of hours for a year-long demand fore-
cast with hourly fidelity. Attributes that di↵erentiate our model from those
in the literature include: (i) nonlinearities associated with modeling battery
discharge and lifetime; (ii) realistic procurement technology sets with varying
sizes and quantities, i.e., a design; and (iii) an unbiased dispatch strategy re-
flective of demand in each time period. Without loss of generality, we apply
our model to forward operating bases (FOBs), though it could also be used in
a variety of other microgrid situations.

The remainder of this paper is organized as follows: Section 2 presents the

formulation of the MINLP, which includes the linearization techniques and
subsequent re-formulation of the problem as a MIP, while Section 3 introduces
heuristics to reduce the size of the problem and produce an initial feasible
solution. In Section 4, we discuss instances for di↵erent demand profiles used
to test the MINLP and MIP models; examine the mathematical characteristics
and methods to solve both problems; and discuss their results and accuracy.
Section 5 concludes.
4 Scioletti et al.

2 Model

Our model includes two types of variables: design and dispatch, i.e., the levels
at which the procured technologies operate to meet a prescribed demand profile
for one year at an hourly time fidelity. We minimize procurement, fuel, and
lifecycle costs subject to load, capacity and system interoperability constraints.

2.1 Model Overview

A hybrid system incurs capital costs that originate from commercial prices and
availability, and operational costs, including those for fuel. We assume that the
hybrid system operates independently of a commercial grid (see Figure 1) and
can consist of multiple component sizes within each technology and/or of more
than one of the same size technology. PV panels form an array, while generators
are located adjacent to each other. Batteries of like type comprise a bank that
operates as a single unit to preclude modeling individual cells, which would
increase the size of the problem. PV technologies first connect to a DC-to-DC
converter for the purpose of maximum power point tracking, which links to
a bi-directional converter, while battery technologies connect directly to the
bi-directional converter. Generator technologies connect directly to the AC
bus, which connects to the power demand. We only implicitly model the bus
system and the bi-directional converter through their efficiencies.

Fig. 1: A topological layout of the proposed hybrid system that includes

generator, PV, and battery technologies. All technologies are physically
located in close proximity to each other.
Hybrid Power System Design and Dispatch 5

The model acquires technologies and then assigns a fraction of the load
to each in order to supply enough power to meet the demand in each time
period. To allow for the acquisition of a variety of generator types, we do not
constrain generators to equally share the load or to operate in droop. Power
from the generators both meets demand and charges the battery so long as the
generator operates within bounds prescribed by the manufacturer. We model
the lifetime of the generator by counting the number of hours it is in operation.
A generator’s fuel consumption is related quadratically to its power output,
but manufacturer data often implies a linear relationship (see Figure 2), in
which the intercept is greater than zero.

Fig. 2: Manufacturer-provided fuel consumption data points for four

di↵erently sized generator technologies [3]. We use afg , bfg , and cfg to model
the quadratic fit.

We model PV power output using NREL’s PVWatts panel calculator. A

PVWatts simulation model maps solar irradiation based on location, and sim-
ulates power output of a PV panel at hourly fidelity for time horizons up
to one year [23]. Although PV panel power output is a direct current (DC),
PVWatts outputs an alternating current (AC) that accounts for power con-
version losses from the hybrid system. If stored, which occurs when PV power
is greater than demand, PV power is subject to efficiency losses related to
battery charging. A PV array requires space, which we restrict by area, or
number of panels. We account for the intermittent nature of PV power by
maintaining a spinning reserve capability through a battery’s state of charge
(SOC), i.e., a measure of its available capacity, and/or by operating genera-
tors at less than their rated power. Spinning reserve constraints in design and
dispatch problems often consider the economics associated with buying and
selling electricity to a grid [45, 32, 44, 47]; however, the remote hybrid system
problem is grid-independent, so we model spinning reserve as a fraction of PV
power output to account for the variability of the latter within each hour-long
time period.
6 Scioletti et al.

Aside from the load, charging the batteries is the only power draw on the
system. Batteries provide power to meet the load, but are also employed as a
reserve for the renewable technologies. We do not allow a battery to charge and
discharge in the same time period. Power output from a battery is a function of
the nonlinear relationship between current and voltage; models that consider
this relationship are more accurate than those that do not for many battery
chemistries [51]. The current depends on a battery’s SOC. Batteries show a
rate-capacity e↵ect, in which the available capacity based on the SOC decreases
with higher current draw.
Figure 3a displays battery voltage as a function of SOC for a fixed discharge
and charge current [1]. By slightly restricting the SOC operating range, we
can model the voltage using a linear relationship between SOC and current.
Common to most batteries is a rate-capacity e↵ect, which implies that as the
magnitude of the discharge current increases, the available capacity decreases.
Peukert’s equation is often used to describe this behavior [24]. This concept is
also employed by [39]’s kinetic energy battery model, which relates the change
in capacity to the charge and discharge rates using a two-tank model. While
the rate-capacity e↵ect is nonlinear over a large current range, especially at
high currents, our hourly time step allows us to use a linear approximation
over the relevant current range (see Figure 3b).

Fig. 3: A graphical comparison of a) battery voltage for a battery b,

soc soc
“twin” k, and time period t, Vbkt as a function of state of charge Bbkt , for
v v
which we use ab and bb to represent the slope and intercept, respectively,
where the intercept is augmented by constants representing Ohm’s law
(see constraint (5g)); and b) the normalized capacity, which we model
as the quotient of the maximum realized current IbU and capacity cref b
of a battery given discharge time.

We account for battery chemistry characteristics such as the voltage and

rate-capacity behavior, but exclude chemistry-specific aspects such as over-
Hybrid Power System Design and Dispatch 7

charge and out gassing of lead-acid batteries [11, 22], and performance as a
function of temperature [30, 11, 22].
A battery’s lifetime is a function of how it operates and the SOC level at
the time of use [28, 40, 35, 34, 26, 56]. In Figure 4a, we present three di↵erent
use profiles: (A), (B), and (C). Regime (A) shows full discharge and charge
cycles, (B) depicts short charge and discharge cycles at a high SOC, and (C)
depicts short charge and discharge cycles at a low SOC. Given identical charge
throughput, life expectancy decreasing by regime is generally: (B), (A), and
(C). A cycle counting method would not distinguish between these cases as
each small charge and discharge would count as a full cycle. Instead of counting
cycles as defined by current reversal, we present an amp-hour (Ah) assessment
method (see Figure 4b); that is, we sum the total amount of current (Ah) that
passes through the battery for both charge and discharge. The quotient of this
value and twice the reference capacity yields the fraction of a cycle completed.

Fig. 4: a) A depth of discharge comparison between three di↵erent

charge and discharge strategies illustrates the shortcomings of cycle
counting. b) A weighting function from which we derive dsoc b and asoc
b ,
the slope and intercept of the line, respectively, accounts for variable bat-
tery aging at di↵erent states of charge for an A123 Lithium-ion battery
[1].

2.2 Mathematical Formulation

We now present the mathematical formulation of our problem, henceforth

referred to as (P). In general, we use lower-case letters for parameters and
upper-case letters for variables. We also use lower-case letters for indices and
upper-case script letters for sets. Superscripts and accents distinguish between
parameters and variables that utilize the same base letter, while subscripts
identify elements of a set. Some parameters and variables are only defined for
8 Scioletti et al.

certain set elements, which are listed in each definition. A plus sign (+) sig-
nifies power going into a technology, while a minus sign ( ) indicates power
leaving. The units of each parameter and variable are provided in brackets
after its definition. We use the term “twins” to denote a tuple or a multiple
of a certain technology type to distinguish the operational patterns of and,
hence, wear on each unit.

Sets

t2T set of all time periods [hours]

j2J set of all battery and generator technologies
g2G⇢J set of all generator technologies
b2B⇢J set of all battery technologies
s2S set of all PV panel types
k 2 J˜j set of identical twins of technology j, given by size, type,
and manufacturer
k 2 G̃g set of all generator twins of type g
k 2 B̃b set of all battery twins of type b

Timing Parameters

⌧ length of one time period [hours]

⌫ ratio of base operation duration to time horizon length
[fraction]

Optimization Model Penalty Parameters

c̃j cost of procuring one twin of technology type j [$/twin]

cs cost of procuring one panel of technology type s [$/panel]
f
t fuel cost penalty in time period t [$/gal]
"j cycle cost penalty for technology type j [$/(hours, cycles)]

Power System Parameters

dP t steady-state power demand in time period t [W]

k̄ overage load coefficient [fraction]
ks spinning reserve required relative to PV power [fraction]

Technology Parameters

¯lj maximum lifetime of technology type j [generator hours,

battery cycles]
⌘j+ , ⌘j electric efficiency of power flow into and out of technology
type j, respectively [fraction]
Hybrid Power System Design and Dispatch 9

pj , p̄j minimum and maximum power rating, respectively, of

technology type j [W]

Generator Parameters

afg , bfg , cfg fuel consumption coefficients for generator g [ Wgal gal gal
2h , W h , h ]

PV Parameters

W
st power output of technology type s in time period t [ panel ]
n̄s maximum allowable number of PV panels of technology type
s [panels]

Battery Parameters

avb , bvb battery b linear voltage model slope and intercept

coefficients, respectively [V]
dsoc soc
b , ab battery b linear lifetime model slope and intercept
coefficients, respectively [unitless]
b0b battery b state of charge used in initial condition constraints
[fraction]
cref
b battery b manufacturer-specified capacity [Ah]
c+
b , cb battery b charge and discharge capacity rate coefficients,
respectively [hours]
rbint battery b internal resistance [Ohms]
iavg
b typical current expected from battery b for both charge and
discharge activities [A]
sb , s̄b battery b state of charge minimum and maximum
operational bounds, respectively [fraction]
iL U
b , ib battery b discharge current lower and upper bound,
respectively [A]
iL+ U+
b , ib battery b charge current lower and upper bound,
respectively [A]

where, for our application, the above parameter values are computed as:
iL
b = 0 8b 2 B
cref
iU
b = b
8b 2 B
cb + ⌧

iL+
b = 0 8b 2 B
cref
iU
b
+
= b
8b 2 B
c+
b
10 Scioletti et al.

Continuous Variables

Power System Variables

Ljk number of expended life cycles for technology type j,

twin k [generator hours, battery cycles]
+
Pjkt , Pjkt aggregate power into and out of technology type j,
twin k in time period t, respectively [W]
PV
Pst aggregate power out of PV technology type s in time
period t [W]

Generator Variables

F̃t amount of fuel used in time period t [gal]

Battery Variables

soc
Bbkt state of charge of battery type b, twin k in time period
t [fraction]
+
Ibkt , Ibkt battery b, twin k current for charge and discharge,
respectively, in time period t [A]
soc
Vbkt battery b, twin k voltage in time period t [V]

Binary and Integer Variables

Power System Procurement Variables

Wjk 1 if technology j, twin k is procured, 0 otherwise

Xs integer number of PV panels of technology type s
procured [panels]

Generator Variables

Ggkt 1 if technology type g, twin k is operating in time

period t, 0 otherwise

Battery Variables
+
Bbkt 1 if battery type b, twin k is charging in time period t,
0 otherwise
Bbkt 1 if battery type b, twin k is discharging in time period
t, 0 otherwise
Hybrid Power System Design and Dispatch 11

Problem (P)
(see §2.3.1 Objective Function)

Minimize
0 1
X X X X X X f
c̃j Wjk + c s Xs + ⌫ @ "j Ljk + t F̃t
A (1)
j2J k2J˜j s2S j2J k2J˜j t2T

subject to

(see §2.3.2 System Operations)

X X XX X
+ PV
⌘j Pjkt Pbkt + Pst (1 + k̄)dP
t 8t 2 T (2a)
j2J k2J˜j b2B k2B̃b s2S
XX X X X
soc
⌘b p̄b Bbkt + p̄g Ggkt Pgkt ks PV
Pst 8t 2 T (2b)
b2B k2B̃b g2G k2G̃g s2S

Wj,k 1 Wjk 8j 2 J , k 2 J˜j : k > 1 (2c)

(see §2.3.3 Generator Operations)

pg Ggkt  Pgkt  p̄g Ggkt 8g 2 G, k 2 G̃g , t 2 T (3a)

X X
F̃t ⌧ (afg (Pgkt )2 + bfg Pgkt + cfg Ggkt ) 8t 2 T (3b)
g2G k2G̃g

Ggkt  Wgk 8g 2 G, k 2 G̃g , t 2 T (3c)

Gg,k 1,t  Ggkt 8g 2 G, k 2 G̃g , t 2 T : k > 1 (3d)
Pg,k 1,t  Pgkt 8g 2 G, k 2 G̃g , t 2 T : k > 1 (3e)

(see §2.3.4 PV Operations)

PV
Pst  st Xs 8s 2 S, t 2 T (4a)
Xs  n̄s 8s 2 S (4b)

(see§ 2.3.5 Battery Storage Operations)

+ soc +
Pbkt = Vbkt Ibkt 8b 2 B, k 2 B̃b , t 2 T (5a)
soc
Pbkt = Vbkt Ibkt 8b 2 B, k 2 B̃b , t 2 T (5b)
12 Scioletti et al.

soc soc ⌧ (⌘b+ Ibkt

+
Ibkt )
Bbkt = Bbk,t 1 + 8b 2 B, k 2 B̃b , t 2 T : t > 1 (5c)
cref
b

soc
sb Wbk  Bbkt  s̄b Wbk 8b 2 B, k 2 B̃b , t 2 T (5d)

soc soc
Bbkt  Bb,k 1,t + (1 Wbk ) 8b 2 B, k 2 B̃b , t 2 T : k > 1 (5e)
soc soc
Bbkt Bb,k 1,t (1 Wbk ) 8b 2 B, k 2 B̃b , t 2 T : k > 1 (5f)

soc
Vbkt = avb Bbk,t
soc
1
+
+ bvb (Bbkt + Bbkt ) + iavg int +
b rb (Bbkt Bbkt )
8b 2 B, k 2 B̃b , t 2 T : t > 1 (5g)

pb Bbkt  Pbkt  p̄b Bbkt 8b 2 B, k 2 B̃b , t 2 T (5h)

+ + +
pb Bbkt  Pbkt  p̄b Bbkt 8b 2 B, k 2 B̃b , t 2 T (5i)

Ibkt  iU soc
b Bbk,t 1 8b 2 B, k 2 B̃b , t 2T :t>1 (5j)
iL
b Bbkt  Ibkt  iU
b Bbkt 8b 2 B, k 2 B̃b , t 2 T (5k)
iL+ +
b Bbkt  +
Ibkt  iU + +
b Bbkt 8b 2 B, k 2 B̃b , t 2 T (5l)

+
Bbkt + Bbkt  Wbk 8b 2 B, k 2 B̃b , t 2 T (5m)

+
Bbkt + Bb0 k0 t  1 8b, b0 2 B; k, k 0 2 B̃b ; t 2 T : b 6= b0 , k 6= k 0 (5n)

(see §2.3.6 Lifecycle)

X
Lgk ⌧ Ggkt 8g 2 G, k 2 G̃g (6a)
t2T

!
X +
Ibkt + Ibkt
Lbk ⌧ (asoc
b dsoc soc
b Bbk,t 1) 8b 2 B, k 2 B̂b (6b)
t 2 2cref
b

¯lj
Ljk  Wjk 8j 2 J , k 2 J˜j (6c)
⌫
Hybrid Power System Design and Dispatch 13

(see §2.3.7 Non-negativity and Integrality)

+
Pjkt , Pjkt 0 8j 2 J , k 2 J˜j , t 2 T (7a)
Ljk 0 8j 2 J , k 2 J˜j (7b)
F̃t 0 t2T (7c)
PV
Pst 0 8s 2 S, t 2 T (7d)
soc + soc
Bbkt , Ibkt , Ibkt , Vbkt 0 8b 2 B, k 2 B̃b , t 2 T (7e)

Xs 0 integer 8s 2 S (7f)

Wjk binary 8j 2 J , k 2 J˜j (7g)

Ggkt binary 8g 2 G, k 2 G̃g , t 2 T (7h)
+
Bbkt , Bbkt binary 8b 2 B, k 2 B̃b , t 2 T (7i)

2.3 Detailed Discussion of Formulation

We model the optimal design and dispatch problem as a nonconvex, mixed-

integer, nonlinear program. Instances of this problem are challenging to solve,
because of the nonconvex relationships between variables, and the lengthy
time horizon, i.e., annual with hourly fidelity. The presence of battery state-
of-charge relationships and of battery lifecycle constraints implies that the
model does not decompose readily by time step. Below, we comment on each
of these characteristics of our model in turn before we suggest procedures to
expedite solutions.

2.3.1 Objective Function

The objective function (1), minimizes the sum of four terms: (i) the cost as-
sociated with procuring various battery and generator technologies; (ii) the
cost associated with procuring various PV panels; (iii) an arbitrarily weighted
measure of the life cycles used by each technology over the total length of
operation; and (iv) a weighted measure of the cost of fuel. Our application
pertains to forward operating bases with a maximum lifetime of one year. To
reconcile time horizon lengths of other than a year, we apply the parameter
⌫, which adjusts operational costs accordingly.

2.3.2 System Operations

Constraint (2a) ensures that the hourly dispatch strategy meets demand. The
first term represents the power from the generators and batteries, accounting
for power system losses; the second term captures the power to charge the
batteries, and the third term reflects the contributions of PV power. The
14 Scioletti et al.

right-hand side is the product of the forecasted demand for the time period
and an overage load factor. Due to the intermittence of solar power, constraint
(2b) enforces “spinning reserves,” which ensure that a backup power source,
either batteries and/or generators, is available to meet a fraction of the load
supplied by PV. Constraint (2c) breaks symmetry and forces the procurement
of twins of technology j to occur in a fixed order [54]. These constraints do
not guarantee a decrease in computation time in every instance we solve, but
they do tend to minimize long solution times (see Section 4).

2.3.3 Generator Operations

If a generator is running, constraint (3a) bounds output power between a

minimum and maximum manufacturer-specified level. Constraint (3b) deter-
mines the amount of fuel used during time period t, which, if afg = 0, is
linear. Constraint (3c) connects procurement to dispatch. Constraints (3d)
and (3e) prioritize the use of technology twins to reduce symmetry [54]. These
constraints force the dispatch of generators in lexicographic order, which pro-
duces unequal wear and is therefore contrary to their likely dispatch method;
in a real dispatch situation, an equal-wear strategy could be pursued without
compromising the objective function value.

2.3.4 PV Operations

We limit the PV output power per panel to st in constraint (4a). The an-
ticipated solar panel output results from a PVWatts simulation run a priori,
which accounts for performance characteristics such as location, panel effi-
ciency, tilt, and angle. Constraint (4b) limits the number of panels considered
for procurement given the expected land area available.

2.3.5 Battery Storage Operations

Constraints (5a) and (5b) represent the nonlinear relationship between voltage,
current, and the power associated with charging and discharging the battery,
respectively. Constraint (5c) updates the battery SOC, which is a function of
its previous SOC and the discharge and charge currents. An efficiency parame-
ter associated with the second term signifies that when the battery charges, the
state of charge receives a fraction of the incoming power due to the conversion
from AC to DC power. For time period t = 1, the constraint is:
!
soc 0 ⌘b+ Ibkt
+
Ibkt
Bbkt = bb Wbk + ⌧ 8b 2 B, k 2 B̃b , t = 1 (8)
cref
b

Constraint (5d) bounds the SOC of a battery to a minimum and maximum

level. Constraints (5e) and (5f) ensure that the batteries operate in droop,
rather than individually, to avoid the situation in which one battery is used
to charge another. When considering only one battery for procurement, these
Hybrid Power System Design and Dispatch 15

constraints are redundant and may be removed. Constraint (5g) models the
battery voltage as a function of its previous state of charge and the direction
of current flow, which, for state of charge levels between a certain range, is
linear (see Figure 3a).
Constraints (5h) and (5i) bound the net power flow of each battery per
time period, while constraints (5j) through (5l) similarly constrain current
flow. For time period t = 1, constraint (5j) is:

Ibkt  iU 0
b bb Bbkt 8b 2 B, k 2 B̃b , t = 1 (9)

Constraints (5m) and (5n) prevent simultaneous charge and discharge for
a given battery, and for di↵erent battery-twin combinations, respectively.

2.3.6 Lifecycle

Constraint (6a) counts the number of operational hours of a generator. Con-

straint (6b) counts weighted equivalent cycles for batteries. A battery’s life-
cycle is a function of both the amount of the charge and discharge currents
as well as the SOC level at which the charge or discharge occurs. Because
the lifecycle constraint considers both charge and discharge, i.e., two opposite-
direction-operations to which together we refer as a round trip, we divide by
two. Constraint (6c) limits technology lifetime.

2.3.7 Non-negativity and Integrality

Finally, constraints (7a) - (7e) ensure that the appropriate variables in our
formulation assume continuous, non-negative values. Constraints (7f) - (7i)
enforce integer and binary restrictions, as appropriate.

2.4 Linearization

Model (P) is nonlinear in that there is one quadratic term (see constraint
(3b)), and bilinear terms exist within constraints (5a), (5b), and (6b). To
increase tractability of the corresponding model instances, we present (U ), a
linearization of (P) which corresponds to an under-estimation of the original
problem.
We can approximate a quadratic function by using piecewise linear func-
tions; however, in our case, the data provided by the manufacturers corre-
sponds to a line (see Figure 2), so we set afg equal to 0, thereby eliminating
the quadratic term. The bilinear terms assume one of two forms: (i) the prod-
uct of a binary variable and a continuous variable, and (ii) the product of two
continuous variables. We provide an exact method to linearize the former, and
use a convex under-estimation technique for the latter. We do not explicitly
16 Scioletti et al.

present the constraints for the case in which t = 1 because the only di↵erence
is that for this case, b0b replaces Bbk,t
soc
1 (which occurs when t > 1).
Substituting the voltage constraint (5g) directly into the power constraints
(5a) and (5b), we obtain:

✓ ◆
+ + avg int + +
Pbkt = avb Bbk,t
soc v
1 + bb (Bbkt + Bbkt ) + ib rb (Bbkt Bbkt ) Ibkt

8b 2 B, k 2 B̃b , t 2 T : t > 1 (10)

✓ ◆
+ avg int +
Pbkt = avb Bbk,t
soc
1 + b v
b (B bkt + B bkt ) + i b r b (B bkt B bkt ) Ibkt

8b 2 B, k 2 B̃b , t 2 T : t > 1 (11)

We can simplify equations (10) and (11) by distributing the respective cur-
rent variable and removing the irrelevant charge or discharge binary variable
in each equation. For example, if a battery is charging during a time period,
it cannot be discharging, so we remove the discharge binary variables Bbkt .

+
Pbkt = avb Bbk,t
soc +
1 Ibkt + (bvb + iavg int + +
b rb )Bbkt Ibkt

8b 2 B, k 2 B̃b , t 2 T : t > 1 (12)

Pbkt = avb Bbk,t
soc
1 Ibkt + (bvb iavg int
b rb )Bbkt Ibkt

8b 2 B, k 2 B̃b , t 2 T : t > 1 (13)

We distribute the terms on the right hand side of the lifecycle constraint
(6b) to identify bilinear terms consisting of SOC and current:

+ +
!
X asoc
b Ibkt dsoc soc
b Bbk,t 1 Ibkt + asoc
b Ibkt dsoc soc
b Bbk,t 1 Ibkt
Lbk ⌧
t 2 2cref
b

8b 2 B, k 2 B̂b (14)

Auxiliary Variables
Equations (12), (13), and (14) contain two sets of bi-linear terms, for each
of which we define a nonnegative continuous variable:

+
Ybkt , Ybkt battery b, twin k exact linearization variable representing
the product of a binary and continuous variable for charge
and discharge, respectively, in time period t [A]
+
Zbkt , Zbkt battery b, twin k linear approximation variable representing
the product of two continuous variables for charge and
discharge, respectively, in time period t [A]
Hybrid Power System Design and Dispatch 17

+ +
Ybkt , Ybkt , Zbkt , Zbkt 0 8b 2 B, k 2 B̃b , t 2 T (15)

+ + +
Ybkt = Bbkt Ibkt 8b 2 B, k 2 B̃b , t 2 T (16)
Ybkt = Bbkt Ibkt 8b 2 B, k 2 B̃b , t 2 T (17)

+ soc +
Zbkt = Bbk,t 1 Ibkt 8b 2 B, k 2 B̃b , t 2 T : t > 1 (18)
soc
Zbkt = Bbk,t 1 Ibkt 8b 2 B, k 2 B̃b , t 2 T : t > 1 (19)

We then substitute these variables directly into (12), (13), and (14):

+
Pbkt +
= avb Zbkt + (bvb + iavg int +
b rb )Ybkt 8b 2 B, k 2 B̃b , t 2 T (20)
Pbkt = avb Zbkt + (bvb iavg int
b rb )Ybkt 8b 2 B, k 2 B̃b , t 2 T (21)
!
X asoc +
dsoc + soc
dsoc
b Ibkt b Zbkt + ab Ibkt b Zbkt
Lbk ⌧
t 2 2cref
b

8b 2 B, k 2 B̃b (22)

Constraint (6b) presents a symmetric function that penalizes both charge

and discharge operations equally as a fraction of capacity. Given our defini-
tion of SOC per constraint (5c), which implies the battery needs to charge in
order to discharge, we can simplify constraint (22) by multiplying it by two,
which cancels the 2 in the denominator, and by removing either the charge or
discharge variables. We choose to remove the discharge variables because our
+
approximation for Zbkt is more accurate (see Table 10).

!
X asoc +
dsoc +
b Ibkt b Zbkt
Lbk ⌧ 8b 2 B, k 2 B̃b (23)
t 2 cref
b

+ +
All constraints involving bilinear terms contain Ybkt , Ybkt , Zbkt , and/or
Zbkt . We execute an exact technique to linearize the bilinearities associated
+
with Ybkt and Ybkt ; we invoke an approximation to eliminate the nonlinearities
+
associated with Zbkt and Zbkt .
18 Scioletti et al.

2.4.1 Exact Linearization

+
We apply the operational bounds of charge and discharge current to Ybkt and
Ybkt , respectively:

iL+ + + U+ +
b Bbkt  Ybkt  ib Bbkt 8b 2 B, k 2 B̃b , t 2 T (24a)
iL
b Bbkt  Ybkt  iU
b Bbkt 8b 2 B, k 2 B̃b , t 2 T (24b)
We can further constrain Ybkt by the upper bound on the battery’s dis-
soc
charge current and the SOC, because Bbkt  Bbkt 8b 2 B, k 2 B̃b , t 2 T
when the battery is discharging, i.e., Bbkt = 1; this tightens the upper bound
imposed by constraint (24b). If the battery is not discharging, then Bbkt = 0,
which forces Ybkt = 0.

Ybkt  iU soc
b Bbk,t 1 8b 2 B, k 2 B̃b , t 2 T : t > 1 (24c)
Note that constraint (24c) is similar to (5j). We do not further constrain
+
Ybkt because, for the parameters in our application, i.e., iU +
b , the bound is
sufficiently tight.
+
We then relate Ybkt and Ybkt to the respective current variable using an
+
exact relationship. For example, when a battery charges, Bbkt is 1, which
+ + + + U+ +
implies that Ibkt = Ybkt . If Bbkt is 0, Ybkt is 0 by (24a), and ib  Ibkt  iU
b
+

(which is redundant). The same logic holds for the discharge case.

iU +
b (1
+
Bbkt +
)  Ibkt +
Ybkt  iU +
b (1
+
Bbkt ) 8b 2 B, k 2 B̃b , t 2 T (24d)
iU
b (1 Bbkt )  Ibkt Ybkt  iU
b (1 Bbkt ) 8b 2 B, k 2 B̃b , t 2 T (24e)
By substituting constraints (24a) through (24e) for constraints (16) and
+
(17) and adding nonnegativity of Ybkt , Ybkt , we achieve an exact reformulation
of the product of a binary and continuous variable.

2.4.2 Approximate Linearization

+
Zbkt and Zbkt represent the product of two continuous variables, which is both
a nonlinear and nonconvex relationship; however, [42] and [12] provide an
approximation technique using the convex envelope of the terms comprising
the bilinear relationship to obtain a lower bound. We depict this linearization
in constraints (25a) through (25h), which replace constraints (18) and (19) in
our reformulation.

+
Zbkt iU + soc
b Bbk,t 1
+
+ s̄b Ibkt s̄b iU
b
+
8b 2 B, k 2 B̃b , t 2 T : t > 1 (25a)
+
Zbkt iL+ soc +
b Bbk,t 1 + sb Ibkt sb iL+b 8b 2 B, k 2 B̃b , t 2 T : t > 1 (25b)
+
Zbkt  iU + soc +
b Bbk,t 1 + sb Ibkt s b iU
b
+
8b 2 B, k 2 B̃b , t 2 T : t > 1 (25c)
+
Zbkt  iL+ soc +
b Bbk,t 1 + s̄b Ibkt s̄b iL+
b 8b 2 B, k 2 B̃b , t 2 T : t > 1 (25d)
Hybrid Power System Design and Dispatch 19

Zbkt iU soc
b Bbk,t 1 + s̄b Ibkt s̄b iU
b 8b 2 B, k 2 B̃b , t 2 T : t > 1 (25e)
Zbkt iL b
soc
Bbk,t 1 + sb Ibkt s b iLb 8b 2 B, k 2 B̃b , t 2 T : t > 1 (25f)
U soc U
Zbkt  ib Bbk,t 1 + sb Ibkt s b ib 8b 2 B, k 2 B̃b , t 2 T : t > 1 (25g)
Zbkt  iLb
soc
Bbk,t 1 + s̄b Ibkt s̄b iLb 8b 2 B, k 2 B̃b , t 2 T : t > 1 (25h)

Thus, (U ) removes constraints (5a), (5b), (5g), and (6b) and replaces them
with constraints (20) through (25h), and adds a non-negativity constraint
(15). The quality of solutions from (U ), compared to (P), is directly related
+
to the tightness of the convexified bounds for Zbkt and Zbkt (see Section 4.4),
which are a function of sb , s̄b , ib , ib , ib , and iU
L+ U + L
b and originate from the
rate-capacity e↵ect of the battery (see constraints (5j) through (5l)).

3 Heuristics

We present a heuristic H that produces an initial feasible solution to mod-

els (U ) and (P) quickly by: (i) limiting the set of candidate designs for in-
stances, and (ii) using the better of two myopic dispatch strategies. The heuris-
tic H possesses the following benefits:
– reduces solution time by eliminating dominated design decisions from the
feasible region;
– supplies a branch-and-bound solver with an initial feasible solution, known
as a “warm start,” which can reduce computation time if the initial solution
is close to optimal; and,
– provides a dispatch strategy that is easier to implement in a microgrid
controller than following ideal dispatch decisions from model (U ), as the
solution does not look ahead to future time periods and is feasible for the
nonlinear model (P).
That is, not only can we mitigate the unpredictable performance issues
associated with mixed-integer (nonlinear) programs by using a fast strategy to
produce an initial solution, but we can also demonstrate that such a solution,
while not having clairvoyance regarding the demand in future time periods,
still yields a near-optimal objective function value.

3.1 Technology Selection

To reduce the size of the problem, we limit generator technologies and num-
ber of twins, i.e., the cardinality of sets G̃g , using Algorithm 1. Similarly, we
limit the number of battery technologies, i.e., the cardinality of B̃b , using Al-
gorithm 2. We fix the cardinality of the sets of generators rated less than
20 Scioletti et al.

100kW to two generators total. We justify our algorithms by the following

Pareto analysis of the objective function terms. Specifically, a solution with
fewer higher-rated generators dominates a solution with more lower-rated gen-
erators, because: (i) the cost of generators per rated power is not linear, e.g., a
15kW generator costs nearly 70% of a 100kW generator (see Table 2); (ii) the
objective function penalizes a dispatch strategy that employs a higher num-
ber of technologies to meet demand; and (iii) fuel consumption of generators is
similar for the four we consider (see Figure 2) regardless of the percent loading
at which we operate them.

Algorithm 1 Determines the generator technologies and respective number

of twins, i.e., cardinality of G̃g = {g1, g2, g3, g4}, per instance: see Table 2 for
details.
procedure GeneratorTechs&Twins
MaxDemand max dP
t2T t
M axDemand
if p̄g1
0.9 then
⇣ ⌘
|G̃g1 | max 1, b M axDemand
p̄
c
g1
|G̃g2 | 2
|G̃g3 | 2
|G̃g4 | 2
else
|G̃g1 | 0
|G̃g2 | 1
|G̃g3 | 1
|G̃g4 | 1

We limit battery procurement to one, but consider up to three sizes per

instance based on peak demand (see Figure 6) and Algorithm 2. Thus, for an
instance with a maximum demand of 239kW, we would consider battery sizes
of 200kW, 150kW, and 100kW.

Algorithm 2 Determines the battery technologies considered, i.e., cardinality

of B̃b = {b1, b2, b3, b4, b5, b6} per instance: see Table 3 for details.
procedure BatteryTechs
MaxDemand max dP
t2T t
if MaxDemand 100kW then
for n 2 {1,...,4} do
if MaxDemand 300kW 50n then
|B̃bn | 1
|B̃b1+n | 1
|B̃b2+n | 1
else
|B̃b5 | 1
|B̃b6 | 1
Hybrid Power System Design and Dispatch 21

3.2 Initial Feasible Solution

To produce an initial feasible solution for instances of (U ), our heuristic, which

we term H, uses technologies chosen by Algorithms 1 and 2. That is, H chooses
the better of two di↵erent, myopic dispatch strategies for each possible combi-
nation of technologies; we term such a combination a design decision and the
design and dispatch solution a design-dispatch pair. Each design decision incor-
porates diesel generators, batteries, and PV systems. Both dispatch strategies:
(i) attempt to run the diesel generators as close to their rated capacities as
possible in order to maximize their operating fuel efficiency and longevity;
(ii) use the maximum amount of energy from purchased PV systems in order
to reduce the fuel cost of running generators and to maximize the return on
the solar panels’ fixed cost; and (iii) employ batteries primarily to balance
load requirements and to provide spinning reserve. For each time period, both
strategies initialize all diesel generators to be o↵, and incrementally add ca-
pacity by exchanging the smallest-sized online generator for the next greater
size if it is o✏ine, and turning on the smallest generator if no larger o✏ine gen-
erator exists. The first dispatch strategy, which we term H1 , increases diesel
generator capacity in this way until load and spinning reserve can be met for
that time period. The second strategy, which we term H2 , is identical to H1 ,
but adds diesel capacity as long as generators can be run at their rated capac-
ities. While the former strategy causes batteries to operate at a lower state
of charge, the latter strategy typically operates batteries at a higher state of
charge owing to its policy of only adding a generator if its capacity is reached.
Neither strategy clearly dominates the other for our instances, and neither re-
quires more than a few seconds of computational e↵ort to produce a solution
under our implementation. After enumerating all design decisions and subse-
quently executing H1 and H2 on all such decisions, H chooses the lowest-cost
option as the initial solution to (U ). Figure 5 provides a flowchart for heuristic
H1 ; the figure for H2 is similar. Algorithm 3, given in the Appendix, provides
pseudocode for H1 ; the pseudocode for H2 is similar.

4 Numerical Results

We solve instances of (P) as a MINLP and (U ) as a MIP on a Sun Fire x2270

m2 with 24 processors (2.93 GHz each), 48 GB RAM, 1 TB HDD, using
GAMS 24.1.3 for fourteen di↵erent instances, each with specific technological
parameter values and system considerations.

4.1 Demand Profile and Technology Information

FOBs (Table 1), which are critical to the Department of Defense’s ability to
deploy combat forces throughout the world and numbered over 700 during the
peak of the Afghanistan and Iraq wars [52], would benefit from hybrid power
22 Scioletti et al.

Input: Purchase decision; net load (PV output Given the state of batteries,
subtracted from load) for all t; initial state of each calculate the upper bound on t ! t+1
battery; T (horizon period). their charge and discharge.
t ! 1.

Upgrade the set

of dispatched
generators.
No
No Can PV, batteries, and
dispatched generators meet
load requirements?
Yes
Yes
Are there any Is t=T?
purchased
generators that
are not currently Determine the net power flow through the assets
dispatched? required to meet load, maximizing output from PV
first and then generators.

Dispatch the
No No Are spinning reserve Yes assets as Yes
requirements met under
previously
the given power flow?
determined.
Output: Infeasible solution.
Output: Dispatch
solution and cost.

1
Fig. 5: Heuristic H determines a myopic dispatch strategy for a given
design decision.

because currently, power planning is not optimized for efficiency [55]. Mili-
tary commanders prioritize power reliability over energy efficiency and fuel
consumption, which is problematic because resupply operations are danger-
ous and expensive. In this subsection, we present instance-specific parameters
including demand data; technology data, which considers procurement quan-
tities; and a description that details how power flows from the hybrid system
to meet demand.

Table 1: We focus our research on remote locations with the characteristics

listed in this table.

Characteristic Requirement
Personnel <150 People
Land Area (n̄s ) <100m by 100m (flat surface)
Infrastructure Type <8 tents (200 by 400 )
Peak Power Demand (max dP
t ) <300kW
Time Horizon (|T |) one year (|T |=8,760 hrs)

4.1.1 Demand

We utilize EnergyPlus-simulated [6] FOB demand profiles for fourteen loca-

tions throughout the world based on an experiment conducted at the Base
Camp Integration Lab (BCIL) at Fort Devens, MA (see Figure 6) [27]. The
variation in both the maximum demand and the general shape of the demand
profiles provide a robust set for use in testing and validating our models.
Within each instance, demand appears to be relatively variable, i.e., there are
no discernible, repeating patterns. Because our instances are FOB-specific and
Hybrid Power System Design and Dispatch 23

therefore have a lifetime of one-year, ⌫ is equal to one. We set k s and k̄ to 0.3,

which is a common fraction used in power planning.

Fig. 6: EnergyPlus simulated year-long (⌫ = 1) power demand forecasts

for fourteen di↵erent locations around the world [27] at hourly time
fidelity (⌧ = 1) serve as our set of instances.

4.1.2 Technology and Supply Parameters

The generator technologies we consider are currently in the military’s inven-

tory, which implies that the associated maintenance and service parts are as
well. There are no hybrid technologies in the military inventory so the PV and
battery technologies selected are typical within each of the industries.
Generators
The generators considered in these instances have power ratings ranging
from 15kW to 100kW with estimated procurement costs based on a market
24 Scioletti et al.

analysis of similar technologies. We consider a fully burdened cost of fuel, which

is an estimate based on the sum of total cost of all personnel and equipment
necessary to move and, when necessary, protect it from the point-of-purchase
to point-of-use [52]. We assume an initial fully burdened cost of fuel of $50 per
gallon (see Subsection 2.3.1). We do not model the fluctuations in oil price,
but assume a small inflation rate. Table 2 displays the characteristics of the
generators.

Table 2: Generator technology characteristics used. (See Section 2.2 for defi-
nitions and units.)

g2G p̄g c̃g "g l̄g ⌘g bfg cfg

g1 100,000 $37,691 1 9,000 1 0.0644 0.95
g2 60,000 $31,967 1 9,000 1 0.0645 0.59
g3 30,000 $29,376 1 9,000 1 0.0593 0.54
g4 15,000 $25,573 1 9,000 1 0.0547 0.25

PV

We consider a 1-kW rated mono-crystalline standard PV panel (40 by 60 )

with an 18% efficiency in converting irradiance to power. We limit the number
of panels (n̄s ) to 75 given spatial restrictions associated with remote hybrid
applications. We assume that the procurement cost of each solar panel is $2000,
which is equivalent to $2 per Watt. PV panels designed for FOBs need to
maintain the expeditionary characteristics of ease of transport and setup so,
unlike some PV systems that track the sun, we assume PV panels are fixed-tilt-
and-angle panels that rely on a user to erect and position on the appropriate
azimuth.

Batteries

We use data from lithium-ion batteries manufactured by A123 [1], which

we assume cost $500 per kWh. Battery lifetime parameters result from a linear
fit to battery test data from [35]. We scale battery performance parameters to
the desired size for stationary applications (see Table 3).

Table 3: Battery technology parameters used. SOC limit s̄b is 1; ⌧ ·iavg

b = cref
b ;
all other parameter values are 0. (See Section 2.2 for definitions and units.)

b2B p̄b c̃b "b l̄b ⌘b+ ⌘b rbint avb bvb cref
b cb c+
b asoc
b dsoc
b
b1 250,000 $125,000 1 1,647 0.95 0.95 0.00336 10.62 214.69 1129 0.0401 3 0.801 -0.801
b2 200,000 $100,000 1 1,647 0.95 0.95 0.0042 10.62 214.69 904 0.0401 3 0.801 -0.801
b3 150,000 $75,000 1 1,647 0.95 0.95 0.0056 10.62 214.69 678 0.0401 3 0.801 -0.801
b4 100,000 $50,000 1 1,647 0.95 0.95 0.0084 10.62 214.69 452 0.0401 3 0.801 -0.801
b5 50,000 $25,000 1 1,647 0.95 0.95 0.01679 10.62 214.69 226 0.0401 3 0.801 -0.801
b6 25,000 $12,500 1 1,647 0.95 0.95 0.03358 10.62 214.69 113 0.0401 3 0.801 -0.801
Hybrid Power System Design and Dispatch 25

Power System

We assume that the hybrid system connects to the power demand through
the Power Distribution Illumination System Electrical (PDISE), which has an
internal bus capable of system frequency regulation. PDISE is the principal
distribution system of the U.S. Army and is compatible with military generator
sets from 5kW to 200kW [4, 5, 7]. Preliminary simulation runs have shown
our results to be feasible from a power flow standpoint.

Technologies and Twin Decisions

To reduce the size of the problem, we limit generator technologies and

number of twins, i.e., the cardinality of sets G̃g , using Algorithm 1. Similarly,
we limit the number of battery technologies, i.e., the cardinality of B̃b , using
Algorithim 2. We seed the linear solver with a solution obtained from the
heuristics coded in Python 2.7.4 [49] and given in Section 3.

Table 4: Cardinality of sets G̃g and B̃b using Algorithm 1 and Algorithm 2,
respectively (see Section 2.2 for definitions).

Instance |G̃g1 | |G̃g2 | |G̃g3 | |G̃g4 | |B̃b1 | |B̃b2 | |B̃b3 | |B̃b4 | |B̃b5 | |B̃b6 |
Bagram 2 2 2 2 0 1 1 1 0 0
Bamako 1 2 2 2 0 0 0 1 1 1
Boston 3 2 2 2 1 1 1 0 0 0
Brazzaville 0 1 1 1 0 0 0 0 1 1
Buenos Aires 2 2 2 2 0 0 1 1 1 0
Dili 0 1 1 1 0 0 0 0 1 1
Dushanbe 2 2 2 2 0 1 1 1 0 0
Gangneung 3 2 2 2 0 1 1 1 0 0
Istanbul 2 2 2 2 0 1 1 1 0 0
Kuwait 2 2 2 2 0 0 1 1 1 0
Mexico City 1 2 2 2 0 0 0 1 1 1
San Salvador 0 1 1 1 0 0 0 0 1 1
Springfield 3 2 2 2 1 1 1 0 0 0
Tallinn 3 2 2 2 1 1 1 0 0 0

4.2 Solving (P)

We attempt to solve (P) for a smaller time horizon of 24 hours using existing
MINLP solvers at their default settings that accept models coded in GAMS
version 24.1.3 [21], which include BARON version 12.5 [50] and Couenne
(Couenne Library 0.4) [18], as well as the open-source solver BONMIN (BON-
MIN Library 1.7) [20]. BONMIN solves a continuous nonlinear program at
each node of the search tree, while BARON attempts to underestimate the
objective function. Couenne, similarly, uses linear relaxations to bound the
26 Scioletti et al.

problem. Any feasible solution to (P) provides an upper bound to the prob-
lem. We set termination conditions as the minimum of a 5% optimality gap
and a time limit of three hours per instance.

Table 5: The size of (P) for a time horizon of one day (|T |=24) over all
instances ranges based on the set of technologies. “Maximum” represents the
instances that consider the highest number of technologies, while “Minimum”
depicts the contrary.

Constraints Variables
Linear Nonlinear Continuous Binary
Maximum 2,145 504 1,176 491
Minimum 834 336 638 198

The nonconvexity of (P) challenges these solvers (Table 6). BARON solves
(P) for four of the fourteen instances within the prescribed criteria, while
BONMIN solves only two given the same criteria. Interestingly, the former
solver provides tighter gaps than the latter, despite the latter being only a
local solver and therefore providing only local lower bounds. Couenne yields
feasible solutions for all instances, but none within the desired gap. By con-
trast, when we solve (U ) for |T | = 24 hours, we obtain solutions to all fourteen
instances within the desired gap in less than one second. Seeding (P) with the
procurement decision from (U ) and solving with any of the three pieces of
nonlinear software we use did not improve performance; seeding the nonlin-
ear model with more than the procurement decision from the linear model
renders the constraints associated with battery state-of-charge relationships
and battery lifecycle infeasible. The poor performance of these MINLP solvers
on the majority of these small instances suggests that (P) is not a tractable
formulation; therefore, we instead focus our e↵orts on (U ).

4.3 Solving (U )

Our model (U ) is a MIP for which we solve all fourteen instances using
CPLEX version 12.5.1.0 [2], a commercial state-of-the-art solver that employs
the branch-and-bound algorithm coupled with heuristics to improve the best
integer solution and cuts to improve bounds. We also employ our own heuris-
tics (see Section 3), the purpose of which is to provide our linear-integer solver
with a “warm start”; however, because they do so in a myopic manner, not
having clairvoyance regarding the variability in the load still allows us to pro-
duce solutions within approximately 5% of optimality with designs that can
be implemented in the field, thus mitigating the potential, detrimental e↵ects
on our solution of not having solved a stochastic model.
Because operational decisions at the beginning of the year likely have little
impact on those at the end of the year, we attempt to reduce model size
Hybrid Power System Design and Dispatch 27

Table 6: Solutions from (P) for each instance given a shortened time horizon
(|T | = 24 hrs). Termination criteria: min(3 hours, optimality gap  5%).
†Model did not find a feasible solution.

BARON BONMIN Couenne

Gap Time Gap Time Gap Time
Instance
(%) (HR:MIN) (%) (HR:MIN) (%) (HR:MIN)
Bagram 5.24 3:00 † 3:00 88.00 3:00
Bamako 8.50 3:00 17.44 3:00 71.52 3:00
Boston 5.86 3:00 † 3:00 85.18 3:00
Brazzaville 0.80 0:42 18.76 3:00 64.93 3:00
Buenos Aires 7.39 3:00 23.57 3:00 66.05 3:00
Dili 11.93 3:00 24.75 3:00 59.54 3:00
Dushanbe 5.11 3:00 † 3:00 86.81 3:00
Gangneung 0.06 0:02 3.44 0:26 83.30 3:00
Istanbul 9.21 3:00 9.99 3:00 70.72 3:00
Kuwait 17.35 3:00 17.35 3:00 69.21 3:00
Mexico City 17.24 3:00 17.24 3:00 67.30 3:00
San Salvador 6.99 3:00 0.00 1:29 63.30 3:00
Springfield 4.76 0:11 11.56 3:00 73.27 3:00
Tallinn 0.29 0:01 † 3:00 79.97 3:00

by aggregating instances into three-hour and twelve-hour time periods for all
14 locations with demand, PV power output per system, and fuel cost by
period set to the mean of each time period. However, because these decisions
are strongly linked by the design, the aggregated scenarios produce solutions
insufficiently robust to handle di↵erent operating circumstances at di↵erent
times of the year. Specifically, these solutions are associated with designs that
tend to have less battery capacity when compared to those obtained when
solving for hourly dispatch. Optimized dispatch with one-hour time periods
uses the battery to balance load while running diesel generators at or near the
rated capacity; however, longer time periods limit the battery’s maximum rate
of charge or discharge over a single time period, which curtails the battery
utility overall. Furthermore, the model sees less variability and lower peaks
with the aggregated values, which can favor designs with diminished diesel
and battery capacity that would be infeasible under problems with hourly
time periods. Therefore, we consider all 8,760 hours in our instances, which
yields large problems (see Table 7), and an incentive to develop strategies to
expedite solutions.
To this end, we use the non-default CPLEX setting “Threads 15” to facil-
itate concurrent optimization and “MemoryEmphasis 1,” which attempts to
reduce the memory storage requirements of the problem. We set branching
priorities based on complexity in the following, decreasing order: battery pro-
curement and generator procurement by rated power. Similar to (P), we set
an optimality gap termination condition of 5%; however, we set the time limit
to ten hours, because the year-long time horizon greatly increases the size of
the problem compared to (P).
28 Scioletti et al.

Table 7: Size of (U ) for each instance (|T |=8,760).

Constraints Variables
Linear Continuous Binary
Bagram 1,375,333 499,347 175,214
Bamako 1,340,290 499,346 175,214
Boston 1,778,296 621,995 262,821
Brazzaville 700,806 280,334 70,086
Buenos Aires 1,375,333 499,347 175,214
Dili 700,806 280,334 70,086
Dushanbe 1,375,333 499,347 175,214
Gangneung 1,778,296 621,995 262,821
Istanbul 1,375,333 499,347 175,214
Kuwait 1,375,333 499,347 175,214
Mexico City 1,340,290 499,346 175,214
San Salvador 700,806 280,334 70,086
Springfield 1,778,296 621,995 262,821
Tallinn 1,778,296 621,995 262,821

With the heuristic, we are able to find solutions using model (U ) for all
instances within a 5% optimality gap in three hours or fewer; results are re-
ported using the symmetry-breaking constraints (3e), (5e), and (5f); while
their e↵ect on performance is not uniform, we retain these redundant con-
straints to minimize long solve times. Table 8 displays solutions and solve
times for the fourteen instances with and without a warm start; those with a
higher maximum demand take longer to solve because the number of allowable
procurement decisions is greater (see Table 4).

Table 8: Solutions from (U ) for each instance (|T | = 8,760 hrs). Termination criteria: min(10 hours, optimality gap  5%)
(see Section 2.2 for definitions).

OBJ Value1 Fuel Use Gap Time Procurement2 Gen Total3 Max Demand4 Heuristic5
($) (gal) (%) (hrs:min) (kW) (kW) (hrs:min)
Bagram 2,092,916 29,048 4.93 6:31 g1 ,g1 ,g2 ,g3 ,g4 ,g4 ,b2 ,s1 305 240 1:31
Bamako 1,062,946 13,789 5.00 3:01 g1 ,g4 ,b4 ,s1 115 125 1:55
Boston 3,536,683 56,192 3.34 8:24 g1 ,g1 ,g1 ,g4 ,b3 ,s1 315 334 0:11
Brazzaville 1,270,267 17,677 3.28 0:47 g2 ,g4 ,b5 ,s1 75 84 3:09
Buenos Aires 1,678,616 24,015 4.98 7:27 g1 ,g1 ,b3 ,s1 200 185 0:10
Dili 1,478,706 22,133 4.89 0:36 g2 ,b5 ,s1 60 87 0:17
Dushanbe 2,205,276 31,652 3.53 7:57 g1 ,g2 ,g3 ,g4 ,g4 ,b2 ,s1 220 231 0:18
Gangneung 2,654,282 40,560 4.97 6:00 g1 ,g1 ,g1 ,b2 ,s1 300 255 0:17
Istanbul 2,243,992 33,530 3.99 5:51 g1 ,g1 ,g4 ,b2 ,s1 215 213 0:17
Kuwait 1,690,109 24,322 4.45 4:10 g1 ,g4 ,b3 ,s1 115 148 0:17
Mexico City 1,178,090 15,811 4.88 7:40 g1 ,g4 ,b4 ,s1 115 122 0:39
San Salvador 980,405 12,940 1.95 1:03 g2 ,g4 ,b5 ,s1 75 87 0:09
Springfield 2,655,202 39,569 2.69 7:50 g1 ,g1 ,g2 ,g4 ,b1 ,s1 275 315 0:18
Tallinn 4,047,348 64,557 3.24 6:12 g1 ,g1 ,g1 ,g4 ,b3 ,s1 315 309 0:17
1 OBJ Value: represents sum of lifecycle, fuel, and procurement costs; see Expression 1.
2 Procurement: note s1 represents 75 panels. Common to all instances is the purchase of the maximum number of solar panels.
3 Gen Total represents the sum of the power ratings for all procured generators in kW:
X X
p̄g Wgk .
g2G k2G̃g

4 max
Max Demand is the maximum demand in kW per time period per instance over the year-long time horizon: t2T
(1 + k̄)dP
t [kW].
5 Heuristic represents computation time to solve (U ) with a “warm start.”

The procurement decisions, measured by the sum of generator power, rely

heavily on the maximum power demand of the instance. When purchased,
Hybrid Power System Design and Dispatch 29

solar panels reduce the e↵ective demand, but we limit their procurement due
to the large area they occupy. Generally, the restrictions on the number of
panels limit the total PV output to a fraction of an instance’s peak demand so
PV rarely provides more power than demanded, thus increasing the necessity
for generators.
Generators operating at less than 30% of their rated power use more fuel
and require more maintenance. Our solutions demonstrate that generators
operate at close to their rated power whenever possible. Figure 7 depicts the
quotient of the total number of hours a generator operates at greater than 30%
of its capacity and the total number of hours it operates cumulatively over the
time horizon. On average throughout all instances, generators operate at or
above 30% of their rated power more than 96% of the time. The model chooses
to procure batteries for all cases.

Fig. 7: A graphical representation of generator performance, measured

as the quotient of the number of hours a generator operates at greater
than 30% rated power and the total number of hours it runs.

Figure 8 is a continuous-time depiction of a discrete 20-hour extract from

a year-long dispatch solution to (U ) involving two 100kW generators and a
150kW battery. For the battery, positive power values represent discharging,
while negative values indicate charging. In Region 1), the demand is greater
than 100kW, the maximum rating for the first generator. Instead of turning
on the second generator, the model chooses to discharge the battery to meet
the load. In Region 2), the load reaches a threshold where the second 100kW
generator turns on to meet the demand, but also charges the battery. Lastly,
in Region 3), the load drops below the threshold and the second generator
turns o↵ while the battery supplies some of the load. This short-term load
30 Scioletti et al.

shifting allows both generators to operate at high efficiency, which reduces fuel
consumption and demonstrates the usefulness of the battery in this situation.

Fig. 8: An example, extracted from the solution to the Buenos Aires

instance (see Table 8), of how two 100kW-generators and a battery dis-
patch power to meet demand for a 20-hour time interval.

Table 9 displays fuel consumption di↵erences between three procurement

options solved using (U ): (i) hybrid (generator, PV, and battery), (ii) generator-
only, and (iii) generator- and PV-only systems. Hybrid systems average 30%
fuel savings across all instances compared to a comparably sized generator-
only system. In one instance, hybrid systems provide a 50% reduction in fuel
use. Although the generator-and-PV-only system is a promising procurement
option that also reduces fuel consumption, the addition of the battery further
increases fuel savings by nearly 10% across all instances. These savings point
to the benefit of employing a method that considers design and dispatch si-
multaneously. Specifically, were we to consider design in isolation, we would
be forced to size the generator capacity for maximum demand to guarantee a
feasible solution; PV might not be available to help shave this peak while the
battery might not be charged to its maximum capacity. Because fuel use con-
tributes significantly to the objective function value, we could expect design
and dispatch solutions optimized in isolation to increase the objective function
value correspondingly.

4.4 Solution Quality

In this section, we present an analysis of the quality of solutions from (U ) by

identifying both where our approximations are imperfect and the magnitude
of these errors compared to the nonlinear model (P). We define the following
metrics:
Hybrid Power System Design and Dispatch 31

Table 9: Fuel consumption results for (U ) resulting from three procurement options, which include (i) a hybrid system (generator, PV, and
battery), (ii) a generator-only system, and (iii) a generator- and PV-only system, for each instance (|T | = 8,760 hrs). Termination criteria:
min(10 hours, optimality gap  5%).

Hybrid System Generator-Only Generator & Solar

Fuel Increase Fuel Fuel Increase
Fuel Fuel Fuel Increase Fuel Increase
vs Consumption vs.
Consumption Consumption / /
Hybrid System Hybrid System
(gallons) (gallons) Fuel Consumption Fuel Consumption
(gallons) (gallons) (gallons)
Bagram 29,048 43,555 14,507 33% 34,265 5,217 15%
Bamako 13,789 25,623 11,834 46% 17,008 3,219 19%
Boston 56,192 67,484 11,292 17% 58,188 1,996 3%
Brazzaville 17,677 27,405 9,728 35% 19,301 1,624 8%
Buenos Aires 24,015 34,283 10,268 30% 25,974 1,959 8%
Dili 22,133 34,758 12,625 36% 23,353 1,220 5%
Dushanbe 31,652 43,555 11,903 27% 35,508 3,856 11%
Gangneung 40,560 48,939 8,379 17% 42,864 2,304 5%
Istanbul 33,530 42,513 8,983 21% 35,548 2,018 6%
Kuwait 24,322 35,994 11,672 32% 26,268 1,946 7%
Mexico City 15,811 27,231 11,420 42% 18,179 2,368 13%
San Salvador 12,940 26,113 13,173 50% 16,263 3,323 20%
Springfield 39,569 50,934 11,365 22% 42,955 3,386 8%
Tallinn 64,557 70,736 6,179 9% 67,085 2,528 4%

+
Pbkt , Pbkt power into and out of battery b, twin k, in time period t,
respectively, computed from a dispatch solution obtained
from (U )
+
P̂bkt , P̂bkt power computed via nonlinear constraints (12) and (13),
respectively, into and out of battery b, twin k, in time
+
period t given values of Ibkt and Ibkt , respectively, and
soc
Bbk,t 1 obtained from (U )

+
bkt , bkt the di↵erence between the actual power into and out of
battery b, twin k, in time period t as modeled in (U ) (see
constraints (20) and (21)), respectively, and the
corresponding theoretical power as modeled in (P) [kW]

dˆpbk the sum over the time horizon of the quotient of the
di↵erence between the total amount of battery b, twin k
power over-estimated and the total amount
underestimated, and the demand converted to [kW] [%]

L̂bk theoretical lifecycles consumed by battery b, twin k over

the time horizon, as modeled in (P) per constraint (6b),
calculated post run
l
bk the quotient of the theoretical and actual lifecycles of
battery b, twin k (see constraints (6b) in (P) and (23) in
(U ), respectively)

bkt = (Pbkt P̂bkt ) 8b 2 B, k 2 B̃b , t 2 T (26)

+ + +
bkt = (Pbkt P̂bkt ) 8b 2 B, k 2 B̃b , t 2 T (27)

X✓ + ◆
dˆpbk = bkt bkt
100% 8b 2 B, k 2 B̃b (28)
t2T
dP
t

l L̂bk
bk = 100% 8b 2 B, k 2 B̃b (29)
Lbk

The results given in Table 10 demonstrate that across all fourteen instances,
relative to a corresponding, hypothetical solution from (P): (i) our model over-
estimates Pbkt by as much as 2.8kW per time period, but this is less than 5% of
+
the average demand; (ii) our model under-estimates Pbkt by as much as 1kW;
32 Scioletti et al.

and (iii) the combination of (i) and (ii) over the time horizon contributes to less
than 0.02% of the total demand summed over the time horizon. It is possible
that changes to the dispatch in one time period a↵ect dispatch in subsequent
time periods and, hence, the quality of our approximations; however, we seek
a dispatch solution at hourly time fidelity, which assumes a steady-state de-
mand and implies that small perturbations due to approximation error are not
particularly consequential. Lastly, (iv) our model over-approximates lifecycles
consumed by roughly 25%. This is acceptable because lifecycles provide only
a small contribution to the objective function value; furthermore, this over-
approximation results in a conservative assessment, which is desirable to o↵set
our omission of temperature e↵ects, which may age the battery more rapidly
than estimated in a solution to (U ).

The approximation error of our model results from the bounds correspond-
+
ing to the linearization associated with auxiliary variables Zbkt and Zbkt . Our
+
model seeks to minimize Zbkt as a means to conserve power generated either
to enter the battery or for direct use in meeting demand. On the other hand,
the higher the value of Zbkt , the more power we are able to withdraw from the
battery. The tightness of the coefficients on the variables on the right hand
+
side of constraints (25a) and (25b) in the case of Zbkt , and of constraints (25g)
and (25h) in the case of Zbkt , a↵ects the accuracy of the approximation. We do
note, however, that for all instances, the procurement strategy we obtain from
solving (U ) yields a feasible completion to the resulting mixed integer nonlin-
ear problem for a 24-hour instance. Longer horizons result in tractability (but
not necessarily feasibility) issues.

Table 10: An analysis of the linearized constraints in (U ) versus the nonlinear constraints in (P).
Positive values represent an over-estimation, while negative values represent an under-estimation.

Percent of Lifecycle
Purchased Discharge2 Charge3
max max + Demand4 Approximation5
Battery1 ( b2B,k2 ) ( b2B,k2 bkt ) max
B̃ ,t2T bkt
b B̃ ,t2T
b ( max dˆp )
b2B,k2B̃b bk
( b2B,k2 l )
B̃ bkb
Bagram b2 2.30 -0.80 0.031 90
Bamako b4 1.15 -0.40 0.050 99
Boston b3 1.73 -0.60 0.004 70
Brazzaville b5 0.58 -0.20 0.005 71
Buenos Aires b3 1.73 -0.60 0.030 76
Dili b5 0.57 -0.20 -0.004 69
Dushanbe b2 2.23 -0.80 0.010 67
Gangneung b2 2.33 -0.80 0.023 72
Istanbul b2 2.31 -0.80 0.024 74
Kuwait b3 1.73 -0.60 0.033 78
Mexico City b4 1.16 -0.40 0.025 79
San Salvador b5 0.57 -0.20 0.006 71
Springfield b1 2.89 -1.00 0.021 71
Tallinn b3 1.74 -0.60 0.013 76
1 Purchased Battery details the type of battery purchased as part of the procurement solution for each instance.
2 Discharge represents the difference between the actual and theoretical power discharged by the battery in kWs (over-approximation 0) per
Equation (26).
3 Charge represents the difference between the actual and theoretical power received by the battery in kWs (under-approximation  0) per Equation
(27).
4 Percent of Demand is the sum of the quotient of the difference of total amount of batter power over-estimated minus the total amount underesti-
mated and the demand over the time horizon per Equation (28).
5 Lifecycle Approximation represents the quotient of the theoretical and actual total lifecycles per Equation (29).
Hybrid Power System Design and Dispatch 33

5 Conclusions

We present an optimization model that determines the procurement and dis-

patch strategy for a year-long demand profile at hourly time fidelity. Our for-
mulation handles up to three choices of battery technologies. We use fourteen
year-long demand profiles at the hourly fidelity for FOBs located in di↵erent
climate zones, and solve all instances using a mixed-integer, linear approxi-
mation of the mixed-integer nonlinear program well within a time limit of 10
hours to an optimality gap of less than 5% for a prescribed set of technolo-
gies; using a heuristic warm start reduces solution times to within under an
hour in most instances we test. Results suggest a hybrid system such as ours
reduces fuel consumption 30%, on average, compared to a generator-only so-
lution. Our solutions indicate a design and dispatch strategy that charges the
battery when demand is low and then discharges the battery to prevent op-
erating generators at a low-power rating. The implicit benefit of this strategy
is that generators maintain levels greater than 30% of their rated power, on
average, 96% of the time they are in use.
We evaluate the quality of our approximation by comparing solutions from
(U ) against the nonlinear representations of power and lifecycle variables in
(P). We find that although our model overestimates battery discharge power,
especially at low SOCs, the total quantity over-estimated is less than 0.02% of
the total demand. This approximation error stems from the bounds on SOC,
which are 0 and 1 (see Section 4.4). Bounds associated with partitioning on
SOC or applying operational logic could reduce this error; however, it may be
at the expense of increased solve times and/or reduced solution quality.
Rather than minimizing costs, our model could easily incorporate objec-
tives such as minimizing environmental impact or total volume of the tech-
nologies procured. The battery parameter calculations employed for our model
are applicable to other chemistries, such as lead acid and nickel cadmium. Be-
cause our model solves for both current and voltage, results from (U ) would be
useful in relating design and dispatch solutions to more detailed dispatch and
power flow models that consider finer-grain time fidelity. Specifically, future
work entails determining dispatch decisions at minute-level fidelity by fixing
the design decisions and introducing greater operational detail, including: (i)
ramp-up and ramp-down time of generators, (ii) minimum up- and down-time
of generators, (iii) rules of thumb by which some controls systems operate,
and (iv) more accuracy in battery performance, e.g., performance factors as
a function of temperature. Such a model is designed to operate using a one-
to two-day “look-ahead” window, consisting of between 1,440 and 2,880 time
periods. While the number of variables would be smaller than in (P) because
of the fixed procurement decisions, the minute-fidelity model contains more
constraints. An alternative approach would use our strategies that determine
the design influenced by a coarse dispatch strategy, and then simulate the dis-
patch using a rolling-horizon approach; a drawback results from the inability
to obtain gradients with respect to the design variables out of that simula-
34 Scioletti et al.

tion, so we would have to use derivative-free optimization techniques. Another

extension of our model, possibly addressed through the simulation approach,
would allow for the construction of a stochastic program to incorporate the
variability of both solar irradiance and power demand.

Acknowledgements The authors would like to thank Dr. Mark Spector, Office of Naval
Research (ONR) for full support of this research e↵ort under contract award #N000141310839.
We appreciate the consult of Dr. Paul Kohl of Georgia Tech, on whom we relied on heav-
ily for battery expertise. We are also grateful to Evan Jones and Kimberly Fowler, Pacific
Northwest National Lab (PNNL), for providing the FOB load data from EnergyPlus. Thanks
is due to Gavin Goodall for his help in conducting model runs. Lastly, we acknowledge the
support of the National Renewable Energy Lab (NREL) for its involvement in this project,
particularly in addressing PV applications and specifically to Stephen Frank for helpful
comments on prior versions of this paper.
Hybrid Power System Design and Dispatch 35

References

1. A123 battery manufacturer. http://www.a123systems.com/. Accessed:

2015-03-15.
2. CPLEX 12. http://www.gams.com/dd/docs/solvers/cplex/. Accessed:
2015-03-15.
3. Fuel Consumption for Diesel Generators. http://
greenmountaingenerators.com/2012/09/. Accessed: 2015-07-10.
4. Intelligent Power Management Distribution System (IPMDS). http:
//www.dtic.mil/ndia/2011power/Session2_12093_Whitmore.pdf. Ac-
cessed: 2015-05-27.
5. Power Distribution Illumination System Electrical (PDISE). http:
//www.fidelitytech.com/military-and-aerospace-manufacturing/
products-and-services/pdise/. Accessed: 2015-05-27.
6. EnergyPlus: creating a new-generation building energy simulation pro-
gram. Energy and Buildings, 33(4):319 – 331, 2001.
7. Army TM 9-6150-226-13. http://www.liberatedmanuals.com/
TM-9-6150-226-13.pdf, 2008. Accessed: 2015-05-27.
8. Micropower system modeling with HOMER. http://homerenergy.com,
2015. Accessed 08/18/2015.
9. C. Abbey and G. Joos. Energy management strategies for optimization of
energy storage in wind power hybrid system. In Power Electronics Spe-
cialists Conference, 2005. PESC’05. IEEE 36th, pages 2066–2072. IEEE,
2005.
10. C. Abbey and G. Joos. A stochastic optimization approach to rating of
energy storage systems in wind-diesel isolated grids. Power Systems, IEEE
Transactions on, 24(1):418–426, 2009.
11. N. Achaibou, M. Haddadi, and A. Malek. Modeling of lead acid batteries
in PV systems. Energy Procedia, 18:538–544, 2012.
12. I. Androulakis, C. Maranas, and C. Floudas. ↵BB: A global optimization
method for general constrained nonconvex problems. Journal of Global
Optimization, 7(4):337–363, 1995.
13. M. Ashari and C. Nayar. An optimum dispatch strategy using set points
for a photovoltaic (PV) diesel and battery hybrid power system. Solar
Energy, 66(1):1 – 9, 1999.
14. S. Ashok. Optimised model for community-based hybrid energy system.
Renewable Energy, 32(7):1155 – 1164, 2007.
15. B. Bala and S. Siddique. Optimal design of a PV-diesel hybrid system for
electrification of an isolated island in Sandwip, Bangladesh using a genetic
algorithm. Energy for Sustainable Development, 13(3):137 – 142, 2009.
16. T. Barbier, M. Anjos, and G. Savard. Optimization of diesel, wind, and
battery hybrid power systems. Submitted, 2014.
17. C. D. Barley and C. B. Winn. Optimal dispatch strategy in remote hybrid
power systems. Solar Energy, 58(46):165 – 179, 1996.
18. P. Belotti. Couenne: a user’s manual. Technical report, Lehigh University,
2009.
36 Scioletti et al.

19. J. L. Bernal-Agustin, R. Dufo-Lopez, and D. M. Rivas-Ascaso. Design of

isolated hybrid systems minimizing costs and pollutant emissions. Renew-
able Energy, 31(14):2227 – 2244, 2006.
20. P. Bonami, L. T. Biegler, A. R. Conn, G. Cornuéjols, I. E. Grossmann,
C. D. Laird, J. Lee, A. Lodi, F. Margot, and N. Sawaya. An algorithmic
framework for convex mixed integer nonlinear programs. Discrete Opti-
mization, 5(2):186–204, 2008.
21. A. Brooke, D. Kendrick, and A. Meeraus. GAMS – A User’s Guide (Re-
lease 2.25). Boyd & Fraser Publishing Company, Danvers, Massachusetts,
1992.
22. J. Copetti and F. Chenlo. Lead acid batteries for photovoltaic applications:
Test results and modeling. Journal of Power Sources, 47(1):109–118, 1994.
23. A. P. Dobos. PVWatts Version 1 technical reference. National Renewable
Energy Laboratory, Report Number TP-6A20-60272, 2013.
24. D. Doer↵el and S. A. Sharkh. A critical review of using the Peukert
equation for determining the remaining capacity of lead-acid and lithium-
ion batteries. Journal of Power Sources, 155(2):395–400, 2006.
25. R. Dufo-Lopez and J. L. Bernal-Agustin. Design and control strategies of
PV-diesel systems using genetic algorithms. Solar Energy, 79(1):33 – 46,
2005.
26. R. Dufo-López, J. M. Lujano-Rojas, and J. L. Bernal-Agustı́n. Comparison
of di↵erent lead–acid battery lifetime prediction models for use in simu-
lation of stand-alone photovoltaic systems. Applied Energy, 115:242–253,
2014.
27. M. Engels, P. A. Boyd, T. M. Koehler, S. Goel, D. R. Sisk, D. D. Hatley,
V. V. Mendon, and J. C. Hail. Smart and Green Energy (SAGE) for
base camps final report. Technical report, Pacific Northwest National
Laboratory (PNNL), Richland, WA (US), 2014.
28. I. J. Fernndez, C. F. Calvillo, A. Sanchez-Miralles, and J. Boal. Capacity
fade and aging models for electric batteries and optimal charging strategy
for electric vehicles. Energy, Sustainability and Society, 60:35–43, 2013.
29. A. Flores-Tlacuahuac and L. T. Biegler. A robust and efficient mixed-
integer non-linear dynamic optimization approach for simultaneous design
and control. Computer Aided Chemical Engineering, 20:67–72, 2005.
30. L. Gao, S. Liu, and R. A. Dougal. Dynamic lithium-ion battery model for
system simulation. IEEE: Transactions on Components and Packaging
Technologies, 25(3):495–505, 2002.
31. T. Givler and P. Lilienthal. Using HOMER software, NREL’s micropower
optimization model, to explore the role of gen-sets in small solar power
systems. Technical Report NREL/TP-710-36774, National Renewable En-
ergy Laboratory, May 2005.
32. H. Gooi, D. Mendes, K. Bell, and D. Kirschen. Optimal scheduling of
spinning reserve. Power Systems, IEEE Transactions on, 14(4):1485–1492,
1999.
33. F. Huneke, J. Henkel, J. A. B. González, and G. Erdmann. Optimisa-
tion of hybrid o↵-grid energy systems by linear programming. Energy,
Hybrid Power System Design and Dispatch 37

Sustainability and Society, 2(1):1–19, 2012.

34. M. R. Jongerden and B. R. Haverkort. Which battery model to use?
Software, IET, 3(6):445–457, 2009.
35. M. Jun, K. Smith, E. Wood, and M. C. Smart. Battery capacity esti-
mation of low-earth orbit satellite application. International Journal of
Prognostics and Health Management, 3:81, 2012.
36. S. Kamel and C. Dahl. The economics of hybrid power systems for sus-
tainable desert agriculture in Egypt. Energy, 30(8):1271 – 1281, 2005.
37. Y. Katsigiannis and P. Georgilakis. Optimal sizing of small isolated hybrid
power systems using tabu search. Journal of Optoelectronics and Advanced
Materials, 10(5):1241, 2008.
38. E. Koutroulis, D. Kolokotsa, A. Potirakis, and K. Kalaitzakis. Methodol-
ogy for optimal sizing of stand-alone photovoltaic-wind-generator systems
using genetic algorithms. Solar Energy, 80(9):1072 – 1088, 2006.
39. J. F. Manwell and J. G. McGowan. Lead acid battery storage model for
hybrid energy systems. Solar Energy, 50(5):399–405, 1993.
40. V. Marano, S. Onori, Y. Guezennec, G. Rizzoni, and N. Madella. Lithium-
ion batteries life estimation for plug-in hybrid electric vehicles. IEEE
Vehicle and Power Propulsion Conference, pages 536–543, 2009.
41. N. C. McCaskey. Renewable energy systems for forward operating bases:
A simulations-based optimization approach. Technical report, Colorado
State University, 2010.
42. G. P. McCormick. Computability of global solutions to factorable noncon-
vex programs: Part I: Convex underestimating problems. Mathematical
Programming, 10(1):147–175, 1976.
43. H. Morais, P. Kadar, P. Faria, Z. A. Vale, and H. Khodr. Optimal schedul-
ing of a renewable micro-grid in an isolated load area using mixed-integer
linear programming. Renewable Energy, 35(1):151 – 156, 2010.
44. M. A. Ortega-Vazquez and D. S. Kirschen. Optimizing the spinning re-
serve requirements using a cost/benefit analysis. Power Systems, IEEE
Transactions on, 22(1):24–33, 2007.
45. M. A. Ortega-Vazquez and D. S. Kirschen. Estimating the spinning reserve
requirements in systems with significant wind power generation penetra-
tion. Power Systems, IEEE Transactions on, 24(1):114–124, 2009.
46. K. A. Pruitt, R. J. Braun, and A. M. Newman. Evaluating shortfalls in
mixed-integer programming approaches for the optimal design and dis-
patch of distributed generation systems. Applied Energy, 102:386–398,
2013.
47. Y. Rebours and D. Kirschen. What is spinning reserve? The University
of Manchester, pages 1–11, 2005.
48. S. Rehman and L. M. Al-Hadhrami. Study of a solar PV diesel battery
hybrid power system for a remotely located population near Rafha, Saudi
Arabia. Energy, 35(12):4986 – 4995, 2010.
49. G. Rossum. Python reference manual. Technical report, Amsterdam, The
Netherlands, The Netherlands, 1995.
38 Scioletti et al.

50. N. V. Sahinidis. BARON: A general purpose global optimization software

package. Journal of Global Optimization, 8:201–205, 1996.
51. M. S. Scioletti, J. K. Goodman, P. A. Kohl, and A. M. Newman. A
physics-based integer-linear battery modeling paradigm. Applied Energy,
176:245–257, 2016.
52. SERDP. Sustainable forward operating bases. https://www.
serdp-estcp.org/content/download/.../FOB_Report_Public.pdf,
2010.
53. S. Shaahid and I. El-Amin. Techno-economic evaluation of o↵-grid hybrid
photovoltaic diesel battery power systems for rural electrification in Saudi
Arabia as a way forward for sustainable development. Renewable and
Sustainable Energy Reviews, 13(3):625 – 633, 2009.
54. H. D. Sherali and J. C. Smith. Improving discrete model representa-
tions via symmetry considerations. Management Science, 47(10):1396–
1407, 2001.
55. J. Vavrin. Power and energy considerations at forward oper-
ating bases. http://e2s2.ndia.org/pastmeetings/2010/tracks/
Documents/9874.pdf, 2010.
56. J. Wang, J. H.-G. P. Liu, E. Sherman, S. Soukiazian, M. Verbrugge,
H. Tataria, J. Musser, and P. Finamore. Cycle-life model for graphite-
LiFePO4 cells. Power Sources, 196(8):3942–3948, 2011.
Hybrid Power System Design and Dispatch 39

Appendix

Algorithm 3 Attempts to determine a myopic dispatch strategy for a given

design decision: a set of generators g 2 G and their twins k 2 G̃g , batteries
b 2 B and their twins k 2 B̃b , and PV systems s 2 S in the design, with G̃g
and B̃b provided by Algorithms 1 and 2, respectively. Let g 2 Ĝ denote the set
of all generators in the design decision, including twins, in decreasing order by
rated capacity.
procedure H1
t 1
while t  |TP
| do
Rest s
st k Xs . spinning reserve requirement, if using max PV
s2S
Loadt (1 + k̄)dP
t . load requirement
Ggt 0, 8g 2 Ĝ . all generators are o↵ to start; Ĝ ordered by max output
Gent 0 . capacity
⇢ of generators that are turned on
ref
cb
BOutt ⌘b · (avb Bb,t
soc
1 + bvb iavg
b rbint ) · min iU
b ,
soc
Bb,t 1
cb +⌧
8 9 . max discharge
< ref
avg int
ref
+ cb c (1 soc
Bb,t )=
(av soc v
b Bb,t 1 +bb +ib rb ) min iU , + , b 1
: b c ⌧ ;
b
BInt . max charge
⌘b+
BRest ⌘ p̄ B soc . spinning reserve from idle battery
Pb b b,t 1
PVt st Xs . maximum PV power output
s2S
while Gent + BOutt + PVt < Loadt or Gent (Loadt PVt ) + BRest < Rest do
ĝ max{g 2 Ĝ : Ggt = 0}
if Gĝ 1,t = 0 then
Gĝ 1,t 1 . turn o↵ smallest running generator
Gĝt 0 Gent Gent ⌘ĝ p̄ĝ + ⌘ĝ 1 p̄ĝ 1 . turn on next smallest
else if G|Ĝ|t = 0 then . turn on smallest running generator
G|Ĝ|t 1; Gent Gent + ⌘ p̄
|Ĝ| |Ĝ|
else return Infeasible . all generators on, load or spinning reserve unmet
if PV t >
P PV Load t + BIn t then . use only PV, max charge battery
+
Pst Loadt + BInt ; Pbt 0; Pbt BInt ; Pgt 0 8g 2 Ĝ
s2S
40 Scioletti et al.

else if PVt + Gent > Loadt + BInt then . max charge battery
PV
Pst st Xs 8s 2 S; Pbt 0
+ P
Pbt BInt ; Pgt Loadt + BInt PVt
g2Ĝ
else if PVt + Gent > Loadt then . charge battery
PV
Pst st Xs 8s 2 S; Pbt 0
+
Pbt PVt + Gent Loadt ; Pgt p̄g Ggt 8g 2 Ĝ
else . discharge battery
PV (Loadt PVt Gent ) +
Pst st Xs 8s 2 S; Pbt ; Pbt 0
⌘b
Pgt p̄g Ggt 8g 2 Ĝ
+
soc soc Pbt Pbt
Bbt Bb,t 1 + ref avg int ref avg int
cb (av B soc +bv
b b,t 1 b
+ib rb ) cb (av B soc +bv
b b,t 1 b
ib rb )
. update battery SOC
t t+1 . continue until time horizon completed or infeasible
P V 8s 2 S, t 2 T ; P + 8t 2 T ; P
return Pst bt bt 8t 2 T ; Pgt 8g 2 Ĝ, t 2 T ;
soc 8t 2 T ; G
Bbt gt 8g 2 Ĝ, t 2 T . calculate cost
This material is based upon work supported by the U.S. Department of Energy, Office of Science, under contract
number DE-AC02-06CH11357.

The submitted manuscript has been created by UChicago Argonne, LLC, Operator of Argonne National Laboratory
(“Argonne”). Argonne, a U.S. Department of Energy Office of Science laboratory, is operated under Contract No.
DE-AC02-06CH11357. The U.S. Government retains for itself, and others acting on its behalf, a paid-up
nonexclusive, irrevocable worldwide license in said article to reproduce, prepare derivative works, distribute copies
to the public, and perform publicly and display publicly, by or on behalf of the Government. The Department of
Energy will provide public access to these results of federally sponsored research in accordance with the DOE
Public Access Plan. http://energy.gov/downloads/doe-public-access-plan.