Accepted Manuscript

Title: Optimization of multipurpose process plant operations:

a multi-time-scale maintenance and production scheduling
approach

Authors: Matteo Biondi, Guido Sand, Iiro Harjunkoski

PII: S0098-1354(17)30008-X
DOI: http://dx.doi.org/doi:10.1016/j.compchemeng.2017.01.007
Reference: CACE 5663

To appear in: Computers and Chemical Engineering

Received date: 4-7-2016

Revised date: 8-11-2016
Accepted date: 4-1-2017

Please cite this article as: Biondi, Matteo., Sand, Guido., & Harjunkoski,
Iiro., Optimization of multipurpose process plant operations: a multi-time-
scale maintenance and production scheduling approach.Computers and Chemical
Engineering http://dx.doi.org/10.1016/j.compchemeng.2017.01.007

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
Optimization of multipurpose process plant operations: a multi-
time-scale maintenance and production scheduling approach
Matteo Biondi,a,b Guido Sand,a,c Iiro Harjunkoskia,*
a
ABB Corporate Research Germany, Wallstadter Str. 59, 68526 Ladenburg, Germany
b
Chair of Production Management, University of Mannheim, 68131 Mannheim, Germany
c
University of Applied Science Pforzheim, Tiefenbronner Str. 65, 75175 Pforzheim, Germany
*[email protected]

Highlights:

A novel optimization based approach for the integrated maintenance and production
scheduling of process plants is presented.
The key for the integration are the concepts of residual useful life and operation modes of
plant units which are motivated by a well-known example from the process industry.
The degradation of unit performance due to aging effects is introduced and the impact on the
scheduling solution modeled and tested on representative examples.

Abstract
Scheduling of production and maintenance plays a fundamental role in the effective operation
of process plants. Frequently the two decision processes are independently addressed,
overlooking the tight relation existing between the way the plant is operated to produce the
required goods and the appearance of maintenance requirements. The presence of degradation
phenomena affecting the performance of plant units and limiting the operational choices makes
the integration of the two decision processes even more important. In this paper an industrial
framework for the integration of maintenance and production scheduling of process plants is
presented as well as some considerations on how the presence of plant unit degradation impacts
on the scheduling problem. The proposed approach ties industrial key-components such as asset
management and production scheduling closer together and represents therefore a contribution
to the smart manufacturing revolution. The integrated maintenance and production scheduling
problem is formulated as a mixed integer linear program (MILP) and tested on a generic process
plant described as a State Task Network (STN).

Keywords: plant operations, maintenance, scheduling, optimization, performance degradation

1. Introduction
The scheduling of production and maintenance is a critical decision process for the profitable
management of any process plant. While the first ensures reaching the production targets to
satisfy customer demands, the second makes sure that plant assets are available and in the
condition to perform the required tasks when needed by the production. The two decision
processes are highly interdependent since they share a clear common denominator: the assets
of the plant, which are used/consumed by production tasks and restored by maintenance
activities.

Nomenclature
Indices
i task
j unit
k operation mode
t time interval
ts scheduling time interval
tp planning time interval
n natural number
Sets
I production tasks
Ij tasks that can be performed on unit j
Is tasks fed by state s
Īs tasks producing output material in state s
J plant units
Ji units that can perform task i
Kj operation modes of unit j
S initial, intermediate and final states
T time intervals
Ts scheduling time intervals
Tp planning time intervals
Parameters
pijk production time of task i on unit j operating in mode k
τj duration of a maintenance task on unit j
Vijmin lower bound on batch size of task i on unit j
Vijmax upper bound on batch size of task i on unit j
proportion of input of task i from state s
proportion of output of task i to state s
dst demand of product in state s at the end of time t
Cs maximum storage capacity for product in state s
Rjmax maximum life of unit j
Dijk wear on unit j caused by the processing of task i in mode k
H length of each planning horizon time interval
csstorage storage cost for one unit of material in state s
mk threshold on the percentage of remaining useful life of a unit to use mode k
Binary variables
Wijkt =1 if unit j starts processing task i in mode k at the beginning of time interval ∈
Mjt =1 if a maintenance task start on unit j at the beginning of time interval ∈
Modejkt =1 if unit j operates in mode k within planning time interval ∈
Qijktn binary for the 2-based representation of integers ∈

Integer variables
Nijkt number of copies of task i produced by unit j in mode k within the time interval ∈

Continuous variables
Bijkt amount of material which starts undergoing task i on unit j in mode k at time ∈
Sst amount of material stored in state s, at the beginning of time period ∈
Ssfin inventory level of state s at the end of the scheduling horizon
Rjt residual life of unit j at the end of time interval ∈
Fjt residual life restoration on unit j due to a maintenance task starting at time ∈
Aijt total amount of material undergoing task i in unit j within time interval ∈
Oijktn auxiliary variable for linearization of bilinear terms ∈
Asset management is the engineering practice that deals with the optimization of plant assets
usage with the ultimate objective of reducing production costs. In real industrial plants, it is
responsibility of the asset management system (NAMUR, 2009) to:
 manage the assets over the whole life cycle, in particular regarding their reliability and
efficiency
 optimize utilization and cost-effective maintenance of the assets
 generate and provide information regarding the development and prognosis of the “asset
health” to support decision making of the enterprise and production management.

Fig. 1 Screenshot from the Ventyx (ABB) Asset Health Center system

Figure 1 shows a screenshot of the summary page of an asset management system from a real
industrial plant. Assets are here represented grouped by classes to have a better overview of the
overall system health. As mandated by the NAMUR recommendation, other critical
information to effectively manage the overall plant are also summarized, for example the
indication of the assets health (in terms of risk of failure) with a trend over the last time periods
to support the prognosis of the remaining asset life. The information contained in the asset
management system is of vital importance to support the maintenance team to effectively plan
repair, cleaning and maintenance actions.
One of the most striking aspects appearing from the asset management system described above
is the lack of connection with the actual process conditions, i.e. with information on the kind of
products that are being processed or will be processed in the plant in the near future, which
ultimately determines the asset wear-out rates. This kind of information is managed by another
key system for the plant operations, the planning and scheduling, which vice versa has typically
no interconnection with the asset management system.
Using asset health information to generate an optimal production plan is a viable solution to
better integrate the two systems and to increase the overall performance (e.g. in terms of costs)
of plant operations.

Fig. 2 Workflow of the integrated maintenance and production scheduling process

The objective of this paper is twofold: first, to propose a joint scheduling approach for the
production and maintenance of process plants that explicitly keeps track of the assets life cycle.
Figure 2 shows a workflow of the integrated decision process. The scheduling system includes
a simple model of the asset wear that can be based on the concept of residual useful life (RUL)
or of probability of failure ( ). In the following, we will focus on a deterministic
formulation of the integrated maintenance and production scheduling problem and therefore
consider the residual useful life as a unique indicator of the asset health. The asset monitoring
system is responsible of providing two types of information to the scheduling system: on the
one hand, an estimation of the parameters describing the wear caused by the production on the
asset. On the other hand, if an extraordinary condition of the asset is detected, it is responsible
of updating the current RUL in the asset wear model of the scheduling system. Assets health
information, along with the production orders, is managed by the scheduling system that takes
care of the sequencing and timing of production tasks on the plant and triggers a maintenance
action on the assets whenever this is required. Furthermore, the system can select an optimal
production or operation mode to influence the equipment wear in an optimal manner. The
typically slow dynamics of the asset wear-out process (and therefore the low frequency of
maintenance requirements) poses significant challenges to the integrated maintenance and
production approach. For this reason, in Section 4 a multi-time scale framework is introduced
to account for the short term decisions (e.g. maintenance and production scheduling) as well as
for the medium-long term ones (e.g. maintenance and production planning).
The second objective of the paper is to investigate the effects of asset aging on the integrated
maintenance and production model presented before. If the aging has influence on the asset
performance, the scheduling model should take that into account to generate the optimal
schedule and eventually anticipate a maintenance action on an asset, if this improves its
performance. In Section 0, two ways to account for performance degradation due to asset aging
are introduced and the influence on the integrated maintenance and production approach is
modeled and tested.
With these characteristics included in the models, the presented approaches show not only
elements of novelty from the scientific point of view, but also an important step forward from
an industrial perspective. As a matter of fact, the methods make an effective use of units’ health
information to generate a feasible plan for production and maintenance. They represent
therefore a significant interconnection between the domains of asset management and
production planning, two critical components for the management of operations of process
plant that are still frequently considered independently in practical applications.
Before entering into the modeling details, Section 2 summarizes the integrated maintenance
and production approaches found in literature that inspired this work. In Section 3, the modeling
framework for the scheduling approach as well as the key elements to realize the integration
between maintenance and production planning are introduced and described.

2. Maintenance and production scheduling: a literature review

Different problems involving the joint scheduling of maintenance and production tasks have
been already described in literature. In some cases, the set of maintenance operations to be
performed during the scheduling horizon is known in advance. Dedopoulos et al. (1995) address
this problem for the scheduling of a multipurpose batch plant. Maintenance tasks are provided
with earliest and latest start times as well as their expected durations and a penalty is paid in
case the maintenance is not performed. Aggoune (2004) considers a flow shop setting with
availability constraints. A set of maintenance activities is provided for each machine of the
plant with two different schemes: fixed maintenance starting time and maintenance starting
time variable within a time window. Naderi et al. (2011) address the problem on a flexible flow-
shop, even though the problem is not solved as a real joint schedule problem: the production
schedule is obtained first and the maintenance operations are inserted into the schedule when
needed. Cheung at al. (2004) address the short-term scheduling of maintenance of a chemical
production site including several types of production and utility plants. The sections of the site
requiring maintenance within each scheduling horizon are supposed to be known in advance as
well as the utilities and material demand profile required during their shutdown, overhaul and
startup.
Another category of approaches addresses the joint scheduling problem with maintenance
requirements depending on the production. Bock et al. (2012) describe a single machine
scheduling problem with job-dependent machine degradation. The life of the machine called
maintenance level can be restored by means of a maintenance activity, whose duration is
proportional to the amount of life restored. Hazaras et al. (2012) extend the classical global
event based scheduling model of Maravelias et al. (2003) to account for fixed and flexible
maintenance tasks. Beyond maintenance requirements provided as input to the problem (with
fixed or interval timing constraints), recurrent maintenance tasks based on the cumulated
processing time are described (i.e. after a given number of processing hours the plant is stopped
for maintenance). Castro et al. (2014) describe a maintenance scheduling problem for a gas
engine power plant in which each unit has to be maintained after a given amount of time spent
online. The presence of a time dependent power demand for the plant poses stringent
requirements on the number of active engines and on maintenance timing. Each plant unit can
be turned into an idle mode (i.e. neither in production mode nor offline for maintenance),
providing the plant with the necessary flexibility and allowing to postpone the maintenance
activity to a more convenient point in time. Nie et al. (2014) recently describe an RTN-based
scheduling problem with a resource, called unit capacity (CAP), that is consumed depending
on the amount of product processed. Maintenance activities are planned to replenish the
capacity along with the production of a mixed batch/continuous process.
A third category of approaches addresses explicitly the effects that the life of processing units
has on their performance, as well as, the impact that the operation of units has on the appearance
of maintenance requirements. Jain et al. (1998) address the problem of cyclic scheduling of a
continuous parallel-process. The performance of processing units is assumed to decay
exponentially based on the time since the last cleaning activity. Alle et al. (2004) extended the
classical Economic Lot Sizing Problem to account for performance decay: the performance of
the (single) unit of the plant, measured in term of process yield for a single product, decreases
linearly with time. Liu et al. (2014) address the integrated maintenance and production planning
problem for a biopharmaceutical industry with the assumption that the yield of one of the
production stages decay based on the number of batches produces. Varnier et al. (2012) address
the problem of scheduling maintenance and production on a flow-shop. Each machine of the
plant is assumed to be capable of working in a normal or degraded mode impacting on the
residual life of the machine and on the processing time of a production job. The switch between
the two modes is only possible at the beginning of the scheduling period or after a maintenance
occurrence, thus limiting the effective flexibility of operations.
The approaches proposed in this paper differ from the one summarized above under several
aspects. With respect to the first category of work, the most important difference is in the
modeling of maintenance requirement that are not fixed (i.e. not provided as input to the
scheduling problem) but dependent on the production. From this point of view, this work is
more along the research line of the second and third categories, but with a significant difference,
being a very flexible definition of operation mode of plant units (see following section for
details) that allow a tighter integration of the production and maintenance scheduling processes.

Fig. 3 STN representation of a chemical process (Kondili et al., 1993)

3. Integrated maintenance and production

The current section is structured as follows: first, the key elements of operation mode and
residual useful life to realize the integration of the two decision processes are described; then
the modeling framework used for the formulation of the integrated problem is introduced.
Finally, two examples from the steel making industry motivating the modeling elements
described above are shown.
3.1. Elements of the model
In order to effectively close the loop between the production and maintenance scheduling
decision processes, we introduce a measure of the need for asset maintenance called Residual
Useful Life (RUL) as described by Bock et al. (2012) for the scheduling of production and
max
maintenance on a single machine. Each asset j is assumed to be capable of working R j

equivalent time units before requiring a maintenance. The processing of a task on a unit wears
it by a certain predefined amount, thereby reducing the unit residual life. When the residual life
of a unit reaches a critical threshold, it can be restored to its maximum level by a maintenance
activity involving fixed costs and downtime.

Fig. 4 Asset health represented as an energy stored in a battery

Fig. 5 Impact of the operation mode k on the equipment wear and on the processing time
This can be visualized by associating to each asset a limited processing energy (represented by
a battery in Fig. 4). In order to be able to perform a task, an asset requires a certain amount of
energy. If the energy-level is not sufficient, the asset’s battery can be “recharged” by means of
a maintenance. An asset property with similar characteristics, called unit capacity (CAP), has
been described by Nie et al. (2014) and Velez et al. (2015), where the reduction of unit life
depends only on the amount of production processed by the unit. In our work, we introduce
more complex dependencies and let the amount of wear caused by the processing of a task on
a unit depend on the type of task processed and on the operation mode of the unit, i.e on the
way the unit processes a task. The impact of operation mode on the process and production is
two-fold (Fig. 5). On the one hand, the wear caused by a task on a unit can be modulated and
therefore the time before the next maintenance prolonged. On the other hand, the operation
mode is assumed to influence the productivity, e.g. the throughput for a continuous process or
the processing time for a batch production. From an industrial perspective, the estimation of the
dependencies between the operation modes and the production time/wear of batches on units
represents a very interesting challenge; this can be addressed by means of big data methods on
the production data to update, eventually online, the critical parameters for the problem.

Fig. 6 Example layout of a common stainless steel production plant

Fig. 7 Argon oxygen decarburization (AOD) unit for the production of stainless steel
The operation mode and the units’ residual life model described above are the key elements to
strengthen the integration of maintenance and production scheduling. This allows an optimal
balancing between production and maintenance.
3.2. Modeling framework
The State Task Network (STN) is a very effective representation of multipurpose process plants
introduced in Kondili et al. (1993) and will be used for the development of the integrated
maintenance and production scheduling framework. The process is represented by a directed
graph with two types of nodes: namely, state nodes corresponding to feed material, intermediate
and final products and task nodes denoting the operations that transform one state into another.
Arcs of the graph between states and tasks (or vice versa) represent the flow of material and
describe the recipe to obtain the final products from the material feeds (Fig. 3). Several MILP
formulations for the scheduling of batch production on STN have been proposed in literature;
while the process specifications do not substantially change between the approaches, it’s the
time representation used in the MILP problem that significantly differentiates one approach
from another. Kondili et al. (1993) and Shah et al. (1993) propose two MILP formulations based
on a predefined discretization of the scheduling horizon, Maravelias and Grossmann (2003)
allow the time points defining the time grid to be also a variable of the problem to give more
flexibility to the solution, Janak et al. (2004) target an even higher flexibility by allowing a
variable time grid for each processing unit (see Méndez et al. (2006) for an extensive review).
The scheduling models presented in this work are based on a STN process representation and
a uniform time discretization as described by Shah et al. (1993) which allows handling the
material balance constraints in a very natural way. Having modeled the residual useful life as a
material/energy required by the asset to process any production task motivated the choice of
the discrete time representation.

3.3. Example from the steel industry

The formalization of operation modes and of non-constant asset degradation as described above
has been inspired by the observation of real production plants in the process industry. As a
representative example, let us consider the steel making process outlined in Fig. 6 (the
production data and plant layout have been slightly modified from the original to preserve
industrial secrecy). The steel making process can be decomposed into four consecutive steps:
the electric arc furnace (EAF), where steel scraps are molten down; the argon oxygen
decarburization (AOD), where the steel carbon content is reduced by injection of oxygen; the
ladle furnace (LF), where the exact chemical composition is adjusted and the right temperature
for the next step achieved; the continuous caster (CC), where the molten steel is casted to the
final requested form (e.g. slabs, billets). Here we focus on the EAF and AOD stages (Fig. 7).
The steel making process in the EAF starts with a small amount of steel scraps charged in the
furnace ladle; furnace electrodes are then lowered and the melting process starts. After a certain
amount of time, the electricity flow is temporarily stopped to add another load of metal scrap
to the bath. The process is repeated until the target weight (around 100-120 tons) and
temperature (1300°C-1400°C) of the molten steel are met. In order to melt down the metal and
to achieve the required final temperature, an amount of roughly 80MWh of energy has to be
provided. The energy is provided over time according to different possible power profiles,
called melting curves (Fig. 8) which can be selected by the operator according to the plant
requirements. The selection of the melting curve has a two-fold impact of the unit: on the one
hand, the time required to fully melt down the metal scrap and to reach the final target
temperature can be reduced (see tA, tB, tC in Fig. 8). On the other hand, the higher power peaks
wear the furnace electrodes faster reducing their lifetime and forces the furnace to undertake
more frequent maintenance. The behavior of the EAF unit in a scheduling context can therefore
be very well described with the concept of operation mode described above in this section.

Fig. 8 Melting curves for the EAF stage of a stainless steel production

The AOD process poses another challenging requirement to the scheduling solution. The
typical lifecycle of the unit is around 50-80 heats (i.e. batches of molten steel); when it reaches
the end of its life, the unit must be taken offline and the refractory material substituted. Having
a spare AOD unit that can be brought online while the main one undertakes maintenance can
reduce the impact on the production process; nonetheless, the substitution of the AOD unit can
last several hours and cause significant disruption to the production process. The impact of
products on the AOD lining depends significantly on the steel type (e.g. austenitic, ferritic…)
and on the steel grade (e.g. 304, 316, 409, 440…).For very special steel grades there are some
“security constraints” that only allow a narrow positioning of the heat in the AOD production
sequence, e.g. between heat positions 40-60 in the life cycle of the unit after maintenance. This
allocation is just a rule of thumb, while the one used in practice is based on dynamically
monitored condition data from the unit and statistical look-up tables to predict the future health
development. In the framework we introduced here, this production constraint can be translated
to the following scheduling requirement: heats belonging to the special steel grades mentioned
above require an operation mode that can only be executed during certain phases of the
equipment lifetime. The problem of available operation modes depending on the asset health
will be addressed in section 5.2 by extending the base model that will be introduced in the next
section.

4. Multi-time scale model

The integrated production and maintenance scheduling problem considered in this section can
be formulated as follows. Given a production process described by an STN representation with
production units capable of working in different operation modes as described in Section 3, the
residual life of each process unit, maintenance costs and reparation times for each unit, a
scheduling horizon and a demand profile of final products over the scheduling horizon,
determine the optimal schedule of production and maintenance activities in order to reach the
production targets and minimize total operation costs (e.g. costs for products storage,
maintenance costs…).
To address the problem discussed in the introduction of different time dynamics between the
production and maintenance decision processes, a multi-time scale approach is presented in the
following. Multi-time scale methods have demonstrated their capability to solve medium/long
term scheduling problems of complex production systems since their introduction in
Papageorgiou et al. (1996), where the scheduling of a generic multipurpose plant working in
campaign mode is addressed. The key idea of these methods is to reduce the complexity of the
problem by considering a decomposition of the scheduling problem into two parts; the first
taking care of high-level decisions (e.g. the number of campaigns, the type and number of task
involved in each campaign) and the second taking care of the finer ones (e.g. detailed operation
schedules within each campaign). In the following model we take a similar approach but solve
on a detailed level only the first part of the scheduling horizon (short-term), while considering
only high-level decision for the rest (long-term). A non-uniform time grid is used for the
scheduling/planning horizon and the problem is formulated as a single MILP problem,
addressing the short term and medium/long term decisions with different granularities.
For the short term, a detailed scheduling model based on Biondi et al. (2015) will be introduced
(see sections 4.1.1 to 4.1.4). While most of the equations remain unchanged, a different
formulation of the residual life update is introduced. The model is based on a STN process
representation and on the MILP model presented by Shah et al. (1993) that uses a discrete time
grid representation.
For the medium/long term, an aggregated STN representation inspired by the master problem
of Maravelias and Grossmann (2004) is introduced. The basic idea is not to focus on each
individual batch, but to consider all the batches of the same kind (i.e. same task performed on
the same unit) produced within each planning interval in an aggregated way. For each planning
time interval only the following aggregated constraints will be considered: assignment of tasks
and maintenance to units, i.e. no sequencing or timing constraints involved; aggregated batch
size constraints, i.e. written only once for all batches of the same kind; material balance and
residual life update constraints for the whole production involved. Within each planning time
interval, no further time point or sub-interval is considered; therefore the last three types of
constraints (batch size, material balance and life update) are imposed only once for the
aggregate production at the end of each planning interval.
For each planning period, it will also be assumed that each unit can work in only one production
mode, independently from the type of task processed. Without any loss of generality, it will
also be assumed that the processing of a task on a unit is not allowed to span over two
consecutive planning time intervals, i.e. each task is completed in the same planning time period
where it started. The effects of this assumption are clearly mitigated by the rolling horizon
nature of the proposed approach as soon as jobs enter the short term scheduling part of the
model. All the standard constraints of the STN representation are then imposed for the
aggregated production at the borders of the medium/long time intervals (see section 4.1.5 to
4.1.9). Special constraints are finally written at the interface between the scheduling and
planning horizon to guarantee the feasibility of the solution (4.1.10).

4.1. Mathematical model

Let Ts be the set of time intervals defining the short term scheduling horizon, Tp the
medium/long term planning horizon and let T  Ts  T p (see Fig. 9). Following the classical

STN formulation, let I be the set of tasks, J set of units, Ij set of tasks that can be performed on
unit j, Ji set of units that can perform task i, S set of initial, intermediate and final states, Is set
of tasks fed by state s and I s set of tasks producing output material in state s. Let furthermore

Kj be the set of operation modes in which unit j can operate. In the following, the key variables
and model constraints (for the scheduling, planning and for the interface between the two) as
well as the objective function are introduced and discussed.

Fig. 9 Non-uniform time discretization of the scheduling horizon

4.1.1. Allocation of tasks to units (scheduling)

In order to allocate production tasks to processing unit and to set the processing start time, a
binary variable Wijkt is introduced. The variable takes value 1 if the processing of task i on unit
j operating in mode k starts at time t. At any time t, a unit cannot start the processing of a task
if another production task is active or if the unit is undergoing a maintenance activity (and vice
versa). By introducing another binary variable Mjt (taking value 1 if a maintenance activity on
unit j starts at time t), the previous constraints can be expressed by:
t t

  W
kK j iI j t 't  pijk 1
ijkt '   M
t 't  j 1
jt '  1 j , t  TS (1)

where pijk is the processing time of task i on unit j in mode k and  j is the duration of the

maintenance task on unit j.

4.1.2. Batch capacity limitation (scheduling)
Let Bijkt be the amount of material (i.e. the batch dimension) starting to be processed in task i
on unit j in mode k at time t. This amount has to be bounded by the minimum and maximum
capacity of the unit itself. Furthermore, if no batch starts to be processed ( W ijkt  0 ) the amount

of material should also be zero. If Vijmin and Vijmax are the minimum and maximum batch

dimension of unit j performing task i, then:

VijminWijkt  Bijkt  VijmaxWijkt j , i  I j , k  K j , t  TS (2)

4.1.3. Material balance (scheduling)

Let I s be the set of tasks receiving material from state s and I s the set of tasks producing

material in state s. For each state s and task i  I s , let is be the proportion of input of task i

from state s and  is the proportion of output of task i to state s. Let Sst be the amount of material

stored in state s, at the beginning of time period t. Then at any time t, the following set of
material balance constraints must be verified for each state s:

S st  S s ,t 1   is  B ijk ,t  pijk    is  B ijkt  d s ,t 1 s , t  TS (3)

iI s jJ i kK j iI s jJ i kK j

where d s,t-1 is the demand of product in state s at time t-1. The constraints impose that the net
increase of product in state s ( S st  S s ,t 1 ) is equal to the difference between the amount

produced in that state and the amount used to start new tasks or to fulfill external demand.
The amount of material stored in state s at any time t cannot exceed the storage capacity Cs
dedicated to state s, therefore:

0  S st  Cs s, t  TS (4)

4.1.4. Unit residual life (scheduling)

Let Rjt be the continuous variable representing the residual useful life of unit j at time t, R max
j

the maximum life of unit j and Dijk the parameter describing the amount of life of unit j
consumed by task i processed with operation mode k. The residual life of unit j at time t should
be decreased by Dijk if task i starts to be processed at time t and restored to its maximum if a
maintenance activity is started. Then the following equations hold:

R jt  R j ,t 1    DijkWijkt  F jt j , t  TS (5)
iI j k K j

where F jt is a continuous variable representing the unit residual life restoration during a

maintenance activity.
<Variables Fjt can only take two values: R max
j  R j ,t 1 or 0, depending on whether a

maintenance starts at time t or not (i.e. if Mjt=1 or Mjt=0). This relation between variables Fjt
max
and Mjt is expressed by the non-linear equation F jt  M jt ( R j  R j ,t 1 ) , that can be linearized

into the following set of linear constraints:

F jt  R max
j M jt j , t  TS (6)

F jt  R max
j  R j ,t 1 j , t  TS (7)

F jt  R max
j M jt  R j ,t 1 j, t  TS (8)
Finally, bounds on the variable Rjt have to be specified:

0  R jt  R max
j j, t  TS (9)

The lower bound on variable Rjt, here set to 0, is the threshold triggering the need for a
maintenance on unit j according to equations (5).
Consider again equations (5) and assume that unit j starts processing task i in mode k at time t.
Even though the production of the task lasts for more than one time unit, the residual life is
reduced at time t (i.e. at the beginning of the task) by the whole wear Dijk and then remains
constant until the end of the processing. A similar consideration also holds for a maintenance
activity starting at time t on machine j: equations (5) in combination with (6)-(8) restore the
residual life of the unit to its maximum already at time t while then keeping it constant until the
end of the maintenance, during which the unit cannot start any other production task due to
constraints (1).
4.1.5. Allocation of task to units (planning)
As we do not count for individual batches in the planning model, the number of times each task
is repeated must be tracked. Let Nijkt be the number of copies of task i produced by unit j in
mode k within the time interval t. Let us furthermore assume that at most one maintenance
activity per unit can be performed within a planning time interval. It is therefore possible to
write a relaxed version of the assignment constraints enforcing that the sum of task durations
(production and maintenance) assigned to a unit within a time interval does not exceed the
duration of the interval itself:

p
iI j kK j
ijk N ijkt   j M jt  H j , t  TP\{t p } ( 10 )

where H is the duration of the time interval t TP and t p is the first interval of the long term

planning horizon (special constraints are required for this interval, see Section 4.1.10).

4.1.6. Capacity constraints (planning)

Let Aijt be the total amount of material undergoing task i in unit j within the planning time
interval t. The variable Aijt is the extension of variable Bijkt defined in the scheduling horizon to
the aggregated planning model. As it is not possible to differentiate between the single batches
of the same task, there is no way to guarantee that each batch respects the unit capacity
constraints. On the other hand, it is at least possible to impose that the overall production of a
task on a unit within a time interval does that:
Vijmin  N ijkt  Aijt  Vijmax  N ijkt j, i  I j , k  K j , t  TP ( 11 )
kK j kK j

4.1.7. Material balance (planning)

Material balance constraints for the planning horizon do not significantly differ from the ones
introduced for the scheduling horizon:

S st  S s,t 1   is  Aijt   is  Aijt  dt s, t  TP\{t p } ( 12 )

iI s jJ i iI s jJ i

0  Sst  Cs s, t TP ( 13 )

Note that, as these constraints are only imposed at the border of the time interval, it is not
guaranteed that the inventory level of state s will not follow outside the feasible bounds within
the time interval. As a matter of fact, this depends only on the effective sequence and timing of
production and maintenance tasks within each time interval, which is not controlled in the
aggregate planning formulation.
4.1.8. Operation mode (planning)
One of the assumptions introduced to simplify the planning of the production in the long term
horizon is that each unit works in a single mode within a time interval. This means that the unit
cannot switch mode between consecutive tasks as in the scheduling horizon, but has to wait
until the start of the next time interval to do that. The selection of which mode a unit uses within
each time interval remains anyway a decision variable of the problem. By defining the binary
variable Modejkt that takes value 1 if unit j operates in mode k within time period t, these
constraints can be expressed as:

N ijkt  U * Mode jkt  j , i  I j , k  K j , t  TP ( 14 )

 Mode
k K j
jkt  1 j , t TP ( 15 )

where U is a large enough constant.

4.1.9. Unit residual life (planning)
Residual life update constraints for the planning horizon can be easily derived from the ones
introduced for the scheduling horizon:

R jt  R j ,t 1   D ijk N ijkt  F jt j , t  TP ( 16 )
iI j kK j

0  R jt  R max
j j, t  TP ( 17 )

A significant difference lies instead in the definition of the restoration variable Fjt. As the
correct sequence of production and maintenance tasks within a single time interval is not
known, also the exact profile of the residual life Rjt cannot be tracked. As a consequence, it is
not possible to write similar constraints like in the scheduling section (Eq. (6)-(8)) and to
guarantee that the life of the unit is restored to its maximum after a maintenance. On the other
hand, it is at least reasonable to impose that the restoration has to be 0 if no maintenance takes
place and that it cannot exceed the maximum life of the unit:

Fjt  R max
j M jt j , t  TP ( 18 )

As for the material balance constraints (12)-(13), it is worth to note that bounds (17) are only
imposed at the border of each planning time interval. It is therefore not possible to guarantee
that the residual life respects these bounds within each of them. As a matter of fact, this only
depends on the effective sequence of production and maintenance tasks which is not determined
by the aggregated solution.

Fig. 10 Handling of the material balance constraints across the scheduling and planning border

4.1.10. Interface between scheduling and planning horizon

Material balance and allocation constraints require a special treatment at the interface between
scheduling and planning horizon. Let t s be the last interval of the scheduling horizon and t p

the first one of the planning horizon. Let us first consider the material balance constraints and
the two cases shown in Fig. 10. In the first case, unit j1 finishes the processing of task i1 during
t s and the output materials of the task are provided to the intermediate/final product storages

before the end of the scheduling horizon. By denoting with S sfin the inventory level of state s at

the end of time interval t s :

S sfin  S s ,t s    is   Bijk ,t s 1 pijk  d s ,t s s ( 19 )

iI s jJ i k K j

As for the variables Sst , bounds for the new variable have to be imposed:
0  S sfin  C s s ( 20 )

In the second case, unit j2 finishes processing task i2 during the first planning time interval. As
a consequence, the produced material has to be accounted in the aggregated formulation of the
material balance constraint for time interval t p :

ts
Ss ,t p  S sfin   is   B ijkt '   is  Aij ,t p   is  Aij ,t p  d s ,t p s ( 21 )
iI s jJ i kK j t '  t s  2  pijk iI s jJ i iI s j J i

The allocation constraints for time interval t p follow a similar logic. Tasks starting on units

during the scheduling horizon and crossing the border between scheduling and planning
horizon, have to be taken into account while evaluating the duration of tasks active within time
interval t p such as in constraints (5). Then the following constraints hold:
ts ts

  W  p ijkt '
k K j iI j t '  t s  2  p ijk
ijk  t s  t '1    M 
t 't s  2 
jt ' ijk  t s  t '1  
j ( 22 )
 p ijk Nijk ,t p   j M j , t p  H j
iI j k K j

where the first two terms of the inequality represent the residual duration of production and
maintenance tasks that started during the scheduling horizon and are still active during t p .

4.1.11. Objective function

As mentioned in the beginning of this section, the target of the optimization problem is the
minimization of operation costs. In the following, different cost components are introduced
and the formal objective function of the optimization problem is written as a linear combination
of these terms:
 c sstorage ( S sfin   S st ) represents the storage costs for material in state s. The inventory
tT p

level is evaluated at the end of the scheduling horizon and at the end of each time
storage
interval of the planning horizon. cs represents the cost for the storage of one unit of
material in state s
  a M
jJ ,tT
j jt  b j F jt R max
j  represents the sum of the costs for each maintenance

occurrence on unit j. The cost of a single maintenance is given by a base cost a j

discounted by at most bj (with b j  a j ) depending on the amount of restoration Fjt. The

presence of this discount on the maintenance cost enforces the solution to plan
maintenance towards the end of the unit residual life and therefore to reduce
unnecessary maintenance costs
  D W
jJ ,tT
ijk ijkt  Z ijkt  represents the sum of the overall wear caused by the production on

the plant. This term enforces units to work in the operation mode causing less impact
on their residual life

Table 1 Feasible task-unit assignments

Unit Tasks
Heater Heating
Reactor1 Reaction1, Reaction2, Reaction3
Reactor2 Reaction1, Reaction2, Reaction3
Still Separation

Table 2 Plant units characteristics

Unit Min Max Max Initial Maintenance duration
capacity capacity life life [hours] [time slots]
[kg] [kg]
Heater 40 100 80 30 15 5
Reactor 1 32 80 150 50 21 7
Reactor 2 20 50 160 120 24 8
Still 80 200 100 40 15 5

Table 3 Plant states characteristics

State Capacity Cs [kg] Initial capacity [kg] Storage cost csstorage
Feed A ∞ ∞ 0
Feed B ∞ ∞ 0
Feed C ∞ ∞ 0
HotA 100 0 1
IntAB 200 0 1
IntBC 150 0 1
ImpureE 100 0 1
Product1 ∞ 0 5
Product2 ∞ 0 5
 Table 4 Demand profile
Time Demand Demand Time Demand Demand
Product 1 Product 2 Product 1 Product 2
[kg] [kg] [kg] [kg]
T56 150 200 W13 234 197
W2 88 150 W14 64 242
W3 125 197 W15 103 220
W4 67 296 W16 77 342
W5 166 191 W17 132 355
W6 203 193 W18 186 320
W7 90 214 W19 174 335
W8 224 294 W20 239 298
W9 174 247 W21 124 252
W10 126 313 W22 194 222
W11 66 226 W23 91 324
W12 119 121 W24 228 337

Table 5 Tasks processing times pijk [hours/time slots] and wear Dijk
Unit j Mode k Task i
Heating Reaction1 Reaction2 Reaction3 Separation
pijk Dijk pijk Dijk pijk Dijk pijk Dijk pijk Dijk
Heater Slow 9/3 1
Normal 6/2 2
Fast 3/1 3
Reactor1 Slow 27/9 4 36/12 1 30/10 3
Normal 15/5 5 21/7 3 18/6 7
Fast 9/3 8 15/5 5 6/2 8
Reactor2 Slow 30/10 4 33/11 2 24/8 2
Normal 18/6 5 18/6 4 21/7 5
Fast 12/4 10 12/4 4 12/4 9
Still Slow 15/5 2
Normal 9/3 5
Fast 6/2 6

4.2. Example and computational results

As an illustrative example, we will consider the STN representation in Fig. 3 derived from the
example proposed in Kondili et al. (1993). The plant has four units (Heater, Reactor1,
Reactor2, Still) capable of performing the five required task, according to the task-unit
assignment shown in Table 1. In Table 2 and Table 3, all other plant data are provided: in Table
2, for each unit the minimum and maximum capacity, the maximum life, the residual life at the
beginning of the scheduling horizon and the duration of a maintenance task are indicated; Table
3 introduces all the necessary states information like maximum and initial capacity as well as
the cost for the storage of one unit of material s. We consider a scheduling horizon Ts of one
week discretized in three hours slots (i.e. Ts={T1, T2,…,T56}) and a planning horizon Tp with
a discretization of one week. 23 intervals are chosen for the planning horizon (i.e. Tp={W2,
W3,…,W24}) so that the total planning and scheduling horizon results in almost six months. A
demand profile of final products Product 1 and Product 2 (expressed in kg) is provided in Table
4 for each week of the scheduling and planning horizon. Without any loss of generality, it will
be assumed that each unit can work in the same three operations mode (i.e. Kj=K): slow, normal
and fast. The processing time pijk as well as the wear Dijk for each combination of task, unit and
mode are reported in Table 5.

Table 6 Problem and model statistics for the example in section 4.2
Problem statistics
Number of units, tasks and modes (for each unit and task) 4–5–3
Number of scheduling and planning intervals 56 – 23
Model statistics:
Number of variables 5369
Discrete variables 2488
Continuous variables 2881
Number of constraints 5817

Table 7 Test runs results for the example in section 4.2

Test run ID CPUs Gap [%] Best solution
Run 1a 900 12.16 30362.34
Run 1b 3600 7.09 28786.16
Run 1c 10000 6.85 28732.46

The model (see Table 6 for basic statistics) has been implemented in GAMS 24.4.3 and solved
using CPLEX 12.6.1 on the NEOS Server (Czyzyk et al.,1998): 2x Intel Xeon X5660 at
2.8GHz, 64GB RAM. We let the optimization problem run with a time limit of 900 seconds
(Run 1a), 3600 seconds (Run 1b) and 10000 seconds (Run1c) and obtained the results
summarized in Table 7. As there is no significant difference in term of optimality gap between
Run 1b and Run 1c, the former will be used for further discussion. To fulfill the demand
requirements, 19 production tasks are scheduled (completely or to start) in the short term
horizon and over 400 over the planning horizon (Figure 11 shows a Gantt chart of the solution
obtained from Run 1b). The tasks are identified by their color according to the legend below
the picture; for each task, the selected operation mode is identified by the pattern used to fill
the rectangle, according to the legend. The diagram below the Gantt chart, shows the evolution
of the residual life (normalized to 1) of each unit. The histogram at the bottom shows (resp. on
the left and on the right of each planning period) the amount of final products available at the
beginning (if any) and at the end of each planning period (before fulfilling the demand).
Production and maintenance tasks are scheduled over the time horizon in a detailed way for the
scheduling horizon and in an aggregate manner in the planning one. Note that for
Fig. 11 Gantt chart, residual life of units and final production (Product 1 and Product 2) of the example in section 4.2
each single planning time period, the sequence of production and maintenance tasks is not
representative: no sequencing constraints have been imposed while only aggregate constraints
on the amount of production are set at the border of the time period. According to the
expectations, tasks scheduled on machine not running at full capacity (e.g. on Heater and Still)
are always operated in slow mode in order to minimize the wear on the units. Another
interesting observation can be done by looking at the distribution of the reactions tasks
(Reaction 1, Reaction 2 and Reaction 3) over the two reactors. Although both reactors can
process any reaction of the problem, the load of Reactor 1 is much higher compared to Reactor
2 (in the presented example 40% more tasks are scheduled on the first reactor). The explanation
to this phenomenon is that the average performance of Reactor 1 (in terms of production time
and wear caused by tasks on the unit) is better than the one of Reactor 2, in particular at Normal
and Fast modes. For this reason, as many tasks as possible are assigned to Reactor 1 and
operated in these modes. Similar considerations explain also why most of the Reaction 3 tasks
are scheduled on Reactor 2 and operated in slow mode.
The results of the example presented in the current section, show how the model correctly
behaves while integrating maintenance and production scheduling and planning decisions. The
presented example is a relatively big instance both in terms of the number of scheduling and
planning time interval as well as in terms of the resulting number of production and maintenance
tasks: the proposed approach is nonetheless capable of producing a “good” solution in a
reasonable amount of time. On the other hand, it is not surprising that the final solution quality
(e.g. the integrality gap) suffers from the problem dimension and can be improved for example
by changing the time discretization scheme and/or by reducing the length of the scheduling and
planning horizon. The effects of these changes on the performance of the model can be clearly
seen in the following two examples. In Run1d (and in Run1e which differs only on the
maximum CPU time) the same problem setting of the original problem is considered but with
different timing assumptions: the scheduling horizon is reduced to four days and the planning
one to 11 times interval of four days each (resulting in a total scheduling and planning horizon
of over one and a half months); the demand is proportionally reduced to account for the reduced
time available within the scheduling horizon and each planning time interval.
In the second example (Run1f and Run1g), the scheduling horizon is kept to one week but the
time discretization has been increased to 6 hours. In order to do that, the tasks production and
maintenance durations have been rounded up to the nearest multiple of the basic scheduling
interval. For the planning horizon, a base interval of two weeks has been chosen and the demand
has been scaled accordingly from the original problem (i.e. the demand of the scheduling
horizon is kept as in the original problem, for the first planning interval the demand is given by
the sum of demands of week 2 and 3 of the original problem, for the second planning interval
by the sum of demands of week 4 and 5…). Table 8 contains the key results of the above
describe models, showing the good performance of the approach for smaller instances of the
problem.

Table 8 Model statistics for smaller problem instances

Test run CPUs Scheduling Horizon Planning Horizon Gap Optimal
ID % solution
Total #Intervals Length of Total #Intervals Length of
length interval length interval
Run 1d 900 4 days 32 3h 92 days 23 4 days 6.5 4155.968
Run 1e 3600 4 days 32 3h 92 days 23 4 days 5.19 4155.968
Run 1f 900 1 week 28 6h 22 weeks 11 2 weeks 5.68 18356.929
Run 1g 3600 1 week 28 6h 22 weeks 11 2 weeks 4.47 18310.214

5. Assets degradation
The model introduced in Section 4 effectively describes the interaction between the production
and maintenance planning processes for the short and medium term horizon. Production tasks
are assigned to plant units and consume the unit life that can be restored by means of a
maintenance task. One of the underlying assumptions of the model is that the health of a unit
(e.g. measured in term of residual useful life) does not influence neither the capacity of a unit
to perform a task nor its performance. This assumption does not always hold in real industrial
applications. Plant assets frequently suffer from aging and wear directly caused by the
production and these phenomena can have very high impact on plant operations. In this section,
we consider two common degradation effects. Their influence on the scheduling model
presented above are described and discussed.

Fig. 12 Effect of scaling on the reduction of the maximum batch dimension

5.1. Batch dimension

The capacity of a unit represents typically the upper bound to the dimension of batches if other
upstream and downstream production constraints are not active. In most of the scheduling
approaches, the maximum capacity of a unit is assumed to be a constant eventually depending
on the type of batch processed but not on time or on the previous production history of the unit.
This assumption can be very stringent to be satisfied for many production units suffering e.g.
of scale formations: several layers of residual material / scale can deposit during the production
of consecutive batches, thereby reducing the effective volume available for the batch (Fig. 12).
In order to take these effects into account, the model presented in Section 4 will be now
extended. The first step is to introduce a model for the unit capacity degradation. It is assumed
in the following that the degradation impacts only on the maximum dimension of the batches,

while the minimum ( Vij

min
) is not influenced. Let  0,1 and assume that
Vijtmax  (1   )Vijmax   ( R j ,t 1 / R max
j )Vijmax . In this formulation the maximum unit capacity

depends on the residual life in a linear way and the parameter α describes the impact of
degradation on the maximum unit capacity. Let us consider for example the two extremes: if
  0 the degradation has no effect on the maximum unit capacity which remains constant
throughout the unit life ( Vijtmax  Vijmax ); if   1 the maximum capacity decreases linearly from

Vijmax to 0 as the unit approaches the end of its life. Without any loss of generality, we can fix

  1 / 2 so that the maximum capacity is equal to Vijmax when the unit has just undertaken a
max max
maintenance ( R jt  R j ) and to Vij / 2 when the residual life is almost zero. In this way, the

maximum batch capacity can be written as:

Vijmax  R j ,t 1 
V max
 1   ( 23 )
2  R max 
ijt
j 
This expression can now be used in the capacity constraints introduced for the scheduling and
planning horizon. Constraints (2) are substituted in a straightforward way by the following two
sets of constraints:

VijminWijkt  Bijkt  VijmaxWijkt j , i  I j , k  K j , t  TS ( 24 )

Vijmax  R j ,t 1 
Bijkt  1   j , i  I j , k  K j , t  TS ( 25 )
2  R maxj


The set of constraints (24) (equal to the original ones) guarantees that the batch dimension is
equal to zero if no batch starts to be processed ( Wijkt  0 implies Bijkt  0 ) and that is greater
than the minimum Vijmin if a production starts. In case a batch starts to be processed in unit j at

time t, then its maximum dimension is bounded by the second set of constraints according to
the degradation model introduced before.
For the planning horizon, the starting points for the derivation of the constraints are the original
batch capacity constraints (11) and the expression for the maximum batch capacity defined in
(23). Note that only the upper bound on Aijt in (11) has to be modified, as the minimum capacity
is not influenced by the degradation according to the model introduced before. It is then possible
to write:

Aijt  Vijtmax N ijkt

k K j

Vijmax  R 
 1  j ,t 1   N ijkt
2  R max
j
k K
 j
Vijmax  R j ,t 1 
  ijkt ijkt R max  j, i  I j , t  TP

2 kK j 
N  N
j 
The bilinear term N ijkt R j , t  1 that appeared in the previous expression has to be eliminated to

preserve the overall linearity of the model. In order to do that, it is possible to use the 2-based
binary representation of integers (Harjunkoski et al.,1999) to obtain:
N
N ijkt   2 n 1 Qijktn ( 26 )
n 1

where Qijktn is a binary variable and N is a big enough integer. The value of N has to be chosen
using some process knowledge to estimate the maximum value of Nijkt, i.e the maximum number
of batches (of type i performed on unit j in mode k) that can be processed within each time
interval tp. In this way, the previous constraints become:

Vijmax    R ,t 1 
Aijt    N    2n 1Qijktn  jmax
 R j 
ijkt
2 k K j   n
( 27 )
max
Vij  
   N ijkt   2n 1Oijktn  j, i  I j , t  TP
2 k K j  n 
R j ,t 1
where an additional continuous variable Oijktn  Qijktn has been introduced. Variable Oijktn
R max
j

is the product of a binary and of a continuous variable between zero and one and can therefore
be linearized (with the same technique used in section 4.1.4 for the residual life update) by the
set of linear constraints:
Oijktn  Qijktn j , i  I j , k  K j , t  TP , n  N ( 28 )

R j ,t 1
Oijktn  j , i  I j , k  K j , t  TP , n  N ( 29 )
R max
j

R j ,t 1
Oijktn  Qijktn   1 j , i  I j , k  K j , t  TP , n  N ( 30 )
R max
j

Constraints (26)-(30) with the addition of the original lower bound on Aijt ( Aijt  Vijmin N ijkt )
k K j

substitute the batch capacity constraints (11) for the planning horizon in case of reduction of
batch dimension due to degradation.
The current model has been tested on the same problem described in section 4.2 with only minor
modifications: the planning horizon has been reduced to three months (i.e. one week of detailed
schedule as in the previous example and 11 weeks for the aggregate planning) and the demand
has been decreased by 20% over the whole scheduling and planning horizon. Key model
statistics and computational results for the three test runs (with 900s, 3600s and 10000s time
limit) are summarized in Table 9 and Table 10. As before, results from the test run with
computational time limit 3600s will be discussed (Run 2b); the demand is fulfilled by 22 tasks
scheduled in the short term and 165 in the long term planning horizon. Even though the gap is
still relatively high (around 15%), the solution presented in Fig. 14 shows some interesting
properties. Reactor1 is operated until almost half of the scheduling horizon without performing
a maintenance, even though its remaining useful life is significantly low and the maximum
batch dimension is reduced almost at half of the maximum capacity (the maximum dimension
of the batch just before the maintenance is scheduled is only 58% of the total). Right after the
maintenance two batches of Reaction 2 and Reaction 3 are run at very high capacity and
fast/normal modes in order to reach the production targets of Product 1 and of Impure E (which
is then immediately used to produce the required target of Product 2).

Table 9 Model statistics for the example in section 5.1

Problem statistics
Number of units, tasks and modes (for each unit and task) 4–5–3
Number of scheduling and planning intervals 56 - 11
Model statistics
Number of variables 6173
Discrete variables 2800
Continuous variables 3373
Number of constraints 9109
 Table 10 Test runs results for the example in section 5.1
Test run ID CPUs Gap [%] Best solution
Run 2a 900 18.98 3286.77
Run 2b 3600 14.56 3128.01
Run 2c 10000 13.39 3087.88

5.2. Available operation modes

In the model presented in Section 4, it is assumed that an idle unit can always start the
processing of a task (if all other constraints are satisfied) in any operation mode independently
from its health condition in terms of residual life. If the unit is subject to degradation, this
assumption might not always be valid. Let us consider again the example of degradation due to
scaling introduced in the previous paragraph, but this time applied to a pipeline (unit) that has
to transport some material (task) at a constant pressure within the plant. If the scale significantly
reduces the capacity of the pipe (Fig. 13), the amount of time required to finish the task will be
much longer. In the modeling framework introduced here, this can be interpreted by assuming
that the pipeline can only be operated in slow or normal mode, but

Fig. 13 Section of a pipe after as good as new (left) and affected by scaling (right)

not in fast mode. In order to model this behavior within the framework introduced in Section 4,
it is assumed that at any time t a unit j can start the processing of task i only if its residual life
Rjt is greater than a given threshold mijk . To simplify the notation, it will be assumed in the

following that each unit j can work in the same set of operation modes (Kj=K) and that the
threshold allowing the start of a task on a unit is expressed in percentage of the unit residual
life and is independent from the type of task and unit ( mijk  mk ). With these assumptions, a

unit can start the processing of a task in mode k only if the percentage of residual life is greater
that mk. This constraint can be straightforwardly expressed through the model variables for the
scheduling and planning horizon:

R jt R max
j
Wijkt  j, i  I j , k  K j , t  TS ( 31 )
mk
 R j ,t 1 R max
j
Mode jkt  j , i  I j , k  K j , t  TP ( 32 )
mk
These constraints have been integrated in the complete model of section 4.1 and tested on the
reduced dataset of the previous section (fewer planning time intervals and lower demand).
Thresholds mk on the percentage of remaining useful life allowing to use mode k (or higher
modes) are set as in Table 11. The key model statistics and results data for three test runs (with
900s, 3600s and 10000s time limit) are shown in Table 12 and Table 13. The solver manages
to find a very good solution (around 2% gap) in less than 900 seconds of computation and no
better solutions are found even with a higher time limit (see Run3b and Run3c). A Gantt chart
of the solution is shown in Fig. 15 in which 16 tasks are scheduled in the short term and 140 in
the long term planning horizon. Once again, production and maintenance tasks are scheduled
over the scheduling and planning horizon to reach production targets while minimizing plant
operation costs. As imposed by constraints (31)-(32), the operation mode of each unit is
controlled by the residual useful life. For example, Reactor 1 is capable of running in Normal
mode only for the first batch of the production. After that, the residual life falls under the 30%
threshold allowing the unit to run only in Slow mode. Only after a maintenance has been planned
(during time period W4), the unit is capable of running in faster modes according to the
production requirements (like during W6, W8 and W9). This approach is also applicable to the
earlier discussed AOD case from steel production.

Table 11 Percentage of remaining useful life (R%) and available modes

Threshold Available modes
R%  mFast  0.6 Slow, Normal, Fast
R%  m Normal  0.3 Slow, Normal
R%  mSlow  0 Slow

Table 12 Model statistics for the example in section 5.2

Problem statistics
Number of units, tasks and modes (for each unit and task) 4–5–3
Number of scheduling and planning intervals 56 - 11
Number of variables 4589
Discrete variables 2008
Continuous variables 2581
Number of constraints 6021
Fig. 14 Gantt chart, residual life and final production (Product 1 and 2) of the example in section 5.1. The coloring follows the legend included in Fig. 11

Fig. 15 Gantt chart, residual life and final production (Product 1 and 2) of the example in section 5.2. The coloring follows the legend included in Fig. 11
 Table 13 Test runs results for the example in section 5.2
Test run ID CPUs Gap [%] Best solution
Run 3a 900 1.45 5284.95
Run 3b 3600 0.65 5284.95
Run 3c 10000 0.51 5284.95

6. Comparison and limitation

In the previous sections three models to tackle the integrated maintenance and production
scheduling problem have been introduced and described. The objective of this section is to
briefly highlight strengths and limitations of the basic multi-time model introduced in section
4. The model represents the basis for the extensions introduced in the previous section; the
considerations summarized in the following are therefore valid for all models presented in the
paper. One of the first noticeable advantages is that owing to the long term view of the approach
it is possible to have a reasonable estimation of when the next maintenance will take place on
each asset; in this way, it is possible to plan in advance all the necessary maintenance related
activity (e.g. assignment of maintenance personnel, gathering of required
material/equipment…) with positive impact on the company organization. Another advantage
of the proposed approach is the possibility to anticipate future demand fluctuation and to
schedule plant operations accordingly, e.g. by increasing the final product stock to leave
capacity for an upcoming maintenance activity or vice versa to decrease asset utilization (by
using lower operations modes) to postpone its need.
The presented models show also some limitations, both from the structural and from the
performance point of view, leaving space for future research and improvement. On the one
hand, one of the most stringent modeling assumptions is the utilization of a single operation
mode for each time interval of the planning horizon. This assumption, necessary to reduce the
computational complexity of the model, is particularly striking with the behavior of the model
in the short term where a great alternation of operation modes is shown (see for example the
short term scheduling part of the three Gantt charts presented in the paper). The second
limitation, especially evident in the extended model presented in section 5.1, is the relatively
high integrality gap even after long computation time. This behavior calls for the development
and application of advanced optimization techniques (e.g. Branch&Cut, Branch&Price,…) in
order to get faster to a good solution and/or to prove its distance to optimality.
7. Summary and conclusion
In this paper, a framework for the effective integration of maintenance and production in a
generic process plant is introduced. Production and maintenance requirements are tightened
together by the concepts of residual life and operation modes of plant units that have been
inspired by the observation of real production processes. In a realistic environment,
maintenance requirements might appear only few times within the whole planning horizon;
nonetheless, their impact on plant operations is so high, that they have to be taken into account
while generating a feasible production schedule. An effective multi-time-scale approach has
been proposed to account for this realistic situation: while maintenance and production tasks
are scheduled on a very detailed level for the short-time horizon, an aggregated model of plant
operations is proposed for the long term horizon to be able to account for maintenance related
decisions.
In real applications, plant units’ aging is frequently associated with the deterioration of
performance. Two examples of how this phenomenon can impact the scheduling problem have
been described, i.e. the reduction of the maximum capacity of units and of the number of
available operation modes that can be used by units to perform the required tasks.
The validity and the performance of the introduced models have been tested on examples that
are representative for a great number of plants from the process industry. The presented
examples highlighted on the one hand the effectiveness of the models to tackle the integrated
maintenance and production scheduling problem (with and without asset performance
degradation); on the other hand, they call for the need of further research as described in the
previous section.
The use of unit health information to generate a feasible production plan and to trigger the need
for a maintenance activity represents a new approach to describe plant operations in an
integrated way and to bridge the gap between the domains of production scheduling and asset
management in process plants.

Acknowledgements
The Marie Curie FP7-ITN research project "ENERGY-SMARTOPS", Contract No: PITN-GA-
2010-264940 is acknowledged for financial support.
References

Aggoune, R., 2004. Minimizing the makespan for the flow shop scheduling problem with
availability constraints. European Journal of Operational Research 153, 534–543

Alle, A., Pinto, J. M., Papageorgiou, L. G., The economic lot scheduling problem under
performance decay. Industrial & Engineering Chemistry Research 2004, 43, (20), 6463-6475

Biondi, M., Sand, G., Harjunkoski, I., 2015. Tighter integration of maintenance and production
in short-term scheduling of multipurpose process plants. Computer Aided Chemical
Engineering 37, 1937-1942

Bock, S., Briskorn, D., Horbach, A., 2012. Scheduling flexible maintenance activities subject
to job-dependent machine deterioration. Journal of Scheduling 15 (5), 565-578

Castro, P. M., Grossmann, I. E., Veldhuizen, P., Esplin, D., 2014. Optimal maintenance
scheduling of a gas engine power plant using generalized disjunctive programming. AIChE
Journal 60, 2083–2097

Cheung, K.-Y., Hui, C.-W., Sakamoto, H., Hirata, K., O’Young, L., 2004. Short-term site-wide
maintenance scheduling. Computers and Chemical Engineering 28, 91–102

Czyzyk, J., Mesnier, M. P., Moré, J. J., 1998. The NEOS Server. IEEE Journal on
Computational Science and Engineering 5 (3), 68-75

Dedopoulos, I.T., Shah, N., 1995. Optimal short-term scheduling of maintenance and
production for multipurpose plants. Industrial & Engineering Chemistry Research 34 (1), 192–
201

Harjunkoski, I., Westerlund, T., Pörn, R., 1999. Numerical and environmental considerations
on a complex industrial mixed integer non-linear programming (MINLP) problem. Computers
and Chemical Engineering 23, 1545–1561.
Hazaras, M.J., Swartz, C.L.E., Marlin, T.E., 2012. Flexible maintenance within a continuous-
time state-task network framework. Computers and Chemical Engineering 46, 167–177

Jain, V., Grossmann, I. E., Cyclic Scheduling of Continuous Parallel Process Units with
Decaying Performance, AICHE Journal,1998, Vol. 44, No. 7, pp. 1623-1636

Janak, S. L., Lin, X., Floudas, C. A., 2004. Enhanced continuous time unit-specific event-based
formulation for short-term scheduling of multipurpose batch processes: Resource constraints
and mixed storage policies. Industrial and Engineering Chemistry Research 43, 2516– 2533.

Kondili, E., Pantelides, C. C., Sargent, W. H., 1993. A general algorithm for short-term
scheduling of batch operations-I. MILP formulation. Computers and Chemical Engineering 2,
211–227

Liu, S., Yahia, A., Papageorgiou, L.G., (2014). Optimal Production and Maintenance Planning
of Biopharmaceutical Manufacturing under Performance Decay. Industrial & Engineering
Chemistry Research , 53 (44) pp. 17075-17091

Maravelias, C. T., Grossmann, I. E., 2003. New general continuous-time state-task network
formulation for short-term scheduling of multipurpose batch plants. Industrial and Engineering
Chemistry Research, 42, 3056–3074

Maravelias, C. T., Grossmann, I. E., 2004. A hybrid MILP/CP decomposition approach for the
continuous time scheduling of multipurpose batch plants. Computers and Chemical engineering
28, 1921-1949

Méndez, C.A., Cerdá, J., Grossmann, I.E., Harjunkoski, I., Fahl, M., 2006. State-of-the-art
review of optimization methods for short-term scheduling of batch processes. Computers and
Chemical Engineering 30, 913-946.

Naderi, B., Zandieh, M., Aminnayeri, M., 2011. Incorporating periodic preventive maintenance
into flexible flowshop scheduling problems. Applied Soft Computing 11 (2), 2094–2101
NAMUR (2009). NAMUR-Recommendation NE129: Plant asset management. User
association of automation technology in process industries (NAMUR)

Nie, Y., Biegler, L. T., Wassick, J. M., Villa, C. M., 2014. Extended Discrete-Time Resource
Task Network Formulation for the Reactive Scheduling of a Mixed Batch/Continuous Process.
Industrial & Engineering Chemistry Research 53, (44), 17112-17123

Papageorgiou, L. G., Pantelides, C. C., 1996. Optimal campaign planning scheduling of

multipurpose batch semicontinuous plants. 1. Mathematical Formulation. Industrial &
Engineering Chemistry Research 1996, 35, (2), 488-509

Shah, N., Pantelides, C. C., & Sargent, W. H., 1993. A general algorithm for short-term
scheduling of batch operations-II. Computational issues. Computers and Chemical Engineering,
2, 229–244

Varnier, C., Zerhouni, N., 2012. Scheduling predictive maintenance in flow-shop. Prognostics
& System Health Management Conference

Velez, S., Merchan, A. F., Maravelias, C. T., 2015. On the solution of large-scale mixed integer
programming scheduling models. Chemical Engineering Science 136, 139-157