Professional Documents
Culture Documents
Stochastic Modelling
in Production Planning
Methods for Improvement
and Investigations on Production
System Behaviour
Stochastic Modelling
in Production Planning
Alexander Hübl
Stochastic Modelling
in Production Planning
Methods for Improvement
and Investigations on Production
System Behaviour
Alexander Hübl
Steyr, Austria
Springer Gabler
© Springer Fachmedien Wiesbaden GmbH 2018
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with
regard to jurisdictional claims in published maps and institutional affiliations.
The applied simulation model was published in Hübl et al. (2011), Fel-
berbauer et al. (2012) and Altendorfer et al. (2013). Preliminary work
has been published in the following conference papers: Hübl and Gmainer
(2008); Hübl and Jodlbauer (2008); Hübl et al. (2009, 2010); Hübl (2014);
Hübl and Altendorfer (2015).
Alexander Hübl
Wels, March 2015
The PhD thesis has been published by Springer within the FTI (Förde-
rung der wirtschaftlich – technischen Forschung, Technologieentwicklung
und Innovation in Oberösterreich) policy of the federal government of Up-
per Austria for the reserach field ”Supply Chain Planning” of Logistikum,
University of Applied Sciences Upper Austria. I’d like to give special thanks
to my co-authors of the published articles for their inspiration and to Erika
for proofreading this book.
Alexander Hübl
Wels, April 2017
Abstract
In this thesis models for production planning are developed and perfor-
mance indicators are analysed to investigate production system behaviour.
Existing literature is extended by considering uncertainty of customer re-
quired lead time and processing times and by increasing the complexity of
multi-machine multi-items production models.
Results are on the one hand a decision support system for determining
capacity and the further development of the production planning method
Conwip and on the other hand the JIT intensity has been developed and the
effects of dispatching rules on production lead time are analytically proven.
The decision support system for determining the capacity provided is based
on the capacity demanded, whereby process and/or customer uncertainty
can be included. The production planning method Conwip has been ex-
tended by including safety stock and by extending the inventory constraint
to sum of WIP (Work-In-Process) and FGI (Finished Goods Inventory) and
not only WIP. JIT-intensity measures the degree of fulfilment of the ”seven
zeros” in a production system. A high JIT intensity leads to an increase of
excess capacity and to a reduction of utilisation and means that the produc-
tion system follows the customer demand fluctuations without any ”seven
zeros”-relevant losses. Consequently, management has the opportunity to
increase sales by selling more customer required items or to reduce excess
capacity. Finally, the effects of dispatching rules on average production
lead time are investigated. An analytic relationship between ”processing
time weighted average production lead time” and covariance between pro-
cessing time and queuing time has been identified. This relationship has
been proven analytically for single-machine production systems. It has been
found out that the ”processing time weighted average production lead time”
for a multi-machine production system is not invariant with respect to the
applied dispatching rule.
Contents
Preface V
Abstract VII
Contents IX
1 Introduction 1
1.1 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research Questions . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Literature Review 9
2.1 JIT Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Hierarchical Production Planning Models . . . . . . . 9
2.1.2 Utilisation in Production Planning . . . . . . . . . . . 12
2.1.3 Impact of JIT Activities on Performance . . . . . . . . 12
2.2 Capacity Setting Methods . . . . . . . . . . . . . . . . . . . . 14
2.3 Conwip . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4 Dispatching and Production Lead Time . . . . . . . . . . . . 19
3 Utilisation Concept 23
3.1 Seven Zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.1 Utilisation . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.2 JIT Intensity . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.1 Long Term Level . . . . . . . . . . . . . . . . . . . . . 31
3.3.2 Medium Term Level . . . . . . . . . . . . . . . . . . . 32
3.3.3 Short Term Level . . . . . . . . . . . . . . . . . . . . . 32
3.4 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . 35
5 Conwip 87
5.1 Model Description . . . . . . . . . . . . . . . . . . . . . . . . 88
5.1.1 Production System . . . . . . . . . . . . . . . . . . . . 88
5.1.2 Production Planning and Control Applying
Conwip . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.1.3 Simulation Model . . . . . . . . . . . . . . . . . . . . . 90
5.1.4 Simulation Study . . . . . . . . . . . . . . . . . . . . . 91
5.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.2.1 Discussion of WAW and Wipcap . . . . . . . . . . . . 93
5.2.2 Wipcap Method Comparison . . . . . . . . . . . . . . 95
5.3 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . 97
5.4 Appendix Conwip . . . . . . . . . . . . . . . . . . . . . . . . 97
7 Conclusions 121
Bibliography 123
List of Figures
1.1 SCP matrix (Rohde et al., 2000; Stadtler, 2005; Stadtler and
Kilger, 2008) . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
3.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 Decreasing utilisation by realising JIT goals in bottleneck
situations of the equipment considered . . . . . . . . . . . . . 29
3.3 Effects of parameters on total variable costs and revenue by
short term decisions . . . . . . . . . . . . . . . . . . . . . . . 34
Figure 1.1: SCP matrix (Rohde et al., 2000; Stadtler, 2005; Stadtler and Kilger,
2008)
Scheduling defines the sequence of production orders for each resource. Dis-
tribution Planning and Transport Planning optimise the transport costs and
holding costs of finished goods in the supply chain which results in detailed
plans for vehicle routing. The main tasks of Purchasing & Material Re-
quirements Planning are to generate purchase orders and to verify material
availability based on the Demand Fulfilment & ATP module.
Graves (2011) states that three major types of uncertainties affect a
production plan of manufacturing companies:
Mula et al. (2006) highlight that further research is needed in the area of
production planning under uncertainty: (I) Investigation of new approaches
to modelling of uncertainty, (II) development of new models that consider
additional sources and types of uncertainty, such as Customer Required
Lead Times (CRL), transport times, quality uncertainty, failure of pro-
duction systems and changes to product structure, (III) investigation of
incorporating all types of uncertainty in an integrated manner, (IV) devel-
opment of empirical work that compares the different modelling approaches
with real case studies, (V) development of a comparative evaluation of the
existent models for the different manufacturing systems. Following the hi-
erarchical planning concept, the reviews of Fleischmann and Meyr (2003),
Meyr (2004) and Missbauer and Uzsoy (2011) argue that little research has
been conducted on the interdependence between the planning hierarchies
and therefore further research is needed.
A further complication arises if the customer has the option of revising
the order specifications, which is a typical situation for automotive suppli-
ers. The integration of such stochastic behaviour into the planning process
4 Chapter 1. Introduction
In this section the research gap and research questions are identified, where-
by the literature review will be presented in Chapter 2.
The reviewed literature focusing on utilisation in a hierarchical produc-
tion planning structure ignores an investigation of firm’s performance of JIT
activities. However, it is known that opposing goals exist on each hierarchi-
cal level. This research gap will be presented in Section 2.1.
None of the reviewed models for capacity setting assume general distri-
bution of customer required lead time and processing times. Multi-machine
setting is not considered in the reviewed models. Moreover, the investigated
models in literature lack a rolling time horizon. Section 2.2 will cover this
research gap in more detail.
The pull production planning method Conwip will be reviewed in Section
2.3. It will be identified that the implementation of Conwip focusing on
improvements of service level and tardiness by changing Wipcap definitions
itself is rarely discussed. Moreover, it is shown that the implementation of
safety stock in a Conwip system has not been dealt with so far.
After reviewing the literature for dispatching rules influencing average
production lead time, it can be stated that the relevant literature ignores the
analytical link between average production lead time and the ”processing
time weighted production lead time” for multi-stage production systems.
Moreover, the influence of different dispatching rules on average production
lead time has not yet been proven theoretically. Section 2.4 will consider
this research gap thoroughly.
To refine the research field, the following specific research questions are
answered in Chapters 3 to 6 respectively:
2. How can a decision support system for short and medium capacity set-
ting under a stochastic customer behaviour and stochastic processing
process be configured? (Chapter 4)
1.3 Methodology
Discrete event simulation and basic concepts of queuing theory are applied
in this thesis to answer the research questions. Mula et al. (2006) have found
out that most of the analytical models address only one type of uncertainty
by assuming a simple structure of the production process. Moreover, for
more complex structures of the production process and more types of un-
certainty the problems are solved by the use of simulation and artificial
intelligence. Since production planning has to deal with a huge amount of
different parameters and variables and a lot of stochastic effects, simulation
is used in recent literature (Kim and Kim, 1994; Huang et al., 1998b; Ku-
tanoglu and Sabuncuoglu, 1999; Gilland, 2002; Abdul-Kader and Gharbi,
2002; Arakawa et al., 2003; Li Sun et al., 2007; Jodlbauer and Huber, 2008;
Hübl et al., 2011; Felberbauer et al., 2012; Altendorfer et al., 2013; Gansterer
et al., 2014) to identify optimal parameters and methods for production
planning.
The most time consuming phase for discrete event simulation studies
is the input data collection and model development (Trybula, 1994; Perera
and Liyanage, 2000; Randell and Bolmsjo, 2001). Traditional simulation
models are created for a certain case only and adapting these models is
both personnel intensive and time consuming (Randell and Bolmsjo, 2001).
Therefore, a scalable simulation model for analysing production systems
with discrete event simulation has been developed (Hübl et al., 2011). The
simulation model is suitable for the three hierarchical planning levels ac-
cording to MRPII (Hopp and Spearman, 2008). The model has also been
applied in some case studies for practitioners (Felberbauer et al., 2012;
Altendorfer et al., 2014). This scalable simulation model is based on a
database, where the structure of the production system is defined in order
to avoid throw away solution (Thompson, 1994). Basically, the structure
6 Chapter 1. Introduction
1.4 Outline
Chapter 2 provides an overview of relevant and recent literature focusing
on the defined research field in Chapter 1.
In Chapter 3, which is based on the working paper Hübl and Jodlbauer
(2013), an utilisation concept of the three planning levels (long, medium,
short) of the SCP matrix based on the JIT philosophy is developed and
discussed analytically. The ”seven zeros” as JIT philosophy are introduced
and the optimal utilisation on each hierarchical production level concerning
opposing goals is discussed.
Based on the result in Chapter 3, representing a predefined optimal
utilisation level, a decision support for short and medium term capacity
setting is developed in Chapter 4. This periodical capacity setting method
for MTO multi-machine production systems includes stochastic customer
required lead time and stochastic processing times to improve service level
and tardiness. The basic idea of this chapter is that the cumulated capacity
provided is greater than or equal to the cumulated capacity demanded for
the planning horizon. Based on the information available, processing time
distribution, customer required lead time distribution and a rolling plan-
ning horizon are considered on the periodical capacity setting method for
1.4. Outline 7
Literature Review
Section 2.1 of this review covers the literature on key performance indi-
cator utilisation and JIT philosophy in a hierarchical production planning
setting. The reviewed literature lacks of a differentiated investigation on
utilisation for each hierarchical production planning level for a production
system avoiding waste in terms of a JIT perspective.
Demand
Forecast
Job Routing
Short term Release Data
Job
Dispatching
Figure 2.1: Manufacturing resource planning MRPII (Wight, 1984; Hopp and
Spearman, 2008)
A fixed capacity is assumed and if the demand exceeds the fixed capacity
the orders are delayed. Moreover, an additional capacity is introduced by
hiring additional personnel.
Decision problems with capacity expansion and/or reduction are mod-
elled in most cases as dynamic programs (Mieghem and Rudi, 2002; Bradley
and Glynn, 2002; Li et al., 2009). Bradley and Glynn (2002) develop an an-
alytic model for a single-machine and single product system which describes
the optimal long term balance between capacity and inventory. They show
that optimal inventory policy varies with capacity investment and that
higher capacity invested allows less inventory. Moreover, the authors de-
scribe how inventory should be optimally substituted for capacity in order
to minimise costs when the capacity level varies. Mieghem and Rudi (2002)
and Angelus and Porteus (2002) deal with this issue in a more general situ-
ation and gain similar results. Obviously there is a trade off between capital
invested in capacity and costs of the capital employed in inventories. Defreg-
ger and Kuhn (2006) develop an MTO order acceptance model with limited
inventory capacity, where it is possible to adjust the inventory levels. In Li
et al. (2009), capacity allocation methods with mixed integer programming
methods are compared for supply chain optimisation. The authors iden-
tify that an integrated planning approach achieves better results than an
approach where each subproblem is treated separately.
The MTO ability of production systems is evaluated in Jodlbauer (2008c)
depending on the capacity provided, the customer required lead time dis-
tribution, and the demand fluctuation. The result of this evaluation shows
that applying a capacity adjustment method which enables the reaction to
short term peaks can decrease the capacity provided for MTO environment.
Jodlbauer and Altendorfer (2010) present a concept for optimising the over-
all capacity provided, for which they use the customer required lead time
distribution too. The result of this paper indicates that flexible capacity on
a short and medium term basis can lead to a cost decrease.
The papers of Balakrishnan et al. (1996) and Balakrishnan et al. (1999)
discuss the capacity rationing problem for a two-product production sys-
tem whereby one product class leads to higher profits per unit. An order
rejection policy for the lower profit products to maximise company profit is
described. In Kok (2000), capacity allocation is discussed where capacity
has to be allocated to different product groups while minimising a total cost
function. In this model the production is triggered by an order-up-to policy
for each product group.
Yang and Fung (2014) present an ATP decision support for order accep-
tance/selection, due date assignment and order scheduling in a multi-site
MTO production supply chain system. They propose a two-stage hierarchi-
cal structure.
16 Chapter 2. Literature Review
2.3 Conwip
Section 2.3 summarises literature about the production planning method
Conwip, whereby the literature is distinguished between analytical models
and simulation. The production planning and control method Conwip is
firstly introduced by Spearman et al. (1990) and Spearman and Zazanis
(1992). The basic parameters Wipcap, work-ahead-window and capacity
trigger are already introduced in these papers. The method has since then
been further developed and compared to the most dominant production
planning and control methods MRP and Kanban either analytically or with
simulation. For a detailed review on Conwip and its applications see Frami-
nan et al. (2003). Prakash and Chin (2014) review 15 modified Conwip sys-
tems and classify them according to: (I) the feedback signal from demand
source triggers production, (II) the products are pushed between worksta-
tions and (III) the WIP level (Wipcap) is limited within workstations.
Looking at the literature discussed in the review below shows that Con-
wip performs well in comparison to other traditional production planning
and control methods. However, the implementation of Conwip in an MTO
system is yet rarely discussed as mostly average inventory and throughput
are compared. Especially research on the question of improving service level
and tardiness measures by changing the Wipcap definition itself and the
evaluation of a safety stock implementation in Conwip is still not available.
Analytical models usually describe the behaviour and performance of
Conwip in comparison to Kanban (or other production planning and control
methods) or they deliver approximations for basic logistical key performance
indicators, such as throughput, WIP and production lead time when Conwip
is applied to a more complex production system.
The first paper introducing Conwip, Spearman et al. (1990) model a
multi-stage serial production system as closed loop queuing network. More-
over, the authors compare its performance to an open loop queuing network
representing the push production control method.
Similarly the superior performance of Conwip in comparison to Kanban
and push methods as MRP is proven in Spearman and Zazanis (1992).
2.3. Conwip 17
Spearman et al. (1990) and Hopp and Spearman (2008) to apply First-In-
System-First-Served (FISFS) within a Conwip production system.
Hübl et al. (2011) develop a simulation model where Conwip is embedded
in a hierarchical production planning structure as proposed by Hopp and
Spearman (2008).
Chong et al. (2013) investigate an integrated Kanban and Conwip pro-
duction system by the use of discrete event simulation. They use total
output, average WIP, average production lead time and average bottleneck
utilisation as performance metrics.
Huang et al. (2015) consider a Conwip assembly production system for
mass production with multi-loop, multi-products, low volume and one-of-a-
kind production environments. The authors develop eight Conwip loop con-
trol policies based on five basic design patterns. The result of the simulation
study shows that the multi-loop design performs better than a single-loop.
Moreover, they have also proposed a WIP upper bound heuristic algorithm
for searching the optimal Wipcap.
are combined to routes such that idle (empty) travel time is minimised.
2.4. Dispatching and Production Lead Time 21
and Earliest Due Date (EDD) are good dispatching rules to reduce average
production lead time in semiconductor wafer fabrication. Land et al. (2014)
focus on overcoming the conflict between order release and dispatching. The
authors find out that operating due date dispatching rules are beneficial.
Jayamohan and Rajendran (2000), El-Bouri et al. (2008) and Chen and
Matis (2013) compare new dispatching rules with standard rules with regard
to different performance measures such as average, maximum and variance
of production lead time and tardiness. The results of Jayamohan and Rajen-
dran’s (2000) simulation study show that SPT is the rule with the shortest
average production lead time for flow shops.
Whenever simulation is applied to discuss the effect of dispatching rules
on the average production lead time, the focus is mainly on comparing and
evaluating the performance of dispatching rules and developing new dis-
patching rules. Very little literature is available on the development of ana-
lytical models based on simulation and empirical studies. One such stream is
based on the application of the funnel model for single-stage models (Wien-
dahl et al., 1994; Wiendahl and Breithaupt, 1999) where approximation
functions for production lead time and inventory are identified applying
empirical data for their parametrisation. Based on this funnel model, Ny-
huis and Wiendahl (2009) derive approximations for average production
lead time using SPT and Longest Processing Time (LPT) as dispatching
rules.
From the research work reviewed above, it is obvious that most litera-
ture dealing with the interaction between dispatching rules and production
lead time either applies simulation or empirical data and approximation
equations. Nevertheless, queuing theory can also be applied.
An analytical model based on queuing theory shows that the SPT rule
minimises the expected production lead time in a static job shop with de-
terministic processing times (Buzacott and Shanthikumar, 1993; Hopp and
Spearman, 2008).
The work by Nyhuis and Wiendahl (1999) and Nyhuis and Wiendahl
(2009) introduce the range, which is the ”processing time weighted average
production lead time” in comparison to the arithmetic average production
lead time. They show that the range is independent of the dispatching
rule applied for a single-stage production system. The papers by Jodlbauer
(2005) and Jodlbauer and Stöcher (2006) extend the framework for contin-
uous input and output functions in a single-stage production system.
Chapter 3
Utilisation Concept for Hierarchical
Production Planning
Finding the optimal capacity for the machinery is a standard goal in ac-
counting and finance. Opposing goals on each hierarchical planning level
lead to different utilisation levels. In this chapter, a utilisation concept is
developed in which the JIT goals are taken into account. The concept shows
that JIT capability enables the production system to reduce the excess ca-
pacity by disinvestment or to increase sales without additional investment.
It is shown that low utilisation at the short term level is positively cor-
related to JIT intensive practices. Consequently, low utilisation results in
potential higher profits. Furthermore, there is no reason to maximise the
short term machine utilisation, other than to sell more machine capacity in
the sense of selling more products or producing for external companies.
Zero Defects
No scrap losses and rework are acceptable. One method to approach ”zero
defects” is implementing TQM (Daugherty et al., 1994). The aim of TQM is
to continuously improve and sustain the quality of products and processes.
In addition, TQM has a strong customer as well as process orientation.
Rehder (1989) states that JIT and TQM are mutually supportive. Flynn
et al. (1995) empirically show that TQM practices improve JIT performance
and JIT practices have a positive effect on quality performance. According
to Deming (2000), TQM also reduces fluctuations in all system components.
This is considered to be a prerequisite for a successful implementation of
JIT.
Zero Set-Up
No changeovers are necessary or, at least, set-up times are vanishing. Small
lot sizes are only possible if the set-up times are very short. Finch (1986)
addresses a variety of techniques to reduce and simplify set-ups. One key
concept is to distinguish between internal and external set-up. Internal set-
up time causes an interruption of the production process while the external
set-up can be accomplished parallel to the production. Single Minute Ex-
change of Die (SMED) is one of the most widely known methods to reduce
the set-up time (Shingo, 1985). Combining ”zero lot size” and ”zero set-
ups” means that the total time used for changeovers decreases even though
the number of set-ups increases.
Zero Breakdown
No machine failures are acceptable to ensure maximum equipment avail-
ability (Daugherty et al., 1994). ”Zero breakdown” means that there is no
idle time caused by machine or tool failure. TPM is a method to max-
imise equipment effectiveness throughout its entire life (Nakajima, 1988).
Schonberger (1986) emphasises the importance of introducing JIT, TQM
and preventive maintenance practices simultaneously based on his concept
of world class manufacturing. Cua et al. (2001) conducted a survey of TQM,
JIT and TPM practices and their overlaps. In addition their study clearly
demonstrates the importance of joint implementation of JIT, TQM and
TPM.
Zero Handling
Unnecessary movements of the material or unnecessary motions conducted
by the machine operator is not acceptable according to Hines and Rich
(1997). The items go directly from one workstation to the next and double
handling or excessive movements should be avoided, according the pioneer
of TPM (Ohno, 1988). The goal ”zero handling” yields short processing
times as well as short transportation times and no queuing time between
two workstations.
3.2. Model 25
Zero Surging
Only small fluctuations in the production plan are acceptable for fulfilling
”zero surging”. In general a smooth final assembly or production schedule
(nearly constant production rate and production mix) is applied in JIT
systems. Otherwise, a high excess inventory or a high excess capacity is
needed.
Maybe the most challenging task for a JIT environment is to transform
high customer demand fluctuations into a smooth required production rate.
Some authors have developed methods to manage demand uncertainty or
demand peaks (Rees et al., 1987; Bartezzaghi and Verganti, 1995; Verganti,
1997; Co and Sharafali, 1997).
There is a trade-off between excess inventory and excess capacity (Bradley
and Glynn, 2002; Jodlbauer and Altendorfer, 2010). The general trend is
to prefer excess capacity to excess inventory, especially for various finished
goods (Schragenheim and Ronen, 1991). Thus, real JIT systems need some
excess capacity to cope with high demand fluctuations and to ensure contin-
uously high service levels. One way to manage demand peaks is to provide
extra capacity, such as an additional shift or overtime work. The classical
way of Toyota to manage the situation on which the production falls behind
the desired rate is two-shifting, whereby two shifts are scheduled per day
separated by down periods for catching-up the desired production rate, for
preventive maintenance or cross training (Schonberger, 1982).
3.2 Model
In this chapter, a single-machine multi-item production system is assumed.
The concept can be extended easily to a multi-machine setting. The sched-
uled working time T of the production system including preventive mainte-
nance is predefined as the maximum capacity available. Idle time tI repre-
sents machine breakdowns, waiting for an order, tool or worker unavailabil-
ity.
26 Chapter 3. Utilisation Concept
Symbol Description
U Machine utilisation (U < 1)
Umin Minimum machine utilisation (Umin < U )
EC Excess capacity
ECmax Maximum excess capacity (customer enabled excess capac-
ity)
IJIT JIT intensity (see equation 3.6)
tP,i Minimal possible processing time per item of production or-
der i
tS,i Set-up time per changeover of production order i
tAP,i Additional processing time per item of production order i
due to waste
tI Idle time
T Scheduled working time
nC,i Number of items needed for customer orders of production
order i without any production on stock
nS,i Number of items produced on stock of production order i
nQ,i Number of scrapped or obsolete items of production order i
nO Number of production orders
For the cost model, fix as well as variable costs are considered. Fix
costs include personnel costs for permanent workers and depreciation of
the equipment used. The variable costs are divided into variable processing
costs, variable material costs, holding costs and costs for rework or disposal.
The variable processing costs include variable costs for energy, spare parts,
operational supplements, expendables and set-ups and variable costs for
variable capacity (for instance: overtime, leasing workers). Variable mate-
3.2. Model 27
rial costs include the items to produce the parent item. Holding costs occur
for items which are on stock. If rework or disposal is necessary for items,
then costs for rework or disposal will occur. If improvements based on the
”seven zeros” are discussed, then the throughput has to be unchanged.
The total costs for the number of items produced on stock are assumed
to be higher than for the number of items produced for customer orders
because of the holding costs. Assuming enough excess capacity to meet
demand peaks, it is beneficial not to produce on stock because of lower costs
and earlier effective revenue. For the trade-off between excess capacity and
excess inventory, see for example Mieghem and Rudi (2002), Bradley and
Glynn (2002) or Jodlbauer and Altendorfer (2010).
3.2.1 Utilisation
Machine utilisation describes how intensively a machine is being used to
produce items. Thus, machine utilisation equals the ratio time used (run
time plus set-up time) to time available (Kuik and Tielemans, 1997; Cox
and Blackstone, 2002).
The time used is the sum of the run time for all produced items during
one time period (for instance one year) plus the set-up times for all pro-
duction orders during the same time period. All produced items consist of
items for customers, items to be produced on stock (these items go first on
stock and after some queuing time to the customer or are obsolete and are
fault free) and scrapped items.
The time available is the sum of all scheduled working time during the
time period. Summarising all these ideas yields the following relationship
for the utilisation:
n O
(tP,i + tAP,i )(nC,i + nS,i + nQ,i ) + tS,i
i=1
U := (3.1)
T
Idle time (no matter for what reason) of the equipment leads to a de-
crease of utilisation and idle time meets the non-negativity constraint. Since
100 percent utilisation equals to 100 percent minus the percentage of time
lost due to the unavailability of orders, machines and/or workers, the fol-
lowing relationship is introduced (Cox and Blackstone, 2002):
nO
tI := T − [(tP,i + tAP,i )(nC,i + nS,i + nQ,i ) + tS,i ] >= 0 (3.2)
i=1
28 Chapter 3. Utilisation Concept
Based on idle time and utilisation the excess capacity EC can be defined
as the ratio idle time to time available (scheduled working time):
n O
T− (tP,i + tAP,i )(nC,i + nS,i + nQ,i ) + tS,i
tI i=1
EC := = (3.3)
T T
In the following paragraphs, Equations 3.1 to 3.3 are analysed by iden-
tifying the effects on them when the ”seven zeros” methodology is applied.
By achieving ”zero defects”, scrap and rework should be avoided and there-
fore the scrapped items nQ,i are reduced whereby nC,i and nS,i remain
unaffected. This results in a decrease of utilisation.
The two JIT goals ”zero lot size” and ”zero set-ups” are considered to-
gether because changing the lot size influences the amount of set-ups imme-
diately. ”Zero lot size” influences the parameter number of production or-
ders nO because the higher nO the more customer lot sizes are applied in the
production system. ”Zero set-ups” affects the set-up time per changeover
tS,i because it
is assumed that a set-up is necessary after each production
nO
order. Hence, i=1 tS,i is minimised by applying these two ”seven zeros”
and results in a decreased utilisation.
”Zero handling” minimises the additional processing time tAP,i by avoid-
ing unnecessary movements, performance losses and transportations. The
application of ”zero handling” results in a decreased utilisation as well.
According to Little’s Law (Little, 1961), ”zero lead time” influences all
unnecessary inventories such as number of scraped items nQ,i and number
of items produced on stock nS,i . ”Zero lead time” results in a reduction of
unnecessary inventories and reduces the metrics utilisation.
”Zero breakdown” reduces the idle time tI occurred by broken equip-
ment. However, this gain is transferred to idle time due to waiting for order,
since it is assumed that the same numbers of items needed for customer or-
ders are produced. Therefore, this ”zero” has no immediate consequence on
the utilisation metrics. But it allows potential for reducing the scheduled
working time T or the sales department to increase the customer orders
nC,i .
If the JIT goal ”zero surging” is focused, then the idle time tI is reduced
because a production plan with low fluctuation is applied to avoid over and
under load. Again, no immediate consequence on the utilisation metrics
occurs but there is potential for reducing the scheduled working time T or
for the sales department to increase the customer orders nC,i .
All ”seven zeros” themselves result in an increase of idle time tI and
lead to a decreased or unchanged utilisation metrics (see Table 3.2 column
U ). However, all ”seven zeros” cause ”further potential”, which can be
(I) a reduction of the scheduled working time (T ↓) or (II) an increased
3.2. Model 29
production of more customer required items (see Table 3.2 column ”Further
potential”). Even for ”zero breakdown” and ”zero surging” this further
potential occurs since the idle time can be used for producing more customer
required products or reduced after their implementation. The effects of the
”seven zeros” as JIT goal on utilisation are defined in Equation 3.1 and 3.2
and are summarised in Table 3.2. In this table, the ”seven zeros” and the
respective parameters influenced are listed.
n O
Zero lot size tS,i ,nO min tS,i dec.a yes
tS,i ,nO i=1
combined with
zero set-ups
n O
min tAP,i (nC,i
Zero handling tAP,i tAP,i i=1 dec.a yes
+nS,i + nQ,i )
n O
min (tP,i + tAP,i )
Zero lead nS,i ,nQ,i nS,i ,nQ,i i=1 dec.a yes
time ·(nS,i + nQ,i )
n O
tP,i nC,i
i=1
Umin := (3.4)
T
The minimum utilisation is the customer required utilisation without
capacity wasted for not optimised processing times, scrapped items, over-
production, set-ups or idle times. Moreover, the minimum utilisation (Equa-
tion 3.4) is less than the utilisation U (Equation 3.1). As discussed earlier
in this section, ”zero breakdown” and ”zero surging” have no influence on
utilisation. Therefore, the maximum excess capacity ECmax is introduced
in Equation 3.5 based on Equation 3.1 and 3.2. Excess capacity, which has
been defined as idle time tI , is according to Cox and Blackstone (2002)
the capacity that is not used to either produce or protect the creation of
throughput.
n O
T− tP,i nC,i
i=1
ECmax := (3.5)
T
ECmax can be interpreted as the customer enabled excess capacity for
producing the customer required items nc,i without any waste. If all wastes
based on the ”seven zeros” are avoided in production two scenarios are
possible: (I) divestment of ECmax since it is not used or (II) sales increase
by producing more customer required products so that ECmax equals to
zero.
Managerial Insight
Excess capacity increases or utilisation decreases by applying all ”seven
zeros” to achieve the customer enabled excess capacity ECmax or minimum
utilisation Umin respectively. The developed metrics JIT intensity provides
the management with a metrics in how well the ”seven zeros” are achieved,
since JIT activities have a positive impact on performance. Management
has two possibilities to deal with the customer required excess capacity: (I)
divestment to reduce the scheduled working time or (II) increase sales by
producing more customer required items.
3.3 Discussion
The influence of the opposing hierarchical goals on utilisation is discussed
for the three planning levels.
1. Minimising the variable costs for workers (leasing workers and over-
time) by fulfilling customer requirements
2. Inventory increases or more idle time for permanent workers
variable processing costs are increasing. Moreover, holding costs or costs for
”rework or disposals” are increasing respectively. Items produced on stock
can be sold later as soon as customer orders arrive. Since there is a risk
for obsolescence, the increase of revenue is not guaranteed. By producing
scrap items, the risk for violating due dates is high. Therefore, a risk for a
decrease in revenue exists.
By increasing the total time needed for set-up, the total set-up costs
might increase when additional capacity is needed or when it is not possible
to produce all items needed for customer orders nC,i . By increasing the total
time needed for set-up, the risk for violating due dates is high. Therefore,
a risk for a decrease in revenue exists.
By increasing the ”additional processing time” tAP,i , the total variable
costs might increase when additional capacity is needed or when it is not
possible to produce all items needed for customer orders nC,i . By increasing
the processing time, the risk for violating due dates is high and therefore
carries a risk for a decrease in revenue.
Increasing the idle time ti means that the time for machine breakdowns,
waiting for an order or tool or worker unavailability increases. By increasing
the idle time, the risk for violating due dates is high. This results again in
a risk for a decrease in revenue.
Only the parameter ”numbers of items produced for customer orders”
nC,i leads to an increase of revenues as indicated in Table 3.3. Also the
increase of ”number of items produced on stock” nS,i accounts for potential
revenues only in future. All parameters increase the total variable costs
whereby tI and tAP,i only increase the total variable costs if additional
capacity is required. The expression ”possible decrease of revenue” in Ta-
ble 3.3 means that due to lost capacity caused by quality problems, too
long set-up times, or unavailability and/or ”additional processing times”,
customer due dates may not be met. Missed due dates cause customer dis-
satisfaction and may lead to reduced revenue. Summarising the short term
view, the following statements are equivalent:
1. Maximising profit (maximising revenue with minimal variable material,
processing, holding, rework, disposal and change over costs)
2. Minimising number of items produced on stock nS,i , number of scrapped
items nQ,i , total time needed for change-overs and idle time tI and
additional processing time tAP I,i
3. Minimising the machine utilisation to the minimum utilisation Umin
4. Maximising excess capacity
In the short term view, minimising the utilisation means producing effi-
ciently because losses according to the ”seven zeros” philosophy are avoided.
34 Chapter 3. Utilisation Concept
Table 3.3: Effects of parameters on total variable costs and revenue by short
term decisions
Managerial Insight
Combining the JIT philosophy, utilisation concep,t and costs and revenue,
two interesting strategies arise. If a high JIT intensity is reached, this means
that the production system can follow the fluctuations of the customer de-
mand without any losses according to the ”seven zeros”. Thus it is possible
to reduce the excess machine capacity as well as the flexibility needed for
medium term capacity adjustments. A high JIT intensity allows lower ca-
pacity invested and requires less working time flexibility. Alternatively, it
is possible to increase revenues by increasing sales, whereas no additional
investment is necessary.
respectively between capacity invested and potential for meeting sales re-
quirements in high fluctuation markets. Total variable costs based on short
term decisions are decreased with low utilisation.
Summarising, the production view in terms of low utilisation or high
excess capacity by fulfilling the seven JIT goals means efficient production.
Chapter 4
Capacity Setting Methods
In this chapter different methods for an MTO environment for a periodi-
cal capacity setting, multi-machine production systems with stochastic cus-
tomer required lead times and stochastic processing times to improve service
level and tardiness are presented. These methods are developed as decision
support when capacity flexibility exists, such as a certain range of possible
working hours a week. The methods differ in the amount of information used
whereby all are based on the cumulated capacity demand at each machine.
In a simulation study the methods’ impact on service level and tardiness is
compared to a constant capacity provided for a single- and a multi-machine
setting. It is shown that the capacity setting methods tested can lead to an
increase in service level and a decrease in average tardiness contrasted to a
constant capacity provided. The methods using information on processing
time and customer required lead time distribution perform best. In the
end of this chapter for a simplified M/M/1 queuing system two different
production rates are introduced. The results found in this chapter can help
practitioners to make efficient use of their flexible capacity.
ily changeable for companies due to, for instance, laws limiting working
times, contract workers and investment/de-investment in machinery and
equipment. These tools allow companies to deal with customer demand
uncertainty by adjusting their capacities. Furthermore, capacity cannot be
adjusted instantaneously, but a periodic capacity adjustment is realistic.
This reflects the situation of a rolling horizon planning in real production
systems where capacity can, for example, be defined for the next week on a
daily basis taking into account the currently known orders.
Based on the assumption that processing time distribution and customer
required lead time distribution are known, these two sources of information
are used to develop different capacity setting methods which are tested in
a discrete event simulation study. Since many manufacturing companies
are dealing with bottlenecks in the production process, a single-machine
framework is considered in this chapter, which is finally extended to a multi-
machine setting.
In this chapter a periodical decision support for short and medium
term capacity setting to improve service level and tardiness is developed.
Stochastic customer behaviour (information uncertainty: not all informa-
tion is available when the decision is taken), the integration of stochastic
production processes (process uncertainty: the processing time of orders is
not deterministic) and the rolling horizon effects of a planning system (re-
evaluation effects: decisions taken in the past influence the current decision
space and not the full length of the decided plan) are considered by the
different methods.
The cumulated capacity concept used by Segerstedt (1996) is applied.
The M/M/1 queuing state dependent capacity adjustment models of Mincso-
vics and Dellaert (2009) and Buyukkaramikli et al. (2013) are extended by
general distributions for the arrival process and processing times. Moreover,
the capacity setting methods are extended to a multi-machine production
system. The proposed framework also covers a distribution of the customer
required lead time. Therefore, the model of Jodlbauer (2008c) is extended
by a rolling horizon and is implemented for calculating the capacity pro-
vided. A limitation of the approaches presented is that no analytical eval-
uation of their performance is possible and therefore simulation has to be
applied. After determining the capacity provided, in a next step a schedul-
ing approach could be considered which is left for further research.
1 3
no proc dist no proc dist
no CRL dist CRL dist
2 4
proc dist proc dist
no CRL dist CRL dist
Figure 4.1: Matrix of stochastic information for demanded capacity – four meth-
ods
Figure 4.2 shows the basic model where order i with the due date Di is
stated by a customer. The customer required lead time represents the time
gap between the point of time when the order is stated and the due date of
the customer order. The production system can either fulfil the customer
order on time or the customer order is delayed. If an order is accomplished
earlier, then it is considered to be on time. This circumstance is measured
by the two metrics service level and tardiness. No lost sales are considered
in this model. If the machine is idle, the order is released into production;
otherwise, the order is waiting in the buffer for processing. Only one order
can be processed at a time. Customer required lead times and processing
times are assumed to be stochastic.
The production system is assumed to work on an MTO basis. This
means production orders are created based on already known demand and
not based on forecast. Only the capacity demand is forecasted in some of
the following methods using the distribution of customer required lead time.
40 Chapter 4. Capacity Setting Methods
4.3 Model
In Table 4.1 the variables and in Table 4.2 the parameters for the single-
machine production system are presented.
To fulfil all orders on time, the demanded cumulated capacity ãx (t) has
to be lower than or equal to the cumulated provided capacity cy t at all
points in time as illustrated in Figure 4.3. A combination of (x, y) (see
Table 4.2) is defined as periodical capacity setting method, whereby x and
y indicate the methods used for capacity demanded respectively provided.
Notation Description
Di Random variable for due date of order i
D̂i Random variable for pushed up due date of order i
Ai Random variable for demanded capacity of order i ; A0 is the
current backlog
Ã(t) Random variable for cumulated demanded capacity until
time t
Â(t) Random variable for pushed up cumulated demanded capac-
ity until time t
L Random variable for capacity weighted customer required
lead time of all orders
E[.] Expected value of a random variable
V ar[.] Variance of a random variable
FG (.) CDF (cumulated distribution function) of random variable
G
FG−1 Inverse of the CDF of random variable G
ãx (t) Variable for approximating cumulated demanded capacity at
the machine at time t based on method x
cy Variable of bounded provided capacity after each capacity
setting cycle based on method y
c̃y Variable of unbounded provided capacity after each capacity
setting cycle based on method y
c(l) Variable for capacity provided in past period l
ĉ Variable for capacity account
Notation Description
x Method parameter of cumulated demanded capacity:
x = 1: Deterministic processing times; no CRL distrib.
x = 2: Processing time distribution; no CRL distrib.
x = 3: Deterministic processing times; CRL distribution
x = 4: Processing time and CRL distributions
y Method parameter of provided capacity:
y = 1: Full utilisation
y = 2: Maximum safety
y = 3: Service level target
κmin Parameters for lower bound for provided capacity cy
κmax Parameters for upper bound for provided capacity cy
κavg Parameter for average provided capacity per period
Δ Parameter for period of time for which cy is set
−ψ Parameter for time where the system has (re)started
θ Service level parameter for capacity setting method c3
β Capacity feasibility parameter for capacity setting methods
a2 and a4
OC(t) Operation characteristics of the machine until time t
δ Parameter for evaluation time window
' G
Ã(t) = Ai (4.2)
{i|Di ≤t}
Equation 4.6 shows how the due date of each order based on the OC
is pushed up to ensure some safety capacity for the anticipated demand
(therefore D̂i ≤ Di holds).
Di
D̂i = OC(τ )dτ (4.6)
0
−1
ã4 (t) = FÂ(t) (β) (4.7)
Figure 4.7: Comparison of the three developed methods calculating the capacity
provided
Full Utilisation: y = 1
For this capacity setting method the capacity provided is set to the average
capacity demanded within an evaluation time window δ. Therefore, the
cumulated demanded capacity at the end of the evaluation time window
is divided by the time where the last customer order reaches its due date.
48 Chapter 4. Capacity Setting Methods
The maximum over all i is responsible for searching the due date of the
last order within the evaluation time window.
Maximum Safety: y = 2
Method y = 2 works as follows: A straight line starting at zero with mini-
mum slope being greater than or equal to the cumulated demanded capacity
is constructed. The fraction in Equation 4.10 can be interpreted as the slope
of cumulated capacity demand at due date Di . The max operator identifies
the maximum slope over all due dates Di leading to a maximum safety in
capacity provided. As an assumption at least one order has to be in the
system. Note that capacity backlogs are ignored since they would lead to
an infinite increase, however, they are included in ãx (Di ) when the first due
date is evaluated.
ãx (Di )
c̃2 = max (4.10)
{i|Di >0} Di
f (Di , c̃3 )
{Di |Di ≤δ}
≥θ
1
{Di |Di ≤δ}
4.3. Model 49
The function f (.) defined in Equation 4.12 delivers 1 if a due date (one
or more orders can have the same due date) is on time and 0 if a certain
due date does not hold. c̃δ3 is the slope of the capacity provided.
c̃3
f (Di , c̃3 ) = 1 for ãx (Di ) ≤ Di (on time) (4.12)
δ
c̃3
f (Di , c̃3 ) = 0 for ãx (Di ) > Di (late)
δ
If y = 3 is applied in combination with methods x = 3 and x = 4, Di has
to be substituted by D̂i for Equations 4.11 to 4.12, because the customer
required lead time distribution is included in the approximations.
cy − κavg ≤ ĉ (4.14)
−1
ĉ = (κavg − c(l))
l=−ψ
The candidate for capacity provided c˜y from methods y = 1...3 is con-
strained by Equation 4.13 and 4.14 in the following Equation 4.15.
50 Chapter 4. Capacity Setting Methods
−1
κmin ≤ cy ≤ min[κmax , κavg + (κavg − c(l))] (4.15)
l=−ψ
Variable Description
Di,j Random variable for machine depended due date of order i at
machine j. Di,j on the last machine equals to Di stated by
the customer
D̂i,j Random variable for pushed up due date of order i at machine
j
Ri,j Random variable of remaining capacity demand for order i
after machine j
Ni,j Set of machines for remaining processing steps of order i after
machine j
Ai,j Random variable for demanded capacity of order i at machine
j
Lj Random variable for capacity weighted customer required
lead time at machine j
OCj (t) Operation characteristics of machine j until time t
Ri,j = Ai,k (4.17)
{k∈Ni,j }
R̃i,j = FR−1
i,j
(β) (4.18)
Comparing the sum over the means of the capacity provided in Equa-
tion 4.17 with the inverse of the CDF of the processing time distribution
for the remaining processing steps of an order in Equation 4.18 leads to
a difference which is equivalent to the safety time concept also applied in
MRP. This safety time accounts for the possibility that a certain order has
a higher processing time at a certain machine than planned. However, the
approach presented in Equation 4.18 does not consider the queue length
as discussed for example in the Workload Control (Bechte, 1988) literature
where also the queuing between the processing steps is considered. Com-
pared to the safety time concept in MRP, which also includes some safety
52 Chapter 4. Capacity Setting Methods
time for queuing, this safety time only includes the processing instabilities.
This is assumed since only the workloads and their latest possible date Di,j
for each single machine are needed but no production plan with start dates
is generated.
Lj is calculated in Equation 4.19 based on L. For simplification reasons
in the calculation of Lj , all Ai,j are assumed to have a common stochastic
distribution and therefore E[Rj ] = E[Ri,j ] holds.
Lj = L − E[Rj ] (4.19)
Note that this assumption only influences the OCj (t) applied in the
methods x = 3, x = 4a and x = 4b.
The planned utilisation is set to the same level for all machines. The
following values for planned utilisation are tested: 0.7, 0.75, 0.8, 0.85, 0.875,
0.9, 0.91, 0.92, 0.93, 0.94, 0.95, 0.96, 0.97, 0.98, 0.99.
Ten different scenarios for single- and multi-machine settings are tested
where parameters for processing time, customer required lead time, upper
and lower bound for provided capacity and period for setting the provided
capacity are predefined as shown in Table 4.5. A basic scenario is defined as
the basis for all other scenarios with a variation coefficient for the processing
time distribution of 1, mean customer required lead time of 20 time periods
for the single-machine production system and 100 time units for the multi-
machine production system. The upper and lower bound for the capacity
provided is given by +/- 20% and Δ is set to 10 periods.
V ar[Ai ]
Description E[Ai ] E[L] κmin , κmax Δ
basic 1 20/100 0.8, 1.2 10
high processing time variation 1.5 20/100 0.8, 1.2 10
low processing time variation 0.5 20/100 0.8, 1.2 10
high customer required lead time 1 50/250 0.8, 1.2 10
low customer required lead time 1 10/50 0.8, 1.2 10
high flexibility 1 20/100 0.5, 1.5 10
low flexibility 1 20/100 0.95, 1.05 10
indefinite flexibility 1 20/100 0, ∞ 10
long capacity setting period 1 20/100 0.8, 1.2 50
short capacity setting period 1 20/100 0.8, 1.2 5
The capacity setting methods for the capacity provided are implemented
in the simulation model by changing the machine speed. Whenever the
average machine speed (which is equivalent to the capacity account) is below
or equal to 1, the provided capacity can be set – if necessary – higher than
1. The average machine speed is cleared at each system restart. Relating
the simulation periods to one shift of a working day and assuming that two
shifts per day make the working time for five days a week, then the capacity
is set every week (Δ = 10 shifts). The capacity account in the beginning is
0. The simulation is conducted in AnyLogic 6.4.1.
4.6. Results 55
4.6 Results
In the first step, the best capacity setting method (x, y) for the basic sce-
nario is evaluated. Based on this evaluation, the results of the best method
(x, y) are compared to the situation with constant provided capacity for the
two production systems. This is followed by an analysis over all scenarios
defined. For all investigations, the performance measures service level in-
crease and tardiness decrease in comparison to average capacity provided
are used.
Table 4.6: Service level increase and tardiness decrease potential in basic scenario
machine production system. This circumstance may result from the higher
complexity and higher uncertainty of the multi-machine production system
based on the stochastic compensation effect.
(see Table 4.5), the reason for the low performance in the multi-machine
production system is conjectured to be an increasing demand uncertainty
with long periods Δ, for which capacity has to be set. Taking the example
introduced above into account, production systems with Δ1 waste capacity
of area 1 if no additional orders arrive until the next capacity setting. The
production system with Δ2 , however, wastes capacity of area 1 and 2 if no
orders arrive until the next capacity setting. Since it is necessary to fulfil
Equation 4.15 too much wasted capacity in some periods leads to a reduced
speed of the equipment in later periods.
Figure 4.10 illustrates an example of two production systems at a cer-
tain point in time of the timeline (see also Figure 4.4). In both cases the
production system is capable to satisfy cy . Hence, they differ from capacity
setting period Δ, whereby Δ1 assumes a short capacity setting period and
Δ2 a long capacity setting period.
A further result based on Table 4.8 is that for any change in capacity
flexibility the performance on both scenario types is worse than in the basic
scenario. A too high range of capacity flexibility can lead to overreactions
60 Chapter 4. Capacity Setting Methods
' '
The production system can either fulfil the customer order on time or
the customer order is delayed. If an order is accomplished earlier, then it
is considered to be on time and stored in buffer FGI. This circumstance is
measured by the two metrics service level s and tardiness C. If the machine
is idle, the order is released into production; otherwise, the order is waiting
in the buffer for processing. Only one order can be processed at a time. The
production unit consists of a buffer for waiting orders and a processing step
that requires a random processing time. The random variable of production
lead time needed for one order from release until its completion is denoted
by W , and fW (.) denotes the PDF of the production lead time after release
to completion.
Whenever more than k orders are in the WIP of the production system,
it runs at speed μH ; otherwise, μL is used. Let Y denote the random WIP
and let G denote the random finished goods inventory. Backorders B are
charged when orders are late.
62 Chapter 4. Capacity Setting Methods
4.7.1 Assumptions
The following assumptions are taken into consideration to create the model:
• M/M/1 model is assumed with two levels of μ → {μL , μH }
• Processing times are exponentially distributed and only one order can
be processed at time. Whenever more than k orders are in the WIP,
μH , otherwise μL is used as production rate and no transient be-
haviour at the switching point is assumed
• Customer orders for single items are stated piecewise from many differ-
ent customers which supports the exponentially distributed customer
order interarrival times with rate λ for single items of order i
• Customer required lead time L is exponentially distribut-ed with its
parameter β
• Customer orders are not changed once they are stated and have be
accepted and released into the production system
• Customer required lead time cannot be influenced by the production
system
• Nothing is produced without a customer order (MTO system)
• FIFO is applied as dispatching rule after the customer orders are re-
leased
• Finished items are delivered to the FGI after processing
• Machine capacity cannot be stored
λ
P0 λ = P1 μL ⇒ P1 = P0 (4.20)
μL
P1 λ + P1 μL = P2 μL + P0 λ ⇒ P1 λ = P2 μL
2
λ λ
⇒ P2 = P1 = P0 (4.21)
μL μL
P2 λ + P2 μL = P3 μH + P1 λ ⇒ P2 λ = P3 μH
2
λ λ λ
⇒ P3 = P2 = P0 (4.22)
μH μH μL
2 2
λ λ λ
⇒ P4 = P3 = P0 (4.23)
μH μL μH
4.7. Extension 63
∞
∞
min(k,n) max(0,n−k)
λ λ
Pn = 1 ⇔ P0 =1 (4.25)
n=0 n=0
μL μH
μkL (μL − λ) (μH − λ)
⇔ P0 = (4.26)
μk+1
L (μH − λ) − λk+1 (μH − μL )
for n ≤k
n
λ
⇒ Pn (n ≤ k) =P0
μL
n
μL k (μL − λ) (μH − λ) λ
=
μL k+1 (μH − λ) − λk+1 (μH − μL ) μL
μL k−n λn (μL − λ) (μH − λ)
= (4.27)
μL k+1 (μH − λ) − λk+1 (μH − μL )
for n >k
k n−k
λ λ
⇒ Pn (n > k) =P0
μL μH
μL k (μL − λ) (μH − λ)
=
μL k+1 (μH − λ) − λk+1 (μH − μL )
k n−k
λ λ
μL μH
λn (μL − λ) (μH − λ)
= n−k
μH (μL k+1 (μH − λ) − λk+1 (μH − μL ))
n
λ μH k (μL − λ) (μH − λ)
= (4.28)
μH μL k+1 (μH − λ) − λk+1 (μH − μL )
64 Chapter 4. Capacity Setting Methods
k
k n
λ
PH = 1 − Pn = 1 − P0
n=0 n=0
μL
μL k+1 − λ k+1
= 1 − P0
μL k (μL − λ)
k+1
1
fW (t) ≈
1 − P0
k ∞
Pn (n ≤ k) fWn (t) + Pn (n > k) fWn (t) (4.30)
n=1 n=k+1
(μL − λ) (μH − λ)
=
μkL (μH − λ) − λk (μH − μL )
−(μL −λ)t k Γ (k, λt) −(μH −λ)t k Γ (k, λt)
e μL +e μH 1 − (4.31)
Γ (k) Γ (k)
4.7. Extension 65
∞
E W̃ = fW (τ ) τ dτ
0
(μL − λ) (μH − λ)
=
μL k (μH − λ) − λk (μH − μL )
μ L k − λk λk k
−
(μL − λ)
2 μL (μL − λ)
λk λk k
+ + (4.32)
(μH − λ)
2 μH (μH − λ)
The exact value of the expected WIP (where N is a sufficiently high
value to reach the desired precision) can be calculated by summing up all
probability states available based on Medhi (1991):
∞
N
E [Y ] = nPn ≈ nPn (4.33)
n=0 n=0
An evaluation of the approximation error from the approximated PDF
is possible
as shown
in Figure 4.12, applying Little’s Law to Equation 4.32
for E Ỹ = λE W̃ .
The results from Figure 4.12 show that the expected WIP has an error
of up to 7%, however, the average error is rather small and therefore this
approximation is applied in the remainder. μ presents the amount of the
flexible capacity μH = μL + Δμ.
Based on Equation 4.32 also the CDF of the approximated production
lead time can be calculated as:
τ
FW (τ ) = fW (t) dt = νL ϑL (τ ) − νH ϑH (τ ) (4.34)
0
66 Chapter 4. Capacity Setting Methods
˜
=1, L =0.9 ˜
=1, L =1.1
(EY-EY)/EY (EY-EY)/EY
0.10 0.10
0.05 0.05
0.00 0.00
0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8
k=0 k=0
-0.05 k=1 -0.05 k=1
k=5 k=5
k=20 k=20
-0.10 -0.10
Figure 4.12: Approximation error of the PDF of the production lead time
X
E [G] = λ FW (τ ) (1 − FL (τ )) dτ (4.35)
0
νL λk+1 μk−1 Γ (k, X (μL + β))
E [G] = L
k
1−
(μL + β) Γ (k)
k −(μL +β)X k
νL λμL e (Xλ)
− + e−βX
μL − λ + β Γ (k) X (μL + β)
1 − e−βX (μL − λ) λk Γ (k − 1, X (μL + β))
− − 1 −
β (μL + β)
k Γ (k − 1)
νL λk+1 e−(μL +β)X (XμL )
k−1
Γ (k, Xλ)
−e−(μL −λ+β)X +
Γ (k) β Γ (k − 1) X (μL + β)
Γ (k − 1, XμL )
+e−βX 1 −
Γ (k − 1)
μLk−1
Γ (k − 2, X (μL + β))
− 1−
(μL + β)
k−1 Γ (k − 2)
e−(μH +β)X (Xλ)
k
νH λμkH
+
μH − λ + β Γ (k) X (μH + β)
λk Γ (k − 1, X (μH + β))
− 1−
(μH + β)
k Γ (k − 1)
Γ (k, Xλ)
+e−(μH −λ+β)X 1 −
Γ (k)
k+1 −(μH +β)X k−1
νH λ e (XμH )
−
β Γ (k − 1) X (μH + β)
k−1
βμH Γ (k, X (μH + β))
+ k
1 −
(μH + β) Γ (k)
Γ (k − 1, XμH )
+e−βX 1 −
Γ (k − 1)
μHk−1
Γ (k − 2, X (μH + β))
− 1− (4.36)
(μH + β)
k−1 Γ (k − 2)
Note that based on this solution for E[G] (see last term of Equation 4.36)
only k values greater than 2 can be studied, which means the flexible ca-
pacity cannot be applied when 1, 2 or 3 pieces are in the system.
68 Chapter 4. Capacity Setting Methods
X
E [B] = E [G] + E Ỹ − λ (1 − FL (τ )) dτ
0
1 − e−βX
= E [G] + E Ỹ − λ (4.37)
β
X ∞
s= FW (τ ) fL (τ ) dτ + FW (X) fL (τ ) dτ
0 X
X
= FW (τ ) fL (τ ) dτ + FW (X) (1 − FL (τ )) (4.38)
0
λk μk−1
L β Γ (k, X (μL + β))
s =νL k
1−
(β + μL ) Γ (k)
λk μkL β e−(μL +β)X X k−1 1
− + k
μL − λ + β Γ (k) (μL + β) λ
1 Γ (k − 1, X (μL + β))
− 1− + μkL
(μL + β)
k−1 Γ (k − 1)
−(μL +β)X k−1
k k−1 e X e−(μL +β)X X k−2
+ λ μL +
Γ (k) Γ (k − 1) (μL + β)
1 Γ (k − 2, X (μL + β))
− 1−
(μL + β)
k−1 Γ (k − 2)
4.7. Extension 69
μkL (μL − λ) e−(μL −λ+β)X Γ (k, Xλ)
−
μL − λ + β Γ (k)
k k−1
λ μH β Γ (k, X (μH + β))
− νH k
1 −
(μH + β) Γ (k)
λk μkH β e−(μH +β)X X k−1 e−(μH −λ+β)X
− +
μH − λ + β Γ (k) (μH + β) λk
1 Γ (k − 1, X (μH + β))
− 1−
(μH + β)
k Γ (k − 1)
−(μH +β)X k−2
e X e−(μH +β)X X k−1
+ λk μk−1 +
H
Γ (k − 1) (μH + β) Γ (k)
1 Γ (k − 2, X (μH + β))
− 1−
(μH + β)
k−1 Γ (k − 2)
μkH e−(μH −λ+β)X (μH − λ) Γ (k, Xλ)
+ μkH e−(μH −λ+β)X − (4.39)
μH − λ + β Γ (k)
holds true. λ (arrival rate) is set to 1 and β (parameter for customer required
lead time distribution) is set to 0.05 respectively. Backorders are charged a
backorder cost b = 20 currency units per unit of time an order is late. The
holdings costs are set to h = 0.5 for WIP and f = 1 for FGI per order per
time unit.
For the following investigations two scenarios for the switching point
k = {5, 10} are analysed and compared to a constant provided capacity
situation, whereby the optimal WAW X based on Altendorfer and Minner
(2011) is applied. The results are presented in Figure 4.13. Each of these
charts shows the influence on the particular logistic key figure when (I)
flexible capacity is used, (II) maximum capacity μH is always applied and
(III) a constant capacity situation where the average of the flexible capacity
situation is provided.
It is desirable that the results of the flexible capacity setting are close the
maximum capacity provided situation. By increasing k the gap increases
70 Chapter 4. Capacity Setting Methods
but the flexible capacity situation has in most cases a better performance
than the constant provided capacity scenario.
When the gap between μL and μH is too intense (key performance in-
dicators service level decreases, backorders and FGI increases respectively),
then overreactions occur by providing too much capacity as explained in
Figure 4.10 for capacity flexibility.
In further reserach the developed equations in this section are used to
optimise (minimise costs) the production system parameters μL , μH , k and
X.
=1, average =1, k=5, b=20, f=1, h=0.5, =0.05 =1, average =1, k=10, b=20, f=1, h=0.5, =0.05
SL SL
1.0 1.0
0.8 0.8
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
=1, average =1, k=5, b=20, f=1, h=0.5, =0.05 =1, average =1, k=10, b=20, f=1, h=0.5, =0.05
E[B] E[B]
50 50
40 40
30 30
0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
=1, average =1, k=5, b=20, f=1, h=0.5, =0.05 =1, average =1, k=10, b=20, f=1, h=0.5, =0.05
E[G] E[G]
10 10
Constant capacity
8 8 Flexible capacity
Maximum capacity
6 6
4 4
Constant capacity
2 Flexible capacity 2
Maximum capacity
0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
(4.40)
∞
Pn =1
n=0
∞ min(k,n)
max(0,n−k)
λ λ
⇔ P0 =1
n=0
μL μH
λ
k n k ∞ n−k
λ λ
⇔P0 + P0 =1
n=0
μL μL μH
n=k+1
k n k ∞ n
λ λ λ λ
⇔P0 + P0 =1
n=0
μL μL μH n=0 μH
λ
k n k
λ λ μH
⇔P0 + P0 =1
n=0
μ L μ L μ H μ H −λ
k
λ μλL − μL k
λ λ
⇔P0 + P0 =1
λ − μL μL μH − λ
k k
μL λ λ λ λ
⇔P0 − P0 + P0 =1
μL − λ μL μL − λ μL μH − λ
k
μL λ λ λ
⇔P0 + P0 − =1
μL − λ μL μH − λ μL − λ
μL λk λ (μH − μL )
⇔P0 − P0 k =1
μL − λ μL (μL − λ) (μH − λ)
μL (μH − λ)
k+1
λk+1 (μH − μL )
⇔P0 − =1
μkL (μL − λ) (μH − λ) μkL (μL − λ) (μH − λ)
μk+1 (μH − λ) − λk+1 (μH − μL )
⇔P0 L
=1
μkL (μL − λ) (μH − λ)
μkL (μL − λ) (μH − λ)
⇔ = P0
μk+1
L (μH − λ) − λk+1 (μH − μL )
4.9. Appendix 73
∞
1
k
fW (t) ≈ Pn (n ≤ k) fWn (t) + Pn (n > k) fWn (t)
1 − P0 n=1 n=k+1
1
k
μnL tn−1
= Pn (n ≤ k) e−μL t
1 − P0 n=1
(n − 1)!
∞
n n−1
−μH t μH t
+ Pn (n > k) e
(n − 1)!
n=k+1
k
1 λ
n
μn tn−1
= P0 e−μL t L
1 − P0 n=1 μL (n − 1)!
∞
λ
k
λ
n−k n n−1
−μH t μH t
+ P0 e
μL μH (n − 1)!
n=k+1
(μL − λ) (μH − λ)
=
k
λ (μH − λ) − (λ/μL ) (μH − μL )
k
λ n μn tn−1
e−μL t L
n=1
μL (n − 1)!
∞ n−k
λ
k
λ n n−1
−μH t μH t
+ e
μL μH (n − 1)!
n=k+1
(μL − λ) (μH − λ)
=
k
λ (μH − λ) − (λ/μL ) (μH − μL )
k ∞
n n−1 λn−k λk −μH t μnH tn−1
−μL t λ t
e + e
(n − 1)! μn−k μkL (n − 1)!
n=1 n=k+1 H
(μL − λ) (μH − λ)
=
k
λ (μH − λ) − (λ/μL ) (μH − μL )
∞
k
λn tn−1 μkH −μH t λn tn−1
−μL t
e + ke
n=1
(n − 1)! μL (n − 1)!
n=k+1
(μL − λ) (μH − λ)
=
k
λμkL (μH − λ) − (λ/μL ) (μH − μL )
74 Chapter 4. Capacity Setting Methods
∞
k
λn tn−1 λn tn−1
μkL e−μL t + μkH e−μH t
n=1
(n − 1)! (n − 1)!
n=k+1
(μL − λ) (μH − λ)
= k
μL (μH − λ) − λk (μH − μL )
Γ (k, λt) Γ (k, λt)
e−(μL −λ)t μkL + e−(μH −λ)t μkH 1 − (4.41)
Γ (k) Γ (k)
P0
Auxiliary calculation for 1−P0 is presented as:
μk
L (μL −λ)(μH −λ)
P0 μk+1 (μH −λ)−λk+1 (μH −μL )
= Lk
1 − P0 (μL λ)(μH −λ)−λk+1 (μH −μL )
μk+1
L (μH −λ)−λk+1 (μH −μL )
(μL − λ) (μH − λ)
= (4.42)
k
λ (μH − λ) − (λ/μL ) (μH − μL )
∞
E W̃ = fW (τ ) τ dτ
0
∞
(μL − λ) (μH − λ) Γ (k, λτ )
= e−(μL −λ)τ μL k
μL k (μH − λ) − λk (μH − μL ) Γ (k)
0
Γ (k, λτ )
+e−(μH −λ)τ μH k 1 − τ dτ
Γ (k)
⎛
∞
(μL − λ) (μH − λ) ⎝ Γ (k, λτ )
= k μL k
e−(μL −λ)τ τ dτ
μL (μH − λ) − λk (μH − μL ) Γ (k)
0
⎞
∞ ∞
Γ (k, λτ )
+ μH k e−(μH −λ)τ τ dτ −μH k e−(μH −λ)τ τ dτ ⎠
Γ (k)
0 0
4.9. Appendix 75
(μL − λ) (μH − λ)
=
μL k (μH − λ) − λk (μH − μL )
⎛ k
μL k μL −k λk μL μλLk − 1 − k (μL − λ)
⎝
2
μL (λ − μL )
k ⎞
μH k μH k μH −k λk μH μλHk − 1 − k (μH − λ)
⎠
+ 2 − 2
(λ − μH ) μH (λ − μH )
(μL − λ) (μH − λ)
=
μL k (μH − λ) − λk (μH − μL )
μL k − λk λk k λk λk k
2 − μ (μ − λ) + 2 + μ (μ − λ)
(μL − λ) L L (μH − λ) H H
(4.43)
τ
FW (τ ) = fW (t) dt
0
τ
(μL − λ) (μH − λ)
=
μL k (μH − λ) − λk (μH − μL )
0
−(μL −λ)t k Γ (k, λt) −(μH −λ)t Γ (k, λt)
e μL +e μH k
1− dt
Γ (k) Γ (k)
(μL − λ) (μH − λ)
= k
μL (μH − λ) − λk (μH − μL )
τ
Γ (k, λt) Γ (k, λt)
e−(μL −λ)t μL k +e−(μH −λ)t μH k 1 − dt
Γ (k) Γ (k)
0
⎛
τ
(μL − λ) (μH − λ) ⎝ μL k
= k e−(μL −λ)t Γ (k, λt) dt
μL (μH − λ) − λk (μH − μL ) Γ (k)
0
⎞
τ k
τ
μ
+μH k e−(μH −λ)t dt − e−(μH −λ)t Γ (k, λt) dt⎠
H
(4.44)
Γ (k)
0 0
76 Chapter 4. Capacity Setting Methods
With:
τ
μL k
e−(μL −λ)t Γ (k, λt) dt
Γ (k)
0
It follows:
(μL − λ) (μH − λ)
FW (τ ) =
μL k (μH − λ) − λk (μH − μL )
μL k−1 e−μL τ (λτ ) μL k e−(μL −λ)τ Γ (k, λτ )
k
−
Γ (k) τ (μL − λ) (μL − λ) Γ (k)
μL k − λk λk Γ (k − 1, μL τ )
+ +
(μL − λ) (μL − λ) Γ (k − 1)
(μL − λ) (μH − λ) 1 − e−(μH −λ)τ
+ μH k
μL (μH − λ) − λ (μH − μL )
k k μH − λ
(μL − λ) (μH − λ)
+
μL k (μH − λ) − λk (μH − μL )
μH k−1 e−μH τ (λτ ) μH k e−(μH −λ)τ Γ (k, λτ )
k
−
Γ (k) τ (μH − λ) (μH − λ) Γ (k)
μH − λ
k k
λ k
Γ (k − 1, μH τ )
+ +
(μH − λ) (μH − λ) Γ (k − 1)
4.9. Appendix 77
(μH − λ)
=
μL k (μH − λ) − λk (μH − μL )
k k−1 k−1 −μL τ
λ μL τ e Γ (k, λτ ) k
− μL k e−(μL −λ)τ + μ L − λk
Γ (k) Γ (k)
Γ (k − 1, μL τ )
+λk
Γ (k − 1)
k k−1 k−1 −μH τ
(μL − λ) λ μH τ e
−
μL k (μH − λ) − λk (μH − μL ) Γ (k)
k −(μH −λ)τ Γ (k, λτ )
k
k Γ (k − 1, μH τ )
−μH e + μH − λ + λ k
Γ (k) Γ (k − 1)
−(μH −λ)τ
(μL − λ) μH 1 − e
k
+
μL k (μH − λ) − λk (μH − μL )
(μH − λ)
= k
μL (μH − λ) − λk (μH − μL )
k k−1 k−1 −μL τ
λ μL τ e Γ (k, λτ ) k
− μL k e−(μL −λ)τ + μ L − λk
Γ (k) Γ (k)
Γ (k − 1, μ L τ ) (μ L − λ)
+λk −
Γ (k − 1) μL (μH − λ) − λk (μH − μL )
k
k −(μH −λ)τ Γ (k, λτ )
μH e 1−
Γ (k)
k−1 k−1 −μH τ
k
λ μH τ e Γ (k − 1, μH τ )
+ −λ 1−
k
Γ (k) Γ (k − 1)
For simplifying the statement the following auxiliary functions are de-
fined:
(μH − λ)
νL =
μL (μH − λ) − λk (μH − μL )
k
(μL − λ)
νH = k
μL (μH − λ) − λk (μH − μL )
λk μL k−1 τ k−1 e−μL τ Γ (k, λτ )
ϑL (τ ) = − μL k e−(μL −λ)τ
Γ (k) Γ (k)
k
Γ (k − 1, μ L τ )
+ μ L − λk + λk
Γ (k − 1)
k −(μH −λ)τ Γ (k, λτ )
ϑH (τ ) =μH e 1−
Γ (k)
k−1 k−1 −μH τ
k
λ μH τ e Γ (k − 1, μH τ )
+ − λk 1 − (4.45)
Γ (k) Γ (k − 1)
78 Chapter 4. Capacity Setting Methods
FW (τ ) = νL ϑL (τ ) − νH ϑH (τ ) (4.46)
X
E [G] =λ FW (τ ) (1 − FL (τ )) dτ
0
X
=λ (νL ϑL (τ ) − νH ϑH (τ )) (1 − FL (τ )) dτ
0
X X
=λνL ϑL (τ ) e−βτ dτ − λνH ϑH (τ ) e−βτ dτ
0 0
X k−1 −μL τ
λk μk−1
L τ e Γ (k, λτ )
=λνL − μkL e−(μL −λ)τ
Γ (k) Γ (k)
0
Γ (k − 1, μL τ ) −βτ
+ μkL − λk + λk (k − 1) e dτ
Γ (k)
X
k −(μH −λ)τ Γ (k, λτ )
− λνH μH e 1−
Γ (k)
0
k−1 −μH τ
λk μk−1
H τ e Γ (k − 1, μH τ )
+ − λk 1 − e−βτ dτ
Γ (k) Γ (k − 1)
X
νL λk+1 μk−1
= L
τ k−1 e−(μL +β)τ dτ
Γ (k)
0
X
λνL μkL
− e−(μL −λ+β)τ Γ (k, λτ ) dτ
Γ (k)
0
X X
λk+1 νL
+ λνL μkL − λk e−βτ dτ + Γ (k − 1, μL τ ) e−βτ dτ
Γ (k − 1)
0 0
X X
λνH μkH
− λνH μkH e−(μH −λ+β)τ dτ + e−(μH −λ+β)τ Γ (k, λτ ) dτ
Γ (k)
0 0
4.9. Appendix 79
X
νH λk+1 μk−1
− H
τ k−1 e−(μH +β)τ dτ
Γ (k)
0
X X
λk+1 νH
+ λk+1 νH e−βτ dτ − Γ (k − 1, μH τ ) e−βτ dτ
Γ (k − 1)
0 0
(4.47)
X
−k
τ k−1 e−(μL +β)τ dτ =X k (X (β + μL )) [Γ (k) − Γ (k, X (μL + β))]
0
X
e−(μL +β)X (Xλ)
k
Γ (k)
e−(μL −λ+β)τ Γ (k, λτ ) dτ = +
X (μL + β) (μL − λ + β) μL − λ + β
0
−k
λk (μL + β) Γ (k)
−
μL − λ + β
−k−1
Γ (k − 1, X (μL + β)) (μL + β) λk
+
μL − λ + β μL − λ + β
Γ (k, Xλ) e−(μL −λ+β)X
((k − 1) (μL + β)) −
μL − λ + β
X
1 − e−βX
e−βτ dτ =
β
0
X
1 − e−(μH −λ+β)X
e−(μH −λ+β)τ dτ =
(μH − λ + β)
0
X
e−(μL +β)X (XμL )
k−1
Γ (k − 1)
Γ (k − 1, μL τ ) e−βτ dτ = +
Xβ (μL + β) β
0
−e−βX Γ (k − 1, XμL )
β
1−k
(k − 2) μk−1 (μL + β)
− L
β
Γ (k − 2) − Γ (k − 2, X (μL + β))
(4.48)
β
80 Chapter 4. Capacity Setting Methods
Leads to:
νL λk+1 μk−1 −k
E [G] = L
X k (X (β + μL )) (Γ (k) − Γ (k, X (β + μL )))
Γ (k)
e−(μL +β)X (Xλ)
k
λνL μkL
−
Γ (k) X (μL + β) (μL − λ + β)
−k
Γ (k) λk (μL + β) Γ (k)
+ −
μL − λ + β μL − λ + β
−k−1
Γ (k − 1, X (μL + β)) (μL + β) λk
+ (k − 1) (μL + β)
μL − λ + β
Γ (k, Xλ) e−(μL −λ+β)X
1 − e−βX λk+1 νL
− + λνL μkL − λk +
μL − λ + β β Γ (k − 1)
−(μL +β)X k−1 −βX
e (XμL ) Γ (k − 1) − e Γ (k − 1, XμL )
+
Xβ (μL + β) β
1−k
(k − 2) μk−1 (μL + β) [Γ (k − 2) − Γ (k − 2, X (μL + β))]
− L
β β
1 − e−(μH −λ+β)X e−(μH +β)X (Xλ)
k
λνH μkH
− λνH μkH +
(μH − λ + β) Γ (k) X (μH + β) (μH − λ + β)
−k
Γ (k) λk (μH + β) Γ (k)
+ −
μH − λ + β μH − λ + β
−k−1
Γ (k − 1, X (μH + β)) (μH + β) λk
+ ((k − 1) (μH + β))
μH − λ + β
Γ (k, Xλ) e−(μH −λ+β)X νH λk+1 μk−1 −k
− − H
X k (X (β + μH ))
μH − λ + β Γ (k)
1 − e−βX λk+1 νH
(Γ (k) − Γ (k, X (β + μH ))) + λk+1 νH −
β Γ (k − 1)
e−(μH +β)X (XμH ) Γ (k − 1) − e−βX Γ (k − 1, XμH )
k−1
+
Xβ (μH + β) β
1−k
(k − 2) μk−1
H (μH + β) (Γ (k − 2) − Γ (k − 2, X (μH + β)))
−
β β
(4.49)
4.9. Appendix 81
−k Γ (k, X (β + μL ))
E [G] =νL λk+1 μk−1
L (β + μL ) 1−
Γ (k)
e−(μL +β)X (Xλ)
k
− λνL μkL
Γ (k) X (μL + β) (μL − λ + β)
−k
1 − λk (μL + β) e−(μL −λ+β)X Γ (k, Xλ)
+ −
μL − λ + β μL − λ + β Γ (k)
−k k
(μL + β) λ Γ (k − 1, X (μL + β))
+
(μL − λ + β) Γ (k − 1)
1−e −βX
1
+ λνL μkL − λk + λk+1 νL
β β
e−βX Γ (k − 1, XμL ) e−(μL +β)X (XμL )
k−1
− +
β Γ (k − 1) Γ (k − 1) Xβ (μL + β)
μLk−1
Γ (k − 2, X (μL + β))
− 1−
β (μL + β)
k−1 Γ (k − 2)
1 − e−(μH −λ+β)X
− λνH μkH
(μH − λ + β)
−k
e−(μH +β)X (Xλ)
k
1 − λk (μH + β)
+ λνH μkH +
Γ (k) X (μH + β) (μH − λ + β) μH − λ + β
−k
(μH + β) λk Γ (k − 1, X (μH + β))
+
μH − λ + β Γ (k − 1)
−(μH −λ+β)X
e Γ (k, Xλ) 1 − e−βX
− + λk+1 νH
μH − λ + β Γ (k) β
−k Γ (k, X (β + μH ))
− νH λk+1 μk−1
H (β + μ H ) 1 −
Γ (k)
−(μH +β)X
e−βX Γ (k − 1, XμH )
k−1
e (XμH ) 1
− λk+1 νH + −
Γ (k − 1) Xβ (μH + β) β β Γ (k − 1)
1−k
μk−1 (μH + β) Γ (k − 2, X (μH + β))
− H 1− (4.50)
β Γ (k − 2)
82 Chapter 4. Capacity Setting Methods
X ∞
s= FW (τ ) fL (τ ) dτ + FW (X) fL (τ ) dτ
0 X
X
k−1 −μL τ
(μH − λ) λk μk−1
L τ e
=
μkL (μH − λ) − λk (μH − μL ) Γ (k)
0
Γ (k, λτ ) k
Γ (k − 1, μL τ )
−μkL e−(μL −λ)τ + μL − λk +λk
Γ (k) Γ (k − 1)
(μL − λ) k −(μH −λ)τ Γ (k, λτ )
− k μ H e 1 −
μL (μH − λ) − λk (μH − μL ) Γ (k)
k k−1 k−1 −μH τ
λ μH τ e Γ (k − 1, μH τ )
+ − λk 1 − βe−βτ dτ
Γ (k) Γ (k − 1)
+ (νL ϑL (X) − νH ϑH (X)) e−βX
⎛X
k k−1 k−1 −μL τ
⎝ λ μL τ e
=νL βe−βτ dτ
Γ (k)
0
X X
Γ (k, λτ ) k
⎛
X
λk μk−1
L β
=νL ⎝ τ k−1 e−(μL +β)τ dτ
Γ (k)
0
X
μkL β
− + μkL − λk 1 − e−βX
μL − λ + β
e−(μL +β)X (XμL )
k−1
λk β
+
Γ (k − 1) Xβ (μL + β)
Γ (k − 1) − e−βX Γ (k − 1, XμL )
+
β
1−k
(k − 2) μk−1 (μL + β) Γ (k − 2) − Γ (k − 2, X (μL + β))
− L
β β
−(μH −λ+β)X
e−(μH +β)X (Xλ)
k
1−e k
μ β
− νH μkH β − H
(μH − λ + β) Γ (k) X (μH + β) (μL − λ + β)
84 Chapter 4. Capacity Setting Methods
−k
Γ (k) λk (μH + β) Γ (k)
+ −
μH − λ + β μH − λ + β
−k−1
Γ (k − 1, X (μH + β)) (μH + β) λk
+ ((k − 1) (μH + β))
μH − λ + β
−(μH −λ+β)X
Γ (k, Xλ) e
−
μH − λ + β
λk μk−1
H β −k
+ X k (X (β + μH )) (Γ (k) − Γ (k, X (μH + β)))
Γ (k)
− λk 1 − e−βX
e−(μH +β)X (XμH )
k−1
λk β
+
Γ (k − 1) Xβ (μH + β)
Γ (k − 1) − e−βX Γ (k − 1, XμH )
+
β
1−k
(k − 2) μk−1
H (μH + β) (Γ (k − 2) − Γ (k − 2, X (μH + β)))
−
β β
+ (νL ϑL (X) − νH ϑH (X)) e−βX (4.51)
λk μk−1
L β Γ (k, X (μL + β))
s =νL k
1−
(β + μL ) Γ (k)
λk μkL β e−(μL +β)X X k−1 1
− + k
μL − λ + β Γ (k) (μL + β) λ
1 Γ (k − 1, X (μL + β))
− 1−
(μL + β)
k−1 Γ (k − 1)
−(μL +β)X k−1
e X e−(μL +β)X X k−2
+ μkL + λk μk−1 +
L
Γ (k) Γ (k − 1) (μL + β)
1 Γ (k − 2, X (μL + β))
− 1−
(μL + β)
k−1 Γ (k − 2)
μkL (μL − λ) e−(μL −λ+β)X Γ (k, Xλ)
−
μL − λ + β Γ (k)
k k−1
λ μH β Γ (k, X (μH + β))
− νH k
1 −
(μH + β) Γ (k)
4.9. Appendix 85
Wipcap
The Wipcap constrains the input of materials or orders into the produc-
tion system. In the current implementation, the WIP is measured in final
production units which are released to the shop floor. Whenever an order
is released to the shop floor, the WIP-counter is increased by the number
of finished goods of this order. Each order is assumed to refer to only one
type of finished goods. As stated above, two different types of Wipcap are
implemented. In type 1, which is the definition directly derived from the
5.1. Model Description 89
6201 6202 6203 6211 6212 6213 6214 6221 6222 6223 6224 6225
M4
M3
4101 4102
M1
3101
M2
2101 2102
M1
1301 1303
Conwip literature (Framinan et al., 2003; Hopp and Spearman, 2008), the
WIP-counter is reduced by the number of finished goods, whenever an order
is finished at machine group M4. In type 2, which is specifically applica-
ble for MTO production systems, the WIP-counter is reduced whenever an
order leaves the FGI. In both implementations orders are only released to
the production system when the WIP-counter including the next order to
be released is lower or equal the Wipcap.
Work-Ahead-Window (WAW)
In the current implementation the WAW sets a time window into the future,
e.g. 20 periods of time. An order can only be released to the production
system when its due date is within this WAW. This is the typical imple-
mentation of a WAW in MTO production systems (Framinan et al., 2003;
Jodlbauer and Huber, 2008; Hopp and Spearman, 2008).
90 Chapter 5. Conwip
Capacity Trigger
The capacity trigger is an additional functionality of Conwip which indi-
cates if enough capacity is available or if additional capacity is needed. For
simplification reasons no such functionality is implemented in the current
study and the capacity provided is constant.
Safety Stock
In addition to the five usually applied parameters in a Conwip system, a
safety stock for finished goods is added in this Conwip implementation. For
each finished good a certain percentage of the average periodic demand is
kept as safety stock to increase service level and decrease tardiness.
To identify which Wipcap leads to a better result for each scenario, all
combinations of the following values for Wipcap and WAW are enumerated.
The ranges and values have been identified in a set of preliminary experi-
ments and are presented in Table 5.3. The Wipcap is incremented by steps
of five, and the WAW by steps of one.
Additionally, the influence of implementing a safety stock in both meth-
ods has been identified by including the following safety stock values for each
finished good as percentage of its monthly average demand as presented in
Table 5.4.
For all parameter combinations tested, 40 replications are conducted, for
which 720 time periods are simulated while a warm-up period of 240 time
periods is excluded (2 years, 480 time periods are evaluated). Overall, this
simulation study consists of 3, 321, 720 simulation runs being performed in
AnyLogic 6.9. To hold an item of FGI costs 1 monetary unit per time period
and a WIP item costs 0.75 monetary units per time period. Tardiness costs
are penalised by 19 monetary units per period.
5.2. Results 93
Wipcap WAW
Parameter Wipcap Lower Upper Lower Upper
demand method level level level level
0.875 1 400 600 6 21
0.875 2 430 630 6 21
1 1 400 600 6 21
1 2 430 630 6 21
1.125 1 400 600 6 27
1.125 2 430 920 6 27
5.2 Results
In this section the results of the simulation study are presented. Firstly, the
influence of Wipcap and WAW on overall costs for Wipcap method 1 and
Wipcap method 2 respectively is discussed in the basic setting. Secondly,
the performance of the two Wipcap methods without implementing a safety
stock is compared for the nine scenarios (3 demand scenarios, 3 customer
required lead time scenarios). Thirdly, the effect of implementing a safety
stock in Conwip is discussed based on the nine scenarios.
Table 5.4: Introduced safety stock percentage of the periodic demand for FGI
the two methods tested has been observed. Hence, this circumstance leads
to the assumptions that the two Wipcap methods investigated have no
significant influence on overall costs. Additionally, a short WAW leads to
higher costs since the tardiness costs are increasing because the due dates
cannot be hold by the production system. The main purpose of the WAW is
to avoid producing items whose due date is too far in the future (Spearman
et al., 1990; Jodlbauer, 2008a). A WAW between 10 and 21 time unites for
the sets tested leads almost to the same overall costs, because the production
is limited by the Wipcap leading to the minimal overall costs and not by
the WAW.
2500
- - Wipcap Method 2
+
2000 + Wipcap Method 1
1500 +
-
Costs
+
- - + +
-
- - - - - - - + -
+ -
-
+ +
1000 +
- +
+ + + +
+
500
0
5 10 15 20
WAW
Figure 5.3 illustrates the minimum overall costs, when the Wipcap changes.
Again, to allow a comparison between both Wipcap methods, the best WAW
for each Wipcap leading to minimal overall costs is selected. For the cases
5.2. Results 95
tested, Wipcap method 1, where only the WIP is constrained by the Wip-
cap, leads to lower overall costs then Wipcap method 2, where WIP+FGI
are constrained by the Wipcap parameter. This circumstance holds for a
confidence interval with a probability of 95% except for three values where
no significant difference has been observed.
4000
- - Wipcap Method 2
-- + Wipcap Method 1
3000
-
-
- ---
Costs
2000 --
---
-
---
------
+++ ------- - --
1000 +++++++++ +++ - ----
-++
+ +++++++++++++++++++++++
0
400 450 500 550 600 650
Wipcap
Observation 5.2.1 (System load). Overall costs, FGI costs, WIP costs,
tardiness costs, production lead time and WAW increases by increasing sys-
tem load
As a first intuitive finding, overall costs, FGI costs, WIP costs, tardi-
ness costs, production lead time and WAW increase when the system load
(bottleneck utilisation) of the investigated production systems is increased
since more orders have to be produced at the same simulation time period.
Increasing demand at the same capacity provided results in a higher pro-
duction lead time and WIP according to Little (1961). If more orders are
stated the risk, that more items are produced too early (FGI costs increase)
and also too late (tardiness costs increase), increases.
Observation 5.2.2 is the main result of this simulation study and indi-
cates that for the scenarios tested a constrained WIP is beneficial compared
to Wipcap method 2 where WIP+FGI are constrained. For the medium
and high load scenario it has been observed that Wipcap method 2 leads
to more tardiness costs. This observation holds only when no safety stock
is applied. Moreover, there is a tendency that method 2 leads to a longer
production lead time than method 1 for medium and high load scenarios.
An explanation for Observation 5.2.2 could be that classic Conwip tries to
balance the workload independently of the customer behaviour. When FGI
is included in the Wipcap (method 2), then the customer behaviour is part
of the production planning process. It seems that the violation of the basic
Conwip principle leads to a worse performance.
Observation 5.2.3 (Saftey stock). Safety stock is not needed for the in-
vestigated system configurations
For the high demand scenarios, Wipcap method 2 (WIP+ FGI) requests
an introduction of safety stock in terms of overall costs. Nevertheless, Wip-
cap method 1 turns out to be the better one and requests no safety stock.
When safety stock is applied then it could be observed that the Wipcap
decreases.
5.3. Concluding Remarks 97
Table 5.5: Analysis of Wipcap method comparison without safety stock [Values for Wipcap method 1
/ Values for Wipcap method 2]
Parameter Results
U* CRL WAW Wipcap Costs Costs Costs Costs Production
[%] FGI WIP tardiness lead time
0.5 9\9 545\590 579\570 156\151b 231\227 192\192 19.3\19
72 1 11\11 565\580 589\634 262\201d 231\232 96\201a 19.3\19.3
2 10\9 460\615 583\540 207\154d 225\227 151\159 18.9\19.1
0.5 12\15 515\625 786\929 306\188d 279\295c 200\446b 20.8\22.1d
82 1 13\9 485\590 771\933a 360\123d 275\282 136\529d 20.4\20.9b
2 13\18 595\620 791\965 369\191d 291\290 131\485c 21.6\21.7
0.5 21\21 585\905 1259\1814b 684\264d 371\445d 204\1104d 24.6\29.8d
92 1 20\18 575\905 1281\1750b 630\257d 366\451d 284\1042c 24.4\30.4d
2 21\21 610\915 1243\1618 704\279d 382\442d 157\897c 25.2\29.8d
a
90% confidence interval
b
95% confidence interval
c
99% confidence interval
d
99.9% confidence interval
*
Bottleneck utilisation
Chapter 5. Conwip
Table 5.6: Analysis of Wipcap method comparison with safety stock [Values for Wipcap method 1 / Values
for Wipcap method 2]
Parameter Results
U* CRL WAW Wipcap Safety Costs Costs Costs Costs Prod.
5.4. Appendix Conwip
Table 5.7: Analysis Wipcap without safety stock: confidence interval 0.1
Costs Costs FGI Costs WIP Costs tardiness Prod. lead time
U* Bound CRL Method Method Method Method Method
[%] 1 2 1 2 1 2 1 2 1 2
72 lower 0.5 520 510 154 149 227 224 133 132 19.1 18.8
upper 638 632 158 153 234 231 251 254 19.5 19.2
lower 1 545 577 260 197 228 228 53 144 19.1 19.1
upper 634 689 264 204 235 235 140 257 19.5 19.5
lower 2 530 492 205 152 223 224 97 111 18.8 18.9
upper 637 590 209 156 228 231 205 208 19.1 19.3
82 lower 0.5 720 843 300 181 276 290 131 358 20.7 21.8
upper 856 1021 313 195 283 300 273 539 21.0 22.4
lower 1 719 832 354 119 272 277 82 428 20.3 20.7
upper 822 1040 366 127 278 286 189 636 20.6 21.2
lower 2 732 816 363 185 286 285 71 335 21.3 21.5
upper 848 1113 374 197 296 294 189 632 21.9 22.0
92 lower 0.5 1185 1498 638 244 365 430 118 784 24.3 29.0
upper 1338 2158 730 285 377 461 296 1452 24.8 30.6
lower 1 1199 1448 585 236 361 435 174 737 24.1 29.4
upper 1370 2080 675 279 372 467 402 1376 24.6 31.4
lower 2 1189 1281 663 260 376 428 85 557 24.9 29.0
upper 1299 1987 745 298 389 456 230 1269 25.6 30.5
*
Bottleneck utilisation
Chapter 5. Conwip
Table 5.8: Analysis Wipcap without safety stock: confidence interval 0.05
Costs Costs FGI Costs WIP Costs tardiness Prod. lead time
U* Bound CRL Method Method Method Method Method
[%] 1 2 1 2 1 2 1 2 1 2
72 lower 0.5 509 498 154 148 227 223 122 121 19.1 18.8
5.4. Appendix Conwip
upper 649 644 159 154 234 231 262 266 19.6 19.2
lower 1 537 566 259 196 227 227 44 133 19.1 19.1
upper 642 700 265 205 236 236 148 268 19.6 19.5
lower 2 520 482 205 152 222 223 87 102 18.8 18.9
upper 647 599 209 156 228 231 215 217 19.1 19.3
82 lower 0.5 707 825 299 180 275 289 118 341 20.6 21.8
upper 869 1039 314 196 284 301 287 557 21.0 22.4
lower 1 709 812 353 118 271 276 72 408 20.3 20.7
upper 832 1060 367 127 279 287 200 656 20.6 21.2
lower 2 721 788 362 184 285 284 60 307 21.3 21.4
upper 859 1141 376 199 297 295 201 661 21.9 22.0
92 lower 0.5 1171 1435 629 240 364 427 101 720 24.2 28.8
upper 1353 2221 738 289 378 464 313 1516 24.9 30.8
lower 1 1182 1388 577 232 360 432 152 675 24.0 29.2
upper 1387 2140 683 283 373 470 424 1437 24.7 31.6
lower 2 1179 1213 656 256 374 425 71 488 24.8 28.9
upper 1310 2055 753 302 390 459 244 1338 25.6 30.7
*
Bottleneck utilisation
101
102
Table 5.9: Analysis Wipcap without safety stock: confidence interval 0.01
Costs Costs FGI Costs WIP Costs tardiness Prod. lead time
U* Bound CRL Method Method Method Method Method
[%] 1 2 1 2 1 2 1 2 1 2
72 lower 0.5 487 475 153 148 226 222 100 98 19.0 18.7
upper 671 667 159 155 236 233 284 288 19.6 19.3
lower 1 521 545 258 195 226 226 28 112 19.0 19.0
upper 659 721 265 207 237 237 165 289 19.6 19.6
lower 2 500 464 204 151 221 222 67 84 18.7 18.8
upper 667 618 210 157 229 232 236 236 19.1 19.4
82 lower 0.5 681 792 297 178 274 287 91 307 20.5 21.7
upper 895 1072 316 199 285 303 313 590 21.1 22.6
lower 1 690 773 350 117 270 275 51 369 20.2 20.6
upper 851 1099 370 129 280 289 220 695 20.7 21.3
lower 2 699 733 360 181 283 282 38 251 21.2 21.3
upper 880 1196 378 201 298 297 223 717 22.0 22.1
92 lower 0.5 1142 1311 612 232 362 421 68 595 24.1 28.5
upper 1381 2345 756 297 380 469 346 1642 25.0 31.1
lower 1 1150 1270 560 224 358 426 109 556 23.9 28.8
upper 1419 2259 700 291 375 476 467 1557 24.8 32.0
lower 2 1158 1081 640 249 372 420 44 355 24.7 28.6
upper 1330 2187 769 309 393 464 271 1471 25.8 31.0
*
Bottleneck utilisation
Chapter 5. Conwip
Table 5.10: Analysis Wipcap without safety stock: confidence interval 0.001
Costs Costs FGI Costs WIP Costs tardiness Prod. lead time
U* Bound CRL Method Method Method Method Method
[%] 1 2 1 2 1 2 1 2 1 2
72 lower 0.5 461 449 152 147 224 220 74 71 18.9 18.6
5.4. Appendix Conwip
upper 697 694 160 156 237 234 310 315 19.7 19.4
lower 1 501 521 257 193 224 225 9 87 18.9 18.9
upper 678 745 266 208 239 239 184 314 19.7 19.7
lower 2 477 443 203 151 220 220 43 63 18.7 18.7
upper 690 639 211 158 231 234 259 257 19.2 19.5
82 lower 0.5 651 753 294 175 272 285 61 268 20.5 21.6
upper 925 1111 319 202 287 305 344 630 21.2 22.7
lower 1 668 727 348 115 268 273 28 323 20.1 20.5
upper 873 1144 372 130 282 291 243 740 20.8 21.4
lower 2 674 668 357 179 281 280 12 186 21.1 21.2
upper 905 1261 380 204 301 299 248 781 22.1 22.2
92 lower 0.5 1109 1168 592 223 359 415 29 450 24.0 28.2
upper 1414 2488 776 306 383 476 385 1787 25.1 31.5
lower 1 1113 1132 541 214 356 419 60 417 23.8 28.4
upper 1456 2396 720 300 377 483 516 1696 24.9 32.4
lower 2 1135 927 622 240 369 414 13 200 24.6 28.2
upper 1354 2340 786 317 396 471 303 1626 25.9 31.3
*
Bottleneck utilisation
103
104
Table 5.11: Analysis Wipcap with safety stock: confidence interval 0.1
Costs Costs FGI Costs WIP Costs tardiness Prod. lead time
U* Bound CRL Method Method Method Method Method
[%] 1 2 1 2 1 2 1 2 1 2
72 lower 0.5 487 475 153 148 226 222 100 98 19.0 18.7
upper 671 667 159 155 236 233 284 288 19.6 19.3
lower 1 521 545 258 195 226 226 28 112 19.0 19.0
upper 659 721 265 207 237 237 165 289 19.6 19.6
lower 2 500 464 204 151 221 222 67 84 18.7 18.8
upper 667 618 210 157 229 232 236 236 19.1 19.4
82 lower 0.5 681 792 297 178 274 287 91 307 20.5 21.7
upper 895 1072 316 199 285 303 313 590 21.1 22.6
lower 1 690 773 350 117 270 275 51 369 20.2 20.6
upper 851 1099 370 129 280 289 220 695 20.7 21.3
lower 2 699 733 360 181 283 282 38 251 21.2 21.3
upper 880 1196 378 201 298 297 223 717 22.0 22.1
92 lower 0.5 1185 1443 638 620 365 431 118 347 24.3 29.3
upper 1338 1664 730 666 377 465 296 577 24.8 31.4
lower 1 1199 1521 585 809 361 450 174 203 24.1 30.4
upper 1370 1678 675 870 372 495 402 372 24.6 33.2
lower 2 1189 1500 663 729 376 440 85 285 24.9 29.9
upper 1299 1738 745 777 389 476 230 530 25.6 31.9
*
Bottleneck utilisation
Chapter 5. Conwip
Table 5.12: Analysis Wipcap with safety stock: confidence interval 0.05
Costs Costs FGI Costs WIP Costs tardiness Prod. lead time
U* Bound CRL Method Method Method Method Method
[%] 1 2 1 2 1 2 1 2 1 2
72 lower 0.5 487 475 153 148 226 222 100 98 19.0 18.7
5.4. Appendix Conwip
upper 671 667 159 155 236 233 284 288 19.6 19.3
lower 1 521 545 258 195 226 226 28 112 19.0 19.0
upper 659 721 265 207 237 237 165 289 19.6 19.6
lower 2 500 464 204 151 221 222 67 84 18.7 18.8
upper 667 618 210 157 229 232 236 236 19.1 19.4
82 lower 0.5 681 792 297 178 274 287 91 307 20.5 21.7
upper 895 1072 316 199 285 303 313 590 21.1 22.6
lower 1 690 773 350 117 270 275 51 369 20.2 20.6
upper 851 1099 370 129 280 289 220 695 20.7 21.3
lower 2 699 733 360 181 283 282 38 251 21.2 21.3
upper 880 1196 378 201 298 297 223 717 22.0 22.1
92 lower 0.5 1171 1422 629 616 364 428 101 325 24.2 29.1
upper 1353 1685 738 671 378 469 313 599 24.9 31.6
lower 1 1182 1506 577 803 360 446 152 187 24.0 30.1
upper 1387 1693 683 876 373 500 424 388 24.7 33.4
lower 2 1179 1477 656 725 374 437 71 261 24.8 29.7
upper 1310 1761 753 782 390 480 244 554 25.6 32.1
*
Bottleneck utilisation
105
106
Table 5.13: Analysis Wipcap with safety stock: confidence interval 0.01
Costs Costs FGI Costs WIP Costs tardiness Prod. lead time
U* Bound CRL Method Method Method Method Method
[%] 1 2 1 2 1 2 1 2 1 2
72 lower 0.5 487 475 153 148 226 222 100 98 19.0 18.7
upper 671 667 159 155 236 233 284 288 19.6 19.3
lower 1 521 545 258 195 226 226 28 112 19.0 19.0
upper 659 721 265 207 237 237 165 289 19.6 19.6
lower 2 500 464 204 151 221 222 67 84 18.7 18.8
upper 667 618 210 157 229 232 236 236 19.1 19.4
82 lower 0.5 681 792 297 178 274 287 91 307 20.5 21.7
upper 895 1072 316 199 285 303 313 590 21.1 22.6
lower 1 690 773 350 117 270 275 51 369 20.2 20.6
upper 851 1099 370 129 280 289 220 695 20.7 21.3
lower 2 699 733 360 181 283 282 38 251 21.2 21.3
upper 880 1196 378 201 298 297 223 717 22.0 22.1
92 lower 0.5 1142 1380 612 607 362 421 68 282 24.1 28.7
upper 1381 1726 756 679 380 475 346 642 25.0 31.9
lower 1 1150 1477 560 792 358 437 109 155 23.9 29.6
upper 1419 1722 700 887 375 508 467 419 24.8 33.9
lower 2 1158 1433 640 716 372 430 44 215 24.7 29.3
upper 1330 1805 769 791 393 486 271 600 25.8 32.5
*
Bottleneck utilisation
Chapter 5. Conwip
Table 5.14: Analysis Wipcap with safety stock: confidence interval 0.001
Costs Costs FGI Costs WIP Costs tardiness Prod. lead time
U* Bound CRL Method Method Method Method Method
[%] 1 2 1 2 1 2 1 2 1 2
72 lower 0.5 487 475 153 148 226 222 100 98 19.0 18.7
5.4. Appendix Conwip
upper 671 667 159 155 236 233 284 288 19.6 19.3
lower 1 521 545 258 195 226 226 28 112 19.0 19.0
upper 659 721 265 207 237 237 165 289 19.6 19.6
lower 2 500 464 204 151 221 222 67 84 18.7 18.8
upper 667 618 210 157 229 232 236 236 19.1 19.4
82 lower 0.5 681 792 297 178 274 287 91 307 20.5 21.7
upper 895 1072 316 199 285 303 313 590 21.1 22.6
lower 1 690 773 350 117 270 275 51 369 20.2 20.6
upper 851 1099 370 129 280 289 220 695 20.7 21.3
lower 2 699 733 360 181 283 282 38 251 21.2 21.3
upper 880 1196 378 201 298 297 223 717 22.0 22.1
92 lower 0.5 1109 1332 592 597 359 414 29 232 24.0 28.3
upper 1414 1774 776 689 383 482 385 692 25.1 32.4
lower 1 1113 1442 541 778 356 427 60 118 23.8 29.0
upper 1456 1757 720 901 377 518 516 456 24.9 34.5
lower 2 1135 1381 622 705 369 422 13 162 24.6 28.9
upper 1354 1857 786 801 396 494 303 653 25.9 32.9
*
Bottleneck utilisation
107
Chapter 6
Dispatching Rules
In this chapter the influence of different dispatching rules on the average
production lead time is investigated. Two theorems based on covariance be-
tween processing time and production lead time are formulated and proved
theoretically. Theorem 1 links the average production lead time to the
”processing time weighted production lead time” for multi-stage production
systems analytically. The influence of different dispatching rules on average
production lead time, which is well-known from simulation and empirical
studies, can be proved theoretically in Theorem 2 for a single-stage produc-
tion system. A simulation study has been conducted to gain more insight
into the influence of dispatching rules on average production lead time in a
multi-stage production system. It has been found out that the ”processing
time weighted average production lead time” for a multi-stage production
system is not invariant with the applied dispatching rule and can be used
as an indicator independent of dispatching rules for single-stage production
systems.
Production lead time of an order is the time span between entering the
buffer in front of the processing unit and leaving the processing unit. The
queuing time denotes the production lead time minus the processing time
and can be interpreted as the waiting time (while not being processed) in
the buffer stock. Processing time, production lead time and queuing time
of order i at stage α are random variables denoted with Pi,α , Li,α and Qi,α ,
respectively. The terminology used in this chapter is presented in Table 6.1.
There is a lot of literature available examining the proof of Little’s Law
(Little, 1961; Eilon, 1969; Medhi, 1991; Hopp and Spearman, 2008). How-
ever, all of this literature discusses the relationship between expected pro-
duction lead time and expected inventory related to the processing rate, but
none discusses the ”processing time weighted production lead time”. For
the expected production lead time, simulation studies show that relationship
between expected production lead time and expected inventory depends on
the applied dispatching rule. The above mentioned proofs for Little’s Law
are extended by a proof delivered in Jodlbauer and Stöcher (2006) which
shows that, for continuous input and output processes, the ”processing time
weighted average production lead time” is independent of the dispatching
rule in a single-stage production system. They also show that Little’s Law
still holds true for the ”processing time weighted average production lead
time”. Furthermore, their proofs are valid for monotonic step functions
which means that it also holds for a setting with discrete customer orders.
In Theorem 6.1.1, the relationship between average production lead time
and ”processing time weighted average production lead time” is formulated
using the covariance of production lead time and processing time.
Cov[L, P ]
E[L] = l − (6.1)
E[P ]
Cov[P, Q] + V ar[P ]
=l− (6.2)
E[P ]
Cov[L, P ] =Cov[P + Q, P ]
=Cov[P, P ] + Cov[Q, P ]
=V ar[P ] + Cov[Q, P ] (6.4)
If the input stream of the production system (see Figure 6.1) is changed,
then Cov[P, Q], l and finally E[L] changes subsequently. Therefore, gener-
ally E[Linput1 ]
= E[Linput2 ] holds true. If another dispatching rule is ap-
plied, the Cov[P, Q] changes and therefore
E[Ldispatchingrule1 ]
= E[Ldispatchingrule2 ] holds true. If dispatching rules
are applied, which do not use processing time information for priority cal-
culation, Cov[P, Q] is equal to 0. Therefore, E[Lrandom ] is defined:
V ar[P ]
E[Lrandom ] := lrandom − (6.5)
E[P ]
E[Lrandom ] can be interpreted as the expected production lead time of
dispatching rules which do not use any information about the processing
time. Nyhuis and Wiendahl (1999)(see Nyhuis and Wiendahl (2009) for
the English version) developed the same formula as Equation 6.5 in 1999,
in which the average production lead time equals the mean range minus
the squared coefficient of variation of the processing time multiplied by its
expected value (for details see Nyhuis and Wiendahl (2009) Equation 4.40).
Theorem 6.1.3 introduces dispatching rules to the model framework
based on the property that the ”processing time weighted average produc-
tion lead time” is independent of the dispatching rule applied. Moreover,
Theorem 6.1.3 holds true only for single-stage production systems and is
formulated for (i) random-like, (ii) LPT-like and (iii) SPT-like dispatching
rules.
Theorem 6.1.3 (Dispatching rules). For any dispatching rule D of a single-
stage production system with:
(i) Cov[P, Qrandlike ] = 0 is applied, E[LD ] = E[Lrandom ] holds true.
Especially, FIFO does not use any processing time information for
prioritising and therefore E[LF IF O ] = E[Lrandom ] holds true.
(ii) Cov[P, QLP T like ] < 0 is applied, E[LD ] > E[Lrandom ] holds true.
Especially, LPT prioritises long processing times and therefore
E[LLP T ] > E[Lrandom ] holds true.
(iii) Cov[P, QSP T like ] > 0 is applied, E[LD ] < E[Lrandom ] holds true.
Especially, SPT prioritises short processing times and therefore
E[LSP T ] < E[Lrandom ] holds true.
112 Chapter 6. Dispatching Rules
1
n
n Pi L i
E[LP ] i=1
l= ≈
n (6.6)
E[P ] 1
n Pi
i=1
According to Jodlbauer and Stöcher (2006) l does not depend on the dis-
patching rule used for single-stage production systems.
(i)
Cov[P, QD ] + V ar[P ]
E[LD ] = l −
(6.2) E[P ]
V ar[P ]
= l− = E[Lrandom ] (6.7)
Cov[P,QD ]=0 E[P ] (6.5)
(ii)
Cov[P, QD ] + V ar[P ]
E[LD ] = l −
(6.2) E[P ]
V ar[P ]
> l− = E[Lrandom ] (6.8)
Cov[P,QD ]<0 E[P ] (6.5)
(iii)
Cov[P, QD ] + V ar[P ]
E[LD ] = l −
(6.2) E[P ]
V ar[P ]
< l− = E[Lrandom ] (6.9)
Cov[P,QD ]>0 E[P ] (6.5)
(iii) Jobs with short processing times are prioritised and queuing time is
short:
Table 6.2 summarises the results for the four-machine case, where l1 to
l4 indicate the ”processing time weighted average production lead time” for
each single stage of the four-machine production system. l12 , l123 and l1234
are the ”processing time weighted average production lead times” for the
stages one to two, one to three and the whole production system respectively.
ρ[P, Q] for stage 1 is the correlation coefficient between the processing time
at stage 1 and the queuing time at stage 1. ρ[P, Q] for the stages 1234 is
the correlation coefficient between the sum of processing times of all stages
and the sum of the waiting time of all stages.
The ”processing time weighted average production lead time” of the first
stage (single-machine production system) is independent of the dispatching
rule as shown in Nyhuis and Wiendahl (2009), Jodlbauer (2005) and Jodl-
bauer and Stöcher (2006). For all other machines in the flow shop (using
the same dispatching rule at each stage), it has been observed that the
”processing time weighted average production lead time” depends on the
dispatching rule used and is increasing over each production state (except
for the dispatching rule SRPT). Especially for LPT an interesting gap be-
tween machine 1 and machine 4 can be observed. The results explicitly
show for all dispatching rules that l1234 is not the sum of l1 to l4 .
For group (i) l is approximately equal at each stage because no processing
time information is used. Moreover, ρ[Pα , Qα ] is 0 for dispatching rules of
group (i) for the single-machine stages. For the multi-stage systems the
covariance term is unequal to 0, which, indeed, makes the analytical proof
of Theorem 6.1.3 for the multi-machine case unfeasible in the current setting.
Category (iv) in Table 6.2 presents the SPT rule in combination with a
slack rule which increases the practical relevance of the SPT dispatching rule
because due date is included in the priority calculation of orders. Based on
this combination it fits into none of the groups discussed in Theorem 6.1.3.
The correlation coefficient for stages 1 and 2 and the multi-stage production
system 12 are positive, which shows a dominant SPT-like behaviour. For
the multi-stage production system 123 and 1234 the slack behaviour gets
dominant in order to fulfil the due dates. The switch between slack and
SPT-like behaviour depends on the switching point between SPT and least
slack rule. In this case the rule switches if orders have exceeded the due
date but any other scenario is possible.
It is conjectured that the input stream of the downstream sub-production
system is influenced by the dispatching of the upstream sub-production sys-
tem. If, for example, LPT is chosen, in the beginning the whole system is
empty and the orders arrive according to a Poisson process at the queue of
116 Chapter 6. Dispatching Rules
the first machine. The order with the longest processing time is released
into the processing at the first machine. However, on the second process-
ing unit no orders are waiting and the finished order at processing unit 1
has no waiting time in front of processing unit 2. Due to the fact that the
processing times are independent, the processing time of the finished order
at processing unit 1 is, with a high probability, higher than the processing
time of processing unit 2. As defined in Figure 6.1, the production system
boundaries include the input stream. If another dispatching rule is applied,
the inputstream changes and a different production system results.
Table 6.5 and Table 6.4 indicate expected value and coefficient of vari-
ation of L for single- and multi-stage production systems. cv [L] for stage 1
for the dispatching rule Random is 2.59. E[L] for the first two stages of the
use of SRPT as dispatching rule is 3.71.
According to Theorem 6.1.3, E[Li ] is equal for all dispatching rules,
where no information about the processing time is used for prioritising jobs
based on a single-stage setting, which is indeed the case for dispatching rules
of group (i). The same behaviour is observed for the multi-stage cases.
For the tested single- and multi-stage production systems where LPT-
like dispatching rules are applied, E[L(ii) ] > E[L(i) ] holds true. Moreover,
if SPT-like dispatching rules are used then E[L(iii) ] < E[L(i) ] holds true for
single- and multi-stage production systems.
cv [L] for a random dispatching rule is higher than for FIFO because the
orders are sorted randomly which has no effect on the expected value but
on the variance and therefore the coefficient of variation is increasing due to
the sorting process. The variance of the customer required lead time effects
the results of cv [L] for the dispatching rule EDD. However, in this case only
little effect on cv [L] compared to FIFO can be observed.
It is not advisable to apply dispatching rules with a high cv [L] (Random,
LPT, LPT-o, SPT, SPT-o, SRTP) because the reliability of the production
schedule is low. Moreover, the more cv [L] is increasing, the more the oper-
ating curves deviate from the ideal operating curves (Nyhuis and Wiendahl,
2009).
Based on Observation 6.2.2 it is assumed that Theorem 6.1.3 holds true
in general for a multi-stage setting and that Theorem 6.1.3 can therefore
be extended. Hence, the analytic proof needs to be investigated in further
research.
6.3. Concluding Remarks 117
Symbol Description
n Number of orders leaving the system
m Number of machines in the production system
Pα Random variable of processing time with its
respective realisation Pi,α of order i at ma-
chine α including additional random times
(e.g. set-up times) but no sequence based
times
Lα Random variable of production lead time with
its respective realisation Li,α of order i at ma-
chine α
Qα := Lα − Pα Random variable of queuing time with its re-
spective realisation Qi,α of order i at machine
α
E[Lα Pα ]
lα := ”Processing time weighted average production
E[Pα ]
lead time” at machine α according to Jodl-
bauer and Stöcher (2006)
m
P := Pα Random variable of overall processing time
α=1
m
L := Lα Random variable of overall production lead
α=1 time
m
Q := L − P = Qα Random variable of overall queuing time
α=1
E[LP ]
l := ”Processing time weighted average production
E[P ]
lead time” according to Jodlbauer and Stöcher
(2006) extended to a general production sys-
tem (single- and multi-stage production sys-
tem)
E[.] Expected value of a random variable
V ar[.] Variance of a random variable
V ar[.]
cv [.] := Coefficient of variation of a random variable
E[.]
ρ[., .] Correlation coefficient of two random variables
Cov[., .] Covariance of two random variables
6.3. Concluding Remarks 119
Deming, W. E., 2000: Out of the crisis. 1st ed., Cambridge Univ. Pr, Cam-
bridge Mass.
Edwards, J., 1983: Mrp and kanban - american style. Twenty-sixth annual
international conference proceedings, 586–603.
Emde, S., M. Fliedner, and N. Boysen, 2012: Optimally loading tow trains
for just-in-time supply of mixed-model assembly lines. IIE Transactions,
44 (2), 121–135, doi: 10.1080/0740817X.2011.575442.
Hines, P. and N. Rich, 1997: The seven value stream mapping tools. In-
ternational Journal of Operations & Production Management, 17 (1),
46–64, doi: 10.1108/01443579710157989.
Hübl, A., 2014: Selected topics of the supply chain matrix under customer
and process uncertainties. Proceedings of the 2014 Winter Simulation
Conference, 4023–4024.
Hung, Y.-F. and I.-R. Chen, 1998: A simulation study of dispatch rules
for reducing flow times in semiconductor wafer fabrication. Production
Planning & Control, 9 (7), 714–722, doi: 10.1080/095372898233722.
Kerkhofs, M., H. Chung, and P. Ester, 2008: Working time flexibility across
europe: a typology using firm-level data. Industrial Relations Journal,
39 (6), 569–585, doi: 10.1111/j.1468-2338.2008.00506.x.
Meal, H. C., 1984: Putting production decisions where they belong. Harvard
Business Review, 62 (2), 102–111.
134 Bibliography
Medhi, J., 1991: Stochastic models in queueing theory. 1st ed., Academic
Press, Boston u.a.
Merschmann, U. and U. W. Thonemann, 2011: Supply chain flexibility, un-
certainty and firm performance: An empirical analysis of german manu-
facturing firms. International Journal of Production Economics, 130 (1),
43–53, doi: 10.1016/j.ijpe.2010.10.013.
Meyr, H., 2004: Supply chain planning in the german automotive industry.
OR Spectrum, 26 (4), doi: 10.1007/s00291-004-0168-4.
Mieghem, J. A. v. and N. Rudi, 2002: Newsvendor networks: Inven-
tory management and capacity investment with discretionary activities.
Manufacturing & Service Operations Management, 4 (4), 313–335, doi:
10.1287/msom.4.4.313.5728.
Mincsovics, G. Z. and N. P. Dellaert, 2009: Workload-dependent capacity
control in production-to-order systems. IIE Transactions, 41 (10), 853–
865, doi: 10.1080/07408170802369391.
Missbauer, H., 2002: Aggregate order release planning for time-varying
demand. International Journal of Production Research, 40 (3), 699–718,
doi: 10.1080/00207540110090939.
Missbauer, H. and R. Uzsoy, 2011: Optimization models of production plan-
ning problems. Planning Production and Inventories in the Extended En-
terprise, K. G. Kempf, P. Keskinocak, and R. Uzsoy, Eds., Springer US,
Boston and MA, International Series in Operations Research & Manage-
ment Science, Vol. 151, 437–507, doi: 10.1007/978-1-4419-6485-4\ 16.
Muchiri, P. and L. Pintelon, 2008: Performance measurement using overall
equipment effectiveness (oee): literature review and practical application
discussion. International Journal of Production Research, 46 (13), 3517–
3535, doi: 10.1080/00207540601142645.
Mula, J., R. Poler, J. P. Garcı́a-Sabater, and F. C. Lario, 2006: Models for
production planning under uncertainty: A review. International Journal
of Production Economics, 103 (1), 271–285, doi: 10.1016/j.ijpe.2005.09.
001.
Nakajima, S., 1988: Introduction to TPM: Total Productive Maintenance.
Productivity Press, Cambridge Mass. u. a.
Neely, A., M. Gregory, and K. Platts, 1995: Performance measurement
system design: A literature review and research agenda. International
Journal of Operations & Production Management, 15 (4), 80–116, doi:
10.1108/01443579510083622.
Bibliography 135
Prakash, J. and J. Chin, 2014: Modified conwip systems: a review and clas-
sification. Production Planning & Control, 1–12, doi: 10.1080/09537287.
2014.898345.
Rohde, J., H. Meyr, and M. Wagner, 2000: Die supply chain planning
matrix. PPS Management, (1), 10–15.
Wiendahl, H.-P. and J.-W. Breithaupt, 1999: Modelling and controlling the
dynamics of production systems. Production Planning & Control, 10 (4),
389–401, doi: 10.1080/095372899233136.