Sie sind auf Seite 1von 134

Produktion und Logistik

Herausgegeben von
B. Fleischmann, Augsburg, Deutschland
M. Grunow, München, Deutschland
H.-O. Günther, Berlin, Deutschland
S. Helber, Hannover, Deutschland
K. Inderfurth, Magdeburg, Deutschland
H. Kopfer, Bremen, Deutschland
H. Meyr, Stuttgart, Deutschland
Th. S. Spengler, Braunschweig, Deutschland
H. Stadtler, Hamburg, Deutschland
H. Tempelmeier, Köln, Deutschland
G. Wäscher, Magdeburg, Deutschland
Diese Reihe dient der Veröffentlichung neuer Forschungsergebnisse auf den Gebieten
der Produktion und Logistik. Aufgenommen werden vor allem herausragende
quantitativ orientierte Dissertationen und Habilitationsschriften. Die Publikationen
vermitteln innovative Beiträge zur Lösung praktischer Anwendungsprobleme der
Produktion und Logistik unter Einsatz quantitativer Methoden und moderner
Informationstechnologie.

Herausgegeben von
Professor Dr. Bernhard Fleischmann Professor Dr. Herbert Meyr
Universität Augsburg Universität Hohenheim

Professor Dr. Martin Grunow Professor Dr. Thomas S. Spengler


Technische Universität München Technische Universität Braunschweig

Professor Dr. Hans-Otto Günther Professor Dr. Hartmut Stadtler


Technische Universität Berlin Universität Hamburg

Professor Dr. Stefan Helber Professor Dr. Horst Tempelmeier


Universität Hannover Universität Köln

Professor Dr. Karl Inderfurth Professor Dr. Gerhard Wäscher


Universität Magdeburg Universität Magdeburg

Professor Dr. Herbert Kopfer


Universität Bremen

Kontakt
Professor Dr. Hans-Otto Günther
Technische Universität Berlin
H 95, Straße des 17. Juni 135
10623 Berlin
Imke Mattik

Integrated Scheduling
of Continuous Casters
and Hot Strip Mills
A Block Planning Application
for the Steel Industry

Foreword by Prof. Dr. Hans-Otto Günther


Imke Mattik
Berlin, Germany

Dissertation Technische Universität Berlin, 2013

D 83

ISBN 978-3-658-03774-1 ISBN 978-3-658-03775-8 (eBook)


DOI 10.1007/978-3-658-03775-8

The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie;


detailed bibliographic data are available in the Internet at http://dnb.d-nb.de.

Library of Congress Control Number: 2013949763

Springer Gabler
© Springer Fachmedien Wiesbaden 2014
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, compu-
ter software, or by similar or dissimilar methodology now known or hereafter developed.
Exempted from this legal reservation are brief excerpts in connection with reviews or schol-
arly analysis or material supplied specifically for the purpose of being entered and executed
on a computer system, for exclusive use by the purchaser of the work. Duplication of this
publication or parts thereof is permitted only under the provisions of the Copyright Law of
the Publisher’s location, in its current version, and permission for use must always be obtained
from Springer. Permissions for use may be obtained through RightsLink at the Copyright
Clearance Center. Violations are liable to prosecution under the respective Copyright Law.
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.
While the advice and information in this book are believed to be true and accurate at the date
of publication, neither the authors nor the editors nor the publisher can accept any legal re-
sponsibility for any errors or omissions that may be made. The publisher makes no warranty,
express or implied, with respect to the material contained herein.

Printed on acid-free paper

Springer Gabler is a brand of Springer DE.


Springer DE is part of Springer Science+Business Media.
www.springer-gabler.de
Foreword V

Foreword

The global steel industry remains the backbone of the industrialized economy. One still
observes a rapid evolution of product offerings and technological innovations with
immediate applications in the steel processing industries.

In a market environment with constantly decreasing steel prices the industry is facing
two major challenges - the cost increase for energy and raw materials and the growing
concern towards CO2 emissions in developed countries. The energy costs alone, for
instance, amount to about ~ 30% of the entire steel production costs.

This book focuses on a production planning model aiming at optimizing the liquid-
solid-interface of the continuous casting and hot strip milling stage. It presents
improvement levers aiming at optimizing lead times – and therefore reducing energy
costs.

The production planning model presented in this book offers support to align the
scheduling of continuous casting and hot strip milling. The approach is practice oriented
and has the advantage of a low model complexity. Incorporating results from research
of make and pack production the block planning principle is leveraged to consider
sequence dependencies of production orders. The resulting production schedules reduce
the waiting times of orders between the continuous casters and the hot strip mills.

Prof. Dr. Hans-Otto Günther


Preface VII

Preface

As a management consultant working on production improvement projects in the steel


industry, I worked diligently to reduce energy costs. This experience gave me the
opportunity to gain insight into the industry and experience the increasing significance
of energy costs. These projects have broadened my understanding of the issues at play.
Indeed, in addition to using technological changes to significantly reduce energy costs,
improved production planning can play a major role in saving energy due to a better
integration of processes. However, the current common practices and applied systems
do not support better integration. This gave me the motivation to focus my research
project on the topic of improved production planning with the core objective to develop
a mathematical model that would lead to a better integration of two major production
facilities in steel production and test it with realistic data to prove its practical
applicability.

This work would not have been possible without the continuous support of my doctoral
supervisor, Prof. Hans-Otto Günther, head of the institute for production management at
the Technical University of Berlin. His feedback and supportive dialogue broadened my
ideas and helped me to overcome all roadblocks. I also would like to thank Dr. Pedro
Amorim for his valuable advice related to and support for programming the
optimization model. Thankfully, I was able to work directly with a large European steel
producer, which chooses to remain anonymous; this collaboration gave me first-hand
practical insight into the challenges of production planning, especially the integration of
production planning. I would like to thank the many employees of that company that
supported my work with long consultations and production tours.

Moreover, I would especially like to express my gratitude to my family, i.e., my parents


Renate and Roland for their continuous support and motivation and to my husband Tako
for accompanying me on this journey and being an excellent sparring partner whenever
I needed help. Last but not least, I would like to thank my daughter Clara for the
necessary distraction and joy that she provided during the last 1.5 years.

Imke Mattik
Contents IX

Contents

Foreword V

Preface VII

List of Figures XIII

List of Tables XV

List of Abbreviations and Akronyms XVII

1 Introduction 1

1.1 Motivation 1

1.2 Dissertation outline 3

2 Production planning and steel industry 5

2.1 Supply chain management and production planning 5

2.1.1 The concept of supply chain management 5

2.1.2 Hierarchical structures in production planning 7

2.1.3 Advanced planning systems 9

2.2 Production planning and detailed scheduling 11

2.2.1 Production planning and scheduling process 11

2.2.2 Relevant planning characteristics 13

2.2.3 Production planning principles 15

2.2.4 Quantitative models of lot sizing and scheduling 16

2.2.5 Sequencing rules 22

2.3 Steel production 22

2.3.1 Industry characteristics 22


X Contents

2.3.2 Current challenges 25

2.3.3 Production setup and processes 30

3 Scheduling concepts with focus on the steel industry 34

3.1 Planning challenges in the steel industry 34

3.1.1 Equipment specific challenges 34

3.1.2 Product portfolio challenges 35

3.1.3 Demand and scheduling characteristics 36

3.1.4 Requirements for quantitative scheduling optimization 37

3.2 Review of batching and scheduling literature with


steel industry focus 39

3.2.1 Classification of batching and scheduling models with steel


industry focus 39

3.2.2 Review of individual publications 44

3.3 Integrated and separate batching and scheduling 52

3.4 Continuous and discrete representation of time 53

3.5 Block planning principle 54

4 Scheduling of CCs and HSMs in the steel industry 57

4.1 Problem description 57

4.2 Decomposition approach 60

4.3 Model formulations 64

4.3.1 Push-Principle 64

4.3.2 Pull-principle 70

5 Computational Experiments 77
Contents XI

5.1 Experimental design 77

5.2 Results and discussion 82

5.2.1 Results for the push principle 83

5.3 Results for the pull-principle 88

5.4 Comparative analysis of the push and pull-principles 92

5.4.1 Maximum and average waiting times 93

5.4.2 Total production and setup times 97

5.4.3 Analysis of production stops 104

5.4.4 Computational effort 105

6 Conclusion 107

Appendix 111

Appendix 1: Discrete time based model formulation 111

References 113
List of Figures XIII

List of Figures

Figure 1: Global crude steel production (WSA 2012b) 1


Figure 2: Inter-company supply chain 6
Figure 3: Software modules of Advanced Planning Systems (Günther et al. 2005: 10) 10
Figure 4: Production planning and scheduling tasks (Based on Schuh (2006: 29) and
Lödding (2008: 5) 11
Figure 5: Customer order decoupling points 14
Figure 6: World crude steel production (WSA 2012b) 23
Figure 7: Top five steel producing countries and companies (WSA 2012a) 24
Figure 8: Steel consumption and customer segments (WSA 2009, King 2010) 24
Figure 9: Steel products and applications (WSA 2012b, King 2012) 25
Figure 10: Energy intensity and CO2 emissions per ton of steel (VdEh 2012) 27
Figure 11: Industry prices for energy excluding value added tax (BMWi 2012) 28
Figure 12: Supply chain of ThyssenKrupp Steel Americas
(TKSA 2010, TKSA 2012) 31
Figure 13: Steel production processes 33
Figure 14: ABC analysis of the steel grade portfolio of a large European steel
company 36
Figure 15: Decomposition structure of publications 45
Figure 16: Flexible block planning concept 55
Figure 17: Production process from CC to HSM (Lopez et al. 1998) 57
Figure 18: "Coffin rule" for HSM scheduling 58
Figure 19: Temperature-time diagram of steel slabs based on Tata Steel
(Wachter 2011: 74) 59
Figure 20: Decomposition approach – The push-principle 61
Figure 21: Decomposition approach – The pull-principle 62
Figure 22: Process flow from the CC to the HSM (i.e., the push-direction) 64
Figure 23: Distribution of T(e) 79
Figure 24: Determination of T(b) and B(t) 80
Figure 25: Production breaks at the hot strip mill (pull principle) 81
XIV List of Figures

Figure 26: Gantt-Chart of one problem instance (push-principle) 88


Figure 27: List of analyses 93
Figure 28: Analysis of the maximum waiting times 94
Figure 29: Deviation of maximum waiting times 95
Figure 30: Analysis of the average waiting times 96
Figure 31: Makespan analysis for the continuous caster (push-principle) 97
Figure 32: Makespan analysis of the CC for high workload scenarios
(push-principle) 98
Figure 33: Makespan analysis for the continuous caster (pull-principle) 99
Figure 34: Summary of the makespan analyses of the CC stage 100
Figure 35: Makespan analysis of the hot strip mill (push-principle) 101
Figure 36: Makespan analysis of the hot strip mill (pull-principle) 102
Figure 37: Summary of makespan analyses of the HSM stage 103
Figure 38: Analysis of production stops 105
Figure 39: Analysis of computation times 106
List of Tables XV

List of Tables

Table 1: Batching and scheduling literature with steel industry focus 41


Table 2: Estimated energy savings potential 59
Table 3: Overview of all scenarios 82
Table 4: Number of variables in the CC model (push-principle) 83
Table 5: Number of variables HSM model (push-principle) 84
Table 6: Results of the continuous caster model (push-principle) 85
Table 7: Results of the hot strip mill model (push-principle) 87
Table 8: Number of variables in the HSM model (pull-principle) 88
Table 9: Number of variables for the CC model (pull-principle) 89
Table 10: Numerical results for the HSM model (pull-principle) 90
Table 11: Numerical results for the CC model (pull-principle) 92
List of Abbreviations and Akronyms XVII

List of Abbreviations and Akronyms

AOD Argon Oxygen Decarbonization


APS Advanced Planning System
BF Blast Furnace
BOF Basic Oxygen Furnace
CC Continuous Caster
CLSP Capacitated Lot Sizing Problem
DLSP Discrete Lot Sizing and Scheduling Problem
EAF Electric Arc Furnace
ETS Emissions Trading Scheme
HSM Hot Strip Mill
MILP Mixed-Integer Linear Programming
MRP Material Requirements Planning
MS Melt Shop
OEM Original Equipment Manufacturer
PP/DS Production Planning/Detailed Scheduling
SCM Supply Chain Management
1 Introduction 1

1 Introduction

1.1 Motivation
The steel industry is a significant industry with global revenues of approximately
US$3,000 billion in 2011 (Global-Insights 2012). Regionally, steel production is located
mostly in Asia (i.e., especially China, Japan, and Korea), North America, and Europe as
illustrated in figure 1. Interestingly, steel production in North America and the European
Union is basically stable or slowly shrinking, whereas Asia’s steel production,
especially in China, is growing rapidly. This trend is also evidenced by the nation of
origin for the largest steel producers; i.e., of the top five global steel producers, three are
from China, and the remaining two are from Europe and South Korea, respectively
(WSA 2012a).

Global crude steel production on regional level1


In Megatonnes (mt), 2011, (CAGR from 2000)

CIS
112 mt (1%)
North America
119 mt (-1%) EU-27
Other Europe
177 mt (-1%)
15 mt (8%)

Middle East
20 mt (6%)
Asia & Oceania
961 mt (10%)
Africa
South America 14 mt (0%)
48 mt (2%)

1 Total of 64 countries that account for approximately 98% of total global crude steel production

Figure 1: Global crude steel production (WSA 2012b)

The steel industry is the backbone for other key industries, such as the construction,
engineering, and automotive industries (King 2010). For example, in Germany, the steel
industry has the highest multiplying effect of all other industries; i.e., 1 EUR of demand

I. Mattik, Integrated Scheduling of Continuous Casters and Hot Strip Mills, Produktion und Logistik,
DOI 10.1007/978-3-658-03775-8_1, © Springer Fachmedien Wiesbaden 2014
2 1 Introduction

translates into 2.73 EUR of total demand for the steel industry and suppliers (Booz&Co
2012, RWI 2011).
However, today’s steel industry, especially in Western Europe, faces significant
challenges. First of all, since Western European steel producers do not have access to
native sources of most of the necessary raw materials, they are dependent on other
sources for iron ore, and coke, all of which have high and still rising raw material costs
(BCG 2009). The other major cost contributor in steel production, i.e., the energy costs,
are also increasing, especially in Germany (BMWi 2012) and the rest of Europe due to
CO2 regulation (Coelho 2010). Due to the former cost contributor, steel producers are
forced to defend their profit margin against the growing demands of the raw material
industry, which benefits from the current oligopoly in this arena (i.e., a low number of
suppliers with many customers in the raw material market). On the other hand, one of
the main customers for European steel companies, i.e., the automotive industry, expects
innovative and high quality products, on time delivery (Cowling et al. 2004: 179), and
minimal costs so that they in turn can reduce the material costs for each car that they
produce. Nevertheless, the raw material costs for steel production increased
significantly with price increases for iron ore of more than 100% in 2011 compared to
2008; in contrast, coking coal and scrap prices increased moderately by 16% and 9%,
respectively, in 2011 as compared to that of 2008 (SBB 2012). Moreover, the price for a
hot rolled coil decreased by 16% in 2011 as compared to that of 2008 (MEPS 2012).

With the steel industry facing challenges on both the raw material supply side and the
customer side, this industry clearly needs to improve operations in order to lower costs
(Melouk et al. 2013: 269), e.g., by developing computer integrated manufacturing
systems (Tang et al. 2000a, Kumar et al. 2006, Bellabdaoui and Teghem 2006).
Moreover, one study indicated a clear demand for scheduling tools (Cowling 2003)
because planning practices often still rely on manual operations. As previously implied,
both of the major cost contributors in steel production, i.e., raw material and energy
costs, have significantly increased. Over the last two decades (i.e., 2011 compared to
1990), the steel industry has already improved specific energy consumption and CO2
emissions by 21% and 22%, respectively (VdEh 2012) partially due to the introduction
of the continuous casting technology (Box and Herbe Jr 1988) and larger and more
1 Introduction 3

efficient blast furnaces (BF) (VdEh 2011). Despite these large improvements, additional
approaches are still needed in order to reduce the consumption of raw materials and
energy even more. An energy saving opportunity that has not been implemented in large
parts of the industry to date is an improved integration of the two production steps,
continuous casting and hot strip milling. Shorter waiting times for the continuously
casted slabs before they are charged into the hot strip mill (HSM) furnaces does not
allow the slabs to fully cool down, thereby avoiding the need to fully reheat the slab for
the HSM process and hence saving energy costs (Cowling and Rezig 2000, Tang et al.
2001).

The objective of this work is to present a new modeling approach to improve the
integration of continuous casters (CCs) and HSMs, thereby reducing the overall energy
costs. The mixed-integer linear programming (MILP) models that are presented herein
leverage the block planning principle proposed by Günther et al. (2006), which reduces
the number of necessary variables to a level that still allows the application of standard
optimization software and fast computation times. To date, integrated approaches are
still not implemented often due to the challenges facing production planning when
attempting to align the scheduling of the two production steps, each of which imposes
different conditions affecting the development of a feasible production schedule.
Moreover, research has often focused on the separate approaches of either continuous
casting scheduling or HSM scheduling or the separate sequencing approaches for the
HSM (e.g., Lopez et al. (1998) and Tang et al. (2000b)) and for the CC process (e.g.,
Tang et al. (2000a) and Dong et al. (2012)).

1.2 Dissertation outline


This dissertation consists of 5 chapters beyond the introduction. Chapter 2 gives an
overview of the basic concepts that build the foundation of production planning and
scheduling research, such as the concept of supply chain management (SCM), the
hierarchical structure of production planning, and the increasingly important role of
advanced planning systems (APSs). After establishing this foundation, the role of
production planning and scheduling is reviewed; more specifically, the process and
production planning characteristics and the two major production planning principles,
4 1 Introduction

i.e., the push and the pull principle are discussed. Additionally, the development of
quantitative models for lot sizing and scheduling is presented, and major sequencing
rules and their effect on production planning are discussed. Lastly, an industry overview
of the steel industry is presented, including today’s challenges and an introduction to the
production setup and processes.

Chapter 3 deals with scheduling concepts in the steel industry. First, the current
scheduling challenges facing the steel industry are discussed, followed by a review of
36 models from the literature the focus on the applicability of these models in practice.
A more detailed discussion of two specific modeling approaches is given. More
specifically, the advantages and disadvantages of an integrated or separate approach to
batching and scheduling is presented, and the choice of the representation of time, i.e.,
either continuous or discrete, is reviewed. Furthermore the in this work applied block
planning principle is introduced and discussed.

In chapter 4, the problematic situation of an integrated scheduling approach for CCs and
HSMs is described. In order to address the challenges present in this situation and the
steel industry in general, two MILP models are developed and presented. The chosen
decomposition approach suggests two opposed model formulations, i.e., one on the
basis of the push-principle and the other based on the pull-principle.

Chapter 5 covers the numerical design with which both MILP formulations are tested
and presents the results of both models. Additionally, comparative analyses of the push
and the pull model are carried out and discussed. Moreover, conclusions and the
summary of the key findings are presented in the last chapter, i.e., chapter 6.
2 Production planning and steel industry 5

2 Production planning and steel industry

2.1 Supply chain management and production planning

2.1.1 The concept of supply chain management


Various definitions for the concept of a supply chain and the corresponding SCM exist.
Common themes in all of these definitions are that a supply chain usually spans
companies or organizational units of a single company (i.e., from the raw material
supplier to the manufacturer to retailers) and the overarching aim of each supply chain
is customer satisfaction (Chandra and Grabis 2007: 20, Govil and Proth 2002: 7,
Christopher 2005: 5, Chopra and Meindl 2007: 3). Stadtler (2005, P.576) defined SCM
as “the task of integrating organizational units along a supply chain and coordinating
materials, information, and financial flows in order to fulfill customer demands with the
aim of improving competitiveness of the supply chain as a whole. Most authors define
the organizational units of an inter-company supply chain as the suppliers,
manufacturers, warehouses, distributors, and retailers (Chandra and Grabis 2007: 18,
Chopra and Meindl 2007: 3). However, in today’s large international companies, which
face such issues as multi-site production planning, intra-company supply chains can
bear the same amount of challenges (Günther 2005: 5). Other scholars (e.g., Werner
2010: 6) also include the disposal and recycling of the resulting waste after production
in the supply chain. An example of an inter-company supply chain is illustrated in
figure 2. Note that a supply chain is often actually considered a supply chain network
(Christopher 2005: 5) due to the number of entities at each step that leads to diverging
product and information flows.

I. Mattik, Integrated Scheduling of Continuous Casters and Hot Strip Mills, Produktion und Logistik,
DOI 10.1007/978-3-658-03775-8_2, © Springer Fachmedien Wiesbaden 2014
6 2 Production planning and steel industry

End-product Distribution
n-tier supplier … 1st-tier supplier Customer
production site center

Figure 2: Inter-company supply chain


A supply chain is a global network of organizations that cooperate to
A company’s supply chain can be a source of competitive advantage since individual
businesses often do not compete against each other as standalone entities but rather as
supply chains (Christopher 2005: 284). Additionally, a world where production and
demand becomes increasingly globally fueled by factor cost differences, high growth in
emerging markets, lower transaction costs, less trade barriers (Jacob and Strube 2008:
6), and increasing product variety and complexity, the task of SCM becomes
increasingly important and complex. According to Chopra and Meindl (2007: 9), the
decisions of SCM can be delineated into the three following categories:

1. Supply chain strategy and design: In this phase, a company decides on the
structure and setup of a supply chain for a specific product or a product family
for the next several years or the product lifecycle. Moreover, decisions on what
is carried out in-house and what is outsourced are made. If the decisions are
made to outsource, the locations and capacities of production facilities are
chosen, and the warehouses and distribution network are setup. Additionally,
decisions about the information and planning systems are also made. Aim of
cooperation is defined in this phase in that each partner benefits at least on most
terms; therefore, transaction prices play a crucial role (Stadtler 2005: 577).
Although all supply chain partners hope to benefit from the relationship, one
player, i.e., often the most powerful or biggest company, typically dominates.
Looking to the automotive industry as an example, the automotive original
2 Production planning and steel industry 7

equipment manufacturer (OEM) usually dominates the whole supply chain of


raw material suppliers, component suppliers, and sales agencies (Govil and
Proth 2002: 8).

2. Supply chain planning: This phase has a much shorter timeframe, i.e., usually up
to a year, and considers the configuration decisions that are carried out in the
strategic phase as fixed. This phase aims at maximizing the supply chain’s profit
over the planning horizon via the implementation of effective planning
decisions. In this phase, the demand is forecasted to make decisions on which
markets are supplied from which locations, if work needs to be subcontracted,
and which inventory policies are necessary. The demand scenario is also the
basis for marketing and prize promotions. This illustrates that the planning phase
establishes the parameters within which the supply chain will function. The
result of this phase is a set of operating policies that govern short-term
operations.
3. Supply chain operation: This phase has a very short timeframe of a week or a
day, and during this phase, companies make decisions regarding single customer
orders. The aim of this phase is to handle customer orders using the most
effective means, i.e., by effective production planning, scheduling of production,
and timely distribution to the customer.

This study focuses on creating effective scheduling models for steel production, and it
clearly falls into the category of supply chain operation. Although the decisions made in
this phase have a short time horizon, they are crucial to the success of the supply chain.
Without a high level of operational excellence, a supply chain cannot be competitive.

2.1.2 Hierarchical structures in production planning


Considering the plurality of topics that can be subsumed under production planning
activities (e.g., investment/divestments of plants, clustering of product families,
allocation of products to plants, and detailed scheduling), we can readily see why
companies usually deconstruct the overall production planning problem into different
groups or organizational units. While companies can adopt many deconstruction
approaches, the production planning is often categorized based on their time horizon
8 2 Production planning and steel industry

and their importance to the company. In 1975, Hax and Meal introduced a production
planning approach that split up the planning problem of a company into four
hierarchical levels of decision making. Their first hierarchical level comprises the
assignment of products to multiple plants and long term investment decisions to
increase capacity. At the second hierarchical level, the capacity of each plant is
allocated to product families, and the accumulation of seasonal stock is planned based
on a monthly demand forecast. The third hierarchical level comprises the determination
of monthly schedules for each product type considering stock levels. At the fourth and
last level, the time allocation per product family is shared among the individual products
of the family with a planning horizon of one month (Hax and Meal 1975: 59f).

Although Hax and Meal (1975) developed one of the first approaches, many other
approaches for the deconstruction of the production planning problem have since been
developed (Schneeweiß 1995, Kletti 2007, Günther and Tempelmeier 2012, Dyckhoff
and Spengler 2010). The deconstruction into hierarchical levels is driven by the
problem’s high level of complexity when solved as a single optimization question. Such
a problem size is usually not manageable for an optimization model or a single person.
With no other alternative, the potential risk of a developing a less than optimal solution
when the problem is tackled in successive steps is irrelevant. Additionally, companies
already have hierarchical structures; therefore, such an approach often fits well into
existing organizational and decision making structures and triggers management
involvement (Hax and Meal 1975: 59, Schneeweiß 1995: 9).

Nowadays, the framework that is most commonly applied to the deconstruction of


planning activities is the use of three hierarchical levels that range from strategic to
operational planning (Günther and Tempelmeier 2012: 26, Dyckhoff and Spengler
2010: 30). In the context of production planning, these planning levels can be defined as
follows:

x Strategic planning focuses on creating a competitive position for the company


and preserving that position in a changing market environment. The planning
horizon of strategic planning is about 5 years. Part of strategic planning includes
the development of a global production footprint, including investment and
divestment decisions.
2 Production planning and steel industry 9

x Tactical planning focuses on developing strategic initiatives for implementation.


The time horizon for tactical planning activities is about one to five years.
Typical questions considered include the allocation of products to plants or
capacity adjustments within the existing production network.

x Operational planning focuses on the efficient operation of the existing


production system. Planning activities have a planning horizon of up to one year.
Typically, these activities include the detailed production scheduling, inventory
control, and lot sizing.

Lin and Moodie (1989) and Missbauer et al. (2009) developed hierarchical planning
approaches for steel mills to optimize operational production planning. In contrast,
Nickel et al. (2005) developed a mathematical model to aid in strategic decision making
regarding the supply network design of a leading international steel company.
Publications that focus on operational planning activities are classified in section 3.2.1.
Since this work develops an approach to optimize scheduling in the steel industry, it
clearly focuses on the operational level of production planning

2.1.3 Advanced planning systems


Advanced Planning Systems (APS) aid companies with SCM and production planning
from the strategic to operational levels. The main characteristics of an APS, which are
also the main differences from classic material requirements planning (MRP) systems,
are as follows (Stadtler and Kilger 2008: 83):
x Integral planning signifies that the APS encompasses the supply chain as a
whole.
x True optimization is carried out by defining objectives and constraints properly
and applying either exact optimization techniques or heuristics.
x A hierarchical planning system (i.e., as described in section 2.1.2) of the APS is
necessary to achieve true optimization for the whole supply chain. Therefore, the
APS defines a number of modules along the supply chain and with varying
planning horizons that are interdependent; i.e., they are linked by information
flows.
In contrast to the modern APS, traditional MRP systems do not have any of the above
properties. They are in most cases restricted to the production and procurement area, are
10 2 Production planning and steel industry

not focused on optimization, and are uncapacitated based on lead times and production
lots that neglect the workload and plan successively from the level of the master
production schedule to detailed scheduling (Stadtler and Kilger 2008: 84, Günther et al.
2006: 3720). Furthermore, unlike APSs, MRP systems usually neglect industry specific
characteristics, such as shelf life or batching restrictions (Fleischmann 2012: 18,
Knolmayer et al. 2009: 105). Figure 3 illustrates the planning modules that are usually
found in an APS.

Procurement Production Distribution Sales

Long-term Strategic Network Design

Demand
Planning
Mid-term Supply Network Planning

Production Transportation
External Planning/ Planning/ Order
Short-term Fulfillment
Procurement Detailed Vehicle
Scheduling Scheduling and
ATP/CTP

Figure 3: Software modules of Advanced Planning Systems (Günther et al. 2005: 10)

This work focuses on planning activities that relate to the module of production
planning/detailed scheduling (PP/DS); therefore, this module is explained in more
detail. The module’s main objective is to determine feasible production plans; thus, it
focuses on operative planning and is restricted to an individual plant in most cases. The
module determines the lot sizes if they were not addressed in mid-term planning,
followed by the determination of the exact sequences and timing (Knolmayer et al.
2009: 100, Stadtler and Sürie 2012: 149) using a finite and thus capacity restricted
scheduling policy (Günther et al. 2006: 3721).

Günther et al. (2006) applied the block planning approach using the APS system from
SAP, SAP APO©. Their investigation focused on a make-and-pack problem from the
consumer goods industry. They showed that the PP/DS module of SAP APO© is well-
suited to support the block planning approach, but their additionally presented MILP
optimization was computationally even more efficient (Günther et al. 2006).
2 Production planning and steel industry 11

Additionally, there are reports about successful implementations of APSs in the steel
industry. The company Deutsche Edelstahlwerke, a mid-size producer of stainless steel
with a yearly production of about 700,000 tons in 2010 (DEW 2012), implemented the
factory planner and supply chain planner by JDA (i.e., formerly known as i2
technologies) and observed benefits regarding a reduction in the order processing time
by 80%, reduction in work-in-progress inventories and raw materials by 20%,
improvement of on-time deliveries from 60% to 85%, and improvement in the lead time
by 20% (JDA 2012).

2.2 Production planning and detailed scheduling

2.2.1 Production planning and scheduling process


Within the scope of operational planning, production planning and scheduling is
concerned with the planning and control of production processes related to the timely
production of products in the appropriate quantities within capacity restrictions (Schuh
2006: 28, Günther and Tempelmeier 2012: 147). To illustrate the tasks of production
planning, the RWTH Aachen used extensive experience with consulting projects to
develop a production planning and scheduling framework intended to illustrate relevant
interdependencies (Schuh 2006: 11). Figure 4 illustrates this framework, which divides
production planning and scheduling into core and cross-sectional tasks.

Core tasks Cross-sectional tasks

Master production schedule


Inventory management

Material requirements
Order processing

planning
Controlling

Outsourced Inhouse:
production: Planning &
Planning & scheduling
scheduling

Data management

Figure 4: Production planning and scheduling tasks (Based on Schuh (2006: 29) and
Lödding (2008: 5)
12 2 Production planning and steel industry

According to this framework, the core tasks include the master production schedule, the
material requirements planning, and the planning and scheduling of outsourced and in-
house production. Master production schedule determines which products in what
quantity need to be produced during the upcoming planning periods. Material
requirements planning calculates the material and resources required for production,
including the secondary requirements of components and parts, based on the output of
the master production schedule and current inventories. It plans the production orders
and determines the capacity need. The next step, production planning and scheduling, is
divided into outsourced and in-house production. The planning and scheduling tasks for
outsourced production usually consist of determining the order quantities and supplier
selection. Planning and scheduling for in-house production usually involves the
determination of lot sizes and production sequences with detailed start and end times
depending on the production characteristics (Lödding 2008: 6, Schuh 2006: 37f, Slack
et al. 2010: 422f).

The cross-sectional tasks include order processing, inventory management, and


controlling of planning and scheduling. Order processing coordinates and monitors the
customer orders across different business units, e.g., production, distribution, and
accounts payable, until they are sent to the customers. Inventory management tracks
products and components in stock and ensures that stock levels are in line with company
requirements, e.g., seasonal demand. Controlling monitors the target achievement with
a set of key performance indicators. Core and cross-sectional activities depend on
effective and efficient data management, which includes planning systems (Lödding
2008: 6, Schuh 2006: 58ff).

The objectives of production planning and scheduling may vary along the triangle of
quality, time, and costs, but some criteria are common to many production entities,
including the following (Dickersbach 2006: 238, Schuh 2006: 28):

x meeting demands within production deadlines,


x high and balanced resource utilization,
x short lead times with low setup efforts,
x low stock and work in progress levels, and
2 Production planning and steel industry 13

x high flexibility.
Obviously, the values of performance indicators and objectives may vary depending on
the characteristics of the production system in question. A short lead time for one
production system, e.g., the production of silicon wafers with a lead time of 35 days,
may be extensively long as compared to another, e.g., the production of cars within 3 to
5 days. Therefore, the characteristics of a production system obviously need to be
clearly considered in order to be able to understand the associated production planning
and scheduling challenges.

2.2.2 Relevant planning characteristics


In order to meet the production planning objectives discussed in the previous section,
numerous characteristics of the production planning problem must be considered.
Hence, no one production planning solution works for all kinds of production
environments. The criteria that must be considered can be grouped into product specific
characteristics, production specific characteristics, and demand specific characteristics
(Günther and Tempelmeier 2012: 10f, Lödding 2008: 98ff, Dickersbach 2006: 239).

Product specific characteristics focus on the specification of the product, i.e., whether it
consists of multiple parts or just one, if it is mobile or not, and the shape of the product
(e.g., fluids or parts). These characteristics significantly influence the structure of the
production system and also the applicability of chosen production planning methods.

Production specific characteristics are concerned with the type of production system
used in the manufacture of the product. The following aspects determine the chosen
production system and influence the planning methods:

x The number of variants that are produced in the same production run. This
impacts production planning, particularly if the setup times are long and/or
sequence dependent.

x The production quantity of each variant. Generally, as the production quantity


increase and the number of variants decrease, the production system in place is
increasingly more automated, e.g., the mass production of razor blades versus the
production of prototypes in a job shop environment.
14 2 Production planning and steel industry

x The complexity of the materials or parts flow for the product system. On a mass
or serial production line, the production steps are usually connected to allow for
continuous production flow, whereas in a job shop environment, various
production facilities are grouped into workshops (e.g. stamping and painting), and
the product needs to be transported from one workshop to the next; therefore,
each product may have a different flow rate. The materials for the parts flow
significantly influence the inventory levels and lead times of a production
process. Transport modes may range from one-piece flow in an automated serial
production process to lot by lot transport in a job shop environment.

x Time constraints between production steps (e.g., necessary waiting times from
one step to the next) where production is not allowed to stop in between the steps.

Demand specific characteristics describe how the production system relates to the
customer. With the help of the customer order decoupling point, different types of
customer relationships are classified. The customer order decoupling point splits the
upstream processes, which are forecast-driven, from the downstream processes, which
react to the customer orders (Stadtler and Fleischmann 2012: 31) as illustrated in figure
5. If the customer orders goods from stock, the production system is called “make-to-
stock” because the production is completely forecast-based and customer orders are
fulfilled from stock. If products are assembled to order, parts are pre-produced based on
demand forecasts and assembled based on customer orders. If a product is made to
order, production starts after the customer order arrives, and if a product is engineered
to order, the product is designed based on a customer order, e.g., ship building.
Customer order
decoupling point

Engineering Production Assembly Customer

Make to
stock

Assemble
to order

Make
to order

Engineer
to order

Figure 5: Customer order decoupling points


2 Production planning and steel industry 15

Apparently, the position where the production becomes order based depends largely on
the lead time and customer expectations. For production planning purposes, this is
highly relevant because this position determines to what extent production planning
depends on forecasts or real customer orders.

2.2.3 Production planning principles


There are two differing basic production planning principles, i.e., the push and the pull
system. The push principle is the underlying concept for most production planning and
scheduling processes and tools as described in Section 2.2.1. The described process
starts with the master production schedule, which determines the quantities of products
to be produced in which planning period. The next step, i.e., the material requirements
planning, calculates the required materials as quoted in the bill of materials and
resources, including the secondary requirements of components and parts, based on the
output of the master production schedule for each production step. The last step, i.e.,
production planning and scheduling, determines the lot sizes and production sequences
with detailed start and end times within a very short planning horizon (e.g., on a daily
basis). Ultimately, production planning using the push principle attempts to pre-plan the
exact start and end times for all production operations in a given planning horizon
(Günther and Tempelmeier 2012: 343) and implies that the production quantities at each
level are pushed to the next level based on the chosen mode of transportation. The push
method is largely criticized since it often leads to high inventory levels in practice
(Christopher 2005: 124, Günther and Tempelmeier 2012: 343, Kletti and Schumacher
2011: 57).

A fundamentally different approach to production planning is the so-called pull


principle. With this concept, the downstream production and logistic processes are
directly or indirectly initiated by an internal or external order of an end product
(Gackstatter et al. 2011: 208). Each downstream process sends an internal production
order to the next downstream process when the former removes a part that is necessary
for production out of inventory. The downstream process will itself start to produce this
part to restock the inventory. This process continues on to the most downstream process
step, e.g., the raw material inventory, within a company or supply chain. This concept
ensures low inventory levels, makes process errors transparent, and is deeply rooted in
16 2 Production planning and steel industry

Japanese production philosophy, which considers holding inventory to be wasteful and


developed pull systems to reduce that waste (Christopher 2005: 129f). The most
renowned implementation of a pull system is called the kanban method. This method
was developed in the 1960’s as part of the Toyota Production System (Kletti and
Schumacher 2011: 20) and works as a communication method within a production
system. The kanban sign indicates the next downstream process to start production.
Very often, this communication sign is a card that is transported as a production order
from one production step to the previous downstream production step; therefore, this
method was named kanban, which is the Japanese word for card, sign, or placard. The
great advantage of this technique is the low need for controlling and systems support as
soon as the production environment is organized as a process flow (Kletti and
Schumacher 2011: 20, Liker and Meier 2006: 95f). Hybrid production planning
strategies combine advantages of the push and pull principles. A hybrid approach is e.g.
developed by Karrer et al. (2012), who design a production control system that
outperforms pure push or pull strategies especially in environments with high demand
variability and forecast error.

2.2.4 Quantitative models of lot sizing and scheduling


Operations research focuses on the development of models and algorithms for decision
support and optimization. This stream of research was founded in Great Britain and the
USA in the 1940’s in order to optimize logistics during the Second World War
(Domschke and Drexl 2005: 2f, Wöhe 2002: 133). Usually operations research is split
into the branches of optimization models and simulation models. Simulation models are
often complex optimization problems that cannot be solved with analytical algorithms;
therefore, simulation runs of alternatives are conducted. This work focuses on
optimization models that focus on short term production planning.

One of the first models in this area was the famous Wagner-Whitin model published in
1958 that optimizes the lot sizes of an uncapacitated single product production system
with time-varying dynamic demand. The model formulation is as follows ( as cited from
Jans and Degraeve (2008: 1622) and Aggarwal and Park (1993: 550f):

T
Min ¦ (sct yt  hct st ) (1)
t 1
2 Production planning and steel industry 17

subject to

st 1  xt d t  st tϵT (2)

xt d sdtk ˜ yt tϵT (3)

xt , st t 0; yt  ^0;1` tϵT (4)

Similar to the economic order quantity formula, the Wagner-Whitin model optimizes
the trade-off between setup and inventory holding costs and, therefore, has three key
variables in each period t, i.e., the production level xt, the binary setup variable yt, and
inventory variable st. The key variables yt, and st. have an associated cost. Indeed, sct the
cost of a setup operation yt, and hct the holding cost of each item st in stock. T is the
number of periods in the planning horizon. The demand of each period is known (i.e.,
dt), and sdtk is the cumulative demand for period t until k. The stock at the beginning of
the first period is zero (s0=0), and the final inventory at the end of the planning horizon
is also zero (sT=0). The objective is to minimize the total cost of production, setup, and
inventory (1). Excess production quantities are carried over as inventory to the next
period (2). According to (3) the lot size is enforced to zero if not corresponding setup
operation is performed. Constraint (3) also ensures that production is limited by the
cumulative demand up to that period. The inventory at the end of the planning horizon
must be 0. The variable domains of the production and inventory variables are positive,
and the setup variable is binary (4).

Of course, this early optimization model of lot sizes lacked practical applicability since
companies usually produce more than one product and rarely have unlimited production
capacity. The following two later models addressed these shortcomings: the Capacitated
Lot Sizing Problem (CLSP) and the Discrete Lot Sizing and Scheduling Problem
(DLSP).

In contrast to the Wagner-Whitin model, the CLSP allows for a multi-product


production system and includes capacity restrictions. The notation is as follows (as cited
in Drexl and Kimms (1997: 224) and Jans and Degraeve (2008: 5):

J T
Min ¦¦ ( sc j y jt  hc j s jt ) (5)
j 1 t 1
18 2 Production planning and steel industry

subject to

s jt 1  x jt d jt  s jt j ϵ J, t ϵ T (6)

vc j ˜ x jt d Ct ˜ y jt j ϵ J, t ϵ T (7)

¦ vc
j 1
j ˜ x jt d Ct tϵT (8)

x jt , s jt t 0; y jt  ^0;1` j ϵ J, t ϵ T (9)

Like the Wagner-Whitin model, the CLSP minimizes the total costs, i.e., the setup and
inventory holding costs of production (5). Moreover, with the CLSP, multi-product
environments can be optimized with the index j to identify the item indicated. For each
item, the demand per period must be fulfilled either from stock or the actual production
quantity (6). The parameter vcj is the time associated with the production of one item of
xjt and Ct is the capacity of the production system per period. The setup constraint (7)
ensures that setup variable y jt equals 1 if the production quantity x jt is greater than

zero. Constraint (8) represents the main difference from the uncapacitated Wagner-
Whitin model, i.e., the addition of a capacity restriction; hence, the production time of
all products produced within a given time period must be smaller or equal to the
capacity of that period. The variable domains of the production and inventory variables
are positive, and the setup variable is binary (9).

The CLSP is a typical example of a so-called big bucket model, which means that each
machine can produce several items or products in the same period. Hence, big bucket
models only allow conclusions on optimal lot sizes for each product in each period but
do not allow any conclusions on optimal sequences since the sequences within a period
are not determined. A different kind of model, the small bucket model, allows for
schedule optimization since only one setup operation is allowed per period.

The DLSP is a typical example of the small bucket model. The notation is as follows (as
cited in Drexl and Kimms (1997: 225) and Jans and Degraeve (2008: 5f):

J T
Min ¦¦ ( sc j z jt  hc j s jt ) (10)
j 1 t 1
2 Production planning and steel industry 19

subject to

s jt 1  x jt d jt  s jt j ϵ J, t ϵ T (11)

vc j ˜ x jt Ct ˜ y jt j ϵ J, t ϵ T (12)

¦y
j 1
jt d1 tϵT (13)

z jt t y jt  y jt 1 j ϵ J, t ϵ T (14)

s jt , x jt t 0; y jt , z jt  ^0;1` j ϵ J, t ϵ T (15)

The DLSP minimizes total costs, i.e., the setup and inventory holding costs, of a multi-
product production system (10). Since the index j is used to identify the item or product,
production systems with multiple products can be optimized. Each item needs to fulfill
the demand balance equation (11). The setup constraint (12) ensures that if production
occurs within a given period, it must be at full capacity. In each period, not more than
one product type can be produced (13). The new variable z jt is a start-up variable that

indicates if a setup state yjt has changed from one period to the next (14). The variable
domains of the production and inventory variables are positive, and the setup variables
are binary (15).

Since the DLSP only allows the production of one type of product in one period, the
optimization results also provide for scheduling optimization because the sequence is
determined at the same time. Hence, the DLSP allows the more detailed modeling of
capacitated lot sizing (Fleischmann 1990: 347). Although this model generation already
includes more realistic conditions, the overall research aim was to include even more
operational and scheduling issues that could exist in industrial environments (Jans and
Degraeve 2008: 1625f). A move towards more practical applicability was the
introduction of continuous time grids, which allowed events to take place at any time on
a pre-defined planning horizon instead of only at pre-defined time points with a discrete
time grid (Méndez et al. 2006: 918).

Other general model characteristics can be categorized into the following four topics:
the objective function, the planning horizon, the production levels, and setups.
20 2 Production planning and steel industry

Objective function

The choice of the objective function for any model must be carefully considered. At the
detailed planning level, the basic model formulations presented earlier in this section,
i.e., the Wagner-Whitin, CLSP, and DLSP, each chose an objective function that
minimized setup and inventory holding costs. Both of these cost positions differ from
situation to situation due to the lack of clear definition associated with each. For
example, the setup costs could be defined as the cost of the wages for the workers that
are needed to carry out the setup. Another option would be to include the capital costs
of the machines or the opportunity costs due to the contribution margin of the lost
production. Another difficulty appears if the setup length depends on the production
sequence; hence, the exact costs cannot be estimated before the production sequence of
the model is determined. Apparently, cost objective functions are not easily defined for
short-term production planning models; therefore, rather technical aims are often
considered. Amongst others, these include the minimization of makespan, production as
late as possible in order to minimize inventories, and minimization of unfilled demand
(Günther et al. 2006: 3719).

Planning horizon
The length of the planning horizon largely depends on the type of model considered.
Long-term optimization questions (e.g., supply chain network optimization) apparently
need a longer planning horizon in order to be able to evaluate potential investment
decisions properly. For short-term planning optimizations, such as lot sizing and
scheduling optimizations, a much shorter planning horizon (e.g., a few days up to a
month) is usually sufficient. The exact length depends on several factors, including how
long in advance the company has secured customer orders, the size of the product
portfolio, and the difference between the minimum batch size and the average order
quantity. As the product portfolio increases, the planning horizon of short-term models
usually increases to allow batching of orders, especially if the difference between the
minimum batch size and the average order quantity is large.

Production levels
Real world production facilities generally use multiple stages to process a product from
the raw materials or components to the end product stage. In the case of an integrated
2 Production planning and steel industry 21

steel mill, all stages from iron ore and coal up to the end product (i.e., the final cold-
rolled coil), are present in one location. However, a detailed scheduling model that
encompasses multiple stages is a large combinatorial problem demanding computational
effort that exponentially increase with the number of processing stages included
(Marchetti and Cerdá 2009: 2733). Therefore, we must consider only the most relevant
production levels in an optimization model without reducing the practical applicability
of the model. Sometimes, scheduling only the bottleneck stage, if one exists, and using
the optimized sequence for the bottleneck stage on every other stage is a sensible
approach (Marchetti and Cerdá 2009: 2733); other options include only integrating the
scheduling of the production steps that offer the highest benefits.

Setups
In many industries, the length of a setup operation depends on the sequence of products.
This is especially true in the chemical industry, the make-pack processes of the
consumer goods industry, and some steel production facilities. In a make-pack
production situation, a product with various flavors (e.g., fruit yogurt) is packed on the
same filling lines. If a product change takes place (e.g., a yogurt with a mild taste and
color to a yogurt with a strong taste and color), changes need to be made to the cleaning
process for the filling line (Lütke Entrup et al. 2005: 5075). Models in these
environments often distinguish between major and minor setup processes. A major
setup process can take several hours, e.g., the cleaning and sterilization of a filling line
after the produced product has a dark color and strong taste. In contrast, a minor setup
takes place if production is changed to a product with a high compatibility and only
includes, e.g., the provision of materials so that the setup length can also be neglected in
some cases (Bilgen and Günther 2009: 9). The importance of sequence dependent
setups in practice has led to extended basic model formulations; e.g., Jordan and Drexl
(1998) extended the basic DLSP to include sequence dependent setup times. Another
strategy to reduce major setups is to pre-define campaigns for specific product families
and include major setups between campaigns (Grunow et al. 2002); moreover, product
sequences can be pre-defined to be carried out in the block planning methodology,
which is described in more detail in Section 3.5.
22 2 Production planning and steel industry

2.2.5 Sequencing rules


Sequencing rules determine the order in which the group of waiting orders is processed
based on a set of criteria. Based on a definition by Lödding (2008: 443ff), the objective
of sequencing rules can be to improve on-time delivery, the service level, or the
performance or output of the production system.

Sequencing rules to improve on-time delivery include the first-in-first-out (FIFO)


principle in which the orders are executed in the same sequence as they arrived at the
production step. Other rules to improve on-time delivery include the earliest planned
start date or the earliest planned end date rule. In these cases, the order with the earliest
planned start date or the earliest planned end date is executed next.

The next group of sequencing rules improves the service level in make-to-stock
production environments by considering how long the safety stock will potentially last
through monitoring incoming customer orders versus forecasted demand for the average
replenishment time. Hence, this approach prioritizes orders for products that are at risk
of being out of stock.

Sequencing rules to increase performance are often based on pre-defined sequences that
minimize the overall setup time for a product group that has sequence dependent setups.
This is also the basis for the block planning approach, which is discussed in more detail
in Section 3.5. Another possible approach involves prioritizing the waiting order with
the shortest processing time. Simulations show that this approach leads to low in-
process inventories, short average makespans, and a high on-time delivery rate
(Lödding 2008: 452).

2.3 Steel production

2.3.1 Industry characteristics


The steel industry is often seen as the backbone for overall economic growth since it
supplies the economy with one of the most important raw materials for such engineering
industries as the automotive and construction industries, among others. Therefore most
of the biggest steel producing countries also have some of the world’s largest
economies. In the last decade, China easily rose to the number one steel producing
2 Production planning and steel industry 23

country with a growth of more than 300% as compared to Europe, where steel
production remained stable for much of that time with exception to a steep dip due to
the financial crisis in 2009 (figure 6).

Global crude steel production


in megatonnes (mt)
World excl. China and EU-27 EU 27 China

1.500 1.429 1.490


1.347 1.341
1.249 1.236
1.147
1.061 619 630
904 967 647 631
1.000 851 621 520
595 177
552 586 173
512 534 210 198 139
500 196 207
193 202 637 683 +349%
187 188 490 512 577
273 356 421
152 182 222
0
2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011

Figure 6: World crude steel production (WSA 2012b)

Today, three Chinese companies are among the world’s top five steel producers (WSA
2012a) as illustrated in figure 7, which accounts for the overall growth in steel
production in China. Another interesting aspect is that the steel industry is not yet as
consolidated as other mature industries in the world. As compared to the automotive
industry, where the top five producers together hold a market share of 46% (OICA
2012),1 the combined market share for the top five steel producers is only 18%.2

1 Global vehicle production amounted to almost 80 million automobiles in 2011, and the top five companies (i.e.,
GM, Volkswagen, Toyota, Hyundai, and Ford) produced 37 million vehicles.
2 Global crude steel production amounted to 1490 megatons in 2011, and the top five producers named in Figure 7
produced 262 megatons.
24 2 Production planning and steel industry

Top-5 Steel producing countries or regions Top-5 Steel producing companies


in megatonnes (mt), 2011 in megatonnes (mt), 2011

ArcelorMittal
China 683 97
(LU)

Hebei Group
EU-27 177 44
(CN)

Baosteel
Japan 108 43
(CN)

Posco Steel
USA 86 39
(KR)

Wuhan Group
India 72 38
(CN)

Figure 7: Top five steel producing countries and companies (WSA 2012a)

The financial crisis in 2009 significantly impacted the overall world economy and was
particularly damaging to the steel industry. In Europe, the production volume of crude
steel decreased by about a third from 2007 to 2009 (see figure 6). Furthermore, global
steel production decreased despite the fact that China’s steel production kept growing.
The deep impact of the financial crisis on the steel industry is indicative of the
significance of overall economic growth on the steel sector. This direct effect is quite
strong because one of the global steel industry’s main customers with a share of 39% is
the construction industry, which is itself quite volatile, as illustrated in figure 8.

Steel consumption and customer segments


100%=1140 Mt, 2009
Other
Oil & Gas 10
6
Transport - Construction
other 9 39

Transport - 11
automotive
25
Engineering

Figure 8: Steel consumption and customer segments (WSA 2009, King 2010)
Note: Each pie represents the percentage of the market share of steel going to that particular industry.
2 Production planning and steel industry 25

The main steel products produced for the industry’s customers can be categorized into
long, flat, and tube products. Long products are mainly wire, rail tracks, or structures
and rebars for the construction industry. Flat products are hot rolled either into coils or
plates. A share of the coils is later cold rolled, and some are even coated. The resulting
products are typically used in the automotive and machinery industry and for white
goods and consumer electronics. Tube products find their application in the transport of
gases and liquids in the chemical and oil and gas industries. An overview of steel
products and their applications is provided in figure 9.

Products Application examples


1,490
▪ Sections/ ▪ Steel structures in construction industry
bars ▪ Railway track material

▪ Rebars ▪ Reinforcement of concrete structures

Long 695 ▪ Wire rod ▪ Processed to steel wire, e.g., tire rod

▪ Plate ▪ Shipbuilding
▪ Heavy mining and construction machinery

▪ Hot rolled coil ▪ Trucks and trailers, agricultural machinery


▪ Structural parts of machinery
Flat 667
▪ Cold rolled coil ▪ White and brown goods
▪ Air-conditioning systems

▪ Coated sheet ▪ Automotive industry


▪ White goods

Tube 128 ▪ Seamless or ▪ Oil and gas industry for transportation


welded tubes ▪ Chemical industry, e.g., for tubing in plants
2011

Figure 9: Steel products and applications (WSA 2012b, King 2012)

2.3.2 Current challenges


Although the production quantities of the steel industry have recovered to an extent
after the financial crisis of 2009, the steel industry still faces significant challenges,
especially in Europe, including the following:

x Production of a high share of industry-wide CO2-emissions,

x High energy consumption and dependency on fossil fuels, and


26 2 Production planning and steel industry

x Dependency on raw materials from overseas.

The iron and steel industry is highly energy intensive and emits large amounts of CO2.
The industry accounts for about 19% of the world’s final energy consumption and about
25% of the world’s direct CO2-emissions from the industry sector (IEA 2007). This is
becoming increasingly important due to expected price increases related to increasing
regulations on CO2 emissions (Coelho 2010). In the European Union, the Emissions
Trading Scheme (ETS) sets forth ambitious benchmark levels and allocates emissions
allowances up to this benchmark starting in 2013. The number of allowances is
equivalent to the total European CO2 emissions and is designed to decrease;
consequently, in 2020, the total CO2 emissions will be reduced by 21% as compared to
2005 levels (EU 2012). Today, most allowances are still distributed up to an industry-
specific benchmark level for free, but in the future, the free distribution of allowances to
companies will decrease (EuropeanCommission 2012a). This will establish a market3 in
which companies trade their emission certificates and hence decide to buy additional
certificates if their emissions are above benchmark levels or invest in technology to
reduce emissions and sell excess certificates when possible. The trading of CO2-
emissions certificates ensure that the cheapest CO2 saving options are carried out first
(Müller et al. 2009: 9).
As illustrated in figure 10, the production of one ton of steel emits about 1.5 tons CO2
on average and consumes 19.3 GJ of energy. These are average values for steel
produced in Germany in 2011 and thus contain steel produced in a basic oxygen furnace
(BOF) and steel produced in an electric arc furnace (EAF). Today, about 70% of
German steel is produced using the BOF route, and the rest is produced using the EAF
route (VdEh 2011: 308). The choice of production process is important because steel
produced using the EAF route only uses electricity and, therefore, does not emit direct
CO2. In contrast, steel produced using the BOF route emits large amounts of CO2

3 In 2012, Germany decided to appoint the European Energy Exchange AG (EEX) in Leipzig as its marketplace for
emission certificates and notified the European Commission of this decision. The European Commission is
currently assessing whether the platform satisfies the rules and regulation (EuropeanCommission 2012b.
'Preparing for the 3rd trading period.').
2 Production planning and steel industry 27

because the BF used in the BOF route processes coal. Further details of the production
processes are presented in Section 2.3.3.

Energy use and CO2- emissions per ton of steel


2011

19.3
17.4 1.9
Energy
intensity
(GJ/t)

1,480
1,336 144
CO2-
Emissions
(kg CO2/t)

Crude Hot rolled Total


steel steel

Figure 10: Energy intensity and CO2 emissions per ton of steel (VdEh 2012)

Steel production is highly energy intensive; in fact, raw material and energy costs
contributed equally to about 60% of the steel price in mid-2008 (BCG 2009: 3). Based
on the price development of energy in major European countries over the last few years,
electricity and gas prices in Europe will potentially continue to increase if they continue
this recent trend as illustrated in figure 11.
28 2 Production planning and steel industry

Industry prices for natural gas Germany France Italy


EUR-Cent/KWh

0
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
116 M. kWh; service on 330 days; 8000 hours - (incl tax, excl. VAT)

Industry prices for electricity


EUR-Cent/KWh

15

10

0
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
50 000 000 kWh; max. consumption: 10 000 kW; annual service: 5 000 hours

Figure 11: Industry prices for energy excluding value added tax (BMWi 2012)

For steel companies in Germany, the challenges are even more severe due to a
government act that forces industries to abandon nuclear energy and increase use of
renewable energy sources (DeutscherBundestag 2011). Expectations indicate that
investment in renewable energy sources will increase by about 60%, and this will lead
to an increase of about 10% in electricity prices in 2020 as compared to prices in 2011
(McKinsey 2012: 7). This price increase for electricity will of course also result in an
increase in steel production costs.

In addition to the mounting energy costs, the other major portion of production costs,
i.e., the raw material costs, is also increasing; in particular, the price of iron ore has
increased by 105% in 2011 as compared to 2008 (SBB 2012). This price increase is
largely linked to the rise in steel production in China. The main raw materials for steel
production are iron ore, coking coal, and scrap. Iron ore deposits are mainly found in the
former Soviet republics, Russia, South America, China, and India (VdEh 2011: 305).
The new role of China and India as large steel producing countries, thereby causing
them to reduce their export of iron ore in order to use it in their own steel production,
has led to a steep increase in the price of iron ore. Consequently, demand has risen with
2 Production planning and steel industry 29

stable supplies of iron ore. Although the deposits of iron ore are not rare, the mining of
the deposits is very capital intensive, and only a few companies own the deposits.
Furthermore, the real capacity driver for mining is often not the mining equipment itself
but rather the rail and shipping capacities. The logistic capacities are often not owned by
the mining companies, and a capacity increase is not feasible because it is very capital
intensive, especially since prices would probably decline if capacity increased. The
situation for coking coal and scrap is not as severe since Europe still has its own coal
deposits and related mining operations. Moreover, Europe also has a large scrap market,
but scrap demand is growing due to an increased need for scrap in growing economies
and trade barriers (e.g., export taxes to protect the domestic steel industry4) (VdEh
2011: 277).

These three main areas of challenges are only valid for European steel companies
because steel companies in South America, Asia, and the US are not currently facing
government regulations that are aimed at reducing CO2 emissions; moreover, Chinese
and South American steel companies do not need to import as many raw materials due
to native deposits within these respective countries. Therefore, European steel industry
experts are concerned about the possible relocation of a significant share of steel
production to countries with lower emission standards. So far, the fear of carbon
leakage reduces the burden for the steel industry. In Germany, for example, the energy
intensive steel industry has been released from the requirements of the green tax and
from cost allocations to finance renewable energy sources. All of these tax and cost
allocation exemptions total up to €8 billion per year (Rieseberg and Wöhrlen 2012: 8).
In Europe, the EU ETS allocates comparatively more certificates to sectors that have a
significant risk of carbon leakage than to other sectors in order to reduce that risk (EU
2012). Nonetheless, raw material and energy prices will continue to increase, and these
categories are already among the biggest cost categories in steel production as previous

4 In particular, Russia has tried to protect their domestic steel industry via the implementation of export taxes and
other trade barriers; consequently, the Russian exports to the EU-27 states were reduced by 76% in 2010 as
compared to 2004 (VdEh 2011. 'Jahrbuch Stahl 2012.' Verlag Stahleisen GmbH, Düsseldorf. 291)
30 2 Production planning and steel industry

implied. Therefore, the iron and steel industry must become more energy efficient to
save on direct energy costs and costs related to CO2 emissions.

2.3.3 Production setup and processes


Two main processes are used to produce steel, i.e., the process involving the BOF and
the process using the EAF. The process route using the BOF produces steel from iron
ore, coke, and small amounts of scrap. In contrast, the EAF produces steel only from
scrap.

Additionally there are three different types of plant setups. An integrated steel mill
covers all production steps from raw materials to at least the hot rolled coil. Other steel
mills are not fully integrated and only process steel up to the continuous casting stage
because the semi-finished product slab can either be sold or further processed at the
third type, the single steel facility. The single steel facility only covers some production
stages after metal solidification, e.g. a HSM or cold rolling. A practical example is the
steel mill investment made by ThyssenKrupp, which is often discussed in press due to
cost increases and deviation from the planned start of production; this investment
consisted of a steel mill that is located in Brazil and covers all production steps from the
BF to continuous casting. The slabs are then shipped either to the ThyssenKrupp plant
in the USA that focuses on hot rolling and further processing or to a German location
that produces hot rolled coils (e.g., Duisburg, Bochum) as illustrated in figure 12.
2 Production planning and steel industry 31

Supply chain of ThyssenKrupp Steel Americas

Duisburg, Bochum
(Germany)
Alabama (USA)
▪ Hot rolling
▪ Hot rolling ▪ Further processing
▪ Further processing

Slabs
Santa Cruz (Brazil)
▪ Blast furnace to
continuous casting

Figure 12: Supply chain of ThyssenKrupp Steel Americas (TKSA 2010, TKSA 2012)

Steel production following the BOF route starts with the preparation of the iron ore.
Some iron ore in the form of lump ore or pellets is readily supplied for the BF and hence
does not require further processing before input into the BF. Smaller grain iron ore
needs to be sintered to increase the grain size for preparation as input material for the
BF. The other input material for the BF, i.e., the coke, must be preprocessed before it is
ready to be charged. A coking plant grinds the coal and then heats up the coal to about
1000°C with air exclusion. This process reduces the gaseous parts in the coal, thereby
increasing the carbonate and reducing brimstone (Münch 2012).

A BF is a shaft furnace that is 30 to 50 m high and can handle capacities up to 15


megatons per day. In the BF, the temperatures increase up to 1600°C, and through a
chemical reaction, the coke extracts oxygen from the iron ore to produce pure iron.
After the reaction, two components of the iron are transferred to the steel mill, and the
by-product slag is mostly sold for cement, road work, or fertilizer or sent to the landfill
(VdEh 2011: 328, Münch 2012).

The aim of the steel production process is to reduce the carbon content of the iron in
order to produce steel; indeed, steel has a carbon content of below 2%, and the carbon
32 2 Production planning and steel industry

content of iron is significantly higher (Bargel and Schulze 2010: 147). The reduction of
the carbon content and the removal of other elements, like phosphorus or brimstone, are
carried out in a converter with temperatures that reach more than 1500°C. With the help
of a water-cooled lance, oxygen is blown into the liquid mixture of iron and scrap metal.
At the end of the process, additional alloying elements are often put into the liquid steel
in order to create the desired characteristics (Münch 2012). When the liquid steel has
reached the correct chemistry it is transferred to the CC. At the CC, it is poured into the
tundish, which is basically a cooled mold that shapes the steel into rectangular bars (i.e.,
slabs) and cools them to the point where they solidify. Slabs are usually 180-260 mm
thick, 800-2000 mm wide, and up to 10 m long and weigh up to 30 tons. Some slabs
need further processing, like surface conditioning or trimming, before they are ready for
the next production step.

In the next production step, the HSM rolls the slabs into coils, i.e., the final product of
the primary steel-making process. The hot rolling begins with a furnace, which warms
the slabs above the recrystallization temperature, i.e., about 1200°C. After reaching the
correct temperature, they are rolled in usually one roughing and several finishing mills
in order to create the desired dimensions. Many options are available for the further
processing of hot rolled coils (e.g. cold rolling, coating, or trimming of coils) depending
on the desired characteristics of the final product.

Steel production involving the use of an EAF is much shorter. The only raw material for
this process is scrap, which is liquefied in an EAF that reaches temperatures of above
3500°C (Münch 2012). To ensure the correct chemistry in the produced steel, an argon-
oxygen mixture is blown into the liquefied steel in the Argon Oxygen Decarbonization
(AOD) converter and alloying materials are also added before the steel is transferred to
the CC. A big advantage of this method is that it does not emit direct CO2; nevertheless,
it requires the use of a lot more electricity as compared to the BOF route. Both
production processes are illustrated in figure 13.
2 Production planning and steel industry 33

Blast furnace process Electric arc furnace process

Metallurgical Coke oven Coke Scrap metal


coal Iron ore

Limestone

Natural gas,
mineral oil, or
coal
Electric arc
furnace
Basic Blast furnace
oxygen
converter

Steel Scrap

Oxygen Argon Oxygen


Decarbonization
Pig iron (hot converter
metal)
Steel

Figure 13: Steel production processes


34 3 Scheduling concepts with focus on the steel industry

3 Scheduling concepts with focus on the steel industry

3.1 Planning challenges in the steel industry


As pointed out in Section 2.3 the steel industry has a unique set of challenges that must
be considered when attempting a quantitative scheduling optimization. The variety of
challenges can be clustered into three groups, i.e., challenges due to equipment and
production setup, product portfolio and demand challenges, and the characteristics of
quantitative scheduling tools.

3.1.1 Equipment specific challenges


The steel industry is usually one of the first industries to prosper when countries begin
to industrialize. Therefore, most production sites in Western Europe are about 50 years
old and have plant layouts that were common at that time. Thus, the material flow
concepts that have been implemented in newer, state-of-the-art steel plants (e.g., in Asia
or South America) usually have improved connections between different production
steps. Old-fashioned material flow concepts often complicate fast transfers from one
production step to the next. Thus, many reports of successful implementations of hot
charging techniques are often based on case examples in Asia (Tang et al. 2001). Since
the equipment used in steel production (e.g., milling stands and furnaces) is very capital
intensive,5 most of the equipment in Western Europe is several decades old but has been
well maintained and updated.

The equipment used for steel production also presents certain unique challenges.
Indeed, furnaces play a large part in steel production shutdowns, but stops and speed
changes are not easily implemented. The inflexibility at different production stages
varies. The primary steel processes from the BF to the HSM are usually more inflexible
in terms of production rate changes as compared to the following production steps in
which furnaces no longer play a major role. For example the BF is probably the most

5 A new furnace (i.e., using the walking beam technique) for a HSM currently costs about €50 million. A HSM with
production speeds of about 500-600 t/h uses about three furnaces of that type.

I. Mattik, Integrated Scheduling of Continuous Casters and Hot Strip Mills, Produktion und Logistik,
DOI 10.1007/978-3-658-03775-8_3, © Springer Fachmedien Wiesbaden 2014
3 Scheduling concepts with focus on the steel industry 35

challenging equipment in terms of production rate flexibility. Although options to


reduce the production speed by reducing the amount of oxygen and coke do exist, the
BF cannot be stopped for a short amount of time. If the furnace were to shut down for
maintenance or a reduction in production, this shut down usually lasts at least more than
two weeks, i.e., until it is cooled down, maintained, and heated up again. In particular,
shutting a BF down due to a reduction in production is usually a very tough decision
since a shutdown often only saves costs if it lasts for more than a month. Indeed, a
temporary decreasing demand is still usually met by leaving the BF running at a
minimum output.

In contrast, the CC involves a continuous production speed; in other words, no speed


changes can be made. In the case of cast breaks, setting the CC up again and starting
production usually takes one hour, whereas the HSM usually has high production speed
flexibility and also allows for some amount of shutdowns. The most difficult question
for HSM planners is whether the furnaces should be kept warm to allow for a quick
production start or shutdown, which will lead to a significantly longer break to heat up
the furnace again.

As previously pointed out, the equipment characteristics, especially the production rate
flexibility of the considered production steps, must be considered in any planning
application.

3.1.2 Product portfolio challenges


A steel company usually has a large variety of product types that are produced at one
location. This often involves more than 300 steel grades, and an infinite number of
dimensions, and requirements for further processing (e.g., cold rolling and coating). An
ABC analysis of the steel grades produced at one location for a large European steel
company reveals that about 30% of all steel grades account for 80% of the production
output; moreover, 10% of all steel grades amount to about 50% of the production output
(see figure 14).
36 3 Scheduling concepts with focus on the steel industry

ABC analysis of the steel grade portfolio of a large European steel company
in Percent

100

90

80
Share of volume

70

60
in Percent

50
40
30

20

10
0
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Share of Products
in Percent

Figure 14: ABC analysis of the steel grade portfolio of a large European steel company

In addition to the complexity of the product portfolio, steel demand is closely linked to
economic development; therefore, planners are faced with many challenges related to
the equipment characteristics in volatile economic situations as pointed out in Section
3.1.1.

3.1.3 Demand and scheduling characteristics


Performance indicators for production planning in the steel industry usually indicate
long lead times and a poor on-time customer delivery rate as compared to other
industries. In addition to the complex product portfolio, these performance anomalies
can be attributed to production runs in long campaigns, a high number of production
levels, and large differences between orders and batch sizes. In addition, a high number
of orders are usually processed each day.

Production runs are planned in long campaigns, if possible, to reduce setup times.
Setups need to be avoided not only to save production time but also to reduce
significant material costs. For example, each setup at the HSM requires that the rolls are
maintained and reconditioned; hence, the planner’s aim is to use the rolls closely until
they are worn out and need reconditioning. Conditions at the CC are similar since a new
mold is used after each setup.
3 Scheduling concepts with focus on the steel industry 37

An average order size is about 100 tons of steel, but the required batch size in the CC is
a minimum of 250 tons (Cowling and Rezig 2000: 187). To increase production
efficiency and avoid additional setup, planners try to at least double or triple the
minimum batch size, i.e., up to 750 tons. Hence, batching of orders is essential; for
example, orders for steel grades of the B or C category have a long waiting time before
production starts. Additionally, technological planning requirements differ at each of the
production steps. Indeed, the production steps up to the CC require large batches of the
same steel grades, and the HSM requires production runs of similar dimensions and
hardness. Hence, the orders do not stay in the same sequence from steelmaking to the
HSM. Overall, the lead time for orders is usually a few weeks, but priority orders can
also be produced in much less time. Since products often need to wait until a specific
steel grade is cast, the ordered dimensions are hot rolled and, if appropriate, sent for
coating.

To increase planning flexibility and timely order fulfillment, a high level of work is
usually in progress. Most of the inventory is at the slab stage between continuous
casting and hot strip milling. The reason for that is twofold. On one hand, a share of the
slabs are produced in a make-to-stock process to reduce customer delivery times;
nevertheless, most production quantities are made to order at least in companies with
large product portfolios since slabs still have a lot of flexibility in terms of dimensions
and further product characteristics requested by the customer. On the other hand, these
slabs are made to order, but kept in the slab yard to reduce planning complexity due to
the different technological planning requirements mentioned above.

In summary, the steel industry faces significant challenges especially due to its large
product portfolio, the minimum batch sizes for continuous casting, and the campaign
operations for rolling and further processing.

3.1.4 Requirements for quantitative scheduling optimization


The challenges pointed out in the previous sections must be considered when setting up
quantitative scheduling tools. The requirements for scheduling tools for the steel
industry can be summarized by the following characteristics:
38 3 Scheduling concepts with focus on the steel industry

Incorporation of relevant technological requirements

As pointed out in Section 3.1.1, a set of equipment specific or technological


requirements must be considered, such as product dependent major setup times at the
CC or the production speed flexibility of specific equipment.

Balance between single level and multilevel optimization

A scheduling tool with practical applicability finds the right balance between single-
level and multi-level scheduling optimization. Due to model complexity and resulting
computation times, building a scheduling tool for all production steps in a steel mill that
considers all relevant technological requirements is probably not possible. Yet,
concentrating only on a single production step is not always useful if integration with
the previous or the next step is beneficial. A good example is the production steps of
continuous casting and hot strip milling. Both steps could be scheduled separately using
single level optimization, but if their optimization is integrated or at least coupled, large
energy cost benefits can be realized.

Low level of manual intervention possible

Sequencing decisions sometimes require extensive expertise that can hardly be


incorporated into a quantitative tool. Therefore, allowing at least some human
intervention for detailed sequencing tasks is beneficial, especially in milling operations
where slabs are sequenced according to a variety of criteria. Furthermore, the option for
manual intervention might improve the technology acceptance.

Fast computation times with large problem instances

Applications in a real industry environment require fast computation times with large
problem instances in order to create the capability to react quickly to sudden events,
such as machine breakdowns or unplanned product changes (e.g., if the steel mill
accidentally produces steel with a different chemical compatibility than planned).

Continuous representation of time

A continuous representation of time generally has a higher practical applicability since


production can start and end anytime during the planning horizon and not just occur
3 Scheduling concepts with focus on the steel industry 39

between pre-defined time points. Hence, for a detailed scheduling model formulation, a
continuous time based model formulation is preferred. For further discussion, see
Section 3.4.

Simultaneous batching and scheduling

Batching and scheduling optimization should ideally be done simultaneously since this
is the only approach that allows for true optimization. Further discussion of this topic
can be found in Section 3.3.

Product portfolio

The model formulation must be able to handle the complexity of a large production
portfolio, and certain characteristics, such as minimum batch sizes and campaign
modes, must be included.

3.2 Review of batching and scheduling literature with steel industry focus

3.2.1 Classification of batching and scheduling models with steel industry


focus
As the starting point for a thorough literature review, a classification of the existing
literature on scheduling and batching models with a focus on the steel industry is
presented in table 1. The following criteria are used to classify the different models
from the literature:
x Decomposition at planning level – The classification of the decomposition
approach is based on Dong et al. (Dong et al. 2012: 5747), where models either
integrate batch planning and scheduling (IBPS) or focus on either batch planning
and are thus called integrated batch planning (IBP) models or scheduling models
and are thus called integrated batch scheduling (IBS). Some authors split batching
and scheduling in two separate models (IBP and IBS).
x Focus – Some of the models focus on a single production step in the primary
steelmaking processes, while others cover more than one process. The primary
production processes covered are the BOF, the EAF, the melt shop (MS), the
40 3 Scheduling concepts with focus on the steel industry

continuous caster CC, and the hot rolling (HR) stage. Additionally, one
publication focuses on a galvanizing line.
x Number of levels – In this column, the number of production levels is given. This
number can deviate from the number of production steps given in the category
focus if the model considers intermediate steps (e.g., a slab yard).
x Decomposition at plant level – Similar to the deconstruction approach at the
planning level, deconstruction at the plant level is also carried out. The table
indicates if more than one production step is covered (otherwise n/a); moreover,
the table indicates if the model covers all production steps separately or
combined. In some cases, separate models are coordinated with a solution
algorithm or solver heuristic, and in those cases, the indication of “coordinated” is
provided.
x Objective – Some models have a single objective (S), whereas others have a
multi-objective approach (M).
x Solution method – The solution techniques of the models differ, and some models
are solved using solution algorithms (SA) specifically designed for the
optimization problem. Others are solved using commercial solvers (CS) or
heuristics (H). Additionally, a commercial decision support system (DSS) and
simulation (S) have also been used.
x Representation of time – Some models split the planning horizon into discrete
periods (D), whereas other models use a continuous representation of time (C).
x Computation time – In this category, an indication of the computation times based
on the numerical tests in the publication is given if it was reported; otherwise,
“NR” is reported.
x Planning horizon – To evaluate the given computation time, a planning horizon in
time units is also given; otherwise “NR” is reported. If a planning horizon in time
units is not applicable, an equivalent level of another relevant factor (e.g., jobs) is
given.
3 Scheduling concepts with focus on the steel industry 41

Table 1: Batching and scheduling literature with steel industry focus

at plant level
Focus
publication year

Solution method

Computation
Decomposition

Decomposition

of time

Planning horizon
Author and

at planning level

Objective

Representation

time
As’ad and Number of levels < 180 5
Demirli IBP HR 1 n/a M H D
sec. days
(2010a)
As’ad and
< 35 3
Demirli IBS HR 1 n/a M CS D
sec. days
(2010b)
Assaf et al. 10-20
IBS HR 1 n/a M SA C 8 hrs
(1997) min.
Bellabdaoui
MS com- <8
and Teghem IBS 3 S CS C 2-6 sec.
CC bined hrs
(2006)
<1
Busch (2008) IBS HR 1 n/a S H C > 1 hr
day
2000
Chang et al.
IBP CC 1 n/a S H n/a ≤ 2 sec. char-
(2000)
ges
Chen et al. <1
IBS HR 1 n/a M SA n/a ≤ 1 hr
(2008) day
Cowling "few <1
IBS HR 1 n/a M DSS n/a
(2003) min." day
Cowling et IBP CC coordi ≤1
2 M SA n/a NR
al. (2004) S HR -nated day
2-3
Cowling and CC coordi "few
IBS 2 S SA D wee
Rezig (2000) HR -nated min."
ks
Dong et al. ≤1 540
IBP MS 1 n/a M SA n/a
(2010) min. slabs
Dong et al. ≤ 10 1
IBP MS 1 n/a M SA n/a
(2012) sec. day
42 3 Scheduling concepts with focus on the steel industry

Table 1(continued): Batching and scheduling literature with steel industry focus

at plant level
Focus
publication year

Solution method

Computation
Decomposition

Decomposition

of time

Planning horizon
Author and

at planning level

Objective

Representation

time
Harjunkoski
IBP EAF Number of levels 1
and com-
and MS 4 M CS C > 2 hr wee
Grossmann bined
IBS CC k
(2001)
Huegler and BOF
com- ≤5 100
Vasko IBS MS 3 M SA C
bined min. heats
(2007) CC
Kangbok et IBP
CC 1 n/a S SA n/a NR NR
al. (2004) S
Kosiba et al. ≤ 15 ≤8
IBS HR 1 n/a M H n/a
(1992) sec. hrs
Kumar et al. MS com- 10
IBS 3 M SA C NR
(2006) CC bined hrs
MS
Li et al. IBP coordi 1
CC 5 S SA C ≤3 min.
(2006) S -nated day
HR
Lin and IBP BOF
com-
Moodie and MS 3 M SA D NR NR
bined
(1989) IBS CC
BOF
250
Liu et al. MS com- ≤45
IBP 6 M SA C orde
(2006) CC bined min.
rs
HR
Lopez et al.
IBS HR 1 n/a M H n/a 15 min. 8 hrs
(1998)
EAF
Melouk et al. IBP com- 30
MS 4 M S C 25 min
(2013) S bined days
HR
Missbauer et MS com- 72
IBS 5 M H C 25 sec.
al. (2009) CC bined hrs
3 Scheduling concepts with focus on the steel industry 43

Table 1(continued): Batching and scheduling literature with steel industry focus

at plant level
Focus
publication year

Solution method

Computation
Decomposition
Decomposition

of time

Planning horizon
Author and

at planning level

Objective

Representation

time
Mohanty and IBP
HR Number of levels
1 n/a M H D NR NR
Singh (1992) S
IBP MS
Park et al. sepa- ≤2 7
and CC 3 S SA n/a
(2002) rate min. days
IBS HR
Petersen et ≤ 60 ≤1
IBS HR 1 n/a M SA C
al. (1992) sec. day
IBP MS
Sato et al. com- 10
and CC 5 M S n/a NR
(1977) bined days
IBS HR
Tang et al. IBP MS com- ≤1
3 M CS C 20 sec.
(2000a) S CC bined day
Tang et al. <1
IBS HR 1 n/a M SA n/a 40 min.
(2000b) day
Tang et al. MS com- <4
IBS 3 M H D 8 hrs
(2002) CC bined min.
ca.
Tang and MS com- <5
IBP 3 M H n/a 3000
Wang (2008) CC bined min.
slabs
gal-
Valls vani
IBP < 35 12
Verdejo et al. - 1 n/a M SA C
S sec. days
(2009) zing
line
Vanhoucke IBP MS 3
com- < 10
and Debels and CC 8 M SA C wee
bined sec.
(2009) IBS HR ks
Vonderem-
1
bse and < 35
IBP CC 1 n/a M CS n/a wee
Haessler sec.
k
(1982)
44 3 Scheduling concepts with focus on the steel industry

Table 1(continued): Batching and scheduling literature with steel industry focus

at plant level
Focus
publication year

Solution method

Computation
Decomposition

Decomposition

of time

Planning horizon
Author and

at planning level

Objective

Representation

time
Wang and
IBS HR Number of levels
1 n/a M H C
< 700 10
Tang (2008) sec. days
MS
Xu and Sand IBP coordi ≤ 15 10
CC 8 S H C
(2012) S -nated min. jobs
HR

3.2.2 Review of individual publications


As presented in the previous section, the majority of the publications deal with specific
stages of the overall steel making process and often focus either on batch planning or
scheduling. As illustrated in figure 15, publications can differ in their focus on
production steps and also their optimization task (i.e., to either optimize batch planning,
scheduling, or integrate both questions). For instance, if two publications focus on the
continuous casting–hot rolling stage, they can still differ in terms of their deconstruction
approach at the planning level (e.g., integrate batch planning and scheduling, i.e.,
example 1, or focus only on the batch planning stage, i.e., example 2).
3 Scheduling concepts with focus on the steel industry 45

Integrated batch Integrated batch Integrated batch


planning and planning (IBP) scheduling (IBS)
scheduling (IBPS)

Blast furnace (BOF)


Electric arc furnace
(EAF)

Melt shop
(MS)

Continuous
caster (CC) 1 2

Hot rolling
(HR)

Figure 15: Decomposition structure of publications

Scheduling approaches that focus on the MS-continuous casting stage usually contain
model formulations that determine the sequence, timing, and device for the allocation of
molten steel at the different production stages in order to satisfy customer demand. In
contrast, publications that focus on batch planning group orders into casts and determine
the production sequence of the casts. The authors of these publications intend to
maximize throughput via effective batching and minimizing the number of casts and
changeovers (Tang et al. 2000a, Missbauer et al. 2009, Tang and Wang 2008, Dong et
al. 2010, Box and Herbe Jr 1988, Bellabdaoui and Teghem 2006, Tang et al. 2002).
They also focus on the liquid stage because the CC is usually the bottleneck in steel
production; therefore, the maximization of throughput at the CC is essential (Tang et al.
2002).

Missbauer et al. (2009) created algorithms for three planning levels in order to optimize
the scheduling of charges from the melt shop to the CC. The first planning level that
was considered schedules the CCs under the capacity restrictions of the upstream stages
and limited hot metal supply. The second planning level schedules the downstream
converter and refining facilities. The third planning level tries to improve the schedules
generated in the first two planning levels. Their project was implemented in a steel plant
in Austria to support the manual planning process (Missbauer et al. 2009).
46 3 Scheduling concepts with focus on the steel industry

Tang and Wang (2008) focus on the same production steps as Missbauer et al. (2009)
and develop a decision support system to optimize the batch planning of orders into
casts to allow efficient operations at the continuous casting stage. They claim that their
decision support system has been implemented at Chinese steel companies (Tang and
Wang 2008).

In contrast, Harjunkoski and Grossmann (2001) include the EAF in addition to the MS
and the CC in their research focus and deconstruct their scheduling optimization into
three MILPs and one LP solved using CSs instead of one large and complex MILP-
problem. Their deconstruction strategy includes presorting into product families and
charges, followed by a scheduling step and a scheduling improvement step at the end.
Although tests with small scale random data seem promising, the tests with industrial-
scale examples show long calculation times of about 3 hours (Harjunkoski and
Grossmann 2001).

Dong et al. (2010) solved an integrated batch planning problem with focus on the MS to
minimize the number of charges, thereby reducing the costs spent for tundishes and
maximizing the throughput (Dong et al. 2010). In a later publication, Dong et al. (2012)
focused on the same problem with consideration of the production constraints for
downstream processes, such as hot rolling, and the use of an improved algorithm to find
a solution (Dong et al. 2012). In both publications, their solution algorithms have short
computation times to solve problems that typically take the equivalent of about one
production day to solve using traditional methods.

Bellabdaoui and Teghem (2006) developed an MILP formulation to solve the MS-
continuous casting scheduling problem. Unlike most other authors who use heuristics,
their model formulation uses a commercial solver, and problems that typically require
the equivalent to one shift of production to solve with traditional means are solved in
less than a minute (Bellabdaoui and Teghem 2006).

To minimize waiting times and stoppages, thereby reducing energy consumption


between the MS and continuous casting stages, Tang et al. (2000a) developed a model
that can be solved using commercial solvers. Batch planning and scheduling were
integrated, and computation times seem adequate for industry applications (Tang et al.
2000a).
3 Scheduling concepts with focus on the steel industry 47

Similar to publications that only optimize the liquid stage of steel production, many
examples in the literature only focus on the HSM. These publications try to capture the
knowledge of human scheduling experts and use exact methods or heuristics to
automate and optimize scheduling decisions. To automate schedule generation, a
travelling salesman problem may be solved to determine a sequence of slabs with the
intent to avoid hardness jumps and to comply with the coffin shape of feasible milling
programs as further explained in section 4.1 (Chen et al. 2008, Kosiba et al. 1992, Tang
et al. 2000b). Other authors doubt the possibility of a smooth implementation of a fully
automated tool and, therefore, develop decision support tools (Cowling 2003).

Lopez et al. (1998) develop a heuristic approach that generates HSM sequences. To
reduce energy consumption, the residence time of the slabs in the furnace is calculated
based on the charging temperature of the slab. Amongst other aims related to
technological sequencing requirements, a penalty objective function tries to ensure that
the slabs only stay in the furnace the minimum amount of time actually needed to reach
the rolling temperature. The approach was tested with real data and compared to actual
production schedules of a steel company and produced superior results in both the
lengths of production runs and the inclusion of priority slabs (Lopez et al. 1998).

Tang et al. (2000b) created an algorithm that creates multiple production sequences to
avoid the greedy method, in which one excellent solution requires that the remaining
orders can only be sequenced in inferior production runs. This algorithm was
implemented at a Chinese steel company and the authors report a 20% improvement in
schedule quality compared to the previous manual process (Tang et al. 2000b).

Algorithms to create sequences for hot rolling production have also been developed by
Assaf et al. (1997) and Chen et al (2008). Assaf et al. developed an algorithm that
evaluates sequences based on sub-models that calculate the heating time of each slab
and the rolling feasibility. Numerical results with real data were compared against
production sequences at a Canadian steel company. Reported results show large
improvements in fuel consumption and makespan (Assaf et al. 1997). Chen et al. (2008)
developed an algorithm based on swarm optimization in order to optimize hot rolling
sequences and show that the approach works with real production data (Chen et al.
2008).
48 3 Scheduling concepts with focus on the steel industry

Wang and Tang (2008) developed an MILP model that simultaneously creates multiple
production sequences. Moreover, unlike other studies that only determine sequences,
this MILP model also determines the exact timing of each slab (Wang and Tang 2008).

Petersen et al. (1992) attempted to reduce the energy consumption of the hot rolling
process by synchronizing the furnace residence time and the production time at the mill
for each slab. Furthermore, the authors maximized the utilization of the heating area in
the furnace by implementing a new method to charge the slabs into the furnace; i.e., the
slabs are placed next to each other without gaps. The heuristic procedure has been tested
at a Danish steel mill, and the authors observed large improvements in the production
capacity, i.e., by 10 to 20%. A large share of this improvement seems to be related to
their method of minimizing the gaps between the slabs in the furnace, and as reported,
the utilization of the heating area in the furnace is increased by about 12%, which can
be directly translated into a capacity increase (Petersen et al. 1992).

As’sad and Demirli (2010a) chose a different approach in which they develop a lot-
sizing model at the master production schedule level. Their investigated rolling mill
uses billets as raw materials and produces steel bars. Therefore, their production
sequences do not have as many constraints as rolling mills that produce coils from slabs.
The authors solved a lot-sizing problem based on a deterministic demand rate to
minimize costs; in this approach orders with the same end product dimensions were
batched into production lots (As’ad and Demirli 2010a). In a second publication, the
authors have the same research focus but also consider the volatility of demand
forecasts by only generating exact schedules for the next few periods and relax
constraints for the later periods (As’ad and Demirli 2010b).

To date, all of the approaches have either focused on the liquid stage of steel production
or on the HSM. Though very complex, integrating the CC and the HSM offers many
advantages, including lower working capital, reduced logistical effort, high energy
savings, and reduced material losses through oxidation (Tang et al. 2001, Cowling and
Rezig 2000). In fact, as pointed out in section 2.3.2, only about 10% of the total energy
consumption for steel production is spent at the hot rolling stage; in contrast to the
energy consumption, the CO2 emissions of this steel production stage could be reduced
to zero. At the liquid steel stage, a large portion of the energy consumption and the
3 Scheduling concepts with focus on the steel industry 49

resulting CO2 emissions are technologically required (e.g. due to the amount of coking
coal required to produce molten iron in the BF). An evaluation of the potential benefits
surprisingly indicates that only a few previous studies have provided model
formulations and case studies for an integrated scheduling of CCs and HSMs. In their
review of methods for integrated steel production, Tang et al. (2001) described the
following four basic stages of integrating the production of CCs and HSMs, i.e., from
no to full integration (Tang et al. 2001: 3f):

1. Continuous casting-cold charge rolling: The continuous casting and HSM


processes are separated by a slab yard. This method is still widely used,
especially in companies with a large product portfolio.

2. Continuous casting-hot charge rolling: Slabs are kept warm using an insulation
measure (e.g., a preservation pit) for a few hours and then charged into the
furnace of the HSM to reach the rolling temperature again. Charge temperature
ranges from 400 to 800°C.

3. Continuous casting-direct hot charge rolling (i.e., also simply called hot
charging): Hot slabs are directly charged from the CC into the heat furnace of
the HSM. This is facilitated by transfer rails or cars that connect the CC and the
HSM. Charge temperature ranges from 700-1000°C.

4. Continuous casting-direct hot rolling: This is a hot charging technique, where


CCs and HSMs are fully integrated; therefore, no heat furnace is necessary. This
requires slabs to have a charge temperature of above 1100°C.

Clearly, the increasing level of integration allows higher energy savings and shorter
production lead times. Although Tang et al. (2001) described practice cases that
involved the full integration of CCs and HSMs, the integration of these production
stages becomes more difficult as the product portfolio becomes increasingly more
complex because the difference between the order and batch sizes makes the creations
of stock slabs almost inevitable (Cowling and Rezig 2000: 187). Nine authors from the
classification in table 1.focus on the CC and HSM stage (Cowling and Rezig 2000,
Cowling et al. 2004, Li et al. 2006, Liu et al. 2006, Melouk et al. 2013, Park et al. 2002,
Sato et al. 1977, Vanhoucke and Debels 2009, Xu et al. 2012), but only Cowling and
Rezig (2000) and Xu et al. (2012) include an energy saving aim in their model
50 3 Scheduling concepts with focus on the steel industry

formulation. The other authors attempted to improve processes via simulation (Melouk
et al. 2013); determine feasible and robust batches by heuristically optimizing penalty
objective functions, including work in progress penalties and high service levels (Liu et
al. 2006); develop an algorithm to determine and optimize schedules with simulation or
solution algorithms, including a high number of technological requirements (Park et al.
2002, Li et al. 2006, Sato et al. 1977); focus on the master production schedule level
(Vanhoucke and Debels 2009); or develop a dynamic multi-agent architecture that
mostly is designed for better coordination between the plants and the slab yard
(Cowling et al. 2004).

Xu et al. (2012) developed a heuristic to coordinate the solution process of two


scheduling problems, i.e., the MS-CC scheduling problem and the HSM scheduling
problem. The coordination heuristic was intended to minimize the waiting time for slabs
between the production stages, and the overall model was designed to minimize the
makespan. To reduce the complexity of the problem to a manageable size, some
metallurgical constraints and production grouping compatibility constraints were
neglected. Only small test instances were used to obtain numerical results due to the
model complexity, but a case study report from an implementation carried out by the
technology company ABB showed that this concept is quite practical and manages to
reduce slab yard storage times by about 20% and energy consumption by about 10%
(Xu et al. 2012).

Cowling and Rezig (2000) developed a method for integrating the planning of CCs and
HSMs. Their method involved generating rolling schedules that considered the
technological requirements of the continuous casting stage for orders that need to be
produced within the next few days. Hence, they allowed the HSM to exert a pull-
influence on the CC. They claimed that their MILP model reduced the inventory levels
in the slab yard; increased the length of the HSM production run, thereby leading to less
material costs for work rolls; and reduced energy consumption due to charging warm
slabs. Furthermore, they claimed to increase the throughput, but the text did not clearly
state whether this statement is valid for only the HSM or also valid for the CC stage.
Due to the influence of the hot strip milling stage on the continuous casting stage in
their methodology, the claim that both plants can increase their throughput seems
improbable (Cowling and Rezig 2000).
3 Scheduling concepts with focus on the steel industry 51

In contrast to the publications mentioned earlier, Vanhoucke and Debels (2009)


developed a scheduling algorithm for the master production schedule level with a
middle-term planning horizon. The obtained results are broken down into daily
production buckets without determined sequences. Therefore, the technological
requirements for detailed scheduling or sequencing decisions were not considered in
their algorithm (Vanhoucke and Debels 2009).

A simulation based method to optimize the planning and scheduling of the whole value
chain from the EAF to the finishing lines was developed by Melouk et al. (2013). Their
simulation included an optimization component that takes the 30 day production
schedule as an input and aims at minimizing the inventory holding costs and downtime
costs of the facility where the coils are cut to length. Furthermore, this approach also
determines the optimal inventory levels of the slab yard and the coil inventory (Melouk
et al. 2013).

In addition to all of the potential benefits of integration, two barriers must be overcome.
An integrated planning approach could lead to a reduction in throughput and would
require an improved ability to react to product mix changes. Throughput reduction
could occur if the HSM is allowed to exert an excessive influence on the CC’s planning.
On one hand, if the caster is produced in the same or nearly the same sequence as the
HSM, the mold of the caster would need to be adjusted for width more often to model
the coffin profile of a milling program. Adjusting the width reduces the casting speed
and hence the throughput. On the other hand, if the caster would need to reduce batch
sizes to produce more steel grades of varying hardness due to an increase in setup
operations, an improved ability to quickly react to product mix changes is necessary
because differences in time, volume, and quality often exist between real slabs and
prescheduled virtual slabs (Tang et al. 2001). Due to these operational challenges, most
European steel companies with complex product portfolios have not yet integrated their
processes and only use hot charging if it is necessary for quality reasons.

The complexity of integrated planning and the relevance for the steel industry are
clearly potential areas of further research with practical applicability. To date, only very
few reports address the scheduling of both production steps. Cowling et al. (2004)
developed a multi-agent architecture to schedule virtual slabs and to integrate
52 3 Scheduling concepts with focus on the steel industry

continuous casting and hot strip milling and claimed that this approach has been
implemented at several steel mills across the world. They did not establish a hot
charging connection based on Tang et al. (2001) and also stated that too little research
has addressed the integration of the scheduling systems for CCs and HSMs (Cowling et
al. 2004, Cowling and Rezig 2000). Current approaches either optimize a decoupled
process or use heuristics or expert systems to optimize an integrated process (Cowling
et al. 2004, Xu et al. 2012). Hence, modeling approaches that reduce the problem
complexity without reducing the practical applicability by over-simplifying the problem
are necessary.

Other studies have also considered approaches that optimize the slab sizes to fulfill
customer orders and maximize the throughput of the casters (Vonderembse and
Haessler 1982, Dawande et al. 2004). Furthermore, the increasingly complex operations
on the slab yard have also found their way into the literature. Dohn and Clausen (2010)
developed an algorithm that optimizes the way the slab yard cranes are planned and
scheduled to ensure a HSM production without delays. Singh et al. (2004) optimized the
slab selection in the slab yard in a situation where several equivalent slabs could fulfill a
production order, but each of them is shuffled in a stack of other slabs. Hence, the slab
that takes the least shuffling effort to secure is selected (Singh et al. 2004).

3.3 Integrated and separate batching and scheduling


A large variety exists in batching or lot sizing models. Dynamic lot sizing problems
determine the size of production lots for products in each period of a finite planning
horizon to meet a deterministic dynamic demand. Quite often, those models use a
discrete time scale. Indeed, scheduling models essentially determine the start and finish
times, machine allocation, and the order of sequence of the production lots and thus the
jobs within a pre-defined planning horizon (Jans and Degraeve 2008: 1621).

Based on the classification of the batching and scheduling models in Section 3.2.1, the
majority of the models clearly focus on either scheduling or batching. The authors
justified their deconstruction of both problems based on the growing complexity of
integrating batching and scheduling decisions in a single model (Dong et al. 2012:
5747). Due to the difficulty associated with obtaining optimality when using separate
optimizations for batching and scheduling, the current trend in the literature is for lot
3 Scheduling concepts with focus on the steel industry 53

sizing and batching models to incorporate more scheduling aspects (e.g., sequence-
dependent setup time and technological scheduling constraints) to improve the practical
applicability of the models’ results. This gives rise to fading boundaries between lot
sizing and scheduling models (Jans and Degraeve 2008: 1620ff). According to Stadtler
and Fleischmann (2012: 153), lot sizing and scheduling should ideally be done
simultaneously due to the large interdependencies between both tasks. Other authors
concluded that rigorous separate approaches often do not determine a good, feasible
solution for real-world multistage batch scheduling problems within reasonable CPU
times (Marchetti and Cerdá 2009: 2733). Consequently, acceptable model approaches
must allow for the integration of batching and scheduling decisions without steep
increases in the modeling complexity while simultaneously keeping computation times
at a manageable level.

3.4 Continuous and discrete representation of time


One of the most distinguishing differences for modeling approaches is their choice of
time scale. If events are allowed to take place only between pre-defined time points
within the planning horizon, a discrete representation of time is chosen. In contrast, a
continuous representation of time allows events to take place at any moment during the
time horizon.

Discrete time models divide the planning horizon into a finite number of time intervals
with pre-defined and equal duration. Furthermore, a task (e.g., the start or end of the
production and setup operations) is only allowed at the boundaries of the time intervals
(Méndez et al. 2006: 918, Mouret et al. 2011). Discrete models are either called big-
bucket models or small-bucket models. In big-bucket models, a bucket size, i.e., the
period length, is selected that allows the production of several products requiring
different setup states on each machine (Suerie and Stadtler 2003: 1039). In most cases,
setup states are only valid for a single period, but some approaches allow the setup state
to be preserved from one period to the next (Suerie and Stadtler 2003). In contrast, in
most cases, small-bucket models allow only up to one setup activity per time period.
Those models determine a clearly defined sequence and, therefore, can be applied as
scheduling models.
54 3 Scheduling concepts with focus on the steel industry

As long as big bucket periods are chosen and the number of periods and products is low,
those models are less complex because scheduling constraints only need to be
monitored at the boundaries of the time periods (Méndez et al. 2006: 918). The trend
towards small bucket sizes to obtain exact plans (Suerie and Stadtler 2003: 1039) also
highlights a major disadvantage of the discrete time scale. The size of the mathematical
model and the computational efficiency depends on the number of time intervals
(Méndez et al. 2006: 919, Ierapetritou and Floudas 1998: 4341); therefore, the choice of
time interval length in discrete models is crucial. To obtain exact and applicable results,
especially for high production speeds, a tight time grid is needed, and time intervals
should take the length of the shortest task processing time. Unfortunately, such a tight
time grid leads to a steep increase in model size and computational effort. On the other
hand, if a less tight time grid is applied with period lengths greater than the processing
time for the shortest task, approximations are necessary, which could ultimately lead to
overproduction, idle times, or infeasibility (Méndez et al. 2006: 919, Ierapetritou and
Floudas 1998: 4341).

In contrast to a discrete time scale, the use of a continuous representation of time to


overcome the limitations of discrete models is possible. In these model formulations,
continuous timing variables define the exact times at which the events occur. In many
cases, this allows for a significant reduction of model variables and simultaneously
permits the formation of exact plans without any approximations (Méndez et al. 2006).

3.5 Block planning principle


The block planning principle is based on the idea that a natural sequence of production
orders often exists due to technological requirements (Günther et al. 2006). For CCs,
this is true because of the chemical compatibility of steel grades. Therefore, steel grades
can be cast in a pre-determined sequence requiring only a minor setup operation.
Otherwise, incompatible steel grades must be separated by a major setup operation,
which increased the time required by about three fold. Each block reflects a pre-defined
sequence of production orders (i.e., steel grades). This means that each product type
occurs once at a pre-defined position in the sequence, and the corresponding batch is of
variable sizes. Depending on customer demand, the size of the production order varies
from no production up to very large batch sizes. The block is cyclically repeated every
3 Scheduling concepts with focus on the steel industry 55

period (e.g., one block per day). This pre-defined sequence of production orders within
a block drastically reduces the complexity of the scheduling problem and complies with
normal practices, particularly in the process industries.

Two main concepts regarding the flexibility of this approach should be noted. Flexible
block planning does not fix the earliest start and latest finish times of blocks. These
times are rather loosely defined by a time window comprising several periods. Time
windows of blocks overlap depending on the degree of flexibility. On the other hand,
for rigid block planning, one block is scheduled within the assigned period (e.g., a day,
week, or month). Although production quantities may vary, the production run is not
allowed to reach into the next period or to start in the previous one. Demand, either
based on customer orders or forecasts, is aggregated into demand elements. Each
demand element is defined by the product characteristics, the quantity, and the due date.
In figure 16, a flexible block planning approach is illustrated; in this approach, the
quantities of three demand elements for one specific product m can be assigned to one
of the blocks or split up into a number of them. The block planning approach is based
on a continuous representation of time. This has the advantage of a reduced number of
variables and constraints as pointed out by (Günther et al. 2006). The block planning
approach has currently gained much attention, especially in the consumer goods
industry, because solutions are highly applicable in practice and optimization models
are very efficiently solved by using standard mixed-integer solvers (Lütke Entrup et al.
2005, Günther et al. 2006).

Demand element e1 e2 e3
(product m)

Production order
1 … m … M 1 … m … M 1 … m … M
Setup operation

Block 1 Block 2 Block 3


Time

Figure 16: Flexible block planning concept

A classic discrete time scale model formulation has been compared to a model
formulation based on the block planning approach. As a production environment, the
production batching and scheduling approach of the CC that has sequence dependent set
56 3 Scheduling concepts with focus on the steel industry

up times was chosen, and a discrete model formulation similar to the DLSP presented in
Section 2.2.4 was compared to the block planning approach presented in section 4.3.
This discrete model formulation can be found in Appendix 1. The size of the discrete
time scale model depends on the number of periods, products (i.e., steel grades), and
CCs. In the production setting of the steel industry a realistic value for the length of a
time period is the time that is needed to produce the minimum batch size (e.g. 250 tons)
including a minor setup operation. Depending on the continuous caster this takes about
1.5 hours. For a production plan with a planning horizon of 28 days and a length of a
time period of 1.5 hours, the number of periods equals T= 28∙24÷1.5= 448. Other
realistic index values are M=200 steel grades, and N=2 CCs. This production setting
leads to M·N·(M·T+T) = 36,019,200 binary variables, M·T·(N+1)+1 = 268,801
continuous variables, and T∙(M2·N+M·N+M+N) = 36,109,696 constraints. This is
clearly a model formulation that can no longer be solved with standard optimization
software.

The block planning approach that optimizes the same production environment is
presented in section 4.3. To estimate the number of variables and constraints, let T be
the number of blocks and one block for each period of 24 hours, T(e) the number of
blocks from which a demand element can be filled, N the number of casters, M the
number of steel grades, E the number of demand elements, and E(n) the number of
demand elements that can be produced on caster n. Then, the number of binary variables
can be determined as M·N·T. Additionally, the number of continuous variables can be
written as N·E(n)·T(e)+3·N·T+1, and the number of constraints equals N·(1+E(n)
·T(e)+5·T+M·T)+E. Based on realistic values from a steel industry example, i.e., M=200,
E=700, T=28, E(n)=350, T(e)=9, and N=2, the number of binary variables equals
11,200, while the number of continuous variables and the number of constraints equal
6,469 and 18,482, respectively. Despite the high number of products and demand
elements, the model formulation would still be solvable with standard optimization
packages. This comparison highlights the fact that the block planning method is a
powerful modeling approach with comparatively low complexity and no loss in
practical applicability.
4 Scheduling of CCs and HSMs in the steel industry 57

4 Scheduling of CCs and HSMs in the steel industry

4.1 Problem description

This work focuses on two production steps continuous caster (CC) and hot strip mill
(HSM), as illustrated in figure 17. Interestingly, production scheduling can have a major
impact on energy efficiency at these production steps.

Continuous caster Hot strip mill

Steel Slab Steel


coils

Furnaces Roughing mill Finishing mills Coiler

Figure 17: Production process from CC to HSM (Lopez et al. 1998)

In practice, these production steps are often decoupled, which means that the output
from the CC (i.e., the slabs) completely cools down in the slab yard before being pushed
to further processing at the HSM, where they are again heated. This scenario is
necessary due to different technological conditions that determine the sequence of
production activities at the CC and the HSM and the planner’s aim to reduce the overall
scheduling complexity. The leading criterion for CC scheduling is steel grade; indeed,
production planners seek to schedule large batches or casts of similar steel grades in an
effort to load the converter in the steel mill multiple times. On the contrary, the HSM
requires a production run based on the dimensions and the hardness of the final coils.
Within a production run at the HSM, the sequence of the slabs is characterized by
increasing widths of the final coil at the beginning of a program with the aim to warm
up the rolls of the mill. After approximately 1/3 of the production run, the milling
program processes coils with decreasing widths to reduce the contact of the steel strips
with the worn out edges of the rolls. Due to the shape of this sequence, it is mostly
referred to as the “coffin rule” (figure 18). In addition to the “coffin shape” of a
production run, steep changes in the hardness from one slab to the next should be
avoided.

I. Mattik, Integrated Scheduling of Continuous Casters and Hot Strip Mills, Produktion und Logistik,
DOI 10.1007/978-3-658-03775-8_4, © Springer Fachmedien Wiesbaden 2014
58 4 Scheduling of CCs and HSMs in the steel industry

Width of the coil


(ca. 600 – 2000 mm)

Length of the
production run
(ca. 5000 t)

Figure 18: "Coffin rule" for HSM scheduling

Although complex to achieve, a closer connection that leads to reduced waiting times
for production orders in the slab yard has several advantages. A direct result is a
reduction in logistic efforts since a reduced waiting time reduces the number of slabs
stored between production steps. For example, if the average waiting time is reduced
from 5 days to one day, the average number of slabs in the slab yard is also reduced by
80%. An even more interesting aspect is the potential reduction in energy consumption
by charging warm slabs into the furnaces of the HSM. A first estimation, illustrated in
table 2, shows the potential cost savings. To heat up a slab from room temperature to a
milling temperature of about 1,300°C requires between 1.5-1.9 GJ/ton depending on the
furnace efficiency (Rentz et al. 1999, VdEh 2012), hence a slab with a slab weight of 20
tons requires about 30 GJ. The average price of natural gas, mostly used as input energy
source for hot strip mill furnaces, of 9.5 EUR/GJ (BMWi 2012) leads to energy costs of
285 EUR/slab if the slab is cold charged. Slabs that are stored in the slab yard for about
a day usually retain a temperature of 300 – 400°C as illustrated in figure 19. Hence, we
assumed a charging temperature of 300°C instead of room temperature (i.e., 25°C). If
the slab is already hot charged with a temperature of 300°C, about 20% in energy
savings can be achieved, since the necessary temperature increase reduces from 1275°
'W
to 1000° and the specific thermal capacity ( c ) shows a linear connection
m ˜ 'T
4 Scheduling of CCs and HSMs in the steel industry 59

between the necessary energy (ΔW) to obtain a required temperature increase (ΔT)
depending on the mass (m). The cost saving of a hot charging versus a cold charging
process therefore accounts for about €61 per slab. With an average production of about
300,000 tons per month, the saving would amount to about €915,000. This illustrates
that improving the connection between the production processes is a worthwhile task.

Table 2: Estimated energy savings potential


Price of natural gas (€/GJ) 9.5
Average slab weight (tons/slab) 20
Energy consumption of the HSM furnaces
1.5
(GJ/t)
Slab temperature - cold charging (°C) 25
Slab temperature - hot charging (°C) 300
Milling temperature (°C) 1,300
Energy consumption - cold charging (GJ/slab) 30
Energy consumption - hot charging (GJ/slab) 23.5
Energy costs - cold charging (€/slab) 285
Energy costs - hot charging (EUR/slab) 224

Figure 19: Temperature-time diagram of steel slabs based on Tata Steel (Wachter
2011: 74)
60 4 Scheduling of CCs and HSMs in the steel industry

As previously indicated, integration has many advantages and also provides financial
benefits, but some disadvantages are also associated with this approach. For example,
the scheduling becomes far more complex if the technological scheduling requirements
of both production steps must be considered. Another disadvantage is a higher risk for
production downtimes due to the lesser availability of safety or buffer stocks. Hence, if
the CC breaks down and needs maintenance, the HSM will also need to be stopped
when the amount of slabs in the slab yard is too low to build valid production runs;
otherwise, the HSM must be slowed down to avoid this scenario.

As implied in section 3.1, a variety of challenges, e.g., the technological inflexibility of


production plants, the huge number of customer orders, and the large variety of product
options, faces the steel industry. To develop a production schedule for a short-term
planning period, e.g., one month, a flexible and efficient planning approach is needed.
In this section, an approach to integrate these orientations with consideration of an
industrial application from the steel industry is presented

4.2 Decomposition approach

The scheduling problem at hand is divided into two sub-models, i.e., the CC and the
HSM model. The overarching aim is to reduce the slab waiting time between both
production steps to enable a hot charging connection between the CC and the HSM and
consequently reduce the energy consumption of the HSM furnaces. The approach was
tested with two decomposition approaches in order to investigate different planning
directions. The first decomposition approach works with the push principle. In other
words, the first production step, i.e., the CC, is scheduled first, and the second
production step, i.e., the HSM, must react to the caster schedule and is thus scheduled
afterwards (figure 20).
4 Scheduling of CCs and HSMs in the steel industry 61

Push-Direction

Predefined sequence of casts based on


technological requirements

Continuous caster model (CC)


Determines the quantity and timing
of production runs and assigns them
to the CC

Start and end time of blocks


at the continuous caster

Hot strip mill model (HSM)


Assigns output from the
CC to milling programs

Final sequence of slabs

Figure 20: Decomposition approach – The push-principle

The second decomposition approach works via a pull system. In other words, the second
production step, i.e., the HSM, is scheduled first, and the schedule for the first
production step, i.e., the CC, must react to the schedule for the HSM (figure 21).
62 4 Scheduling of CCs and HSMs in the steel industry

Pull-Direction

Predefined sequence of milling programs to


reduce energy consumption

Hot strip mill model (HSM)


Determines the quantity and timing
of production runs and assigns them
to the HSM

▪ Start and end time of blocks at the


hot strip mill
▪ Predefined sequence of casts based
on technological requirements

Continuous caster model (CC)


Assigns virtual production quantities
from the HSM to casts

Feasible schedule on both


production steps

Figure 21: Decomposition approach – The pull-principle

Both planning directions, i.e., push or pull, require changes in the model formulations.
The first model scheduled (i.e., push-direction: CC, pull-direction: HSM) creates a
production schedule with the objective to minimize the makespan of all production
orders by reducing energy-intensive, unproductive time. The second model scheduled
(i.e., push-direction: HSM, pull-direction: CC) leverages the output from the first model
to schedule production orders with the objective to reduce the waiting time of slabs
between the CC and the HSM. Additional model characteristics that are valid for both
planning directions are described subsequently.

Sequences within blocks in CC and HSM

In the CC model, the sequence of product types within each block is predefined
according to the chemical compatibility of the steel grades, thereby reducing the number
of major setup operations, which reduces the risk for casting breaks. To save energy and
to minimize changeover times in the HSM, the milling programs should be sequenced
according to the required furnace temperature for that grade. Since milling programs are
designed to group coils with similar dimensions and hardness, the required furnace
4 Scheduling of CCs and HSMs in the steel industry 63

temperature within one milling program stays in the same range but differs from the
other milling programs. Without steep temperature increases or decreases, the furnace
runs more efficiently, and unproductive time is decreased due to fewer slabs requiring
more time to reach the milling temperature.

Demand

The customer order quantities for the same steel grade that clearly use the same milling
program and have the same due date are grouped into demand elements. Each demand
element can be satisfied from a number of pre-defined blocks for which the latest finish
time precedes the due date of the demand element.

Production windows

Start and finish times for each block are not limited to period boundaries. They are
rather loosely defined by a window of time indicating the earliest start and latest finish
time of each block.

Connection of both production steps

To connect both models and fulfill the objective to reduce the waiting time between the
production steps, the production blocks from the model are numbered consecutively
according to sequence. Figure 22 illustrates this assignment for the push direction. For
the push direction, Index b=1, 2,…, B indicates blocks that are consecutively numbered
irrespective of the specific caster. Cast blocks are assigned to blocks at the HSM where
each block is processed by three different milling programs (i.e., p=1, 2, 3). For
instance, block 5 from the continuous caster is the only one completed before the end of
the first period and thus assigned to block 1 at the hot strip mill for processing in the
second period. Similarly, blocks 1 and 2 from the continuous caster, which are
completed in period 2, are assigned to block 2 at the hot strip mill for processing in
period 3. It should be noted that, depending on the customer order specifications, the
cast blocks are merged into the different milling programs available at the hot strip mill.
64 4 Scheduling of CCs and HSMs in the steel industry

Continuous Casters

Block 1 Block 2 Block 3 Block 4

Caster 1 b=1 b=2 b=3 b=4


Various steel grades/casts per block
Caster 2 b=5 b=6 b=7 b=8

b =5 b = 1,2 b = 6,7 b = 3,8 b=4

Hot strip mill Block 1 Block 2 Block 3 Block 4 Block 5


Milling
p = 1,2,3 p = 1,2,3 p = 1,2,3 p = 1,2,3 p = 1,2,3
programs

Period 1 Period 2 Period 3 Period 4 Period 5 Period 6


Time

Figure 22: Process flow from the CC to the HSM (i.e., the push-direction)

For the pull direction, Index b=1, 2,…, B indicates blocks that are also consecutively
numbered according to sequence and contain production quantities produced via up to 3
milling programs. With the help of the identifier index b, the CC model identifies the
HSM blocks and assigns them into feasible CC blocks.

4.3 Model formulations

4.3.1 Push-Principle
With a decomposition approach based on the push principle, the CC schedule is created
first, followed by the scheduling of the HSM on the basis of the output from the CC
step. Further assumptions are as follows:

x The CC model groups demand elements of the same steel grade into casts to
reduce idle time and setup time. thereby reducing the makespan.

x The HSM model groups casts that need the same milling program with a
production run in order to reduce the waiting time of casts between the CC and
the HSM.

x The production quantities from the CC model are identified by their block
number (i.e., b=1, 2,…, B), the total slab weight, the quantities that have to be
processed by a specific milling program, and their arrival time at the HSM.
4 Scheduling of CCs and HSMs in the steel industry 65

CC model
The following notation is used for formulas throughout this section:

Indices and sets


tϵT time period (block), (t=1, 2,…,T)
nϵN CC
mϵM sequence of specific steel grades over all blocks (m=1, 2,…,M)
eϵE demand elements
t ϵ T(e) preceding blocks from which demand element e can be satisfied
m(e) steel grade to which demand element e refers
E(m) set of demand elements e that belong to steel grade m
E(n) set of demand elements e that can be produced on caster n
Parameters
cs nc casting speed of CC n (tons/hour)
st c setup time between casts of different steel grades (hours)
dmec quantity of demand element e (tons)
mlnc minimum size of a cast on CC n (tons)
es ntc earliest start time of block t on CC n (hours)
lf ntc latest finish time of block t on CC n (hours)
Decision variables
c
Ymnt 1 steel grade m processed at CC n in block t (0, otherwise)
c
Z net quantity of demand element e satisfied from production in block t at CC
n (tons)
pt ntc processing time of block t at CC n (hours)
startntc start time of the block at CC n in time period t ( startnc0 0 )
End ntc finish time of block t at CC n
W c
makespan
The model formulation for the CC is given as follows:

Min W c (16)

W c t End nT
c
nϵN (17)

The objective function (16) minimizes the makespan, which is equivalent to the finish
time of the last cast/block defined in (17).

c
Z net d dmec ˜ Ymc(e) nt n ϵ N; e ϵ E(n); t ϵ T(e) (18)
66 4 Scheduling of CCs and HSMs in the steel industry

c
Constraint (18) models the relationship between the product flow Z net from block t at

caster n into demand element e and the binary setup variable Ymc( e) nt . The flow quantity

is forced to zero if no corresponding setup operation is performed, i.e., Ymc(e) nt 0.

¦ ¦ Z net
c
dmec eϵE (19)
n tT ( e )

Constraint (19) ensures that the quantities of all demand elements are satisfied from
production runs at the casters during the relevant preceding periods.

§ ¦ Z net c ·
¨ eE ( m) c ¸
ptntc ¦¨ c
 st ˜ Ymnt ¸ n ϵ N; t ϵ T (20)
m
¨ csnc ¸
© ¹

Equation (20) determines the processing time of block t at caster n based on the weight
of the different produced steel grades and the average casting speed of the CC. Before
the production of a specific steel grade, a setup time is incurred.

startntc 1  ptntc 1 d startntc n ϵ N; t ϵ T: t ≥2 (21)

According to (21), a block is allowed to start as soon as the preceding block has been
completed.

startntc t es ntc n ϵ N; t ϵ T (22)

End ntc startntc  pt ntc n ϵ N; t ϵ T (23)

End ntc d lf ntc n ϵ N; t ϵ T (24)

Constraints (22) to (24) refer to the time window for each block given by the parameter
es ntc (earliest start time) and lf ntc (latest finish time).

¦Z
eE ( m )
c
net t mlnc ˜ Ymnt
c
m ϵ M; n ϵ N; t ϵ T (25)

According to (25), the size of a cast is not allowed to drop below the minimum size of a
converter load.

Finally, the variable domains are defined as follows:


4 Scheduling of CCs and HSMs in the steel industry 67

c
Ymnt  ^0;1` m ϵ M; n ϵ N; t ϵ T (26)

c
Z net t0 n ϵ N; e ϵ E(n); t ϵ T(e) (27)

pt ntc t 0 n ϵ N; t ϵ T (28)

start ntc t 0 n ϵ N; t ϵ T (29)

End ntc t 0 n ϵ N; t ϵ T (30)

Wc t0 (31)

HSM model

The following notation is used for formulas in this section:

Indices and sets


pϵP sequence of specific milling programs at the HSM (p=1, 2,…,P)
tϵT time period (block), (t=1, 2,…,T)
bϵB blocks (steel slabs) from the CC
t  T (b) Ž T feasible blocks at the HSM for the assignment of caster block b; for
t≤25: t  ¬at bh ,0.9 ˜ lf t h ¼ and t>25: t  ¬at bh ,0.95 ˜ lf t h ¼
b  B(t ) Ž B caster blocks that can be assigned to block t at the HSM, i.e., blocks
with at bh  ¬es th ,0.9 ˜ lf t h ¼ for t≤25 and at bh  ¬es th ,0.95 ˜ lf t h ¼ for t>25
Parameters
sp h milling speed of the HSM (tons/hour)
h
st setup time for the HSM (hours)
le hp minimum size of a production run for milling program p (tons)
esth earliest start time of block t
lf t h latest finish time of block t
h
Qbp quantity produced in block b at the CC to be processed by milling
program p at the HSM (tons)
S bp
h
binary parameter that equals 1 if casts from block b have to be
processed by milling program p at the HSM (0, otherwise), note that
one cast block can be processed by several milling programs
h
at b arrival time of block b at the HSM (finishing time of the cast block at
the CC including a transportation time)
68 4 Scheduling of CCs and HSMs in the steel industry

Decision variables
Ypth 1 milling program p is processed at the HSM in block t (0, otherwise)
U bth 1 if block b from the caster is processed at the HSM in block t (0,
otherwise)
ptth processing time of block t at the HSM (hours)
startth start time of block t at the HSM
startcast bth start time of production of block b from the CC at the HSM in block t
' maximum waiting time of a block between the CC and the HSM

The model formulation for the HSM is as follows:

Min ' (32)

' t startcast bth  at bh b ϵ B; t ϵ T (33)

The objective function (32) along with constraint (33) minimizes the maximum cast
waiting time between their arrival at the HSM from the CC and the start time of
processing at the HSM.

¦ U bth 1 bϵB (34)


tT (b )

Constraint (34) ensures that each block from the CC is assigned to exactly one block at
the HSM for which the earliest start time does not precede the arrival time of the caster
block.

S bp
h
˜Ubth d Ypth b ϵ B; p ϵ P; t ϵ T(b) (35)

According to constraint (35), milling program p must be set up in block t enforcing


Ypth 1 if block b from the CC is assigned to it; i.e., U bth 1 , and the caster block

requires the respective milling program (i.e., parameter S bp


h
has a value of 1).

¦ ¦ Qbph ˜U bth
bB (t ) pP
ptth  ¦ st h ˜ Ypth tϵT (36)
sp h pP

startth1  ptth1 d startth t ϵ T: t ≥2 (37)


4 Scheduling of CCs and HSMs in the steel industry 69

The processing time of each block is determined in (36) based on the total slab weight
assigned to block t at the HSM, the average rolling speed, and the setup time for all
milling programs activated in the block. According to (37), a block cannot start before
the end of the preceding block.

startth t esth tϵT (38)

startth  ptth d lf t h tϵT (39)

Constraints (38) and (39) limit the start time of a block to a given window of time.

startth t startcastbth  lf Th ˜ (1  U bth ) b ϵ B; t ϵ T(b) (40)

startcastbth t startth  lf Th ˜ (1  U bth ) b ϵ B; t ϵ T(b) (41)

startcastbth t atbh b ϵ B; t ϵ T(b) (42)

Constraints (40) and (41) ensure that the start time, startth , of a block at the HSM equals

the start time, startcastbth , for processing a cast from the CC only if the corresponding

assignment is active, i.e., U bth 1 . Otherwise, these constraints are not binding due to

the subtraction of the large positive value, lfTh , on the right hand side of (25) and (26).
In addition, the start time for the processing of a cast is determined by its arrival time at
the HSM (42). It should be noted that, in case several cast blocks, say b and b', are
assigned to one block t at the hot strip mill, the start time of this block startth is

assumed as the start time for all the assigned cast blocks, i.e. startcastbth = startcastbh' t =

startth . This assumption is justified by the need to prepare the entire input material once
the block at the hot strip mill is started.

startcastbth d lf t h  ptth  lf Th ˜ (1  U bth ) b ϵ B; t ϵ T(b) (43)

The start time of a block from the caster needs to be earlier or equal to the latest finish
time of the assigned block at the HSM minus the production time provided that the
corresponding assignment is active, i.e., U bth 1 (43).

¦ Qbph ˜U bth t lehp ˜ Ypth p ϵ P; t ϵ T (44)


bB (t )
70 4 Scheduling of CCs and HSMs in the steel industry

Constraint (44) enforces a minimum size for each milling program and block.

Finally the variable domains are defined as follows:

Ypth  ^0;1` p ϵ P; t ϵ T (45)

U bth  ^0;1` b ϵ B; t ϵ T(b) (46)

pt th t 0 tϵT (47)

startth t 0 tϵT (48)

startcastbth t 0 b ϵ B; t ϵ T(b) (49)

't0 (50)

Average calculation of waiting times between the two processes

The average waiting time of each cast was calculated using the following formula
( startcast bth  atbh )
average . Additionally the constraints (51) and (52) were added to
B
ensure that the variable startcastbth equaled the arrival time at the HSM, at bh , if U bth 0.

Hence, the subtraction yields startcastbth  atbh 0 if U bth 0.

startcastbth t atbh  lf Th ˜U bth b ϵ B; t ϵ T(b) (51)

atbh t startcastbth  lf Th ˜ U bth b ϵ B; t ϵ T(b) (52)

4.3.2 Pull-principle
With a decomposition approach based on the pull principle, the HSM schedule is
created first, followed by the scheduling of the CC based on the output of the HSM
model. Therefore, the HSM model is presented first, followed by the CC model. The
following assumptions also apply:

x The HSM model is designed to group casts requiring the same milling program
to a production run in an attempt to reduce the makespan.
4 Scheduling of CCs and HSMs in the steel industry 71

x The CC model is designed to group casts that belong to the same steel grade to a
production run in an attempt to reduce the waiting time for the following
production step, i.e., the HSM.

x The production quantities from the HSM are identified as they leave the CC by
their virtual block number (b=1, 2,…,B), total slab weight, quantities that have
to be processed by a specific steel grade, and the latest release time to the HSM.

HSM model

The following notation is used in this section:

Indices and sets


tϵT time period (block), (t=1, 2,…,T)
pϵP sequence of specific milling programs at the HSM (p=1, 2,…,P)
eϵE blocks (i.e., steel slabs) from the CC
t ϵ T(e) preceding blocks from which demand element e can be satisfied
p(e) milling program to which demand element e refers
E(p) set of demand elements e that belong to milling program p
Parameters
sp h milling speed of the HSM (tons/hour)
st h setup time for the HSM (hours)
le hp minimum size of a production run for milling program p (tons)
mle h maximum size of a production run p (tons)
esth earliest start time of block t
h
lf t latest finish time of block t
dmeh quantity of demand element e (tons)
Decision variables
Ypth 1 milling program p processed at the HSM in block t (0, otherwise)
h
Z et 1 fraction of demand element e produced in block t (tons) (0, otherwise)
h
pt t processing time of block t at the HSM (hours)
startth start time of block t at the HSM
End th finish time of production of block b at the HSM in block t
W makespan
The model formulation for the HSM is as follows:

Min W (53)
72 4 Scheduling of CCs and HSMs in the steel industry

W t End Th (54)

The objective function (53) minimizes the makespan, which is equivalent to the finish
time of the last block as defined in (54).

Z eth d Y ph( e)t e ϵ E; t ϵ T(e) (55)

Constraint (55) models the relationship between the product flow Z eth from block t into

demand element e and the binary setup variable Y ph( e )t . The flow quantity is forced to

zero if no corresponding setup operation is performed, i.e., Yph( e )t 0.

¦Z
tT ( e )
h
et 1 eϵE (56)

Constraint (56) ensures that quantities of all demand elements are satisfied from
production runs at the HSM in the relevant preceding periods.

¦Z h
et ˜ dmeh
 ¦ st h ˜ Y pth
h eE ( p )
pt t tϵT (57)
sp h p

startth1  ptth1 d startth t ϵ T: t ≥2 (58)

The processing time for each block is determined in (57) based on the total slab weight
assigned to block t at the HSM, the average rolling speed, and the setup time for all
milling programs activated in the block. According to (58), a block cannot start before
the end of the preceding block.

startth t esth tϵT (59)

End th startth  ptth tϵT (60)

End th d lf t h tϵT (61)

A window of time for each block is given in (59) and (61) by the parameter esth (i.e.,

earliest starting time) and lf t h (i.e., latest finish time). Constraint (60) defines the finish
time of a block.

¦Z
eE ( p )
h
et ˜ dmeh t le hp ˜ Ypth p ϵ P; t ϵ T (62)
4 Scheduling of CCs and HSMs in the steel industry 73

¦Z
eE ( p )
h
et ˜ dmeh d mle h ˜ Y pth p ϵ P; t ϵ T (63)

Constraints (62) and (63) enforce a minimum and maximum size for each milling
program and block.

Finally the variable domains are defined as follows:

Ypth  ^0;1` p ϵ P; t ϵ T (64)

Z eth  {0;1} e ϵ E; t ϵ T(e) (65)

pt th t 0 tϵT (66)

startth t 0 tϵT (67)

End th t 0 t ϵ T) (68)

W t0 (69)

Continuous caster model

The following notation is used for formulas in this section:

Indices and sets

tϵT time period (block), (t=1, 2,…,T)


nϵN CC
mϵM sequence of specific steel grades over all blocks (m=1, 2,…,M)
bϵB virtual blocks in which the slab is planned for rolling
t  T (b) Ž T feasible blocks at the CC for the assignment of virtual rolling block b,
i.e., t  ¬1.1 ˜ rtbc , lf ntc ¼ for t≤25 and t  ¬1.1 ˜ rtbc ,1.05 ˜ lf ntc ¼ for t>25
b  B(t ) Ž B virtual rolling blocks that can be assigned to block t at the CC, i.e.,
rtbc  ¬1.1 ˜ es ntc , lf ntc ¼ for t≤25 and rtbc  ¬1.1 ˜ es ntc ,1.05 ˜ lf ntc ¼ for t>25
m ϵ M(n) set of steel grades m that can be produced on caster n
Parameters
cs nc casting speed of CC n (tons/hour)
st c setup time between casts of different steel grades (hours)
mlnc minimum size of a cast on CC n (tons)
74 4 Scheduling of CCs and HSMs in the steel industry

es ntc earliest start time of block t on CC n (hours)


lf ntc latest finish time of block t on CC n (hours)
c
Qbm quantity to be produced in block b at the HSM that belongs to steel
grade m (tons)
S bm
c
binary parameter that equals 1 if coils from block b belong to steel
grade m (0, otherwise), note that one HSM block can include several
steel grades
rtbc Latest release time of block b to the HSM (start time of block b at the
HSM including a transportation time)
Decision variables
c
Ynmt 1 CC n processes steel grade m in block t (0, otherwise)
c
U bnt 1 block b from the HSM processed at the CC n in block t (0, otherwise)
pt ntc processing time of block t at CC n (hours)
startntc start time of the block at CC n in time period t ( startnc0 0)
End ntc finish time of steel grade m in block t at CC n
c
endcast bnt finishing time of the production of virtual block b from the HSM at CC
n in block t
' maximum waiting time for a block between the CC and the HSM

The model formulation for the CC is given as follows:

Min ' (70)

' t rtbc  endcastbnt


c
b ϵ B; n ϵ N; t ϵ T(b) (71)

The objective function (70) along with constraint (71) minimizes the maximum waiting
time for casts between the ending time at the CC and the release time to the HSM.

¦U
tT ( b )
c
bnt 1 b ϵ B; n ϵ N (72)

Constraint (72) ensures that each virtual block from the HSM is assigned to exactly one
block at the CC for which the earliest start does not precede the latest leaving time for
the HSM.

S bm
c
˜ U bnt
c
d Ynmt
c
b ϵ B, m ϵ M(n) ; n ϵ N; t ϵ T(b) (73)
4 Scheduling of CCs and HSMs in the steel industry 75

c
According to constraint (73), steel grade m must be set up in block t enforcing Ymnt 1 if
c
virtual block b from the HSM is assigned to it, i.e., U bnt 1 . Moreover, the HSM block

requires the respective steel grade; i.e., parameter S bm


c
has a value of 1.

§ ¦ Qbm c
˜ U bnt
c
·
¨ bB (t ) ¸
pt c
nt ¦ ¨
mM ( n ) ¨ cs c
 st ˜ Ynmt ¸
c c
n ϵ N; t ϵ T (74)
n ¸
© ¹

The processing time of each block is determined in (74) based on the total slab weight
assigned to block t at the CC, the average casting speed, and the setup time for all steel
grades activated in the block. According to (75), a block cannot start before the end of
the preceding block.

startntc 1  ptntc 1 d startntc n ϵ N; t ϵ T: t ≥2 (75)

startntc t es ntc n ϵ N; t ϵ T (76)

End ntc startntc  pt ntc n ϵ N; t ϵ T (77)

End ntc d lf ntc n ϵ N; t ϵ T (78)

Constraints (76) and (78) limit the start time of a block to the given window of time.
Constraint (77) defines the ending time as the start time plus the production time.

End ntc t endcastbnt


c
 lf nTc ˜ (1  U bnt
c
) b ϵ B; n ϵ N; t ϵ T(b) (79)

c
endcast bnt t End ntc  lf nTc ˜ (1  U bnt
c
) b ϵ B; n ϵ N; t ϵ T(b) (80)

c
endcast bnt d rtbc b ϵ B; n ϵ N; t ϵ T(b) (81)

Constraints (79) and (80) ensure that the start time End ntc of a block at the CC equals
c
the ending time endcast bnt for processing a virtual block from the HSM only if the
c
corresponding assignment is active, i.e., U bnt 1 . Otherwise, these constraints are not

binding due to the subtraction of the large positive value, lf nTc , on the right hand side of
(79) and (80). In addition, the ending time for processing a cast is determined by its
latest release time at the CC (81).
76 4 Scheduling of CCs and HSMs in the steel industry

c
endcastbnt d lf t c  lf nTc ˜ (1  U bnt
c
) b ϵ B; n ϵ N; t ϵ T(b) (82)

The ending time of a virtual block from the HSM at the caster must be smaller or equal
to the latest finish time at the caster provided that the corresponding assignment is
c
active, i.e., U bnt 1 (82).

¦Q
bB ( t )
c
bm ˜ U bnt
c
t mlnc ˜ Ynmt
c
m ϵ M(n); n ϵ N; t ϵ T (83)

According to (83), the size of a cast is not allowed to drop below the minimum size of a
converter load.

Finally, the variable domains are defined as follows:


c
Ymnt  ^0;1` n ϵ N; m ϵ M(n); t ϵ T (84)

c
U bnt  ^0;1` b ϵ B; n ϵ N; t ϵ T(b) (85)

pt ntc t 0 n ϵ N; t ϵ T (86)

start ntc t 0 n ϵ N; t ϵ T (87)

End ntc t 0 n ϵ N; t ϵ T (88)

c
endcast bnt t0 b ϵ B; n ϵ N; t ϵ T(b) (89)

't0 (90)

Average calculation of waiting times between the two processes

The average waiting time for each cast was calculated using the following formula:
(rtbc  endcasts bnt
c
)
average . Additionally, constraints (91) and (92) were added to
B˜ N
c
ensure that the variable endcast bnt equaled the latest release time to the HSM, rtbc , if
c
U bnt 0 . Hence, rtbc  endcasts bnt
c c
0 if U bnt 0.

c
endcastbnt t rtbc  lf nTc ˜ U bnt
c
b ϵ B; n ϵ N; t ϵ T(b) (91)

rtbc t endcastbnt
c
 lf nTc ˜ U bnt
c
b ϵ B; n ϵ N; t ϵ T(b) (92)
5 Computational Experiments 77

5 Computational Experiments

5.1 Experimental design


To test the proposed integrated scheduling approach in both scheduling sequences,
problem instances were randomly generated. Corresponding parameter values and the
characteristics of the scheduling tasks, e.g., the sequencing conditions or minimum
batch sizes, were based on the real-life scheduling settings of a leading European steel
company. We consider a steel plant with two CCs that differ in terms of production
speed and product quality. Therefore, each steel grade is assigned to one specific caster
based on product quality restrictions. Although the number of possibly produced steel
grades is high due to minimum batch sizes of 250 tons and 500 tons at the CC and the
HSM, respectively, a limit of up to 35 steel grades can be produced on both casters per
day. Throughout the experiments, a short-term planning horizon of 28 days was
considered. The plant operated 7 days per week and 24 hours per day.

Basic scenarios

In our experiments, we consider two capacity load scenarios of 80% and 90%, which
reflect the conditions of average and high workload, respectively. In addition, two
product portfolio scenarios were investigated, i.e., one scenario with a product portfolio
of 100 and the second scenario with 200 different steel grades. This reflects a current
trend in the industry, i.e., companies with large product portfolios and high operational
complexity and specialized companies with smaller product portfolios and thus easier
operating conditions. Moreover, three different milling programs were considered
resulting in a total of 300 and 600 product options, respectively.

Determining the total production volume

Based on industrial experience, we assumed a capacity loss of 15% due to the need for
the setup operations at the CC to react to the HSM schedule. Given a production speed
of 300 t/hour at the first CC and 250 t/hour at the second CC and 24 hours in an
operational day, the total production volume at the CC would be 302,940 tons.
Although the production speed of the HSM is 500 t/hour, this facility does not
experience an overall bottleneck due to significantly fewer setup operations for the
HSM. Based on the total production volume of the CC scheduled with the pull-

I. Mattik, Integrated Scheduling of Continuous Casters and Hot Strip Mills, Produktion und Logistik,
DOI 10.1007/978-3-658-03775-8_5, © Springer Fachmedien Wiesbaden 2014
78 5 Computational Experiments

principle, the total production volume of the 80% capacity load scenario and the 90%
capacity load scenario would be 242,352 tons and 272,646 tons, respectively.

Generation of demand elements

Demand elements are characterized by their size, steel grade, milling program, and due
date. In the random generation of example problems, these specifications were
determined as follows:

x The size of a demand element was drawn from the uniform distribution
U[250,500], resulting in 727 and 646 demand elements for the 90% and the 80%
capacity load scenarios, respectively.

x The respective steel grade was determined randomly with equal probability for
the 100 and 200 steel grades in the small and large product portfolio scenarios,
respectively. Similarly, one of the three milling programs was assigned.

x A share of 25% of the demand elements was assigned to periods 7, 14, 21, and
28.

Production setup

As previously indicated, the considered industrial application contains two CCs that
differ in terms of production speed and product quality and one HSM. Due to the
differing product surface quality on the CCs, each steel grade can only be produced on
one of the two casters. From the entire portfolio of steel grades, only a limited number is
produced on each caster per day. The mix of product types and quantities for each block
are not necessarily the same since decision variables for each block indicate if the
production process is set up for a particular product, e.g., steel grade or milling
program, and which quantities of the respective products are to be produced. At the
HSM, setup times are product-independent because the required setup operation always
involves the exchange of the work rolls that requires approximately the same amount of
time and does not depend on the production program. At the CC, the product-dependent
major setup times are considered in the pre-defined sequence of steel grades, which is
based on their chemical compatibility. The remaining minor and product-independent
setup times are included in the production time for each block.

Determination of T(e)
5 Computational Experiments 79

In order to realistically mimic the build-up of inventories, the CC model (i.e., if the push
strategy is used) and the HSM model (i.e., if the pull strategy is used) utilize an index
set, T(e), which comprises the feasible production periods for each of the demand
elements. The side-effect of keeping this index set small is that the size of the
optimization model can be accordingly reduced. In a real application, overlapping time
windows are necessary to allow the batching of demand elements in an attempt to save
setup operations and reduce energy consumption. In the generation of the example
problems, index sets T(e) always overlaps by 4 days, as illustrated in Figure 18.

Demand elements

25% t = 1-7 +2

T(e)

25% -2 t = 8-14 +2

T(e)

25% -2 t = 15-21 +2

T(e)

25% -2 t = 22-28

T(e)
t

Figure 23: Distribution of T(e)

Determination of B(t) and T(b)

The data sets B(t) and T(b) indicate feasible blocks in which the HSM (i.e., in the push-
principle) is allowed to allocate production quantities that are transferred from the CCs
as illustrated in figure 24. They reduce the number of blocks that the model needs to
check for the assignments of production quantities while simultaneously reducing the
computation time. The data sets are based on the idea that a feasible block in which the
HSM might allocate production quantities to the CC requires a production window
interval that spans from the arrival time of the caster block at the HSM to the finish time
of the production window. Since the cast blocks still need enough time for processing in
the HSM when they are allocated to a specific production window, the finish time is
reduced by a factor of 0.9 or 0.95. For the last production periods, this factor is higher
because a reduction of 10% would reduce the solution space too much due to the higher
digits of the latest finish time (e.g., 632 hours).
80 5 Computational Experiments

PUSH PRINCIPLE

Arrival time too close


Arrival time within pro- to latest finish time of
duction window interval Cast Cast production window –
– allocation possible Block Block allocation not possible
b=1 b=2
atb atb

esnt lfnt
Production window

Figure 24: Determination of T(b) and B(t)

A similar procedure for creating T(b) and B(t) for the CC when the model is scheduled
using the pull-principle is carried out. The only differences are the digits of the factors
to account for the specific needs. The earliest starting time is increased by a factor of 1.1
since the production quantities at the CC still need to be processed before they are
transferred to the HSM. Additionally, for the last periods, the latest finishing time is
increased by a factor of 1.05, because the release time is greater than the latest finish
time for the last period at the CC. This is due to the longer production time at the HSM
(i.e., 672 hours compared to the 648 hours of the CC). This helps to reduce the model
complexity without significantly reducing the solution space.

Determination of production windows

According to the block planning principle, the number of blocks and the corresponding
time window during which a block can be executed must be defined in advance. In our
case study oriented investigation, one block per period (i.e., 24 hours) was assumed.
The proposed modeling approach was tested with two different time window scenarios.
For both scenarios, the initially scheduled production step has less flexible production
windows than the production step that was scheduled second because the reacting
process needs more flexibility to schedule orders from the first process due to slightly
differing production speeds and setup times.

In the flexible scenario, adjacent time windows for the initially scheduled process (i.e.,
pull-principle: HSM, push-principle: CC) overlap by 6 hours; i.e., the time window
comprises 36 hours in total. The flexible scenario for the production step scheduled
5 Computational Experiments 81

second (i.e., pull-direction: CC, push-direction: HSM) features adjacent time windows
that overlap by 32 hours; i.e., the time window comprises 88 hours in total.

In the less flexible scenario, the time window is set at 32 hours per block with an
overlap of 4 hours between adjacent blocks for the process scheduled first. For the
second scheduling step, the time window is set at 84 hours per block with an overlap of
30 hours.

To avoid end-of-horizon effects for models that optimize the makespan of all orders, we
assumed that the blocks for all periods greater or equal to 22 have the same earliest start
time. To ensure that no extensively long production runs are scheduled during periods
greater or equal to 22 if the HSM is scheduled with the pull-principle, a maximum
program length of 12,000 tons was included.

Owing to the two-stage production system, the last day of production at the CC has to
be eliminated from the planning horizon because no output could be finalized at the
HSM. Therefore, the time window of the final block at the CC ends after 27 days.
Similarly, for the HSM scheduled with the push-principle, the first day of production at
the HSM was eliminated because no output from the CC would be available yet. Hence,
the earliest start times of the first block are set at time zero for the CC and at 24 hours
for the HSM. Likewise, for the HSM scheduled with the pull-principle, breaks were
included at the beginning of each production week to compensate for the lesser amount
of time spent on setup operations as compared to the CC production. Hence, the earliest
start time of the first block is set to 50 hours, and for each start of a production week in
t=8, 15, 22, a break of 20 hours was included as illustrated in figure 25.

HOT STRIP MILL


Production t=1
windows
t=2
t=3
t=4
t=5
t=6
t=7
t=8

t=55 Δt=20 t

Figure 25: Production breaks at the hot strip mill (pull principle)
82 5 Computational Experiments

Determination of E(n)

At the considered steel plant, each steel grade has different surface quality restrictions;
therefore, all products are assigned to one specific CC. To account for this technical
requirement for the steel grades and consider the production speeds of both CCs, we
assumed that 55% of all steel grades were assigned to CC 1 (m=1…109 for the large
product portfolio and m=1…55 for the small product portfolio), while the remaining
45% of the steel grades were allocated to CC 2 (m=110…200 for the large product
portfolio and m=56…100 for the small product portfolio). In our experiments with the
push-principle, a description of the corresponding index sets E(n) follows subsequently.

With two workload scenarios of 80% and 90%, two product portfolio scenarios, and two
production window scenarios, eight different scenarios in total were investigated (see
table 3 for an overview). To facilitate the identification of the scenario characteristics in
the subsequent section, we introduced a scenario code as indicated in the final column
of table 3. For each scenario, five problem instances were randomly generated.

Table 3: Overview of all scenarios

Product
Scenario Workload Time window Code
portfolio
1 large (M=200) high (90%) flexible (200, 90%, f)

2 large (M =200) high (90%) less flexible (200, 90%, l)

3 large (M =200) average (80%) flexible (200, 80%, f)

4 large (M =200) average (80%) less flexible (200, 80%, l)

5 small (M =100) high (90%) flexible (100, 90%, f)

6 small (M =100) high (90%) less flexible (100, 90%, l)

7 small (M =100) average (80%) flexible (100, 80%, f)

8 small (M =100) average (80%) less flexible (100, 80%, l)

5.2 Results and discussion


Numerical investigations were conducted to test the suitability of the models for
industry application. The numerical tests were performed on a PC with an Intel(R) Core
5 Computational Experiments 83

i5 CPU with 2.4 GHz and 3 GB RAM. ILOG’s OPL Studio and its incorporated solver
CPLEX 12.1 were used as a modeling environment. The data set used to assess the
model was based on real data from a European steel company.

5.2.1 Results for the push principle


The variables and constraints (i.e., average values over five replications) are listed in
table 4 and table 5. The product portfolio of scenarios 1 through 4 is twice the size of
that of scenarios 5 through 8. This doubles the number of binary variables and increases
c
the number of constraints. The number of the continuous decision variables, Z net , in the
CC model depends on the data sets E(n), which are randomly generated, and this results
in a slightly different number of variables among the scenarios with a high workload
(1/2/5/6) and those with an average workload (3/4/7/8).

Table 4: Number of variables in the CC model (push-principle)


Binary Continuous Total no. of No. of
Scenario variables variables variables constraints

1/2 (200,90%, f/l) 11,200 7,441 18,641 19,478

3/4 (200,80%, f/l) 11,200 6,630 17,830 18,586

5/6 (100,90%, f/l) 5,600 7,441 13,041 13,878

7/8 (100,80%, f/l) 5,600 6,630 12,230 12,986

In the HSM, the binary decision variable, U bth , and the continuous variable, startcast bth ,
depends on the index set T(b), which is created based on the results of the CC model.
Hence, all scenarios of the HSM model also have a slightly differing number of
variables.
84 5 Computational Experiments

Table 5: Number of variables HSM model (push-principle)


Binary Continuous Total no. of No. of
Scenario variables variables variables constraints

1 (200, 90%, f) 274 248 522 2149

2 (200, 90%, l) 270 244 515 2115

3 (200, 80%, f) 273 247 521 2145

4 (200, 80%, l) 266 240 507 2075

5 (100, 90%, f) 272 246 517 2127

6 (100, 90%, l) 269 243 512 2103

7 (100, 80%, f) 271 245 515 2117

8 (100, 80%, l) 265 239 504 2061

The presented models for CC and HSM scheduling were tested separately. For the CC
model, we observed that a first feasible solution was obtained fairly quickly, followed
by a rather slow convergence of the CPLEX solver. Hence, we assessed this model with
a stopping criterion based on the relative solution quality. The test run was stopped after
the identification of the first solution with an MIP-gap smaller or equal to 1%. The MIP-
gap is the difference in percentage between the actual objective function value and a
theoretical lower boundary for the optimal objective function value. This theoretical
lower boundary is determined by a linear programming (LP) relaxation of the problem.
The solver then leverages branch-and-bound procedures to add the missing integrality
constraints to the LP relaxation. Therefore, the MIP-gap is the maximum improvement
potential of the objective function value.

The results of the test runs were evaluated based on the following three criteria: the
resulting makespan, the MIP-gap, and the computational time. The results of the CC
model are summarized in table 6. All values represent the average over five replications.
The general conclusion from these results is that nearly optimal results were achieved in
calculation times of less than a second for all scenarios. Specific observations are as
follows:
5 Computational Experiments 85

x An evaluation of the makespan values shows that flexible production windows


allow slightly shorter makespans as compared to that of less flexible production
windows (i.e., comparing scenario 1 with 2, 3 with 4, 5 with 6, and 7 with 8).
This effect achieves an improvement of about 0.3% in the scenario pairs 1/2,
3/4, 5/6, and 7/8.

x Comparing the corresponding pairs of high workload scenarios, i.e., 1/5 and 2/6,
shows that reducing the portfolio of steel grades from 200 to 100 reduces the
makespan by about 1%. This reduction in the portfolio of steel grades reduces
the average workload scenarios, i.e., 3/7 and 4/8, to 0.5%. These results
generally show that the block planning approach can handle the great variety of
product options and the large number of customer orders.

x As expected, the makespan results mirror the different workload levels (i.e.,
comparing scenario 1 with 3, 2 with 4, 5 with 7, and 6 with 8).

x Regarding calculation times, all scenarios are solved in less than a second, and
the differences are minor.

Table 6: Results of the continuous caster model (push-principle)


Makespan CPU MIP-Gap
Scenario
(h) time (s) (%)

1 (200, 90%, f) 595.73 0.33 0.04

2 (200, 90%, l) 598.06 0.36 0.14

3 (200, 80%, f) 579.72 0.25 0.02

4 (200, 80%, l) 581.63 0.21 0.00

5 (100, 90%, f) 590.88 0.34 0.01

6 (100, 90%, l) 592.88 0.27 0.01

7 (100, 80%, f) 576.79 0.35 0.04

8 (100, 80%, l) 578.66 0.26 0.02

In the second series of experiments, the HSM was evaluated. Due to the reduction in the
solution space, optimum solutions with closed MIP-gaps were obtained in up to 3
86 5 Computational Experiments

minutes. Although all scenarios were optimally solved, the stopping criterion was also
the first solution found with a MIP-gap of smaller or equal to 1% to ensure consistency
in the numerical tests of the pull-principle.

The general objective of the HSM is to close the time gap to the preceding CC stage as
much as possible in order to avoid excessive reheating of the casts. The summarized
results in table 7.(i.e., reported as an average over five replications) reveal that this
objective was achieved with average and maximum waiting times on the order of 10 and
20 hours, respectively, over all scenarios. All scenario were solved to optimality with
MIP-gaps of 0.

The less flexible production windows seem to positively impact solution quality for all
scenarios.

x Scenario 1 as compared to scenario 2 shows a reduction of 9.9% and 32.6% of


the maximum and average waiting times, respectively.

x The biggest differences in maximum and average waiting times were achieved
in a comparison of scenario 3 to scenario 4, i.e., 25.8% and 38.9%,
respectively.

x A comparison of scenario 5 and scenario 6 shows a reduction of 26.2% and


35.7% in the maximum and average waiting times, respectively. This is the
largest relative reduction in maximum waiting times.

x Comparing scenario 7 and scenario 8 shows relatively small differences, i.e.,


13.9% and 22.7% for the maximum and average waiting times, respectively.

Generally, the performance of the HSM was affected by the arrival times and
production quantities of the CC model. Detailed results show that as the batch sizes in
the CC increased, the maximum waiting time at the HSM also increased. Hence, the
maximum waiting times for the small product portfolio scenarios 5/6 and 7/8 are longer
than those of scenarios 1/2 and 3/4 with a large product portfolio.

Regarding the computational effort, the production windows prove to be the main
driver. All scenarios with less flexible time windows were optimally solved
significantly faster than their counterpart scenarios with flexible time windows (i.e.,
comparing scenario 1 and 2, 3 and 4, 5 and 6, and 7 and 8).
5 Computational Experiments 87

The computation times of scenarios 3 and 7 are about 8 to 10 times longer than the
computation times of the other scenarios. This is due to a single replication that proved
to be an outlier as compared to the other test runs.

Table 7: Results of the hot strip mill model (push-principle)

Maximum Average waiting CPU MIP-Gap


Scenario
waiting time (h) time (h) time (s) (%)

1 (200, 90%, f) 22.84 10.19 13.89 0.00

2 (200, 90%, l) 20.79 7.68 6.67 0.00

3 (200, 80%, f) 23.31 10.79 188.01 0.00

4 (200, 80%, l) 18.54 7.77 7.29 0.00

5 (100, 90%, f) 26.39 10.79 27.70 0.00

6 (100, 90%, l) 20.91 7.95 9.49 0.00

7 (100, 80%, f) 24.91 11.15 78.06 0.00

8 (100, 80%, l) 21.87 9.09 15.48 0.00

Overall, the CC model produced near-optimal solutions with low computational effort
(table 6). The results of the HSM model (table 7) illustrate that the integrated planning
of both production steps is in fact possible. The waiting times for casts were a lot
shorter than the average current industry practice. Total computational times for the CC
and the HSM are less than a minute in most cases and less than 190 seconds in all cases.
The results of one problem instance of scenario 1 are illustrated in a Gantt-Chart in
figure 26. The processing time also includes minor setup operations and with the help of
the block indicator it is apparent when the production quantities of the CC are processed
at the HSM.
88 5 Computational Experiments

b Block indicator b

t 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27

2 2 2 2 2
CC 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
3 4 5 6 7

54 56

CC 2 3 5 5
29 30 31 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52
2 3 5

9,
HSM 1, 2, 3, 4, 5, 6, 7, 8, 10, 11, 12, 14, 15,
16, 18, 20,
22,
23,
25-27
13 41 17, 45, 19 47 48 21, 24,
29 30 31 32 33 34 35 36 37, 39 40 42 43 50 52-56
44 46 49 51
38,

168 336 504 648

Figure 26: Gantt-Chart of one problem instance (push-principle)

5.3 Results for the pull-principle


The number of variables and constraints (i.e., reported as average values over five
replications) are presented in table 8 and table 9. In the HSM model, the high workload
in scenarios 1, 2, 5, and 6 increased the number of demand elements, which resulted in a
higher number of binary variables ( Z eth ). The number of continuous variables is
determined by the variables that indicate the start and end production times; hence, the
same number exists for all scenarios.

Table 8: Number of variables in the HSM model (pull-principle)


Binary Continuous Total no. of No. of
Scenario variables variables variables constraints

1/2/5/6 (100/200,90%, f/l) 7,355 86 7,441 8,306

3/4/7/8 (100/200,80%, f/l) 6,544 86 6,630 7,414

c c
In the CC model, the number of decision variables, i.e., U bnt and endcast bnt , depends on
the index set T(b), which was created based on the results of the HSM model. Hence, all
5 Computational Experiments 89

scenarios of the CC model have slightly differing number of variables. The scenarios for
the small product portfolio, i.e., scenarios 5 through 8, have less binary variables since
c
Ymnt depends on the data set M(n), and this is only half the size for the scenarios with
the small product portfolio.

Table 9: Number of variables for the CC model (pull-principle)


Binary Continuous Total no. of No. of
Scenario variables variables variables constraints

1 (200, 90%, f) 5,800 370 6,169 27,289

2 (200, 90%, l) 5,796 366 6,163 26,947

3 (200, 80%, f) 5,811 381 6,192 28,406

4 (200, 80%, l) 5,803 373 6,176 27,674

5 (100, 90%, f) 3,000 370 3,371 14,519

6 (100, 90%, l) 2,794 364 3,157 14,167

7 (100, 80%, f) 3,010 380 3,391 15,125

8 (100, 80%, l) 3,003 373 3,376 14,692

To allow comparability, both model tests, i.e., the push- and pull-principle, were
conducted in a consistent manner. Hence, both models for CC and HSM scheduling
were tested separately and were assessed with a stopping criterion based on the relative
solution quality. More specifically, all test runs were stopped after the first solution with
an MIP-gap that was smaller or equal to 1% was found.

The results of the test runs were also evaluated based on the following three criteria: the
resulting makespan, the MIP-gap, and the computational time. The results from the
HSM model are summarized in table 10. All reported values represent the average over
five replications. The general conclusion from these results is that nearly optimal results
were achieved in calculation times of less than 2 minutes for all scenarios. Specific
observations of the test results for the HSM model are as follows:

x An evaluation of the makespan values shows that flexible production windows


allow slightly shorter makespans of about 0.1% as compared to less flexible
90 5 Computational Experiments

production windows for the scenarios with a small product portfolio (i.e.,
comparing scenario 5 with 6 and 7 with 8). However, this effect is negligible for
the scenario pairs with a large product portfolio.

x Comparing the makespans of the large product portfolio scenarios with the
corresponding small product portfolio scenarios shows no effect since the steel
grade is not a complexity driver for the HSM model and the number of milling
programs is the same.

x As expected, the makespan results mirror the different workload levels (i.e.,
comparing scenarios 1with 3, 2 with 4, 5 with 7, and 6 with 8).

x Regarding calculation times, a considerable effect was observed for the different
product portfolio scenarios. The scenarios with the large product portfolios are
solved faster than the corresponding scenarios with small product portfolios (i.e.,
comparing scenarios 1 with 5, 2 with 6, 3 with 7, and 4 with 8).

Table 10: Numerical results for the HSM model (pull-principle)


Makespan CPU MIP-Gap
Scenario
(h) time (s) (%)

1 (200, 90%, f) 665.73 39.20 0.67

2 (200, 90%, l) 666.06 35.22 0.73

3 (200, 80%, f) 606.67 63.48 0.74

4 (200, 80%, l) 606.55 26.33 0.67

5 (100, 90%, f) 665.95 39.38 0.71

6 (100, 90%, l) 666.19 113.28 0.74

7 (100, 80%, f) 604.38 116.83 0.64

8 (100, 80%, l) 605.14 53.63 0.76

In the second series of experiments, the CC was evaluated. Solutions with MIP-gaps
smaller than 1% were obtained in less than 10 seconds.

The general objective of the CC model when scheduled with the pull-principle is to
close the time gap to the following HSM stage as much as possible in order to avoid
5 Computational Experiments 91

excessive reheating of the casts. The summarized results in table 11 (i.e., reported as an
average over five replications) reveal that this objective was achieved with average and
maximum waiting times on the order of 12 and 30 hours, respectively, for all scenarios.
For six of the eight scenarios, the flexible production windows seem to positively
impact the solution quality for the maximum waiting times. This effect diminishes for
the average waiting times.

x When comparing scenario 1 with scenario 2, we find that a reduction of 21.9%


for the maximum waiting times was achieved. The effect on the average waiting
times is an increase of 5.9%.

x A comparison of scenarios 3 and 4 reveals the maximum positive effect on the


average waiting time. More specifically, a positive effect of 14.8% for the
maximum waiting times and 41.7% for the average waiting times were observed.

x Slightly larger reductions of 17.1% for the maximum waiting times were
achieved in scenario 5 as compared to scenario 6. The effect on the average
waiting times was slight, i.e., a 1% reduction.

x For scenarios 7 and 8, no effect was observed.

Generally, the waiting times were shorter for the scenarios with the small product
portfolio because the CC requires less time for setup and hence can react better to the
HSM plan.

Regarding computational effort, the workload proves to be the main driver. When
the high workload scenarios are compared to their corresponding average workload
scenario, differences in computational time ranging from 38% to 190% were
observed (i.e., in a comparison of scenario 1 with 3, 2 with 4, 5 with 7, and 6 with 8).
Overall, all scenarios were solved quickly with computational times between 2 and 6
seconds.
92 5 Computational Experiments

Table 11: Numerical results for the CC model (pull-principle)


Scenario Maximum Average CPU MIP-
waiting time (h) waiting time (h) time (s) Gap
(%)

1 (200, 90%, f) 31.33 13.29 2.90 0.00

2 (200, 90%, l) 38.17 12.51 5.69 0.16

3 (200, 80%, f) 26.59 9.90 2.11 0.00

4 (200, 80%, l) 30.52 14.03 1.96 0.00

5 (100, 90%, f) 28.30 12.03 4.15 0.00

6 (100, 90%, l) 33.13 12.16 3.44 0.00

7 (100, 80%, f) 28.72 12.89 2.48 0.00

8 (100, 80%, l) 25.31 11.79 2.16 0.17

Overall, the pull-principle produces inferior results as compared to the push-principle


regarding average and maximum waiting times between production steps. Regarding the
computational effort, results can be obtained quickly, i.e., on average in about a minute,
for both planning principles.

5.4 Comparative analysis of the push and pull-principles


In the following sections, the results of both models, each scheduled with a different
production planning principle (i.e., either the push- or pull-principle), are compared as
illustrated in figure 27. The resulting maximum and average waiting times and the
computation times were analyzed for all scenarios and both models.

The next step is deeper and examines the results on the plant level. The makespan of the
CCs and the HSM are individually compared in order to analyze the resulting difference
of push- and pull scheduling. Finally, the number of production stops was analyzed on a
plant level in Section 5.4.3 in order to evaluate whether the resulting production plans
would be applicable in practice.
5 Computational Experiments 93

Level of comparison Analysis

Push and pull model ▪ Maximum and average waiting times


▪ Computational effort

Continuous casters and ▪ Makespan


hot strip mill ▪ Production stops and idle time

Figure 27: List of analyses

5.4.1 Maximum and average waiting times


The most important comparative analysis for both scheduling directions is the
differences in the resulting maximum and average waiting times because they indicate
the feasibility of the model’s results in terms of establishing a hot charging connection.
The maximum waiting times for all scenarios for both scheduling directions are
illustrated in figure 28, while figure 30 displays the average waiting times for all
scenarios for both scheduling directions. As illustrated in figure 28, the average of all
maximum waiting times over all scenarios of the push principle is 22.4 hours, whereas
the equivalent for the pull principle is 30.3 hours. Thus, the average of all maximum
waiting times over all scenarios of the push principle is 7.9 hours or 26% lower than
that of the pull principle. In fact, no scenario of the pull-principle has a maximum
waiting time that is superior to the results of the push principle.
94 5 Computational Experiments

Maximum waiting times


hours

Push-Principle
22.8
Scenario 1 Pull-Principle
31.3
20.8
Scenario 2
38.2
23.3
Scenario 3
26.6
18.5
Scenario 4
30.5
26.4
Scenario 5
28.3
20.9
Scenario 6
33.1
24.9
Scenario 7
28.7
21.9
Scenario 8
25.3

Ø= Ø=
22.4 30.3

Figure 28: Analysis of the maximum waiting times

Looking at the resulting maximum waiting times for each scenario, the results of the
pull-principle seemingly deviate more from the average than the results of the push-
principle. Figure 29 reveals that the maximum deviation from the average of a scenario
scheduled with the pull-principle is 26%, whereas the maximum deviation is 18% for
the scenarios scheduled with the push-principle.
5 Computational Experiments 95

Maximum waiting times


hours (deviation from average, %)
Push-Principle
Pull-Principle

22.8 (2%) Scenario 1 31.3 (4%)

20.8 (-7%) Scenario 2 38.2 (26%)

23.3 (4%) Scenario 3 26.6 (-12%)

18.5 (-17%) Scenario 4 30.5 (1%)

26.4 (18%) Scenario 5 28.3 (-6%)

20.9 (-7%) Scenario 6 33.1 (9%)

24.9 (11%) Scenario 7 28.7 (-5%)

21.9 (3%) Scenario 8 25.3 (-16%)

Ø 22.4 Ø 30.3

Figure 29: Deviation of maximum waiting times

In addition to the analysis of the maximum waiting times, the average waiting times
were calculated in order to present another tangible result of the model tests. The
difference in average waiting times for both models is similar to the difference in
maximum waiting times. As displayed in figure 30, the average of all average waiting
times for the scenarios scheduled in the push-direction is 2.9 hours or 24% less than the
average of all scenarios scheduled in the pull-direction. In contrast to the resulting
maximum waiting times, one scenario with the pull-principle, i.e., scenario 3, is
superior to that of the push-principle.
96 5 Computational Experiments

Average waiting times


hours

Push-Principle
10.2
Scenario 1 Pull-Principle
13.3
7.7
Scenario 2
12.5
10.8
Scenario 3
9.9
7.8
Scenario 4
14.0
10.8
Scenario 5
12.0
7.9
Scenario 6
12.2
11.1
Scenario 7
12.9
9.1
Scenario 8
11.8

Ø= Ø=
9.4 12.3

Figure 30: Analysis of the average waiting times

Overall, the results show that the push-principle produces superior results as compared
to the pull-principle. An explanation is that with the push-principle, the less complex
production step reacts to the more complicated production step (i.e., a product portfolio
of up to 200 steel grades), i.e., the CC. Hence, the HSM can more easily assign caster
blocks to its 3 milling programs as compared to the process of splitting the milling
program for the CC into many steel grades and still fulfilling the minimum batch size
restrictions. To obtain feasible results for the scenarios scheduled with the pull-
principle, slack time was included in the HSM schedule to ensure that the CC has
enough time to pre-produce the supply and compensate for the higher number of setup
operations for the CC. This additional slack time is also reflected in the waiting time of
the orders.
5 Computational Experiments 97

A comparison of the resulting waiting time with the temperature-time diagram of figure
19 in section 4.1 shows that the potential charging temperature of the slabs after about
25 hours is between 300 to 400°C. Hence, a hot charging connection is feasible, and the
potential energy cost reduction is in the area of 20%.

5.4.2 Total production and setup times


To do a thorough investigation of the results, we now take a deeper look and compare
the makespans of the CCs and the HSM individually for both scheduling directions. In
the first step, a comparison of the makespan of the CC for the push and pull principle
was carried out, followed by the same analysis for the HSM.

The makespan was broken up into production time, setup operations, breaks after start
of production, and the resulting makespan. For the analysis, the averages of all scenarios
and for each scenario were reported over five numerical test replications. For the
makespan analysis of the CC, only one caster per scenario was included. This was
usually the caster with the maximum makespan. If both casters had equal makespans,
the caster with the higher production and setup time was included. The makespan
analysis for the CC scheduled with the push-principle is displayed in figure 31.

Makespan analysis of continuous caster scheduled with the push-principle


hours

586.8

45.4
(8%)

71.3
(12%)
470.1
(80%)

Production time Setup time Idle time Makespan


excl. setups
and breaks

Figure 31: Makespan analysis for the continuous caster (push-principle)


98 5 Computational Experiments

The analysis of the makespan of the CC scheduled with the push-principle reveals that
80% of the makespan is spent in production time, and another 12% is necessary for
setups. The remaining 8% of the total makespan is filled with idle time after the start of
production. This is surprising since the model objective is to minimize the makespan.
Upon further investigation, the reason is caused by the data set T(e). As indicated in
Section 5.1, T(e) indicates the periods in which the demand elements are allowed to be
produced to reduce slab inventory. Moreover, if the CC has completed the production of
the demand elements that are allowed to be produced, it must wait until the next 25% of
all demand elements are allowed to be produced. To further investigate this explanation,
we examined the same analysis carried out only for the high workload scenarios (figure
32). Indeed, only 4% of the makespan is spent in idle time after the start of production.
Hence, the breaks were reduced by 50%, but it still exists. This shows that with a higher
workload or a T(e) that allows the production of demand much earlier, the breaks would
reduce to zero.

Makespan analysis of continuous caster for all high workload scenarios


hours

594.4
20.9
(4%)

77.0
(13%)
496.5
(84%)

Production time Setup time Idle time Makespan


excl. setups
and breaks

Figure 32: Makespan analysis of the CC for high workload scenarios (push-principle)

After investigating the makespan analysis for the CC scheduled with the push-principle,
we also analyzed the makespan of the CC scheduled with the pull-principle.
5 Computational Experiments 99

Makespan analysis of continuous caster scheduled with the pull-principle


hours

610.3

65.4
(11%)

76.2
(12%)
468.7
(77%)

Production time Setup time Idle time Makespan


excl. Setups
and breaks1

Figure 33: Makespan analysis for the continuous caster (pull-principle)

As shown by figure 33, the share of the production time equals 77%; moreover, the
setup time amounts to 12%, and idle time after the start of production amounts to 11%.
As illustrated in figure 34, comparing the pull and push-principles shows that the total
makespan of the pull-principle is 23.5 hours, i.e., about 4% higher than that of the push-
principle. The relative share of production time is lower for the pull-principle, i.e., 77%
as compared to 80%, and the share of setup time is similar to that of the push-principle
at 12%. Most notably, the share of idle time is 3% higher as compared to that of the
push-principle.
100 5 Computational Experiments

Summary of the makespan analyses of the CC stage


%

100% = 586.8 610.3

Production time 77
80
excl. Setup and idle time

Setups 12
12
Idle time 8 11

CC - Push- CC - Pull-
principle principle

Figure 34: Summary of the makespan analyses of the CC stage

The makespan analyses for the CC show that the share of idle time leads to a higher
makespan and a less efficient use of equipment if the caster is scheduled with the pull-
principle. Although the share of idle time decreases for the high workload scenarios
when the CC is scheduled with the pull principle, i.e., from 11% to 9%,6 the values
remain inferior to the makespan result of the push-principle.

The differences among the planning directions are caused by two main factors. On the
one hand, the CC model, which is scheduled with the pull-principle, does not optimize
the makespan and instead optimizes the waiting time. Hence, it produces later, including
pauses, to finish in a timely fashion for the HSM. Yet, the CC needs more overall slack
time to be able to react to the HSM. If the HSM is scheduled first, it sets a time grid
during which the caster needs to pre-produce the supply for the HSM. The feasibility of
the time grid relies to a great extent on the mix of a production block at the HSM. Since
the steel grades are allocated to the CC based on the caster’s product speed ratio,
difficulties arise as soon as a production block occurs at the HSM that causes

6 If we investigate only the high workload scenarios as in the analysis in Figure 22, the share of production time is
78%, while the share of setup time and the share of breaks after the start of production is 13% and 9%,
respectively. The makespan adds up to 634.2 hours.
5 Computational Experiments 101

disproportionate loads for one of the CCs. Thus, the CC with the higher ratio of the
production must use more time than the HSM to produce its share of the orders. To
produce feasible solutions even when this ratio is not perfectly met, more slack time is
needed for the CCs and must be implemented in the design of the production windows.

After taking a look at the continuous casting stage, the same analyses for the HSM were
carried out. Similar to that of the CC, the makespan analysis for the HSM was broken
down into production time, setup operations, postponed start of production, breaks and
idle time after the start of production, and the resulting makespan. For the analysis, the
averages for all scenarios were reported for five replications in each scenario. First, the
makespan analysis of the HSM scheduled with the push-principle was considered. As
illustrated in figure 35, the makespan of the HSM scheduled with the push-principle
added up to 627.5 hours. The production time accounted for 82%, whereas 4%, 5%, and
9% were spent on setup, postponed start of production, and 9 breaks and idle time after
the start of production, respectively. As compared to the continuous casting stage, the
makespan was of course higher, i.e., 40.5 hours, and the relative time spent on setup
was much less due to the significantly less complex product portfolio (i.e., 3 milling
programs) instead of up to 200 steel grades for the CC.

Makespan analysis of the hot strip mill scheduled with the push-principle
hours

627.5

55.4
(9%)

33.2
(5%)

514.7 24.2
(82%) (4%)

Production time Setup Postponed Breaks and Makespan


excl. Setups time start of idle time after
and breaks1 production start of
production

Figure 35: Makespan analysis of the hot strip mill (push-principle)


102 5 Computational Experiments

The postponed start of production was necessary because the HSM waits for the first
production block from the CC to start production. Breaks and idle time after the start of
production were caused by the objective function that is designed to reduce the waiting
time instead of the makespan. Hence, the model accepts idle time if the production of a
specific block at the HSM is faster than the production of the upcoming block at the CC.
If we compare the share of breaks and idle time after production for the HSM scheduled
with the push-principle to the share of breaks for the CC scheduled with the pull-
principle, we find that the HSM spends relatively less time on pauses, i.e., 9% and 11%
for breaks and idle time, respectively. This is relevant because it shows again that the
HSM seems to handle the role as the reacting facility better than the CC.

The makespan analysis of the HSM scheduled with the pull-principle is illustrated in
figure 36. It shows that the makespan amounts to 635.8 hours, i.e., 8.3 hours or 1.3%
higher than that scheduled with the push-principle. The makespan is split up into the
production time (81%), setup time (2%), breaks before the start of production (8%),
and breaks after the start of production (10%).

Makespan analysis of the hot strip mill scheduled with the pull-principle
hours

635.8

61.6
(10%)

50.0
(8%)
514.2
(81%) 10.0
(2%)

Production time Setup Postponed Breaks and Makespan


excl. setups time start of idle time after
and breaks production start of
production

Figure 36: Makespan analysis of the hot strip mill (pull-principle)


5 Computational Experiments 103

Initially, the makespan of the HSM scheduled with the pull-principle is surprisingly
higher than that scheduled with the push-principle since the objective of the HSM
model with the pull-principle is to minimize the makespan. Nevertheless, this is due to
the design of the production windows. To allow the CC enough time to pre-produce the
slab supply for the HSM at the start of a production week (t=8, 15, 21), a break of 20
hours is included in the production window setup, thereby reaching a total of 60 hours.
The start of production is also postponed by 50 hours. Hence, only 1.6 hours of idle
time after the start of production cannot be explained by the design of the production
windows.

These remaining 1.6 hours are caused by the data set T(e), i.e., similar to the idle time in
the makespan analysis of the CC as explained previously in this section. If the
makespan of the HSM scheduled with the pull-principle is compared to its equivalent
scheduled with the push principle as illustrated in figure 37, the relative amount of setup
time is noticeably reduced by 50%. A setup time reduction seems likely since the first
scheduled production step allows the HSM to optimize its batching of orders. Regarding
the pauses, the values of the push principle are superior, but the values are influences by
the design of the production windows in which pauses of 60 hours after the start of
production are included.

Summary of the makespan analyses of the HSM stage


%

100% = 627.5 635.8

Production time
82 81
excl. Setup and idle time

Setups
Postponed start of production
Breaks and idle time after 4 2
5 8
start of production 9 10

HSM - Push- HSM - Pull-


principle principle

Figure 37: Summary of makespan analyses of the HSM stage


104 5 Computational Experiments

Overall, the values of the push scheduling for the HSM are obviously superior to the
values of the pull scheduling. The main reason is clearly the required production
window design that was necessary to obtain feasible solutions. Clearly, a production
window design that includes slack time to balance production speed and setup time
differences always includes a little buffer slack time in order to ensure feasible solutions
for all scenarios.

The analyses of the makespans of both production steps and both scheduling directions
clearly showed that the overall equipment utilization when both production steps are
scheduled in the push-direction is superior to that of the productions steeps scheduled in
the pull-direction. The makespan of the continuous casting stage is 4% lower if it is
scheduled with the push-principle, and the makespan of the HSM stage is 1.3% lower if
it is scheduled with the push-principle. These differences could be even greater if we
consider the continuous casting stage where about 45 hours were spent in idle time
instead of production; moreover, this could be reduced with a higher workload or am
amended T(e). This clearly shows that both models scheduled in the push-direction are
able to significantly increase their production, especially in high demand situations,
whereas the models scheduled in the pull-direction do not allow a higher production
quantity.

5.4.3 Analysis of production stops


In addition to the duration of production stops after the start of production, an analysis
of the number of production stops after the start of production was conducted. The
number of productions stops was calculated by considering the average number of
production stops for each replication and scenario. This analysis was included to
consider equipment specific characteristics as indicated in Section 3.1.1. This section
highlighted the fact that the equipment for steel production does not include a lot of
flexibility regarding production speed changes and short production stops. In particular,
the CC is known for its inflexibility because it cannot handle speed changes and
production stops and requires a longer setup operation afterwards. In contrast to the CC,
the HSM is much more flexible and has the ability to change production speeds and
handle short stops. Hence, the analysis illustrated in figure 38 is especially relevant for
the CC. Indeed, if the push-principle is applied, the number of stops is decreased as
5 Computational Experiments 105

compared to that of the pull-principle. With the push-principle, each CC and HSM has
on average 6.8 stops and 6.3 stops, respectively. On the other hand, if the pull-principle
is used, the number of stops for each CC and for the HSM is 7.7 stops and 8.4 stops,
respectively. This is an increase of 14% and 32% for the CC and the HSM, respectively.

Average number of production stops of each scenario


8.4

7.7

+14% +32%
6.8

6.3

Push- Pull- Push- Pull-


principle principle principle principle

Continuous Caster Hot Strip Mill

Figure 38: Analysis of production stops

5.4.4 Computational effort


The computation time needed to obtain results that are optimal or near optimal is
another relevant criterion for the evaluation of both scheduling directions. Figure 39
displays the sum of calculation times for the continuous casting stage and the HSM for
both the pull- and push-principle. Additionally, the relative deviation from average is
given in the bubbles next to the sum of the calculation time. A comparison of the
average computational effort over all scenarios reveals that the push-principle is in total
about 46% faster, i.e., an average of 43.6 seconds, than the pull-principle at 64 seconds.
The next observation is that the continuous casting stage in both cases was computed
much faster than the hot strip milling stage, i.e., up to 6 versus up to 188 seconds,
respectively.

Another interesting aspect is the deviation from the average computation time.
Regarding the pull-principle, the values deviate much less with a maximum deviation of
-56% to 86% from the average of 64 seconds. In contrast to the pull-principle
106 5 Computational Experiments

deviations, the computation time for the push-principle deviates a lot more. The
maximum deviations range from -84% to 332% from the average of 43.6 seconds. The
influence of the outliers also becomes clear if we look at the median instead of the
average. The median of all scenarios for the pull-principle is 49.7 seconds, which
signifies a reduction of 14.3 seconds from the average. Whereas the median of all
scenarios of the push-principle equals 15 seconds, which results in a difference between
the average and the median of 28.6 seconds.

Hot Strip Mill


Continuous Caster
Computation times of both production steps for all scenarios
Seconds, in % xx Deviation from average in %

Pull-Principle Push-Principle

Scenario 1 3 39 42 -34 0 14 14 -67

Scenario 2 6 35 41 -36 7 7 -84


0
Scenario 3 2 63 66 2 0 188 188 332

Scenario 4 2 26 28 -56 7 7 -83


0
Scenario 5 4 39 44 -32 0 28 28 -36

Scenario 6 3 113 117 82 0 9 10 -78

Scenario 7 2 117 119 86 0 78 78 80

Scenario 8 2 54 56 -13 0 15 16 -64

Ø 64.0 Ø 43.6

Figure 39: Analysis of computation times

Overall, a result close to optimum is achieved in usually about a minute and does not
take longer than 3.5 minutes, which shows that both models have a high applicability in
industrial environments in terms of their computation time.
6 Conclusion 107

6 Conclusion

Most steel companies operate integrated steel mills that process the steel from the raw
material to the hot rolled coil. This production setup allows a more efficient production
due to a better connection between all of the production steps if properly managed as
compared to steel production facilities that only cover the MS to the continuous casting
stage or from the HSM and further processing stages. However, most integrated steel
mills do not fully exploit the cost savings opportunities offered by improved connection
between stages, especially between the continuous casting and hot rolling stages.
Following are some of the reasons for this lack of operational excellence:

x Significant differences in the planning characteristics of CCs and HSMs, which


lead to an increase in complexity if the planning is integrated;

x Higher risk of production downtime if both production steps are coupled and only
a low level of inventory exists; and

x Insufficient decision support tools that support integrated planning approaches.

The insufficient exploitation of energy cost savings is especially critical based on the
current situation in the steel industry in Europe, especially in Germany; indeed, the
German steel industry finds itself increasingly under cost pressure from both the
demand and supply side. Cost pressure from the supply side comes particularly from the
increasing raw material costs and rising energy costs in Germany. Likewise, cost
pressure from the demand comes from the customer’s demands for increasingly
innovative, yet cheaper products (e.g., the automotive OEMs demanding lower steel
prices). Furthermore, certain important customer groups, i.e., the automotive and
machinery industries, demand improved service levels and high quality products, which
also calls for tools that support production planning and scheduling decisions.

The objective of this work has been to develop optimization models to reduce the
waiting time for slabs between the continuous casting and the hot strip milling stages,
thereby reducing the energy cost associated with the operation of the HSM furnaces. To
ensure the practical applicability of the developed optimization tool and the relevance of
the optimization objective, both were developed in cooperation with a large European

I. Mattik, Integrated Scheduling of Continuous Casters and Hot Strip Mills, Produktion und Logistik,
DOI 10.1007/978-3-658-03775-8_6, © Springer Fachmedien Wiesbaden 2014
108 6 Conclusion

steel company that operates a large integrated steel mill in Germany among other
locations.

To lay the foundation, the short term production planning and scheduling were
classified in the overarching concept of SCM and in the production planning hierarchy.
Furthermore, the role of APS for short term production planning was discussed in
Section 2.1. Section 2.2 then examined the process, characteristics, and basic principles
of production planning. Furthermore, an introduction on the development of quantitative
models for lot sizing and scheduling was given. Additionally, the specific characteristics
of steel production, including an introduction into current industry challenges, were
described in Section 2.3.

Building on that, Section 3.1 focused on the scheduling challenges of the steel industry
by reviewing the associated equipment, product portfolio, and demand specific
challenges and presented recommendations on requirements for a quantitative
scheduling optimization for the steel industry. A classification of batching and
scheduling literature with a steel industry focus and a review of individual publications
revealed that most models focused only on either the liquid or the solid stages of
production or focused on the planning results at the master production schedule level,
which prohibits any proposed energy cost reductions. Furthermore, the advantages of an
integrated approach to batching and scheduling as well as the continuous representation
of time were reviewed in Sections 3.3 and 3.4. Lastly, the applied block planning
methodology used in this research was discussed, and an evaluation of the number of
variables and constraints in a block planning model was compared to that of a discrete
time based model in Section 3.5.

Considering the deficiencies of current optimization models revealed in Section 3.2,


mixed-integer programming models were developed in Section 4 as the main
contribution of this work. This supports the scheduling decisions of CCs and HSMs
using two separate optimization models that consider the relevant planning
characteristics and focus on reducing the waiting time of semi-finished products, slabs,
while still allowing hot charging to a degree. This leads to significant energy costs
savings. Major features of the models include the following:
6 Conclusion 109

x The deconstruction approach with a separate but interdependent MILP model


formulations for the CCs and the HSMs,

x Pre-defined production sequences for the CC model that include the chemical
compatibility of steel grades to minimize major set ups and keep the number of
variables at an acceptable level,

x Pre-defined production sequences at the HSM that balance the temperature


requirements of product groups to minimize steep temperature changes to allow
for a more efficient furnace operation, and

x Two decomposition approaches (i.e., one based on the push principle and the
other on the pull principle) permit a comparative evaluation of results.

The numerical design reflects the real complexity of customer orders and the product
portfolio of a large steel company (i.e., the product portfolio contains more than 100
steel grades and numerous dimensions). Furthermore, all other production parameters
are based on the parameter values in the industry. Section 5.2 presented the results of
numerical tests and demonstrated that the two MILP model formulations with both
deconstruction approaches can be solved in a short amount of time using the standard
optimization software ILOG CPLEX 12.1.

Finally, in Section 5.4, the comparative analyses between the results of the model
formulations based on the push principle and the pull principle were presented. The
main findings of the comparative analyses are as follows:

x The results for the model formulation that is based on the push-principle were
clearly superior regarding the maximum and average waiting times. More
specifically, the maximum waiting time was on average 26% lower and the
minimum waiting time was on average 24% lower than the equivalent pull-model
formulation.

x The makespan analyses show that the production schedules from the push-model
formulation are more efficient than the pull-model formulations. The makespans
of the CC and the HSM are shorter with the push-model formulation, which gives
more flexibility for production quantity increases.
110 6 Conclusion

x The push-model formulations have a higher practical applicability due to fewer


production stops in the schedule. Production stops are triggered by the models’
objective to minimize waiting time; hence, the models stop production rather than
waiting for the best start time.

x Both model formulations are computationally efficient. The push-models arrive at


a solution in 44 seconds on average in the numerical performance tests. In
contrast, the pull-models arrive at a solution in 64 seconds on average. Both
computation times demonstrate high practical applicability.

Overall, the push model formulation demonstrated clear superiority in all analyses for
the objective, the production efficiency, the practical applicability, and the required
computation time. The results for the push-model show waiting times of less than 25
hours and the possible average slab charge temperature of between 300° and 400°C,
which leads to energy cost savings of about 20% per slab. Section 3.2 shows that only
Xu et al. (2012) and Cowling and Rezig (2000) focus on an integrated batching and
scheduling approach with an energy saving aim for CCs and HSMs. The approach of
Xu et al. (2012) lacked practical applicability since it only solved small test instances
within a reasonable amount of time due to a very complex model formulation. However,
they discuss how their concept can be implemented in practice. Cowling and Rezig
(2000) focused on better CC and HSM integration mainly with the intent to increase the
size of production runs and save on setup costs and wasted energy during setup.
However, they allow the HSM scheduling to exert a greater influence on the CC
planning to improve HSM production runs without evaluating the potential lower
performance of the CC as demonstrated in this dissertation.
Appendix 111

Appendix

Appendix 1: Discrete time based model formulation

A discrete time based model formulation for continuous caster scheduling is used to
evaluate the potential number of variables of a discrete model formulation as compared
to the block planning approach. The notation is as follows.

Indices and sets


tϵT Time period
nϵN Number of continuous casters
mϵM Number of specific steel grades
Parameters
c
cs mn Capacity consumption of steel grade m on CC n (hours/ton)
c
stlm Setup time for changing over from steel grade l to m (hours)
c
dmmt Demand of steel grade m in period t (tons)
c
mln Minimum size of a cast on CC n (tons)
C ntc Capacity of CC n in period t (hours)
Decision variables
I mtc Inventory of steel grade m at the end of period t
c
Z mnt Quantity of steel grade m produced in period t at CC n (tons)
c
xlmnt 1 If a changeover from steel grade l to m takes place in period t on CC n (0
otherwise)
c
y mnt 1 If the production system is set up for product m in period t on CC n (0
otherwise)
Wc Makespan
The model formulation for the CC is given as follows:

Min W c (93)

W c t t ˜ y mnt
c
mϵ M, n ϵ N, t ϵ T (94)

1  ¦ Z mnt  d mt
c c c c
I mt I mt m ϵ M; t ϵ T; Ij0=0 (95)
nN

I. Mattik, Integrated Scheduling of Continuous Casters and Hot Strip Mills, Produktion und Logistik,
DOI 10.1007/978-3-658-03775-8, © Springer Fachmedien Wiesbaden 2014
112 Appendix

¦Z
mM
c
mnt ˜ cs mn
n
¦ ¦ st
lL mM |l z m
c
lm ˜ xlmnt
c
d C ntc n ϵ N, t ϵ T (96)

c
cs mn ˜ Z mnt
c
d Cntc ˜ y mnt
c
mϵ M, n ϵ N, t ϵ T (97)

¦y
mM
c
mnt 1 n ϵ N; t ϵ T (98)

c
xlmnt t ylnc t 1  y mnt
c
1 l  M ; m  M | l z m; n  N ; t  T (99)

c
Z mnt t mlnc ˜ y mnt
c
mϵ M, n ϵ N, t ϵ T (100)

Finally, the variable domains are defined as follows:


c
xlmnt  ^0;1` l  M ; m  M | l z m; n  N ; t  T (101)

c
y mnt  ^0;1` m  M ; n  N; t T (102)

c
Z mnt t0 m  M ; n  N; t T (103)

c
I mt t0 m  M ;t T (104)

Wc t0 (105)
References 113

References
Aggarwal, A. and Park, J. K. (1993) 'Improved algorithms for economic lot size
problems.' Operations Research, Vol. 41 No. 3, pp. 549-571.
As’ad, R. and Demirli, K. (2010a) 'A bilinear programming model and a modified
branch-and-bound algorithm for production planning in steel rolling mills with
substitutable demand.' International Journal of Production Research, Vol. 49 No.
12, pp. 3731-3749.
As’ad, R. and Demirli, K. (2010b) 'Production scheduling in steel rolling mills with
demand substitution: Rolling horizon implementation and approximations.'
International Journal of Production Economics, Vol. 12 No. 2, pp. 361-369.
Assaf, I., Chen, M. and Katzberg, J. (1997) 'Steel production schedule generation.'
International Journal of Production Research, Vol. 35 No. 2, pp. 467-477.
Bargel, H.-J. and Schulze, G. (2010) Werkstoffkunde. 10th edn. Heidelberg: Springer.
BCG (2009) 'Sustainable Steelmaking.' The Boston Consulting Group.
Bellabdaoui, A. and Teghem, J. (2006) 'A mixed-integer linear programming model for
the continuous casting planning.' International Journal of Production Economics,
Vol. 104, pp. 260-270.
Bilgen, B. and Günther, H. O. (2009) 'Integrated production and distribution planning in
the fast moving consumer goods industry: a block planning application.' OR
Spectrum.
BMWi (2012) 'Industry prices for energy.' [Data File]. Federal Ministry of Economics
and Technology. Retrieved from: http://www.bmwi.de/BMWi/Navigation/
Energie/Statistik-und-Prognosen/Energiedaten/energiepreise-energiekosten.html.
Booz and Co (2012) 'Die Stahlindustrie in Deutschland „Rückgrat des Industrie-
standorts Deutschland“'. Booz and Company.
Box, R. E. and Herbe Jr, D. G. (1988) 'A Scheduling Model for LTV Steel's Cleveland
Works' Twin Strand Continuous Slab Caster.' Interfaces, Vol. 18 No. 1, pp. 42-56.
Busch, H. (2008) 'Stochastische Hill-Climbing-Algorithmen zur Produktionssteuerung
von Grobblechwalzstraßen.' Automatisierungstechnik, Vol. 56 No. 4, pp. 197-207.
Chandra, C. and Grabis, J. (2007) Supply Chain Configuration - Concepts, Solutions,
and Application. New York: Springer.
Chang, S. Y., Chang, M.-R. and Hong, Y. (2000) 'A lot grouping algorithm for a
continuous slab caster in an integrated steel mill.' Production Planning & Control,
Vol. 11 No. 4, pp. 363-368.
Chen, A. L., Yang, G. K. and Wu, Z. M. (2008) 'Production scheduling optimization
algorithm for the hot rolling processes.' International Journal of Production
Research, Vol. 46 No. 7, pp. 1955-1973.
114 References

Chopra, S. and Meindl, P. (2007) Supply Chain Management - Strategy, Planning &
Operations. 3rd edn. New Jersey: Pearson Prentice Hall.
Christopher, M. (2005) Logistics and Supply Chain Management. 3rd edn. Harlow:
Prentice Hall.
Coelho, J. (2010) 'Prompt EUAs to spike on tighter ETS cap: analysts', Point Carbon
News, Vol. 82.
Cowling, P. (2003) 'A flexible decision support system for steel hot rolling mill
scheduling.' Computers & Industrial Engineering, Vol. 45, pp. 307-321.
Cowling, P., Ouelhadj, D. and Petrovic, S. (2004) 'Dynamic scheduling of steel casting
and milling using multi-agents.' Production Planning & Control, Vol. 15 No.2, pp.
178-188.
Cowling, P. and Rezig, W. (2000) 'Integration of continuous caster and hot strip mill
planning for steel production.' Journal of Scheduling, Vol. 3, pp. 185-208.
Dawande, M., Kalagnanam, J., Ho Soo, L., Reddy, C., Siegel, S. and Trumbo, M.
(2004) 'The Slab-Design Problem in the Steel Industry.' Interfaces, Vol. 34 No. 3,
pp. 215-225.
Deutscher Bundestag (2011) 'Bundestag beschließt Atomausstieg und Energiewende.'
Retrieved from: http://www.bundestag.de/dokumente/textarchiv/2011/34938007
_kw26_de_energiewende/index.html.
DEW (2012) 'Company Profile.' Deutsche Edelstahlwerke. Retrieved from: www.dew-
stahl.com/fileadmin/files/dew-stahl.../company-profile.ppt.
Dickersbach, J. (2006) Supply Chain Management with APO. 2nd edn. Heidelberg:
Springer.
Dohn, A. and Clausen, J. (2010) 'Optimising the Slab Yard Planning and Crane
Scheduling Problem using a two-stage heuristic.' International Journal of
Production Research, Vol. 48 No. 15, pp. 4585-4608.
Domschke, W. and Drexl, A. (2005) Einführung in Operations Research. 6th edn.
Heidelberg: Springer.
Dong, H., Huang, M., Ip, W. H. and Wang, X. (2010) 'On the integrated charge
planning with flexible jobs in primary steelmaking processes.' International
Journal of Production Research, Vol. 48 No. 21, pp. 6499-6535.
Dong, H., Huang, M., Ip, W. H. and Wang, X. (2012) 'Improved variable
neighbourhood search for integrated tundish planning in primary steelmaking
processes.' International Journal of Production Research, Vol. 50 No. 20, pp.
5747-5761.
Drexl, A. and Kimms, A. (1997) 'Lot sizing and scheduling -- Survey and extensions.'
European Journal of Operational Research, Vol. 99 No. 2, pp. 221-235.
References 115

Dyckhoff, H. and Spengler, T. (2010) Produktionswirtschaft. 3rd edn. Heidelberg:


Springer.
EU (2012) 'Emissions Trading System (EU ETS).'Retrieved from: http://ec.europa.eu
/clima/policies/ets/index_en.htm.
European Commission (2012a) 'Benchmarks for free allocation.' Retrieved from:
http://ec.europa.eu/clima/policies/ets/benchmarking/index_en.htm.
European Commission (2012b) 'Preparing for the 3rd trading period.' Retrieved from:
http://ec.europa.eu/clima/policies/ets/auctioning/third/index_en.htm.
Fleischmann, B. (1990) 'The discrete lot-sizing and scheduling problem.' European
Journal of Operational Research, Vol. 44 No. 3, pp. 337-348.
Fleischmann, B. (2012) 'The Frutado Case.' In H. Stadtler, B. Fleischmann, M. Grunow,
H. Meyr and C. Sürie (Eds.) Advanced Planning in Supply Chains: pp. 11-18.
Heidelberg: Springer.
Gackstatter, U., Gönner, T. and Kurth, H. (2011) 'Business Excellence in Produktion,
Logistik und Servicebereichen.' In W. Wenger, M. J. Geiger and A. Kleine (Eds.)
Business Excellence in Produktion und Logistik: pp. 203-226. Wiesbaden: Gabler.
Global-Insights (2012) 'Total sales of steel industry.' IHS Global Insights (Ed.).
Govil, M. and Proth, J.-M. (2002) Supply Chain Design and Management. San Diego:
Academic Press.
Grunow, M., Günther, H.-O. and Lehmann, M. (2002) 'Campaign planning for multi-
stage batch processes in the chemical industry.' OR Spectrum, Vol. 24, pp. 281-
314.
Günther, H. O. (2005) 'Supply Chain Management und Advanced Planning Systems: A
Tutorial.' In H. O. Günther, D. C. Mattfeld and L. Suhl (Eds.) Supply Chain
Management und Logistik: pp. 3-41. Heidelberg: Physica-Verlag.
Günther, H. O., Grunow, M. and Neuhaus, U. (2006) 'Realizing block planning
concepts in make-and-pack production using MILP modelling and SAP APO ©.'
International Journal of Production Research, Vol. 44 No. 18/19, pp. 3711-3726.
Günther, H. O., Mattfeld, D. C. and Suhl, L. (2005) 'Supply Chain Management und
Logistik - Optimierung, Simulation, Decision Support.' Physica Verlag,
Heidelberg.
Günther, H. O. and Tempelmeier, H. (2012) Produktion und Logistik. 9th edn. Berlin:
Springer.
Harjunkoski, I. and Grossmann, I. E. (2001) 'A decomposition approach for the
scheduling of a steel plant production.' Computers & Chemical Engineering, Vol.
25, pp. 1647-1660.
116 References

Hax, A. C. and Meal, H. C. (1975) 'Hierarchical integration of production planning and


scheduling.' Logistics: TIMS Studies in Management Sciences: pp. 53-69.
Amsterdam: North-Holland.
Huegler, P. A. and Vasko, F. J. (2007) 'Metaheuristics for meltshop scheduling in the
steel industry.' Journal of the Operational Research Society, Vol. 58 No. 6, pp.
791-796.
IEA (2007) 'Tracking Industrial Energy Efficiency and CO2 Emissions - Executive
Summary.' International. Eenergy Agency (Ed.).
Ierapetritou, M. G. and Floudas, C. A. (1998) 'Effective Continuous-Time Formulation
for Short-Term Scheduling. 1. Multipurpose Batch Processes.' Industrial &
Engineering Chemistry Research, Vol. 37 No. 11, pp. 4341-4359.
Jacob, F. and Strube, G. (2008) 'Why Go Global? - The Multinational Imperative' In E.
Abele, T. Meyer, U. Näher, G. Strube and R. Sykes (Eds.) Global Production.
Heidelberg: Springer.
Jans, R. and Degraeve, Z. (2008) 'Modeling industrial lot sizing problems: A review.'
International Journal of Production Research, Vol. 46, pp. 1619-1643.
JDA (2012) 'Improving On-Time Delivery at Deutsche Edelstahlwerke.'Retrieved from:
http://www.jda.com/company/display-collateral/pID/2008.
Jordan, C. and Drexl, A. (1998) 'Discrete lotsizing and scheduling by batch sequencing.'
Management Science, Vol. 44 No. 5, pp. 698-713.
Kangbok, L., Chang, S. Y. and Yushin, H. (2004) 'Continuous slab caster scheduling
and interval graphs.' Production Planning & Control, Vol. 15 No. 5, pp. 495-501.
Karrer, C., Alicke, K. and Günther, H. O. (2012) 'A framework to engineer production
control strategies and its application in electronics manufacturing.' International
Journal of Production Research, Vol. 50 No. 22, pp. 6595-6611.
King, J. F. (2010) 'Engineering Steel (SBQ) Industry Report.' James F. King
Steelonthenet.com.
King, J. F. (2012) 'World Steel consumption 2011.' James F. King Steelonthenet.com.
Kletti, J. (2007) Konzeption und Einführung von MES-Systemen. Heidelberg: Springer.
Kletti, J. and Schumacher, J. (2011) Die perfekte Produktion. Heidelberg: Springer.
Knolmayer, G., Mertens, P., Zeier, A. and Dickersbach, J. (2009) Supply Chain
Management based on SAP Systems. Heidelberg: Springer.
Kosiba, E. D., Wright, J. R. and Cobbs, A. E. (1992) 'Discrete event sequencing as a
Traveling Salesman Problem.' Computers in Industry, Vol. 19 No. 3, pp. 317-327.
Kumar, V., Kumar, S., Tiwari, M. K. and Chan, F. T. S. (2006) 'Auction-based
approach to resolve the scheduling problem in the steel making process.'
International Journal of Production Research, Vol. 44 No. 8, pp. 1503-1522.
References 117

Li, J., Li, L., Tang, L. and Wu, H. (2006) 'A case of rule-based heuristics for scheduling
hot rolling seamless steel tube production.' Expert Systems, Vol. 23 No. 3, pp.
145-158.
Liker, J. K. and Meier, D. (2006) The Toyota Way Fieldbook. New York: McGraw-
Hill.
Lin, C.-W. and Moodie, C. L. (1989) 'Hierarchical production planning for a modern
steel manufacturing system.' International Journal of Production Research, Vol. 27
No. 4, pp. 613-628.
Liu, S., Tang, J. and Song, J. (2006) 'Order-planning model and algorithm for
manufacturing steel sheets.' International Journal of Production Economics, Vol.
100 No. 1, pp. 30-43.
Lödding, H. (2008) Verfahren der Fertigungssteuerung. Heidelberg: Springer.
Lopez, L., Carter, M. W. and Gendreau, M. (1998) 'The hot strip mill production
scheduling problem: A tabu search approach.' European Journal of Operational
Research, Vol. 160 No. 2/3, pp. 317-335.
Lütke Entrup, M., Günther, H. O., Van Beek, P., Grunow, M. and Seiler, T. (2005)
'Mixed-Integer Linear Programming approaches to shelf-life-integrated planning
and scheduling in yoghurt production.' International Journal of Production
Research, Vol. 43 No. 23, pp. 5071-5100.
Marchetti, P. A. and Cerdá, J. (2009) 'An approximate mathematical framework for
resource-constrained multistage batch scheduling.' Chemical Engineering Science,
Vol. 64 No. 11, pp. 2733-2748.
McKinsey (2012) 'Die Energiewende in Deutschland - Zusammenfassung.'
McKinsey&Company, Inc.
Melouk, S. H., Freeman, N. K., Miller, D. and Dunning, M. (2013) 'Simulation
optimization-based decision support tool for steel manufacturing.' International
Journal of Production Economics, Vol. 141 No. 1, pp. 269-276.
Méndez, C. A., Cerdá, J., Grossmann, I. E., Harjunkoski, I. and Fahl, M. (2006) 'State-
of-the-art review of optimization methods for short-term scheduling of batch
processes.' Computers and Chemical Engineering, Vol. 30 No. 6–7, pp. 913-946.
MEPS (2012) 'Price trend of hot rolled coil in different countries.' MEPS international
Ltd. (Ed.).
Missbauer, H., Hauber, W. and Stadler, W. (2009) 'A scheduling system for the
steelmaking-continuous casting process. A case study from the steel-making
industry.' International Journal of Production Research, Vol. 47 No. 15, pp. 4147-
4172.
Mohanty, R. P. and Singh, R. (1992) 'A Hierarchical Production Planning Approach for
a Steel Manufacturing System.' International Journal of Production and Operation
Management, Vol. 12 No. 5, pp. 69-78.
118 References

Mouret, S., Grossmann, I. E. and Pestiaux, P. (2011) 'Time representations and


mathematical models for process scheduling problems.' Computers and Chemical
Engineering, Vol. 35 No. 6, pp. 1038-1063.
Müller, E., Engelmann, J., Löffler, T. and Strauch, J. (2009) Energieeffiziente Fabriken
planen und betreiben. Springer, Heidelberg.
Münch, A. (2012) 'Stahlherstellung.' Retrieved from: http://www.think-ing.de/
specials/stahl-special/stahlherstellung.
Nickel, S., Velten, S. and Weimerskirch, G. (2005) 'Strategische Supply-Chain
Entscheidungen in der Stahlindustrie - Eine Fallstudie.' In H. O. Günther, D. C.
Mattfeld and L. Suhl (Eds.) Supply Chain Management und Logistik -
Optimierung, Simulation, Decision Support: pp. 163-177. Heidelberg: Physica
Verlag.
OICA (2012) '2011 production statistics by manufacturer and by country.' International
Organization of Motor Vehicle Manufacturer (Ed.). Retrieved from:
http://oica.net/category/production-statistics/.
Park, H., Hong, Y. and Chang, S. Y. (2002) 'An efficient scheduling algorithm for the
hot coil making in the steel mini-mill.' Production Planning & Control, Vol. 13
No. 3, pp. 298-306.
Petersen, C. M., Soerensen, K. L. and Vidal, R. V. V. (1992) 'Inter-process
Synchronization in steel production.' International Journal of Production Research,
Vol. 30 No. 6, pp. 1415-1425.
Rentz, O., Jochum, R. and Schultmann, F. (1999) 'Report on Best Available Techniques
(BAT) in the German Ferrous Metals Processing Industry.' Karlsruhe: Deutsch-
Französisches Institut für Umweltforschung (DFIU).
Rieseberg, S. and Wöhrlen, C. (2012) 'Befreiung der energieintensiven Industrie in
Deutschland von Energieabgaben.' Rosa-Luxemburg Stiftung.
RWI (2011) 'Die volkswirtschaftliche Bedeutung einer Grundstoffindustrie am Beispiel
der Stahlindustrie.' Rheinisch-Westfälisches Institut für Wirtschaftsforschung.
Sato, S., Yamaoka, T., Aoki, Y. and Ueda, T. (1977) 'Development of integrated
production scheduling system for iron and steel works.' International Journal of
Production Research, Vol. 15 No. 6, pp. 539-552.
SBB (2012) 'Raw material prices.' Steel business briefing (Ed.).
Schneeweiß, C. (1995) 'Hierarchical structures in organisations: A conceptual
framework.' European Journal of Operational Research, Vol. 86 No. 1, pp. 4-31.
Schuh, G. (2006) Produktionsplanung und -steuerung. 3rd edn. Heidelberg: Springer.
Singh, K. A., Srinivas and Tiwari, M. K. (2004) 'Modelling the slab stack shuffling
problem in developing steel rolling schedules and its solution using improved
Parallel Genetic Algorithms.' International Journal of Production Economics, Vol.
91 No. 2, pp. 135-147.
References 119

Slack, N., Chambers, S. and Johnston, R. (2010) Operations Management. 6th edn.
Harlow: Prentice Hall.
Stadtler, H. (2005) 'Supply chain management and advanced planning--basics, overview
and challenges.' European Journal of Operational Research, Vol. 163 No. 3, pp.
575-588.
Stadtler, H. and Fleischmann, B. (2012) 'Hierarchical Planning and the Supply Chain
Planning Matrix.' In H. Stadtler, B. Fleischmann, M. Grunow, H. Meyr and C.
Sürie (Eds.) Advanced Planning in Supply Chains: pp. 21-34. Heidelberg:
Springer.
Stadtler, H. and Kilger, C. (2008) Supply chain management and advanced planning.
4th edn. Heidelberg: Springer.
Stadtler, H. and Sürie, C. (2012) 'Production Planning and Detailed Scheduling
(PP/DS).' In H. Stadtler, B. Fleischmann, M. Grunow, H. Meyr and C. Sürie (Eds.)
Advanced Planning in Supply Chains: pp. 149-190. Heidelberg: Springer.
Suerie, C. and Stadtler, H. (2003) 'The Capacitated Lot-Sizing Problem with Linked Lot
Sizes.' Management Science, Vol. 49 No. 8, pp. 1039-1054.
Tang, L., Liu, J., Rong, A. and Yang, Z. (2000a). 'A mathematical programming model
for scheduling steelmaking-continuous casting production.' European Journal of
Operational Research, Vol. 120 No. 2, pp. 423-435.
Tang, L., Liu, J., Rong, A. and Yang, Z. (2000b) 'A multiple traveling salesman
problem model for hot rolling scheduling in Shanghai Baoshan Iron & Steel
Complex.' European Journal of Operational Research, Vol. 124 No. 2, pp. 267-
282.
Tang, L., Liu, J., Rong, A. and Yang, Z. (2001) 'A review of planning and scheduling
systems and methods for integrated steel production.' European Journal of
Operational Research, Vol. 133 No. 1, pp. 1-20.
Tang, L., Luh, P. B., Liu, J. and Fang, L. (2002) 'Steel-making process scheduling using
Lagrangian relaxation.' International Journal of Production Research, Vol. 40 No.
1, pp. 55-70.
Tang, L. and Wang, G. (2008) 'Decision support system for the batching problems of
steelmaking and continuous-casting production.' Omega, Vol. 36 No. 6, pp. 976-
991.
TKSA (2010) 'Welcome to the World of ThyssenKrupp Steel USA.' Retrieved from:
http://www.thyssenkruppsteelusa.com/. ThyssenKrupp Steel USA.
TKSA (2012) 'Day to day steel.' Retrieved from: http://www.thyssenkrupp-
csa.com.br/en/complex/day-to-day-steel.html. ThyssenKrupp Steel Americas.
Valls Verdejo, V., Alarcó, M. A. P. and Sorlí, M. P. L. (2009) 'Scheduling in a
continuous galvanizing line.' Computers & Operations Research, Vol. 36 No.1, pp.
280-296.
120 References

Vanhoucke, M. and Debels, D. (2009) 'A finite-capacity production scheduling


procedure for a Belgian steel company.' International Journal of Production
Research, Vol. 47 No. 3, pp. 561-84.
VdEh (2011) 'Jahrbuch Stahl 2012.' Düsseldorf: Verlag Stahleisen GmbH.
VdEh (2012) Statistisches Jahrbuch der Stahlindustrie. Wirtschaftsvereinigung Stahl.
Vonderembse, M. A. and Haessler, R. W. (1982) 'A mathematical programming
approach to schedule master slab casters in the steel industry.' Management
Science, Vol. 28 No. 12, pp. 1450-1461.
Wachter, J. C. (2011) 'A simulation study aimed at the increase of hot charging and
average charging temperature at the hot strip mill of Tata Steel Ijmuiden by using
hot boxes.' School of Management and Governance. University of Twente,
Enschede.
Wang, X. and Tang, L. (2008) 'Integration of batching and scheduling for hot rolling
production in the steel industry.' Internation Journal of Advanced Manufacturing
Technology, Vol. 36, pp. 431-444.
Wöhe, G. (2002) Einführung in die Allgemeine Betriebswirtschaftslehre. 21st edn.
München: Verlag Vahlen.
WSA (2009) 'World Steel in Figures 2009.' World Steel Association (Ed.).
WSA (2012a). 'Top Steel producing companies 2011.' World Steel Association (Ed.).
Retrieved from: http://worldsteel.org/statistics/top-producers.html.
WSA (2012b) 'World crude steel production - summary.' World Steel Association
(WSA). Retrieved from: http://worldsteel.org/statistics/statistics-archive/annual-
steel-archive.html.
Xu, C., Sand, G., Harjunkoski, I. and Engell, S. (2012) 'A new heuristic for plant-wide
schedule coordination problems: The intersection coordination heuristic.'
Computers & Chemical Engineering, Vol. 42, pp. 152-167.