Integrated Production and Energy Supply Planning on an Industrial Site
by Mixed-Integer Linear Programming
Ruiwen Liao and C
´
eline Gicquel
Laboratoire Interdisciplinaire des Sciences du Num
´
erique, Universit
´
e Paris Saclay, France
Keywords:
Production Planning, Lot-Sizing, Renewable Energy, Energy Supply, Mixed-Integer Linear Programming.
Abstract:
Industrial production sites are nowadays faced with two major concerns: the need to reduce the environmental
impact of their processes and the economic difficulties caused by the rising energy prices. Both challenges may
be partially tackled by powering industrial processes with electricity generated on-site from renewable sources.
However, the volatility of the electricity prices and the intermittence of the locally generated renewable energy
sources result in the need to solve an integrated industrial production and energy supply planning problem.
This work investigates a single-machine multi-product lot-sizing problem for an industrial process powered by
both grid electricity and on-site renewable energy. We propose a new extension of the Proportional Lot-sizing
and Scheduling Problem relying on a two-level structure for time discretization. The first level is related to
the product demand satisfaction, the second one is used for both production and energy supply planning. The
proposed extension is compared to a previously published extension of the General Lot-sizing and Scheduling
Problem dealing with a similar problem. Our preliminary numerical results show that in most cases, our model
provides a production plan of the same cost, but with a significantly reduced computational effort.
1 INTRODUCTION
Industrial companies are increasingly under pressure
to mitigate the CO2 and pollution emissions linked
to the manufacturing of industrial products. They are
also confronted with a sharp rise of the price of con-
ventional energy sources (gas, grid electricity...) so
that the availability and affordability of energy is be-
coming a critical parameter in manufacturing. One
way to deal with these two challenges consists in pow-
ering industrial processes with electricity generated
on-site from renewable sources.
In an industrial production plant, the energy man-
agement system is usually limited to two elements:
the energy bought from an outside supplier on the na-
tional market and the energy consumed by the manu-
facturing processes. As conversion devices (e.g. wind
turbines, photovoltaic panels) able to produce elec-
tric power from renewable energy sources are becom-
ing less expensive, it is now possible for an industrial
plant to build its own environmental friendly energy
system and to use this ’green’ electricity to power its
processes. For example, more than 30000 solar panels
have been installed on the rooftop of Bentley’s factory
in Crewe as well as on the roof of its carport, which
provide up to two third of the factory’s energy demand
with a power generation capacity of 7.7kW. In Cal-
ifornia, the brewery factory Anheuser-Busch’s Bud-
weiser installed 6500 solar panels as well as wind tur-
bines, which provide 30% of the factory’s electricity
use. Batteries may also be installed to store electric-
ity. However, the intermittence of renewable energy
sources (wind, sun...) makes it impossible to fully re-
place gas or grid electricity by on-site generated elec-
tricity to power an industrial process: both types of
energy should thus be used in combination. More-
over, the time-of-use pricing scheme widely used by
electricity providers means that it is necessary to ac-
curately track the timing and quantity of grid electric-
ity bought from these providers. In this circumstance,
making sure that energy supply and consumption are
balanced at all time is challenging. Thus, an inte-
grated energy supply and industrial production plan-
ning problem needs to be solved.
With all these challenges mentioned above,
energy-efficient production planning has attracted
great research attention in the last decades (Terbrack
et al., 2021; Gao et al., 2020). Some works con-
sider that all the needed energy comes from electricity
bought from the main electric grid at a time-varying
price and introduce energy consumption costs in the
objective function of the production planning model:
Liao, R. and Gicquel, C.
Integrated Production and Energy Supply Planning on an Industrial Site by Mixed-Integer Linear Programming.
DOI: 10.5220/0011602900003396
In Proceedings of the 12th International Conference on Operations Research and Enterprise Systems (ICORES 2023), pages 119-126
ISBN: 978-989-758-627-9; ISSN: 2184-4372
Copyright
c
2023 by SCITEPRESS Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
119
see e.g. (Johannes et al., 2019). Other works study
the case where energy can be obtained from both the
main electric grid and a non-adjustable on-site elec-
tricity generation system based on renewable energy
sources. Most of these studies focus on the short-
term scheduling of a job shop (Golp
ˆ
ıra et al., 2018)
or a flow shop (Biel et al., 2018; Wang et al., 2020).
However, lot-sizing in the presence of intermittent re-
newable energy sources has been rarely studied in the
literature. (Rodoplu et al., 2019) addressed a single-
item multi-machine lot-sizing problem. Chance con-
straints were introduced to deal with the fluctuation of
renewable energy sources. A lot-sizing and schedul-
ing problem was investigated in (Wichmann et al.,
2019), with consideration of renewable energy re-
sources and energy storage system.
The present work focuses on a short-term produc-
tion planning problem, namely a multi-item single-
machine lot-sizing problem, and seeks to integrate
the energy supply planning into the problem mod-
eling. Basically, a lot-sizing problem is a produc-
tion planning problem in which we seek to find the
optimal trade-off between the startup and inventory
holding costs. Startup costs are fixed costs occur-
ring every time the machine’s status changes in or-
der to process of a new type of product, while in-
ventory holding costs represent the costs of keeping
finished products in inventory between the time they
are produced and the time they are used to satisfy the
customers’ demand. In our case, the energy supply
system comprises three main elements: on-site power
generation devices producing a time-varying amount
of green electricity from renewable energy sources, an
on-site energy storage system and the main electric
grid with which electricity may be traded at a time-
varying price. See Figure 1 for a graphical represen-
tation of the studied integrated energy supply and in-
dustrial production system.
Inventory
Power generation devices using
renewable energy sources
Wind turbine
PV panel
Main grid
Energy system
External
demands
Integrated system
Production
system
Figure 1: Integrated energy supply and industrial produc-
tion system.
One modelling difficulty here comes from the fact
that the time discretization needed to manage prod-
uct demand satisfaction, production planning and en-
ergy supply may significantly differ. Indeed, follow-
ing the terminology introduced by (Copil et al., 2017),
we have to handle two exogenous time structures, i.e.
two sets of points in time at which externally given
events, that are defined by the data of the model, are
considered. The first time structure is imposed by the
timing of the external demand and usually uses rather
large time buckets (typically days or weeks). The sec-
ond one is imposed by the discrete time grid used to
track the availability of the on-site generated electric-
ity and the varying price of grid electricity. This time
grid usually uses much smaller time buckets (typi-
cally hours or 10-minutes intervals). In between these
two exogenous time structures, to plan production, we
have to handle a third endogenous time structure rep-
resenting the points in time at which internal events
are captured by decision variables.
As mentioned above, the problem of determining
optimal production lot sizes under intermittent renew-
able energy has been scarcely studied in the litera-
ture. A noticeable exception can be found in (Wich-
mann et al., 2019). This paper namely investigates
an integrated industrial production and energy sup-
ply planning problem for a manufacturing site pow-
ered by a decentralized power system using renew-
able energy sources. Their problem modeling relies
on a three-level time structure: a coarse time dis-
cretization to track the customers’ demands, a fine
energy-oriented time discretization to track the varia-
tions in the energy price and availability, and another
flexible production-oriented fine discretization to de-
termine the production plan. The resulting model can
be seen as an extension of the General Lot-sizing and
Scheduling Problem (GLSP) introduced by (Fleis-
chmann and Meyr, 1997) and takes the form of a
large-size mixed-integer linear program (MILP) for-
mulation involving many big-M constraints. Conse-
quently, numerical difficulties arise when trying to
solve medium-size instances of the problem with a
mathematical programming solver.
This is why we propose in this paper a new model
for the integrated energy supply and industrial pro-
duction planning problem. This model relies on a
two-level time structure. This one consists in a coarse
time discretization to track the demand satisfaction
and in a single fine time discretization based on the
exogenous time structure imposed by the energy-
related input data. This fine time discretization is used
both to track the variations in the energy price and
availability and to plan the industrial production. The
use of a fixed time grid involving short planning pe-
ICORES 2023 - 12th International Conference on Operations Research and Enterprise Systems
120
riods to determine the production lot sizes leads to
the formulation of a small bucket’ lot-sizing prob-
lem, i.e. to the formulation of a lot-sizing problem
in which there are some restrictions on the number
of item types that may be produced in each planning
period. More precisely, we consider here the Pro-
portional Lot-sizing and Scheduling Problem (PLSP)
in which at most two different types of product may
be produced in each period: see (Drexl and Haase,
1995). The main advantage of the proposed modeling
approach is that it leads to an MILP formulation much
smaller that the one obtained with the model proposed
in (Wichmann et al., 2019).
Our contributions are thus twofold. First, we pro-
pose a new extension of the PLSP in order to simulta-
neously plan the industrial production and the energy
supply in a manufacturing plant powered, at least par-
tially, by a decentralised energy system using renew-
able energy sources. Second, we carry out numerical
experiments to compare the proposed model with the
GLSP-based model previously published by (Wich-
mann et al., 2019). Our numerical results show that
in most cases, our model provides a production plan
of the same cost than the one provided by the model
introduced by (Wichmann et al., 2019) but with a sig-
nificantly reduced computational effort.
The paper is organized as follows. Section 2 pro-
vides a detailed description of the optimization prob-
lem under study. Section 3 presents the proposed
PLSP-based model based on a two-level time struc-
ture. Section 4 discusses the results of the compu-
tational experiments conducted on medium-size ran-
domly generated instances. Conclusion and research
perspectives are given in Section 5.
2 PROBLEM DESCRIPTION
We seek to plan production over a finite horizon in an
industrial plant producing J types of item. The set of
items is denoted as J = {1, 2,..., J}. A first exoge-
nously imposed time grid is introduced to track the
points in time at which the external customers’ de-
mand occurs. The planning horizon is thus divided
into a set T = {1,2,. ..,T } of macroperiods of fixed
length l: the end of each macroperiod corresponds to
a point in time at which demand may occur. Let d
j,t
denote the demand for item j to be satisfied at the end
of macroperiod t. Backlogging is not allowed, i.e., all
the demand should be satisfied on time. At the end
of each macroperiod, the finished products remaining
after the demand has been satisfied are held in inven-
tory in the warehouse for later use. Each unit of item
j held in inventory between two macroperiods incurs
a unit cost h
j
. The initial inventory of item j is given
by
¯
I
j,0
.
We assume that the production process comprises
a single production resource. This resource may be in
three distinct kinds of states: idle, setup for a given
item, and manufacturing a given item. The resource
is idle when it is turned on but not ready for produc-
tion. It may also be setup for a specific item: in this
case, it is ready to produce this type of item but is not
producing it. Finally, the resource is in a item-related
manufacturing state when it is actually producing this
specific type of item: this is however possible only
if the resource is already set up for this type of item.
Let k
j
be the amount of time needed to produce one
unit of item j on the resource. A startup is the action
of changing the machine status to prepare it to pro-
duce a new type of item. A startup for item j thus
occurs whenever the state of the resource is switched
from any other idle, setup or manufacturing state to
the setup state corresponding to this item. Let f
j
be
the fixed startup cost to be paid each time a startup for
item j happens. All startup times are assumed to be
negligible.
Energy is consumed by startups and manufactur-
ing. The amount of energy consumed during a startup,
denoted by e
f
j
, depends on the product that the ma-
chine will be setup for. The energy consumed during
manufacturing is proportional to the number of items
produced. Let e
j
denote the energy needed to produce
one unit of item j. We assume that no energy is con-
sumed when the machine is idle or when it is setup
for a given product but not producing.
The energy is supplied either from the main grid
or from the on-site renewable energy conversion de-
vices. The necessity to track the time variations of
the main grid energy price and of the amount of en-
ergy generated by the on-site power system leads us
to consider a second exogenously imposed time grid.
This one is based on planning periods sufficiently
short to assume that the energy price and the green
energy power generation are constant during a pe-
riod. Each macroperiod t = 1, ...,T of length l is
thus split into n energy-oriented microperiods of fixed
length l
e
such that l = nl
e
. Let R = {1, ...,nT } be
the set of all energy-oriented microperiods and R
t
=
{(t 1)n + 1, ...,tn} be the set of energy-oriented mi-
croperiods belonging to macroperiod t. We denote
by e
P
r
(resp e
S
r
) the purchasing (resp. selling) price
of the energy traded with the main grid during mi-
croperiod r. Since the voltage of electricity used for
long-distance transmission by the grid operator is not
the same as the one used in the manufacturing plant,
a grid transformer is needed to adapt the electricity to
production use. This causes energy losses mostly due
Integrated Production and Energy Supply Planning on an Industrial Site by Mixed-Integer Linear Programming
121
to the electrical current flowing in the coils and the
magnetic field alternating. We denote by η
G
the effi-
ciency of the grid transformer. The amount of green
electricity generated on-site by the local energy sys-
tem during micro-period r is denoted by p
R
r
.
Finally, the energy storage system consists of a
battery which can store the residual electricity gener-
ated by the on-site power conversion devices and the
energy bought from the main grid. Let c
B
be the max-
imum amount of energy that can be stored in this sys-
tem. The amount of energy which can be charged into
(resp. discharged from) the battery during an energy-
oriented microperiod is denoted by m
C
(resp. m
D
).
Battery charging/discharging incurs some energy loss
due to the internal resistance: let η
C
/η
D
be the charg-
ing/discharging efficiency of the battery.
The optimization problem under study here con-
sists in building an integrated energy supply and pro-
duction plan. This plan should simultaneously deter-
mine when and how many finished products to pro-
duce and when and how much energy to trade with
the grid and to charge into/discharge from the battery.
This plan should ensure that the demand for the fin-
ished products is satisfied at all time and that there
is always enough energy to supply the manufacturing
process while minimizing the total production and en-
ergy procurement costs.
3 PLSP-BASED MODEL
As will be shown by the numerical experiments to be
provided in Section 4, using the GLSP-based model
previously published by (Wichmann et al., 2019) to
build a production and energy supply plan is pos-
sible only for small instances. Significant compu-
tational difficulties are encountered when trying to
solve medium to large size instances. These difficul-
ties mainly come from the sharp increase in the size of
the MILP formulation with the number of items and
horizon length. In order to overcome this, we propose
a new PLSP-based model for the problem relying on
the use of small buckets to plan production.
3.1 Model Description
The main idea underlying the proposed PLSP-based
model consists in aligning the production-related time
discretization on the energy-related time discretiza-
tion and to define the production plan using the ex-
ogenously defined energy-oriented microperiods. The
use of a fixed time grid involving short planning pe-
riods to determine the production lot sizes leads to
the formulation of a small bucket’ lot-sizing prob-
lem, i.e. to the formulation of a lot-sizing problem
in which there are some restrictions on the number
of item types that may be produced in each planning
period. More precisely, we consider here the Propor-
tional Lot-sizing and Scheduling Problem (PLSP) in
which at most two different types of item may be pro-
duced in each energy-oriented microperiod. The use
of a fixed time grid (rather than a flexible one as done
in the GLSP) to plan production significantly sim-
plifies the mathematical formulation of the problem
as it eliminates the need to manage the intersections
between energy-oriented and production-oriented mi-
croperiods, which requires a large number of addi-
tional binary variables and big-M type constraints.
In order to formulate the problem as an MILP, we
introduce the following decision variables:
I
j,t
: amount of item j held in inventory at the end
of macroperiod t,
Y
j,r
: setup state of the resource at the end energy-
oriented micro-period r; Y
j,r
= 1 if the resource is
setup to produce item j at the end of r, Y
j,r
= 0
otherwise.
X
j,r
: startup indicator. X
j,r
= 1 if there is a startup
for item j during r, X
j,r
= 0 otherwise,
Q
j,r
: quantity of item j produced in energy-
oriented microperiod r.
P
C
r
/P
D
r
: amount of energy charged into / dis-
charged from the battery during energy-oriented
microperiod r,
P
B
r
: state of charge of the battery at the end of
energy-oriented microperiod r,
P
GP
r
/P
GS
r
: amount of electricity purchased from /
sold to the main grid during energy-oriented mi-
croperiod r.
3.2 MILP Formulation
Objective Function. The objective function aims at
minimizing the total startup, inventory holding and
energy costs over the planning horizon.
min
R
r=1
J
j=1
f
j
· X
j,r
+
T
t=1
J
j=1
h
j
· I
j,t
+
R
r=1
(e
P
r
· P
GP
r
e
S
r
· P
GS
r
).
(1)
Production-Oriented Constraints. We first focus
on the set of production-oriented constraints. These
constraints aim at building a feasible production plan.
I
j,t
= I
j,t1
+
rR
t
Q
j,r
d
j,t
, j J ,t T . (2)
ICORES 2023 - 12th International Conference on Operations Research and Enterprise Systems
122
k
j
Q
j,r
6 l(Y
j,r1
+Y
j,r
), j J , r R . (3)
J
j=1
k
j
Q
j,r
6 l, r R . (4)
J
j=1
Y
j,r
6 1, r R . (5)
X
j,r
> Y
j,r
Y
j,r1
, j J , r R . (6)
X
j,r
+Y
j,r1
6 1, j J ,r R . (7)
Y
j,0
= 0, j J . (8)
I
j,0
=
¯
I
j,0
, j J . (9)
I
j,T
> I
j,0
, j J . (10)
Constraints (2)-(6) define the production plan ac-
cording to the assumptions used in the PLSP. Con-
straints (2) are the inventory balance equations. In-
equalities (3) ensure that item j can be produced in
microperiod r only if the machine is set up for j at
the beginning or/and at the end of the period. Con-
straints (4) make sure that the total capacity consumed
by the production taking place during each micrope-
riod r stays below the capacity available in r. Equa-
tions (5) guarantee that the machine is set up for at
most one item at the end of each microperiod. To-
gether with Constraints (3), they ensure that at most
two items may be produced during a microperiod r.
The relationship between startup and setup decision
variables is described by (6): if the machine is setup
for item j at the end of r but was not setup for j at
the end of the r 1, a startup for j must occur during
r. (7) are basic valid inequalities that may be used to
strengthen the MILP formulation: see (Belvaux and
Wolsey, 2001). They indicate that if the machine was
already set up for j at the end of r 1, then there is
no need to carry out a start up for item j during r. In-
versely, if a startup for product j occurs in r, it means
the machine was not set up for this product in r 1.
Constraints (8)-(10) define the initial and final condi-
tions.
Energy-Oriented Constraints. We now focus on
the set of energy-oriented constraints aiming at build-
ing a feasible energy supply plan in each energy-
oriented micro-period.
P
U
r
+
P
GS
r
η
G
+
P
C
r
η
C
= η
G
·P
GP
r
+ p
R
r
+η
D
·P
D
r
, r R .
(11)
P
B
r
= P
B
r1
+ P
C
r
P
D
r
, r R . (12)
0 6 P
B
r
6 c
B
, r R . (13)
P
C
r
6 m
C
, r R . (14)
P
D
r
6 m
D
, r R . (15)
Equalities (11) ensure that the energy supply and
demand is balanced within each microperiod. The
energy demand consists in the energy consumed by
the manufacturing process P
U
r
, the energy sold to the
main grid P
GS
r
/η
G
and the energy charged into the
battery P
C
r
/η
C
. Note how, since part of the energy
is lost within the grid transformer, the amount of en-
ergy needed to sell P
GS
r
to the main grid is given
by P
GS
r
/η
G
. Similarly, the actual energy needed to
charge P
C
r
into the battery is computed by P
C
r
/η
C
due
to the losses during the battery charging process. The
energy supply consists in the energy purchased from
the grid η
G
· P
GP
r
, the energy generated from the re-
newable resources p
R
r
and the energy discharged from
the battery η
D
·P
D
r
. Note that, when we buy P
GP
r
from
the main grid, only η
G
· P
GP
r
energy is really available
for the plant. The same applies for the energy dis-
charging process. The energy balance in the battery
is defined by Equations (12). The amount of energy
stored in the battery at the end of microperiod r is
equal to the amount stored at the beginning of r plus
the amount charged into the battery during r minus the
amount discharged from s. Constraints (13) impose
that the amount of energy stored in the battery does
not exceed its capacity. Inequalities (14) and (15) are
the battery maximum charging and discharging rate
constraints.
Coupling Constraints. The final set of constraints
links together the decision variables related to produc-
tion and the one related to energy supply.
P
U
r
=
J
j=1
e
f
j
· X
j,r
+
J
j=1
e
j
· Q
j,r
, r R . (16)
Equations (16) guarantee that the energy con-
sumption by the factory equals to the sum of the en-
ergy consumed for the startups and for the production
on the machine.
Domain Definition Constraints. Finally, the defi-
nition domain of the decision variables are given by
Constraints (17)- (19).
X
j,r
,Y
j,r
{0, 1}, j J ,r R , (17)
Integrated Production and Energy Supply Planning on an Industrial Site by Mixed-Integer Linear Programming
123
I
j,t
,Q
j,r
> 0, j J ,t T ,r R (18)
P
C
r
,P
D
r
,P
GP
r
,P
GS
r
,P
B
r
,P
U
r
> 0, r R . (19)
4 COMPUTATIONAL
EXPERIMENTS
In this section, numerical experiments are conducted
to assess the performance of the proposed PLSP-
based model and compare it with the one of the
GLSP-based model proposed in (Wichmann et al.,
2019). The procedure used to randomly generate a set
of instances is described in Subsection 4.1. Prelimi-
nary computational results are given and discussed in
Subsection 4.2.
4.1 Instances
The parameter setting used to randomly generate in-
stances is mostly based on the numerical values pro-
vided in (Wichmann et al., 2019). As this paper did
not involve energy losses in the grid transformer nor
during the battery charging/discharging process, we
use the values provided in (Zhang et al., 2017) and
(Golp
ˆ
ıra et al., 2018) for these parameters.
The production system comprises a single ma-
chine producing a set of products on a finite plan-
ning horizon spanning several days. Each day is di-
vided into two macroperiods, each one with a length
of l =480 minutes: these two macroperiods corre-
spond respectively to a morning and evening eight-
hour shift. We consider three set of instances of vari-
ous sizes. Small-size instances involve J = 3 products
and a two-day planning horizon of T = 4 macroperi-
ods. Medium-size instances involve J = 5 products
and eight days, i.e., T = 16 macroperiods. Finally,
large-size instances involve J = 10 products and a 16-
days planning of T = 32 macroperiods.
For any item j = 1, ...,J, a setup of the machine
costs f
j
= 200e and consumes e
f
j
= 10kWh. More-
over, producing one unit of this item takes k
j
= 0.05
minutes and consumes e
j
= 0.1kWh. The production
capacity of the machine within one macroperiod can
be computed as Cap = l/k
j
= 480/0.05 = 9600. The
demand for each item j = 1,...J is randomly gen-
erated as follows. We consider a utilization rate of
the resource of ρ = 0.8 and set the mean value of
the total demand to be satisfied in each macroperiod
t = 1, ,..., T to ρCap = 0.8 9600 items. For each
item j = 1, ...,J, the demand d
j,t
is generated accord-
ing to a Normal distribution of mean
ρCap
J
and stan-
dard deviation
ρCap
3J
. In case the randomly generated
value is negative, we replace it by 0. We thus have
d
j,t
max(int(N (
ρCap
J
,
ρCap
3J
)),0). The initial inven-
tory of item j is randomly generated following the
uniform distribution using
¯
I
j,0
randint(0, 2 d
j,1
).
The unit inventory holding cost of item j is set to
h
j
= 0.05e per macroperiod. In the PLSP model,
the production-oriented microperiods are identical to
the energy-oriented ones. In the GLSP model, each
macroperiod t may be flexibly divided into a prede-
fined number |S
t
| of production-oriented microperi-
ods of variable length: we set |S
t
| = 7 for each t.
Regarding the energy supply and consumption,
each macroperiod t is split evenly into |R
t
| = 8
energy-oriented microperiods of l
e
=60min length.
The energy price is assumed to display intra-day vari-
ations but to be otherwise daily periodic. The unit
energy price for each hour (indexed from 1 to 16)
of a given day in our reference scenario is displayed
in Table 1. Moreover, we consider in our experi-
ments two additional scenarios: a low (resp. ex-
treme low) energy price scenario in which the ref-
erence unit price in every period is divided by 10
(resp. by 100). The amount of on-site generated
electricity depends on the weather and thus varies
both within the day and from one day to the next.
A microperiod indexed by r corresponds to the hour
indexed r
d
= r(mod16) of the day. The value of
p
R
r
for microperiod r is randomly generated using
a Normal distribution with an expected value and
a standard deviation equal to Gen[r
d
]: see Table 1.
Note that the values of Gen[r
d
] are consistent with a
power generation by PV panels. We thus have p
R
r
max(N (Gen[r(mod16)],Gen[r(mod16)]),0). The
energy storage system has a capacity of c
B
=
500kWh. Within each microperiod r, the maxi-
mum amount of energy that can be charged into or
discharged from the battery is set to m
C
= m
D
=
250kWh. The charging and discharging efficiencies,
as well as the grid transformer efficiency, are set to
η
C
= η
D
= η
G
= 0.95.
4.2 Preliminary Numerical Results
For each considered instance size and energy price
level, we randomly generated 10 instances, resulting
in a total number of 90 instances. Each instance was
solved with the MILP solver CPLEX 20.10 using ei-
ther the GLSP-based model described in (Wichmann
et al., 2019) or the PLSP-based model described in
Section 3. The implementation was done in Python.
The computational experiments were carried out on a
laptop running under Windows 10 with an Intel(R)
Core(TM) i5-6300HQ CPU @ 2.30GHz processor
and 16 GB RAM. The time limit was set to 20 min-
ICORES 2023 - 12th International Conference on Operations Research and Enterprise Systems
124
Table 1: Energy prices and generated electricity in one day.
Time of a day 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Energy price 4.8 6.1 6.3 6.0 5.6 4.0 3.7 3.8 4.5 5.1 5.4 5.9 6.4 6.3 5.5 4.5
Gen[r
d
] 1 4 10 18 25 27 30 30 25 15 5 2 0 0 0 0
Table 2: Optimization results for small-size instances with J = 3, T = 4.
Energy price level initial low extremely low
Model PLSP GLSP PLSP GLSP PLSP GLSP
# VAR 563 7779 563 7779 563 7779
# BINVAR 259 3875 259 3875 259 3875
# CONS 497.9 9629.9 500.1 9627.9 498.7 9627.7
Z
best
15567.60 15567.60 3515.14 3515.10 2732.36 2732.25
Gap
LP
4.37% 4.86% 13.74% 14.04% 18.91% 19.01%
Gap
MIP
0.01% 0.18% 0.01% 0.31% 0.00% 0.02%
Computation time 1.53 1029.92 1.23 1179.20 1.41 924.67
Table 3: Optimization results for medium-size instances with J = 5, T = 16.
Energy price level initial low extremely low
Model PLSP GLSP PLSP GLSP PLSP GLSP
# VAR 3035 175163 3035 175163 3035 175163
# BINVAR 1541 87621 1541 87621 1541 87621
# CONS 2890.4 204144 2913.4 204144 2915 204144
Z
best
59912.29
No solution
14209.03
No solution
9999.33
No solution
Gap
LP
3.84% - 10.63% - 14.38% -
Gap
MIP
1.41% - 2.47% - 2.62% -
Computation time 1200.68 - 1152.15 - 1097.88 -
Table 4: Optimization results for large-size instances with J = 10, T = 32.
Energy price level initial low extremely low
Model PLSP GLSP PLSP GLSP PLSP GLSP
# VAR 10069 1271349 10069 1271349 10069 1271349
# BINVAR 5642 636234 5642 636234 5642 636234
# CONS 12022.5 1387775 12088.4 1387775 14428.2 1387775
Z
best
130087.44
No solution
41935.94
No solution
32371.44
No solution
Gap
LP
5.71% - 13.75% - 16.29% -
Gap
MIP
4.74% - 10.98% - 12.39% -
Computation time 1201.63 - 1201.71 - 1201.78 -
utes.
The results for small, medium, and large-size in-
stances are displayed in Tables 2, 3, and 4, respec-
tively. Each column provides the average results over
the 10 instances corresponding to the given energy
price level. We provide, for each set of 10 instances:
# VAR/ # BINVAR, the number of variables and
binary variables involved in the formulation,
# CONS, the number of constraints in the formu-
lation,
Z
best
, the value of the best feasible integer solution
found by the solver in 20min,
Gap
LP
, the integrality gap, i.e., the relative differ-
ence between the lower bound provided by the lin-
ear relaxation of the problem and the value of the
best integer feasible solution found by the solver,
Gap
MIP
, the optimality gap, i.e., the gap be-
tween the best lower bound and the best upper
bound found by the solver before the time limit is
reached (note that this gap is equal to 0% in case
Integrated Production and Energy Supply Planning on an Industrial Site by Mixed-Integer Linear Programming
125
a guaranteed optimal solution could be found),
the computation time (in seconds).
We first note that the computational effort needed
to solve the problem with a mathematical program-
ming solver is drastically reduced when using the
PLSP-based model instead of the GLSP-based model.
Thus, for the small instances, the proposed PLSP-
based model can be solved to optimality within less
than 2s, while the GLSP-based model cannot con-
verge to the optimal solution after 20 minutes of
computation: see Table 2. Moreover, for the set
of medium and large instances, no feasible solution
can be found by the solver in 20min when using
the GLSP-based model. In contrast, the PLSP-based
model is able to provide feasible solutions of accept-
able quality. The integrality gaps seem to be simi-
lar for both formulations (see in particular Table 2).
Thus, the observed improvement in the computational
efficiency may be mainly explained by the significant
decrease in the size of the MILP formulation obtained
when using the PLSP-based model.
In terms of solution quality, the GLSP-based for-
mulation is more flexible regarding the production
plan than the PLSP-based formulation. Namely, the
GLSP-based model allows the production of more
than 2 different types of items in a given energy-
oriented microperiod, which is forbidden in the
PLSP-based model. Thus, theoretically, the GLSP-
based model may provide less expensive production
and energy supply plans. However, in practice, as
can be seen in the results reported in Table 2, the
value of the optimal solution found by the PLSP-
based model is equal or very close to the one found
by the GLSP-based model. Thus, the GLSP-based
model is able to provide a lower cost production plan
than the PLSP-based model in only four instances
out of the 30 corresponding instances. Moreover, the
maximum observed relative cost difference over these
four instances is only 0.05%. This indicates that in
most cases, the optimal solution obtained by the PLSP
model is of the same cost as the one obtained by the
GLSP model.
5 CONCLUSION AND
PERSPECTIVES
This paper investigated an integrated lot-sizing and
energy supply planning problem in a single-machine
multi-product setting. A new MILP model based on
an extension of the PLSP was proposed to handle
this problem. Our numerical results showed that the
proposed model outperforms a previously published
GLSP-based model with respect to the computation
times. However, the availability of the renewable
energy resources heavily depends on weather condi-
tions, so that the corresponding power generation can-
not be predicted precisely. Therefore, future work
should focus on explicitly taking into account these
uncertainties in the problem modeling.
REFERENCES
Belvaux, G. and Wolsey, L. A. (2001). Modelling practical
lot-sizing problems as mixed-integer programs. Man-
agement Science, 47(7):993–1007.
Biel, K., Zhao, F., Sutherland, J. W., and Glock, C. H.
(2018). Flow shop scheduling with grid-integrated on-
site wind power using stochastic MILP. International
Journal of Production Research, 56(5):2076–2098.
Copil, K., W
¨
orbelauer, M., Meyr, H., and Tempelmeier, H.
(2017). Simultaneous lotsizing and scheduling prob-
lems: a classification and review of models. OR Spec-
trum, 39(1):1–64.
Drexl, A. and Haase, K. (1995). Proportional lotsizing and
scheduling. International Journal of Production Eco-
nomics, 40(1):73–87.
Fleischmann, B. and Meyr, H. (1997). The general lotsizing
and scheduling problem. OR Spectrum, 19(1):11–21.
Gao, K., Huang, Y., Sadollah, A., and Wang, L. (2020).
A review of energy-efficient scheduling in intelligent
production systems. Complex & Intelligent Systems,
6(2):237–249.
Golp
ˆ
ıra, H., Khan, S. A. R., and Zhang, Y. (2018). Ro-
bust smart energy efficient production planning for a
general job-shop manufacturing system under com-
bined demand and supply uncertainty in the presence
of grid-connected microgrid. Journal of Cleaner Pro-
duction, 202:649–665.
Johannes, C., Wichmann, M. G., and Spengler, T. S.
(2019). Energy-oriented production planning with
time-dependent energy prices. Procedia CIRP,
80:245–250.
Rodoplu, M., Arbaoui, T., and Yalaoui, A. (2019). Single
item lot sizing problem under renewable energy un-
certainty. IFAC-PapersOnLine, 52(13):18–23.
Terbrack, H., Claus, T., and Herrmann, F. (2021). Energy-
oriented production planning in industry: a systematic
literature review and classification scheme. Sustain-
ability, 13(23):13317.
Wang, S., Mason, S. J., and Gangammanavar, H. (2020).
Stochastic optimization for flow-shop scheduling with
on-site renewable energy generation using a case in
the United States. Computers & Industrial Engineer-
ing, 149:106812.
Wichmann, M. G., Johannes, C., and Spengler, T. S. (2019).
Energy-oriented lot-sizing and scheduling considering
energy storages. International Journal of Production
Economics, 216:204–214.
Zhang, H., Cai, J., Fang, K., Zhao, F., and Sutherland, J. W.
(2017). Operational optimization of a grid-connected
factory with onsite photovoltaic and battery storage
systems. Applied Energy, 205:1538–1547.
ICORES 2023 - 12th International Conference on Operations Research and Enterprise Systems
126