0 Stimmen dafür0 Stimmen dagegen

1 Aufrufe263 Seitenétudes des cas de la logistique

Sep 12, 2019

© © All Rights Reserved

PDF, TXT oder online auf Scribd lesen

études des cas de la logistique

© All Rights Reserved

Als PDF, TXT **herunterladen** oder online auf Scribd lesen

1 Aufrufe

études des cas de la logistique

© All Rights Reserved

Als PDF, TXT **herunterladen** oder online auf Scribd lesen

Sie sind auf Seite 1von 263

Herausgegeben von

C. Bierwirth, Halle, Deutschland

B. Fleischmann, Augsburg, Deutschland

M. Fleischmann, Mannheim, Deutschland

M. Grunow, München, Deutschland

H.-O. Günther, Bremen, Deutschland

S. Helber, Hannover, Deutschland

K. Inderfurth, Magdeburg, Deutschland

H. Kopfer, Bremen, Deutschland

H. Meyr, Stuttgart, Deutschland

K. Schimmelpfeng, 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 Gebie-

ten der Produktion und Logistik. Aufgenommen werden vor allem herausragende

quantitativ orientierte Dissertationen und Habilitationsschriften. Die Publikatio-

nen vermitteln innovative Beiträge zur Lösung praktischer Anwendungsprobleme

der Produktion und Logistik unter Einsatz quantitativer Methoden und moderner

Informationstechnologie.

Herausgegeben von

Professor Dr. Christian Bierwirth Professor Dr. Herbert Kopfer

Universität Halle Universität Bremen

Universität Augsburg Universität Hohenheim

Universität Mannheim Universität Hohenheim

Technische Universität München Technische Universität Braunschweig

Technische Universität Berlin Universität Hamburg

Universität Hannover Universität Köln

Universität Magdeburg Universität Magdeburg

Kontakt

Professor Dr. Thomas S. Spengler

Technische Universität Braunschweig

Institut für Automobilwirtschaft

und Industrielle Produktion

Katharinenstraße 3

38106 Braunschweig

Thomas Kirschstein

Integrated Supply

Chain Planning in

Chemical Industry

Potentials of Simulation

in Network Planning

Foreword by Prof. Dr. Claudia Becker

und Prof. Dr. Christian Bierwirth

Thomas Kirschstein

Halle (Saale), Germany

ISBN 978-3-658-08432-5 ISBN 978-3-658-08433-2 (eBook)

DOI 10.1007/978-3-658-08433-2

Springer Gabler

© Springer Fachmedien Wiesbaden 2015

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or

part of the material is concerned, speci¿cally the rights of translation, reprinting, reuse of illus-

trations, recitation, broadcasting, reproduction on micro¿lms or in any other physical way, and

transmission or information storage and retrieval, electronic adaptation, computer software, or by

similar or dissimilar methodology now known or hereafter developed.

The use of general descriptive names, registered names, trademarks, service marks, etc. in this

publication does not imply, even in the absence of a speci¿c statement, that such names are

exempt from the relevant protective laws and regulations and therefore free for general use.

The publisher, the authors and the editors are safe to assume that the advice and information in this

book are believed to be true and accurate at the date of publication. Neither the publisher nor the

authors or the editors give a warranty, express or implied, with respect to the material contained

herein or for any errors or omissions that may have been made.

Springer Fachmedien Wiesbaden is part of Springer Science+Business Media

(www.springer.com)

V

Foreword

In literature on operations management, chemical industry is primarily dealt with focused

case studies. The limited ﬂexibility of production processes as well as the volatility of

customer demand were identiﬁed as the most prominent challenges for chemical supply

chain management. Most chemical production assets are complex, inﬂexible, and incur

high set-up costs. Hence, such assets can only be used in an economically reasonable way

if operated continuously. Therefore, logistical processes oﬀer substantial contributions to a

chemical company’s value added. E.g. logistical processes ensure the supply of downstream

assets in case of asset break downs. More important, however, logistical processes allow

balancing demand variations within chemical production networks. Therefore, transport

and distribution planning is essential in both, intra-site logistics, mostly relying on pipeline

transportation, and inter-site logistics, mostly relying on rail and ship transportation.

With this book, Thomas Kirschstein covers this topic in an encompassing, profound,

and general way. Starting with basic chemical production processes, which are modelled

with time series methods, established and new logistical planning problems for distributed

chemical production networks are presented. Finally, these components are integrated in a

simulation-based planning framework which is implemented in a decision support system.

The developed decision support system is validated by means of case studies relying on

historical records of a real world chemical production network.

The present work is an important contribution to scientiﬁc literature from a methodolo-

gical and application-oriented point of view. It develops systematically the basic elements

for modelling complex chemical production networks and illustrates the beneﬁts of advan-

ced decision support systems in chemical supply chain management. We wish the book

continued success and wide acceptance.

VII

Preface

This book is the product of a process which started in 2006 as a cooperation seminar

between the Dow Oleﬁnverbund GmbH and the chairs of Production & Logistics, Stati-

stics, and Operations Research. At this time I couldn’t imagine a result like this. During

this time not only the project has developed and changed, also my experiences, skills, and

expectations have grown. These developments would have been impossible without the

support of a bunch of persons. The space provided here does not suﬃce to duly thank all

these persons. So, the following thanks are rather exemplary than encompassing.

Special thanks are due to my supervisors Prof. Dr. Claudia Becker, Prof. Dr. Christian

Bierwirth, and Prof. Dr. Taïeb Mellouli which oﬀered invaluable support by their advices

and constant commitment in all those years.

No less important was the support of all employees of Dow Oleﬁnverbund GmbH which

allowed me to study not only Dow’s production and logistics processes but also their day-

to-day business. In particular, I thank Wolfgang Schnabel, Andreas Kroupa, and Wubbe

Prins for organizing and coordinating the practical part of this project. Without their

support, this book would not exist.

Besides my direct supporters, I’d like to thank all colleagues at the department for

Economics and Business Administration. Thanks to them research and teaching at the

university was always a pleasant and (almost) always a productive challenge. Represen-

tative for all colleagues I thank Heidrun Rudolph, Lisiane Schnegelsberg, and Dr. Steﬀen

Liebscher from the chair of Statistics as well as Ute Lorenz, Dorota Mańkowska, Jens

Kuhpfahl, and Prof. Dr. Frank Meisel from the chair of Production & Logistics.

An important contribution to the successful completion of this book oﬀered my family

and friends. Not only by numerous discussions about the contents of the projects but also

by the constant interest in its progress, I never lost sight of the goal of this project.

But, it is unthinkable that this book would exist without the support of my wife Susanne

and my daughters Antonia and Emilia. All three of them are constant sources of joy,

inspiration, and recuperation in particular in times of frustration and self-doubts. For

their support, I’m deeply grateful.

Thomas Kirschstein

IX

Contents

List of Tables XI

List of Figures XV

1 Introduction 1

1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2.1 Characterization of chemical production processes . . . . . . . . . . . . . . . 7

2.2 Modelling chemical production processes . . . . . . . . . . . . . . . . . . . . . 14

2.2.1 Chemical kinetics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.2.2 Modelling & simulation of chemical processes . . . . . . . . . . . . . 18

2.2.3 Process identiﬁcation & control . . . . . . . . . . . . . . . . . . . . . . 22

2.3 Time series methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.3.1 ARIMA models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.3.2 GARCH models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.3.3 Multivariate time series models . . . . . . . . . . . . . . . . . . . . . . 31

2.3.4 Data preparation, model speciﬁcation and residual checking . . . . . 33

3.1 Characteristics of chemical industry logistics . . . . . . . . . . . . . . . . . . 52

3.2 Planning problems for pipeline operations . . . . . . . . . . . . . . . . . . . . 54

3.2.1 Technical and organizational prerequisites . . . . . . . . . . . . . . . 54

3.2.2 Single-product pipelines . . . . . . . . . . . . . . . . . . . . . . . . . . 56

3.2.3 Multi-product pipelines . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

3.2.3.1 Batch ﬂow pipelines . . . . . . . . . . . . . . . . . . . . . . . 68

3.2.3.2 Batch split pipelines . . . . . . . . . . . . . . . . . . . . . . . 81

3.2.3.3 Multi-source pipeline systems . . . . . . . . . . . . . . . . . 85

3.3 Planning problems for rail operations . . . . . . . . . . . . . . . . . . . . . . . 87

X

3.3.2 A short-term rail transportation problem . . . . . . . . . . . . . . . . 90

3.3.2.1 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . 90

3.3.2.2 Components for modelling rail transports . . . . . . . . . . 91

3.3.2.3 Components for modelling turnover processes . . . . . . . . 92

3.3.2.4 Components for modelling the objective function . . . . . 93

3.3.2.5 Mathematical model . . . . . . . . . . . . . . . . . . . . . . . 94

3.4 Planning problems for ship operations . . . . . . . . . . . . . . . . . . . . . . 108

3.4.1 Technical and organizational prerequisites . . . . . . . . . . . . . . . 108

3.4.2 Maritime inventory routing problems . . . . . . . . . . . . . . . . . . 110

3.4.3 Maritime inventory shipping problems . . . . . . . . . . . . . . . . . . 113

4.1 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

4.2 Sources and eﬀects of uncertainty in chemical industry . . . . . . . . . . . . 141

4.3 A framework for simulation-based integrated planning of supply chains in

chemical industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

4.3.1 Conceptual modelling & data analysis . . . . . . . . . . . . . . . . . . 151

4.3.2 Components of chemical supply chain simulation models . . . . . . . 159

4.3.3 Veriﬁcation & validation . . . . . . . . . . . . . . . . . . . . . . . . . . 167

4.3.4 Planning of simulation experiments . . . . . . . . . . . . . . . . . . . 170

4.3.4.1 Performance measures in (chemical) supply chain models . 172

4.3.4.2 Experimental designs . . . . . . . . . . . . . . . . . . . . . . 175

4.3.4.3 Simulation optimization . . . . . . . . . . . . . . . . . . . . . 185

4.3.5 Decision support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196

Bibliography 207

Appendix 229

XI

List of Tables

2.2 Characteristics of production processes in chemical industry . . . . . . . . . 10

2.3 EACF for T = 100 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.4 EACF for T = 1000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.5 theoretical EACF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.6 Information criteria for both sampled time series and various model speci-

ﬁcations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.7 Estimated parameters for the Naphtha time series . . . . . . . . . . . . . . . 43

2.8 Coeﬃcients of the initial V ARX(3) model for de-alkylation plant . . . . . 46

2.9 Coeﬃcients of the V ARX(3) model with outlier correction and variable

selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

3.2 Transition matrices and production modes for providers and consumers . . 64

3.3 All combinations of plant production modes . . . . . . . . . . . . . . . . . . . 65

3.4 Transition matrix Q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

3.5 Comparison of problem features for Magatão et al. (2004) and Relvas et al.

(2006) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

3.6 Set of parameters and decision variables for the sELSP . . . . . . . . . . . . 73

3.7 Transition costs ctrans

st for the PIG insertion scenario . . . . . . . . . . . . . . 78

3.8 Interface quantities dst and transition costs ctrans

st for the interface scenario 79

3.9 Net demand rate ωs and holding cost rates chold s for both scenarios . . . . . 79

3.10 Resulting optimal pumping cycles for the sELSP-BP and sELSP-BP-IF . . 80

3.11 Pumping times and idle times of the optimal schedules for the sELSP-BP

and sELSP-BP-IF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

3.12 Classiﬁcation of literature on scheduling of one-to-many pipeline systems . 83

3.13 Sets, parameters, variables, and decision variables for the MC-RTP . . . . . 98

3.14 Exemplary assignment of RTCs to two trains . . . . . . . . . . . . . . . . . . 100

3.15 Technical parameters for the rail operations planning example . . . . . . . . 103

3.16 Consumption rate ωis , stock capacities sCap Ini

is , initial stocks sis , and target

Tar

stock levels sis for all sites i and chemicals s . . . . . . . . . . . . . . . . . . 104

3.17 Sets, parameters, variables, and decision variables for the MISP-STA . . . 115

3.18 Technical speciﬁcation of tankers and tanks (nCap Cap Trv

vk , qsk , cv ) . . . . . . . . 117

XII

planning in chemical industry . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

4.2 Classiﬁcation of literature on integrated SC management planning in chem-

ical industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

4.3 Classiﬁcation of sources of uncertainties and examples (per dimension) . . 145

4.4 Material balances (in t/hour) and storage capacities (in t) per site and

chemical . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

4.5 Examples for processors according to processor type and processing attributes160

4.6 Relation between aggregation levels and types of processors . . . . . . . . . 162

4.7 Pseudo-code for chemical SC simulation model . . . . . . . . . . . . . . . . . 166

4.8 Classiﬁcation of V&V techniques . . . . . . . . . . . . . . . . . . . . . . . . . 168

4.9 Attributes of variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172

4.10 Examples and classiﬁcation of performance measures in SCM . . . . . . . . 174

4.11 Values of the control variables for the experimental design . . . . . . . . . . 179

4.12 Eﬀects of input variables on responses in example 11 . . . . . . . . . . . . . 182

4.13 Overview on simulation optimization methods . . . . . . . . . . . . . . . . . 187

4.14 Optimal and baseline values for inventory parameters . . . . . . . . . . . . . 200

4.15 Performance measures for optimal and baseline values of inventory parameters202

A.1 Information criteria for VARX models of order 1 to 6 for the de-alkylation

plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

A.2 Coeﬃcients of the V ARX(3) model with outlier correction . . . . . . . . . 230

A.3 ANOVA table for the initial V ARX(3) model (without outlier correction) 230

A.4 ANOVA table for the initial V ARX(3) model with outlier correction and

variable selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232

A.5 Stationary ﬂow rates (in t/h) for exemplary chemical production network . 234

A.6 Coeﬃcients of the time series models for cracker at site 1 (V ARX(2)+AR(3))234

A.7 Coeﬃcients of the time series models for cracker at site 2 (V ARX(2)+AR(3))234

A.8 Coeﬃcients of the time series models for hydrogenation plant at site 1

(V ARX(3) + AR(2)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

A.9 Coeﬃcients of the time series models for hydrogenation plant at site 2

(V ARX(3) + AR(3)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

A.10 Coeﬃcients of the time series models for Butex plant at site 2 (V ARX(1)+

AR(4)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

A.11 Coeﬃcients of input and output time series models of SISO and MISO plants235

A.12 Un-/loading capacities, initial stock of empty RTCs, target and initial stock

levels per transported chemical at both production sites . . . . . . . . . . . 236

A.13 Cost rates for the MC-RTP instances . . . . . . . . . . . . . . . . . . . . . . . 236

A.14 Distributions of inter-arrival time and deliver quantities for external cus-

tomers/suppliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236

XIII

A.16 Transition matrices for plants at site 2 . . . . . . . . . . . . . . . . . . . . . . 237

A.17 Resolution V design for 6 dichotomous variables ("-1" encodes the variable’s

lower level and "1" the upper level) . . . . . . . . . . . . . . . . . . . . . . . . 238

A.18 Estimated responses for all possible conﬁgurations (dominated conﬁgura-

tion gray coloured) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239

A.19 Pseudocode for inventory system model . . . . . . . . . . . . . . . . . . . . . 243

A.20 Summary for the logistic regression model of responses of eﬃcient settings 244

XV

List of Figures

2.2 Production quantities of basic chemicals in 2008-2010 . . . . . . . . . . . . . 6

2.3 Exemplary production network of Naphtha derivatives . . . . . . . . . . . . 11

2.4 Scheme of a distillation column . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.5 Overview on chemical operations . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.6 Steps in chemical process modelling & control . . . . . . . . . . . . . . . . . 14

2.7 Topics in chemical kinetic and key words/methods . . . . . . . . . . . . . . . 15

2.8 Schematic ﬂow sheet of a steam cracker . . . . . . . . . . . . . . . . . . . . . 21

2.9 Chemical process control scheme . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.10 Theoretical and empirical ACF and PACF for an ARMA(2,1) process . . . 38

2.11 ACF of residuals, ACF of squared residuals and QQ-plot of residuals for

ARMA(2,0) model (small sample, T = 100) and for ARMA(2,1) model

(large sample, T = 1000) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

2.12 Time series plot of Naphtha inﬂow rate . . . . . . . . . . . . . . . . . . . . . 42

2.13 ACF and PACF of corrected Naphtha time series . . . . . . . . . . . . . . . 43

2.14 Flowsheet of the de-alkylation plant . . . . . . . . . . . . . . . . . . . . . . . 44

2.15 Raw data for a de-alkylation plant . . . . . . . . . . . . . . . . . . . . . . . . 45

2.16 Residual diagnostic plots for V ARX(3) model of the de-alkylation plant. . 47

2.17 Scatterplot of residuals for outlier corrected V ARX(3) model of the de-

alkylation plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

2.18 Real and ﬁtted time series of both outﬂow rates (outlier indices superimposed) 50

3.2 Modal split for chemical products in Germany in 2009 (based on total

transported quantity in mill. tons) . . . . . . . . . . . . . . . . . . . . . . . . 52

3.3 Scheme of an exemplary serial multi-access pipeline . . . . . . . . . . . . . . 57

3.4 Cumulative distribution function and loss function for Y (4) . . . . . . . . . 66

3.5 Inventory pattern in a batch ﬂow system . . . . . . . . . . . . . . . . . . . . 71

3.6 Illustration of interface calculation in batch pipeline systems. . . . . . . . . 76

3.7 Illustration of interface calculation in batch pipeline systems . . . . . . . . . 81

3.8 One-to-many pipeline types . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

XVI

3.10 Optimal stock levels and network ﬂows for periods 1 to 8 of the MC-

RTP example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

3.11 Optimal stock levels for the MC-RTP instance of example 6 . . . . . . . . . 107

3.12 Optimal stock levels and chemical ﬂows for periods 1 to 8 of the MISP-

STA example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

3.13 Optimal stock levels for the MISP-STA instance of example 7 . . . . . . . . 120

4.2 Adapted SCM matrix (based on Stadtler (2005)) . . . . . . . . . . . . . . . . 127

4.3 Daily median inﬂow rates of aromatic hydrocarbons for a de-alkylation plant147

4.4 Hourly average inﬂow rate of aromatic hydrocarbons (including conﬁdence

intervals at a conﬁdence level of α = 0.1%) . . . . . . . . . . . . . . . . . . . . 148

4.5 Diagnostic plots of trimmed inﬂow rates of aromatic hydrocarbons for a

de-alkylation plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

4.6 Steps in the simulation-based planning projects . . . . . . . . . . . . . . . . 150

4.7 Development scheme for simulation studies . . . . . . . . . . . . . . . . . . . 152

4.8 Exemplary chemical supply chain with two sites . . . . . . . . . . . . . . . . 154

4.9 Flow chart of a part of the exemplary chemical SC using the notation of

Table 4.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

4.10 Development scheme for simulation studies . . . . . . . . . . . . . . . . . . . 165

4.11 Steps in experimental planning . . . . . . . . . . . . . . . . . . . . . . . . . . 171

4.12 Simulated and ﬁtted responses per experimental conﬁguration . . . . . . . . 183

4.13 Loss functions for Naphtha consumption during pipeline inspection and

their relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191

4.14 Density and loss function for total Naphtha consumption during order lead

time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192

4.15 Scatterplot of Pareto front (β-service level in grey scale) . . . . . . . . . . . 194

4.16 Grey-scaled levelplot of estimated Pareto front . . . . . . . . . . . . . . . . . 195

4.17 Example of Delaunay triangulation-based sub-sample generation . . . . . . 199

4.18 Diagnostic plots of ARDs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

4.19 Graphical determination of optimal performance vector . . . . . . . . . . . . 202

A.1 Diagnostic plots for the ARX(1) model of the Naphtha time series . . . . . 229

A.2 Residual diagnostic plots for V ARX(3) model with outlier compensation. 233

A.3 QQ-plots of residuals for models (4.19)-(4.21) . . . . . . . . . . . . . . . . . . 240

A.4 Density function of the Weibull distribution with k = 1.5 and λ = Γ( 365

5

)

. . . 241

3

A.5 Loss functions for both sites during pipeline inspection . . . . . . . . . . . . 242

A.6 Diagnostic plots for (4.22) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245

XVII

List of Abbreviations

AB agent-based

ACF auto-correlation function

AIC Akaike information criterion

AR(X) auto-regressive [model] (with exogenous variables)

auto-regressive conditional heteroscedasticity [model]

ARCH(X)

(with exogenous variables)

auto-regressive integrated moving average [model]

ARIMA(X)

(with exogenous variables)

auto-regressive moving average [model] (with

ARMA(X)

exogenous variables)

ARD average relative deviation

CV corporate value

DE discrete event

DSS decision support system

EACF extended auto-correlation function

(s)ELSP(-BP- (sequence-dependent) economic lot sequencing problem

IF) (with global stocks and interface handling)

FIR ﬁnite impulse-response

generalized auto-regressive conditional

GARCH(X)

heteroscedasticity [model] (with exogenous variables)

GA genetic algorithm

GDP gross domestic product

h hour

HQIC Hannan-Quinn information criterion

IRP inventory routing problem

IPPSP integrated production planning and scheduling problem

(s)LP (stochastic) linear program

MA(X) moving average [model] (with exogenous variables)

MAD median absolute deviation (from median)

MC-RTP multi-chemical rail transportation problem

(s)MILP (stochastic) mixed-integer linear program

MIMO multiple input, multiple output

MINLP mixed-integer non-linear program

MIP mixed-integer program

MIRP maritime inventory routing problem

MIRSP maritime inventory routing and scheduling problem

XVIII

MISP(-STA)

and tank assignment)

MISO multiple input, single output

Nomenclature statistique des activités économiques

NACE

dans la Communauté européenne

NLP non-linear program

N P-hard non-deterministic polynomial-time hard

NPV net present value

NSGA non-dominated sorting genetic algorithm

OLAP on-line analytical processing

PACF partial auto-correlation function

PE(T) polyethylene (terephthalate)

PDP pick-up and delivery problem

PIG pipeline inspection gauges

PP polypropylene

PS polystyrene

QQ-plot quantile-quantile plot

RMSE root mean squared error

ROV real-options-based value

RSA/RSM response surface approximation/methodology

RTC rail tank car

SBR styrene-butadiene rubber

SC supply chain

SCM supply chain management

SCOR Supply Chain Operations Reference

SD system dynamics

SDE stochastic diﬀerential equations

SIC Schwarz information criterion

SIMO single input, multiple output

SISO single input, single output

t ton

TSM time series methodology

TSP travelling salesman problem

V&V veriﬁcation and validation

vector auto-regressive [model] (with exogenous

VAR(X)

variables)

vector auto-regressive moving average [model] (with

VARMA(X)

exogenous variables)

VLE vapour-liquid-equilibrium

vector moving average [model] (with exogenous

VMA(X)

variables)

VRP vehicle routing problem

XIX

List of Notation

General notation

a, ..., z, α, ..., ω vectors

A, ..., Z, Γ, ..., Ω matrices

A, ..., Z sets

i, j, k, b indices

t, T time index, time horizon

E(⋅)/P (⋅)/V ar(⋅)/Cov(⋅, ⋅) expectation/probability/variance/covariance of ⋅

Γ(⋅) Gamma function

ˆ

x̂/x̂ estimate of x / estimate of estimate of x

Bin (n, p) Binomial distribution with n trials and probability p

N (μ, σ 2 ) Normal distribution with mean μ and variance σ 2

WB (k, λ) Weibull distribution with shape k and scale λ

Notation of Chapter 2

N index and number of lags for exogenous variables

L/M number of output/exogenous variables

K number of parameters

p/q/P /Q order of AR/MA/ARCH/GARCH process

α, β coeﬃcients of (G)ARCH models

γ/ρ, Corr(⋅, ⋅) partial correlation/ correlation

φ, θ AR and MA coeﬃcients

Φ, Θ VAR and VMA coeﬃcient matrices

coeﬃcient scalars/vectors/matrices for exogenous

υ/υ/Υ

regressors

δ(B), ω(B), θ(B), φ(B) univariate lag polynomials

Φ(B), Θ(B), Υ(B) multivariate lag polynomials

σ2 variance

Σ covariance matrix

μ/μ mean scalar/vector

, ε, ξ, η/, ξ/Ξ error scalars/vectors/matrices

B back-shift operator

C(t) concentration function

X(t) turnover function

XX

Cs starting concentration

A reactant

EA activation energy

R universal gas constant

tR Damköhler number

m reaction order

r reaction rate

k reaction rate constant

W (t) Wiener process

Notation of Chapter 3

K, H set of state combinations

I set of lot positions

N set of access points

O set of states

S set of modes

ρCap pumping rate capacity

ρ pumping rate

Q transition matrix

qst transition probability from state s to state t

π steady state vector

ωs /ωs ﬂow level of state s/state combination s

Xj ﬂow rate at access point j

yh total material balance of state combination h

ph probability of state combination h

μC expected net deﬁcit

α, β service level

r stock level

V (⋅) expected loss function

L/L̄ stock level / maximum stock level

T stock /T f ill stock-up time/pipeline ﬁll time

τ time for pipeline transport

s /cs

cset hold set-up / holding costs for product s

ctrans

st transition costs for a change from product s to t

binary, 1 if at position i = 0, ..., I chemical s is

bsi

scheduled

Ts cycle time of chemical s ∈ S

tbi starting time of position i

binary, 1 if from position i − 1 to position i a transition

xsti

from chemical s to t occurs

XXI

osij

before i

binary, 1 if position i is the last position where

rsi

chemical s is scheduled

T cycle fundamental cycle time

Notation of Section 3.2 summarized in Table 3.13

Notation of Section 3.3 summarized in Table 3.17

Notation of Chapter 4

D set of Delaunay simplices

I set of sites

O set of states

P set of products/chemicals

Y set of performance vectors

set of system conﬁgurations / eﬃcient system

Z/Z ef f

conﬁgurations

ωs ﬂow level of state s/state combination s

Q transition matrix

π steady state vector

ρCap pumping rate capacity

ρ pumping rate

q (opt) (optimal) order quantity

t(opt) (optimal) order interval

cbatch /chold batch-injection / holding cost for product

y vector of dependent measures

n number of dependent measures (length of y)

x vector of control variables

m number of control variables (length of x)

z vector of performance measures (typically z = (x, y))

ˆ real/simulation/meta-model mapping function linking

H(⋅)/Ĥ(⋅)/Ĥ(⋅)

x and y

C(⋅) cost function

U (⋅) proﬁt/performance function

V (⋅) (ﬁrst order) loss function

mi number of levels of control variable i

b binary vector, coding a discrete system conﬁguration

μ intercept vector/basic performance

γ/γ/Γ coeﬃcient scalar/vector/matrix of linear model

relative average target stock deviation at site i for

rip

product p

XXII

r

and all products

sTar

ip target stock level for product p at site i

sCap

ip inventory capacity for product p at site i

sitp stock level for product p at site i in period t

tr total number of trains dispatched

trijt number of trains dispatched from i to j in period t

βip service level for product p at site i

β average service level for all products and sites

real

ωipt realized chemical ﬂow of product p at site i in period t

plan planned/required chemical ﬂow of product p at site i in

ωipt

period t

(si , Si ) inventory parameters (re-order level and stock-up level)

lit (Naphtha) stock level at location i in period t

¯l average total stock level ( ∑t ∑T i lit )

δ number of shipments

talarm time interval between successive pipeline inspections

tinsp duration of pipeline inspections

tship order lead time

Delaunay simplex (set of points constituting the

d

simplex)

relative performance deviation vector / average relative

e/e

performance deviation

1

1 Introduction

1.1 Motivation

The chemical industry is a key industry in developed countries. Not only its size but also

its role as a raw material supplier for almost all other industries establishes the chemical

industry as one of the most important industries in a nation’s economy. The chemical

industry is deﬁned by the type of production process. Chemical industry focuses on the

modiﬁcation of a substance’s chemical and/or physical properties.1

In this thesis the chemical industry and, particularly, the basic chemical industry is

the subject of interest. The basic chemical industry provides industrial goods (e.g. basic

plastics, coatings, fertilizer) as well as consumer products (e.g. detergents, tires, insulat-

ing materials). This industry can be seen as an indicator for a nation’s economy since

shocks in economic development (or the expectation of it) typically ﬁrst aﬀect the basic

industries. A general change in (expected) economic development induces a revision of

raw material stock levels to match the expected production plans in a broad range of com-

panies in various industries. Since almost all industries use various chemical products as

raw materials, the chemical industry is confronted with the accumulated "revision eﬀect"

manifesting itself either in an increase or decrease of total customer demand. This is typi-

cally one of the ﬁrst measurable and tangible eﬀects of a change in economic development.

Hence, companies in chemical industry are faced with a volatile and hardly inﬂuenceable

customer demand. Moreover, most chemical products are commodities, i.e. there are only

few possibilities for product diversiﬁcation such that the market for chemical products is

highly competitive with a strong focus on prices.

Another crucial characteristic is that many chemical production processes are only cost-

eﬃcient if operated continuously. Continuously operated chemical production plants are

expensive and complex technical systems. Typically, they are optimized for operation

at a speciﬁc conversion rate such that operational costs per output unit are minimized.

Deviations from the optimal production rates typically cause additional operational costs

stressing the small contribution margins.

Chemical transformation processes are typically split into multiple stages that are or-

ganizationally separated but technologically interdependent. Chemical production plants

are interconnected via direct energy and product ﬂows but managed locally. Due to diverg-

1

See Schönsleben (2011).

T. Kirschstein, Integrated Supply Chain Planning in Chemical Industry, Produktion und Logistik,

DOI 10.1007/978-3-658-08433-2_1, © Springer Fachmedien Wiesbaden 2015

2

ing production processes in basic chemicals’ production as well as the processes’ continuity

and interconnectedness, chemical production plants are often clustered at large-scaled in-

tegrated chemical production sites. This form of organization is economically preferable

since co-products can be processed further with little logistical eﬀort. However, managing

such industrial complexes is a challenging task since all product ﬂows have to be kept in

balance simultaneously to maintain all production processes. This is particularly compli-

cated if unforeseen events occur. Therefore, local production and consumption as well as

material imports and exports have to be planned accordingly. This perspective is called

the intra-site perspective.

Basic chemical companies often operate a certain number of such integrated production

sites such that a geographically disperse network of production sites exists. To keep all

materials balanced at all sites, material transports between sites have to be considered.

This perspective is called the inter-site perspective.

In summary, basic chemical companies face a challenging combination of diﬃcult market

conditions, inﬂexible and complex production processes, and a wide-spread, interdepen-

dent production network. Hence, there is an inevitable strain for minimizing the oper-

ational costs of production from a company-wide perspective. To identify and exploit

such a network’s cost saving potentials, elaborate decision support tools are necessary

taking into account not only the complex production network of each site but also the

interrelations between sites. Such a company-wide perspective typically addresses tac-

tical and strategical decisions since, on the operational level, the local management of

the production sites is responsible. However, operational processes have to be modelled

accurately to be able to anticipate the network’s reactions on an aggregated level. On the

tactical and strategical level, environmental inﬂuences have to be incorporated to reﬂect

the uncertain nature of future developments.

At this level of aggregation, multiple objectives are typically pursued. Beside the op-

erational costs of a certain network conﬁguration, e.g. investment costs to realize such a

conﬁguration as well as robustness aspects and service measures come into scope.2 Thus,

tactical/strategical decision support tools have to provide

• a reliable model of the production sites (including the production processes and the

local logistical processes),

2

Here, robustness can be understood as the ability of a system conﬁguration to perform well under

various environmental scenarios such that the risk of a critical state of the network is minimized or,

at least, restricted to some upper bound.

3

• the capacities of the logistical system (for transport, turn-over, and stock holding),

This thesis aims at providing a general framework for building a tactical/strategical deci-

sion support model to optimize (large-scaled) chemical production networks. This frame-

work consists of three core parts:

1. To model local chemical production processes at production sites, time series mod-

els are used which are able to accurately represent the dynamic, time-dependent

structure of chemical production processes.

(linear) planning models are used to decide about the logistical activities within

the network on an operational level. Hence, the behaviour of the logistic systems is

described dynamically considering varying system states.

ages the interactions between the production system and the logistical system.

Moreover, additional stochastic inﬂuences can be incorporated.

This thesis is organized as follows: Chapter 2 introduces an overview on chemical pro-

duction processes. Relevant chemical and physical concepts are introduced which are

essential to understand the dynamic nature of chemical reactions. It is shown that time

series models are adequate methods to describe the behaviour of chemical production

plants.

Chapter 3 provides an overview on logistical planning models typically occurring in

chemical production networks. These models are categorized by means of the transport

mode.

Chapter 4 constitutes the main chapter of this thesis. A literature review on integration

concepts for network planning in chemical industry is provided. A categorization of risk

sources aﬀecting the performance of a network conﬁguration is provided that subsequently

allows modelling of stochastic environmental disturbances and their eﬀects. The integra-

tion of local production site models and logistical interaction models in a simulation

framework is described. Experimental designs and simulation optimization techniques

are described. These techniques are used to ﬁnd conﬁgurations of the simulated network

which improve its performance.

The thesis concludes with Chapter 5 where the main results of this work are reviewed,

discussed, and summarized.

4

Throughout this thesis, all methods and techniques described are illustrated by small-

scaled examples or case studies based on real-world data. In Chapter 4, a complex case

study inspired by a cooperation project with a chemical company is presented, integrating

the very most of the aforementioned methods. Figure 1.1 shows a graphical overview on

the structure of the thesis and the associated examples.

Chapter 1: Introduction

section example

Motivation —

Outline —

section example Chapter 3: Distribution planning

Characteristics —

Classiﬁcation — section example

Pipeline

(TSM) Ex. 5: Batch pipeline sequencing

Modelling Ex. 2: applied univariate TSM Rail Ex. 6: Rail distribution planning

Ex. 3: applied multivariate TSM Ex. 7: Maritime inventory

Ship

shipping

section example

Literature —

Uncertainties Ex. 8: Modelling plant states

Ex. 9: Conceptual model of a chemical SC

Case study

Simulation Ex. 11: Simulation optimization by experimental designs

Ex. 12: Simulation optimization by a genetic algorithm

Ex. 13: Post-optimization for decision support

Chapter 5: Conclusion

5

This chapter provides an overview on basic deﬁnitions, terms, concepts, and techniques

to describe and model chemical production processes. This allows modelling of the core

components in chemical production networks. Figure 2.1 shows an exemplary chemical

production network where the production plants are highlighted.

intra-site inter-site

S1 C1

intra-site

intra-site

S2 C2

From Figure 2.1, it becomes obvious that modelling the production plants provides the

basic data to describe the material ﬂows within the whole network. Before the chemical

production processes are described in detail, some characteristics and key ﬁgures about

chemical products are provided.

Basically, two major groups of chemicals can be distinguished: inorganic and organic

chemicals. The latter subsumes all chemicals containing at least one bond of a hydrogen

and carbon atom. The former group encompasses all chemicals without such a bond.1

This work primarily focuses on organic chemicals which are produced by the (basic)

chemical industry. For the industrial production of organic chemicals, three basic natural

resources are available: natural gas, coal, and crude oil. Of these three materials crude oil

1

In particular, pure carbon, e.g. in the form of diamonds is handled as an inorganic chemical. See

e.g. Seager and Slabaugh (2007) for more details.

T. Kirschstein, Integrated Supply Chain Planning in Chemical Industry, Produktion und Logistik,

DOI 10.1007/978-3-658-08433-2_2, © Springer Fachmedien Wiesbaden 2015

6

is the exceedingly most important raw material for the production of organic chemicals.

The (basic) chemical industry is an intermediary who transforms these raw materials

into basic and intermediate (organic) chemicals. Three main production phases can be

distinguished:

1. raw substance splitting: raw materials are split into (short-chained) basic chemicals

The category of basic chemicals comprises about 20-30 chemical substances building the

basis for all subsequent substances. This class summarizes basic organic substances as

well as basic gases and inorganic basic substances (such as Chlorine or Ammonia).2 These

substances are purely intermediate and not sold to consumer markets.

Intermediate chemicals are usually simply structured chemicals most often composed

by basic chemicals that are only exceptionally made for (private) consumer markets.

This class comprises, e.g., alcohols and many kinds of acids.3 Examples for marketable

intermediate chemicals are ammonia compounds used as basic fertilizers in agriculture.

The classes of basic and intermediate chemicals consist of a fairly small number of sub-

stances which are mainly ﬂuids or gases. But these substances are used in vast quantities

to produce ﬁnal chemical products. To give an impression, Figure 2.2 shows the produced

quantities of basic and intermediate chemicals in 2008-2010 for Germany (grouped as

stated above).4

production quantities in 1,000 tons

100000

organic chemicals

inorganic chemicals

basic gases

80000

60000

40000

20000

0

2

See e.g. Behr et al. (2010, p. 10 ﬀ.) or Baerns et al. (2006, ch. 16).

3

See e.g. Behr et al. (2010, p. 10 ﬀ.) or Baerns et al. (2006, ch. 17).

4

These production statistics are calculated based on the European goods classiﬁcation scheme NACE.

The three classes "basic gases", "inorganic chemicals", and "organic chemicals" correspond to the

NACE classes 20.11, 20.13, and 20.14, see Eurostat (2011).

7

The group of organic chemicals represents about 50% of the total quantity of produced

chemicals with a total quantity of about 40 million tons per year in Germany. These prod-

ucts are typically crude oil derivatives. Thereof, approximately 20 million tons account

for organic basic chemicals such as alkenes or aromatic compounds where the remain-

ing 20 million tons comprise organic intermediate chemicals such as alcohols or chlorine

derivatives.5

Final products in chemical industry are produced by chemical reactions of intermediate

and basic chemicals. These ﬁnal products are widely used in almost all other industries.6

Final chemicals can be categorized into

1. polymers/plastics

5. pharmaceuticals.

Among these categories, pharmaceuticals and body care products set up own sub-

industries due to the special characters of their products and production processes.

These companies directly serve (private) consumer markets. The remaining three cat-

egories comprise "classic" chemical companies mainly providing intermediate products

for other industries such as mechanical engineering industry, building industry, textile

industry, and plastics industry.

In the next section chemical production processes are characterized and categorized.

The theoretical modelling of the underlying chemical reactions is described subsequently.

Chemical production processes realise chemical reactions in industrial scale in chemical

plants. Based on models of chemical reactions, methods are provided to describe the

behaviour of chemical production processes.

Production in chemical industry is in many aspects diﬀerent from common industries.

In most industries ﬁnal products are produced by mechanical transformation processes

such as assembling or machining. In contrast, in chemical industry the chemical and/or

physical properties of substances are altered. Most chemical production processes rely on

chemical reactions aiming at the transformation of reactants into substances of interest.

The term chemical reactions refers to changes of the reactants’ molecular structures.7 Be-

5

See Eurostat (2011).

6

However, especially in the last two decades a trend towards a deeper vertical integration and special-

ization is visible in chemical industry. This aims at product portfolios containing a higher share of

consumer goods which promise higher contribution margins and less market risks.

7

See e.g. Baerns et al. (2006, p. 24 ﬀ.).

8

side chemical reactions, also physical transformation processes (so-called basic operations)

are used to alter speciﬁc properties of the reactants.

Usually, chemical reactions require an initiation, i.e. they only take place under speciﬁc

circumstances. The list of reaction parameters is vast. Basic physical parameters are

pressure, temperature, electricity, or light. Speciﬁc constellations of these parameters

inﬂuence the reaction rate, i.e. how fast or slow a reaction takes place.8 If a reaction

can only take place by means of auxiliary chemicals, it is called catalytic and such an

auxiliary reactant is called catalyst.9 Another important measure of chemical reactions

is the conversion rate of a reaction, i.e. how many percent of the input reactants’ mass

is transformed into the substances of interest.10 This measure is important to decide

whether a reaction can be realized in an economically proﬁtable way.

The molecular structure of reactants can be changed in numerous ways:11 First the

reactants’ molecules can be combined which is called synthesis. A prominent example is

the hydrogenation of carbon dioxide to produce methanol.12

Second, a molecule or molecular fragment can also be split which is called decomposition.

To recycle bottles made of polyethylene terephthalate (PET) catalytic depolymerization

is used to split the PET in valuable components.13

In a substitution, a molecular fragment of a reactant is replaced by a fragment of another

reactant. A prominent example is the alkylation where an alkyl group is transferred from

one molecule to another. E.g., the production of Ethylbenzene from Benzene and Ethylene

by the so-called Friedel-Crafts alkylation is a standard process in chemical industry.14

If more than one molecular fragment is substituted, this is categorized as a metathesis.

A recent industrial application is the oleﬁn metathesis to produce e.g. Propene from

Ethene and 2-Butene.15

The types of reactions presented above can be combined with the ordinary classiﬁ-

cation of production processes in convergent, divergent, and transformation processes.

Depending on the number of input and output products this process classiﬁcation can be

enhanced as Table 2.1 shows.

A SISO process is a single input-single output process and corresponds to transforma-

tion processes or substitution reactions. MISO processes (multiple inputs-single output)

comprise convergent processes and SIMO processes (single inputs-multiple output) com-

prise divergent processes similar to decomposition and synthesis, respectively. MIMO

8

E.g. see Baerns et al. (2006, p. 32).

9

The catalyst is not part of the molecular reorganization, i.e. no part of the catalyst is part of the

resulting chemicals. See Behr et al. (2010, ch. 12) for more information.

10

E.g. see Baerns et al. (2006, ch. 11.2) for more details.

11

A similar classiﬁcation scheme particularly addressing organic reactions can be found in Jones and

Bunnett (1989).

12

See Bill (1997).

13

The resulting components depend on the speciﬁc depolymerization process applied, see e.g. Mishra

et al. (2002) or Paszun and Spychaj (1997).

14

E.g. see Degnan et al. (2001).

15

See Mol (2004).

9

```

``` # outputs

``` single multiple

# inputs ```

```

single SISO SIMO

multiple MISO MIMO

vergent and divergent parts. They correspond to metatheses or coupled decomposition

reactions.

Based on this classiﬁcation of chemical reactions, corresponding characteristics of the

production systems can be deduced. The production of basic chemicals is in almost all

cases a split of long-chained raw materials (crude oil, coal, natural gas) into short-chained

substances (such as alkenes). These production processes are primarily divergent and can

be categorized as SIMO processes. The cracking of mixtures of long-chained substances

into their short-chained components is usually performed by thermal chemical reactions

(such as steam-cracking). The resulting mixture of short-chained substances is usually

separated by distillation processes. Such processes are in general continuous and hardly

interruptible. Typically, they are single-purpose assets i.e. assets designed to produce a

ﬁxed set of chemicals.16 Depending on the raw materials’ composition and the operating

parameters (temperature, pressure, reaction time, etc.) the production coeﬃcients can be

controlled under certain restrictions. The set-up and control of the production coeﬃcients

depends on the precedence relation of the produced products. Not in all cases a focal main

product exists.17

The production of intermediate chemicals requires more manifold types of production

processes. Similar to the production of basic chemicals, chemical reactions are typically

accompanied by separation processes such that most of these reactions can be catego-

rized as SIMO or MIMO processes. Because intermediate chemicals are required for the

production of ﬁnal chemical products in huge quantities, they are usually produced by

continuously operated plants. Typically, the production plants are specialized to perform

a speciﬁc reaction and, hence, are single-purpose plants.

Final chemical products are typically produced on multi-purpose plants which are de-

signed for a speciﬁc product family. Such production processes are usually convergent.

The composition of raw materials to produce a ﬁnal chemical is called a recipe. Multi-

purpose plants are capable to handle multiple recipes, i.e. reactants and products handled

vary in both type and quantity. These processes can be mainly categorized as MISO or

MIMO processes.

16

However, the production coeﬃcients may vary.

17

This depends primarily on the further use of the output products and/or their market prices. For

example steam crackers are usually optimized for Ethylene production because Ethylene is used in a

wide variety of ﬁnal products.

10

Depending on the product portfolio as well as the structure of product demand, multi-

purpose plants are operated either in batch mode or continuously. There are plenty of

deﬁnitions about both terms.18 Here, a technological point of view is used, i.e. batch

processes are characterized by a ﬁxed production capacity which is deﬁned as the quan-

tity of produced chemicals after which a process interruption is required. In contrast,

continuous processes are characterized by a production rate, which is deﬁned as the quan-

tity of goods produced in a given time. The time between process interruptions is not

technically limited. The former especially occurs for specialty chemicals and pharmaceu-

ticals, where the latter is typical for polymers/plastics and some agrochemicals.19 For

body care products the production technology is mixed depending on the product variety.

Low-volume products with many product variations and often changing recipes, such as

cosmetics, are usually produced in batch mode, whereas high-volume products with few

product variations, such as detergents, are usually produced continuously. In both cases,

the production processes are interruptible. Table 2.2 summarizes the above-mentioned

characteristics.20

ﬁnal chemicals

basic chemicals intermed. chemicals

commodity speciality

mode continuous continuous continuous batch

vergence divergent di-/convergent convergent/transform.

purpose single single multiple

of continuously processed plants. This principally leads to an advantage for horizontally

coupled production processes and, hence, horizontally integrated chemical companies.

To exploit these economies of scope, locally concentrated production sites are necessary

to avoid logistical eﬀorts. These integrated production sites comprise a great variety

of production plants which are interconnected by product and energy ﬂows.21 A typi-

cal ﬂow sheet example for sites based on cracking of Naphtha is depicted in Figure 2.3.

Main products are the simple alkenes Ethylene, Propylene, and Butadiene. Beside these

pure alkenes, a fraction called pyrolysis gasoline (Pygas) is extracted which is a mix-

ture of acyclic and, mainly, cyclic hydrocarbons (aromatics) such as Xylene, Toluene,

and Benzene. Benzene is the most important cyclic hydrocarbon and raw material for

18

See e.g. Loos (1997, p. 48 ﬀ.) and references therein.

19

Low-volume products such as specialized pesticides are often produced in batch mode, see e.g. Loos

(1997, p. 70 ﬀ.).

20

Far more typical characteristics of production processes could be included, see e.g. Loos (1997, sec. 3.1).

However, the chosen characteristics are the most important with respect to logistical implications.

21

The head sites of BASF in Ludwigshafen and Dow Chemical in Midland are popular, large-scaled

examples of such integrated production sites.

11

e.g. Ethylbenzene which in turn can be de-hydrogenated to Styrene. Styrene can then be

used as raw material for Styrene-Butadiene rubber which is an important raw material

e.g. for the production of tires. Other branches of Benzene application are the production

of Aniline (which is widely used in polyurethane production) and Cumene which is a

composition of Benzene and Propylene and is mainly processed to resins.22

Naphta

cracking

Pygas

nitration

dehydrogenation hydrogenation

co-polymerization polymerization

Note that the provided ﬁnal products depicted in Figure 2.3 rather represent a set of sub-

products than a uniform substance. These sub-products have the same basic molecular

structure but can diﬀer e.g. in certain physical characteristics or colour.23

The vessels containing chemical reactions are called (chemical) reactors. Their design

and size depends on the intended reaction. The performance of a reactor is measured in

terms of the conversion rate, the purity of the substances of interest, and economic aspects

such as resistance, energy consumption etc. Basic types of reactors can be categorized

by diﬀerent characteristics.24 The vessel design is a ﬁrst category. Roughly, one can

distinguish between (stirred-)tank reactors and pipe reactors.

To eﬀectively execute chemical reactions, the involved reactants have to be provided in

suﬃcient purity. Moreover, many chemical reactions result not only in one pure output

product but in a mixture of output products. Thus, preparation and post-production

22

See e.g. Baerns et al. (2006, ch. 16) for details.

23

This can be obtained e.g. by various additives or process settings, see Behr et al. (2010).

24

E.g. see Baerns et al. (2006) or Trambouze and Euzen (2004) for more detailed classiﬁcation schemes.

12

processes have to be carried out to provide valuable reactants for subsequent chemical

reactions. These operations are called basic operations and can be categorized in thermic

and mechanic operations.25

The most prominent thermic basic operation is the rectiﬁcation or distillation to sepa-

rate individual components from a mixture.26 For separation, the mixture is evaporated

completely and successively cooled down. The components can be eﬀectively separated if

they have diﬀerent boiling points or dew points, respectively. In the case of close boiling

points for the components in the feed mixture, distillation is still applicable if an auxiliary

chemical (so-called solvent) is available changing the boiling or dew points of at least one

component (so-called fractional distillation).27

A distillation column is a metal tube which is separated in compartments by so-called

trays or plates. The mixture to be separated is fed into the middle of the column. In-

side the column temperature and pressure are variable depending on the height. At the

column’s top temperature is maximal and pressure minimal. Conversely, at the bottom

temperature is minimal and pressure maximal. Depending on the boiling points of the

mixture’s components, the composition of liquids and gases diﬀers in each compartment

and at each tray, respectively. Ideally, at each tray a single fraction/component of the

feed mixture can be obtained.28 To guarantee constant conditions regarding pressure and

temperature, a surplus of residue liquids at the bottom of the column is (re-)boiled and

fed back into the column (so-called reﬂux). Similarly, a surplus of gases at the top of col-

umn is condensed and fed back. Figure 2.4 shows a schematic overview of a prototypical

column.29

Separation processes are designed and optimized for a speciﬁc mixture to be separated

which determines e.g. the number and position of trays as well as the atmospheric con-

ditions. Such processes are single-purpose plants for SIMO or SISO30 processes. The

production rate often can be varied in certain ranges without causing serious variations

of the separation accuracy.

The distinction in reactions and basic operations to prepare and post-process chemical

reactions is valid in most cases. However, there is a principal advantage to integrate both

chemical process steps. Despite organizational and technical drawbacks, such an integra-

tion is physically and chemically advantageous due to more favourable energy balances.

One example is the reactive distillation where chemical reactions take place inside the

25

E.g. see Baerns et al. (2006) for details.

26

For a more general overview on (thermal) separation processes see e.g. Seader et al. (2011) or Baerns

et al. (2006, ch. 9).

27

E.g. see Behr et al. (2010, p. 88-89),Hoﬀman (1977) or Seader et al. (2011) for details.

28

Regarding the determination of the maximal or optimal number of trays a large body of literature has

been expanded and is still growing, e.g. see Seader et al. (2011) for an overview and Yeomans and

Grossmann (2000) or Viswanathan and Grossmann (1993) for more detailed insights.

29

In practice many adaptations and subtypes of this basic form have been derived depending on the

speciﬁc processes and circumstances the distillation column is intended for. See e.g. Smith (2005,

ch. 11 and 12) for an overview.

30

If only one component is considered as valuable.

13

condenser

gas

Prod. D

reﬂux

Prod. C

mixture/feed

trays

Prod. B

vapor

liquid

Prod. A

reboiler

distillation column.31 With the exception of such special techniques, chemical operations

can be categorized as displayed in Figure 2.5.

operations

Chemical production processes can be divided in chemical reactions and basic opera-

tions (i.e. physical transformations). In chemical production plants, multiple processes

from both classes are combined and take place in sub-plants which are closely intercon-

nected. The planning and conﬁguration of such plants is very complex and expensive.

Hence, a detailed modelling of the underlying chemical and physical processes is necessary

to avoid misinvestments. The next section outlines an overview on the steps necessary to

31

For an overview on reactive distillation see e.g. Baerns et al. (2006, p. 322 ﬀ.) and for details see

e.g. Taylor and Krishna (2000).

14

model and control a chemical production plant beginning with a brief introduction into

the mathematical modelling of chemical reactions.

A chemical production process is a combination of physical and chemical transformation

processes. The behaviour of such a transformation system can be described in mathemat-

ical terms. To describe the behaviour of a chemical reaction system two general questions

have to be answered:

• What are the requirements and outcomes of the intended reaction(s) regarding en-

ergy and reactants?

the intended reaction, whereas the answers to the second question are typically subsumed

under the term (chemical) kinetics.

Based on thermodynamical and kinetic descriptions of the individual process steps, a

meta-model can be developed which is able to describe and predict the behaviour of a

whole chemical production process. Such a process model can be developed for diﬀerent

purposes and at diﬀerent levels of detail: To design a chemical production process, a de-

tailed model of the potential plant(s) necessarily includes the description of the system’s

dynamics. In contrast, once the production process is designed, a model is necessary to de-

scribe the dependency of the system’s output w.r.t. certain control parameters. Figure 2.6

depicts a prototypical procedure in chemical process modelling.

Thermodynamics Kinetics

In the next subsection, kinetic deﬁnitions and concepts are introduced. Subsequently,

relevant properties of chemical operations w.r.t their modelling are outlined. Special

15

existing chemical production processes, subsumed under the term "system identiﬁcation".

The theory that addresses problems regarding the inter-temporal description of chemical

process parameters is subsumed under the term kinetics of chemical processes. The ki-

netics of chemical reactions include the temporal description of a reaction, i.e. the way

the concentrations/shares of reactants and products of a chemical reaction are developing

during the reaction. This requires the description of mass transport processes and the

description of heat/energy transport processes. Figure 2.7 displays this trisection and

introduces some keywords/methods explained below.32

chemical kinetics

concentration diﬀusion convection

Arrhenius eq. concentration conduction

In reaction kinetics, the object of interest is the progress of a chemical reaction measured

in terms of the change in concentration of the corresponding reactants and products.33 The

change of a component’s amount in time is deﬁned as the reaction rate of this component.34

Assuming constant volume this equals the concentration of this component. The reaction

rate rj of component j can be expressed as the change in concentration cj over time:

dcj

rj = . (2.1)

dt

The change in concentration typically depends on the reaction time, the concentration

of the other components, the temperature, and reaction speciﬁc properties. Temperature

32

In the strict sense the term kinetics only refers to reaction kinetics, i.e. the temporal development of

chemical reactions. Mass and energy transport processes are usually categorized as a pure physical

phenomenon. However, here a classiﬁcation scheme based on the topics’ time domain is chosen. For

an extensive introduction in chemical kinetics, see e.g. Levine (2005).

33

In the following, the term component subsumes reactants and products as well.

34

There exist also some other deﬁnitions of reaction rates based on reaction-speciﬁc characteristics,

e.g. volume, mass or activated catalyst surface, see e.g. Baerns et al. (2006, p. 59-60) or Behr et al.

(2010, p. 41-42).

16

where c is the vector of concentrations and T is the temperature. The temperature eﬀect

is usually assumed to be constant (also called reaction rate constant) and obtained by the

Arrhenius equality:

EA

f (T ) = k = k0 ⋅ e R⋅T (2.3)

where k0 is a reaction speciﬁc factor, EA is the energy necessary to start the reaction

and R is the universal gas constant.35 In the simplest case, the reaction rate depends

multiplicatively on the powers of the reactant’s concentrations:

rj = k ⋅ f (c) = k ⋅ cm

1 (t) ⋅ ... ⋅ cN (t)

1 mN

(2.4)

where mi is the reaction order of reactant i and ∑N i=1 mi is the order of the whole reaction.

36

The type of reaction orders depends on the structure of the reaction. Assume a simple

decomposition of one reactant, e.g. in cracking reactions of hydrocarbons in gasoline

production:

ν1 ⋅ A1 → ν2 ⋅ A2 + ... + νn ⋅ An (2.5)

where A1 is the (only) reactant and A2 , .., An are the products. The stoichiometric co-

eﬃcients νj refer to the number of molecules of chemical j consumed/produced during

the reaction. For all components their concentration depends on the concentration of the

reactant, i.e. for reactions of order m1 = 1 follows for the reaction rate of chemical j

rj = νj ⋅ k ⋅ c1 (t). (2.6)

If one molecule of A1 is split into the products, it follows ν1 = −1. For the change in

concentration of the reactant follows:

r1 = dc1

dt = −k ⋅ c1 (t) (2.7)

which is the simplest form of a homogeneous linear diﬀerential equation of order 1 with

the solution:

where cs,1 is the concentration of A1 at the beginning of the reaction. Here, (2.8) implies

an exponentially decaying concentration of the reactant over time.

35

E.g. see Behr et al. (2010, p. 42).

36

There are also other forms of reaction rate equations, such as hyperbolic relations. See Baerns et al.

(2006, p. 62 ﬀ).

17

in reaction networks that are much more diﬃcult to describe.37 An interesting elemen-

tary reaction is a 2-step sequential reaction which is relevant e.g. in modelling thermal

separation processes:

k1 k2

A1 → A2 → A3

reaction for both reactions, it follows that

dc1 (t)

r1 = = −k1 ⋅ c1 (t) (2.9)

dt

dc2 (t)

r3 = = k1 ⋅ c1 (t) − k2 ⋅ c2 (t) (2.10)

dt

dc3 (t)

r3 = = k2 ⋅ c2 (t). (2.11)

dt

This simple system of diﬀerential equations can be solved iteratively. The solution of

(2.9) is given in (2.8) and, hence, can be substituted in (2.10):

dc2 (t)

r2 = = k1 (cs,1 ⋅ e−k1 ⋅t ) − k2 ⋅ c2 (t). (2.12)

dt

Assuming k1 ≠ k2 and cs,2 = 0, (2.12) can be solved using standard results for homogeneous,

linear diﬀerential equations of ﬁrst order:

c2 (t) = (e −e ) (2.13)

k2 − k1

cs,1

c3 (t) = cs,1 + (k2 ⋅ e−k1 ⋅t − k1 ⋅ e−k2 ⋅t ) . (2.14)

k2 − k1

These examples of some simple types of chemical reactions show the general solution

methodology for determining the behaviour of chemical reactions over time. Especially

the modelling of catalytic and reversible reactions requires more sophisticated mathemat-

ical methods to determine the concentration rates.38 However, in the end more or less

complicated systems of diﬀerential equations have to be solved.

Another prominent thermodynamical relation often relevant in describing chemical pro-

cesses is the heat conduction formalized in Fourier’s law.39 This partial diﬀerential equa-

tion describes the diﬀusion of heat (energy) in time depending on (contact) area and

temperature gradient. This partial diﬀerential equation leads to some other prominent

37

For more detailed information about the deduction of reaction rates for other elementary reactions and

reaction networks e.g. see Levine (2005) or Baerns et al. (2006, ch. 4).

38

See e.g. Baerns et al. (2006, p. 65-72) or Levine (2005) for details.

39

See e.g. Behr et al. (2010), p. 49 ﬀ.

18

thermodynamical equations such as the heat and diﬀusion equation.40 The diﬀusion and

behaviour of energy is the main controller of most reactions. E.g. an energy-consuming

chemical reaction system is arranged such that the energy loss due to the chemical trans-

formation process is compensated by external energy supply (e.g. heating). Distillation

processes are typical examples for such processes.41

Another class of equations subsumes kinetic and phase transformations of all involved

reactants. Such equations describe how the reactants’ molecules are transformed and

distributed in the reactor depending on time and other environmental parameters. De-

pending on the kind of chemical processes under consideration, both classes of equations

are of varying importance for modelling. E.g. for catalytic packed-bed reactors the chem-

icals’ reaction rates heavily depend on local physical conditions at the (solid) catalyst

material. The precise modelling of the local physical conditions and the mixture of chem-

icals ﬂowing is important and complex in this case. In contrast, for classic stirred-tank

reactors kinetic and phase transformations are comparatively easy to model.

Most mass balances are deduced from the chemical transformation process directly,

whereas most energy equations are mainly deduced from thermodynamic relations. The

kinetics of a chemical reaction subsume relationships determining how the reaction rate

depends on conditions such as pressure, temperature and catalysts.

Combining all these thermodynamic and kinetic relations, theoretical mass and heat equa-

tions can be used to calculate (theoretical) reaction time and the size of a reactor to

generate a desired quantity of a certain product.42 This requires a model for the chemical

reaction quantifying the transformation rates in time depending on the concentration of

the involved reactants and catalysts as well as the physical parameters (e.g. pressure and

heat). There is usually a wide range of possibilities to process chemical reactions depend-

ing on the physical parameters. If a chemical reaction is isolated from the environment

(i.e. no energy transfer to the surroundings), this is called adiabatic. In contrast, if a

reaction is processed under constant physical conditions, it is called isothermal/isobaric

(i.e. usually heat has to be transferred to the environment).

Under isothermal/isobaric conditions the transformation processes of a chemical reac-

tion can be modelled depending on the behaviour of the concentrations in time given a

certain temperature.43 To get an idea, assume the simple splitting reaction

A1 → A2 + ... + An

40

There are a lot of standard textbooks regarding this topic see e.g. Sandler (2006), Koretsky (2004), or

Letcher (2004).

41

See e.g. Seader et al. (2011).

42

See Behr et al. (2010, p. 59 ﬀ.) or Baerns et al. (2006, p. 145 ﬀ.).

43

See e.g. Behr et al. (2010, p. 64-66).

19

A1 is given in (2.8). Now, deﬁne the turnover ratio X(t) = 1− c(t)

cs as the part of reactant A1

that is transformed at a given time t relative to the initial concentration at the beginning

of the reaction. Plugging in (2.8) yields

Assuming a constant volume of the reactants and processing the reaction in a stirred-

tank reactor one important question is to determine the time needed to achieve a desired

turnover rate. To answer this question, (2.15) is solved exactly to t = − k1 ln(1 − X) for a

given turnover ratio X. Normalizing this result by the reaction constant k leads to the

so-called Damköhler number for this speciﬁc class of reactions:

desired concentration/turnover rate. To calculate the Damköhler numbers for nth-order

reactions, the concentration diﬀerential equation can be derived directly by the following

steps:

dc1 (t)

= − k1 ⋅ cn1 (t) (2.17)

dt

d(1 − X)

⇔ = − k ⋅ (1 − X)n ⋅ cns,1 (2.18)

dt

dX

⇔ − = − k ⋅ (1 − X)n ⋅ cns,1 (2.19)

dt

1

⇔ dt = dX (2.20)

k ⋅ (1 − X)n ⋅ cns,1

1 XR 1

⇒ tR = n

cs,1 ⋅ k ∫0

dY (2.21)

(1 − Y )n

1 1 1

⇔ tR = ⋅( ( − 1)) (2.22)

cns,1 ⋅ k n − 1 (1 − X)n−1

For more complex reactions the derivation of Damköhler numbers is much more compli-

cated.44

Under adiabatic conditions heat ﬂows have to be modelled accordingly with respect to

the mass balances. For exothermic reactions usually a stimulation of the chemical reaction

is required due to the higher energy level at the beginning of the reaction. In general,

the interactions between heat and mass balances are more diﬃcult to model.45 More

complicated cases occur when both conditions are mixed, i.e. under polytropic conditions.

Here, solutions for heat and mass balances are hardly available analytically. Instead,

44

See e.g. Baerns et al. (2006, ch. 5.2) for more details.

45

E.g. see Behr et al. (2010, p. 67).

20

Given a speciﬁc reaction and type of reactor, it is usually possible to derive a theo-

retic concentration model which is able to anticipate the reaction’s behaviour over time

depending on the speciﬁc environmental conditions such as pressure, temperature, and

reactant inﬂow rates. But although such a theoretical approach for a chemical process can

be derived more or less easily, the relationships of the technical realization in terms of a

chemical plant are far more complicated due to the fact that a theoretical model necessar-

ily requires simpliﬁcations and assumptions. For example, a homogeneously distributed

mixture inside a reactor is often assumed which is rarely the case in a real reactor. This

typically leads to a stochastiﬁcation of theoretic relations, e.g. the diﬀerential equations

may turn into stochastic diﬀerential equations (SDEs). This turns the concentration as

calculated in (2.7) into a stochastic variable with associated distribution.46 For example,

(2.7) can be turned into

dc1 (t) = −k ⋅ c1 (t) ⋅ dt + σ ⋅ dW (2.23)

where dW denotes a stochastic process (the so-called Wiener process, which is described in

section 2.3) and σ is some variability parameter indicating the inﬂuence of the stochastic

component. Equation (2.23) constitutes a speciﬁc Ornstein-Uhlenbeck process.47 For this

speciﬁc class of SDEs an analytic solution can be derived as48

t

c1 (t) = cs,1 ⋅ e−k⋅t + σ ⋅ ∫ ek⋅(s−t) dW (s) (2.24)

0

which extends (2.8) by the addition of a stochastic term. Note that (2.23) can be solved

since the stochastic component is constant in time and independent from the concentration

process c1 (t). For most other types of SDEs or systems of SDEs analytic solutions are

not available.49

The design of chemical plants usually combines a bundle of chemical reactions and

basic operations, where the chemical reaction of interest is rather the core of a network

surrounded by preparation and follow-up operations. Predicting the behaviour of such

a network is diﬃcult and usually not manageable by analytic models even if theoreti-

cal models are available for all components. Therefore, simulation approaches are often

applied.50

To give an example, Figure 2.8 shows the schematic sketch of a typical steam cracker

splitting Naphtha (or similar hydrocarbons) into short-chained components.

The focal product is usually the C2 -fraction, known as Ethylene.52 The chemical process

46

See e.g. Behr et al. (2010, p. 68).

47

Here, the parameter μ = 0, see Uhlenbeck and Ornstein (1930).

48

See e.g. Hassler (2007).

49

See e.g. Hassler (2007).

50

See e.g. Behr et al. (2010).

51

See Behr et al. (2010, p. 179).

52

Ethene or Ethylen is the simplest alkene consisting of only two carbon atoms.

21

"#$

!

!

!

of splitting the feed’s long-chained molecules into smaller ones is processed in a pipe

reactor (located in a furnace to provide the necessary temperature). To separate the

produced mixture of short-chained chemicals a bundle of separation steps is processed in

a system of distillation columns.

Modelling the complete production process as a SIMO process requires the modelling of

the splitting process ﬁrst. This process can be characterized as a continuously processed

decomposition reaction in a pipe reactor.53 A model to describe the complete chemical

splitting process contains 680 single chemical reactions describing the dependencies of 37

substances.54 This decomposition describes the kinetic behaviour of the reaction depen-

dent on energy (heat) and pressure.55 Additionally, the environmental parameters have

to be modelled in terms of energy and mass equations.56 This model simulates the output

of the splitting process depending on certain control variables. In this case, mainly the

residence time of the feed in the pipe (between 0.1 and 2 seconds), the composition of the

feed, and the cracking temperature (between 800 and 900 °C) determine the composition

of the output mixture. Depending on the costs associated to energy and feed as well

as the value of the output mixtures, it is still an ongoing eﬀort to derive optimization

models maximising the proﬁt by ﬁnding the optimal control setting.57 However, not only

the composition of the cracking mixture but also the quality of the separation processes

is necessary to predict the outcome of the whole plant. For these components additional

models were developed.58

53

Classiﬁcation according to Figure 2.5.

54

See Edwin and Balchen (2001).

55

Details can be found in Sundaram and Froment (1978),Willems and Froment (1988a), and Willems

and Froment (1988b). The latter references gives information about the necessary activation energies

of the speciﬁc reactions.

56

Here speciﬁc models were developed, see e.g. Ghashghaee and Karimzadeh (2007) for a comprehensive

one.

57

E.g. see Edwin and Balchen (2001).

58

See e.g. Qi et al. (2002) for a reactive distillation model for C4 separation process or specialized

22

Hence, to model a chemical plant precisely requires a lot of eﬀort. This eﬀort is typically

expended when a new plant is planned. However, such a precise process model is not

necessary or useful for other purposes. In short-term control of chemical production plants,

so-called prediction control models are used which focus on the measurement of system

responses to certain changes of control variables on an empirical basis. These models use

basic relations of the theoretical models introduced above. The next subsection outlines

a brief overview of such models.

In general, the processes controlled in chemical industry are not hazardous but still dan-

gerous. Hence, the monitoring and control of chemical production plants is an important

topic in theory and practice. A process control scheme for chemical processes can be

depicted as shown in Figure 2.9.

controller chem. process

inputs

disturbances

estimator

estimates/model

The chemical process can be controlled by various variables which aﬀect physical con-

ditions as well as the product and energy inﬂows of the chemical plant. Output variables

comprise the physical properties of the plant’s outﬂows. In case of continuously oper-

ated plants, outputs are measured as ﬂow rates. Beside the output quantity, the output

quality e.g. in terms of the composition of the output mixture is an important measure.

To link inputs/controls with output measures, an estimator quantiﬁes the relationship

between both components. This estimator relies on a predeﬁned model representing the

theoretical background of the estimated process. This model is completed by estimated

parameters which are based on historical records of the process’ performance. Note that

the estimator has to be able to distinguish between undesired, uncontrollable disturbances

and desired previous control eﬀects. The resulting parametrized model can then be used

to predict the future behaviour of the process depending on the control settings. This

model is reported to the controller unit. The controller unit varies the control variables

to meet the desired target conﬁguration of the production process. The optimal sequence

of control variables may depend on economic parameters (e.g. input prices) resulting in

a simultaneous optimization of controls and outputs. This implies a dynamic process

59

Based on Darby et al. (2009).

23

control. Usually, however, predeﬁned static optimal target values for output measures are

to be met.60

Since this work deals with the aggregated simulation and planning of chemical produc-

tion processes, the focus is laid upon methods to determine estimations of the process

models. For process control this task is the crucial one as the estimations’ accuracy de-

termines the accuracy of the whole control process. The task to ﬁnd an accurate process

model is often called process identiﬁcation.61 To describe the input-output behaviour of

(continuously operated) chemical production plants ﬁnite impulse response (FIR) models

are widely used. These models can be seen as regression models where the historical

records of input/control measures determine the output measure. The term "ﬁnite" indi-

cates that a ﬁnite number of historical records is used to predict the process’ outputs.62

Often, chemical processes show a signiﬁcant time-dynamic behaviour which is typically

reﬂected in auto-correlated and cross-correlated process measures. However, classic re-

gression models do not incorporate auto-correlation explicitly which in turn leads to a loss

in estimation eﬃciency or, even worse, biased estimates.63 Therefore, time series meth-

ods can be applied to incorporate auto-correlation eﬀects. According to the classiﬁcation

shown in Table 2.1 four basic types of FIR models can be distinguished.

SISO model

For the simplest case of univariate input variables and univariate output measures the

following model describes a linear dependency between the input (or control) variable

x = (x1 , ..., xT )′ and the output variable y = (y1 , ..., yT )′

N

yt = μ + ∑ υb ⋅ B b xt + ξt (2.25)

b=0

where T denotes the time horizon and B denotes the so-called back-shift operator

B b xt = xt−b . The parameters υb reﬂect the (lagged) inﬂuence of the control variable

on the output variable. μ is the average output level. The noise vector ξ = (ξ1 , ..., ξT )′

contains the disturbances that cannot be explained by the model. Note that ξ can be

derived from some coloured noise model as well as from a classic white noise model.

White noise refers to a stochastic process with zero mean and no auto-correlation, while

coloured noise accounts for auto-correlation. If the process under study fulﬁls the white

noise assumptions, (2.25) can be interpreted as a simple regression model and can be

60

About the conception of model predictive control schemes see e.g. the brief overview in Darby et al.

(2009) or textbooks such as Camacho and Bordons (2004).

61

E.g. see the overview on chemical process modelling provided by Lewin et al. (2002) or textbooks about

the topic such as Ljung (1999).

62

This ﬁnite number of regressors can be chosen quite large which is indicated by the term "non-

parsimonious" FIR models, see Dayal and MacGregor (1996).

63

See the seminal papers Yule (1921) and Yule (1926) or some standard textbook on time series analysis,

e.g. Cryer and Chan (2008, pp. 260-265).

24

residuals, more elaborate time series models should be estimated.65 These are discussed

in the next section.

SIMO model

For multiple outputs the univariate control variable approach (2.25) has to be re-

formulated using matrix notation

N

yt = μ + ∑ υ b ⋅ B b xt + ξ t (2.26)

b=0

where yt = (y1,t , ..., yL,t )′ is the vector of L output variables, υ b = (υ1,b , ..., υL,b )′ the vector

of regression parameters, and ξ t = (ξ1,t , ..., ξL,t )′ the vector of output errors. Similarly,

Ξ = (ξ1,1 , ..., ξL,1 , ..., ξ1,T , ..., ξL,T ) is an L×T matrix which is derived from an L-dimensional

stochastic process either white or coloured. μ = (μ1 , ..., μL )′ denotes the vector of output

levels.

MISO model

The MISO model is used to describe converging production processes. Here, the structure

is similar to a SISO model, where the input is some (say) M -dimensional vector such that

N

yt = μ + ∑ υ ′b ⋅ B b xt + ξt (2.27)

b=0

with xt = (x1,t , ..., xM,t )′ the vector of control variables and υ b = (υ1,b , ..., υM,b )′ the vector

of regression parameters for lag b. For the noise term the same remarks hold as for the

SISO model.

MIMO model

The last and most general class of processes combines both previous models. Here, inputs

and outputs are multivariate leading to the following multivariate regression model

N

yt = μ + ∑ Υb ⋅ B b xt + ξ t (2.28)

b=0

with x = (x1,t , ..., xM,t )′ the vector of control variables, yt = (y1,t , ..., yL,t )′ the vector of

output variables, Υb = (υ 1,b , ..., υ L,b ) = (υ1,1,b , ..., υ1,L,b , ..., υM,L,b ) the matrix of regression

parameters for lag b, and ξt = (ξ1,t , ..., ξL,t )′ the vector of noise. Again Ξ is some L × T

matrix sampled from a coloured or white L-dimensional stochastic process.

64

E.g. see Dayal and MacGregor (1996) for more information.

65

Note that there is still the option to ignore the time-correlation in favour of increasing the input lag

N . It can be shown that this model yields the correct estimates asymptotically, see Ljung (1999).

However, the estimates based on ﬁnite samples may be poor, see Dayal and MacGregor (1996).

25

The models (2.25)-(2.28) can be interpreted as regression models if the corresponding noise

processes are white. Otherwise, the time-dependency structures of the processes have to

be taken into account to obtain accurate estimates of the desired regression coeﬃcients.66

Therefore, (2.25)-(2.28) can be transformed into standard time series models for which

established estimation procedures already exist. For the single output cases ((2.25) and

(2.27)), this results in so-called autoregressive, moving-average models with exogenous

regressors (or brieﬂy ARMAX models) whereas multiple output processes lead to so-called

vector ARMAX models (VARMAX).

The remainder of this section introduces the relevant notation with an additional focus

on the extension to heteroscedastic models (so-called (G)ARCH and ARMA-GARCH

models) as these can be seen as the discrete-time counterpart of continuous stochastic

processes formulated in terms of SDEs.

This section describes the class of the most common ARMA models and some of their

extensions. The term ARMA combines both basic types of time-dependencies, the auto-

regressive (AR) model and the moving average (MA) model. Suppose a time series y =

′

(y1 , .., yT ) collected over T periods with zero mean. Autoregressive dependency means

that any observation yt depends on previous observations yt−i of this time series with

i = 1, ..., p such that

p

yt = ∑ φi ⋅ yt−i + t (2.29)

i=1

with
t ∼ wn(0, σ 2 )67 for all t ∈ {p+1, ..., T }. This denotes an autoregressive process of order

p (in short AR(p)). Typically, the stochastic nature of the analysed process is assumed

to be stable in time. I.e. the stochastic characteristics (such as variance, autocorrelations

and mean) of the recorded process do not change in time. This is called the stationarity

premise which implies some restrictions for the parameters φi . E.g. suppose an AR(1)

process with φ = 1.68 It follows

t

yt = yt−1 +
t = yt−2 +
t−1 +
t = ... = ∑
i . (2.30)

i=0

66

Alternatively, the noise type can be ignored applying so-called non-parsimonious ﬁnite impulse-response

(FIR) models by increasing N , the number of lagged observations of the control variable(s). This

method has some serious drawbacks e.g. biased and instable estimates/forecasts in case of ﬁnite

samples. See e.g. Dayal and MacGregor (1996) and references therein.

67

The term wn(μ, σ 2 ) refers to a white noise process with zero mean and (constant) variance σ 2 .

68

This deﬁnes a so-called random walk.

26

t t

E(yt ) = E (∑
i ) = ∑ E(
i ) = 0 (2.31)

i=0 i=0

t t

V ar(yt ) = V ar (∑
i ) = ∑ V ar(
i ) = t ⋅ σ 2 (2.32)

i=0 i=0

t s t s

Cov(yt , ys ) = Cov (∑
i , ∑
i ) = ∑ ∑ Cov(
i ,
j ) = t ⋅ σ 2 with s ≥ t. (2.33)

i=1 i=1 i=1 j=1

Hence, this process is not stationary due to the time-varying, unbounded variance and

covariance.69 Only AR(1) processes with ∣φ∣ < 1 are stationary. This result can be gener-

alized for any AR(p) model for which the following conditions must hold:70

p

∑ φi < 1 (2.34)

i=1

∣φi ∣ < 1 i = 1, ..., p. (2.35)

In the moving average (MA) model yt depends on the previous errors
t−i instead of

previous observations. An MA model of order q (MA(q)) is represented by

q

yt = ∑ θj ⋅
t−j +
t (2.36)

j=1

with
t ∼ wn(0, σ 2 ) for all t ∈ {1, ..., T }. For MA processes no stationarity conditions have

to be considered, even an inﬁnite MA(∞) process remains stationary.71 MA processes

have another special characteristic: their non-uniqueness with respect to the (ﬁrst lag)

auto-correlation ρ. Suppose a time series generated by an MA(1) process with parameter

θ: yt = θ ⋅
t−1 +
t . The autocorrelation ρ = ρ (yt , yt−1 ) can be expressed as

ρ= . (2.37)

V ar(yt )

V ar(yt ) = (1 + θ2 ) ⋅ σ 2 , (2.38)

Cov (yt , yt−1 ) = θ ⋅ σ 2 , (2.39)

θ

ρ = . (2.40)

1 + θ2

69

The covariance expression can be understood intuitively as the sum of variances due to the error terms

commonly shared by both variables yt and ys . For more detailed deductions see Cryer and Chan

(2008, pp. 12-13) or Brockwell and Davis (2002, pp. 16-17).

70

These conditions are necessary but not suﬃcient. Precisely, for the characteristic equation φ(x) = 0 all

roots must have absolute value greater 1 (with the characteristic polynomial φ(x) = 1 − φ1 ⋅ x − φ2 ⋅ x2 −

... − φp ⋅ xp ), see Schlittgen and Streitberg (2004, pp. 121-132) or Cryer and Chan (2008, pp. 72-77).

71

See Cryer and Chan (2008, pp. 55-56).

72

See e.g. Cryer and Chan (2008).

27

However, setting θ′ = 1θ yields exactly the same auto-correlation ρ. Hence, there are at least

two possible values for θ producing exactly the same time series w.r.t. the auto-correlation

structure.73 This problem is related to the stationarity condition of AR processes. To solve

this problem the invertibility condition is introduced. An MA process must be invertible

into an inﬁnite AR process. This holds if and only if the characteristic equation for the

characteristic polynomial θ(x) = 1 + θ1 ⋅ x + θ2 ⋅ x2 + ... + θq ⋅ xq has roots with absolute value

larger than 1.74 Given a special MA process with known order and unknown parameter

(set), there exists only one parameter (set) such that this MA process is invertible.75

Both types of auto-correlation models are rarely found in real world problems in genuine

form, but in combination they build a huge class of time series patterns summarized as

so-called ARMA models. An ARMA(p,q) model can be formalized as:

p q

yt = ∑ φi ⋅ yt−i + ∑ θj ⋅
t−j +
t (2.41)

i=1 j=1

with
t ∼ wn(0, σ 2 ) for all t ∈ {1, ..., T }. To assure stationarity and invertibility the process

has to satisfy stationarity conditions from its AR part and invertibility conditions from

its MA part.76 Basic ARMA models only include ex ante stationary models but, by little

extensions, this class can be extended to comprehend also diﬀerence stationary time series.

Consider again a random walk model (which is clearly non-stationary):

yt = yt−1 + t . (2.42)

Taking the ﬁrst diﬀerences results in a stationary time series since the white noise process

remains

Δyt = yt − yt−1 =
t . (2.43)

The time series Δyt is stationary. Hence, the original series yt is called diﬀerence station-

ary.77 Diﬀerencing (i.e. the concept of taking diﬀerences) extends the ARMA class to the

most popular ARIMA class where ARIMA stands for autoregressive integrated moving

average. The concept of diﬀerencing can be extended to higher orders leading to complex

mathematical formulation in traditional notation.78 Therefore, the general ARIMA model

is formulated compactly as

φ(B)(1 − B)d yt = θ(B)
t (2.44)

where φ(B) and θ(B) denote the characteristic polynomials for the AR and MA part as

73

For details see Schlittgen and Streitberg (2004, p. 117).

74

See Schlittgen and Streitberg (2004, pp. 124-132).

75

See e.g. Cryer and Chan (2008, p. 80).

76

For details see Shumway and Stoﬀer (2006, pp. 93-97).

77

Alternatively, also the term integrated of order 1 is used. This term suggests that this concept can be

applied to diﬀerenced time series again, see e.g. Pfaﬀ (2006, pp. 23-24).

78

For more details about higher-order diﬀerencing see e.g. Shumway and Stoﬀer (2006, pp. 98-103) or

Cryer and Chan (2008, pp. 87-98).

28

described above.

A useful extension to classic ARIMA models are models incorporating exogenous ex-

planatory variables, so called ARIMAX models. As the name suggests, additional param-

eters are added to the ARIMA model. The use of explanatory variables usually results

from knowledge about the process’ external dependencies, but sometimes it is also reason-

able to model outliers or structural changes of the time series using auxiliary exogenous

variables.79

Assume the SISO model (2.25) where ξ is coloured noise and follows an ARMA process

with lag polynomials φξ (B) and θξ (B). Without loss of generality μ is assumed to be

zero and (2.25) changes to

N

φξ (B)yt = ∑ υb B b xt + θξ (B)
t (2.45)

b=0

models leads to some complications. Basically, there are two distinct cases: First, the

regression variable x = (x1 , ..., xT )′ is deterministic. In this case, a recursive estimation

procedure starting with a standard regression analysis can be applied leading to the

maximum likelihood estimates of the regression parameters under assumption of Gaussian

(white) noise.80

In the second case, the regression variable x is a (stochastic) time series, i.e. φx (B)xt =

θ (B)εt such that (2.45) changes to

x

N

θx (B)

φξ (B)yt = ∑ υb B b εt + θξ (B)
t (2.46)

b=0 φx (B)

where εt is a white noise series.

The ﬁrst approach to solve such a problem is to perform a regression analysis be-

tween the time series y and the regression variable x ignoring the fact of auto-correlated,

stochastic variables. Afterwards, the residuals from this ﬁrst step could be obtained and

a time series could be ﬁtted for these residuals, e.g. using ARIMA models. Unfortunately,

this approach can lead to biased and ineﬃcient estimates even if the sample is large.81

Due to the time series characteristics of the dependent and independent variables, the

cross-correlations between x and y might be overlaid by the individual temporal depen-

dencies of x and y. To obtain (reasonable) estimates for the impulse response parameters

′

υ = (υ0 , ..., υN ) , the time series y has to be "cleaned" from the time series eﬀects of the re-

gressor variables x. This procedure is called prewhitening and performed in the following

steps:

79

See Cryer and Chan (2008, ch. 11).

80

See Shumway and Stoﬀer (2006, p. 293 ﬀ.) for details.

81

See Cryer and Chan (2008, p. 40) and referred literature for details.

29

2. transform yt and xt by the estimates φ̂ and θ̂ : ỹt = y

θ̂x (B) t

and ε̂t = x

θ̂x (B) t

b=0 υ̂b B xt

In the case of the MISO model with multiple, say M , input variables the prewhitening

approach can be easily adapted by iteratively performing steps 1 and 2 of the prewhitening

procedure for all input variables to obtain estimates for υ i , i = 1, ..., M . The residual series

is ﬁnally calculated by ξˆt = yt − ∑N

b=0 υ̂1,b B x1,t − ... − ∑b=0 υ̂M,b B xM,t .

1 b NM b 83

Chemical processes can be modelled in detail as a bunch of equations and diﬀerential

equations based on chemical and physical laws. These laws rely on static assumptions

about the environment where the corresponding processes take place. In practice, all com-

ponents are inﬂuenced by stochastic factors that can inﬂuence the static process behaviour

and/or the dynamic process characteristics. Incorporating continuous stochastic processes

to a diﬀerential equation leads to stochastic diﬀerential equations. Linear stochastic dif-

ferential equations (SDEs) are usually formulated in the following general form using the

Wiener process W (t):

where W (t) is the Wiener process which is characterized by the following properties:84

• P (W (0) = 0) = 1,

• W (t) has normally distributed increments W (t) − W (s) ∼ N (0, t − s) with 0 ≤ s < t,

It can be shown that for a linear SDE with coeﬃcient functions ci (t) (which are con-

tinuous in t) an explicit, unique solution always exists and the SDE has ﬁnite ﬁrst and

second-order moments.85

Time series models and SDEs deal with the same sort of stochastic process. Both diﬀer

only in the domain of variables, which are either discrete or continuous. For instance,

chemical processes are continuous by nature. In practice, however, the condition of a

82

Details about prewhitening can be found e.g. in Box et al. (2008, p. 417 ﬀ.) or Cryer and Chan (2008,

p. 265 ﬀ.).

83

See e.g. Wei (1990, p. 328 ﬀ.).

84

See Hassler (2007, p. 117) and Iacus (2008, p. 18 ﬀ.)

85

See Hassler (2007) and Øksendal (2003).

30

chemical reaction is measured by various sensors in discrete time intervals. Hence, the

question arises whether discrete time series models can be seen as discretized SDEs.

It can be shown that ARCH and GARCH models are able to approximate stochastic

diﬀerential processes if the latter fulﬁl certain properties.86 Albeit the goodness of ﬁt

is limited,87 both types of methods are related and can be converted into each other.88

Moreover, simple stochastic processes show quite simple auto-correlation structures simi-

lar to basic ARMA models. For instance, the Ornstein-Uhlenbeck process can be seen as

the continuous equivalent of the AR(1) process. In other words, an Ornstein-Uhlenbeck

process measured in discrete intervals can be interpreted/modelled as an AR(1) process

(see also (2.23), (2.60), and (2.61)).89

This subsection brieﬂy introduces GARCH models as discrete counterpart of continuous

stochastic processes. In contrast to ARMA models, the basic idea is that the variance/

volatility in time is no longer deterministic and constant but depends on previous errors

and volatility, i.e.

y t = μ +
t (2.48)

σt2 = f (σt−1

2 2

, ..., σt−p ,
t−1 , ...,
t−q ) (2.49)

t ∼ wn(0, σt2 ). (2.50)

The model above is called a GARCH(p,q) model if f (⋅) is a linear function such that

(2.49) changes to

If the volatility depends only on the previous errors
i (i.e. q = 0), the corresponding

models are called ARCH(p). (G)ARCH processes are known to be weakly stationary if

q p

∑ αi + ∑ βi < 1.90

i=1 i=1

(G)ARCH models have become very popular in empirical economic research. In the

past decades a vast amount of extensions and generalizations have been published.91

These numerous adaptations of the basic model consisting of (2.48), (2.50) and (2.51)

are able to incorporate very speciﬁc time series characteristics and are usually driven by

empirical observations and experiences.92 A relevant extension in the context of chemical

86

See Nelson (1990) and Duan (1997).

87

See Wang (2002).

88

E.g. see Fornari and Mele (2001) for an application of GARCH models as diﬀusion approximations

ﬁtted to ﬁnancial data sets.

89

See e.g. Hassler (2007, p. 236) about the similarity between both processes and Uhlenbeck and Ornstein

(1930) or Vasicek (1977) for details about the Ornstein-Uhlenbeck process.

90

E.g. see Hassler (2007, p. 96). To assure stationarity for higher moments (>2) more restrictive condi-

tions have to be met, see Ling and McAleer (2002).

91

See Engle (2002) for a review.

92

E.g. see Bauwens et al. (2006) for a review of recent multivariate GARCH models and Bollerslev (2008),

Hansen and Lunde (2005) or Degiannakis and Xekalaki (2004) for a review of univariate (G)ARCH

31

production processes are so-called GARCH-M and ARMA-GARCH models. For this class

of models the mean process in (2.48) is not a constant, but a stochastic process itself.

The GARCH-M(p,q) mean process is modelled as function of σ:93

yt = μ ⋅ σt + t . (2.52)

GARCH(P ,Q)(p,q) the mean process follows an ARMA(P ,Q) process whereas the volatil-

ity is modelled by a GARCH(p,q) process:

P Q

yt = ∑ φi ⋅ yt−i + ∑ θj ⋅
t−j +
t (2.53)

i=1 j=1

2

+ ... + βp σt−p

2

+ (2.54)

t ∼ wn(0, σt2 ). (2.55)

This class of models allows formulating a great variety of dependency patterns and is used

in many empirical applications.94 Its most appealing advantage is the possibility to model

both a time-dependency structure for the observed series and for their volatility combined

in one model. For modelling chemical production processes, the latter property is useful

for describing the variations in a chemical process’ equilibrium. In normal operation,

chemical reactions take place under controlled physical and chemical conditions. The

parameters of these reactions such as conversion rates and reaction times depend on

these conditions and determine the stochastic processes of output and input measures.

Variations in process control variables aﬀect these conditions and have an immediate

impact on the parameters of the underlying reactions. This may change the structures

of the input and output processes. As a consequence, a destabilization of the chemical

process might occur resulting in an increase in the measures’ volatility.

The incorporation of exogenous regressors (leading to ARMAX-GARCH models) is less

emphasized in the literature since it is similar to ARMAX models. In this case (2.53) is

substituted by (2.45). For estimation of the corresponding parameters, prewhitening can

be applied to eliminate the inﬂuence of the exogenous time series variables. Afterwards,

the ARMA-GARCH model can be estimated on the residuals.95

The methodology for multivariate time series models is similar to the univariate cases

except for the fact that all notation is changed into vectors and matrices such that most

models.

93

See Hassler (2007).

94

See Li et al. (2002) for an extensive review and references therein for applications.

95

It has to be noted that the estimation of ARMA-GARCH models requires some more sophisticated

methods compared to simple ARMA models, see Francq and Zakoian (2004).

32

concepts can be transferred more or less easily. This section brieﬂy introduces the key

models for multivariate time series models, so-called vector autoregressive moving average

processes (VARMA) and their extensions with exogenous regressors (VARMAX). Multi-

variate GARCH processes (MGARCH) are not discussed in detail here.96

Suppose a multivariate time series of dimension L collected at T points of time, say

Y ∈ RL×T . The basic types of dependency, auto-regression and moving average, are used to

formulate a model explaining the behaviour of this multivariate time series. Suppose that

the observations of Y at a given time t, say yt ∈ RL , depend on the previous observations

yt−i with i = 1, ..., p and previous error terms, say t−j with j = 1, ..., q. This relation can

be formulated as a VARMA(p,q) model

p q

yt = ∑ Φi yt−i + ∑ Θj t−j + t (2.56)

i=1 j=1

p q

⇔ yt = ∑ Φi B i yt + ∑ Θj B j t + t (2.57)

i=1 j=1

Θq (B) = I + Θ1 B + ... + Θq B q denote the corresponding lag polynomials. It is required

that all matrices are non-singular and the covariance matrix of t is positive deﬁnite.

Similar to the univariate case, the process is stationary and invertible if all roots of the

determinantal polynomial of ∣Φp (B)∣ and ∣Θq (B)∣ are outside of the unit circle.97

Model (2.58) can easily be extended to a VARMAX model by incorporating exogenous

variables. Let X ∈ RN ×T be the matrix of explanatory variables, then (2.58) is reformulated

to

Φp (B)yt = μ + ΥN (B)xt + Θq (B)t (2.59)

model is equal to the MIMO model (2.28) if Ξ is assumed to be a VARMA process of

order (p, q).98 If N = 1, (2.59) reduces to the SIMO model (2.26).99 For the estimation

of VARMAX models integrated estimation procedures are available.100 The problem of

time series being used as exogenous variables does not pose a direct problem since these

variables can be handled as endogenous101 such that their dependency structure is simul-

taneously estimated. However, such a procedure complicates the model and increases the

96

For an introduction see e.g. Bauwens et al. (2006).

97

About the stationarity and invertibility conditions see e.g. Wei (1990, pp. 335 ﬀ.).

98

For an application of an MIMO model in chemical process modelling see Barceló et al. (2011).

99

For more details about the relationship between multivariate ﬁnite impulse response models and (par-

simonious) VARMAX models in chemical process analysis, see e.g. Seppala et al. (2002).

100

See e.g. Hall and Nicholls (1980) for a direct maximum likelihood estimation or Metaxoglou and Smith

(2007) for estimation based on a transformation into state-space models.

101

I.e. these variables can be included in Y.

33

number of parameters to be estimated and, hence, may hamper the model’s feasibility.102

If exogeneity can be assumed (e.g. in case of control variables), a direct modelling of the

exogenous variables’ stochastic models is the preferred option. Similar to the prewhiten-

ing approach, a time series model for the exogenous variables has to be ﬁtted which can

be used to improve the estimates of the original VARMA model.103

The previous section points out which structure of a time series model to choose depending

on the general type of the production plant to be modelled. Once a model class is selected,

the corresponding model’s parameters have to speciﬁed based on historical records of the

investigated plant. The model ﬁtted to the historical records has then to be checked for

adequacy, i.e. the residuals of the model have to be checked for the assumptions inherent

to the chosen model (typically white noise).

In a ﬁrst step, data describing the input and output ﬂows has to be acquired. In

basic chemical industry, most inputs and outputs are ﬂow rates recorded in discrete time

intervals. For short-term plant control, these intervals are very short.104 For the purpose

of simulation such a short interval is not in all cases advantageous as it determines the

granularity of the subsequent simulation model and, hence, its complexity. Moreover,

the time-dependency structure is immediately aﬀected by the choice of the recording

interval. In principle, the shorter the recording intervals are, the larger the order of the

process is to be expected.105 In the other extreme, a too long recording interval will not

show signiﬁcant time-dependencies most probably as short-term dependencies diminish

or cancel out by summation/averaging.

To give an example for the inﬂuence of the length of the recording interval, let y(t)

denote the value of an Ornstein-Uhlenbeck process as introduced in (2.23) at time t. For

this speciﬁc process the autocorrelation ρ for t → ∞ can be expressed as106

Let Δt denote the recording interval and let h = m ⋅ Δt be a multiple of Δt. Then

yt+m = y(t + m ⋅ Δt) is the discrete equivalent of y(t + h). For the autocorrelation of two

measures holds

m

ρ(y(t), y(t + m ⋅ Δt)) = ρ(yt , yt+m ) = e−k⋅m⋅Δt = (e−k⋅Δt ) = φm . (2.61)

Note that (2.61) represents the autocorrelation function of an AR(1) process with pa-

102

See e.g. Horváth (2003).

103

In the state-space reformulation of the VARMAX model, the time series model of the exogenous

variables is used to predict initial values of the regressor matrix, see Metaxoglou and Smith (2007).

104

Intervals about or shorter than one second are frequently found, see Darby et al. (2009).

105

I.e. the parameters p and q in the ARIMA models introduced before, see e.g. (2.41).

106

For the underlying expressions for asymptotic variance and autocovariance, see Hassler (2007, p. 237).

34

words, although the type of the discrete time series process remains unchanged, the cor-

responding parameter changes with Δt. From an empirical point of view, analysing the

process structure by means of the empirical autocorrelation of a ﬁnite sample will not

allow detecting this pattern if Δt is too large (since empirical autocorrelations will be too

small to be distinguished from noise). A reasonable choice with respect to the use of time

series models within the simulation environment needs to be deﬁned. From numerous

experiments with samples from multiple chemical plants, an interval length of 30 minutes

up to multiple hours is appropriate in most cases.

In principle, the measured ﬂow rates may vary either due to ﬂuctuations of the underly-

ing production process or due to measurement errors. For the analysis of the production

system, one is interested in the former variation only. However, the latter cannot be

excluded since there is no inﬁnitely precise and reliable sensor. Flow measures can be

aﬀected by turbulences or pollutions. Hence, implicitly it has to be assumed that the

sensors’ measures of the corresponding ﬂow rates are unbiased and that the measurement

error does not dominate the error process of the production process. Since these sen-

sors are also used for the automatic control systems, continuous eﬀorts are expended for

checking their accurateness. Hence, it can be assumed that the sensors work properly

in the sense that they are unbiased and have measurement errors that are as small as

(technically) possible.107

Once the data of a chemical production plant is collected, the basic type of model is

speciﬁed, i.e. SISO, SIMO, MISO or MIMO. When deciding on the basic model type the

number of relevant measures has to be determined. A lot of variables may aﬀect the

performance of a chemical production plant (e.g. product ﬂows, atmospheric conditions,

energy ﬂows). Among these, the relevant variables need to be extracted. Relevance refers

to the use of time series models within the simulation environment and prerequisites to

build an appropriate model of the production process. For the ﬁnal simulation model,

main chemicals (raw, intermediate, and ﬁnal chemicals) of the studied production system

are ﬁxed parts of the time series models. From the remaining variables (such as energy

ﬂows or auxiliary chemical ﬂows), variables are included which yield a relevant improve-

ment of the accuracy of the ﬁnal time series model. If a variable cannot improve the ﬁnal

model’s accuracy, it should be dropped from the analysis to avoid over-speciﬁcation.108

To decide on a model’s accuracy, various criteria are available quantifying the model’s

goodness of ﬁt. These information criteria (IC) are based on the model’s likelihood

typically assuming a Gaussian error process.109 The very most information criteria can

107

However, there is typically no possibility to check these assumptions explicitly.

108

In case of over-speciﬁcation (synonymously over-ﬁtting) additional variables or parameters are incor-

porated in the model which show a spurious correlation with the error terms. This may lead to less

eﬃciently estimated parameters and poor prediction accuracy, see Cryer and Chan (2008, pp. 185-188)

or Schlittgen and Streitberg (2004, pp. 347-349).

109

See e.g. Schlittgen and Streitberg (2004, pp. 332-345).

35

be expressed as

K

IC = −2 ⋅ ln (LT (K)) +

⋅ κ(T ) (2.62)

T

where K denotes the number of parameters ﬁtted in the model, κ(T ) is a function of the

number of observations T (this function distinguishes the various criteria), and LT (K)

denotes the model’s likelihood. The number K of parameters to be estimated depends

on both, the order of the model and the number of exogenous variables. For example,

assuming a VARX(p) of dimension d with non-zero mean and k = 1 exogenous variable

(which has a contemporary eﬀect only), the number of parameters to be estimated is

K = p ⋅ d2 + d + d.

From (2.62) it can be taken that the model’s goodness of ﬁt and the number of parame-

ters used are counterbalanced. Among a set of model speciﬁcations, the speciﬁcation with

minimal IC value is recommended. In other words, information criteria aim at minimiz-

ing the residuals’ variance with as few parameters as possible. Often used information

criteria for time series models are the Akaike information criterion (AIC),110 the Schwarz

information criterion (SIC)111 or the Hannan-Quinn information criterion (HQIC)112 with

the following κ functions:

⎧

⎪

⎪

⎪ 2 for AIC

⎪

⎪

κ(T ) = ⎨ ln (T ) for SIC . (2.63)

⎪

⎪

⎪

⎪

⎩ ln (ln (T ))

⎪ for HQIC

Among these IC, the AIC has a non-zero asymptotic probability for over-ﬁtting.113 To

avoid an over-ﬁtting of the model order, a corrected AIC (AICc ) can be calculated by

AICc = AIC + 2⋅(K+1)⋅(K+2)

T −K−2 .114

The SIC is deduced from Bayesian arguments. It consistently estimates the true order

of ARMA(p, q) processes and is probably the most widely used information criterion in

univariate time series analysis.115 The HQIC is the most recent IC and especially designed

for multivariate time series models.116 In practice, multiple ICs are simultaneously cal-

culated which allows the analyst to cross-check the recommendations of the various ICs.

Strongly deviating recommendations may indicate an inappropriate model structure.

One source of mis-speciﬁcation is the type of the error process. Homogeneity of vari-

ances is typically assumed by standard time series models. However, in the context of

chemical production processes e.g. an adjustment of a plant’s production rate may lead

to an imbalance of the underlying chemical reaction(s). This instability may materialize

in ﬂuctuations of the output ﬂow rate(s) which may also aﬀect ﬂow rates in subsequent

110

See Akaike (1974).

111

This criterion is often called Bayesian information criterion (BIC), see Schwarz (1978).

112

See Hannan and Quinn (1979).

113

In other words it is biased, see Cryer and Chan (2008, p. 131).

114

See Hurvich and Tsai (1989).

115

See Cryer and Chan (2008, p. 131).

116

See Quinn (1980).

36

model for the error component of the model, e.g. a GARCH model. The presence of

time-varying variances can be uncovered by checking the squared and absolute residuals

of an already ﬁtted time series model for auto-correlation.

To decide on the order of time series models as well as to check residuals for the white

noise assumptions, auto-correlations of time series and residuals need to be calculated and

analysed. Standard metrics to analyse for time-dependent correlation structures are the

autocorrelation function (ACF), partial ACF (PACF) and extended ACF (EACF). The

ACF estimates the empirical auto-correlations between lagged observations

Corr (yt , yt−k ) = = ρk (2.64)

V ar(Y )

of an empirical time series by the standard sample correlation formula.117 These values

are usually displayed in a correlogram. If the auto-correlation for a speciﬁc lag exceeds

some critical value,118 a signiﬁcant auto-correlation is to be assumed. Hence, the ACF

allows detecting the order of pure MA processes which can be easily seen from the formal

deﬁnition of MA processes in (2.36). I.e. the ACF of an MA(q) process theoretically shows

signiﬁcant values for the ﬁrst q auto-correlations.119

To detect AR-like correlations, the PACF corrects the ACF values by the eﬀect that is

caused by preliminary auto-correlations120

cast) is applied, the so-called Durbin-Levinson-algorithm.121 Again, if the PACF value

for a speciﬁc lag exceeds a critical bound,122 a signiﬁcant AR-like correlation is to be

presumed.

Unfortunately, neither PACF nor ACF lead to directly interpretable results for ARMA

processes. The extended ACF tries to overcome this drawback by jointly providing infor-

mation about the order of both components. For each AR order tested, the EACF ﬁrst

determines estimates of the AR coeﬃcients by a sequence of regression models. After-

wards, the residuals’ ACF is calculated. The results are presented in a table indicating

signiﬁcant or non-signiﬁcant auto-correlations (typically denoted by an x and o, respec-

tively). In such a table, the rows represent the AR order p whereas columns represent

117

E.g. see Cryer and Chan (2008, pp. 109-112).

118

For details about this bound, see e.g. Cryer and Chan (2008, p. 112).

119

Contrarily, pure AR processes show an exponentially decaying ACF. For a more precise description see

e.g. Cryer and Chan (2008, pp. 109-112).

120

For details see e.g. Schlittgen and Streitberg (2004, pp. 194-201), Cryer and Chan (2008, pp. 112-115)

or Brockwell and Davis (2002, pp. 94-96).

121

For details see Schlittgen and Streitberg (2004, pp. 196-198).

122

This bound is deduced by assuming a Gaussian√AR(p) process. Then, sample auto-correlations are

approximately normal and the critical value is 2/T , see Cryer and Chan (2008, p. 115).

37

the MA order q. To determine a time series’ ARMA order, the coordinates of the upper

left vertex of a triangle of os correspond to the order (p, q) of the underlying ARMA

process.123

ACF, PACF and EACF are applied to time series data to initially estimate the type and

order of the underlying time series model. Once an initial model is ﬁtted, ACF, PACF and

EACF are applied to the corresponding residuals to check the white noise assumptions. If

signiﬁcant auto-correlations are found, the model type or order is adjusted accordingly and

the procedure is repeated until the residuals do not show any serious auto-correlations.

A normality test such as Jarque-Bera’s test or graphical tools such as QQ-plots are also

part of the standard procedure of residual checks. Although normality is often assumed to

derive arguments for theoretical deductions, this is not a necessary condition for consistent

estimates.124 Nonetheless, if normality can be assumed, an identical distribution of the

residuals and variance homogeneity follows immediately. However, if normality cannot be

conﬁrmed, the weaker concept of identical variances can be checked by calculating ACF,

PACF and EACF for the squared and absolute residuals of a model.125 If noticeable

correlations occur, heterogeneous variances must be presumed and the time series model

should be extended by a sub-model for the variances (i.e. an ARCH or GARCH model).

The following examples illustrate how the above mentioned methods can be used to

ﬁnd appropriate time series models for empirical samples. The ﬁrst example shows the

application of the methods by means of a simulated univariate time series. This example

emphasizes the eﬀect of the length of time series and gives an impression how to interpret

the described graphical tools. The second example describes the analysis of a sample of

the input ﬂow rate of a cracker plant. This example focuses on practical problems (such as

outliers) in analysing even simple univariate time series of chemical production processes.

The third example shows how the transformation process of a chemical production plant

can be modelled using multivariate time series methodology based on a real-world sample

from a de-alkylation plant. Again practical problems such as the determination of the

model order and multivariate outliers are discussed.

123

For details about the deduction of the EACF test statistic, see Tsay and Tiao (1984) and for a practical

overview see e.g. Cryer and Chan (2008, pp. 116-118) or Tsay (2002, pp. 51-53).

124

Consistent estimates can be obtained for independent and identically distributed residuals with quasi-

maximum likelihood estimators and Gaussian MLE, see Yao and Brockwell (2006).

125

See Cryer and Chan (2008, pp. 277-280).

38

model with sizes T = 100 and T = 1000 (assuming Gaussian errors ∼ N (0, 4) and zero

mean).126 The parameters are φ = (0.7, −0.2) for the AR part and θ = (0.4) for the MA

part. The empirical ACF and PACF are calculated for both samples next to the model’s

theoretical ACF and PACF. The corresponding correlograms are depicted in Figures 2.10a

- 2.10f.

1.0

1.0

1.0

0.5

0.5

0.5

ACF

ACF

ACF

0.0

0.0

0.0

−0.5

−0.5

−0.5

5 10 15 20 5 10 15 20 5 10 15 20

Lag Lag Lag

(a) ACF for T = 100 (b) ACF for T = 1000 (c) Theoretical ACF

1.0

1.0

1.0

0.5

0.5

0.5

Partial ACF

Partial ACF

partial ACF

0.0

0.0

0.0

−0.5

−0.5

−0.5

5 10 15 20 0 5 10 15 20 25 30 5 10 15 20

Lag Lag Lag

(d) PACF for T = 100 (e) PACF for T = 1000 (f) Theoretical PACF

Figure 2.10: Theoretical and empirical ACF and PACF for an ARMA(2,1) process

The number of observations available for estimation considerably inﬂuences the em-

pirical ACF and PACF. For the small sample with T = 100 the ACF correlogram (Fig-

ure 2.10a) does not show any similarity to the corresponding theoretical ACF correlogram

(Figure 2.10c). None of the ACF correlograms shows an easily interpretable pattern. In

contrast, for all PACFs (Figures 2.10d - 2.10f) the ﬁrst two lags show the highest cor-

relation pointing to a second order AR component. Tables 2.3 - 2.5 show the results of

the sample EACFs and the theoretical EACF for this example ("x" denotes a signiﬁcant

correlation and "o" denotes a non-signiﬁcant one).

For the small sample, the EACF values (Table 2.3) show an ambiguous pattern. Proba-

bly, an ARMA(1,3) or ARMA(3,1) model would be chosen as initial choices. In contrast,

the large sample (Table 2.4) shows a good ﬁt to the theoretical EACF (Table 2.5) and an

ARMA(2,1) model seems an appropriate initial choice. Since the analysis of ACF, PACF

and EACF cannot reveal an unambiguous favourite model for both samples, all possible

order speciﬁcations with p ∈ {0, ..., 4} and q ∈ {0, ..., 4} are estimated for both samples.127

The corresponding AICc , SIC, and HQIC values are displayed in Table 2.6 where the min-

imum for each criterion and time series is set in bold font and the true model order is

coloured in grey.128

126

The samples are simulated by means of the arima.sim function with R 2.15.1.

127

For all models no intercept is estimated.

128

Only a subset of all tested orders is displayed in the table for the sake of brevity (q ≤ 1).

39

0 1 2 3 4 5 6 0 1 2 3 4 5 6 0 1 2 3 4 5 6

0 x x o x x x o 0 x x o x x o o 0 x x x x x x x

1 x x o o o x o 1 x x x x o o o 1 x x x x x x x

AR lag

AR lag

AR lag

2 o x o x o o o 2 x o o o o o o 2 x o o o o o o

3 x o o o o o o 3 x x o o o o o 3 x x o o o o o

4 x x o o x o o 4 x x x o o o o 4 x x x o o o o

5 x x x o o o o 5 x o o x o o o 5 x x x x o o o

6 x x o x o o o 6 x o o x o o o 6 x x x x x o o

Table 2.3: EACF for T = Table 2.4: EACF for T = Table 2.5: theoretical

100 1000 EACF

AR MA T = 100 T = 1000

order order AICc SIC HQIC AICc SIC HQIC

1 0 4.273 4.175 4.160 4.486 4.479 4.476

2 0 4.141 3.943 3.912 4.272 4.258 4.252

3 0 4.330 3.987 3.941 4.272 4.246 4.237

4 0 4.556 4.022 3.960 4.291 4.250 4.238

0 1 4.258 4.160 4.145 4.463 4.456 4.453

1 1 4.197 3.999 3.968 4.277 4.262 4.256

2 1 4.330 3.987 3.940 4.269 4.244 4.234

3 1 4.566 4.032 3.970 4.291 4.250 4.238

4 1 4.835 4.062 3.984 4.317 4.257 4.242

Table 2.6: Information criteria for both sampled time series and various model speciﬁca-

tions

For both samples all three criteria receive their minimum for a unique model order. In

the small sample case, all criteria suggest an ARMA(2,0) whereas in the large sample

case an ARMA(2,1) is suggested (which is the correct order). Fitting these models to both

time series and analysing the corresponding residuals results in diagnostic plots without

remarkable features as Figures 2.11a - 2.11f show.

For the residuals’ ACFs only spurious signiﬁcant auto-correlations occur at lag 9 for

the small sample and lag 30 for the large sample. Hence, it can be concluded that no auto-

correlation remains and the model successfully captures the time-dependency pattern.129

For the squared residuals’ ACFs, slightly signiﬁcant auto-correlations occur at lag 11 for

the small sample and at lags 6, 17, and 26 for the large sample. Therefore, also identical

variances of the residuals can be assumed.130 In summary, identically and independently

distributed residuals can be assumed for both models. This is also conﬁrmed by the QQ-

plots (Figure 2.11c and Figure 2.11f) which show a good ﬁt to the normal distribution for

both models. This conﬁrms the assumption of independently and identically distributed

residuals indicating adequate models. Note that for the small sample the true model order

129

Since the ACFs show no attractions, the PACF need not to be analysed.

130

Also the ACFs of absolute residuals show no serious attractions.

40

is not identiﬁed despite the fact that the residual analysis reveals no hints for an inaccurate

model speciﬁcation.

0.4

0.4

4

2

0.2

0.2

0

0.0

0.0

ACF

ACF

−2

Sample Quantiles

−0.2

−0.2

−4

−0.4

−0.4

5 10 15 20 5 10 15 20 −2 −1 0 1 2

Lag Lag Theoretical Quantiles

(a) ACF of residuals for T = 100 (b) ACF of squared residuals for T = 100 (c) QQ-plot of residuals for T = 100

0.4

0.4

5

0.2

0.2

0

0.0

0.0

ACF

ACF

Sample Quantiles

−0.2

−0.2

−5

−0.4

−0.4

5 10 15 20 25 30 5 10 15 20 25 30 −3 −2 −1 0 1 2 3

Lag Lag Theoretical Quantiles

(d) ACF of residuals for T = 1000 (e) ACF of squared residuals for T = 1000 (f) QQ-plot of residuals for T = 1000

Figure 2.11: ACF of residuals, ACF of squared residuals and QQ-plot of residuals for ARMA(2,0) model (small sample, T = 100) and for

ARMA(2,1) model (large sample, T = 1000)

41

42

Example 2 (Univariate time series analysis for chemical production data). An application

of univariate time series analysis is the modelling of univariate input ﬂows of chemical

production processes (as observed for SISO or SIMO processes). To apply prewhitening

procedures, the auto-correlation structure of the (exogenous) time series of the input ﬂow

has to be modelled. In this example, the input ﬂow rate of a Naphtha cracker is analysed.131

Figure 2.12 shows the average hourly input ﬂow rate for one week (T = 168).

41.62

input flow rate in t/hour

41.60

41.58

41.56

0 50 100 150

Time

The time series shows a stationary pattern with a homogeneously oscillating ﬂow rate

around a constant level. However, an additive outlier occurs at time index 128. To

analyse the time dependency pattern of the Naphtha time series, the eﬀect of the outlier

has to be removed. Therefore, the outlier’s eﬀect (i.e. the raise beyond the mean level)

is estimated by modelling a simple linear regression model and the corresponding value of

the time series is corrected by this estimate. For the corrected Naphtha time series, ACF

and PACF are calculated (Figure 2.13a and Figure 2.13b).

The ACF shows a decaying cyclic pattern whereas the PACF clearly indicates a signiﬁ-

cant partial auto-correlation at lag 1. Therefore, an AR(1) model is chosen as the initial

model. To incorporate the outlier at index 128, an ARX(1) model is ﬁtted to the original

Naphtha time series where a dummy variable xt accounts for the eﬀect of the outlier132

yt = μ + φ ⋅ yt−1 + β ⋅ xt + t . (2.66)

The estimated parameters of the ARX(1) model are displayed in Table 2.7.

The corresponding residuals’ standard deviation is estimated as σ̂ = 0.002 and the

model’s AICc is calculated as -8.73. An analysis of the residuals reveals no hints for a vio-

lation of the underlying assumption of independent and identically distributed residuals.133

131

Data is provided by Dow Chemical. The scale of the time series has been transformed due to reasons

of privacy protection.

132

The dummy variable is binary, i.e. xt = 1 for t = 128 and zero otherwise.

133

Diagnostic plots can be found in the Appendix in Figure A.1.

43

1.0

0.8

0.6

ACF

0.4

0.2

0.0

0 5 10 15 20

Lag

(a) ACF of corrected Naphtha time series

0.8

0.6

Partial ACF

0.4

0.2

0.0

5 10 15 20

Lag

(b) PACF of corrected Naphtha time series

Jarque-Bera’s test for normality conﬁrms that the hypothesis of normally distributed resid-

uals cannot be rejected (at a 5% level), i.e. it is assumed that
ˆ ∼ N (0, 0.0022 ). Due to

the already convincing results of the ARX(1) model (which has minimal order) no other

ARMA models of higher order are estimated and (2.66) is taken as the best model for the

Naphtha time series.

parameter μ̂ φ̂ β̂

estimate 41.6 0.79 0.0564

standard error 0.001 0.046 0.002

44

carbons are one of the major chemicals resulting from cracking of Naphtha. This class of

chemicals comprises Benzene, Toluene and Xylene. For a post-processing to ﬁnal chemi-

cals, Benzene is the most important aromatic hydrocarbon. By de-alkylation, Toluene and

Xylene can be transformed to Benzene. The de-alkylation of Toluene and Xylene is a cat-

alytic exothermal reaction where the alkyl-group(s) in the Toluene and Xylene molecules

are removed by reacting with hydrogen to Benzene and Methane. Methane can be reduced

to hydrogen and carbon black, where the hydrogen can in turn be used as input for the de-

alkylation.134 The conversion rate of this transformation is about 60 - 90%.135 To operate

eﬃciently, not transformed Xylene and Toluene is separated from Benzene and cycled

back to the input mixture. The stream of input chemicals also contains longer-chained

hydrocarbons (C9 ) resulting from the previous cracking process and imperfect separation.

These hydrocarbons are also contained in the output ﬂow and have to be separated from

Benzene. The ﬂow sheet of the chemicals involved in this production process is depicted

in Figure 2.14.

Toluene + Xylene

Methane

Benzene

H2

inﬂow mixture

All chemicals entering or leaving this plant are recorded. Figure 2.15a shows the average

hourly inﬂow and outﬂow rates for two weeks (T = 336).137 Although hydrogen is a

necessary component to undergo the relevant chemical reactions, the external inﬂow of

hydrogen has shown no individual eﬀect on the outﬂow rate. Thus, external hydrogen

supply has been dropped from the further analysis. The outputs are pure Benzene and the

134

See e.g. Ma and Trimm (1996).

135

See Ozokwelu (2006).

136

See e.g. Ozokwelu (2006).

137

This data is provided by Dow Chemical. The scales of all variables are transformed due to privacy

protection reasons.

45

rate and outﬂow rates can be presumed. To verify this ﬁnding, Figure 2.15b shows the

scatter plots for Benzene outﬂow rate and C9 outﬂow rate dependent on the (contemporary)

inﬂow rate.

● ●

● ●

●●●

● ●

●

● ●●

●●

●

●●

●

●

●● ●●

120

●●● ●

●

●●●

● ●●

●

●

● ●

●●

130

● ● ●●

●●

●

●● ●

●

●

●

●

● ●●●

● ●●

●

●●

●

●● ● ●

● ● ● ●● ●

●

●●

●

●

● ● ●

●

● ●

input rate

● ● ●●

●

●●

●●● ●

●● ●●● ●

●●

●

● ●●

● ●● ●

● ● ●

●●●

●●●● ●

●●● ●●

●●

● ●

●● ● ●

Benzene rate

●●●● ● ● ● ●● ●● ●

120

● ●● ●● ● ●●

●● ●

●

●

●●

● ●●●

110

●

●●

● ● ●

● ● ● ●

●

●

●● ●●

● ●

●

● ● ● ●

● ●

●

●●

● ●

●●

●

●

●

●●

● ●

●

●●●

110

● ●

● ●●

●

● ● ● ● ●

● ●

● ●

● ● ●

● ● ●

100

●

●

● ●

120

●

● ●

Benzene rate

● ●●

● ●

●●

90

●●

●

●●●

●

110

● ● ●

3.0 ●

● ●● ●

100

● ●

● ●●

● ● ●●

●●

● ●

●

●

● ●●

● ●

●●

●

●●

● ●

● ● ●●●

● ●

●●

●●● ●

●●

●

● ● ●

●●

●

●

● ●●●

● ● ● ●

●

90

●●

●●

● ●

●● ● ● ● ●

●● ●

●●

●●● ●●●

●●●●●● ●●

● ●

●

●

●

●

●

● ●●

●●

● ●

● ●● ●●

C9 rate

● ● ●●

●

●●●

●● ● ●● ●

●

2.5

●●

●● ●

●● ● ●

●

● ● ● ●

●●

3.0

●●● ● ● ●● ● ● ●

● ● ● ●●●●

● ● ● ● ●

●

●

●● ● ●●

●

●

●●● ●● ● ●● ●

●●

●●

●

● ● ● ● ●

●

●● ● ●

● ● ● ●●

C9 rate

●● ●

●

● ●● ●● ●

●● ●

●

●● ● ● ●●

●

●

● ● ● ●

2.5

●●

●

● ● ●

● ● ● ●

●●● ●

●

● ●

●

2.0

● ●

●

● ●

● ● ● ● ●

●

● ●

●● ●

●

2.0

●● ● ●

●

●

● ●

●●

●

●

0 50 100

200 150

250 300 110 120 130

Time input rate

(a) Inﬂow and outﬂow rates for de-alkylation plant (b) Scatter plots for Benzene and C9 outﬂow rates

dependent on inﬂow rate

From Figure 2.15b it can be observed that the output rate of Benzene depends linearly

on the input stream, while the C9 outﬂow rate more likely shows a slightly quadratic

dependency pattern. Furthermore, from the general nature of chemical production plants,

an auto-correlation structure of both output components can be presumed.138 Let Y be

the matrix of both output ﬂow rates with dimension 336 × 2. yt ∈ R2 is the vector of both

output rates at time index t. The inﬂow rate x is a vector of length 336 and xt denotes

the inﬂow rate at time index t. To model the outﬂow rates yt , a vector auto-regression

(VARX) model is supposed with inﬂow rate x and squared inﬂow rate x2 as exogenous

variables. Figure 2.15a clearly indicates a non-stationary time series of inﬂow rates which

is primarily characterized by multiple changes of its mean. This can be interpreted as

actions to control the plant’s performance. An analysis of the stationary parts of the

inﬂow time series shows no indication for auto-correlation structures. It is assumed that x

follows a white noise process with varying means and is handled as an exogenous regressor.

138

A raw initial analysis by checking the residuals of a simple regression of the input rate on both output

rates conﬁrms this presumption. The results are not shown here for the sake of brevity.

46

To select an appropriate order for the supposed VARX model, models of various orders are

ﬁtted and compared based on standard information criteria.139 The information criteria

suggest a model order of p = 3 or p = 4.140

For reasons of parsimony a V ARX(3) model is chosen, i.e.

0.12 1.67 0.03 0.78 −0.20 −2.91 0.96 −0.00023

( ) ( ) ( ) ( ) ( )

0.04 0.82 0.01 −0.21 −0.02 0.15 −0.03 −0.00004

Table 2.8: Coeﬃcients of the initial V ARX(3) model for de-alkylation plant

Σ̂ = .

⎝ 0.014 0.005 ⎠

Afterwards, the residuals are analysed for deviations from the underlying assumptions

by checking for remaining correlations and normality. Checking for remaining correla-

tions among the residuals is performed by analysing the auto-correlation for both residual

components and the cross-correlation between both components. These are shown in Fig-

ure 2.16a. Checking for normality is done by analysing the quantile-quantile plots (QQ-

plots) for both residual components (shown in Figure 2.16b) and a Jarque-Bera-Test for

normality.

From Figure 2.16a no clearly signiﬁcant auto- and cross-correlations are obvious. At

lag 16 of the ACF of the C9-residuals and at lags 0 and -6 of the residuals’ CCF slightly

signiﬁcant correlations can be found (at a 5% level). It is to conjecture that the signiﬁcant

ﬁndings are spurious since it is hardly reasonable to assume a delay of 6 or 16 hours

before the outﬂow rates are aﬀected.

The QQ-plots displayed in Figure 2.16b suggest a good ﬁt to the normal distribution.

However, the multivariate Jarque-Bera-test rejects the normality hypothesis (at a 5% α-

level). This result is probably caused by some ﬂuctuations in the outﬂow rates occurring

at changes of the production level (e.g. see time intervals 40-70, 100-170 and 270-310).

These outliers might be caused either by process instabilities or measurement errors. It

is to conjecture that a better ﬁt to the normal distribution can be achieved by explicitly

139

Computations are performed using the function VARselect from the vars package with R 2.15.1. See

Pfaﬀ (2008).

140

The detailed results are kept for the appendix in Table A.1.

141

Computations are performed using the VAR function from the vars package with R 2.15.1, see Pfaﬀ

(2008). The model’s ANOVA can be found in Table A.3 in the appendix.

0.3

Benzene residuals

6

Benzene residuals

0.2

4

0.1

2

−0.1

0

autocorrelation

−2

0.3−0.3

C9 residuals

Sample Quantiles

−4

0.2

0.1

−6

C9 residuals

0.2

−0.1

autocorrelation

0.1

−0.3

0 5 10 15 20 25

Lag

0.3

Benzene & C9 residuals

0.0

0.1

Sample Quantiles

−0.1

−0.1

crosscorrelation

−0.2

−3 −2 −1 0 1 2 3

−0.3

−20 −10 0 10 20 Normal Quantiles

Lag

(a) Autocorrelation and cross-correlation for residuals of both outﬂow (b) QQ-plots of residuals’ marginal distributions for both outﬂow rates

rates

Figure 2.16: Residual diagnostic plots for V ARX(3) model of the de-alkylation plant.

47

48

199

●

● ordinary residual

0.2

outlying residual

64

● 61

158 ● ●

●

● ●

●

●

●

● ●

● ● ● ● ●● 302

249 ●

0.1

●

● ●● ●

● ● ●

● ● ● ● ●● ● ●

residuals (C9)

● ● ● ●

● ●

●● ●

● ●

● ● ●

● ● ●

●● ●● ●●● ●

●● ● ● ● ●

● ●

51 ● ●● ●● ●

● ●● ●

● ● ● ● ● ●

● ● ● ● ● ●● ●●●

●●

● ●●

● ●

● ●

●● ●● ● ●● ●● ● ● ● ● ●

● ● ● ●●

●

● ●●●●● ● ● ● ●

●

● ● ● ● ●● ● ● ● ● ●● ●

0.0

●

●●● ● ● ●● ● ● ●●● ● ●

● ● ●●

● ● ● ● ● ●● ● ●● ● ● ●

●

● ●

● ●●●● ● ● ● ● ● ● ●

● ● ● ● ● ● ● ●● ●

● ● ●● ● ●●

●● ● ● ● ● ● ●

● ● ● ● ● ●● ● ● ●●● ● ● 155

●●

● ● ●

●● ● ●● ●● ●

●●

●

● ●

●● ● ● ●

● ● ● ● ● ● ●

207

● ● ●● ● ● ● ●● ●● ● ●

● ● ●

● ● ●●● ● ● ●● ●

● ● ●●●

● ● ●

−0.1

297 ●

●

●●

● ● ●● ● ●

●● ●

●

● ● ●

●

●

● ●

20 ● ●

●

65 ●●

●

●

188 202 156 109

● ●

−0.2

● ●

−6 −4 −2 0 2 4 6

residuals (Benzene)

Figure 2.17: Scatterplot of residuals for outlier corrected V ARX(3) model of the de-

alkylation plant

based on a robust estimation of the residuals’ covariance matrix142

Σ̂ = .

⎝ 0.008 0.005 ⎠

Based on the estimated robust covariance matrix, the Mahalanobis distances for all residu-

als are calculated. All residuals whose Mahalanobis distances exceed a critical value based

on the 97.5% quantile of the χ22 -distribution are handled as outliers.143 For each identi-

ﬁed outlier, a dummy variable is constructed which has entry 1 at the time index of the

identiﬁed outlier (and is zero otherwise).144 Figure 2.17 shows the pairs of residuals at

each time. Grey-coloured ◾-marked points refer to identiﬁed outliers whose time indices

are superimposed.

Subsequently, (2.67) is enhanced by the dummy variables for the outliers. To provide the

142

The robust estimator is the minimum covariance determinant estimator (MCD), carried out by covMcd

function with default settings from the robustbase package, see Rousseeuw and van Driessen (1999).

143

See e.g. Maronna et al. (2006) or Liebscher et al. (2012).

144

I.e. so-called innovative outliers are modelled. The procedure can be seen as a multivariate adaptation

of the univariate standard procedure for outlier detection as described e.g. in Cryer and Chan (2008,

pp. 257-260).

49

most parsimonious model, only parameters with signiﬁcant eﬀects are ﬁnally selected.145

The corresponding re-ﬁtted coeﬃcients of the selected model can be found in Table 2.9.146

−− 2.51 −− −− −0.15 −2.67 1.05 −0.00028

( ) ( ) ( ) ( ) ( )

0.04 0.85 0.01 −0.25 −0.02 0.18 −0.03 −0.00003

Table 2.9: Coeﬃcients of the V ARX(3) model with outlier correction and variable selec-

tion

While the estimates of the autocorrelation coeﬃcients for the C9 time series (lower rows

in Φ̂1 to Φ̂3 ) only change slightly, the estimates the autocorrelation coeﬃcients for the

Benzene time series (upper rows in Φ̂1 to Φ̂3 ) are clearly aﬀected since three parameters

are dropped from the model. The remaining coeﬃcients are aﬀected, too. In particular,

the lagged cross-correlations to the C9 time series change from 1.67 to 2.51 and from

-2.91 to -2.67 (right upper entries in Φ̂1 and Φ̂3 ). This conﬁrms the serious eﬀect of even

unobtrusive outliers in multivariate times series analysis. By incorporating the outliers’

eﬀects, the model’s AIC decreases from -4.22 to -4.72. Similarly, SIC decreases from -4.05

to -4.17. The analyses of residuals show a similar pattern as for the initial model and

reveal no serious hints for cross- or auto-correlation.147 Now, the multivariate Jarque-

Bera test does not reject the hypothesis of multivariate normally distributed variables (at

a 5% level). The residuals’ empirical covariance matrix is ﬁnally estimated as

Σ̂ = .

⎝ 0.011 0.004 ⎠

Using this model, the outﬂow rates can be ﬁtted quite well, as Figure 2.18 shows.

145

The estimated coeﬃcients for the V ARX(3) model with outlier compensation can be found in the

appendix, Table A.2.

146

The model’s ANOVA can be found in Table A.4 in the appendix.

147

The corresponding correlograms can be found in the appendix in Figure A.2.

50

125

115

105

Benzene rate

real

fitted

outlier

90 95

3.0

2.5

C9 rate

real

2.0

fitted

outlier

Time

Figure 2.18: Real and ﬁtted time series of both outﬂow rates (outlier indices superimposed)

51

industry logistics

The preceding chapter deals with methods describing the behaviour of product ﬂows in

chemical production plants to model the core components of chemical production networks

in detail. The next step in modelling product ﬂows in chemical production networks

is to describe product ﬂows between chemical production sites and plants. At a site,

intermediate chemicals are produced by some plants and consumed by some other plants

whereby raw chemicals are only consumed and ﬁnal chemicals are only produced. To

buﬀer temporal imbalances of chemical ﬂows, inventories are hold at the sites. Figure 3.1

shows the schematic chemical production network with added inventories symbols and

highlighted chemical ﬂows.

intra-site inter-site

S1 C1

intra-site

intra-site

S2 C2

This chapter deals with methods to plan the distribution of chemicals in chemical

production networks from a focal point of view. Speciﬁc planning problems occurring for

the transport modes pipeline, rail and ship are described.

T. Kirschstein, Integrated Supply Chain Planning in Chemical Industry, Produktion und Logistik,

DOI 10.1007/978-3-658-08433-2_3, © Springer Fachmedien Wiesbaden 2015

52

Logistics in basic chemical industry is in many aspects diﬀerent from other industries.

This is mainly due to the special properties of the production processes and handled

products. Intermediate and basic chemicals are typically ﬂuids or gases and often (at

least) environmentally hazardous. Hence, transport and storage processes are strongly

focused on reliability and safety. Furthermore, large quantities of these chemicals have to

be shipped to a relatively small set of destinations. This leads to a tendency of transport

consolidation. On the other hand, material distribution planning has to face high out-of-

stock costs. These are caused by high set-up costs for continuously operated plants and a

high risk of demand losses. This has led to high average stock levels in chemical industry

(compared to other industries).1 This fact promises opportunities for cost reductions by

an improved distribution planning and an eﬃcient inventory management.

To exploit these cost reduction potentials, the special requirements of chemical products

and production processes have to be considered in distribution planning. Due to the

vast quantities of raw and intermediate chemicals, high-capacity transport systems are

advantageous. High-capacity transport systems include pipelines, rail cars, and ships/

barges. For ﬁnal chemical distribution road transports are dominating because the set of

destinations is disperse and transport quantities are often suﬃciently small. This becomes

apparent in Figure 3.2 where the modal splits are depicted (in terms of the total quantity

of transported chemical products in Germany in 2009 categorized by transport mode).

road

[43.6%]

road

[70.2%]

pipeline

[14.6%]

open sea open sea

[7.7%] [8.8%]

short sea

[5.8%] short sea

rail [9.6%]

pipeline [8.4%]

rail

[23.4%]

[8.0%]

(a) Modal split - all companies (total: 313 (b) Modal split - chemical companies (total: 107

mill. tons) mill. tons)

Figure 3.2: Modal split for chemical products in Germany in 2009 (based on total trans-

ported quantity in mill. tons)2

Figure 3.2a shows the modal split for all transports of chemical products in Germany

in 2009. This includes all transports along the complete supply chain, in particular ﬁnal

chemical transports to customers as well as raw and intermediate chemical transports

1

See Shah (2005).

2

See VCI (2011).

53

between chemical companies. The modal split is fairly similar to the total modal split

over all products.3 Road transports are dominating with a share of about 70% of the

total transported quantity. This picture changes if only transports performed by chemical

companies are considered (Figure 3.2b). Here, the share of road transports is considerably

smaller (~43%). Most transports are handled via high-capacity transporters. E.g. 70%

of the total amount of transported alkenes such as Ethylene and Propylene are shipped

via pipeline.4 This shows the special relevance of these transport modes for the chemical

industry. It is to presume that this picture is even more biased in favour of high-capacity

transports in basic chemical industry.

For all considered means of transport the problem is to plan transport ﬂows and local

stocks such that total costs for transports and stock holding in the production network are

minimized. Depending on the mean of transport, various technical restrictions have to be

considered. In the case of rail cars or ships, transport ﬂows have to be accompanied by a

corresponding ﬂow of transport carriers such that carrier ﬂows are planned simultaneously.

Depending on the modelling of the carriers, the problem structure is similar either to

inventory-routing problems (IRP)5 or to capacitated network design problems (CNDP).

In the former case, routes for each carrier have to be determined whereas in the latter case,

carrier ﬂows are modelled as dynamic arc capacities.6 The literature on IRPs considering

ship transports constitutes the sub-class of maritime IRPs which is brieﬂy reviewed in

subsection 3.4.2. No approach is known addressing IRPs for rail transports.7 This is

probably because in the rail context transports are not performed by a single transport

carrier (such as a ship). Rail transports require a joint motion of locomotives and rail

cars. The routing of both types of equipment would complicate the problem considerably.

Therefore, in subsection 3.3.2 a hybrid CNDP-IRP model for rail transports in chemical

industry is formulated as a multi-layer multi-commodity network design model. Here, the

transport traction, oﬀered by locomotives, is modelled as a binary decision incurring a

ﬁxed charge whereby the transport capacity, oﬀered by rail cars, is modelled by rail car

ﬂows routed through the network.

In contrast to all other means of transport, pipeline transports do not require mobile

transport devices. Therefore, the planning of pipeline transports is not concerned with

the matching of transport ﬂows and transport carrier ﬂows. Nonetheless, other problems

come into scope such as the planning of pumping sequences when the pipeline is used by

multiple products.

3

See Statistisches Bundesamt (2012).

4

Further 20% are transported by ships, see Association of Petrochemicals Producers in Europe (2004).

5

For a recent and encompassing overview on IRPs see Andersson et al. (2010).

6

These capacities are typically associated with ﬁxed charges, see e.g. Gendron et al. (1999) or SteadieSeiﬁ

et al. (2014).

7

See Andersson et al. (2010).

54

Some facts distinguish pipeline systems from all other modes of transport:

The last fact induces that investment and management of pipeline systems are com-

pletely organized by the participating companies and (almost) all costs are internalized.

Investments in pipeline infrastructure make sense under speciﬁc circumstances only:

These requirements are given e.g. for the transport of raw and intermediate chemicals

within a chemical production site. For trans-regional transports, most prominent exam-

ples for pipeline transports are crude oil and natural gas pipelines. However, also some

basic chemicals such as Ethylene or Ammonia are transported via pipelines over long

distances to connect partners in a chemical supply chain.9

Pipelines are used for transports of ﬂuids or gases in a tube over long distances. The

tube is segmented by pumping stations into more or less regular parts. In contrast to all

other modes of transport, the transport infra-structure does not need a (mobile) carrier to

operate the transports. Operating costs for pipeline transports are comparatively small

and are mainly driven by maintenance and energy costs. In combination with the high

transport capacity, this leads to small transport cost rates measured on a ton-kilometre

basis.10 This cost advantage comes at the expense of spatial and organisational inﬂexi-

bility. Pipelines can only be used for liqueﬁable substances and are designed for a given

set of products. Because most substances to be transported have certain chemical prop-

erties, pipelines are built with regard to these properties, i.e. to minimize deterioration

8

Nonetheless, there exist several projects where the public sector supports pipeline investments in coop-

eration with private companies. See e.g. Association of Petrochemicals Producers in Europe (2004).

9

E.g. see Association of Petrochemicals Producers in Europe (2004) for details about the European

network of Ethylene and Propylene pipelines.

10

E.g. the energy consumption rate for pipeline transports is 0.14 mega-joules [MJ]/t-km compared with

0.35 MJ/t-km and 0.45 MJ/t-km for ship and rail transports. See van Essen et al. (2003).

55

and corrosion. Using the pipeline for other chemicals is usually not possible or associated

with high preparation and/or repair costs.

To categorize pipeline systems from an organizational point of view, two basic criteria

can be used: The number of products to be transported and the number of access points.

Commonly, pipelines are designed for one product only, e.g. in the case of crude oil trans-

ports. This allows optimizing the pipeline’s technological conﬁguration with respect to

the energy consumption. However, in chemical industry also multi-product pipelines are

met when the intended chemicals to be transported are chemically similar.11

The structure of the pipeline network is another aspect to be considered. Since in

pipelines the product ﬂow is unidirectional in very most cases,12 the network structure

can be categorized according to the number of sources and sinks. In the simplest case,

there is a single source and a single sink (one-to-one network). Similarly, if there are

multiple sinks to be supplied this is labelled as a one-to-many network. Conversely, there

exist many-to-one networks and, ﬁnally, for multiple sources and sinks also many-to-many

networks. Table 3.1 displays this categorization scheme.

system structure

many-to-

one-to-one one-to-many many-to-one

many

# products

single

ﬂow split junction ﬂow

batch batch

multiple batch ﬂow batch split

junction network ﬂow

Of special relevance among these types are the one-to-many and many-to-many net-

works as these show the most challenging planning problems. In the single product case,

the main problem is to keep the ﬂow in balance along the pipeline or pipeline segments

such that inﬂow and outﬂow are equal at any time.

A more complicated problem is met in the multi-product case where batches of products

have to be routed through the pipeline (network) whereby a batch may also be split into

sub-batches at pipeline forks. In the cases of multiple sources, additionally a coordinated

pumping planning is required under consideration of the pumping capacity, the due dates

of the batches, and the segment-wise ﬂow balances.

Pipelines are typically operated with a unidirectional material ﬂow. In typical set-

tings the associated partners along the pipeline can be categorized distinctly in (pure)

11

A necessary condition is that chemicals to be transported do not react with each other. See e.g. Asso-

ciation of Petrochemicals Producers in Europe (2004).

12

However, there exist exceptions as described in Magatão et al. (2004).

56

providers/sources and (pure) consumers/sinks. Here, the providers are located at the

head of the pipeline and the consumers are located at the bottom. This is called a seri-

ally operated pipeline.

In contrast, if the partners along the pipeline can act as providers and consumers, each

partner has to be able to receive from and feed in the pipeline. These situations are met

in local pipeline systems, e.g. at large-scaled integrated chemical production sites, where

multiple plants and storage facilities are connected to a single-product pipeline. Here,

the storage facilities buﬀer the total pipeline ﬂow and serve as receivers and suppliers of

products at the same time. This requirement is met by building orbital or cyclic pipelines

connecting all partners. Still the material ﬂow is unidirectional, but all partners can take

or feed material from/to the pipeline as long as the (segment-wise) total ﬂow is in balance.

There might be additional chemical or technical restrictions that have to be taken into

account for pipeline transport planning such as the interruptibility of pipeline operation.

E.g. if a product remains too long in a pipeline without movement, chemical reactions

may take place which may cause corrosion or polymerization processes. This would lead

to a failure and cause expensive maintenance and/or cleaning operations. However, the

most obvious technical restriction is probably the pump rate of a pipeline. The pump rate

measures the transport speed on a volume-per-time basis. Typically, there is an upper

bound for the tolerable pressure inside a pipeline determining the maximum pump rate

(say ρcap ). The realized pump rate might be smaller and depends on the power of pumps

at the pumping stations.13

A relevant restriction to be fulﬁlled for operating pipelines is the ﬂow balance. I.e., the

input ﬂow equals the output ﬂow implying constant physical conditions (such as pressure)

for all pipeline segments. For instance, assume a serially operated multi-access pipeline

where N is the set of access points consisting of N out destination points and N in feeding

locations such that N out ⋂ N in = ∅ and N out ⋃ N in = N . The ﬂow rate ρ is measured

in volume-per-time units and ρi is the ﬂow rate at location i measured at a point of

time.14 For i ∈ N in , ρi is an inﬂow and for i ∈ N out it is an outﬂow. Assuming constant

pressure and a completely ﬁlled pipeline, the inﬂow-outﬂow condition ∑i∈N in ρi = ∑j∈N out ρj

holds at each point in time. The ﬂow rates ρi can be controlled via valves. For multi-

access pipelines, at the access points heterogeneous ﬂow rates may occur. Figure 3.3

shows an example for a serial multi-access pipeline with N = {1, ..., 4}, N in = {1, 2} and

N out = {3, 4}.

Due to the ﬂow balance and pipeline capacity it follows ∑i∈N in ρi = ∑j∈N out ρj ≤ ρcap and

13

Typically the pumping power can be controlled. In a reasonably planned pipeline system typically the

pumps’ maximal power suﬃces to exploit the maximum pump rate ρcap .

14

A time indexing of ﬂow rates is dropped for the sake of simplicity. In principle, ﬂow rates are time-

continuous variables, i.e. ρi (t) is the precise notation of ﬂow rate functions.

57

i=2

i=1 ρ2

ρ1

j=3 j=4 ρ4

ρ3

j ≤ ρcap .

The planning and control of pipelines is a technically challenging task since the ﬂow

parameters have to be controlled continuously to ensure immediate reactions in case of

leakages or similar problems. Therefore, a pipeline model is necessarily representing the

reaction of (outﬂow) ﬂow rate and pressure depending on certain control parameters such

as inﬂow pressure and pumping power. Modelling the pipeline ﬂow leads to a system of

non-linear partial diﬀerential equations which cannot be solved analytically.15 However,

numerical methods are available16 that ﬁnd their way into modern pipeline management

systems.17

Despite the considerable challenges in pipeline control, diﬃcult operational planning

problems do not occur for single-product pipelines from a logistical point of view. An ex-

ception is the planning of maintenance operations which, for instance, include the planning

of replacement investments to prevent technical short-falls. When maintained carefully,

short-falls due to technical problems occur rarely.18 The structural integrity of a pipeline

has to be checked in regular intervals by pipeline inspection gauges (PIGs) that are used

e.g. to scan for structural weaknesses. These PIGs most often are passive, i.e. they have

no driving unit and are transported within the usual material ﬂow. Beside the termina-

tion of these PIG runs (in accordance to legal regulations), there is no planning problem

associated as long as the usual pipeline operation does not have to be interrupted.

Beside the control of material ﬂow and the planning of maintenance activities, there

are no operational planning problems involved in running a pipeline system that could

be labelled as a logistical problem. The only parameters with logistical relevance to be

planned operationally are the ﬂow rates. Flow rates have to assure that all plants along

the pipeline can work properly, i.e. it has to be ensured that enough material is available

to supply the consuming plants. If the pipeline cannot be stopped (ρ > 0), it has to be

15

See e.g. Matko et al. (2000) for an overview or Herrán-González et al. (2009) for the special case of gas

pipelines.

16

See e.g. Blaẑiĉ et al. (2004).

17

E.g. see Cameron et al. (2001).

18

See e.g. Association of Petrochemicals Producers in Europe (2004) for empirical ﬁgures.

58

assured additionally that suﬃcient storage capacity is available to maintain a minimal ﬂow

rate at any time. As the producing plants along the pipeline are not working constantly

at the same level, buﬀers have to be integrated in the pipeline system to assure material

availability. A crucial question arising is the distribution of stocks along the pipeline. A

typical setting consists of tanks/inventories at each access point.

In contrast to common assumptions in inventory management, the demands of chem-

icals at chemical production sites are composed by the input and output ﬂows of the

related plants instead of a total number of orders within a period. This implies that

the uncertainty covered by safety stocks arises rather from technological sources than

from external markets. Moreover, make-to-stock production is predominating in chemical

industry because almost all products are commodities.

In the case of continuously operated single-purpose plants the set of input and output

chemicals is ﬁxed (i.e. there is only one recipe used). However, the throughput of the plant

can vary in most cases. I.e. there is a ﬁnite set of production modes each one associated

to speciﬁc ﬂow levels of all chemicals. As such plants are typically designed for maximal

eﬃciency, the production mode associated with the maximum throughput is the intended

state of production. A change to another mode is often forced by disturbances such as

technical failures.19

To sum up, chemical production plants are characterized by a ﬁnite number of produc-

tion modes (say ∣S∣ = k) corresponding to a ﬁnite number of ﬂow levels ω(s) with s ∈ S

for the associated chemicals.20 Each production mode s lasts for some time (reﬂecting

e.g. repair times).

The unintended change of a plant’s production mode can be interpreted as a stochastic

process. Two components are necessary to model such a process: If a plant enters a

certain production mode, the sojourn time (the time until the mode changes again) must

be expressed. When the production mode changes, a transition model determines which

mode is entered next. Since a ﬁnite number of production modes exists, the transition

from one mode to another is a discrete process with ﬁnite state space. The sojourn time

of a production mode is a continuous variable in principle. However, the analysis and

modelling of sojourn times as continuous variables requires a lot of historical records to ﬁt

an accurately parametrized sojourn time distribution for each mode.21 In most cases, this

empirical problem can be circumvent by considering a discrete time scale. A stochastic

model in discrete time assumes that a plant is in exactly one mode/state per period of

time. The sojourn time of a mode can then be reﬂected as the length of a sequence of

periods without a mode change. Thus, a joint sojourn time and transition model can be

19

The shut-down of a plant can also be seen as a speciﬁc production mode with zero throughput. Beside

unintended changes of the optimal production mode, sometimes planned shut-downs occur e.g. for

regular technical inspections or due to a severe drop in demand.

20

Note that the number of ﬂow levels of a certain chemical is smaller than the number of production

modes if multiple modes have the same ﬂow level of the speciﬁc chemical.

21

This is because some modes last only for a few days and rarely occur, e.g. plant break downs.

59

expressed by transition probabilities ruling the transition from one production mode to

another at the beginning of each time period. This constitutes a discrete Markov process

or a Markov chain.22

Assume that there are probabilities that a plant’s production mode changes in a period

from mode s to mode t denoted by qst ≥ 0. Transitions from mode s to s are allowed:

qss ≥ 0 reﬂecting the probability that the plant remains in mode s in the next period

(sojourn time). Assume that transition matrices Q exist for all plants which are squared

matrices composed by the mode transition probabilities qst . Such a matrix is called

a stochastic matrix if the sum of each row or each column equals 1, respectively. By

convention, it is assumed that the rows sum up to 1, i.e. ∑kt=1 qst = 1, s = 1, ..., k. A

stochastic matrix constitutes a discrete time Markov process/chain.23 Let u0 be a binary

(row) vector of length k which has entry 1 at some starting position s and zero entries

otherwise, i.e. u0 represents the starting state of the Markov process in period 0.24 Then

the probability vector for period 1 is calculated by u1 = u0 ⋅ Q whose entries represent the

probabilities that the process enters the corresponding state. If Q constitutes a regular

Markov chain, a steady state vector π = {π1 , ..., πs , ..., πk } exists where πs represents the

long-term probability that the modelled process is in mode s.25 The steady state vector

π can be obtained by Π = lim Qn where each row of Π equals the steady state vector

n→∞

π.26

Another property of the steady state vector is its temporal invariance: If the Markov

process starts with π as the initial distribution (u0 = π) the starting distribution remains

unchanged for all following periods: u1 = πQ = πQn = un = π for n > 0.27

Let X jm denote the ﬂow rate of the considered chemical of plant m ∈ {1, ..., M j } at

location j. Each plant has a ﬁnite number of production states/modes S jm with ∣S jm ∣ =

k jm . Each state is associated to a ﬂow rate of the considered material that is consumed

or produced per period denoted by ω(s) with s ∈ S jm or, more compact, ωsjm ∈ Ojm . The

corresponding transition matrix Qjm is of dimension k jm ×k jm . The transition probability

jm

from state s to state t is denoted by qst for plant m at location j.

j

Let X be the total ﬂow rate at location j in a period of time (i.e., X j = ∑M

j

m=1 X

jm ).

The set of states of X , denoted by O , depends on the combination of the states of all

j j

j

j

location j. Its cardinality is given by ∣S ∣ = k = ∏M

j j

m=1 k

jm . Hence, an element sj ∈ S j is

a vector of states for all plants at location j, sj = (sj1 , ..., sjM ) with sjm ∈ S jm . Under

j

22

Since the term "chain" refers to ﬁnite state space, the term "Markov chain" is probably more precise,

see e.g. Ibe (2008, ch. 2).

23

See Grinstead and Snell (1997, pp. 405-407).

24

To ease calculus all vectors are row vectors in the following.

25

The term regular Markov chain, simply spoken, refers to the property that all possible states can be

reached in a ﬁnite number of subsequent periods independent from the starting distribution. For

more information about deﬁnition, conditions, and properties of regular and ergodic Markov chains

see e.g. Grinstead and Snell (1997, pp. 433 ﬀ.).

26

See e.g. Grinstead and Snell (1997, pp. 435 ﬀ.)

27

More formally, π is the eigenvector of Q with eigenvalue 1.

60

j

independence of the production plants and their states,28 the transition probability qst

from state combination s to state combination t is given by

Mj

j

qst = ∏ qsjm

m tm . (3.1)

m=1

(with dimension k j × k j ) and constitute regular discrete Markov processes. Hence, steady

state vectors π j for each location j exist which represent the long-term probabilities of a

site’s state combinations.29

At integrated chemical production sites, however, plants might not be entirely indepen-

dent e.g. due to a common energy supply. Under such circumstances the states of diﬀerent

plants are interdependent since the transition probabilities also depend on the states of

other plants. In this case, X j and S j are the same as described above but the transition

matrix Qj has to be derived directly by analysing the entire production network as a more

complex stochastic process, not as a simple convolution of independent Markov chains.

The associated total material balance of a state combination sj ∈ S j at location j is

deﬁned by

Mj

ωsjj = ω(sj ) = ∑ ω(sjm )

m=1

with ω j ∈ Oj . Note that there might exist state combinations with identical total

material balance, i.e. ∣Oj ∣ ≤ ∣S j ∣.30 Without loss of generality deﬁne that if X j > 0 there is

a (net) demand, whereas if X j < 0 the considered material is surplus at location j. Often,

for a given location j, X j is either exclusively positive (Oj ⊂ R+ ) or exclusively negative

(Oj ⊂ R− ).

The total material balance over all locations associated to the considered pipeline is the

sum of all X j : X = ∑j∈N X j . Again, X depends on the combination of the locations’ total

production states. Let S denote the set of the locations’ state combinations and O the

set of associated total production balances of X. To calculate the transition probabilities

qst from a total state combination s ∈ S to a total state combination t ∈ S the same

procedure as mentioned before is applied assuming independence across locations. As

diﬀerent locations typically do not share a direct technical relation (in contrast to the

plants at a location), it can be assumed that independence is not a crucial assumption here.

Therefore, s, t ∈ S are vectors of ∑N j

j=1 M elements assigning each plant at each location

a state, i.e. s = (s , ..., s ) with s ∈ S and N = ∣N ∣. Consequently, Q is constructed

1 N j j

28

I.e. that a mode change at a plant does not depend on the states of all other plants. In other words,

changes of a plant’s production mode are not inﬂuenced by the other plants’ modes.

29

This follows immediately for regular Markov chains because the corresponding transition matrices

must contain positive entries only. For more information about the summation of Markov chains see

Rozhkov (2010).

30

Equality holds if all state combinations result in unique total material balances.

61

Markov process whose steady state vector is denoted by π. Supposed the pipeline supply

corresponds to the regular transport mode for all destinations, then E(X = ∑j∈N Xj ) = 0

is an appropriate assumption.31

A critical situation occurs when the supply system is out of balance, i.e. if X ≠ 0. Either

the material is surplus (X < 0) and has to be stored or it is in deﬁcit (X > 0) and has to

be provided from stock. Hence, stock capacities and safety stocks are required to cover

such critical situations.

Since the pipeline ﬂow rates in most pipeline systems can be adjusted easily, the precise

distribution of stock capacities and safety stock levels can be compensated by adjusting

the pipeline ﬂow rates accordingly. It suﬃces to determine a common stock capacity and

safety stock level. The technical and economic impact of such situations depend on the

height of the imbalance and its duration. In supply systems, strategic safety stocks are

designed to cover the demand of a certain number of periods of time with a predeﬁned

probability.32 Typically, this time span is the lead time required to replenish inventory,

e.g. by placing an order with an external supplier. In the case of pipeline supply systems,

the supply should be organized by the partners along the pipeline which act as suppliers

and customers. Hence, a desired time span can be assumed during which the supply

system should be operated autarkical (with a certain probability). Let b denote this

critical time span. The demand during this time span is denoted by Y (b) and is composed

as the total sum of total material balances over b periods. The set of all possible state

sequences of length b out of domain S is denoted as H = S × ⋯ × S with ∣H∣ = k b where

%&& & & & & & & & '&& & & & & & & & &*

b

k = ∣S∣. An element h ∈ H is a vector of b state combinations, i.e. h = (s1 , ..., sb ) with

si ∈ S. The probability ph of a state sequence h can be calculated straight forwardly by

using the transition matrix Q and the steady state vector π as follows33

b

ph = πs1 ⋅ ∏ qsl−1 sl . (3.2)

l=2

lated by

b

yh = ∑ ω(sl ) (3.3)

l=1

where the total sum over all production quantities at all locations and plants is ω(sl ) =

31

Otherwise, the supply system is dis-balanced in expectation which would be an indication for a badly

managed system or not properly modelled suppliers/consumers.

32

See e.g. Tempelmeier (2005, pp. 397 ﬀ.).

33

This again implies independence. This can be seen as a special case of a hidden Markov model with

certain observations, see e.g. Ephraim and Merhav (2002). Although the number of sequences grows

exponentially, calculating the corresponding probabilities can be reduced considerably (O(k 2 b)) by

iteratively calculating the probabilities of sub-sequences (dynamic programming). In the context of

hidden Markov models this is called forward-backward algorithm, see e.g. Yu and Kobayashi (2003).

62

N j

∑j=1 ωsj .

l

Let yh(g) denote the total material balances in ascending order (yh(1) = min yh ) and ph(g)

h∈H

the associated probability vector. For a given total safety stock level along the pipeline r,

−1

the α-service level of the pipeline storage system is deﬁned by ∑fg=1 ph(g) with f determined

such that r > yh(f ) and r ≤ yh(f +1) . To deﬁne a safety stock level satisfying a desired

34

α-service level (say rα ), the quantile y α with 0 < α ≤ 1 needs to be deﬁned by y α = yh(f )

−1

with f determined such that ∑fg=1 ph(g) ≥ α and ∑fg=1 ph(g) < α. It follows that setting the

safety stock to

leads to stock-out situations at most with probability 1 − α (within the critical time span

b).

Moreover, the safety stock level can be determined based on the β-service level. I.e. at

least β ⋅100 percent of the average net deﬁcit occurring along the pipeline within b periods

b

can be provided from stock. Here, the average net deﬁcit is deﬁned by μC = ∑kg=f ph(g) ⋅ yh(g)

(with f determined such that yh(f ) > 0 and yh(f −1) ≤ 0). For the expected loss in dependence

b

of a certain stock level r follows V (r) = ∑kg=f (yh(g) − r) ⋅ ph(g) (with f determined such that

yh(f ) > r and yh(f −1) ≤ r). Then, 1 − Vμ(r)

C deﬁnes the β-service level of a stock level r for the

pipeline storage system.35 The total safety stock level satisfying a predeﬁned β-service

level can be determined by

rβ ≥ V −1 ((1 − β) ⋅ μC ) . (3.5)

The setting of a global safety stock level based on the lower bounds introduced above

implies that

of locations into material (net) providers and (net) consumers such that there is one

unique ﬂow direction within the pipeline, i.e. Xj ≥ 0 or Xj ≤ 0 for all locations j. I.e. if

the pipeline is serially operated, the total safety stock has to be distributed among the

provider locations. Another situation occurs for cyclic pipelines. Here, the material ﬂow is

unidirectional, but each location can act as consumer or provider as long as the total ﬂow

is balanced. The safety stock can be distributed among all participants along the pipeline

as long as all participants can feed surplus material from stock to starving partners.

Safety stocks at the consumer locations have to cover the risks of supply (i.e. pipeline

short-falls) and/or the risk of consumption rates higher than the pipeline’s maximum

34

This deﬁnition is similar to the usual α-service level deﬁnition, see e.g. Tempelmeier (2005, p. 397 ﬀ.).

35

Again see Tempelmeier (2005, p. 397 ﬀ.) for the usual β-service level deﬁnition.

63

transport capacity. The former risk is neglected in the deductions since pipelines are

typically highly reliable transporters. In the latter case, the described procedure can be

used to determine the total safety stock to be held at the consumer locations.

A further assumption is that the maximum ﬂow rate ρcap is not exceeded at a single

location, i.e. X j ≤ ρcap

j ≤ ρcap at all locations j. Otherwise, a local safety stock level has

to be incorporated covering situations when the local demand X j exceeds the maximum

ﬂow rate ρcapj . To calculate the local safety stock level, the procedure described above can

be adapted. Consider the transition matrix Qj and recall the set of state combinations

S j as well as the corresponding set of local material balances Oj ∋ ωsj = ω(sj ) with

sj ∈ S j . To incorporate the local ﬂow restriction, deﬁne a critical material balance by

j ). Subsequently, deﬁne the local critical demand surplus yh

ω C (sj ) = max (0, ω(sj ) − ρcap jC

over b periods for a state combination h = (s1 , ..., sb ) by yh = ∑i=1 ω (si ). The probability

j j jC b C j

of a state sequence remains as pjh = πsjj ⋅ ∏bl=2 qsjj sj . The required local safety stock rj can

1 l−1 l

∣Hj ∣

j = ∑g=f πh(g) ⋅ yh(g)

j jC

be determined e.g. by assuming that the expected demand surplus is μC

(with f determined such that yhjC(f ) > 0 and yhjC(f −1) ≤ 0).36 The local α- or β-service level

constraints can be deduced by means of (3.4) and (3.5) accordingly. Note that since

the calculation for the total safety stock along the pipeline does not diﬀerentiate states

(w.r.t. the local pipeline capacity), local safety stock levels serve as a part of the global

safety stock as well and are implicitly included in the total safety stock level calculation.

Conversely, the pipeline capacity at location j also restricts the feed rate. Hence, the

distribution of safety stock shares among the provider locations has to take into account

these constraints as well. In general, the number of covered periods should be as equal as

possible for all local safety stocks. I.e. the share of the total safety stock r for a particular

provider location j can be assigned by requiring an identical number of periods covered,

ρcap

i.e. rj = r ⋅ ∑ min(ρj cap ,ρcap ) for all j ∈ N in .

j j

The following example exercises the aforementioned procedure to derive safety stock

levels for a serially operated pipeline system with two provider locations and two consumer

locations

36

Here, yhjC(g) denotes the surplus demand in ascending order and πhj (g) its corresponding probability.

64

Example 4 (Serial pipeline supply network). Suppose that the pipeline system depicted

in Figure 3.3 connects four locations consuming or providing Ethylene. The pipeline is

serially operated with two providers of Ethylene at its head and two consumer plants at its

bottom. Provider P 1 is an Ethylene producing plant (e.g. a steam cracker) with a capacity

of 3,000 tons per day. Provider P 2 is a seaport where Ethylene can be unloaded from

tanker ships into tanks feeding the pipeline. Feeding the pipeline is operated in batch mode.

I.e. when the Ethylene stock level is suﬃciently high, Ethylene is fed into the pipeline with

a constant rate of 1,000 tons per day. Otherwise, there is no feeding. The consumer plants

are supposed to be continuously operated (e.g. to produce Ethylbenzene) with consumption

rates of 2,500 tons and 1,500 tons per day, respectively. All production plants are restricted

to two production modes, either full capacity working (s = 2) or a breakdown (s = 1) with

no production/consumption. Tables 3.2a- 3.2d show the transition matrices of providers

(QP 1 and QP 2 ) and consumers (QC1 and QC2 ) as well as the corresponding production

modes and associated Ethylene quantities given in 100 tons.

s 1 2 s 1 2 s 1 2 s 1 2

ωsP 1 0 -30 ωsP 2 0 -10 ωsC1 0 25 ωsC2 0 15

0 0.60 0.40 0 0.85 0.15 0 0.60 0.40 0 0.90 0.10

-30 0.20 0.80 -10 0.05 0.95 25 0.10 0.90 15 0.10 0.90

(a) QP 1 (b) QP 2 (c) QC1 (d) QC2

Table 3.2: Transition matrices and production modes for providers and consumers

1 2 1 3 1 4 1 1

πP 1 = ( , ) , πP 2 = ( , ) , π C1 = ( , ) , π C2 = ( , ) . (3.6)

3 3 4 4 5 5 2 2

Altogether, there are 24 = 16 state combinations. The state space of the total Ethylene

balance X is O =(-40, -30, -25, -15, -10, -5, 0, 5, 10, 15, 25, 30, 40) with ∣O∣ = 13. The

system is in balance (ωs = 0) if all locations are working (s = (2, 2, 2, 2)) or if all locations

are oﬀ (s = (1, 1, 1, 1)). Tables 3.3a and 3.3b show the state combinations constituting

all unbalanced total states.

For the total material balance X the corresponding transition matrix Q is calculated

using (3.1). For instance, q99 (i.e. that there is no transition from state s9 with ωs9 = 0)

is calculated as q99 = ∏4j=1 qljj lj = q22 q22 q22 q22 = 0.8 ⋅ 0.95 ⋅ 0.9 ⋅ 0.9 = 0.62. Analogously, the

P 1 P 2 C1 C2

From Q the steady state vector can be derived as π =(1/20,1/60,1/20,1/5,1/60,

1/40,1/15,1/120,1/5,1/40,1/15,1/10,1/120,1/30,1/10,1/30). Note that E(X) = ∑16 t=1 πt ⋅

ω(st ) = 0, i.e. in the long run the supply system is expected to be in balance.

Assume that the critical period during which the supply system should be kept au-

65

1 -40 (2, 2, 1, 1) 10 5 (1, 2, 1, 2)

2 -30 (2, 1, 1, 1) 11 10 (2, 1, 2, 2)

3 -25 (2, 2, 1, 2) 12 15 (1, 2, 2, 1)

4 -15 (2, 2, 2, 1) 13 15 (1, 1, 1, 2)

5 -15 (2, 1, 1, 2) 14 25 (1, 1, 2, 1)

6 -10 (1, 2, 1, 1) 15 30 (1, 2, 2, 2)

7 -5 (2, 1, 2, 1) 16 40 (1, 1, 2, 2)

(a) Total surplus states (b) Total deﬁcite states

t 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

ωst -40 -30 -25 -15 -15 -10 -5 0 0 5 10 15 15 25 30 40

-40 0.41 0.02 0.05 0.27 0.00 0.10 0.01 0.01 0.03 0.01 0.00 0.07 0.00 0.00 0.01 0.00

-30 0.06 0.37 0.01 0.04 0.04 0.02 0.24 0.09 0.00 0.00 0.03 0.01 0.01 0.06 0.00 0.01

-25 0.05 0.00 0.41 0.03 0.02 0.01 0.00 0.00 0.27 0.10 0.01 0.01 0.01 0.00 0.07 0.00

-15 0.07 0.00 0.01 0.62 0.00 0.02 0.03 0.00 0.07 0.00 0.00 0.15 0.00 0.01 0.02 0.00

-15 0.01 0.04 0.06 0.00 0.37 0.00 0.03 0.01 0.04 0.02 0.24 0.00 0.09 0.01 0.01 0.06

-10 0.21 0.01 0.02 0.14 0.00 0.31 0.01 0.02 0.02 0.03 0.00 0.21 0.00 0.01 0.02 0.00

-5 0.01 0.06 0.00 0.10 0.01 0.00 0.55 0.02 0.01 0.00 0.06 0.02 0.00 0.14 0.00 0.02

0 0.03 0.18 0.00 0.02 0.02 0.05 0.12 0.28 0.00 0.01 0.01 0.03 0.03 0.18 0.00 0.02

0 0.01 0.00 0.07 0.07 0.00 0.00 0.00 0.00 0.62 0.02 0.03 0.02 0.00 0.00 0.15 0.01

5 0.02 0.00 0.21 0.02 0.01 0.03 0.00 0.00 0.14 0.31 0.01 0.02 0.02 0.00 0.21 0.01

10 0.00 0.01 0.01 0.01 0.06 0.00 0.06 0.00 0.10 0.00 0.55 0.00 0.02 0.02 0.02 0.14

15 0.03 0.00 0.00 0.31 0.00 0.05 0.02 0.00 0.03 0.01 0.00 0.46 0.00 0.02 0.05 0.00

15 0.00 0.02 0.03 0.00 0.18 0.01 0.01 0.03 0.02 0.05 0.12 0.00 0.28 0.02 0.03 0.18

25 0.01 0.03 0.00 0.05 0.00 0.01 0.28 0.05 0.01 0.00 0.03 0.07 0.01 0.41 0.01 0.05

30 0.00 0.00 0.03 0.03 0.00 0.01 0.00 0.00 0.31 0.05 0.02 0.05 0.00 0.00 0.46 0.02

40 0.00 0.00 0.01 0.01 0.03 0.00 0.03 0.01 0.05 0.01 0.28 0.01 0.05 0.05 0.07 0.41

tonomous is set to b = 4 days. Hence, all ∣S∣ = 65, 536 sequences of state combinations are

evaluated. The corresponding probabilities and total material balances are calculated using

(3.2) and (3.3). Figure 3.4a shows the states of the total material balances Y (4) and the

associated cumulated probabilities. The corresponding loss function V (r) is depicted as a

function of the safety stock level r in Figure 3.4b.

To assure material availability with a probability of at least α = 0.95, a safety stock level

of at least rα = 95 ⋅ 100 tons is required. This safety stock level also ensures a β-service

level of at least β = 0.95 where μC = 22.45 (as deﬁned above). Both values are depicted in

ﬁgures Figure 3.4a and Figure 3.4b, respectively.

66

1.0

●

95% ● ●

● ● ● ●

● ● ● ● ● ● ●

● ● ●

●

● ●

0.8

cumulated probabilities

● ●

●

●

●

●

●

●

●

0.6

●

●

●

●

●

0.4

●

●

●

●

● ●

●

0.2

●

●

●

● ●

●

●

●

● ● ●

● ●

0.0

● ● ●

● ● ● ● ● ●

total material balance Y(4) y = 95

0.95

20

●

15

expected loss V(r)

●

10

●

5

●

●

●

μC ⋅ 0.05 = 1.122

●

●

●

●

●

● ● ● ● ● ● ●

0

● ● ● ● ●

0 50 100 150

V−1 ⋅ (μC ⋅ 0.05) = 95 safety stock level r

(b) Loss function V (r)

Figure 3.4: Cumulative distribution function and loss function for Y (4)

67

Material ﬂows along multi-product pipelines are far more complicated to plan. The most

prominent ﬁeld of application for multi-product pipelines is the distribution of reﬁnery

products among a network of distribution terminals.37 The operating principle is essen-

tially the same as for single product pipelines. However, as multiple materials pass the

pipeline sequentially, the diﬀerent materials have to be separated in some sense. Basically

there are two options to separate batches of diﬀerent materials

• separation PIGs or

• interfaces.

For the ﬁrst option, special PIGs are used. Separation PIGs are technical devices acting

as an impenetrable membrane. This separates subsequent chemicals perfectly at a point

of destination. A disadvantage is the eﬀort to plug-in the PIGs when a transition occurs.

Beside the investment costs for PIGs, they have to be re-distributed to the feed points

and must be constantly maintained. This is economically reasonable only under spe-

ciﬁc conditions, e.g. when the transported materials are of high value and comparatively

heterogeneous such that a mixing of materials induces high re-processing costs.

The second option is simply to allow a mixture of successively injected materials. This

transition mixture is often labelled as interface.38 This option is preferable as in many

cases the set of materials to be transported is chemically homogeneous (e.g. crude oil

derivatives). However, the interface is still a problem as it typically does not meet the

chemical speciﬁcations of one of the parental materials. There are two ways to deal with

the interface: Either the interface is added to one of both parental materials where it

does not harm in the further production processes39 or the interface has to be extracted

(e.g. at the end of a serial pipeline) for a special treatment.40 No matter how interfaces are

treated, transition eﬀorts have to be faced. Hence, minimizing the number of interfaces

is one objective to aim at when planning multi-product pipeline schedules.

If transition eﬀorts can be expressed in terms of costs, the total costs for pipeline

transport can be used as the planning objective. In this context, total costs comprise

transition costs, transport costs, and inventory costs. Transport costs account for oper-

ational pipeline costs, such as the energy costs for pumping. These costs are aﬀected by

the pipeline’s pump rate which needs to be explicitly controlled and planned when the

transported materials diﬀer in viscosity or other physical properties.41 Inventory costs are

37

E.g. see Rejowski and Pinto (2008) and references therein.

38

See e.g. Hull (2005).

39

Especially as its concentration declines to 0 when getting mixed with large amounts of the parental

material.

40

E.g. see Hull (2005).

41

See Rejowski and Pinto (2008) for an example. However, often exists a regular pump rate for which

the operational costs per ﬂow unit are minimized. Deviations from this optimized pump rate usually

incur additional operational costs due to increased (relative) energy consumption. See Hane and

Ratliﬀ (1995) or Hull (2005).

68

distinguished in stock holding costs and shortage costs. Note that in literature primarily

the ﬁrst category is taken into account.42 According to the classiﬁcation presented in

Table 3.1 the next subsection deals with the batch ﬂow case, i.e. multi-product pipelines

with only two access points.

In the batch ﬂow system, the pipeline connects two locations where a set of products

is produced and consumed. The subset of products that are used at both locations is

denoted by S. These products have to be balanced between both locations by pipeline

transports.

In the literature on batch ﬂow planning, primarily planning problems from the oil in-

dustry are prevalent. Here, pipeline systems are more extensively used than in all other

industries where pipelines mainly distribute chemicals from a reﬁnery to customer mar-

kets. The problem is to meet the customer demands in time and replenish local inventories.

These planning problems are typical operational planning problems where the production

quantities and customer demands are given for a certain time horizon. At both locations

tanks are used to buﬀer demand and supply over time. The connecting pipeline is used

to transport the materials to the consuming location and/or to the producing location.

The aim is to ﬁnd a pumping schedule that meets the demand at the consuming loca-

tion by minimizing the operating costs. Technical restrictions to be considered encompass

e.g. settling periods, tank pumping restrictions or interface expenditure approximations.43

In the literature, basically two directions can be distinguished. A concise comparison of

the problem features tackled in both directions is contained in Table 3.5.

Pipeline system

complete ﬁll state complete

range ﬂow rate ﬁxed

reversible ﬂow direction ﬁxed

one-at-a-time source/destination one-at-a-time

Tank system

ﬁnite capacities ﬁnite

one-at-a-time inﬂow/outﬂow one-at-a-time

Demand

range quantity per period ﬁxed

Table 3.5: Comparison of problem features for Magatão et al. (2004) and Relvas et al.

(2006)

42

This is because in most studied planning problems stock outs are not allowed at all and, hence, shortages

are explicitly prohibited. See MirHassani (2008) or Cafaro and Cerdá (2008) for examples where both

costs are considered. Note that in Cafaro and Cerdá (2008) shortages are deﬁned as tardily delivered

batches.

43

E.g. see Relvas et al. (2006) or Rejowski and Pinto (2008).

69

Both directions commonly assume that the pipeline remains completely ﬁlled over the

total time horizon. The pipeline is fed by only one tank at a time and can only feed one

tank at a time. Conversely, a tank can either be ﬁlled or emptied at a time. It is assumed

that for each product a set of tanks is available and it has to be determined when to use

which tank for pumping.

Magatão et al. (2004) assume a one-to-one pipeline system with reversal ﬂow option.44

The objective is to minimize interfaces and tank change-overs. The problem is modelled

as a large-scaled, time-discrete mixed-integer program (MILP). Magatão et al. (2005)

provide a solution approach relying on the decomposition of the monolithic MILP into

two sub-MILPs and a main MILP.

In contrast, Relvas et al. (2006) assume a unidirectional pipeline ﬂow system. The

corresponding problem is formulated as a time-continuous MILP. Time-discrete customer

demands are dynamically assigned to certain batches depending on the batches’ arrival

times. The objective is to maximize the utilization of the pipeline, i.e. to maximize the

product ﬂow from the provider to the customer depot. Originally, the pipeline’s ﬂow

rate is ﬁxed and the pipeline has to be operated continuously. In Relvas et al. (2007)

the original model is extended by relaxing both restrictions. To deal with the increased

complexity, Relvas et al. (2009) propose an eﬃcient heuristic for determining desirable

product sequences.

In principle both types of operational pipeline planning problems can be applied to

chemical pipeline systems, too, as the technical speciﬁcations in both industries are sim-

ilar. However, in chemical industry the distribution of ﬁnal chemicals is rarely based on

multi-product pipelines. More likely, supply processes of raw and intermediate materials

are performed by pipeline transports. This implies that chemical production sites exist

at both locations. As most chemical production sites are continuously operated, it can be

presumed that in normal operation a constant demand or surplus per period of time is

observed. I.e. there is a ﬁxed (net) consumption and a ﬁxed (net) production rate for each

product and site. The ﬂow rate between both sites is then the minimum of consumption

and production rate. The remaining surplus or deﬁcit has to be handled diﬀerently, not

aﬀecting the pipeline operations planning at all. If the production system is in balance,

both rates are equal for each product. In any case, it can be concluded that local con-

sumption and production rates for each product are assumed to be equal. Hence, in the

long run a pumping sequence has to be determined minimizing the total costs consisting

of stock holding and transition costs.

Let ω denote the production and consumption rate for a certain product and ρ denote

the pump capacity at which the pipeline is operated. Furthermore, let T denote the cycle

time between two successive batches of the considered product and τ denote the time

necessary to transport a batch of a product from provider to consumer location.45 The

44

A many-to-many pipeline network with reversal ﬂows is provided in Moura et al. (2008).

45

I.e. it is assumed that the pipeline is continuously operated at a constant pump rate for all products.

70

for τ periods in the pipeline. Assuming that ρ > ω, the time span necessary to ﬁll the

pipeline is T f ill = Fρ = T ⋅ ωρ . Hence, the stock-up time is T stock = T − T f ill = T ⋅ (1 − ωρ )

and the maximum stock level L obtained at both locations is calculated by multiplying

the stock-up time T stock with the production rate ω, thus, L = ω ⋅ T stock = ω ⋅ T ⋅ (1 − ωρ ).

The stock levels obtained in pipeline and inventories at both locations are depicted in

Figure 3.5.

Since production and consumption rates are equal, the total stock in the system is

constant at the level L̄(T ) = L + τ ⋅ ω = ω ⋅ (τ + T ⋅ (1 − ωρ )) and depends only on the cycle

length T . The holding costs decrease with decreasing cycle length up to the theoretic

minimum of τ ⋅ ω in case of a permanent material ﬂow.46 The total stock level splits into

the average stock level at the sites with L2 and the average stock in the pipeline given by

τ ⋅ ω.

However, as a counterbalance, the smaller the batches the higher the number of product

transitions and related costs. Since each interface is associated with some ﬁnancial eﬀort,

smaller batches increase transition costs. Hence, there is the typical structure of lot-sizing

problems and a strong analogy to the economic lot scheduling problem (ELSP).47

In general the ELSP aims at identifying the optimal recurring sequence and size of

product batches on a single machine in order to minimize the sum of holding and set-

up/transition costs. More formally, for a given set of products S there are constant

consumption rates ωs , production rates ρs , set-up cost rates cset

s , and holding cost rates

chold

s .48 The total expected costs per period are calculated by

cset chold ⋅ ωs ⋅ Ts ωs

T C ELSP = ∑ ( s

+ s ⋅ (1 − )) (3.7)

s∈S Ts 2 ρs

determining the cycle times Ts is caused by the feasibility constraint, i.e. that the batches

scheduled on the machine are not overlapping at any time. In fact, it can be shown that

ﬁnding the optimal solution and checking a schedule for sequence feasibility is both N P-

hard.49 Hence, there is a lot of literature about heuristics for this problem.50 Sequence

feasibility refers to the fact that based on the cycle times Ts and production/pumping

times Ts ⋅ ωρss a schedule can be build. This fundamental schedule/cycle has a total length

of T cycle = max Ts and is repeated inﬁnitely. I.e. every T cycle periods this schedule starts

s∈S

anew. Hence, it has to be assured that each product is exactly scheduled every Ts periods

also across subsequent fundamental cycles.

46

Though, the theoretical minimum is possible in the one-product case only.

47

See e.g. Narro Lopez and Kingsman (1991) for a review.

48

In analogy to the production states of production plant introduced above, a product s transported via

the pipeline deﬁnes the current state of the pipeline.

49

I.e. there is no algorithm known that can solve this problem in polynomial time. See Hsu (1983).

50

See e.g. Chatﬁeld (2007) for a concise review.

Provider

12

L

8

Δ

Δω − Δ(ρ − ω)

Δ

4

stock level

● ●

0

00 ⎛ ω⎞ 6

1T ⎛ ω⎞ 12

2T Time

T 1− T 2−

⎝ ρ⎠ ⎝ ρ⎠

Pipeline

12

F

10

L

8

6

Δρ − Δρ

4

Δ Δ

stock level

2

0

● ●

0 6 12 Time

0 τ ⎛ ω⎞ 1T τ+T ⎛ ω⎞ 2T τ + 2T

T 1− T 2−

⎝ ρ⎠ ⎝ ρ⎠

Consumer

12

10

L

8

6

Δ

Δ(ρ − ω) − Δω

Δ

stock level

2

0

● ●

0 ω 6 12 Time

0 τ 1T τ + T⎛1 − ω⎞ τ + T 2T τ + T⎛2 − ω⎞ τ + 2T

τ−T

ρ ⎝ ρ⎠ ⎝ ρ⎠

71

72

The proposed problem with equal consumption and production rates can be seen as

a special case of the classic ELSP by assuming constant pumping rates for all products,

i.e. ρs = ρ and no set-up times. Additionally, in this case the cost function has to be

adjusted according to the total stock level in the entire supply system (provider, pipeline,

and consumer stock). Hence, (3.7) changes to

cset ωs

T C ELSP −BP = ∑ ( s

+ chold

s ⋅ ωs ⋅ (τ + Ts ⋅ (1 − ))) (3.8)

s∈S Ts ρ

where τ denotes the pumping time to transport the material through the pipeline and

L̄ = ωs ⋅ (τ + Ts ⋅ (1 − ωρs )) is the total stock level along the pipeline. If ∑s∈S ωs ≤ ρ, a

feasible schedule always exists.51 If sequence feasibility is neglected, the optimal cycles

Ts∗ can be derived from (3.8) with respect to Ts which results in

1

2 cset

Ts∗ =2

2

3

s

. (3.9)

chold

s ⋅ ωs ⋅ (1 − ωρs )

If the set of independent solutions results in a schedule feasible w.r.t. sequence and ca-

pacity, this is the optimal solution for (3.8) at the same time. Otherwise, heuristics for

the classic ELSP can be applied.52

If subsequent batches in the pipeline are separated by interfaces, the transition costs

may depend on the product sequence as e.g. reprocessing costs may diﬀer for diﬀerent

mixtures. Therefore, a sequence-dependent economic lot sequencing problem (sELSP)

has to be formulated to represent the case of interface separation. To reformulate (3.8),

let ctrans

st denote the transition costs for subsequent batches of products s and t.

Then, the set-up cost part per period of time is replaced by the total transition costs

occurring in a fundamental cycle divided by the fundamental cycle’s total length

1

min → T C sELSP −BP = ( ∑ ∑ ∑ (ctrans ⋅ xsti )) (3.10)

T cycle s∈S t∈S i∈I st

ωs

+ ∑ (chold

s ⋅ ωs ⋅ (τ + Ts ⋅ (1 − ))) .

s∈S ρ

hold

Note that replacing the stock holding costs part in (3.10) by ∑s∈S cs 2ωs Ts (1 − ωρs ) leads

to the classic sELSP.53

To assure sequence feasibility, a mixed-integer non-linear program (MINLP) is formu-

lated for the ELSP-BP and sELSP-BP. Table 3.6 contains the required set of parameters

and variables.

Sequence feasibility requires that the lot sizes are determined such that the reach of

51

This can be seen as an adaptation of the so-called capacity feasibility constraint of the classic ELSP,

see e.g. Chatﬁeld (2007).

52

For example Dobson’s approach (Dobson, 1987) or Haessler’s approach (Haessler, 1979).

53

See Wagner and Davis (2002) or Dobson (1992).

73

Sets

S = 1, ..., S set of chemicals

I = 1, ..., I set of lot positions

Parameters

M large number

τ pumping time

ρ pump rate

ωs demand rate of chemical s ∈ S

ctrans

st transition costs for a transition from chemical s ∈ S to t ∈ S (e.g. due to

PIG insertion or interface handling)

chold

s holding costs for chemical s ∈ S

Decision variables

bsi binary, 1 if at position i = 0, ..., I a chemical s ∈ S is scheduled

Ts cycle time of chemical s ∈ S

tbi starting time of position i ∈ I

Variables

xsti binary, 1 if from position i−1 to position i a transition from chemical s ∈ S

to t ∈ S occurs

osij binary, 1 if chemical s ∈ S is lastly scheduled j positions before i ∈ I

rsi binary, 1 if position i ∈ I is the last position where chemical s ∈ S is

scheduled

rrs binary, 1 if chemical s is the ﬁnally scheduled chemical in the fundamental

cycle

T cycle fundamental cycle time

Table 3.6: Set of parameters and decision variables for the sELSP

a certain chemical’s lot covers the time span until the production of the next lot of this

chemical starts. Therefore, assume that for a set of chemicals S a set of lot positions I is

available, with ∣S∣ ≤ ∣I∣. If the number of lot positions equals the number of chemicals, the

inequality Ts ≥ ∑t∈S Tt ⋅ ωρt = T cycle holds for each chemical’s cycle time Ts . In this case, each

chemical is scheduled at exactly one position in the fundamental cycle. The sequence with

minimum transition costs can be determined by solving a Travelling Salesman problem

where the transition cost rates ctrans

st serve as distances between cities. However, such a

solution implies similar production rates ωs and similar cost rates for each product. If

these parameters vary, an optimal fundamental cycle can be composed by heterogeneous

individual cycle lengths Ts . As a consequence, chemicals with short cycles have to be

scheduled more than once in the fundamental cycle. Hence, either ∣I∣ > ∣S∣ or ∣I∣ ≫ ∣S∣

must hold for the number of positions. Sequence feasibility in such a schedule requires that

for each chemical the time span between two successive lots of this chemical is equal to

its cycle time Ts . Formally, the set of constraints (3.11)-(3.27) assures sequence feasibility

74

ωs

tbi+1 ≥ tbi + Ts ⋅ − M ⋅ (1 − bsi ) ∀i ∈ I, s ∈ S (3.11)

ρ

ω s

T cycle ≥ tbi + Ts ⋅ − M ⋅ (1 − bsi ) ∀i ∈ I, s ∈ S (3.12)

ρ

Ts ≥ tbi + T cycle − tbj − M ⋅ (3 − osii − bsi − rsj ) ∀i, j ∈ I, j ≥ i, s ∈ S (3.13)

Ts ≥ tbi − tbi−j − M ⋅ (3 − osij − bsi−j − bsi ) ∀i, j ∈ I, j < i, s ∈ S (3.14)

Ts ≤ tbi + T cycle − tbj + M ⋅ (3 − osii − bsi − rsj ) ∀i, j ∈ I, j ≥ i, s ∈ S (3.15)

Ts ≤ tbi − tbi−j + M ⋅ (3 − osij − bsi−j − bsi ) ∀i, j ∈ I, j < i, s ∈ S (3.16)

I

rsi ≥ bsi − ∑ bsl ∀s ∈ S, i ∈ I (3.17)

l=i+1

∑ rsi = 1 ∀s ∈ S (3.18)

i∈I

∑ bsi ≥ 1 ∀s ∈ S (3.19)

i∈I

∑ bsi ≤ 1 ∀i ∈ I (3.20)

s∈S

bs0 = 1 ∀s ∈ S (3.21)

∑ bsi ≤ ∑ bsi−1 ∀i ∈ I (3.22)

s∈S s∈S

bsi−1 + bti

xsti ≤ ∀i ∈ I, i > 1, s, t ∈ S (3.24)

2

xst1 ≥ rrs + bt1 − 1 ∀s, t ∈ S (3.25)

rrs + bt1

xst1 ≤ ∀s, t ∈ S (3.26)

2

i−1

osij ≥ bsi−j − ∑ bsl ∀s ∈ S, i, j ∈ I (3.27)

l=i−j+1

i

∑ osij = 1 ∀s ∈ S, i ∈ I (3.28)

j=1

I

rrs ≥ rsi − ∑ ∑ rtj ∀s ∈ S, i ∈ I, i < I (3.29)

t∈S j=i+1

Inequalities (3.11) assure that the starting time of a lot at position tbi+1 is not smaller

than the previous lot’s starting time tbi plus the time required to pump the quantity of the

scheduled chemical at position i. Constraints (3.12) restrict the fundamental cycle time

T cycle of the fundamental cycle to the maximum of all starting times tbi (plus correspond-

ing pumping time). The time span between two successive lots of a certain chemical s is

exactly Ts . There are two cases to be distinguished: First, when a new fundamental cycle

starts, the time span between the last position, say i, at which this chemical is scheduled

in the old fundamental cycle (rsi = 1) and the ﬁrst position in the new fundamental cycle,

75

say i, at which the chemical is scheduled has to be restricted. Then, osii = 1 and bsi = 1

indicate that chemical s is scheduled at position i and has been lastly scheduled i periods

before, which is in period 0 or in the previous fundamental cycle. This case is modelled

by constraints (3.13) and (3.15).

Second, when a chemical is scheduled at least at two positions in a fundamental cycle,

say i and k with i > k, and nowhere in between, then this chemical is previously scheduled

j = i − k positions before position i. Hence, osi(i−k) = osij = 1 and bsi = bsk = bs(i−j) = 1.

Thus, constraints (3.14) and (3.16) ensure that tbk − tbi = Ts .

Constraints (3.17) and (3.18) deﬁne the last position at which each chemical s is sched-

uled. Constraints (3.19)-(3.20) assure that each chemical is at least scheduled once in the

fundamental cycle. No more than one chemical is scheduled at each position. Constraints

(3.21) indicate that all chemicals are scheduled in the (previous) fundamental cycle.

Constraints (3.22) avoid intermediate idle positions within the fundamental cycle. Con-

straints (3.23)-(3.26) set the transition variables xsti = 1 iﬀ bsi−1 = 1 and bti = 1. Here,

constraints (3.25)-(3.26) deﬁne xst1 = 1 iﬀ chemical s is the last chemical and chemical t

is the ﬁrst chemical scheduled in a fundamental cycle.

Constraints (3.27)-(3.28) set osij = 1 iﬀ chemical s is previously scheduled at position

i − j (i.e. bsi−j = 1). Finally, constraints (3.29) set rrs = 1 iﬀ chemical s is the chemical

scheduled at last in a fundamental cycle.

The sELSP-BP constituted by (3.10)-(3.28) refers to the situation when transition costs

are caused by eﬀorts for PIG injection to separate the batches. However, if no PIG is

used for separation, an interface is built. This interface has to be treated like one of the

two parental chemicals or is specially treated.

In the former case, some adaptations are necessary. First, let dst represent the quantity

of material which is subtracted from (dst < 0) or added to (dst > 0) a batch due to an

interface of the chemical sequence s → t. Note that this quantity typically does not depend

on the quantity of the parental batches but on the pressure and diameter of the pipeline.54

As both parameters are assumed to be constant, dst is a parameter. Moreover, it holds

that dst = −dts as the loss of one parental chemical is the surplus of the other one.55

The size of a batch received at the consumer location depends now on its predecessor

and successor. In turn, the injected batch size has to compensate the correction of the

received batch size due to the interfaces. Let qsi denote the total batch size correction at

the consumer location for the batch at position i and material s. Then, qsi = ∑t∈S dts ⋅xtsi +

∑u∈S dsu ⋅ xsu(i+1) holds. Note that due to the speciﬁc structure of dst also ∑s∈S ∑i∈I qsi = 0

holds, i.e. no quantity is lost. Figure 3.6 illustrates a situation for three subsequent batches

of chemicals t, s and u at positions i − 1 to i + 1, respectively.

54

This holds as long as the batch is larger than the critical mixing quantity. See e.g. Hall and Nicholls

(1980) or Cafaro and Cerdá (2004).

55

Note that dst = −dts is not a necessary condition as it implies equal parental shares constituting the

interface. It suﬃces to force sign(dst ) = −sign(dts ) where sign(⋅) denotes the sign function. However,

this generalization requires some more notation and, therefore, is not discussed here.

76

chemical t chemical s chemical u

batch - dts

batch + dst − dsu batch + dus

As Figure 3.6 shows, the interface between chemicals t and s is added to the batch of

chemical s (i.e. dts < 0 and dst > 0). In contrast, the interface between chemicals s and u is

added to the batch of chemical u (i.e. dus < 0 and dsu > 0). It holds bt(i−1) = bsi = bu(i+1) = 1

and it follows from (3.23) and (3.24) that xtsi = xsu(i+1) = 1. Consequently, qsi = dts −dsu < 0

holds since ∣dts ∣ < ∣dsu ∣ and the batch of chemical s at position i is reduced by qsi units.

The batch size of a chemical s is given by Ts ⋅ ωs , i.e. due to the interfaces the realized

batch size is Ts ⋅ ωs + qsi . Due to the correction of the batch sizes, the pumping times have

to adjusted accordingly. I.e. for qsi > 0 pump times can be reduced such that the pumping

time expression changes to Ts ⋅ωρs −qsi . Thus, (3.11) and (3.12) change to

Ts ⋅ ωs − qsi

tbi+1 ≥ tbi + − M ⋅ (1 − bsi ) ∀i ∈ I, s ∈ S (3.30)

ρ

Ts ⋅ ωs − qsi

T cycle ≥ tbi + − M ⋅ (1 − bsi ) ∀i ∈ I, s ∈ S (3.31)

ρ

which implies that the pumping times depend on the pumping sequence, too. Additionally,

the objective function is aﬀected since the batch size corrections depend on the sequence.

I.e. (3.10) changes to

1

min → T C sELSP −BP −IF = ( ∑ ∑ ∑ (ctrans ⋅ xsti )) (3.32)

T cycle s∈S t∈S i∈I st

ωs

+ ∑ (chold

s ⋅ (∑ qsi + ωs ⋅ (τ + Ts ⋅ (1 − )))) .

s∈S i∈I ρ

The remaining constraints (3.13)-(3.28) remain unchanged. This constitutes the sELSP

model with global stocks and interface handling (sELSP-BP-IF).

A third option to handle interfaces is extraction and external post-processing. This

implies that all batches reduce to a certain amount due to the associated interfaces. To

reﬂect this speciﬁc interface handling procedure in the sELSP-BP-IF, the values of dst

have to be strictly negative such that qsi are negative, too. Note that this implies a total

loss of material due to the pumping process.

Note further that the MINLPs introduced above are hard to solve to optimality even

for very small problem sizes. Presumably, heuristics developed for the classic ELSP

77

can be adopted to the sELSP, sELSP-BP, and sELSP-BP-IF with a little eﬀort, e.g. by

solving a TSP-like sub-problem to account for the sequence dependency.56 However, the

development of eﬃcient heuristics is not in the scope of this work. For the relevant

problem sizes with a few chemicals discussed here, optimization is still possible within

reasonable time (i.e. within some minutes at most). To show the applicability of the

sELSP-BP and sELSP-BP-IF, the following example 5 describes a case study from basic

chemical industry.

56

Another approach which promises a good prospect for adoption is the genetic optimization approach

of Chatﬁeld (2007).

78

Example 5 (Batch pipeline sequencing). Suppose a one-to-one serial pipeline that con-

nects two chemical sites. The provider site produces the raw and intermediate materials

for the consumer site. Suppose that the consumer site produces Styrene-based chemicals

such as Polystyrene and Styrene-Butadiene rubber. Raw materials for Styrene production

are Benzene and Ethylene. Among others, a steam cracker provides Ethylene and Pygas,

which is subsequently reﬁned to Benzene. It is assumed that the capacities of these plants

are unbalanced such that the deﬁcit of Pygas and Benzene has to be imported. Addition-

ally, the cracker feed, Naphtha, has to be imported.

These three chemicals are mixable, chemically related liquids and can be transported

via the same pipeline. Pygas and Naphtha are mixtures of hydrocarbons, whereby Pygas

can be interpreted as a sub-mixture primarily consisting of aromatic chemicals with main

component Benzene. At the provider site these three chemicals are surplus and, thus, are

transported via a pipeline. Ordering these chemicals to increasing purity57 leads to the

sequence Naphtha, Pygas, Benzene. Assume that two possibilities for pipeline operation

are discussed: Either the batches are separated by PIGs such that no interfaces occur or

the interfaces are "downgraded" to the less pure parental chemical.

The transition costs in case of PIG separation depend on the monetary eﬀort for PIG

handling, e.g. redistribution of PIGs as well as depreciation of the PIG due to capital

commitment. Since Benzene is a highly reactive chemical it is assumed that two kinds of

PIGs are used. For separation of Naphtha and Pygas a PIG with lower investment costs

can be used whereas for the separation of Benzene a more robust device has to be used

which induces higher investments and, thus, higher capital commitment costs. Table 3.7

shows the transition costs ctrans

st given for this scenario.

st [e]

s Naphtha Pygas Benzene

Naphtha 0 500 1,000

Pygas 500 0 1,000

Benzene 1,000 1,000 0

st for the PIG insertion scenario

For the downgrading option the interfaces are re-processed. E.g. an interface consisting

of Pygas and Naphtha is handled as Naphtha and stored in the Naphtha tanks where

the concentration of aromatics increases due to the Pygas inﬂow. The cracking process

separates these components again. This implies in turn, that downgraded ingredients

are re-processed leading to additional processing costs. I.e. an interface of Benzene and

Naphtha has to be downgraded to Naphtha and the Benzene part is re-processed twice.

Hence, sequence-dependent transition costs ctrans

st are calculated by the quantity to be re-

57

Purity refers to the fact that Naphtha and Pygas are mixtures of multiple hydrocarbons whereby

Pygas is a sub-fraction of Naphtha (obtained after cracking) and Benzene is a pure component of

Pygas (obtained after distillation).

79

processed multiplied with the processing cost rate and the number of re-processing steps.

The processing cost rate is assumed to be 100 e per ton for cracking of Naphtha and

separation of Benzene from Pygas as well. Table 3.8 shows the interface quantities dst

and associated transition costs ctrans

st for this scenario.

st [e]

s Naphtha Pygas Benzene Naphtha Pygas Benzene

Naphtha 0 10 50 0 1,000 10,000

Pygas -10 0 10 1,000 0 1,000

Benzene -50 -10 0 10,000 1,000 1,000

st for the interface scenario

Typically, it can be assumed that the market value of (basic) chemicals increases with

passing more processing steps. Hence, stock holding costs increase, too. The net demand

rates provided in tons per day for Naphtha, Pygas, and Benzene under normal processing

conditions can be found in Table 3.9 accompanied by stock holding cost rates in e per ton

and day.

s chold

s [e/(tons⋅day)] ωs [tons/day]

Naphtha 0.10 3,000

Pygas 0.15 1,000

Benzene 0.20 2,000

Table 3.9: Net demand rate ωs and holding cost rates chold

s for both scenarios

ρ = 7, 000 tons per day. Obviously, the pipeline capacity suﬃces to supply the consumer site

with all three raw and intermediate chemicals. The pipeline is assumed to be operated at

maximum capacity (as operating costs are optimized for this pump rate) or to be completely

inactive.

In this example two scenarios for pipeline operation are investigated:

• interface downgrading.

To determine the optimal pumping schedule, for the ﬁrst scenario an sELSP-BP and

for the second scenario an sELSP-BP-IF has to be solved. To compare these scenarios,

both sELSP-BP and sELSP-BP-IF are solved to optimality under the parameter setting

provided in Table 3.9 as well as Table 3.7 or Table 3.8, respectively. Up to 15 lot positions

are allowed in each model (i.e. ∣I∣ = 15).58 Both solutions diﬀer substantially with total

58

Both MINLPs are solved on an Intel Core 2 Q6700 at 2.66 GHz with BONMIN in about 140 seconds, see

Bonami and Lee (2007).

80

costs per period of 3,270 e and 3,737 e for the sELSP-BP and sELSP-BP-IF, respectively.

Table 3.10 shows the cycle times as well as starting and pump times for both models.

model

Naphtha Pygas Benzene 1 2 3 4

sELSP-BP 2.066 2.066 2.066 0 0.417 1.023 –

sELSP-BP-IF 2.769 1.385 2.769 0 0.428 1.385 1.583

Table 3.10: Resulting optimal pumping cycles for the sELSP-BP and sELSP-BP-IF

It can be taken from Table 3.10 that for the sELSP-BP a unique cycle time for all

products is the optimal solution such that each chemical is pumped every 2.066 days.

For the sELSP-BP-IF the fundamental cycle time is 2.769 days which is probably caused

by the generally higher level of transition costs forcing larger batches. The solution of

the sELSP-BP-IF yields more diverse pumping times. Here, Pygas is scheduled twice in a

fundamental cycle. This is caused by the exceedingly high transition costs between Naphtha

and Benzene which induces an intermediate batch of Pygas between Naphtha and Benzene

batches. Table 3.11 shows the pumping times of the chemicals and the idle times for each

batch position.

model

Naphtha Pygas Benzene 1 2 3 4

sELSP-BP-IF 1.186 0.198 0.794 0.231/2 0.162/3 0.001/2 0.001/1

Table 3.11: Pumping times and idle times of the optimal schedules for the sELSP-BP and

sELSP-BP-IF

It can be observed in Table 3.11 that for the sELSP-BP-IF solution pumping times of

Naphtha and Benzene are prolonged in favour of a decrease of the Pygas batch size since

Pygas is scheduled twice. Note that the pumping times for the sELSP-BP-IF solution also

reﬂect the correction due to the interfaces. Figure 3.7 illustrates the fundamental cycle

and the associated interface quantities.

It can be observed that q31 = −20 and q13 = 20 which results in an increase of the

pumping time of Benzene and a decrease of the pumping time of Naphtha by approximately

700 ⋅ 24 ⋅ 60 ≈ 4 minutes which is almost negligible.

2

Comparing both scenarios merely based on the period cost, the PIG separation shows a

cost advantage. However, not all costs are included in this model since the redistribution

of the PIGs is not considered in the sELSP-BP. Moreover, investment and maintenance

of the PIGs causes additional costs. In result, the moderate advantage in period costs

59

Behind the idle time the index of the chemical scheduled at position i is provided separated by a slash.

Indices of chemicals are: Naphtha...1; Pygas...2; Benzene...3.

81

3 2 1 2

compared to the interface downgrading option may vanish if an encompassing cost analysis

is conducted.

As the literature on pipeline scheduling has its roots in the petrochemical industry, a

straight-forward extension from batch ﬂow pipeline systems to batch split systems is

natural. Here, typically a single source has to serve a set of distinct sinks with a set

of (petro-)chemicals. Typically, the source represents a reﬁnery and the sinks represent

local fuel depots from which the local customer demand is served. Conversely, a harbour

may also serve as a supplier of diﬀerent types of raw materials for a set of reﬁneries.60

These sinks are connected to the source via a serial pipeline system. At the source and

the sinks a set of storage tanks is available for each product. Common assumptions for

the proposed models in this branch of literature are summarized as follows:

• General characteristics

– ﬁnite time horizon

– interface re-processing

– known and deterministic demands for each product at each depot

– known and deterministic production rates for each product at the source

• Pipeline system

– unidirectional ﬂow

– pump rate range

– initial pipeline ﬁlling

– constant pressure

• Tank system

– tank level ranges

– initial tank levels.

60

See e.g. Más and Pinto (2003).

82

In the following, the literature focusing on batch split pipeline scheduling is reviewed.

Integrative approaches containing pipeline transports in a supply chain context are out

of scope in this paragraph.61 Such approaches provide crude pipeline models to reﬂect

the transports in production networks over time but simplify technical details of pipeline

operation (such as the precise material tracking in time or interface handling) for the sake

of solvability of the entire model.62

Regarding the objectives, all approaches try to minimize the total costs of pipeline

operations over the planning horizon. Four cost categories are typically considered:

• Pumping costs (in terms of energy consumption multiplied with energy cost rate)

• shortage costs

The proposed approaches diﬀer in the considered cost categories and in the type of

formulation which is either a linear or a non-linear model. Inventory costs depend on the

duration and level of stock holding. Hence, if stock levels and the time periods are decision

variables, inventory cost calculation is non-linear as both variables have to be multiplied.

The pumping costs primarily depend on the energy consumption for pumping. Typically,

pipelines are assumed to be operated at constant pressure denoted by the constant pump

rate. In some cases, however, an adjustment of the pump rate is technologically possible

and economically reasonable, e.g. to speed up the transport of the chemicals. If this is the

case, the pump rate determines the duration of pumping. As the size of an injected batch

is in any case a decision variable, the time for pumping is represented by the ratio of batch

size and pump rate which determines the total energy consumption and, consequently,

the pumping costs. Hence, pumping costs are non-linear if the pump rate is a decision

variable. Given these criteria the literature classiﬁcation scheme is set up as follows:

The indicator κ ⊆ (p, h, b, i) is a tuple representing the considered cost categories where p

indicates pumping costs, h indicates inventory holding costs, b indicates backorder costs,

and i indicates interface costs. Note that not all cost categories are considered in all

61

Examples for such integrative approaches can be found in Pitty et al. (2008); Neiro and Pinto (2004);

Pinto et al. (2000).

62

See Rejowski and Pinto (2008) and the remarks about the literature review therein.

83

references, i.e. ∣κ∣ ≤ 4. The tuple λ ∈ {n, l}∣κ∣ indicates whether the corresponding cost

category is linearly or non-linearly modelled.

The second classiﬁcation category distinguishes between approaches with continuous

or discrete formulation for time and pipeline. In a continuous formulation, continuous

coordinates are used to describe e.g. the position of a batch or a pipeline branch. The in-

dicator ν ∈ {c, d}2 indicates whether a continuous or discrete formulation is chosen, where

the ﬁrst component refers to the time aspect and the second refers to the pipeline. In

general, time-discrete formulations are more intuitive, but to the costs of larger model di-

mensions in terms of the number of equations and variables. In contrast, time-continuous

formulations are more compact but less intuitive.63

For the pipeline a discrete formulation implies that the pipeline is represented as a

sequence of discrete segments. This eases the tracking of product batches along the

pipeline and the formulation of product removals at the depots. Table 3.12 summarizes

the relevant literature according to the proposed classiﬁcation scheme.

Reference κ λ ν

Rejowski and Pinto (2003) (p, h, i) (l, l, l) (d, d)

Cafaro and Cerdá (2004) (p, h, i) (l, l, l) (c, c)

MirHassani (2008) (i) (l) (d, d)

Cafaro and Cerdá (2008) (p, h, b, i) (l, l, l, l) (c, c)

Rejowski and Pinto (2008) (p, h, i) (n, n, l) (c, d)

MirHassani and Fani Jahromi (2011) (p, h, i) (l, l, l) (c, d)

The literature referred to in Table 3.12 aims at minimizing the total operational costs

for a planning period. Most works take stock holding costs, pumping costs, and costs

for reprocessing the interface into account. Two exceptions are noteworthy: MirHassani

(2008) and Cafaro and Cerdá (2008).

MirHassani (2008) proposes a comparatively simple pipeline scheduling model consid-

ering only interface reprocessing costs to determine optimal schedules. In this model the

pipeline is subdivided into equally-sized segments. Similarly, the time horizon is subdi-

vided into equal periods of time. However, in contrast to most other models, a branching

pipeline is modelled. I.e. the depots are not located at a serial main pipeline, but con-

nected via a sub-pipeline that branches from the main pipeline. Figure 3.8 depicts both

types of pipeline systems with one source (S1) and three depots (D1, D2, D3).

Albeit the problem structure for branching pipelines is generally similar to a pipeline

structure with multiple depots located along one main pipeline, it complicates the

mathematical formulation by forcing to track batches along the pipeline branches

which equals the modelling of additional pipelines. In the follow-up paper (MirHas-

63

I.e. they contain less decision variables. See e.g. Rejowski and Pinto (2008) or Maravelias and Gross-

mann (2003) for a brief or more extensive discussion, respectively.

84

D1

S1 D1 D2 D3 S1 D3

D2

sani and Fani Jahromi (2011)), the mathematical formulation is improved towards a

time-continuous model showing a better performance which is able to handle larger

problem instances.

Most of the literature on pipeline transportation considers customer demands for the

products at the depots to be served at the end of the planning horizon. An exception is

proposed in Cafaro and Cerdá (2008) where customer demands at the depots are associ-

ated with due-dates such that backorder costs are incorporated in the objective function.

Moreover, the work proposes a rolling horizon model for updating and re-scheduling pre-

viously determined schedules according to updated demand characteristics.

Beside the incorporation of backorder costs in multi-period models, the handling of

pipeline operation costs diﬀers. In general, pipeline operation costs are driven by the

energy consumption which, in turn, depends on the quantity and speed of pumping. As

stated before, most pipeline models are adapted for a speciﬁc pump rate, i.e. models are

based on a ﬁxed pump rate or pump rate range (see Rejowski and Pinto (2003); Cafaro

and Cerdá (2004, 2008)). However, these works incorporate time-varying energy cost

rates depending on the time of day. I.e. in speciﬁed periods (typically when total energy

consumption in a region is high) higher energy cost rates have to be faced. Hence, pipeline

operation is shifted in tendency to less expensive periods.

In contrast, in Rejowski and Pinto (2008) the speed of pumping is a decision variable

which leads to a non-linear term in the objective function since both pump rate and

batch length have to be determined. Furthermore, this work is noteworthy as it explicitly

accounts for the inventory holding costs, but at the expense of another non-linear term

in the objective function. Note that the time-continuous formulation requires that both

the period’s length and the tank level in a period are decision variables. The product of

these terms determines the holding costs. Cafaro and Cerdá (2004, 2008) approximate

the holding costs by averaging the stock levels at the depots.

Despite the complexity of the scheduling decisions, all proposed models can be solved

with standard solvers in a reasonable amount of time.64 This is because only small problem

instances are solved w.r.t. the number of products to be scheduled and the number of

64

Either e.g. CPLEX in the MILP case (see e.g. Rejowski and Pinto (2003); Cafaro and Cerdá (2008))

or e.g. CONOPT in the MINLP case (see Rejowski and Pinto (2008)).

85

hardly more than ten products. While the number of products to be transported is

limited, the time domain is not. Most approaches are rather short-term models that have

to be applied to a rolling horizon environment.65

To tackle more realistic planning problems of real-world pipeline systems, models for

multi-source pipeline systems have been developed recently. This extension of one-to-

many systems considerably increases complexity. Hence, ﬁrst approaches provide a heuris-

tical decomposition that decomposes a problem into three blocks:66

1. Allocation decision: determine which locations are potential candidates for injecting

and receiving batches.

3. Batch scheduling: determine the exact starting and receiving dates for each batch.

the neglected interdependency of the decomposed decisions.67 Therefore, monolithic,

time and volume-continuous MILP formulations are proposed recently to integrate these

decisions.68 These models are straight-forward adaptations of one-to-many approaches.69

The basic setting is adapted from the one-to-many case (see the previous paragraph),

i.e. a serially operated main pipeline is assumed with a reﬁnery at its head and depots

along this pipeline. One key alteration is that depots now can receive and inject batches

in the pipeline. Most other assumptions are inherited from the one-to-many systems. In

particular, the product ﬂow is still unidirectional and only one location at a time can

inject a batch in the pipeline.70

However, the one-at-a-time restriction blocks all depots except for the injecting and the

receiving location. Only pipeline segments connecting these depots are active, whereas

the remaining pipeline segments are idle. In general, however, there are no technical

restrictions prohibiting simultaneous product ﬂows on idle pipeline segments. Hence, si-

multaneous pump runs may cause a drastic increase in the pipeline schedule’s eﬃciency.

Simultaneous pump runs are only possible if the used pipeline segments do not overlap,

i.e. simultaneous ﬂows do not use the same pipeline segments. Hence, restrictions for as-

suring "non-overlappedness" of simultaneous pump runs have to be incorporated. Despite

65

See e.g. Cafaro and Cerdá (2008).

66

See e.g. Boschetto et al. (2008). Another decomposition approach is proposed by Moura et al. (2008).

Here, in the ﬁrst stage injection and destination locations as well as batch sizes are determined which

are subsequently scheduled in the second stage.

67

See Boschetto et al. (2008).

68

See Cafaro and Cerdá (2009) and Cafaro and Cerdá (2010).

69

To be precise, these are extensions of Cafaro and Cerdá (2004, 2008).

70

See Cafaro and Cerdá (2009).

86

this increase in complexity, the resulting extended MILP can still solve realistic problem

instances to optimality in a reasonable amount of time. Surprisingly, for some examples

computation times are decreasing compared to the one-at-a-time approach.71 More im-

portant, the resulting schedules are more compact (increasing the pipeline utilization)

and reduce total operation costs.

In general, many-to-many systems include all other pipeline system conﬁgurations as

special cases. Since computational complexity of the models proposed in Cafaro and

Cerdá (2009, 2010) is limited and the general characteristics of pipeline system are the

same as for the previously described systems, the model proposed by Cafaro and Cerdá

(2010) can be seen as a generic basic model encompassing the previously published works

on one-to-many and many-to-many pipeline systems.

With respect to the technical characteristics of pipeline transports in basic chemical

industry the diﬀerences to petrochemical industry are limited. However, multi-product

pipelines in basic chemical industry are typically less widespread. Here, single-product

pipelines are dominating. For local material transports at chemical production sites,

multi-product pipelines do not make sense because a constant ﬂow of materials among

the set of local production plants has to be realized. For trans-regional transports, multi-

product pipelines might be used although most intermediate and basic chemicals are

chemically not very similar. This induces that ﬁnding a common technical standard

conﬁguration of a multi-product pipeline is a challenging task. Incompatible chemicals

can not be transported without physical separation. Costs for separation or interface

reprocessing are often very high. Consequently, if multi-product pipelines are used in the

basic chemical industry, the number of diﬀerent chemicals transported is usually smaller

than in petrochemical industry. Moreover, the number of partners along the pipeline can

be assumed to be smaller, since no depots for serving customer markets are required.

Instead, the pipelines are used for balancing supply and demand of raw and intermediate

chemicals among a set of (interdependent) chemical production sites. As a consequence,

the resulting scheduling problems are expected to be less complex than in petrochemical

industry. Available operational scheduling models, which have been originally designed

for petrochemical industry, can be adapted here. Since pipelines in chemical industry are

used to interconnect chemical production sites, the quantities to be distributed via the

pipeline system can be assumed to be rather constant. Hence, not an operational reactive

pumping schedule but a long-term pumping schedule (as provided by the sELSP-BP) is

more likely to be useful for this speciﬁc industry.

71

See Cafaro and Cerdá (2010).

87

While pipelines are used for (almost) continuous transports of chemicals between and

inside chemical production sites, rail transports are (typically) used for smaller and more

sporadic shipments between chemical production sites and to customers with rail road

access. Rail transports are of special interest in basic chemical industry because large

quantities of material can be transported at once on a dense infrastructure compared

to waterway or pipeline networks.72 Particularly locations without access to navigable

waterways often depend on this mode of transport. Moreover, travelling times for rail

transports are often easier to calculate and uncertainty is smaller in contrast to waterway

transports. Hence, rail transports are typically a reliable and well predictable mode of

transport.

However, transport capacities of ships and barges are generally larger than train capac-

ities.73 Hence, transport cost rates per volume and distance unit are typically higher for

rail transports. Since most large-scaled chemical production sites have waterway access,

raw materials are typically supplied via ship transports, but intermediate and minor basic

chemicals are often transported via rail e.g. because the distribution structure is disperse

and/or not all customers/suppliers have waterway access.

Compared to pipeline and ship, rail transports are more ﬂexible,74 but require more or-

ganizational and technical eﬀorts. Hence, the next subsection sheds light on technical and

organizational details considering rail transports of chemical products. Subsequently, an

operational planning model is proposed aiming to support short-term transport decisions

in a chemical production network connected by rail links.

Since rail transports depend on the availability of rail cars, transport capacities have to

be managed in accordance to the chemical quantities planned to be shipped. This (often)

forces chemical companies relying on rail transports to manage ﬂeets of rail cars and

organize transports by themselves. Most often these rail car ﬂeets are rented or bought.75

But although the rail cars are managed by chemical companies, the transport itself is

performed by rail transport service providers (or rail operators). Hence, the organization

of rail transports requires the coordination of product demands, transport capacities, and

72

At least in most developed countries (such as most countries in Europe and the USA) a denser railway

network can be assumed.

73

For transcontinental transports ships and rail transports are not competing. However, in Europe

open sea and rail transports may compete. E.g. transport relations to and from middle and western

European countries passing the Mediterranean or Black Sea can also be performed by trans-balkan

rail transports.

74

In the sense that the railway network is denser and, hence, more locations can be reached by rail

transports.

75

For instance, chemical companies operating in the US often own huge rail car ﬂeets for domestic rail

transports, see Closs et al. (2003).

88

considered for planning rail transport operations.

Special prerequisites are needed to exchange chemicals by rail. Necessarily, a link to

the common rail road network must exist. Such an access link often exists or can be built

easily. To be able to properly organize rail transports, shunting activities have to take

place. This includes the movement of rail cars for de-/composing and rearranging trains

as well as movements of rail cars to and from loading and unloading stations. The ability

to move and arrange rail cars depends on the available number and layout of shunting rail

tracks as well as the available manpower and shunting equipment (e.g. locomotives). The

interplay of these factors deﬁnes the shunting capacity, which can be measured e.g. by the

maximum number of trains to be de-/composed in a certain amount of time. However,

shunting activities are hard to plan and depend on a variety of organizational details,

such as the exact position of rail cars and their intended positions after shunting. Hence,

shunting yards are often bottlenecks when organizing rail transports in chemical industry.

The planning of shunt yard operations is a challenging task.76

The vehicles required to transport products by rail are called rail cars. A vast number

of rail car types exists, often specialized for certain products.77 Since most chemicals are

ﬂuids or gases, the speciﬁc type of rail cars considered in chemical industry is the rail tank

car (RTC). Various sub-types of RTCs can be categorized, where typically the volume to

be transported varies between 20 and 120 m3 . Hence, the payload (in tons) carried by a

certain RTC depends on its volume and the density of the transported chemical.

Gases are transported liqueﬁed in so-called gas RTCs either under pressure or super-

cooled. Fluids are transported in ﬂuid RTCs that can be emptied on top or bottom or

both.78 Transported liquids are typically not under pressure. Depending on the volume,

the payload for both types of RTCs varies between approximately 12 and 70 tons.

The unloading procedure is similar for both RTC types. RTCs are either pumped out

or pressed out by pumping in an inert gas (such as nitrogen). In practice, gas RTCs

are typically pumped out. However, the unloading procedure leads to an equalization of

pressure. Hence, a rest of the transported chemical remains in the RTC if it is not pressed

out.79 Since the removal of the rest is time-consuming, shippers of chemicals often decide

to keep the rest inside the RTC in favour of faster turnover times. However, this implies

that if this RTC should be used to transport another chemical, the rest left inside causes

a contamination. Hence, in practice an RTC is often dedicated to a speciﬁc chemical.

In general, RTCs can interchange between chemical-speciﬁc ﬂeets if they are rinsed

e.g. by nitrogen. Another opportunity is to simply accept the contamination incurred by

76

See e.g. Boysen et al. (2012) for an overview on shunting yard optimization models.

77

See Schenker (2007).

78

To get an impression about RTC types, see the list of RTC types oﬀered by lessors such as GATX

(GATX (2009)) or VTG (VTG (2011)).

79

The unloaded rest may account for approximately 10% of the total payload depending on the speciﬁc

unloading technique, see Compressed Gas Association (1999, pp. 104-115) for details.

89

transferring RTCs from one ﬂeet to another.80 In this case, no rinsing costs occur for

rearranging the RTC ﬂeets but re-processing costs for the contaminated chemical have to

be faced. It has to be noted that this option is only possible among ﬂeets of the same

RTC type (i.e. gas or ﬂuid RTC).

For transferring the relevant chemicals to and from RTCs, special technical devices

are required. For loading and unloading activities, the RTCs have to be connected to

the local pipeline system at so-called transfer arms. This is done by shunting the RTCs

to the transfer stations. Typically, an RTC has one access valve, i.e. either unloading or

loading can take place at a time. The transfer time depends on the technical speciﬁcations

of the local pipeline system and the RTC. Once the transfer procedure is ﬁnished, the

RTC is decoupled from the local pipeline system, shunted, and transported to its next

destination. The total process time for unloading/loading of RTCs consists of

1. preparation time for shunting and connecting the RTC to the transfer station,

Steps 1) and 3) depend on the availability of manpower and the layout of the transfer

stations. Since transfer stations are often serially located along dead-end rail tracks, RTCs

can interfere with one another. Hence, the planning of turnover activities may have severe

inﬂuence on the total turnover time and on the RTCs’ total cycle time.81 After shunting

and transfer, RTCs have to be transported from some point of origin to some point of

destination.

Often, chemical companies act as shippers of chemicals and manage the RTCs while

the transport is organized by a rail operator.82 Basically, the shipper has two transport

options:

In the ﬁrst case, the rail operator has to oﬀer the locomotive(s) for hauling and has to

determine route as well as schedule of the block train on a rail network. In the second case,

the RTCs handed over have to be consolidated at central shunting yards with other rail

cars in order to form block trains. Afterwards, block trains have to be routed, scheduled,

and re-composed. Obviously, the latter option is far more complex to plan and organize.

Hence, higher shipping fees can be expected on an RTC basis in contrast to the block train

80

This is similar to interface "downgrading" in pipeline batch planning, see subsection 3.2.3.

81

On the impact of cycle times on supply chain performance see Closs et al. (2003).

82

This diﬀers from the typical organization in road freight transport where the transport operator man-

ages the transport and owns/manages the transporters.

90

option. To save transport costs, chemical companies often prefer to organize complete

block train transports. This option is the more useful the higher and more frequent

transport volumes are.

To sum up, a lot of technical and organizational restrictions have to be considered

for planning rail transports in the chemical industry. Beside technical constraints, the

planning of transport processes is integrated in the production and replenishment planning

procedures since the consolidation of transport volumes is often organized already at the

chemical production sites. Moreover, large chemical companies often manage a wide-

spread network of production sites. Hence, RTCs can be used for a variety of diﬀerent

transport tasks. To support decisions on when and how many chemicals to be shipped

from which origin to which destination, the following subsection proposes a short-term

rail transportation model.

As stated above, rail transports in chemical industry are often used for distribution of

intermediate and basic chemicals with moderate demand. Often the transported chemicals

serve as raw materials for other chemical plants which either belong to a customer’s or the

shipper’s production network. Since rail transports require cargo consolidation to operate

economically eﬃciently, at large-scaled chemical production sites bundles of chemicals are

handled. I.e. rail transports are a viable option even if the individual demands of each

chemical cannot be shipped eﬃciently by rail.

multiple interdependent plants are located. At each site multiple chemicals are produced

and/or consumed. For each chemical and at each site, tanks are available for intermediate

storage. Local deﬁcits or surpluses of chemicals have to be balanced by imports or exports

of chemicals. It is assumed that all sites are capable to use rail transports for balancing,

i.e. all sites have a rail road access as well as shunting and turnover facilities for RTCs. For

each chemical, a ﬂeet of RTCs is available for transportation. Transports are performed

by rail operators oﬀering train capacities in regular intervals. Therefore, complete trains

are composed at the sites which are handed over to the rail operator.

On the operational level it is assumed that parameters are known and deterministic. It

is to decide about the short-term distribution of the considered chemicals in the network

such that the total operational costs for transport, turnover, and storage are minimized.

More precisely, it is to decide about

91

design model.83 The following basic notation is used: Suppose a set of chemical production

sites I is given for a set of (basic) time periods T . Among all chemicals consumed and

produced at all sites, a subset of these chemicals S exists such that each chemical s ∈ S

is surplus at least at one site and in one period and is in deﬁcit at least at one other

site and period. The local total balance of chemical s in period t at site i is denoted

by ωits . If there is a total surplus, it holds ωits < 0. In case of a (net) deﬁcit, it follows

ωits > 0. For each chemical s and site i ∈ I, tanks are available for storage with maximum

total storage capacity sCapis given in tons. To assure against plant breakdowns due to

material shortages, a local target stock level sTaris is deﬁned. Note that if customers are

supplied, these can be modelled as production sites with zero target stock levels where

order quantities are modelled as sporadic deﬁcits in periods when orders are due. Non-

zero inventory capacities can be set to reﬂect an intermediate storage option and allow

premature deliveries. Otherwise, the transports are forced to arrive exactly in time.

To distribute chemicals among the sites, RTCs are required. Therefore, each chemical

s ∈ S has a ﬂeet of RTCs distributed among the sites. Each ﬂeet is assumed to be

composed by a homogeneous RTC type with maximum payload capacity rsCap , length rsLe ,

and tare weight rsWe . At each site, these RTCs can be consolidated to trains, can be

unloaded, and/or can be loaded.

For transports between sites, trains can be composed and dispatched.84 The set of rail

links available for transport is denoted by L ⊆ I × I. Note that each site has access to at

least one rail link. Each site can dispatch a maximum number of trains yitCap in period t

which acts as a proxy for a site’s shunting capacity. The maximum number of trains to

be chartered for a rail link (i, j) ∈ L, ȳijt

Cap

, reﬂects varying contingents negotiated between

the shipper and the rail operator(s). 85

Le We

Similarly, individual maximum train lengths trij and train weights trij are assumed

for each rail link (i, j) ∈ L reﬂecting technical train speciﬁcations such as the power and

number of locomotive(s) committed by the rail operator as well as legal restrictions e.g. for

maximum train length. Typically, either the length or the weight constraint is restrictive.

The former restriction is often restrictive if only empty RTCs are hauled whereby the

latter is restrictive if only loaded RTCs are hauled.

83

See Newman and Yano (2000) for a similar intermodal transportation planning problem.

84

Loaded and empty RTCs are shunted and composed to a train load. The train load is handed over to

the rail operator who oﬀers the traction (i.e. the locomotive) and performs the transport.

85

This allows incorporating periods without available train dispatches, e.g. on weekends or holidays.

92

The discretized travelling time of a train operating on rail link (i, j) ∈ L is denoted by

tTrv

ij which is an integer value and reﬂects the number of basic periods elapsing during

shipment. Obviously, this is a notable simpliﬁcation of real-world conditions and a draw-

back of all time-discrete models. In this case, however, the simpliﬁcation does not severely

interfere the modelling since travelling times in trans-regional rail freight transports are

typically relatively large.86 Transport times include the pure travelling times as well as

waiting and inspection times, e.g. when borders are crossed or when cargo trains have to

wait for passenger trains to pass. Hence, even for comparatively short distances to bridge,

travelling time is counted in hours. As a rule of thumb, the basic period can be set to

the maximum over the greatest common divisor of all time parameters including train

travel times as well as e.g. turnover times. The total planning horizon is restricted since

demand and production rates have to be updated frequently. Planning horizons of 7 to

14 days seem appropriate for most problem instances in practice.

An additional advantage of a time-discrete model is that e.g. work shifts for the turnover

staﬀ are discrete in time. Under regular conditions turnover and transfer activities do not

take place around the clock but e.g. in two shifts with 8 hours each (or one shift with 12

hours). Hence, a time-discrete formulation allows to easily incorporate such regulations.

Therefore, the unloading and loading capacities, eCap Cap

its and lits , are expressed in tons of

chemical s to be handled at site i in period t. Both parameters reﬂect the work force

availability and technical restrictions such as the number and capacity of transfer arms. It

is assumed that exclusive turnover capacities are provided for each chemical at a site. This

implies single-product transfer stations for each product. This is reasonable if handled

chemicals are dissimilar and a joint unloading/loading is too dangerous or expensive.87

However, in case of similar chemicals, commonly used transfer stations may exist at the

sites.88 In this case, a slightly adjusted turnover capacity formulation is required. In basic

chemical industry, however, individual transfer stations are predominant.

To properly model the RTC handling processes at the sites, the time necessary to un-

load/load the RTCs has to be considered. The total processing time for RTC handling

is summarized in parameter tTrn

is which is the number of basic periods required to make

an RTC available for transport (after loading or unloading). I.e. if RTCs are unload-

ed/loaded, tTrn

is is the minimum time span an RTC has to spend at site i. In most cases,

a turnover time span of one work shift, i.e. one basic period, is assumed before an RTC

86

A reasonable choice for the basic period is e.g. 24, 12 or 8 hours. In practice, more granular planning

accuracy is seldom required since travelling times in trans-regional rail transports are not much shorter

than 12 hours. See Newman and Yano (2000) for a more detailed discussion.

87

Dissimilar chemicals are likely to react with each other which is typically unintended as it bears the

risk of debris, hazardous reactions or even explosions.

88

Such stations are often used in crude oil industry when diﬀerent types of fuel or other oil derivatives

are handled.

93

is available for transport. Note that RTCs which are not unloaded/loaded are available

for transport immediately.89

If the number of RTCs to be unloaded/loaded in a period at a site exceeds the respective

capacities eCap Cap

its and lits , the remaining number of RTCs have to wait at the site’s shunting

L

yard. Loaded RTCs, denoted by srits , may wait on rail tracks ready for forwarding.

This equals an additional inventory denoted by sLits which induces stock holding costs

for committed capital and costs for supervising the RTCs. Since most chemicals are

hazardous, storing requires permanent supervision and control. In regular tanks this is

typically assured by sensor systems. Although RTCs are designed for recurring transports

of hazardous chemicals, they typically do not meet e.g. legal requirements for a permanent

technical supervision system.

To evaluate the operational eﬀort for realizing transports and keeping stocks, three types

of costs have to be considered:

• transport costs

• turnover costs.

Stock holding costs are calculated as follows: At each site safety stocks of each handled

chemical are hold. They protect against extraordinary demand peaks e.g. caused by plant

breakdowns. With respect to the inventory holding costs, the stock level sits should meet

the desired safety stock level sTar

is . An overshooting of the safety stock levels causes un-

necessary stock holding costs (e.g. due to capital commitment) whereas an undershooting

increases expected shortfall costs since in emergency cases plant shutdowns menace due

to a shortfall of supply. Because the costs induced by a missing or surplus ton of a certain

chemical depend on the target stock level sTar is , over- and undershooting (oits and uits )

are measured as relative deviations from sTar Os

is . The corresponding cost rates cis and cis

Us

are measured in monetary units per relative deviation of the stock level of chemical s

from its target stock level at site i. Often it holds cUsis ≫ cis because a plant shutdown

Os

90

causes much higher costs compared to stock holding. Additionally, for chemicals stored

intermediately in (loaded) RTCs, the cost rate cAdd reﬂects the additional supervision and

maintenance eﬀort measured in monetary units per ton and period.

Transport costs reﬂect the monetary eﬀort to be faced when trains are dispatched, i.e. a

rail operator performs a transport service. Therefore, cost rate cTrv

ij accounts for the charge

89

E.g. if empty RTCs arrive in a period at a site and are not to be loaded, they can be forwarded to

another site in the same period.

90

In general, expected shortfall costs increase non-linearly with increasing undershooting of target stock.

However, for the sake of solvability this non-linearity is dropped in this model.

94

paid when a train is dispatched on rail link (i, j) ∈ L. These charges are typically based

on general agreements with the responsible rail operator and depend on the distances to

be travelled, borders crossed, train speciﬁcations and other details.91

The turnover cost rate cTrn

is comprises direct costs for unloading and loading RTCs. This

cost rate reﬂects labour costs for shunting, unloading/loading, and supervision activities.

To sum up, a rail transportation model for multiple products shipped by trains between

chemical production sites can be formulated as a multi-layer, multi-commodity time-

space expanded network ﬂow model. In other words, an operational multi-chemical rail

transportation problem is provided (in short MC-RTP).

In the network model each chemical corresponds to a main layer representing the infras-

tructure of this chemical. The main layer consists of four sub-layers which are dedicated

to

Flows between two nodes of the same sub-layer correspond to storing or shipping ﬂows

of RTCs or chemicals, respectively. Flows between nodes of diﬀerent sub-layers indicate

loading or unloading ﬂows. At each node inﬂows and outﬂows are in balance. Each ﬂow is

restricted by stock, RTC or train capacities. Transport ﬂows of chemicals are restricted by

the accompanying ﬂow of loaded RTCs, i.e. ﬂows in the second and third sub-layer match

each other. Hence, transport capacities are dynamically modelled as in a classic IRP. In

diﬀerence to classic IRPs, no routes for individual carriers are determined. Instead, the

RTC ﬂow balances ensure that the correct transport capacity is modelled. The available

transport capacity, oﬀered by RTCs, can only be exploited by chartering trains. This

component assigns arc-based ﬂow capacities incurring ﬁxed costs which is a characteristic

of ﬁxed-charge network design models. Hence, the proposed model can be seen as a hybrid

model bearing characteristics of both prominent model classes. Figure 3.9 illustrates an

exemplary ﬂow network for ﬁve periods and three sites.

Grey arrows indicate (passive) ﬂows of chemicals and (empty) RTCs. Marked by black

arrows is the following sequence of activities: In period 1 empty RTCs are shipped from

site 2 to site 1 where they are needed to load a certain quantity of the chemical in period

2. I.e. a ﬂow between forth and third sub-layer indicates a change in the RTCs’ states.

91

Note that additional ﬂexibility can be incorporated by time varying charges e.g. for weekend transports.

95

layer

r)

b- s

su TC

ye

1,2 2,2 3,2 4,2 5,2

la

R

ship

(4 pty

1,1 ping 2,1 3,1 4,1 5,1

em

th

loading

unloading

1,3 2,3 3,3 4,3 5,3

b- Cs

r)

ye

1,2 2,2 3,2 4,2 5,2

su RT

la

(3 ded

g

1,1 2,1 3,1 shippin 4,1 5,1

a

rd

lo

1,3 2,3 3,3 4,3 5,3

ye s

la C

b- T

r)

su in R

1,2 2,2 3,2 4,2 5,2

nd ls

g

shippin 4,1

(2 ica

1,1 2,1 3,1 5,1

em

ch

un

lo

ad

in

1,3 2,3 3,3 4,3 5,3

g

g

ye s

din

la nk

r)

b- ta

loa

su in

e

sit

st s

storing

(1 ical

1,1 2,1 3,1 4,1 5,1

em

time

Correspondingly, a quantity of the chemical is transferred from the tanks (ﬁrst sub-layer)

to the mobile stock (second sub-layer). In period 3 the loaded RTCs depart by train from

site 1 carrying the chemical to site 2 where they arrive in period 4 (second and third

sub-layer). In this period the RTCs are unloaded such that the local chemical stocks are

replenished and, simultaneously, the RTCs are emptied. In this sequence two shipments

are performed, i.e. (at least) two trains are chartered.

To formalize the sketched network ﬂow model, the remaining notation is introduced

as follows: A shipment ﬂow of chemical s from site i to site j in period t is denoted by

xijts (second sub-layer). Shipment ﬂows of loaded and empty RTCs are denoted by rxLijts

and rxEijts , respectively (third and forth sub-layer). Unloading ﬂows and loading ﬂows of

E L

chemical s in period t at site i are described by zits and zits (ﬂows between ﬁrst and second

sub-layer). These ﬂows are accompanied by the corresponding RTC ﬂows (between third

E L

and forth sub-layer) rzits and rzits , respectively. To enable a shipment of loaded and/or

empty RTCs from node i to node j a number of trains (yijt ) has to be dispatched.

The local chemical stock in tanks is denoted by sits (ﬁrst sub-layer) whereas the mo-

bile stock in RTCs is denoted by sLits (second sub-layer). The number of loaded and

L E

empty RTCs at hand are described by srits and srits . Table 3.13 summarizes all notation

introduced above (and some more).

96

Sets

I = {1, ..., I} set of sites

T = {1, ..., T } set of periods

L⊆I ×I set of rail links

Parameters

Cost rates

cOs

is over-shooting cost rate for chemical s ∈ S

cUs

is under-shooting cost rate for chemical s ∈ S

cTrn

s turnover costs for chemical s ∈ S

cTrv

ij costs for chartering a train on rail link (i, j) ∈ L

(artiﬁcial) cost rate for carrying an empty RTC of chemical s ∈ S

c̃Tr−E

ijs

on rail link (i, j) ∈ L

(artiﬁcial) cost rate for carrying a loaded RTC of chemical s ∈ S

c̃Tr−L

ijs

on rail link (i, j) ∈ L

cAdd cost rate for storing chemical in RTCs

Utilization parameters

ωits

ωits > 0 there is a net deﬁcit, otherwise a surplus

rsWe tare weight of RTCs for chemical s ∈ S

rsLe length of RTCs for chemical s ∈ S

sTar

is target stock level for chemical s ∈ S at site i ∈ I

Capacities

We

trij maximum train weight for trains on rail link (i, j) ∈ L

Le

trij maximum train length for trains on rail link (i, j) ∈ L

Cap maximum number of trains to be dispatched on rail link (i, j) ∈ L

ȳijt

in period t ∈ T

maximum number of trains to be dispatched at site i ∈ I in period

yitCap

t∈T

maximum number of empty RTCs of chemical s ∈ S to be carried

nCap−E

ijs

by a train on rail link (i, j) ∈ L

maximum number of loaded RTCs of chemical s ∈ S to be carried

nCap−L

ijs

by a train on rail link (i, j) ∈ L

eCap

its unloading capacity of chemical s ∈ S at site i ∈ I in period t ∈ T

Cap

lits loading capacity of chemical s ∈ S at site i ∈ I in period t ∈ T

97

sCap

is inventory capacity for chemical s ∈ S at site i ∈ I

Initialization parameters

sIni

is initial stock level for chemical s ∈ S at site i ∈ I

sL−Ini

is initial stock level stored in RTCs for chemical s ∈ S at site i ∈ I

L−Ini

ris initial loaded RTCs for chemical s ∈ S available at site i ∈ I

E−Ini

ris initial empty RTCs for chemical s ∈ S available at site i ∈ I

quantity of chemical s ∈ S dispatched in period τij ∈ {1 − tTrv

ij , ..., 0}

xIni

ijτij s

on rail link (i, j) ∈ L

number of loaded RTCs dispatched in period τij ∈ {1 − tTrvij , ..., 0}

rxL−Ini

ijτij s

on rail link (i, j) ∈ L

number of empty RTCs dispatched in period τij ∈ {1 − tTrvij , ..., 0}

rxE−Ini

ijτij s

on rail link (i, j) ∈ L

L−Ini

rziτis s

number of RTCs loaded in period τis ∈ {1 − tTrn

is , ..., 0} at site i ∈ I

E−Ini

number of RTCs unloaded in period τis ∈ {1 − tTrn

is , ..., 0} at site

rziτis s

i∈I

Time parameters

tTrv

ij integer, travelling time for trains on rail link (i, j) ∈ L

tTrn

is integer, turnover time for chemical s ∈ S at site i ∈ I

Decision variables

integer, ﬂow of RTCs loaded with chemical s ∈ S on rail link

rxLijts

(i, j) ∈ L in period t ∈ T

integer, ﬂow of empty RTCs for chemical s ∈ S on rail link

rxEijts

(i, j) ∈ L in period t ∈ T

E

zits unloading ﬂow of chemical s ∈ S in period t ∈ T at site i ∈ I

L

zits loading ﬂow of chemical s ∈ S in period t ∈ T at site i ∈ I

E

integer, RTCs of chemical s ∈ S unloaded in period t ∈ T at site

rzits

i∈I

L

rzits integer, RTCs of chemical s ∈ S loaded in period t ∈ T at site i ∈ I

integer, number of trains dispatched on rail link (i, j) ∈ L in

yijt

period t ∈ T

Variables

sLits stock of chemical s ∈ S stored in RTCs in period t ∈ T at site i ∈ I

L

number of loaded RTCs for chemical s ∈ S stored in RTCs in

srits

period t ∈ T at site i ∈ I

98

E

number of empty RTCs for chemical s ∈ S in period t ∈ T at site

srits

i∈I

relative over-shooting in % of the target stock of chemical s ∈ S in

oits

period t ∈ T at site i ∈ I

relative under-shooting in % of target stock of chemical s ∈ S in

uits

period t ∈ T at site i ∈ I

Table 3.13: Sets, parameters, variables, and decision variables for the MC-RTP

Using the aforementioned notation, the total operational costs for balancing the chem-

ical production network by rail transports is expressed in (3.33) which constitutes the

MC-RTP’s objective function to be minimized.

T C = ∑ ∑ ∑ (oits ⋅ cOs

is + uits ⋅ cis ) +

Us

i∈I t∈T s∈S

∑ ∑ yijt ⋅ cTrv

ij +

(i,j)∈L t∈T

∑ ∑ ∑ (rzits

E

+ rzits

L

) ⋅ cTrn

s +

i∈I t∈T s∈S

i∈I t∈T s∈S

The ﬁrst part of (3.33) accounts for the stock holding costs arising from deviations of

stock levels from the desired target stock levels. The second part sums up the total train

charges over all rail links by multiplying the number of dispatched trains yijt with the

corresponding train cost rate cTrv

ij . Turnover costs for turnover activities taking place at

the sites are summed up in the third part of (3.33). The last part accounts for costs for

supervision and control of loaded RTCs waiting at the sites which can be interpreted as

costs for dynamic storage capacity extensions.

The considered restrictions can be categorized in balancing and capacity constraints.

Constraints (3.34)-(3.37) represent the balancing constraints.

E

− zits

L

− ωits ∀i ∈ I, t ∈ T , s ∈ S (3.34)

E

+ zits

L

− ∑ xijts + ∑ xji(t−tTrv

ji )s

∀i ∈ I, t ∈ T , s ∈ S (3.35)

j∈I j∈I

L

srits = sri(t−1)s

L

− rzits

E

+ rzi(t−t

L

Trn )s

is

ji

j∈I j∈I

E

srits = sri(t−1)s

E

− rzits

L

+ rzi(t−t

E

Trn )s

is

99

ji

j∈I j∈I

Here, (3.34) and (3.35) track the quantities of chemical s available at site i in period t

E L

in tanks (3.34) and in loaded RTCs (3.35). Unloading and loading ﬂows (zits and zits )

connect both inventories inversely, e.g. unloading RTCs (i.e. zits E

> 0) reduces sLits and

increases sits . Furthermore, tank inventories buﬀer the net consumption or surplus ωits .

In contrast, the stock loaded in RTCs sLits absorbs incoming chemical ﬂows from other

sites (xji(t−tTrv

ji )s

) and provides outgoing chemical ﬂows to other sites (xijts ).

Equations (3.36) and (3.37) represent stock balances for loaded and unloaded RTCs.

E L

Both stocks are interconnected by RTC transfer ﬂows rzits and rzits . E.g. unloading of

E

rzits RTCs in period t at site i immediately reduces the number of loaded RTCs. However,

tTrn

is periods are required for unloading and shunting before these RTCs can be handled

as empty RTCs again. Both stocks are replenished by incoming RTCs from other sites

(rxLji(t−tTrv )s and rxEji(t−tTrv )s ) and provide outgoing RTC ﬂows to other sites (rxLijts and

ji ji

rxEijts ).

Capacity constraints can be separated into two subgroups: Static capacities which refer

to infrastructural conditions (such as tank, train, or turnover capacities) and dynamic

capacities referring to restrictions that are planned simultaneously (such as RTC stocks

and ﬂows). Constraints (3.38)-(3.42) express static capacity constraints.

E

zits ≤ eCap

its ∀i ∈ I, t ∈ T , s ∈ S (3.38)

L

zits ≤ lits

Cap

∀i ∈ I, t ∈ T , s ∈ S (3.39)

sits ≤ sCap

is ∀i ∈ I, t ∈ T , s ∈ S (3.40)

yijt ≤ ȳijt

Cap

∀(i, j) ∈ L, t ∈ T (3.41)

∑ yijt ≤ yit ∀i ∈ I, t ∈ T

Cap

(3.42)

j∈I

We

⋅ yijt ∀(i, j) ∈ L, t ∈ T (3.43)

s∈S s∈S

Le

⋅ yijt ∀(i, j) ∈ L, t ∈ T (3.44)

s∈S

Constraints (3.38) and (3.39) restrict the unloading and loading ﬂows to the unloading and

loading capacities of the corresponding periods and sites. Similarly, (3.40) ensures that in

all periods stock levels are less or equal to the tank capacities. Constraints (3.41) require

that the maximum number of trains dispatchable on a rail link per period is not exceeded.

Similarly, constraints (3.42) ensure that the sum of all trains composed on site i in period

t is smaller than the corresponding maximum shunting capacity. Constraints (3.43) and

100

(3.44) refer to technical train speciﬁcations assuring that maximum weight and length

are not exceeded. On the right-hand side, the total available capacities are calculated as

the product of maximum train weight and length multiplied with the number of trains

dispatched on a link (i, j) ∈ L. On the left-hand side, the trains’ total hauling weight is

calculated as the sum of all chemicals loaded in RTCs (∑s∈S xijts ) plus tare weights of all

RTCs attached to the train(s). The total length of the trains is the sum of the lengths of

all RTCs attached.

If yijt > 1 for a speciﬁc trip (i, j, t) and only one of both constraints is restrictive, an

integer feasible assignment of RTCs to individual trains is assured by constraints (3.43)

and (3.44). However, if both constraints are (almost) restrictive92 and yijt > 1, an integer

feasible assignment of RTCs to trains only exists if the number of RTCs is a multiple of

the number of trains yijt . To illustrate this situation, consider a speciﬁc trip with y = 2.

For simplicity assume that a homogeneous ﬂeet of RTCs exist with equal tare weight

rWe = 2, equal length rLe = 1, and equal capacity rCap = 8. Let trLe = 20 and trWe = 100.

Le

Then, the maximum number of RTCs due to (3.44) is trrLe ⋅y = 20 1 ⋅2 = 40 with an associated

tare weight of 40 ⋅ rWe = 80. According to (3.43) the total remaining (weight) capacity for

payload is then y⋅trWe −80 = 120 which implies a total number of r120 Cap = 15 of loaded RTCs.

In consequence, 25 empty RTCs and 15 loaded RTCs with a maximum payload of 120

could be carried by two trains according to (3.43) and (3.44). However, the numbers of

loaded and empty RTCs are not multiples of two. The most balanced integer assignment

of loaded and empty RTCs to two trains is displayed in Table 3.14.

# empty loaded length weight

1 13 7 20 40 + 7 ⋅ 8 = 96

2 12 8 20 40 + 8 ⋅ 8 = 104

total 25 15 40 200

From Table 3.14 it can be taken that for train # 1 a feasible RTC assignment is made

while for train # 2 the maximum total weight of 100 is exceeded. In case when multiple

trains can be dispatched on a trip, a reformulation of (3.43) and (3.44) would overcome

this problem. However, to reformulate both constraints additional variables are necessary

indicating whether both constraints are (almost) restrictive for a speciﬁc trip. This would

lead to a more complicated model. Since the problem only occurs if the total payload is

close to its maximum, it suﬃces to introduce a safety buﬀer for the total train weight.

We

I.e. a train’s allowed total weight trij We

is calculated as trij ˜ We

= tr ij − max rs

Cap ˜ We

where tr ij

s∈S

92 Le

To be precise, it has to hold trij ⋅ yijt − ∑s∈S rsLe ⋅ (rxE L We

ijts + rxijts ) = 0 and trij ⋅ yijt − ∑s∈S xijts +

We E L Cap

∑s∈S rs ⋅ (rxijts + rxijts ) ≤ max rs .

s∈S

101

rsCap

is the technical maximum of the train’s total weight. Note that in practice the ratio ˜ We

tr ij

is quite small (approximately 1-5 %) such that no serious loss in train utilization is to be

expected.

Dynamic capacity constraints are formulated in (3.45)-(3.49).

E

zits ≤ rzits

E

⋅ rsCap ∀i ∈ I, t ∈ T , s ∈ S (3.47)

L

zits ≤ rzits

L

⋅ rsCap ∀i ∈ I, t ∈ T , s ∈ S (3.48)

sLits ≤ srits

L

⋅ rsCap ∀i ∈ I, t ∈ T , s ∈ S (3.49)

L

− 1) ⋅ rsCap ∀i ∈ I, t ∈ T , s ∈ S (3.50)

Here, (3.45) and (3.46) ensure that the shipped quantity of a chemical s matches the

total payload capacity of all loaded RTCs to be shipped.93 Constraints (3.47) and (3.48)

restrict the loading and unloading ﬂows to the payload capacities of the simultaneously

loaded and unloaded RTCs. Constraints (3.49) and (3.50) enforce that the quantity of

chemicals designated to be stored in RTCs matches the total payload capacity of loaded

RTCs. Constraints (3.45) and (3.46) in combination with (3.49) and (3.50) also restrict

loading and unloading ﬂows to be matched with the numbers of loaded and unloaded

RTCs in a period.

To assess relative under- and over-shooting of target stock levels, constraints (3.51)

assign the deviations of the ratio ssTar

its

from 1 to the variables oits and uits . If sits > sTar

is ,

is

it follows that sTar − 1 > 0. Because oits and uits are non-negative variables and both are

sits

is

associated to positive cost rates, the cost-minimal values for oits and uits satisfying (3.51)

are obtained by oits = ssTar

its

− 1 and uits = 0. Otherwise, uits = ssTar

its

− 1 and oits = 0. Finally,

is is

(3.52) and (3.53) deﬁne the variables’ domains.

sits

oits − uits = −1 ∀i ∈ I, t ∈ T , s ∈ S (3.51)

sTar

is

L

xijts , zits E

, zits , sits , sLits , oits , uits ∈ R+ ∀i ∈ I, j ∈ I, t ∈ T , s ∈ S (3.52)

E L

, rzits L

, srits E

, srits ∈ N0 ∀i ∈ I, j ∈ I, t ∈ T , s ∈ S (3.53)

Constraints (3.33)-(3.53) constitute the MC-RTP. The optimal solution of the MC-

93

Note that dropping (3.46) would circumvent the distinction in empty and loaded RTCs. This would

allow sending empty RTCs declared as loaded ones to another site where these could be loaded

immediately. Hence, the turnover times would not be considered.

102

RTP minimizes the total costs for transports, turnover activities, and stock holding. The

solution provides ﬂows of RTCs and chemicals within the considered time horizon. At

the end of the planning horizon, no transports are planned because the arrival periods

are beyond the time horizon.94 In practice, therefore, the MC-RTP has to be applied in

a rolling horizon environment with overlapping time horizons to update schedules with

new information e.g. about consumption and production rates.95 I.e. starting with an

initial MC-RTP instance and for ﬁxed a length of the planning horizon, trains and RTCs

are scheduled in accordance to the corresponding (optimal) solution for the ﬁrst peri-

ods (the re-planning interval).96 After the re-planning interval has elapsed, an updated

MC-RTP instance is set up and solved. Therefore, updated initial values for stocks and

available RTCs etc. and additional production/consumption rates for the new periods to

be planned are incorporated. Additionally, trains that have departed but not arrived yet

have to be considered. To use the MC-RTP in a rolling horizon environment, additional

initialization equations have to be added.

At ﬁrst, initial stock levels in tanks sIni L−Ini

is , quantities available in RTCs sis , the corre-

L−Ini E−Ini

sponding numbers of loaded RTCs ris , and the available numbers of empty RTCs ris

for each chemical s and site i at the beginning of the planning horizon are incorporated

(see (3.54)-(3.57)).

si0s = sIni

is ∀i ∈ I, s ∈ S (3.54)

E

sri0s = ris

E−Ini

∀i ∈ I, s ∈ S (3.55)

L

sri0s = ris

L−Ini

∀i ∈ I, s ∈ S (3.56)

sLi0s = sL−Ini

is ∀i ∈ I, s ∈ S (3.57)

If RTCs and chemicals are dispatched in previous periods (and arrive during the planning

horizon), these have to be considered. This aﬀects variables xijτij s , rxLijτij s , and rxEijτij s

with 1 − tTrv ij ≤ τij ≤ 0 which have to be set to the corresponding initial values denoted by

xIni

ijτij s , rx L−Ini E−Ini

ijτij s , and rxijτij s as shown in (3.58)-(3.60).

xijτij s = xIni

ijτij s ∀(i, j) ∈ L, s ∈ S, τij ∈ {1 − tTrv

ij , ..., 0} (3.58)

rxEijτij s = rxE−Ini

ijτij s ∀(i, j) ∈ L, s ∈ S, τij ∈ {1 − tTrv

ij , ..., 0} (3.59)

rxLijτij s = rxL−Ini

ijτij s ∀(i, j) ∈ L, s ∈ S, τij ∈ {1 − tTrv

ij , ..., 0} (3.60)

94

A similar problem occurs in pipeline operations planning when batches planned at the end of the

planning horizon are solely injected to keep the pipeline working. Such batches do not arrive at

depots during the planned time horizon. Hence, they cannot satisfy demands at depots.

95

For a similar rolling horizon approach for pipeline operations planning see Cafaro and Cerdá (2008).

96

Note that the re-planning interval should be considerably smaller than the planning horizon, see

e.g. Cafaro and Cerdá (2008).

103

Similarly, RTCs already shunting at the yards at the beginning of the planning horizon

have to be considered by setting the variables rziτ L

is s

E

and rziτis s

for 1 − tTrn

is ≤ τis ≤ 0 to

L−Ini E−Ini

their corresponding initial values rziτis s

and rz iτis s

E

rziτis s

= rziτ

E−Ini

is s

∀i ∈ I, s ∈ S, τij ∈ {1 − tTrn

is , ..., 0} (3.61)

L

rziτis s

= rziτ

L−Ini

is s

∀i ∈ I, s ∈ S, τij ∈ {1 − tTrn

is , ..., 0} (3.62)

environment. Otherwise, the corresponding variables are set to 0.

The following example illustrates the applicability of the MC-RTP for an artiﬁcial prob-

lem instance.

sisting of three sites where three commonly produced/consumed chemicals are planned for

14 periods. A period equals 12 hours. At the beginning of each day, a single train can

be dispatched from every site. Each site is connected with any other site via a rail link,

i.e. L = {(i, j) ∶ i, j ∈ I ∧ i ≠ j}. Table 3.15a shows the travelling times tTrv

ij for all rail

links (i, j) ∈ L.

1 2 3 1 2 3 1 2 3

1 - 2 4 1 20 20 25 rsCap 15 15 20

site i

site i

2 2 - 1 2 20 20 20 rsWe 5 5 10

3 4 1 - 3 25 20 25 rsLe 20 20 25

(a) tTrv

ij

E−Ini

(b) ris (c) rsCap , rsWe , rsLe

Table 3.15: Technical parameters for the rail operations planning example

For each chemical a ﬂeet of RTCs is available. No previous transports are assumed

to be on track such that the initialization variables in constraints (3.58)-(3.62) are set to

zero. Similarly, all available RTCs at the sites at the beginning of the planning horizon

are empty and ready for loading. Hence, initialization variables in constraints (3.56) and

(3.57) are zero, too. Tables 3.15b and 3.15c show the distribution of the RTC ﬂeets among

the sites and the technical conﬁgurations for the RTCs in the chemicals’ ﬂeets, respectively.

All turnover times tTrn

is are set to one period. Furthermore, train speciﬁcations are given

We

by trij = 800 and trij

Le

= 1, 000 for all (i, j) ∈ L. The unloading and loading capacities eCap its

Cap

and lits are set to 700 tons such that a complete train’s load can be (un-)loaded within a

period.

For simplicity it is assumed that the chemicals’ consumption/production rates ωits are

constant over the planning horizon. Table 3.16 displays the assumed consumption/pro-

duction rates ωis as well as stock capacities sCap Ini

is , initial stocks sis , and target stock levels

104

sTar

is for each chemical s and site i.

site i 1 2 3

chemical s 1 2 3 1 2 3 1 2 3

ωis -50 25 100 -50 25 -200 100 -50 100

sCap

is 3000 2500 2500 1500 1500 5000 5000 5000 5000

sIni

is 400 500 700 500 300 1000 700 500 1000

sTar

is 500 250 500 500 250 1000 750 500 1000

is , initial stocks sis , and target

stock levels sTar

is for all sites i and chemicals s

Finally, cost rates are deﬁned as follows: cOs Us Trv Trn =

cAdd = 10 for all i, j ∈ I and s ∈ S. Note that overshooting the target stock levels induces

only a thousandth of the cost for undershooting. The train cost rate cTrvij is comparatively

small forcing extensive rail activities.

Solving this instance of the MC-RTP leads to the optimal solution after 8 seconds with

an objective value of 20,845.97 Note that network design problems are N P − hard prob-

lems.98 However, the computational complexity depends on the relation of the parts of the

objective function and the size of the network (particularly the number of links). Here,

a comparatively small problem instance with a favourable ratio of costs99 is given which

eases solving to optimality.

The optimal transport ﬂows and stock levels are displayed in Figure 3.10 for the ﬁrst

eight periods by means of a time-space expanded network. Each cell shows the stock levels

at a site in a particular period. Dispatched trains are indicated by arrows between two

nodes where the associated transport quantities (in tons of chemicals) and/or the number

of empty RTCs are assigned to each arrow.100

In t = 1, from sites i = 1 and i = 3 repositioning transports are dispatched enabling

chemical transports in later periods. For example, empty RTCs of chemicals s = 1 and

s = 3 are dispatched in period t = 1 from site i = 3 to site i = 2 arriving there in t = 2.

In period t = 3, empty RTCs of chemical s = 1 are partially forwarded101 (4 out of 20)

to site i = 1 together with some loaded RTCs of chemical s = 3.102 Similar sequences of

re-positioning and supply transports can be found for the remaining chemicals and sites.

These sequences are repeated if the production/consumption rates do not change in time

which results in recurring re-positioning cycles.

97

Calculations are performed on a 2.6 GHz machine with IBM CPLEX 12.5, see IBM (2010).

98

See e.g. Crainic (2000).

99

To be precise, the inventory costs dominate transport costs such that network ﬂow primarily determines

the total costs.

100

For clarity, the indices of variables annotated at the arcs are reduced and indicate the number of the

chemical only.

101

The rest is used for transports of chemical s = 1 back to site i = 3 in periods t = 5 and t = 7.

102

To be precise, 26 loaded RTCs of chemical s = 3 are dispatched in period t = 3 carrying 520 tons from

site i = 2 to site i = 1. The train’s maximum weight is totally exploited.

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

450 500 430 480 410 460 510 560

600 500 400 300 720 620 520 420

i =1

475 450 425 400 375 350 325 300

E E E

rx s=3 = 13 rx s=3 = 12 rx s=3 = 26

x s=1 = 120 x s=1 = 120

E

x s=3 = 520 rx s=2 = 12

E

rx s=1 =4

550 600 650 700 510 560 445 495

1200 1400 1080 1280 1160 1360 1280 1480

i =2

275 250 225 320 295 270 245 310

x s=3 = 320 x s=3 = 280

x s=1 = 240 x s=2 = 90 x s=1 = 285

E

rx s=3 = 20 x s=2 = 120 x s=2 = 90

E E E

rx s=1 = 25 rx s=3 =5 rx s=1,3 = 16

600 500 400 300 200 340 360 545

900 800 700 600 500 720 620 800

i =3

550 600 530 580 540 590 550 600

target stock chemical 1 chemical 2 chemical 3

1000

Figure 3.10: Optimal stock levels and network ﬂows for periods 1 to 8 of the MC-RTP example

105

106

For settings with realistically sized ﬁnite planning horizons, only the beginnings of such

cycles are visible. At the end of the planning horizon no transports are reasonable as

the savings in stock holdings become eﬀective in periods beyond the planning horizon

only. Therefore, stock levels are the less balanced the closer the planning horizon ap-

proaches. This end-of-horizon eﬀect becomes strikingly apparent for chemical s = 1 (dark-

grey curves) at sites i = 1 and i = 3 as it is illustrated by Figure 3.11.103 The ﬁgure shows

the stock levels for all chemicals, periods, and sites including the corresponding target

stocks (as dashed lines).

It can be observed that the stock levels meet the corresponding target stocks well in most

cases. For chemical s = 2 (medium grey curves) the stock levels are close above the target

stocks at all sites and in (almost) all periods.104 Chemicals s = 3 (light grey curves) and

s = 1 are oscillating around their target stock levels at sites i = 1 and i = 2 whereas at

site i = 3 both chemicals are clearly below their target stocks. Transports of chemical s = 3

from site i = 2 to i = 3 could reduce holding costs at both sites.105 However, this option

is not realized either due to a lack of RTCs or due to economic reasons (e.g. the related

transport costs).

This example provides some ﬁrst insights into the interplays of re-positioning and stock-

piling transports crucially depending on both the availability of RTCs and the heights

of target stock levels. Moreover, it becomes apparent that in practical applications the

MC-RTP has to be applied in a rolling horizon environment to overcome end-of-horizon

problems like stock decreases and undone repositioning transports.106

In reality, the production networks typically consist of more than three sites107 whereby

a time horizon of one or two weeks is reasonable in most cases. The time horizon depends

on the forecasting stability of production/consumption estimates as well as the transport

times. Hence, the model’s complexity increases with increasing numbers of nodes and

periods. In cases of very large instances, a heuristical procedure can be set up like this:

1. Build a relaxed MC-RTP by replacing the train cost part ∑(i,j)∈L ∑t∈T yijt ⋅ cTr

ij in the

objective function (3.33) with

∑ ∑ ∑ (rxLijts ⋅ c̃Tr−L

ijs + rxijts ⋅ c̃ijs )

E Tr−E

103

But also the eﬀect is obvious for chemical s = 2 (medium grey curves) at sites i = 1 and i = 3.

104

Except for periods t = 3 and t = 7 at site i = 2 where a slight undershooting is observed.

105

Since this chemical is in deﬁcite at site i = 3 but surplus at site i = 2.

106

For a more detailed discussion of end-of-horizon problems see Hughes and Powell (1988).

107

For example the production network of Dow Chemical consists of more than 20 sites.

chemical s = 1

location i = 1 stock level

chemical s = 2 target stock

chemical s = 3

750 1000

500

stock level

250

0

location i = 2

1250

750

stock level

250

0

location i = 3

750

stock level

250

0

0 2 4 6 8 10 12 14

period

Figure 3.11: Optimal stock levels for the MC-RTP instance of example 6

107

108

A considerable part of the model’s complexity is caused by the integer decisions to deter-

mine the number of dispatched trains. Since these decision variables are associated with

ﬁxed charges per train, a variabilization of these ﬁxed costs simpliﬁes the problem consider-

ably. Variable transport costs depend linearly on the number of RTCs dispatched on a rail

link. To deﬁne variable transport cost rates c̃Tr−L

ijs and c̃Tr−E

ijs , calculate the maximal num-

trWe trLe

ber of empty and loaded RTCs to be carried, i.e. nCap−L

ijs = ⌊ rWe +r

ij

Cap ⌋ and nijs

Cap−E

= ⌊ rLe

ij

⌋,

s s s

respectively. Then, the transport cost rates can be calculated as

cTr

ij

c̃Tr−E

ijs = Cap−E

(3.63)

nijs

cTr

ij

c̃Tr−L

ijs = Cap−L

. (3.64)

nijs

The proposed heuristic reduces the computation time by approximately 50% for the in-

stance described in this example.108 The heuristical solution yields total costs of 21,117.75

which is about 1.3 % above the optimal value.

Ship transportation is one of the most attractive transport modes for raw chemical pro-

curement in chemical industry. Raw materials and basic chemicals are required constantly

and in huge quantities to supply continuously operating chemical production plants. The

ideal transport mode is the pipeline transport. However, pipelines dedicated for basic and

raw chemicals are seldom available due to large distances to bridge, high investment costs,

and the inﬂexibility of sourcing. In most cases, raw and basic chemicals are transported

by ships if a sea or river port is available.

Transporting chemicals in ships (or tankers) leads to a similar technological classiﬁcation

of transport carriers as for RTCs. Depending on the phase of the chemical to be trans-

ported, two main classes can be subdivided: For the transport of liquids chemical tankers

are used whereas gas tankers are used for gases.

Chemical tankers are similar to oil tankers but typically considerably smaller. Large-

scaled chemical tankers for open sea transports have a deadweight of about 40,000 tons

at maximum, whereas the largest class of (crude) oil tankers has a deadweight range from

320,000 up to 550,000 tons.109 Chemical tankers carry a number of tanks that can be

108

Note that the relaxed MC-RTP was solved to optimality although this is not necessary for a heuristic.

The procedure can be fastened by terminating optimization at a threshold for computation time or

optimality gap.

109

Chemical tankers for inland waterway transports have typically a deadweight of at most 10,000 tons.

109

the chemicals to be transported, there exist diﬀerent classes of chemical tankers which vary

in their safety features.110 Unloading and loading takes place at speciﬁc port terminals by

pumping the chemicals in or out. At modern chemical tankers, each tank can be handled

separately. To avoid explosion or other chemical reactions the tanks are always ﬁlled

either by the chemicals to be transported or by inert gas (e.g. nitrogen).

Gas tankers are technologically more sophisticated vessels. Similar to transports in

RTCs, gaseous chemicals can only be transported eﬃciently in a liqueﬁed state. Hence,

gas tankers have to keep the chemicals either under high pressure or under super-cooled

conditions.111 In particular, maintaining super-cool conditions in all tanks is technologi-

cally challenging for long trips e.g. over open sea. Despite sophisticated tank insulation,

temperature of tanks and cargo increases while travelling. This leads to a vaporization of

the super-cooled cargo and an increase of pressure inside the tanks. The resulting gas (so-

called boil oﬀ ) has to be released from tank to keep the pressure in an acceptable range.

It is either used to produce energy on board or is re-liqueﬁed. Hence, there is either a

loss of cargo or additional energy is required to keep all cargo liqueﬁed.112 Therefore, the

costs for long-haul transports of liqueﬁed gases over open sea are noteworthy. Generally,

liquefaction by pressure is economically preferred at small-sized vessels (up to 10,000 m3 ),

whereas liquefaction by refrigeration is carried out on large-scaled tankers (up to 125,000

m3 ).113 The loading procedure is more complicated compared to chemical tankers since

diﬀerences in pressure and temperature have to be harmonized before loading can take

place. While unloading gas tankers, the chemicals on board vaporize gradually. Hence,

similar to gas RTCs, a rest of the chemical remains in the tanks if it is not pumped out

(e.g. by spray pumps). However, since the transport quantities are much larger compared

to RTCs, the tanks are typically emptied completely. Most gas tankers have a transport

capacity of about 125,000 m3 .114 The payload in tons depends on the density of the

liqueﬁed chemical.

The waterway transports of raw and basic chemicals are organized diﬀerently from

RTC transports. Chemical companies usually do not own or rent tankers (for long peri-

ods) themselves, but place transport orders e.g. on spot markets.115 Hence, routing and

110

See Stopford (2009) for details.

111

Supercooled means here temperature well below 0 ○ C depending of the gas to be liqueﬁed, e.g. liqueﬁed

natural gas (LNG) is liqueﬁed at about -160 ○ C, whereas Ethylene liqueﬁes at about -100 ○ C, see

Stopford (2009).

112

For instance, the transport of liqueﬁed natural gas requires 10-25% of the energy transported in form

of cargo to maintain super-cooled conditions and re-liquefy the vaporized natural gas, see Schumacher

(2011).

113

This classiﬁcation is neither distinctive nor complete. There are also tankers providing both types of

liquefaction techniques (so-called semi-refrigerated tankers) as well as small-scaled fully refrigerated

tankers or large-scaled only pressurized tankers. See Stopford (2009) for details.

114

Most recently larger ships are laid down with a capacity of about 250,000 m3 . See Colton (2012) for

an encompassing list of LNG carriers under service or construction.

115

An exception are large petrochemical companies that often own a number of tankers to deliver their

110

scheduling of tankers is not integrated with inventory and replenishment planning per

se. Transports are organized by ship operators which own ﬂeets with diﬀerent types of

tankers. Similar to rail transports, the contracts are either quantity-based or tanker-

based, i.e. either the transport fee depends on the quantity to be shipped or a complete

tanker is chartered.116 The former option is typically used for small- to medium-sized

transport orders (say up to 10,000 tons) and organized on large-scaled tankers with mul-

tiple tanks. Typically, such large-scaled tankers operate on predeﬁned routes acquiring

numerous transport orders of various materials at various ports.117 Due to scale eﬀects,

such tankers can operate at comparatively low cost rates per m3 ⋅ km, but they are in-

ﬂexible due to the ﬁxed routes. If a single transport order has suﬃcient size, chartering

a complete (smaller-scaled) tanker is often economically preferable. Chartering smaller

tankers generates comparatively high cost rates but oﬀers more ﬂexibility.

Both types of transport contracts imply a distinction between the shipper (a chemical

company) placing the transport orders and the ship operator realizing these transport

orders. Despite the fact that the chemical companies do not own or rent tankers for long

periods, an integration of transportation and inventory planning is reasonable and can

be organized by a more or less close cooperation between chemical companies and ship

operators. From the perspective of chemical companies, a close cooperation with ship

operators requires a collaborative planning of tanker routes visiting a set of production

sites under consideration of the stock levels at the sites. Such problems are categorized

as maritime inventory routing problems. In contrast, if the cooperation is less close,

only transport orders are placed e.g. on a spot market for chemical shipments. Here, the

problem is to determine start and end node, timing, and composition of transport orders

under consideration of the stock levels at the nodes. Such problems are categorized as

maritime inventory shipping problems. The next two subsections brieﬂy review available

literature in both classes.

Maritime inventory routing and scheduling problems correspond to the class of shipping

problems for industrial operations where the shipper owns or manages the ships used for

the transports.118 In chemical industry, this class primarily addresses petrochemical com-

panies with a global network of production sites and customer terminals. The transport

quantities are huge and well predictable, justifying an integrated collaborative planning

of inventory management and routing. Maritime inventory routing problems (MIRP) are

116

For rail transports either contracts are based on the number of RTCs (wagonload contract) or complete

block trains.

117

This corresponds to the so-called liner problem, see e.g. Christiansen et al. (2004).

118

See e.g. Christiansen et al. (2004).

111

subtypes of the classic inventory routing which are characterized by the following facts:119

• There is a set of ports, where at least one product is consumed or produced (at a

constant rate).

• Ships are used to transport products between the ports in a ﬁnite time horizon.

If multiple periods are considered, the routes have to be coordinated in space and

time. Such problems are also categorized as maritime routing and scheduling problems

(MIRSP). Typically, the objective is to minimize the total operational costs over a ﬁnite

time horizon consisting of

• transport costs for operating the ships (mainly fuel and manpower),

Note that inventory costs are omitted by many approaches121 with the argument that all

inventories are owned by the same company, i.e. the total inventory among all nodes of the

network is assumed as independent from the routing and scheduling decisions. However,

this argument is not entirely correct in the long run since the frequency of transports

aﬀects the total stock among the network.122 In other words, the higher the transport

frequency (i.e. the transport costs) the lower the total stock in the network (and, hence,

total inventory holding costs). In the short run, however, inventory holding costs (in the

sense of capital commitment costs) are probably suﬃciently small to be disregarded as an

optimization criterion.

Contributions for MIRSPs mostly diﬀer in two categories: the number of products

considered and the solution method(s). Basically, the underlying problem structure is

a vehicle routing problem (VRP) and/or a pick-up and delivery problem (PDP) under

maritime-speciﬁc constraints such as port and turnover capacities as well as inventory

constraints. Since VRP and PDP are hard to solve to optimality even in their basic

form, solution methods for MIRSP mainly rely on sophisticated and tailored heuristical

procedures. First contributions (Christiansen and Nygreen (1998); Christiansen (1999))

119

For a comparison with "classic" IRPs for other modes of transport see e.g. Christiansen et al. (2004)

or Andersson et al. (2010).

120

This is not a necessary condition, however most work assumes heterogeneous ﬂeets as this is typical

for real world problems.

121

See Christiansen (1999); Al-Khayyal and Hwang (2007); Siswanto et al. (2011).

122

Also compare the discussion about pipeline scheduling, e.g. see Figure 3.5.

112

focus on one-product problems and model the underlying real-world problem as an Inven-

tory PDP with Time Windows (IPDPTW). It can be shown that inventory constraints

can be reformulated to time windows. To solve the IPDPTW, a column generation ap-

proach is proposed. Therefore, the IPDPTW is reformulated allowing a decomposition

into ship routing and harbour visiting sub-problems.123 By solving the sub-problems,

feasible ship routes and visiting sequences for all ships and harbours are generated.124

In the master problem, a combination of the identiﬁed routes and sequences has to be

determined. Therefore, a tailored branch-and-bound procedure is proposed to ﬁnd an

optimal combination of routes and sequences.

MIRSPs for multiple products require a sophisticated modelling of the tanker capacity

which has to be modelled as a set of tanks dedicated to a speciﬁc product on a speciﬁc trip.

In Persson and Gothe-Lundgren (2005) it is shown that in such a formulation inventory

constraints cannot be converted into time windows. In Al-Khayyal and Hwang (2007)

the problem is simpliﬁed by ﬁxing the assignment of tanks to products. Siswanto et al.

(2011) show that for MIRSPs considering non-dedicated tanks and multiple products even

small instances are much harder to solve to optimality than the single product problems

with dedicated tanks. Contributions also diﬀer in the technological restrictions at ports,

e.g. how many ships or products can be served simultaneously. Christiansen and Nygreen

(1998); Christiansen (1999) and Siswanto et al. (2011) allow at most one ship to be handled

at a time in a harbour, whereas Al-Khayyal and Hwang (2007) allow multiple ships to be

docked at the same time.

Almost all contributions to MIRSP use a time-continuous formulation and consider con-

stant consumption and/or production rates for the products at the sites. One exception

is Persson and Gothe-Lundgren (2005) using a time-discrete formulation. Moreover, pro-

duction planning decisions are incorporated besides routing and inventory management.

The second exception propose Christiansen et al. (2011) where a solution procedure for a

MIRSP with time-varying consumption rates is described (but no mathematical model).

The literature on MIRSPs is large and still growing, particularly addressing speciﬁc

aspects of industrial applications. E.g. in Christiansen et al. (2011) and Stålhane et al.

(2012) applications of MIRSPs in cement industry and for LNG transports are proposed.

A related branch of literature focuses on the problem of scheduling shipments when the

ships are not owned/managed by the shipper. These maritime inventory shipping prob-

lems are discussed in the next subsection.

123

For each ship a routing model is formulated which is independent from all other ships. Similarly, for

each harbour a visiting model is proposed which is independent from all other harbours.

124

Promising routes and sequences are identiﬁed by solving the LP-relaxation of the master problem

iteratively to determine columns with least reduced costs.

113

Maritime inventory shipping models face the same general problem structure as MIRSPs.

However, the organization of shipments is diﬀerent for this class of problems. In contrast

to MIRSPs, maritime inventory shipping problems (MISP) do not explicitly consider the

routing of a particular ship. Instead, it is assumed that some ships are available for

transport orders, e.g. by hiring tankers on a spot market. Hence, the planning of routes

is left to the ship operator and not integrated in the production and inventory planning.

From the shipper’s perspective, it has to be decided about

These decisions are made based on the inventory and production constraints as well as

the cost rates oﬀered for diﬀerent types of ships by the ship operators.

Ronen (2002) proposes a MISP for a single ship type where the ships’ capacities are

not modelled in detail, e.g. no tanks are modelled and the transport capacity is only

restricted by a port-based upper bound. Only one ship can leave a departure port per

period but multiple ships may arrive at a destination port in a period. The objective

function minimizes the total shipping and inventory costs where the inventory costs are

measured as the weighted sum of absolute deviations from a desired safety stock level,

i.e. no adjustments according to the target stock level are made. The shipping costs

consist of a ﬁxed part per dispatched ship and a variable part depending on the quantity

shipped. To overcome some lacks of detailedness of the approach presented by Ronen

(2002), the following extensions are made:

and charter rates.

• Diﬀerent types of tanks on a ship are modelled with varying sets of loadable prod-

ucts.

• Transport costs consist of a daily charter rate (depending on ship type) and ﬁxed

port fees (depending on the departure and arrival port and the ship type).

• Inventory costs are measured as the weighted sum of relative deviations from a

target stock level.

Due to the great variety of ship types available for chemical transports (in particular

among gas tankers) it seems reasonable to reﬂect this option by modelling a heterogeneous

114

set of tanker types. The main features characterizing tankers from an economic point of

view are the oﬀered transport capacity and the associated costs. The transport capacity

depends on the size and type of tanks on board. The type of tank also determines the

set of products that can be loaded. In the case of chemical tankers, this set is mainly

determined by the safety class of a tank. For gas tankers the liquefaction technology

plays an additional role.125 Hence, the number and the types of tanks installed on a

tanker determines its ﬂexibility and capacity.

Altogether, a MISP is proposed that minimizes the total costs for shipping and stock

holding. It has to be decided on the transport volumes to be shipped, the ship types

used for the shipments and, subsequently, the assignment of chemicals to tanks on a trip.

The following maritime inventory shipment problem with ship type and tank assignment

(MISP-STA) uses the following notation given in Table 3.17.

Sets

I = {1, ..., I} set of ports/sites

T = {1, ..., T } set of periods

K = {1, ..., K} set of tank types

V = {1, ..., V } set of tanker types

L⊆I ×I set of links

Parameters

cOs

is over-shooting cost rate

cUs

is under-shooting cost rate

cTrn

is turnover costs for chemical s ∈ S at port i

cTrv

v cost rate for chartering a tanker of type v for one period

cPort

iv port fee for tanker type v at port i

local net balance of chemical s ∈ S at port i ∈ I in period t ∈ T ; if

ωits

ωits > 0 there is a net deﬁcit, otherwise a surplus

Cap

qsk capacity of tank type k for chemical s ∈ S

nCap

vk number of tanks of type k installed on tanker type v

sCap

is inventory capacity for chemical s ∈ S at port i ∈ I

sTar

is target stock level for chemical s ∈ S at port i ∈ I

sIni

is initial stock level for chemical s ∈ S at port i ∈ I

initial ﬂows of chemical s ∈ S from i to j in period t ∈ T in period

xIni

ijτ s

τ ∈ {1 − tTrv

ij , ..., 0}

tTrv

ij integer valued travelling time on link (i, j) ∈ L

Decision variables

125

E.g. Ethylene can be transported eﬃciently under super-cooled conditions only. For most other basic

chemicals also liquefaction by pressure is possible. See Stopford (2009).

115

yijtv number of tankers of type v used on trip (i, j, t) (integer)

number of tanks of type k dedicated to chemical s on trip (i, j, t)

zijtsk

(integer)

Variables

relative over-shooting in % of the target stock of chemical s ∈ S in

oits

period t ∈ T at port i ∈ I

relative under-shooting in % of target stock of chemical s ∈ S in

uits

period t ∈ T at port i ∈ I

Table 3.17: Sets, parameters, variables, and decision variables for the MISP-STA

Since it is assumed that tankers are chartered for a speciﬁc transport order, it is rea-

sonable to assume that a time-dependent charter rate is charged when a tanker is booked.

Hence, the charter costs on link (i, j) ∈ L are calculated by a charter rate per period cTrv

v

multiplied by the trip’s duration tTrv

ij .

126 Furthermore, port fees are to be expected de-

pending on the handling time (for unloading or loading) as well as the size of the ships.127

These together constitute the transport costs depending on the travel time, ship type,

and ports (see ﬁrst two rows of (3.65)).

As a counterpart, inventory costs have to be faced. As discussed for the MC-RTP model,

it is assumed that target stock levels (in the sense of safety stocks) are predeﬁned for each

port/production site and product. Deviations from these target levels are to be minimized

and are included in the objective function multiplied by speciﬁc cost rate equivalents.

These cost rate equivalents are higher for negative deviations (i.e. undershooting) and

lower (or zero) for positive deviations (i.e. overshooting). As before, undershooting cost

rate equivalents reﬂect the expected costs for a plant shutdown due to raw material short-

age whereas overshooting cost rate equivalents reﬂect unnecessary stock holding costs. As

in the MC-RTP, stock deviations are measured relative to the target stock levels taking

into account varying target stock levels.128 The total costs of transports and inventory

holding are minimized and formulated as

T C = ∑ ∑ ∑ (yijtv ⋅ (tTrv

ij ⋅ cv + civ

Trv Port

jv )) +

+ cPort

(i,j)∈L t∈T v∈V

∑ ∑ ∑ ∑ (zijtsk ⋅ (cTrn

is + cjs )) +

Trn

126

To reﬂect distance-dependent discounts the cost rates can be assumed to be negotiated for each trip

individually, i.e. cTrv

ijv .

127

See Stopford (2009) for more details.

128

See the discussion in subsection 3.3.2.

116

∑ ∑ ∑ (cOs

is ⋅ oits + cis ⋅ uits ) .

Us

(3.65)

i∈I s∈S t∈T

ji )s

j∈I

− ∑ xijts ∀i ∈ I, s ∈ S, t ∈ T (3.66)

j∈I

Cap

) ∀(i, j) ∈ L, s ∈ S, t ∈ T (3.67)

k∈K

Cap

(3.68)

s∈S v∈V

sits

− 1 = oits − uits ∀i ∈ I, s ∈ S, t ∈ T (3.69)

sTar

is

si,0,s = sIni

is ∀i ∈ I, s ∈ S (3.70)

sits ≤ sCap

is ∀i ∈ I, t ∈ T , s ∈ S (3.71)

xijτ s = xIni

ijτ s ∀(i, j) ∈ L, τ ∈ {−tTrv

ij ...0}, s ∈ S (3.72)

xijts ∈ R+ ∀i ∈ I, j ∈ I, t ∈ T , s ∈ S (3.73)

Constraints (3.66) constitute the inventory balance equations consisting of the previous

period’s stock, local consumption/production (ωits ), and in-/outﬂows from/to other ports.

Constraints (3.67) assure that a suﬃcient number of tanks is assigned to chemical s in

order to store the shipping quantity xijts on board. Similarly, constraints (3.68) restrict

the total number of dedicated tanks on a trip to the number of tanks installed on board

of the chartered ships. Relative stock deviations from the desired target stock levels

are calculated by constraints (3.69). Constraints (3.70) initialize the stock levels at the

beginning of the planning horizon whereas (3.71) assure that stock capacities are not

exceeded. Finally, constraints (3.72) incorporate ﬂows from previous periods which allows

applying the model in a rolling horizon environment.

The following example shows the applicability of the MISP-STA by means of the same

artiﬁcial setting as described in example 6.

Example 7 (Maritime inventory shipping). Assume three ports with associated production

sites where three chemicals are produced/consumed. Shipments are planned for a 14-

117

conﬁgurations are set as stated in Tables 3.15a and 3.16. Three types of tankers are

available for transportation. Diﬀerent numbers and types of tanks are installed on the

tankers. Two types of tanks are distinguished. Table 3.18 shows the number of tanks

installed on the tankers and tank capacities for all products.

1 2 3 1 2 3

tank k

1 1 2 4 0 500 0

2 1 4 8 450 100 600

cTrv

v 200 300 400 – – –

Table 3.18: Technical speciﬁcation of tankers and tanks (nCap Cap Trv

vk , qsk , cv )

v are increasing non-linearly with increasing tanker size to reﬂect

economies of scale for larger-sized tankers. Two tank types are used, k = 2 is an all-round

tank that can be used for all chemicals, whereas type k = 1 is specialized for chemical s = 2.

The remaining cost rates are set as follows: cPort

iv = 10 and cTrn

is = 1 for all i ∈ I, v ∈ V,

and s ∈ S, respectively. Cost rates for over- and undershooting target stock levels are set

is = 1000 and cis = 1 for all i ∈ I and s ∈ S. This setting was solved to

as in example 6: cUs Os

optimality in about 5.7 hours, the corresponding minimal total cost is 13,514.130 Again, the

MISP-STA can be categorized as a speciﬁc network design problem where arc capacities

are associated with ﬁxed costs for chartering ships. Therefore, the same remarks w.r.t. the

model’s complexity hold as for the MC-RTP. In this problem instance, however, the ratio

between transport and stock holding cost is less favourable as the transport cost rates are

considerably higher than in the MC-RTP instance. Moreover, the transport capacity is

not restricted by a predeﬁned set of RTCs but purely depends on the number of chartered

ships.

Figure 3.12 shows the resulting optimal transport quantities dispatched in the ﬁrst eight

periods.131

Mainly tanker type v = 1 is chartered since handled quantities are suﬃciently small.

Exceptions can be observed in periods t = 1, t = 4 and t = 7 where chemicals s = 1 and s = 3

are shipped jointly which requires ship type v = 2. Primarily, chemical s = 3 is transported

from port/site i = 2 to i = 1 and i = 3. Chemical s = 1 is shipped from i = 1 and i = 2 to

i = 3. The transports from i = 1 to i = 3 are routed via i = 2. Note that direct transports

from i = 1 to i = 3 last four periods whereas indirect transports via port i = 2 last only

129

Due to the longer transport and handling times, a rougher time scale can be expected in the mar-

itime/waterway context compared to rail context. I.e. a period’s length is expected to be one or two

day(s).

130

The calculation was performed on a 2.6 GHz machine with IBM CPLEX 12.5, see IBM (2010).

131

Associated indices refer to the transported chemical and the used ship type on a trip (i, j, t).

118

t =8 t =2 t =3 t =4 t =5 t =6 t =7 t =9

450 500 550 600 300 350 400 450

600 500 400 600 500 400 700 600

i =8

475 450 425 400 375 350 325 300

x s=3 = 400

x s=3 = 300 y v=8 = 8

y v=8 = 8 x s=8 = 350

y v=8 = 8

400 450 500 400 450 500 400 450

100 900 8000 8000 900 8000 900 8000

i =2

275 250 225 350 325 300 275 250

x s=8 = 850 x s=2 = 850 x s=8 = 850 x s=3 = 400 x s=2 =

y v=2 = 8 y v=8 = 8 y v=2 = 8 y v=2 = 8 y v=8 = 8

100 8800 8000 100 8000 100 900 8800

i =3

550 600 500 550 600 650 700 500

target stock chemical 8 chemical 2 chemical 3

8000

Figure 3.12: Optimal stock levels and chemical ﬂows for periods 1 to 8 of the MISP-STA example

119

three periods. Hence, indirect transports induce lower transport costs. Chemical s = 2 is

transported from the only producer site i = 3 to site i = 2.132

Comparing Figure 3.12 and Figure 3.10, it can be observed that the total quantity shipped

in the ﬁrst periods is higher for the maritime setting than for the corresponding rail setting

(example 6). This is caused by the restricted number of RTCs in the rail setting. The

RTCs have to be re-positioned to perform material transports. This leads to comparatively

more balanced stock levels in the maritime setting. To illustrate this fact, Figure 3.13

shows the stock levels and target stocks of all chemicals at all ports over the complete

planning horizon.

In the maritime setting, all stock levels are closer to their corresponding target stock

levels compared to Figure 3.11. This is particularly apparent for chemicals s = 1 and s = 3

at site i = 3. In the rail setting the stocks of both chemicals were below their target stocks

(almost) over the complete planning horizon. In the maritime setting, both stocks are

considerably closer to their target stocks. This corresponds to the stock patterns of both

chemicals at site i = 2 where the corresponding stocks’ overshooting is reduced. I.e. in

the maritime setting transports of chemicals s = 1 and s = 3 from site i = 2 to i = 3 are

economically preferable (in contrast to the rail problem).133

In general, the maritime setting oﬀers more transport ﬂexibility in the sense that ships

can be dispatched independent of the availability of RTCs. On the other hand, the freight

consolidation on ships is restricted by the technical constraints of the tanks available on

board. In the given examples, the ﬁrst fact prevails resulting in smoother stock patterns.

In contrast to the MC-RTP, the MISP-STA is more complex due to the larger number

of integer variables to be determined. To solve larger instances, a heuristical procedure

can be set up as follows:

1. Build a relaxed MISP-STA by dropping the ship ﬂow variables yijtv and (3.68). Re-

place the ﬁrst part of (3.65) by

∑ ∑ ∑ ∑ zijtsk ⋅ tTrv

ij ⋅ c̃k

Trv

(3.75)

(i,j)∈L t∈T s∈S k∈K

Similar to the MC-RTP, a considerable part of the model’s complexity is inherited from

the integer decisions determining the number of tankers. For variabilization, the cost rates

132

Transports to site i = 1 take place in later periods.

133

Compare also the transport ﬂows displayed in Figure 3.12 and Figure 3.10.

120

chemical s = 1

location i = 1 stock level

chemical s = 2 target stock

chemical s = 3

750 1000

500

stock level

250

0

location i = 2

1250

750

stock level

250

0

location i = 3

750

stock level

250

0

0 2 4 6 8 10 12 14

period

Figure 3.13: Optimal stock levels for the MISP-STA instance of example 7

121

c̃Trv

k approximate the shipment charges on a tank basis and are calculated as follows

Cap

c̃Trv

k = v

. (3.76)

V ⋅ ∑k∈K nCap

vk

The heuristic delivers a solution for the described problem instance within 8 seconds. The

solution’s total cost is 14,379 which is approximately 6% above the optimum.

123

supply chains

In the previous chapters isolated planning problems in chemical industry are described, re-

viewed and modelled. These approaches allow analysts to model typical chemical produc-

tion processes and logistical planning problems in chemical production networks. Chemi-

cal production networks consist of many chemical plants clustered at chemical production

sites. Such networks can be seen as an important part of chemical supply chains (SC). In

the scientiﬁc literature, there is no unique and concise deﬁnition what a SC is, but some

common features are prevalent in most deﬁnitions:

at a market.

the product or service can be oﬀered.

• multiple units: Several business units (companies) are responsible for the transfor-

mation processes.

Hence, a chemical’s SC can be deﬁned as all (business) units and processes involved in

the production and distribution of this chemical. From this point of view, chemical pro-

duction networks of large-scaled chemical companies are major parts of the SCs of many

chemicals. A particular property of chemical SCs is that a comparatively small number of

independent companies is involved in the SC. This is because chemical companies are of-

ten deeply integrated, i.e. many subsequent production steps are processed by a chemical

company. On the supply side, raw material suppliers such as oil-producing companies and

reﬁneries are to be taken into account. Moreover, logistical service providers managing

transport and distribution processes play a vital role in chemical SCs.

To manage SCs, a broad body of (scientiﬁc) literature has emerged since the 1980s.1

Supply chain management (SCM) aims at the planning, execution, and control of all

transformation processes in a deﬁned SC such that the SC’s performance is optimized.2

The main focus of SCM is on the coordinated planning of the SC processes. For planning,

1

See Croom et al. (2000) for an overview.

2

This is often operationalized as minimizing the SC’s total costs or maximizing the SC’s total proﬁt.

See Stadtler (2005) for a more detailed discussion and deﬁnition of SCM.

T. Kirschstein, Integrated Supply Chain Planning in Chemical Industry, Produktion und Logistik,

DOI 10.1007/978-3-658-08433-2_4, © Springer Fachmedien Wiesbaden 2015

124

SCM draws many approaches from related disciplines (such as logistics and operations

management), but focuses on their integration and interactions.

In the management of chemical SCs, product ﬂows within chemical production networks

are to be planned from a focal point of view. I.e. local and network-wide processes have to

be managed such that the production network’s total performance is optimized. Figure 4.1

shows the schematic overview of the exemplary chemical SC where relevant elements for

integrated planning are highlighted.

intra-site inter-site

S1 C1

intra-site

intra-site

S2 C2

Figure 4.1: Chemical SC scheme for integrated planning (relevant elements highlighted)

For integrated planning approaches, all elements in the considered chemical SC are

relevant including transport relations to (external) customers and suppliers. Only external

suppliers or customers themselves are not modelled in detail.

Core elements of chemical SCs considered in integrated planning approaches at the

tactical/strategical level are capacities and parameters of the production systems and

the logistical systems (e.g. turnover or inventory capacities). Since sites and plants are

interconnected, local adjustments more or less immediately aﬀect the remaining sites/-

plants. Integrated approaches capture these spill-over eﬀects by a combined modelling of

the interdependent components.

Integration not only focuses on the spatial or temporal dimension of an SC but also

on separated planning (sub-)problems. These (sub-)problems and their interdependencies

are modelled in a common framework. Basically, two options for integration can be

distinguished:3 The ﬁrst option is to merge the sub-problems into a monolithic model

(also called deep integration). The second option is to stay with the decomposition into

3

See Dolk and Kottemann (1993) or Geoﬀrion (1999).

125

called functional integration). In the latter case, a transfer of data has to be organized

between the sub-problems in order to anticipate spill-over eﬀects between related sub-

problems. Often a monolithic optimization model can be formulated, but its complexity

prohibits a direct solution using standard techniques. To tackle such a problem basically

three options are available:

1. The decomposition of the monolithic model into suitable sub-problems and setting

up a functional integration scheme.4

of the monolithic model.

The following subsection reviews the literature on integrated planning problems with

special focus on (basic) chemical industry. The papers reviewed are categorized according

to methodical features and problem characteristics.

The management of SCs consists of manifold aspects. These aspects encompass traditional

planning tasks such as distribution planning or vehicle routing as well as network-speciﬁc

problems such as the coordination of partners along the SC. Usually, the planning tasks

are solved individually where the results of a speciﬁc task serve as input for subsequent

tasks. This is called a sequential planning approach. To provide a categorization for SCM-

related planning tasks, the so-called supply chain planning matrix was developed.5 It cat-

egorizes SCM planning problems according to the planning level (i.e. strategical, tactical,

or operational problems) and the progress of the value-adding process (i.e. procurement,

production, distribution or sales).

In this concept, the strategic network planning comprises network design and structure

decisions e.g. about the location of SC facilities and their corresponding layout.

At the tactical level, the master planning module comprises aggregated models of the

considered SC taking into account capacities of the production and logistics system as

well as demand forecasts. Here, aggregated material ﬂows between the SC facilities are

planned for time horizons up to one year. A rough time discretization (e.g. in months

or weeks) is used to reﬂect e.g. varying demand patterns. Demand forecasts are made at

4

Note that the suitable sub-problems do not need to be the same as the original sub-problems.

5

See Meyr et al. (2008) and Stadtler (2005). Note that there is a lot of work from numerous scientiﬁc

perspectives labelled with SCM, see Croom et al. (2000) for a detailed analysis and categorization of

SCM approaches. Based on this classiﬁcation scheme this work reviews SCM approaches at a network

level modelling material ﬂows.

126

the demand planning module which comprises models to compute demand forecasts for

diﬀerent regions and time horizons by assuming speciﬁc types of demand processes.

At the operational level, problems are considered on a detailed perspective dealing with

the short-term planning and control of basic SC operations. The aggregated material ﬂows

determined at the tactical level are broken down to the local level on a short-period basis

like a daily or hourly time frame. For the pre-determined production quantities assigned

to a speciﬁc production site on the tactical level, plans and schedules for all plants at this

site are derived in the production planning & scheduling modules. Similarly, replenishment

and transport decisions are determined in the inventory management and distribution &

transport planning module, respectively.

The general SCM matrix can be adapted to better reﬂect the speciﬁc characteristics of

basic chemical industry.6 In the traditional SC planning matrix, the material requirement

planning modules are concerned with determining order quantities of raw and intermediate

materials necessary to realize a pre-deﬁned production plan. These tasks are particularly

challenging if the variety of handled materials is large.7 In basic chemical industry, how-

ever, the variety of raw and intermediate chemicals is comparatively small while their

demand is high and mostly determined by the plants’ technical speciﬁcations.

Moreover, the means of transport diﬀer for procurement and distribution transports.

For raw and intermediate transports often pipeline, ship, and rail are used. In contrast,

for the distribution of ﬁnal chemicals rail and road transports are prevalent. As pipeline,

ship, and rail are comparatively inﬂexible and diﬀerently organized compared to road

transports, an independent planning of procurement and customer transports is justiﬁed.

Therefore, the procurement planning module encompasses the planning of procurement

transports of raw and intermediate chemicals and the planning of their local stocks. In

contrast, the distribution transports module comprises the tasks of the classic distribution

& transport planning module for ﬁnal chemicals.

The transport capacities provided along a chemical SC have to be planned on the tac-

tical level. In the case of ships or RTCs, the corresponding ﬂeet sizes are determined. To

enlarge ﬂeets, equipment can be bought or rented. To reduce ﬂeets, equipment can be sold

or renting contracts may not be prolonged. Such ﬂeets are sometimes used commonly by

multiple chemical companies e.g. by short-term renting contracts or by exchanging RTCs

or ships (so-called swaps). Therefore, the original master production planning module is

complemented by the master transportation planning module capturing all tactical de-

cisions about the capacities of the logistical system in the SC.8 Pipeline capacities are

determined by technological details and, thus, are subjects to strategical issues as they

cannot be adapted easily in the short run.

6

See Zoryk-Schalla et al. (2004) for a report from an Aluminium-producing company about implementing

an advanced planning system based on an adapted SCM matrix.

7

See Meyr et al. (2008).

8

Note that this module also contains planning problems determining turnover capacities at the sites if

they limit the transport quantities.

127

The distinction into a strategical, tactical, and operational level is retained but re-

named into design, conﬁguration, and operations to better reﬂect the terms most used in

literature.9 Figure 4.2 shows the adapted SCM matrix.

guration transportation planning production planning planning

tions planning planning planning scheduling fulﬁllment

The basic planning problems corresponding to the speciﬁc modules are brieﬂy described

as follows:

– Strategic planning: This block comprises planning problems determining the

general structure of the considered SC, e.g. location and structure of produc-

tion sites, type of production processes or types and capacities of logistical

facilities.10

• conﬁguration:

– Master production planning: This module encompasses decisions about the

usage of production capacities of the considered SC, e.g. the determination of

production quantities and their assignment to production sites.11

– Master transportation planning: Determination of aggregated transport ﬂows

among the participants of the SC and provided transport capacities.12

– Demand planning: Determination of demand forecasts for a ﬁnite planning

horizon. They are based on historical records of former periods, external pre-

9

See e.g. Melo et al. (2009).

10

See e.g. Vidal and Goetschalckx (1997) or Tsiakis et al. (2001) for an overview.

11

In chemical industry this is often encompassed by campaign planning problems. See Kallrath (2005) or

Grunow et al. (2002).

12

Examples are e.g. rail car ﬂeet sizing models (Cheon et al., 2012), container ﬂeet sizing (Dong and

Song, 2009), or ship ﬂeet sizing (Ronen, 1993).

128

dictors (e.g. economic growth indicators), and assumptions about the demand

process.13

• management

– Production planning: Determination of local production quantities and assign-

ment to local plants.14 For multi-product batch processes, the aim is to deﬁne

batch sizes for diﬀerent chemicals and to assign them to available plants.15 In

case of multi-product continuous processes, the aim is to determine the produc-

tion mode, i.e. the physical conditions and raw material composition specifying

the product mixture.16

– Production scheduling: Determination of mode sequences/schedules of batches

assigned to the plants. The changeovers between production modes typically

induce costs for cleaning, maintenance, and/or lost material.17

– Procurement planning: Determination of transport modes and transport ca-

pacities for procurement transports as well as planning of inventories of raw

and intermediate chemicals at a site.18

– Distribution planning: Planning of distribution processes, e.g. management of

distribution stocks, routing, and scheduling ﬁnal chemical transports to cus-

tomers.19

– Demand fulﬁlment: Management of customer orders, e.g. tracking of orders

along the production process, acceptance of new orders, setting of initial order

due dates.20

The modules displayed in Figure 4.1 address diﬀerent planning problems and associated

planning models. There are plenty of interdependencies between the blocks as the out-

come of a particular module is typically the basis for subsequent lower-level modules. In

hierarchical planning, long-term/strategic decisions form the basis for tactical planning

whose outcome typically guides operational plans. In a hierarchical planning approach,

top-level decisions anticipate the low-level decisions to a certain extent.21 However, the

13

See Stadtler (2005).

14

The distinction between production planning and scheduling problems is not always consistent, see

Kallrath (2002) or Berning et al. (2004).

15

For an overview on integrated planning and scheduling, a detailed description in the context of multi-

product plants can be found in Maravelias and Sung (2009).

16

This problem is extensively studied for reﬁnery operations planning leading to non-linear process mod-

els, see Zhang et al. (2001), Li et al. (2005), or Alhajri et al. (2008).

17

See e.g. Méndez et al. (2006) for an extensive overview and classiﬁcation for scheduling of chemical

batch processes.

18

This comprises e.g. inventory routing models for procurement via ship or pipeline (see Ronen (2002)

or Moura et al. (2008)) and local tank management models (see e.g. Saharidis et al. (2009)).

19

This block comprises e.g. the branch of MIRSPs (see e.g. Christiansen et al. (2004)) and most pipeline

scheduling models (see e.g. Cafaro and Cerdá (2010) or Neiro and Pinto (2004)).

20

See Kilger and Meyr (2008).

21

See e.g. Fleischmann and Meyr (2003).

129

anticipation is never perfect such that the ﬁnal solution of top- and base-level problems is

typically not (globally) optimal. To overcome this deﬁciency, integrated planning models

are often build by aggregating planning problems from adjacent blocks either horizontally

or vertically.

To categorize the literature on integrated planning, the modelling technique is an im-

portant criterion. Diﬀerent types of models are suitable depending on the planning prob-

lem’s requirements. In general, when addressing long-term problems, assumptions about

the occurrence of future events have to be made. However, the future is known for its

uncertainty. Hence, techniques capable to handle uncertainty are prevalent. Such ap-

proaches can be further distinguished according to the stochastic elements considered

(which is typically the demand). For short-term problems, uncertainty is (most often)

not an obstacle. For operational problems, mathematical optimization is widely used,

albeit simulation is still a reasonable option.22 Mathematical optimization models are

categorized w.r.t. linearity (linear vs. non-linear), the domain(s) of decision variables (in-

teger/binary vs. continuous), and stochasticity (stochastic vs. deterministic). Simulation

techniques come to the fore when at least one of the following criteria prevails:23

In general, these criteria often hold for planning problems with a rather long planning

horizon. Therefore, simulation-based models are prevalent especially for SC conﬁguration

problems.24 Special classes of simulation techniques are distinguished.

In this work the term simulation refers to stochastic and dynamic models.25 If all

components are modelled continuously, this is typically called a system dynamics model.26

Such a model is reasonable either at a high aggregation level27 or when all modelled

processes are indeed continuous.28 The system studied is tracked continuously and its

(aggregated) behaviour can be described at any point in time.

22

See Papageorgiou (2009) for details and Adhitya and Srinivasan (2010) for an application of simulation

for a detailed process modelling.

23

See Law (2007) or Carson (2004) for a more detailed discussion about these prerequisites. The sub-class

of deterministic simulation is typically referred to as computer experiments, see Santner et al. (2003)

for more details. In this work the focus is on stochastic simulation models.

24

See Kleijnen and Smits (2003) for an overview.

25

I.e. there exist stochastic process elements to be modelled and the system is studied over time, see Law

(2007) for a more detailed classiﬁcation.

26

See Ogata (2003) for an introduction.

27

E.g. when (discrete) objects handled in a network can be aggregated to continuous ﬂows.

28

For more details about system dynamics models in SCM see e.g. Kleijnen (2005). For applications see

e.g. Rabelo et al. (2007) or Venkateswaran et al. (2004).

130

In contrast, if the processes of the modelled system are discrete, the corresponding

model is called a discrete-event simulation. Here, the system’s states are calculated at a

ﬁnite number of points in time. This type of simulation is often used for atomic simula-

tions, i.e. when a system is modelled in detail.29

A third category are so-called agent-based simulation models. This sub-class of sim-

ulation models is characterized by a network of interacting agents. Each agent reacts

to stimuli from other agents and environmental variables depending on mathematically

formulated decision rules.30 This type of simulation is particularly useful to model inter-

actions between interrelated, but independent entities such as business units, companies,

or market actors.31

To categorize the literature on integrated SC planning in chemical industry the following

criteria are used:

• SCM matrix refers to the modules of the adapted SCM matrix (Figure 4.2) addressed

by the reviewed reference. Capital letters indicate the planning level (design, con-

ﬁguration, and operation) with at most two superscripts. The ﬁrst superscript is

p, t or d referring to transportation, production, or demand. The potential sec-

ond superscript speciﬁes the sub-problem on the operational level with p, s or d

referring to procurement/production, scheduling or distribution/demand fulﬁlment.

E.g. the combination Otp refers to the procurement planning module whereas the

combination Opp refers to the production planning module.

• Modelling technique refers to the technique used to formalize the planning problem

and comprises mathematical optimization approaches such as (mixed-integer) linear

programs [(MI)LP], (mixed-integer) non-linear programs [(MI)NLP], or stochastic

programs [s(MI)LP] as well as simulation techniques such as agent-based models

[AB], discrete-event models [DE], or system dynamics models [SD]. If mathematical

optimization models are embedded in a simulation framework, this is indicated by

a connector. E.g. DE-MILP indicates that a MILP is embedded in a discrete-event

simulation framework.

• Objective refers to the objective pursued (such as costs, net present value [NPV],

corporate value [CV], or proﬁt). If multiple objectives are considered, this is stated

by [mu].

continuous [co] or batch [ba] production is modelled. For aggregated models the

technology might not be modelled in detail, then the production technology is un-

speciﬁed [us].

29

See e.g. Law (2007), chap. 1 for more details.

30

See e.g. Sterman (2000) for an encompassing textbook or Kleijnen (2005) for a brief introduction.

31

Therefore, agent-based simulation models are often applied to model market or social interaction

networks, see e.g. Axelrod (2001).

131

• Time scale describes the way the time dimension is represented. Continuous [co] or

discrete [di] models are distinguished if time is modelled explicitly. Otherwise, the

model refers to a single period [si].

• Transport mode refers to the transport mode(s) considered (ship [sh], pipeline [pi],

road [ro], or rail [ra]). For aggregated models only ﬂows might be relevant. Then,

the mode is unspeciﬁed [us].

(e.g. demand [de], prices [pr], or multiple [mu]).

• Solution method refers to the method applied to solve the proposed model e.g. (com-

mercial) standard solvers [standard] (such as CPLEX or CONOPT), a speciﬁc op-

timization method [speciﬁc], a hierarchical decomposition approach [decomp.], or a

heuristical procedure [heur.]. For simulation models, often a ﬁnite set of scenarios

is evaluated and compared [scen.]. But also genetic algorithms are sometimes used

for simulation optimzation [GA].32

If a criterion is not included in the proposed model, this is indicated by a horizontal line.

In general, simulation approaches are advantageous if a static framework of entity-

processing units exists, i.e. the system’s general structure is static. But for strategic

problems the structure of a SC is typically the subject to be altered. As this work aims at

a framework relying on simulation, literature primarily focusing on strategical problems

is out of focus.

At the tactical level, however, more simulation-based approaches can be found. Ta-

ble 4.1 shows a classiﬁcation of literature for integrated SC conﬁguration and operations

planning where 7 out of 20 reviewed articles propose simulation-based approaches.

Among the simulation approaches Garcia-Flores and Wang (2002), Mele et al. (2006),

and Puigjaner and Guillén-Gosálbez (2008) propose agent-based simulation models where

the agents (partially) utilize MILPs to determine their decisions (hybrid model).33 These

approaches primarily investigate the organizational structure of an SC, i.e. the interac-

tion of involved decision making units. In Garcia-Flores and Wang (2002) a chemical SC

producing paints and coatings is modelled by using six classes of agents: retailers, lo-

gistic service providers, warehouses, purchasing departments, plants, and suppliers. The

production processes at the plants are organized as multi-purpose, multi-product batch

processes. The precise operational planning and scheduling of batches is carried out by a

MILP implemented in a standard software package. Decision ﬁelds are the parameters of

the inventory policy and the parameters of the negotiation scheme between the warehouse

in cases of stock-outs. I.e. in case of a stock-out at a warehouse, the requested order is

32

This classiﬁcation encompasses the solution methods applied in the reviewed literature. A more detailed

classiﬁcation (e.g. addressing meta-heuristics, math-heuristics, etc.) is omitted for the sake of brevity.

33

E.g. site managing agents decide about the assignment of production orders to plants whereas ﬁnancial

agents decide about loans to be taken.

132

type of

time transport uncer- solution

Reference SCM matrix modelling objective produc-

scale mode tainty method

tion

McDonald and Karimi (1997) (Opp , Ops , C p , C t ) M ILP cost ba di us de standard

Vidal and Goetschalckx (2001) (C p , C t ) N LP proﬁt co si us — heur.

Gjerdrum et al. (2002) (C p , C t ) M IN LP multiple us di us — speciﬁc

pp

Gupta and Maranas (2003) (O , Ops , C p , C t ) sM ILP cost ba di us de decomp.

Jackson and Grossmann (2003) (C p , C t ) N LP proﬁt co di us — decomp.

Gupta and Maranas (2004) (C p , C t ) sM ILP NPV/ROV us di us de standard

Ryu et al. (2004) (C p , C t ) sLP proﬁt us si us de speciﬁc

Chen and Lee (2004) (C p , C t ) M IN LP multiple ba di us de+pr standard

Oh and Karimi (2006) (C p , C t ) LP proﬁt co di us — standard

Yi and Reklaitis (2007) (Opp , Ops , C p , C t ) M IN LP cost ba co us — speciﬁc

Amaro and Barbosa-Póvoa

(Otd , Opp , Ops , C t , C p ) M ILP proﬁt ba di mu — standard

(2008)

Al-Othmann et al. (2008) (C t , C p ) sLP proﬁt co di us de+pr standard

pp

Kim et al. (2008) (O , C p , C t , D) M IN LP proﬁt co si mu — standard

Garcia-Flores and Wang (2002) (Opp , C t , C p ) AB + M ILP cost ba di us mu scen.

Jung et al. (2004) (Opp , Ops , C p ) DE + M ILP cost ba di us de speciﬁc

Mele et al. (2006) (Ops , Opp , C t , C p ) AB + M ILP proﬁt ba di us mu GA

Puigjaner and Guillén-Gosálbez

(C t , C p ) AB + M ILP multiple us co us mu GA

(2008)

ps pp t p

Jung et al. (2008) (O , O , C , C ) DE + M ILP inventory ba di us de scen.

Pitty et al. (2008) (Otp , Opp , C t , C p ) DE proﬁt co di pi/sh mu scen.

Adhitya and Srinivasan (2010) (Opp , Ops , C t , C p ) DE proﬁt ba co us mu scen.

abbr.: ba...batch; co...continuous; de...demand; di...discrete; mu...multiple; pi...pipeline; pr...price; ra...rail; ro...road; sh...ship;

si...single; us...unspeciﬁed

Table 4.1: Classiﬁcation of literature on integrated SC conﬁguration and management planning in chemical industry

133

delivered from another warehouse and the charge of additional transport costs is subject

of the negotiation process between the agents. However, logistical details determining the

additional transport cost are out of scope.34 The optimization is based on an evaluation

of a pre-deﬁned set of scenarios. Mele et al. (2006) propose a similar simulation approach

with the same assumptions for the logistical processes and production processes as well

as the same optimization model for operational planning and scheduling of batches. This

work proposes an optimization procedure based on a genetic algorithm which seeks for an

inventory parameter constellation maximizing the SC’s expected total proﬁt. Puigjaner

and Guillén-Gosálbez (2008) extend this work by dealing with multiple objectives. The

applied multi-objective genetic algorithm is NSGA-II.35

Another hybrid simulation model is proposed by Jung et al. (2004, 2008). Here, a

discrete-event simulation is chosen to model the stochastic environment. The focus is on

the organization of production processes, i.e. optimization of operations is in focus. In

Jung et al. (2004) a single-stage production-distribution problem is considered where the

production planning and scheduling of multiple batch plants is embedded in a stochastic

rolling horizon environment.36 The objective is to maximize the expected proﬁt by decid-

ing on the safety stocks at the plants without falling below a minimum customer service

level. The customer demand is the only stochastic component. For optimization a stochas-

tic gradient approach is applied which iteratively increases the safety stocks for each com-

bination of product and plant until the expected proﬁt cannot be increased any further.

Jung et al. (2008) extend this work by considering a multi-stage production-distribution

system with pure inventory holding facilities and combined facilities for production and

inventory holding. For both types of facilities the inﬂuence of inventory parameters on

performance measures is analysed by simulation experiments. Thee approaches incorpo-

rate operational decisions by MILPs. Hence, optimal reactions on varying system states

are modelled on the operational level. However, both models omit logistical details on

the operational level (only ﬂows between facilities are considered).

Pure simulation approaches are proposed by Pitty et al. (2008) and Adhitya and Srini-

vasan (2010). Pitty et al. (2008) propose a discrete-event simulation model for a reﬁnery

supply chain. Operational decisions such as unloading schedules and production planning

are made based on simple priority rules. Various conﬁgurations of the modelled SC are

studied and compared to reveal optimization potentials. This approach explicitly consid-

ers some details of ship and pipeline transports. Adhitya and Srinivasan (2010) describe a

discrete-event simulation model for an SC producing and distributing lubricant additives.

Here, batch production is modelled. Again, operational production decisions are made by

priority rules and a scenario analysis is conducted to evaluate the eﬀects of other priority

34

A classic transportation problem is solved to determine the total transport eﬀort.

35

See Deb et al. (2007).

36

For production planning the MILP proposed by McDonald and Karimi (1997) is used. Production

schedules are determined by heuristics to ease solvability.

134

Core decisions on the tactical level are the assignment of production quantities to the

available production sites and the distribution of raw, intermediate, and ﬁnal products

among the various participants in the SC. Of particular importance is the integration of

production planning and ﬁnal product distribution. Numerous approaches are devoted to

this kind of production-distribution problem.38 Vidal and Goetschalckx (2001), Gjerdrum

et al. (2002), and Chen et al. (2003); Chen and Lee (2004) study eﬀects of ﬁnancial

instruments and restrictions on production-distribution problems, in particular in multi-

national companies. Vidal and Goetschalckx (2001) study the problem of determining

transfer prices of intermediate products which are transported between subsidiaries of

a multi-national company. This induces tax costs leading to quadratic terms in the

objective function due to the multiplication with the production and transport quantities

to be determined simultaneously. To solve the resulting NLP a sequential ﬁx-and-relax

procedure is proposed. Transports are represented as ﬂows between SC facilities. I.e. no

details of the logistics system are modelled. The same holds for Gjerdrum et al. (2002)

where transfer prices between multiple companies are studied. Here, the problem is to

ﬁnd a "fair" proﬁt allocation by determining transfer prices. The resulting MINLP is

solved by a branch-&-bound algorithm.39 A fair distribution of the total expected proﬁt

is also pursued in Chen and Lee (2004) next to customer service levels, safety stocks,

and the robustness of solutions. Details of the logistical system are not modelled, but

the transport costs depend on the transport quantity reﬂecting economies of scale of

diﬀerent transport modes. This model considers discrete scenarios for customer demand

and product prices.40 A fuzzy approach is proposed which relies on a component-wise

normalization of the vector of multiple objectives. To merge the normalized objective

vector into a single measure, diﬀerent metrics are discussed such as the minimum of the

normalized objective components or the product of all objective components. None of the

reviewed approaches incorporates transportation aspects in detail.

As tactical decisions aﬀect an SC’s performance for a comparatively long time span, the

uncertainty of various system parameters has to be considered in the planning process.

Stochasticity of demand is considered by Gupta and Maranas (2003, 2004) as well as Ryu

et al. (2004); Al-Othmann et al. (2008). All these approaches do not consider details of

the logistical system and determine transport ﬂows between facilities only.41 Gupta and

Maranas (2003, 2004) rely on the tactical planning model proposed by McDonald and

Karimi (1997). In Gupta and Maranas (2003) this model is decomposed into the pro-

duction assignment problem and distribution planning problem. While the distribution

37

The inventory policy rules the generation of transport orders which are characterized by a random lead

time and pre-determined transport costs.

38

For a recent review of integrated production and distribution planning not restricted to the chemical

industry see Mula et al. (2010).

39

The integer variables reﬂect a discrete number of transfer price levels.

40

The deterministic version of this model can be found in Chen et al. (2003).

41

Transport ﬂows are associated with a ﬁxed transport cost rate.

135

planning depends on the realized demand, production planning is based on the estimated

demand. A two-stage stochastic linear program is created seeking for a conﬁguration

with minimum expected total costs. In Gupta and Maranas (2004) the objective of the

model is modiﬁed by considering future payments. For discounting future payments, two

approaches are discussed: The net present value (NPV) and the real-options-based value

(ROV).42 In Ryu et al. (2004) the hierarchy of production and distribution planning is

inverted, i.e. the distribution problem rules the production planning problem. Further-

more, both components are aﬀected by the uncertainty of demand. To solve the problem,

a parametric programming approach is described. Al-Othmann et al. (2008) present a

scenario-based sLP for the optimization of a petroleum SC consisting of companies from

the crude oil sector as well as the reﬁnery, petroleum, and basic chemical sector. The

objective is to ﬁnd a conﬁguration maximizing the expected proﬁt. A two-stage stochas-

tic linear program is proposed whereby decisions about the production quantities in the

crude oil sector constitute the ﬁrst stage problem43 whereas the decisions for the other

sectors are made based on a discrete number of demand scenarios.

A special part of the chemical industry is the oil-producing industry. Core production

units in this industry are reﬁneries. When reﬁnery operations are modelled, typically

non-linear models are used due to the non-linear production yield curves of distillation

units.44 Approaches for reﬁnery production-distribution planning are proposed by Kim

et al. (2008), Amaro and Barbosa-Póvoa (2008), Jackson and Grossmann (2003), Oh

and Karimi (2006), and Yi and Reklaitis (2007). Among these, only Kim et al. (2008)

and Amaro and Barbosa-Póvoa (2008) incorporate some logistical aspects by considering

multiple transport modes with varying transport cost rates. In Kim et al. (2008) the

production network of a Korean petroleum company is modelled. The model considers

multiple, non-speciﬁed modes of transport (with varying cost rates). Amaro and Barbosa-

Póvoa (2008) also deal with unspeciﬁc transport modes but the proposed model considers

batch production processes. Jackson and Grossmann (2003) use non-linear expressions

for modelling continuous production processes (similar to Kim et al. (2008)). To solve the

production-distribution problem, a decomposition approach is proposed by splitting the

model into the spatial and temporal dimension. A simpliﬁed production process model

is used by Oh and Karimi (2006) where the product fractions are handled as constants

and no reﬂux streams are considered. This work focuses on ﬁnancial impacts of duty

drawbacks for planning production and distribution in multi-national companies. An

LP is presented which seeks for the production-distribution plan maximizing the total

after-tax proﬁt. The inﬂuence of tax diﬀerences and exchange rates between diﬀerent

currencies is considered by Yi and Reklaitis (2007). Here, batch processes are assumed for

production and a periodic continuous-time model is formulated. An optimal conﬁguration

42

Both diﬀer in the interest rate used for discounting and the deﬁnition of the expected return. In case

of the NPV the expected rate of return is used whereby the ROV utilizes the risk-free rate of return.

43

I.e. the decisions are independent from demand ﬂuctuations by utilizing expectations.

44

See Li et al. (2005).

136

of the studied network is exhibited by deriving analytical expressions for production and

order lot-sizes.

For tactical production planning roughly assigning production quantities to production

sites suﬃces for deriving reliable mid-term plans.45 On the operational level, however,

a more detailed view on the involved production processes is needed. In particular, the

planning of batch processes is a challenging task as it involves

Planning of continuous processes faces a similar problem structure except that the batch

size is not limited by technological restrictions. Beside production planning, the planning

of logistical activities needs to be involved. Since (operational) production planning typi-

cally focuses on one production site, intra-site transports (which are typically based on the

pipeline mode) need to be considered but also inter-site transports or customer deliveries

are of importance in daily business in some cases.

Table 4.2 shows a literature classiﬁcation for integrated approaches primarily addressing

operational planning problems in chemical SC management.

All approaches summarized in this table contain a production planning and scheduling

component and are deterministic. The integrated production planning and scheduling

problem (IPPSP)49 is the core component of most approaches. The approaches diﬀer in

details regarding the production system and the incorporation of logistical aspects. The

reviewed approaches can be subdivided into approaches modelling batch and continuous

production.

Batch production is considered by the ﬁrst seven approaches shown in Table 4.2. Timpe

and Kallrath (2000) propose a time-discrete MILP for the IPPSP with multiple production

sites/plants, multiple production modes, intermediate storages, sales points, and trans-

ports between sites, storages, and sales points. The model uses diﬀerent time scales for

production and sales planning which allows a more precise planning of production. In a

production period, a site can produce in diﬀerent modes whereby changeover times and

costs are considered. To store intermediate and raw materials, one-product and multi-

product tanks are available and an assignment of products to tanks is implemented. The

45

Rough assignment refers to an assignment to aggregated production facility groups such as similar

plants or even complete sites.

46

This is due to the fact that typically batch processes have limited capacity, e.g. the volume of the

reactor is limited.

47

This is necessary in case of multi-plant networks which are typical for producing e.g. speciality chemi-

cals.

48

This might be necessary due to sequence-dependent set-up eﬀorts and due dates.

49

See Maravelias and Sung (2009), Li et al. (2010a) or Shao et al. (2009) for an introduction and literature

review. For an application of the ELSP in the pure batch scheduling context in chemical industry see

Cooke and Rohleder (2006).

type of

model- time transport uncer- solution

Reference SCM matrix objective produc-

ling scale mode tainty method

tion

Timpe and Kallrath (2000) (Otd , Opp , Ops ) M ILP sales ba di us — standard

Grunow et al. (2002) (Otd , Opp , Ops ) M ILP cost ba di us — decomp.

Neumann et al. (2002) (Opp , Ops ) M ILP other ba co — — heur.

Romero et al. (2003) (Opp , Ops ) M ILP revenue ba di — — standard

Berning et al. (2004) (Opp , Ops ) other proﬁt ba di — — GA

Guillén et al. (2007) (Otd , Opp , Ops ) M ILP equity ba di us — standard

Sung and Maravelias (2009) (Opp , Ops ) M IP cost ba di — — decomp.

tp

Bok et al. (2000) (O , Otd , Opp , Ops ) M ILP proﬁt co di us — decomp.

Pinto et al. (2000) (Otd , Opp ) M ILP proﬁt co si pi — standard

Neiro and Pinto (2004) (Otp , Opp , Ops ) M IN LP cost co di pi — standard

Erdirik-Dogan and Grossmann

(Opp , Ops ) M ILP proﬁt co co — — decomp.

(2006, 2008)

tp pp ps

Shah and Ierapetritou (2011) (O , O , O ) M ILP proﬁt co co pi — standard

DE +

own approach (Otp , Opp , C t , C p ) mu co di pi/ship/rail mu speciﬁc/GA

M ILP

abbr.: ba...batch; co...continuous; de...demand; di...discrete; mu...multiple; pi...pipeline; pr...price; ra...rail; ro...road; sh...ship;

si...single; us...unspeciﬁed

137

138

model is solved with standard solvers for multiple objectives. Maximization of sales is

ﬁnally recommended which best ﬁts to the intended application in practice.

Grunow et al. (2002) propose a sequential planning approach consisting of the pro-

duction planning step (called campaign planning) as well as the assignment and schedul-

ing step (called assignment model). The focus is on one production site with multiple

multi-purpose processing units. Transports between plants are mentioned but not ex-

plicitly modelled.50 The individual models are solved with standard solvers seeking for a

minimum-cost solution. A similar decomposition scheme is proposed by Neumann et al.

(2002). Here, the batch planning problem is formulated as a MILP whereas the schedul-

ing problem is solved by a heuristic considering limited plant capacities. This way, a

re-scheduling step is avoided and intermediate storages are introduced. A quite simi-

lar but integrated model for batch sizing and scheduling is proposed by Berning et al.

(2004). Due to the model’s complexity, a genetic algorithm is used to derive near-optimal

solutions. Sung and Maravelias (2009) propose an algorithm to determine the set of fea-

sible solutions for a production planning problem by evaluating infeasible polytopes of

the underlying scheduling problem. The resulting linear inequalities can be added to the

production planning model to ﬁnd an optimal and feasible production plan (w.r.t. to the

scheduling constraints).

In Romero et al. (2003) a simpliﬁed IPPSP for chemical batch production is described

which is enhanced by budget constraints. A one-stage batch production on parallel,

identical plants is considered. The goal is to maximize the total revenue over the planning

horizon without violating budget and resource constraints. Due to the simple production

model, the proposed MILP can be solved with standard solvers. This approach is extended

by Guillén et al. (2007) where multiple plants as well as transports between plants are

incorporated next to cash ﬂow and budget constraints. The objective is to maximize the

accumulated equity during the planning horizon, i.e. the diﬀerence between accumulated

cash and outstanding liabilities.

For continuous production planning and scheduling, there is no technological production

capacity limit. The number of production modes is typically smaller for continuous pro-

duction plants which eases the scheduling problem. Erdirik-Dogan and Grossmann (2006,

2008) consider a single production site with continuous single-stage production. Erdirik-

Dogan and Grossmann (2006) present the "basic" production planning and scheduling

problem for a single production unit. The production unit can be set up to multiple

products but only one product can be produced at a time. Deterministic demand ﬁg-

ures serve as lower bounds for the portfolio of chemicals to be produced. The model

is formulated in continuous time51 whereby a rough division of the planning horizon in

time periods is used. This allows the introduction of demand due dates and eases the

50

To overcome resource conﬂicts in the sequential procedure, a post-optimization step performs a re-

scheduling of the schedule provided by the assignment model.

51

I.e. production times are continuous variables.

139

calculation of inventory levels. The objective is to maximize total proﬁt over the planning

horizon by considering approximated inventory costs as well as changeover and process-

ing costs. To solve the proposed MILP, an iterated decomposition approach is described

by splitting the planning model into a master problem52 and a sub-problem.53 Here,

the master problem provides an upper bound for the original problem whereas the sub-

problem provides a lower bound. The iteration is stopped if the diﬀerence between both

bounds is suﬃciently small. Erdirik-Dogan and Grossmann (2008) extend this model by

considering multiple parallel production units. This model additionally requires to assign

production lots to the production units which complicates the problem considerably. A

similar iterated decomposition approach as used by Erdirik-Dogan and Grossmann (2006)

is presented to solve larger instances of this problem.

In practice, large-scaled chemical companies operate multiple diﬀerent plants at mul-

tiple production sites. Such a situation requires to model materials transports between

plants/sites and is considered by Bok et al. (2000), Neiro and Pinto (2004), Pinto et al.

(2000), and Shah and Ierapetritou (2011). Bok et al. (2000) present a discrete-time pro-

duction planning and scheduling model with multiple plants. Each plant can be set-up

to a speciﬁc production mode once at the beginning of each period. Transports between

sites aﬀect the provider’s and the receiver’s stock levels.54 Raw material transports are

associated with ﬁxed ordering costs. However, further logistical details such as turnover

capacities are not considered. The objective is to maximize the total proﬁt over the plan-

ning horizon. To solve larger problem instances, a two-stage decomposition procedure

is described which partitions the monolithic model into a master problem and a sub-

problem. The master problem determines a production and transport plan by ignoring

changeover calculations. In the sub-problem the production plan is determined given the

transport ﬂows from the solution of the master problem. Based on the solution of the

sub-problem, cuts are added to the master problem in the next iteration.

A related problem for the operational production and distribution planning of reﬁneries

is presented by Pinto et al. (2000). Here, a single-period planning model is formulated

which determines the blending and distribution of reﬁnery products via a network of

pipelines and tanks. Multiple distillation units are assumed to provide a continuous ﬂow

of multiple (raw) materials being transported by multi-product pipelines to tanks. Each

tank is designated to one product type only. Transition processes for switching from one

product to another are modelled.55 The model aims at determining a pumping schedule

maximizing the proﬁt. The resulting MILP is solved with standard solvers. Neiro and

Pinto (2004) extend this work by modelling petroleum supply and production planning

52

The master model determines production quantities and inventories per period.

53

The sub-model determines the precise schedule using the restricted set of products planned by the

master problem.

54

Customer transports are not considered because products are sold ex factory.

55

Each customer market is supplied via a multi-product pipeline from these tanks. The distribution

pipelines can be fed simultaneously from multiple tanks to allow a blending of the (raw) products to

ﬁnal products.

140

for multiple periods. Transition processes of multi-product pipelines are neglected. The

production is formulated as a generic transformation model that accounts for multiple

production modes. For each unit exists an intended production mode. Deviations from

the intended mode generate additional processing costs. Albeit the model is formulated as

a proﬁt maximization model, the demand is deterministic and no stock-outs are allowed

such that, in fact, the model is a cost minimization problem considering production,

transport, and storage costs.

Shah and Ierapetritou (2011) present an integrated planning model for reﬁnery pro-

duction based on distillation and blending units. Production planning is simpliﬁed by

assuming semi-continuous production modes.56 Changeovers of production modes are

penalized in the objective function.57 The proposed MILP is formulated in continuous

time and can be solved with standard solvers for realistically sized problem instances.

The objective is to maximize a proﬁt-like measure considering cost-inducing performance

measures such as the number of setups or the plants’ utilization rates. The weights for

the performance measures have to be interpreted rather as preference measures than as

cost rates. The setting of these weights signiﬁcantly aﬀects the model’s computational

complexity.

To sum up, the review of existing literature reveals that a lot of eﬀort has been spent in

developing integrated approaches to tackle deﬁciencies of sequential planning processes.

Prevalent in scientiﬁc literature are analytical optimization approaches allowing a high

degree of both portability and conciseness. On the tactical level, however, stochastic

components have to be considered to enable an accurate modelling of real systems. Op-

timization approaches, typically, handle only one source of uncertainty.58 But for an

encompassing model focusing on multiple system components, usually multiple sources

of uncertainty have to be considered. Simulation approaches oﬀer this capability and,

therefore, are prevalent in literature on integrated tactical problems (see Table 4.1).

The literature review also reveals that most simulation-based approaches lack a system-

atic procedure for deriving recommendations of improvements. Instead, most approaches

still bear the stigma of rather descriptive models where a (pre-)selected number of scenar-

ios is evaluated and compared. Potentials for improvement are not revealed directly this

way. However, a systematic procedure for post-analysis of simulation models can lead to

an exploration of the space of possible system conﬁgurations which allows deriving rec-

ommendations for improvements on a (more) reliable basis. This topic will be discussed

in section 4.3.4.

A further fact is that most integrated approaches do not consider transport processes

accurately, in particular at the operational level. Most approaches focus on production

planning decisions and handle logistical processes as auxiliary components. Exceptions

56

I.e. for the set of production tasks the product recipes can be varied in certain ranges only

57

Changeovers are also considered for multi-purpose tanks whereby the changeover implies a downgrade

of the interface of mixed products.

58

This is most often the demand, see Peidro et al. (2009) for more detailed information.

141

such as Pinto et al. (2000), Neiro and Pinto (2004), and Shah and Ierapetritou (2011)

consider reﬁnery operations where the logistical system diﬀers from basic chemical indus-

try in the sense that primarily the pipeline management and tank management constitute

critical logistical components. In basic chemical industry, however, also rail transports

constitute a relevant logistical aspect, in particular for inter-site transports. Also other

logistical details such as turnover capacity are often not modelled in detail.

These drawbacks are overcome by the simulation framework proposed in this thesis by

integrating MILPs for operational distribution planning considering transport-mode spe-

ciﬁc restrictions in a simulation environment. It aims at integrating both the production

and logistics system in basic chemical industry at the same level of aggregation. As the

proposed framework is designed to support tactical decisions, multiple sources of uncer-

tainty are handled. The characteristics of the proposed framework are brieﬂy summarized

in the last row of Table 4.2. The next section categorizes and discusses typical sources of

uncertainty relevant for integrated supply chain planning in chemical industry.

industry

Planning a network of integrated chemical production sites causes various problems. Due

to the interdependency of the production processes, local disturbances at a certain level

of a particular production site can aﬀect the site’s performance entirely or at least down-

stream (inter-site eﬀect). Moreover, due to the interconnectedness of production sites

in a chemical production network, such a local disturbance can spread out on the en-

tire SC (intra-site eﬀect). This typically leads to increased risk costs. The organization

and central control of chemical production networks aims at limiting the eﬀects of such

disturbances.

In basic chemical industry spill-over eﬀects are intensiﬁed by the inﬂexibility of contin-

uous production processes. Even small disturbances may lead to plant break-downs. In

general, continuously operated plants have to face high set-up or changeover costs.59 This

is because most chemical reactions need some running-in time before the process and pro-

duced products meet the required (quality) speciﬁcations. The products produced during

a changeover from one production mode to another are not marketable or require some

post-processing (so-called oﬀ-spec products). Additionally, after a complete shut-down

technical inspection and cleaning routines have to take place before the restart. Due to

the high costs of re-start and changeover procedures, plant shutdown and changeover costs

dominate most other costs like e.g. external supply or storage costs. In the worst case,

ﬁnal-product plants have to be shut down which causes a loss of demand if the asset is

operated at maximum capacity such that losses cannot be caught up leading to lost sales.

59

See e.g. Kallrath (2002) or Oh and Karimi (2004).

142

networks, it is diﬃcult to quantify the eﬀects arising from certain sources of uncertain-

ties. This is caused by the complexity of production processes, the interconnectedness

of the production network, and the limited capacities of commonly used logistical in-

frastructures. However, quantitative information is necessary to determine economically

reasonable risk management decisions.

To avoid shutdowns and to increase the local network’s robustness against disturbances,

buﬀers of raw and intermediate materials have to be at hand or at least immediately acces-

sible. Therefore, one of the most crucial tasks in SC planning is the management of stocks

and material ﬂows in the network.60 This is also necessary for perfectly balanced produc-

tion sites, i.e. when all local plants are coordinated in such a way that no intermediate

material is in deﬁcit or runs in surplus during normal operation. Stocks of intermediate

products are kept to buﬀer plant shortfalls or unintentional changeovers. However, most

integrated production sites are gradually grown and not planned in one step. Hence, in

normal operation, there typically exists a set of intermediate chemicals which are surplus

or in deﬁcit. Regular logistical activities are necessary to keep the production network

running properly.

Beside uncertainties occurring from the production system, other internal sources also

aﬀect or disturb the network’s performance. Basically, there are two dependent techno-

logical structures in chemical plants. Production sites cannot be operated without an

infrastructure supplying these plants with raw materials and energy. At the same time,

produced materials have to be transferred, stored, or further processed. Intra-site trans-

ports of intermediate and basic chemicals are usually handled via local pipeline systems

where dedicated pipelines for each chemical exist. Most raw and intermediate chemicals

are liquids or gases and are stored in tanks. This allows the decoupling of interconnected

plants to a certain extent by buﬀering material imbalances. Most ﬁnal and some intermedi-

ate chemicals are solid substances that are either directly transferred to customer-oriented

packagings or, in the bulk case, transferred into outbound storages such as silos.

Energy in chemical production sites comprises two main sources: electric power and

heat. Electric power is distributed via a cable network. Either a local power plant exists

or the production site is supplied by an external partner. At a production site often a

backup system exists to overcome short-falls of the primary source of electric power (at

least for some time). Hence, disturbances due to a power short-fall occur quite seldom at

chemical production sites.

Heat is often distributed as steam, which is centrally generated and distributed via

heat transfer pipelines at diﬀerent pressures. Especially the steam supply is a vulnerable

process in each integrated chemical production. Typically, the set of local plants can be

sub-divided into steam providers and consumers. Steam can result as a by-product (e.g. of

60

See Park et al. (2006) for a complex inventory management model for a single site.

143

is required by endothermic reactions or for distillation processes. I.e. if a steam-supplying

plant shuts down, additional external steam supply is required to properly supply steam-

consuming plants at a site. Otherwise, steam-consuming plants are forced to shut down

as well. This supply can either be organized by local backup boilers or by an external

producer such as a power plant. However, if the backup access to an external steam

supplier is not available, a backup boiler needs to be ready for steam supply in an instant

which constitutes an expensive option. The backup boiler requires permanent heating

and, thus, generates also remarkable continuous energy costs. Therefore, it is diﬃcult to

decide whether to risk a shutdown of steam-consuming plants or to maintain a backup

boiler permanently.

The local pipeline system of chemical production sites needs to be carefully supervised

because disturbances immediately aﬀect all associated plants. Possible disturbances are

manifold. However, pipeline systems are exceptionally reliable from a technological point

of view if they are carefully maintained with respect to corrosion and material weaknesses.

Problems may occur from operating a plant under inappropriate conditions. For example,

if pressure and/or temperature are not within the allowed ranges, chemical reactions may

take place inside pipelines. This may aﬀect the quality of the transported material and

can lead to congestions or damages in the pipeline.

Infrastructure also encompasses all facilities dedicated to material handling and trans-

ferring from external sources. This includes inter-site pipelines, roads, railways, or berths

as well as all facilities needed for turnover processes.

Beside technological problems, there may also arise disturbances due to managerial

mistakes. The management of integrated production sites encompasses e.g. decisions

about ﬁnancial means, the staﬀ, and stocks of all materials relevant for the production

processes. Financial disturbances include mis-estimations of prices and costs and may

lead to ineﬃcient production or sourcing decisions. Uncertainty of workforce occurs from

ﬂuctuations in personnel capacity.62 Disturbances of material supply can occur from two

sources: Either the local stocks are inappropriate and/or the procurement and distribution

processes do not match the realized requirements, e.g. orders are delivered too late or in

the wrong quantity/quality.

An important part of external uncertainty is caused by customer demands.63 However,

in basic chemical industry almost all ﬁnal chemicals are produced to stock and, thus,

are decoupled from direct market demands. Hence, variations of the customer demand

do not aﬀect the upstream production network immediately. Due to the exceptionally

high re-start and changeover costs for plants, demand variations are mostly reﬂected by

stock increases or sales price declines. The adaptation of production capacities by means

61

A prominent example are steam crackers (see Figure 2.8) where the quenching process provides high-

pressure steam, see Behr et al. (2010, p. 178).

62

E.g. due to strikes, illnesses etc.

63

See Applequist et al. (2000).

144

customer demand, e.g. in a recession. For these reasons, demand uncertainty is not in the

main focus of this work.

Another type of external uncertainty emerges from the suppliers’ side. Suppliers com-

prise all partners providing raw and intermediate materials which are not part of the con-

sidered chemical production network. From these external partners materials are supplied

to keep the production network working. Uncertainties aﬀect the network’s performance

when orders cannot be fulﬁlled in the expected way. E.g. shipments arrive too late or

not in the expected quantities. Such uncertainty can hardly be avoided, but has to be

reﬂected in safety stocks and strategic supply planning.

As a last type, external uncertainties may arise from governmental authorities or service

providers. Here, the potential realizations can be manifold depending on the speciﬁc

source. Changing governmental regulations may force a re-organization or re-conﬁguration

of the chemical SC, e.g. due to restrictive environmental regulations, duty drawbacks, or

changed tax rates.64 However, such events typically have a long-term character and,

thus, have no immediate eﬀect on SC operations. In contrast, service contracts support

SC operations on regular basis. A prominent example are logistical services provided

by external partners, so-called 3rd party logistics (3PL). Here, logistical processes are

outsourced and only supervised by the SC management. Uncertainty results e.g. from

unknown delivery dates or shortfall times (in case of outsourced maintenance activities).

These variations have to be considered in planning the SC operations.

In general terms, uncertainty refers to the unknown realization of at least one (random)

variable. In SCM, objects considered and planned have diﬀerent dimensions categorized

into quantitative, qualitative, and temporal measures.65 Examples of uncertain events can

be found for each category discussed above.66 Table 4.3 shows a classiﬁcation of sources

of uncertainty and examples w.r.t. quantity, quality, and time.67

Depending on the system to model, speciﬁc sources and dimensions of uncertainty are

more or less relevant. Typically, the quality dimension is less relevant in the SCM context

because its measures are diﬃcult to be incorporated in quantitative models. In contrast,

quantitative measures are crucial in planning SC operations. Temporal aspects become

especially relevant in the logistical context, e.g. when capacities of transport resources are

to be planned and estimations about turn-over times are necessary.

To model uncertainty, knowledge about the nature of the underlying stochastic process

is required. More precisely, assumptions have to be made on how this process is ruled,

which outcomes are possible,68 and what their corresponding probabilities are.69 If the set

64

See e.g. Vidal and Goetschalckx (2001) or Oh and Karimi (2004).

65

See e.g. Van der Vorst and Beulens (2002).

66

For a similar categorization see Klibi et al. (2010).

67

For related examples see also Applequist et al. (2000).

68

The so-called sample space.

69

The probability of an outcome may depend on other variables or decisions made in advance, see e.g. Law

145

internal external

technological

produc-

ﬁnan- person- materi- infra-

tion demand supply 3PLs

cial nel ally structure

system

work- demand transport

quan- credit stock pumping yield avail-

force varia- capaci-

tity line levels capacity rates ability

size tion ties

qua- quali- reli- oﬀ-spec product product reli-

interests decay

lity ﬁcation ability rate specs specs ability

evapo- pro-

due pumping due delivery transport

time speed ration cessing

date rates dates dates times

rate times

of outcomes and the process’ basic structure are known, information about the underlying

probabilities can be deduced.

For example, the output rate of a simple SISO reactor depends on various conditions.

To model the transformation of input to output, knowledge about the chemical reaction

and the chemical reactor can be used. E.g., a linear model might be used to describe this

relationship properly on an aggregated level (e.g. the hourly production rate). Neglecting

minor inﬂuences70 leads to a simpliﬁcation of the process model. Additionally, measure-

ment errors may hinder a perfect description of the process and lead to uncertainty in

the observed process measures.71 This uncertainty is expressed e.g. by a (normal) error

process. The resulting linear regression model can be veriﬁed using historical records of

the process. Often historical records allow analysts to deduce a proper stochastic model

of such a process. For more complex production processes more sophisticated stochastic

models (as described in section 2.3) can be necessary.

However, not all relevant stochastic processes are typically known at such a level of

detail. For external sources of uncertainty, typically, the underlying processes are not

known (entirely) or not describable in detail.72 Even the set of outcomes might not be

perfectly known. Modelling such processes is often based on historical records and rough

assumptions about the modelled process. E.g. the total customer demand faced by a

company or entire SC builds a conglomerate of a manifold of individual orders placed in

a given interval of time. However, neither models for each individual ordering process can

be derived nor an agglomerative model describing the ordering processes of all potential

customers could be handled. Hence, to model and estimate demands, one has to rely

on historical records about the total customer orders in history. A model to predict the

total demand to be faced in a period relies on indicators (such as GDP growth rate)

(2007).

70

E.g. the composition of input ﬂows or precise temperature and pressure within the reactor.

71

For a more detailed discussion and deﬁnition about uncertainty in the context of operations manage-

ment see Zimmermann (2000).

72

According to Zimmermann (2000) this refers to a lack of information causing the uncertainty.

146

and on common assumptions about the customers’ ordering processes.73 However, there

is no reliable information about the uncertainty inherent in the estimation procedure

since the individual ordering processes cannot be modelled adequately. Hence, no reliable

information about the degree of uncertainty is available.74 This phenomenon often occurs

in long-term planning problems, e.g. when an SC conﬁguration is to be found which ﬁts

best to a fuzzy set of scenarios.75

For the purpose of modelling SC operations on a detailed level, precise knowledge about

the stochastic processes to be modelled is required. For an existing SC the status quo

of operations is known and documented in most cases. Hence, internal processes can

be described and information about the processes’ stochasticity is (at least) partly avail-

able. This allows the proper modelling of internal uncertainty. E.g. chemical production

processes can be modelled using the methods described in section 2.3 and illustrated in

example 3. However, such a process model is insuﬃcient to model the production pro-

cess as a whole since only the output streams can be described with respect to the input

stream. To set up a proper model of the entire production process the stochasticity of

the input stream(s) has to be modelled as well. Typically, there exists a ﬁnite set of

operation states for a plant which are overlaid by stochastic noise. To model the states

of a plant, the Markov methodology introduced in subsection 3.2.2 can be used, enriched

by a stochastic model for the noise process. The following example extends the analysis

of the de-alkylation plant introduced in example 3 by analysing and modelling the inﬂow

process.

de-alkylation plant introduced in example 3. This chemical production process requires

streams of aromatic hydrocarbons (mainly Toluene and Xylene) and hydrogen to produce

Benzene (and a stream of long-chained hydrocarbons as a by-product). The limiting input

is the stream of aromatic hydrocarbons. Nonetheless, the supply of hydrogen has to be

assured, although it is typically not limiting. To assess the states of operation, the inﬂow

of the aromatic hydrocarbon stream is analysed. Because the set of general operation

states of this plant should be analysed, the median inﬂow rate of aromatic hydrocarbons is

recorded on a daily basis.76 Figure 4.3a shows the daily median inﬂow rates of aromatic

hydrocarbons recorded on 182 days and Figure 4.3b shows the corresponding histogram.77

It can be observed that two basic states of this plant are prevalent: Either the plant

produces at a level of about 141 units or it is not producing. However, there are some

ﬂuctuations primarily at the beginning of the recording interval. Investigating the rea-

73

See Tsiakis et al. (2001).

74

A prominent example is the occurrence of so-called rare events or hazards such as atomic accidents or

stock exchange crashes. See Taleb (2007) for a more detailed discussion.

75

See Klibi et al. (2010).

76

This is in contrast to the modelling of the short-term process model where more granular data is

required. See example 3 where hourly averages are used.

77

Note that the data used for estimating the detailed process model corresponds to the ﬁrst 14 observa-

tions depicted here in Figure 4.3a.

147

140

100

120

median inflow rate per day

80

100

frequency

80

60

60

40

40

20

20

0

0

Time (days) median inflow rate per day

(a) Time series of daily median inﬂow rates of aro- (b) Histogram of daily median inﬂow rates of aro-

matic hydrocarbons matic hydrocarbons

Figure 4.3: Daily median inﬂow rates of aromatic hydrocarbons for a de-alkylation plant

sons of these perturbations reveals that they are caused by technical re-adjustments due

to preceding repair and maintenance activities. Hence, two general states of operation

are considered as relevant, namely: full capacity production or a complete shutdown. Let

O = {0, 141} denote the set of these two states.

To model the state of the plant, a discrete Markov process is used. To calculate the

transition matrix Q of a discrete Markov process, the transition probabilities between both

states have to be estimated. All transitions of the recorded inﬂow data is used. The time

series of plant states ωt are calculated by

⎧

⎪

⎪ 141 ft > 0

ωt = ⎨ (4.1)

⎪

⎪ ft = 0

⎩ 0

where ft is the median inﬂow rate at day t depicted in Figure 4.3. The estimated transition

matrix Q̂ is then given by

0 141

0 ⎛ 0.875 0.125 ⎞

Q̂ = .

141 ⎝ 0.002 0.998 ⎠

The corresponding steady state vector is π = (0.087, 0.913). In other words, in the long

run, on 8.7% of days, the de-alkylation plant suﬀers from an (unintended) break-down.

The probability to face a break-down at a day which started in normal operation is 0.2%.

When the plant operates at full capacity, there are still variations in the realized inﬂow

rates. To model these ﬂuctuations, the time series of average hourly inﬂow rates is anal-

ysed in an appropriate interval where the plant operated at full capacity. As in Example

148

Figure 4.4 shows the time series of the recorded ﬂow rates.

144

average inflow rate per hour

143

142

141

140

Time (hours)

Figure 4.4: Hourly average inﬂow rate of aromatic hydrocarbons (including conﬁdence

intervals at a conﬁdence level of α = 0.1%)

In general, the inﬂow rates oscillate slightly around an average of 141 units. However,

some outliers occur which are probably caused by measurement errors rather than real vari-

ations of the ﬂow rate. To account for these observations, mean and standard deviation

are robustly estimated by calculating median and median absolute deviation from median

(MAD). This yields the following robust estimates: μ̂median = 141 and σ̂M AD = 0.18. As-

suming a normal distribution of the inﬂow rate, both estimates are used to calculate a

conﬁdence interval with a conﬁdence level of α = 0.1%. The corresponding quantiles are

displayed in Figure 4.4 by the dashed lines. Observations exceeding these intervals are

handled as outliers.

Having removed the outlying observations, the implied normal distribution is checked by

Shapiro-Wilk’s test. Figure 4.5a shows the QQ-plot and Figure 4.5b the autocorrelation

values for the trimmed time series of inﬂow rates.

The trimmed time series shows a good ﬁt to the normal model and no hints for auto-

correlation.

149

141.4

141.2

Sample quantiles

141.0

140.8

140.6

140.4

3 2 1 0 1 2 3

Theoretical quantiles

(a) QQ-Plot of trimmed inﬂow rates of aromatic hydrocarbons

1.0

0.8

0.6

ACF

0.4

0.2

0.0

0 5 10 15 20 25

Lag

(b) Autocorrelations of trimmed inﬂow rates of aromatic hydrocar-

bons

Figure 4.5: Diagnostic plots of trimmed inﬂow rates of aromatic hydrocarbons for a de-

alkylation plant

150

planning of supply chains in chemical industry

From the review of literature it can be concluded that simulation-based approaches estab-

lish an opportunity to tackle integrated planning problems when analytic optimization

is not able to properly reﬂect all aspects of the considered problem. However, ﬁnding

an optimal solution for an integrated planning problem by means of a simulation model

is more complicated compared to analytic optimization approaches. This is because the

development of a simulation model is a complex multi-step process. Figure 4.6 shows a

simple sketch of the steps in a simulation study.

simulation model

perform

experiments

meta-model

decision support

I.e. the relevant elements of the real system as well as their relations are represented at an

appropriate level of detail. It deﬁnes the general structure of the simulation model and

determines the requirements for data analysis.

In the data analysis step historical data of the real system to be modelled are analysed to

extract relevant parameters for the simulation model and the involved stochastic processes.

The simulation model is the implementation of the conceptual model accompanied by the

required parameters obtained from the data analysis step as some sort of a computer

program.

To derive recommendations for improving the system performance, simulation experi-

ments have to be conducted. I.e. control parameters of the simulation model are altered

in a structured way to investigate their eﬀects on the performance measures. A speciﬁc

setting of the control parameters is called a conﬁguration of the system.

As a simulation model typically contains stochastic elements, a simulation run is a

stochastic process. Hence, the outcome of simulation runs are stochastic variables. There-

fore, a certain number of simulation runs has to be conducted to handle the stochasticity

151

of the simulation outcomes. Keeping run times in acceptable ranges and deriving concise

information about the system performance at the same time, is a challenging task. Var-

ious approaches are provided in the literature to tackle this problem (often labelled as

simulation optimization).78

In the case of single-objective problems, the aim is to ﬁnd the optimal conﬁguration of

the system under study. Once found, this conﬁguration is the sole recommendation for

the management. The objective typically represents an expected performance measure.

However, since stochastic events may inﬂuence the performance measure, it is advisable

to incorporate measures of variability (like the variance of a performance measure) as a

further decision criterion. This immediately leads to a multi-objective problem.79

In the multi-objective case, a set of Pareto-optimal conﬁgurations has to be found (so-

called Pareto set). In many cases not all possible Pareto-optimal conﬁgurations can be

determined. Instead, a subset is returned by a multivariate simulation optimization ap-

proach. Based on this subset, analytical models can be ﬁtted to this data describing the

relationship between the control parameters and performance measures. These analytical

models are often lower-polynomial models and are called meta-models, response surface

models, or surrogate models.80 In other words, meta-models are mathematical functions

representing interdependencies of variables in a (simulated) system. Meta-models can be

used to derive Pareto-optimal conﬁgurations without the need to simulate these conﬁgu-

rations. Meta-models can help to ﬁnd the most suitable system conﬁguration in an easy

way (e.g. by weighting input and performance measures with costs and/or income rates).

In the remainder of this section the outlined planning process is described step-by-step.

Along these theoretical deductions, an exemplary simulation model of a chemical SC is

successively described, developed, and analysed in the example blocks.

Building a conceptual model is the ﬁrst step in a simulation project. The main idea is to

generate a blue-print of the structure of the system to be modelled. Based on the problem

deﬁnition, the scope of the study, and the real-world system, a concept is derived how to

model and solve the problem stated. The conceptual model builds the back-bone of the

subsequent simulation model. Figure 4.7 shows a prototypical development scheme for

simulation projects.

At least two parties are involved in the development of a simulation study. In SCM,

simulation projects are instruments for process optimization. Such projects are initiated

by the perception of process deﬁciencies and weaknesses. Overcoming these problems is

78

See e.g. Law (2007), Banks et al. (2005) or Kleijnen (2007).

79

In this context Pareto-optimal conﬁgurations are often called stochastically dominating, see Klibi et al.

(2010).

80

See e.g. Box and Draper (2007) or Kleijnen (2007).

81

Following Law (2007) and Banks et al. (2005).

152

project scope

modelling

programming

conceptual simulation

model veriﬁcation

model

the responsibility of SC managers. The perceived deﬁciencies deﬁne the aims and scope of

the project as well as the system components to be improved. Modellers have to propose

a concept for identifying the reasons why the perceived deﬁciencies occur. In a ﬁrst

phase this concept description is called a conceptual model or assumptions document.82

In this phase of the project all stakeholders commonly discuss and argue about the scope

and capabilities of the model to be developed. From the modeller’s point of view, the

conceptual model is a simpliﬁcation of the real-world system.83 Simpliﬁcation can be

achieved by reducing complexity or scope of the model:84

• complexity reduction

– subsume system components

– drop randomness

– approximate dependencies

• scope reduction

– drop system components

– drop variables

– restrict domains of variables.

Together, managers and modellers have to agree upon simpliﬁcations and implicit assump-

tions made, ﬁnally leading to the deﬁnition of a suitable conceptual model. As Figure 4.7

suggests, the development of simulation models is an iterative process, where the appro-

priateness of model speciﬁcations is checked in a structured way. Errors in implementing

the conceptual model in a computer program are checked by veriﬁcation.

82

See Law (2007) or Banks et al. (2005) for the latter term.

83

See Robinson (2006).

84

See Robinson (2006).

153

In the validation phase the behaviour of the simulation model is compared with the real

system’s behaviour. If errors are revealed, the modelling and programming steps have to

be re-started to ensure the required level of accuracy.

All sub-processes of the development process, beginning with the perception of deﬁcien-

cies, require the analysis of process data. The data characterizes the system’s performance

and behaviour. This is necessary to parametrize the simulation model and build a credible

basis for comparing the model’s behaviour and reality. A common database maintained

by both modellers and managers allows linking necessary and available information. A

lack of data may lead to uncertainty, more complexity, and a loss in accuracy. Uncertainty

is reﬂected e.g. by stochasticity at variables which in turn may lead to a larger number

of simulation experiments and more volatile performance measures.85 On the other hand,

the information collected in the project database needs to be relevant for the project.

Otherwise, a ﬂood of data to be evaluated and analysed hinders an eﬃcient development

process as it is time and resource consuming.

The analysis of available and relevant data typically requires a set of statistical methods.

They are applied to extract deeper knowledge about the processes to be modelled. The

set of methods is vast and the choice of methods depends on the needs of the speciﬁc

project. Time series methodology is one prominent branch of methods.

The following example describes the conceptual model of an artiﬁcial chemical SC

analysed in this section.

framework, in this example an exemplary chemical SC basically consisting of a core part

of two chemical production sites is described. The example is inspired by the production

network of Dow Chemical in Europe. The underlying real-world processes are not described

in detail here.86 Instead, the resulting conceptual model after analysing the real world

system is presented. All assumptions and models used are described in the following.

Figure 4.8 visualizes the production network.

Both production sites produce basic and intermediate chemicals based on Naphtha. How-

ever, both sites focus on diﬀerent ﬁnal products. They also vary in their plants’ capacities.

The Naphtha is steam-cracked at both sites with similar plant conﬁgurations. Hence, the

ratios of output ﬂows are similar but the heights of the ﬂow rates are not.

Grey-coloured chemicals and associated arrows indicate auxiliary chemicals which are

not further processed at either site. As common in steam-cracking, the fraction of C1

hydrocarbons produced as a by-product is immediately used for heating the cracker fur-

naces. Similarly, the fraction of long-chained hydrocarbons (C9+ ) is used as fuel for steam

production at both sites.87 Hydrogen (H2 ) is produced and consumed by various plants.

85

See Zimmermann (2000).

86

This is partly due to privacy protection reasons and partly because the proposed example does not

match the real SC entirely.

87

Additionally, natural gas is admixed if the local fuel production is insuﬃcient. In general, cracking is

154

site 1 PE C1

H2

C2

Benz. Styr. PS C2

C5+

C9+

H2

C3

C1 Cum. C3

C4

Naphtha

C4

S1 rail transports

Naphtha site 2 PE C5

H2

C2

Benz. Styr. C6

C5+

C9+ H2

C3

C1 PP C7

C4

Buta. SBR C8

Raf.

supplier storage customer production plant

Imbalances are compensated either by heating with H2 or by external supply from producers

of industrial gases nearby both sites.88

Steam crackers produce a fraction of C4 hydrocarbons which is a mixture of diﬀerently

structured chemicals all containing four carbon atoms.89 In this mixture, 1,3-Butadiene

is the largest sub-fraction with the highest applicability in downstream processes. There-

fore, this sub-fraction is extracted by rectiﬁcation.90 The remaining sub-fractions of C4

hydrocarbons (so-called raﬃnate) is sold to the market ex factory.

At both sites imbalances of the processed intermediate chemicals occur when all plants

produce at full capacity. Unbalanced intermediate chemicals are indicated by a coloured

storage symbol when inter-site transports are possible. Correspondingly coloured arrows

between both sites indicate the direction of the product ﬂow. Table 4.4 quantiﬁes these

imbalances and provides an overview on the chemicals’ associated storage capacities.

A net demand (indicated by negative values in the balance columns in Table 4.4) of

intermediate chemicals has to be satisﬁed by product ﬂows between both sites or from

a highly endothermic reaction which in turn leads to a net demand of natural gas.

88

Note that this is a typical situation since most chemical production sites have demand for industrial

gases not only for chemical reactions but also e.g. for cooling (N2 ).

89

See Sun and Wristers (2006).

90

Since all C4 -sub-fractions possess similar boiling points, extractive distillation has to be applied, i.e. a

solvent chemical is used to drive the boiling points apart. See Sun and Wristers (2006).

155

chemical site 1 site 2 site 1 site 2

Naphtha -3000 -2000 30,000 25,000

Ethylene (C2 ) -200 200 20,000 20,000

Propylene (C3 ) 250 -300 10,000 8,000

Crude C4 hydrocarbons (C4 ) 300 -400 4,000 3,500

Pygas (C5+ ) 0 0 15,000 10,000

Pyoil (C9+ ) 60 40 2,500 2,000

1,3-Butadiene (Butadiene) — 0 — 2,500

C4 hydrocarbons (Raﬃnate) — 360 — 5,000

Benzene 0 0 25,000 20,000

Styrene -400 400 7,500 10,000

Cumene 585 — 25,000 —

Polyethylene (PE) 1,200 380 55,000 10,000

Polypropylene (PP) — 600 — 12,500

Polystyrene (PS) 600 — 17,500 —

Styrene-Butadiene rubber (SBR) — 320 — 18,000

Table 4.4: Material balances (in t/hour) and storage capacities (in t) per site and chemical

precise overview of the product ﬂow rates per site and plant is displayed in Table A.5 in

the appendix. In addition, the coeﬃcients of the time series models generating the product

ﬂows in detail are provided in Tables A.6 - A.11 in the appendix.

Sources of disturbances considered in this example are categorized in three classes. First,

the production plants are stochastic transformers, i.e. the transformation processes are

modelled by stationary time series models with normally distributed errors. The plants’

states are modelled by Markov models as introduced before. The corresponding transition

matrices are provided in the appendix in Table A.15 and Table A.16. Additionally, nor-

mally distributed errors are added to simulate the inﬂow rates with ∼ N (0, 10

ω

) where ω

is the current state of the plant.91

From Table 4.4 it can be taken that inter-site transports of Ethylene, Propylene, C4 hy-

drocarbons, and Styrene suﬃce to buﬀer local imbalances and increase plant utilizations.92

Therefore, rail transports are considered between both sites. For each product a centrally

managed ﬂeet of RTCs is available for transports. Each RTC provides a transport capacity

of 50 tons.

For inter-site transports a regular train shuttle is organized. A train can be chartered

departing in accordance with a predeﬁned schedule. The train is operated by a rail ser-

vice provider. If a train is chartered, the dedicated RTCs are gathered at the departing

site’s shunting yard where they are fetched by the rail service provider at the scheduled

91

In the case of of plants with multiple inputs, for each input stream its variation is calculated based on

its current state.

92

This is because reciprocal imbalances exist at both sites for all these chemicals. E.g. Ethylene is short

at site 1 but surplus at site 2. Hence, transports of Ethylene from site 2 to site 1 are reasonable.

156

time. After an expected travelling time of approximately 24 hours the train arrives at the

arrival site’s shunting yard where the attached RTCs are decoupled and further processed

(i.e. loaded or unloaded). The train’s travelling time is a normally distributed random

variable with N(24, 1). Loaded RTCs become available for transport the day after un-

loading. During this time, turnover processes as well as security checks and shunting

operations take place.

Unloading and loading times (in hours) are normally distributed random variables with

N(2, 0.25). The total turnover capacity depends on the number of transfer arms available

and the number of operating hours. Turnover operations take place 10 hours per day from

Monday to Saturday. Due to strikes, technical failures, etc. operating hours may reduce

to 5 or 0 hours. A discrete Markov model is used to model these disturbances constituted

by the following transition matrix:

10 5 0

10 ⎛ 0.99 0.005 0.005 ⎞

⎜ ⎟

Qturn−over = 5 ⎜⎜ 0.15 0.80 0.05 ⎟⎟

.

0 ⎝ 0.20 0.05 0.75 ⎠

Since both sites are geographically separated and do not share a common workforce, the

Markov model is applied for both sites independently.

The number of transfer arms and corresponding turnover capacities can be found in

Table A.12 in the appendix.93

To dynamically generate plans for train departures and RTC ﬂows on the operational

level, the MC-RTP is used. The interval of time for discretization is a 24-hour rhythm.

The target stocks for the corresponding chemicals are calculated such that the maximal

demand of a chemical can be supplied for approximately 10 days from stock. The target

and initial stock levels as well as the number of initially available empty RTCs can be

found in Table A.12. Cost rates for the MC-RTP instances are shown in Table A.13 in

the appendix. Note that the costs for additional storage in RTCs are given as cAdd = 1.94

For Styrene, C4 hydrocarbons, Naphtha, and the ﬁnal chemicals the network is imbal-

anced which requires exports and imports to and from external partners. Naphtha is sup-

plied via pipeline which connects the sites with the exclusive supplier (a reﬁnery nearby).

The supply is organized in batches injected by the supplier when an order is placed by a

production site.

Batch injection and transport is associated with costs for pipeline operation (denoted by

cbatch ). Stock holding at the sites is associated with a stock holding cost rate chold . The

ﬁxed pipeline transport capacity ρ is 7,500 tons per day (which exceeds both consumption

rates ωi ), optimal batch sizes q opt and order intervals topt can be determined by solving the

93

The turnover capacities are symmetric, i.e. loading capacity at the sending site equals unloading ca-

pacity at the receiving site. In an asymmetric setting both capacities do not match.

94

This allows a virtually cost-neutral storage extension.

157

1

2 2 ⋅ cbatch ⋅ ωi

qiopt = 2

2

3 hold and (4.2)

c ⋅ (1 − ωρi )

1

2 2 ⋅ cbatch

ti = 2

opt

2

3 hold . (4.3)

c ⋅ ωi ⋅ (1 − ωρi )

To calculate the optimal order quantities qiopt and corresponding order intervals topt

i requires

information about holding and batch injection costs, chold and cbatch .

Holding costs chold typically consist of the costs for maintaining storage facilities and

committed capital. Maintenance costs of storage facilities mainly encompass labour costs

for the maintenance staﬀ and the facilities’ depreciations. Both are independent from

the stock levels and, hence, do not aﬀect the stock level decision. In contrast, costs for

committed capital are typically assumed to grow linearly with the stock level as they reﬂect

the opportunity costs for investments which cannot be realized. With a market price for

Naphtha of about 700 e per ton and a return on assets of about 5% in basic chemical

industry, the storage cost rate can be approximated by chold = 35 e per ton and year.96

Batch injection costs are comparatively hard to assess as they depend on the techni-

cal speciﬁcation of the pipeline. Pipeline operations primarily induce costs for energy

consumption. A pipeline transport generates energy costs which are proportional to the

transport volume. For modern basic chemical pipelines an energy consumption of about

42 kWh per ton and kilometre is estimated.97 Depending on the energy supply contract,

the costs of a batch transport can be calculated. However, with a contract charging energy

costs proportional to the energy consumption, batch injection and transport costs grow

linear with increasing batch size. Hence, the ﬁxed cost part for pipeline operations cbatch

only consists of operational costs for preparing pipeline operation which are comparatively

small. Hence, it is concluded that qiopt → ωi which implies a continuous operation of the

pipeline at the level of the consumption rate. This result is reasonable as long as no con-

siderable ﬁxed costs for providing a Naphtha batch occur.98 In this situation, both sites are

supplied continuously, i.e. the pipelines’ feed rate is 5,000 tons a day with both crackers

operating at full capacity. The ﬂow rate is reduced accordingly when one or both of the

95

Note that this implies that both pipelines can be operated independently, see e.g. Günther and Tem-

pelmeier (2004).

96

Prices for Naphtha are partially transparent on the spot markets and can be found e.g. in Kellermann

(2012). However, most quantities are traded based on long-term contracts. Hence, the spot mar-

ket prices are potentially slightly over-estimated. The return on assets is a key ﬁgure reported in

companies’ economic reports, see e.g. BASF (2012) for one speciﬁc example or Fortune (2006) for an

industry-wide ﬁgure.

97

See van Essen et al. (2003).

98

Note that typically reﬁneries are operating continuously where Naphtha results as a direct by-product.

Hence, production-speciﬁc set-up costs due to a variation of the production process are not to be

expected.

158

crackers are shut-down. The safety stocks to be held at both sites are deﬁned such that for

5 days the crackers can be supplied from stock. I.e. safety stock levels are set to 15, 000

and 10, 000 tons for site 1 and 2, respectively. Details on the supply system will be studied

in example 12.

Final chemicals (such as Polystyrene, Polyethylene, etc.) are sold to customers ex

factory (i.e. customer transports are not considered). Styrene and C4 hydrocarbons are

transported via RTCs. To plan these transports the MC-RTP instance is extended by an

artiﬁcial sink and an artiﬁcial source representing customers and suppliers of Styrene and

C4 hydrocarbons, respectively. These ﬁctional nodes are always able to absorb or provide

a required quantity of chemicals. They represent the spot market for both products where

orders are placed and a certain company accepts (at a certain price). Transports from/to

this source/sink represent transports to ex ante unknown partners and, hence, transport

times are stochastic. Here, a normally distributed transport time is assumed with N(5, 1)

(given in days).

Logistical disturbances are incorporated reﬂecting RTC defects and variations in the

customers’ pick-up quantities and frequencies (see Table A.14 in the appendix). RTC

defects occur with probability p = 0.025 per day and RTC. I.e. the number of defect RTCs

per day is binomial distributed with sritsf ail

∼ Bin (srits , p) where srits is the number of

RTCs of chemical s checked at site i in period t.99 The time for repair is given in days

and normally distributed with N (7, 1).

Auxiliary processes may fail e.g. steam or electric power supply breaks down which

causes plant breakdowns or a site-wide breakdown. For simplicity only two events are

modelled here: First, a breakdown of the steam network which forces all plants with energy-

consuming reactions to shut-down. This set includes the cracker, Butadiene extraction,

Benzene production (hydrogenation plant) as well as Cumene and Styrene production.100

To model such an event, a Markov approach is used with the following transition matrix

where state ω = 1 indicates normal operations and ω = 0 indicates a breakdown:

1 0

1 ⎛ 0.995 0.005 ⎞

Qsteam = .

0 ⎝ 0.35 0.65 ⎠

The second auxiliary process aﬀecting the production network is a breakdown of the hy-

drogen network which immediately aﬀects all plants consuming or producing hydrogen.

This includes the cracker, Benzene production, and Styrene production. The following

99

Note that any RTC is inspected before it leaves a site or is loaded. Hence, this encompasses all RTCs

E

waiting for further processing (srits ).

100

The alkylation reaction taking place in Cumene and Styrene production is exothermic (see Vora et al.

(2006)) which induces a production of steam. The same holds for the cracker where the quenching

process produces steam. Hence, these plants are connected to the site’s steam network and aﬀected

when steam distribution cannot be managed.

159

1 0

1 ⎛ 0.99 0.01 ⎞

Qhydro = .

0 ⎝ 0.50 0.50 ⎠

Note that both auxiliary processes are assumed to be highly reliable and failures can be

bypassed or repaired in comparatively short time.

In the next step, all components and (statistical) models described in this example are

implemented in an R program. The implementation is described in the next example.

The components necessary to model a chemical SC depend on the type of simulation

model chosen. Basically, the simulation of SCs is focused on material ﬂows among a set of

processors whereby also ﬂows of money or information might be reasonable.101 Processors

transform and manage the ﬂow of materials depending on speciﬁc rules and stimuli from

other model components.

In agent-based simulation models, processors are agents which are interrelated and

directly interact depending on their internal decision rules, stimuli from related agents,

and environmental stimuli. Beside material ﬂows, ﬂows of information are often used

modelling the transport of stimuli between agents. Incorporating asymmetric information

is a general advantage of this simulation technique such that this model type is particularly

applicable for systems with a decentralized management.

In system dynamics simulation, processors reﬂect stocks. Flows between stocks are

continuous in time. Processors are controlled by setting ﬂow rates which are managed

by control cycles as exemplarily depicted in Figure 2.9. I.e. ﬂow rates are adjusted

w.r.t. variations of the realized ﬂow rates from their corresponding target levels.

Discrete-event simulation is probably the most ﬂexible simulation technique. Flows are

here discretely modelled, i.e. at a discrete number of time indices ﬂows are controlled.

Processors can represent various system elements depending on the type of ﬂows con-

trolled and the processors’ internal decision rules. Processors of material ﬂows represent

e.g. production facilities where ﬂows are transformed and (possibly) delayed to reﬂect

the time consumption for production. Moreover, processors may also reﬂect transport or

storage facilities where ﬂows are controlled depending on the available capacities.

Typically, the management of ﬂows is implicitly determined by local rules of the pro-

cessors. To improve the system’s performance, however, it is sometimes advisable to

coordinate these local decisions, i.e. the processors are centrally managed. This implies

that the simulation model requires an additional processor reﬂecting central information

101

See Kleijnen (2005).

160

as optimization programs) such that a hybrid simulation model is formed.

For modelling chemical SCs using discrete-event simulation, the following types of pro-

cessors are included to reﬂect material ﬂows:

• production facilities,

– storage (tanks, silos, etc.),

– turnover (transfer stations, quays, etc.),

– transport (rail yards, pipelines, ships, etc.),

– energy supply (electric power, steam, etc.),

– auxiliary product supply (water, industrial gases, etc.),

– workforce supply.

ﬂows according to local restrictions. Processors can be characterized according to general

attributes about the handling of in- and outgoing material ﬂows. Typical categories are

the vergence and volume of ﬂows as well as the handling of time. Table 4.5 displays

examples of production and logistical processors w.r.t. these categories.

processors storage turnover transport

multiple simultaneous multi-sink

ﬂow split separation

consumers loading pipelines

ver- multiple simultaneous multi-source

gence merge synthesis

sources unloading pipelines

route transformation one-to-one one-at-a-time ship/rail/truck

batch

delay ✓ ✓ ship/rail

production

time non- continuous

— — pipeline

delay production

conser-

ﬂow regular regular regular regular

ving

quan-

redu- oﬀ-spec product

tity decay rest in RTC

cing product interface

Table 4.5: Examples for processors according to processor type and processing attributes

Processors may merge, split, or route ingoing ﬂows. For production processors this

corresponds to the general classiﬁcation of chemical production processes (see Table 2.1).

161

For logistical processors these attributes correspond to the technical details of the mod-

elled equipment. Splitting storages can be used to supply multiple consuming processors

at the same time. Similarly, splitting turnover processors can handle multiple carriers at

the same time or may feed multiple storages at the same time. Transport processors that

are able to split ﬂows simultaneously are e.g. one-to-many pipeline. Analogous examples

are given for merging and routing processors.

Depending on the real system’s speciﬁcations, a crucial attribute of processors is their

handling of time deciding whether ingoing ﬂows are delayed or not. An example of

delaying production processors are batch processes where typically some time elapses until

the chemical transformation process is ﬁnished. Storage and turnover processors are also

delaying whereby in the former case the delay is an intended feature and ﬂexible. In the

latter case, the delay occurs due to the processing times for turnover preparation (e.g. for

RTC shunting and connection to the transfer system). The time delay for transport

processors accounts for the transport time which is typical for e.g. rail and ship transports.

In contrast, pipeline transport can be non-delayed if the pipeline is completely ﬁlled once

and continuously operated under constant conditions.

Another characteristic is whether a processor conserves the inﬂowing quantity or not.

If non-usable or invaluable materials are produced during a transformation process, these

can be dropped from the simulation model for the sake of brevity such that a total loss

of material is modelled. Examples of such losses are oﬀ-spec materials102 produced after

a mode changeover in a multi-purpose plant and interfaces in multi-product pipeline

operation.

The kind of modelling of auxiliary processes depends on their structure and importance

for the entire system. Often auxiliary processes are modelled on an aggregated level as

(global) environmental variables aﬀecting the attributes (such as capacities) of processors.

Such processes are often not modelled explicitly by means of processors or ﬂows but rather

as (environmental) events which are ruled by stochastic processes.

Similar to SC planning problems, the elements of simulation models (i.e. ﬂows, pro-

cessors, and environmental variables) can be structured according to the level of detail

involved. Based on the basic structure of a chemical SC as depicted in Figure 4.10, three

aggregation levels are distinguished:

• plant level

• intra-site level

• inter-site level.

Table 4.6 shows a brief overview on the related model components and aggregation levels.

At the plant level one or more production processors form a material transformation

process. Chemical production plants can be modelled as processors with technologically

102

See page 141 for the deﬁnition.

162

logistic

production auxiliary

storage turnover transport

plant ✓ — — — (✓)

intra-site — ✓ ✓ ✓ ✓

inter-site — (✓) — (✓) —

ﬁxed rules of operation (e.g. a ﬁxed material ﬂow inside the plant). Here, static models

are used to represent the transformation process. Considered stochastic elements are

e.g. yield rates, general production mode, and processing times. Operating rules at the

plant level are e.g. priority rules for material processing (such as ﬁrst-in-ﬁrst-serve).

The intra-site level is characterized by a set of multiple plants which are interconnected

by logistical processors (such as pipelines and tanks). Logistical processors enable inter-

mediate storage and a connection to external sources of materials. Examples for local

processor rules are inventory policies for each storage processor. At this level, a coordina-

tion of production processors addresses e.g. production planning and scheduling problems

for multi-product batch processes. In this case, the central control of production proces-

sors tackles an operational planning problem. A similar intra-site coordination problem

occurs in the management of local stocks within a multi-product pipeline system.103 Such

coordination problems can either be solved by applying heuristics locally or by an em-

bedded analytical optimization model.104 Considered stochastic inﬂuences in modelling

logistical processors reﬂect e.g. variations in their capacity due to technical failures or the

processing times of handled objects.

Coordination problems also occur on the inter-site level where multiple integrated pro-

duction sites are interconnected by (material) ﬂows. At the inter-site level, only logistical

processors are added to the model if any at all.105 Here, the main focus is on the man-

agement of inter-site transports by utilizing the logistical facilities located at the sites.

To utilize these facilities, transport carriers have to be modelled, i.e. trucks, rail cars, or

ships. The ﬂow of these entities forms the capacity for inter-site material ﬂows. Hence,

at the inter-site level primarily ﬂows (e.g. of materials, transport carriers, or information)

have to be planned for which operational distribution planning and routing models can

be used. Stochastic inﬂuences often aﬀect the transport times but also the quality of

materials which may decay during transport.

If the modelled real system relies on discrete objects, a discrete-event simulation is

the natural choice for simulation. However, in chemical industry production processes

are continuously operated. Hence, material ﬂows are usually continuous with variable

103

See Neiro and Pinto (2004).

104

See Garcia-Flores and Wang (2002), Jung et al. (2004) or Jung et al. (2008) for examples of the latter

case.

105

An example are inter-site pipelines.

163

ﬂow rates. In contrast, the associated logistical sub-systems (e.g. for rail and ship trans-

ports) rely on discrete entities which requires discrete-event simulation. This implies that

continuous processes have to be simulated for discrete time intervals. Hence, the ideal

simulation model is a continuous time discrete-event model which integrates continuous

chemical process models. If operational planning models are included in the simulation

model, this forces a continuous time formulation of these optimization models. However,

typical operational planning models are complex and time continuous formulations are

often not adequate to solve such problems eﬃciently. Often a time-discrete formulation

is preferred to keep the problems solvable for instances of practical size.106 Therefore,

time-discrete, discrete-event formulations are often appropriate for hybrid simulation ap-

proaches.107 Fortunately, continuous chemical production processes can be approximated

by discrete time series models quite accurately as long as the time increment is chosen to

be not too large, see chapter 2.

The following example describes the implementation of the conceptual model described

in example 9.

tual model of the chemical SC described in example 9 is implemented as a hybrid time-

discrete discrete-event simulation model. It is implemented as an R program. The model

uses two discrete time scales. The basic time scale is given in days whereby each day is

subdivided into 24 hours. The conceptual model’s graphical representation from Figure 4.8

can be extended using the processor notation introduced in Table 4.5. Figure 4.9 shows a

graphical representation of the implemented simulation model for site 1.108

Flows of each production processor are generated in each period depending on the stock

levels of the adjacent stock processors and speciﬁc time series models.109 Flows between

production processors use unidirectional single-product pipelines. As non-delayed transport

processors, these pipelines are omitted from the ﬂow chart of Figure 4.9 for the sake of

simplicity. Transport and turnover processors handle RTC and material ﬂows. Both

depend on each other and have to be planned simultaneously.

To generate plans for inter-site rail transport and turnover processes, the MC-RTP model

is used in a rolling horizon environment where the individual MC-RTP instances are solved

by CPLEX 12.3.110 The planning horizon is set to seven days.111 The re-planning interval

is set to three days such that ﬂow decisions are ﬁxed for three days.112 The general struc-

106

See e.g. Pinto et al. (2000).

107

See Garcia-Flores and Wang (2002), Jung et al. (2004) or Jung et al. (2008) for examples.

108

Site 2 and information ﬂows are omitted for the sake of brevity.

109

Details are described in example 9 and Table A.6-Table A.11. Flows from sources and sinks (repre-

senting suppliers and customers) are generated as described in example 9 and Table A.14.

110

See IBM (2010).

111

For the conﬁgurations used in the simulation study, all MC-RTP instances are solved to optimality in

less than two seconds.

112

To determine the re-planning interval, simulation experiments with varying re-planning intervals are

conducted. In tendency, it holds that the shorter the replanning interval is, the better the system

164

C2 PE

Benz. Styr. PS

C5+

C3

Cum.

C4

storage processor

production processor turnover processor

storage processor for empty RTCs

ﬁctional sink/source transport processor

storage processor for loaded RTCs

Figure 4.9: Flow chart of a part of the exemplary chemical SC using the notation of

Table 4.5

ture of the implemented simulation model is displayed in Figure 4.10. Numbers annotated

to the blocks refer to the lines in the pseudo-code provided in Table 4.7.

In each iteration (i.e. for each day), it is initially checked whether a new transport

plan is to be determined. If this is the case, initial data is gathered reﬂecting the current

state of the SC. This comprises the stocks of chemicals and RTCs at the sites (loaded or

empty) as well as the currently ﬂowing chemicals and RTCs. Additionally, the expected

production rates are estimated by linearly extrapolating the last week’s (hourly) production

rates. The capacities of the storage facilities as well as expected values for turnover and

transport times are constants and, hence, ﬁxed parameters for each MC-RTP instance.

The turnover and train capacities follow regular train and workforce schedules which are

assumed to be deterministic and, hence, known for all instances. The cost rates ruling the

MC-RTP are ﬁxed for all instances (see Table A.13).

After solving an MC-RTP instance, the resulting optimal ﬂows of materials and RTCs

are passed to the simulation model. Whether these ﬂows can be realized depends on the

availability of resources which are aﬀected by stochastic elements. Hence, scheduled ﬂows

can be realized only if suﬃcient

• RTCs,

• turnover capacities,

performance and the longer the computation time is. An interval length of three days yields a

reasonable compromise between run time and solution quality.

165

set-up matrices

2-10

initialize inidicators

t=1

12

is t re-plan- yes generate

13

ning period? MC-RTP-instance

no

solve 14

instance

15-16

random events 19-28 RTC ﬂows

RTC ﬂows

no

end of

horizon?

yes

evaluate

performance

are available. Breakdowns of turnover capacities are ruled by a Markov approach and

immediately aﬀect all ﬂows. A loss of empty RTCs due to unexpected maintenance activ-

ities can aﬀect scheduled ﬂows if the remaining number of empty RTCs is insuﬃcient to

transport the scheduled ﬂow. Moreover, dispatched trains may arrive later than estimated

such that scheduled turnover processes are delayed. A breakdown of production plants may

induce e.g. stock-outs such that planned chemical ﬂows cannot be realized.

Anyway, when a scheduled ﬂow cannot be realized as intended, it is postponed to the

next period. Hence, stochastic inﬂuences typically lead to a postponement of transport

activities. In the next re-planning period all ﬂow schedules are re-set to the new optimal

solution such that formerly postponed ﬂows are integrated in the new schedule. Table 4.7

shows the detailed pseudo-code of the implemented simulation model.

Conﬁguring the SC can be done by a set of parameters of the simulation model reﬂecting

the adaptation of capacities (e.g. RTC ﬂeet sizes, inventory, or turnover capacities) and

of operational rules (e.g. target stocks and cost rates for the MC-RTP). Note that the cost

rates ruling the MC-RTP instances reﬂect the operational costs for transports, turnover,

166

initial rail cars srini , cost rates, inventory capacities & target stocks, turnover

capacities, train & handling capacities)

2 generate result & auxiliary matrices ;

3 initialize inventory & RTC stock matrices ;

4 for all sites i do

5 generate state sequence vector of length T for hydrogen (ωithydro ), steam

(ωitsteam ) and transfer system (ωitturn ) ;

6 for all plants j do

7 generate state sequence vector ωitj ;

8 assign plant states w.r.t. ωitj , ωithydro and ωitsteam ;

9 end

10 end

11 for t = 1 → T do

12 if t is planning period then

13 generate MC-RTP instance for next 7 days ;

14 solve MC-RTP instance for next 7 days ;

15 generate transport times for scheduled trains ;

16 update ﬂow schedules according to new solution ;

17 end

18 for all sites i do

19 for all RTC ﬂeets do

20 determine loss of empty RTCs for maintenance ;

21 assign maintenance durations for each RTC under repair ;

22 receive repaired RTCs ;

23 update empty RTC stock ;

24 end

25 for τ = 1 → 24 do

26 for all plants j do

27 generate hourly product ﬂows depending on ωitj (w.r.t. to

available stocks & inventory capacities) ;

28 end

29 for all products p do

30 realize scheduled transport & transfer ﬂows (w.r.t. to available

RTCs & turnover capacities) ;

31 update stocks sit ;

32 end

33 end

34 end

35 end

36 end

167

and stock holding. They can be seen as policy parameters as they control the level of

preference for stock holding or transportation. In particular, the setting of inventory cost

rates for over- and undershooting determines the smoothness of stocks and can be seen as

an alternative for the variation of target stock levels.

To evaluate the performance of a simulation run, the ﬂow and stock matrices are re-

turned by the model. These matrices can be used to calculate a desired number of perfor-

mance measures (e.g. the number of dispatched trains, average stock levels, etc.). These

depend on the purpose of the study. It has to be noted that each simulation run is a stochas-

tic experiment. Hence, multiple replications are necessary to evaluate the performance of

a certain SC conﬁguration properly.

Veriﬁcation and validation are instruments to ensure the quality of modelling. Veriﬁca-

tion particularly focuses on the technical aspect ensuring that the programmed model

behaves as intended.113 Validation focuses on the accuracy of the ﬁnal simulation model

which should reﬂect the behaviour of the corresponding real-world system suﬃciently

precisely.114 The veriﬁcation and validation process (V&V process) intends to establish

conﬁdence of the managers in the developed simulation model such that the ﬁnal simula-

tion model can be used for resolving the system’s initial deﬁciencies eﬃciently.

Basically, three diﬀerent groups may supervise the development process: the modellers,

the managers, or independent consultants.115 In any development process the modellers

constantly check the quality of modelling and programming process by logical checks,

back-tracing, and debugging. However, even a technically perfect implementation of a

mis-speciﬁed model cannot compensate the mis-speciﬁcation. Additionally, modellers

might not be objective with respect to the accuracy of their "product" and, in principle,

there might be a tendency to conceal own mistakes.

Therefore, the managers typically have to take a vital part in the V&V process. In

particular, the opinion of the project managers is inevitable to decide on the applicability

of the ﬁnal model.116 However, the managers are typically neither directly involved in the

programming process nor experts in the ﬁeld of programming. Hence, they may have a

lack of knowledge and expertise for a quick and qualiﬁed judgement about the accuracy

of modelling. As a third option, external consultants can be involved in the V&V process

to judge about the accuracy of modelling.117 These consultants should be experts in the

ﬁeld and should not be involved in the development process.

In the V&V process, various techniques are available relying on the comparison of a

113

See Law (2007).

114

See Law (2007).

115

See Sargent (2005).

116

See Sargent (2005).

117

See Balci (1997).

168

model with a reference system (e.g. the real system or the conceptual model). They can

be categorized according to the type of test and the type of criterion used. Basically,

the model structure can be evaluated by testing the internal logic of the model. I.e. it

is tested whether the relations ruling the model match the assumed relations deduced

from the reference system. On the other hand, the output of a model can be checked for

similarity with the reference system.

The criteria used for V&V can be qualitative or quantitative. Qualitative criteria

involve a subjective assessment on the logic or similarity of the tested model. In contrast,

quantitative criteria measure the similarity or consistency by a numerical comparison

of a reference value from the reference system with a (calculated) test value obtained

from the tested model.118 Table 4.8 shows examples of V&V techniques according to the

categorization criteria discussed.119

type of test

logical similarity

quali-

type of criterion

tative

• trace analysis • path testing

quanti- • extreme value test • statistical techniques

tative

• sensitivity analysis • trace-driven testing

These techniques are more or less applicable for the diﬀerent V&V activities. For

verifying the implemented simulation model, software engineering techniques are most

widely used which refer to the techniques displayed in the upper left cell in Table 4.8.

Cause-eﬀect-diagrams visualize relations between events and system outcomes. They

can be compared with commonly accepted assumptions about relations between events

and system outcomes to detect unintended model behaviour. Similarly, animation and

trace analysis often help modellers and managers to check the plausibility of the model’s

behaviour over time. However, not all programming mistakes uncover themselves in an

animation or in the analysed traces. Moreover, rare events are hard to detect since

animation is applied for rather short time frames and trace analysis is applied for a

118

Similarly, both extremes can be labelled with subjective and objective. See Rabe et al. (2008) for a

continuous-scale classiﬁcation on the degree of subjectivity.

119

For an extensive list of V&V techniques see Balci (1998).

169

A more objective way to check the logical structure of a model is to use quantitative

criteria. Fixed-value and extreme-value tests aim at the evaluation of the model’s be-

haviour under speciﬁc conditions for which a reference value can be derived ex ante. In

the former case, a ﬁxed setting is created omitting all stochastic inﬂuences. Hence, the

deterministic logic of the model can be tested. Similarly, the model can be tested under

extreme conditions.121 Sensitivity analysis carries out a systematic test of the model re-

sponse(s) for a restricted set of conﬁgurations to estimate general relations between input

parameters and output variables. These relations can be compared with ex ante known

relations. Note that for V&V purposes typically not the precise estimation of parameter

eﬀects is important, but the general relation (e.g. the sign of an eﬀect or the order among

the variables’ eﬀects).122

If data of the real system is available, the developed simulation model can be tested for

similarity with the real system in a quantitative way (bottom-right cell in Table 4.8). For

this purpose, a lot of statistical procedures can be applied depending on the speciﬁc object

to be tested. Typically, regression techniques, distribution tests, or time series analysis

methods are used.123 A reliable quantitative approach is to generate a forecast of the near

future by means of the simulation model which is then compared with the real systems

behaviour after the forecast period has expired. This is called predictive validation.124

A mixture of trace analysis and ﬁxed-value test is the trace-driven simulation where a

historical situation is simulated. The model’s output is compared with the historical

records then.

If explicit tests for similarity cannot be carried out, rather vague qualitative techniques

should be applied (top-right cell in Table 4.8). The Turing test is carried out by experts

or managers. It evaluates whether the output of a model can be distinguished from real

historical records.125 If not, this is an indicator that the model represents the reality

suﬃciently accurately.126 Similarly, face validity is a discussion process where simulation

outputs and/or model components are jointly discussed by the developers, managers,

and/or experts.127

If used solely, each technique has drawbacks and restrictions such that, typically, a set

of V&V techniques is used to verify the simulation model and to check validity. Con-

ﬁdence in the appropriateness of the model is rather created by a well-structured V&V

120

See Sargent (2005) and Rabe et al. (2008) for more information.

121

I.e. conditions which are barely realistic (e.g. when parameters are set to values which are judged as

just tolerable in reality), see Sargent (2000).

122

See Kleijnen and Sargent (2000) and Kleijnen (1995).

123

See e.g. Sargent (2005); Kleijnen (1995); Rabe et al. (2008).

124

See e.g. Rabe et al. (2008).

125

Therefore, real historical records and simulated records are presented and the expert/manager is asked

to identify the simulated record.

126

See Schruben (1980).

127

See Sargent (2000).

170

process than by a speciﬁc technique.128 Once the model development has passed all V&V

steps successfully, the simulation model is assumed to accurately reﬂect the real system’s

behaviour (w.r.t. the scope of the project). In the next steps, simulation experiments are

planned.

By deﬁnition, simulation models are rather descriptive than normative. Hence, to reveal

relevant information for decision makers, simulation experiments have to be conducted.

The way how to plan experiments and how to extract concise information from experi-

mental results is subject of the broad ﬁeld of experimental design. The general aim is to

extract as much information as possible about the simulated system with as little com-

putational eﬀort as possible. As both aims are contradictory, general assumptions about

the information to be extracted are made and the best way to extract this information

w.r.t. the associated computational eﬀort is sought for. This is typically done by assum-

ing a special type of relation between the variables of the studied system constituting a

mathematical model (e.g. a linear regression model).

The deﬁnition and classiﬁcation of variables depends on the speciﬁc questions to be

answered by the experiments. After speciﬁcation of relevant variables and their type of

relation, an experimental design is set up and simulation runs are performed. Analysing

the simulation results allows the analyst to parametrize the mathematical model. This

results in a meta-model of the simulated system. This simple mathematical model is

ﬁnally used to extract quantitative, concise information about the system’s behaviour.

Figure 4.11 reﬁnes Figure 4.6 by adding details about the experimental phase.

At the beginning of the planning process, it has to be clariﬁed which questions should

be answered by the study. In most cases, the research questions are rather qualitative,

unspeciﬁc statements than precise algebraic formulations. Hence, such statements have

to be operationalized into mathematically manageable terms. I.e. measures have to be

deﬁned describing the aspects of the system under study that should be investigated.

Such measures are called variables. In the context of experimental studies variables have

a lot of attributes depending on their purposes.130 Table 4.9 shows an overview on the

most common attributes of variables.

Basically, variables are distinguished in response variables and explanatory variables

depending on the role deﬁned by the study’s questions and hypotheses. Explanatory

128

See Rabe et al. (2008).

129

Based on Hinkelmann and Kempthorne (2005a).

130

See e.g. Law (2007).

171

statistical model

Experimental design

optimization

meta-modelling

decision support

this case, they are constructed by observable variables using a predeﬁned model and are

called latent variables.133 To investigate impacts of explanatory variables on response

variables, they have to be controllable. If they are uncontrollable, but measurable, they

are usually handled as nuisance variables. Eﬀects of uncontrollable, latent variables are

typically subsumed in error terms. In simulation experiments variables are observable and

measurable by deﬁnition.134 Nuisance variables are incorporated by stochastic processes

in a simulation model.

From the underlying questions of a study and their operationalization it follows how

variables are deﬁned and which role they can have. When designing and analysing sim-

ulation experiments typically multiple response and explanatory variables exist. When

131

Note that in structural equation models measurable indicator variables are grouped into factors. These

factors are unobservable/latent and inﬂuence latent/unobservable response variables. Such models

are typically designed and estimated to conﬁrm a theory about the construction and relations of the

latent factors, see Pearl (2000).

132

I.e. the existence of such a variable is assumed, but it cannot be measured precisely e.g. intelligence or

trustworthiness.

133

See e.g. Borsboom et al. (2003) for an example and deﬁnition.

134

In principle, also latent variables might be considered in simulation studies. However, in a simulation

study the subject studied is explicitly modelled in detail and not as a black box. Hence, the intro-

duction of latent variables in a simulation model yields typically no additional insights in the general

relation of its elements.

172

numeric

explanatory

(continuous, measurable controllable

(independent)

values

discrete)

response

ordinal

latent (dependent) uncontrollable

nominal nuisance

simulating a business system, response variables are usually measures of the system’s per-

formance variables. These variables are assumed to contribute to the overall success of

the business system. The contributions of performance variables to the overall success

are often operationalized as costs or revenue rates.

A prominent measure of overall success is the proﬁt obtained with a given set of resources

in a given time span. Basically, the proﬁt depends on costs, revenues and assets employed,

whereby those depend on system-internal and -external inﬂuences. In SC simulation the

aim is to investigate the behaviour of the system and explain the relations between its

components. Here, the focus is on studying internal relations under external disturbances.

Let x ∈ Rm denote the vector of controllable independent variables and y ∈ Rn denote the

vector of dependent response variables deﬁned in the model building phase. Furthermore,

e ∈ Ro denotes the vector of uncontrollable variables which are not considered in detail

in the model but are captured as environmental eﬀects. The function H(⋅) describes the

real relation between dependent and independent as well as environmental variables

y = H(x∣e). (4.4)

Typically, the function H(⋅) is unknown and reﬂects the behaviour of the real system.

Since the controllable variables x are often strategic decisions, experiments in reality are

often too expensive. Hence, simulation models are developed mimicking the real system’s

behaviour. Formally, the simulation model can be seen as a function Ĥ(⋅) such that

ŷ = Ĥ(x∣e). (4.5)

All these components inﬂuence the proﬁt u of the real system, i.e.

173

where U (⋅) is a function describing the relation between inﬂuencing variables and obtained

proﬁt u. Typically, U (⋅) is unknown in all details, but at least the direction of inﬂuences

of the variables are known.135 As the vector of controllable variables x reﬂects the set of

options to take, it is sought for an optimal conﬁguration xopt that either maximizes U (⋅)

or, if the focus is purely on internal processes, minimizes the associated total costs where

C(⋅) denotes the cost function of the system under study. I.e.

xopt = arg max U (x, H(x)∣e) or xopt = arg min C (x, H(x)∣e) . (4.7)

x x

x̂opt = arg max U (x, Ĥ(x)∣e) or x̂opt = arg min C (x, Ĥ(x)∣e) . (4.8)

x x

model. This is referred to as a scenario. In common sense, a scenario comprises both,

stochastic processes reﬂecting environmental conditions and the general structure of the

modelled system. Typically, the focus of simulation is on the internal processes of a supply

chain under a more or less speciﬁc environmental regime. E.g. in chemical industry an

SC’s revenues depend mainly on product prices that can be realized. Due to the highly

competitive market for basic chemicals and the inﬂexibility of (continuous) production

processes in combination with immense capital commitment for production plants, the

focus for optimization is on the internal processes of a chemical supply chain.

To evaluate the (total) costs of a conﬁguration for a speciﬁc scenario e requires knowl-

edge about the measures inducing costs (cost drivers) and a formal description of the re-

lation between costs and measures.136 However, a precise mathematical relation between

(performance) measures and costs is often hard to obtain because interdependencies ex-

ist among cost drivers. Moreover, cost rates are unknown and/or vary in time as they

e.g. depend on discounts, quantities, or interest rates. Therefore, the direct evaluation of

the costs of a conﬁguration is not reasonable as it implies deterministic and time-invariant

assumption about the underlying cost function C(⋅).

Instead, the direct evaluation of cost drivers is in focus for optimization. Cost drivers

are performance measures driving costs or proﬁt but their immediate contribution to

total costs/proﬁt cannot be expressed explicitly.137 For assessing the SC performance

diverse performance measurement systems are proposed with many of them relying on the

balanced scorecard approach138 in combination with the SCOR model.139 Performance

measures can be categorized according to the dimensions of the balanced scorecard or

135

I.e. the signs of the partial derivatives of w.r.t. x and y are known at least in some ranges.

136

In case of a linear relation, only one parameter (the cost rate) is required. But also more complex

relations are imaginable.

137

See Sürie and Wagner (2008).

138

See Kaplan et al. (1992) for the seminal work on the balanced scorecard and Bhagwat and Sharma

(2007) for performance measurement systems in SCM based on the balanced scorecard approach.

139

See Sürie and Wagner (2008) or Gunasekaran et al. (2004, 2001) and Agami et al. (2012) for an overview

of performance measurement systems.

174

SCOR model as well as their attributes.140 Table 4.10 shows a classiﬁcation of performance

measures according to aggregation level and type of measure.

costs for

costs per labour

cost inventory cost deprecation &

hour

ﬁnancial

amortization

revenue/

cash ﬂow total proﬁt corporate value

value-based

cost-to-income return on

relational cost-to-asset ratio

ratio investment

cash ﬂow cycle product

time order lead times

non-ﬁnancial

capacity forecasting customer service

reliability

utilization accuracy level

quality of

quality delivery quality process quality

produced goods

Depending on the scope of the simulation study, performance measures should be chosen

w.r.t. to the aggregation level as well as managerial impact. Ideally, an SC evaluation

system containing an already existing system of performance measures can be adapted

to the simulation environment.141 However, this procedure bears the risk that too many

criteria have to be evaluated which increases the computational eﬀort for evaluating SC

conﬁgurations. Moreover, traditional performance measurement systems aim at reﬂecting

the current situation of an SC whereas simulation aims at ﬁnding a future conﬁguration.142

Hence, the choice of performance measures to be evaluated in the simulation model should

account for cost and proﬁt drivers instead of a direct evaluation of ﬁnancial measures

since they typically imply assumptions about environmental ﬁnancial drivers such as raw

material prices or interest rates. This information, however, is uncertain and complicates

the system’s/model’s analysis. Furthermore, such measures distract from the underlying

processes to be optimized. In general, a conﬁguration of the SC is pursued that works well

under almost every environmental state. This ability is often labelled as ﬂexibility in SC

performance measurement systems.143 To operationalize this dimension in SC planning,

a robust conﬁguration is sought.144

Robustness is often an important feature, in particular when an objective’s variation

depends on the conﬁguration.145 Robustness can be measured in various ways, e.g. by con-

140

For the latter case see e.g. Beamon (1998).

141

See Kleijnen and Smits (2003).

142

See Kleijnen and Smits (2003).

143

For a discussion on ﬂexibility see Gunasekaran et al. (2001).

144

See Kleijnen and Smits (2003).

145

I.e. a conﬁguration x not only inﬂuences the mean level of responses H(x) but also their variance.

175

sidering the inverse coeﬃcient of variation instead of the pure expected performance146 or

by considering both expectation and variance of a measure as independent objectives.147

Another possibility to invoke robustness is to optimize a certain quantile instead of the

expected performance. This has the advantage that also measures with a skewed distri-

bution can be handled.148

From the deductions above it can be concluded that simulation studies on SC con-

ﬁguration put focus on multiple objectives/performance measures which depend on the

conﬁguration x. The conﬁguration and its performance aﬀect the total cost and/or total

proﬁt. The combination of a conﬁguration and its performance z = (x, y) is called a

constellation. In simulation studies on SC conﬁguration typically a set of eﬃcient con-

stellations is sought.149 Per deﬁnition an eﬃcient constellation is not dominated by any

other constellation. A constellation z = (x, y) is (strictly) dominated by a constellation

z̃ (formally z̃ ≻ z) if z̃ is (at least) indiﬀerent for all variables (z̃j ≿ zj for each j with

j = 1, ..., m + n) and strictly preferred for at least one variable (∃j ′ such that z̃j ′ ≻ zj ′ ).150

The set of eﬃcient constellations is denoted by Z ef f .

To sum up, there are two cases to be distinguished in simulation optimization: If a

unique objective can be deﬁned and calculated, an optimal conﬁguration is searched for.

In case of multiple objectives, the set of eﬃcient constellations has to be determined.

Both tasks are hindered by the following properties of simulation models:

Also, the selection of xopt or Z ef f is a challenging task and much research has been devoted

to this topic.

From a historical point of view, experimental designs were developed to provide methods

for planning real-world experiments.152 Here, typically the set of controllable variables

is large but their domains are restricted. Since most variables are considered as discrete

variables, they are often simpliﬁed to dichotomous variables or their domain is reduced

146

The so-called Taguchi approach see e.g. Shang et al. (2004).

147

See e.g. Kleijnen (2007, sec. 4.6) for an example.

148

See e.g. Law (2007, sec. 9.4).

149

Sometimes also called Pareto front.

150

The notation a ≻ b denotes that a is better than b, whereas a ≿ b declares that a is at least as good

as b. This notation is chosen to allow ﬂexibility in ordering relation. E.g. if for all components of z

holds "the smaller the better", ≻ could be replaced by <. However, using e.g. service level measures or

output measures as response variables, requires a converse ordering relation. Due to this, the common

order relations are generalized by preference relations.

151

The sample space depends on the number and domains of the control variables.

152

See e.g. the seminal work of Fisher (1971).

176

to a small number of values. Assume that for all control variables xi with i = 1, ..., m

such that x = (x1 , ..., xm ), a ﬁnite number of states mi exist. Then a conﬁguration x is

constituted by a combination of m active states, where each control variable has exactly

one level. Let bi denote a binary vector of length mi reﬂecting the level of variable

xi by a 1-entry at the corresponding position. Hence, the set of vectors bi is a binary

representation of conﬁguration x. Furthermore, let y denote the vector of responses of

length n. This allows modelling the outcome of conﬁguration x by a simple linear model

(multivariate analysis of variance):

m

ˆ

ˆ x = E (ŷx ) = μ + ∑ Γi ⋅ bi = E (Ĥ

ŷ (x∣μ, Γ1 , ..., Γm )) (4.9)

i=1

where Γi are matrices of dimension n × mi such that the entry in the kth row and lth

column (γkli ) is the eﬀect of the lth level of control variable xi on response variable yk .

Furthermore, μ ∈ Rn denotes the mean vector.

The aim is to (consistently) estimate the coeﬃcients μ and Γi with as little compu-

tational eﬀort as possible. I.e. only the minimal number of conﬁgurations153 necessary

to estimate these coeﬃcients should be evaluated. A set of conﬁgurations constitutes an

experimental design. A simple experimental design is the full-factorial design generated

by evaluating all possible conﬁgurations in the sample space. In this case, ∏m i=1 mi conﬁg-

urations have to be simulated. Obviously, this approach can only be realized in situations

when the sample space is small. For most realistic simulation models this approach is

computationally intractable since the number of conﬁgurations grows exponentially and

computational power is still limited.

From (4.9) it can be seen that for a particular response variable yj the number of

coeﬃcients to be estimated is 1 + ∑mi=1 mi .

154 Hence, there are only

m

1 + ∑ mi − m (4.10)

i=1

are generated such that all pairs of conﬁgurations are orthogonal.156 This constitutes a

resolution III design which is the simplest form of a fractional factorial design.157 If all

response measures yj are assumed to be independent,158 the corresponding coeﬃcients

153

In literature primarily the term "setting" is used instead of "conﬁguration".

154

This is for the mean and one row of all Γi .

155

This is because one coeﬃcient of each control variable is redundant as it equals the overall mean, see

Hinkelmann and Kempthorne (2005b) for details.

156

Orthogonality is useful as it implies pairwise uncorrelated conﬁgurations such that the classic estimators

for independent observations (e.g., the least squares estimator) can be applied, see e.g. Kleijnen (2007,

p. 33-35).

157

The term fractional refers to the fact that a subset of the full factorial design is extracted. See

e.g. Hinkelmann and Kempthorne (2005b) or Kleijnen (2007).

158

This is a reasonable assumption since otherwise a response variable could be calculated by the other

responses.

177

can be estimated simultaneously by a single design’s results such that multiple objectives

do not induce more conﬁgurations to be evaluated.159

Resolution III designs are applicable when all interactions between control variables

are assumed to be 0. If only mean and main eﬀects are to be estimated despite the

presence of non-zero two-variable interactions, resolution III designs have to be enhanced

to consistently estimate mean and main eﬀects.160 For enhancing the resolution III design

in the case of dichotomous variables the so-called foldover theorem can be applied.161 The

main idea is that any conﬁguration carries information of both the variables’ main eﬀects

and interaction eﬀects. To separate both eﬀects the mirrored conﬁguration is evaluated

which carries the same interaction eﬀects but the inverse main eﬀects. Hence, the number

of conﬁgurations to be evaluated doubles if two-variable interactions are assumed and a

so-called resolution IV design is created.162

If two-variable interactions are assumed to be non-negative and should be estimated,

the MANOVA model (4.9) changes to

m

⎛ b′i ⋅ Γ1(ij) ⋅ bj ⎞

ˆ ⎜ ⎟

ŷx = E (ŷx ) = μ + ∑ Γi ⋅ bi + ∑ ⎜ ⋮ ⎟ (4.11)

i=1 i,j∣j>i

⎜ ⎟

⎝ b′i ⋅ Γn(ij) ⋅ bj ⎠

where Γk(ij) is the matrix of interaction eﬀects between control variables xi and xj of

response variable yk . Matrix Γk(ij) has dimension mi × mj , but the ﬁrst row and the ﬁrst

column is always zero.163 Then, the number of coeﬃcients to be estimated for a particular

response variable is given by

m

1 + ∑ ((mi − 1) + ∑ (mi − 1)(mj − 1)). (4.12)

i=1 j>i

For example, assuming m = 6 dichotomous variables requires 7 coeﬃcients to be esti-

mated for each response variable for a main-eﬀect model as (4.9). If no interaction eﬀects

are assumed, for estimating these coeﬃcients a resolution III design with 8 conﬁgurations

needs to be evaluated.165 If two-way interactions are assumed to be non-zero but are

not to be estimated, the corresponding resolution IV design consists of 16 conﬁgurations.

The explicit modelling of all two-variable interaction eﬀects increases the number of coef-

159

See Kleijnen (2007) and the references therein.

160

This is because main and interaction eﬀects are confounded, see Kleijnen (2007, sec. 2.6) for details.

161

See Box and Wilson (1951) or Hinkelmann and Kempthorne (2005b, sec. 13.6.4).

162

Note that due to this generation scheme the resulting design is still orthogonal, see also Kleijnen (2007,

sec. 2.6).

163

The zero-entries correspond to constellations when at least one control variable is set at their base

value such that no interactions occur.

164

For details how such designs are generated see Hinkelmann and Kempthorne (2005b) or Kleijnen (2007).

165

To ensure orthogonality of a design, the number of settings has to be a power of 2 for dichotomous

variables.

178

implying that the corresponding resolution V design has 32 conﬁgurations.166 To generate

an explicit design a lot of methods and algorithms are proposed which are standard pro-

cedures implemented in statistical software packages and, hence, are not discussed here

in detail.167

Fractional factorial designs are useful when the domain of control variables is discrete

(binary at best). However, in case of continuous variables whose inﬂuence on the responses

is assumed to be linear and which are restricted to speciﬁc intervals, fractional factorial

designs can be applied, too. In fact, (4.9) changes to

ˆ x = E (ŷx ) = μ + Γ ⋅ x

ŷ (4.13)

where Γ is the n × m matrix of regression coeﬃcients. In such a case the control variables’

domains can be reduced to a binary set containing only the corresponding intervals’ ex-

tremes.168 This is because γki (the regression coeﬃcient of control variable xi on response

variable yk ) equals γki ⋅ (xk1 − xk0 ) = γk1i − γk0i where xkl are the chosen interval points of

the control variable and γkli are the coeﬃcients at both levels according to (4.9). Due to

the assumed linearity, eﬃcient constellations ẑˆ = (x, ŷ) ˆ can only exist on the boundary

of the simplex spanned by the control variables’ intervals. Hence, after estimating all

ˆ

coeﬃcients, i.e. μ̂ and Γ̂i , the meta-model Ĥ (x∣μ̂, Γ̂1 , ..., Γ̂m ) can be used to derive the

eﬃcient conﬁgurations by checking for dominance of the vectors ẑˆ = (x, ŷ ˆ x ) of all ∏m mi

i=1

possible conﬁgurations.

designs for planning simulation experiments, the chemical SC described in Examples 9

and 10 is considered again. Logistical core components in this model are the chemical

transports between both production sites. Hence, logistical key decisions have to determine

the corresponding capacities for inter-site transports. Relevant capacities are the RTCs at

hand per chemical transported, the transfer arms, the inventory capacities, and the target

levels as well as the number of available trains and the number of working days for RTC

handling.

To apply fractional factorial designs, the control variables have to be of nominal or

ordinal scale or the relations to the performance measures are assumed to be linear. Here,

the inﬂuence of the number of available trains, the number of working days for RTC

handling, and the number of transfer arms is investigated. As status quo, the data provided

in example 9 is used: From Monday to Saturday trains can be dispatched and RTCs

can be processed. For each product an equal number of transfer arms is available at

both sites according to Table A.12 in the appendix. As an alternative conﬁguration, the

166

For these numbers see e.g. Hinkelmann and Kempthorne (2005b, p. 535).

167

See Kleijnen (2007, ch. 2) for some generators for dichotomous variables.

168

Or any other distinct pair of values from the interval.

179

eﬀect of making Sunday available for train dispatching and RTC processing should be

tested to investigate whether an improvement in the network’s robustness and inventory

balancing can be achieved. Furthermore, the impacts of doubling the turnover capacity for

all chemicals is investigated which might reduce RTC cycle times and, hence, reduce the

number of required RTCs. In brief, Table 4.11 shows the values for each variable.

trains days C2 C3 C4 Styrene

status quo 6 6 2 3 4 3

extension 7 7 4 6 8 6

Table 4.11: Values of the control variables for the experimental design

Since for all variables two states are considered, in total 26 = 64 possible conﬁgurations

constitute the corresponding full factorial design. It is assumed that a variable’s eﬀects

on the responses depend on the values of the other variables. E.g. it can be assumed that

the individual eﬀect of extending the turnover capacity of a certain chemical potentially

reduces turnover times. In this case, RTCs might be repositioned faster and, hence, in

a given amount of time transport quantities could be increased. This might manifest

in more smoothed inventory levels and a reduced risk for inventory shortfalls. However,

these potentials might only be exploited if the transport capacities are extended, too. I.e. to

realize these eﬀects, train dispatches on Sunday might be necessary.

In other words, interaction eﬀects between the variables are considered to reﬂect the

systems’ dynamics. To ﬁnd a balance between modelling accuracy and computational

eﬀort, only two-variable interactions are considered. Note that, since mi = 2 for all

variables, main eﬀects and two-variable interaction eﬀects correspond to 22 coeﬃcients to

be estimated according to (4.12). Hence, a resolution V design is constructed with 25 = 32

conﬁgurations. Table A.17 shows the 32 evaluated conﬁgurations. For each conﬁguration

a time span of 180 days is simulated. To assess the responses’ variability, 100 replications

are conducted per conﬁguration.

The system’s reactions under speciﬁc constellations of variables manifest themselves in

various measures. Three classes are distinguished here: First, inventory-related measures

reﬂect costs incurred by holding stocks. On the other hand, stock-out costs occur when

supplied plants have to be shut down due to a shortage of material. These costs typically

reﬂect the loss in revenues when customer orders cannot be fulﬁlled due to a shortage

of ﬁnal products. It is assumed that the target stock levels reﬂect the points where stock

holding costs and expected stock-out costs are in balance.169 Hence, any deviation from the

desired target stock level implies an increase in total inventory-related costs. Therefore,

169

Note that a re-conﬁguration of the logistical system may lead to an improved logistical performance

(e.g. faster cycle time) and may inﬂuence the determination of target stock levels. However, this eﬀect

is not investigated here.

180

1

2 ∑T (sitp − sTar )2

1 2

3 t=1 ip

rip = ⋅ (4.14)

sTar

ip T

ip denote the stock level and target stock level of chemical p in period

t at site i, respectively.170 In total, eight inventory-related measures are calculated for

each simulation run of this example.171 In (4.14) over- and under-shooting of the target

level is a squared measure. This implies that over-shooting and under-shooting have the

same negative eﬀect by changing costs in the same rates. This might not be true for any

real-world problem. If over- and under-shooting should be handled individually, rip has to

be calculated with diﬀerent weights for under- and over-shooting.

A second category of measures reﬂects the logistical eﬀort inherited by a speciﬁc sys-

tem conﬁguration. Since capacities of the logistical system are handled as input variables

or kept ﬁx, responses should also measure the utilization of the capacities and, hence,

reﬂect the (additional) operational costs. In this example, operational logistic costs en-

compass RTC handling and maintenance costs, turnover costs, and charges for dispatched

trains. All costs except for the trains’ charges primarily depend on the labour costs of

the workforces. However, in the very most cases these costs are ﬁxed for a given stock of

personnel and, hence, do not depend on the utilization. Therefore, only the total number

of dispatched trains tr is kept as a response variable

tr = ∑ ∑ ∑ trijt , (4.15)

i∈I j∈I t∈T

where trijt denotes the number of trains dispatched in period t on link (i, j).

As a third category production-related performance measures reﬂect the monetary cash-

ﬂow resulting from production operations. This includes costs for raw and auxiliary mate-

rial consumption, labour costs of operators, opportunity costs for tied capital, and revenues

due to sales of ﬁnal and intermediate chemicals. A conﬁguration of the network’s logisti-

cal system aﬀects the production system by providing prerequisites for normal production

operations. I.e. a conﬁguration assures that raw materials are available and that ﬁnal and

intermediate chemicals can be stored and delivered to customers. In this simulation study,

these eﬀects result in the availability or storability of intermediate chemicals. If material is

unavailable or not storable (e.g. due to fully utilized stock capacities), a disturbance of the

production processes results such that plants have to reduce their production outcome or

shut down. An appropriate measure to account for production disturbances is to measure

the gap between planned and realized production for all plants associated to the considered

170

Note that rip can be seen as a relative root mean squared error (RMSE) for chemical p at site i.

171

Since there are two sites and four products in focus.

181

chemicals. This measure equals the β-service level of each chemical considered:

∑t∈T ωipt

real

βip = plan

. (4.16)

∑t∈T ωipt

real plan

In (4.16), ωipt and ωipt denote the realized and planned balance of chemical p in period t

at site i.172 Note that the planned balance ωipt plan

of a chemical also incorporates variations

of the plants’ production states due to technical or other reasons such that the diﬀerence

real plan

between ωipt and ωipt only accounts for deviations due to logistical reasons.

To sum up, for the chemicals considered in inter-site transport in total 16 responses

are deﬁned by (4.14) - (4.16).173 To simplify the following analyses it is assumed that all

chemicals at both sites have equal priority. Service level and inventory-related responses

can be merged into average measures by

r= (4.17)

∣I∣ ⋅ ∣P∣

and

∑i∈I ∑p∈P βip

β= . (4.18)

∣I∣ ⋅ ∣P∣

Then, a scenario comprises six input variables whose eﬀects on three response variables

are evaluated. To calculate the eﬀects, linear models with two-way interaction eﬀects

are estimated for each response variable. The average β-service level is restricted to the

interval [0, 1]. An ordinary linear model is not an appropriate model for probabilistic

measures since its prediction space is not restricted. Therefore, a logit model is used to

describe the input variables’ eﬀects on the average β-service level. I.e.174

6 6

β

log ( ) = μβ + ∑ (γiβ + ∑(γiβ γjβ )ij ) (4.19)

1−β i=1 j>i

where γiβ is the (main) eﬀect of setting input variable xi to its higher value and (γiβ γjβ )ij

is the two-way interaction eﬀect of setting variables xi and xj to their higher values. Note

that μβ corresponds to the basic conﬁguration where all input variables are set to their

lower/initial values.

For both other responses ordinary linear models are estimated, i.e.

6 6

tr = μtr + ∑ (γitr + ∑(γitr γjtr )ij ) (4.20)

i=1 j>i

6 6

r = μr + ∑ (γir + ∑(γir γjr )ij ) (4.21)

i=1 j>i

172 plan

Note that in the considered scenarios ∑t∈T ωipt ≠ 0 for all sites, periods, and products.

173

For C4 at site 1 no consuming plant exist. Hence, no local β-service level exists such that only seven

service levels are calculated.

174

For detailed information about logit models see e.g. Faraway (2005).

182

where the coeﬃcients (γ) are deﬁned as described above. Note that the relative average

sCap

inventory deviation r is restricted to the interval [0, max (1, ip

sTar

)]. This violates the as-

i∈I,p∈P ip

sumption of classic linear models that variables have continuous and unbounded support.

However, an explicit modelling of truncated variables does not reveal a signiﬁcant advan-

tage of this model as the simulated values are suﬃciently far away from their extremes.

Hence, a linear model is used since the violation of model assumptions is expected to be

ignorable. Beside the input variables, a main inﬂuence on the responses is the number

and duration of plant break-downs due to technological reasons. To take this inﬂuence

into account, the total quantity of the corresponding chemicals produced at both sites is

incorporated as an additional explanatory variable.175 Figures 4.12a - 4.12c show the box-

plots for the three responses. The estimated responses based on the models (4.19)-(4.21)

are superimposed by the symbol ▲.

The corresponding estimated coeﬃcients γ can be found in Table 4.12.

variable Estimate Pr(> ∣t∣) Estimate Pr(> ∣t∣) Estimate Pr(> ∣t∣)

(Intercept) -13.00 0.00∗∗∗ 205.94 0.00∗∗∗ 0.51 0.00∗∗∗

# arms C2 -0.02 0.77 -3.07 0.00∗∗∗ -0.02 0.01∗∗∗

# arms C3 0.04 0.51 -0.95 0.33 0.01 0.23

# arms C4 0.05 0.40 -0.88 0.37 0.00 0.88

# arms Styrene -0.04 0.49 -0.44 0.66 0.00 0.62

# trains 0.20 0.00∗∗∗ 18.66 0.00∗∗∗ -0.03 0.00∗∗∗

# handling days 0.07 0.20 -0.88 0.37 0.00 0.71

arms.c2:arms.c3 -0.09 0.07† 0.48 0.54 0.00 0.99

arms.c2:arms.c4 0.02 0.72 -0.05 0.95 0.01 0.23

arms.c2:arms.sty 0.05 0.28 -0.21 0.79 0.00 0.83

arms.c2:days.train -0.05 0.34 0.15 0.86 0.00 0.67

arms.c2:days.hand 0.07 0.12 -0.48 0.55 0.00 0.85

arms.c3:arms.c4 -0.02 0.65 -0.50 0.53 0.00 0.68

arms.c3:arms.sty 0.06 0.23 0.34 0.67 0.00 0.77

arms.c3:days.train -0.01 0.88 0.72 0.37 -0.01 0.17

arms.c3:days.hand 0.01 0.87 -0.17 0.83 0.00 0.70

arms.c4:arms.sty 0.02 0.74 0.14 0.86 0.00 0.88

arms.c4:days.train -0.07 0.13 0.17 0.83 0.00 0.90

arms.c4:days.hand -0.04 0.36 0.99 0.22 0.00 0.47

arms.sty:days.train 0.01 0.83 -0.73 0.36 0.00 0.79

arms.sty:days.hand -0.10 0.03∗ 1.40 0.08† -0.01 0.23

days.train:days.hand -0.06 0.22 0.63 0.43 0.00 0.97

adj. R2 /pseudo-R2 0.61 0.45 0.05

∗∗∗ ∗∗ ∗ †

signiﬁcance codes: ...≤ 0.001; ...≤ 0.01; ...≤ 0.05; ...≤ 0.1

The corresponding QQ-plots for all models’ residuals are depicted in Figures A.3a-A.3c.

The QQ-plots displayed in Figure A.3b and Figure A.3c show a good ﬁt to the normal

175

Note that this variable is independent from the experimental design such that no collinearity exists.

183

1.00

0.95

●

●

● ●

● ●

● ● ● ● ●

●

● ● ● ● ●

● ●

● ● ● ●

●

● ● ● ● ● ● ●

● ● ● ● ●

● ●

● ● ● ● ● ● ● ●

●

● ● ● ● ● ● ● ● ● ● ●

● ●

● ● ● ● ● ●

●

● ● ● ●

● ●

● ● ● ● ● ●

β−service level

● ● ● ● ● ● ●

● ● ● ● ● ● ● ● ● ●

● ● ● ● ● ●

● ●

● ● ● ● ● ● ●

● ● ● ● ● ●

● ●

● ● ● ● ● ● ●

● ● ● ● ● ●

● ● ●

0.90

● ● ● ● ● ● ● ● ● ●

● ● ● ● ● ●

● ● ● ● ● ●

● ● ●

● ● ● ● ● ●

● ●

● ● ●

● ●

● ● ● ●

● ●

● ● ● ●

● ● ●

● ●

● ● ● ● ● ●

● ● ●

●

● ● ● ●

● ● ● ●

● ●

0.85

● ● ●

● ● ●

● ●

●

● ●

●

●

●

● ● ●

● ●

0.80

0.75

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

scenario

(a) Boxplots of simulated and ﬁtted β-service levels per scenario

320

● ●

●

●

●

●

300

●

280

# trains

260

● ●

● ● ●

● ● ●

● ●

240

● ● ● ●

●

●

● ● ●

● ●

● ● ●

220

●

●

●

●

fitted value per scenario

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

scenario

(b) Boxplots of simulated and ﬁtted number of trains per scenario

●

●

0.7

● ●

● ●

●

inventory deviation

0.6

0.5

0.4

●

●

●

●

● ●

● ●

●

0.3

● ● ●

● ● ● ●

●

● ●

●

●

fitted value per scenario

●

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

scenario

(c) Boxplots of simulated and ﬁtted inventory deviation per scenario

184

distribution.176 For the logit model (4.16) the QQ-plot has only limited relevance, since the

residuals are not assumed to be normally distributed.177 Thus, Figure A.3a shows a typical

pattern for truncated variables with increasing concentration towards the truncation. From

the QQ-plots no serious violations of the model assumptions can be detected for all three

ﬁtted models. All models signiﬁcantly reduce the corresponding residuals’ variances178

whereby the explained variance is about 50% for (4.16) and (4.20) but only 5% for (4.21).

The low level of explained variance for (4.21) is an indicator that either not all relevant

inﬂuences are included in the regression or the investigated variables only slightly aﬀect

this measure.

From Figure 4.12 it can be deduced that the variances of the response variables are

similar in the experimental conﬁgurations. This presumption is supported by Bartlett’s test

for variance homogeneity for all three responses (at a signiﬁcance level of 1%).179 Hence,

an eﬀect of the input variables on the variation of the responses cannot be veriﬁed such

that a discussion on the robustness of an experimental conﬁguration is obsolete w.r.t. to

the selected three responses. Table 4.12 contains the estimates and t-test probabilities

of the coeﬃcients of models (4.19)-(4.21). It can be observed that the number of trains

dispatchable within a week (shipment capacity) is the only variable that has a signiﬁcant

eﬀect on all responses. As expected, the eﬀect on the β-service level and the total number

of trains dispatched is positive, whereas the inventory deviation can be signiﬁcantly reduced

by additionally available train transports. An extension of the turnover capacity for C2

has signiﬁcant negative eﬀects on the number of trains and the stock deviation.

Only few (slightly) signiﬁcant interaction eﬀects can be found. A negative impact on

the β-service level occurs for a simultaneous extension of turnover capacities for C2 and

C3 as well as the simultaneous extension of Styrene’s turnover capacity in combination

with extended RTC processing times. Negative two-way interaction eﬀects are not to be

expected whereas positive ones indicate an eﬀective exploitation of extended capacities.

Due to the fact that only 100 replications were performed it is presumed that the (slightly)

signiﬁcant two-way interaction eﬀects are spurious and reﬂect a noise eﬀect.

A positive interaction eﬀect on the total number of dispatched trains is observed for the

simultaneous extension of Styrene’s turnover capacity in combination with an extended

RTC processing time. I.e. more trains are dispatched as RTCs can be handled faster.

This might be an indication that the ﬂeet of Styrene RTCs is too small. However, this

176

However, Jarque-Bera tests do not support this founding for both models. A closer inspection of the

QQ-plots reveals indications for slightly fat-tailed distributions of the residuals. Hence, the models

were re-ﬁtted for Box-Cox-transformed dependent variables (with λtrain = 1.4747 and λstock = 1.6363).

For these models Jarque-Bera tests conﬁrm the normality of residuals. For the sake of brevity the

detailed model speciﬁcations are not shown here as the general results are the same as for the original

model.

177

See Faraway (2005).

178

This founding is veriﬁed by F-tests for models (4.20) and (4.21) as well as Wald’s test for model (4.16)

at a signiﬁcance level of 1%.

179

Note that the left-skewed distribution of the simulated β-service levels as depicted in Figure 4.12a is

compensated by the logit transformation.

185

Based on the models (4.16)-(4.21) with estimated coeﬃcients as stated in Table 4.12 the

expected outcomes of the remaining 32 (not simulated) conﬁgurations can be estimated.

The results are depicted in Table A.18 in the appendix. Among all possible 64 conﬁg-

urations, 15 can be dropped as these are dominated by at least one other conﬁguration.

Note that the basic conﬁguration (# 1) where all variables are at their initial values is

eﬃcient, but the extreme conﬁguration with all variables at their maximum values (# 64)

is dominated.180 I.e. a simultaneous extension of all variables is not advisable. Based on

the signiﬁcant main eﬀects it can be concluded that an extension of the turnover capacity

for C2 results in a reduction of the logistical eﬀort measured in terms of the number of

dispatched trains.181 On the other hand, if the focus is rather on the logistical reliability

the extension of the shipment capacity has a positive eﬀect on both reliability measures

(β-service level and stock deviation) which comes at the cost of 18 additional trains in a 6-

month period on average.182 An extension of both capacities improves the stock reliability

only slightly and reduces the additional number of dispatched trains.183

Considering a situation, when a re-conﬁguration of the existing SC should be evalu-

ated, the decision for a potential re-conﬁguration depends on the costs of the potential

re-conﬁgurations184 as well as the predicted beneﬁt of the corresponding performance ef-

fects.185

other methods have to be applied to ﬁnd eﬃcient constellations. The "classic way" is

the so-called response surface methodology (RSM).186 Here, local linear meta-models are

estimated iteratively using resolution III or IV designs (this step in the RSM is called

response surface approximation (RSA).187 Local refers to a (small) sub-space of the con-

trol variables’ sample space. In the subsequent iteration the investigated sub-space is

re-located according to a given step-width and a gradient-based direction (depending on

the preference relation of the dependent variables).188 If an optimum exists in some region

of the sample space, local linear meta-models will show a serious lack of ﬁt which can be

180

Conﬁguration # 64 is e.g. dominated by conﬁguration # 60.

181

For details see e.g. conﬁguration 3 in Table A.18.

182

See e.g. conﬁguration # 17 in Table A.18.

183

On average by three trains, see e.g. conﬁguration # 18 in Table A.18.

184

Re-conﬁguration costs are e.g. investments for turnover capacity extensions.

185

Financial impacts may realize e.g. as cost savings due to more smooth stock levels.

186

See Barton and Meckesheimer (2006) for details on meta-modelling using RSM.

187

See Kleijnen (2007).

188

Classic RSM applies a simple steepest descent approach, see Law (2007) or Kleijnen (2007). Typically,

RSM assumes an unconstrained experimental space (except for upper and lower bounds). However,

an extension to incorporate (additional) constraints on the experimental space is provided by Kleijnen

(2008).

186

quantiﬁed by various measures.189 If the lack of ﬁt exceeds a critical value, a local poly-

nomial of higher-order (i.e. second-order) is used to approximate the local responses.190

If the local meta-model ﬁts suﬃciently accurately, this model is used to determine the

optimal conﬁguration.191

Unfortunately, in classic RSM the local RSA relies on local polynomial relations and

requires that all general assumptions of linear models have to be fulﬁlled. This includes

homogeneous variances among the conﬁgurations and normally distributed errors. In

many simulation experiments these assumptions are violated which hinders a consistent

ﬁt of linear models and interferes the optimization process. Therefore, in advanced RSM,

local linear meta-models are replaced by other functions known as surrogate functions.

In the literature, ﬁve main classes of surrogate functions are distinguished: Kriging mod-

els,192 neural networks,193 regression splines,194 support vector regression,195 and radial

basis functions.196 The general procedure of the RSM remains unchanged, but the gra-

dient deﬁning the local sub-space to be explored in the next iteration is calculated based

on the aforementioned local meta-model functions.197

As an alternative to RSM, simulation responses can be used directly to explore the sam-

ple space of control variables. To do so, a lot of combinatorial optimization approaches

were adapted for simulation optimization. In general, there are four main classes of

methods that have shown a particular applicability in (multi-objective) simulation opti-

mization: Meta-heuristics, gradient-based procedures,198 random search,199 and sample

path optimization.200 Of particular interest are meta-heuristics as they have shown a

good performance for a wide range of combinatorial optimization approaches. Therefore,

commercial simulation software primarily uses these techniques to incorporate simulation

optimization routines.201 Among meta-heuristics, tabu search, scatter search, and genetic

algorithms are most widely used.202 Table 4.13 provides an overview on all aforementioned

techniques.

In non-commercial applications often genetic algorithms are proposed.203 In general,

189

Such as the adjusted R2 or the Karesh-Kuhn-Tucker conditions, see Kleijnen (2007).

190

The most prominent design to ﬁt a second-order polynomial is the so-called central-composite design,

see e.g. Kleijnen (2007). Here, a resolution V design is enhanced by 1 + 2 ⋅ m conﬁgurations to be able

to estimate quadratic eﬀects, too.

191

See Wilson et al. (2001) for details on applying RSM to ﬁnd eﬃciency frontiers.

192

See e.g. Kleijnen (2009) for details.

193

See e.g. Sabuncuoglu and Touhami (2002) or Fonseca et al. (2003).

194

See Li et al. (2010b) or Barton and Meckesheimer (2006).

195

See e.g. Clarke et al. (2005).

196

See e.g. Barton and Meckesheimer (2006) or Barton (2009) for an overview and Shin et al. (2002) for

an application of radial basis functions.

197

See Li et al. (2010b) for a comparison of these meta-models.

198

See e.g. Fu (2006).

199

See Andradóttir (2006).

200

See Fu et al. (2005) for an overview.

201

See e.g. Law (2007) or Banks et al. (2005).

202

See e.g. the commercial build-in optimizer OptQuest Laguna and Martí (2003) and Laguna (2011), or

WITNESS optimizer (Lanner Group, 2002).

203

Albeit, there is also commercial simulation software using this type of meta-heuristics, see e.g. Barton

187

Response surface

direct optimization

approximation

• meta-heuristics

• linear & quadratic regres-

sion (classic RSA) – tabu search

– scatter search

• regression splines

– genetic algorithms

• support vector regression

• gradient-based methods

• radial basis functions

• random search

• neural networks

• sample path optimization

the subset of non-dominated conﬁgurations with best ﬁtness scores

5. repeat step 2 and 4 until the set of non-dominated conﬁgurations is suﬃciently large.

to be achieved:

set of eﬃcient conﬁgurations (i.e. the Pareto front) as entirely as possible.205

While the ﬁrst goal comes along with single-objective optimization, the second goal is

characteristic for multi-objective optimization problems. Hence, multi-objective optimiza-

tion procedures try to spread the pool of non-dominated conﬁgurations along the Pareto

front. To do so, the diversity of non-dominated conﬁgurations has to be incorporated in

the solution procedure.

204

See e.g. Coello et al. (2007).

205

See e.g. Srinivas and Deb (1994) or Coello et al. (2007).

188

the ﬁtness value of a speciﬁc conﬁguration is the higher, the more other conﬁgurations it

dominates and the more separated this conﬁguration is. Originally, the measure of sepa-

rateness depends on the distance between two conﬁgurations and a sharing parameter.207

In the improved version (NSGA-II), this measure is replaced by the so-called crowding

distance which is the average distance to a conﬁguration’s component-wise nearest neigh-

bours.208 The NSGA approach has achieved a wide ﬁeld of applications, particularly in en-

gineering and SCM applications.209 To further improve the algorithm’s performance, the

set of non-dominated conﬁgurations is stored in a central, iteration-independent archive

which assures that no non-dominated conﬁguration is lost during the optimization pro-

cess.210

The following example shows an application of the NSGA-II algorithm for a subsystem

of the chemical SC example introduced before.

of Naphtha was introduced as a continuous supply of both crackers via pipeline from a

reﬁnery. This implies a direct dependency on the supplying reﬁnery. An alternative is

to assume that the supplier in Figure 4.8 represents a sea ship terminal where Naphtha

can be unloaded from ships. At the terminal, tanks are available for intermediate storage

of Naphtha. Transports of Naphtha via ship are now associated with shipment costs to

be charged for a speciﬁc transport relation. Often reﬁneries are located ashore or have a

direct connection to a sea ship terminal in order to reduce the dependency from a single

supplier.211 The inventory management of Naphtha is crucial in the conﬁguration of the

SC because Naphtha stocks constitute an important part of the total quantity of stored

chemicals. Hence, Naphtha stocks cause a relevant contribution to the SC’s total stock

holding costs. Therefore, the inventory management primarily aims at keeping the stock

holding and supply costs in balance.

The inventory policy controls the stocks at the three locations (the harbour and both

production sites) and the number of Naphtha shipments. The total demand faced by the

terminal equals the sum of the Naphtha consumption of both sites which is ω max = 5, 000

tons per day in regular operation. The crackers at both sites are modelled as described in

in = 20, 000 tons a day.

example 9. It is assumed that the harbour’s unloading capacity is ρcap

206

See Deb et al. (2002) and Srinivas and Deb (1994).

207

See Srinivas and Deb (1994).

208

See Deb et al. (2002).

209

See e.g. Deb et al. (2007) or Bin et al. (2010) for engeneering applications and Puigjaner and Guillén-

Gosálbez (2008) or Mele et al. (2006) for applications in the SCM context.

210

Otherwise, the maximal number of eﬃcient conﬁgurations to be found equals the population size in

each iteration. This loss of non-dominated conﬁgurations is called Pareto drift and the improved

algorithm is called NSGA-IIa, see Goel et al. (2007).

211

However, typically contracts with only few suppliers will be made to cover the regular demand of

Naphtha. Nonetheless, the sea ship terminal provides the opportunity to order additional supplies

e.g. in case of a shortfall of the regular supplier(s).

189

Unloading is possible every day. The tank capacities at both sites are shown in Table 4.4

whereas the tank capacity at the sea port is assumed to be 75, 000 tons.

The inventory management follows local (si , Si ) policies at both production sites (i =

{1, 2}) and the harbour (i = h). Six parameters have to be determined to fully specify the

SC’s raw material inventory management. If the available stock at location i ∈ {1, 2, h} in

period t (say lit′ ) falls below the corresponding order point si , an order is placed with the

order quantity qit = Si −lit′ . The available stock comprises the local inventory level (lit ) and

all non-delivered orders (∑Tτ=t qiτ ). If orders are placed at the production sites, they are

immediately fulﬁlled via pipeline as long as a suﬃcient quantity of Naphtha is available

at the harbour tanks.

The pipeline is assumed to be highly reliable. In certain intervals, however, the pipeline

control system reports failures which are partly caused by misestimations of pressure but

sometimes also due to micro-leakages. In any case, an inspection of the identiﬁed pipeline

segment is necessary which forces an interruption of pipeline operations. It is assumed

that the time span between two alarms talarm (given in days) is Weibull distributed with

talarm ∼ WB (k = 1.5, λ = Γ( 5 ) such that E(t

365

)

alarm ) = 365. I.e. one year is expected to

3

elapse before a new alarm occurs.

The inspection time before the pipeline can be used again (tinsp ) is also assumed to be

Weibull distributed with (tinsp − 1) ∼ WB (k = 1.5, λ = Γ(35 ) ) with a minimum inspection

3

time of 1 day in case of a false alarm. If the pipeline has to be repaired, the time span

expands considerably. The formula of the density function and a plot of the density for

talarm are given in the appendix in (A.1) and Figure A.4, respectively.

At the harbour, Naphtha orders are placed according to the same inventory policy. Here,

however, the ordered quantity is delivered after some transport time when the ship arrives

at the harbour. This time span depends on the supplier and the shipper organizing the

transport. The order lead time (tship ) is considered as a Weibull-distributed random vari-

able with (tship − 2) ∼ WB (k = 2, λ = Γ(1.5)

5

). It is assumed that under optimal conditions

(i.e. closest supplier and immediately available shipment capacities) at least two days

elapse before an order can be received.

The objective is to minimize the inventory holding costs which comprises a) ordering/-

shipment costs, b) stock holding costs and c) shortfall costs. The drivers of these costs

are a) the number δ of shipments, b) the average total inventory ¯l and c) the plant uti-

lization β.212 As discussed above, the precise cost rates of these three drivers are assumed

to be unknown and time-varying. The aim is to determine eﬃcient conﬁgurations of the

inventory parameters and to quantify the trade-oﬀ between the three cost drivers.

From example 9 it is known that the pipeline transport is assumed to induce negligible

ﬁxed costs such that a continuous supply of both sites minimizes the cycle stocks. Hence,

it is concluded that Si = si + ωimax for i = 1, 2 such that at both sites only safety stocks

212 ∑t ∑i lit

β is calculated analogously to (4.18) and (4.16). ¯l is calculated as ¯l = T

. δ simply counts the

number of non-zero orders at the harbour (qht > 0).

190

are at hand to cover shortfalls of the pipeline supply but no cycle stocks due to batch

building.213 If a pipeline is under repair or inspection, the demand to be covered from

stock depends on the inspection time tinsp (which is Weibull distributed) and production

modes of the crackers during this time (which follow a discrete Markov process). Hence,

a (closed-form) mathematical expression of the distribution of Naphtha consumption at a

site during a pipeline inspection is hard to derive. However, Monte-Carlo simulation can

be used to approximate the distributions of this measure at both sites.

Based on the density estimates, the corresponding loss functions are calculated straight-

forwardly. Since the Markov processes of both crackers are very similar, the loss func-

tions show a very similar pattern, diﬀering primarily in the scale (which corresponds to

the crackers’ capacities). Figures 4.13a-4.13c show the estimated density functions and a

QQ-plot depicting the re-order levels to be held at both sites to achieve a speciﬁc β-service

level.214 I.e. each point in Figure 4.13c corresponds to a speciﬁc β-service level that is

achieved by setting the associated re-order levels at the sites.215 The loss functions for

both sites are shown in Figure A.5 in the appendix.

Based on the estimated density functions, the mean and the 95%-quantile are superim-

posed by the dashed lines in Figure 4.13a and Figure 4.13b whereby the latter coincides

with the re-order level ensuring a 95% α-service level. For both sites the density functions

of total consumption are bimodal and skewed to the right. The ﬁrst mode is at zero con-

sumption and is inherited from the Markov chain of the production models. It corresponds

to situations when a pipeline inspection coincides with a cracker shut-down. The second

mode is inherited from the Weibull distribution determining the pipeline inspection time

which also causes the skewness.

From Figure 4.13c it can be taken that a linear relation between the re-order levels at

both sites exists. This relation can be used to reduce the set of parameters to be explored by

deﬁning s2 = s1 ⋅ 23 whereby the slope of the ﬁtted regression line in Figure 4.13c corresponds

almost exactly to the ratio of cracker capacities.

Accordingly, the relation between the local re-order level si and the expected loss during a

pipeline inspection period can be quantiﬁed.216 Local stocks protect against the shortfall risk

due to a pipeline break-down. Additionally, the variability of supply has to be examined.

Since it is generally advantageous to consolidate inventories (risk pooling), stocks at the

harbour should be held to buﬀer supply and production risk. Therefore, the parameters

(sh , Sh ) have to be determined controlling the Naphtha availability of both crackers and

the number of shipments for replenishment. The density of the total Naphtha consumption

during the order lead time is a convolution of Markov processes and the Weibull distributed

213

Here, ωimax denotes the maximum consumption rate of cracker i.

214

The density functions are derived by Gaussian kernel estimation with the default settings of the density

function from the stats package based on a 10,000-replicates sample, see R Core Team (2012). The

loss functions are calculated straightforwardly from the density functions’ estimates.

215

The leftmost point implies a β-service level of 5% whereas the rightmost point constitutes a β-service

level of 99.5%.

216

This corresponds to the expected cracker utilization and, hence, expected shortfall costs.

191

mean 95%−quantile

0.00006

0.00004

density

0.00002

0.00000

consumption (in t)

mean 95%−quantile

0.00008

density

0.00004

0.00000

consumption (in t)

●

15000

●

safety stock site 2

●

●

●

●

●

10000

●

●

●

●

●●

●●

●●

●●

●●

●●

●●

●●

●●●●●

●●●

5000

●●●

●●●

●●●

●●●

●●●●

●●●●

●●

●●●●●

●●●

●

●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●

0

safety stock site 1

(c) Comparison plot of re-order levels

Figure 4.13: Loss functions for Naphtha consumption during pipeline inspection and their

relation

lead time. Figure 4.14a and Figure 4.14b show the simulated density and loss function of

192

this measure.217

mean 95%−quantile

0.000030

0.000020

density

0.000010

0.000000

consumption (in t)

(a) Density function for total Naphtha consumption

V−1⎛⎝E(X) ⋅ (0.05)⎞⎠

25000

expected loss V(r) (in t)

15000

5000

E(X) ⋅ (0.05)

0

re−order point sh (in t)

Figure 4.14: Density and loss function for total Naphtha consumption during order lead

time

Figure 4.14a shows no (clear) bimodality whereby the density for a zero-consumption is

positive. This is because the independent Markov chains of both crackers form a collapsed

Markov process whose state space and steady state vector is smoother. In both ﬁgures,

95%-service levels are superimposed by dashed lines (the α- service level in Figure 4.14a

and the β-service level in Figure 4.14b). Using this information about the central re-order

level at the harbour in combination with the information about the local re-order level, the

system’s total expected backlog could be calculated using the density and loss functions of

each stage. However, the interaction of local safety stocks and the central safety stock at

the harbour needs to be formulated explicitly which is quite cumbersome. Since the local

217

Again a Monte-Carlo simulation with a sample size of 10,000 is used. The estimation procedure is the

same as described before.

193

safety stocks also buﬀer lacks of supply due to delayed shipments, an independent setting

of the parameters si based on the individual loss functions underestimates the real capacity

utilization (and the β-service level). Furthermore, the deduction of the average stock level

at all three locations cannot be expressed in a closed form.218 Therefore, the complete

inventory system is simulated.

The system is modelled on a daily basis and implemented as an R program. Due to the

simplicity of the model, the program’s pseudocode is kept for the appendix (Table A.19).

The NSGA-II algorithm is used for optimization with three parameters specifying a con-

ﬁguration: s1 , sh , and Sh . The remaining parameters are set according to the relations

discussed above. I.e. the re-order point of location 2 is given by s2 = s1 ⋅ 2/3 whereas the

order levels are expressed as S1 = s1 + 3, 000 and S2 = s2 + 2, 000. The shipment quantity

Sh − sh is restricted to be at least 15,000 tons which avoids too frequent deliveries and

ensures a suﬃcient utilization of the tanker ﬂeets. Furthermore, lower bounds for the

re-order points are set to s1 , sh ≥ 5, 000 to ensure a minimal safety stock at both crackers.

The parameters’ upper bounds are given by the local tank capacities. The implementation

of the NSGA-II algorithm in the nsga2 function contained in the mco package of the R

environment is used.219 In this optimization procedure, the population size is set to 1,024

which is a suﬃciently large sample for further statistical analysis. The number of gener-

ations is set to 50 which results in a fairly good convergence to the Pareto front.220 The

remaining parameters are set to the default values.221

Each conﬁguration is simulated for a 10-years period to account for the rather long

intervals between pipeline inspections. No replications of a conﬁguration are performed

since the variability of the performance measures primarily depends on the length of the

simulated time span. Averaging over a number of replications would in turn overweight

the warm-up phase of the simulation. In the warm-up phase the performance measures

are quite stable due to ﬁxed initial stocks.222 The optimization using NSGA-II generates a

Pareto front consisting of 1,024 observations. Figure 4.15 visualizes the found solutions.

In Figure 4.15 the average total stock in tons per period is plotted against the number

of shipments ordered per period. Each Pareto-optimal solution is represented by a grey-

scaled point whereby the shade indicates the realized β-service level of this solution. Bright

solutions show a β-service level of about 40% whereas dark-coloured refer to a β-service

level close to 100%.

This sample can be used to describe the Pareto front by an appropriate surrogate function

218

See Schneider et al. (1995) for an approximation.

219

See Trautmann et al. (2010).

220

Experiments with 100 and 200 generations and smaller population size reveal no signiﬁcant shifts in

the Pareto front.

221

The defaults are e.g. a mutation probability of 0.2 and a crossover probability of 0.7, see Trautmann

et al. (2010).

222

Initial stocks are 15,000, 10,000, and 50,000 tons for locations 1,2, and the harbour stocks, respectively.

In general, for such a simulation study there is a trade-oﬀ between multiple replications and one long

simulation run, see Law (2007, sec. 9.5).

194

● ●● ● ● ●

●● ●●● ● ●

● ● ●● ●● ●

β−service

● ● ● ● ●●

●●

●● ●● ● ●●

●●●

●

● ● ●

●● ● ● ●●● ● ● ●● ●

●

●●

● ●● ● ● ● ●● ● ●●

●

● ● ● ● level●

●

number of shipments (per period)

● ● ● ●●● ● ● ●

● ●● ●●● ● ●● ● ●

● ● ● ● 100 %

0.20

● ● ● ●● ● ●

●●●

● ● ● ● ●● ●

●

● ●● ● ●

●●●●●

● ● ● ● ● ●●●●● ●

●● ● ●

●● ●● ●●● ●● ● ●

●● ●●●●● ● ●●● ● ● ●●

●

●● ● ● ● ● ● ● ● ●

● ● ● ● ● ● ●●● ● ● ● ●● ● ●● 84 %

●●● ● ●● ● ● ● ● ●

●

●● ● ● ● ● ● ● ●

● ● ●● ● ● ● ● ● ●

● ●● ● ●● ● ● ●

●● ●●● ●●

● ●● ●● ● ● ● ●

●● ● ●● ●

●● ● ● ● ● ● ●

●●●● ●● ● ●● ● ●● ● ●● ●● ●

● ●

● ●● ● ● ● ● ● ● ● 46 %

0.15

●

● ● ● ●● ● ● ●

● ● ● ●● ● ● ● ● ● ●

● ● ●

●

● ●

● ● ● ● ● ● ● ●● ●

●● ●● ●●● ● ● ●● ● ● ● ●

●● ● ●●● ● ●

● ● ● ● ●

●

●● ●●●

● ● ● ●

●● ● ● ● ●

●

●●●●●● ●●● ●● ● ●

● ● ●●

● ●● ● ● ● ● ● ●

● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ●●

● ●●●●●

●●

●●● ●● ● ●●

● ●●●

● ●●

● ●● ● ● ● ●● ●● ●● ●

● ● ● ●● ●

●

●

●

●

●●●

●●●

●●●●●●

● ● ●●●● ● ● ● ●● ● ●

●●●

●●

● ● ●

● ● ● ● ●● ● ●

●

●●●

●

●

●

●

●● ●

● ●● ●

● ●● ● ● ●●

●

● ● ●

● ● ●

●

●●

●

●●

●

●●●

● ●

●●●● ●

●

●●● ● ●●● ● ●

● ● ●●●● ● ●

●●

●● ● ● ● ● ● ● ●

●

●

●●

●

●● ●● ● ● ● ● ● ●

●● ● ●● ●● ● ●●●

0.10

●

●●●

●

● ● ● ●● ●

●

●

●●

●●●●●

● ●● ●●● ● ● ●●

●

●●●

●●

●

●●●●●● ● ● ● ● ● ● ●

●

●

● ●

●●

●●●● ●● ● ● ●

● ●●

● ●● ● ●● ●

● ●●● ●● ●

● ●

●●

●●●

● ●●●● ●

● ● ● ● ● ●● ●

● ● ●

●

● ●●● ● ● ● ●●● ● ●

●● ● ●● ●

●●

●●●●●

●●

●● ● ●● ● ● ● ● ● ●

●●●

●● ●

●●

●●●

●● ● ●

●●

●

●● ●● ● ● ● ● ● ●

● ●●● ● ● ● ● ● ● ●

●●●●● ● ●●● ● ● ● ●

● ● ● ● ●● ● ●

●●●● ●● ● ● ● ● ● ●● ● ●

● ● ●● ●● ●

●

●●●●● ●●

● ●● ● ● ● ●●●●●

●

●●● ● ● ●

●● ●● ●● ●● ●●

●●

●

●● ● ● ● ● ●● ●● ● ●●●●●

●●

●●● ●●●● ● ●● ● ●● ● ●

●

0.05

●●●●●●

●● ●

●

●●●

● ●

● ●● ●●

●

●●●●

● ●●● ●●

average total stock level

which eases the determination of optimal conﬁguration in the following decision support

steps.

Among the three performance measures, the realized β-service level depends on both

other performance measures. Let i be the index of Pareto optimal solutions and deﬁne by

¯li , δi and βi the corresponding performance measures. Then, the following model can be

used to approximate the observed conﬁgurations

βi

log ( ) = μ + γ11 ⋅ ¯li + γ12 ⋅ δi + γ21 ⋅ ¯li2 + γ22 ⋅ δi2 + (4.22)

1 − βi

γ31 ⋅ log(¯li ) + γ32 ⋅ log(δi ) + γ1 ⋅ δi ⋅ ¯li +
i .

The model summary of (4.22), including parameter estimates, can be found in Ta-

ble A.20. The model shows a good overall ﬁt which is indicated by unsuspicious residual

plots (see Figure A.6) and McFadden’s pseudo-R2 of about 0.99. Using this model, the

Pareto front can be described continuously which is depicted in Figure 4.16.

In Figure 4.16 the Pareto front is evaluated over the area spanned by the sample depicted

in Figure 4.15. The grey scale, representing the β-service level, is the same as before.

Additionally, contour lines are superimposed indicating levels of constant β-service level.

Figure 4.16 quantiﬁes the trade-oﬀ between stock holding and the frequency of shipping

w.r.t. the β-service level. The shape of the substitution lines remains stable for all β-

service levels but the substitution area (i.e. the length of the contour line) decreases for

high β-service levels (β > 99.5%). This implies that a decrease in average stock cannot

be compensated arbitrarily by an increase of the shipping frequency. The minimum stock

level to achieve a speciﬁc β-service level forms the upper end of each contour line. For

195

β−service

level

1.0

number of shipments (per period)

0.20

0.9

99

0.9

0.9

0

97

0.9

5

950

0.8

0.9

900

0.15

0.9

750

0.7

0.9

0.5000

500

0.9

000

0.6

0.60

0.10

0.8

00

00

0.7

0

00

0.5

0

average stock level

instance, the average stock level to achieve a β-service level of 95% varies between 20,000

and 38,000 tons approximately.

Based on the individual loss functions depicted in Figure A.5a and Figure A.5b local

re-order levels s1 and s2 of 18,000 and 13,000 tons are indicated, respectively. Setting

the re-order level sh at the harbour to about 42,000 tons (as Figure 4.14b suggests) and

the minimal order-up-to level Sh to 57,000 tons, results in an average stock level of ap-

proximately 50,000 tons accompanied with a β-service level close to 100% and a shipment

frequency of about 0.28 ships per day. As 0.28 exceeds the maximum of Pareto-optimal

solutions, this conﬁguration is clearly ineﬃcient. Obviously, a component-wise setting of

inventory parameters does not lead to a suﬃciently accurately modelled system perfor-

mance. Based on the estimated relation between the performance measures depicted in

Figure 4.16, however, the optimal conﬁguration for given weights of all three measures

can be easily determined.

196

Decision support systems aim at helping the management of an organizational unit to

ﬁnd decisions for planning problems.223 Information about the relevant processes has to

be provided and an evaluation of potential solutions has to be possible.

Various types of DSS can be distinguished of which for the present work mainly model-

driven DSS are relevant. Model-driven DSS make use of a model (optimization, simulation

or statistical) of some (sub)-system of the company to provide decision-relevant data. In

contrast, e.g. knowledge- or communication-driven DSS provide topic-speciﬁc expertises

or support shared working processes.224

Typically, DSS are integrated software packages providing three core components: 1)

a database, 2) a model-base and 3) a user interface.225 The database component is

designed to store and analyse system-relevant data. Statistical analysis routines need

to be applicable to huge volumes of data. Data and its fast analysis show a strong

impact on calculating key performance indicators, e.g. allowing to detect bottlenecks.226

Furthermore, data builds the basis of the modelling component for which suitable model

parameters have to be extracted.

The model base contains programs which form a decision/problem-speciﬁc represen-

tation of the system under study.227 Based on this core model, the performance of the

system can be evaluated under various conﬁgurations. Moreover, this component also

provides programs to facilitate the decision process e.g. by providing tools for criteria

ranking in case of multi-criteria problems228 and tools for analysing a speciﬁc conﬁgu-

ration in higher detail.229 The latter particularly addresses the comparison of multiple

conﬁgurations by a set of performance measures.

Finally, the user interface component allows managers to use the DSS without detailed

knowledge about its internal structure.230 Basically, the user interface eases the data

transfer between users and DSS whereby the users’ inputs reﬂect e.g. parameter ranges and

preferences of criteria. The outputs of a DSS contain a set of solutions (e.g. Pareto-optimal

solutions) in the multi-criteria case or the optimal solution in the single-objective case.

Often information is exchanged in a spreadsheet format but also graphical representations

of solutions are provided.231 In the end, the usability of a software program is crucial for

the implementation of a DSS in a business organization.

To illustrate the post-processing of generated information from a simulation model for

223

See Power (2002, ch. 1).

224

For an encompassing classiﬁcation see e.g. Power (2002, ch. 1).

225

See Power (2002, ch. 1).

226

The fast access to (statistical) analyses of huge databases is often referred to as on-line analytical

processing (OLAP), see e.g. Shim et al. (2002).

227

See Power and Sharda (2007).

228

Such as an analytic hierarchy process, see Power and Sharda (2007).

229

So-called decision analyses, see Power and Sharda (2007).

230

See Power and Sharda (2007).

231

See Shim et al. (2002).

197

lation study either a ﬁnal recommendation (e.g. in the form of an optimal conﬁguration)

or a set of potential alternatives for the system considered has to be provided. In exam-

ple 12 a meta-model for the Pareto-front was ﬁtted which provides the set of potential

system states from which a speciﬁc solution has to be selected. After the management

has selected a speciﬁc solution, the corresponding conﬁguration has to be extracted, ide-

ally without forcing additional computational eﬀort due to additional simulation runs.

From the Pareto-front meta-model, however, the parameter conﬁguration corresponding

to a speciﬁed point cannot be deduced directly. Instead an approach is necessary which

responds quickly to varying preferences.

Let xi denote the ith conﬁguration of the determined sample of the Pareto-front and

yi the associated performance measures such that yi = Ĥ(xi ) where Ĥ(⋅) represents the

simulation model. Hence, the set of zi = (xi , yi ) constitutes a sub-sample of the set of

eﬃcient constellation zi ∈ Z̄ ef f ⊂ Z ef f .232 Based on the range and the relations among

the y-components a Pareto-front meta-model (4.22) is estimated which can be used to

describe the set of eﬃcient constellations Z ef f more precisely.

Let ŷj = (β̂j , δj , ¯lj ) denote an estimated performance vector whereby β̂j = f (δj , ¯lj ) and

f (⋅) denotes the Pareto-front model (4.22). Then, an inverse mapping function Ĥ −1 is

required to obtain the corresponding xj . However, the inverse of a simulation model cannot

be derived directly. Therefore, an approximation of Ĥ −1 is required.

Since simulation models are typically applied when an analytical model is unknown,

ﬁnding an appropriate analytic function that accurately mimics Ĥ −1 can be expected to be

impossible.233 Instead, RSM can be adapted. Originally, RSM is designed to ﬁnd (Pareto-

)optimal conﬁgurations by iteratively approximating the simulation model’s local behaviour

using simple analytic functions (such as linear regression models). The idea is that simple

functions can provide a suﬃciently accurate local approximation of the simulation model’s

behaviour as long as the region covered is pretty small. This idea can also be applied to

the presented problem.

In classic RSM the region for which the simulation model is approximated iteratively

moves through the sample space (controlled by a gradient rule) to explore the performance

function of the system. Here, the region depends on the performance vector to be tested

(ŷ). However, no additional simulation runs shall be performed. Instead, the already

evaluated constellations (xi , yi ) can be analysed. In this set, conﬁgurations with a perfor-

ˆ ˆ

mance similar to ŷ can be used to ﬁt a local approximation of Ĥ −1 (say Ĥ −1 ). Ĥ −1 can

232

Note that in this example only the y-component of an evaluated constellation z determines the eﬃ-

ciency, since all x values are stock control parameters and are ﬁnally reﬂected by the average stock

level which is a part of y.

233

Otherwise, it would be very likely that a closed-form analytic function also exists to approximate Ĥ

which would be a representation of H(⋅) (the real system). Hence, a simulation study would be

superﬂuous and this function would be a better analytic model.

198

ˆ

then be used to estimate the conﬁguration x̂ that corresponds to ŷ by evaluating Ĥ −1 for

ˆ −1

ˆ = Ĥ

ŷ, i.e. x̂ (ŷ). The procedure is brieﬂy described as follows:

ˆ

3. estimate a local (linear) model based on (xi , yi ) = zi ∈ Z̄ŷef f , i.e. xi = Ĥŷ−1 (yi ) +
i

ˆ

4. estimate a desired conﬁguration x̂ by x̂ = Ĥŷ−1 (ŷ).

In the described procedure, two components have to be speciﬁed: First, which type of model

ˆ

to use for the local model Ĥŷ−1 and, second, how to determine Z̄ŷef f . In this example for

ˆ

the ﬁrst component Ĥŷ−1 a simple ﬁrst order linear regression model with interaction terms

is estimated for each performance measure. I.e. for all (xi , yi ) ∈ Z̄ŷef f holds

(dimension 3 × 3), Γ2 is the matrix of ﬁrst-order interaction eﬀects (dimension 3 × 3),

(yi yi ) is the vector of all three ﬁrst-order interactions (i.e. (yi yi ) = (βi ⋅ ¯li , βi ⋅ δi , δi ⋅ ¯li )),

γ is the vector of second-order interaction coeﬃcients, (yi yi yi ) is the scalar of second-

order interaction (i.e. (yi yi yi ) = βi ⋅ ¯li ⋅ δi ), and i is the error vector. Note that for each

performance measure, eight parameters have to be estimated which requires ∣Z̄ ef f ∣ > 8. In

the presented example this restriction was always fulﬁlled.234 The parameters are estimated

by ordinary least squares regression.

To deﬁne a local subset of Pareto-optimal solutions close to ŷ the Delaunay triangulation

is calculated for the set Z̄ = Z̄ ef f ∪ {ŷ}. The Delaunay triangulation subdivides the convex

hull of a set of points into disjunct simplices. Each simplex consists of d + 1 points

whereby d denotes the dimension of the data set.235 A speciﬁc property of the Delaunay

triangulation is that for each simplex the circumhypersphere constituted by its points is

empty which implies that the Delaunay triangulation is unique.236 Let D denote the set of

Delaunay simplices where each simplex dk is a set of d+1 points, i.e. dk = {zk1 , ..., zk(d+1) ∈

Z̄}. Now select the subset D′ ⊂ D whose simplices all contain ŷ: D′ = {dk ∣ ŷ ∈ dk }.

Finally, retain all zi ∈ Z̄ ef f such that ∃d′k ∈ D′ with zi ∈ d′k , i.e. Z̄ŷef f = {zi ∣ ∃d′k ∈

D′ ∶ zi ∈ d′k }. Selecting Z̄ŷef f in this way has the appealing property that neighbouring

points are comparatively equally distributed around ŷ which facilitates an accurate ﬁt

234

If this restriction would be violated, either a simpler model (e.g. by dropping interaction eﬀects) had

to be chosen or the neighbourhood Z̄ ef f had to be enlarged.

235

See Delaunay (1934) for the original paper.

236

As long as the data set is in general position, i.e. no co-circular or degenerated subsets exist, see De Berg

et al. (2008, ch. 9) for a precise deﬁnition.

199

ˆ

of Ĥŷ−1 .237 In the two-dimensional case this fact is supported by the property that the

Delaunay triangulation maximizes the minimum angle of all simplices.238 Alternatively,

the procedure could be brieﬂy described as ﬁnding all adjacent points of ŷ in the Delaunay

graph.239

To illustrate the determination of Z̄ŷef f , Figure 4.17 shows a Delaunay triangulation for

an artiﬁcial data set drawn from a bivariate uniform distribution.

point to be tested ⎛⎝^ sub−sample points ⎛⎝Z ^yef f⎞⎠

10

● ●

●

●

●

● ●●

●

●

●

● ●

●

●

● ●

●

8

●

●

●

●

●

●

● ● ●

●

●

● ●

●

●

6

●

●

●

●

● ●

● ●

●

● ● ●●

●

●

● ●

●

4

●

●

●

●

● ●

●

●

●

●

●

●

## Viel mehr als nur Dokumente.

Entdecken, was Scribd alles zu bieten hat, inklusive Bücher und Hörbücher von großen Verlagen.

Jederzeit kündbar.