This action might not be possible to undo. Are you sure you want to continue?

**Trondheim, October 2007
**

Norwegian University of

Science and Technology

Faculty of Social Science and Technology Management

Department of Industrial Economics and Technology

Management

Kjetil Trovik Midthun

Optimization models for

liberalized natural gas markets

NTNU

Norwegian University of Science and Technology

Thesis for the degree of philosophiae doctor

Faculty of Social Science and Technology Management

Department of Industrial Economics and Technology Management

©Kjetil Trovik Midthun

ISBN 978-82-471-4512-8 (printed ver.)

ISBN 978-82-471-4526-5 (electronic ver.)

ISSN 1503-8181

Theses at NTNU, 2007:205

Printed by Tapir Uttrykk

Acknowledgements

This thesis is the result of my work for the degree of Philosophiae Doctor at the

Department of industrial economics and technology management, at the Norwe-

gian University of Science and Technology (NTNU).

First of all I would like to thank my supervisor Asgeir Tomasgard. His contin-

ued support and guidance has been very important for my work. Some of the best

and most inspiring moments in the work with this thesis were our discussions.

My co-supervisor, Mette Bjørndal, has also been important in the work with this

thesis. I really appreciate the time we spent working together.

From November 2005 to March 2006 I visited Professor Yves Smeers at the

Center for Operations Research and Econometrics (CORE), in Louvain-la-Neuve

(Belgium). Working with Professor Yves Smeers was interesting and inspiring. In

addition, the stay allowed me to meet many nice people at CORE and introduced

me to Belgian culture.

I also want to thank good colleagues at the Department, especially Peter Schütz

who has faced many of the same problems and challenges as I have since we started

on our PhDs back in autumn 2002. I am also grateful to Matthias Nowak for

always taking the time to answer my questions.

My friends in Trondheim have been very important for me in this period. I want

to thank Håvard Berland, Hugo Hammer, Geir Solgaard and Steinar Kragset for

many good memories.

I am grateful to my family at home for always providing support and encour-

agement. I want to thank Astrid for putting things in perspective when my

frustration with the work was at a maximum. Finally I want to thank Katrine

for putting up with me through a lot of hectic periods and for cheering me up

when I am down. I look forward to spending more time with you all in the time

to come.

Kjetil Trovik Midthun

Trondheim, August 2007

iii

Contents

1 Introduction 1

1.1 The natural gas value chain . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Operations research modeling framework . . . . . . . . . . . . . . . 7

1.3 Existing literature and research contribution . . . . . . . . . . . . . 12

1.4 Papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2 Optimization Models for the Natural Gas Value Chain 29

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.2 The Natural Gas Value Chain . . . . . . . . . . . . . . . . . . . . . 30

2.3 A Natural Gas Transportation Model . . . . . . . . . . . . . . . . . 34

2.4 Management of Natural Gas Storages . . . . . . . . . . . . . . . . 48

2.5 Value Chain Optimization and Portfolio Management . . . . . . . 52

2.6 The Expected Gas Value Function (EGV) . . . . . . . . . . . . . . 61

2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

2.A Notation and Deﬁnitions . . . . . . . . . . . . . . . . . . . . . . . . 66

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

3 Modeling optimal economic dispatch and ﬂow externalities 75

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

3.2 Natural gas ﬂows . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

3.3 Max ﬂow and network externalities - two numerical examples . . . 87

3.4 Optimal dispatch with economic objective functions . . . . . . . . 93

3.5 Optimal economic dispatch - two numerical examples . . . . . . . . 97

3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

3.A Linearization of the Weymouth equation . . . . . . . . . . . . . . . 106

3.B Unconstrained equilibrium . . . . . . . . . . . . . . . . . . . . . . . 107

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

4 An operational portfolio optimization model for a natural

gas producer 113

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

4.2 Modeling natural gas networks . . . . . . . . . . . . . . . . . . . . 115

4.3 The portfolio optimization model . . . . . . . . . . . . . . . . . . . 122

4.4 Test case from the Norwegian continental shelf . . . . . . . . . . . 128

v

Contents

4.5 Computational results and discussion . . . . . . . . . . . . . . . . . 134

4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

5 Capacity booking in a Transportation Network with stochastic

demand and a secondary market for Transportation Capacity 147

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

5.2 Problem description and assumptions . . . . . . . . . . . . . . . . 149

5.3 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

5.4 Model properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

5.5 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . 165

5.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172

5.A The equilibrium conditions . . . . . . . . . . . . . . . . . . . . . . 174

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

vi

Chapter 1

Introduction

This thesis presents diﬀerent optimization models for the natural gas value chain.

It focuses on the new challenges faced by the participants in the value chain for

Norwegian gas after the liberalization of the European gas markets. Most of

the models have a producer perspective and are designed to help analyze the

value chain for natural gas and give decision support for the gas producers. The

modelling framework in this thesis consists of linear programming, mixed inte-

ger programming, quadratic programming, stochastic programming and mixed

complementarity problems. The thesis consists of this introductionary summary

and four papers. Some of the work in this thesis has been sponsored by the

VENOGA and the RAMONA project. VENOGA is a project involving Statoil,

the Research Council of Norway, NTNU and SINTEF. The goal of the project

is to build decision support models and competence for eﬃcient operation and

coordination in value chains where Norwegian gas is central. The RAMONA

project involves the University in Stavanger, the University in Bergen, NTNU,

SINTEF, Statoil, Gassco and CognIT. The intention with the project is to de-

velop methods to optimize regularity and security of supply for the Norwegian

gas production- and transportation system.

Part 1 of the thesis consists ﬁve sections. In this ﬁrst section a short intro-

duction to the thesis is given, in section 2 the background for the thesis as well

as a presentation of the natural gas value chain is presented. Section 3 gives

an introduction to the model types I have worked within. In section 4, relevant

literature is presented and I indicate where the papers in this thesis extend the

existing literature. Lastly, in section 5, a summary of the papers included in this

thesis is given.

Part 2 consists of the four papers included in this thesis. The ﬁrst paper, ‘Opti-

mization Models for the Natural Gas Value Chain’, gives a thorough introduction

to the natural gas value chain, and modeling techniques for the diﬀerent parts of

the value chain. The paper is meant as a tutorial on modeling natural gas value

chains.

In the second paper, ‘Modeling optimal economic dispatch and ﬂow exter-

nalities in natural gas networks’, a model for economic analysis in natural gas

transportation networks is presented. The model includes a presentation of the

1

Chapter 1 Introduction

pressure constraints and system eﬀects in natural gas networks; in addition it

uses economic objective functions (such as maximization of social surplus).

The third paper is ‘An operational portfolio optimization model for a natural

gas producer’. The model presented here has a system perspective, where we

assume that all decisions in the value chain is made by a central planner. The

value of actively using the pipelines in the network as storage (line-pack) to

maximize proﬁts is examined using a stochastic model and real market data from

three European market hubs.

The last paper, ‘Capacity booking in a Transportation Network with stochas-

tic demand and a secondary market for Transportation Capacity’, provides an

equilibrium model of the booking system in the North Sea. The eﬀects of diﬀer-

ent objective functions for the network operator are discussed. In addition, the

importance of system eﬀects and stochasticity is analyzed.

1.1 The natural gas value chain

In this section I describe the background for the papers in this thesis and give an

introduction to the natural gas value chain. The focus will be on the North Sea,

and especially on the Norwegian interests.

In 2006, the petroleum industry accounted for approximately 36 % of the total

income for the Norwegian economy. Natural gas is increasing in importance in the

petroleum industry, and the production of natural gas is expected to reach 42%

of total petroleum production in Norway by 2010. In a European context, the

Norwegian production of natural gas is signiﬁcant and accounts for approximately

15% of the natural gas consumption. Most of the gas is transported to Germany

and France, where Norwegian gas accounts for approximately 30% of the total

consumption.

The liberalization process

An important part of the background for this thesis is the liberalization process

in the natural gas industry in Europe. The process has been ongoing for more

than twenty years, starting in Great Britain. Great Britain is also today the

country that is most advanced in the liberalization process, measured by market

opening and liquidity in the short-term markets. For a discussion of the lib-

eralization process in Great Britain, see Roeber (1996) and Weir (1999). Also

in other landing points for natural gas in Europe, such as Zeebrugge and the

TTF, short-term markets have emerged. Financial markets with natural gas as

the underlying commodity is developing in these market hubs. This indicates a

growing trust in the liquidity of the spot markets. Neumann et al. (2006) gives

an example of econometric analysis of the convergence of European spot market

2

1.1 The natural gas value chain

prices for natural gas. The paper shows that the Interconnector (a pipeline which

connects the market hubs in NBP and Zeebrugge) has led to almost perfect price

convergence between the NBP (National Balancing Point in Great-Britain) and

Zeebrugge in the time period considered in the paper.

The liberalization process has to a large degree been driven by the European

Union. This is in contrast to the process in the US and UK which were mainly

market driven. The European Union has decided on two gas directives that

speciﬁes the development of one internal European market (European Union

1998, 2003, European Commission & Transport 2002). The most important part

of the directives is the decision to open transportation infrastructure for third

party access.

Implications for the value chain in the North Sea

The supply side in the North Sea is dominated by large companies. Before the

liberalization process, coordination of production and transportation on the Nor-

wegian continental shelf was accomplished through inter company groups like the

Gas Supply Committee (Forsyningsutvalget, FU) and the Gas Negotiating Com-

mittee (Gassforhandlingsutvalget, GFU). In 2001 the GFU was abandoned and

replaced by individual company sales. The large production companies also own

the transportation infrastructure. After the liberalization process, the ownership

of the infrastructure was given to a newly formed company: Gassled (owned by

the original owners of the infrastructure). Gassled has ownership of all infrastruc-

ture open for third party access. New infrastructure facilities will be incorporated

in Gassled when used by a third party.

The routing in the network in the North Sea is the responsibility of Gassco.

Gassco is an independent company responsible for ensuring nondiscriminatory

access to the infrastructure owned by Gassled. The tariﬀs in the network are

regulated by the ministry of petroleum and energy. The intention is to ensure that

the proﬁts are generated in the production ﬁelds and not in the transportation

network. The access to the network is decided by allocation rules (Ability to

Use and Capacity Allocation Key). For more information on the tariﬀ system,

see Gassco (2006). For details on the liberalization process in Norway, see Austvik

(2003) and Dahl (2001).

Elements in the natural gas value chain

Natural gas is formed naturally from organic material: plant and animal remains.

Subjected to high pressure and temperature over millions of years, the organic

material changed into coal, oil and natural gas. Natural gas is a mixture of

hydrocarbon and non-hydrocarbon gases (such as helium, hydrogen sulﬁde, and

3

Chapter 1 Introduction

nitrogen). The gases are found in porous geological formations (reservoirs) be-

neath the earth’s surface. In these reservoirs, the gas can be in gaseous phase or

in solution with crude oil. Unlike other fossil energy sources, the natural gas is

a relatively clean fuel, meaning that it emits lower levels of harmful byproducts

when burnt than for instance oil and coal. Useful sites for information on natural

gas are NGSA (2007), Gassco (2006) and EIA (2006).

A simpliﬁed picture of the oﬀshore natural gas value-chain in the North Sea

is shown in Figure 1.1. The gas is transported from the production ﬁelds to

processing plants, or directly to the market hubs in Europe. There are storage

possibilities along the transportation route. In addition, the transportation net-

work itself can be considered as a storage facility since there are large volumes

of gas contained in the pipelines at all times. In the following I will go shortly

through some of the important characteristics of the natural gas value chain.

Figure 1.1: Illustration of the natural gas value chain.

Exploration

Before any gas is produced, the gas must be located and wells must be drilled.

A common method for exploration on the North-Continental shelf is seismology.

4

1.1 The natural gas value chain

The methods have been developed from simple two-dimensional seismology in the

early 1970s to the latest technology where four dimensions are being recorded

(the fourth being shear waves). In addition to seismology, methods such as

magnetometers and gravimeters can also be used. The magnetometers measures

small diﬀerences in the Earth’s magnetic ﬁeld, while the gravimeters measure

the Earth’s gravitational ﬁeld. In addition, mathematical models are used to

predict the underground geological structures and conditions. The models give a

hypothetical picture of the subsurface.

The procedures developed so far can only indicate where the gas deposits are

located. Exploratory wells must be drilled in order to prove the existence and

actual characteristics of the deposits. The cost of drilling such exploratory wells

is high, and the location of these wells is important for further development of

the ﬁelds. After a well is drilled, new information is available to the developers.

The new information will inﬂuence the optimal total allocation of wells on the

ﬁeld.

Production and processing

The gas is produced from the reservoirs. The driving forces are pressure from

the expanding gas as well as water which causes the gas to ﬂow into the well.

The gas production depends on the pressure in the reservoirs. High pressure in

the reservoir gives a high production rate. In order to increase the pressure in

the reservoir, and thus increase production capacity, compressors are sometimes

used.

The natural gas sold to Europe consists mainly of methane (dry gas). The gas

produced at the ﬁelds can however contain other components with market value,

such as associated hydrocarbons (for instance ethane, propane and butane). Gas

containing both dry gas and associated hydrocarbons is called rich gas. The

rich gas is transported to processing plants where the dry gas and wet gas (the

associated hydrocarbons) are separated. The wet gas is then heated in order to

separate the diﬀerent components, which in turn is sold in component markets.

Modeling of processing plants is not within the scope of this thesis, but the

interested reader can consult Bullin (1999) for examples.

Figure 1.2 illustrates the production of gas in Norway and in the world in

total. As we can see from this ﬁgure, the Norwegian production has increased

drastically the last ten years.

Transportation

In the North Sea, the gas is transported from the production ﬁelds oﬀshore to

processing plants on the Norwegian mainland or to market hubs on the European

mainland and in Britain through long, sub sea pipelines operated at high pressure

5

Chapter 1 Introduction

Figure 1.2: The development of gas production in Norway and in the world in

total.

levels. The gas molecules ﬂow in the pipeline from high pressure points to low

pressure points. At the production ﬁelds, pressure is increased with compressors

in order to create a pressure diﬀerence that is suﬃcient for the gas to ﬂow to the

landing points. With the completion of Langeled (pipeline), the network consists

of 7800 km of pipelines. For details on the infrastructure and topology in the

North Sea, see OED (2006).

Natural gas ﬂow is a multi-commodity ﬂow, where the diﬀerent components

have diﬀerent market value. The components give the gas ﬂow diﬀerent proper-

ties, and factors such as caloriﬁc value and corrosion of pipelines depend on the

mixture of components in the ﬂow.

Storage

The demand for natural gas shows strong seasonal patterns and large short-term

volatility (day-to-day variation in the prices). Both these factors give a large value

to optimal storage utilization. Since the sale of natural gas is limited both by

production capacity in the ﬁelds and the transportation network connecting the

ﬁelds to the market nodes, there is a large value in storing gas close to the market

nodes in the summer (when demand is low) in order to sell more gas in winter

(when demand is high). In the same manner, gas can be injected to the storages

during days with low price / low demand and extracted from the storage when

the price is favorable. There are many diﬀerent forms of storages that are used

for storing natural gas: abandoned oil- and gas ﬁelds, aquifers, LNG-storages and

salt caverns. The storages are diﬀerent with respect to capacity, injection- and

extraction capabilities and cost of operating. For more information on storages,

see EIA (2002). In addition, the pipeline network can also be used as storage

(line-pack). This is due to the fact that there are large volumes of natural gas

6

1.2 Operations research modeling framework

contained in the pipelines. This volume is needed for gas to ﬂow in the pipelines.

By injecting more gas into the network than is extracted in a given time period,

the line-pack is increased. If however more gas is extracted than injected, the

line-pack is decreased.

Markets

Traditionally, the gas from the North Sea has been sold in long-term take-or-

pay contracts (TOP). In the TOP-contracts, the price is determined based on a

formula which main components are the prices of competing fuels (for instance

oil). A yearly volume is decided, and then the buyers have ﬂexibility with respect

to nomination on shorter time-periods (within certain limits). If the total volume

during a year is lower than the agreed upon volume, the buyers still pays for the

agreed volume. Originally, these contracts where established for ﬁeld and market

combinations. This means that it was determined which ﬁeld should deliver in

which contract. After the liberalization process in Norway, this has changed, and

the companies are now free to deliver the contracted volumes from the ﬁeld of

their choice.

In the market hubs, short-term markets have emerged. The price in the short-

term markets is volatile. The markets have so far had low liquidity, but the

situation is improving. The increasing liquidity in the markets is indicated by the

development of forward and future contracts with natural gas as the underlying

commodity. For a discussion of the development of the spot market for natural gas

in the UK, see Roeber (1996). For a discussion of price models for commodities,

see for instance Schwarz (1997).

Remaining reserves

No one knows exactly how much gas is left in the ground for us to use, but we

do have some estimates. The volume of proved reserves has increased over the

years. Today the reserves are suﬃcient for approximately 65 years (given that

today’s production level is kept constant). If we look at Norway, the reserves

to production ratio is approximately 30 (see Figure 1.3). All data used in this

section is provided by BP (2006).

1.2 Operations research modeling framework

This thesis is within the ﬁeld of operations research (OR). Operations research

can be deﬁned as an interdisciplinary science which uses quantitative methods

to support decision making. In this thesis the main focus is on mathematical

programming, which can be deﬁned as the study of problems where one seeks to

7

Chapter 1 Introduction

Figure 1.3: Reserves to production ratio for Norway.

maximize or minimize a real function by choosing the values of real or integer

variables from an allowed set.

In the following I will give a short introduction to the model types that I have

worked with in this thesis. The overview is meant for the non-expert reader and

it does not give a comprehensive introduction to the model types. References to

literature where more information can be found is given for each model type.

An introduction to the ﬁeld of operations research can be found in textbooks

such as, for instance Hillier & Lieberman (2001). Webpages, such as the home-

page of the Institute for Operations Research and the Management Sciences (IN-

FORMS 2007), provides information and overviews of available resources in the

ﬁeld of operations research.

Linear programming (LP)

In linear programming, a linear objective function is optimized (maximized or

minimized) over a convex polyhedron speciﬁed by linear and non-negativity con-

straints. George B. Dantzig is often considered as the founder of LP (Dantzig

1949, Wood & Dantzig 1949). Other important contributors are, for instance,

John von Neumann and Leonid Kantorovich. For a nice overview of the history

of the development of Linear Programming, see Dantzig (1991).

The general LP problem can be formulated in the following way:

max c

T

x (1.1)

s.t. Ax ≤ b (1.2)

x ≥ 0 (1.3)

The decision variables in the problem is given by x. c

T

, A and b are parameters.

Equation (1.1) is the objective function, while Equation (1.2) gives the constraint

on the decision variables x. Equation (1.3) is the non-negativity constraints on

8

1.2 Operations research modeling framework

x. Equations (1.2) and (1.3) deﬁne the feasible set for x. The x variables could

also have been unrestricted in sign.

Mixed integer programming (MIP)

When some of the variables in the optimization problem are required to take

an integer value, the model type is called Mixed Integer Programming. If all

variables in the problem are restricted to be integer, the model type is called

Integer Programming (IP). IP and MIP problems are by deﬁnition non-convex

problems. The simplex method can not be used directly on this problem class.

A nice introduction to integer programming is given in Wolsey (1998).

An example of a MIP problem is given by:

max c

T

x +d

T

y (1.4)

s.t. Ax +Dy ≤ b (1.5)

x ≥ 0 (1.6)

y integer (1.7)

The decision variables are given as x and y, where y is required to be integer. c

T

,

d

T

, A, D and b are parameters. The objective function is given by Equation (1.4).

Equation (1.5) is a constraint in the problem, and Equations (1.6) and (1.7) gives,

respectively, the non-negativity constraint on x and the integer constraint for y.

Equations (1.5)-(1.7) deﬁne the feasible set for x and y.

Quadratic programming (QP)

Quadratic programming is a special case of non-linear programming, where the

objective function is quadratic while the constraint set consists of linear equa-

tions. Since the quadratic programs have much in common with the LP programs,

powerful dual and complementarity slackness properties allow specialized algo-

rithms for many cases. Examples of textbooks with more information on QP

are Bazaraa et al. (1993) and Hillier & Lieberman (2001).

An example of a QP problem is given by:

max c

T

x +x

T

Qx (1.8)

s.t. Ax ≤ b (1.9)

x ≥ 0 (1.10)

The decision variables are given by x, while c

T

, Q, A and b are parameters.

Equation (1.8) gives the objective function for the problem, while Equations (1.9)

9

Chapter 1 Introduction

and (1.10) deﬁne the feasible set for x. The second paper in this thesis use

quadratic programming to represent economic objective functions (social surplus,

producer surplus and consumer surplus) in situations where there are supply and

demand curves present in, respectively, production nodes and market nodes.

Stochastic programming (SP)

In many situations uncertainty is an important characteristic of the problem we

try to model. Examples are uncertainty in prices and demand for a commodity.

Stochastic programming is a problem class that allows the modeler to take the

uncertainty into account when building the model. This way, the model will put a

value on ﬂexibility in decisions. When everything is known for certain, ﬂexibility

is of no value (you do not have to change decisions since you know what will

happen).

The ﬁrst papers on SP were Dantzig (1955) and Beale (1955). Examples of

textbooks that give a thorough introduction to the ﬁeld of stochastic program-

ming are Kall & Wallace (1994) and Birge & Loveaux (1997). In addition Higle

(2005) gives a nice tutorial on the ﬁeld. The home page of the Stochastic Pro-

gramming Community (COSP 2007) has a lot of resources available within the

ﬁeld of SP.

An often used approach to represent the uncertainty in the model is a scenario

tree, see for example (Birge & Loveaux 1997). In the scenario tree, a discrete

representation of the uncertainty is created. There are many diﬀerent ways of

obtaining this representation, for a nice discussion, see Kaut & Wallace (2007).

Figure 1.4 shows an example of a scenario tree. In each leaf node in the tree there

is a realization of the stochastic parameter (for instance price). In addition, there

is a set of decision variables in each node of the scenario tree. The decisions in the

root node are the same for all scenarios, and are called the ﬁrst-stage decisions.

In the second stage, the decisions will depend on the realization of the stochastic

parameters.

An example of a SP problem is given by:

max c

T

x +

¸

s∈S

p

s

d

T

s

y

s

(1.11)

s.t. Ax = b (1.12)

T

s

x +W

s

y

s

= h

s

, s ∈ S (1.13)

x ≥ 0 (1.14)

y

s

≥ 0, s ∈ S (1.15)

The ﬁrst stage decision variables are given by x. In each scenario s in the set

10

1.2 Operations research modeling framework

Figure 1.4: An example of a scenario tree.

of scenarios S there is a decision variable y

s

. The objective function, in Equa-

tion 1.11, gives the proﬁt from the ﬁrst stage decisions and the expected proﬁts

for the second stage decisions. The probability for each scenario is given by p

s

.

The feasible set for the problem is deﬁned by Equations (1.12)-(1.15).

Mixed complementarity problems (MCP)

Mixed complementarity problems refer to a wide range of problems where the

deﬁning equations consist of both complementarity conditions and equality con-

straints. For a nice introduction to complementarity problems, see Billups &

Murty (2000) and Cottle et al. (1992). In the last paper in this thesis we present

a linear mixed complementarity problem.

An example of a linear mixed complementarity problem is given by:

a +Au +Cv = 0 (1.16)

b +Du +Bv ≥ 0 (1.17)

v ≥ 0 (1.18)

v

T

(b +Du +Bv) = 0. (1.19)

The decision variables are given by u and v. A, C, b, D and B are parameters.

Equation (1.19) is the complementarity condition in this problem. The product

11

Chapter 1 Introduction

of variable v and the left hand side in constraint (1.17) must be zero. Equa-

tions (1.17), (1.18) and (1.19) gives a linear complementarity problem (LCP).

By including variable u and constraint (1.16) we get a mixed complementarity

problem.

The KKT-conditions of an optimization problem can be formulated as a mixed

complementarity problem. By aggregating the KKT-conditions of several players

(several optimization problems) in a multiplayer game, equilibrium problems can

be solved. The MCPs can be used to ﬁnd Nash equilibria (Nash 1950), and

Generalized Nash Equilibria (Debreu 1952, Arrow & Debreu 1954). In a Nash

Equilibrium no player has incentive to deviate from his decisions given that the

other players do not deviate. Generalized Nash Equilibria is found in the situation

where the players can inﬂuence the feasible region of each others’ optimization

problems.

1.3 Existing literature and research contribution

In the following I will give a short introduction to the literature on diﬀerent

aspects of the natural gas, and in addition indicate where this thesis extends the

existing literature.

The petroleum industry has been a pioneer in the application of operations re-

search, and the literature is therefore extensive. In Bodington & Baker (1990), an

interesting overview of the history of mathematical programming in the petroleum

industry is given.

Investment models

In this model class, the goal is to give decision support for strategic decisions such

as ﬁeld investments and sequencing of investments. There are a large number of

publications within this ﬁeld. This is not surprising given the large risks and

costs associated with oﬀshore investments.

There exist a number of deterministic investment models. Sullivan (1988)

presents some applications of Mathematical Programming methods to invest-

ment problems in the petroleum industry. A new MIP model with a detailed

description of reservoir production is also presented in the paper. In Haugland

et al. (1988) existing models for early evaluations of petroleum ﬁelds are pre-

sented. The paper also presents a MIP model which proposes platform capacity,

where and when wells should be drilled and production from the wells. Nygreen

et al. (1998) presents a MIP model used by The Norwegian Petroleum Direc-

torate. The model is a multiperiod model and is used for investment planning

for ﬁelds in the North Sea which contain a mixture of oil and gas. In van den

Heever & Grossmann (2001) a model for design and planning of oﬀshore ﬁeld

12

1.3 Existing literature and research contribution

infrastructures is presented. The model is a multiperiod mixed-integer nonlinear

programming model and incorporates complex ﬁscal rules, such as tariﬀ, tax and

royalty calculations. The net present value of projects is discussed in light of the

ﬁscal rules.

There are also some models which incorporate uncertainty. Jørnsten (1992)

presents an integer model for sequencing oﬀshore oil and gas ﬁelds, where the

objective is to maximize total economic beneﬁt. In addition to the deterministic

model, a stochastic version with uncertainty in future demand for natural gas

is presented. In Haugen (1996) a stochastic dynamic programming model is

constructed to analyze a supplier’s problem of scheduling ﬁelds and pipelines in

order to be able to meet contractual agreements. The uncertainty in the model

is in the resources (production proﬁles). In Jonsbraten (1998) a stochastic MIP

model for optimal development of an oil ﬁeld is presented. The objective of

the model is to maximize the expected net present value of the oil ﬁeld given

uncertain future oil prices. Goel & Grossmann (2004) presents a stochastic MIP

model for planning of oﬀshore gas ﬁeld developments. The expected net present

value is maximized under uncertainty in reserves.

Value chain models

The upstream value chain of natural gas consists of several components, such

as production, transportation, processing, storage and markets. Because of the

special properties of the transportation network a value chain approach to op-

timizing the system is important. In the value chain approach, the complete

network is considered and optimized simultaneously. The value chain approach

has become even more valuable after the liberalization process, which meant an

increase in ﬂexibility for the participants in the value chain.

In Ulstein et al. (2007) planning of oﬀshore petroleum production is studied on

a tactical level. The model has a value chain approach where production plans,

network routing, processing of natural gas and sales in the markets is considered.

In addition, multi-commodity ﬂows and quality restrictions in the markets are

considered. The pressure constraints in the network is however not included in

the model. The non-linear splitting for chemical processing is linearized with

binary variables. The resulting model is mixed integer programming model.

Selot et al. (2007) presents an operational model for production and routing

planning in the natural gas value chain. The model combines a detailed infras-

tructure model with a complex contractual model. There is no market for natural

gas included in the model. The infrastructure model includes non-linear equa-

tions for relating pressure and ﬂow in wells and pipelines, multi-commodity ﬂows

and contractual agreements in the market nodes (delivery pressure and quality of

the gas). The contractual model is based on a set of logical conditions for produc-

13

Chapter 1 Introduction

tion sharing and customer requirements. The combined model is a mixed integer

nonlinear programming model (MINLP). In addition, the model is non-convex

due to the pressure-ﬂow relationship and the modeling of multi-commodity ﬂows.

In the ﬁrst paper in this thesis, we present a tutorial for modeling of the natural

gas value chain. This is the ﬁrst publication (to our knowledge) which presents

a portfolio optimization model with markets, contracts, production planning,

multi-commodity ﬂow and handling of the pressure constraints in the transporta-

tion network. The framework for analysis presented in this model is primarily

aimed at a tactical level. In the third paper in this thesis, we present an opera-

tional stochastic model for portfolio optimization. In this paper we include the

storage in the pipelines in the transportation network (line-pack) in the analysis.

By using real market data from three European hubs, we give an estimate of the

value of actively using the line-pack to maximize proﬁts for the value-chain.

Transportation models

The transportation of natural gas is one of the key elements when studying the

natural gas industry. Because of the interdependence between ﬂows in pipelines,

it is important to ﬁnd a tradeoﬀ between accurately describing the properties

of the transportation network, and being able to solve the model. A simpliﬁed

representation leads to an inaccurate model of the transportation (and may lead

to wrong conclusions), while a too detailed presentation makes the model non-

linear and non-convex. In the following some examples of steady-state models,

as well as more technical models and simpliﬁed economical models are given.

De Wolf & Smeers (2000) presents a model for optimizing gas ﬂow through a

network, with cost minimization. The ﬂow in the network is steady-state, and

the resulting problem is solved by an extension to the simplex algorithm. Also

in O’Neill et al. (1979) a steady-state representation of the gas ﬂow is used in a

model for allocation of natural gas. Westphalen (2004) gives a nice presentation

of stochastic optimization in gas transportation. The model presented in Selot

et al. (2007) provides an accurate description of the steady-state ﬂow using a non-

convex MINLP model. In the ﬁrst paper in this thesis we present a linearization of

the Weymouth equation which enables analysis of large networks and stochastic

problems.

There are a large number of publications with a technical approach to gas

transportation. The models are detailed and accurate in their description of

the physics of gas transportation, such as transient ﬂow and interaction with

compressors. A discussion of transient ﬂows is given in Kelling et al. (2000),

while the homepage of the Pipeline Simulation Interest Group (2007) gives a

comprehensive overview on modeling, simulation and optimization of natural gas

ﬂows. In Ehrhardt & Steinbach (2005) a model for operational planning in natural

14

1.3 Existing literature and research contribution

gas networks is presented. A transient ﬂow model is used to control the network

load distribution for the next 24 to 48 hours. Martin et al. (2006) presents a model

to optimize ﬂow in a network, and minimize the costs of the compressors in the

network. The model gives a detailed representation of the physical properties

of natural gas transportation, and oﬀers linearization techniques for the non-

linearities in the model. Nowak & Westphalen (2003) presents a linear model for

transient ﬂow modeling.

There also exist some economical models with a simpliﬁed representation of

the transportation networks. In these models, the capacities in the pipelines is

normally represented with a ﬁxed, static capacity limit. Examples of such models

are Cremer & Laﬀont (2002) and Cremer et al. (2003). With the simpliﬁed

representation of the transportation network, the system eﬀects of natural gas

transportation are neglected.

In the second paper in this thesis, we show that it is diﬃcult, if not impos-

sible, to determine appropriate static capacities in a natural gas network. We

discuss the system eﬀects in natural gas networks and provide a framework for

economic analysis in natural gas networks. The paper uses the linearization of

the Weymouth equation presented in paper one, and in addition it uses economic

objective functions such as maximization of social surplus. This is the ﬁrst ex-

ample of economic analysis in a gas transportation network where the ﬂows are

determined based on pressure constraints. The discussion of system eﬀects is

similar to the discussion of externalities in the electricity networks. For a nice

discussion of externalities in electricity networks, see Wu et al. (1996).

Equilibrium models

Equilibrium models are used to study situations where more than one player acts

strategically. The models are formulated as complementarity problems. A nice

overview of complementarity problems in natural gas markets is given in Gabriel

& Smeers (2005). The paper both gives a survey of some of the existing models,

as well as develops relevant models for the restructured natural gas markets.

In Wolf & Smeers (1997) a stochastic version of the Stackelberg-Nash-Cournot

model is presented. A market leader is deciding on his production level under

uncertainty in demand, the followers then reacts to the production level after the

uncertainty is resolved. The model is used on the European natural gas market.

In Boots et al. (2004) the downstream market for natural gas in Europe is studied

in a successive oligopoly approach. The players in the network include upstream

producers and downstream traders. A mixed nonlinear complementarity prob-

lem (NCP) to study natural gas markets is presented in Gabriel et al. (2005).

The model includes producers, storage reservoir operators, peak gas operators,

pipeline operators and consumers. The KKT conditions are used to formulate the

15

Chapter 1 Introduction

NCP model. Zhuang & Gabriel (2006) presents a stochastic equilibrium model

for deregulated natural gas markets. The ﬁrst stage decisions in the MCP model

are commitments in long-term contracts, while the second-stage decisions are

spot market activities.

In the last paper in this thesis, we examine the booking procedure in the

North Sea as a stochastic mixed complementarity problem. We formulate the

problem as a Generalized Nash game and we then use theory from variational

inequality to show existence of solution and to solve our problem. This is the

ﬁrst study of a booking system similar to the one implemented in the North Sea

and, to our knowledge, also the ﬁrst study of a booking system in a natural gas

transportation network. It is also one of few examples of applications of stochastic

mixed complementarity problems.

16

1.4 Papers

1.4 Papers

In the following I will give a short presentation of the papers, as well as a de-

scription of my contribution on each of them.

Paper 1: Optimization Models for the Natural Gas Value

Chain

The paper gives an overview of the modeling of a natural gas value chain, and

is meant to be a tutorial on the subject. The importance of the value chain

perspective as well as the portfolio perspective is discussed in the paper. The

model includes spot markets, forward markets, long-term contracts, production

plans, storage utilization, pressure constraints, compressors and multi-commodity

ﬂows. A linearization of the Weymouth equation, which allows studies of large

scale networks, is presented. No numerical examples are presented in the paper.

The content is based on the experience from work done in SINTEF and NTNU.

The modeling and analysis in the paper has been done on SINTEF and NTNU.

My contribution is in structuring and writing large parts of the paper.

The paper is published in G. Hasle, K.-A. Lie, E. Quak (eds.): Geometric Mod-

elling, Numerical Simulation and Optimization, Springer Verlag, 2007. Minor

changes in the references have been made in the version included in this thesis.

Co-authors: Senior Researcher at SINTEF, Frode Rømo, Researcher at SINTEF,

Marte Fodstad and my supervisor, Associate Professor Asgeir Tomasgard.

Paper 2: Modeling optimal economic dispatch and ﬂow

externalities in natural gas networks

In this paper we combine the modeling framework developed in paper 1 with eco-

nomic analysis. In the existing literature there are a number of economic models

that disregard the system eﬀects in natural gas transportation networks. Also,

there exist a number of more technical models without economic analysis. In this

paper we combine the two approaches, and provide a framework for economic

analysis in natural gas transportation networks. We also examine the eﬀects of

ignoring the system eﬀects when doing economic analysis.

I have done the implementation of the model. In addition, I have had an equal

part in modeling, the analysis and in writing the paper.

17

Chapter 1 Introduction

The paper is submitted to an international journal.

Co-authors: co-supervisor, Associate Professor Mette Bjørndal and my super-

visor, Associate Professor Asgeir Tomasgard.

Paper 3: An operational portfolio optimization model for a

natural gas producer

In this paper we present an operational optimization model for a natural gas

producer. The model has a system perspective. We use the modeling framework

presented in paper 1, and extends the model with pipeline storage (line-pack).

To our knowledge, this is the ﬁrst study of an operational stochastic portfolio

optimization model for natural gas production and sales. We provide numerical

examples based on real market data from three European hubs. Especially, we

evaluate the commercial value of actively using the line-pack in the pipelines to

maximize proﬁts for the producers. We also examine the value of using a stochas-

tic model compared to a deterministic model.

I have done the implementation of the model. In addition, I have had an equal

part in modeling, the analysis and in writing the paper.

The paper is submitted to an international journal.

Co-authors: Post doc. Matthias P. Nowak and my supervisor, Associate Pro-

fessor Asgeir Tomasgard.

Paper 4: Capacity booking in a Transportation Network with

Stochastic Demand and a Secondary Market for

Transportation Capacity

In this paper we study allocation of transportation capacity in a system that re-

sembles the one implemented in the North Sea. In papers 1, 2 and 3 we have used

a system perspective on the value chain, but now we study the situation when

more than one player is making decisions in the value chain. We look at diﬀerent

objective functions for the network operator, discuss the importance of modeling

the pressure constraints in the network, and look at the eﬀects stochasticity has

on the solutions. The model is formulated as a Generalized Nash game. To our

knowledge, this is the ﬁrst time a booking system in a natural gas transportation

network is studied using this approach.

I have done the implementation of the model. In addition, I have had an equal

18

1.4 Papers

part in modeling, the analysis and in writing the paper.

Co-authors: co-supervisor, Associate Professor Mette Bjørndal, Professor Yves

Smeers and my supervisor, Associate Professor Asgeir Tomasgard.

19

Bibliography

Arrow, K. J. & Debreu, G. (1954), ‘Existence of an equilibrium for a competitive

economy’, Econometrica 22, 265–291.

Austvik, O. G. (2003), Norwegian Natural Gas. Liberalization of the European

Gas Market, Europa-programmet, Oslo, Norway.

Bazaraa, M. S., Sherali, H. D. & Shetty, C. M. (1993), Nonlinear Programming,

Theory and Algorithms, John Wiley & Sons, Inc., New York.

Beale, E. M. L. (1955), ‘On minimizing a convex function subject to linear in-

equalities’, Journal of the Royal Statistical Society, Series B 17, 173–184.

Billups, S. C. & Murty, K. G. (2000), ‘Complementarity problems’, Journal of

Computational and Applied Mathematics 124, 303–318.

Birge, J. R. & Loveaux, F. V. (1997), Introduction to Stochastic Programming,

Springer.

Bodington, C. E. & Baker, T. E. (1990), ‘A history of mathematical programming

in the petroleum industry’, Interfaces 20, 117–127.

Boots, M. G., Rijkers, F. A. M. & Hobbs, B. F. (2004), ‘Trading in the down-

stream european gas market: A successive oligopoly approach’, The Energy

Journal 25.

BP (2006), ‘British petroleum homepage’. www.bp.com.

URL: www.bp.com

Bullin, K. A. (1999), Economic Optimization of Natural Gas Processing Plants

Including Business Aspects, PhD thesis, Texas A&M University.

COSP (2007), ‘Stochastic programming community’, Home page.

Cottle, R., Pang, J. & Stone, R. (1992), The linear complementarity problem,

Academic press, Boston.

Cremer, H., Gasmi, F. & Laﬀont, J. (2003), ‘Access to pipelines in competitive

gas markets’, Journal of Regulatory Economics 24(1), 5–33.

21

Bibliography

Cremer, H. & Laﬀont, J. (2002), ‘Competition in gas markets’, European Eco-

nomic Review 46, 928–935.

Dahl, H. J. (2001), Norwegian Natural Gas Transportation Systems. Operations

in a Liberalized European Gas Market, PhD thesis, NTNU, Trondheim, Nor-

way.

Dantzig, G. B. (1949), ‘Programming of interdependent activities. II. mathemat-

ical model’, Econometrica 17, 200–211.

Dantzig, G. B. (1955), ‘Linear programming under uncertainty’, Management

Science 1, 197–206.

Dantzig, G. B. (1991), Linear programming, in J. K. Lenstra, ed., ‘History of

Mathematical Programming: A Collection of Personal Reminiscences’, North-

Holland Publishing Co, Amsterdam, The Netherlands.

De Wolf, D. & Smeers, Y. (2000), ‘The gas transmission problem solved by an

extension of the simplex algorithm’, Management Science 46(11), 1454–1465.

Debreu, G. (1952), A social equilibrium existence theorem, in ‘Proceedings of the

National Academy of Sciences’.

Ehrhardt, K. & Steinbach, M. (2005), Nonlinear optimization in gas networks,

in H. E. Bock, ed., ‘Modeling, Simulation and Optimization of Complex Pro-

cesses’, Springer-Verlag, Berlin - Heidelberg - New York, pp. 139–148.

EIA (2002), ‘The basics of underground natural gas storage’.

http://www.eia.doe.gov/pub/oil_gas/

natural_gas/analysis_publications/storagebasics/storagebasics.pdf.

EIA (2006), ‘Energy information administration homepage’.

URL: http://www.eia.doe.gov/

European Commission, D.-G. f. E. & Transport (2002), ‘Opening up to choice:

Launching the single european gas market’, Oﬃce for Oﬃcial Publications of

the European Communities.

European Union (1998), ‘Directive 98/30/EC of the european parliament and of

the council’.

European Union (2003), ‘Directive 2003/55/EC of the european parliament and

of the council’.

Gabriel, S., Kiet, S. & Zhuang, J. (2005), ‘A mixed complementarity-based equi-

librium model of natural gas markets’, Operations Research 53(5), 799–818.

22

Bibliography

Gabriel, S. & Smeers, Y. (2005), ‘Complementarity problems in restructured

natural gas markets’, CORE Discussion Paper No. 2005/37.

Gassco (2006), ‘Gassco homepage’. www.gassco.no.

URL: www.gassco.no

Goel, V. & Grossmann, I. E. (2004), ‘A stochastic programming approach to

planning of oﬀshore gas ﬁeld developments under uncertainty in reserves’, Com-

puters & Chemical Engineering 28, 1409–1429.

Haugen, K. K. (1996), ‘A stochastic dynamic programming model for scheduling

of oﬀshore petroleum ﬁelds with resource uncertainty’, European Journal of

Operational Research 88, 88–100.

Haugland, D., Åsa Hallefjord & Asheim, H. (1988), ‘Models for petroleum ﬁeld

exploitation’, European Journal of Operational Research 37, 58–72.

Higle, J. L. (2005), Stochastic programming: Optimizatin when uncertainty mat-

ters, in P. Gray, ed., ‘TutORials in Operations Research’, INFORMS.

Hillier, F. S. & Lieberman, G. J. (2001), Introduction to operations research,

McGraw-Hill Book Co, Singapore.

INFORMS (2007), ‘Institute for operations research and the management sci-

ences’, Home page. http://www.informs.org/.

Jonsbraten, T. W. (1998), ‘Oil ﬁeld optimization under price uncertainty’, The

Journal of the Operational Research Society 49, 811–818.

Jørnsten, K. O. (1992), ‘Sequencing oﬀshore oil and gas ﬁelds under uncertainty’,

European Journal of Operational Research 58, 191–201.

Kall, P. & Wallace, S. W. (1994), Stochastic Programming, John Wiley & Sons,

Chichester.

Kaut, M. & Wallace, S. W. (2007), ‘Evaluation of scenario generation methods

for stochastic programming’, Paciﬁc Journal of Optimization 3, 257–271.

Kelling, C., Reith, K. & Sekirnjak, E. (2000), A practical approach to transient

optimization for gas networks, Technical report, PSIG.

Martin, A., Möller, M. & Moritz, S. (2006), ‘Mixed integer models for the sta-

tionary case of gas network optimization’, Mathematical Programming 105(2-

3), 563–582.

Nash, J. (1950), Equilibrium points in n-person games, in ‘Proceedings of the

national academy of sciences’.

23

Bibliography

Neumann, A., Siliverstovs, B. & von Hirschhausen, C. (2006), ‘Convergence of

european spot market prices for natural gas? a real-time analysis of market

integration using the kalman ﬁlter’, Applied Economics Letters 13, 727–732.

NGSA (2007), ‘Natural gas supply association homepage’. www.naturalgas.org.

URL: www.naturalgas.org

Nowak, M. & Westphalen, M. (2003), A linear model for transient gas ﬂow,

in ‘Proceedings of ECOS’, pp. 1751–1758. available as SINTEF-report STF

38S03601, Trondheim, Norway.

Nygreen, B., Christiansen, M., Bjørkvoll, T., Haugen, K. & Kristiansen, Ø.

(1998), ‘Modelling norwegian petroleum production and transportation’, An-

nals of Operations Research 82, 251–267.

OED (2006), ‘Fakta norsk petroleumsverksemd 2006’.

O’Neill, P., Williard, M., Wilkins, B. & Pike, R. (1979), ‘A mathematical pro-

gramming model for allocation of natural gas’, Operations Research 27(5), 857–

873.

Pipeline Simulation Interest Group (2007), ‘PSIG homepage’.

URL: www.psig.org

Roeber, J. (1996), ‘The development of a UK natural gas spot market’, Energy

Journal 17, 1–15.

Schwarz, E. (1997), ‘The stochastic behavior of commodity prices: Implications

for valuation and hedging’, The Journal of Finance 52(3), 923 – 973.

Selot, A., Kuok, L. K., Robinson, M., Mason, T. L. & Barton, P. I. (2007), A

short term operational planning model for upstream natural gas production

systems. to appear in AIChE Journal.

Sullivan, J. (1988), ‘The application of mathematical programming methods to oil

and gas ﬁeld development planning’, Mathematical Programming 42, 189–200.

Ulstein, N., Nygreen, B. & Sagli, J. (2007), ‘Tactical planning of oﬀshore

petroleum production’, European Journal of Operational Research 176, 550–

564.

van den Heever, S. A. & Grossmann, I. E. (2001), ‘A lagrangean decomposition

heuristic for the design and planning of oﬀshore hydrocarbon ﬁeld infrastruc-

tures with complex economic objectives’, Ind. Eng. Chem. Res 40, 2857–2875.

24

Bibliography

Weir, C. (1999), ‘Regulation and the development of competition in the U.K. gas

supply industry’, Review of industrial organization 15, 135–147.

Westphalen, M. (2004), Anwendungen der Stochastischen Optimierung im

Stromhandel und Gastransport, PhD thesis, University Duisburg-Essen (Ger-

many).

Wolf, D. D. & Smeers, Y. (1997), ‘A stochastic version of a stackelberg-nash-

cournot equilibrium model’, Management Science 43(2).

Wolsey, L. A. (1998), Integer Programming, John Wiley & Sons, Inc.

Wood, M. K. & Dantzig, G. B. (1949), ‘Programming of interdependent activities.

I. general discussion’, Econometrica 17, 193–199.

Wu, F., Varaiya, P., Spiller, P. & Oren, S. (1996), ‘Folk theorems on transmis-

sion access: Proofs and counterexamples’, Journal of Regulatory Economics

10(1), 5–23.

Zhuang, J. & Gabriel, S. (2006), A complementarity model for solving stochastic

natural gas market equilibria. January 2006, Energy Economics, accepted.

25

Paper I

Asgeir Tomasgard, Frode Rømo, Marte Fodstad and Kjetil Midthun:

Optimization Models for the

Natural Gas Value Chain

Chapter in G. Hasle, K.-A. Lie, E. Quak (eds.): Geometric Modelling,

Numerical Simulation and Optimization, Springer Verlag, 2007

Chapter 2

Optimization Models for the Natural Gas

Value Chain

2.1 Introduction

The models in this chapter are based on the authors experience from making

decision support tools for the Norwegian gas industry. Our focus is on describ-

ing modeling techniques and important technological issues, rather than a very

detailed representation needed for commercial models. We study the natural gas

value chain seen from the point of view of an upstream company with a portfolio

of production ﬁelds. Such a company should plan its operations considering long

term contract obligations, the short term markets and transportation capacity

booking. In particular we describe how the operations and planning are inﬂu-

enced by the existence of spot markets and forward markets. For the models to

make sense it is also critical to include the technological characteristics of natural

gas transportation and processing. We therefore give a set of models where the

interplay between the technological characteristics of natural gas and the markets

are highlighted. In these models the economical content and the understanding

of gas markets is essential.

We structure the paper by gradually introducing the diﬀerent levels of the

supply chain. We start by describing the most important components of the

natural gas value chain in Section 2. Then in Section 3 we focus on how to

model natural gas transportation in a steady-state situation. This is the type

of transportation models suitable for planning problems with time resolution

weeks, months or years. In Section 4 we introduce gas storages and in Section 5

we describe a portfolio perspective and start investigating the integrated supply

chain view. Here we introduce short term markets. In Section 6 we see how

the spot-markets can be used to price natural gas storage capacity and indicate

how to estimate the terminal value of natural gas still in storages or in reservoirs

at the end of the planning horizon using the concept of an Expected Gas Value

Function. An appendix describing all notation used in the paper is included at

the end. All together these sections will give a supply chain optimization model

with an integrated view of the value chain, from production, via transportation

29

Chapter 2 Optimization Models for the Natural Gas Value Chain

PHYSICAL PROCESS

Production

MARKETS

Component

market

Spot market

Financial

market

Contract

market

Upstream

market

Storage

Processing

plant

Transportation

Figure 2.1: Important objects in the natural gas value chain

and processing to contract management and gas sales.

2.2 The Natural Gas Value Chain

Here we give a brief description of the diﬀerent elements of the natural gas value

chain on the Norwegian continental shelf: production, transportation, processing,

contract management and sales. The ﬁrst action is to transport the natural gas

from production ﬁelds to processing plants or transportation hubs where gas

from diﬀerent ﬁelds is mixed. Rich gas components are extracted and sold in

separate markets. The remaining dry gas is transported to the import terminals

in UK or on the European continent. In these hubs bilateral contracts and spot-

trades are settled. Also upstream markets exist, where the gas is sold before it is

transported to the import terminals. We focus on the value chain of a producing

company, hence the issues of transmission and distribution to end customers are

not considered.

In Figure 2.1 we show the main components of the natural gas export value

chain. Before we go in detail on these we give a short summary of the main

eﬀects of liberalization and regulation in the European gas market.

30

2.2 The Natural Gas Value Chain

Production

Production of natural gas takes place in production ﬁelds. Often these ﬁelds

have several owners, and each owner has production rights that are regulated by

lifting agreements. Typically a producer’s rights allow him to produce between

a minimum level of production and a maximum level of production within a set

of time periods of diﬀerent length. This production band may be ﬂexible so that

gas can be transferred between periods within predeﬁned limits. Normally such

production intervals are deﬁned for single days, for years, and for intermediate

periods in between like weeks and months.

Much of the natural gas produced is traditionally committed to take-or pay

contracts where the buyer has agreed to take a volume in a given import terminal

for a sequence of years. Again there is ﬂexibility on when to take the gas within a

year (or other time periods) and typically the daily oﬀtake is within a minimum

and maximum level. The customer nominates volumes within the take-or-pay

agreements, and the producer has to deliver. These nominations are often done

weekly, with ﬁnal nomination the day before production. In take-or-pay contracts

the price is usually indexed to other commodities like oil, to temperature and

several other parameters.

Transportation and Processing

Natural gas is transported in pipelines by using compressors to create a higher

pressure in the originating end of a pipeline, so that molecules will ﬂow towards

the end. Several pipelines may meet in a node in the transportation network.

They may have diﬀerent pressure at the end of the pipeline, but the input pres-

sure of all pipelines going out of a transportation node must be smaller than

the smallest end pressure of pipelines coming into the node, unless there is a

compressor in the node.

An example of an export network for natural gas is the one you ﬁnd at the

Norwegian continental shelf which consist of 6600 km of pipelines. Here natural

gas from diﬀerent ﬁelds have diﬀerent quality, in terms of energy content and

its chemical composition (methane, ethane, propane and several more). Hence

when natural gas from diﬀerent ﬁelds is blended in the transportation network,

it is critical to either keep track of the energy content of the blend or the total

content of each natural gas component.

Some of the components can be extracted from the rich gas in processing

plants. Processing facilities separate the rich gas into its various components.

The components are liqueﬁed petroleum gases like ethane, propane and butanes,

which are exported by ship to separate commodity markets. The remaining dry

gas (methane and some ethane) is transported in pipelines to import terminals

in the UK, France, Belgium and Germany.

31

Chapter 2 Optimization Models for the Natural Gas Value Chain

The organization of transportation markets varies a lot from region to region.

One will often ﬁnd that existing transportation rights already accounts for much

of the available transportation capacity in the network. The Gas directive (Euro-

pean Union 1998) enforces undiscriminating third party access to the remaining

capacity (see Section 2.2 for a discussion of The Gas directive). One way of re-

solving this is to introduce primary markets for transportation capacity where

capacity can be booked. In some cases a ﬁxed tariﬀ is used for zones or for

pipelines, in other cases bids are given for capacity and the market settled by

some auction mechanism. In all cases the market is cleared and capacity is al-

located by given transparent rules. In a secondary market with shorter time

horizons transportation capacity is balanced with transportation needs for the

diﬀerent shippers.

In this paper we will only focus on the utilization of transportation capacity,

while the capacity allocation regime and tariﬀ regime is not discussed. For a

further discussion on these topics see Dahl et al. (2003).

Storage

There exist several types of natural gas storages. Abandoned oil and gas ﬁelds

have high capacity and thereby a cost advantage. They also have low risk as geo-

logical data are known. In aquifers water is replaced with gas. They have higher

risk as seismic investigation is necessary. Salt caverns are underground storage

tanks washed out from salt layers. They typically have high costs. Injection

rates, capacities, withdrawal rates and characteristics depending on ﬁlling rate

vary between the types. Storages are important in planning models because they

allow us to store natural gas close to the market and thereby use them to exploit

spot-market variations. They also allow producers to produce in time periods

where demand is low and to thereby utilize the available transportation capac-

ity. Also they can be used as seasonal storages to smooth out seasonal eﬀects.

Whether storage is used as to avoid bottlenecks in the system in high demand

periods or to utilize market possibilities, today’s storage capacity is very limited

when compared to the total production volumes.

Import Terminals and Markets

The import terminals are landing facilities for natural gas where the export

pipelines end. Natural gas is delivered here according to speciﬁcation on mini-

mum and maximum pressure and energy content. These characteristics are often

speciﬁed by the contracts as terms of delivery. Further transportation from the

import terminals are taken on by the buyer using a transmission network to

distribute the gas to the end customers.

32

2.2 The Natural Gas Value Chain

Originally these terminals were the points of deliverance for the many take-or-

pay contracts. Recently these terminals have also been the location of the growing

spot markets for natural gas and for ﬁnancial derivatives on the spot market. The

leading European hubs in terms of liquidity are the National Balancing Point in

Great Britain, TTF in the Netherlands and Zeebrugge in Belgium.

A diﬀerent type of markets is also emerging upstream in the pipeline network.

The main idea here is to have standardized trading mechanisms for natural gas

at some important locations in the network to be able to provide an additional

ﬂexibility for the producers. Upstream markets are used to perform trades of

natural gas before transportation takes place. They are useful because there is

a need for having standardized mechanisms to exchange gas between producers.

They include additional ﬂexibility for producers in terms of being able to stay

within the limits of their own lifting agreements, transportation capacities and

contract commitments in case of unexpected events or in case overselling or un-

derselling of natural gas has occurred. The buyer of gas upstream also has the

responsibility to transport the gas to downstream markets. Upstream markets

are not as well developed as the other markets. Still the idea is old and the former

variant was the less standardized bilateral long term swing agreements between

diﬀerent producers, allowing ﬁelds with little ﬂexibility an option to draw gas

from ﬁelds with more ﬂexibility in volumes.

Liberalization and Regulation

The European natural gas industry has developed rapidly over the past thirty

years. The European Commission has worked toward strengthening the compe-

tition within the complete gas- and energy value chain. A breakthrough in this

process came on the 22nd of June 1998 when the gas directive was passed in the

European Commission (European Commission & Transport 2002). In the direc-

tive a stepwise liberalization of the European gas market is described. The key

components of the gas directive are third party access to all transportation in-

stallations, division of activities within the ﬁrms in the value chain (physically or

by accounting) and the possibility for certain consumers to obtain their gas from

the supplier of their choice. The directive was followed by a second gas directive

in 2003 (European Union 2003) which moved another step towards liberalization.

Another implication of the gas directive and of EU competition laws was the

2002 closing down of the Gas Negotiation Committee (GFU), the forum for co-

ordinated gas sales from the Norwegian continental shelf. The GFU formerly

coordinated the supply of Norwegian natural gas producers Statoil and Hydro.

Now the sales are company based and rarely linked to a speciﬁc production ﬁeld.

An expected result from these changes is that short-term markets will evolve for

natural gas. Though liquidity is still low, there are already clear signs indicating

33

Chapter 2 Optimization Models for the Natural Gas Value Chain

that short-term contracts and spot-trades will play an important role in the

future. The prior market structure is dominated by long-term agreements and

thus minimizes the uncertainty for the participants. In practice the producers

take the price risk, as prices are ﬁxed towards various indexes, while the buyers

take the volume risks by going into long term agreements. The new markets will

include short-term bilateral contracts, spot markets and ﬁnancial markets. The

introduction of short-term markets will most likely also lead to higher volatility

and thus higher uncertainty.

Abolishment of the GFU-system and the introduction of a system where the

individual companies are responsible for disposal of their own gas reserves called

for a new access and tariﬀ regime in the transportation network. The ﬁrst step in

the Norwegian transportation system was taken with the creation of Gassco AS in

May 2001 under the provisions of a Norwegian White Paper. Gassco is assigned

all the operator’s responsibilities warranted in the Norwegian Petroleum Law

and related Regulations. As a State owned company, Gassco AS should operate

independently and impartially and oﬀer equal services to all shippers. Systems

operated by Gassco are the rich and dry gas systems previously operated by

Statoil, Norsk Hydro and TotalFinaElf.

The models presented in this paper are simpliﬁed variants of models developed

in co-operation with Gassco and Statoil to deal with the changes mentioned

above.

2.3 A Natural Gas Transportation Model

When modeling natural gas pipeline ﬂow it is important to have a conscious

view on how time and the dynamics of gas ﬂow should be handled. The model-

ing of natural gas ﬂow in continuous time has clear links to the process control

paradigm (Hofsten 2000). Within this paradigm one normally uses active control,

to operate the system according to a predetermined load and supply, ﬁnding a

sequence of control actions which leads the system to a target state. The control

regime often focuses on single processes or single components in the network.

For our purpose we need to model a system of pipelines with a set of production

ﬁelds, processing plants and markets. The natural choice is to look at mixed

integer programming models from the modeling paradigm of mathematical pro-

gramming. Here time is discretized. If the resolution of time periods is minutes

or hours there is a need to model the transient behavior of natural gas. Some at-

tempts on optimizing the transient behavior of a system of natural gas pipelines

are Westphalen (2004) and Nowak & Westphalen (2003), but only systems of

limited size and complexity can be handled. To be able to handle the complexity

needed for our models, we leave the concept of modeling the transient behavior

of natural gas and approximate the time dimension by discrete time periods of

34

2.3 A Natural Gas Transportation Model

such length that steady-state descriptions of the ﬂow will be adequate. When

the time resolution of the model are months, weeks, and maybe days, rather than

minutes and hours, we can assume that the system is in a steady-state in each

time period. The mathematical optimization models used to describe natural gas

ﬂow in the case of steady-state models are typical non-linear and non-convex. An

approach using a non-linear formulation of the mathematical models is illustrated

in De Wolf & Smeers (2000). We present here a linearized model based on mixed

integer programming to optimize routing of natural gas in pipeline networks. We

base our presentation on work done on linearization from Rømo, Tomasgard &

Nowak (2004). Several examples on linearization of natural gas ﬂow exist in the

literature. For a recent PhD thesis on linearization of natural gas ﬂow see Van der

Hoeven (2004).

In this paper we describe the essential constraints needed to model the tech-

nological characteristics of natural gas ﬂow in a steady-state setting. Issues like

pressure, gas quality and gas components are dealt with from a pipeline trans-

portation perspective. More detailed models very similar to the one we present

here are today in use by Statoil and Gassco in the software package GassOpt

developed by SINTEF. GassOpt is mainly used by the operator of the gas trans-

portation system in the Norwegian sector of the North Sea. They are obliged to

verify the delivery capabilities and robustness of the pipeline system transporting

natural gas to European markets.

In the model presented in this section, we will focus on the transportation alone

with the main purpose to meet demand for transportation generated by planned

production proﬁles for the diﬀerent ﬁelds. This typically represents the situation

facing the neutral operator. The pipeline system is a natural monopoly, and is

controlled by the authorities. This veriﬁcation is also of strategic importance

for the independent producers and customers in Germany, Belgium and France.

The security of supply will inﬂuence the price possible to achieve for long term

contracts, and contribute to infrastructure investment decisions, and GassOpt is

one of the tools used to ensure maximum utilization of the infrastructure.

GassOpt itself focuses on analyzes of transportation possibilities. It can be

used for optimal routing decisions from a ﬂow maximization perspective. Also it

is used to reroute natural gas when unexpected incidents lead to reduced capacity

(in production units or pipeline). Thirdly, it can be applied at more tactical/op-

erational level by a commercial player in capacity planning and capacity booking.

In this section we present a static model of one period. Demand for natural

gas in the import terminal is assumed to be aggregated over the contracts in the

terminals and planned production volumes given as constants to represent the

license holders’ production plans. So the main task of this model is to operate the

transportation network to make sure demand is met by the planned production.

In Section 4 we extend the model with several time periods and storage capabil-

35

Chapter 2 Optimization Models for the Natural Gas Value Chain

Figure 2.2: Network presentation in GassOpt

ities. In Section 5 we include contracts, markets and a portfolio perspective on

managing the natural gas supply chain with stochastic prices and demand.

The GassOpt Modeling Interface

In GassOpt, the underlying physical network is represented in a graphical mod-

eling environment with nodes and arcs. The modeling tool is hierarchical and

applies to general network-conﬁgurations. Figure 2.2 indicates the network com-

plexity for the North Sea network. The squared nodes contain subsystems with

further nodes and pipelines. When modeling the North Sea system we need

approximately 75 nodes and 100 arcs to represent the network.

36

2.3 A Natural Gas Transportation Model

The GassOpt Mathematical Model

This network model includes ﬂow balances, blending of diﬀerent gas qualities

from diﬀerent ﬁelds, processing nodes for extracting components of the natural

gas, compressor nodes, node pressures and the nonlinear nature of pressure drop

in pipelines. The model describes a steady-state situation where the network

is in equilibrium in terms of pressures and natural gas mix. It is typically the

kind of model used to model situations where ﬂows are aggregated over a given

time period. When the time period gets short enough, for example hours or

minutes, this steady-state description will not be good enough because of the

need to describe the transient behavior of natural gas ﬂow. The objective for the

optimization model is to ensure optimal routing and mixing of natural gas.

The model should make sure the nominated volumes are delivered to the import

terminals within a time period. This objective can be achieved in several ways.

Penalties are introduced in the objective function to inﬂuence the impact of the

following goals:

1. Maintain planned production from the producers, where this is physically

possible.

2. Deliver natural gas which meets quality requirements in terms of energy

content.

3. Deliver within the pressure requirements in the contracts.

4. Minimize the use of energy needed in order to deliver the natural gas to the

customers by minimizing the pressure variables.

A typical optimization case describes a speciﬁed state of the network, includ-

ing expected production and demand (characterized by volume and quality),

shutdown situations and turn-up capacity (additional available but unplanned

production capacity) from production ﬁelds. In a normal situation, there will

be several possible strategies to deliver the maximum amount of gas to the cus-

tomers. To make the model generate and report these realistic ﬂows, we have

introduced penalty costs in the objective function on deviation from planned pro-

duction, quality requirements, pressure agreements and the energy use. These

penalty costs can of course theoretically interfere with and prevent us to achieve

the main goal, to deliver in accordance with the demand of the customers. The

tests we have performed on the full North Sea network, show that this ‘multi-

criteria’ aspect does not sacriﬁce much of the maximal ﬂow potential, but is

rather used to choose between alternative solutions with about the same ﬂow.

In a fault situation, for example if a ﬁeld or pipeline is down, the model will

prioritize to deliver the nominated volumes in the import terminals. For more

information about multi-criteria decision making, see for instance Rardin (1998).

37

Chapter 2 Optimization Models for the Natural Gas Value Chain

Seen from an operator’s point of view the model tries to meet the customer’s

requirements for a given state of the network: either by optimal routing of gas or

by turning up production in ﬁelds with ﬂexibility on the production side. In the

last case we say that we use turn-up capacity, which is available in some ﬁelds

with ﬂexible production characteristics.

Sets

Below the sets used in the mathematical description of the model is presented.

N The set of all nodes in the network.

B The set of nodes where gas ﬂows are splitted into two or more

pipelines.

M Nodes with buyers of natural gas: typically import terminals.

I(n) The set of nodes with pipelines going into node n.

O(n) The set of nodes with pipelines going out of node n.

R The set of nodes with processing capabilities.

S The set of nodes with storage facilities.

K(b) The set of contracts in node b ∈ B.

C The set of components deﬁning the chemical content of the

natural gas.

T The set of time periods included in the model.

L The set of breakpoints used to linearize the Weymouth equation.

Z The set of split percentages used to discretize possible split

fractions in split-nodes of the network.

Y The number of discretized storage and injection rate levels used to

linearize storage characteristics.

Objective Function

Our goal is to route the gas ﬂow through the network, in order to meet demand

in accordance with contractual obligations (volume, quality and pressure). In

the formulation given below, variable f

im

is the ﬂow of gas from node i into

market node m, p

in

ij

is the pressure into the pipeline going from node i to j,

+

m

and

−

m

is the positive and negative deviation from the contracted pressure level

respectively, ∆

+

g

and ∆

−

g

represents underproduction and the use of turn-up in

relation to the planned production in ﬁeld g, δ

l−

m

is the negative deviation from

the lower quality level limit, and δ

u+

m

is the positive deviation from the upper

quality level limit in market node. The value of the ﬂow to the customer nodes is

given by the constant ω. Furthermore, κ is the penalty cost for pressure level,

is the penalty cost for deviation from contracted pressure level, χ is the penalty

38

2.3 A Natural Gas Transportation Model

cost for deviation from contracted quality to customers and ι for use of turn-up.

max Z =

¸

i∈I(m)

¸

m∈M

ω

m

f

im

−

¸

i∈N

¸

j∈N

κp

in

ij

−

¸

m∈M

+

m

+

−

m

−

¸

g∈G

ι

∆

+

g

+ ∆

−

g

−

¸

m∈M

χ

δ

l−

m

+δ

u+

m

(2.1)

Energy consumption for transporting the natural gas is minimized through

making the penalty cost (κ) insigniﬁcant in size as compared to the value of the

natural gas transported. This contributes to reduce the necessary build up of

pressure to a minimum, without interfering with the correct volume, quality and

pressure to the customer terminals. The penalty on using turn-up capacity will

make sure that planned production in the ﬁelds is prioritized ﬁrst, as long as

it does not inﬂuence the throughput of the pipeline system. For most practical

cases the contracted pressure level is not a soft constraint, and will then rather

be put into a hard constraint instead of being penalized in the objective function.

Constraints

Production capacity The following constraint says that the total ﬂow out of

a production node g cannot exceed the planned production of the ﬁeld in that

node. Here f

gj

is the ﬂow from production ﬁeld g to node j:

¸

j∈O(g)

f

gj

≤ G

g

, g ∈ G. (2.2)

Demand This constraint says that the total ﬂow into a node with customers

for natural gas must not exceed the demand of that node:

¸

j∈I(m)

f

jm

≤ D

m

, m ∈ M. (2.3)

Mass balance for node j The following constraint ensures the mass balance in

the transportation network. What ﬂows into node j must also ﬂow out of node

j:

¸

i∈I(j)

f

ij

=

¸

n∈O(j)

f

jn

, j ∈ N. (2.4)

Pressure constraints for pipelines Oﬀshore transportation networks often con-

sist of very long pipelines without compression, where it is crucial to describe

the pressure drops in the pipeline system. We use the Weymouth equation to

39

Chapter 2 Optimization Models for the Natural Gas Value Chain

describe the ﬂow in a pipeline as a function of input and output pressure. The

Weymouth equation is described in e.g., Campbell (1992). In the Weymouth

equation W

ij

(p

in

ij

, p

out

ij

) is the ﬂow through a pipeline going from node i to node

j as a consequence of the pressure diﬀerence between p

in

ij

and p

out

ij

:

W

ij

(p

in

ij

, p

out

ij

) = K

W

ij

p

in

ij

2

−p

out

ij

2

, j ∈ N, i ∈ I(j). (2.5)

Here K

W

ij

is the Weymouth constant for the pipeline going from i to j. This

constant depends among others on the pipelines length and its diameter and is

used to relate the correct theoretical ﬂow to the characteristics of the speciﬁc

pipeline. Figure 2.3 illustrates the Weymouth equation. The ﬁgure shows that

the function in the interesting area (positive pressure levels) is one fourth of a

cone. The cone starts in origo, and is limited by the inlet pressure axis, and the

45

◦

line between the inlet pressure and outlet pressure axes.

Through Taylor series expansion it is possible to linearize Equation (2.5) around

a point (PI, PO) representing ﬁxed pressure into the pipeline and ﬁxed pressure

out of the pipeline respectively:

W

ij

(p

in

ij

, p

out

ij

) ≤W

ij

(PI, PO) +

∂W

ij

∂p

in

ij

(p

in

ij

−PI)

+

∂W

ij

∂p

out

ij

(p

out

ij

−PO), j ∈ N, i ∈ I(j).

(2.6)

We introduce a set of points to linearize this expression, (PI

l

, PO

l

), where

l = 1, . . . , L. Then we replace for each pipeline the nonlinear function (2.5) with

L linear constraints of the type:

f

ij

≤K

W

ij

PI

l

PI

2

l

−PO

2

l

p

in

ij

−K

W

ij

PO

l

PI

2

l

−PO

2

l

p

out

ij

, j ∈ N, i ∈ I(j), l = 1, . . . , L.

(2.7)

For any given pipeline ﬂow, only one of these L constraints will be binding,

namely the one that approximates the ﬂow best. The planes described in (2.7)

will be tangent to the cone at the line where the ratio between pressure in and

out of the pipeline is equal to the ratio between PI

l

and PO

l

. Together the

planes give an outer approximation of the cone. This approximation will consist

of triangular shapes deﬁned by these planes.

Pipelines without pressure drop For physical pipelines between nodes where

the distances are very limited it is not necessary to model pressure drops by the

40

2.3 A Natural Gas Transportation Model

80

100

120

140

160

180

200

50

100

150

0

50

100

150

200

Pressure in

Pressure out

F

l

o

w

Figure 2.3: A three-dimensional illustration of how the Weymouth relates pres-

sure at the inlet and outlet points to the capacity in the pipeline.

41

Chapter 2 Optimization Models for the Natural Gas Value Chain

Weymouth equation. In this case a simple maxﬂow restriction is:

f

ij

≤ F

ij

, j ∈ N, i ∈ I(j), (2.8)

where F

ij

is the capacity. In this case there is no pressure drop, so:

p

out

ij

= p

in

ij

, j ∈ N, i ∈ I(j). (2.9)

Relationship between pressures into a node and out of a node To achieve a

relevant ﬂow pattern, it is sometimes preferable to model the pressure out of all

the pipelines going into the same node homogenously:

p

out

in

= p

out

jn

, n ∈ N, i ∈ I(n), j ∈ I(n). (2.10)

Another important issue is the relationship between pressure in ingoing pipelines

and the outgoing. In general for a node n the input pressure of all pipelines going

out of n must be lower than the lowest pressure out of any pipeline going into

node n, see Figure 2.4. There is one exception, and that is the case where a

pipeline into the node has 0 ﬂow. The end pressure of this arc is neglected. In

the Equation (2.11) the variable ρ

ij

is 0 for pipelines without ﬂow and 1 for the

others. M is a number which is large enough to not restrict the pressures when

the ﬂows are 0. Then the following constraints make sure that the input pressure

of a pipeline leaving n is less than the output pressure of a pipeline ending in n

as long as both pipelines have a ﬂow larger than 0.

p

in

nj

−p

out

in

+M(ρ

nj

+ρ

in

−1) ≤ M, n ∈ N, i ∈ I(n), j ∈ O(n) (2.11)

f

nj

≤ Mρ

nj

, n ∈ N, j ∈ O(n) (2.12)

ρ

nj

=

**i if ﬂow from node n to node j
**

0 otherwise.

(2.13)

The Weymouth equation used gives an upper bound on the ﬂow in a pipeline.

This means that even if there is a pressure diﬀerence in a pipeline the ﬂow can

be zero. Because of this property it is not necessary to explicitly model the

possibility of shutting down a pipeline. The model can simply put the ﬂow to

zero, and still keep the desired pressure. If omitting the constraints presented

above one has to be aware of this when interpreting the results from the model.

Modeling bidirectional pipelines For pipelines designed to handle ﬂows in both

directions, the ρ

ij

variable deﬁned in the previous paragraph is used to determine

the direction of ﬂow. Equations (2.14) and (2.15) make sure that there only ﬂows

gas in one direction in the pipeline.

f

ij

≤ Mρ

ij

, i ∈ N, j ∈ O(i), (2.14)

ρ

jn

= 1 −ρ

nj

, n ∈ I(j), j ∈ I(n). (2.15)

42

2.3 A Natural Gas Transportation Model

Node n

Node i

1

Node i

N

Node j

1

Node j

N

Figure 2.4: Example of a split node with the possibility to shut down operation

of one of the pipelines. The upper nodes, i, have pipelines going to

node n. The lower nodes, j, have pipelines coming from node n. The

index on i and j goes from 1 to N, where N is the total amount of

nodes.

Nodes with compression or pressure drop In some cases we allow the pressure

to increase in a node by using a compressor, or we force a pressure drop in the

node. We here present a simpliﬁed formulation for modeling compression nodes

where pressure can be build up or forced down. The compressor characteristics

includes a compressor factor Γ used to limit how much the gas can be compressed

in a node. If there is no compressor, this factor is 1. If there is a compressor,

this Γ is a function of the ﬂow f

n

=

¸

j∈I(n)

f

jn

into the node:

Γ

n

(f

n

) =

W

max

η(K

a

−1)

100K

a

f

n

+ 1

K

a

K

a

−1

, n ∈ N (2.16)

In this expression, the parameter K

a

is the adiabatic constant for a certain gas

type, W

max

is the power output capacity of the compressor, and η is the com-

pressor eﬃciency, (Campbell 1992). Here we simplify this by using a constant

compression factor independent of the ﬂow. Then the pressure out of the com-

pressor node n is limited by the compressor factor times the pressure into the

node n:

Γ

n

p

out

jn

≥ p

in

ni

, n ∈ N, j ∈ I(n), i ∈ O(n). (2.17)

43

Chapter 2 Optimization Models for the Natural Gas Value Chain

Pressure drop is modeled in the same way, but with a reduction factor Θ

n

instead

of a compressor factor:

Θ

n

p

out

jn

≥ p

in

ni

, n ∈ N, j ∈ I(n), i ∈ O(n). (2.18)

Here Θ

n

and Γ

n

are constants, where 0 < Θ

n

≤ 1 and 1 ≤ Γ

n

. The formulation

is only meaningfull if at most one of the factors is diﬀerent from 1 in a node.

Contracted pressure It may be necessary to model the contracted pressure in

nodes with customers. Most import terminals have a limited range around a

target pressure P

m

which they accept for incoming gas:

p

out

im

+

−

m

−

+

m

= P

m

, m ∈ M, i ∈ I(m). (2.19)

Here

−

m

and

+

m

are negative and positive deviations from the target pressure.

These deviations are penalized in the objective at a level reﬂecting how hard the

pressure constraint is in practice.

It is also possible to specify restrictions for each pipeline for example for the

pressure into and out of a given pipeline. Pressure restrictions often apply to

nodes with compression or nodes where processing of the gas is being performed.

These constraints are called technical pressure constraints. Examples are min-

imum and maximum pressure out of pipeline (represented by (2.20) and (2.21)

respectively).

p

out

ij

≥ P

min

ij

, j ∈ N, i ∈ I(j). (2.20)

p

in

ij

≤ P

max

ij

, j ∈ N, i ∈ I(j). (2.21)

Gas quality and energy content In this model, gas quality can be speciﬁed

in two diﬀerent ways, focusing on combustion value (GCV) of the natural gas,

or the content of CO

2

. These properties are both technically and economically

important for the customer. When dealing with CO

2

, the customer accept a

maximum content in terms of [mol %]. This is typically due to environmental

taxes or to requirements related to avoiding corrosion in pipelines. If we focus

on GCV, the customer accepts deliveries between a minimum and maximum

combustion value. High GCV is in itself tractable as the energy content is higher,

but in practice the plants using the natural gas are technically calibrated for a

certain GCV-range. The quality is then measured in [MJ/Sm

3

]. Here we only

give the formulation for GCV:

Q

min

m

≤ q

im

≤ Q

max

m

, m ∈ M, i ∈ I(m), (2.22)

where q

im

is gas quality (GCV ) in a pipeline going from node i to market

node m. In practice we need more ﬂexibility in the model by allowing reduced

44

2.3 A Natural Gas Transportation Model

quality in order to increase the ﬂow. Modeling this as hard constraints could lead

to situations where unexpected shutdowns of production ﬁelds or pipelines may

lead to a complete stop in deliveries to a customer due to the contractual quality.

If it is an alternative to get some deliverances, outside the contracted limits, but

within what is technically acceptable the latter will be chosen. This tradeoﬀ will

be valued in economical terms as reduction in the customer price. We need the

variables δ

l+

m

and δ

l−

m

to indicate the positive and negative deviation from the

lower quality limit Q

min

m

of customer node m. Likewise we need δ

u+

m

and δ

u−

m

to

indicate the positive and negative deviation from the upper quality limit Q

max

m

:

q

im

+δ

l−

m

−δ

l+

m

= Q

min

m

, m ∈ M, i ∈ I(m), (2.23)

q

im

+δ

u−

m

−δ

u+

m

= Q

max

m

, m ∈ M, i ∈ I(m). (2.24)

Gas quality and blending Gas quality is a complicating element because we have

to keep track of the quality in every node and pipeline, and this depends on the

ﬂow. Where two ﬂows meet, the gas quality out of the node to the downstream

pipelines depends on ﬂow and quality from all the pipelines going into the node.

The ﬂow in each pipeline is a decision variable in the model, and so is the quality

out of each node. We assume that the resulting blending quality is common for

all the downstream pipelines being connected to a node, and that it is decided

by the convex combination of inﬂow qualities to the node:

q

ij

=

¸

n∈N

q

ni

f

ni

¸

n∈N

f

ni

, i ∈ N, j ∈ O(i), (2.25)

or:

q

ij

¸

n∈N

f

ni

−

¸

n∈N

q

ni

f

ni

= 0, i ∈ N, j ∈ O(i). (2.26)

This equation has two quadratic terms on the form q

ni

f

ni

. These terms can

easily be reformulated in the following way: Deﬁne α = q

ni

−f

ni

and β = q

ni

+f

ni

.

Then q

ni

f

ni

= 1/4(α

2

− β

2

). Linearizing α

2

and β

2

is straightforward using

Special Ordered Sets of type 2 (SOS2, see for instance Williams (1999)). In the

SOS2 set at most two variables can be non-zero, and the two variables must

be adjacent. Still this means that we need to move into solution techniques

from integer programming, in particular branch and bound, so solution time will

increase exponentially with the numbers of SOS2 sets needed.

Modeling multi component ﬂows If we model the ﬂow of C components of the

natural gas we require that the split fractions of the components going into the

diﬀerent pipelines out of the node n is equal for all components. For simplicity

let us assume we always have only two pipelines out of a split node n ∈ N going

45

Chapter 2 Optimization Models for the Natural Gas Value Chain

Noden

Nodej

2

Nodej

1

Figure 2.5: The ﬂow is split in node n to node j

1

and j

2

.

to node j

1

and j

2

(see Figure 2.5). Let us also denote the ﬁrst component in the

set C of components for c

1

. All components are indexed from c

1

, . . . , c

C

. Then

the relation of the volume split between j

1

and j

2

is equal for all components:

f

c

1

nj

1

f

c

1

nj

2

=

f

c

nj

1

f

c

nj

2

, n ∈ N, c ∈ C. (2.27)

This is a quadratic expression, and we reformulate it using the equations (2.28)

to (2.32). We need a set of binary variables ϑ

nz

where z = 1, . . . , Z, each repre-

senting the choice of a split percentage for the share of natural gas going to node

j

1

. The ϑ

nz

variable is modeled as a special ordered set of type 1 (SOS1), where

only one variable can be non-zero (Williams 1999). For each ϑ

nz

we deﬁne a

constant E

z

giving the percentage related to the z. We also deﬁne a new variable

e

nz

representing the ﬂow through node n of component c if ϑ

nz

= 1.

The ﬁrst constraint says that the ﬂow from n to j

1

of component c equals the

percentage E

z

multiplied with the total ﬂow through node n of the component c.

f

c

nj

1

=

Z

¸

z=1

E

z

e

c

nz

, n ∈ B. (2.28)

The set B consists of all split nodes in the network. Then we need to restrict the

formulation so that only one ϑ

nz

is positive for each node:

Z

¸

z=1

ϑ

nz

= 1, z ∈ {1, . . . , Z}, n ∈ B. (2.29)

The e

c

nz

variables giving the ﬂow through the node of each component is con-

strained by the capacity of the node, corresponding to the active ϑ

nz

.

¸

c∈C

e

c

nz

≤ F

n

ϑ

nz

, z ∈ {1, . . . , Z}, n ∈ B. (2.30)

46

2.3 A Natural Gas Transportation Model

We also require that what ﬂows through the node of each component either goes

to node j

1

or to node j

2

:

Z

¸

z=1

e

c

nz

= f

c

nj

1

+f

c

nj

2

, n ∈ B, c ∈ C. (2.31)

And to make sure that there does not ﬂow more out of the node of each component

than what comes in:

f

c

nj

1

+f

c

nj

2

=

¸

i∈N

f

c

in

, c ∈ C, n ∈ B. (2.32)

Processing plants Some of the gas components are extracted and sold in sep-

arate component markets. The extraction is handled in processing plants in the

network. In the modeling of this process it is assumed that the volume of each

component extracted is a constant fraction of the total volume of that component

in a processing plant (A

c

r

). Hence, no decision on the conﬁguration of the pro-

cessing plant is made, but pressures and gas ﬂows through a processing plant can

be modeled by several processing nodes in sequence or parallel. This is expressed

in equation (2.33). The mass balance for the processing plant nodes can then be

formulated as in equation (2.34). The variable a

c

r

is used to keep track of how

much of component c is extracted from the ﬂow in processing plant r.

a

c

r

= A

c

r

¸

i∈N

f

c

ir

, c ∈ C, r ∈ R (2.33)

¸

i∈N

f

c

ir

=

¸

j∈N

f

c

rj

+a

c

r

(2.34)

Modeling turn-up: ﬂexibility in the production ﬁelds Turn-up is an expression

used for the ﬂexibility present in some production ﬁelds. For example reduced

transport capacity due to a shutdown in one part of the network may be com-

pensated by turning up the planned production from other gas ﬁelds not directly

aﬀected by the reduced capacity. When modeling this turn-up capacity it is im-

portant to keep in mind that even if one are free to utilize this ﬂexibility, it is

not acceptable from a practical point of view that the model presents a ﬂow al-

location where ﬁelds with signiﬁcant turn-up capacity will take over production

from minor ﬁelds, which basically is not aﬀected by the shutdown. The turn-up is

only used to take over production from ﬁelds that for some reason are prevented

to deliver. Hence, our ﬁrst priority is to meet demand in the network and our

second priority is to produce in accordance with the planned production at the

ﬁelds.

47

Chapter 2 Optimization Models for the Natural Gas Value Chain

We model this by adding a penalty cost for using turn-up in the objective

to avoid turn-up to be used at the expense of normal production capacity in

other ﬁelds. This works because not delivering gas to customers would generate

a loss which is considerably higher than the small penalty put on using turn-up

capacity.

The variables ∆

−

g

and ∆

+

g

represent underproduction and the use of turn-up

in relation to the planned production of G

g

for ﬁeld g. As before f

gj

is the ﬂow

from g to j:

¸

j∈O(g)

f

gj

+ ∆

−

g

−∆

+

g

= G

g

, g ∈ G (2.35)

2.4 Management of Natural Gas Storages

As a consequence of the liberalization process in the natural gas industry, the

natural gas markets have become more dynamic. The spot markets and the

possibility to trade gas in forward markets have increased the importance of gas

storages. In this section we discuss models for gas storage operations in a market

with uncertain demand.

In order to discuss the management of natural gas storages, a couple of terms

need to be established (see Figure 2.6 for an illustration of the terms):

Storage capacity gives the maximal volume of natural gas in the storages facility.

The storages capacity is limited by the physical properties of the storage.

Volume of natural gas in the storage is the total volume of natural gas in a

given storage at a given time.

Cushion gas is the amount of gas needed to create necessary pressure in order

to lift gas from the storage. The amount of cushion gas needed varies with

the type of storage and the geological conditions at the storage location.

For some types of storages the cushion gas requirement is as high as 80%

of the total gas volume in the storage.

Working gas is the gas volume available during normal operation of the storage.

This corresponds to the total amount of gas in the storage subtracted the

cushion gas.

Storage Facilities

The most common storage facilities are abandoned oil- and gas reservoirs, aquifers,

salt caverns and LNG-storages. In the following, a short overview of advantages

and disadvantages of these possibilities will be given. For further discussion of

storage facilities, see EIA (2002).

48

2.4 Management of Natural Gas Storages

Cushion gas

Working gas

Storage

capacity

Injection

Extraction

Figure 2.6: The complete square is the total storage capacity. The lower part

of the ﬁgure is the cushion gas needed for operation of the storage,

and the upper part of the ﬁgure is the gas currently available for

extraction from the storage.

49

Chapter 2 Optimization Models for the Natural Gas Value Chain

Abandoned oil- and gas reservoirs are the most common storage facility. One

reason for this is the relatively low startup costs. The storage facility is

already in place, and so is most of the surface installations needed. Another

advantage of this type of storage is the fact that infrastructure is normally

already in place. One major drawback is the amount of cushion gas needed

for operation.

Aquifer is a porous, underground water-bearing layer which can be transformed

into a storage facility by replacing the water with natural gas. When using

abandoned oil- and gas reservoirs the geological properties are known, this

is not the case when using aquifers. This adds risk to the development

of this type of storages. Cushion gas in the amount of 80 to 90 % is

needed for operation, and the development takes time and is costly. These

storages are normally only used in locations where no oil- and gas reservoirs

are available. One advantage of this type of storage is the relatively high

delivery rate.

Caverns are created from underground salt or rock formations. In the salt cav-

erns, water is used to dissolve halite and to shape cavities in natural salt

formations. These cavities have the properties of a high-pressure gas con-

tainer, with impenetrable walls. The storages have a high delivery capacity,

and a cushion gas requirement of only approximately 25 %. The process

of dissolving halite and shaping the cavities makes this alternative more

expensive than the previous two alternatives.

LNG-storages are, in contrast to the previously presented alternatives, above-

ground facilities. These storages consist of tanks containing liqueﬁed natu-

ral gas (LNG) or liqueﬁed petroleum gas (LPG). The capacity of these tanks

is normally very limited compared to the other alternatives presented.

Motivation for Utilization of Storages

The possibility of storing natural gas gives the participants increased ﬂexibility

with regards to production and transportation decisions. One important use of

natural gas storages is to take advantage of the strong seasonal pattern in prices.

Since the primary use of natural gas is for heating and production of electricity,

the fundamental price determinant in the markets is the weather conditions. The

demand is normally higher in winter than in summer, and the production capac-

ity is also lower than the peak demand. This means that the monthly demand for

natural gas may be much higher than the possible changes in production level can

satisfy. The diﬀerence between production capacity and peak demand can to a

certain degree be satisﬁed through utilization of storages. The use of storages can

substitute for investments in new production ﬁelds and transportation capacity.

50

2.4 Management of Natural Gas Storages

I

n

f

l

o

w

r

a

t

e

Storage level

X1 X2

X3

H1

H

2

H3

Figure 2.7: Illustration of the linearization of the injection rate of a storage.

Traditionally the storages have been used in order to ensure a high security of

supply. When problems occurred either in the production or transportation facili-

ties, storages could be used to supply the downstream participants. The storages

operate as a security buﬀer in this case. With the development of short-term

markets and volatile spot prices, the storages will be important for participants

wanting to utilize the price ﬂuctuations. Especially for gas producers not having

a reservoir close to the market this will be important. It can take several days

before a decision to change production level at the ﬁeld will result in increased

delivery in the market.

Modeling Storages

The maximum in- and outﬂow rates of the storage varies with the current storage

level. The maximal injection rate is a strictly decreasing convex function of the

storage level. Likewise the outﬂow rate can be given as a strictly increasing

convex function of the storage level. To be able to realistically represent the in-

and outﬂow rates, the use of special ordered sets of type 2 is chosen (Williams

1999). An illustration of the implementation of the SOS2 is shown in Figure 2.7

for the injection rate. The storage levels are discretized by a set of constants

X

1

, . . . , X

Y

, the corresponding injection rates are H

1

, . . . , H

Y

and the variables

ν

1

, . . . , ν

Y

are used to give a convex combination of two of the points. This means

that if ν

y

has a value diﬀerent from 0, then only one additional variable can be

non-zero. The only two candidates in this case are ν

y−1

or ν

y+1

. The storage

51

Chapter 2 Optimization Models for the Natural Gas Value Chain

level at a given time t is represented by x

t

s

.

¸

i

f

t

is

≤

Y

¸

y=1

ν

t

ys

H

ys

, s ∈ S, (2.36)

Y

¸

y=1

ν

t

ys

= 1, SOS2, s ∈ S, (2.37)

x

t

s

=

Y

¸

y=1

ν

t

ys

X

y

, s ∈ S, (2.38)

x

t

s

= x

t−1

s

+

¸

i∈I(s)

f

t

is

−

¸

i∈O(s)

f

t

si

, s ∈ S. (2.39)

The maximum and minimum levels of storage are modeled implicitly with

the representation given. The maximal level (equal to the total capacity of the

storage) is restricted by the inﬂow function. When the storage reaches the upper

capacity level, the associated inﬂow rate is equal to zero. The minimum level

(coming from the requirement of a certain level of cushion gas in the storage)

is handled in a similar way: when the minimum storage level is reached, the

associated outﬂow rate will be equal to zero.

2.5 Value Chain Optimization and Portfolio

Management

We will here give a short description on how to include markets and portfolio

optimization in the natural gas value chain. For more advanced models on port-

folio optimization in the natural gas value chain see Rømo, Tomasgard, Fodstad

& Midthun (2004) from which most of the ideas presented here originate. Other

relevant references are Nygreen et al. (1998) which considers portfolio optimiza-

tion for oil and gas ﬁelds in a strategic horizon and Ulstein et al. (2004) which

considers tactical value chain coordination, but without stochasticity and without

pressure constraints in the transportatin network.

Diﬀerent Levels of Portfolio and Value Chain Integration

The models presented here have both a portfolio and a value chain perspective.

These are important properties of a natural gas optimization model. The impor-

tance of these perspectives can be realized when considering the complexity of

the transportation system. Due to the technical nature of the gas network, sev-

eral physical and technical threshold-values exist. If such values are trespassed,

52

2.5 Value Chain Optimization and Portfolio Management

only minor incremental deliveries in one part can cause signiﬁcant unintended

reductions elsewhere. The bottlenecks in the transportation system make the

ﬂexibility incorporated in a system perspective valuable. We will not give all

the previous formulations of the transportation network again, but each time pe-

riod t in a value chain model will include transportation network constraints and

variables like the ones from Section 3 with an additional index t on all variables.

The motivation behind the portfolio and value chain perspectives can be sum-

marized by considering four levels of planning:

1. Traditional production planning: In this ﬁrst level the model ensures bal-

ancing of the production portfolio with the contract portfolio. Stochastic

demands and prices that are not perfectly correlated motivate a portfolio

perspective on the planning, as the portfolio variation will be lower than

the variation of the stochastic parameters of separate ﬁelds or contracts.

2. Production and market optimization: At this level markets are used to

supplement the physical production in order to gain more from the physical

production capabilities. The market can be used to resolve bottlenecks in

the transportation network or on the production side. The purpose is to

maximize the proﬁt from production and contract obligations using also

spot markets. At this level geographical swaps and time swaps of gas can be

performed using the market, and they are used to fully utilize the ﬂexibility

in the system.

3. Trading: At this level contracts and ﬁnancial instruments are traded in-

dependently of the physical production and contract obligations based on

market opportunities. The trading is similar to the previous level in terms

of using the spot market and ﬁnancial instruments like futures and options,

but the motivation is now speculation, not solving bottleneck problems.

These trades are in no way connected to the physical production and con-

tract obligations, unless the producer has market power.

4. Risk management: So far we have assumed the producer is risk neutral and

tries to maximize expected proﬁt. In that case it is enough to supplement

physical production with trades in the spot market at level 2. If the pro-

ducer is risk averse hedging the portfolio outcome using futures, forwards

or options may be optimal.

The distinction between level 2 and 3 is clear in theory, but in practice the

transition will be gradual.

53

Chapter 2 Optimization Models for the Natural Gas Value Chain

Field A FieldB

Emden Zeebrugge

Upstream

market

Storage

Figure 2.8: Example of a natural gas network

Utilization of Short-Term Markets in Value Chain

Optimization

The use of short-term markets allows for considerable ﬂexibility in the system.

Consider the network in Figure 2.8. In a situation where ﬁeld B needs to produce

and the company has an obligation to deliver in a bilateral contract in Emden

several possibilities exist:

• Field A supplies Emden, while ﬁeld B sells spot in Zeebrugge

• The company may buy spot in Emden and the procution from ﬁeld B can

be sold in the spot market in Zeebrugge.

• The company buys spot in Emden, while it sells the production from B

spot in the upstream market.

• Storage might be used to supply Emden, while the production from ﬁeld B

is sold elsewhere.

These simple geographical swaps makes the system more ﬂexible and gives the

company the possibility to maximize the ﬂow of natural gas (and the value of

their production) beyond what traditional transportation planning would have

done. For example bottlenecks in the production or in the transportation may

be resolved or moved using the markets actively.

54

2.5 Value Chain Optimization and Portfolio Management

A diﬀerent reason to use the markets is time swaps. Consider Figure 2.8 again.

This time ﬁeld B needs to produce in time 1, and the company has an obligation

to deliver in time 2. Several options are then available to the company:

• In period 1 ﬁeld B may supply storage, and in period 2 the storage supplies

Emden.

• In period 1 ﬁeld B can sell spot in Zeebrugge, and in period 2 either use a

forward contract or buy spot in Emden.

• In period 1 ﬁeld B can sell spot upstream, and then use either a forward

contract or the spot market to supply Emden

This is just some of many possibilities that exist for geographical swaps and

time swaps. The network considered is also very small. When expanding the

network to, for instance, 20 ﬁelds, 80 pipelines and 10 markets, the number of

possible routing decisions gets very large and the ﬂexibility increases. It is this

ﬂexibility we try to capture when modeling the portfolios of production ﬁelds

and contracts. The ﬂexibility further increases when perfect spot markets are

added. The need for ﬂexibility comes from the fact that demands and prices

are stochastic. The gain from portfolio thinking increases because they are not

perfectly correlated. We assume the company is a price taker. For simplicity of

notation, we assume there is only one company in the markets. If not, we would

also need to model the other companies’ transportation needs.

Including Markets and Contracts

In Section 3 only aggregated deliveries to take-or-pay contracts in the diﬀerent

customer nodes m ∈ M were considered. When including market transactions in

the model a representation of the uncertainty in the price process is important.

Based on this representation scenarios describing the uncertainty can be gener-

ated and optimal decisions in the interaction between the physical system and the

market can be made. In this description some simpliﬁcations have been made.

Only one company is considered, so no upstream market exists, the possibility

of delaying production through lifting agreements will be disregarded, and only

trades in the spot market will be considered. The possibility of trading forward

contracts is only interesting for a risk adverse company. This will be discussed

shortly at the end of this section.

Figure 2.9 illustrates how the market nodes are included in the model. The

arrows show that gas might ﬂow from the transportation network to the market.

There is no ﬂow from the market to the network (as would be the case for an

upstream market). In addition, transactions within the market node can be

performed. In the spot market the company can purchase or sell volumes of

55

Chapter 2 Optimization Models for the Natural Gas Value Chain

Naturalgasnetwork

Bilateral

contracts

Spot

Market Forward

Figure 2.9: The market node

56

2.5 Value Chain Optimization and Portfolio Management

natural gas. Obligations in the take-or-pay contracts can be fulﬁlled either by

ﬂow from the network to the market node, or by transactions within the market

node.

Modeling Stochasticity

We use the modeling paradigm of stochastic programming to represent uncer-

tainty in the models, see for example Kall & Wallace (1994). Uncertainty is then

represented in a scenario tree, see Figure 2.10. The nodes in the scenario tree

represent decision points, and uncertainty is resolved along the arcs going out

of a node with several branches. In practice decisions are only made when new

information becomes known. A stage is the set of time periods elapsing between

each time information is learned by the decision maker. Each stage in the tree

typically consists of several time periods, but only nodes after a branching are

decision points, as they are the only time periods when new information about

the future is resolved. Still, decision variables are present in time periods where

information is not resolved, hence the time periodization using time periods t

reﬂect in which time period the decision has eﬀect. In Figure 2.10 there are 3

time periods. Time periods 1 and 2 are in stage 1, starting with the decision in

node 0 and ending just before the new decisions at stage 2 (in nodes 2, 6 and 10).

In a two-stage stochastic programming model we deﬁne a set of time periods t ∈

T

1

= {t

1

, . . . , T

1

} belonging to the ﬁrst stage where information is deterministic,

and a set of time periods t ∈ T

2

= {T

1

+ 1, . . . , T

2

} where some parameters

are stochastic (as seen from t ∈ T

1

). When time passes on and one enters the

ﬁrst t ∈ T

2

, uncertainty is resolved and also the remaining time periods can be

considered deterministic.

In the model presented here we use a two-stage formulation for ease of notation.

Several parameters are stochastic in reality. We will consider stochasticity in:

contractual demands, contract prices and spot prices. We denote the stochastic

contract price for contract k in customer node m at time period t ∈ T

2

as

˜

φ

t

mk

.

Stochastic demand for contract k in customer node m at time period t is ˜ µ

t

mk

.

The stochastic spot price is represented with

˜

ψ

t

m

. The vector of all stochastic

variables in time period t is

˜

ξ = (

˜

ψ

t

,

˜

φ

t

, ˜ µ

t

).

We use a tilde over the variable to reﬂect that it is stochastic (as seen from

t ∈ T

1

) and remove the tilde when the variable is deterministic. We then get the

following:

˜

ξ

t

m

Stochastic variables for customer node m in time period t ∈ T

2

seen from a

point in time t

∈ T

1

.

ξ

t

m

Deterministic parameters for customer node m in t ∈ T

1

, or t ∈ T

2

after

uncertainty is resolved (Seen from a point in time t

where t

∈ T

2

).

57

Chapter 2 Optimization Models for the Natural Gas Value Chain

0 1

2

6

10

3

7

11

Time

Scenarios

Figure 2.10: An example of a scenario tree

A scenario tree can be constructed for example using price processes for natural

gas or descriptions of the dynamic aspects of stochastic demand. We will not go in

detail on how to do this here, but assume the scenario tree exists in the remaining

part of this paper.

The Objective

We describe the supply chain portfolio optimization model as a two-stage stochas-

tic program with relatively complete recourse (Kall & Wallace 1994). The only

stochasticity that is present is in the right hand side and in the objective. The

typical length of a time period for a tactical planning model is one month, and

the planning horizon would typically be 12 months, where for example the ﬁrst 6

months would belong to T

1

and the last 6 months to T

2

in a two-stage formula-

tion. The objective is to maximize expected proﬁt taken into consideration cash

ﬂows and shortfall costs. Hence the objective can be described by summarizing

the expected cash ﬂow of the time periods. The cash ﬂow of each time period

t can be described as a function Π

t

(x

t−1

; ξ

t

) (or

˜

ξ

t

if stochastic) where x

t−1

is

the storage level in the start of the time period. The decision variables and con-

straints are equal in all time periods, except for initialization in time period 0

where only initial storage levels are deﬁned x

0

ns

and for the terminal conditions

at the end of the model horizon. We use the vector x

0

to denote the initial level

of all storages and x

t

to denote the level of all storages in time period t. The

proﬁt function for time period t ∈ T

1

∪ T

2

can be formulated as:

Π

t

(x

t−1

; ξ

t

) =

¸

m∈M

¸

k∈K

φ

t

mk

µ

t

mk

+

¸

m∈M

ψ

t

m

(ζ

t−

m

−ζ

t+

m

), (2.40)

58

2.5 Value Chain Optimization and Portfolio Management

where the ﬁrst term is the income from delivery in contract k in market m at

time t and the second term gives the proﬁt from trading in the spot market in

node m in time period t.

The two-stage stochastic program with ﬁxed relatively complete recourse is:

max

¸

t∈T

1

Π

t

(x

t−1

) +Q(x

T

1

), (2.41)

where

Q(x

T

1

) = max E

˜

ξ

[

¸

t∈T

2

Π

t

(x

t−1

;

˜

ξ

t

) + EGV

x

T

2

], (2.42)

subject to a set of constrains representing transportation, production and mar-

kets. These constraints are mainly the constraints descried earlier in this paper,

but we will look closer at the ones changing because of the introduction of markets

and contracts. The constraint sets are identical for all time periods t ∈ T

1

∪ T

2

.

For the last time period the objective includes the terminal value of the natural

gas in storages expressed by the Expected Gas Value function, EGV(x

T

2

). This

function is described in more detail in Section 6.

The solution resulting from maximizing expected proﬁts will normally be diﬀer-

ent from the solution reached with the objective function presented in Section 2.3.

This means that the solution does not necessaritly maximize the throughput in

the network, or minimize the cost of achiving a given throughput. The solution

will however show how the network should be managed in order to achieve the

highest possible expected proﬁt.

Constraints Including Markets and Contracts

The mass balance in the market node for each time period and each scenario is

expressed as:

¸

i∈I(m)

f

t

im

+ζ

t+

m

= ζ

t−

m

+

¸

k∈K(m)

µ

t

mk

, ∀m ∈ M, ∀t ∈ T . (2.43)

In (2.43), ζ

t

m

represent transactions in the spot market in node m in time period

t. The + sign indicates purchases of natural gas whilst the − sign indicates sales.

Delivery in contract type k in the node m in time period t are included in µ

t

mk

.

The mass balance equation illustrates the ﬂexibility gained by including markets

in the model. It is no longer necessary to ship the gas to the market node in

order to fulﬁll the contractual agreements, since the spot market can be utilized

for this. This means that geographical and time swaps are now available to the

company.

59

Chapter 2 Optimization Models for the Natural Gas Value Chain

Risk Aversion

In the discussion so far only the possibility for trading natural gas through the

spot market has been discussed. For a risk neutral company that is maximizing

expected proﬁts this is an adequate approach. Since the forward price is a good

predictor of the expected future spot price, trading in the forward market would

on average be approximately equal to trading on the spot market (this is based

on a simple arbitrage argument, see for instance Hull (2003). The fact that

natural gas is a commodity makes the argument less obvious, but under some

assumptions still valid. In the case where the company is risk averse however the

situation changes and some tools to handle risk management are needed. The

inclusion of a forward market then gives the company the possibility to hedge,

that is: to reduce the risk of their position. By trading forward contracts a given

price can be locked in on advance.

In this case the company will no longer maximize expected proﬁts from their

operations, but rather maximize a utility function that incorporates the risk

aversion of the company. Another way of doing this is to introduce a penalty

function that will put extra cost in the objective function on deviations from

some target proﬁt value. In addition to the change in the objective function, the

mass balance in the market node (see (2.43)) will be changed to incorporate the

possibility to trade in the forward market.

Solution Times

The complexity of the models introduced in this paper to a large extent depends

on the modeling of the gas components. The inclusion of gas components adds

a large number of integer variables to the problem. When excluding the gas

components, a stochastic model with a network consisting of approximately 80

nodes and 1000 scenarios, can be solved within an hour. This problem will have

approximately one million rows, one and a half million columns, four million

non-zero elements and fourteen thousand binary variables. When including gas

components the solution time increase signiﬁcantly, and it is diﬃcult to ﬁnd

an optimal solution. For a physical system similar to the one above, with 100

scenarios and 10 breakpoints (see Section 2.3), a solution with an integrality

gap of 4% to 5 % typically can be reached within 12 hours. If the objective

is only to maximize ﬂow in a static model, solution times are within minutes

when components are omitted and increases correspondingly when components

are added.

60

2.6 The Expected Gas Value Function (EGV)

2.6 The Expected Gas Value Function (EGV)

So far no considerations have been made with respect to how the ﬁnal period in

the planning horizon will be handled. The model presented so far will most likely

end up with a very low storage level, and the production might also be higher

than optimal when considering a longer horizon (since the value of the gas still

contained in the reservoirs is neglected).

In order to handle the end-of-horizon problem, several possibilities exist. One

way of handling the storage problem is to set a target value for the storage level

at the end-of-horizon, for instance the starting level.

x

T

s

≥ x

0

s

(2.44)

This level might also be made dependent on various factors, such as the season in

which the end-of-period belongs. This way of modeling the end-of-period however

allows for limited ﬂexibility and also neglects the true value of the gas contained

in the storage. A way of determining the optimal level for the storages in the last

period is by using the expected gas value function.

The Expected Gas Value function (EGV) gives an estimate of the value of

a unit of gas in storage at some point in time t, based on expectations for the

future development of the spot price of natural gas. When the EGV is used as

a boundary value, the alternative value of the natural gas in storage is thereby

included. This alternative value comes from the opportunities present after the

end of the model horizon. Hence for each end-of-horizon storage level, the EGV

must reﬂect the value of an optimal out-of-horizon strategy for injecting gas in

the storage and selling gas from the storage.

If high prices are expected in the future, the EGV will encourage a high storage

level in ﬁnal time period T

2

, whilst if lower prices are expected the optimal level

in period T

2

may be lower. Figure 2.11 illustrates how the EGV is included in

the existing optimization model. As the ﬁgure shows, the estimation of EGV is

performed independently from the value chain model and the purpose is to give

a boundary condition for the value of gas.

An important element in the model used to estimate EGV is the development

of the natural gas spot price represented through spot price curves. These can

be modeled using stochastic processes. Several versions of such models exist,

for an overview of some of them, see Schwarz (1997). Based on the chosen price

model, scenarios describing possible future outcomes can be constructed (see Fig-

ure 2.12). Hence, for any given initial storage level a strategy is found for injection

and withdrawal of natural gas based on a stochastic process for the gas price. In

practice this is a real-options approach used to value the value of gas in the

storage. The option value in gas storages comes from the operational ﬂexibility.

The company can switch between injection, withdrawal or idle modes, depending

61

Chapter 2 Optimization Models for the Natural Gas Value Chain

EGV

Field A FieldB

Emden Zeebrugge

Upstream

market

Storage

Portfoliomanagementmodel

StochasticProgrammingModel

EGVvaluation

T

o

t

a

l

v

a

l

u

e

o

f

s

t

o

r

a

g

e

Storagelevel

Figure 2.11: The estimation of the EGV is performed in a stochastic optimiza-

tion model that is independent of the existing optimization model.

The EGV is then used in the value-chain model as a boundary value

on the gas in storage and reservoirs.

62

2.6 The Expected Gas Value Function (EGV)

on the price development. For references on real-options, see for instance Hull

(2003). It is possible to do this estimation both for individual storages, and also

for the combination of all or some of the storages in the network. In the latter

case a more complicated model is needed for estimation of the EGV.

Time

S

p

o

t

p

r

i

c

e

Figure 2.12: Representation of the development of the spot price of natural gas.

In this case a recombining trinomial tree. The arcs in the ﬁgure

represent price movements, while the nodes represent diﬀerent price

scenarios.

In the following, an example of how the EGV can be calculated is given.

The procedure is based on Scott et al. (2000) and Manoliu (2004), and use a

stochastic dynamic programming framework. For references to similar work in

hydro power, see for instance Pereira et al. (1999), Pereira & Pinto (1991). After

choosing a stochastic model to represent the price of natural gas, a discrete

approximation of the storage facility state space is made. A tree similar to the

one constructed for the spot price (Figure 2.12) can be constructed also for the

storage level. In this case the nodes represent diﬀerent storage levels, while the

arcs represent injection and withdrawal of natural gas in the storage. A multilevel

tree representing the spot price and the amount in storage is then created. The

valuation is performed by backward induction through the tree. The option value

is calculated in each node by taking the maximum of the decision values of hold,

injection and withdrawal. The hold decision value is equal to the expectation of

the option value of the next steps, when storage level is unaltered. The injection

value is the negative value of gas injected in this period, plus the expected value

of increased storage level in future nodes. The withdrawal value is then the value

of releasing gas in this period, plus the expectation of option values of decreased

storage levels in coming nodes. This can be illustrated by (2.45), which shows

63

Chapter 2 Optimization Models for the Natural Gas Value Chain

the value in a given node in the tree:

I

t

(τ

t

) = π

t

ϕ

t

+I

t+1

τ

t+1

. (2.45)

I

t

(τ

t

) is the value of storage level τ in time period t in the node considered.

This is determined by the value of ﬂow π

t

(ϕ

t

), (where ϕ

t

is the volume injected

or withdrawn in period t) in the node in period t plus value of the storage level

τ

t+1

in the next time period (in nodes following the considered one). The storage

level is updated according to (2.46):

τ

t+1

= τ

t

+ϕ

t

. (2.46)

An illustration of a gas value function is given in Figure 2.13. The challenge is

in ﬁnding the appropriate total value for each level of storage, as well as ﬁnding

the breakpoints.

T

o

t

a

l

v

a

l

u

e

o

f

s

t

o

r

a

g

e

Storagelevel

MV

1

MV

2

MV

3

MV

4

Figure 2.13: An example of an expected gas value function. The MVs shows the

expected marginal value of gas for various levels of storage. This is

the additional value of one more unit of natural gas in the storage.

Even though short-term markets for natural gas are developing in Europe,

the liquidity in these markets is still very limited. This lack of liquidity makes

estimation of the spot-price process diﬃcult, and therefore also estimation of the

64

2.7 Conclusions

EGV diﬃcult. Given that a spot-price model can be modeled for any given time

horizon, a time-horizon of a couple of years may be appropriate for estimating

the EGV. As the time-horizon for estimation is increased, the discount rate will

make the gas value in the last periods decrease strongly.

2.7 Conclusions

In this paper we have gradually introduced the complexity of a stochastic op-

timization model for the natural gas value chain. We focus on coordination of

the diﬀerent levels of the chain and on a portfolio perspective. We started out

by deﬁning necessary constraints for a steady-state formulation of the underlying

transportation network, supporting multi commodity ﬂows and pressures. Next

we introduced the time aspect and the use of storages. Thereafter we introduced

stochasticity in demands and prices and gave a stochastic programming formula-

tion for a portfolio optimization model. Natural extensions of this model would

be contract selection and more advanced modeling of the production ﬂexibility

reﬂected by lifting agreements. Finally we deﬁned the Expected Gas Value func-

tion and explained its use for giving the terminal value of stored natural gas and

indicated how to calculate it.

Most of the model formulations presented here are simpliﬁed variants of models

that are implemented for commercial use on the Norwegian continental shelf. In

this paper we have taken the position of a large producer, but many of the

formulations would be relevant for more general models focusing on other parts

of the natural gas value chain.

65

Chapter 2 Optimization Models for the Natural Gas Value Chain

Appendix

2.A Notation and Deﬁnitions

Sets

N The set of all nodes in the network.

G The set of nodes in the network with production ﬁelds .

B The set of nodes where gas ﬂows are splitted into

two or more pieplines.

M Nodes with buyers of natural gas: typically import terminals.

I(n) The set of nodes with pipelines going into node n.

O(n) The set of nodes with pipelines going out of node n.

R The set of nodes with processing capabilities.

S The set of nodes with storage facilities.

K(b) The set of contracts in node b ∈ B.

C The set of components deﬁning the chemical content of the natural

gas.

T The set of time periods included in the model.

L The set of breakpoints used to linearize the Weymouth equation.

Z The set of split percentages used to discretize possible split fraction

in split-node of the network.

Y The number of discretized storage and injection rate levels used to

linearize storage characteristics.

Indexes

n Used for nodes in general. n ∈ N. When more indexes are needed,

i and j will be used.

g Used for nodes with production ﬁelds, g ∈ G.

b Split nodes, m ∈ M.

m Customer nodes b ∈ B.

r Used for nodes with processing plants.

s Storage facility s ∈ S.

k Contract k ∈ K.

c Component c ∈ C.

t Time period t ∈ T .

l Breakpoits in linearized Weymouth restrictions.

z Breakpoints in linearization of split percentages in split nodes.

y Breakpoints for linearization of injection rate levels in storages.

66

2.A Notation and Deﬁnitions

Constants

G

g

Planned production [Sm

3

/|t|] in ﬁeld g ∈ G.

F

ij

Upper limit for ﬂow through the pipeline from

node i ∈ N to node j ∈ N.

F

n

Upper limit for ﬂow through node n ∈ N.

P

max

n

Max pressure [bar]into a node n ∈ N.

P

max

ij

Max pressure [bar] into the pipeline from

node i ∈ N to node j ∈ N.

P

min

ij

Min pressure [bar] out of the pipeline from

node i ∈ N to node j ∈ N.

P

b

Target pressure [bar] for deliverances to a customer node b ∈ B.

Q

max

n

Max energy content requirement for gas deliverances to node n ∈ N.

Energy quality is given by (GCV or CO

2

) or

[MJ/Sm

3

]

[mol%]

, where GCV

is the (Gross Caloric Value).

Q

min

n

Min energy content requirement for gas deliverances to node n ∈ N.

D

b

Demand in standard cubic meter pr time unit [Sm

3

/|t|] for natural

gas in node b ∈ N.

S

l

Storage capacity [Sm

3

/|t|]in node s ∈ S.

K

W

ij

The Weymouth constant is used as a constant in an empirical

expression for linking ﬂow and pressure in pipelines.

A

c

r

Fraction of component c in processing plant r that is extracted

from the ﬂow.

PI Fixed point for pressure into a pipeline.

PO Fixed point for pressure out of a pipeline.

Γ

n

Compressor factor in node n ∈ N.

Θ

n

Pressure reduction factor in node n ∈ N.

η Compressor eﬃciency.

K

a

Adiabatic constant for a certain gas type.

W

max

Power output capacity of the compressor.

ω

b

Value of gas to customer b.

κ Penalty cost for pressure level.

Penalty cost for deviation from contracted pressure level.

ι Penalty cost for use of turn-up.

χ Penalty cost for deviation from contracted quality to customers.

E

z

Gives the split percentage related to a given z in linearization of

split nodes.

X

y

Discrete representations of storage level in linearization of storages.

H

y

Discrete representations of injection rates in storages.

67

Chapter 2 Optimization Models for the Natural Gas Value Chain

Decision Variables

f

c

ij

Flow from node i ∈ N to node j ∈ N of component c.

In some cases index c is omitted when we do not consider

multi commodity ﬂow.

f

n

Total ﬂow into node n.

e

nz

Flow through node n of component c used for linearization in

splitting nodes m ∈ M.

p

in

ij

Pressure [bar] into the pipeline going from node i to node j.

p

out

ij

Pressure [bar] out of the pipeline going from node i to node j.

q

ij

Gas quality (GCV or CO

2

) in pipeline going from node i to node j.

ν

y

Give convex combinations of X

y

and H

y

.

σ

in

Equal to 1 if ﬂow from i to n, otherwise 0.

ϑ

nz

Binary variable representing split percentage in node n.

a

c

r

Amount extracted of component c in plant r.

ρ

ij

Equal to 1 if ﬂow goes from i to j, otherwise 0.

ζ

t−

m

Volume sold in spot market m in time period t.

ζ

t+

m

Volume bought in spot market m in time period t.

δ

l+

b

Positive deviation from the lower quality limit Q

min

b

of customer

node b.

δ

l−

b

Negative deviation from the lower quality limit Q

min

b

of customer

node b.

δ

u+

b

Positive deviation from the upper quality limit Q

max

b

of customer

node b.

δ

u−

b

Negative deviation from the upper quality limit Q

max

b

of customer

node b.

x

t

s

The storage level at a given time t in a storage s ∈ S.

+

b

Positive deviation from the contracted pressure to customer b.

−

b

Negative deviation from the contracted pressure to customer b.

∆

+

g

Positive deviation from the planned production in ﬁeld g.

∆

−

g

Negative deviation from the planned production in ﬁeld g.

Functions

EGV

t

(x

s

) Expected gas value in time period t as a function of the

storage level in storage s.

W

ij

(PI, PO) Flow resulting from pressure diﬀerence between Pressure in,

PI and

pressure out, PO, of a pipeline according to the

Weymouth equation.

Γ(f

n

) Compressor factor as a function of ﬂow f

n

into the

node n ∈ N.

68

2.A Notation and Deﬁnitions

Stochastic Variables

˜

φ

t

bk

Contract price for contract k in customer node b in time period t.

˜ µ

t

bk

Demand for contract k in customer node b in time period t.

˜

ψ

t

m

The spot price in market m in time period t.

˜

ξ

t

The vector of all stochastic variables

˜

φ

t

, ˜ µ

t

and

˜

ψ

t

.

In time periods where these parameters are not stochastic or where uncertainty

is resolved, the tilde is dropped in the variable name.

69

Bibliography

Campbell, J. (1992), The Equipment Modules, Gas Conditioning and Processing,

7 edn, Norman, Okla.

Dahl, H., Rømo, F. & Tomasgard, A. (2003), An optimisation model for rationing-

eﬃcient allocation of capacity in a natural gas transportation network, in ‘Con-

ference Proceedings: IAEE Praha’.

De Wolf, D. & Smeers, Y. (2000), ‘The gas transmission problem solved by an

extension of the simplex algorithm’, Management Science 46(11), 1454–1465.

EIA (2002), ‘The basics of underground natural gas storage’.

http://www.eia.doe.gov/pub/oil_gas/

natural_gas/analysis_publications/storagebasics/storagebasics.pdf.

European Union (1998), ‘Directive 98/30/EC of the european parliament and of

the council’.

European Union (2003), ‘Directive 2003/55/EC of the european parliament and

of the council’.

European Commission, D.-G. f. E. & Transport (2002), ‘Opening up to choice:

Launching the single european gas market’, Oﬃce for Oﬃcial Publications of

the European Communities.

Hofsten, K. (2000), Model-Based Dynamic Control and Optimization of Gas Net-

works, PhD thesis, NTNU, Trondheim.

Hull, J. (2003), Options, Futures and Other Derivatives, ﬁfth edn, Prentice Hall

International, Upper Saddle River, N.J.

Kall, P. & Wallace, S. W. (1994), Stochastic Programming, John Wiley & Sons,

Chichester.

Manoliu, M. (2004 ), ‘Storage options valuation using multilevel trees and calen-

dar spreads’, SunGard Energy Systems, New York . To appear in International

Journal of Theoretical and Applied Finance.

Nowak, M. & Westphalen, M. (2003), A linear model for transient gas ﬂow,

in ‘Proceedings of ECOS’, pp. 1751–1758. available as SINTEF-report STF

38S03601.

71

Bibliography

Nygreen, B., Christiansen, M., Bjørkvoll, T., Haugen, K. & Kristiansen, Ø.

(1998), ‘Modelling norwegian petroleum production and transportation’, An-

nals of Operations Research 82, 251–267.

Pereira, M., Campodónico, N. & Kelman, R. (1999), Application of stochastic

dual DP and extensions to hydrothermal scheduling, Technical report, PSRI.

Pereira, M. & Pinto, L. (1991), ‘Multi-stage stochastic optimization applied to

energy planning’, Mathematical Programming 52, 359–375.

Rardin, R. L. (1998), Optimization in Operations Research, Prentice Hall, Upper

Saddle River, New Jersey.

Rømo, F., Tomasgard, A., Fodstad, M. & Midthun, K. (2004), Stochastic op-

timization of the natural gas value chain in a liberalized market, Technical

report, SINTEF, Trondheim, Norway.

Rømo, F., Tomasgard, A. & Nowak, M. (2004), Steady-state optimization of

natural gas pipeline ﬂow, Technical report, SINTEF, Trondheim, Norway.

Schwarz, E. (1997), ‘The stochastic behavior of commodity prices: Implications

for valuation and hedging’, The Journal of Finance 52(3), 923 – 973.

Scott, T., Brown, H. & Perry, N. (2000), ‘Storing true value?’, Energy and Power

Risk Management .

Ulstein, N., Nygreen, B. & Sagli, J. (2004), Tactical planning of oﬀshore

petroleum production. Working paper, NTNU, Trondheim.

Van der Hoeven, T. (2004), Math in Gas and the Art of Linearization, PhD thesis,

International business school and research center for natural gas, Groningen,

The Netherlands.

Westphalen, M. (2004), Anwendungen der Stochastischen Optimierung im

Stromhandel und Gastransport, PhD thesis, University Duisburg-Essen (Ger-

many).

Williams, H. P. (1999), Model Building in Mathematical Programming, 4th edn,

John Wiley & Sons, Chichester.

72

Paper II

Kjetil T. Midthun, Mette Bjørndal and Asgeir Tomasgard:

Modeling optimal economic

dispatch and ﬂow

externalities in natural gas

networks

Submitted to international journal

Chapter 3

Modeling optimal economic dispatch and

ﬂow externalities in natural gas networks

Abstract:

We present a combined framework for modelling the technological issues of

gas transportation and analysis of natural gas markets. In our framework we

model the optimal dispatch of supply and demand in natural gas networks,

with diﬀerent objective functions, i.e., maximization of ﬂow, and diﬀerent

economic surpluses. The models take into account the physical structure of

the transportation network, and examine the implications it has for economic

analysis. More speciﬁcally there are externalities in pipeline networks due

to pressure constraints and system eﬀects. Incremental increase in produc-

tion in one part of the network may cause signiﬁcant reductions elsewhere.

The proposed network ﬂow model for natural gas takes into account pres-

sure drops and system eﬀects when representing network ﬂows. Pressure

drops are modelled by the Weymouth equation. The Weymouth equation

is linearized such that it is possible to perform economic analysis in large

networks within our framework. The importance of combining economics

with a model for pressure drops and system eﬀects is illustrated by small

examples.

3.1 Introduction

In this paper, we present a model for analyzing the optimal economic dispatch

of natural gas in pipeline networks. Our approach combines a framework for

modeling the technological characteristics of natural gas ﬂows with optimization

modeling of markets. The economic objectives include maximization of social

surplus, consumer surplus and producer surplus. During the last years, the liber-

alization process in Europe and other places in the world, has led to an increased

interest in such market-oriented models, including the spatial demand and sup-

ply for the commodity in the optimization of the transportation system. In the

existing literature, we have found no approach that combines the modeling of

the technology of natural gas ﬂows with economic analysis of the transportation

system.

In our paper, we show that it is important to represent the underlying phys-

ical properties, like the relation between ﬂow and pressure, pressure drops and

75

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

resulting system eﬀects, in the economic dispatch models. The inclusion of the

technology leads to interesting, and in some cases, surprising results when op-

timizing the operation of the network. We show examples of the errors that

will be made if the technology is not modeled, and we illustrate and discuss the

externalities that arise from system eﬀects in the network. The linearization

of the gas ﬂow equations will make it computationally feasible to analyze even

large-scale networks. However, in this paper we focus on the qualitative eﬀects

of including the physical properties of natural gas ﬂows, and illustrate the eﬀects

in a small network example with two production nodes, two market nodes and a

single transportation node.

Some examples of existing models with focus on ﬂow maximization and natural

gas transportation physics are Ehrhardt & Steinbach (2004, 2005), Martin et al.

(2006), Westphalen (2004) and De Wolf & Smeers (2000). A recent PhD thesis on

the linearization of natural gas ﬂows is Van der Hoeven (2004). In An et al. (2003)

a non-linear model for optimization of the combined natural gas and electricity

power ﬂow is presented. The linearization that we use in our modeling framework

is based on Rømo et al. (2004).

When it comes to papers with focus on the economics of natural gas transporta-

tion, the studies we have found have a straightforward representation of gas ﬂows,

without considering the special physical properties of the network ﬂows. Some

examples are for instance Cremer et al. (2003) and Cremer & Laﬀont (2002),

where the optimal allocation of resources has been discussed, and the social wel-

fare maximizing solution is derived in a simpliﬁed setting, where ﬂow externalities

are not taken into account. In Cremer & Laﬀont (2002) the possibility of a par-

ticipant with market power is also considered. Examples of economic equilibrium

models are given in Gabriel & Smeers (2006), Gabriel et al. (2005) and Gabriel

et al. (2001). These models focus mainly on the economic issues and do not

give a detailed description of the engineering aspects underlying transportation

networks.

In the electricity sector on the other hand, the eﬀects on the physical power

ﬂows are typically taken into account when allocating capacity and dispatching

units. Due to Kirchhoﬀs’ laws in meshed electricity networks, there are severe

externalities that complicate the pricing procedures. Combining the physics and

the economics in models and analysis has for years been the standard way of an-

alyzing electricity transportation and markets. A key ingredient of an electricity

market design is the management of congestion or bottlenecks in the transmission

network through locational prices or similar measures. Schweppe et al. (1988)

formulated the optimal economic dispatch problem for electricity markets, and

introduced the concept of optimal nodal prices. Chao & Peck (1996) present a

consistent system of ﬂow gate prices, and Hogan (1993) describes a hedging sys-

tem for locational prices, through transmission congestion contracts. In Wu et al.

76

3.2 Natural gas ﬂows

(1996), some eﬀects of the externalities in electricity networks are investigated.

Common assertions about general network performance and characteristics are

scrutinized, and in some cases, counter examples are given. A thorough discus-

sion of the interaction between the transmission network and the energy markets

is given in Bjørndal (2000).

Developments during the last ten years have made it more important than ever

to better understand the integration of technology and economics in the natural

gas markets. In Europe, the European Commission passed the gas directive in

1998 (European Union 1998, European Commission & Transport 2002), followed

by a second gas directive in 2003 (European Union 2003). The key components of

the gas directive are third party access to all transportation installations, division

of activities within the ﬁrms’ value chain, i.e., vertical separation, and the possi-

bility for certain consumers to obtain their gas from the supplier of their choice.

A consequence of the gas market liberalization is that we may expect to see less

emphasis on long-term sales contracts, emerging short-term contract markets,

and more spot sales. Thus, central coordination of production and infrastructure

utilization may be replaced by market prices as the coordinating mechanism. In

order to achieve an eﬃcient market, it is vital that the prices provide the correct

signals of the value of the commodity to the market participants.

The purpose of this paper is to provide a framework for modeling natural

gas markets that will enable future analysis to capture the technical as well as

economic issues, in the same manner as the models we have seen for the liberalized

electricity markets. We apply a linearized model for the technology, while the

economic models may lead to non-linearities in the objective functions. In Section

2, we describe network ﬂow models for natural gas pipelines, showing both a non-

linear model, using the Weymouth equation, and a linearized approximation of it.

In Section 3, network externalities are discussed under the traditional objective

function, maximizing ﬂow or throughput. In Section 4, the analysis is extended

to alternative objective functions, maximizing social surplus, producer surplus

or consumer surplus. For this, we utilize price responsive supply and demand

curves. In Section 5, two numerical examples are given to illustrate the discussion

in Section 4. Some ﬁnal conclusions are provided in Section 6.

3.2 Natural gas ﬂows

In our presentation, we will use production nodes, transportation nodes and

market nodes, as well as pipelines, to describe a natural gas system. The trans-

portation system for natural gas in the North Sea is chosen as a motivating case

for our work. It is the largest existing oﬀshore network, and consists of long

pipelines, where the modeling of pressure drops along the pipelines is important.

For simplicity, we assume that there is one production ﬁeld in each production

77

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

node. In the production nodes, we also assume that there are compressors. Thus,

from these nodes, gas may be sent into the system at a pressure level that can

be chosen within certain limits, depending on the compressor unit characteristics

and pipeline design.

Natural gas is a mix of diﬀerent hydrocarbons and other gas components like

ethane, methane, butane, carbon dioxide and several others. The gas quality

in terms of its chemical composition and energy content will vary from produc-

tion ﬁeld to production ﬁeld. In our approach, we will not model the natural

gas components, but rather assume that all the gas in the system is equal and

may be viewed as one commodity. Our discussion will not be inﬂuenced by this

assumption, but in reality modeling gas components may introduce even more

externalities into the system.

Usually, several pipelines meet at transportation nodes in the network. The

natural gas is mixed, and a homogeneous gas leaves the transportation node

through one or several other pipelines. We will assume that it is not possible

to increase pressure in the transportation nodes, i.e., there are no compressors

available in the transportation nodes. This reﬂects the present system of the

North Sea, where transportation nodes usually lack compressors due to the high

installation and maintenance cost of such sub-sea installations.

In the market nodes, where the gas is traded, we have only incoming pipelines.

Also in the market nodes, there may be limitations on pressure levels, for instance

due to contractual agreements, which typically take the form of minimum and

maximum pressure requirements.

In the natural gas ﬂow model we present here, we will ﬁrst show how to model

pressure and ﬂow in a single pipeline, thereafter; we will discuss the system eﬀects

of pressure, and then give a full mathematical model.

Flow in a single pipeline

The Weymouth Equation

We model the pressure drops in pipelines based on the Weymouth equation,

which describes the relationship between inlet pressure, outlet pressure and the

amount of gas ﬂowing in a pipeline. The ﬂow through the pipeline is driven by

the pressure diﬀerence between the inlet and the outlet of the pipeline, and the

larger the diﬀerence, the more gas molecules will ﬂow between the two points in

a given period of time (see illustration in Figure 3.1). The Weymouth equation

is given by:

Q =

Tπ

8P

1, 44 ×10

−3

r

2

i

−r

2

j

d

5

R

MT

s

ZLF

0.5

(3.1)

78

3.2 Natural gas ﬂows

80

100

120

140

160

180

200

50

100

150

0

50

100

150

200

Pressure in

Pressure out

F

l

o

w

Figure 3.1: A three-dimensional illustration of how the Weymouth equation re-

lates pressure at the inlet and outlet points to the capacity in the

pipeline.

79

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

For details on this empirical equation, see for example Campbell (1992). The

volume of natural gas is given in standard cubic meters (Sm

3

). The standard cu-

bic meter refers to one cubic meter of gas under normal conditions, where normal

conditions are deﬁned to be 1 atmospheric pressure (1.01325 bar) and 15℃. The

parameters and variables in Equation (3.1) have the following interpretations:

Q The throughput (or ﬂow rate), [MSm

3

/day],

r

i

Inlet pressure in the pipeline, [bar],

r

j

Outlet pressure in the pipeline, [bar],

d The pipeline diameter, [m],

L The pipeline length, [m],

M The molecular weight of the gas, [kg/kmol],

T

S

The ambient temperature of gas, [K],

T The temperature at standard conditions, [K],

P The pressure at standard conditions, [bar],

R The gas constant, [J/(kmol ×K)],

Z The average compressibility factor of the gas,

F The friction factor of the pipeline.

We often aggregate the constants of the Weymouth equation into a Weymouth

constant K

W

ij

for the pipeline going from network node i to node j. The Wey-

mouth equation can then be expressed as in Equation (3.2), where W

ij

(r

i

, r

j

) is

the ﬂow through a pipeline going from node i to node j as a consequence of the

pressures r

i

and r

j

:

W

ij

(r

i

, r

j

) = K

W

ij

r

2

i

−r

2

j

. (3.2)

If there are limitations on the pressure level in the nodes, the ﬂow through the

pipeline will be constrained by these limits, since the ﬂow must be less than or

equal to the ﬂow that may be obtained if the inlet pressure is at maximum level,

R

i

, and the outlet pressure is at minimum level, R

j

. Consequently,

W

ij

(r

i

, r

j

) ≤ K

W

ij

R

i

2

−R

j

2

. (3.3)

There may be other capacity constraints limiting ﬂow, due to for instance

pipeline design parameters or capacity limitations in nodes. However, in this

paper, we focus on capacity constraints following from restrictions on pressure

levels only.

Linearization of the Weymouth Equation

Through Taylor series expansion it is possible to linearize Equation (3.2) around

a point (RI

i

, RO

j

), representing ﬁxed pressures into and out of the pipeline:

80

3.2 Natural gas ﬂows

W

ij

(r

i

, r

j

) ≤ W

ij

(RI

i

, RO

j

) +

∂W

ij

∂r

i

(r

i

−RI

i

) +

∂W

ij

∂r

j

(r

j

−RO

j

) .

After some calculations, the Taylor expansion around (RI

i

, RO

j

) will take the

following form for a pipeline between nodes i and j (for a detailed description,

see Appendix A):

f

ij

≤ K

W

ij

RI

i

RI

2

i

−RO

j

2

r

i

−K

W

ij

RO

j

RI

2

i

−RO

2

j

r

j

,

(3.4)

where f

ij

is the ﬂow variable. When we use the linearization to compute ﬂow

for each pipeline, around 20 of these constraints are added to the problem, rep-

resenting diﬀerent pairs of RI

ij

and RO

ij

(where RI

ij

is larger than RO

ij

).

Network ﬂows and system eﬀects

In a natural gas network, the practical capacities that can be utilized in one part

of the network depend on the pressures and ﬂows elsewhere in the system. For

a node with more than one pipeline connected, the chosen pressure level in the

node inﬂuences the capacity in all pipelines connected to the node. There are

two alternative ways of modeling the system eﬀects of pressure.

First, we can choose to model the situation, where in a network node, the

outlet pressure in all incoming pipelines and inlet pressure of all outgoing arcs,

are equal. Under this assumption, the Weymouth equation must be modeled as

an equality, as presented in its original non-linear form in Equation (3.1). The

linearization in Equation (3.4) is not valid, as modeling ﬂow using an inequality

in this constraint, basically corresponds to allowing a pressure drop at the inlet

of the pipelines.

Alternatively, we can model a more advanced node with valve arrangements,

that makes it possible to reduce the inlet pressure in some of the outgoing

pipelines. In practice, this corresponds to modeling the pipeline capacity as

an inequality, allowing the ﬂow in the pipeline to be lower than the ﬂow following

from a given node pressure. In this case, the linearization in Equation (3.4) may

be used.

The linearized variant is used in models developed for the system operator in

the North Sea for analyzing network ﬂow. The main reason is that it better

reﬂects the typical situation of many real transportation networks, including the

one in the North Sea. Another reason is that the formulation using Equation (3.1)

leads to a non-convex problem, limiting the size of problems that may be ana-

lyzed. One should note that the externalities due to system eﬀects, that we will

81

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

describe in the following sections, will be present also in the ﬁrst case, when an

equality is used in Equation (3.1), as it gives an even stronger link between the

diﬀerent pipeline ﬂows.

In the following, we will illustrate the system eﬀects of pressure with the ex-

ample network in Figure 3.2. There are two production nodes (A and B), two

market nodes (D and E) and a transportation node (C). The maximum pres-

sure in production node A is assumed to be larger than in production node B.

Assuming similar design parameters and similar lengths for the two pipelines AC

and BC, this gives the pipeline from A to C larger capacity than the pipeline

going from B to C. In this example, we will assume that any pressure limits

in node C are not restrictive. There are also minimum pressure requirements in

the market nodes D and E. Since the minimum pressure in node E is assumed

to be lower than the corresponding ﬁgure for node D, the capacity of pipeline

CE is larger than the capacity of CD, assuming the two pipelines are otherwise

equal. The pipeline ﬂows depend on the Weymouth constants of the correspond-

ing pipelines, and the pressure diﬀerences along the pipelines, as given by the

Weymouth equation.

The relationship between the pressure levels in the pipeline system is illustrated

in Figure 3.3. The x-axis represents the location of the nodes, while the y-axis

represents the pressure. Maximum and minimum pressure requirements in the

nodes are given by horizontal lines in the ﬁgure. The downward sloping lines

connecting two nodes represent a pipeline and show the inlet and outlet pressures

and the pressure drop along the pipeline. To increase the ﬂow in a pipeline, either

the outlet pressure needs to be reduced, or the inlet pressure must increase, or

in other words, the absolute value of the slope of the corresponding line in the

ﬁgure must increase.

A B

C

D E

Figure 3.2: Example of a transportation network consisting of two production

nodes, a transportation node and two market nodes.

82

3.2 Natural gas ﬂows

In the upper part of Figure 3.3, we have assumed that the pressure in node B

is at the upper limit, while it is still possible to increase the pressure in node A.

Moreover, the pressure in node E is at the lower limit.

Consider now a situation where production ﬁeld A is oﬀered a new contract

for delivery to market E. To accommodate this increase in ﬂow from A to E,

the pressure in node A must be increased relative to the pressure in node C.

Besides, the pressure diﬀerence between node C and E must increase in order

to increase the ﬂow through this pipeline. Since the pressure in node E is at

its lower limit, in practice, this means that the pressure in node C has to be

increased, and since there are no compressors in node C, it means that pressure

must increase in the production ﬁeld to accommodate this. The new situation

is shown in the lower part of Figure 3.3. More gas ﬂows from A to E (the line

joining A and E has become steeper), but less gas now ﬂows from B to D (the

line joining B and D is ﬂatter) because of the higher pressure in node C, and the

fact that pressure in node B cannot increase. In the new situation, more capacity

is available from node C to the markets, but less capacity is available from the

production nodes to node C. The available transportation capacity from node B

has been reduced as a consequence of the increased transportation from node A.

The fact that there are no compressors to increase pressure in the transportation

node, introduce externalities, and there is a trade-oﬀ between the capacity of the

upper pipelines (before the transportation node) and lower pipelines (after the

transportation node).

A natural gas network model

In the following, we present a linear programming model for physical ﬂow max-

imization in the transportation network, based on work at SINTEF and NTNU

Rømo et al. (2006), and in Section 3.4 we will extend this model to include

economic objectives.

Sets

N The set of all nodes in the network.

G The set of nodes in the network with production ﬁelds.

M The set of market nodes.

T The set of transportation nodes.

I(n) The set of nodes with pipelines going into node n.

O(n) The set of nodes with pipelines coming from node n.

83

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

Figure 3.3: The top ﬁgure shows the original state of the network, while the

lower ﬁgure shows the state of the network after the ﬂow from A to

E is increased.

84

3.2 Natural gas ﬂows

Constants

R

i

Max pressure [bar] in node i ∈ N.

R

i

Min pressure [bar] in node i ∈ N.

Decision variables

k

g

Production in ﬁeld g ∈ G.

f

ij

Flow from node i ∈ N to node j ∈ O(i).

r

i

Pressure [bar] in node i.

q

m

Volume of natural gas in market m ∈ M.

p

d

m

Price of natural gas in market m ∈ M.

p

s

g

Price of natural gas in production node g ∈ G.

Objective function

The mathematical formulation for maximizing ﬂow is straightforward:

max

¸

i∈I(m)

¸

m∈M

f

im

, (3.5)

where f

im

is the ﬂow from node i to market node m, M is the set of all market

nodes and I(m) is the set of all nodes with pipelines going into m.

Constraints

The ﬁrst set of constraints ensures that mass is conserved in the network. Pro-

duction k

g

in node g ∈ G must equal the amount of gas f

gj

transported from the

production node g into nodes j in its set of downstream nodes O(g) :

¸

j∈O(g)

f

gj

= k

g

, g ∈ G. (3.6)

For the transportation nodes, the amount of gas that ﬂows into node j must

also ﬂow out of node j:

¸

i∈I(j)

f

ij

=

¸

n∈O(j)

f

jn

, j ∈ T . (3.7)

In the market node m we need to make sure that the quantity of gas available

in the market, q

m

, is equal to the sum of the gas ﬂowing into the market:

85

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

¸

j∈I(m)

f

jm

= q

m

, m ∈ M. (3.8)

Moreover, we need to make sure that the maximum and minimum requirements

for the pressure in the nodes are satisﬁed:

r

i

≥ R

i

, i ∈ N. (3.9)

r

i

≤ R

i

, i ∈ N. (3.10)

We will present two alternatives for modeling pipeline capacity constraints

due to pressure limitations. The ﬁrst is the one that we advocate, where pres-

sure and system eﬀects are modeled properly, based on the Weymouth equation.

The linear constraint in (3.4) corresponds to an outer linearization of the Wey-

mouth equation for the pipeline ﬂow. A more precise approximation is obtained

by using a set of linearizations around the pairs of inlet and outlet pressures

(RI

l

, RO

l

), l = 1, . . . , L, for each pipeline. We denote this constraint set the

Weymouth formulation (WF). The second alternative is the formulation usually

found in economic models of gas systems, like Cremer et al. (2003) and Cremer &

Laﬀont (2002), which we will denote Independent Static Flow constraints (ISF).

When we discuss the economic modeling of gas networks, we will compare the

ISF formulation with the WF formulation that we suggest should be used.

The Weymouth formulation is given by:

f

ij

≤ K

W

ij

RI

il

RI

2

il

−RO

2

jl

r

i

−K

W

ij

RO

jl

RI

2

il

−RO

2

jl

r

j

,

i ∈ N, j ∈ O(i), l = 1, ..., L,

(3.11)

where we use L linear constraints for each pipeline. The system eﬀects of pressure

are implicitly modeled by the nodal pressure variables. If the pressure in a node

must be uniform for all connected pipelines, the inlet pressure of a pipeline out

of transportation node n is equal to the outlet pressure in a pipeline going into

the node. Thus, the capacity of the pipelines connected to the node is dependent

on the nodal pressure, and therefore also dependent on one another. With the

linearization of the Weymouth equation, the nodal pressure formulation ensures

that pressure is not built up at nodes without compressors, but allows for pressure

drops, thus reducing ﬂow in a given pipeline relative to its pressure potential.

Still, pressure levels in one part of the network will depend on the chosen pressure

values in other parts of the network.

The alternative model for pipeline capacities, is the ISF-formulation,

86

3.3 Max ﬂow and network externalities - two numerical examples

f

ij

≤ C

ij

, i ∈ N, j ∈ O(i), (3.12)

where C

ij

is the static ﬂow-independent capacity for the pipeline connecting

nodes i and j. Note that in this case, the system eﬀects disappear from the

model, as the ﬂow capacity of one pipeline is not inﬂuenced by other pipelines’

ﬂows. When replacing the Weymouth constraints with ISF constraints, it is an

open question how to determine pipeline capacities. One possibility is to use the

maximum pressure diﬀerence for the pipeline to determine its static capacity:

C

ij

= K

W

ij

R

i

2

−R

j

2

. (3.13)

In this case, the system eﬀect is not properly modeled, and the capacity is set

at its best possible value, which can be achieved in theory, but not necessarily

in practice. In the next section, we will illustrate the diﬀerence in two numerical

examples.

3.3 Max ﬂow and network externalities - two

numerical examples

In this section, we present two examples that illustrate the importance of includ-

ing pressure drops in the modeling of natural gas ﬂows. Network externalities

are present both in operational max ﬂow problems and in long-term investment

problems. In Section 3.4 we extend the analysis to situations with alternative

economic objectives.

Example 1: Externalities from network operations

Returning to the example network in Figure 3.2, Section 3.2, let us investigate

how an increase in ﬂow from node A to node E can inﬂuence the total through-

put. We analyze the network using pressure constraints and pressure drops and

compare it to an approach with ISF capacities.

The design parameters in Table 3.1 will be used. The pressure constraints of the

pipelines are due to the design parameters in the network, as well as compressor

capacities at the production nodes, and speciﬁcations in the delivery contracts.

The Weymouth constants are chosen such that they resemble pipelines in real

installations in the North Sea. Maximization of throughput in this system, gives

a ﬂow from A to E of 36.99 MSm

3

/d, and 31.17 MSm

3

/d from B to D, i.e., a

total ﬂow to the market nodes of 68.16 MSm

3

/d.

Now, consider a situation where the producer in node A wants to deliver

41 MSm

3

/d to the market in node E. How will this inﬂuence the ﬂow from

87

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

A to C B to C C to D C to E

Max pressure into pipeline, [bar] 200 150 140 140

Min pressure out of pipeline, [bar] 120 120 100 70

Weymouth-constant 0,63 0,41 0,38 0,34

Table 3.1: Design parameters

B to D? Imposing this new constraint and maximizing throughput using the

model from Section 3.2 gives the following result: ﬂow from A to E is, as desired,

41 MSm

3

/d, while the ﬂow from B to D is reduced to 22.68 MSm

3

/d. Thus,

when the ﬂow from A to E increases by 4 MSm

3

/d, the ﬂow from B to D is

reduced by more than 8 MSm

3

/d. The pressure in node C has increased from

129 bar to 139 bar, and this change has decreased the transportation capacity

from the production nodes to node C. The results are illustrated in Figure 3.4.

It is evident that the relationship between ﬂows in the network is far from lin-

ear, and that rather small changes in one part of the network can have a large

inﬂuence on other parts.

Figure 3.4: Example of a transportation network consisting of two production

nodes, a transportation node and two market nodes. The ﬁgure to

the left gives the original solution, and the ﬁgure to the right the ﬂow

when the ﬂow from A to E is increased.

Using Equation (3.13) to compute ISF capacity constraints on the pipeline

ﬂows, based on speciﬁed max and min pressures, gives the following ﬂow con-

straints: C

AC

= 100.8 MSm

3

/d, C

BC

= 36.9 MSm

3

/d, C

CD

= 37.23 MSm

3

/d,

and C

CE

= 41.22 MSm

3

/d. The maximal throughput is then equal to min(100.8+

36.9, 37.23+41.22) MSm

3

/d = 78.45 MSm

3

/d, while if we consider the contrac-

tual paths, assuming production node A trades with market E and production

node B with D, the capacities for trades are C

AE

= min(100.8, 41.22) MSm

3

/d

88

3.3 Max ﬂow and network externalities - two numerical examples

and C

BD

= min(36.9, 37.22) MSm

3

/d. The capacities are illustrated in Fig-

ure 3.5.

Figure 3.5: The ﬁgure on the left shows the pipeline network with ISF con-

straints. The ﬁgure on the right shows the ISF capacities of contract

paths.

Using any of these capacities either fails to recognize the dependency of ﬂows

completely (the contractual capacities), or exhibit a linear relationship between

ﬂows, where, for instance, reducing ﬂow from node A to any of the market nodes

by one unit, makes it possible to increase the ﬂow from node B by one unit.

Another issue is that the ISF constraints calculated from Equation (3.13), based

on maximum and minimum pressure limits, grossly overestimates the practical

capacities of the pipeline network, with a max ﬂow of 78.45 MSm

3

/d as compared

to the 68.16 MSm

3

/d if pressure levels and limits are taken into account.

Example 2: Externalities and investment decisions

Understanding the physics of the network system is important in order to make

good infrastructure and capacity decisions. In the following, we show an example

of network externalities due to investment decisions. Assume that we have a

simple basis network to start with, consisting of production node A, market node

D and two intermediate nodes CP1 and CP2. The network is illustrated in

Figure 3.6.

A single producer in node A is transporting gas to the market node D through

a single pipeline. However, a small ﬁeld is under development (node B). The

ﬁeld in node B will deliver its production to a new market in node E. Instead of

building a new pipeline going from B to E, the nodes can be connected to the

existing pipeline from A to D. By using the pipeline from A to D, the additional

pipelines are shorter than if a completely new pipeline should cover the whole

distance. Assume that the pipeline going to market node E can be connected at

89

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

Figure 3.6: The original network with the two potential junction points.

one of two intermediate positions on the existing pipeline from A to D: CP1 or

CP2. The investment possibilities are illustrated in Figure 3.7. This is a typical

realistic investment situation on the Norwegian continental shelf.

Figure 3.7: Two diﬀerent junction points.

Assuming that A is responsible for delivering to node D and that B is respon-

sible for the volume going to node E, we solve the max ﬂow problems for the

two cases using the model in Section 3.2. The design parameters are given in

Table 3.2, and we assume that the pipeline from B to A is not a bottleneck in

the system, and does not inﬂuence the pressure at node A.

With these values, the ﬂow capacity of the pipeline between A and D is

51.3 MSm

3

/d, before any investments are done. After investment, the two al-

ternative junction points lead to diﬀerent transportation capacities between A

90

3.3 Max ﬂow and network externalities - two numerical examples

Max pressure into pipeline from A 160 bar

Min pressure out of pipeline into D 80 bar

Min pressure out of pipeline into E 70 bar

Weymouth-constant from A to D 0.37

Weymouth-constant from A to CP1 0.44

Weymouth-constant from A to CP2 0.62

Weymouth-constant from CP1 to D 0.69

Weymouth-constant from CP1 to E 0.62

Weymouth-constant from CP2 to D 0.46

Weymouth-constant from CP2 to E 0.52

Table 3.2: Design parameters

and D. We assume in both cases that the volume between B and E is ﬁxed to

10 MSm

3

/d, and then maximize the throughput between A and D. The result

of the numerical example is illustrated in Figure 3.8. The graph on the top left

hand shows the original situation with volumes and pressures. The graph on the

top right hand shows the solution if CP1 is chosen, and the one underneath is

the solution when using CP2 as a junction point. With CP1, the capacity C

AD

is 44.1 MSm

3

/d, and with CP2, the capacity is 47.5 MSm

3

/d. The reduction

in capacity between A and D is 7.2 MSm

3

/d and 3.8 MSm

3

/d respectively, so

there is a daily diﬀerence of 3.4 MSm

3

/d between the two alternatives.

Comparing the solution for the extended network with the original network, we

see that the pressures in the nodes CP1 and CP2 have decreased. This decrease

is due to the need for higher delivery between A and the junction point. Since

the pressure in node A was at the maximum level in the original solution, this

increase can only be obtained by reducing the pressure in the junction points.

The diﬀerence in the solutions for the two alternative junction points, can be ex-

plained by looking at the Weymouth equation: when the length of the pipeline is

increased, the Weymouth constant decreases (see Equation (3.1)), and therefore,

less gas is transported for a given pressure diﬀerence

1

.

Another aspect to take into consideration is the investment cost of the new

pipelines. In the alternatives presented here, the pipeline from CP2 to E is

approximately 40 % longer than the pipeline from CP1 to E.

If the pressure constraints had not been included in the model, the location of

the junction point would not have inﬂuenced the capacity between A and D. The

1

The lengths used in this example is 840 km between A and D, 600 km from A to CP1, 240

km from CP1 to D and 300 km to E, 300 km from A to CP2, 540 km from CP2 to D

and 424 km to E. The Weymouth-constant is dependent on the square of the length of the

pipeline (see Equation (3.1)), and therefore the diﬀerence between the constants may be

non-intuitive.

91

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

Figure 3.8: Flow and pressure values for the original network and for the two

investment possibilities.

92

3.4 Optimal dispatch with economic objective functions

example shows that there are several considerations that need to be accounted

for when we are looking at expanding a natural gas network. The owners of

ﬁeld A must be compensated somehow for the loss of capacity after the new

investments. Moreover, there is a trade-oﬀ between the additional investment

costs of building a longer pipeline and the possibility to get more capacity in the

network. The optimal solution will depend on several factors, such as investment

costs, ownership in the ﬁelds and in the network, regulation of the network and

market conditions (supply and demand).

3.4 Optimal dispatch with economic objective

functions

Incorporating pressure constraints and pressure drops when modeling natural

gas ﬂows, will inﬂuence the solutions also when using diﬀerent economic objec-

tive functions. From the max ﬂow examples, we saw that the externalities in

natural gas ﬂows inﬂuence the practical capacities of the pipeline networks. By

introducing economic objectives, we will investigate the validity of economic rea-

soning based on simple ISF constraints as opposed to the more detailed modeling

of the network in our approach.

In the following, we assume that there exist supply functions in the ﬁeld nodes

and demand functions in the market nodes, representing marginal production

cost and marginal beneﬁts from consumption, respectively. We focus on the

short-term optimal operation of the natural gas system, given the existing in-

frastructure, with production, transportation and market nodes, pipelines, and

compressors with predetermined properties. We will assume that all the relevant

short-term costs are reﬂected in the supply functions. This is a simpliﬁcation, as

there may be some ﬂow dependent and pressure dependent costs associated with

for instance using gas for building up pressure. Flow dependent or pressure de-

pendent costs could be taken into account by adding corresponding cost terms in

the objective function, however, as we want to focus on the eﬀect of the capacity

constraints following from pressure limitations, we have not considered these cost

terms. Our approach follows along the same lines as the analyses of congestion in

electricity markets, where congestion management is frequently studied assuming

lossless approximations of power ﬂows.

We start by deﬁning diﬀerent economic objectives, before we summarize well-

known results for the single pipeline case, and ﬁnally move on to our contribu-

tion in the analysis of economic objectives in a pipeline system. We consider

the maximization of the following objectives: social surplus, consumer surplus

and producer surplus. In our presentation, we use linear demand and supply

functions. With the linearized Weymouth equations to model pipeline ﬂows, the

93

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

resulting optimization problems are quadratic optimization problems with linear

constraints.

Economic dispatch in a single pipeline

Consider the situation in Figure 3.9, with a system consisting of a production

node, a market node, and a single pipeline connecting the two nodes. The trans-

portation capacity is restricted by the design parameters of the pipeline. Clearly,

in this case, there is no pipeline system to interact with, so an ISF constraint for

the pipeline ﬂow, such as Equation (3.13), is suﬃcient to represent the maximum

amount of gas possible to transport. Overviews of existing theory for analyzing

single transportation links with capacity constraints exist, for example in Tirole

(1988). For completeness, we give a short introduction here (for more details, see

also Appendix 3.B)).

Figure 3.9: Example of a transportation network consisting of a single pipeline.

With a congested pipeline with capacity f

cap

< f

unc

, as in Figure 3.10, max-

imizing social surplus corresponds to maximizing the trapezoid limited by the

supply function, the demand function and the capacity constraint, i.e., the sum of

the areas CS, PS and A. We call this solution the constrained optimal economic

dispatch. At quantity f

cap

the two nodes will have diﬀerent marginal values, and

one possibility is that a system operator can bring along the equilibrium by giving

the producer a price of p

s

and charging a price of p

d

from the market. The income

to the system operator would then be equal to the area A in Figure 3.10. Note

that there may be other ways to implement the optimal solution, for instance

by managing capacity limitations through redispatching in secondary markets,

or through various two-part tariﬀs. These diﬀerent alternatives for dealing with

transmission constraints will typically result in diﬀerent allocations of total sur-

plus between suppliers, consumers and a possible system operator. We will not

discuss these issues further in this paper, however, independent of how capacity

constraints are managed, in the optimal solution, maximizing social surplus, the

marginal values should be consistent with the prices p

s

and p

d

.

In the unconstrained solution, maximizing social surplus gives an optimal gas

ﬂow, f

∗

, equal to f

unc

. When maximizing producer or consumer surplus, the

optimal unconstrained quantity f

∗

is usually less than f

unc

(refer Appendix 3.B).

Since the optimization problems have concave objective functions and convex

94

3.4 Optimal dispatch with economic objective functions

Figure 3.10: Situation when pipeline capacity is restricting the solution.

feasible regions, the pipeline will be used to the capacity limit in the constrained

cases where f

cap

< f

∗

, i.e., the optimal quantity produced and transported over

the pipeline is equal to min(f

cap

, f

∗

). This means that for suﬃciently constrained

networks, the optimal solution will be identical and equal to f

cap

for all three

objective function alternatives. For details on the single pipeline optimization

see Appendix 3.B. In the following, we will assume that when maximizing either

consumer surplus or producer surplus, the consumer or producer respectively,

will take all the proﬁt in area A. This means that maximizing consumer surplus

is done by maximizing the area CS +A. Likewise, maximizing producer surplus

is the same as maximizing the area PS+A. An interpretation of this assumption

is that when maximizing consumer or producer surplus, we assume either the

consumer or production side to be in control of the pipeline system, and the

network operator function.

Economic dispatch in a pipeline system

In this section we formulate the optimal economic dispatch in a network of

pipelines, with several production and market nodes. We focus on the diﬀer-

ence between using the modeling framework described in Section 3.2 as opposed

to applying ISF capacities throughout the network.

Supply and demand functions for the production and market nodes are assumed

95

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

to be deterministic and linear. Thus, the demand function in market node m ∈ M

is given by:

p

d

m

= α

d

m

−β

d

m

q

m

, (3.14)

where p

d

m

is the price, depending on volume q

m

, and α

d

m

and β

d

m

are positive

constants. For simplicity, we assume that the supply functions go through the

origin, such that supply in production node g ∈ G is given by:

p

s

g

= β

s

g

k

g

, (3.15)

where p

s

g

is the price in production node g, k

g

is the volume produced, and β

s

g

is

a positive constant.

The objective functions then, have the following formulations, and are gener-

alizations of the two-node situation, which is illustrated in Figure 3.10:

• Max social surplus:

max

¸

m∈M

α

d

m

q

m

−

1

2

β

d

m

q

2

m

−

¸

g∈G

1

2

β

s

g

k

2

g

. (3.16)

• Max consumer surplus (monopsony):

max

¸

m∈M

α

d

m

q

m

−

1

2

β

d

m

q

2

m

−

¸

g∈G

β

s

g

k

2

g

, (3.17)

• Max producer surplus (monopoly):

max

¸

m∈M

α

d

m

q

m

−β

d

m

q

2

m

−

¸

g∈G

1

2

β

s

g

k

2

g

. (3.18)

When comparing models with the Weymouth formulation to models with ISF

capacities, we analyze all the diﬀerent objective function variants; maximization

of social surplus, consumer surplus and producer surplus. Hence, we will compare

the results obtained from two classes of optimization models, using:

1. WF pressure constraints: {max (3.16) or (3.17) or (3.18) s.t. (3.6)-(3.11)}

2. ISF capacity constraints: {max (3.16) or (3.17) or (3.18) s.t. (3.6)-(3.8)

and (3.13)}

96

3.5 Optimal economic dispatch - two numerical examples

3.5 Optimal economic dispatch - two numerical

examples

In this section, we provide numerical examples illustrating the eﬀects that pres-

sure constraints have on optimal economic dispatch in a natural gas network.

We start with a moderately constrained network that illustrates the diﬃculties

in ﬁnding ISF capacities. Thereafter, we move on to a strongly constrained net-

work. This example shows that even in a highly constrained network, where all

objective function alternatives give the same solution with ISF constraints, the

solutions are diﬀerent for the alternative objectives when including WF pres-

sure constraints. This is due to the fact that modeling pressure in a network,

introduces some ﬂexibility that is not incorporated with the ISF capacities.

Example 3: Optimal dispatch and ISF capacities

This example illustrates how the ISF capacities introduced in Equation (3.13)

overestimates the true capacity in the network. If this approach is used to model

ﬂow constraints due to pressure limits throughout a network, we obtain a re-

laxation of the optimal dispatch problem, where ﬂow and ﬂow constraints are

modeled by means of the Weymouth equations. This is evident from noting that

Equation (3.13) gives the highest possible ﬂow between two nodes i and j with

pressure limits R

i

and R

j

, but that Equation (3.13) fails to recognize the depen-

dencies between pipeline pressures. In a network, it will not normally be possible

to run every pipeline with maximum pressure diﬀerence between the connected

nodes. If this is possible, the optimal solution is equal to the solution from the

ISF capacities model, if not, the ISF formulation based on capacity constraints

(3.13) is a relaxation of the true model.

In the following, consider a network with the same conﬁguration as in Fig-

ure 3.2. The design parameters are shown in Table 3.3, while the ISF capacities

are calculated from Equation (3.13) and displayed in Table 3.4.

A to C B to C C to D C to E

Maximum pressure into pipeline 160 180 150 150

Minimum pressure out of pipeline 125 125 90 75

Weymouth-constant 0.40 0.45 0.30 0.35

Table 3.3: Design parameters

We then use the supply and demand functions in Table 3.5 to compute the

optimal dispatch under the three diﬀerent objective functions. As can be seen

from the ﬂows in Figure 3.11 and the prices (or marginal values) in Table 3.11,

97

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

Capacity from A to C 42.33

Capacity from B to C 60.37

Capacity from C to D 39.69

Capacity from C to E 41.57

Table 3.4: ISF capacities in the network

the results depend a great deal on which model is used to represent network ﬂows.

As expected, we see that the solutions for the ISF capacities give higher total ﬂow

for all the objective functions. The objective function value is also increased for

all the objective functions under the ISF capacities. However, the ﬂow pattern

given by the model with ISF constraints is naturally not necessarily feasible for

the pressure constraint (WF) model.

Supply from node A p

s

1

= k

1

Supply from node B p

s

2

= 3k

2

Demand in node D p

d

1

= 200 −2q

1

Demand in node E p

d

2

= 200 −3q

2

Table 3.5: Supply and demand functions.

Objective

Social Consumer Producer

surplus surplus surplus

WF ISF WF ISF WF ISF

Price in node A 34.6 42.3 38.7 42.3 35.5 42.3

Price in node B 83.7 96 54.6 62.1 54.6 54.6

Price in node D 140 120.6 145.8 124.4 141 127.4

Price in node E 102.5 96.2 110.6 124.4 127.4 127.4

Objective value 8255.21 9058.9 6820.88 7145.03 6115 6316.57

Table 3.6: Prices and objective function values.

It may be argued that it is unrealistic to apply the capacities from Equa-

tion (3.13) directly, but rather make some reductions to allow for the system ef-

fects of the pressure constraints. However, it is an open question how this should

be accomplished, as the necessary capacity adjustments will depend on supply

and demand parameters as well as the design parameters of the transportation

network.

98

3.5 Optimal economic dispatch - two numerical examples

Figure 3.11: Solution with pressure constraints (on the left) and ISF constraints

(on the right).

99

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

Example 4: Optimal dispatch in a strongly constrained

network

We continue to use the example network in Figure 3.2, and the same supply and

demand functions as in the previous example (see Table 3.5). First, note that

in the unconstrained case, the competitive solution, maximizing social surplus,

would have resulted in a price of 76.9 in every node in the network, and the

total production (and consumption) would be 102.5 MSm

3

/d. As expected, the

production is highest in node A while sales are highest in node D. The solutions

maximizing consumer surplus and producer surplus both give lower production

volumes. When maximizing consumer surplus, the market price is 111.1 and the

total production is 74 MSm

3

/d. Maximizing producer surplus gives a market

price of 123.8 and a total production of 63.5 MSm

3

/d. The size of the CS, A,

and PS areas in Figure 3.10) are given in the ISF column in Table 3.10 and are

equal independent of which objective is used (CS +A, PS +A or CS +A+PS.

In order to analyze a strongly constrained network, we choose design parame-

ters such that none of the ﬂows from the unconstrained solutions are feasible. The

design parameters are displayed in Table 3.7, and the ISF capacities calculated

by Equation (3.13) are shown in Table 3.8.

When maximizing the three diﬀerent surpluses in this constrained network, the

ISF network model gives the same value for the decision variables, independent

of the objective function. The resulting ﬂow pattern is illustrated in Figure 3.12,

showing the production in the ﬁeld nodes and the volumes consumed in the

market nodes. The identical results for the three objectives, using the ISF for-

mulation, is similar to the single pipeline case in Section 3.4. Since the optimal

unconstrained solution for every objective function is beyond the capacity of the

network, the network will be used to the capacity limit, which is here equal to

min(39.4 + 12.7, 21.2 + 21.2) = 42.4.

A to C B to C C to D C to E

Maximum pressure into pipeline 172 135 135 135

Minimum pressure out of pipeline 130 130 115 115

Weymouth-constant 0.35 0.35 0.30 0.30

Table 3.7: Design parameters

For the network model with WF pressure constraints and pressure drops, how-

ever, the optimal ﬂow patterns are no longer identical for the diﬀerent objective

functions. I.e., even in this very strongly constrained network, the dependencies

between ﬂow in diﬀerent parts of the system, inﬂuence the solutions. The ﬂow

patterns are illustrated in Figure 3.13, and Table 3.10 displays the allocation of

surpluses between consumers (CS), producers (PS), grid revenue (A) and the

100

3.5 Optimal economic dispatch - two numerical examples

Figure 3.12: The solution when maximizing social surplus, consumer surplus and

producer surplus with ISF capacities.

Capacity from A to C 39.4

Capacity from B to C 12.7

Capacity from C to D 21.2

Capacity from C to E 21.2

Table 3.8: The ISF capacities.

101

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

total social surplus CS +A +PS) for the diﬀerent objective functions (see Fig-

ure 3.10). This is contrasted with the allocation of surplus in the ISF model. The

total surplus in the ISF model is higher; however, this is once more due to the

fact that the ISF solution is not feasible when we take into account the pressure

limits and the externalities in the network.

Consider now the WF solutions. In the monopsony case, maximizing consumer

surplus (CS + A), we note that compared to the solution with maximal social

surplus, some of the production has been switched from production node A to

production node B . This reduces the size of the traditional consumer surplus CS,

however, the resulting increase in A is larger, such that total monopsony surplus

(CS + A) increases. This result is due to two factors: the capacity constraints

on the network, and the special properties of the pressure constraints. With

the pressure constraints, the eﬀective capacity of each pipeline is determined by

the operation of the other pipelines in the system. This adds ﬂexibility to the

operator of the system, and gives possibilities of moving and creating bottlenecks

in the system. In this example, the monopsony has decreased the pressure in node

C, and therefore increased the capacity from node B to node C at the expense

of the capacity from node C to nodes D and E.

Objective

Social Consumer Producer

surplus surplus surplus

Production in node A 37.56 34.7 37.17

Production in node B 4.3 6.3 4.56

Sale in node D 20.93 20.5 20.87

Sale in node E 20.93 20.5 20.87

Table 3.9: Result from optimization.

Objective

Social Consumer Producer Surplus

surplus surplus surplus with ISF

CS 1095.1 1053.5 1089.2 1123.6

PS 733.1 663.6 722.1 674.16

A 4715.4 4777.2 4726.6 4884.48

CS +PS +A 6543.6 6494.3 6537.9 6682.24

Table 3.10: Results from the optimization. See Figure 3.10 for deﬁnition of the

areas A, CS and PS.

To summarize the results, we start by noting that in the unconstrained case,

102

3.5 Optimal economic dispatch - two numerical examples

Figure 3.13: The solution when maximizing social surplus, consumer surplus and

producer surplus with pressure constraints.

103

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

the three objective functions will in general lead to diﬀerent ﬂow patterns in the

system. When modeling the physical properties of the network using only ISF

capacity limits to restrict the solutions, the diﬀerent objectives will give identical

solutions in terms of ﬂow when the network is suﬃciently constrained. For the

very constrained network, this means that when using the ISF formulation, the

monopsony/monopoly cannot do better than the social surplus solution in terms

of changing the size of the areas CS, A and PS. Only the wealth distribution

changes with the objective in this case, due to diﬀerent assumptions on who is to

receive the grid revenue A. On the other hand, when modeling the network ﬂows

with pressure limits and pressure drops, the solutions may diﬀer with diﬀerent

objectives. The possibility of moving bottlenecks in the system when modeling

pressure and system eﬀects, allows the monopsony or monopoly to get a higher

surplus than they would in the solution with maximal social surplus, and at the

expense of the other party.

3.6 Conclusions

Due to the physical properties of natural gas transportation networks, there exist

severe externalities in the operation and development of these networks. Flows

are constrained by pressure limits and driven by pressure drops. This means that

changes in one part of the system will inﬂuence capacities and performance in

other parts of the system. An analysis of bottlenecks and threshold values in

the transportation network therefore must have a system perspective rather than

a pipeline perspective. Thus, in order to analyze a natural gas transportation

system, it is necessary to take into consideration the entire network.

In this paper, we provide a modeling framework that takes into account these

externalities. A linearization of the Weymouth equation, that describes ﬂow in a

pipeline, is proposed in order to allow for analysis of ﬂows and economic surpluses

even in large networks. The eﬀects of the network externalities are investigated

by means of numerical examples in a small network consisting of two production

nodes, two market nodes and an intermediate transportation node.

From these simple examples, it is demonstrated that the relation between ﬂows

in the system is highly non-linear and non-additive. The network externalities

may inﬂuence to a large extent eﬀective capacities, optimal ﬂow patterns and

marginal values or prices. Moreover, it is shown that ﬂexibility in operations

arise from these system and pressure eﬀects, in the sense that it is possible to

move bottlenecks in the system with a proﬁt for either producers or consumers.

The externalities arising from system eﬀects and pressure constraints thus inﬂu-

ence even the more qualitative aspects of the solutions, such as the eﬀects from

diﬀerent assumptions on the market design, and the competitive structure of the

markets.

104

3.6 Conclusions

Thus, we argue that modeling the physical properties of natural gas networks,

taking into account the technology in terms of physics and system eﬀects, is

important for economic analysis in natural gas networks.

105

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

Appendix

3.A Linearization of the Weymouth equation

The Weymouth equation is given by:

W

ij

= K

r

2

i

−r

2

j

. (3.19)

By using a ﬁrst order Taylor series expansion around a ﬁxed point RI , RO,

this can be expressed as:

W

ij

(r

i

, r

j

) ≤ W

ij

(RI , RO) +

δW

δr

i

(r

i

−RI ) +

δW

δr

j

(r

j

−RO) . (3.20)

We start by ﬁnding the partial derivatives:

δW

δr

i

=

2Kr

i

2

r

2

i

−r

2

j

, (3.21)

δW

δr

j

= −

2Kr

j

2

r

2

i

−r

2

j

. (3.22)

Putting these expressions into Equation (3.20) and using the ﬁxed point RI ,

RO we obtain the following result:

W

ij

(RI , RO) =K

RI

2

−RO

2

+

KRI

RI

2

−RO

2

(r

i

−RI ) (3.23)

−

KRO

RI

2

−RO

2

(r

j

−RO) .

This can be written as:

W

ij

(RI , RO) =K

RI

2

−RO

2

−

RI

2

−RO

2

RI

2

−RO

2

+

RI

RI

2

−RO

2

r

i

(3.24)

−K

RO

RI

2

−RO

2

r

j

.

which ﬁnally translates into the following approximation for the ﬂow through

a pipeline with pressure diﬀerence r

i

−r

j

:

106

3.B Unconstrained equilibrium

W

ij

(RI , RO) = K

RI

RI

2

−RO

2

r

i

−

RO

RI

2

−RO

2

r

j

. (3.25)

The ﬂow surface given by equation 3.19 is the upper quarter of a cone starting

in the origin. The linearization constraints are planes passing through the origin.

Each such plane linearized around (RI, RO) is tangent to the cone in all points

where

r

i

r

j

=

RI

RO

. For more details, see Westphalen (2004).

3.B Unconstrained equilibrium

Consider a single pipeline where the capacity of the pipeline is not binding. Also

assume demand and supply functions are known and linear. For a more details

regarding these calculations, see for instance Tirole (1988):

Demand: p

d

= α

d

−β

d

k

g

, (3.26)

Supply: p

s

= α

s

+β

s

q

m

. (3.27)

When looking at a system with a single pipeline the production, k

g

, and sold

volume, q

m

, are identical and equal to the ﬂow in the pipeline, f. The highest

possible social surplus is achieved for the system in a competitive market, where

the prices are p

d∗

and p

s∗

, and the traded quantity is f

∗

corresponding to the

intersection between the supply- and demand curves:

f

∗

=

α

d

−α

s

β

s

+β

d

⇒ p

d∗

= p

s∗

=

α

d

β

s

+α

s

β

d

β

s

+β

d

. (3.28)

Maximization of producer surplus can be done by implementing the monopoly

solution. In order to ﬁnd the monopoly price and quantity, the intersection

between marginal revenue and marginal cost is found. This gives the following

solution:

MC = α

s

+β

d

f, (3.29)

MR = α

d

−2β

d

f, (3.30)

MC = MR ⇒ f

∗

=

α

d

−α

s

β

s

+ 2β

d

⇒p

d∗

=

α

d

β

s

+α

d

β

d

−α

s

β

d

β

s

+ 2β

d

,

p

s∗

=

2α

s

β

d

+α

d

β

s

β

s

+ 2β

d

. (3.31)

107

Chapter 3 Modeling optimal economic dispatch and ﬂow externalities

The social surplus is reduced since the traded quantity in this solution is dif-

ferent from the competitive markets’ solution. In this case the surplus of the

producers’ is maximized.

In the opposite case, total consumer surplus is maximized through the monop-

sony solution. In order to ﬁnd the monopsony solution, ﬁrst the marginal expen-

diture curve (ME) must be found. The marginal expenditure curve is equal to

the derivative of the cost function (C) for the monopsony:

C = α

s

f +β

s

f

2

, (3.32)

ME = α

s

+ 2β

s

f, (3.33)

The optimal solution is found at the intersection of this curve and the marginal

revenue product (MRP) curve (assumed equal to the demand curve):

ME = MRP ⇒ f

∗

=

α

d

−α

s

2β

s

+β

d

⇒p

d∗

=

2α

d

β

s

+α

s

β

d

β

d

+ 2β

s

,

p

s∗

=

α

s

β

d

+α

s

β

s

+α

d

β

s

β

d

+ 2β

s

. (3.34)

108

Bibliography

An, S., Qing, L. & Gedra, T. (2003), ‘Natural gas and electricity optimal power

ﬂow’, Proceedings of the IEEE/PES Transmission and Distribution Conference,

Dallas .

Bjørndal, M. (2000), Topics on Electricity Transmission Pricing, PhD thesis,

Norwegian School of Economics and Business Administration, Bergen.

Campbell, J. (1992), The Equipment Modules, Gas Conditioning and Processing,

7 edn, Norman, Okla.

Chao, H. & Peck, S. (1996), ‘A market mechanism for electric power transmis-

sion’, Journal of Regulatory Economics 10(1), 25–59.

Cremer, H., Gasmi, F. & Laﬀont, J. (2003), ‘Access to pipelines in competitive

gas markets’, Journal of Regulatory Economics 24(1), 5–33.

Cremer, H. & Laﬀont, J. (2002), ‘Competition in gas markets’, European Eco-

nomic Review 46, 928–935.

De Wolf, D. & Smeers, Y. (2000), ‘The gas transmission problem solved by an

extension of the simplex algorithm’, Management Science 46(11), 1454–1465.

Ehrhardt, K. & Steinbach, M. (2004), ‘KKT systems in operative planning for

gas distribution networks’, Proceedings in applied mathematics and mechanics

4(1), 606–607.

Ehrhardt, K. & Steinbach, M. (2005), Nonlinear optimization in gas networks,

in H. E. Bock, ed., ‘Modeling, Simulation and Optimization of Complex Pro-

cesses’, Springer-Verlag, Berlin - Heidelberg - New York, pp. 139–148.

European Commission, D.-G. f. E. & Transport (2002), ‘Opening up to choice:

Launching the single european gas market’, Oﬃce for Oﬃcial Publications of

the European Communities.

European Union (1998), ‘Directive 98/30/EC of the european parliament and of

the council’.

European Union (2003), ‘Directive 2003/55/EC of the european parliament and

of the council’.

109

Bibliography

Gabriel, S., Kydes, A. & Whitman, P. (2001), ‘The national energy modeling sys-

tem: A large-scale energy-economic equilibrium model’, Operations Research

49(1), 14–25.

Gabriel, S. & Smeers, Y. (2006), Recent advances in optimization. lecture notes

in economics and mathematical systems, in A. Seeger, ed., ‘Complemenatar-

ity Problems in Restructured Natural Gas Markets’, Springer-Verlag, Berlin

Heidelberg, pp. 343–373.

Gabriel, S., Kiet, S. & Zhuang, J. (2005), ‘A mixed complementarity-based equi-

librium model of natural gas markets’, Operations Research 53(5), 799–818.

Hogan, W. (1993), ‘Markets in real electric networks require reactive prices’, The

Energy Journal 14(3), 171–200.

Martin, A., Möller, M. & Moritz, S. (2006), ‘Mixed integer models for the sta-

tionary case of gas network optimization’, Mathematical Programming 105(2-

3), 563–582.

Rømo, F., Fodstad, M., Tomasgard, A., Hellemo, L. & Nowak, M. (2006), An

optimization system using steady state model for optimal gas transportation

on the norwegian continental shelf, Technical report, SINTEF, Trondheim,

Norway.

Rømo, F., Tomasgard, A., Fodstad, M. & Midthun, K. (2004), Stochastic op-

timization of the natural gas value chain in a liberalized market, Technical

report, SINTEF, Trondheim, Norway.

Schweppe, F., Caramanis, M., Tabors, R. & Bohn, R. (1988), Spot Pricing of

Electricity, Kluwer Academic Publishers, Norwell, Massachusetts.

Tirole, J. (1988), The Theory of Industrial Organization, MIT Press, London

Cambridge.

Van der Hoeven, T. (2004), Math in Gas and the Art of Linearization, PhD thesis,

International business school and research center for natural gas, Groningen,

The Netherlands.

Westphalen, M. (2004), Anwendungen der Stochastischen Optimierung im

Stromhandel und Gastransport, PhD thesis, University Duisburg-Essen (Ger-

many).

Wu, F., Varaiya, P., Spiller, P. & Oren, S. (1996), ‘Folk theorems on transmis-

sion access: Proofs and counterexamples’, Journal of Regulatory Economics

10(1), 5–23.

110

Paper III

Kjetil T. Midthun, Matthias P. Nowak and Asgeir Tomasgard:

An operational portfolio

optimization model for a

natural gas producer

Submitted to international journal

Chapter 4

An operational portfolio optimization

model for a natural gas producer

Abstract:

We present a short-term portfolio optimization model for a large natural

gas producer. The time horizon in the model is one week with daily reso-

lution. In this time-span the model includes spot market sales, production

plans, storage management and fulﬁllment of long-term contracts. The pa-

per discusses the value of actively using the storage capacity provided by

the line-pack of the pipelines to maximize proﬁt for the producer. We also

study the value of using a stochastic model as compared to a deterministic

model. The model is tested over two 60-days periods using real market data

and realistic production and transportation capacities.

4.1 Introduction

We present an operational portfolio planning model for a large oﬀshore producer

of natural gas. The objective for the producer is to maximize the value of the

produced natural gas within a week. The production targets are given as daily

and weekly volumes that must meet upper and lower production requirements

for each ﬁeld.

The entities that we model in the natural gas networks are: production ﬁelds

present in production nodes, pipelines between nodes, junction nodes where

pipelines meet and gas is blended, compressors used to lift the pressure in the

network at given nodes, and market nodes at which gas is sold through take-or-

pay contracts or at the spot market. In addition we look at storage in the pipeline

system. The producer strives to send more gas to the market on days with high

prices and less on days with low prices, while the overall produced volume does

not change within the week. When maximizing the value of production, the gas

producer must consider both physical limitations as well as guidelines from tac-

tical/ strategic models and ﬁeld lifting agreements (e.g. production limits). The

portfolio planning covers coordination of production in a set of ﬁelds, transporta-

tion in pipelines, management of storage and sales through contracts and spot

markets.

To our knowledge this is the ﬁrst study of a stochastic portfolio optimization

113

Chapter 4 An operational portfolio optimization model...

model for natural gas production and sales. We examine the value of using a

stochastic model, as compared to a deterministic approach. In addition, we eval-

uate the commercial value of actively using the storage (line-pack) present in the

export pipelines in order to increase the average price of the sold gas. Any other

storage capacity with close proximity to the market hubs like aquifers, caverns,

LNG-storages and abandoned oil- and gas reservoirs, can easily be included in

the model as well. The modelling of pipelines as storages and the interaction

with compressors to achieve this is non-trivial and our approach goes beyond the

methods reported in the current literature. Our analysis is motivated by the gas

production on the Norwegian continental shelf. The analysis is also relevant for

other regions where contract markets and spot-markets exist together and where

storage of gas may be used to increase the proﬁt.

Our model is a continuation of the work presented in Tomasgard et al. (2007)

that gives an introduction to modelling the value chain for natural gas covering

both system eﬀects, multi-commodity aspects of natural gas ﬂow and portfolio

optimization. That model is at a tactical level with a monthly resolution, no com-

putational results are presented, no price model is given and storage in pipelines

is not included. In related literature Nygreen et al. (1998) considers deterministic

investment optimization for oil and gas ﬁelds in a strategic horizon with focus

on how to develop ﬁelds and build pipelines. In Iyer et al. (1998) a multi-period

mixed-integer linear programming model for scheduling investment and opera-

tion in oﬀshore oil ﬁeld facilities is presented. Stochastic models for planning

oﬀshore gas ﬁeld developments under uncertainty in reserves is presented in Goel

& Grossmann (2004) and in Haugen (1996). Ulstein et al. (2007) give a model

for deterministic tactical value chain coordination including the multi-commodity

aspects of natural gas. That approach does not include a physical model for the

natural gas ﬂow and thereby neglects the important system eﬀects in the trans-

portation networks (Midthun et al. 2006) and also does not allow the modelling

of storage in pipelines. In Selot et al. (2007) a MINLP model for operational

planning for upstream natural gas production systems is presented. The model

gives a detailed description of the physical properties of natural gas production

and transportation.

In Thompson et al. (2003) an algorithm for the valuation and optimal operation

of a single natural gas storage facility is presented, and Byers (2006) present a

valuation methodology for commodity storage facilities.

In our test case we use real market prices from 3 European hubs and realistic

production and transportation capacities from the Norwegian continental shelf.

This is the world’s largest oﬀshore pipeline network for natural gas and produces

about 15 % of the European consumption of natural gas. We simplify the analysis

by assuming that one player can control both all production ﬁelds and the trans-

portation network. Our analysis here is on the value of portfolio optimization,

114

4.2 Modeling natural gas networks

coordination and storage, not on how the proﬁt should be shared between play-

ers or how the market should be organized. For analysis of the organization of

multi player behavior in the transportation market, see Gabriel & Smeers (2005)

and Midthun et al. (2007). Examples of computational equilibrium models for

natural gas markets are Gabriel et al. (2005) and Zhuang & Gabriel (2006).

We build on a linear steady-state model for the transportation network based

on Tomasgard et al. (2007). An alternative mixed integer approach is presented

in Martin et al. (2006). An overview of linearization of natural gas ﬂow is given

in Van der Hoeven (2004). All papers mentioned so far are steady-state models.

Work has also been done on modeling of the transient behavior of gas ﬂow. An

overview can be found in Kelling et al. (2000). In Westphalen (2004) and Nowak

& Westphalen (2003) models for transient ﬂow in a system of gas transportation

pipelines are presented. A deterministic version of the model in Nowak & West-

phalen (2003) is implemented in Nørstebø (2004). Here the time resolution is

in hours. Our work on modelling storage in pipelines builds on and extends the

approaches in Westphalen (2004) and Kelling et al. (2000).

In Section 4.2 we discuss important aspects of natural gas transportation. A

mathematical formulation is given in Section 4.3. In Section 4.4 we present the

market models and more details of the computational cases. We then present the

results in Section 4.5 and concludes.

4.2 Modeling natural gas networks

An important implication of the liberalization process of the European gas in-

dustry is the development of short-term markets in Europe. Now the producers

of natural gas have the possibility to sell their gas through, for instance, week-

ahead, day-ahead and within-day contracts, in addition to long-term contracts.

The majority of produced gas is still sold through long-term take-or-pay con-

tracts (TOP) where the producer has little opportunity to increase proﬁtability.

The structure of the take-or-pay contracts calls for careful planning. The buyer

in the take-or-pay contracts has ﬂexibility with respect to the weekly, as well as

daily, nominated volume. Consequently, volumes sold in the short-term markets

are the only factor the producer may inﬂuence, and these deliveries will never be

prioritized above the TOP-contracts. Spot markets therefore represent an im-

portant opportunity for increased proﬁt as an increase of the average spot price

achieved will be reﬂected directly in the company’s proﬁt, but also a challenge in

terms of the security of supply for TOP contracts. This means that the planning

decisions have to be ﬂexible and able to meet varying demands.

We here discuss how to model pressure constraints, and move on to discuss the

importance of transient behavior of the gas versus steady-state modelling. Then

we discuss the representation of line-pack as storage in the portfolio optimization

115

Chapter 4 An operational portfolio optimization model...

80

100

120

140

160

180

200

50

100

150

0

50

100

150

200

Pressure in

Pressure out

w

o

l

F

[Bar]

[Bar]

[

S

m

]

3

Figure 4.1: The Weymouth equation.

together with the representation of compressors used to control this storage.

Modelling pressure and system eﬀects

In a natural gas transportation network the capacity in a pipeline depends on

the design parameters of the pipeline (such as length and diameter), as well as on

the pressure at the inlet and the pressure at the outlet. The Weymouth-equation

is often used to describe ﬂow in a pipeline as a function of the inlet and outlet

pressure as well as design parameters (see Figure 4.1). For more details on the

Weymouth-equation, see for instance Campbell (1992). F

ij

(p

i

, p

j

) approximates

the ﬂow through a pipeline going from node i to node j as a function of the input

pressure p

i

and output pressure p

j

:

F

ij

(p

i

, p

j

) = K

W

ij

p

i

2

−p

j

2

, j ∈ N, i ∈ I(j). (4.1)

K

W

ij

is the Weymouth constant for the pipeline going from i to j. This con-

stant depends on the pipelines length, its diameter and other pipeline speciﬁc

parameters.

Consider the network in Figure 4.2. The ﬂow in the pipeline AC from A to C

depends on the pressure level in node A and C. The pressure in node C inﬂuences

116

4.2 Modeling natural gas networks

A B

C

D E

Figure 4.2: Example of a transportation network consisting of two production

nodes, a transportation node and two market nodes.

the ﬂow both in AC and CD. If the pressure in node C increases, the ﬂow from A

to C decreases while the ﬂow from C to D increases. Flow in one pipeline in the

network inﬂuences the potential ﬂow in the rest of the network. If ﬁxed capacities

are used for the pipelines instead of pressure constraints, the system eﬀects are

neglected. For a more detailed motivation for our steady state linearization and

discussion of system eﬀects we refer to Midthun et al. (2006).

Transient natural gas ﬂow and pipelines as storage

Steady-state models assume that inﬂow and outﬂow of gas is equal within a

period. However, long transportation pipelines may also be used as storages.

The line-pack in a pipeline is the gas contained in the pipeline. In the graph

to the left in Figure 4.3 a possible use of line-pack is shown for a case where

the pipeline is 90% full and a given price development. In the graph to the

right we see the development when the pipeline is already ﬁlled to max capacity

(Figure 4.3). The line-pack can be increased when prices are expected to increase

and decreased when prices are favorable.

Often, in order to model time and storage capacity, the pipeline is divided into

several segments with steady-state conditions in each segment. Applying well

known time-length criteria to long sub sea pipelines will result in a high number

of segments (Osiadacz 1983, Streeter & Wylie 1970). In an operational model

with daily resolution and routing ﬂexibility we need to check whether a simple

steady-state formulation will still be meaningful, or whether we need to use one of

the more complicated models partitioning the pipelines into a set of steady-state

segments.

117

Chapter 4 An operational portfolio optimization model...

Figure 4.3: Line-pack: The horizontal line shows the starting level of line-pack,

the dotted line shows the level of line-pack and the full line shows

the price development.

When the amount of gas taken out of the pipeline is less than the amount of gas

sent into it (in practice additional gas is being pushed in) the pressure and density

at the inlet is rising. A sudden increase of inﬂow will cause a pressure wave that

goes through the pipeline at the speed of sound. The intensity of the pressure

wave is decreasing since the energy is used to push gas particles toward the outlet,

causing both a slight increase of velocity and density also there. This eﬀect is

shown in Figures 4.4 to Figure 4.5. We use Matlab to simulate the development

of velocity and density of the gas molecules in a gas pipeline for given input and

output proﬁles. The simulation is performed in the same manner as described

in Nowak & Westphalen (2003), where diﬀerential equations are used to describe

the transient gas ﬂow in a pipeline. An increase in the outﬂow will lead to a similar

pressure wave, this time going from the outlet toward the inlet and sucking gas

out of the whole pipeline. At each point in time the pressure decreases along

the pipeline, while the velocity increases, since the energy for the gas transport

is stored in the compression. The ﬁgures illustrate that short term variations in

input and outtake have little impact on both velocity and density in the middle

of the pipeline or the other end of the pipeline.

We then look at a situation where we use a time resolution of 24-hours (the ﬂow

rate is constant within each 24-hour period, but in the simulation we show hourly

eﬀect on pressure and velocity). The chosen input/output-pattern, as well as the

resulting inlet and outlet pressure is illustrated in Figure 4.6. The ﬁgure shows

that the inlet and outlet pressure in the pipeline is an approximately piecewise

linear function. The ﬁgure also illustrates that there is transient behavior in the

118

4.2 Modeling natural gas networks

0

1

2

3

4

5

6

7

0

1

2

3

4

5

6

7

8

9

1.4

1.6

1.8

2

Distance [100000 m] Time [100000 s]

Velocity [m/s]

Figure 4.4: Simulation of velocity for a small variation in input and output.

0

1

2

3

4

5

6

7

0

1

2

3

4

5

6

7

8

9

158

160

162

164

166

168

170

172

Distance [100000 m]

Time [100000 s]

Rho [kg/m ]

3

ρ

Figure 4.5: Simulation of density (ρ) for a small variation of ﬁrst input and then

output.

119

Chapter 4 An operational portfolio optimization model...

pipeline in the ﬁrst hours after a change in input and/or output in the pipeline.

This can be seen from the non-linear behavior of the inlet and outlet pressure in

the hours after a changed ﬂow pattern. With a time-resolution of less than an

hour, this non-linear behavior will be important. We argue however, that with a

time-resolution of 24-hours, a modiﬁed steady-state approach is suﬃcient.

0

50

100

150

200

250

p

r

e

s

s

u

r

e

[

b

a

r

]

p

r

e

s

s

u

r

e

[

b

a

r

]

25 50 75 100 125 150

hours hours

0

20

40

60

80

ﬂ

o

w

[

S

m

3

]

ﬂ

o

w

[

S

m

3

]

inlet pressure

outlet pressure

input

output

Figure 4.6: Resulting inlet and outlet pressure from given input/output data.

The modiﬁed steady-state approach is based on steady-state equations, but

allows for diﬀerent in- and outﬂow quantities in the pipelines. We extend this

formulation with a storage model that depends on a weighted average of the

pressures at the inlet and outlet points of the pipeline. In Westphalen (2004) the

line-pack is calculated as a function of average pressure in the pipeline (p

avg

ij

=

1

2

(p

i

+p

j

), where p

i

is the pressure at the inlet and p

j

is the pressure at the

outlet). We have used a formula that in addition considers the shape of the

pressure drop in the pipeline (Kelling et al. 2000):

p

avg

ij

=

2

3

p

i

+p

j

−

p

i

p

j

p

i

+p

j

. (4.2)

This expression can be linearized in the following way (by using a ﬁrst-order

Taylor expansion):

120

4.2 Modeling natural gas networks

p

avg

ij

=

2

3

p

i

+p

j

−

P

avg

i

P

avg

j

P

avg

i

+P

avg

j

−

P

avg

j

(p

i

−P

avg

i

)

P

avg

i

+P

avg

j

−

P

avg

i

p

j

−P

avg

j

P

avg

i

+P

avg

j

,

(4.3)

where P

avg

i

and P

avg

j

are constants, representing the approximated average inlet

and outlet pressure in the pipeline. In addition, we need an Equation of State

in order to relate the average pressure in the pipeline to the average density in

the pipeline. For a discussion on diﬀerent versions of the Equation of State,

see Modisette (2000). We use the following form:

p = ρ

R

m

Tz (p, T) , (4.4)

where ρ is the density in the pipeline, R is the Gas constant, m is the molec-

ular weight and z (p, T) is the compressibility factor as a function of pressure p

and temperature T. The value of z is chosen to be 0.7 in this paper based on

experience from the pipelines at the Norwegian continental shelf (Dahl 2001).

For a constant temperature T, z is still varying in pressure, but for our relevant

pressure levels the variations are small.

When using this Equation of State, we can ﬁnd the density in the pipeline

as a function of the average pressure in the pipeline. We can then estimate the

line-pack by multiplying the density with the volume of the pipeline:

LP

ij

=

m

RTz

p

avg

ij

A

ij

L

ij

, (4.5)

where LP

ij

is the line-pack in the pipeline between i and j, A

ij

is the area of a

cross-section of the pipeline and L

ij

is the length of the pipeline. The complete

formulation is given in Section 4.3, Equations (4.16)-(4.18).

Compressor modeling

For production ﬁelds we do not consider the physics in the reservoirs. The model-

ing starts when gas enters the compressors that build up pressure for the pipeline

transportation. The cost of running a compressor depends on the inlet and outlet

pressure, and the gas ﬂow. The compressor work, as an isentropic process is given

by (Katz & Lee 1990):

W = Aq

¸

P

out

P

in

χ

χ−1

−1

¸

, (4.6)

where W is the work of the compressor, P

in

is the pressure of the gas entering

the compressor and P

out

is the gas leaving the compressor (corresponds to the

121

Chapter 4 An operational portfolio optimization model...

inlet pressure at the connected pipeline). Here χ is the adiabatic constant and A

represents the aggregation of various constants in the equation. In the following,

the compressor work will be replaced with the cost of running the compressor

assuming a constant energy price.

The function is neither strictly convex nor concave. This makes it diﬃcult

to represent the function in an optimization model. Including the original com-

pressor cost function leads to a non-convex problem. One approach is presented

in Nowak (2006). This method works well for a deterministic model (Nørstebø

2004), and gives a very good approximation of the actual compressor costs. How-

ever, the method is not eﬃcient enough to be used for the model sizes that we are

considering. The solution times for stochastic network models with 3-5 stages,

and from 500 to 1000 scenarios will be too high for practical use. Therefore, to

simplify Equation (4.6) we assume constant input pressure to the compressor.

The compressor cost is linearized with the following constraints:

c

gts

≥ a

1

+b

1

d

gts

−d

g

+c

1

p

gts

−p

g

, g ∈ G, t ∈ T , s ∈ S, (4.7)

c

gts

≥ a

2

−b

2

d −d

gts

−c

2

p

g

−p

gts

, g ∈ G, t ∈ T , s ∈ S, (4.8)

c

gts

≥ a

3

+b

3

d

gts

−d

g

−c

3

p

g

−p

gts

, g ∈ G, t ∈ T , s ∈ S, (4.9)

c

gts

≥ a

4

−b

4

d

g

−d

gts

+c

4

p

gts

−p

g

, g ∈ G, t ∈ T , s ∈ S, (4.10)

Here d

gts

and p

gts

is the production and outlet pressure in ﬁeld g at time t

in scenario s and the overlined and underlined values give the maximum and

minimum values of the variables, respectively. Then c

g

is the approximated

compressor cost at production ﬁeld g. Constants a, b, c are positive and describe

the plane used for approximation.

4.3 The portfolio optimization model

We model the uncertainty in prices and volumes using scenario trees in a 3-stage

stochastic program (see e.g. Kall & Wallace (1994)). It has seven time periods

and daily resolution. We start here with the description of the scenario trees and

the time structure of the model and then give the mathematical formulation.

Scenario structure and rolling horizon

Assume that our model starts on a Monday. The model horizon then comprises

the weekdays from Monday to Sunday. We implement the ﬁrst stage decisions

(the decisions for Monday) and run the model again on Tuesday (with the week-

days from Tuesday to next Monday). The parameters regarding production levels

122

4.3 The portfolio optimization model

for each ﬁeld, remaining production limits for the week and line-pack levels are

updated between the runs. Also the scenario tree structure depends on which

weekday the model is run, see Figure 4.7. The reason for this is the special

structure of the weekend where all markets for Friday, Saturday and Sunday are

cleared on Friday.

Figure 4.7: The diﬀerent scenario trees

Notation

Sets

N The set of all nodes in the network.

G The set of production nodes in the network.

B The set of junction nodes in the network.

M The set of market nodes in the network.

I(n) The set of originating nodes with pipelines going into node n.

O(n) The set of end nodes for pipelines going out of node n.

T The set of time periods.

S The set of scenarios.

Z The set of constraints used to linearize the Weymouth equation.

Indexes

n Index used for nodes in general. When more indexes are needed,

i and j will be used.

g Index for production nodes.

b Index for junction nodes.

m Index for market nodes.

t Time period index.

s Scenario index.

z Index for linearized Weymouth constraints.

123

Chapter 4 An operational portfolio optimization model...

Constants

G

g

The maximum daily production level in ﬁeld g.

G

g

The minimum daily production level.

H

g

The maximum weekly production level.

H

g

The minimum weekly production level.

P

n

The maximum pressure in node n.

P

n

The minimum pressure in node n.

K

ij

The Weymouth constant for the pipeline going from i to j.

PI

z

Fixed point for pressure into a pipeline.

PO

z

Fixed point for pressure out of a pipeline.

A Constant in the compressor work function.

χ Adiabatic constant.

R

m

Price in the take or pay contract in market m.

B

g

Constant to convert compressor work to compressor cost.

P

avg

i

Approximated average pressure into the pipelines in node i.

P

avg

j

Approximated average pressure in the pipelines entering node j.

Decision variables

d

g

Production in ﬁeld g.

q

m

Spot sale in market m.

v

m

Delivery in take or pay contract in market m.

f

ij

The ﬂow between node i and j.

p

n

The pressure in node n.

p

avg

ij

The average pressure in the pipeline going from node i to j.

LP

ij

Line-pack in the pipeline going from i to j.

c

g

Cost of compressor in ﬁeld g.

q

in

ij

Volume inserted to the pipeline going from i to j.

q

out

ij

Volume extracted from the pipeline going from i to j.

Stochastic variables and probabilities

v

m

Nomination in long-term contracts in market m.

r

m

Spot price in market m.

π Probability of a given scenario.

Functions

F (p

i

, p

j

) Flow in a pipeline with inlet pressure p

i

and

outlet pressure p

j

.

W (q, P

in

, P

out

) Work of a compressor.

124

4.3 The portfolio optimization model

The mathematical model

The model maximizes expected income from sales in the spot market and de-

liveries in long-term contracts minus the costs of using the network. The costs

incorporate the compressor costs and production costs:

Π =

¸

s∈S,t∈T

π

ts

¸

m∈M

(r

mts

q

mts

+R

m

v

mts

) −

¸

g∈G

(c

gts

+κ

g

d

gts

)

¸

¸

. (4.11)

Here π

ts

is the probability of scenario s in time t, q

mts

is the volume of gas sold

in market m, while r

mts

is the price obtained for this gas in market m. The

second term gives the revenues from the long-term contracts: v

mts

is the volume

delivered in long-term contracts in market m while R

m

is the price obtained.

The last term gives the cost of the compressors and the production cost. The

parameter κ

g

is the cost per unit of production d

gts

, in ﬁeld g, and c

gts

is the cost

of the compressor in ﬁeld g. The compressor cost is modeled with constraints (4.7)

to (4.10).

At each ﬁeld there is a minimum and a maximum volume that can be produced

by the company. These limitations are due to both the properties of the ﬁeld and

the guidelines from a tactical plan. The limits are given both on a daily and a

weekly level. The total production should be within a given interval. The daily

limits are given by:

G

g

≤ d

gts

≤ G

g

, g ∈ G, t ∈ T , s ∈ S, (4.12)

where G

g

is the lower production limit, G is the upper limit and d

gts

is the

production in the ﬁeld. The weekly limits are formulated as:

H

g

≤

¸

t∈T

d

gts

≤ H

g

, g ∈ G, s ∈ S, (4.13)

where H

g

is the lower limit on weekly production and H

g

is the upper limit on

weekly production.

As discussed in Section 4.2, we model the transportation system as a steady-

state system. The Weymouth-equation is used to relate the pressure diﬀerence

and design parameters of the pipelines to the ﬂow in the pipeline. In order to get a

linear model, we use the same linearization as Tomasgard et al. (2007) and Rømo

et al. (2007). This gives the following constraints for the model without line-pack:

125

Chapter 4 An operational portfolio optimization model...

f

ijts

≤ K

W

ij

PI

z

PI

2

iz

−PO

2

jz

p

its

− K

W

ij

PO

z

PI

2

iz

−PO

2

jz

p

jts

, j ∈ N, i ∈ I(j), z = 1, ..., Z,

(4.14)

where (PI

z

, PO

z

) are break-points for the linearization representing pressure at

both ends of the pipeline. We use Z such constraints to linearize the Weymouth

equation, each with a diﬀerent ratio

PI

z

PO

z

, see Tomasgard et al. (2007). The

inlet pressure in the pipeline is given by p

its

, the outlet pressure is p

jts

and the

resulting ﬂow is f

ijts

. K

W

ij

is the Weymouth constant for the pipeline going from

i to j.

For the model with line-pack it is the net daily change of the volume in the

pipeline that matters, hence we substitute the variable f

ij

with two new variables:

q

in

ij

and q

out

ij

, representing input and output. The Weymouth-equation can then

be written as:

1

2

q

in

ijts

+q

out

ijts

≤ K

W

ij

PI

z

PI

2

iz

−PO

2

jz

p

its

− K

W

ij

PO

z

PI

2

iz

−PO

2

jz

p

jts

, j ∈ N, i ∈ I(j), z = 1, ..., Z.

(4.15)

As discussed in Section 4.2 the line-pack in a pipeline is found by multiplying

the approximated density with the volume of the pipeline.

LP

ijts

=

m

RTz

p

avg

ijts

A

ij

L

ij

, i ∈ N, j ∈ O(i), t ∈ T , s ∈ S, (4.16)

where LP

ij

is the line-pack in the pipeline between i and j, A

ij

is the area of a

cross-section of the pipeline and L

ij

is the length of the pipeline. The line-pack

inventory is given by:

LP

ij,t+1,s

= LP

ijts

+q

in

ijts

−q

out

ijts

, i ∈ N, j ∈ O(i), t ∈ T \{T}, s ∈ S. (4.17)

As a business constraint we require that the level of line-pack at the end of

horizon is equal to the starting level. This is in order to not deplete the line-pack

which is also used as a buﬀer in order to secure supply:

LP

ij,T,s

+q

in

ijTs

−q

out

ijTs

= LP

ij1s

, i ∈ N, j ∈ O(i), s ∈ S, (4.18)

126

4.3 The portfolio optimization model

where T is the last time period. Constraint (4.18) is necessary to take care of

end-of-horizon eﬀects. If the constraint had not been included, the line-pack

level would have been reduced to a minimum in period T. Alternative ways

of formulating this constraint is to include a value of the gas in period T (for

instance based on an Expected Gas Value Function, as presented in Tomasgard

et al. (2007)).

The pressure in the nodes needs to be within maximum and minimum require-

ments. These requirements come from compressor capacities, design parameters

of the network and contractual agreements:

P

i

≤ p

its

≤ P

i

, i ∈ N, t ∈ T , s ∈ S. (4.19)

In the production nodes we must make sure that the produced quantity, d

gts

,

of gas is ﬂowing into the connected pipelines, f

gits

. The formulation without

line-pack is:

d

gts

=

¸

i∈O(g)

f

gits

, g ∈ G, t ∈ T , s ∈ S, (4.20)

and with line-pack it is:

d

gts

=

¸

i∈O(g)

q

in

gits

, g ∈ G, t ∈ T , s ∈ S. (4.21)

In the junction nodes, the amount of gas that enters the node must be equal

to the amount leaving the node. The formulation without line-pack is:

¸

i∈I(k)

f

ikts

=

¸

j∈O(k)

f

kjts

, k ∈ B, t ∈ T , s ∈ S, (4.22)

and with line-pack it is:

¸

i∈I(k)

q

out

ikts

=

¸

j∈O(k)

q

in

kjts

, k ∈ B, t ∈ T , s ∈ S. (4.23)

B is the set of junction nodes in the network.

In the market nodes the company sell q

mts

in the spot market. Additionally, the

company must deliver long-term contracted volumes (v

mts

). The mass balance

equation for the market nodes for the formulation without line-pack is:

q

mts

=

¸

i∈I(m)

f

imts

−v

mts

, m ∈ M, t ∈ T , s ∈ S, (4.24)

and with line-pack it is:

127

Chapter 4 An operational portfolio optimization model...

q

mts

=

¸

i∈I(m)

q

out

imts

−v

mts

, m ∈ M, t ∈ T , s ∈ S. (4.25)

Since it is assumed that the company cannot inﬂuence the prices in the market

nodes, there are no restrictions on how much the company can buy or sell in the

market. The demand is assumed to be completely inelastic, or alternatively the

volumes that the company is buying/ selling are not large enough to inﬂuence

the market price.

In each event node z where uncertainty is resolved in the scenario tree, we

need to add non-anticipativity constraints (Rockafellar & Wets 1991). Let the

scenarios passing through node z be given by S(z) and let T(z) be the time period

of node z. Then we have the following constraints:

1

|S(z)|

¸

s

∈S(z)

(d

gts

, q

mts

, f

ijts

, p

its

, c

gts

, p

avg

ijts

, LP

ijts

, q

in

ijts

) = (4.26)

(d

gts

, q

mts

, f

ijts

, p

its

, c

gts

, p

avg

ijts

, LP

ijts

, q

in

ijts

), z ∈ Z, s ∈ S(z), t ∈ T(z).

The two models then consist of:

1. Model without line-pack: max (4.11) s.t. (4.7)-(4.14), (4.19)-(4.20), (4.22),

(4.24), (4.26)

2. Model with line-pack: max (4.11) s.t. (4.7)-(4.13), (4.15)-(4.19), (4.21),

(4.23), (4.17), (4.25)-(4.26).

4.4 Test case from the Norwegian continental

shelf

The network chosen as a test case is shown in Figure 4.16. In the network there

are three ﬁelds, two junction points and three market nodes. The transportation

capacity of the pipelines is comparable to the dry-gas transportation system in the

North-Sea, but the structure is simpler. The demand in the take-or-pay volumes

are chosen to reﬂect the dominant position of these contracts in the North-Sea.

The chosen market hubs are large hubs in Europe (Zeebrugge, NBP and TTF).

In reality the Title Transfer Facility (TTF) hub is however not connected to the

transportation system from the North-Sea. This market hub is included because

of the liquidity of the market and the availability of price data as a substitute for

less liquid spot markets in the Emden, Dornum and Dunkerque hubs.

128

4.4 Test case from the Norwegian continental shelf

The spot markets

The price scenarios are generated based on time-series of historical observations.

The price data was supplied by the Heren Energy Ltd. The price tends to be

higher during winter than during summer. In addition to the seasonal variations,

there are large upwards peaks in the time series. Being able to deliver extra

quantities of gas when such a peak occurs is advantageous.

N

O

K

/

S

m

3

Figure 4.8: Prices at the three hubs in the winter period 2005.

In this article we focus on the winter season, which we deﬁne to be from

October 1st to March 31st. The price series are illustrated in Figure 4.8. During

the winter period, the prices are high and the ﬂexibility in the network is low.

During the summer season, the nominated quantities of the long-term contracts

are lower and thus the ﬂexibility is larger. The winter period thus requires careful

planning, and the gain from good planning may be substantial. Figure 4.9 shows

the weekly variation over 7 days, and illustrates that the price volatility within

a week can be large.

Before constructing the prediction models for the time-series, the price peaks

(deﬁned as deviations of more than 50% of the average value) are considered as

outliers and are therefore removed and replaced with the average value of the

price on the same weekday in the previous and coming week.

Forecasting and scenario generation

In each node in the scenario-tree, there is a spot price and a take-or-pay volume

for each of the market hubs (the method to construct scenarios for the take-or-

129

Chapter 4 An operational portfolio optimization model...

Figure 4.9: Illustration of the weekly variations in the last week of February 2005.

pay volumes is given in the next section). We use AR(2)-models to predict the

spot price in all market nodes (see Figure 4.10).

historical

data

predicted

data

x

t−2

x

t−1

x

t

ˆ x

t+1

ˆ x

t+2

ˆ x

t+3

ˆ x

t+4

ˆ x

t+5

ˆ x

t+6

ˆ x

t+j

=

α +

2

i=1

β

i

x

t+j−i

α +β

1

ˆ x

t+j−1

+β

2

x

t+j−2

α +

2

i=1

β

i

ˆ x

t+j−i

if j = 1

if j = 2

if j > 2

Figure 4.10: The AR(2) prediction model.

An example of how one of the AR(2)-models ﬁts the real data is shown in Fig-

ure 4.11. The models seem to represent the time-series development reasonably

good and are unbiased. This is important for the scenario generation method

we have chosen. Such price models will in general track the real data process

by following it closely but in general lag a bit behind downward and upward

movements. The AR(2) model is parameterized so that the expected error is 0,

and then the scenarios are there to describe the historical deviations from the

forecast.

We estimate the ﬁrst four moments (expectation, variance, skewness and kur-

tosis) and correlation of the prediction error distributions for the markets based

on the historical data. We then use a moment matching procedure that was de-

veloped in Høyland et al. (2003) to generate our scenarios. Figure 4.12 shows an

example of a scenario tree for the prediction error. The indexes f

1

and f

2

give

the number of branches in the ﬁrst and second stage, respectively. The value of

stage, branch

t

is zero in all nodes in the scenario tree, except for the nodes in the

ﬁrst period in a new stage (corresponding to period t+1 and t+6 in Figure 4.12).

130

4.4 Test case from the Norwegian continental shelf

Figure 4.11: Real prices and values obtained from the prediction model (here

illustrated for the NBP).

We generate S multivariate scenarios for the prediction error with the correct cor-

relation between the markets and with correct moments for the individual error

terms.

11

t+1

(1,f

1

)

t+1

21

t+6

(2,f

2

)

t+6

(2,S−f

2

)

t+6

(2,S)

t+6

Figure 4.12: The scenario tree for the prediction errors.

Finally, we combine the AR(2) prediction model with the scenario tree for the

prediction errors to one scenario tree. Each scenario presents a path from the

root node to the leaf node (there are S unique paths through the tree). This

is illustrated in Figure 4.13. The predicted values, ˆ x

t+j

can depend on both

historical data and predicted data (ˆ x

t+2

for instance depends on both x

t

and

ˆ x

t+1

) (see Figure 4.10). The value in each node in a path through the scenario

tree can then easily be found by the formula shown in the ﬁgure. Hence we use

the forecasting method to predict the expected price, and scenario generation to

describe the variation (error) around this price.

131

Chapter 4 An operational portfolio optimization model...

An important issue in stochastic models is the information structure. The

nonanticipativity constraints make sure that decisions at time t can only depend

on information available at this time, see for example Ruszczyński (1997). We

have implemented the nonanticipativity constraints in Equation (4.26). In Fig-

ure 4.13 the nonanticipativity constraints are represented by the ellipsoids.

ˆ x

s

t+j

=

α +

2

i=1

β

i

x

s

t+j−i

+

s

t+j

α +β

1

ˆ x

s

t+j−1

+β

2

x

s

t+j−2

+

s

t+j

α +

2

i=1

β

i

ˆ x

s

t+j−i

+

s

t+j

if j = 1

if j = 2

if j > 2

t t + 1 t + 2 t + 3 t + 4 t + 5 t + 6

Figure 4.13: All paths through the scenario tree. The ellipsoids represent the

nonanticipativity equations.

Take-or-pay volumes

For the take-or-pay contracts, we assume that the company’s customers can be

divided into two groups: Group 1 comprises gas purchasers with large delivery

commitments in Europe. Group 2 contains purchasers that have more ﬂexibility

to utilize the TOP-contract as a call option.

The Group 2 customers will nominate the maximum amount given that the

spot price exceeds the TOP-price (plus the transaction costs), and the minimal

amount when the spot price is lower than the TOP-price.

For Group 1 we assume that the nominated volume is a linear function of the

spot price within a certain range. This means that they will nominate a high

volume given that the spot-price is high, and nominate a low volume in case of a

low spot-price. The ﬂuctuations in nomination is however less extreme than for

Group 2. This is illustrated in ﬁgure 4.14.

A large percentage of the gas in the North-Sea is still sold through long-term

contracts. We have reﬂected this in the model by choosing the expected demand

in the take-or-pay contracts to be a large percentage of the total production

capacity in the ﬁelds.

132

4.4 Test case from the Norwegian continental shelf

Figure 4.14: The two diﬀerent customer types.

The supply side

Most gas ﬁelds in the North-Sea produce both oil and gas. There are however

eight pure gas ﬁelds. The total export of natural gas from Norway to Europe

in 2005 were 82.5 billion Sm

3

. In our model, we have included approximately

60% of this volume in a simpliﬁed network structure. The transportation network

consists of long, subsea pipelines that are operated at high pressure. An overview

of the production capacities and pipelines characteristics in the North-Sea can be

found in OED (2006). Currently, as much as 90% of the gas is sold through long-

term contracts. It is however expected that this amount will decrease, and that

the trade in short-term markets will increase. In our case study the TOP-volume

is in average 60% of the total production.

When running the model on a rolling horizon and implementing the ﬁrst stage

decisions (decisions for the ﬁrst day), the model with line-pack may end with a

line-pack inventory that is lower or higher than the starting level. In order to

compare the models with and without line-pack, one has to assign a value to this

line-pack diﬀerence. One option would be to use the price of gas at the last day

in the test period for which the model is run, but since the prices are volatile

this is an unstable method. Changes of one day in the length of the test period

may have substantial eﬀect on the overall return both in positive and negative

direction in this case. Instead, we use the average price in the test period. We use

two measures when comparing the results from the models: proﬁt from the spot

market (after the inventory value adjustment) and the average obtained price in

the spot market. In the following, all gas volumes are given in million standard

cubic meters and the proﬁts are given in million Norwegian kroner (NOK).

133

Chapter 4 An operational portfolio optimization model...

Price Model 1 (PM1) Price Model 2 (PM2)

Rolling horizon T T + 1 T T + 1

AR(2)-parameters T −3 T −3, T −2 T −1 T

Prediction errors T −2, T −1 T −1, T T −1 T

Table 4.1: Time periods used to estimate AR(2)-parameters and the prediction

errors for our two price models.

Test instances

We have tested the model on real price data, with a rolling horizon over 60 plan-

ning days (72 calendar days). Figure 4.15 shows the two diﬀerent time periods

considered in this paper: October 3rd 2005 - December 24th 2005 (period T) and

January 2nd 2006 - March 25th 2006 (period T + 1). We have used two diﬀer-

ent approaches for parameterizations of the price models. In PM1 we use two

diﬀerent time periods for the parameterization of the AR-model and for ﬁnding

the error distribution. See Table 4.1 for an overview. In PM2 we use as recent

data as possible and the same data for both the parameterization and the error

estimation. In order to capture the eﬀect of price peaks, we have in some of the

cases introduced extreme scenarios: one scenario where the price is doubled and

one scenario where the price is halved for the entire week. In addition, shorter

ﬂuctuations are added to some scenarios (doubling or halving of the price for one

or more time periods). We have also tested the eﬀect of updating the prediction

errors on a monthly basis.

Figure 4.15: The time periods considered in this article.

4.5 Computational results and discussion

In order to ensure feasibility for all daily runs on the rolling-horizon, we introduce

a penalty cost for not meeting the take-or-pay requirements on a given day, as well

as for not being able to meet the line-pack-requirements. These penalty costs are

necessary for the deterministic model; otherwise commitments from previous runs

on the rolling horizon can lead to infeasible problems. The stochastic versions

will most often have taken the possibility of a price increase into consideration.

When comparing the results from the models, the penalty costs are not included.

134

4.5 Computational results and discussion

Figure 4.16: The network considered in this paper.

Perfect information

We start the analysis by looking at a situation where the producer knows the exact

values for the price and TOP-volumes for the next seven days. The solution with

this input data gives insight to the value of line-pack, as well as how the solution

of the model depends on the starting level of the line-pack. This solution also

gives a benchmark to compare the results of our model run with stochastic prices.

The results are presented in Table 4.2 (the models with line-pack are started

with the pipelines ﬁlled to the capacity limit). As we can see from these results,

the value of line-pack depends on which period we look at. For autumn 2005,

T, the diﬀerence is 1.72%. For spring 2006 however, the diﬀerence is 13.97%.

The price data for spring 2006 are much more volatile than for autumn 2005.

As expected, the added volatility in the spot prices increases the value of the

inherent ﬂexibility in the line-pack.

We then look at the importance of the starting level of the line-pack in the

system. In Figure 4.17 we have looked at a situation where we initiate the model

with a line-pack level of 90%. In the ﬁgure, the resulting line-pack utilization is

compared with a model where the starting level is at 100%.

As expected, both models show the same behavior: the line-pack increase in

periods with low price and decrease in periods with high prices. The results from

optimizing with the two diﬀerent starting levels are shown in Table 4.3. In our

test-case, the model with a starting line-pack level of 90% of total capacity has

higher proﬁts. The reason is that this model has higher ﬂexibility with respect

to storage utilization (the target level for line-pack at the end of the week is at

least 90% which is a weaker requirement). We get an improved income from the

135

Chapter 4 An operational portfolio optimization model...

Case 1 Case 2

Time period T T

Model type With line-pack Without line-pack

Spot market income 14573.94 14205.71

Average price 2.388 2.388

Average obtained price 2.927 2.833

Adjusted with average price 14450.40 14205.71

Case 3 Case 4

Time period T + 1 T + 1

Model type With line-pack Without line-pack

Spot market income 18806.26 15994.74

Average price 2.927 2.927

Average obtained price 3.777 3.344

Adjusted with average price 18228.61 15994.74

Table 4.2: Spot market income for the optimization model run with perfect

information.

Figure 4.17: Utilization of line-pack in the network for the model with perfect

information (spring 2006).

136

4.5 Computational results and discussion

spot market of 4.0%.

Generally, we can conclude that on a ﬁxed horizon it is beneﬁcial to start

with a high level of line-pack and have a low target level at the end-of-horizon.

When the start- and target-level are the same, a high inventory requirement is

not necessarily beneﬁcial.

Case 5 Case 6

Time period T + 1 T + 1

Line-pack start level 100% 90%

Spot market income 18806.26 18816.68

Average price 2.927 2.927

Average obtained price 3.777 3.978

Adjusted with average price 18228.61 18968.62

Table 4.3: Results when running the model with diﬀerent starting line-pack lev-

els.

Stochastic versus deterministic models

In the following we compare the results of the stochastic model based on fore-

casting and scenario generation with the deterministic model based on forecasting

alone. We use the perfect information case as a benchmark. The stochastic ver-

sion uses 900 scenarios to describe price uncertainty in the 3 markets for the

next 7 days. We tested the diﬀerent combinations of parameter estimations and

error estimations described in Section 4.4. For each of the instances, we present

aggregated proﬁt from both time periods. The results for PM1 are shown in Ta-

ble 4.4. The largest diﬀerence between the stochastic and deterministic version

is found for the model including line-pack. The diﬀerence is here 4.64%, while

the diﬀerence is 0.41% for the model without line-pack. The diﬀerence up to

the benchmarks for these models are however large (11.12% and 5.72% for the

stochastic models).

We then added extreme scenarios to the stochastic model, as well as the option

to update the distribution of prediction errors on a monthly basis. The best result

for the model with line-pack was achieved when only adding outliers, the distance

to the benchmark was 7.94% and the value was 7.05% larger than the result for

the deterministic model. For the model without line-pack, the best combination

turned out to be using both outliers and monthly update of the prediction error.

The distance to the benchmark for this result was 4.76%, and the result was

2.34% larger than the deterministic model.

Further, we estimated the AR(2)-models and prediction errors on the winter

period most recent to the rolling horizon using PM2. The results are given in

137

Chapter 4 An operational portfolio optimization model...

Case 7 Case 8

Price model PM1 PM1

Model type With line-pack With line-pack

Uncertainty Stochastic Deterministic

Spot market income 30317.06 30199.84

Average obtained price 3.022 2.882

Average price 2.658 2.658

Adjusted with average price 29408.23 28104.57

Case 9 Case 10

Price model PM1 PM1

Model type Without line-pack Without line-pack

Uncertainty Stochastic Deterministic

Spot market income 28566.24 28450.87

Average obtained price 2.946 2.934

Table 4.4: Comparison of the deterministic and stochastic model with and with-

out line-pack. The results are aggregated for T and T + 1.

Table 4.5. As we can see from these tables, the stochastic models still give better

results than the deterministic models, but the diﬀerence has decreased. For the

model with line-pack, the diﬀerence is now 0.52%, while for the model without

line-pack the diﬀerence is almost zero (0.013%). The distance to the benchmark

is now, respectively, 7.1% and 5.5%. We see that the distance to the benchmark

has decreased for the model with line-pack, but actually increased for the model

without line-pack.

Again, we introduce outliers as well as monthly updates of the prediction error.

For the model with line-pack, the best result is obtained when the prediction

error is updated monthly. The change is however quite small - the distance to

the benchmark solution is 7.07% and the distance to the deterministic solution

is 0.56%. For the model without line-pack, the best solution was also achieved

when the prediction error was updated monthly. The distance to the benchmark

solution was now 5.36% and the distance to the deterministic solution was 0.21%.

The diﬀerence between the models with and without line-pack is quite clear

in all our results. If we compare the best models with line-pack (PM2, monthly

updated error distribution) with the best model without line-pack (PM1, extreme

scenarios and monthly updated error distributions) we get an advantage of 5.6%,

or 1607 million NOK. For the benchmark models, the same comparison gave a

diﬀerence of 8.21%.

The diﬀerence between the stochastic models and the deterministic version is

less clear. Still, the stochastic version is consistently outperforming the deter-

138

4.6 Conclusions

Case 11 Case 12

Price model PM2 PM2

Model type With line-pack With line-pack

Uncertainty Stochastic Deterministic

Spot market income 30934.25 30795.71

Average obtained price 3.126 3.108

Average price 2.658 2.658

Adjusted with average price 30355.84 30199.84

Case 13 Case 14

Price model PM2 PM2

Model type Without line-pack Without line-pack

Uncertainty Stochastic Deterministic

Spot market income 28525.01 28521.45

Average obtained price 2.942 2.941

Table 4.5: Comparison of the deterministic and stochastic model with and with-

out line-pack. The results are aggregated for T and T + 1.

ministic version. For some combinations of parameter and error estimation we

gain a lot from using the stochastic model. The largest diﬀerence is found for

Price Model 1 in spring 2006, the stochastic solution is 12.84% larger than the

deterministic solution in this case. The stochastic model seems to be more ro-

bust than the deterministic model when it comes to choice of parameters and

time period for error estimation. Also we observed that the deterministic model

has positive values for the penalty functions in some cases, indicating that the

line-pack end-of-horizon constraints were violated. This did not happen for any

of the stochastic models.

When we initiated the system with 90% line-pack level, the results were clearly

inferior to the models started with 100% line-pack level. This result indicates that

the price models are not accurate enough to utilize the added ﬂexibility from the

reduced target storage.

4.6 Conclusions

We present a stochastic portfolio optimization model for operational planning in

natural gas value chains. We compare the results from the model both with and

without storage in the pipelines and with and without stochasticity (prices and

TOP-quantity). We have used real market prices, and tested the model on two

time periods: fall 2005 and spring 2006.

The results show that modelling storage in the model can increase the prof-

139

Chapter 4 An operational portfolio optimization model...

itability of the system substantially when operated optimally. With perfect infor-

mation, the added proﬁts were as high as 13.97% for the 60-day period in spring

2006. This indicates a large commercial value of actively using the storage inherit

in the pipelines (line-pack) to maximize proﬁts.

We have also seen that taking into account the stochasticity in the problem

can, in some instances, lead to large gains in the objective function value. In

addition, the deterministic model was infeasible for some days on the rolling-

horizon (not able to deliver in TOP-contracts and/or not able to meet the target

level for line-pack) for some of the tests we did. We also found that the results

from the deterministic model were more sensitive to the value of the parameters

in the prediction model than the results from the stochastic model.

There are many possible extensions of this work. Firstly, it would be interest-

ing to compare the results from this model with a transient ﬂow model. With

a transient ﬂow model, also shorter time periods (such as hours) could be con-

sidered. Secondly, market power in some, or all, hubs could be introduced (with

uncertainty in the price elasticity in the markets). Thirdly, we can relax the

system perspective on the value-chain, and see how the inclusion of one or more

producers would inﬂuence the results.

140

Bibliography

Byers, J. W. (2006), ‘Commodity storage valuation: A linear optimization based

on traded instruments’, Energy Economics 28, 275–287.

Campbell, J. (1992), Gas Conditioning and Processing: The Equipment Modules,

John M. Campbell & Company, Norman, Oklahoma, USA.

Dahl, H. J. (2001), Norwegian Natural Gas Transportation Systems. Operations

in a Liberalized European Gas Market, PhD thesis, NTNU, Trondheim, Nor-

way.

Gabriel, S. A., Zhuang, J. & Kiet, S. (2005), ‘A large-scale complementarity

model of the North American natural gas market’, Energy Economics 27, 639–

665.

Gabriel, S. & Smeers, Y. (2005), ‘Complementarity problems in restructured

natural gas markets’, CORE Discussion Paper No. 2005/37.

Goel, V. & Grossmann, I. E. (2004), ‘A stochastic programming approach to

planning of oﬀshore gas ﬁeld developments under uncertainty in reserves’, Com-

puters & Chemical Engineering 28, 1409–1429.

Haugen, K. K. (1996), ‘A stochastic dynamic programming model for scheduling

of oﬀshore petroleum ﬁelds with resource uncertainty’, European Journal of

Operational Research 88, 88–100.

Høyland, K., Kaut, M. & Wallace, S. W. (2003), ‘A heuristic for moment-

matching scenario generation’, Computational Optimization and Applications

24(2-3), 169–185.

Iyer, R. R., Grossmann, I. E., Vasantharajan, S. & Cullick, A. S. (1998), ‘Opti-

mal planning and scheduling of oﬀshore oil ﬁeld infrastructure investment and

operations’, Ind. Eng. Chem. Res. 37, 1380 –1397.

Kall, P. & Wallace, S. W. (1994), Stochastic Programming, John Wiley & Sons,

Chichester.

Katz, D. & Lee, R. (1990), Natural Gas Engineering, McGraw-Hill, New York,

USA.

141

Bibliography

Kelling, C., Reith, K. & Sekirnjak, E. (2000), A practical approach to transient

optimization for gas networks, Technical report, PSIG.

Martin, A., Möller, M. & Moritz, S. (2006), ‘Mixed integer models for the sta-

tionary case of gas network optimization’, Mathematical Programming 105(2-

3), 563–582.

Midthun, K. T., Bjørndal, M., Smeers, Y. & Tomasgard, A. (2007), Capacity

booking in a transportation network with stochastic demand and a secondary

market for transportation capacity. Working paper at NTNU, Trondheim,

Norway.

Midthun, K. T., Bjørndal, M. & Tomasgard, A. (2006), Modeling optimal eco-

nomic dispatch and ﬂow externalities in natural gas networks. Working paper,

NTNU, Trondheim, Norway. Submitted to international journal.

Modisette, J. L. (2000), Equation of state tutorial, Technical report, PSIG.

Nørstebø, V. S. (2004), Modeling short term planning in natural gas networks

(in Norwegian), Master’s thesis, NTNU, Trondheim, Norway.

Nowak, M. (2006), Compressor approximation. Working paper, NTNU, Trond-

heim, Norway.

Nowak, M. & Westphalen, M. (2003), A linear model for transient gas ﬂow,

in ‘Proceedings of ECOS’, pp. 1751–1758. available as SINTEF-report STF

38S03601, Trondheim, Norway.

Nygreen, B., Christiansen, M., Bjørkvoll, T., Haugen, K. & Kristiansen, Ø.

(1998), ‘Modelling norwegian petroleum production and transportation’, An-

nals of Operations Research 82, 251–267.

OED (2006), ‘Fakta norsk petroleumsverksemd 2006’.

Osiadacz, A. (1983), ‘Optimal numerical method for simulating dynamic ﬂow

of gas in pipelines’, International Journal for Numerical Methods in Fluids

3, 125–135.

Rømo, F., Tomasgard, A., Fodstad, M., Hellemo, L. & Nowak, M. (2007), An

optimization system for gas transportation on the norwegian continental shelf.

Working paper, NTNU, Trondheim, Norway.

Rockafellar, R. T. & Wets, R. J.-B. (1991), ‘Scenarios and policy aggregation in

optimization under uncertainty’, Mathematics of Operations Research 16, 119–

147.

142

Bibliography

Ruszczyński, A. (1997), ‘Decomposition methods in stochastic programming’,

Mathematical Programming 79, 333–353.

Selot, A., Kuok, L. K., Robinson, M., Mason, T. L. & Barton, P. I. (2007), A

short term operational planning model for upstream natural gas production

systems. to appear in AIChE Journal.

Streeter, V. L. & Wylie, E. B. (1970), ‘Natural gas pipeline transients’, Society

of Petroleum Engineers Journal, December, 357–364.

Thompson, M., Davison, M. & Rasmussen, H. (2003), Natural gas storage valu-

ation and optimization: A real options approach. Working paper, University

of Western Ontario, Canada.

Tomasgard, A., Rømo, F., Fodstad, M. & Midthun, K. (2007), Optimization

models for the natural gas value chain, in G. Hasle, K.-A. Lie & E. Quak,

eds, ‘Geometric Modelling, Numerical Simulation, and Optimization: Applied

Mathematics at SINTEF’, Springer, chapter Optimization Models for the Nat-

ural Gas Value Chain.

Ulstein, N., Nygreen, B. & Sagli, J. (2007), ‘Tactical planning of oﬀshore

petroleum production’, European Journal of Operational Research 176, 550–

564.

Van der Hoeven, T. (2004), Math in Gas and the Art of Linearization, PhD thesis,

International business school and research center for natural gas, Groningen,

The Netherlands.

Westphalen, M. (2004), Anwendungen der Stochastischen Optimierung im

Stromhandel und Gastransport, PhD thesis, University Duisburg-Essen (Ger-

many).

Zhuang, J. & Gabriel, S. (2006), A complementarity model for solving stochastic

natural gas market equilibria. January 2006, Energy Economics, accepted.

143

Paper IV

Kjetil T. Midthun, Mette Bjørndal, Asgeir Tomasgard and Yves Smeers:

Capacity booking in a

Transportation Network with

stochastic demand and a

secondary market for

Transportation Capacity

Chapter 5

Capacity booking in a Transportation

Network with stochastic demand and a

secondary market for Transportation

Capacity

Abstract:

We present an equilibrium model for transport booking in a gas transporta-

tion network. The booking regime is similar to the regime implemented in

the North-Sea. The model looks at the challenges faced by the network op-

erator in regulating such a system. There are some privileged players in the

network, with access to a primary market for transportation capacity. The

demand for capacity is stochastic when the booking in the primary market

is done. There is also an open secondary market for transportation capacity

where all players participate including a competitive fringe. We consider

diﬀerent objective functions for the network operator, and the diﬀerence

between setting ﬁxed capacities and modeling the pressure constraints in a

sub-sea pipeline-network. This is modelled as a Generalized Nash Equilib-

rium using a stochastic complementarity problem.

5.1 Introduction

We study booking of transportation capacity in a natural gas network with sev-

eral large players and a competitive fringe. The oﬀshore pipeline system in the

North-Sea provides a case for our analysis, but the model and results are interest-

ing for natural gas transportation in general. There are two booking stages in the

transport capacity market. In the ﬁrst stage the large producers book capacity

within their predeﬁned capacity rights. In the second stage there is a redistri-

bution of capacity in a bilateral secondary market, where also the competitive

fringe participates. Here the network operator can sell remaining capacity in the

system, and capacity bought in the ﬁrst-stage primary market can be sold by the

producers.

The purpose of the paper is to develop a model that can be used to analyze

how diﬀerent objective functions for the system operator aﬀect the eﬃciency of

the transportation system. We also investigate the eﬀect of using diﬀerent model

147

Chapter 5 Capacity booking in a Transportation Network...

representations of the physical properties of the transport network. Another

interesting topic is how stochasticity in the price for natural gas inﬂuences our

results. The model is based on Generalized Nash Equilibrium and is represented

as a stochastic complementarity problem. To our knowledge this is the ﬁrst time

the booking system for natural gas transportation is studied using this approach.

The network operator inﬂuences the eﬃciency in the network through the rout-

ing. The routing decisions will also determine the capacity sold in the secondary

market. This is diﬀerent from the role of the network operator compared to the

articles studying electricity networks, by for instance Yao et al. (2004) and Hu

et al. (2004) where the network operator choose the production from each pro-

ducer in order to maximize social surplus. In the North-Sea, the network operator

acts as a neutral third party.

We formulate the model as a mixed complementarity problem, see for exam-

ple Ferris & Pang (1997) and Facchinei & Pang (2003). A path-breaking paper for

the use of complementarity problems modelling economic equilibrium was Lemke

& Howson (1964). In the energy sector there are numerous examples of papers

using complementarity problems to model and solve economic equilibria. Gabriel,

Zhuang & Kiet (2005) presents a linear complementarity equilibrium model for

the North American natural gas market. Gabriel, Kiet & Zhuang (2005) presents

a multi-seasonal, multiyear mixed nonlinear complementarity problem of natu-

ral gas markets. Smeers (2003a) and Smeers (2003b) discuss the deregulation of

the electricity markets and the organization of regional electricity transmission.

In Jing-Yuan & Smeers (1999) spatial oligopolistic electricity models are given

and Generalized Nash Equilibria are found in a system with Cournot generators

and regulated transmission prices. Yao et al. (2006) presents a model of two-

settlement electricity markets using an Equilibrium Problems with Equilibrium

Constraints (EPEC). Hu et al. (2004) model strategic bidding by generators to an

ISO that is maximizing social surplus. The loop ﬂow is taken into consideration

and shown to be important for the results. The model is an EPEC solved as an

All-KKT system in PATH. Hobbs (2001) presents Cournot models of bilateral

power markets.

In Section 5.2 we discuss the background for this article, as well as the as-

sumptions we have made. The model formulation as a stochastic Mixed Com-

plementarity Problem is presented in Section 5.3. A more detailed description of

the equilibrium conditions is given in Appendix 5.A. The properties of the model

are discussed in Section 5.4. We then move on to some numerical examples in

Section 5.5. Finally, the conclusions are given in Section 5.6.

148

5.2 Problem description and assumptions

5.2 Problem description and assumptions

We present here the ideas and motivation for our case analysis, the assumptions

we have made and the reason for introducing them.

System in the North-Sea

We study a system with ﬁeld nodes, each with a set of large producers in addition

to a competitive fridge. The producers deliver natural gas into a transportation

network passing through junction nodes and ending in market nodes. The market

for capacity in this network is managed by an independent system operator (ISO)

named Gassco. The producers book transport capacity from ﬁeld to market and

can not determine the actual routing of the gas through the network. The routing

is the responsibility of the ISO. The image on the left in Figure 5.1 illustrates

the point-to-point perspective of the producers. The transportation network can

be considered as a black box for the producers. The system operator operate

the network taking into account the details in the network, as illustrated in the

image on the right in Figure 5.1.

At the Norwegian Continental Shelf (NCS) capacity distribution is done in

a primary market, and the remaining capacity after this initial distribution is

handled through a secondary market. In the secondary market, both transactions

of capacity facilitated by the ISO and bilateral transactions between shippers are

included. The secondary market is open to all qualiﬁed shippers. Only the large

producers book capacity in the primary market limited by predeﬁned capacity

rights. This booking right depends on their need to transport induced by the

TOP contracts. The actual demand for capacity due to the TOP contracts is

uncertain until delivery. In sum, the available capacity in the primary market is

actually larger than the total capacity in the network. If a conﬂict arises with

respect to over-booking, a capacity allocation key is used to resolve these matters.

We have not explicitly modeled this rule in this paper.

In addition to the long term contracts for gas in the markets nodes, there are

short-term markets where gas may be sold. In this article we have assumed that

the producers may act strategically in the transport capacity market, but that

they are price-takers in the spot markets in the market nodes. This is reasonable

as Norway’s overall production is around 15% of the European consumption of

natural gas. The main market hubs are in UK, Germany, France and Belgium.

In the market hubs there are large buyers of natural gas who distribute the gas

further to either the suppliers or end-customers. In our model, the analysis ends

at the market hubs. For details regarding the liberalization of the European gas

market see European Union (1998) and European Union (2003), and for details

on the Norwegian case, see Austvik (2003).

149

Chapter 5 Capacity booking in a Transportation Network...

The purpose of our model is mainly to analyze the eﬀect diﬀerent objectives

of the ISO will have on the operation of the system. The price the ISO can

take is regulated and ﬁxed both in the primary and in the secondary market,

so its decision variables are only routing and secondary market sales of available

capacity. If we represent the ISO with a feasibility problem, the corresponding

game will have an inﬁnite amount of equilibria. For each choice of secondary

market sales from the ISO, a solution satisfying the large producers’ equilibrium

conditions can be found. Hence we focus on the following alternatives: max ﬂow,

max value of ﬂow and max social surplus. In the following we assume that the

ISO does not have economic interests in the routing, and acts as a benevolent

central planner.

We also investigate how the representation of the physical networks as well

as the booking rights in the primary markets will inﬂuence the eﬃciency of the

network.

Figure 5.1: The ﬁeld nodes are denoted by g, junction nodes by j and market

nodes by m. The gas ﬂows from top to bottom.

Second-stage decision structure

Our model is a one level game where each of the producers decison problem

is a stochastic two-stage program with recourse (Kall & Wallace 1994). The

stochastic elements are the spot price in the markets and the quantity in the

150

5.2 Problem description and assumptions

TOP-contracts. The uncertainty is modeled with scenarios (see Figure 5.3). The

decisions in the two stages are illustrated in Figure 5.2.

In the secondary market (in the second stage) we assume that the large produc-

ers and the competitive fringe make simultaneous volume decisions in a Cournot

manner. Each of the large producers recognizes that they will inﬂuence the price

for transportation capacity, but make independent volume decisions. The players

in the competitive fringe are price takers in the capacity market. Their reaction

function is expressed as their demand for transportation capacity at a given trans-

portation price. This demand is positive as long as the market price for natural

gas in a market hub is higher than the marginal production cost for the compet-

itive fringe in a ﬁeld node plus the transportation price from that node to the

market.

Further, we assume that the ISO’s decisions are made simultaneously with the

producers. Hence, the ISO is a Cournot player whose volume decisions cannot

be manipulated by other players strategically. An alternative would be to model

this as a multi-leader one-follower Stackelberg game (Yao 2006) with the ISO as

a follower. A common way of modeling this follower situation, when the ISO has

a convex optimization problem, is by including the KKT-conditions for the ISO’s

routing and capacity release in the other players’ optimization problem. They

will then act strategically because they anticipate the ISO’s reaction to their own

volume decisions. In this case each player solves a mathematical program with

equilibrium constraints (MPEC, Luo et al. (1996)) and the resulting game over

all the players become an EPEC. In our approach we stay within the framework

of Mixed Complementarity Problems as all decisions are simultaneous, and a

common way of modeling this is to merge all the players KKT-conditions into

a large complementarity system. We think that the setting with simultaneous

decisions is closer to the reality of the Norwegian continental shelf. Firstly, the

players are not supposed to act strategically, for example in terms of inﬂuencing

the ISO in the transportation market. Secondly, the other players never know or

get information about the ISO’s routing decisions. This is conﬁdential informa-

tion, and so are the booking requests, sales and production volumes of the other

players.

First-stage decision structure

In the ﬁrst stage each of the large producers decides on a booking volume. This

booking decision is based on maximizing the excepted revenue for the second

stage where production and transportation strategies are made as well as trades

in the secondary market for transportation capacity.

We assume that each player makes his ﬁrst-stage decisions and his second-stage

decisions simultaneously. In practice this means that the second-stage decisions

151

Chapter 5 Capacity booking in a Transportation Network...

will depend on the outcomes of the stochastic variables, but the contingent strat-

egy covering all possible outcomes is made before the player observes the other

players booking. Each producer’s optimization problem is then a stochastic two-

stage program with recourse, given the other players ﬁxed decisions. The overall

problem is still a Mixed Complementarity Problem, often called a Stochastic

Mixed Complementarity Problem because of the stochastic variables and two-

stage structure.

If, on the other hand the booking decisions had been used strategically by the

players, we would need to include the second-stage equilibrium over all the players

as a part of the booking problem in the ﬁrst stage for each player. Normally this is

done by including the KKT-conditions from the second stage equilibrium in each

player’s ﬁrst-stage optimization problem. In such a setting each player’s problem

would be a stochastic MPEC, where the second-stage equilibrium conditions for

each scenario is part of the ﬁrst-stage optimization problem and parameterized

on the ﬁrst stage decisions (Patriksson & Wynter 1999).

When the ﬁrst- and second-stage decisions are made simultaneously we model

the situation where either a player does not know the other players’ booking

decisions when he makes his second-stage decisions, or he has this booking infor-

mation but does not let it inﬂuence his second stage decisions. In the Norwegian

regime with a conﬁdential booking process, we feel that this is a sound model.

Then the only information revealed (or acted on) between the ﬁrst and second

stage is the uncertainty that is resolved. This is a one level game as the scenarios

are independent of the ﬁrst-stage decisions.

Figure 5.2: The sequencing of decisions.

152

5.2 Problem description and assumptions

Figure 5.3: The scenario structure in the large producers’ decision problem

153

Chapter 5 Capacity booking in a Transportation Network...

5.3 Model

We start by introducing the notation. We then move on to a discussion of the price

of transportation capacity in the secondary market. The networks we present are

connected graphs.

Notation

Sets

N The set of all nodes in the network.

G The set of ﬁeld nodes in the network.

J The set of junction nodes in the network.

M The set of market nodes in the network.

I(n) The set of nodes with pipelines going into node n

(predecessor nodes).

O(n) The set of nodes with pipelines going out of node n

(successor nodes).

L The set of large producers in the network.

¯

L

g

The set of all producers in ﬁeld g (including the

competitive fringe).

S The set of scenarios.

Indexes

n Used for nodes in general.

g Index for ﬁeld nodes.

j Index for junction nodes.

m Index for market nodes.

s Scenario index.

l The index used for producers.

154

5.3 Model

Constants

R

n

The maximum pressure in node n.

R

n

The minimum pressure in node n.

K

ij

The Weymouth constant for the pipeline going from i to j.

B

lgm

Booking limit for producer l from ﬁeld g to market m.

P

lm

Price in the long term contracts for producer l in market m.

T

gm

Per unit tariﬀ for transportation between ﬁeld g and market m.

MC

g

Aggregated marginal cost parameter in ﬁeld g.

C

ni

Capacity in the pipeline between node n and i.

c

g

Parameter in the cost function for the competitive fringe in ﬁeld g.

c

lg

Parameter in the cost function for producer l in ﬁeld g.

Decision variables

b

lgm

Booking in the primary market by producer l between ﬁeld g and

market m.

d

lgs

Production in ﬁeld g by producer l in scenario s.

q

lms

Spot sale in market m by producer l in scenario s.

h

lgms

Capacity between g and m traded by producer l in the

secondary market in scenario s.

f

nis

The ﬂow between node n and i in scenario s.

r

ns

The pressure in node n in scenario s.

z

gms

Capacity sold by the ISO in the secondary market between ﬁeld g

and market m in scenario s.

t

gms

Price of transportation capacity between ﬁeld g and market m

in scenarios s.

x

gms

Quantity produced in ﬁeld g and sold in market m in scenario s

by the competitive fringe.

Stochastic variables and probabilities

v

lms

Nomination in long-term contracts in market m.

p

ms

Spot price in market m.

φ

s

Probability of a given scenario.

Functions

C

lg

(d) The cost function for producer l in ﬁeld g.

W

g

(y) Cost function for the competitive fringe in ﬁeld g.

155

Chapter 5 Capacity booking in a Transportation Network...

Price of capacity in the secondary market

The price in the secondary market in a node is given by a demand function from

the competitive fringe in that node. We assume that the competitive fringes

in the diﬀerent ﬁeld nodes are independent. The competitive fringe’s demand

function for transportation capacity between ﬁeld g and market m in scenario s

is then found from the proﬁt maximization problem for the competitive fringe in

ﬁeld g:

Π

gs

= max

¸

m∈M

(p

ms

· x

gms

−t

gms

· x

gms

) −W

g

¸

m∈M

x

gms

, (5.1)

where x

gms

is the quantity traded in spot market m by the competitive fringe in

ﬁeld g in scenario s, t

gms

is the price of transportation capacity between g and m

in the secondary market in scenario s. W

g

is the cost function in ﬁeld g. In order

to ﬁnd the demand function for the competitive fringe, the ﬁrst order condition

for optimality is used:

∂Π

gs

∂x

gms

= p

ms

−t

gms

−

∂W

g

¸

m∈M

x

gms

∂x

gms

= 0, g ∈ G, m ∈ M, s ∈ S. (5.2)

In this article, we assume that W

g

is a quadratic function. For ease of presen-

tation, we will in the following assume that the cost function for the competitive

fringe is:

W

g

¸

m∈M

x

gms

=

1

2

c

g

·

¸

m∈M

x

gms

2

(5.3)

where c

g

is the cost parameter for the competitive fringe in ﬁeld g. Nevertheless,

all results are valid for general quadratic cost functions (and most for a general

cost function).

We model this implicitly in the large producers’ problem as an elastic demand

function. The inverse demand function is given as:

t

gms

= p

ms

−c

g

·

¸

m

∈M

x

gm

s

, g ∈ G, m ∈ M, s ∈ S. (5.4)

The volume bought by the competitive fringe, x

gms

is given as the sum of

capacities sold by the ISO, z

gms

, and the large producers, h

lgms

. The h

lgms

variable is positive when the large producers sell capacity, and negative if the

large producers buy capacity. We then have the following relation between x

gms

,

z

gms

and h

lgms

:

156

5.3 Model

x

gms

= z

gms

+

¸

l∈L

h

lgms

, g ∈ G, m ∈ M, s ∈ S, (5.5)

which leads to the following expression for the price in the secondary market:

t

gms

= p

ms

−c

g

·

¸

m

∈M

z

gm

s

+

¸

l∈L

h

lgm

s

, g ∈ G, m ∈ M, s ∈ S. (5.6)

Since we only allow ﬂow in one direction in our network, we need to make sure

that x

gms

cannot be negative.

z

gms

+

¸

l∈L

h

lgms

≥ 0, g ∈ G, m ∈ M, s ∈ S, (5.7)

where h

lgms

is the secondary market trades of producer l of capacity from g to m.

The inclusion of this constraint means that the decision space for each producer

depends on the other participants decisions (the other producers and the ISO).

The large producers

The income for the large producers (L) in the network comes from deliveries in

the long term contracts, sales in the spot markets and sales in the secondary

market for transportation capacity. The cost for the producers come from the

per unit tariﬀ paid in the primary market for transportation capacity (which we

assume is independent of the large producers’ decisions), the cost of production

and from purchasing additional transportation capacity in the secondary market.

The objective function for producer l can be formulated as:

Π

l

= max −

¸

g∈G

¸

m∈M

T

gm

b

lgm

+

¸

s∈S

φ

s

¸

¸

m∈M

(p

ms

q

lms

+P

lm

v

lms

)

¸

+

¸

s∈S

φ

s

¸

g∈G

¸

m∈M

h

lgms

·

p

ms

−c

g

·

¸

m

∈M

z

gm

s

+

¸

l

∈L

h

l

gm

s

¸

¸

−

¸

s∈S

φ

s

¸

g∈G

C

lg

(d

lgs

)

¸

¸

, (5.8)

where b

lgm

is the booking in the primary market, T

gm

is the tariﬀ in the primary

market, φ

s

is the probability of scenario s, p

ms

is price in the spot market, q

lms

157

Chapter 5 Capacity booking in a Transportation Network...

is volume sold in the spot market, P

lm

is the price in the take-or-pay contracts,

v

lms

is the volume in the take-or-pay contracts, h

lgms

is the capacity traded in

the secondary market (positive when the producer sell capacity, negative when

he buys), the price in the secondary market is given by (5.6), C

lg

is the cost

function for the producer and d

lgs

is the production. z

gms

is the capacity sold

by the ISO in the secondary market.

The booking constraint in the primary market is given as:

b

lgm

≤ B

lgm

, g ∈ G, m ∈ M, (5.9)

where B

lgm

is the ﬁxed upper limit on booking for the producer. For the second

stage the following constraints are needed:

d

lgs

=

¸

m∈M

(b

lgm

−h

lgms

) , g ∈ G, s ∈ S, (5.10)

q

lms

+v

lms

=

¸

g∈G

(b

lgm

−h

lgms

) , m ∈ M, s ∈ S, (5.11)

h

lgms

≤ b

lgm

, g ∈ G, m ∈ M, s ∈ S, (5.12)

z

gms

+

¸

l∈L

h

lgms

≥ 0, g ∈ G, m ∈ M, s ∈ S. (5.13)

Constraint (5.10) make sure that the producer has booked enough capacity

for the production in ﬁeld g. Constraint (5.11) make sure that the producer has

booked enough capacity for the total sale in market m. The two constraints also

make sure that the producer utilizes all the booked capacity. Constraint (5.12)

makes sure that the producer cannot sell more capacity than he has booked in

the primary market, and constraint (5.13) ensures that the producers cannot buy

more capacity than the ISO sells.

Independent system operator

We present three diﬀerent objective function alternatives for the ISO:

• maximize ﬂow (MF):

Π

MF

s

= max

¸

m∈M

¸

i∈I(m)

f

ims

, (5.14)

The network operator will always choose z

gms

in order to maximize the ﬂow

under the constraint that all prices (for ﬁeld-market combinations) must be pos-

itive (see Equation (5.26)). With this objective, the system operator will be

158

5.3 Model

indiﬀerent with regards to prices in the market nodes and cost functions in the

ﬁeld nodes.

• maximize value of ﬂow (MVF):

Π

MVF

s

= max

¸

m∈M

¸

i∈I(m)

p

ms

·

f

ims

−

¸

l∈L

v

lms

, (5.15)

The strength of this formulation, MVF, compared with the MF formulation is

that the ISO now routes the gas according to value. The weakness is that he has

no incentive to route according to marginal cost in the ﬁelds.

If we assume that the network operator has full information regarding the

cost functions of the participants, the ISO can take both value of ﬂow and cost

structure in the ﬁelds into account by maximizing social surplus.

• maximize social surplus (MSS):

Π

MSS

s

=max

¸

m∈M

¸

i∈I(m)

p

ms

·

f

ims

−

¸

l∈L

v

lms

+

¸

m∈M

¸

l∈L

P

lm

v

lms

−

1

2

¸

g∈G

MC

g

·

¸

¸

i∈O(g)

f

gi

¸

2

. (5.16)

MC

g

is the slope of the linear aggregated supply function for ﬁeld g:

MC

g

·

¸

i∈O(g)

f

gi

. (5.17)

The supply function is found by assuming that all producers have a cost function

of the form:

W

g

= c

lg

d

2

lg

, (5.18)

and that no production capacities exist. Under these assumptions, the aggregate

supply function is linear. MC

g

is found in the following manner:

MC

g

=

1

¸

l∈

L

g

1

2c

lg

, (5.19)

159

Chapter 5 Capacity booking in a Transportation Network...

where

¯

L

g

is the set of producers L and the competitive fringe in ﬁeld node g. The

aggregated supply function is found by horizontal summation of the individual

supply functions.

Between the production facilities and the market-hubs there is a transporta-

tion network. The gas molecules are transported from nodes with high pressure

to nodes with lower pressure through pipelines. The design parameters of the

pipelines (length, diameter, roughness) as well as external variables (tempera-

ture) decide how much gas is transported for a given pressure diﬀerence. The

relation between pressure in the nodes and ﬂow in the pipelines are determined

based on the Weymouth equation, see for instance Menon (2005). For a discus-

sion of system eﬀects on capacity related to pressure constraints see Midthun

et al. (2006).We have chosen to linearize this expression with the formulation

used in Tomasgard et al. (2007):

f

ij

≤ K

ij

RI

i

RI

2

i

−RO

j

2

r

i

−K

ij

RO

j

RI

2

i

−RO

2

j

r

j

.

(5.20)

About 20 of these constraints that are approximating the Weymouth constraint

are used for each pipeline in order to linearize the ﬂow around pairs of pressure

in, RI

i

, and pressure out, RO

j

. Here f

ij

is the ﬂow from node i to j and r

n

is

the pressure in node n.

In addition, constraints on the pressure level in each node must satisﬁed:

R

n

≤ r

ns

≤ R

n

n ∈ N, s ∈ S, (5.21)

where R

n

is the smallest allowed pressure in node n, and R

n

is the largest

allowed pressure in node n.

In the numerical analysis, we will also look at an alternative formulation with

ﬁxed capacities. In this case the pressure constraints and the Weymouth equation

are replaced with the following formulation:

f

nis

≤ C

ni

, n ∈ N, i ∈ O(n). (5.22)

In the following we will refer to this formulation as Independent Static Flow

(ISF), while the Weymouth formulation is referred to as WF. It is non-trivial to

determine appropriate values for the ISF capacities. See Midthun et al. (2006) for

a discussion. In this paper we solve an optimization model (with WF formulation)

where the objective is to maximize the throughput in the network. The ISF

capacities are then set equal to the resulting ﬂow pattern in this model. The

WF formulation is a relaxation of this ISF formulation, but it also represents the

real system more precisely as it includes the ﬂexibility of moving bottlenecks by

160

5.3 Model

adjusting pressures. The ISF formulation is more restricted but any increase in

its capacities will allow a solution which is infeasible in the WF formulation.

The system operator must make sure that the mass is conserved in the network.

We assume that each ﬁeld is connected to a junction node, and that each market

is connected to a junction node. In addition for ease of notation, we assume that

no junction nodes are connected to each other. The mass balance equations are

given as:

¸

j∈O(g)

f

gjs

=

¸

m∈M

z

gms

+

¸

l∈L

b

lgm

, g ∈ G, s ∈ S, (5.23)

where O(g) is the set of nodes connected to a pipeline leaving from ﬁeld g. In

the junction nodes, the mass balance can be formulated as:

¸

g∈I(j)

f

gjs

=

¸

m∈O(j)

f

jms

, j ∈ J, s ∈ S, (5.24)

where I(j) is the set of nodes connected to a pipeline entering node j. Finally,

a constraint for the mass conservation in the market nodes must be included:

¸

n∈I(m)

f

nms

=

¸

g∈G

z

gms

+

¸

l∈L

b

lgm

, m ∈ M, s ∈ S. (5.25)

The following constraint is included in the model with maximum ﬂow and

maximum value in order to ensure that the price in the secondary market is

positive:

p

m

−c

g

·

¸

m

∈M

z

gm

s

+

¸

l∈L

h

lgm

s

≥ 0, g ∈ G, m ∈ M. (5.26)

Alternatively, we could have introduced a constraint that ensured that the ISO

income was positive in total (or for all ﬁeld-market combinations).

Benchmark

In Chapter 5.5 we benchmark our solutions with an optimization model where an

independent operator schedules production, routing and sale in order to maximize

the social surplus of all the players in the network. The closer the equilibrium in

our game gets to the benchmark solution, the better the strategy is with respect to

maximizing the social surplus. The mathematical formulation of the benchmark

model is given below.

161

Chapter 5 Capacity booking in a Transportation Network...

Π

BM

s

= max

¸

m∈M

¸

l∈

L

(p

ms

q

lms

+P

lms

v

lms

) −

¸

g∈G

¸

l∈

L

1

2

MC

lg

d

2

lgs

, (5.27)

where MC

lg

is the slope of the linear supply function of producer l in ﬁeld g.

In addition, we need constraints (5.20) and (5.21) from the network operator

problem presented in section 5.3. The mass balance is taken care of by:

¸

l∈

L

d

lgs

=

¸

j∈O(g)

f

gjs

, g ∈ G, s ∈ S, (5.28)

¸

g∈G

f

gjs

=

¸

m∈M

f

jms

, j ∈ J, s ∈ S (5.29)

¸

l∈

L

(q

lms

+v

lms

) =

¸

j∈I(m)

f

jms

, m ∈ M, s ∈ S. (5.30)

5.4 Model properties

Our model is a General Nash Equilibrium game where the feasible regions of the

players depend on the other players’ decisions. Let X

l

∈ R

α

be the strategy set

of player l with decision variables x

l

= (x

l1

, . . . , x

lα

) . We have |L| producers

and 1 ISO, constituting the set of players,

¯

L. Deﬁne β = |L| + 1. The set

X =

¸

l∈

¯

L

X

l

is the full Cartesian product of the strategy sets of individual

players and x = (x

T

1

, . . . , x

T

β

)

T

(in the case that no common constraints existed,

it would be the strategy set of the game). Also deﬁne the vector x

−l

of all players’

decisions except player l’s and correspondingly X

−l

=

¸

j∈

¯

L|j=l

X

j

.

We will deﬁne more formally the dependence between the players through

the common constraints and deﬁne the point to set mapping K

l

: X

−l

⇒ X

l

representing player l’s feasible region, given the actions of the other players.

K

l

(x

−l

) ⊆ X

l

, x ∈ X.

Then a generalized Nash equilibrium (GNE) is deﬁned as a point x

∗

∈ X that

simultaneously optimizes all the players individual decision problems so that:

x

∗

l

∈ K

l

(x

∗

−l

), l ∈

¯

L and Π

l

(x

∗

) ≥ Π

l

(x

l

, x

∗

−l

), x

l

∈ K

l

(x

∗

−l

), l ∈

¯

L where

Π : R

αβ

→ R is the objective function of player l. That is, the Generalized Nash

Equilibrium is reached when no player has incentive to change his strategy given

that the other players do not change their strategy.

Pioneering results on the existence of GNE are presented in the papers of De-

breu (1952) (social equilibrium) and Arrow & Debreu (1954) (abstract economy)

that generalized the results of Nash (1950). Rosen (1965) is an early paper con-

cerning not only existence but also investigating uniqueness of solutions for a

162

5.4 Model properties

restricted class of problems. lchiishi (1983) gave more general results concerning

the existence of such GNE.

It is well known that Nash equilibria (with independent player strategy sets)

can be viewed as Variational Inequalities (VI), see Lions & Stampacchia (1967)

for a nice overview. An early reference formulating the generalized Nash equilib-

rium as a Quasi Variational Inequality (QVI) is Bensoussan (1974). See for exam-

ple Ferris & Pang (1997) or Facchinei & Pang (2003) for more on the relationships

between complementarity problems and Variational Inequalities. This means that

in addition to existence and uniqueness proofs following the Arrow/Debreu/Rosen

tradition, also the theory of VI may be used to analyze this, see Harker & Pang

(1990), Harker (1991) and Pang & Fukushima (2005) for good overviews of this

direction of analysis.

Following the lines of the discussion in Harker (1991), we deﬁne F

l

(x

∗

) =

∇

x

l

Π

l

(x

∗

l

, x

∗

−l

) and F(x

∗

) = (F

0

(x

∗

)

T

, . . . , F

|L|

(x

∗

)

T

)

T

Then the GNE may be

expressed as the Quasi Variational Inequality QV I(F, K(x)) :

F(x

∗

)

T

(x −x

∗

) ≥ 0, x ∈ K(x

∗

), (5.31)

where K(x) =

¸

l∈

¯

L

K

l

(x

−l

).

It may here be noted that a standard Nash equilibrium may be expressed as a

VI(F,K):

F(x

∗

)

T

(x −x

∗

) ≥ 0, x ∈ X. (5.32)

In our case, the x vector consist of the following variables: x = (b, h, d, q, f, r, z).

Theorem 5.2 from Chan & Pang (1982) (Theorem 2 in Harker (1991)) give con-

ditions for existence of a solution. We use notation in accordance with what we

deﬁned above:

Theorem 5.4.1. Let F and K be a point-to-point mapping and point-to-set map-

ping respectively from R

αβ

into itself. Suppose that there exists a nonempty com-

pact set X such that

1. K(x) ⊆ X, x ∈ X,

2. F is continuous on X,

3. K is a nonempty, continuous, closed and convex valued mapping on X.

Then there exists at least one solution to the QV I(F, K(x)) in (5.31).

For our problem this is satisﬁed by the deﬁnitions of F and K. F consists

of continuous, linear expressions since our objective functions are quadratic (see

Equations (5.8) and (5.14)-(5.16)). The mapping in our model is deﬁned by

Equations (5.13), (5.23) and (5.25)-(5.26). Since all these equations are linear,

the conditions in Theorem 5.4.1 are satisﬁed. We then know that our Generalized

Nash Game has at least one solution.

163

Chapter 5 Capacity booking in a Transportation Network...

Common constraints

We deﬁne common constraints as constraints where decision variables for more

than one player appear. In our model, all the common constraints are continuous,

linear functions (see Equations (5.33)-(5.36)) and satisfy the necessary constraint

qualiﬁcations (LICQ). We can therefore apply Theorems 4-6 from Harker (1991)

directly. These theorems state that if F is a continuous function in the VI (F, X)

then the VI solutions are the only points in the solution set of the QVI (F, K(x))

at which the optimal dual variables λ

∗

∈ R

pβ

for the common constraints are such

that λ

∗

0

= λ

∗

j

, j ∈ L. The theorems also state that any strictly interior solution

(for the common constraints) of the QVI (F, K(x)) is a solution to the VI (F, X)

as described in (5.32). In general there will be several GNE in the game, but

only the VI solutions will have a common positive value of an additional unit

of a common resource (if the resource is depleted), or a zero value of a common

resource for all players (if not used in full). Further, if F is strictly monotone

there is a unique solution to the VI over X, Facchinei & Pang (2003), Theorem

2.3.3. This means that if an interior x

∗

is known, the only other GNE may be

found at the boundary of the common constraints, and they will not have equal

λ’s for the common constraints.

We have focused on the VI solution in this article. A discussion of the common

constraints and the implication of requiring equal shadow prices are given in the

next sections. In our model we have the following common constraints (dual

variables belonging to each constraint are given to the right):

z

gms

+

¸

l∈L

h

lgms

≥ 0 τ

gms

, (5.33)

¸

j∈O(g)

f

gjs

=

¸

m∈M

z

gms

+

¸

l∈L

b

lgm

, u

gs

, (5.34)

¸

n∈I(m)

f

nms

=

¸

g∈G

z

gms

+

¸

l∈L

b

lgm

, u

ms

, (5.35)

p

m

−c

g

·

¸

m

∈M

z

gm

s

+

¸

l∈L

h

lgm

s

≥ 0, χ

gms

. (5.36)

Constraint (5.33) gives the balance between capacity sold by the system oper-

ator and capacity traded by the large producers. If this constraint is not binding,

the large producers buy less capacity than the ISO sells. If the constraint is

binding, the large producers are buying all capacity sold by the ISO. For the

producers, the shadow price τ

gms

then gives the value of an additional unit of

capacity bought. For the ISO, the shadow price gives the value of selling one

164

5.5 Numerical examples

additional unit of capacity and thus increasing the ﬂow in the network with one

unit. Constraints (5.34) and (5.35) specify that the booked capacity in the net-

work must be equal to the actual ﬂow in the pipelines. Constraint (5.34) gives

the balance for each ﬁeld node, and constraint (5.35) gives the balance for each

market node. For the producers, the shadow price u

gs

gives the value of booking

one additional unit of capacity out of ﬁeld g in the primary market. For the ISO,

the shadow price gives the value of increasing the diﬀerence between the ﬂow

out of ﬁeld g and the capacity sold, z

gms

. Since the ﬂow variable is part of the

objective function for the ISO, the shadow price gives the value for the ISO of

increasing the ﬂow out of the ﬁeld. The same argument is valid for the shadow

price u

ms

. Constraint (5.36) ensures that the price in the secondary market is

positive. The price depends on the volumes sold by the ISO and the large pro-

ducers. For both the producers and the ISO, the shadow price χ

gms

gives the

value of selling one additional unit of transportation capacity.

For the MVF and MSS formulation, we advocate that the VI solution to the

GNE game is the important one. In this case the ISO will have made routing

decisions which make sure that all players’ marginal value of an additional trans-

portation unit is equal. In the system we have described, the tariﬀ is ﬁxed and

may not be changed in order to give speciﬁc incentives to the players. Hence it is

clear that the ISO has a lot of inﬂuence through the routing decisions, and such a

fair routing policy is preferable. For the MF formulation however, the VI solution

depends on the conversion of 1 Sm

3

to NOK, since we relate objective functions

that are not commensurable with respect to the units. Since the marginal values

are given in diﬀerent units, it may not make sense to require equality in the equi-

librium solution. The equilibrium solution will change if we change the currency

(from NOK to for instance Dollars or Euros).

We have not been able to prove that the F function is strictly monotone,

and the equilibriums we present in the numerical examples may therefore not be

unique.

5.5 Numerical examples

We consider the network illustrated in Figure 5.4. There are two large producers,

each present in both g

1

and g

2

. In addition, there is a competitive fringe in g

1

and g

2

.

In the following sections, we use our model to analyze several cases. We start

with a deterministic setting in which we look at the diﬀerent ISO objective func-

tion alternatives and the diﬀerence between the WF formulation and the ISF

formulation. We then introduce stochasticity to our model to see how it inﬂu-

ences the eﬃciency in the network.

Our model is designed for a situation where both a primary market and a

165

Chapter 5 Capacity booking in a Transportation Network...

Figure 5.4: The network used in the numerical examples.

secondary market is used to allocate capacity in the network. The ISO inﬂuences

the eﬃciency of the network through routing decisions and capacity distribution

in the secondary market, while the large producers inﬂuence the eﬃciency by

booking in the primary market and trading in the secondary market. In the

North-Sea today, the booking in the primary market is limited by predeﬁned

booking limits and in case of overbooking a capacity allocation key is used to

distribute the scarce capacity. In our model we resemble this capacity allocation

key by requiring equal marginal value for all players in our common constraints.

Because of this allocation rule, we can use unlimited booking rights in the primary

market. In reality, the total booking rights in the North-Sea is twice the real

capacity.

In each case we solve the stochastic MPC by formulating the equilibrium con-

ditions for the problem. The equilibrium conditions consist of the aggregated

KKT-conditions for all players (see Appendix 5.A). In order to ﬁnd an equi-

librium, we enter the KKT-conditions to the complementarity problem solver

PATH (Dirkse & Ferris 1995). As we discussed in Section 5.4, we focus on the

VI solution to the problem. All prices and costs are given in

1

100

NOK. Since

we have inelastic demand functions in the market nodes, the social surplus will

be identical with the producer surplus in our network (which is an interesting

setting from a Norwegian perspective).

166

5.5 Numerical examples

Node/pipeline R R K

ij

C

ij

g

1

180 170

g

2

185 170

j

1

170 130

m

1

130 115

m

2

130 100

g

1

-j

1

0.5 38.39

g

2

-j

1

0.6 52.71

j

1

-m

1

0.4 46.11

j

1

-m

2

0.35 44.99

Table 5.1: The design parameters for the network

Case 1: The diﬀerent ISO objective function alternatives (WF

formulation)

We start by illustrating the diﬀerence in the objective function alternatives we

have presented for the ISO. The parameters in the cost functions for the large

producers (see Equation (5.18)) are given as c

lg

: c

11

=

5

2

, c

12

= 6, c

21

= 3, c

22

= 5,

and for the competitive fringe in ﬁeld g (see Equation (5.3)): c

1

= 10, c

2

= 12.

The network parameters are given in Table 5.1. The prices in the two markets

are given as: p

m

1

= 130 and p

m

2

= 160. The tariﬀ in the primary market is 10

for each ﬁeld market combination.

When we solve the benchmark model (see Section 5.3), we get a total surplus for

all the players of 7220.43. This corresponds to the maximal achievable surplus in

the network. The results from the optimization with the three diﬀerent objective

functions for the ISO is given in Table 5.2.

As we can see from the results, the model version where the ISO maximizes

social surplus (MSS) gives the highest total surplus in the network. The total

surplus is only 0.8% lower than the benchmark solution. The total social surplus

obtained in the MVF and MF models are, respectively, 9.41% and 1.59% smaller

than the benchmark solution. We also see that the value of ﬂow is largest in

the MVF formulation, while the social surplus has decreased. The reason for the

decrease in social surplus is that the production costs have increased more than

the income from the spot market. The reason is that the VI solution requires

equal marginal values for all players in the common constraints. Since the ISO

only considers the income from the ﬂow in the network (and not the production

costs), the ISO has a large marginal value of ﬂow and therefore forces ineﬃcient

production decisions from the producers.

The equilibrium for the MF model is, as discussed in Section 5.4, diﬃcult to

167

Chapter 5 Capacity booking in a Transportation Network...

Max social Max value Max ﬂow

surplus (MVF) (MF)

(MSS)

Competitive fringe g

1

(NOK) 258.49 194.79 222.22

Competitive fringe g

2

(NOK) 704.17 704.17 704.17

Producer 1 (NOK) 3085.89 2595.33 3129.77

Producer 2 (NOK) 2435.35 2282.92 2410.73

ISO proﬁt (NOK) 678.15 763.51 638.5

Social surplus (NOK) 7162.05 6540.72 7105.39

Flow (Sm

3

) 80.32 88.85 76.35

Value of ﬂow (NOK) 11668.39 12870.88 11207.20

Table 5.2: Results from the diﬀerent ISO objective functions. WF formulation.

interpret since the units are diﬀerent in the objective functions for the ISO and

the large producers. If we change the currency (corresponds to changing the

weighting of the ﬂow for the ISO), the equilibrium also changes. By using a

currency of

1

100

NOK we put the emphasis on the large producers, and since the

social surplus corresponds to producer surplus in our models, we get a solution

close to the benchmark. In Table 5.3 we see the results from changing the currency

from

1

100

NOK to EUR (this is done by changing the weighting of the ﬂow for the

ISO, so the units are comparable with the results in Table 5.2). While the ﬂow

in the MF formulation was the lowest among the three alternatives in Table 5.2,

it has increased to the maximum possible ﬂow in the network in Table 5.3.

In the MSS and the MVF formulation, the change of currency will not aﬀect

the solutions, and in the remaining examples we will therefore focus on the MSS

and the MVF formulations.

Case 2: ISF versus WF formulation

In this example we look at the diﬀerence between using the WF formulation

and the ISF formulation (see Section 5.3 for a discussion of how the ISF capac-

ities are determined). We use the same parameters as in the previous example

(Section 5.5). Every ﬂow pattern obtained with the ISF formulation is feasible

within the WF formulation. In the ISF formulation the capacity in the network

is therefore more restricted than in the WF formulation (the reason for including

the ISF formulation is that it is a common approach for economic analysis in gas

networks).

The results from this optimization is shown in Table 5.4. We see the same

pattern in these results as we saw for the WF formulation: the MSS formulation

168

5.5 Numerical examples

Max ﬂow

(MF)

Competitive fringe g

1

(NOK) 174.18

Competitive fringe g

2

(NOK) 704.17

Producer 1 (NOK) 2328.40

Producer 2 (NOK) 2097.78

ISO proﬁt (NOK) 785.86

Social surplus (NOK) 6090.39

Flow (Sm

3

) 91.09

Value of ﬂow (NOK) 13191.20

Table 5.3: Results from the MF formulation with a larger weight on the ISO

objective function.

gives the highest social surplus in the network. Compared with the WF formula-

tion, the total surplus is reduced with 2.58% for the MSS formulation and 6.85%

for the MVF formulation.

The importance of using the WF formulation depends on the network structure,

the uncertainty in prices and the volume uncertainty in the TOP-contracts. Large

ﬂuctuations (as is common in natural gas prices) give more value to ﬂexibility

and therefore the WF formulation will improve the eﬃciency in the network. The

correlation between prices is also important. High correlation may result in less

diﬀerence between the ISF and the WF formulation (since the ﬂexibility in the

network is less important in this case).

Max social Max value

surplus

Competitive fringe g

1

(NOK) 174.42 174.42

Competitive fringe g

2

(NOK) 707.18 707.18

Producer 1 (NOK) 2599.82 2329.55

Producer 2 (NOK) 2904.06 2098.80

ISO proﬁt (NOK) 591.71 782.96

Social surplus (NOK) 6977.19 6092.91

Flow (Sm

3

) 71.97 91.10

Value of ﬂow (NOK) 10705.93 13192.28

Table 5.4: Results from the diﬀerent ISO objective functions. ISF formulation.

169

Chapter 5 Capacity booking in a Transportation Network...

Node/pipeline R R K

ij

g

1

190 170

g

2

185 170

j

1

170 130

m

1

130 100

m

2

130 90

g

1

-j

1

0.5

g

2

-j

1

0.6

j

1

-m

1

0.4

j

1

-m

2

0.35

Table 5.5: The design parameters for the network

Case 3: The eﬀect of stochasticity

In this example we look at the eﬀect of stochasticity in our model. We use the

network parameters in Table 5.5, and the following cost parameters for the large

producers c

lg

: c

11

= 3, c

12

= 4, c

21

= 4, c

22

=

7

2

, and for the competitive fringe

in ﬁeld g: c

1

= 9, c

2

= 9. The tariﬀs in the primary market are put at 10 for all

ﬁeld-market combinations.

The eﬀects of stochasticity are largest when the price is volatile, and the cor-

relation between the market prices is low, or negative. If volatility is low, or

correlation is very high, the optimal booking in the ﬁrst stage varies less be-

tween the scenarios. When the optimal booking in the ﬁrst stage is similar in all

scenarios, the eﬀect of stochasticity is reduced.

We have chosen to use negative correlation and uniformly distributed prices

between 75 and 225. Table 5.6 shows the results from the optimization. The

benchmark solution in this case is 9008.59. We see that the total expected social

surplus in the network has been reduced with 3.68% and 5.92% for the MSS and

MV formulation, respectively, compared to the benchmark solution. The reason

for these results is the capacity allocation we have chosen (focus on the VI solu-

tion), and the fact that all booked capacity must be used. In a stochastic setting,

the capacity allocation in the primary market is done such that the marginal unit

goes to the player that has the largest expected marginal value. When prices are

very volatile, this means that the large producers in some scenarios have more

capacity than they ideally would have wanted to have.

We have also looked at the wait-and-see solution (Madansky 1960) and ex-

pected result of using the expected value solution (Birge & Loveaux 1997). In

the wait-and-see solution (WSS), the 15 scenarios are solved independently and

we then ﬁnd the expected value over the 15 scenarios. That is, we assume that

170

5.5 Numerical examples

Booking limit = +∞ Wait-and-see solution

Max social Max value Max social Max value

surplus surplus

Competitive fringe g

1

441.91 420.63 453.98 581.19

Competitive fringe g

2

621.50 637.82 726.52 786.71

Producer 1 (NOK) 3180.73 3275.30 3666.62 3462.53

Producer 2 (NOK) 2927.52 2985.92 3303.10 3124.28

ISO proﬁt (NOK) 1505.52 1155.55 758.84 765.28

Social surplus (NOK) 8677.18 8475.30 8909.06 8719.99

Flow (Sm

3

) 94.43 99.84 84.93 96.89

Value of ﬂow (NOK) 14522.37 15040.02 13660.29 14830.25

Table 5.6: Results from the model with stochasticity. Columns 2-3 shows the

result with unlimited booking for each producer, and each ﬁeld-market

combination, and columns 4-5 shows the wait-and-see solution with

unlimited booking.

the large producers somehow get perfect information of the future before they

make their decisions in the ﬁrst stage. The diﬀerence between the WSS solution

and the solution from the stochastic model is the expected value of perfect in-

formation (EVPI). EVPI tells us how much each player would have been willing

to pay for knowing the outcome in the second stage. The results from this test

(columns 4-5 in Table 5.6) shows that the total surplus in the network has in-

creased drastically in the WSS solution. The total expected social surplus is now

only 1.1% lower than the benchmark solution for the MSS formulation, and 3.2%

for the MVF formulation.

In order to ﬁnd the the expected result of using the expected value solution

(EEV), we ﬁrst solve a deterministic problem where the stochastic variables are

represented with their expected values (EVP). We then use the booking decisions

from the EVP in the stochastic problem. The results from the EEV formulation is

shown in Table 5.7. For the MSS formulation, we see that the stochastic solution

is 1.77% higher than the EEV solution. The diﬀerences are small for the MVF

formulation.

The situation without a primary market

We have also tested the model without a primary market (booking limits equal

to zero), and found that the pricing mechanism in the secondary market was

ineﬃcient in this case. Since the price of capacity is based only on one producer’s

marginal cost (the competitive fringe), we found equilibria with a large distance

171

Chapter 5 Capacity booking in a Transportation Network...

Max social Max value

surplus

Competitive fringe g

1

(NOK) 420.73 420.63

Competitive fringe g

2

(NOK) 613.84 696.77

Producer 1 (NOK) 3262.36 3216.85

Producer 2 (NOK) 2994.11 2892.63

ISO proﬁt (NOK) 1235.14 1236.56

Social surplus (NOK) 8526.18 8463.44

Flow (Sm

3

) 95.84 99.84

Value of ﬂow (NOK) 14587.55 15040.02

Table 5.7: Results form the EEV formulation.

to the benchmark solution. For each of the large producers, a decision to increase

production will lead to an increase in production cost in addition to an increase

in price of transportation capacity (when h

lgms

is increased, the price of capac-

ity increase). It may therefore be beneﬁcial for the large producer to decrease

the production even if the marginal production cost is lower than the marginal

revenue.

In order to represent a situation without a primary market, a diﬀerent market

clearing mechanism in the secondary market is needed. As illustrated in the

numerical examples in this section, the market clearing mechanism we have chosen

works well in the presence of a primary market. The design and tests of new

clearing mechanisms is an interesting topic for future research.

5.6 Conclusions

We have presented a stochastic MCP model based on Generalized Nash Equilib-

rium for analyzing a capacity distribution system with two stages: a primary

market where only privileged players can participate and an open secondary

market. This system is based on the existing capacity distribution system in

the North-Sea. We have compared the results from our model with a benchmark

model where a central planner with full information maximizes social surplus in

the network. We have shown that there exists at least one equilibrium solution

(the VI solution) to our models.

We found that the MSS formulation for the ISO lead to a higher total social

surplus in the network than the alternatives. In the deterministic setting we found

a diﬀerence of 0.8% between the benchmark solution and the MSS solution. The

formulation requires that the system operator has full information regarding the

172

5.6 Conclusions

cost structure of the producers in the ﬁelds.

An alternative that we have considered in this paper is to maximize value of

ﬂow to the market nodes. In this case we only need to assume that the ISO

knows the market prices of natural gas. In the deterministic case, the distance to

the benchmark solution was 9.41% for the MVF formulation. The social surplus

for the MVF formulation was 8.6% lower than the social surplus in the MSS

formulation for the deterministic case, and 2.3% lower in the stochastic case. The

results from the WF formulation were highly dependent on the chosen weighting

in the objective functions.

Secondly, we found that stochasticity is important for our results. The book-

ing rights lead to suboptimal solutions in some of the scenarios when prices are

uncertain. The WSS solution indicated a high value of perfect information (so-

cial surplus increased with 2.67% for the MSS formulation). The EEV solution

illustrated that there was a value of solving the stochastic problem (social surplus

increased with 1.77% for the MSS formulation).

Finally we found that modelling the pressure constraints in the network is

important. In this article we have set the ﬁxed capacities such that the total

throughput of the system is maximized. We still found that the ﬂexibility in the

WF formulation was valuable. In our example, we found that the WF formulation

gave an increase of 2.65 % for the MSS formulation.

Given that the value of the ﬂow in the pipelines in the North-Sea in 2006 was

approximately 130 billion NOK, the relatively low percentage diﬀerences we have

shown in this paper still amounts to a substantial amount of money.

Possible future extensions of the model are other market clearing mechanisms

in the secondary market, inclusion of elastic demand functions in the spot markets

for natural gas, the possibility for the large producers to hold back capacity in

the secondary market and strategic behavior in the primary market.

173

Chapter 5 Capacity booking in a Transportation Network...

Appendix

5.A The equilibrium conditions

In this section we give the equilibrium conditions for our model. Shadow prices

for constraints are introduced directly in the Lagrangian function. The matching

of shadow prices with constraints can also be seen from the KKT-conditions. We

distinguish two types of shadow prices: those that are unrestricted in sign (URS)

and those that are restricted in sign. For the shadow prices that are restricted in

sign, we use the following notation for the complementarity condition with the

belonging constraint: G(x) − a ≤ 0 ⊥ ≥ 0. The complementarity condition

states that either G(x) −a or must be equal to zero.

The large producers

The KKT-conditions for producer l is found through the Lagrangian function:

174

5.A The equilibrium conditions

L

l

=−

¸

g∈G

¸

m∈M

T

gm

b

lgm

+γ

lgm

(B

lgm

−b

lgm

)

+

¸

s∈S

φ

s

¸

¸

m∈M

(p

ms

q

lms

+P

lm

v

lms

)

¸

+

¸

s∈S

φ

s

¸

g∈G

¸

m∈M

h

lgms

p

ms

−c

g

¸

m

∈M

z

gm

s

+

¸

l

∈L

h

l

gm

s

¸

¸

−

¸

s∈S

φ

s

¸

g∈G

C

lg

(d

lgs

)

¸

¸

+

¸

s∈S

φ

s

¸

µ

1lgs

¸

m∈M

(b

lgm

−h

lgms

) −d

lgs

¸

+

¸

s∈S

φ

s

µ

2lms

¸

¸

g∈G

(b

lgm

−h

lgms

) −q

lms

−v

lms

¸

¸

¸

+

¸

s∈S

φ

s

[α

lgms

(b

lgm

−h

lgms

)]

+

¸

s∈S

φ

s

¸

τ

gms

z

gms

+

¸

l∈L

h

lgms

¸

+

¸

s∈S

φ

s

u

gs

¸

¸

m∈M

¸

l∈L

b

lgm

+z

gms

−

¸

j∈O

(

g)

f

gjs

¸

¸

¸

+

¸

s∈S

φ

s

u

ms

¸

¸

g∈G

¸

l∈L

b

lgm

+z

gms

−

¸

j∈I(m)

f

jms

¸

¸

¸

+

¸

s∈S

φ

s

¸

χ

gms

p

m

−c

g

¸

m

∈M

z

gm

s

+

¸

l∈L

h

lgm

s

¸

.

Finding the derivative of the Lagrangian function with respect to the decision

variables we get the KKT-conditions for optimality:

175

Chapter 5 Capacity booking in a Transportation Network...

∂L

l

∂b

lgm

= −T

gm

−γ

lgm

+

¸

s∈S

φ

s

(µ

1lgs

+µ

2lms

+α

lgms

+u

gs

+u

ms

) ≤ 0 ⊥ b

lgm

≥ 0, (5.37)

∂L

l

∂γ

lgm

= B

lgm

−b

lgm

≥ 0 ⊥ γ

lgm

≥ 0, (5.38)

∂L

l

∂q

lms

= p

ms

−µ

2lms

≤ 0 ⊥ q

lms

≥ 0, (5.39)

∂L

l

∂d

lgs

= −

∂C

lg

∂d

lgs

−µ

1lgs

≤ 0 ⊥ d

lgs

≥ 0, (5.40)

∂L

l

∂h

lgms

= p

ms

−c

g

¸

m

∈M

z

gm

s

−c

g

¸

l

∈L

¸

m

∈M

h

l

gm

s

−c

g

¸

m

∈M

h

lgm

s

−c

g

¸

m

∈M

χ

gm

s

−µ

1lgs

−µ

2lms

−α

lgms

+τ

gms

= 0, h

lgms

URS, (5.41)

∂L

l

∂µ

1lgs

=

¸

m∈M

(b

lgm

−h

lgms

) −d

lgs

= 0, µ

1lgs

URS, (5.42)

∂L

l

∂µ

2lms

=

¸

g∈G

(b

lgm

−h

lgms

) −q

lms

−v

lms

= 0, µ

2lms

URS, (5.43)

∂L

l

∂α

lgms

= b

lgm

−h

lgms

≥ 0, ⊥ α

lgms

≥ 0, (5.44)

∂L

l

∂τ

gms

= z

gms

+

¸

l∈L

h

lgms

≥ 0, ⊥ τ

gms

≥ 0, (5.45)

∂L

∂u

gs

=

¸

m∈M

¸

l∈L

b

lgm

+z

gms

−

¸

j∈O(g)

f

gjs

= 0, u

gs

URS, (5.46)

∂L

∂u

ms

=

¸

g∈G

¸

l∈L

b

lgm

+z

gms

−

¸

j∈I(m)

f

jms

= 0, u

ms

URS, (5.47)

∂L

∂χ

gms

= p

m

−c

g

¸

m

∈M

z

gm

s

+

¸

l∈L

h

lgm

s

≥ 0, ⊥ χ

gms

≥ 0. (5.48)

The network operator

For the network operator, we present the KKT-conditions for the three diﬀerent

objective function alternatives. First the maximize ﬂow objective.

176

5.A The equilibrium conditions

Maximize ﬂow The Lagrangian function for the system operator can be formu-

lated as

1

:

L

s

=

¸

s∈S

φ

s

¸

m∈M

¸

i∈I(m)

f

ims

+η

nils

K

1

nil

r

ns

−K

2

nil

r

is

−f

nis

¸

¸

+

¸

s∈S

φ

s

u

gs

¸

¸

m∈M

¸

l∈L

b

lgm

+z

gms

−

¸

j∈O

(

g)

f

gjs

¸

¸

¸

+

¸

s∈S

φ

s

¸

χ

gms

p

m

−c

g

¸

m

∈M

z

gm

s

+

¸

l∈L

h

lgm

s

¸

+

¸

s∈S

φ

s

u

js

¸

¸

m∈O(j)

f

jms

−

¸

g∈I(j)

f

gjs

¸

¸

¸

+

¸

s∈S

φ

s

u

ms

¸

¸

g∈G

¸

l∈L

b

lgm

+z

gms

−

¸

j∈I(m)

f

jms

¸

¸

¸

+

¸

s∈S

φ

s

ω

1ns

R

n

−r

ns

+ω

2ns

(r

ns

−R

n

)

+

¸

s∈S

φ

s

¸

τ

gms

z

gms

+

¸

l∈L

h

lgms

¸

.

KKT-conditions The KKT-conditions:

1

We have simpliﬁed the Weymouth equation such that K

1

nil

and K

2

nil

represents the constants

in the expression

177

Chapter 5 Capacity booking in a Transportation Network...

∂L

∂f

gjs

= −η

gjs

−u

gs

−u

js

≤ 0 ⊥ f

gjs

≥ 0, (5.49)

∂L

∂f

jms

= 1 −η

jms

+u

js

−u

ms

≤ 0 ⊥ f

jms

≥ 0, (5.50)

∂L

∂z

gms

= −c

g

¸

m

inM

χ

gm

s

+u

gs

+u

ms

+τ

gms

≤ 0 ⊥ z

gms

≥ 0, (5.51)

∂L

∂r

gs

=

¸

j∈O(g)

¸

l∈L

η

gjls

r

gs

K

1

gjl

−ω

1gs

+ω

2gs

≤ 0 ⊥ r

gs

≥ 0, (5.52)

∂L

∂r

ms

=

¸

j∈I(m)

−

¸

l∈L

η

jmls

r

ms

K

2

jml

−ω

1ms

+ω

2ms

≤ 0 ⊥ r

ms

≥ 0, (5.53)

∂L

∂r

js

=

¸

g∈I(j)

−

¸

l∈L

η

gjls

r

js

K

1

gjl

+

¸

m∈O(j)

¸

l∈L

η

jmls

r

js

K

2

jml

−ω

1js

+ω

2js

≤ 0 ⊥ r

js

≥ 0, (5.54)

∂L

∂u

gs

=

¸

m∈M

¸

l∈L

b

lgm

+z

gms

−

¸

j∈O(g)

f

gjs

= 0, u

gs

URS, (5.55)

∂L

∂u

js

=

¸

m∈O(j)

f

jms

−

¸

g∈I(j)

f

gjs

= 0, u

js

URS, (5.56)

∂L

∂u

ms

=

¸

g∈G

¸

l∈L

b

lgm

+z

gms

−

¸

j∈I(m)

f

jms

= 0, u

ms

URS, (5.57)

∂L

∂χ

gms

= p

m

−c

g

¸

m

∈M

z

gm

s

+

¸

l∈L

h

lgm

s

≥ 0, ⊥ χ

gms

≥ 0 (5.58)

∂L

∂ω

1ns

= R

n

−r

ns

≥ 0 ⊥ ω

1ns

≥ 0, (5.59)

∂L

∂ω

2ns

= r

ns

−R

n

≥ 0 ⊥ ω

2ns

≥ 0, (5.60)

∂L

∂η

nis

= K

ni

r

2

ns

−r

2

is

−f

nis

≥ 0 η

nis

≥ 0, (5.61)

∂L

l

∂τ

gms

= z

gms

+

¸

l∈L

h

lgms

≥ 0, ⊥τ

gms

≥ 0. (5.62)

178

5.A The equilibrium conditions

Maximize value The Lagrangian function for the system operator can be for-

mulated as:

L =

¸

s∈S

φ

s

¸

m∈M

¸

i∈I(m)

p

ms

f

ims

−

¸

l∈L

v

lms

+η

nils

K

1

nil

r

ns

−K

2

nil

r

is

−f

nis

¸

¸

+

¸

s∈S

φ

s

u

gs

¸

¸

m∈M

¸

l∈L

b

lgm

+z

gms

−

¸

j∈O

(

g)

f

gjs

¸

¸

¸

+

¸

s∈S

φ

s

¸

χ

gms

p

m

−c

g

¸

m

∈M

z

gm

s

+

¸

l∈L

h

lgm

s

¸

+

¸

s∈S

φ

s

u

js

¸

¸

m∈O(j)

f

jms

−

¸

g∈I(j)

f

gjs

¸

¸

¸

+

¸

s∈S

φ

s

u

ms

¸

¸

g∈G

¸

l∈L

b

lgm

+z

gms

−

¸

j∈I(m)

f

jms

¸

¸

¸

+

¸

s∈S

φ

s

ω

1ns

R

n

−r

ns

+ω

2ns

(r

ns

−R

n

)

+

¸

s∈S

φ

s

¸

τ

gms

z

gms

+

¸

l∈L

h

lgms

¸

.

KKT-conditions The KKT-conditions:

179

Chapter 5 Capacity booking in a Transportation Network...

∂L

∂f

gjs

= −η

gjs

−u

gs

−u

js

≤ 0 ⊥ f

gjs

≥ 0, (5.63)

∂L

∂f

jms

= p

ms

−η

jms

+u

js

−u

ms

≤ 0 ⊥ f

jms

≥ 0, (5.64)

∂L

∂z

gms

= −c

g

¸

m

∈M

χ

gms

+u

gs

+u

ms

+τ

gms

≤ 0 ⊥ z

gms

≥ 0, (5.65)

∂L

∂r

gs

=

¸

j∈O(g)

¸

l∈L

η

gjls

r

gs

K

1

gjl

−ω

1gs

+ω

2gs

≤ 0 ⊥ r

gs

≥ 0, (5.66)

∂L

∂r

ms

=

¸

j∈I(m)

−

¸

l∈L

η

jmls

r

ms

K

2

jml

−ω

1ms

+ω

2ms

≤ 0 ⊥ r

ms

≥ 0,

(5.67)

∂L

∂r

js

=

¸

g∈I(j)

−

¸

l∈L

η

gjls

r

js

K

1

gjl

+

¸

m∈O(j)

¸

l∈L

η

jmls

r

js

K

2

jml

−ω

1js

+ω

2js

≤ 0, ⊥ r

js

≥ 0, (5.68)

∂L

∂u

gs

=

¸

m∈M

¸

l∈L

b

lgm

+z

gms

−

¸

j∈O(g)

f

gjs

= 0 u

gs

URS, (5.69)

∂L

∂u

js

=

¸

m∈O(j)

f

jms

−

¸

g∈I(j)

f

gjs

= 0, u

js

URS, (5.70)

∂L

∂u

ms

=

¸

g∈G

¸

l∈L

b

lgm

+z

gms

−

¸

j∈I(m)

f

jms

= 0 u

ms

URS, (5.71)

∂L

∂χ

gms

= p

m

−c

g

¸

m

∈M

z

gm

s

+

¸

l∈L

h

lgm

s

≥ 0, ⊥ χ

gms

≥ 0 (5.72)

∂L

∂ω

1ns

= R

n

−r

ns

≥ 0 ⊥ ω

1ns

≥ 0, (5.73)

∂L

∂ω

2ns

= r

ns

−R

n

≥ 0 ⊥ ω

2ns

≥ 0, (5.74)

∂L

∂η

nis

= K

ni

r

2

ns

−r

2

is

−f

nis

≥ 0 η

nis

≥ 0, (5.75)

∂L

l

∂τ

gms

= z

gms

+

¸

l∈L

h

lgms

≥ 0, ⊥τ

gms

≥ 0. (5.76)

180

5.A The equilibrium conditions

Maximize social surplus The Lagrangian function for the system operator can

be formulated as:

L =

¸

s∈S

φ

s

¸

m∈M

¸

i∈I(m)

p

ms

f

ims

−

¸

l∈L

v

lms

+

¸

m∈M

¸

l∈L

P

lm

v

lms

¸

¸

−

¸

s∈S

φ

s

¸

g∈G

1

2

MC

g

¸

¸

j∈O(g)

f

gjs

¸

2

¸

¸

¸

+

¸

s∈S

φ

s

η

nils

K

1

nil

r

ns

−K

2

nil

r

is

−f

nis

+

¸

s∈S

φ

s

u

gs

¸

¸

m∈M

¸

l∈L

b

lgm

+z

gms

−

¸

j∈O

(g)f

gjs

¸

¸

¸

+

¸

s∈S

φ

s

¸

χ

gms

p

m

−c

g

¸

m

∈M

z

gm

s

+

¸

l∈L

h

lgm

s

¸

+

¸

s∈S

φ

s

u

js

¸

¸

m∈O(j)

f

jms

−

¸

g∈I(j)

f

gjs

¸

¸

¸

+

¸

s∈S

φ

s

u

ms

¸

¸

g∈G

¸

l∈L

b

lgm

+z

gms

−

¸

j∈I(m)

f

jms

¸

¸

¸

+

¸

s∈S

φ

s

ω

1ns

R

n

−r

ns

+ω

2ns

(r

ns

−R

n

)

+

¸

s∈S

φ

s

¸

τ

gms

z

gms

+

¸

l∈L

h

lgms

¸

.

181

Chapter 5 Capacity booking in a Transportation Network...

∂L

∂f

gjs

= −MC

g

¸

j

∈O(g)

f

gjs

+η

gjs

−u

gs

−u

js

≤ 0 ⊥ f

gjs

≥ 0, (5.77)

∂L

∂f

jms

= p

ms

−η

jms

+u

js

−u

ms

+τ

gms

≤ 0 ⊥ f

jms

≥ 0, (5.78)

∂L

∂z

gms

= −c

g

¸

m

∈M

χ

gms

+u

gs

+u

ms

≤ 0 ⊥ z

gms

≥ 0, (5.79)

∂L

∂r

gs

=

¸

j∈O(g)

¸

l∈L

η

gjls

r

gs

K

1

gjl

−ω

1gs

+ω

2gs

≤ 0 ⊥ r

gs

≥ 0, (5.80)

∂L

∂r

ms

=

¸

j∈I(m)

−

¸

l∈L

η

jmls

r

ms

K

2

jml

−ω

1ms

+ω

2ms

≤ 0 ⊥ r

ms

≥ 0, (5.81)

∂L

∂r

js

=

¸

g∈I(j)

−

¸

l∈L

η

gjls

r

js

K

1

gjl

+

¸

m∈O(j)

¸

l∈L

η

jmls

r

js

K

2

jml

−ω

1js

+ω

2js

≤ 0, ⊥ r

js

≥ 0, (5.82)

∂L

∂u

gs

=

¸

m∈M

¸

l∈L

b

lgm

+z

gms

−

¸

j∈O(g)

f

gjs

= 0 u

gs

URS, (5.83)

∂L

∂u

js

=

¸

m∈O(j)

f

jms

−

¸

g∈I(j)

f

gjs

= 0, u

js

URS, (5.84)

∂L

∂u

ms

=

¸

g∈G

¸

l∈L

b

lgm

+z

gms

−

¸

j∈I(m)

f

jms

= 0 u

ms

URS, (5.85)

∂L

∂χ

gms

= p

m

−c

g

¸

m

∈M

z

gm

s

+

¸

l∈L

h

lgm

s

≥ 0, ⊥ χ

gms

≥ 0 (5.86)

∂L

∂ω

1ns

= R

n

−r

ns

≥ 0 ⊥ ω

1ns

≥ 0, (5.87)

∂L

∂ω

2ns

= r

ns

−R

n

≥ 0 ⊥ ω

2ns

≥ 0, (5.88)

∂L

∂η

nis

= K

ni

r

2

ns

−r

2

is

−f

nis

≥ 0 η

nis

≥ 0, (5.89)

∂L

l

∂τ

gms

= z

gms

+

¸

l∈L

h

lgms

≥ 0, ⊥τ

gms

≥ 0. (5.90)

182

Bibliography

Arrow, K. J. & Debreu, G. (1954), ‘Existence of an equilibrium for a competitive

economy’, Econometrica 22, 265–291.

Austvik, O. G. (2003), Norwegian Natural Gas. Liberalization of the European

Gas Market, Europa-programmet, Oslo, Norway.

Bensoussan, A. (1974), ‘Points de nash dans le cas de fonctionnelles quadratiques

et jeux diﬀdrentiels lindaires a N personnes’, SIAM Journal on Control and

Optimization 12.

Birge, J. R. & Loveaux, F. V. (1997), Introduction to Stochastic Programming,

Springer.

Chan, D. & Pang, J. (1982), ‘The generalized quasi-variational inequality prob-

lem’, Mathematics of Operations Research 7, 211–222.

Debreu, G. (1952), A social equilibrium existence theorem, in ‘Proceedings of the

National Academy of Sciences’.

Dirkse, S. P. & Ferris, M. C. (1995), ‘The PATH solver: A non-monotone stabi-

lization scheme for mixed complementarity problems’, Optimization Methods

and Software 5, 123–156.

European Union (1998), ‘Directive 98/30/EC of the european parliament and of

the council’.

European Union (2003), ‘Directive 2003/55/EC of the european parliament and

of the council’.

Facchinei, F. & Pang, J. (2003), Finite-Dimensional Variational Inequalities and

Complementarity Problems: Volume I, Springer series of operations research,

New York.

Ferris, M. C. & Pang, J. S. (1997), ‘Engineering and economic applications of

complementarity problems’, SIREV 39, 669–713.

Gabriel, S. A., Zhuang, J. & Kiet, S. (2005), ‘A large-scale complementarity

model of the North American natural gas market’, Energy Economics 27, 639–

665.

183

Bibliography

Gabriel, S., Kiet, S. & Zhuang, J. (2005), ‘A mixed complementarity-based equi-

librium model of natural gas markets’, Operations Research 53(5), 799–818.

Harker, P. T. (1991), ‘Generalized nash games and quasivariational inequalities’,

European Journal of Operational Research 54, 81–94.

Harker, P. T. & Pang, J. S. (1990), ‘Finite-dimensional variational inequality

and nonlinear complementarity problems: a survey of theory, algorithms and

applications’, Mathematical Programming 48, 161–220.

Hobbs, B. F. (2001), ‘Linear complementarity models of nash cournot competition

in bilateral and poolco power markets’, IEEE Transaction on Power Systems

16(2).

Hu, X., Ralph, D., Ralph, E. K., Bardsley, P. & Ferris, M. C. (2004), Elec-

tricity generation with looped transmission networks: Bidding to an iso,

Technical report, CMI EP Working Paper No. 65. Available at SSRN:

http://ssrn.com/abstract=556809.

Jing-Yuan, W. & Smeers, Y. (1999), ‘Spatial oligopolistic electricity models with

cournot generators and regulated transmission prices’, Operations Research

47(1), 102–112.

Kall, P. & Wallace, S. W. (1994), Stochastic Programming, John Wiley & Sons,

Chichester.

lchiishi, T. (1983), Game Theory for Economic Analysis, Academic Press, New

York.

Lemke, C. E. & Howson, J. T. J. (1964), ‘Equilibrium points of bimatrix games’,

Journal of the Society for Industrial and Applied Mathematics 12, 413–423.

Lions, J. & Stampacchia, G. (1967), ‘Variational inequalities’, Communications

on Pure and Applied Mathematics 20, 493–519.

Luo, Z.-Q., Pang, J.-S. & Ralph, D. (1996), Mathematical programs with equilib-

rium constraints, Cambridge University Press.

Madansky, A. (1960), ‘Inequalities for stochastic linear programming problems’,

Management Science 6, 197–204.

Menon, S. E. (2005), Gas pipeline hydraulics, CRC Press, Boca Raton, USA.

Midthun, K. T., Bjørndal, M. & Tomasgard, A. (2006), Modeling optimal eco-

nomic dispatch and ﬂow externalities in natural gas networks. Working paper,

NTNU, Trondheim, Norway. Submitted to international journal.

184

Bibliography

Nash, J. (1950), Equilibrium points in n-person games, in ‘Proceedings of the

national academy of sciences’.

Pang, J. S. & Fukushima, M. (2005), ‘Quasi-variational inequalities, generalized

nash equilibria, and multi-leader-follower games’, Computational management

science 2.

Patriksson, M. & Wynter, L. (1999), ‘Stochastic mathematical programs with

equilibrium constraints’, Operations Research Letters 25, 159–167.

Rosen, B. (1965), ‘Existence and uniqueness of equilibrium points for concave

n-person games’, Econometrica 33, 520–534.

Smeers, Y. (2003a), ‘Market incompleteness in regional electricity transmission.

Part I: The forward market’, Networks and Spatial Economics 3, 151–174.

Smeers, Y. (2003b), ‘Market incompleteness in regional electricity transmission.

Part II: The forward and real time markets’, Networks and Spatial Economics

3, 175–196.

Tomasgard, A., Rømo, F., Fodstad, M. & Midthun, K. (2007), Optimization

models for the natural gas value chain, in G. Hasle, K.-A. Lie & E. Quak,

eds, ‘Geometric Modelling, Numerical Simulation, and Optimization: Applied

Mathematics at SINTEF’, Springer, chapter Optimization Models for the Nat-

ural Gas Value Chain.

Yao, J. (2006), Cournot Equilibrium in Two-Settlement Electricity Markets: For-

mulation and Computation, PhD thesis, University of California, Berkeley,

USA.

Yao, J., Adler, I. & Oren, S. (2006), Modeling and computing two-settlement

oligopolistic equilibrium in a congested electricity network. Working paper,

UCEI.

Yao, J., Oren, S. S. & Adler, I. (2004), Computing cournot equilibria in two

settlement electricity markets with transmission constraints, in ‘Proceedings

of the 37th Hawaii International Conference on System Sciences’.

185

NTNU Norwegian University of Science and Technology Thesis for the degree of philosophiae doctor Faculty of Social Science and Technology Management Department of Industrial Economics and Technology Management ©Kjetil Trovik Midthun ISBN 978-82-471-4512-8 (printed ver.) ISBN 978-82-471-4526-5 (electronic ver.) ISSN 1503-8181 Theses at NTNU, 2007:205 Printed by Tapir Uttrykk

Acknowledgements

This thesis is the result of my work for the degree of Philosophiae Doctor at the Department of industrial economics and technology management, at the Norwegian University of Science and Technology (NTNU). First of all I would like to thank my supervisor Asgeir Tomasgard. His continued support and guidance has been very important for my work. Some of the best and most inspiring moments in the work with this thesis were our discussions. My co-supervisor, Mette Bjørndal, has also been important in the work with this thesis. I really appreciate the time we spent working together. From November 2005 to March 2006 I visited Professor Yves Smeers at the Center for Operations Research and Econometrics (CORE), in Louvain-la-Neuve (Belgium). Working with Professor Yves Smeers was interesting and inspiring. In addition, the stay allowed me to meet many nice people at CORE and introduced me to Belgian culture. I also want to thank good colleagues at the Department, especially Peter Schütz who has faced many of the same problems and challenges as I have since we started on our PhDs back in autumn 2002. I am also grateful to Matthias Nowak for always taking the time to answer my questions. My friends in Trondheim have been very important for me in this period. I want to thank Håvard Berland, Hugo Hammer, Geir Solgaard and Steinar Kragset for many good memories. I am grateful to my family at home for always providing support and encouragement. I want to thank Astrid for putting things in perspective when my frustration with the work was at a maximum. Finally I want to thank Katrine for putting up with me through a lot of hectic periods and for cheering me up when I am down. I look forward to spending more time with you all in the time to come.

Kjetil Trovik Midthun

Trondheim, August 2007

iii

.

. . . . . . . . . . . . . .5 Value Chain Optimization and Portfolio Management 2. . .3 The portfolio optimization model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . . . .6 The Expected Gas Value Function (EGV ) . . . 2. . . . . . . . .3 Existing literature and research contribution .Contents 1 Introduction 1. . . . . . . . . . . . . . . . . . . . . . .2 Modeling natural gas networks . . . . . . . . . . . . . . . . . . 2.4 Optimal dispatch with economic objective functions . . . . . . . . . . . . . 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 Optimal economic dispatch . . . . . . . . . . 4.6 Conclusions . . . . . . 1. . . . . 2 . . . . . . 7 . . . . . . 1 . . 3 Modeling optimal economic dispatch and ﬂow externalities 3. . .two numerical examples 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3. . . . . . . . . . . 2. . . . . . . . . . .1 The natural gas value chain . . . . . . . . . . . .2 Natural gas ﬂows . . . . . . . . . . . . . . . . . 17 . . . . . . . 3. . . . . . 3. . . . . . . . . 2. .2 The Natural Gas Value Chain . . .1 Introduction . . . . . . . . . . . . . . . . . . . . 4 An operational portfolio optimization model for a natural gas producer 4. . . . . . . . . . . . . .1 Introduction . . . . . . . . . . . . . . . . . . . 3. Bibliography . . . . . . . . . . . 19 . . . .3 A Natural Gas Transportation Model . . . . . . 3. . . . . . . . . . . . . . . . . . . . . . . . Bibliography . . . . . . 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .A Notation and Deﬁnitions . . . . . . . . . . . . . . . .4 Test case from the Norwegian continental shelf . . 2. . .two numerical examples . . . . . . . . . . 4. . v . . . . . . . . . . . . .7 Conclusions . . . . . . . . 4. . . . . . . .4 Papers . . . . . . .B Unconstrained equilibrium . . . . . . . . . . . 1. .3 Max ﬂow and network externalities . . . . 29 29 30 34 48 52 61 65 66 69 75 75 77 87 93 97 104 106 107 108 113 113 115 122 128 2 Optimization Models for the Natural Gas Value Chain 2. Bibliography . .4 Management of Natural Gas Storages . . . . . . . . . . . . . . . . . . . . .1 Introduction . . . . .2 Operations research modeling framework . . .A Linearization of the Weymouth equation . . . . . . . . . . . . . . . . . . . . . . . 3. . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .A The equilibrium conditions . . . . .6 Conclusions . . . . . . . . . . . vi . . . . . . . . . . . . 5. . . . . . . . . . . Bibliography . . . . . . . . . . . . . . . . . . . . . . 139 Bibliography . . . . . . . . . . . . . . . .2 Problem description and assumptions . . . 5. . . . . . . . . . . . . . . . . . . . . . . . 140 5 Capacity booking in a Transportation Network with stochastic demand and a secondary market for Transportation Capacity 5. . . . . 5. . . . . . . . . . . . . . 147 147 149 154 162 165 172 174 183 . . 134 4. . . . . . . . . . 5. . . . . . . . . . . . . . . . .5 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 Computational results and discussion . . . . . .Contents 4.1 Introduction . . . . . . . . . . . . . . 5. . . . . . . . . . . . . . . .6 Conclusions . . . . . . . . . . . . . . . . . . . . .4 Model properties . . . . . . . . . . . . . . . 5. . . . .3 Model . .

Chapter 1 Introduction

This thesis presents diﬀerent optimization models for the natural gas value chain. It focuses on the new challenges faced by the participants in the value chain for Norwegian gas after the liberalization of the European gas markets. Most of the models have a producer perspective and are designed to help analyze the value chain for natural gas and give decision support for the gas producers. The modelling framework in this thesis consists of linear programming, mixed integer programming, quadratic programming, stochastic programming and mixed complementarity problems. The thesis consists of this introductionary summary and four papers. Some of the work in this thesis has been sponsored by the VENOGA and the RAMONA project. VENOGA is a project involving Statoil, the Research Council of Norway, NTNU and SINTEF. The goal of the project is to build decision support models and competence for eﬃcient operation and coordination in value chains where Norwegian gas is central. The RAMONA project involves the University in Stavanger, the University in Bergen, NTNU, SINTEF, Statoil, Gassco and CognIT. The intention with the project is to develop methods to optimize regularity and security of supply for the Norwegian gas production- and transportation system. Part 1 of the thesis consists ﬁve sections. In this ﬁrst section a short introduction to the thesis is given, in section 2 the background for the thesis as well as a presentation of the natural gas value chain is presented. Section 3 gives an introduction to the model types I have worked within. In section 4, relevant literature is presented and I indicate where the papers in this thesis extend the existing literature. Lastly, in section 5, a summary of the papers included in this thesis is given. Part 2 consists of the four papers included in this thesis. The ﬁrst paper, ‘Optimization Models for the Natural Gas Value Chain’, gives a thorough introduction to the natural gas value chain, and modeling techniques for the diﬀerent parts of the value chain. The paper is meant as a tutorial on modeling natural gas value chains. In the second paper, ‘Modeling optimal economic dispatch and ﬂow externalities in natural gas networks’, a model for economic analysis in natural gas transportation networks is presented. The model includes a presentation of the

1

Chapter 1 Introduction pressure constraints and system eﬀects in natural gas networks; in addition it uses economic objective functions (such as maximization of social surplus). The third paper is ‘An operational portfolio optimization model for a natural gas producer’. The model presented here has a system perspective, where we assume that all decisions in the value chain is made by a central planner. The value of actively using the pipelines in the network as storage (line-pack) to maximize proﬁts is examined using a stochastic model and real market data from three European market hubs. The last paper, ‘Capacity booking in a Transportation Network with stochastic demand and a secondary market for Transportation Capacity’, provides an equilibrium model of the booking system in the North Sea. The eﬀects of diﬀerent objective functions for the network operator are discussed. In addition, the importance of system eﬀects and stochasticity is analyzed.

**1.1 The natural gas value chain
**

In this section I describe the background for the papers in this thesis and give an introduction to the natural gas value chain. The focus will be on the North Sea, and especially on the Norwegian interests. In 2006, the petroleum industry accounted for approximately 36 % of the total income for the Norwegian economy. Natural gas is increasing in importance in the petroleum industry, and the production of natural gas is expected to reach 42% of total petroleum production in Norway by 2010. In a European context, the Norwegian production of natural gas is signiﬁcant and accounts for approximately 15% of the natural gas consumption. Most of the gas is transported to Germany and France, where Norwegian gas accounts for approximately 30% of the total consumption.

**The liberalization process
**

An important part of the background for this thesis is the liberalization process in the natural gas industry in Europe. The process has been ongoing for more than twenty years, starting in Great Britain. Great Britain is also today the country that is most advanced in the liberalization process, measured by market opening and liquidity in the short-term markets. For a discussion of the liberalization process in Great Britain, see Roeber (1996) and Weir (1999). Also in other landing points for natural gas in Europe, such as Zeebrugge and the TTF, short-term markets have emerged. Financial markets with natural gas as the underlying commodity is developing in these market hubs. This indicates a growing trust in the liquidity of the spot markets. Neumann et al. (2006) gives an example of econometric analysis of the convergence of European spot market

2

1.1 The natural gas value chain prices for natural gas. The paper shows that the Interconnector (a pipeline which connects the market hubs in NBP and Zeebrugge) has led to almost perfect price convergence between the NBP (National Balancing Point in Great-Britain) and Zeebrugge in the time period considered in the paper. The liberalization process has to a large degree been driven by the European Union. This is in contrast to the process in the US and UK which were mainly market driven. The European Union has decided on two gas directives that speciﬁes the development of one internal European market (European Union 1998, 2003, European Commission & Transport 2002). The most important part of the directives is the decision to open transportation infrastructure for third party access. Implications for the value chain in the North Sea The supply side in the North Sea is dominated by large companies. Before the liberalization process, coordination of production and transportation on the Norwegian continental shelf was accomplished through inter company groups like the Gas Supply Committee (Forsyningsutvalget, FU) and the Gas Negotiating Committee (Gassforhandlingsutvalget, GFU). In 2001 the GFU was abandoned and replaced by individual company sales. The large production companies also own the transportation infrastructure. After the liberalization process, the ownership of the infrastructure was given to a newly formed company: Gassled (owned by the original owners of the infrastructure). Gassled has ownership of all infrastructure open for third party access. New infrastructure facilities will be incorporated in Gassled when used by a third party. The routing in the network in the North Sea is the responsibility of Gassco. Gassco is an independent company responsible for ensuring nondiscriminatory access to the infrastructure owned by Gassled. The tariﬀs in the network are regulated by the ministry of petroleum and energy. The intention is to ensure that the proﬁts are generated in the production ﬁelds and not in the transportation network. The access to the network is decided by allocation rules (Ability to Use and Capacity Allocation Key). For more information on the tariﬀ system, see Gassco (2006). For details on the liberalization process in Norway, see Austvik (2003) and Dahl (2001).

**Elements in the natural gas value chain
**

Natural gas is formed naturally from organic material: plant and animal remains. Subjected to high pressure and temperature over millions of years, the organic material changed into coal, oil and natural gas. Natural gas is a mixture of hydrocarbon and non-hydrocarbon gases (such as helium, hydrogen sulﬁde, and

3

There are storage possibilities along the transportation route. In the following I will go shortly through some of the important characteristics of the natural gas value chain.Chapter 1 Introduction nitrogen). In addition. A common method for exploration on the North-Continental shelf is seismology. or directly to the market hubs in Europe. 4 . the transportation network itself can be considered as a storage facility since there are large volumes of gas contained in the pipelines at all times. Exploration Before any gas is produced.1. meaning that it emits lower levels of harmful byproducts when burnt than for instance oil and coal.1: Illustration of the natural gas value chain. The gas is transported from the production ﬁelds to processing plants. A simpliﬁed picture of the oﬀshore natural gas value-chain in the North Sea is shown in Figure 1. Figure 1. Useful sites for information on natural gas are NGSA (2007). Gassco (2006) and EIA (2006). Unlike other fossil energy sources. the gas can be in gaseous phase or in solution with crude oil. The gases are found in porous geological formations (reservoirs) beneath the earth’s surface. the gas must be located and wells must be drilled. In these reservoirs. the natural gas is a relatively clean fuel.

new information is available to the developers. In order to increase the pressure in the reservoir. the gas is transported from the production ﬁelds oﬀshore to processing plants on the Norwegian mainland or to market hubs on the European mainland and in Britain through long. Figure 1. The rich gas is transported to processing plants where the dry gas and wet gas (the associated hydrocarbons) are separated. and the location of these wells is important for further development of the ﬁelds. The natural gas sold to Europe consists mainly of methane (dry gas). The driving forces are pressure from the expanding gas as well as water which causes the gas to ﬂow into the well. compressors are sometimes used. The gas produced at the ﬁelds can however contain other components with market value. such as associated hydrocarbons (for instance ethane. High pressure in the reservoir gives a high production rate. The procedures developed so far can only indicate where the gas deposits are located. In addition to seismology. Gas containing both dry gas and associated hydrocarbons is called rich gas. the Norwegian production has increased drastically the last ten years. The new information will inﬂuence the optimal total allocation of wells on the ﬁeld. while the gravimeters measure the Earth’s gravitational ﬁeld. which in turn is sold in component markets. methods such as magnetometers and gravimeters can also be used. but the interested reader can consult Bullin (1999) for examples. Modeling of processing plants is not within the scope of this thesis. The gas production depends on the pressure in the reservoirs. The models give a hypothetical picture of the subsurface. Exploratory wells must be drilled in order to prove the existence and actual characteristics of the deposits. In addition. The magnetometers measures small diﬀerences in the Earth’s magnetic ﬁeld.1. and thus increase production capacity. Transportation In the North Sea. propane and butane). The wet gas is then heated in order to separate the diﬀerent components. The cost of drilling such exploratory wells is high.2 illustrates the production of gas in Norway and in the world in total. After a well is drilled. As we can see from this ﬁgure. sub sea pipelines operated at high pressure 5 . Production and processing The gas is produced from the reservoirs. mathematical models are used to predict the underground geological structures and conditions.1 The natural gas value chain The methods have been developed from simple two-dimensional seismology in the early 1970s to the latest technology where four dimensions are being recorded (the fourth being shear waves).

and extraction capabilities and cost of operating. the pipeline network can also be used as storage (line-pack). The components give the gas ﬂow diﬀerent properties.2: The development of gas production in Norway and in the world in total. and factors such as caloriﬁc value and corrosion of pipelines depend on the mixture of components in the ﬂow. In the same manner. In addition. LNG-storages and salt caverns. The gas molecules ﬂow in the pipeline from high pressure points to low pressure points. For more information on storages. There are many diﬀerent forms of storages that are used for storing natural gas: abandoned oil. This is due to the fact that there are large volumes of natural gas 6 . At the production ﬁelds. gas can be injected to the storages during days with low price / low demand and extracted from the storage when the price is favorable. Natural gas ﬂow is a multi-commodity ﬂow. The storages are diﬀerent with respect to capacity. see EIA (2002). see OED (2006). levels. the network consists of 7800 km of pipelines. pressure is increased with compressors in order to create a pressure diﬀerence that is suﬃcient for the gas to ﬂow to the landing points. Since the sale of natural gas is limited both by production capacity in the ﬁelds and the transportation network connecting the ﬁelds to the market nodes.and gas ﬁelds.Chapter 1 Introduction Figure 1. With the completion of Langeled (pipeline). where the diﬀerent components have diﬀerent market value. there is a large value in storing gas close to the market nodes in the summer (when demand is low) in order to sell more gas in winter (when demand is high). aquifers. Storage The demand for natural gas shows strong seasonal patterns and large short-term volatility (day-to-day variation in the prices). Both these factors give a large value to optimal storage utilization. injection. For details on the infrastructure and topology in the North Sea.

If the total volume during a year is lower than the agreed upon volume. A yearly volume is decided. All data used in this section is provided by BP (2006). In the TOP-contracts. In this thesis the main focus is on mathematical programming. this has changed. the buyers still pays for the agreed volume. For a discussion of price models for commodities. short-term markets have emerged. these contracts where established for ﬁeld and market combinations.2 Operations research modeling framework This thesis is within the ﬁeld of operations research (OR). but the situation is improving. the price is determined based on a formula which main components are the prices of competing fuels (for instance oil). Originally. the line-pack is increased. Remaining reserves No one knows exactly how much gas is left in the ground for us to use. This volume is needed for gas to ﬂow in the pipelines. 1. Markets Traditionally. If we look at Norway.1. see for instance Schwarz (1997). By injecting more gas into the network than is extracted in a given time period. Operations research can be deﬁned as an interdisciplinary science which uses quantitative methods to support decision making. Today the reserves are suﬃcient for approximately 65 years (given that today’s production level is kept constant).2 Operations research modeling framework contained in the pipelines. and the companies are now free to deliver the contracted volumes from the ﬁeld of their choice. which can be deﬁned as the study of problems where one seeks to 7 . The volume of proved reserves has increased over the years. but we do have some estimates. In the market hubs. This means that it was determined which ﬁeld should deliver in which contract. The markets have so far had low liquidity. For a discussion of the development of the spot market for natural gas in the UK. The increasing liquidity in the markets is indicated by the development of forward and future contracts with natural gas as the underlying commodity. The price in the shortterm markets is volatile. the reserves to production ratio is approximately 30 (see Figure 1. the gas from the North Sea has been sold in long-term take-orpay contracts (TOP). If however more gas is extracted than injected. After the liberalization process in Norway. and then the buyers have ﬂexibility with respect to nomination on shorter time-periods (within certain limits).3). see Roeber (1996). the line-pack is decreased.

Ax ≤ b x≥0 (1. cT .t. maximize or minimize a real function by choosing the values of real or integer variables from an allowed set. In the following I will give a short introduction to the model types that I have worked with in this thesis. while Equation (1.3) The decision variables in the problem is given by x.3) is the non-negativity constraints on 8 .2) (1. For a nice overview of the history of the development of Linear Programming. for instance Hillier & Lieberman (2001). The general LP problem can be formulated in the following way: max cT x s. Equation (1. References to literature where more information can be found is given for each model type. A and b are parameters. Other important contributors are. for instance. provides information and overviews of available resources in the ﬁeld of operations research.1) is the objective function.Chapter 1 Introduction Figure 1. such as the homepage of the Institute for Operations Research and the Management Sciences (INFORMS 2007). The overview is meant for the non-expert reader and it does not give a comprehensive introduction to the model types. George B. Webpages. see Dantzig (1991). Wood & Dantzig 1949). John von Neumann and Leonid Kantorovich. Dantzig is often considered as the founder of LP (Dantzig 1949.1) (1. a linear objective function is optimized (maximized or minimized) over a convex polyhedron speciﬁed by linear and non-negativity constraints.2) gives the constraint on the decision variables x. An introduction to the ﬁeld of operations research can be found in textbooks such as. Linear programming (LP) In linear programming.3: Reserves to production ratio for Norway. Equation (1.

The objective function is given by Equation (1. Mixed integer programming (MIP) When some of the variables in the optimization problem are required to take an integer value. Equations (1. the model type is called Integer Programming (IP).7) deﬁne the feasible set for x and y.2) and (1. An example of a MIP problem is given by: max cT x + dT y s. The simplex method can not be used directly on this problem class. A nice introduction to integer programming is given in Wolsey (1998). and Equations (1. D and b are parameters. (1993) and Hillier & Lieberman (2001).4) (1. dT .5)-(1.9) (1. Quadratic programming (QP) Quadratic programming is a special case of non-linear programming.7) gives.5) is a constraint in the problem. Q.2 Operations research modeling framework x. The x variables could also have been unrestricted in sign. Ax + Dy ≤ b x≥0 y integer (1.t. A and b are parameters. An example of a QP problem is given by: max cT x + xT Qx s.t. Equation (1. Ax ≤ b x≥0 (1. Equations (1.1. while cT . Equation (1.5) (1.10) The decision variables are given by x. where the objective function is quadratic while the constraint set consists of linear equations. IP and MIP problems are by deﬁnition non-convex problems.7) The decision variables are given as x and y.6) and (1.9) 9 . A.4). Since the quadratic programs have much in common with the LP programs.8) (1. cT .6) (1. If all variables in the problem are restricted to be integer. respectively. Examples of textbooks with more information on QP are Bazaraa et al.8) gives the objective function for the problem.3) deﬁne the feasible set for x. where y is required to be integer. while Equations (1. the model type is called Mixed Integer Programming. the non-negativity constraint on x and the integer constraint for y. powerful dual and complementarity slackness properties allow specialized algorithms for many cases.

The ﬁrst papers on SP were Dantzig (1955) and Beale (1955).t. The decisions in the root node are the same for all scenarios. Stochastic programming is a problem class that allows the modeler to take the uncertainty into account when building the model. In the scenario tree. and are called the ﬁrst-stage decisions.14) (1. s ∈ S x≥0 ys ≥ 0. see Kaut & Wallace (2007). for a nice discussion. Figure 1. ﬂexibility is of no value (you do not have to change decisions since you know what will happen). In addition Higle (2005) gives a nice tutorial on the ﬁeld. production nodes and market nodes.15) s. the model will put a value on ﬂexibility in decisions. there is a set of decision variables in each node of the scenario tree. An example of a SP problem is given by: max cT x + s∈S ps dT ys s (1.10) deﬁne the feasible set for x. The home page of the Stochastic Programming Community (COSP 2007) has a lot of resources available within the ﬁeld of SP. Ax = b Ts x + Ws ys = hs . the decisions will depend on the realization of the stochastic parameters. When everything is known for certain. This way. s ∈ S The ﬁrst stage decision variables are given by x. see for example (Birge & Loveaux 1997). producer surplus and consumer surplus) in situations where there are supply and demand curves present in. Stochastic programming (SP) In many situations uncertainty is an important characteristic of the problem we try to model.12) (1.13) (1. Examples are uncertainty in prices and demand for a commodity. In each leaf node in the tree there is a realization of the stochastic parameter (for instance price). a discrete representation of the uncertainty is created. There are many diﬀerent ways of obtaining this representation.11) (1. Examples of textbooks that give a thorough introduction to the ﬁeld of stochastic programming are Kall & Wallace (1994) and Birge & Loveaux (1997). An often used approach to represent the uncertainty in the model is a scenario tree. In the second stage.4 shows an example of a scenario tree. In each scenario s in the set 10 . In addition.Chapter 1 Introduction and (1. respectively. The second paper in this thesis use quadratic programming to represent economic objective functions (social surplus.

1. see Billups & Murty (2000) and Cottle et al. For a nice introduction to complementarity problems. The feasible set for the problem is deﬁned by Equations (1. An example of a linear mixed complementarity problem is given by: a + Au + Cv = 0 b + Du + Bv ≥ 0 v≥0 v (b + Du + Bv) = 0.16) (1. In the last paper in this thesis we present a linear mixed complementarity problem.12)-(1. (1992).15). Equation (1. A. Mixed complementarity problems (MCP) Mixed complementarity problems refer to a wide range of problems where the deﬁning equations consist of both complementarity conditions and equality constraints.17) (1. D and B are parameters.19) The decision variables are given by u and v.18) (1. The objective function.2 Operations research modeling framework Figure 1. gives the proﬁt from the ﬁrst stage decisions and the expected proﬁts for the second stage decisions.19) is the complementarity condition in this problem. The probability for each scenario is given by ps . of scenarios S there is a decision variable ys . C.11. in Equation 1. T (1. The product 11 .4: An example of a scenario tree. b.

(1. This is not surprising given the large risks and costs associated with oﬀshore investments. A new MIP model with a detailed description of reservoir production is also presented in the paper.19) gives a linear complementarity problem (LCP). By aggregating the KKT-conditions of several players (several optimization problems) in a multiplayer game. Generalized Nash Equilibria is found in the situation where the players can inﬂuence the feasible region of each others’ optimization problems. The paper also presents a MIP model which proposes platform capacity. an interesting overview of the history of mathematical programming in the petroleum industry is given. where and when wells should be drilled and production from the wells.Chapter 1 Introduction of variable v and the left hand side in constraint (1. Nygreen et al. There exist a number of deterministic investment models. In van den Heever & Grossmann (2001) a model for design and planning of oﬀshore ﬁeld 12 . 1. and in addition indicate where this thesis extends the existing literature. Arrow & Debreu 1954). (1998) presents a MIP model used by The Norwegian Petroleum Directorate. In a Nash Equilibrium no player has incentive to deviate from his decisions given that the other players do not deviate.17). equilibrium problems can be solved. In Bodington & Baker (1990).17) must be zero. Investment models In this model class. The KKT-conditions of an optimization problem can be formulated as a mixed complementarity problem.18) and (1. (1988) existing models for early evaluations of petroleum ﬁelds are presented. In Haugland et al. Equations (1. Sullivan (1988) presents some applications of Mathematical Programming methods to investment problems in the petroleum industry. The model is a multiperiod model and is used for investment planning for ﬁelds in the North Sea which contain a mixture of oil and gas. The petroleum industry has been a pioneer in the application of operations research. and the literature is therefore extensive. There are a large number of publications within this ﬁeld.3 Existing literature and research contribution In the following I will give a short introduction to the literature on diﬀerent aspects of the natural gas.16) we get a mixed complementarity problem. By including variable u and constraint (1. The MCPs can be used to ﬁnd Nash equilibria (Nash 1950). the goal is to give decision support for strategic decisions such as ﬁeld investments and sequencing of investments. and Generalized Nash Equilibria (Debreu 1952.

(2007) presents an operational model for production and routing planning in the natural gas value chain. The expected net present value is maximized under uncertainty in reserves. which meant an increase in ﬂexibility for the participants in the value chain. Jørnsten (1992) presents an integer model for sequencing oﬀshore oil and gas ﬁelds. The model combines a detailed infrastructure model with a complex contractual model. In Ulstein et al. The net present value of projects is discussed in light of the ﬁscal rules. Because of the special properties of the transportation network a value chain approach to optimizing the system is important. The model is a multiperiod mixed-integer nonlinear programming model and incorporates complex ﬁscal rules. The infrastructure model includes non-linear equations for relating pressure and ﬂow in wells and pipelines. storage and markets. such as tariﬀ. Selot et al. multi-commodity ﬂows and quality restrictions in the markets are considered. The value chain approach has become even more valuable after the liberalization process. The non-linear splitting for chemical processing is linearized with binary variables. (2007) planning of oﬀshore petroleum production is studied on a tactical level. There is no market for natural gas included in the model. tax and royalty calculations. processing. The pressure constraints in the network is however not included in the model. The model has a value chain approach where production plans. In Haugen (1996) a stochastic dynamic programming model is constructed to analyze a supplier’s problem of scheduling ﬁelds and pipelines in order to be able to meet contractual agreements. a stochastic version with uncertainty in future demand for natural gas is presented. In addition to the deterministic model. In the value chain approach. Goel & Grossmann (2004) presents a stochastic MIP model for planning of oﬀshore gas ﬁeld developments. The objective of the model is to maximize the expected net present value of the oil ﬁeld given uncertain future oil prices. processing of natural gas and sales in the markets is considered. multi-commodity ﬂows and contractual agreements in the market nodes (delivery pressure and quality of the gas). the complete network is considered and optimized simultaneously. where the objective is to maximize total economic beneﬁt. The contractual model is based on a set of logical conditions for produc- 13 . In Jonsbraten (1998) a stochastic MIP model for optimal development of an oil ﬁeld is presented. Value chain models The upstream value chain of natural gas consists of several components.1. network routing. In addition. There are also some models which incorporate uncertainty. The uncertainty in the model is in the resources (production proﬁles). transportation. The resulting model is mixed integer programming model.3 Existing literature and research contribution infrastructures is presented. such as production.

and being able to solve the model. we present a tutorial for modeling of the natural gas value chain. such as transient ﬂow and interaction with compressors. while the homepage of the Pipeline Simulation Interest Group (2007) gives a comprehensive overview on modeling. The models are detailed and accurate in their description of the physics of gas transportation. The model presented in Selot et al. Because of the interdependence between ﬂows in pipelines. (1979) a steady-state representation of the gas ﬂow is used in a model for allocation of natural gas. Transportation models The transportation of natural gas is one of the key elements when studying the natural gas industry.Chapter 1 Introduction tion sharing and customer requirements. (2000). The combined model is a mixed integer nonlinear programming model (MINLP). while a too detailed presentation makes the model nonlinear and non-convex. we present an operational stochastic model for portfolio optimization. In the ﬁrst paper in this thesis. it is important to ﬁnd a tradeoﬀ between accurately describing the properties of the transportation network. In Ehrhardt & Steinbach (2005) a model for operational planning in natural 14 . By using real market data from three European hubs. In the ﬁrst paper in this thesis we present a linearization of the Weymouth equation which enables analysis of large networks and stochastic problems. we give an estimate of the value of actively using the line-pack to maximize proﬁts for the value-chain. A discussion of transient ﬂows is given in Kelling et al. This is the ﬁrst publication (to our knowledge) which presents a portfolio optimization model with markets. with cost minimization. simulation and optimization of natural gas ﬂows. and the resulting problem is solved by an extension to the simplex algorithm. production planning. In the following some examples of steady-state models. In addition. In this paper we include the storage in the pipelines in the transportation network (line-pack) in the analysis. Also in O’Neill et al. De Wolf & Smeers (2000) presents a model for optimizing gas ﬂow through a network. the model is non-convex due to the pressure-ﬂow relationship and the modeling of multi-commodity ﬂows. contracts. multi-commodity ﬂow and handling of the pressure constraints in the transportation network. as well as more technical models and simpliﬁed economical models are given. (2007) provides an accurate description of the steady-state ﬂow using a nonconvex MINLP model. The ﬂow in the network is steady-state. A simpliﬁed representation leads to an inaccurate model of the transportation (and may lead to wrong conclusions). Westphalen (2004) gives a nice presentation of stochastic optimization in gas transportation. In the third paper in this thesis. There are a large number of publications with a technical approach to gas transportation. The framework for analysis presented in this model is primarily aimed at a tactical level.

storage reservoir operators. Examples of such models are Cremer & Laﬀont (2002) and Cremer et al. and minimize the costs of the compressors in the network. pipeline operators and consumers. A market leader is deciding on his production level under uncertainty in demand.1. For a nice discussion of externalities in electricity networks. peak gas operators. (2003). the followers then reacts to the production level after the uncertainty is resolved. (1996). we show that it is diﬃcult. see Wu et al. The players in the network include upstream producers and downstream traders. to determine appropriate static capacities in a natural gas network. and oﬀers linearization techniques for the nonlinearities in the model. The paper both gives a survey of some of the existing models. if not impossible. In Boots et al. the system eﬀects of natural gas transportation are neglected. and in addition it uses economic objective functions such as maximization of social surplus. Equilibrium models Equilibrium models are used to study situations where more than one player acts strategically. There also exist some economical models with a simpliﬁed representation of the transportation networks. (2004) the downstream market for natural gas in Europe is studied in a successive oligopoly approach. (2006) presents a model to optimize ﬂow in a network. as well as develops relevant models for the restructured natural gas markets. static capacity limit. The models are formulated as complementarity problems. the capacities in the pipelines is normally represented with a ﬁxed. Martin et al. A nice overview of complementarity problems in natural gas markets is given in Gabriel & Smeers (2005). In the second paper in this thesis. The model gives a detailed representation of the physical properties of natural gas transportation. With the simpliﬁed representation of the transportation network. This is the ﬁrst example of economic analysis in a gas transportation network where the ﬂows are determined based on pressure constraints. The model is used on the European natural gas market. The KKT conditions are used to formulate the 15 . The model includes producers.3 Existing literature and research contribution gas networks is presented. A transient ﬂow model is used to control the network load distribution for the next 24 to 48 hours. (2005). In Wolf & Smeers (1997) a stochastic version of the Stackelberg-Nash-Cournot model is presented. The discussion of system eﬀects is similar to the discussion of externalities in the electricity networks. The paper uses the linearization of the Weymouth equation presented in paper one. We discuss the system eﬀects in natural gas networks and provide a framework for economic analysis in natural gas networks. A mixed nonlinear complementarity problem (NCP) to study natural gas markets is presented in Gabriel et al. Nowak & Westphalen (2003) presents a linear model for transient ﬂow modeling. In these models.

The ﬁrst stage decisions in the MCP model are commitments in long-term contracts. we examine the booking procedure in the North Sea as a stochastic mixed complementarity problem. In the last paper in this thesis.Chapter 1 Introduction NCP model. to our knowledge. while the second-stage decisions are spot market activities. This is the ﬁrst study of a booking system similar to the one implemented in the North Sea and. It is also one of few examples of applications of stochastic mixed complementarity problems. also the ﬁrst study of a booking system in a natural gas transportation network. 16 . Zhuang & Gabriel (2006) presents a stochastic equilibrium model for deregulated natural gas markets. We formulate the problem as a Generalized Nash game and we then use theory from variational inequality to show existence of solution and to solve our problem.

I have had an equal part in modeling.1. Marte Fodstad and my supervisor. Springer Verlag. as well as a description of my contribution on each of them. We also examine the eﬀects of ignoring the system eﬀects when doing economic analysis. The content is based on the experience from work done in SINTEF and NTNU. pressure constraints. Quak (eds. The model includes spot markets. Researcher at SINTEF. and provide a framework for economic analysis in natural gas transportation networks.4 Papers In the following I will give a short presentation of the papers. The modeling and analysis in the paper has been done on SINTEF and NTNU. which allows studies of large scale networks. Hasle. In this paper we combine the two approaches. E. 2007. Minor changes in the references have been made in the version included in this thesis. there exist a number of more technical models without economic analysis.-A. My contribution is in structuring and writing large parts of the paper. and is meant to be a tutorial on the subject. production plans. A linearization of the Weymouth equation. In the existing literature there are a number of economic models that disregard the system eﬀects in natural gas transportation networks.4 Papers 1. compressors and multi-commodity ﬂows. storage utilization. No numerical examples are presented in the paper. In addition. Numerical Simulation and Optimization. Also.): Geometric Modelling. The importance of the value chain perspective as well as the portfolio perspective is discussed in the paper. is presented. the analysis and in writing the paper. Associate Professor Asgeir Tomasgard. 17 . I have done the implementation of the model. Frode Rømo. K. forward markets. Paper 2: Modeling optimal economic dispatch and ﬂow externalities in natural gas networks In this paper we combine the modeling framework developed in paper 1 with economic analysis. long-term contracts. Paper 1: Optimization Models for the Natural Gas Value Chain The paper gives an overview of the modeling of a natural gas value chain. The paper is published in G. Co-authors: Senior Researcher at SINTEF. Lie.

Nowak and my supervisor. In addition. Matthias P. Especially. The model has a system perspective. In papers 1. Associate Professor Asgeir Tomasgard. I have had an equal part in modeling. I have had an equal 18 . Associate Professor Asgeir Tomasgard. but now we study the situation when more than one player is making decisions in the value chain. Co-authors: Post doc. We provide numerical examples based on real market data from three European hubs. Co-authors: co-supervisor. 2 and 3 we have used a system perspective on the value chain. The paper is submitted to an international journal. discuss the importance of modeling the pressure constraints in the network. Paper 3: An operational portfolio optimization model for a natural gas producer In this paper we present an operational optimization model for a natural gas producer. We look at diﬀerent objective functions for the network operator. and extends the model with pipeline storage (line-pack). I have done the implementation of the model. Paper 4: Capacity booking in a Transportation Network with Stochastic Demand and a Secondary Market for Transportation Capacity In this paper we study allocation of transportation capacity in a system that resembles the one implemented in the North Sea. I have done the implementation of the model. We also examine the value of using a stochastic model compared to a deterministic model. The model is formulated as a Generalized Nash game. we evaluate the commercial value of actively using the line-pack in the pipelines to maximize proﬁts for the producers. We use the modeling framework presented in paper 1. and look at the eﬀects stochasticity has on the solutions. To our knowledge. this is the ﬁrst study of an operational stochastic portfolio optimization model for natural gas production and sales. this is the ﬁrst time a booking system in a natural gas transportation network is studied using this approach. In addition. the analysis and in writing the paper. Associate Professor Mette Bjørndal and my supervisor. To our knowledge.Chapter 1 Introduction The paper is submitted to an international journal.

Associate Professor Mette Bjørndal.1. 19 .4 Papers part in modeling. Co-authors: co-supervisor. Professor Yves Smeers and my supervisor. Associate Professor Asgeir Tomasgard. the analysis and in writing the paper.

.

New York. K. M. ‘British petroleum homepage’. ‘Trading in the downstream european gas market: A successive oligopoly approach’. Billups. Oslo.. Econometrica 22.. 173–184. Gasmi. 5–33. (1999). Journal of Computational and Applied Mathematics 124. E. (1990). (1955). Birge. (2000). 303–318.. & Stone. G. Journal of Regulatory Economics 24(1). Texas A&M University. & Laﬀont. Europa-programmet. Interfaces 20. C. & Baker. Series B 17. J. Cremer. BP (2006). A. M. Nonlinear Programming. Theory and Algorithms. S. R. F. The Energy Journal 25. G. Bazaraa. Liberalization of the European Gas Market. COSP (2007). ‘Access to pipelines in competitive gas markets’. K.bp. Inc. R. F. ‘Complementarity problems’.com.Bibliography Arrow. Cottle. H. E. E. Boots. Pang. & Debreu. M. Academic press. 265–291. B. The linear complementarity problem. Boston. M. G. & Hobbs. www. Journal of the Royal Statistical Society. & Loveaux. PhD thesis. T. Bodington.. J. Beale. Home page. J. J. K. Introduction to Stochastic Programming.bp. O. (1997). H. C. ‘A history of mathematical programming in the petroleum industry’. (2004). C. M. (1992). URL: www. & Shetty. F. 117–127. Rijkers. G. F. Norway. V. R. Austvik. L. ‘Stochastic programming community’. (2003). & Murty. (2003). Springer.com Bullin. (1954). John Wiley & Sons. (1993). D. 21 .. Norwegian Natural Gas. ‘On minimizing a convex function subject to linear inequalities’. A. ‘Existence of an equilibrium for a competitive economy’. S. Sherali. Economic Optimization of Natural Gas Processing Plants Including Business Aspects.

(2002). ‘Programming of interdependent activities. G. J. (1991). E.Bibliography Cremer. & Zhuang. mathematical model’.-G. in J. EIA (2002). http://www. Management Science 46(11). Springer-Verlag. D. f. (2000). K.. Ehrhardt. & Steinbach. H. H. ‘Opening up to choice: Launching the single european gas market’. S. Norwegian Natural Gas Transportation Systems. (1949). G. EIA (2006). & Smeers. ‘The gas transmission problem solved by an extension of the simplex algorithm’. J. Econometrica 17. Debreu. Operations in a Liberalized European Gas Market. ‘Directive 2003/55/EC of the european parliament and of the council’. Kiet. Dantzig. 200–211. ‘The basics of underground natural gas storage’. Dahl. Dantzig. B. URL: http://www. Bock.New York. S. ed..eia. 799–818. (1955). Nonlinear optimization in gas networks.doe. Y. K. 197–206. ‘Linear programming under uncertainty’. & Transport (2002). pp.gov/pub/oil_gas/ natural_gas/analysis_publications/storagebasics/storagebasics. De Wolf. (2005). Oﬃce for Oﬃcial Publications of the European Communities. NTNU. Simulation and Optimization of Complex Processes’. A social equilibrium existence theorem. ‘Energy information administration homepage’. ‘Directive 98/30/EC of the european parliament and of the council’. M. in ‘Proceedings of the National Academy of Sciences’. (2001). Management Science 1. B.. G. Amsterdam. Gabriel. European Union (1998).Heidelberg . & Laﬀont. (1952). E.pdf. The Netherlands. (2005). Lenstra. ‘Modeling. 22 .eia. Norway. European Union (2003). PhD thesis. Berlin .gov/ European Commission. ‘A mixed complementarity-based equilibrium model of natural gas markets’. 1454–1465. B. D. ed. European Economic Review 46. in H. II. Linear programming. NorthHolland Publishing Co. Operations Research 53(5). 928–935. ‘Competition in gas markets’. Trondheim. G. Dantzig.doe. J. ‘History of Mathematical Programming: A Collection of Personal Reminiscences’. 139–148.

Gassco (2006). Åsa Hallefjord & Asheim. & Lieberman.. (2004). J. ‘Gassco homepage’. K. W. PSIG. The Journal of the Operational Research Society 49. URL: www. C. E. 2005/37. S. McGraw-Hill Book Co. M.. ‘Models for petroleum ﬁeld exploitation’. H. Kall. & Sekirnjak. ‘Complementarity problems in restructured natural gas markets’. Jørnsten. S. S. 811–818.informs. Nash.no Goel. Y. P. 1409–1429. (2005). A. (1996). 191–201. & Wallace. ‘Institute for operations research and the management sciences’.Bibliography Gabriel. I. Jonsbraten. INFORMS. A practical approach to transient optimization for gas networks. (2005). Introduction to operations research. Singapore.org/. S. Chichester. Home page. 23 . CORE Discussion Paper No. European Journal of Operational Research 37. 563–582. http://www. European Journal of Operational Research 58. ‘Mixed integer models for the stationary case of gas network optimization’. G. Higle. L. M. Hillier. Gray. 88–100. J. O. F. & Wallace.gassco. Martin. John Wiley & Sons. S. (2007). Kaut. ‘A stochastic dynamic programming model for scheduling of oﬀshore petroleum ﬁelds with resource uncertainty’. Haugland. (1988). ‘A stochastic programming approach to planning of oﬀshore gas ﬁeld developments under uncertainty in reserves’. K. ‘Oil ﬁeld optimization under price uncertainty’. T.gassco. J. (2001). Mathematical Programming 105(23). in P. Möller. W. & Grossmann. E. 257–271. & Smeers. Stochastic programming: Optimizatin when uncertainty matters. Technical report. Paciﬁc Journal of Optimization 3. (1994). Haugen. K. D. ed.. Stochastic Programming. (1950).no. (1992). K. European Journal of Operational Research 88. Kelling. ‘TutORials in Operations Research’. ‘Sequencing oﬀshore oil and gas ﬁelds under uncertainty’. (2006). in ‘Proceedings of the national academy of sciences’. 58–72. Reith. W.. & Moritz. V. Computers & Chemical Engineering 28. ‘Evaluation of scenario generation methods for stochastic programming’. (2000). Equilibrium points in n-person games. INFORMS (2007). (1998). www.

O’Neill. C. Res 40. ‘Modelling norwegian petroleum production and transportation’. & von Hirschhausen. Nygreen. B. (1979). M. L. Chem. pp. Mason. I. Christiansen. & Pike. to appear in AIChE Journal. Applied Economics Letters 13. NGSA (2007). Sullivan. ‘Convergence of european spot market prices for natural gas? a real-time analysis of market integration using the kalman ﬁlter’. 1–15. P. 923 – 973. Nygreen. (2003). & Kristiansen. A.Bibliography Neumann.. I. M.. 189–200. N. E. Williard. J. Ind. A. Annals of Operations Research 82. & Westphalen. & Grossmann. ‘A lagrangean decomposition heuristic for the design and planning of oﬀshore hydrocarbon ﬁeld infrastructures with complex economic objectives’. ‘The stochastic behavior of commodity prices: Implications for valuation and hedging’.naturalgas. B. in ‘Proceedings of ECOS’.. www. Ø. K. Selot. Trondheim. ‘The development of a UK natural gas spot market’.naturalgas. Robinson. Pipeline Simulation Interest Group (2007). URL: www. Wilkins. Bjørkvoll. (2001).. European Journal of Operational Research 176. (1988). Haugen.org. & Barton. ‘PSIG homepage’. (1996). M. L. ‘Natural gas supply association homepage’. 2857–2875. The Journal of Finance 52(3). A short term operational planning model for upstream natural gas production systems. ‘A mathematical programming model for allocation of natural gas’. M. ‘The application of mathematical programming methods to oil and gas ﬁeld development planning’. ‘Tactical planning of oﬀshore petroleum production’.. 857– 873. Siliverstovs. T. (1998). & Sagli.. (1997). Eng. J. 727–732. URL: www.. Schwarz. Norway. (2007). van den Heever. A linear model for transient gas ﬂow. P. OED (2006). 1751–1758. B. (2006).. available as SINTEF-report STF 38S03601.. Kuok. A. R. S. 251–267.org Roeber. Ulstein. T. 550– 564.psig. 24 . K. ‘Fakta norsk petroleumsverksemd 2006’. J. M. (2007). B. Energy Journal 17. E.. Operations Research 27(5).org Nowak. Mathematical Programming 42.

Review of industrial organization 15. & Gabriel. I.K. general discussion’. (1998). C. (1999).. & Dantzig. L.Bibliography Weir. K. S. A. (2004). Y. 25 . Wu. Wolsey. & Smeers. Inc. Econometrica 17. (2006). 5–23. Journal of Regulatory Economics 10(1). A complementarity model for solving stochastic natural gas market equilibria. Zhuang. Wolf. G. Integer Programming. Spiller. (1996). P. Anwendungen der Stochastischen Optimierung im Stromhandel und Gastransport. & Oren. accepted. Westphalen. S. B. John Wiley & Sons. (1997). D. M. January 2006. D. Wood. ‘A stochastic version of a stackelberg-nashcournot equilibrium model’. ‘Folk theorems on transmission access: Proofs and counterexamples’. University Duisburg-Essen (Germany). J. M. Energy Economics. F. 193–199.. gas supply industry’. PhD thesis. ‘Regulation and the development of competition in the U. 135–147. Varaiya. ‘Programming of interdependent activities. (1949). Management Science 43(2). P.

.

Numerical Simulation and Optimization. K.Paper I Asgeir Tomasgard. Hasle. 2007 .-A. E. Frode Rømo. Quak (eds.): Geometric Modelling. Marte Fodstad and Kjetil Midthun: Optimization Models for the Natural Gas Value Chain Chapter in G. Lie. Springer Verlag.

.

All together these sections will give a supply chain optimization model with an integrated view of the value chain. This is the type of transportation models suitable for planning problems with time resolution weeks. For the models to make sense it is also critical to include the technological characteristics of natural gas transportation and processing. Our focus is on describing modeling techniques and important technological issues. An appendix describing all notation used in the paper is included at the end. We start by describing the most important components of the natural gas value chain in Section 2.Chapter 2 Optimization Models for the Natural Gas Value Chain 2. In Section 4 we introduce gas storages and in Section 5 we describe a portfolio perspective and start investigating the integrated supply chain view. We study the natural gas value chain seen from the point of view of an upstream company with a portfolio of production ﬁelds. via transportation 29 . In Section 6 we see how the spot-markets can be used to price natural gas storage capacity and indicate how to estimate the terminal value of natural gas still in storages or in reservoirs at the end of the planning horizon using the concept of an Expected Gas Value Function. We structure the paper by gradually introducing the diﬀerent levels of the supply chain. months or years. from production. We therefore give a set of models where the interplay between the technological characteristics of natural gas and the markets are highlighted. the short term markets and transportation capacity booking.1 Introduction The models in this chapter are based on the authors experience from making decision support tools for the Norwegian gas industry. Such a company should plan its operations considering long term contract obligations. rather than a very detailed representation needed for commercial models. Then in Section 3 we focus on how to model natural gas transportation in a steady-state situation. Here we introduce short term markets. In these models the economical content and the understanding of gas markets is essential. In particular we describe how the operations and planning are inﬂuenced by the existence of spot markets and forward markets.

The remaining dry gas is transported to the import terminals in UK or on the European continent. hence the issues of transmission and distribution to end customers are not considered.1: Important objects in the natural gas value chain and processing to contract management and gas sales. Also upstream markets exist. where the gas is sold before it is transported to the import terminals. Before we go in detail on these we give a short summary of the main eﬀects of liberalization and regulation in the European gas market. Rich gas components are extracted and sold in separate markets. transportation. processing. In Figure 2. The ﬁrst action is to transport the natural gas from production ﬁelds to processing plants or transportation hubs where gas from diﬀerent ﬁelds is mixed. 30 . We focus on the value chain of a producing company. 2. In these hubs bilateral contracts and spottrades are settled.2 The Natural Gas Value Chain Here we give a brief description of the diﬀerent elements of the natural gas value chain on the Norwegian continental shelf: production.1 we show the main components of the natural gas export value chain.Chapter 2 Optimization Models for the Natural Gas Value Chain PHYSICAL PROCESS MARKETS Upstream market Production Transportation Contract market Financial market Storage Spot market Processing plant Component market Figure 2. contract management and sales.

and each owner has production rights that are regulated by lifting agreements. The remaining dry gas (methane and some ethane) is transported in pipelines to import terminals in the UK. These nominations are often done weekly. The customer nominates volumes within the take-or-pay agreements. for years. but the input pressure of all pipelines going out of a transportation node must be smaller than the smallest end pressure of pipelines coming into the node. France. to temperature and several other parameters. Some of the components can be extracted from the rich gas in processing plants.2 The Natural Gas Value Chain Production Production of natural gas takes place in production ﬁelds. Hence when natural gas from diﬀerent ﬁelds is blended in the transportation network.2. Here natural gas from diﬀerent ﬁelds have diﬀerent quality. Again there is ﬂexibility on when to take the gas within a year (or other time periods) and typically the daily oﬀtake is within a minimum and maximum level. Transportation and Processing Natural gas is transported in pipelines by using compressors to create a higher pressure in the originating end of a pipeline. so that molecules will ﬂow towards the end. propane and butanes. ethane. propane and several more). The components are liqueﬁed petroleum gases like ethane. Several pipelines may meet in a node in the transportation network. and for intermediate periods in between like weeks and months. Typically a producer’s rights allow him to produce between a minimum level of production and a maximum level of production within a set of time periods of diﬀerent length. Normally such production intervals are deﬁned for single days. Processing facilities separate the rich gas into its various components. This production band may be ﬂexible so that gas can be transferred between periods within predeﬁned limits. An example of an export network for natural gas is the one you ﬁnd at the Norwegian continental shelf which consist of 6600 km of pipelines. Belgium and Germany. unless there is a compressor in the node. In take-or-pay contracts the price is usually indexed to other commodities like oil. Much of the natural gas produced is traditionally committed to take-or pay contracts where the buyer has agreed to take a volume in a given import terminal for a sequence of years. it is critical to either keep track of the energy content of the blend or the total content of each natural gas component. 31 . They may have diﬀerent pressure at the end of the pipeline. with ﬁnal nomination the day before production. and the producer has to deliver. in terms of energy content and its chemical composition (methane. Often these ﬁelds have several owners. which are exported by ship to separate commodity markets.

Salt caverns are underground storage tanks washed out from salt layers. Also they can be used as seasonal storages to smooth out seasonal eﬀects. Further transportation from the import terminals are taken on by the buyer using a transmission network to distribute the gas to the end customers. These characteristics are often speciﬁed by the contracts as terms of delivery.Chapter 2 Optimization Models for the Natural Gas Value Chain The organization of transportation markets varies a lot from region to region. They also have low risk as geological data are known. In some cases a ﬁxed tariﬀ is used for zones or for pipelines. today’s storage capacity is very limited when compared to the total production volumes. Whether storage is used as to avoid bottlenecks in the system in high demand periods or to utilize market possibilities. Storage There exist several types of natural gas storages. For a further discussion on these topics see Dahl et al. They also allow producers to produce in time periods where demand is low and to thereby utilize the available transportation capacity. Storages are important in planning models because they allow us to store natural gas close to the market and thereby use them to exploit spot-market variations. The Gas directive (European Union 1998) enforces undiscriminating third party access to the remaining capacity (see Section 2. One will often ﬁnd that existing transportation rights already accounts for much of the available transportation capacity in the network. They have higher risk as seismic investigation is necessary.2 for a discussion of The Gas directive). Abandoned oil and gas ﬁelds have high capacity and thereby a cost advantage. Import Terminals and Markets The import terminals are landing facilities for natural gas where the export pipelines end. Natural gas is delivered here according to speciﬁcation on minimum and maximum pressure and energy content. Injection rates. in other cases bids are given for capacity and the market settled by some auction mechanism. while the capacity allocation regime and tariﬀ regime is not discussed. capacities. (2003). In all cases the market is cleared and capacity is allocated by given transparent rules. 32 . One way of resolving this is to introduce primary markets for transportation capacity where capacity can be booked. In aquifers water is replaced with gas. They typically have high costs. In this paper we will only focus on the utilization of transportation capacity. withdrawal rates and characteristics depending on ﬁlling rate vary between the types. In a secondary market with shorter time horizons transportation capacity is balanced with transportation needs for the diﬀerent shippers.

division of activities within the ﬁrms in the value chain (physically or by accounting) and the possibility for certain consumers to obtain their gas from the supplier of their choice. Still the idea is old and the former variant was the less standardized bilateral long term swing agreements between diﬀerent producers. Liberalization and Regulation The European natural gas industry has developed rapidly over the past thirty years. The European Commission has worked toward strengthening the competition within the complete gas. Recently these terminals have also been the location of the growing spot markets for natural gas and for ﬁnancial derivatives on the spot market. The buyer of gas upstream also has the responsibility to transport the gas to downstream markets. Upstream markets are not as well developed as the other markets. In the directive a stepwise liberalization of the European gas market is described. Upstream markets are used to perform trades of natural gas before transportation takes place. TTF in the Netherlands and Zeebrugge in Belgium. Now the sales are company based and rarely linked to a speciﬁc production ﬁeld.2. The leading European hubs in terms of liquidity are the National Balancing Point in Great Britain. The directive was followed by a second gas directive in 2003 (European Union 2003) which moved another step towards liberalization. Another implication of the gas directive and of EU competition laws was the 2002 closing down of the Gas Negotiation Committee (GFU). A diﬀerent type of markets is also emerging upstream in the pipeline network. the forum for coordinated gas sales from the Norwegian continental shelf. there are already clear signs indicating 33 . An expected result from these changes is that short-term markets will evolve for natural gas. The GFU formerly coordinated the supply of Norwegian natural gas producers Statoil and Hydro. A breakthrough in this process came on the 22nd of June 1998 when the gas directive was passed in the European Commission (European Commission & Transport 2002).and energy value chain. The main idea here is to have standardized trading mechanisms for natural gas at some important locations in the network to be able to provide an additional ﬂexibility for the producers. allowing ﬁelds with little ﬂexibility an option to draw gas from ﬁelds with more ﬂexibility in volumes. Though liquidity is still low. The key components of the gas directive are third party access to all transportation installations.2 The Natural Gas Value Chain Originally these terminals were the points of deliverance for the many take-orpay contracts. They include additional ﬂexibility for producers in terms of being able to stay within the limits of their own lifting agreements. They are useful because there is a need for having standardized mechanisms to exchange gas between producers. transportation capacities and contract commitments in case of unexpected events or in case overselling or underselling of natural gas has occurred.

Some attempts on optimizing the transient behavior of a system of natural gas pipelines are Westphalen (2004) and Nowak & Westphalen (2003). processing plants and markets. to operate the system according to a predetermined load and supply. In practice the producers take the price risk. 2. The control regime often focuses on single processes or single components in the network. Gassco AS should operate independently and impartially and oﬀer equal services to all shippers. spot markets and ﬁnancial markets. while the buyers take the volume risks by going into long term agreements. Within this paradigm one normally uses active control. Here time is discretized. If the resolution of time periods is minutes or hours there is a need to model the transient behavior of natural gas. The models presented in this paper are simpliﬁed variants of models developed in co-operation with Gassco and Statoil to deal with the changes mentioned above. but only systems of limited size and complexity can be handled. Abolishment of the GFU-system and the introduction of a system where the individual companies are responsible for disposal of their own gas reserves called for a new access and tariﬀ regime in the transportation network.3 A Natural Gas Transportation Model When modeling natural gas pipeline ﬂow it is important to have a conscious view on how time and the dynamics of gas ﬂow should be handled. as prices are ﬁxed towards various indexes. The modeling of natural gas ﬂow in continuous time has clear links to the process control paradigm (Hofsten 2000). The new markets will include short-term bilateral contracts. The natural choice is to look at mixed integer programming models from the modeling paradigm of mathematical programming. As a State owned company. The prior market structure is dominated by long-term agreements and thus minimizes the uncertainty for the participants. Systems operated by Gassco are the rich and dry gas systems previously operated by Statoil. For our purpose we need to model a system of pipelines with a set of production ﬁelds. To be able to handle the complexity needed for our models. The introduction of short-term markets will most likely also lead to higher volatility and thus higher uncertainty.Chapter 2 Optimization Models for the Natural Gas Value Chain that short-term contracts and spot-trades will play an important role in the future. Gassco is assigned all the operator’s responsibilities warranted in the Norwegian Petroleum Law and related Regulations. we leave the concept of modeling the transient behavior of natural gas and approximate the time dimension by discrete time periods of 34 . Norsk Hydro and TotalFinaElf. ﬁnding a sequence of control actions which leads the system to a target state. The ﬁrst step in the Norwegian transportation system was taken with the creation of Gassco AS in May 2001 under the provisions of a Norwegian White Paper.

Several examples on linearization of natural gas ﬂow exist in the literature. it can be applied at more tactical/operational level by a commercial player in capacity planning and capacity booking. Also it is used to reroute natural gas when unexpected incidents lead to reduced capacity (in production units or pipeline).3 A Natural Gas Transportation Model such length that steady-state descriptions of the ﬂow will be adequate. GassOpt is mainly used by the operator of the gas transportation system in the Norwegian sector of the North Sea. In the model presented in this section. and maybe days. Tomasgard & Nowak (2004). GassOpt itself focuses on analyzes of transportation possibilities. In this section we present a static model of one period. For a recent PhD thesis on linearization of natural gas ﬂow see Van der Hoeven (2004). Issues like pressure. Demand for natural gas in the import terminal is assumed to be aggregated over the contracts in the terminals and planned production volumes given as constants to represent the license holders’ production plans. More detailed models very similar to the one we present here are today in use by Statoil and Gassco in the software package GassOpt developed by SINTEF. This veriﬁcation is also of strategic importance for the independent producers and customers in Germany. It can be used for optimal routing decisions from a ﬂow maximization perspective. In Section 4 we extend the model with several time periods and storage capabil- 35 . In this paper we describe the essential constraints needed to model the technological characteristics of natural gas ﬂow in a steady-state setting. and contribute to infrastructure investment decisions. Belgium and France. We base our presentation on work done on linearization from Rømo. The pipeline system is a natural monopoly. we will focus on the transportation alone with the main purpose to meet demand for transportation generated by planned production proﬁles for the diﬀerent ﬁelds. The mathematical optimization models used to describe natural gas ﬂow in the case of steady-state models are typical non-linear and non-convex. and GassOpt is one of the tools used to ensure maximum utilization of the infrastructure. gas quality and gas components are dealt with from a pipeline transportation perspective. rather than minutes and hours. we can assume that the system is in a steady-state in each time period. We present here a linearized model based on mixed integer programming to optimize routing of natural gas in pipeline networks. An approach using a non-linear formulation of the mathematical models is illustrated in De Wolf & Smeers (2000).2. This typically represents the situation facing the neutral operator. When the time resolution of the model are months. The security of supply will inﬂuence the price possible to achieve for long term contracts. So the main task of this model is to operate the transportation network to make sure demand is met by the planned production. and is controlled by the authorities. Thirdly. weeks. They are obliged to verify the delivery capabilities and robustness of the pipeline system transporting natural gas to European markets.

The modeling tool is hierarchical and applies to general network-conﬁgurations.2 indicates the network complexity for the North Sea network. In Section 5 we include contracts.Chapter 2 Optimization Models for the Natural Gas Value Chain Figure 2. 36 . The GassOpt Modeling Interface In GassOpt. Figure 2. When modeling the North Sea system we need approximately 75 nodes and 100 arcs to represent the network.2: Network presentation in GassOpt ities. markets and a portfolio perspective on managing the natural gas supply chain with stochastic prices and demand. The squared nodes contain subsystems with further nodes and pipelines. the underlying physical network is represented in a graphical modeling environment with nodes and arcs.

A typical optimization case describes a speciﬁed state of the network. The model describes a steady-state situation where the network is in equilibrium in terms of pressures and natural gas mix. compressor nodes. pressure agreements and the energy use. processing nodes for extracting components of the natural gas. there will be several possible strategies to deliver the maximum amount of gas to the customers. In a normal situation. 3. for example hours or minutes.3 A Natural Gas Transportation Model The GassOpt Mathematical Model This network model includes ﬂow balances. for example if a ﬁeld or pipeline is down. where this is physically possible. Deliver natural gas which meets quality requirements in terms of energy content. including expected production and demand (characterized by volume and quality). Penalties are introduced in the objective function to inﬂuence the impact of the following goals: 1. the model will prioritize to deliver the nominated volumes in the import terminals. but is rather used to choose between alternative solutions with about the same ﬂow. see for instance Rardin (1998). Deliver within the pressure requirements in the contracts. 4. These penalty costs can of course theoretically interfere with and prevent us to achieve the main goal. to deliver in accordance with the demand of the customers. This objective can be achieved in several ways. this steady-state description will not be good enough because of the need to describe the transient behavior of natural gas ﬂow. The model should make sure the nominated volumes are delivered to the import terminals within a time period. In a fault situation. Minimize the use of energy needed in order to deliver the natural gas to the customers by minimizing the pressure variables. To make the model generate and report these realistic ﬂows.2. It is typically the kind of model used to model situations where ﬂows are aggregated over a given time period. For more information about multi-criteria decision making. shutdown situations and turn-up capacity (additional available but unplanned production capacity) from production ﬁelds. show that this ‘multicriteria’ aspect does not sacriﬁce much of the maximal ﬂow potential. The tests we have performed on the full North Sea network. 2. Maintain planned production from the producers. quality requirements. we have introduced penalty costs in the objective function on deviation from planned production. blending of diﬀerent gas qualities from diﬀerent ﬁelds. When the time period gets short enough. node pressures and the nonlinear nature of pressure drop in pipelines. 37 . The objective for the optimization model is to ensure optimal routing and mixing of natural gas.

Z The set of split percentages used to discretize possible split fractions in split-nodes of the network. pin is the pressure into the pipeline going from node i to j. K(b) The set of contracts in node b ∈ B. T The set of time periods included in the model. O(n) The set of nodes with pipelines going out of node n. variable fim is the ﬂow of gas from node i into market node m. Sets Below the sets used in the mathematical description of the model is presented. which is available in some ﬁelds with ﬂexible production characteristics. κ is the penalty cost for pressure level. and δm is the positive deviation from the upper quality level limit in market node. in order to meet demand in accordance with contractual obligations (volume. Furthermore. S The set of nodes with storage facilities. The value of the ﬂow to the customer nodes is given by the constant ω. C The set of components deﬁning the chemical content of the natural gas. In the last case we say that we use turn-up capacity. δm is the negative deviation from u+ the lower quality level limit.Chapter 2 Optimization Models for the Natural Gas Value Chain Seen from an operator’s point of view the model tries to meet the customer’s requirements for a given state of the network: either by optimal routing of gas or by turning up production in ﬁelds with ﬂexibility on the production side. ∆+ and ∆− represents underproduction and the use of turn-up in g g l− relation to the planned production in ﬁeld g. quality and pressure). M Nodes with buyers of natural gas: typically import terminals. + m ij and − is the positive and negative deviation from the contracted pressure level m respectively. is the penalty cost for deviation from contracted pressure level. L The set of breakpoints used to linearize the Weymouth equation. N B Objective Function Our goal is to route the gas ﬂow through the network. Y The number of discretized storage and injection rate levels used to linearize storage characteristics. I(n) The set of nodes with pipelines going into node n. χ is the penalty 38 . R The set of nodes with processing capabilities. The set of nodes where gas ﬂows are splitted into two or more pipelines. The set of all nodes in the network. In the formulation given below.

without interfering with the correct volume. quality and pressure to the customer terminals.3 A Natural Gas Transportation Model cost for deviation from contracted quality to customers and ι for use of turn-up. Here fgj is the ﬂow from production ﬁeld g to node j: fgj ≤ Gg . g ∈ G.2) Demand This constraint says that the total ﬂow into a node with customers for natural gas must not exceed the demand of that node: fjm ≤ Dm . j∈O(g) (2. (2. The penalty on using turn-up capacity will make sure that planned production in the ﬁelds is prioritized ﬁrst. and will then rather be put into a hard constraint instead of being penalized in the objective function.2. max Z = i∈I(m) m∈M ωm fim − i∈N j∈N κpin − ij m∈M l− δm u+ + δm + m + − m − g∈G ι ∆+ g + ∆− g − m∈M χ (2. For most practical cases the contracted pressure level is not a soft constraint. What ﬂows into node j must also ﬂow out of node j: fij = fjn .4) i∈I(j) n∈O(j) Pressure constraints for pipelines Oﬀshore transportation networks often consist of very long pipelines without compression.1) Energy consumption for transporting the natural gas is minimized through making the penalty cost (κ) insigniﬁcant in size as compared to the value of the natural gas transported.3) Mass balance for node j The following constraint ensures the mass balance in the transportation network. m ∈ M. as long as it does not inﬂuence the throughput of the pipeline system. Constraints Production capacity The following constraint says that the total ﬂow out of a production node g cannot exceed the planned production of the ﬁeld in that node. where it is crucial to describe the pressure drops in the pipeline system. We use the Weymouth equation to 39 . This contributes to reduce the necessary build up of pressure to a minimum. j ∈ N . j∈I(m) (2.

The ﬁgure shows that the function in the interesting area (positive pressure levels) is one fourth of a cone. (P Il . where l = 1. For any given pipeline ﬂow. pout ) ≤Wij (P I.g.. .6) We introduce a set of points to linearize this expression. pout ) = Kij ij ij pin − pout . ∂pout ij ij (2. P O) + ij ij + ∂Wij in (p − P I) ∂pin ij ij j ∈ N .7) will be tangent to the cone at the line where the ratio between pressure in and out of the pipeline is equal to the ratio between P Il and P Ol . .7) j ∈ N . Then we replace for each pipeline the nonlinear function (2. Campbell (1992). . This constant depends among others on the pipelines length and its diameter and is used to relate the correct theoretical ﬂow to the characteristics of the speciﬁc pipeline.3 illustrates the Weymouth equation.5) with L linear constraints of the type: W fij ≤Kij P Il pin 2 ij P Il2 − P Ol P Ol W − Kij pout . The cone starts in origo. Together the planes give an outer approximation of the cone. i ∈ I(j).5) W Here Kij is the Weymouth constant for the pipeline going from i to j. L. Through Taylor series expansion it is possible to linearize Equation (2. ∂Wij out (p − P O). (2. P O) representing ﬁxed pressure into the pipeline and ﬁxed pressure out of the pipeline respectively: Wij (pin . pout ) is the ﬂow through a pipeline going from node i to node ij ij j as a consequence of the pressure diﬀerence between pin and pout : ij ij W Wij (pin . The planes described in (2. namely the one that approximates the ﬂow best. P Ol ). . l = 1. . 2 ij P Il2 − P Ol (2. ij ij 2 2 j ∈ N . In the Weymouth equation Wij (pin . . This approximation will consist of triangular shapes deﬁned by these planes. i ∈ I(j). Pipelines without pressure drop For physical pipelines between nodes where the distances are very limited it is not necessary to model pressure drops by the 40 .5) around a point (P I. . L. The Weymouth equation is described in e. i ∈ I(j). and is limited by the inlet pressure axis. . and the 45◦ line between the inlet pressure and outlet pressure axes.Chapter 2 Optimization Models for the Natural Gas Value Chain describe the ﬂow in a pipeline as a function of input and output pressure. only one of these L constraints will be binding. Figure 2.

41 .3 A Natural Gas Transportation Model 200 150 Flow 100 50 0 150 160 140 100 100 50 Pressure out 80 Pressure in 120 200 180 Figure 2.3: A three-dimensional illustration of how the Weymouth relates pressure at the inlet and outlet points to the capacity in the pipeline.2.

The end pressure of this arc is neglected. i ∈ I(j). This means that even if there is a pressure diﬀerence in a pipeline the ﬂow can be zero. j ∈ N . Then the following constraints make sure that the input pressure of a pipeline leaving n is less than the output pressure of a pipeline ending in n as long as both pipelines have a ﬂow larger than 0. j ∈ O(i).Chapter 2 Optimization Models for the Natural Gas Value Chain Weymouth equation. M is a number which is large enough to not restrict the pressures when the ﬂows are 0. In general for a node n the input pressure of all pipelines going out of n must be lower than the lowest pressure out of any pipeline going into node n. ij ij (2.15) make sure that there only ﬂows gas in one direction in the pipeline. Equations (2.14) (2. j ∈ I(n). In this case there is no pressure drop. (2. j ∈ O(n) n ∈ N . in jn n ∈ N .10) Another important issue is the relationship between pressure in ingoing pipelines and the outgoing. (2. i ∈ I(j).12) (2. i ∈ N .14) and (2. and still keep the desired pressure. Because of this property it is not necessary to explicitly model the possibility of shutting down a pipeline. In this case a simple maxﬂow restriction is: fij ≤ Fij . j ∈ I(n).15) ρjn = 1 − ρnj . and that is the case where a pipeline into the node has 0 ﬂow.13) i if ﬂow from node n to node j 0 otherwise. it is sometimes preferable to model the pressure out of all the pipelines going into the same node homogenously: pout = pout . n ∈ I(j). ρnj = n ∈ N .8) Relationship between pressures into a node and out of a node To achieve a relevant ﬂow pattern. i ∈ I(n). There is one exception. nj in fnj ≤ M ρnj . the ρij variable deﬁned in the previous paragraph is used to determine the direction of ﬂow. Modeling bidirectional pipelines For pipelines designed to handle ﬂows in both directions. In the Equation (2. The Weymouth equation used gives an upper bound on the ﬂow in a pipeline.11) the variable ρij is 0 for pipelines without ﬂow and 1 for the others.9) (2. j ∈ N . If omitting the constraints presented above one has to be aware of this when interpreting the results from the model. The model can simply put the ﬂow to zero. i ∈ I(n). so: pout = pin . pin − pout + M (ρnj + ρin − 1) ≤ M.4. j ∈ O(n) (2. see Figure 2.11) (2. where Fij is the capacity. fij ≤ M ρij . 42 .

If there is no compressor. have pipelines going to node n. and η is the compressor eﬃciency. i ∈ O(n). The index on i and j goes from 1 to N . (Campbell 1992).2. The upper nodes. jn ni n ∈ N . n∈N (2. this Γ is a function of the ﬂow fn = j∈I(n) fjn into the node: W max η(Ka − 1) +1 100Ka fn Ka Ka −1 Γn (fn ) = . have pipelines coming from node n. If there is a compressor. (2. The compressor characteristics includes a compressor factor Γ used to limit how much the gas can be compressed in a node. i.17) 43 .16) In this expression. this factor is 1. where N is the total amount of nodes.3 A Natural Gas Transportation Model Node i1 Node iN Node n Node j1 Node jN Figure 2. the parameter Ka is the adiabatic constant for a certain gas type. j. Here we simplify this by using a constant compression factor independent of the ﬂow. or we force a pressure drop in the node.4: Example of a split node with the possibility to shut down operation of one of the pipelines. Then the pressure out of the compressor node n is limited by the compressor factor times the pressure into the node n: Γn pout ≥ pin . Nodes with compression or pressure drop In some cases we allow the pressure to increase in a node by using a compressor. j ∈ I(n). We here present a simpliﬁed formulation for modeling compression nodes where pressure can be build up or forced down. W max is the power output capacity of the compressor. The lower nodes.

20) and (2. i ∈ O(n). (2. i ∈ I(m).21) respectively). j ∈ I(n).Chapter 2 Optimization Models for the Natural Gas Value Chain Pressure drop is modeled in the same way. or the content of CO2 . (2. Examples are minimum and maximum pressure out of pipeline (represented by (2. m ∈ M. m m These deviations are penalized in the objective at a level reﬂecting how hard the pressure constraint is in practice. but in practice the plants using the natural gas are technically calibrated for a certain GCV-range. If we focus on GCV. These properties are both technically and economically important for the customer. Gas quality and energy content In this model. These constraints are called technical pressure constraints. High GCV is in itself tractable as the energy content is higher. j ∈ N . Most import terminals have a limited range around a target pressure Pm which they accept for incoming gas: pout + im − m − + m = Pm . Contracted pressure It may be necessary to model the contracted pressure in nodes with customers. the customer accept a maximum content in terms of [mol %]. Here we only give the formulation for GCV: Qmin ≤ qim ≤ Qmax . i ∈ I(j). (2.20) (2. ij j ∈ N . When dealing with CO2 . This is typically due to environmental taxes or to requirements related to avoiding corrosion in pipelines.21) pin ij ≤ max Pij . m m (2. the customer accepts deliveries between a minimum and maximum combustion value. i ∈ I(j).19) Here − and + are negative and positive deviations from the target pressure. min pout ≥ Pij . where 0 < Θn ≤ 1 and 1 ≤ Γn . It is also possible to specify restrictions for each pipeline for example for the pressure into and out of a given pipeline. jn ni n ∈ N . m ∈ M. but with a reduction factor Θn instead of a compressor factor: Θn pout ≥ pin . The quality is then measured in [M J/Sm3 ]. gas quality can be speciﬁed in two diﬀerent ways. The formulation is only meaningfull if at most one of the factors is diﬀerent from 1 in a node. Pressure restrictions often apply to nodes with compression or nodes where processing of the gas is being performed.22) where qim is gas quality (GCV ) in a pipeline going from node i to market node m. In practice we need more ﬂexibility in the model by allowing reduced 44 .18) Here Θn and Γn are constants. focusing on combustion value (GCV) of the natural gas. i ∈ I(m).

and that it is decided by the convex combination of inﬂow qualities to the node: qij = or: n∈N qni fni . j ∈ O(i). If it is an alternative to get some deliverances. These terms can easily be reformulated in the following way: Deﬁne α = qni −fni and β = qni +fni . in particular branch and bound. and so is the quality out of each node. (2. Likewise we need δm and δm to m indicate the positive and negative deviation from the upper quality limit Qmax : m l− l+ qim + δm − δm = Qmin . and this depends on the ﬂow. so solution time will increase exponentially with the numbers of SOS2 sets needed.2. see for instance Williams (1999)). Then qni fni = 1/4(α2 − β 2 ). Modeling multi component ﬂows If we model the ﬂow of C components of the natural gas we require that the split fractions of the components going into the diﬀerent pipelines out of the node n is equal for all components. m ∈ M. but within what is technically acceptable the latter will be chosen.25) qij n∈N fni − n∈N qni fni = 0. n∈N fni i ∈ N . In the SOS2 set at most two variables can be non-zero. This tradeoﬀ will be valued in economical terms as reduction in the customer price.26) This equation has two quadratic terms on the form qni fni . j ∈ O(i). Modeling this as hard constraints could lead to situations where unexpected shutdowns of production ﬁelds or pipelines may lead to a complete stop in deliveries to a customer due to the contractual quality. i ∈ I(m). i ∈ N . For simplicity let us assume we always have only two pipelines out of a split node n ∈ N going 45 . We assume that the resulting blending quality is common for all the downstream pipelines being connected to a node. Where two ﬂows meet. outside the contracted limits. m Gas quality and blending Gas quality is a complicating element because we have to keep track of the quality in every node and pipeline. Still this means that we need to move into solution techniques from integer programming. We need the l+ l− variables δm and δm to indicate the positive and negative deviation from the u+ u− lower quality limit Qmin of customer node m.24) qim + u− δm − u+ δm = Qmax . (2. m m ∈ M.3 A Natural Gas Transportation Model quality in order to increase the ﬂow. i ∈ I(m). Linearizing α2 and β 2 is straightforward using Special Ordered Sets of type 2 (SOS2. (2. The ﬂow in each pipeline is a decision variable in the model. the gas quality out of the node to the downstream pipelines depends on ﬂow and quality from all the pipelines going into the node.23) (2. and the two variables must be adjacent.

Z. cC .29) The ec variables giving the ﬂow through the node of each component is connz strained by the capacity of the node. We need a set of binary variables ϑnz where z = 1.5). c fnj1 = Z z=1 Ez ec . . Then the relation of the volume split between j1 and j2 is equal for all components: c1 c fnj1 fnj = c 1. For each ϑnz we deﬁne a constant Ez giving the percentage related to the z. . We also deﬁne a new variable enz representing the ﬂow through node n of component c if ϑnz = 1. (2.30) 46 .5: The ﬂow is split in node n to node j1 and j2 . ec ≤ Fn ϑnz . All components are indexed from c1 . . . c ∈ C. .32). The ϑnz variable is modeled as a special ordered set of type 1 (SOS1). . . . . Let us also denote the ﬁrst component in the set C of components for c1 . Z}. corresponding to the active ϑnz .27) This is a quadratic expression. where only one variable can be non-zero (Williams 1999). .Chapter 2 Optimization Models for the Natural Gas Value Chain Node n Node j1 Node j2 Figure 2. nz c∈C (2. . and we reformulate it using the equations (2. z ∈ {1. c1 fnj2 fnj2 n ∈ N . (2. . (2.28) The set B consists of all split nodes in the network. The ﬁrst constraint says that the ﬂow from n to j1 of component c equals the percentage Ez multiplied with the total ﬂow through node n of the component c. . nz n ∈ B. z=1 z ∈ {1. .28) to (2. each representing the choice of a split percentage for the share of natural gas going to node j1 . Then we need to restrict the formulation so that only one ϑnz is positive for each node: Z ϑnz = 1. n ∈ B. to node j1 and j2 (see Figure 2. Z}. . . n ∈ B.

The mass balance for the processing plant nodes can then be formulated as in equation (2. This is expressed in equation (2.31) And to make sure that there does not ﬂow more out of the node of each component than what comes in: c c fnj1 + fnj2 = i∈N c fin . nz n ∈ B. Hence.34) Modeling turn-up: ﬂexibility in the production ﬁelds Turn-up is an expression used for the ﬂexibility present in some production ﬁelds. (2. c ∈ C. (2.33). In the modeling of this process it is assumed that the volume of each component extracted is a constant fraction of the total volume of that component in a processing plant (Ac ).32) Processing plants Some of the gas components are extracted and sold in separate component markets. c ∈ C. Hence. The extraction is handled in processing plants in the network.3 A Natural Gas Transportation Model We also require that what ﬂows through the node of each component either goes to node j1 or to node j2 : Z z=1 c c ec = fnj1 + fnj2 . it is not acceptable from a practical point of view that the model presents a ﬂow allocation where ﬁelds with signiﬁcant turn-up capacity will take over production from minor ﬁelds. The turn-up is only used to take over production from ﬁelds that for some reason are prevented to deliver. The variable ac is used to keep track of how r much of component c is extracted from the ﬂow in processing plant r. no decision on the conﬁguration of the pror cessing plant is made.2. 47 . our ﬁrst priority is to meet demand in the network and our second priority is to produce in accordance with the planned production at the ﬁelds. ac = Ac r r i∈N c fir = i∈N j∈N c frj + ac r c fir . r ∈ R (2. which basically is not aﬀected by the shutdown. n ∈ B.34). but pressures and gas ﬂows through a processing plant can be modeled by several processing nodes in sequence or parallel. When modeling this turn-up capacity it is important to keep in mind that even if one are free to utilize this ﬂexibility.33) (2. c ∈ C. For example reduced transport capacity due to a shutdown in one part of the network may be compensated by turning up the planned production from other gas ﬁelds not directly aﬀected by the reduced capacity.

The storages capacity is limited by the physical properties of the storage.6 for an illustration of the terms): Storage capacity gives the maximal volume of natural gas in the storages facility. As before fgj is the ﬂow from g to j: fgj + ∆− − ∆+ = Gg .Chapter 2 Optimization Models for the Natural Gas Value Chain We model this by adding a penalty cost for using turn-up in the objective to avoid turn-up to be used at the expense of normal production capacity in other ﬁelds. a short overview of advantages and disadvantages of these possibilities will be given. In order to discuss the management of natural gas storages. Working gas is the gas volume available during normal operation of the storage.and gas reservoirs. aquifers. the natural gas markets have become more dynamic. 48 .35) g g j∈O(g) 2.4 Management of Natural Gas Storages As a consequence of the liberalization process in the natural gas industry. In this section we discuss models for gas storage operations in a market with uncertain demand. For further discussion of storage facilities. Cushion gas is the amount of gas needed to create necessary pressure in order to lift gas from the storage. This works because not delivering gas to customers would generate a loss which is considerably higher than the small penalty put on using turn-up capacity. In the following. salt caverns and LNG-storages. g∈G (2. The variables ∆− and ∆+ represent underproduction and the use of turn-up g g in relation to the planned production of Gg for ﬁeld g. Storage Facilities The most common storage facilities are abandoned oil. see EIA (2002). This corresponds to the total amount of gas in the storage subtracted the cushion gas. Volume of natural gas in the storage is the total volume of natural gas in a given storage at a given time. The spot markets and the possibility to trade gas in forward markets have increased the importance of gas storages. For some types of storages the cushion gas requirement is as high as 80% of the total gas volume in the storage. The amount of cushion gas needed varies with the type of storage and the geological conditions at the storage location. a couple of terms need to be established (see Figure 2.

The lower part of the ﬁgure is the cushion gas needed for operation of the storage.2. and the upper part of the ﬁgure is the gas currently available for extraction from the storage.4 Management of Natural Gas Storages Injection Working gas Storage capacity Extraction Cushion gas Figure 2.6: The complete square is the total storage capacity. 49 .

Another advantage of this type of storage is the fact that infrastructure is normally already in place. water is used to dissolve halite and to shape cavities in natural salt formations. Cushion gas in the amount of 80 to 90 % is needed for operation. One important use of natural gas storages is to take advantage of the strong seasonal pattern in prices.and gas reservoirs the geological properties are known. 50 . Caverns are created from underground salt or rock formations. One major drawback is the amount of cushion gas needed for operation. the fundamental price determinant in the markets is the weather conditions. These storages consist of tanks containing liqueﬁed natural gas (LNG) or liqueﬁed petroleum gas (LPG). This means that the monthly demand for natural gas may be much higher than the possible changes in production level can satisfy. The storage facility is already in place. These cavities have the properties of a high-pressure gas container. and a cushion gas requirement of only approximately 25 %. underground water-bearing layer which can be transformed into a storage facility by replacing the water with natural gas. The capacity of these tanks is normally very limited compared to the other alternatives presented. aboveground facilities. This adds risk to the development of this type of storages. When using abandoned oil.and gas reservoirs are the most common storage facility. this is not the case when using aquifers. in contrast to the previously presented alternatives. One advantage of this type of storage is the relatively high delivery rate. The use of storages can substitute for investments in new production ﬁelds and transportation capacity. In the salt caverns. One reason for this is the relatively low startup costs. and so is most of the surface installations needed. The storages have a high delivery capacity. Motivation for Utilization of Storages The possibility of storing natural gas gives the participants increased ﬂexibility with regards to production and transportation decisions. The demand is normally higher in winter than in summer. The process of dissolving halite and shaping the cavities makes this alternative more expensive than the previous two alternatives.and gas reservoirs are available. and the production capacity is also lower than the peak demand. Aquifer is a porous. with impenetrable walls. LNG-storages are. Since the primary use of natural gas is for heating and production of electricity. and the development takes time and is costly. These storages are normally only used in locations where no oil.Chapter 2 Optimization Models for the Natural Gas Value Chain Abandoned oil. The diﬀerence between production capacity and peak demand can to a certain degree be satisﬁed through utilization of storages.

. An illustration of the implementation of the SOS2 is shown in Figure 2. . then only one additional variable can be non-zero. . Likewise the outﬂow rate can be given as a strictly increasing convex function of the storage level. . storages could be used to supply the downstream participants. . The storages operate as a security buﬀer in this case. . To be able to realistically represent the inand outﬂow rates. .and outﬂow rates of the storage varies with the current storage level.7: Illustration of the linearization of the injection rate of a storage. .2. HY and the variables ν1 . The storage levels are discretized by a set of constants X1 . . . This means that if νy has a value diﬀerent from 0. When problems occurred either in the production or transportation facilities.7 for the injection rate. Especially for gas producers not having a reservoir close to the market this will be important. With the development of short-term markets and volatile spot prices. XY . The maximal injection rate is a strictly decreasing convex function of the storage level.4 Management of Natural Gas Storages Inflow rate H1 H2 H3 X1 X2 Storage level X3 Figure 2. the corresponding injection rates are H1 . νY are used to give a convex combination of two of the points. It can take several days before a decision to change production level at the ﬁeld will result in increased delivery in the market. the use of special ordered sets of type 2 is chosen (Williams 1999). Traditionally the storages have been used in order to ensure a high security of supply. . The only two candidates in this case are νy−1 or νy+1 . The storage 51 . . the storages will be important for participants wanting to utilize the price ﬂuctuations. Modeling Storages The maximum in.

The importance of these perspectives can be realized when considering the complexity of the transportation system. For more advanced models on portfolio optimization in the natural gas value chain see Rømo. t fis − i∈I(s) i∈O(s) The maximum and minimum levels of storage are modeled implicitly with the representation given. Y y=1 t νys Hys . (2004) which considers tactical value chain coordination. Due to the technical nature of the gas network.39) xt = s xt = xt−1 + s s t νys Xy .37) (2. The minimum level (coming from the requirement of a certain level of cushion gas in the storage) is handled in a similar way: when the minimum storage level is reached. but without stochasticity and without pressure constraints in the transportatin network. Other relevant references are Nygreen et al. the associated outﬂow rate will be equal to zero. SOS2.5 Value Chain Optimization and Portfolio Management We will here give a short description on how to include markets and portfolio optimization in the natural gas value chain. s ∈ S. t fsi . If such values are trespassed. Fodstad & Midthun (2004) from which most of the ideas presented here originate. The maximal level (equal to the total capacity of the storage) is restricted by the inﬂow function. Diﬀerent Levels of Portfolio and Value Chain Integration The models presented here have both a portfolio and a value chain perspective. the associated inﬂow rate is equal to zero. 52 . s ∈ S. 2.36) (2. These are important properties of a natural gas optimization model. When the storage reaches the upper capacity level. s ∈ S. (2. s t fis ≤ i Y y=1 Y y=1 t νys = 1. (2. (1998) which considers portfolio optimization for oil and gas ﬁelds in a strategic horizon and Ulstein et al.38) s ∈ S.Chapter 2 Optimization Models for the Natural Gas Value Chain level at a given time t is represented by xt . several physical and technical threshold-values exist. Tomasgard.

not solving bottleneck problems. We will not give all the previous formulations of the transportation network again. At this level geographical swaps and time swaps of gas can be performed using the market. The motivation behind the portfolio and value chain perspectives can be summarized by considering four levels of planning: 1. but in practice the transition will be gradual. and they are used to fully utilize the ﬂexibility in the system. but each time period t in a value chain model will include transportation network constraints and variables like the ones from Section 3 with an additional index t on all variables. forwards or options may be optimal. Trading: At this level contracts and ﬁnancial instruments are traded independently of the physical production and contract obligations based on market opportunities. 2. If the producer is risk averse hedging the portfolio outcome using futures. but the motivation is now speculation. The purpose is to maximize the proﬁt from production and contract obligations using also spot markets. These trades are in no way connected to the physical production and contract obligations. 3. unless the producer has market power. The bottlenecks in the transportation system make the ﬂexibility incorporated in a system perspective valuable. as the portfolio variation will be lower than the variation of the stochastic parameters of separate ﬁelds or contracts. The market can be used to resolve bottlenecks in the transportation network or on the production side. 53 . The distinction between level 2 and 3 is clear in theory. In that case it is enough to supplement physical production with trades in the spot market at level 2. 4. Risk management: So far we have assumed the producer is risk neutral and tries to maximize expected proﬁt. Stochastic demands and prices that are not perfectly correlated motivate a portfolio perspective on the planning. Production and market optimization: At this level markets are used to supplement the physical production in order to gain more from the physical production capabilities.5 Value Chain Optimization and Portfolio Management only minor incremental deliveries in one part can cause signiﬁcant unintended reductions elsewhere. The trading is similar to the previous level in terms of using the spot market and ﬁnancial instruments like futures and options.2. Traditional production planning: In this ﬁrst level the model ensures balancing of the production portfolio with the contract portfolio.

54 . while ﬁeld B sells spot in Zeebrugge • The company may buy spot in Emden and the procution from ﬁeld B can be sold in the spot market in Zeebrugge. In a situation where ﬁeld B needs to produce and the company has an obligation to deliver in a bilateral contract in Emden several possibilities exist: • Field A supplies Emden.8. while the production from ﬁeld B is sold elsewhere. Consider the network in Figure 2. • The company buys spot in Emden.8: Example of a natural gas network Utilization of Short-Term Markets in Value Chain Optimization The use of short-term markets allows for considerable ﬂexibility in the system. while it sells the production from B spot in the upstream market. For example bottlenecks in the production or in the transportation may be resolved or moved using the markets actively. These simple geographical swaps makes the system more ﬂexible and gives the company the possibility to maximize the ﬂow of natural gas (and the value of their production) beyond what traditional transportation planning would have done. • Storage might be used to supply Emden.Chapter 2 Optimization Models for the Natural Gas Value Chain Field A Field B Storage Upstream market Zeebrugge Emden Figure 2.

Several options are then available to the company: • In period 1 ﬁeld B may supply storage. When including market transactions in the model a representation of the uncertainty in the price process is important. In this description some simpliﬁcations have been made. the number of possible routing decisions gets very large and the ﬂexibility increases. 80 pipelines and 10 markets. The possibility of trading forward contracts is only interesting for a risk adverse company. the possibility of delaying production through lifting agreements will be disregarded. This will be discussed shortly at the end of this section. Consider Figure 2. for instance. • In period 1 ﬁeld B can sell spot upstream. • In period 1 ﬁeld B can sell spot in Zeebrugge.9 illustrates how the market nodes are included in the model. This time ﬁeld B needs to produce in time 1. Based on this representation scenarios describing the uncertainty can be generated and optimal decisions in the interaction between the physical system and the market can be made. In the spot market the company can purchase or sell volumes of 55 . The arrows show that gas might ﬂow from the transportation network to the market. we would also need to model the other companies’ transportation needs.8 again. Including Markets and Contracts In Section 3 only aggregated deliveries to take-or-pay contracts in the diﬀerent customer nodes m ∈ M were considered.2. The gain from portfolio thinking increases because they are not perfectly correlated. In addition. There is no ﬂow from the market to the network (as would be the case for an upstream market). The need for ﬂexibility comes from the fact that demands and prices are stochastic. we assume there is only one company in the markets. and only trades in the spot market will be considered. Figure 2. For simplicity of notation. It is this ﬂexibility we try to capture when modeling the portfolios of production ﬁelds and contracts.5 Value Chain Optimization and Portfolio Management A diﬀerent reason to use the markets is time swaps. and in period 2 either use a forward contract or buy spot in Emden. Only one company is considered. The network considered is also very small. When expanding the network to. so no upstream market exists. 20 ﬁelds. The ﬂexibility further increases when perfect spot markets are added. and then use either a forward contract or the spot market to supply Emden This is just some of many possibilities that exist for geographical swaps and time swaps. and the company has an obligation to deliver in time 2. and in period 2 the storage supplies Emden. transactions within the market node can be performed. If not. We assume the company is a price taker.

9: The market node 56 .Chapter 2 Optimization Models for the Natural Gas Value Chain Natural gas network Market Forward Bilateral contracts Spot Figure 2.

5 Value Chain Optimization and Portfolio Management natural gas. . or by transactions within the market node. contract prices and spot prices. Each stage in the tree typically consists of several time periods. We denote the stochastic ˜ contract price for contract k in customer node m at time period t ∈ T2 as φt . t ξm Deterministic parameters for customer node m in t ∈ T1 . When time passes on and one enters the ﬁrst t ∈ T2 . 6 and 10). but only nodes after a branching are decision points. see Figure 2. T2 } where some parameters are stochastic (as seen from t ∈ T1 ). Time periods 1 and 2 are in stage 1. . and a set of time periods t ∈ T2 = {T1 + 1. .2.10 there are 3 time periods. as they are the only time periods when new information about the future is resolved. The nodes in the scenario tree represent decision points.10. ˜ We use a tilde over the variable to reﬂect that it is stochastic (as seen from t ∈ T1 ) and remove the tilde when the variable is deterministic. . or t ∈ T2 after uncertainty is resolved (Seen from a point in time t where t ∈ T2 ). In practice decisions are only made when new information becomes known. A stage is the set of time periods elapsing between each time information is learned by the decision maker. Modeling Stochasticity We use the modeling paradigm of stochastic programming to represent uncertainty in the models. and uncertainty is resolved along the arcs going out of a node with several branches. uncertainty is resolved and also the remaining time periods can be considered deterministic. . Several parameters are stochastic in reality. decision variables are present in time periods where information is not resolved. We will consider stochasticity in: contractual demands. . The vector of all stochastic t ˜t t ˜ ˜ variables in time period t is ξ = (ψ . starting with the decision in node 0 and ending just before the new decisions at stage 2 (in nodes 2. . see for example Kall & Wallace (1994). Obligations in the take-or-pay contracts can be fulﬁlled either by ﬂow from the network to the market node. We then get the following: ˜t ξm Stochastic variables for customer node m in time period t ∈ T2 seen from a point in time t ∈ T1 . In the model presented here we use a two-stage formulation for ease of notation. ˜mk ˜t The stochastic spot price is represented with ψm . . In Figure 2. T1 } belonging to the ﬁrst stage where information is deterministic. hence the time periodization using time periods t reﬂect in which time period the decision has eﬀect. µ ). 57 . Uncertainty is then represented in a scenario tree. φ . mk Stochastic demand for contract k in customer node m at time period t is µt . In a two-stage stochastic programming model we deﬁne a set of time periods t ∈ T1 = {t1 . Still.

where for example the ﬁrst 6 months would belong to T1 and the last 6 months to T2 in a two-stage formulation. The decision variables and constraints are equal in all time periods. except for initialization in time period 0 where only initial storage levels are deﬁned x0 and for the terminal conditions ns at the end of the model horizon.40) 58 . but assume the scenario tree exists in the remaining part of this paper. The only stochasticity that is present is in the right hand side and in the objective. and the planning horizon would typically be 12 months. Hence the objective can be described by summarizing the expected cash ﬂow of the time periods. We will not go in detail on how to do this here. The objective is to maximize expected proﬁt taken into consideration cash ﬂows and shortfall costs. The typical length of a time period for a tactical planning model is one month. The cash ﬂow of each time period ˜ t can be described as a function Πt (xt−1 .Chapter 2 Optimization Models for the Natural Gas Value Chain Scenarios 2 3 0 1 6 7 10 11 Time Figure 2. The proﬁt function for time period t ∈ T1 ∪ T2 can be formulated as: Πt (xt−1 . The Objective We describe the supply chain portfolio optimization model as a two-stage stochastic program with relatively complete recourse (Kall & Wallace 1994). We use the vector x0 to denote the initial level of all storages and xt to denote the level of all storages in time period t. ξ t ) = m∈M k∈K φt µt + mk mk m∈M t t− t+ ψm (ζm − ζm ). ξ t ) (or ξ t if stochastic) where xt−1 is the storage level in the start of the time period.10: An example of a scenario tree A scenario tree can be constructed for example using price processes for natural gas or descriptions of the dynamic aspects of stochastic demand. (2.

3. The + sign indicates purchases of natural gas whilst the − sign indicates sales. (2. mk ∀m ∈ M. Delivery in contract type k in the node m in time period t are included in µt . The constraint sets are identical for all time periods t ∈ T1 ∪ T2 . since the spot market can be utilized for this. mk The mass balance equation illustrates the ﬂexibility gained by including markets in the model. The solution will however show how the network should be managed in order to achieve the highest possible expected proﬁt. but we will look closer at the ones changing because of the introduction of markets and contracts. or minimize the cost of achiving a given throughput. ∀t ∈ T . It is no longer necessary to ship the gas to the market node in order to fulﬁll the contractual agreements. This means that the solution does not necessaritly maximize the throughput in the network. EGV(xT2 ). Constraints Including Markets and Contracts The mass balance in the market node for each time period and each scenario is expressed as: t t+ t− fim + ζm = ζm + i∈I(m) k∈K(m) µt .43). 59 .41) where Q(xT1 ) = max Eξ [ ˜ ˜ Πt (xt−1 . This function is described in more detail in Section 6. (2.42) t∈T2 subject to a set of constrains representing transportation. production and markets.43) t In (2. ζm represent transactions in the spot market in node m in time period t. For the last time period the objective includes the terminal value of the natural gas in storages expressed by the Expected Gas Value function. (2. The two-stage stochastic program with ﬁxed relatively complete recourse is: max t∈T1 Πt (xt−1 ) + Q(xT1 ). These constraints are mainly the constraints descried earlier in this paper.2. The solution resulting from maximizing expected proﬁts will normally be diﬀerent from the solution reached with the objective function presented in Section 2.5 Value Chain Optimization and Portfolio Management where the ﬁrst term is the income from delivery in contract k in market m at time t and the second term gives the proﬁt from trading in the spot market in node m in time period t. ξ t ) + EGV xT2 ]. This means that geographical and time swaps are now available to the company.

but rather maximize a utility function that incorporates the risk aversion of the company. the mass balance in the market node (see (2.Chapter 2 Optimization Models for the Natural Gas Value Chain Risk Aversion In the discussion so far only the possibility for trading natural gas through the spot market has been discussed. one and a half million columns. but under some assumptions still valid. see for instance Hull (2003). In addition to the change in the objective function. The fact that natural gas is a commodity makes the argument less obvious. Since the forward price is a good predictor of the expected future spot price. If the objective is only to maximize ﬂow in a static model. with 100 scenarios and 10 breakpoints (see Section 2. four million non-zero elements and fourteen thousand binary variables. a solution with an integrality gap of 4% to 5 % typically can be reached within 12 hours. and it is diﬃcult to ﬁnd an optimal solution. a stochastic model with a network consisting of approximately 80 nodes and 1000 scenarios.43)) will be changed to incorporate the possibility to trade in the forward market. In this case the company will no longer maximize expected proﬁts from their operations. can be solved within an hour. When excluding the gas components. Another way of doing this is to introduce a penalty function that will put extra cost in the objective function on deviations from some target proﬁt value. solution times are within minutes when components are omitted and increases correspondingly when components are added. Solution Times The complexity of the models introduced in this paper to a large extent depends on the modeling of the gas components. that is: to reduce the risk of their position. The inclusion of gas components adds a large number of integer variables to the problem. By trading forward contracts a given price can be locked in on advance. In the case where the company is risk averse however the situation changes and some tools to handle risk management are needed. trading in the forward market would on average be approximately equal to trading on the spot market (this is based on a simple arbitrage argument. This problem will have approximately one million rows. For a physical system similar to the one above. The inclusion of a forward market then gives the company the possibility to hedge. When including gas components the solution time increase signiﬁcantly. 60 . For a risk neutral company that is maximizing expected proﬁts this is an adequate approach.3).

This way of modeling the end-of-period however allows for limited ﬂexibility and also neglects the true value of the gas contained in the storage. depending 61 . One way of handling the storage problem is to set a target value for the storage level at the end-of-horizon.6 The Expected Gas Value Function (EGV ) So far no considerations have been made with respect to how the ﬁnal period in the planning horizon will be handled. Figure 2. Hence.6 The Expected Gas Value Function (EGV) 2. The option value in gas storages comes from the operational ﬂexibility. This alternative value comes from the opportunities present after the end of the model horizon. The model presented so far will most likely end up with a very low storage level. In practice this is a real-options approach used to value the value of gas in the storage. based on expectations for the future development of the spot price of natural gas.44) This level might also be made dependent on various factors. the alternative value of the natural gas in storage is thereby included. Hence for each end-of-horizon storage level.12). whilst if lower prices are expected the optimal level in period T2 may be lower. scenarios describing possible future outcomes can be constructed (see Figure 2.2. for an overview of some of them. for instance the starting level. and the production might also be higher than optimal when considering a longer horizon (since the value of the gas still contained in the reservoirs is neglected). These can be modeled using stochastic processes. Several versions of such models exist.11 illustrates how the EGV is included in the existing optimization model. When the EGV is used as a boundary value. several possibilities exist. Based on the chosen price model. In order to handle the end-of-horizon problem. the estimation of EGV is performed independently from the value chain model and the purpose is to give a boundary condition for the value of gas. As the ﬁgure shows. the EGV must reﬂect the value of an optimal out-of-horizon strategy for injecting gas in the storage and selling gas from the storage. A way of determining the optimal level for the storages in the last period is by using the expected gas value function. see Schwarz (1997). xT ≥ x0 s s (2. such as the season in which the end-of-period belongs. The company can switch between injection. The Expected Gas Value function (EGV ) gives an estimate of the value of a unit of gas in storage at some point in time t. If high prices are expected in the future. the EGV will encourage a high storage level in ﬁnal time period T2 . for any given initial storage level a strategy is found for injection and withdrawal of natural gas based on a stochastic process for the gas price. An important element in the model used to estimate EGV is the development of the natural gas spot price represented through spot price curves. withdrawal or idle modes.

11: The estimation of the EGV is performed in a stochastic optimization model that is independent of the existing optimization model.Chapter 2 Optimization Models for the Natural Gas Value Chain Portfolio management model Field A Field B Storage Upstream market Zeebrugge Emden EGV Total value of storage Storage level Stochastic Programming Model EGV valuation Figure 2. The EGV is then used in the value-chain model as a boundary value on the gas in storage and reservoirs. 62 .

Pereira & Pinto (1991). The procedure is based on Scott et al. A tree similar to the one constructed for the spot price (Figure 2. In this case a recombining trinomial tree. This can be illustrated by (2. A multilevel tree representing the spot price and the amount in storage is then created.12: Representation of the development of the spot price of natural gas. and also for the combination of all or some of the storages in the network. The injection value is the negative value of gas injected in this period. In the latter case a more complicated model is needed for estimation of the EGV. The hold decision value is equal to the expectation of the option value of the next steps. For references to similar work in hydro power. In this case the nodes represent diﬀerent storage levels. an example of how the EGV can be calculated is given. plus the expectation of option values of decreased storage levels in coming nodes. while the nodes represent diﬀerent price scenarios. and use a stochastic dynamic programming framework. which shows Spot price 63 .2. a discrete approximation of the storage facility state space is made. when storage level is unaltered. (1999). plus the expected value of increased storage level in future nodes.12) can be constructed also for the storage level. The arcs in the ﬁgure represent price movements. see for instance Pereira et al. see for instance Hull (2003).45). (2000) and Manoliu (2004).6 The Expected Gas Value Function (EGV) on the price development. while the arcs represent injection and withdrawal of natural gas in the storage. In the following. injection and withdrawal. It is possible to do this estimation both for individual storages. The option value is calculated in each node by taking the maximum of the decision values of hold. For references on real-options. The valuation is performed by backward induction through the tree. The withdrawal value is then the value of releasing gas in this period. After choosing a stochastic model to represent the price of natural gas. Time Figure 2.

and therefore also estimation of the 64 . The MVs shows the expected marginal value of gas for various levels of storage. This is determined by the value of ﬂow π t (ϕt ). MV4 MV3 Total value of storage MV2 MV1 Storage level Figure 2. This is the additional value of one more unit of natural gas in the storage. (2. (2. (where ϕt is the volume injected or withdrawn in period t) in the node in period t plus value of the storage level τ t+1 in the next time period (in nodes following the considered one).13. the liquidity in these markets is still very limited. Even though short-term markets for natural gas are developing in Europe.46) An illustration of a gas value function is given in Figure 2.Chapter 2 Optimization Models for the Natural Gas Value Chain the value in a given node in the tree: I t (τ t ) = π t ϕt + I t+1 τ t+1 .13: An example of an expected gas value function.45) I t (τ t ) is the value of storage level τ in time period t in the node considered. This lack of liquidity makes estimation of the spot-price process diﬃcult.46): τ t+1 = τ t + ϕt . The storage level is updated according to (2. as well as ﬁnding the breakpoints. The challenge is in ﬁnding the appropriate total value for each level of storage.

Finally we deﬁned the Expected Gas Value function and explained its use for giving the terminal value of stored natural gas and indicated how to calculate it.7 Conclusions EGV diﬃcult. but many of the formulations would be relevant for more general models focusing on other parts of the natural gas value chain.2. Thereafter we introduced stochasticity in demands and prices and gave a stochastic programming formulation for a portfolio optimization model. In this paper we have taken the position of a large producer. Next we introduced the time aspect and the use of storages. Given that a spot-price model can be modeled for any given time horizon. supporting multi commodity ﬂows and pressures. Natural extensions of this model would be contract selection and more advanced modeling of the production ﬂexibility reﬂected by lifting agreements.7 Conclusions In this paper we have gradually introduced the complexity of a stochastic optimization model for the natural gas value chain. We started out by deﬁning necessary constraints for a steady-state formulation of the underlying transportation network. the discount rate will make the gas value in the last periods decrease strongly. 2. Most of the model formulations presented here are simpliﬁed variants of models that are implemented for commercial use on the Norwegian continental shelf. 65 . We focus on coordination of the diﬀerent levels of the chain and on a portfolio perspective. a time-horizon of a couple of years may be appropriate for estimating the EGV. As the time-horizon for estimation is increased.

Z The set of split percentages used to discretize possible split fraction in split-node of the network. T The set of time periods included in the model. s Storage facility s ∈ S. Y The number of discretized storage and injection rate levels used to linearize storage characteristics. r Used for nodes with processing plants. n ∈ N . C The set of components deﬁning the chemical content of the natural gas. L The set of breakpoints used to linearize the Weymouth equation. g Used for nodes with production ﬁelds. O(n) The set of nodes with pipelines going out of node n. y Breakpoints for linearization of injection rate levels in storages. g ∈ G. The set of nodes where gas ﬂows are splitted into two or more pieplines. M Nodes with buyers of natural gas: typically import terminals.Chapter 2 Optimization Models for the Natural Gas Value Chain Appendix 2.A Notation and Deﬁnitions Sets The set of all nodes in the network. The set of nodes in the network with production ﬁelds . b Split nodes. 66 . I(n) The set of nodes with pipelines going into node n. l Breakpoits in linearized Weymouth restrictions. c Component c ∈ C. K(b) The set of contracts in node b ∈ B. i and j will be used. k Contract k ∈ K. R The set of nodes with processing capabilities. m Customer nodes b ∈ B. When more indexes are needed. m ∈ M. t Time period t ∈ T . N G B Indexes n Used for nodes in general. z Breakpoints in linearization of split percentages in split nodes. S The set of nodes with storage facilities.

Pressure reduction factor in node n ∈ N . Fixed point for pressure out of a pipeline. Max pressure [bar] into the pipeline from node i ∈ N to node j ∈ N . Fraction of component c in processing plant r that is extracted from the ﬂow. Penalty cost for deviation from contracted pressure level. Upper limit for ﬂow through node n ∈ N . Penalty cost for deviation from contracted quality to customers. Power output capacity of the compressor. Value of gas to customer b. where GCV [mol%] is the (Gross Caloric Value). Compressor factor in node n ∈ N . Min energy content requirement for gas deliverances to node n ∈ N . Storage capacity [Sm3 /|t|]in node s ∈ S. Penalty cost for use of turn-up.A Notation and Deﬁnitions Constants Gg Fij Fn max Pn max Pij min Pij Pb Qmax n Qmin n Db Sl W Kij Ac r PI PO Γn Θn η Ka W max ωb κ ι χ Ez Xy Hy Planned production [Sm3 /|t|] in ﬁeld g ∈ G. Min pressure [bar] out of the pipeline from node i ∈ N to node j ∈ N . Demand in standard cubic meter pr time unit [Sm3 /|t|] for natural gas in node b ∈ N . 67 . 3 Energy quality is given by (GCV or CO2 ) or [M J/Sm ] .2. Compressor eﬃciency. Discrete representations of storage level in linearization of storages. Max energy content requirement for gas deliverances to node n ∈ N . Adiabatic constant for a certain gas type. Gives the split percentage related to a given z in linearization of split nodes. Penalty cost for pressure level. Discrete representations of injection rates in storages. The Weymouth constant is used as a constant in an empirical expression for linking ﬂow and pressure in pipelines. Max pressure [bar]into a node n ∈ N . Fixed point for pressure into a pipeline. Upper limit for ﬂow through the pipeline from node i ∈ N to node j ∈ N . Target pressure [bar] for deliverances to a customer node b ∈ B.

Chapter 2 Optimization Models for the Natural Gas Value Chain

Decision Variables

c fij Flow from node i ∈ N to node j ∈ N of component c. In some cases index c is omitted when we do not consider multi commodity ﬂow. fn Total ﬂow into node n. enz Flow through node n of component c used for linearization in splitting nodes m ∈ M. pin Pressure [bar] into the pipeline going from node i to node j. ij pout Pressure [bar] out of the pipeline going from node i to node j. ij qij Gas quality (GCV or CO2 ) in pipeline going from node i to node j. νy Give convex combinations of Xy and Hy . σin Equal to 1 if ﬂow from i to n, otherwise 0. ϑnz Binary variable representing split percentage in node n. ac Amount extracted of component c in plant r. r ρij Equal to 1 if ﬂow goes from i to j, otherwise 0. t− ζm Volume sold in spot market m in time period t. t+ ζm Volume bought in spot market m in time period t. l+ δb Positive deviation from the lower quality limit Qmin of customer b node b. l− δb Negative deviation from the lower quality limit Qmin of customer b node b. u+ δb Positive deviation from the upper quality limit Qmax of customer b node b. u− δb Negative deviation from the upper quality limit Qmax of customer b node b. xt The storage level at a given time t in a storage s ∈ S. s + Positive deviation from the contracted pressure to customer b. b − Negative deviation from the contracted pressure to customer b. b ∆+ Positive deviation from the planned production in ﬁeld g. g ∆− Negative deviation from the planned production in ﬁeld g. g

Functions

EGVt (xs ) Expected gas value in time period t as a function of the storage level in storage s. Wij (P I, P O) Flow resulting from pressure diﬀerence between Pressure in, P I and pressure out, P O, of a pipeline according to the Weymouth equation. Γ(fn ) Compressor factor as a function of ﬂow fn into the node n ∈ N .

68

2.A Notation and Deﬁnitions

Stochastic Variables

˜ φt bk µt ˜bk ˜t ψm ˜ ξt Contract price for contract k in customer node b in time period t. Demand for contract k in customer node b in time period t. The spot price in market m in time period t. ˜ ˜ ˜ The vector of all stochastic variables φt , µt and ψ t .

In time periods where these parameters are not stochastic or where uncertainty is resolved, the tilde is dropped in the variable name.

69

Y.pdf.eia. & Wallace. To appear in International Journal of Theoretical and Applied Finance. H. (2000). Upper Saddle River. De Wolf. J. NTNU. ‘The basics of underground natural gas storage’. M. ‘Storage options valuation using multilevel trees and calendar spreads’. European Union (2003). A. John Wiley & Sons. A linear model for transient gas ﬂow. Prentice Hall International. Trondheim. (2004 ). ‘Directive 98/30/EC of the european parliament and of the council’. & Smeers. (2003). & Transport (2002). Chichester. Futures and Other Derivatives. European Commission. M. (1994). Options. ﬁfth edn. ‘Opening up to choice: Launching the single european gas market’. available as SINTEF-report STF 38S03601. 71 . E. ‘The gas transmission problem solved by an extension of the simplex algorithm’. D. Hull. N. Gas Conditioning and Processing. pp. EIA (2002). Nowak. New York . 7 edn. D. (2003). The Equipment Modules.doe. http://www.gov/pub/oil_gas/ natural_gas/analysis_publications/storagebasics/storagebasics. (2000). Kall. W. f. in ‘Proceedings of ECOS’. S. European Union (1998). Stochastic Programming.Bibliography Campbell. (2003). PhD thesis. M. Okla. SunGard Energy Systems. F. in ‘Conference Proceedings: IAEE Praha’. ‘Directive 2003/55/EC of the european parliament and of the council’. J. P. & Tomasgard. & Westphalen. (1992). 1454–1465. Norman. Oﬃce for Oﬃcial Publications of the European Communities. Management Science 46(11). Model-Based Dynamic Control and Optimization of Gas Networks.J. Hofsten. Manoliu. 1751–1758. K. Rømo. An optimisation model for rationingeﬃcient allocation of capacity in a natural gas transportation network.. Dahl.-G.

A. PhD thesis. 4th edn. ‘Storing true value?’. Schwarz. P. Ulstein. Brown. F. (2004). M.. Norway. (2000). (1997). Optimization in Operations Research.. A. T. & Sagli. ‘Multi-stage stochastic optimization applied to energy planning’. Tomasgard. Rømo. Working paper. Tactical planning of oﬀshore petroleum production. M. Williams. & Kelman. Bjørkvoll... Model Building in Mathematical Programming. SINTEF. International business school and research center for natural gas. R. T. Westphalen. Tomasgard. Math in Gas and the Art of Linearization. Ø.. N. Nygreen. N. 251–267. Application of stochastic dual DP and extensions to hydrothermal scheduling. Technical report. PSRI. ‘Modelling norwegian petroleum production and transportation’. Rømo. M. PhD thesis. R. University Duisburg-Essen (Germany). K. Trondheim. Norway. Mathematical Programming 52. F. & Kristiansen. B. Prentice Hall. Scott. Trondheim. Haugen. 72 . & Pinto. & Nowak.. (2004). B. (1998). New Jersey. SINTEF. J. (2004). N. Groningen. K. The Journal of Finance 52(3). H. Van der Hoeven. 923 – 973. Annals of Operations Research 82. Trondheim.. M.. Chichester. T. Christiansen. Campodónico. M. Technical report. & Midthun. (1999). Rardin. NTNU. Anwendungen der Stochastischen Optimierung im Stromhandel und Gastransport. 359–375. (1991). The Netherlands. Energy and Power Risk Management . (2004).Bibliography Nygreen. (2004).. Pereira. John Wiley & Sons. ‘The stochastic behavior of commodity prices: Implications for valuation and hedging’. (1998). & Perry. Fodstad. M. E. Steady-state optimization of natural gas pipeline ﬂow. Upper Saddle River. L. L. Stochastic optimization of the natural gas value chain in a liberalized market. (1999). H. Technical report. Pereira.

Midthun.Paper II Kjetil T. Mette Bjørndal and Asgeir Tomasgard: Modeling optimal economic dispatch and ﬂow externalities in natural gas networks Submitted to international journal .

.

pressure drops and 75 . 3. and examine the implications it has for economic analysis. The economic objectives include maximization of social surplus.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities in natural gas networks Abstract: We present a combined framework for modelling the technological issues of gas transportation and analysis of natural gas markets. like the relation between ﬂow and pressure. including the spatial demand and supply for the commodity in the optimization of the transportation system. Our approach combines a framework for modeling the technological characteristics of natural gas ﬂows with optimization modeling of markets.e. In our framework we model the optimal dispatch of supply and demand in natural gas networks. In the existing literature. consumer surplus and producer surplus. with diﬀerent objective functions. the liberalization process in Europe and other places in the world. The proposed network ﬂow model for natural gas takes into account pressure drops and system eﬀects when representing network ﬂows. we show that it is important to represent the underlying physical properties. In our paper.1 Introduction In this paper. Incremental increase in production in one part of the network may cause signiﬁcant reductions elsewhere. The Weymouth equation is linearized such that it is possible to perform economic analysis in large networks within our framework. we have found no approach that combines the modeling of the technology of natural gas ﬂows with economic analysis of the transportation system. The models take into account the physical structure of the transportation network. we present a model for analyzing the optimal economic dispatch of natural gas in pipeline networks. i. has led to an increased interest in such market-oriented models. The importance of combining economics with a model for pressure drops and system eﬀects is illustrated by small examples. and diﬀerent economic surpluses. During the last years. Pressure drops are modelled by the Weymouth equation.. More speciﬁcally there are externalities in pipeline networks due to pressure constraints and system eﬀects. maximization of ﬂow.

where ﬂow externalities are not taken into account. Chao & Peck (1996) present a consistent system of ﬂow gate prices. Westphalen (2004) and De Wolf & Smeers (2000). The linearization that we use in our modeling framework is based on Rømo et al. (2003) and Cremer & Laﬀont (2002). in this paper we focus on the qualitative eﬀects of including the physical properties of natural gas ﬂows. (2005) and Gabriel et al. (2004). Due to Kirchhoﬀs’ laws in meshed electricity networks. Schweppe et al. A recent PhD thesis on the linearization of natural gas ﬂows is Van der Hoeven (2004). A key ingredient of an electricity market design is the management of congestion or bottlenecks in the transmission network through locational prices or similar measures. and illustrate the eﬀects in a small network example with two production nodes. (2006). These models focus mainly on the economic issues and do not give a detailed description of the engineering aspects underlying transportation networks. (2001). Some examples of existing models with focus on ﬂow maximization and natural gas transportation physics are Ehrhardt & Steinbach (2004. without considering the special physical properties of the network ﬂows.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities resulting system eﬀects. When it comes to papers with focus on the economics of natural gas transportation. We show examples of the errors that will be made if the technology is not modeled. Combining the physics and the economics in models and analysis has for years been the standard way of analyzing electricity transportation and markets. In An et al. there are severe externalities that complicate the pricing procedures. Gabriel et al. the studies we have found have a straightforward representation of gas ﬂows. 76 . (2003) a non-linear model for optimization of the combined natural gas and electricity power ﬂow is presented. and introduced the concept of optimal nodal prices. and in some cases. and we illustrate and discuss the externalities that arise from system eﬀects in the network. The linearization of the gas ﬂow equations will make it computationally feasible to analyze even large-scale networks. 2005). In Wu et al. and Hogan (1993) describes a hedging system for locational prices. surprising results when optimizing the operation of the network. Some examples are for instance Cremer et al. Examples of economic equilibrium models are given in Gabriel & Smeers (2006). through transmission congestion contracts. In the electricity sector on the other hand. However. two market nodes and a single transportation node. The inclusion of the technology leads to interesting. Martin et al. In Cremer & Laﬀont (2002) the possibility of a participant with market power is also considered. the eﬀects on the physical power ﬂows are typically taken into account when allocating capacity and dispatching units. and the social welfare maximizing solution is derived in a simpliﬁed setting. (1988) formulated the optimal economic dispatch problem for electricity markets. in the economic dispatch models. where the optimal allocation of resources has been discussed.

network externalities are discussed under the traditional objective function.2 Natural gas ﬂows (1996).2 Natural gas ﬂows In our presentation. European Commission & Transport 2002). it is vital that the prices provide the correct signals of the value of the commodity to the market participants. In Section 5. showing both a nonlinear model. counter examples are given. We apply a linearized model for the technology.3. transportation nodes and market nodes. In order to achieve an eﬃcient market. the European Commission passed the gas directive in 1998 (European Union 1998. A consequence of the gas market liberalization is that we may expect to see less emphasis on long-term sales contracts. in the same manner as the models we have seen for the liberalized electricity markets. emerging short-term contract markets. we assume that there is one production ﬁeld in each production 77 . using the Weymouth equation. while the economic models may lead to non-linearities in the objective functions. and a linearized approximation of it. we describe network ﬂow models for natural gas pipelines.e. to describe a natural gas system. and in some cases. division of activities within the ﬁrms’ value chain. maximizing social surplus. The key components of the gas directive are third party access to all transportation installations. In Europe. followed by a second gas directive in 2003 (European Union 2003). For simplicity. some eﬀects of the externalities in electricity networks are investigated. the analysis is extended to alternative objective functions. A thorough discussion of the interaction between the transmission network and the energy markets is given in Bjørndal (2000). Common assertions about general network performance and characteristics are scrutinized. Some ﬁnal conclusions are provided in Section 6. Developments during the last ten years have made it more important than ever to better understand the integration of technology and economics in the natural gas markets. In Section 2. maximizing ﬂow or throughput. and consists of long pipelines. The transportation system for natural gas in the North Sea is chosen as a motivating case for our work. we will use production nodes. 3. producer surplus or consumer surplus. as well as pipelines. two numerical examples are given to illustrate the discussion in Section 4. vertical separation. we utilize price responsive supply and demand curves. The purpose of this paper is to provide a framework for modeling natural gas markets that will enable future analysis to capture the technical as well as economic issues. In Section 4. It is the largest existing oﬀshore network.. Thus. where the modeling of pressure drops along the pipelines is important. For this. In Section 3. and more spot sales. i. central coordination of production and infrastructure utilization may be replaced by market prices as the coordinating mechanism. and the possibility for certain consumers to obtain their gas from the supplier of their choice.

5 (3. we will ﬁrst show how to model pressure and ﬂow in a single pipeline. where the gas is traded. from these nodes. where transportation nodes usually lack compressors due to the high installation and maintenance cost of such sub-sea installations. butane. The ﬂow through the pipeline is driven by the pressure diﬀerence between the inlet and the outlet of the pipeline. We will assume that it is not possible to increase pressure in the transportation nodes. The gas quality in terms of its chemical composition and energy content will vary from production ﬁeld to production ﬁeld. which typically take the form of minimum and maximum pressure requirements.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities node. there may be limitations on pressure levels. but rather assume that all the gas in the system is equal and may be viewed as one commodity. carbon dioxide and several others. Usually. Thus. This reﬂects the present system of the North Sea. In the market nodes. we also assume that there are compressors. but in reality modeling gas components may introduce even more externalities into the system. methane.1). there are no compressors available in the transportation nodes. and then give a full mathematical model. Flow in a single pipeline The Weymouth Equation We model the pressure drops in pipelines based on the Weymouth equation. depending on the compressor unit characteristics and pipeline design.e. which describes the relationship between inlet pressure. Also in the market nodes. In our approach. we will discuss the system eﬀects of pressure. and a homogeneous gas leaves the transportation node through one or several other pipelines. outlet pressure and the amount of gas ﬂowing in a pipeline. The natural gas is mixed. several pipelines meet at transportation nodes in the network. we will not model the natural gas components. we have only incoming pipelines. Natural gas is a mix of diﬀerent hydrocarbons and other gas components like ethane. and the larger the diﬀerence. i. In the natural gas ﬂow model we present here. gas may be sent into the system at a pressure level that can be chosen within certain limits. 44 × 10−3 Q= 8P 2 2 ri − rj d5 R M Ts ZLF 0. thereafter.. In the production nodes. the more gas molecules will ﬂow between the two points in a given period of time (see illustration in Figure 3. The Weymouth equation is given by: Tπ 1. for instance due to contractual agreements.1) 78 . Our discussion will not be inﬂuenced by this assumption.

79 .1: A three-dimensional illustration of how the Weymouth equation relates pressure at the inlet and outlet points to the capacity in the pipeline.2 Natural gas ﬂows 200 150 Flow 100 50 0 150 160 140 100 100 50 Pressure out 80 Pressure in 120 200 180 Figure 3.3.

[kg/kmol]. due to for instance pipeline design parameters or capacity limitations in nodes. L The pipeline length. the ﬂow through the pipeline will be constrained by these limits.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities For details on this empirical equation. Z The average compressibility factor of the gas. W Wij (ri . ri Inlet pressure in the pipeline. see for example Campbell (1992). Ri . [m]. F The friction factor of the pipeline. rj ) = Kij 2 2 ri − rj .2). [m]. [K]. rj Outlet pressure in the pipeline. rj ) is the ﬂow through a pipeline going from node i to node j as a consequence of the pressures ri and rj : W Wij (ri .2) around a point (RIi . in this paper. Consequently. [J/(kmol × K)]. ROj ). TS The ambient temperature of gas. R The gas constant. rj ) ≤ Kij Ri − Rj 2 . we focus on capacity constraints following from restrictions on pressure levels only. [bar ]. The volume of natural gas is given in standard cubic meters (Sm3 ).3) There may be other capacity constraints limiting ﬂow. However.1) have the following interpretations: Q The throughput (or ﬂow rate). and the outlet pressure is at minimum level. (3. T The temperature at standard conditions. Linearization of the Weymouth Equation Through Taylor series expansion it is possible to linearize Equation (3. M The molecular weight of the gas. where normal conditions are deﬁned to be 1 atmospheric pressure (1. The parameters and variables in Equation (3. d The pipeline diameter.01325 bar) and 15 . representing ﬁxed pressures into and out of the pipeline: 80 . The Weymouth equation can then be expressed as in Equation (3. P The pressure at standard conditions. We often aggregate the constants of the Weymouth equation into a Weymouth W constant Kij for the pipeline going from network node i to node j. [K]. [bar ].2) If there are limitations on the pressure level in the nodes. 2 (3. [bar]. where Wij (ri . since the ﬂow must be less than or equal to the ﬂow that may be obtained if the inlet pressure is at maximum level. Rj . [M Sm3 /day]. The standard cubic meter refers to one cubic meter of gas under normal conditions.

When we use the linearization to compute ﬂow for each pipeline. Under this assumption. (3. we can model a more advanced node with valve arrangements. representing diﬀerent pairs of RIij and ROij (where RIij is larger than ROij ). The main reason is that it better reﬂects the typical situation of many real transportation networks. where in a network node. as presented in its original non-linear form in Equation (3. In this case. ROj ) will take the following form for a pipeline between nodes i and j (for a detailed description. including the one in the North Sea. In practice. For a node with more than one pipeline connected. the outlet pressure in all incoming pipelines and inlet pressure of all outgoing arcs. the practical capacities that can be utilized in one part of the network depend on the pressures and ﬂows elsewhere in the system.4) is not valid. limiting the size of problems that may be analyzed. this corresponds to modeling the pipeline capacity as an inequality. are equal. Another reason is that the formulation using Equation (3. as modeling ﬂow using an inequality in this constraint. First. The linearization in Equation (3. ∂ri ∂rj After some calculations.2 Natural gas ﬂows Wij (ri . rj ) ≤ Wij (RIi . There are two alternative ways of modeling the system eﬀects of pressure. around 20 of these constraints are added to the problem. that we will 81 . see Appendix A): W fij ≤ Kij RIi 2 RIi − ROj 2 ri W − Kij ROj 2 2 RIi − ROj rj .3. the Taylor expansion around (RIi .1) leads to a non-convex problem. the Weymouth equation must be modeled as an equality. The linearized variant is used in models developed for the system operator in the North Sea for analyzing network ﬂow. the chosen pressure level in the node inﬂuences the capacity in all pipelines connected to the node.4) may be used.4) where fij is the ﬂow variable. Network ﬂows and system eﬀects In a natural gas network.1). we can choose to model the situation. One should note that the externalities due to system eﬀects. ROj ) + ∂Wij ∂Wij (ri − RIi ) + (rj − ROj ) . Alternatively. basically corresponds to allowing a pressure drop at the inlet of the pipelines. the linearization in Equation (3. that makes it possible to reduce the inlet pressure in some of the outgoing pipelines. allowing the ﬂow in the pipeline to be lower than the ﬂow following from a given node pressure.

In this example. a transportation node and two market nodes. as given by the Weymouth equation. the absolute value of the slope of the corresponding line in the ﬁgure must increase. Since the minimum pressure in node E is assumed to be lower than the corresponding ﬁgure for node D. Maximum and minimum pressure requirements in the nodes are given by horizontal lines in the ﬁgure. we will assume that any pressure limits in node C are not restrictive. A B C D E Figure 3.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities describe in the following sections. and the pressure diﬀerences along the pipelines. The maximum pressure in production node A is assumed to be larger than in production node B. either the outlet pressure needs to be reduced. the capacity of pipeline CE is larger than the capacity of CD. There are also minimum pressure requirements in the market nodes D and E. To increase the ﬂow in a pipeline. Assuming similar design parameters and similar lengths for the two pipelines AC and BC. while the y-axis represents the pressure.2. we will illustrate the system eﬀects of pressure with the example network in Figure 3. will be present also in the ﬁrst case. There are two production nodes (A and B).3. The pipeline ﬂows depend on the Weymouth constants of the corresponding pipelines. or in other words. as it gives an even stronger link between the diﬀerent pipeline ﬂows. this gives the pipeline from A to C larger capacity than the pipeline going from B to C. 82 . In the following. assuming the two pipelines are otherwise equal.2: Example of a transportation network consisting of two production nodes. two market nodes (D and E) and a transportation node (C). when an equality is used in Equation (3.1). or the inlet pressure must increase. The downward sloping lines connecting two nodes represent a pipeline and show the inlet and outlet pressures and the pressure drop along the pipeline. The relationship between the pressure levels in the pipeline system is illustrated in Figure 3. The x-axis represents the location of the nodes.

Sets N G M T I(n) O(n) The The The The The The set set set set set set of of of of of of all nodes in the network. it means that pressure must increase in the production ﬁeld to accommodate this. The fact that there are no compressors to increase pressure in the transportation node. transportation nodes. this means that the pressure in node C has to be increased. and since there are no compressors in node C. nodes with pipelines coming from node n. (2006). Consider now a situation where production ﬁeld A is oﬀered a new contract for delivery to market E. and in Section 3. nodes with pipelines going into node n. A natural gas network model In the following. but less gas now ﬂows from B to D (the line joining B and D is ﬂatter) because of the higher pressure in node C. market nodes.3. while it is still possible to increase the pressure in node A. The new situation is shown in the lower part of Figure 3. introduce externalities. based on work at SINTEF and NTNU Rømo et al. The available transportation capacity from node B has been reduced as a consequence of the increased transportation from node A. the pressure in node A must be increased relative to the pressure in node C. we have assumed that the pressure in node B is at the upper limit. but less capacity is available from the production nodes to node C. Moreover.4 we will extend this model to include economic objectives. To accommodate this increase in ﬂow from A to E. Besides. the pressure in node E is at the lower limit. we present a linear programming model for physical ﬂow maximization in the transportation network. and the fact that pressure in node B cannot increase. and there is a trade-oﬀ between the capacity of the upper pipelines (before the transportation node) and lower pipelines (after the transportation node).3. more capacity is available from node C to the markets. nodes in the network with production ﬁelds. in practice. Since the pressure in node E is at its lower limit. the pressure diﬀerence between node C and E must increase in order to increase the ﬂow through this pipeline.3. In the new situation. 83 .2 Natural gas ﬂows In the upper part of Figure 3. More gas ﬂows from A to E (the line joining A and E has become steeper).

84 . while the lower ﬁgure shows the state of the network after the ﬂow from A to E is increased.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities Figure 3.3: The top ﬁgure shows the original state of the network.

6) For the transportation nodes. Price of natural gas in market m ∈ M. Min pressure [bar] in node i ∈ N . Decision variables kg fij ri qm pd m ps g Production in ﬁeld g ∈ G. the amount of gas that ﬂows into node j must also ﬂow out of node j: fij = i∈I(j) n∈O(j) fjn .3.2 Natural gas ﬂows Constants Ri Ri Max pressure [bar] in node i ∈ N . j∈O(g) (3. (3. Flow from node i ∈ N to node j ∈ O(i). Volume of natural gas in market m ∈ M.5) where fim is the ﬂow from node i to market node m. Objective function The mathematical formulation for maximizing ﬂow is straightforward: max i∈I(m) m∈M fim . Price of natural gas in production node g ∈ G. M is the set of all market nodes and I(m) is the set of all nodes with pipelines going into m. g ∈ G. Pressure [bar] in node i. Production kg in node g ∈ G must equal the amount of gas fgj transported from the production node g into nodes j in its set of downstream nodes O(g) : fgj = kg . qm . j ∈ T . is equal to the sum of the gas ﬂowing into the market: 85 .7) In the market node m we need to make sure that the quantity of gas available in the market. Constraints The ﬁrst set of constraints ensures that mass is conserved in the network. (3.

When we discuss the economic modeling of gas networks. the capacity of the pipelines connected to the node is dependent on the nodal pressure. . ROl ). l = 1. A more precise approximation is obtained by using a set of linearizations around the pairs of inlet and outlet pressures (RIl . Thus. We denote this constraint set the Weymouth formulation (WF). where pressure and system eﬀects are modeled properly. and therefore also dependent on one another. With the linearization of the Weymouth equation. L. we will compare the ISF formulation with the WF formulation that we suggest should be used.. L. pressure levels in one part of the network will depend on the chosen pressure values in other parts of the network.9) (3. like Cremer et al. based on the Weymouth equation. j∈I(m) (3.10) We will present two alternatives for modeling pipeline capacity constraints due to pressure limitations.. j ∈ O(i). . The linear constraint in (3. The alternative model for pipeline capacities. . where we use L linear constraints for each pipeline. (2003) and Cremer & Laﬀont (2002). (3. l = 1.8) Moreover.. . . for each pipeline. The Weymouth formulation is given by: W fij ≤ Kij RIil 2 RIil − 2 ROjl W ri − Kij ROjl 2 2 RIil − ROjl rj .11) i ∈ N .Chapter 3 Modeling optimal economic dispatch and ﬂow externalities fjm = qm . thus reducing ﬂow in a given pipeline relative to its pressure potential. The ﬁrst is the one that we advocate. the nodal pressure formulation ensures that pressure is not built up at nodes without compressors. Still. The system eﬀects of pressure are implicitly modeled by the nodal pressure variables. m ∈ M. we need to make sure that the maximum and minimum requirements for the pressure in the nodes are satisﬁed: ri ≥ Ri . 86 . i ∈ N.4) corresponds to an outer linearization of the Weymouth equation for the pipeline ﬂow. The second alternative is the formulation usually found in economic models of gas systems. is the ISF-formulation. but allows for pressure drops. (3. If the pressure in a node must be uniform for all connected pipelines. i ∈ N. which we will denote Independent Static Flow constraints (ISF). ri ≤ Ri . the inlet pressure of a pipeline out of transportation node n is equal to the outlet pressure in a pipeline going into the node.

consider a situation where the producer in node A wants to deliver 41 MSm 3 /d to the market in node E. gives a ﬂow from A to E of 36. Section 3. i ∈ N . which can be achieved in theory. Network externalities are present both in operational max ﬂow problems and in long-term investment problems. it is an open question how to determine pipeline capacities. we will illustrate the diﬀerence in two numerical examples.1 will be used. the system eﬀect is not properly modeled.17 MSm 3 /d from B to D. the system eﬀects disappear from the model.2. In the next section. as the ﬂow capacity of one pipeline is not inﬂuenced by other pipelines’ ﬂows. we present two examples that illustrate the importance of including pressure drops in the modeling of natural gas ﬂows. j ∈ O(i). In Section 3. 3.two numerical examples fij ≤ Cij .two numerical examples In this section. When replacing the Weymouth constraints with ISF constraints.. let us investigate how an increase in ﬂow from node A to node E can inﬂuence the total throughput. Now. The design parameters in Table 3.3 Max ﬂow and network externalities .16 MSm 3 /d . One possibility is to use the maximum pressure diﬀerence for the pipeline to determine its static capacity: W Cij = Kij Ri − Rj 2 . 2 (3. The pressure constraints of the pipelines are due to the design parameters in the network.12) where Cij is the static ﬂow-independent capacity for the pipeline connecting nodes i and j. and the capacity is set at its best possible value.2. a total ﬂow to the market nodes of 68.4 we extend the analysis to situations with alternative economic objectives.13) In this case. and 31. (3.3 Max ﬂow and network externalities . We analyze the network using pressure constraints and pressure drops and compare it to an approach with ISF capacities.3. Example 1: Externalities from network operations Returning to the example network in Figure 3. Note that in this case. as well as compressor capacities at the production nodes. and speciﬁcations in the delivery contracts. Maximization of throughput in this system. but not necessarily in practice.99 MSm 3 /d . i. The Weymouth constants are chosen such that they resemble pipelines in real installations in the North Sea. How will this inﬂuence the ﬂow from 87 .e.

The maximal throughput is then equal to min(100.13) to compute ISF capacity constraints on the pipeline ﬂows. a transportation node and two market nodes. Thus. the ﬂow from B to D is reduced by more than 8 MSm 3 /d .8 MSm 3 /d .4. The ﬁgure to the left gives the original solution. [bar] Weymouth-constant Table 3. and this change has decreased the transportation capacity from the production nodes to node C.2 gives the following result: ﬂow from A to E is.8+ 36. 37.4: Example of a transportation network consisting of two production nodes. It is evident that the relationship between ﬂows in the network is far from linear. when the ﬂow from A to E increases by 4 MSm 3 /d .38 C to E 140 70 0. CCD = 37. and the ﬁgure to the right the ﬂow when the ﬂow from A to E is increased. while if we consider the contractual paths. while the ﬂow from B to D is reduced to 22.41 C to D 140 100 0.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities A to C 200 120 0.68 MSm 3 /d . as desired. 41. [bar] Min pressure out of pipeline.23 + 41. The pressure in node C has increased from 129 bar to 139 bar . assuming production node A trades with market E and production node B with D. gives the following ﬂow constraints: CAC = 100. the capacities for trades are CAE = min(100.9.22) MSm 3 /d 88 .23 MSm 3 /d .63 B to C 150 120 0.8. CBC = 36. 41 MSm 3 /d . Figure 3.22) MSm 3 /d = 78.22 MSm 3 /d .9 MSm 3 /d .45 MSm 3 /d . and CCE = 41. and that rather small changes in one part of the network can have a large inﬂuence on other parts.34 Max pressure into pipeline. Using Equation (3. The results are illustrated in Figure 3. based on speciﬁed max and min pressures.1: Design parameters B to D? Imposing this new constraint and maximizing throughput using the model from Section 3.

makes it possible to increase the ﬂow from node B by one unit.6.13).3. reducing ﬂow from node A to any of the market nodes by one unit. 37. where. consisting of production node A. Using any of these capacities either fails to recognize the dependency of ﬂows completely (the contractual capacities).3 Max ﬂow and network externalities . the nodes can be connected to the existing pipeline from A to D.5. A single producer in node A is transporting gas to the market node D through a single pipeline. the additional pipelines are shorter than if a completely new pipeline should cover the whole distance. Assume that we have a simple basis network to start with. The ﬁgure on the right shows the ISF capacities of contract paths.45 MSm 3 /d as compared to the 68. However. By using the pipeline from A to D. Instead of building a new pipeline going from B to E.5: The ﬁgure on the left shows the pipeline network with ISF constraints. Example 2: Externalities and investment decisions Understanding the physics of the network system is important in order to make good infrastructure and capacity decisions. with a max ﬂow of 78. The capacities are illustrated in Figure 3. Another issue is that the ISF constraints calculated from Equation (3. for instance. The ﬁeld in node B will deliver its production to a new market in node E. Figure 3.16 MSm 3 /d if pressure levels and limits are taken into account.9.two numerical examples and CBD = min(36. based on maximum and minimum pressure limits. The network is illustrated in Figure 3.22) MSm 3 /d . or exhibit a linear relationship between ﬂows. In the following. market node D and two intermediate nodes CP 1 and CP 2. a small ﬁeld is under development (node B). Assume that the pipeline going to market node E can be connected at 89 . grossly overestimates the practical capacities of the pipeline network. we show an example of network externalities due to investment decisions.

the two alternative junction points lead to diﬀerent transportation capacities between A 90 .Chapter 3 Modeling optimal economic dispatch and ﬂow externalities Figure 3.7. and we assume that the pipeline from B to A is not a bottleneck in the system.7: Two diﬀerent junction points. With these values. Figure 3.6: The original network with the two potential junction points. After investment. one of two intermediate positions on the existing pipeline from A to D: CP 1 or CP 2.3 MSm 3 /d . The investment possibilities are illustrated in Figure 3. before any investments are done.2. and does not inﬂuence the pressure at node A. This is a typical realistic investment situation on the Norwegian continental shelf.2. the ﬂow capacity of the pipeline between A and D is 51. we solve the max ﬂow problems for the two cases using the model in Section 3. The design parameters are given in Table 3. Assuming that A is responsible for delivering to node D and that B is responsible for the volume going to node E.

62 0.8. If the pressure constraints had not been included in the model. 91 . the capacity CAD is 44.2 MSm 3 /d and 3. With CP 1. We assume in both cases that the volume between B and E is ﬁxed to 10 MSm 3 /d .4 MSm 3 /d between the two alternatives.1 MSm 3 /d . the pipeline from CP 2 to E is approximately 40 % longer than the pipeline from CP 1 to E.2: Design parameters and D. The 1 The 160 bar 80 bar 70 bar 0.69 0. and with CP 2.3 Max ﬂow and network externalities . Since the pressure in node A was at the maximum level in the original solution. 600 km from A to CP 1. the Weymouth constant decreases (see Equation (3.1)). The result of the numerical example is illustrated in Figure 3. The graph on the top left hand shows the original situation with volumes and pressures.1)).8 MSm 3 /d respectively. The Weymouth-constant is dependent on the square of the length of the pipeline (see Equation (3. This decrease is due to the need for higher delivery between A and the junction point.52 lengths used in this example is 840 km between A and D. and the one underneath is the solution when using CP 2 as a junction point.62 0.37 0. and therefore the diﬀerence between the constants may be non-intuitive. we see that the pressures in the nodes CP 1 and CP 2 have decreased. 300 km from A to CP 2. The graph on the top right hand shows the solution if CP 1 is chosen. and therefore. The reduction in capacity between A and D is 7. and then maximize the throughput between A and D. In the alternatives presented here. 240 km from CP 1 to D and 300 km to E.44 0.two numerical examples Max pressure into pipeline from A Min pressure out of pipeline into D Min pressure out of pipeline into E Weymouth-constant from A to D Weymouth-constant from A to CP 1 Weymouth-constant from A to CP 2 Weymouth-constant from CP 1 to D Weymouth-constant from CP 1 to E Weymouth-constant from CP 2 to D Weymouth-constant from CP 2 to E Table 3. Comparing the solution for the extended network with the original network. so there is a daily diﬀerence of 3.5 MSm 3 /d . 540 km from CP 2 to D and 424 km to E.46 0. this increase can only be obtained by reducing the pressure in the junction points. the location of the junction point would not have inﬂuenced the capacity between A and D. can be explained by looking at the Weymouth equation: when the length of the pipeline is increased. The diﬀerence in the solutions for the two alternative junction points.3. less gas is transported for a given pressure diﬀerence1 . the capacity is 47. Another aspect to take into consideration is the investment cost of the new pipelines.

8: Flow and pressure values for the original network and for the two investment possibilities.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities Figure 3. 92 .

there is a trade-oﬀ between the additional investment costs of building a longer pipeline and the possibility to get more capacity in the network. as we want to focus on the eﬀect of the capacity constraints following from pressure limitations. we use linear demand and supply functions. will inﬂuence the solutions also when using diﬀerent economic objective functions. This is a simpliﬁcation.4 Optimal dispatch with economic objective functions example shows that there are several considerations that need to be accounted for when we are looking at expanding a natural gas network. ownership in the ﬁelds and in the network. From the max ﬂow examples.4 Optimal dispatch with economic objective functions Incorporating pressure constraints and pressure drops when modeling natural gas ﬂows. Our approach follows along the same lines as the analyses of congestion in electricity markets. Flow dependent or pressure dependent costs could be taken into account by adding corresponding cost terms in the objective function. pipelines. We focus on the short-term optimal operation of the natural gas system. The optimal solution will depend on several factors. transportation and market nodes. and compressors with predetermined properties. We start by deﬁning diﬀerent economic objectives. before we summarize wellknown results for the single pipeline case. the 93 . In our presentation. consumer surplus and producer surplus. given the existing infrastructure. we saw that the externalities in natural gas ﬂows inﬂuence the practical capacities of the pipeline networks. regulation of the network and market conditions (supply and demand). such as investment costs. By introducing economic objectives. as there may be some ﬂow dependent and pressure dependent costs associated with for instance using gas for building up pressure. and ﬁnally move on to our contribution in the analysis of economic objectives in a pipeline system.3. In the following. representing marginal production cost and marginal beneﬁts from consumption. We will assume that all the relevant short-term costs are reﬂected in the supply functions. however. respectively. with production. where congestion management is frequently studied assuming lossless approximations of power ﬂows. we assume that there exist supply functions in the ﬁeld nodes and demand functions in the market nodes. Moreover. The owners of ﬁeld A must be compensated somehow for the loss of capacity after the new investments. With the linearized Weymouth equations to model pipeline ﬂows. we will investigate the validity of economic reasoning based on simple ISF constraints as opposed to the more detailed modeling of the network in our approach. We consider the maximization of the following objectives: social surplus. 3. we have not considered these cost terms.

maximizing social surplus corresponds to maximizing the trapezoid limited by the supply function. At quantity fcap the two nodes will have diﬀerent marginal values. a market node. or through various two-part tariﬀs. and one possibility is that a system operator can bring along the equilibrium by giving the producer a price of ps and charging a price of pd from the market. Figure 3. PS and A. f ∗ . see also Appendix 3.13).Chapter 3 Modeling optimal economic dispatch and ﬂow externalities resulting optimization problems are quadratic optimization problems with linear constraints. for example in Tirole (1988).e. Overviews of existing theory for analyzing single transportation links with capacity constraints exist.. as in Figure 3. however. equal to func . Clearly. in this case. Economic dispatch in a single pipeline Consider the situation in Figure 3. In the unconstrained solution. and a single pipeline connecting the two nodes. With a congested pipeline with capacity fcap < func .B)).10. the marginal values should be consistent with the prices ps and pd . so an ISF constraint for the pipeline ﬂow. We will not discuss these issues further in this paper. with a system consisting of a production node. i.10. there is no pipeline system to interact with. such as Equation (3. for instance by managing capacity limitations through redispatching in secondary markets. We call this solution the constrained optimal economic dispatch. is suﬃcient to represent the maximum amount of gas possible to transport. The transportation capacity is restricted by the design parameters of the pipeline. Note that there may be other ways to implement the optimal solution. independent of how capacity constraints are managed. the sum of the areas CS . consumers and a possible system operator. we give a short introduction here (for more details.9: Example of a transportation network consisting of a single pipeline.9. For completeness. The income to the system operator would then be equal to the area A in Figure 3. maximizing social surplus. in the optimal solution. the demand function and the capacity constraint. maximizing social surplus gives an optimal gas ﬂow. These diﬀerent alternatives for dealing with transmission constraints will typically result in diﬀerent allocations of total surplus between suppliers. the optimal unconstrained quantity f ∗ is usually less than func (refer Appendix 3. When maximizing producer or consumer surplus. Since the optimization problems have concave objective functions and convex 94 .B).

This means that for suﬃciently constrained networks. the consumer or producer respectively. We focus on the diﬀerence between using the modeling framework described in Section 3. the optimal quantity produced and transported over the pipeline is equal to min(fcap . with several production and market nodes. This means that maximizing consumer surplus is done by maximizing the area CS + A. the pipeline will be used to the capacity limit in the constrained cases where fcap < f ∗ . maximizing producer surplus is the same as maximizing the area P S + A. An interpretation of this assumption is that when maximizing consumer or producer surplus. f ∗ ). In the following.e.B.10: Situation when pipeline capacity is restricting the solution. feasible regions. For details on the single pipeline optimization see Appendix 3.4 Optimal dispatch with economic objective functions Figure 3. i. and the network operator function. will take all the proﬁt in area A.3. Likewise. Economic dispatch in a pipeline system In this section we formulate the optimal economic dispatch in a network of pipelines. we will assume that when maximizing either consumer surplus or producer surplus.2 as opposed to applying ISF capacities throughout the network. Supply and demand functions for the production and market nodes are assumed 95 . the optimal solution will be identical and equal to fcap for all three objective function alternatives.. we assume either the consumer or production side to be in control of the pipeline system.

6)-(3.18) s. we will compare the results obtained from two classes of optimization models.18) s. have the following formulations. 2 g g (3.t. and αm and βm are positive m constants.16) or (3. For simplicity.14) d d where pd is the price.8) and (3. depending on volume qm .17) • Max producer surplus (monopoly): max m∈M d d 2 αm qm − βm qm − g∈G 1 s 2 β k .15) s where ps is the price in production node g. Hence.13)} 96 .16) • Max consumer surplus (monopsony): max m∈M 1 d 2 d αm qm − βm qm 2 − g∈G s 2 βg k g . kg is the volume produced.t. and βg is g a positive constant. 2 g g (3.17) or (3. consumer surplus and producer surplus. Thus. g (3.11)} 2. the demand function in market node m ∈ M is given by: d d pd = αm − βm qm .16) or (3.18) When comparing models with the Weymouth formulation to models with ISF capacities.10: • Max social surplus: max m∈M 1 d 2 d αm qm − βm qm 2 − g∈G 1 s 2 β k .Chapter 3 Modeling optimal economic dispatch and ﬂow externalities to be deterministic and linear. The objective functions then. (3. such that supply in production node g ∈ G is given by: s p s = βg k g . (3. which is illustrated in Figure 3.6)-(3.17) or (3. ISF capacity constraints: {max (3. (3. m (3. we assume that the supply functions go through the origin. and are generalizations of the two-node situation. WF pressure constraints: {max (3. maximization of social surplus. we analyze all the diﬀerent objective function variants. using: 1.

4. we move on to a strongly constrained network.3.5 Optimal economic dispatch .13) fails to recognize the dependencies between pipeline pressures. where ﬂow and ﬂow constraints are modeled by means of the Weymouth equations.13) is a relaxation of the true model. 97 .11. we provide numerical examples illustrating the eﬀects that pressure constraints have on optimal economic dispatch in a natural gas network.11 and the prices (or marginal values) in Table 3. but that Equation (3.5 Optimal economic dispatch . This is due to the fact that modeling pressure in a network. we obtain a relaxation of the optimal dispatch problem.45 C to D 150 90 0. If this approach is used to model ﬂow constraints due to pressure limits throughout a network. Example 3: Optimal dispatch and ISF capacities This example illustrates how the ISF capacities introduced in Equation (3. while the ISF capacities are calculated from Equation (3.13) gives the highest possible ﬂow between two nodes i and j with pressure limits Ri and Rj .2. In the following. where all objective function alternatives give the same solution with ISF constraints. consider a network with the same conﬁguration as in Figure 3. This example shows that even in a highly constrained network. the solutions are diﬀerent for the alternative objectives when including WF pressure constraints. In a network. This is evident from noting that Equation (3. We start with a moderately constrained network that illustrates the diﬃculties in ﬁnding ISF capacities. if not.13) overestimates the true capacity in the network. the optimal solution is equal to the solution from the ISF capacities model.30 C to E 150 75 0.two numerical examples In this section. The design parameters are shown in Table 3.3: Design parameters We then use the supply and demand functions in Table 3.35 Table 3. As can be seen from the ﬂows in Figure 3.5 to compute the optimal dispatch under the three diﬀerent objective functions.two numerical examples 3.40 B to C 180 125 0.3. Maximum pressure into pipeline Minimum pressure out of pipeline Weymouth-constant A to C 160 125 0. If this is possible.13) and displayed in Table 3. introduces some ﬂexibility that is not incorporated with the ISF capacities. Thereafter. it will not normally be possible to run every pipeline with maximum pressure diﬀerence between the connected nodes. the ISF formulation based on capacity constraints (3.

3 54.4 127.3 83.5 42.6 141 127.13) directly.9 Producer surplus WF ISF 35. as the necessary capacity adjustments will depend on supply and demand parameters as well as the design parameters of the transportation network.4 110.4: ISF capacities in the network the results depend a great deal on which model is used to represent network ﬂows.8 124.1 145.3 54.5: Supply and demand functions.21 9058.5 96.4 6115 6316.6 102.03 Price in node A Price in node B Price in node D Price in node E Objective value Social surplus WF ISF 34.2 8255.69 41. but rather make some reductions to allow for the system effects of the pressure constraints.57 Table 3. the ﬂow pattern given by the model with ISF constraints is naturally not necessarily feasible for the pressure constraint (WF) model.37 39.6 62.6: Prices and objective function values. it is an open question how this should be accomplished. we see that the solutions for the ISF capacities give higher total ﬂow for all the objective functions. As expected.88 7145. It may be argued that it is unrealistic to apply the capacities from Equation (3.6 124.6 54. Objective Consumer surplus WF ISF 38. However.7 42.7 96 140 120.57 Table 3.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities Capacity Capacity Capacity Capacity from from from from A B C C to C to C to D to E 42. The objective function value is also increased for all the objective functions under the ISF capacities.4 127. 98 .4 6820. However.33 60.6 42. Supply from node A Supply from node B Demand in node D Demand in node E ps 1 ps 2 pd 1 pd 2 = k1 = 3k2 = 200 − 2q1 = 200 − 3q2 Table 3.

99 .5 Optimal economic dispatch .two numerical examples Figure 3.11: Solution with pressure constraints (on the left) and ISF constraints (on the right).3.

and the total production (and consumption) would be 102.7: Design parameters For the network model with WF pressure constraints and pressure drops. independent of the objective function. the production is highest in node A while sales are highest in node D. grid revenue (A) and the 100 . inﬂuence the solutions. note that in the unconstrained case. maximizing social surplus. even in this very strongly constrained network. The design parameters are displayed in Table 3. Maximizing producer surplus gives a market price of 123. and the same supply and demand functions as in the previous example (see Table 3. and Table 3.5 MSm 3 /d .7.2) = 42.4. When maximizing the three diﬀerent surpluses in this constrained network. however. which is here equal to min(39.13) are shown in Table 3.10 displays the allocation of surpluses between consumers (CS). 21.2 + 21. The ﬂow patterns are illustrated in Figure 3. As expected. The resulting ﬂow pattern is illustrated in Figure 3..13. is similar to the single pipeline case in Section 3. and P S areas in Figure 3. The solutions maximizing consumer surplus and producer surplus both give lower production volumes. showing the production in the ﬁeld nodes and the volumes consumed in the market nodes.10 and are equal independent of which objective is used (CS + A.12. A.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities Example 4: Optimal dispatch in a strongly constrained network We continue to use the example network in Figure 3. Since the optimal unconstrained solution for every objective function is beyond the capacity of the network.30 C to E 135 115 0.35 C to D 135 115 0. using the ISF formulation.5 MSm 3 /d . P S + A or CS + A + P S.8. the market price is 111. the network will be used to the capacity limit.8 and a total production of 63. First. I.2. Maximum pressure into pipeline Minimum pressure out of pipeline Weymouth-constant A to C 172 130 0. the dependencies between ﬂow in diﬀerent parts of the system.1 and the total production is 74 MSm 3 /d .e. the competitive solution. In order to analyze a strongly constrained network.7. we choose design parameters such that none of the ﬂows from the unconstrained solutions are feasible.4.5). producers (P S).4 + 12. The identical results for the three objectives. and the ISF capacities calculated by Equation (3. the ISF network model gives the same value for the decision variables. would have resulted in a price of 76. The size of the CS.9 in every node in the network. When maximizing consumer surplus.10) are given in the ISF column in Table 3.30 Table 3.35 B to C 135 130 0. the optimal ﬂow patterns are no longer identical for the diﬀerent objective functions.

101 .2 Table 3.12: The solution when maximizing social surplus.7 21.4 12.3.5 Optimal economic dispatch . consumer surplus and producer surplus with ISF capacities.2 21.two numerical examples Figure 3. Capacity Capacity Capacity Capacity from from from from A B C C to C to C to D to E 39.8: The ISF capacities.

the eﬀective capacity of each pipeline is determined by the operation of the other pipelines in the system.2 663.6 6494. this is once more due to the fact that the ISF solution is not feasible when we take into account the pressure limits and the externalities in the network. The total surplus in the ISF model is higher. This result is due to two factors: the capacity constraints on the network.1 733.10). the resulting increase in A is larger.93 20. See Figure 3.9 CS PS A CS + P S + A Social surplus 1095. and therefore increased the capacity from node B to node C at the expense of the capacity from node C to nodes D and E. In this example. we start by noting that in the unconstrained case. In the monopsony case.6 674. To summarize the results. we note that compared to the solution with maximal social surplus.48 6682. and gives possibilities of moving and creating bottlenecks in the system.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities total social surplus CS + A + P S) for the diﬀerent objective functions (see Figure 3.1 4715.56 20.5 20.87 Production in node A Production in node B Sale in node D Sale in node E Table 3. With the pressure constraints.6 722. 102 . CS and P S. Consider now the WF solutions.10 for deﬁnition of the areas A.93 Objective Consumer surplus 34. and the special properties of the pressure constraints. Objective Consumer Producer surplus surplus 1053.16 4884.3 6537. the monopsony has decreased the pressure in node C. This reduces the size of the traditional consumer surplus CS. maximizing consumer surplus (CS + A).7 6. however.6 Surplus with ISF 1123. Social surplus 37.5 1089. such that total monopsony surplus (CS + A) increases.17 4.24 Table 3.10: Results from the optimization.56 4.5 Producer surplus 37. some of the production has been switched from production node A to production node B .4 6543.87 20.3 20. This is contrasted with the allocation of surplus in the ISF model.3 20. This adds ﬂexibility to the operator of the system.9: Result from optimization. however.1 4777.2 4726.

5 Optimal economic dispatch .two numerical examples Figure 3.3. consumer surplus and producer surplus with pressure constraints. 103 .13: The solution when maximizing social surplus.

allows the monopsony or monopoly to get a higher surplus than they would in the solution with maximal social surplus. The eﬀects of the network externalities are investigated by means of numerical examples in a small network consisting of two production nodes. due to diﬀerent assumptions on who is to receive the grid revenue A. in order to analyze a natural gas transportation system. and at the expense of the other party. this means that when using the ISF formulation. there exist severe externalities in the operation and development of these networks. An analysis of bottlenecks and threshold values in the transportation network therefore must have a system perspective rather than a pipeline perspective. The possibility of moving bottlenecks in the system when modeling pressure and system eﬀects. we provide a modeling framework that takes into account these externalities. This means that changes in one part of the system will inﬂuence capacities and performance in other parts of the system. The externalities arising from system eﬀects and pressure constraints thus inﬂuence even the more qualitative aspects of the solutions. it is demonstrated that the relation between ﬂows in the system is highly non-linear and non-additive. When modeling the physical properties of the network using only ISF capacity limits to restrict the solutions. 104 . Moreover. that describes ﬂow in a pipeline. and the competitive structure of the markets. the solutions may diﬀer with diﬀerent objectives. when modeling the network ﬂows with pressure limits and pressure drops. two market nodes and an intermediate transportation node. optimal ﬂow patterns and marginal values or prices. Thus. On the other hand. the diﬀerent objectives will give identical solutions in terms of ﬂow when the network is suﬃciently constrained. 3. is proposed in order to allow for analysis of ﬂows and economic surpluses even in large networks. For the very constrained network.6 Conclusions Due to the physical properties of natural gas transportation networks. Only the wealth distribution changes with the objective in this case. A linearization of the Weymouth equation. it is shown that ﬂexibility in operations arise from these system and pressure eﬀects. From these simple examples. in the sense that it is possible to move bottlenecks in the system with a proﬁt for either producers or consumers. such as the eﬀects from diﬀerent assumptions on the market design. The network externalities may inﬂuence to a large extent eﬀective capacities. A and P S. it is necessary to take into consideration the entire network. Flows are constrained by pressure limits and driven by pressure drops. In this paper.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities the three objective functions will in general lead to diﬀerent ﬂow patterns in the system. the monopsony/monopoly cannot do better than the social surplus solution in terms of changing the size of the areas CS.

6 Conclusions Thus. taking into account the technology in terms of physics and system eﬀects. is important for economic analysis in natural gas networks.3. 105 . we argue that modeling the physical properties of natural gas networks.

20) and using the ﬁxed point RI . (3.24) which ﬁnally translates into the following approximation for the ﬂow through a pipeline with pressure diﬀerence ri − rj : 106 .22) Putting these expressions into Equation (3.20) We start by ﬁnding the partial derivatives: 2Kri δW .23) (rj − RO) . δri δrj (3.A Linearization of the Weymouth equation The Weymouth equation is given by: Wij = K 2 2 ri − rj . RO.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities Appendix 3. rj ) ≤ Wij (RI . RO) =K − This can be written as: RI 2 − RO 2 RI − RO RO RI 2 − RO 2 rj . RO) + δW δW (ri − RI ) + (rj − RO) .19) By using a ﬁrst order Taylor series expansion around a ﬁxed point RI . Wij (RI . this can be expressed as: Wij (ri .21) (3. RO we obtain the following result: Wij (RI . RO) =K −K RI 2 − RO 2 − + RI RI − RO 2 2 ri (3. = δri 2 2 2 ri − rj δW 2Krj =− . 2 2 RI 2 − RO 2 + KRO RI 2 − RO 2 KRI RI 2 − RO 2 (ri − RI ) (3. δrj 2 2 2 ri − rj (3.

Also assume demand and supply functions are known and linear.19 is the upper quarter of a cone starting in the origin.B Unconstrained equilibrium rj . RO) = K RI ri − RO 3.B Unconstrained equilibrium Consider a single pipeline where the capacity of the pipeline is not binding. (3. (3. and the traded quantity is f ∗ corresponding to the intersection between the supply. qm . The highest possible social surplus is achieved for the system in a competitive market. 2 2 Wij (RI . are identical and equal to the ﬂow in the pipeline. β s + 2β d β s + 2β d 2αs β d + αd β s ps∗ = .29) (3. Supply: ps = αs + β s qm .26) (3. kg . This gives the following solution: f∗ = MC = αs + β d f.28) βs + βd βs + βd Maximization of producer surplus can be done by implementing the monopoly solution.3. The linearization constraints are planes passing through the origin. the intersection between marginal revenue and marginal cost is found. where the prices are pd∗ and ps∗ . RO) is tangent to the cone in all points ri RI where rj = RO . see Westphalen (2004). Each such plane linearized around (RI. For more details. and sold volume. In order to ﬁnd the monopoly price and quantity.31) 107 .30) MC = MR ⇒ f ∗ = (3.25) RI − RO RI 2 − RO 2 The ﬂow surface given by equation 3. f . αd − αs αd β s + αd β d − αs β d ⇒pd∗ = . For a more details regarding these calculations. β s + 2β d (3.27) When looking at a system with a single pipeline the production. see for instance Tirole (1988): Demand: pd = αd − β d kg . MR = αd − 2β d f.and demand curves: αd β s + αs β d αd − αs ⇒ pd∗ = ps∗ = . (3.

The marginal expenditure curve is equal to the derivative of the cost function (C) for the monopsony: C = αs f + β s f 2 . ME = αs + 2β s f.34) 108 .32) (3. total consumer surplus is maximized through the monopsony solution. (3.Chapter 3 Modeling optimal economic dispatch and ﬂow externalities The social surplus is reduced since the traded quantity in this solution is different from the competitive markets’ solution. ﬁrst the marginal expenditure curve (ME ) must be found.33) The optimal solution is found at the intersection of this curve and the marginal revenue product (MRP ) curve (assumed equal to the demand curve): ME = MRP ⇒ f ∗ = 2αd β s + αs β d αd − αs ⇒pd∗ = . 2β s + β d β d + 2β s αs β d + αs β s + αd β s . In order to ﬁnd the monopsony solution. In this case the surplus of the producers’ is maximized. ps∗ = β d + 2β s (3. In the opposite case.

European Union (1998). Ehrhardt. K. ‘Competition in gas markets’. M.. D. ‘Access to pipelines in competitive gas markets’. H. (2002). ‘Natural gas and electricity optimal power ﬂow’. Proceedings of the IEEE/PES Transmission and Distribution Conference. Cremer. & Laﬀont. Cremer. Norman. (1996). E. & Steinbach. 606–607. ‘The gas transmission problem solved by an extension of the simplex algorithm’. Ehrhardt. European Commission. J. ‘Directive 98/30/EC of the european parliament and of the council’. Norwegian School of Economics and Business Administration. European Union (2003). J. M. Nonlinear optimization in gas networks. The Equipment Modules. T. Journal of Regulatory Economics 10(1). (1992). ‘Opening up to choice: Launching the single european gas market’. F. S. & Peck. & Laﬀont. M. Simulation and Optimization of Complex Processes’.New York. (2000). & Gedra. in H. S. 928–935. Berlin . & Steinbach. Springer-Verlag. K. L. (2004). Journal of Regulatory Economics 24(1). Dallas . (2003). & Transport (2002). 139–148. Qing. Y. Chao. 7 edn. (2000). E. ed. ‘A market mechanism for electric power transmission’. (2005). J. Campbell.. D. f.Bibliography An. pp. H.. & Smeers. Gas Conditioning and Processing. Bock. PhD thesis. Oﬃce for Oﬃcial Publications of the European Communities. ‘Modeling. Management Science 46(11). Bjørndal. 5–33. ‘KKT systems in operative planning for gas distribution networks’. Okla. Bergen. 1454–1465. De Wolf. H. Proceedings in applied mathematics and mechanics 4(1). 109 . ‘Directive 2003/55/EC of the european parliament and of the council’.-G. European Economic Review 46. Topics on Electricity Transmission Pricing. (2003).Heidelberg . Gasmi. 25–59.

PhD thesis.. Groningen. The Theory of Industrial Organization. Journal of Regulatory Economics 10(1). ‘A mixed complementarity-based equilibrium model of natural gas markets’. 171–200.. Operations Research 49(1). Mathematical Programming 105(23). F. Rømo. A. P. PhD thesis. pp. (2005).. Spot Pricing of Electricity. Norwell.. Y. M. & Smeers. 799–818. Kydes. (2004). Schweppe. (2001). M. Math in Gas and the Art of Linearization. Gabriel. (1988). A. S. Tomasgard.. Fodstad. M. & Whitman. R. S. Hellemo. Tomasgard. Spiller. Berlin Heidelberg. S. SINTEF.. Caramanis. Westphalen. L. Norway. Trondheim. P. J. M.. (2004).. University Duisburg-Essen (Germany). (2006). R. W. Anwendungen der Stochastischen Optimierung im Stromhandel und Gastransport. Kluwer Academic Publishers. 110 . Technical report. Van der Hoeven. Fodstad. MIT Press. (1996). S. Gabriel. lecture notes in economics and mathematical systems. J. Stochastic optimization of the natural gas value chain in a liberalized market. SINTEF. M. An optimization system using steady state model for optimal gas transportation on the norwegian continental shelf. 343–373. (2006). 14–25. in A. Seeger. 5–23. ‘The national energy modeling system: A large-scale energy-economic equilibrium model’. Tirole. Rømo. & Nowak. Tabors. F. Möller. Massachusetts. A. Hogan. F. The Energy Journal 14(3). Varaiya. The Netherlands. Trondheim. ‘Complemenatarity Problems in Restructured Natural Gas Markets’. Operations Research 53(5). & Bohn. Martin. Springer-Verlag... K. ‘Mixed integer models for the stationary case of gas network optimization’. ed. & Moritz. Wu.. & Zhuang. ‘Markets in real electric networks require reactive prices’. Norway. London Cambridge. Kiet. Recent advances in optimization. & Midthun. S. F. Technical report. (2004).Bibliography Gabriel. M.. A. P.. (2006). T. (1993). ‘Folk theorems on transmission access: Proofs and counterexamples’. S. International business school and research center for natural gas. & Oren. 563–582. (1988).

Nowak and Asgeir Tomasgard: An operational portfolio optimization model for a natural gas producer Submitted to international journal . Midthun.Paper III Kjetil T. Matthias P.

.

Chapter 4 An operational portfolio optimization model for a natural gas producer Abstract: We present a short-term portfolio optimization model for a large natural gas producer. The objective for the producer is to maximize the value of the produced natural gas within a week. production plans.1 Introduction We present an operational portfolio planning model for a large oﬀshore producer of natural gas.g. The producer strives to send more gas to the market on days with high prices and less on days with low prices. The paper discusses the value of actively using the storage capacity provided by the line-pack of the pipelines to maximize proﬁt for the producer. In this time-span the model includes spot market sales. The entities that we model in the natural gas networks are: production ﬁelds present in production nodes. The time horizon in the model is one week with daily resolution. The production targets are given as daily and weekly volumes that must meet upper and lower production requirements for each ﬁeld. In addition we look at storage in the pipeline system. The portfolio planning covers coordination of production in a set of ﬁelds. while the overall produced volume does not change within the week. storage management and fulﬁllment of long-term contracts. production limits). transportation in pipelines. We also study the value of using a stochastic model as compared to a deterministic model. the gas producer must consider both physical limitations as well as guidelines from tactical/ strategic models and ﬁeld lifting agreements (e. To our knowledge this is the ﬁrst study of a stochastic portfolio optimization 113 . The model is tested over two 60-days periods using real market data and realistic production and transportation capacities. and market nodes at which gas is sold through take-orpay contracts or at the spot market. pipelines between nodes. compressors used to lift the pressure in the network at given nodes. junction nodes where pipelines meet and gas is blended. When maximizing the value of production. management of storage and sales through contracts and spot markets. 4.

and Byers (2006) present a valuation methodology for commodity storage facilities. This is the world’s largest oﬀshore pipeline network for natural gas and produces about 15 % of the European consumption of natural gas. Any other storage capacity with close proximity to the market hubs like aquifers. model for natural gas production and sales. multi-commodity aspects of natural gas ﬂow and portfolio optimization. 2006) and also does not allow the modelling of storage in pipelines.and gas reservoirs. as compared to a deterministic approach. In our test case we use real market prices from 3 European hubs and realistic production and transportation capacities from the Norwegian continental shelf. (2003) an algorithm for the valuation and optimal operation of a single natural gas storage facility is presented. Our analysis here is on the value of portfolio optimization. In addition. Our model is a continuation of the work presented in Tomasgard et al. (2007) give a model for deterministic tactical value chain coordination including the multi-commodity aspects of natural gas. We examine the value of using a stochastic model. (1998) a multi-period mixed-integer linear programming model for scheduling investment and operation in oﬀshore oil ﬁeld facilities is presented.. The modelling of pipelines as storages and the interaction with compressors to achieve this is non-trivial and our approach goes beyond the methods reported in the current literature. LNG-storages and abandoned oil. In Thompson et al. 114 .Chapter 4 An operational portfolio optimization model. We simplify the analysis by assuming that one player can control both all production ﬁelds and the transportation network. (2007) a MINLP model for operational planning for upstream natural gas production systems is presented. That model is at a tactical level with a monthly resolution. no price model is given and storage in pipelines is not included.. (2007) that gives an introduction to modelling the value chain for natural gas covering both system eﬀects. caverns. Our analysis is motivated by the gas production on the Norwegian continental shelf. we evaluate the commercial value of actively using the storage (line-pack) present in the export pipelines in order to increase the average price of the sold gas. That approach does not include a physical model for the natural gas ﬂow and thereby neglects the important system eﬀects in the transportation networks (Midthun et al. can easily be included in the model as well. The model gives a detailed description of the physical properties of natural gas production and transportation. In related literature Nygreen et al. In Selot et al. no computational results are presented. In Iyer et al. The analysis is also relevant for other regions where contract markets and spot-markets exist together and where storage of gas may be used to increase the proﬁt. Ulstein et al. Stochastic models for planning oﬀshore gas ﬁeld developments under uncertainty in reserves is presented in Goel & Grossmann (2004) and in Haugen (1996). (1998) considers deterministic investment optimization for oil and gas ﬁelds in a strategic horizon with focus on how to develop ﬁelds and build pipelines.

We then present the results in Section 4.5 and concludes. (2006). For analysis of the organization of multi player behavior in the transportation market.2 we discuss important aspects of natural gas transportation. The majority of produced gas is still sold through long-term take-or-pay contracts (TOP) where the producer has little opportunity to increase proﬁtability. Then we discuss the representation of line-pack as storage in the portfolio optimization 115 . nominated volume. 4. Examples of computational equilibrium models for natural gas markets are Gabriel et al.2 Modeling natural gas networks coordination and storage. Here the time resolution is in hours. In Section 4. see Gabriel & Smeers (2005) and Midthun et al. An overview of linearization of natural gas ﬂow is given in Van der Hoeven (2004). (2005) and Zhuang & Gabriel (2006). (2000). An overview can be found in Kelling et al. day-ahead and within-day contracts. This means that the planning decisions have to be ﬂexible and able to meet varying demands. Our work on modelling storage in pipelines builds on and extends the approaches in Westphalen (2004) and Kelling et al. not on how the proﬁt should be shared between players or how the market should be organized. All papers mentioned so far are steady-state models.3. Consequently. In Westphalen (2004) and Nowak & Westphalen (2003) models for transient ﬂow in a system of gas transportation pipelines are presented. The structure of the take-or-pay contracts calls for careful planning. A mathematical formulation is given in Section 4. Work has also been done on modeling of the transient behavior of gas ﬂow. An alternative mixed integer approach is presented in Martin et al. In Section 4.4 we present the market models and more details of the computational cases. Spot markets therefore represent an important opportunity for increased proﬁt as an increase of the average spot price achieved will be reﬂected directly in the company’s proﬁt. volumes sold in the short-term markets are the only factor the producer may inﬂuence. (2007). A deterministic version of the model in Nowak & Westphalen (2003) is implemented in Nørstebø (2004).2 Modeling natural gas networks An important implication of the liberalization process of the European gas industry is the development of short-term markets in Europe. The buyer in the take-or-pay contracts has ﬂexibility with respect to the weekly. Now the producers of natural gas have the possibility to sell their gas through.4. as well as daily. and these deliveries will never be prioritized above the TOP-contracts. and move on to discuss the importance of transient behavior of the gas versus steady-state modelling. for instance. but also a challenge in terms of the security of supply for TOP contracts. weekahead. (2007). (2000). We here discuss how to model pressure constraints. We build on a linear steady-state model for the transportation network based on Tomasgard et al. in addition to long-term contracts.

The pressure in node C inﬂuences 116 . its diameter and other pipeline speciﬁc parameters. Modelling pressure and system eﬀects In a natural gas transportation network the capacity in a pipeline depends on the design parameters of the pipeline (such as length and diameter). For more details on the Weymouth-equation. 200 150 3 Flow [Sm ] 100 50 0 150 100 100 Pressure out [Bar] 50 80 140 120 200 180 160 Pressure in [Bar] Figure 4. pj ) approximates the ﬂow through a pipeline going from node i to node j as a function of the input pressure pi and output pressure pj : W Fij (pi .2. The ﬂow in the pipeline AC from A to C depends on the pressure level in node A and C.. pj ) = Kij pi 2 − pj 2 . see for instance Campbell (1992).. together with the representation of compressors used to control this storage. i ∈ I(j). This constant depends on the pipelines length. Fij (pi . (4.1) W Kij is the Weymouth constant for the pipeline going from i to j.1: The Weymouth equation. as well as on the pressure at the inlet and the pressure at the outlet.1). j ∈ N .Chapter 4 An operational portfolio optimization model. Consider the network in Figure 4. The Weymouth-equation is often used to describe ﬂow in a pipeline as a function of the inlet and outlet pressure as well as design parameters (see Figure 4.

In the graph to the left in Figure 4. Applying well known time-length criteria to long sub sea pipelines will result in a high number of segments (Osiadacz 1983. Streeter & Wylie 1970).2: Example of a transportation network consisting of two production nodes. in order to model time and storage capacity. Flow in one pipeline in the network inﬂuences the potential ﬂow in the rest of the network. If the pressure in node C increases. The line-pack can be increased when prices are expected to increase and decreased when prices are favorable. Often. the ﬂow from A to C decreases while the ﬂow from C to D increases. In an operational model with daily resolution and routing ﬂexibility we need to check whether a simple steady-state formulation will still be meaningful. the pipeline is divided into several segments with steady-state conditions in each segment.3 a possible use of line-pack is shown for a case where the pipeline is 90% full and a given price development. Transient natural gas ﬂow and pipelines as storage Steady-state models assume that inﬂow and outﬂow of gas is equal within a period. the system eﬀects are neglected. For a more detailed motivation for our steady state linearization and discussion of system eﬀects we refer to Midthun et al. If ﬁxed capacities are used for the pipelines instead of pressure constraints. 117 . (2006). However. long transportation pipelines may also be used as storages. or whether we need to use one of the more complicated models partitioning the pipelines into a set of steady-state segments.4.3). The line-pack in a pipeline is the gas contained in the pipeline. In the graph to the right we see the development when the pipeline is already ﬁlled to max capacity (Figure 4. the ﬂow both in AC and CD.2 Modeling natural gas networks A B C D E Figure 4. a transportation node and two market nodes.

this time going from the outlet toward the inlet and sucking gas out of the whole pipeline. This eﬀect is shown in Figures 4. the dotted line shows the level of line-pack and the full line shows the price development. where diﬀerential equations are used to describe the transient gas ﬂow in a pipeline. but in the simulation we show hourly eﬀect on pressure and velocity). causing both a slight increase of velocity and density also there. The intensity of the pressure wave is decreasing since the energy is used to push gas particles toward the outlet. An increase in the outﬂow will lead to a similar pressure wave. since the energy for the gas transport is stored in the compression. The ﬁgures illustrate that short term variations in input and outtake have little impact on both velocity and density in the middle of the pipeline or the other end of the pipeline..4 to Figure 4. Figure 4.6. We use Matlab to simulate the development of velocity and density of the gas molecules in a gas pipeline for given input and output proﬁles. The ﬁgure also illustrates that there is transient behavior in the 118 .3: Line-pack: The horizontal line shows the starting level of line-pack. A sudden increase of inﬂow will cause a pressure wave that goes through the pipeline at the speed of sound. We then look at a situation where we use a time resolution of 24-hours (the ﬂow rate is constant within each 24-hour period. The simulation is performed in the same manner as described in Nowak & Westphalen (2003). When the amount of gas taken out of the pipeline is less than the amount of gas sent into it (in practice additional gas is being pushed in) the pressure and density at the inlet is rising..5. At each point in time the pressure decreases along the pipeline. The ﬁgure shows that the inlet and outlet pressure in the pipeline is an approximately piecewise linear function. The chosen input/output-pattern.Chapter 4 An operational portfolio optimization model. while the velocity increases. as well as the resulting inlet and outlet pressure is illustrated in Figure 4.

4 9 8 7 6 5 4 3 2 2 1 0 0 1 3 4 5 6 7 Time [100000 s] Distance [100000 m] Figure 4.6 1.5: Simulation of density (ρ) for a small variation of ﬁrst input and then output.8 1.4: Simulation of velocity for a small variation in input and output.2 Modeling natural gas networks Velocity [m/s] 2 1.4. ρ Rho [kg/m ] 172 170 168 166 164 162 160 158 9 8 7 6 5 4 3 2 5 4 3 1 0 2 1 7 6 3 Time [100000 s] Distance [100000 m] Figure 4. 119 .

2000): 2 3 pi pj pi + pj pavg = ij pi + pj − .6: Resulting inlet and outlet pressure from given input/output data.and outﬂow quantities in the pipelines. We extend this formulation with a storage model that depends on a weighted average of the pressures at the inlet and outlet points of the pipeline. We argue however. The modiﬁed steady-state approach is based on steady-state equations. where pi is the pressure at the inlet and pj is the pressure at the outlet).2) This expression can be linearized in the following way (by using a ﬁrst-order Taylor expansion): 120 . We have used a formula that in addition considers the shape of the pressure drop in the pipeline (Kelling et al. In Westphalen (2004) the line-pack is calculated as a function of average pressure in the pipeline (pavg = ij 1 2 (pi + pj ).. this non-linear behavior will be important. that with a time-resolution of 24-hours. With a time-resolution of less than an hour. (4. a modiﬁed steady-state approach is suﬃcient. 250 80 200 pressure [bar] 60 ﬂow [Sm ] ﬂow [Sm3 ] 150 100 inlet pressure outlet pressure input output 25 50 75 hours 100 125 150 40 50 20 0 0 Figure 4. This can be seen from the non-linear behavior of the inlet and outlet pressure in the hours after a changed ﬂow pattern. pipeline in the ﬁrst hours after a change in input and/or output in the pipeline.Chapter 4 An operational portfolio optimization model.. but allows for diﬀerent in.

(4.3) where Piavg and Pjavg are constants.3. z is still varying in pressure. see Modisette (2000). For a discussion on diﬀerent versions of the Equation of State.4. The compressor work. The cost of running a compressor depends on the inlet and outlet pressure.6) where W is the work of the compressor. For a constant temperature T . Equations (4. (4. In addition. T ) . When using this Equation of State. The modeling starts when gas enters the compressors that build up pressure for the pipeline transportation. but for our relevant pressure levels the variations are small. LP ij = Piavg Pjavg Pjavg (pi − Piavg ) Piavg pj − Pjavg − − Piavg + Pjavg Piavg + Pjavg Piavg + Pjavg . The value of z is chosen to be 0. representing the approximated average inlet and outlet pressure in the pipeline.7 in this paper based on experience from the pipelines at the Norwegian continental shelf (Dahl 2001). The complete formulation is given in Section 4.2 Modeling natural gas networks pavg = ij 2 3 pi + pj − (4. We can then estimate the line-pack by multiplying the density with the volume of the pipeline: p=ρ m avg p Aij Lij . Compressor modeling For production ﬁelds we do not consider the physics in the reservoirs. P in is the pressure of the gas entering the compressor and P out is the gas leaving the compressor (corresponds to the 121 .4) m where ρ is the density in the pipeline. and the gas ﬂow.5) RT z ij where LP ij is the line-pack in the pipeline between i and j. R is the Gas constant. we can ﬁnd the density in the pipeline as a function of the average pressure in the pipeline. T ) is the compressibility factor as a function of pressure p and temperature T . (4. m is the molecular weight and z (p. We use the following form: R T z (p.16)-(4. as an isentropic process is given by (Katz & Lee 1990): W = Aq P out P in χ χ−1 −1 .18). Aij is the area of a cross-section of the pipeline and Lij is the length of the pipeline. we need an Equation of State in order to relate the average pressure in the pipeline to the average density in the pipeline.

Chapter 4 An operational portfolio optimization model... inlet pressure at the connected pipeline). Here χ is the adiabatic constant and A represents the aggregation of various constants in the equation. In the following, the compressor work will be replaced with the cost of running the compressor assuming a constant energy price. The function is neither strictly convex nor concave. This makes it diﬃcult to represent the function in an optimization model. Including the original compressor cost function leads to a non-convex problem. One approach is presented in Nowak (2006). This method works well for a deterministic model (Nørstebø 2004), and gives a very good approximation of the actual compressor costs. However, the method is not eﬃcient enough to be used for the model sizes that we are considering. The solution times for stochastic network models with 3-5 stages, and from 500 to 1000 scenarios will be too high for practical use. Therefore, to simplify Equation (4.6) we assume constant input pressure to the compressor. The compressor cost is linearized with the following constraints: cgts ≥ a1 + b1 dgts − dg + c1 pgts − pg , g ∈ G, t ∈ T , s ∈ S, cgts ≥ a2 − b2 d − dgts − c2 pg − pgts , g ∈ G, t ∈ T , s ∈ S, cgts ≥ a3 + b3 dgts − dg − c3 pg − pgts , g ∈ G, t ∈ T , s ∈ S, cgts ≥ a4 − b4 dg − dgts + c4 pgts − pg , g ∈ G, t ∈ T , s ∈ S, (4.7) (4.8) (4.9) (4.10)

Here dgts and pgts is the production and outlet pressure in ﬁeld g at time t in scenario s and the overlined and underlined values give the maximum and minimum values of the variables, respectively. Then cg is the approximated compressor cost at production ﬁeld g. Constants a, b, c are positive and describe the plane used for approximation.

**4.3 The portfolio optimization model
**

We model the uncertainty in prices and volumes using scenario trees in a 3-stage stochastic program (see e.g. Kall & Wallace (1994)). It has seven time periods and daily resolution. We start here with the description of the scenario trees and the time structure of the model and then give the mathematical formulation.

**Scenario structure and rolling horizon
**

Assume that our model starts on a Monday. The model horizon then comprises the weekdays from Monday to Sunday. We implement the ﬁrst stage decisions (the decisions for Monday) and run the model again on Tuesday (with the weekdays from Tuesday to next Monday). The parameters regarding production levels

122

4.3 The portfolio optimization model for each ﬁeld, remaining production limits for the week and line-pack levels are updated between the runs. Also the scenario tree structure depends on which weekday the model is run, see Figure 4.7. The reason for this is the special structure of the weekend where all markets for Friday, Saturday and Sunday are cleared on Friday.

Figure 4.7: The diﬀerent scenario trees

Notation

Sets N G B M I(n) O(n) T S Z Indexes n g b m t s z Index used for nodes in general. When more indexes are needed, i and j will be used. Index for production nodes. Index for junction nodes. Index for market nodes. Time period index. Scenario index. Index for linearized Weymouth constraints. The The The The The The The The The set set set set set set set set set of of of of of of of of of all nodes in the network. production nodes in the network. junction nodes in the network. market nodes in the network. originating nodes with pipelines going into node n. end nodes for pipelines going out of node n. time periods. scenarios. constraints used to linearize the Weymouth equation.

123

Chapter 4 An operational portfolio optimization model... Constants Gg Gg Hg Hg Pn Pn Kij PI z PO z A χ Rm Bg Piavg Pjavg The maximum daily production level in ﬁeld g. The minimum daily production level. The maximum weekly production level. The minimum weekly production level. The maximum pressure in node n. The minimum pressure in node n. The Weymouth constant for the pipeline going from i to j. Fixed point for pressure into a pipeline. Fixed point for pressure out of a pipeline. Constant in the compressor work function. Adiabatic constant. Price in the take or pay contract in market m. Constant to convert compressor work to compressor cost. Approximated average pressure into the pipelines in node i. Approximated average pressure in the pipelines entering node j.

Decision variables dg qm vm fij pn pavg ij LPij cg in qij out qij Production in ﬁeld g. Spot sale in market m. Delivery in take or pay contract in market m. The ﬂow between node i and j. The pressure in node n. The average pressure in the pipeline going from node i to j. Line-pack in the pipeline going from i to j. Cost of compressor in ﬁeld g. Volume inserted to the pipeline going from i to j. Volume extracted from the pipeline going from i to j.

Stochastic variables and probabilities vm rm π Nomination in long-term contracts in market m. Spot price in market m. Probability of a given scenario.

Functions F (pi , pj ) W (q, Pin , Pout ) Flow in a pipeline with inlet pressure pi and outlet pressure pj . Work of a compressor.

124

g ∈ G. The total production should be within a given interval. The second term gives the revenues from the long-term contracts: vmts is the volume delivered in long-term contracts in market m while Rm is the price obtained.3 The portfolio optimization model The mathematical model The model maximizes expected income from sales in the spot market and deliveries in long-term contracts minus the costs of using the network.12) where Gg is the lower production limit. t ∈ T .2. The Weymouth-equation is used to relate the pressure diﬀerence and design parameters of the pipelines to the ﬂow in the pipeline. (4. This gives the following constraints for the model without line-pack: 125 .10). The limits are given both on a daily and a weekly level. s ∈ S. in ﬁeld g. The compressor cost is modeled with constraints (4. The costs incorporate the compressor costs and production costs: Π= s∈S. At each ﬁeld there is a minimum and a maximum volume that can be produced by the company. The weekly limits are formulated as: Hg ≤ t∈T dgts ≤ Hg . we use the same linearization as Tomasgard et al.11) Here πts is the probability of scenario s in time t. we model the transportation system as a steadystate system. and cgts is the cost of the compressor in ﬁeld g.7) to (4. These limitations are due to both the properties of the ﬁeld and the guidelines from a tactical plan. The parameter κg is the cost per unit of production dgts . The daily limits are given by: Gg ≤ dgts ≤ Gg . The last term gives the cost of the compressors and the production cost. (2007).4. s ∈ S. As discussed in Section 4. (4.t∈T (rmts qmts + Rm vmts ) − g∈G πts m∈M (cgts + κg dgts ) . In order to get a linear model. g ∈ G. G is the upper limit and dgts is the production in the ﬁeld. (2007) and Rømo et al. qmts is the volume of gas sold in market m.13) where Hg is the lower limit on weekly production and Hg is the upper limit on weekly production. while rmts is the price obtained for this gas in market m. (4.

14) 2 2 P Iiz − P Ojz where (P Iz .. The line-pack inventory is given by: LP ijts = in out LP ij. i ∈ I(j). z = 1.16) RT z ijts where LP ij is the line-pack in the pipeline between i and j. Aij is the area of a cross-section of the pipeline and Lij is the length of the pipeline. representing input and output. s ∈ S.2 the line-pack in a pipeline is found by multiplying the approximated density with the volume of the pipeline. the outlet pressure is pjts and the W resulting ﬂow is fijts . Z. z = 1.. Kij is the Weymouth constant for the pipeline going from i to j.18) 126 . The z inlet pressure in the pipeline is given by pits . t ∈ T \{T }.T.. (4. i ∈ N . m avg p Aij Lij . . hence we substitute the variable fij with two new variables: in out qij and qij . s ∈ S. j ∈ O(i). .. For the model with line-pack it is the net daily change of the volume in the pipeline that matters. t ∈ T . each with a diﬀerent ratio P Oz . Z. j ∈ O(i)...15) As discussed in Section 4. This is in order to not deplete the line-pack which is also used as a buﬀer in order to secure supply: in out LP ij. i ∈ I(j).. The Weymouth-equation can then be written as: 1 in out q + qijts 2 ijts W ≤ Kij W − Kij P Iz 2 2 P Iiz − P Ojz P Oz 2 2 P Iiz − P Ojz pits pjts . j ∈ N . P Oz ) are break-points for the linearization representing pressure at both ends of the pipeline. (4.Chapter 4 An operational portfolio optimization model.t+1. s ∈ S.17) As a business constraint we require that the level of line-pack at the end of horizon is equal to the starting level. (4.s = LP ijts + qijts − qijts . see Tomasgard et al. i ∈ N . fijts ≤ − W Kij W Kij P Iz 2 P Iiz 2 − P Ojz P Oz pits pjts .. (4. (4. (2007).s + qijT s − qijT s = LP ij1s . We use Z such constraints to linearize the Weymouth PI equation. j ∈ O(i). j ∈ N . i ∈ N .

s ∈ S.21) In the junction nodes. of gas is ﬂowing into the connected pipelines. (4. The formulation without line-pack is: dgts = i∈O(g) fgits . g ∈ G. The formulation without line-pack is: fikts = i∈I(k) j∈O(k) fkjts . k ∈ B. t ∈ T .3 The portfolio optimization model where T is the last time period. If the constraint had not been included. k ∈ B. Constraint (4. s ∈ S. the amount of gas that enters the node must be equal to the amount leaving the node.19) In the production nodes we must make sure that the produced quantity. t ∈ T . s ∈ S. g ∈ G. dgts .20) and with line-pack it is: dgts = i∈O(g) in qgits . i ∈ N . t ∈ T . t ∈ T .18) is necessary to take care of end-of-horizon eﬀects. fgits . (2007)). Alternative ways of formulating this constraint is to include a value of the gas in period T (for instance based on an Expected Gas Value Function. (4. These requirements come from compressor capacities. (4.24) and with line-pack it is: 127 . s ∈ S. The pressure in the nodes needs to be within maximum and minimum requirements. (4. In the market nodes the company sell qmts in the spot market. s ∈ S. The mass balance equation for the market nodes for the formulation without line-pack is: qmts = i∈I(m) fimts − vmts . t ∈ T . (4. Additionally. t ∈ T . as presented in Tomasgard et al.23) B is the set of junction nodes in the network.4. the line-pack level would have been reduced to a minimum in period T . m ∈ M. design parameters of the network and contractual agreements: Pi ≤ pits ≤ Pi . the company must deliver long-term contracted volumes (vmts ). s ∈ S. (4.22) and with line-pack it is: out qikts = i∈I(k) j∈O(k) in qkjts .

. The demand in the take-or-pay volumes are chosen to reﬂect the dominant position of these contracts in the North-Sea. The transportation capacity of the pipelines is comparable to the dry-gas transportation system in the North-Sea. m ∈ M. (4. pavg . t ∈ T (z).22). cgts .24). NBP and TTF).23). pavg .4 Test case from the Norwegian continental shelf The network chosen as a test case is shown in Figure 4. ijts The two models then consist of: 1.. In the network there are three ﬁelds.t. Then we have the following constraints: 1 |S(z)| in (dgts . 128 . qmts = i∈I(m) out qimts − vmts . (4. LPijts . (4.20).11) s. qijts ) = ijts s ∈S(z) (4.19)-(4.15)-(4. Model without line-pack: max (4. but the structure is simpler.11) s. (4. Dornum and Dunkerque hubs. Let the scenarios passing through node z be given by S(z) and let T (z) be the time period of node z.26). qmts . The chosen market hubs are large hubs in Europe (Zeebrugge. (4. (4. (4. (4. In each event node z where uncertainty is resolved in the scenario tree. The demand is assumed to be completely inelastic. LPijts .17). cgts . This market hub is included because of the liquidity of the market and the availability of price data as a substitute for less liquid spot markets in the Emden. (4.26) 2. fijts . z ∈ Z.13). (4. (4. pits . qmts . two junction points and three market nodes. (4.25)-(4.19).25) Since it is assumed that the company cannot inﬂuence the prices in the market nodes.Chapter 4 An operational portfolio optimization model. or alternatively the volumes that the company is buying/ selling are not large enough to inﬂuence the market price. 4.7)-(4. pits .21). In reality the Title Transfer Facility (TTF) hub is however not connected to the transportation system from the North-Sea. qijts ). there are no restrictions on how much the company can buy or sell in the market. s ∈ S(z).16. Model with line-pack: max (4.26) in (dgts .14).7)-(4. fijts . s ∈ S. t ∈ T . we need to add non-anticipativity constraints (Rockafellar & Wets 1991).t.

there are large upwards peaks in the time series. and the gain from good planning may be substantial. The price tends to be higher during winter than during summer. which we deﬁne to be from October 1st to March 31st. Being able to deliver extra quantities of gas when such a peak occurs is advantageous.4. During the winter period. NOK/Sm 3 Figure 4. and illustrates that the price volatility within a week can be large. the prices are high and the ﬂexibility in the network is low. During the summer season. In addition to the seasonal variations. the nominated quantities of the long-term contracts are lower and thus the ﬂexibility is larger. Before constructing the prediction models for the time-series. The price series are illustrated in Figure 4.8: Prices at the three hubs in the winter period 2005.4 Test case from the Norwegian continental shelf The spot markets The price scenarios are generated based on time-series of historical observations.9 shows the weekly variation over 7 days. Forecasting and scenario generation In each node in the scenario-tree. there is a spot price and a take-or-pay volume for each of the market hubs (the method to construct scenarios for the take-or- 129 .8. the price peaks (deﬁned as deviations of more than 50% of the average value) are considered as outliers and are therefore removed and replaced with the average value of the price on the same weekday in the previous and coming week. In this article we focus on the winter season. Figure 4. The price data was supplied by the Heren Energy Ltd. The winter period thus requires careful planning.

We use AR(2)-models to predict the spot price in all market nodes (see Figure 4.10: The AR(2) prediction model. We estimate the ﬁrst four moments (expectation.. branch is zero in all nodes in the scenario tree. skewness and kurtosis) and correlation of the prediction error distributions for the markets based on the historical data. The indexes f1 and f2 give the number of branches in the ﬁrst and second stage. Such price models will in general track the real data process by following it closely but in general lag a bit behind downward and upward movements. Figure 4. Figure 4.9: Illustration of the weekly variations in the last week of February 2005. We then use a moment matching procedure that was developed in Høyland et al.. The value of stage. respectively. This is important for the scenario generation method we have chosen.12). and then the scenarios are there to describe the historical deviations from the forecast. An example of how one of the AR(2)-models ﬁts the real data is shown in Figure 4. except for the nodes in the t ﬁrst period in a new stage (corresponding to period t+1 and t+6 in Figure 4. variance. (2003) to generate our scenarios. The AR(2) model is parameterized so that the expected error is 0.12 shows an example of a scenario tree for the prediction error. pay volumes is given in the next section). The models seem to represent the time-series development reasonably good and are unbiased. historical data xt−2 xt−1 xt predicted data xt+1 xt+2 xt+3 xt+4 xt+5 xt+6 ˆ ˆ ˆ ˆ ˆ ˆ xt+j = ˆ α + i=1 βi xt+j−i α + β1 xt+j−1 + β2 xt+j−2 ˆ 2 α + i=1 βi xt+j−i ˆ 2 if j = 1 if j = 2 if j > 2 Figure 4.11.10).Chapter 4 An operational portfolio optimization model. 130 .

S−f2 ) t+6 (2.S) t+6 (1.10).4. xt+j can depend on both ˆ historical data and predicted data (ˆt+2 for instance depends on both xt and x xt+1 ) (see Figure 4. 11 t+1 21 t+6 (2.11: Real prices and values obtained from the prediction model (here illustrated for the NBP). 131 . The predicted values. We generate S multivariate scenarios for the prediction error with the correct correlation between the markets and with correct moments for the individual error terms. Each scenario presents a path from the root node to the leaf node (there are S unique paths through the tree).13. Hence we use the forecasting method to predict the expected price.f1 ) t+1 Figure 4.f2 ) t+6 (2. Finally. The value in each node in a path through the scenario ˆ tree can then easily be found by the formula shown in the ﬁgure.4 Test case from the Norwegian continental shelf Figure 4.12: The scenario tree for the prediction errors. and scenario generation to describe the variation (error) around this price. This is illustrated in Figure 4. we combine the AR(2) prediction model with the scenario tree for the prediction errors to one scenario tree.

13 the nonanticipativity constraints are represented by the ellipsoids. The Group 2 customers will nominate the maximum amount given that the spot price exceeds the TOP-price (plus the transaction costs). The ﬂuctuations in nomination is however less extreme than for Group 2. The nonanticipativity constraints make sure that decisions at time t can only depend on information available at this time.26). see for example Ruszczyński (1997). 132 . An important issue in stochastic models is the information structure.Chapter 4 An operational portfolio optimization model. Group 2 contains purchasers that have more ﬂexibility to utilize the TOP-contract as a call option.14. We have reﬂected this in the model by choosing the expected demand in the take-or-pay contracts to be a large percentage of the total production capacity in the ﬁelds. Take-or-pay volumes For the take-or-pay contracts. A large percentage of the gas in the North-Sea is still sold through long-term contracts. We have implemented the nonanticipativity constraints in Equation (4. The ellipsoids represent the nonanticipativity equations.. This means that they will nominate a high volume given that the spot-price is high. For Group 1 we assume that the nominated volume is a linear function of the spot price within a certain range. we assume that the company’s customers can be divided into two groups: Group 1 comprises gas purchasers with large delivery commitments in Europe. In Figure 4.13: All paths through the scenario tree. This is illustrated in ﬁgure 4. and nominate a low volume in case of a low spot-price.. and the minimal amount when the spot price is lower than the TOP-price. xs ˆt+j = s α + i=1 βi xs t+j−i + t+j s s α + β1 xt+j−1 + β2 xt+j−2 + ˆ 2 α + i=1 βi xs ˆt+j−i + s t+j 2 s t+j if j = 1 if j = 2 if j > 2 t t+1 t+2 t+3 t+4 t+5 t+6 Figure 4.

The total export of natural gas from Norway to Europe in 2005 were 82. The transportation network consists of long. we have included approximately 60% of this volume in a simpliﬁed network structure. In our case study the TOP-volume is in average 60% of the total production. as much as 90% of the gas is sold through longterm contracts. When running the model on a rolling horizon and implementing the ﬁrst stage decisions (decisions for the ﬁrst day). and that the trade in short-term markets will increase. The supply side Most gas ﬁelds in the North-Sea produce both oil and gas. One option would be to use the price of gas at the last day in the test period for which the model is run. all gas volumes are given in million standard cubic meters and the proﬁts are given in million Norwegian kroner (NOK).5 billion Sm 3 . we use the average price in the test period. Changes of one day in the length of the test period may have substantial eﬀect on the overall return both in positive and negative direction in this case. There are however eight pure gas ﬁelds. subsea pipelines that are operated at high pressure. the model with line-pack may end with a line-pack inventory that is lower or higher than the starting level. An overview of the production capacities and pipelines characteristics in the North-Sea can be found in OED (2006). Currently. In our model. Instead. 133 .14: The two diﬀerent customer types. one has to assign a value to this line-pack diﬀerence. In order to compare the models with and without line-pack. It is however expected that this amount will decrease. We use two measures when comparing the results from the models: proﬁt from the spot market (after the inventory value adjustment) and the average obtained price in the spot market. In the following.4 Test case from the Norwegian continental shelf Figure 4. but since the prices are volatile this is an unstable method.4.

These penalty costs are necessary for the deterministic model.15 shows the two diﬀerent time periods considered in this paper: October 3rd 2005 . the penalty costs are not included. shorter ﬂuctuations are added to some scenarios (doubling or halving of the price for one or more time periods). T − 2 T − 2.15: The time periods considered in this article. T − 1 T − 1. 134 . otherwise commitments from previous runs on the rolling horizon can lead to infeasible problems. We have used two diﬀerent approaches for parameterizations of the price models. we introduce a penalty cost for not meeting the take-or-pay requirements on a given day. 4.Chapter 4 An operational portfolio optimization model. with a rolling horizon over 60 planning days (72 calendar days).5 Computational results and discussion In order to ensure feasibility for all daily runs on the rolling-horizon. Figure 4. In PM2 we use as recent data as possible and the same data for both the parameterization and the error estimation. Price Model 1 (PM1) T T +1 T −3 T − 3.1 for an overview.1: Time periods used to estimate AR(2)-parameters and the prediction errors for our two price models. as well as for not being able to meet the line-pack-requirements. The stochastic versions will most often have taken the possibility of a price increase into consideration. T Price Model 2 (PM2) T T +1 T −1 T T −1 T Rolling horizon AR(2)-parameters Prediction errors Table 4. In addition. When comparing the results from the models.. In order to capture the eﬀect of price peaks.March 25th 2006 (period T + 1). we have in some of the cases introduced extreme scenarios: one scenario where the price is doubled and one scenario where the price is halved for the entire week.. See Table 4.December 24th 2005 (period T ) and January 2nd 2006 . We have also tested the eﬀect of updating the prediction errors on a monthly basis. Test instances We have tested the model on real price data. In PM1 we use two diﬀerent time periods for the parameterization of the AR-model and for ﬁnding the error distribution. Figure 4.

The solution with this input data gives insight to the value of line-pack. The results are presented in Table 4. the resulting line-pack utilization is compared with a model where the starting level is at 100%.16: The network considered in this paper.72%.17 we have looked at a situation where we initiate the model with a line-pack level of 90%. T . The results from optimizing with the two diﬀerent starting levels are shown in Table 4. We then look at the importance of the starting level of the line-pack in the system. This solution also gives a benchmark to compare the results of our model run with stochastic prices.3. the diﬀerence is 1. the diﬀerence is 13. For autumn 2005.5 Computational results and discussion Figure 4. both models show the same behavior: the line-pack increase in periods with low price and decrease in periods with high prices. The price data for spring 2006 are much more volatile than for autumn 2005. For spring 2006 however. the added volatility in the spot prices increases the value of the inherent ﬂexibility in the line-pack. As expected. We get an improved income from the 135 .4.97%. As we can see from these results. In Figure 4. Perfect information We start the analysis by looking at a situation where the producer knows the exact values for the price and TOP-volumes for the next seven days. In our test-case. As expected. the value of line-pack depends on which period we look at. The reason is that this model has higher ﬂexibility with respect to storage utilization (the target level for line-pack at the end of the week is at least 90% which is a weaker requirement).2 (the models with line-pack are started with the pipelines ﬁlled to the capacity limit). In the ﬁgure. the model with a starting line-pack level of 90% of total capacity has higher proﬁts. as well as how the solution of the model depends on the starting level of the line-pack.

927 3..344 15994.17: Utilization of line-pack in the network for the model with perfect information (spring 2006).71 2.2: Spot market income for the optimization model run with perfect information. 136 .74 2.388 2.833 14205.61 Case 2 T Without line-pack 14205.927 3.94 2. Figure 4.71 Case 4 T +1 Without line-pack 15994..74 Table 4.26 2.388 2.40 Case 3 T +1 With line-pack 18806.777 18228. Time period Model type Spot market income Average price Average obtained price Adjusted with average price Time period Model type Spot market income Average price Average obtained price Adjusted with average price Case 1 T With line-pack 14573.Chapter 4 An operational portfolio optimization model.927 14450.

64%.777 18228.927 3. The distance to the benchmark for this result was 4.68 2.41% for the model without line-pack.26 2.76%. The diﬀerence is here 4. the best combination turned out to be using both outliers and monthly update of the prediction error.62 Table 4.34% larger than the deterministic model. Time period Line-pack start level Spot market income Average price Average obtained price Adjusted with average price Case 5 T +1 100% 18806.94% and the value was 7.05% larger than the result for the deterministic model.5 Computational results and discussion spot market of 4. The results are given in 137 .12% and 5.61 Case 6 T +1 90% 18816.0%. we present aggregated proﬁt from both time periods. Generally.978 18968.4. We use the perfect information case as a benchmark.4. We then added extreme scenarios to the stochastic model. The stochastic version uses 900 scenarios to describe price uncertainty in the 3 markets for the next 7 days. as well as the option to update the distribution of prediction errors on a monthly basis.4. For each of the instances. When the start. the distance to the benchmark was 7.3: Results when running the model with diﬀerent starting line-pack levels. while the diﬀerence is 0. a high inventory requirement is not necessarily beneﬁcial. The largest diﬀerence between the stochastic and deterministic version is found for the model including line-pack.927 3.and target-level are the same. and the result was 2. The best result for the model with line-pack was achieved when only adding outliers. For the model without line-pack. The results for PM1 are shown in Table 4. The diﬀerence up to the benchmarks for these models are however large (11. we estimated the AR(2)-models and prediction errors on the winter period most recent to the rolling horizon using PM2.72% for the stochastic models). We tested the diﬀerent combinations of parameter estimations and error estimations described in Section 4. Further. we can conclude that on a ﬁxed horizon it is beneﬁcial to start with a high level of line-pack and have a low target level at the end-of-horizon. Stochastic versus deterministic models In the following we compare the results of the stochastic model based on forecasting and scenario generation with the deterministic model based on forecasting alone.

21%.36% and the distance to the deterministic solution was 0. If we compare the best models with line-pack (PM2.934 Price model Model type Uncertainty Spot market income Average obtained price Average price Adjusted with average price Price model Model type Uncertainty Spot market income Average obtained price Table 4. or 1607 million NOK.946 Case 8 PM1 With line-pack Deterministic 30199.07% and the distance to the deterministic solution is 0. The diﬀerence between the models with and without line-pack is quite clear in all our results.6%. The distance to the benchmark is now. We see that the distance to the benchmark has decreased for the model with line-pack. 7.23 Case 9 PM1 Without line-pack Stochastic 28566. we introduce outliers as well as monthly updates of the prediction error. As we can see from these tables.87 2.21%.06 3.5%.658 28104. Table 4.84 2. The diﬀerence between the stochastic models and the deterministic version is less clear. the best solution was also achieved when the prediction error was updated monthly.882 2.5. The results are aggregated for T and T + 1. the diﬀerence is now 0.56%.. The change is however quite small . For the model with line-pack. monthly updated error distribution) with the best model without line-pack (PM1. but the diﬀerence has decreased. the best result is obtained when the prediction error is updated monthly.4: Comparison of the deterministic and stochastic model with and without line-pack. For the model with line-pack. For the model without line-pack..57 Case 10 PM1 Without line-pack Deterministic 28450. Still. while for the model without line-pack the diﬀerence is almost zero (0. The distance to the benchmark solution was now 5.022 2.013%).52%. respectively.658 29408. For the benchmark models. extreme scenarios and monthly updated error distributions) we get an advantage of 5.1% and 5. Again.the distance to the benchmark solution is 7. the stochastic models still give better results than the deterministic models.Chapter 4 An operational portfolio optimization model. the same comparison gave a diﬀerence of 8.24 2. the stochastic version is consistently outperforming the deter- 138 . but actually increased for the model without line-pack. Case 7 PM1 With line-pack Stochastic 30317.

ministic version. 4.6 Conclusions Case 11 PM2 With line-pack Stochastic 30934. The largest diﬀerence is found for Price Model 1 in spring 2006.658 30199. This result indicates that the price models are not accurate enough to utilize the added ﬂexibility from the reduced target storage.5: Comparison of the deterministic and stochastic model with and without line-pack. the results were clearly inferior to the models started with 100% line-pack level.658 30355. For some combinations of parameter and error estimation we gain a lot from using the stochastic model. We have used real market prices. The results show that modelling storage in the model can increase the prof- 139 .84 Case 14 PM2 Without line-pack Deterministic 28521. and tested the model on two time periods: fall 2005 and spring 2006. the stochastic solution is 12.25 3. The results are aggregated for T and T + 1.45 2. We compare the results from the model both with and without storage in the pipelines and with and without stochasticity (prices and TOP-quantity).01 2.942 Case 12 PM2 With line-pack Deterministic 30795. This did not happen for any of the stochastic models.108 2.4.126 2.84 Case 13 PM2 Without line-pack Stochastic 28525.6 Conclusions We present a stochastic portfolio optimization model for operational planning in natural gas value chains. indicating that the line-pack end-of-horizon constraints were violated. The stochastic model seems to be more robust than the deterministic model when it comes to choice of parameters and time period for error estimation. Also we observed that the deterministic model has positive values for the penalty functions in some cases. When we initiated the system with 90% line-pack level.941 Price model Model type Uncertainty Spot market income Average obtained price Average price Adjusted with average price Price model Model type Uncertainty Spot market income Average obtained price Table 4.71 3.84% larger than the deterministic solution in this case.

Thirdly. the deterministic model was infeasible for some days on the rollinghorizon (not able to deliver in TOP-contracts and/or not able to meet the target level for line-pack) for some of the tests we did. we can relax the system perspective on the value-chain. the added proﬁts were as high as 13. itability of the system substantially when operated optimally. it would be interesting to compare the results from this model with a transient ﬂow model. and see how the inclusion of one or more producers would inﬂuence the results. This indicates a large commercial value of actively using the storage inherit in the pipelines (line-pack) to maximize proﬁts. also shorter time periods (such as hours) could be considered.97% for the 60-day period in spring 2006. market power in some.. in some instances. 140 . We have also seen that taking into account the stochasticity in the problem can. We also found that the results from the deterministic model were more sensitive to the value of the parameters in the prediction model than the results from the stochastic model. With a transient ﬂow model. There are many possible extensions of this work..Chapter 4 An operational portfolio optimization model. lead to large gains in the objective function value. With perfect information. Secondly. In addition. Firstly. hubs could be introduced (with uncertainty in the price elasticity in the markets). or all.

W. Stochastic Programming. & Kiet. Eng. H. Norway. 88–100. E. ‘Commodity storage valuation: A linear optimization based on traded instruments’. A. New York. CORE Discussion Paper No. J. USA. R. 2005/37. (1992). (1996). 37. ‘A heuristic for momentmatching scenario generation’. Kall. W. K. Campbell. Kaut. Operations in a Liberalized European Gas Market. S. Computational Optimization and Applications 24(2-3). (1998). 1409–1429. S. ‘A stochastic dynamic programming model for scheduling of oﬀshore petroleum ﬁelds with resource uncertainty’. Y. McGraw-Hill. 1380 –1397.. John M. Norman. ‘A stochastic programming approach to planning of oﬀshore gas ﬁeld developments under uncertainty in reserves’. Energy Economics 28. & Lee. Ind. J. European Journal of Operational Research 88. 141 . (2005). & Smeers. D. Energy Economics 27. M. J. & Wallace. Res. R. 275–287. Katz. S.. NTNU. R. Iyer. S. K. Natural Gas Engineering. P. E. Vasantharajan. S. Oklahoma. Zhuang.. Høyland. A. ‘Optimal planning and scheduling of oﬀshore oil ﬁeld infrastructure investment and operations’. Goel. PhD thesis. & Wallace. (2005). (2001). ‘A large-scale complementarity model of the North American natural gas market’. (1990). S. & Cullick. I. Gabriel.Bibliography Byers. (2006). K. V. Haugen. Gabriel. Chem. W. Trondheim. J. Chichester. (1994). (2003). Dahl. I. (2004). Campbell & Company. Norwegian Natural Gas Transportation Systems. Grossmann. Gas Conditioning and Processing: The Equipment Modules. & Grossmann. USA. Computers & Chemical Engineering 28. 169–185. ‘Complementarity problems in restructured natural gas markets’. 639– 665. S. John Wiley & Sons..

Nørstebø. ‘Optimal numerical method for simulating dynamic ﬂow of gas in pipelines’. J. V. Working paper. Norway. & Westphalen. R. K. Nowak. T. Working paper. Bjørndal. Rockafellar. Technical report. Reith.. Osiadacz. Working paper.-B. Technical report. Midthun. NTNU. S. 142 . Hellemo.. J. Modisette. K. 251–267. Trondheim. (1991). in ‘Proceedings of ECOS’. A.. Trondheim.. Rømo. OED (2006). F. (2000). Compressor approximation. & Wets. Bjørkvoll. T. (2006). (2007). Norway. 125–135. International Journal for Numerical Methods in Fluids 3. Tomasgard. A linear model for transient gas ﬂow. K. A. Nygreen. T. 1751–1758. & Moritz. (2006). & Kristiansen. M. ‘Scenarios and policy aggregation in optimization under uncertainty’.. M. & Tomasgard. Fodstad. K. Y. R.. Norway. NTNU. A. ‘Modelling norwegian petroleum production and transportation’. (2004). M. Norway.. B. Norway. (2000). available as SINTEF-report STF 38S03601. A. (2007). Christiansen. (2003). Capacity booking in a transportation network with stochastic demand and a secondary market for transportation capacity. L. 119– 147. PSIG. NTNU. ‘Fakta norsk petroleumsverksemd 2006’. Working paper at NTNU. 563–582. pp. & Sekirnjak. Modeling optimal economic dispatch and ﬂow externalities in natural gas networks. Bjørndal. M. M.. Möller. (2006). Master’s thesis. Trondheim... M. S. Trondheim. E. Ø. T. C. A. Martin. L. (1998). M. Nowak. & Nowak. NTNU. PSIG. Equation of state tutorial. Trondheim. & Tomasgard. Modeling short term planning in natural gas networks (in Norwegian). Mathematical Programming 105(23). A practical approach to transient optimization for gas networks. Midthun. Haugen.Bibliography Kelling. Norway. Mathematics of Operations Research 16. (1983). Smeers. Annals of Operations Research 82.. Trondheim. M. An optimization system for gas transportation on the norwegian continental shelf. ‘Mixed integer models for the stationary case of gas network optimization’. M. Submitted to international journal.

. V.. The Netherlands. Hasle. in G. eds. University Duisburg-Essen (Germany).. (2004).. B. Springer... L. 357–364. I. H. Mason. T. & Rasmussen. S. December. (1997). Nygreen. January 2006. M. Streeter. 550– 564. Zhuang. (2003). Kuok. (1970). E. International business school and research center for natural gas. B. A. PhD thesis. P. Canada. Energy Economics. ‘Geometric Modelling. European Journal of Operational Research 176. Working paper. Ulstein. Quak. M. K. (2006). (2004). A short term operational planning model for upstream natural gas production systems.Bibliography Ruszczyński. Society of Petroleum Engineers Journal. & Gabriel. (2007). Fodstad. N. & Wylie. ‘Natural gas pipeline transients’.. A complementarity model for solving stochastic natural gas market equilibria. A. A. 333–353. Lie & E. L. K.-A. Thompson. accepted. (2007). Optimization models for the natural gas value chain. PhD thesis. chapter Optimization Models for the Natural Gas Value Chain. J. & Sagli. K. & Barton. M. J. ‘Tactical planning of oﬀshore petroleum production’. and Optimization: Applied Mathematics at SINTEF’. Robinson. Rømo. Davison. F. Numerical Simulation. & Midthun. T. M. 143 . Anwendungen der Stochastischen Optimierung im Stromhandel und Gastransport. M. Natural gas storage valuation and optimization: A real options approach. to appear in AIChE Journal. ‘Decomposition methods in stochastic programming’. Math in Gas and the Art of Linearization. Mathematical Programming 79. Tomasgard. Westphalen. (2007). Groningen. Selot. L. Van der Hoeven. University of Western Ontario.

.

Mette Bjørndal.Paper IV Kjetil T. Asgeir Tomasgard and Yves Smeers: Capacity booking in a Transportation Network with stochastic demand and a secondary market for Transportation Capacity . Midthun.

.

The model looks at the challenges faced by the network operator in regulating such a system. and capacity bought in the ﬁrst-stage primary market can be sold by the producers. The booking regime is similar to the regime implemented in the North-Sea. The demand for capacity is stochastic when the booking in the primary market is done.Chapter 5 Capacity booking in a Transportation Network with stochastic demand and a secondary market for Transportation Capacity Abstract: We present an equilibrium model for transport booking in a gas transportation network. but the model and results are interesting for natural gas transportation in general. Here the network operator can sell remaining capacity in the system. We also investigate the eﬀect of using diﬀerent model 147 . with access to a primary market for transportation capacity. This is modelled as a Generalized Nash Equilibrium using a stochastic complementarity problem. where also the competitive fringe participates. We consider diﬀerent objective functions for the network operator. In the ﬁrst stage the large producers book capacity within their predeﬁned capacity rights. The purpose of the paper is to develop a model that can be used to analyze how diﬀerent objective functions for the system operator aﬀect the eﬃciency of the transportation system. There is also an open secondary market for transportation capacity where all players participate including a competitive fringe. There are two booking stages in the transport capacity market. and the diﬀerence between setting ﬁxed capacities and modeling the pressure constraints in a sub-sea pipeline-network. 5.1 Introduction We study booking of transportation capacity in a natural gas network with several large players and a competitive fringe. In the second stage there is a redistribution of capacity in a bilateral secondary market. The oﬀshore pipeline system in the North-Sea provides a case for our analysis. There are some privileged players in the network.

To our knowledge this is the ﬁrst time the booking system for natural gas transportation is studied using this approach. In Jing-Yuan & Smeers (1999) spatial oligopolistic electricity models are given and Generalized Nash Equilibria are found in a system with Cournot generators and regulated transmission prices. Yao et al. The model formulation as a stochastic Mixed Complementarity Problem is presented in Section 5. see for example Ferris & Pang (1997) and Facchinei & Pang (2003).Chapter 5 Capacity booking in a Transportation Network. the conclusions are given in Section 5. Finally. The loop ﬂow is taken into consideration and shown to be important for the results. Hu et al.A. the network operator acts as a neutral third party. by for instance Yao et al. as well as the assumptions we have made..6.4. In the North-Sea. representations of the physical properties of the transport network. (2004) where the network operator choose the production from each producer in order to maximize social surplus. 148 .5. (2004) model strategic bidding by generators to an ISO that is maximizing social surplus. In Section 5. The network operator inﬂuences the eﬃciency in the network through the routing. Zhuang & Kiet (2005) presents a linear complementarity equilibrium model for the North American natural gas market. A more detailed description of the equilibrium conditions is given in Appendix 5. Kiet & Zhuang (2005) presents a multi-seasonal. A path-breaking paper for the use of complementarity problems modelling economic equilibrium was Lemke & Howson (1964).3. The routing decisions will also determine the capacity sold in the secondary market. This is diﬀerent from the role of the network operator compared to the articles studying electricity networks. Gabriel. In the energy sector there are numerous examples of papers using complementarity problems to model and solve economic equilibria. (2004) and Hu et al. We formulate the model as a mixed complementarity problem.. (2006) presents a model of twosettlement electricity markets using an Equilibrium Problems with Equilibrium Constraints (EPEC). The model is based on Generalized Nash Equilibrium and is represented as a stochastic complementarity problem. We then move on to some numerical examples in Section 5. The model is an EPEC solved as an All-KKT system in PATH.2 we discuss the background for this article. Gabriel. Smeers (2003a) and Smeers (2003b) discuss the deregulation of the electricity markets and the organization of regional electricity transmission. multiyear mixed nonlinear complementarity problem of natural gas markets. Hobbs (2001) presents Cournot models of bilateral power markets. Another interesting topic is how stochasticity in the price for natural gas inﬂuences our results. The properties of the model are discussed in Section 5.

In our model. The routing is the responsibility of the ISO. 149 . Germany. The transportation network can be considered as a black box for the producers.2 Problem description and assumptions 5. In sum. a capacity allocation key is used to resolve these matters.2 Problem description and assumptions We present here the ideas and motivation for our case analysis. This booking right depends on their need to transport induced by the TOP contracts. The secondary market is open to all qualiﬁed shippers. as illustrated in the image on the right in Figure 5. both transactions of capacity facilitated by the ISO and bilateral transactions between shippers are included. The actual demand for capacity due to the TOP contracts is uncertain until delivery. This is reasonable as Norway’s overall production is around 15% of the European consumption of natural gas. In the secondary market. The producers book transport capacity from ﬁeld to market and can not determine the actual routing of the gas through the network. and the remaining capacity after this initial distribution is handled through a secondary market. the assumptions we have made and the reason for introducing them. If a conﬂict arises with respect to over-booking. In this article we have assumed that the producers may act strategically in the transport capacity market. We have not explicitly modeled this rule in this paper. Only the large producers book capacity in the primary market limited by predeﬁned capacity rights. The producers deliver natural gas into a transportation network passing through junction nodes and ending in market nodes. and for details on the Norwegian case. The market for capacity in this network is managed by an independent system operator (ISO) named Gassco. In the market hubs there are large buyers of natural gas who distribute the gas further to either the suppliers or end-customers. At the Norwegian Continental Shelf (NCS) capacity distribution is done in a primary market. see Austvik (2003). In addition to the long term contracts for gas in the markets nodes. the analysis ends at the market hubs. The system operator operate the network taking into account the details in the network. each with a set of large producers in addition to a competitive fridge.1. the available capacity in the primary market is actually larger than the total capacity in the network.1 illustrates the point-to-point perspective of the producers. there are short-term markets where gas may be sold. The image on the left in Figure 5. France and Belgium. but that they are price-takers in the spot markets in the market nodes.5. System in the North-Sea We study a system with ﬁeld nodes. For details regarding the liberalization of the European gas market see European Union (1998) and European Union (2003). The main market hubs are in UK.

The stochastic elements are the spot price in the markets and the quantity in the 150 . For each choice of secondary market sales from the ISO. The purpose of our model is mainly to analyze the eﬀect diﬀerent objectives of the ISO will have on the operation of the system. The price the ISO can take is regulated and ﬁxed both in the primary and in the secondary market.. Second-stage decision structure Our model is a one level game where each of the producers decison problem is a stochastic two-stage program with recourse (Kall & Wallace 1994).Chapter 5 Capacity booking in a Transportation Network. a solution satisfying the large producers’ equilibrium conditions can be found. The gas ﬂows from top to bottom. and acts as a benevolent central planner. If we represent the ISO with a feasibility problem. Figure 5.. junction nodes by j and market nodes by m. max value of ﬂow and max social surplus. the corresponding game will have an inﬁnite amount of equilibria. We also investigate how the representation of the physical networks as well as the booking rights in the primary markets will inﬂuence the eﬃciency of the network.1: The ﬁeld nodes are denoted by g. Hence we focus on the following alternatives: max ﬂow. so its decision variables are only routing and secondary market sales of available capacity. In the following we assume that the ISO does not have economic interests in the routing.

The uncertainty is modeled with scenarios (see Figure 5. when the ISO has a convex optimization problem. In this case each player solves a mathematical program with equilibrium constraints (MPEC.2. Further. and a common way of modeling this is to merge all the players KKT-conditions into a large complementarity system. the ISO is a Cournot player whose volume decisions cannot be manipulated by other players strategically. (1996)) and the resulting game over all the players become an EPEC. This is conﬁdential information.2 Problem description and assumptions TOP-contracts. sales and production volumes of the other players. They will then act strategically because they anticipate the ISO’s reaction to their own volume decisions. In the secondary market (in the second stage) we assume that the large producers and the competitive fringe make simultaneous volume decisions in a Cournot manner. is by including the KKT-conditions for the ISO’s routing and capacity release in the other players’ optimization problem. This demand is positive as long as the market price for natural gas in a market hub is higher than the marginal production cost for the competitive fringe in a ﬁeld node plus the transportation price from that node to the market. An alternative would be to model this as a multi-leader one-follower Stackelberg game (Yao 2006) with the ISO as a follower. Secondly. The decisions in the two stages are illustrated in Figure 5. In practice this means that the second-stage decisions 151 . Firstly. We assume that each player makes his ﬁrst-stage decisions and his second-stage decisions simultaneously.5. but make independent volume decisions. the other players never know or get information about the ISO’s routing decisions. Hence. Their reaction function is expressed as their demand for transportation capacity at a given transportation price. In our approach we stay within the framework of Mixed Complementarity Problems as all decisions are simultaneous. A common way of modeling this follower situation. for example in terms of inﬂuencing the ISO in the transportation market. we assume that the ISO’s decisions are made simultaneously with the producers. We think that the setting with simultaneous decisions is closer to the reality of the Norwegian continental shelf.3). Luo et al. the players are not supposed to act strategically. This booking decision is based on maximizing the excepted revenue for the second stage where production and transportation strategies are made as well as trades in the secondary market for transportation capacity. First-stage decision structure In the ﬁrst stage each of the large producers decides on a booking volume. The players in the competitive fringe are price takers in the capacity market. and so are the booking requests. Each of the large producers recognizes that they will inﬂuence the price for transportation capacity.

and second-stage decisions are made simultaneously we model the situation where either a player does not know the other players’ booking decisions when he makes his second-stage decisions. will depend on the outcomes of the stochastic variables. 152 . When the ﬁrst. Each producer’s optimization problem is then a stochastic twostage program with recourse. The overall problem is still a Mixed Complementarity Problem. on the other hand the booking decisions had been used strategically by the players. Then the only information revealed (or acted on) between the ﬁrst and second stage is the uncertainty that is resolved.Chapter 5 Capacity booking in a Transportation Network. often called a Stochastic Mixed Complementarity Problem because of the stochastic variables and twostage structure. If. This is a one level game as the scenarios are independent of the ﬁrst-stage decisions. we would need to include the second-stage equilibrium over all the players as a part of the booking problem in the ﬁrst stage for each player. we feel that this is a sound model. In such a setting each player’s problem would be a stochastic MPEC.2: The sequencing of decisions. In the Norwegian regime with a conﬁdential booking process.. given the other players ﬁxed decisions. Figure 5. where the second-stage equilibrium conditions for each scenario is part of the ﬁrst-stage optimization problem and parameterized on the ﬁrst stage decisions (Patriksson & Wynter 1999). or he has this booking information but does not let it inﬂuence his second stage decisions. but the contingent strategy covering all possible outcomes is made before the player observes the other players booking. Normally this is done by including the KKT-conditions from the second stage equilibrium in each player’s ﬁrst-stage optimization problem..

5.2 Problem description and assumptions Figure 5.3: The scenario structure in the large producers’ decision problem 153 .

3 Model We start by introducing the notation. The networks we present are connected graphs. The set of ﬁeld nodes in the network. Index for ﬁeld nodes. We then move on to a discussion of the price of transportation capacity in the secondary market. Index for market nodes. The set of nodes with pipelines going into node n (predecessor nodes). Index for junction nodes. Scenario index. The set of all producers in ﬁeld g (including the competitive fringe). The index used for producers. The set of scenarios. Notation Sets N G J M I(n) O(n) L Lg S The set of all nodes in the network. Indexes n g j m s l Used for nodes in general. The set of large producers in the network. The set of junction nodes in the network. The set of nodes with pipelines going out of node n (successor nodes). 154 . 5.Chapter 5 Capacity booking in a Transportation Network.. The set of market nodes in the network..

Capacity sold by the ISO in the secondary market between ﬁeld g and market m in scenario s. Probability of a given scenario. Spot sale in market m by producer l in scenario s. Capacity in the pipeline between node n and i. Quantity produced in ﬁeld g and sold in market m in scenario s by the competitive fringe. Capacity between g and m traded by producer l in the secondary market in scenario s. Price in the long term contracts for producer l in market m. The ﬂow between node n and i in scenario s. The pressure in node n in scenario s. 155 . Cost function for the competitive fringe in ﬁeld g. The Weymouth constant for the pipeline going from i to j.3 Model Constants Rn Rn Kij Blgm Plm Tgm MC g C ni cg clg The maximum pressure in node n. Functions Clg (d) Wg (y) The cost function for producer l in ﬁeld g. Price of transportation capacity between ﬁeld g and market m in scenarios s. Spot price in market m. The minimum pressure in node n. Production in ﬁeld g by producer l in scenario s. Parameter in the cost function for producer l in ﬁeld g. Per unit tariﬀ for transportation between ﬁeld g and market m. Aggregated marginal cost parameter in ﬁeld g. Decision variables blgm dlgs qlms hlgms fnis rns zgms tgms xgms Booking in the primary market by producer l between ﬁeld g and market m.5. Booking limit for producer l from ﬁeld g to market m. Parameter in the cost function for the competitive fringe in ﬁeld g. Stochastic variables and probabilities vlms pms φs Nomination in long-term contracts in market m.

1) where xgms is the quantity traded in spot market m by the competitive fringe in ﬁeld g in scenario s. m ∈ M. The competitive fringe’s demand function for transportation capacity between ﬁeld g and market m in scenario s is then found from the proﬁt maximization problem for the competitive fringe in ﬁeld g: Πgs = max m∈M (pms · xgms − tgms · xgms ) − Wg m∈M xgms . all results are valid for general quadratic cost functions (and most for a general cost function). and negative if the large producers buy capacity.2) In this article. and the large producers. Price of capacity in the secondary market The price in the secondary market in a node is given by a demand function from the competitive fringe in that node. zgms and hlgms : 156 . The hlgms variable is positive when the large producers sell capacity. g ∈ G. we will in the following assume that the cost function for the competitive fringe is: Wg m∈M xgms 1 = cg · 2 2 xgms m∈M (5.4) The volume bought by the competitive fringe. The inverse demand function is given as: tgms = pms − cg · m ∈M xgm s . We model this implicitly in the large producers’ problem as an elastic demand function. xgms is given as the sum of capacities sold by the ISO. m ∈ M. (5.3) where cg is the cost parameter for the competitive fringe in ﬁeld g. Nevertheless. the ﬁrst order condition for optimality is used: ∂Wg ∂Πgs = pms − tgms − ∂xgms m∈M xgms ∂xgms = 0. we assume that Wg is a quadratic function. zgms . (5. s ∈ S. In order to ﬁnd the demand function for the competitive fringe. g ∈ G. We then have the following relation between xgms . hlgms .Chapter 5 Capacity booking in a Transportation Network. We assume that the competitive fringes in the diﬀerent ﬁeld nodes are independent. For ease of presentation. Wg is the cost function in ﬁeld g.. s ∈ S.. tgms is the price of transportation capacity between g and m in the secondary market in scenario s. (5.

7) where hlgms is the secondary market trades of producer l of capacity from g to m. g ∈ G. (5.3 Model xgms = zgms + l∈L hlgms . The cost for the producers come from the per unit tariﬀ paid in the primary market for transportation capacity (which we assume is independent of the large producers’ decisions). The large producers The income for the large producers (L) in the network comes from deliveries in the long term contracts. Tgm is the tariﬀ in the primary market.6) Since we only allow ﬂow in one direction in our network. m ∈ M. (5. we need to make sure that xgms cannot be negative.8) where blgm is the booking in the primary market. zgms + l∈L hlgms ≥ 0. m ∈ M.5. the cost of production and from purchasing additional transportation capacity in the secondary market. pms − cg · hl gm s − (5. s ∈ S. s ∈ S. g ∈ G. g ∈ G. qlms 157 .5) which leads to the following expression for the price in the secondary market: tgms = pms − cg · m ∈M zgm s + l∈L hlgm s . sales in the spot markets and sales in the secondary market for transportation capacity. (5. pms is price in the spot market. φs is the probability of scenario s. m ∈ M. s ∈ S. The inclusion of this constraint means that the decision space for each producer depends on the other participants decisions (the other producers and the ISO). The objective function for producer l can be formulated as: Πl = max − + s∈S g∈G m∈M Tgm blgm + s∈S φs m∈M (pms qlms + Plm vlms ) zgm s + m ∈M l ∈L φs φs s∈S g∈G g∈G m∈M hlgms · Clg (dlgs ) .

zgms is the capacity sold by the ISO in the secondary market. s ∈ S. The booking constraint in the primary market is given as: blgm ≤ Blgm . (5. negative when he buys). g ∈ G.6). (5.11) (5. m ∈ M. Constraint (5. g ∈ G.10) (5. the price in the secondary market is given by (5.13) qlms + vlms = hlgms ≤ blgm .14) The network operator will always choose zgms in order to maximize the ﬂow under the constraint that all prices (for ﬁeld-market combinations) must be positive (see Equation (5..9) where Blgm is the ﬁxed upper limit on booking for the producer. With this objective. hlgms is the capacity traded in the secondary market (positive when the producer sell capacity.10) make sure that the producer has booked enough capacity for the production in ﬁeld g.11) make sure that the producer has booked enough capacity for the total sale in market m.Chapter 5 Capacity booking in a Transportation Network. g ∈ G. Constraint (5. m ∈ M. Independent system operator We present three diﬀerent objective function alternatives for the ISO: • maximize ﬂow (MF): ΠMF = max s m∈M i∈I(m) fims . s ∈ S. g ∈ G.12) makes sure that the producer cannot sell more capacity than he has booked in the primary market. is volume sold in the spot market. The two constraints also make sure that the producer utilizes all the booked capacity. s ∈ S. g∈G (5. the system operator will be 158 ..13) ensures that the producers cannot buy more capacity than the ISO sells. For the second stage the following constraints are needed: dlgs = m∈M (blgm − hlgms ) . s ∈ S. Constraint (5. Plm is the price in the take-or-pay contracts. and constraint (5. zgms + l∈L hlgms ≥ 0. m ∈ M. m ∈ M. (blgm − hlgms ) .12) (5. vlms is the volume in the take-or-pay contracts. Clg is the cost function for the producer and dlgs is the production.26)).

compared with the MF formulation is that the ISO now routes the gas according to value. The weakness is that he has no incentive to route according to marginal cost in the ﬁelds. • maximize value of ﬂow (MVF): ΠMVF = max s m∈M i∈I(m) pms · fims − l∈L vlms . MVF.5. MC g is found in the following manner: MC g = 1 1 l∈Lg 2clg . • maximize social surplus (MSS): ΠMSS = max s m∈M i∈I(m) pms · i∈O(g) fims − 2 fgi . If we assume that the network operator has full information regarding the cost functions of the participants. the ISO can take both value of ﬂow and cost structure in the ﬁelds into account by maximizing social surplus. the aggregate supply function is linear.19) 159 .15) The strength of this formulation. (5. l∈L vlms + m∈M l∈L Plm vlms 1 − 2 MC g · g∈G (5.18) and that no production capacities exist. Under these assumptions. lg (5. (5.16) MC g is the slope of the linear aggregated supply function for ﬁeld g: MC g · i∈O(g) fgi .3 Model indiﬀerent with regards to prices in the market nodes and cost functions in the ﬁeld nodes. (5.17) The supply function is found by assuming that all producers have a cost function of the form: Wg = clg d2 .

The aggregated supply function is found by horizontal summation of the individual supply functions. and Rn is the largest allowed pressure in node n.We have chosen to linearize this expression with the formulation used in Tomasgard et al. n ∈ N . diameter.22) In the following we will refer to this formulation as Independent Static Flow (ISF). (5. roughness) as well as external variables (temperature) decide how much gas is transported for a given pressure diﬀerence. (2007): fij ≤ Kij RIi 2 RIi 2 ri − ROj − Kij ROj 2 2 RIi − ROj rj . Between the production facilities and the market-hubs there is a transportation network.Chapter 5 Capacity booking in a Transportation Network. In the numerical analysis. In addition. (5. i ∈ O(n). s ∈ S.20) About 20 of these constraints that are approximating the Weymouth constraint are used for each pipeline in order to linearize the ﬂow around pairs of pressure in. we will also look at an alternative formulation with ﬁxed capacities. It is non-trivial to determine appropriate values for the ISF capacities. The gas molecules are transported from nodes with high pressure to nodes with lower pressure through pipelines.. (5. (2006) for a discussion. The ISF capacities are then set equal to the resulting ﬂow pattern in this model. The relation between pressure in the nodes and ﬂow in the pipelines are determined based on the Weymouth equation. see for instance Menon (2005). but it also represents the real system more precisely as it includes the ﬂexibility of moving bottlenecks by 160 . ROj . Here fij is the ﬂow from node i to j and rn is the pressure in node n. In this paper we solve an optimization model (with WF formulation) where the objective is to maximize the throughput in the network. where Lg is the set of producers L and the competitive fringe in ﬁeld node g. In this case the pressure constraints and the Weymouth equation are replaced with the following formulation: fnis ≤ C ni . The design parameters of the pipelines (length. (2006). and pressure out. See Midthun et al. The WF formulation is a relaxation of this ISF formulation. RIi . while the Weymouth formulation is referred to as WF. For a discussion of system eﬀects on capacity related to pressure constraints see Midthun et al.. constraints on the pressure level in each node must satisﬁed: Rn ≤ rns ≤ Rn n ∈ N .21) where Rn is the smallest allowed pressure in node n.

The system operator must make sure that the mass is conserved in the network. (5. the better the strategy is with respect to maximizing the social surplus. g ∈ G. m ∈ M.24) where I(j) is the set of nodes connected to a pipeline entering node j. 161 .5. we assume that no junction nodes are connected to each other. and that each market is connected to a junction node. g ∈ G.25) The following constraint is included in the model with maximum ﬂow and maximum value in order to ensure that the price in the secondary market is positive: pm − cg · m ∈M zgm s + l∈L hlgm s ≥ 0. m ∈ M. Benchmark In Chapter 5. The mass balance equations are given as: fgjs = j∈O(g) m∈M zgms + l∈L blgm . s ∈ S. (5. In the junction nodes. Finally.23) where O(g) is the set of nodes connected to a pipeline leaving from ﬁeld g. the mass balance can be formulated as: fgjs = g∈I(j) m∈O(j) fjms . In addition for ease of notation. (5. routing and sale in order to maximize the social surplus of all the players in the network.3 Model adjusting pressures. We assume that each ﬁeld is connected to a junction node. (5. j ∈ J . we could have introduced a constraint that ensured that the ISO income was positive in total (or for all ﬁeld-market combinations).26) Alternatively. The ISF formulation is more restricted but any increase in its capacities will allow a solution which is infeasible in the WF formulation. s ∈ S.5 we benchmark our solutions with an optimization model where an independent operator schedules production. The closer the equilibrium in our game gets to the benchmark solution. s ∈ S. a constraint for the mass conservation in the market nodes must be included: fnms = n∈I(m) g∈G zgms + l∈L blgm . The mathematical formulation of the benchmark model is given below.

. Pioneering results on the existence of GNE are presented in the papers of Debreu (1952) (social equilibrium) and Arrow & Debreu (1954) (abstract economy) that generalized the results of Nash (1950).. .27) where MC lg is the slope of the linear supply function of producer l in ﬁeld g. 1 β it would be the strategy set of the game). fjms . the Generalized Nash Equilibrium is reached when no player has incentive to change his strategy given that the other players do not change their strategy. . s ∈ S. L. . x ∈ X. xT )T (in the case that no common constraints existed. The set X = l∈L Xl is the full Cartesian product of the strategy sets of individual ¯ players and x = (xT . The mass balance is taken care of by: dlgs = l∈L j∈O(g) fgjs .30) fgjs = g∈G (qlms + vlms ) = l∈L j∈I(m) fjms . s ∈ S. ΠBM = max s m∈M l∈L (pms qlms + Plms vlms ) − g∈G l∈L 1 MC lg d2 . . lgs 2 (5.28) (5.Chapter 5 Capacity booking in a Transportation Network. 5. Also deﬁne the vector x−l of all players’ decisions except player l’s and correspondingly X−l = j∈L|j=l Xj . xl ∈ Kl (x∗ ). ..4 Model properties Our model is a General Nash Equilibrium game where the feasible regions of the players depend on the other players’ decisions. .29) (5. x∗ ). g ∈ G. l ∈ L where l −l −l −l αβ Π : R → R is the objective function of player l.3. Deﬁne β = |L| + 1. We have |L| producers ¯ and 1 ISO. Rosen (1965) is an early paper concerning not only existence but also investigating uniqueness of solutions for a 162 . given the actions of the other players. In addition. Kl (x−l ) ⊆ Xl . . constituting the set of players. ¯ We will deﬁne more formally the dependence between the players through the common constraints and deﬁne the point to set mapping Kl : X−l ⇒ Xl representing player l’s feasible region. j ∈ J .20) and (5.21) from the network operator problem presented in section 5. we need constraints (5. Let Xl ∈ Rα be the strategy set of player l with decision variables xl = (xl1 . Then a generalized Nash equilibrium (GNE) is deﬁned as a point x∗ ∈ X that simultaneously optimizes all the players individual decision problems so that: ¯ ¯ x∗ ∈ Kl (x∗ ). s ∈ S m∈M (5. xlα ) . m ∈ M. That is. l ∈ L and Πl (x∗ ) ≥ Πl (xl .

.16)). K(x)) in (5. linear expressions since our objective functions are quadratic (see Equations (5.31) where K(x) = l∈L Kl (x−l ). Since all these equations are linear.14)-(5.K): F (x∗ )T (x − x∗ ) ≥ 0.13).4. x−l ) and F (x ) = (F0 (x ) .2 from Chan & Pang (1982) (Theorem 2 in Harker (1991)) give conditions for existence of a solution.32) In our case. h. The mapping in our model is deﬁned by Equations (5. x ∈ K(x∗ ). K(x)) : F (x∗ )T (x − x∗ ) ≥ 0. also the theory of VI may be used to analyze this. For our problem this is satisﬁed by the deﬁnitions of F and K. . ¯ It may here be noted that a standard Nash equilibrium may be expressed as a VI(F. continuous. Following the lines of the discussion in Harker (1991).31). 3.26). r. F is continuous on X. q.1.4 Model properties restricted class of problems. Suppose that there exists a nonempty compact set X such that 1.1 are satisﬁed. the conditions in Theorem 5.23) and (5. Then there exists at least one solution to the QV I(F. f. see Lions & Stampacchia (1967) for a nice overview. Let F and K be a point-to-point mapping and point-to-set mapping respectively from Rαβ into itself.8) and (5. K is a nonempty. . This means that in addition to existence and uniqueness proofs following the Arrow/Debreu/Rosen tradition. z). (5. 2. It is well known that Nash equilibria (with independent player strategy sets) can be viewed as Variational Inequalities (VI). x ∈ X. (5. lchiishi (1983) gave more general results concerning the existence of such GNE. Theorem 5. see Harker & Pang (1990). the x vector consist of the following variables: x = (b.25)-(5. F|L| (x ) ) Then the GNE may be expressed as the Quasi Variational Inequality QV I(F. We use notation in accordance with what we deﬁned above: Theorem 5. An early reference formulating the generalized Nash equilibrium as a Quasi Variational Inequality (QVI) is Bensoussan (1974). K(x) ⊆ X. x ∈ X. . (5. F consists of continuous. d.4. Harker (1991) and Pang & Fukushima (2005) for good overviews of this direction of analysis. closed and convex valued mapping on X. 163 . we deﬁne Fl (x∗ ) = ∗ ∗ ∗ ∗ T ∗ T T xl Πl (xl . We then know that our Generalized Nash Game has at least one solution.5. See for example Ferris & Pang (1997) or Facchinei & Pang (2003) for more on the relationships between complementarity problems and Variational Inequalities.

The theorems also state that any strictly interior solution 0 j (for the common constraints) of the QVI (F. This means that if an interior x∗ is known.36) fgjs = j∈O(g) . or a zero value of a common resource for all players (if not used in full). .33) (5. Theorem 2. the only other GNE may be found at the boundary of the common constraints. ≥ 0. Facchinei & Pang (2003). K(x)) is a solution to the VI (F. K(x)) at which the optimal dual variables λ∗ ∈ Rpβ for the common constraints are such that λ∗ = λ∗ .33) gives the balance between capacity sold by the system operator and capacity traded by the large producers. the large producers buy less capacity than the ISO sells. blgm blgm l∈L (5. all the common constraints are continuous.36)) and satisfy the necessary constraint qualiﬁcations (LICQ). These theorems state that if F is a continuous function in the VI (F.Chapter 5 Capacity booking in a Transportation Network. Further. X) as described in (5. linear functions (see Equations (5. j ∈ L..33)-(5. ums . X) then the VI solutions are the only points in the solution set of the QVI (F. the shadow price τgms then gives the value of an additional unit of capacity bought.3. the shadow price gives the value of selling one 164 . In our model we have the following common constraints (dual variables belonging to each constraint are given to the right): zgms + l∈L hlgms ≥ 0 zgms + m∈M l∈L τgms . We can therefore apply Theorems 4-6 from Harker (1991) directly. χgms . In our model.35) (5. For the producers. but only the VI solutions will have a common positive value of an additional unit of a common resource (if the resource is depleted). Common constraints We deﬁne common constraints as constraints where decision variables for more than one player appear. In general there will be several GNE in the game.34) (5. if F is strictly monotone there is a unique solution to the VI over X. the large producers are buying all capacity sold by the ISO. ugs . If this constraint is not binding.3. We have focused on the VI solution in this article.32).. fnms = n∈I(m) g∈G zgms + zgm s + pm − cg · m ∈M hlgm s l∈L Constraint (5. For the ISO. and they will not have equal λ’s for the common constraints. A discussion of the common constraints and the implication of requiring equal shadow prices are given in the next sections. If the constraint is binding.

the tariﬀ is ﬁxed and may not be changed in order to give speciﬁc incentives to the players. Hence it is clear that the ISO has a lot of inﬂuence through the routing decisions. In addition. We have not been able to prove that the F function is strictly monotone. 5. there is a competitive fringe in g1 and g2 .34) and (5. and the equilibriums we present in the numerical examples may therefore not be unique. For the ISO.5 Numerical examples additional unit of capacity and thus increasing the ﬂow in the network with one unit. Constraint (5. since we relate objective functions that are not commensurable with respect to the units. There are two large producers. The price depends on the volumes sold by the ISO and the large producers. We then introduce stochasticity to our model to see how it inﬂuences the eﬃciency in the network.5 Numerical examples We consider the network illustrated in Figure 5.4. For both the producers and the ISO. the shadow price gives the value for the ISO of increasing the ﬂow out of the ﬁeld. Since the ﬂow variable is part of the objective function for the ISO.36) ensures that the price in the secondary market is positive.35) gives the balance for each market node. we advocate that the VI solution to the GNE game is the important one. In the following sections.34) gives the balance for each ﬁeld node. For the MVF and MSS formulation. Constraints (5.5. the VI solution depends on the conversion of 1 Sm3 to NOK. each present in both g1 and g2 . and constraint (5. we use our model to analyze several cases. Constraint (5. The same argument is valid for the shadow price ums . it may not make sense to require equality in the equilibrium solution. For the producers. In the system we have described. The equilibrium solution will change if we change the currency (from NOK to for instance Dollars or Euros). Our model is designed for a situation where both a primary market and a 165 . and such a fair routing policy is preferable. the shadow price ugs gives the value of booking one additional unit of capacity out of ﬁeld g in the primary market. In this case the ISO will have made routing decisions which make sure that all players’ marginal value of an additional transportation unit is equal.35) specify that the booked capacity in the network must be equal to the actual ﬂow in the pipelines. the shadow price χgms gives the value of selling one additional unit of transportation capacity. the shadow price gives the value of increasing the diﬀerence between the ﬂow out of ﬁeld g and the capacity sold. For the MF formulation however. We start with a deterministic setting in which we look at the diﬀerent ISO objective function alternatives and the diﬀerence between the WF formulation and the ISF formulation. Since the marginal values are given in diﬀerent units. zgms .

The equilibrium conditions consist of the aggregated KKT-conditions for all players (see Appendix 5. while the large producers inﬂuence the eﬃciency by booking in the primary market and trading in the secondary market.4: The network used in the numerical examples. the booking in the primary market is limited by predeﬁned booking limits and in case of overbooking a capacity allocation key is used to distribute the scarce capacity. The ISO inﬂuences the eﬃciency of the network through routing decisions and capacity distribution in the secondary market. the total booking rights in the North-Sea is twice the real capacity. we can use unlimited booking rights in the primary market.. In reality. secondary market is used to allocate capacity in the network. In our model we resemble this capacity allocation key by requiring equal marginal value for all players in our common constraints. In each case we solve the stochastic MPC by formulating the equilibrium conditions for the problem. Since we have inelastic demand functions in the market nodes. Because of this allocation rule. the social surplus will be identical with the producer surplus in our network (which is an interesting setting from a Norwegian perspective). As we discussed in Section 5.4.Chapter 5 Capacity booking in a Transportation Network.A). In the North-Sea today. 166 . we enter the KKT-conditions to the complementarity problem solver PATH (Dirkse & Ferris 1995). we focus on the 1 VI solution to the problem. Figure 5. All prices and costs are given in 100 NOK . In order to ﬁnd an equilibrium..

The reason is that the VI solution requires equal marginal values for all players in the common constraints. The tariﬀ in the primary market is 10 for each ﬁeld market combination.99 Table 5.1. c21 = 3. The parameters in the cost functions for the large producers (see Equation (5. c2 = 12.2.8% lower than the benchmark solution.5. Since the ISO only considers the income from the ﬂow in the network (and not the production costs). The results from the optimization with the three diﬀerent objective functions for the ISO is given in Table 5.6 0. When we solve the benchmark model (see Section 5.11 44. diﬃcult to 167 . c12 = 6. we get a total surplus for all the players of 7220. 2 and for the competitive fringe in ﬁeld g (see Equation (5.41% and 1. the model version where the ISO maximizes social surplus (MSS) gives the highest total surplus in the network. as discussed in Section 5.3). As we can see from the results.3)): c1 = 10.5 0. The total surplus is only 0. The reason for the decrease in social surplus is that the production costs have increased more than the income from the spot market. This corresponds to the maximal achievable surplus in the network. 9.39 52. The total social surplus obtained in the MVF and MF models are. We also see that the value of ﬂow is largest in the MVF formulation. The prices in the two markets are given as: pm1 = 130 and pm2 = 160. The equilibrium for the MF model is. c22 = 5.4 0.4.71 46.18)) are given as clg : c11 = 5 .59% smaller than the benchmark solution.5 Numerical examples Node/pipeline g1 g2 j1 m1 m2 g1 -j1 g2 -j1 j1 -m1 j1 -m2 R 180 185 170 130 130 R 170 170 130 115 100 Kij C ij 0. The network parameters are given in Table 5. respectively. the ISO has a large marginal value of ﬂow and therefore forces ineﬃcient production decisions from the producers.43. while the social surplus has decreased.1: The design parameters for the network Case 1: The diﬀerent ISO objective function alternatives (WF formulation) We start by illustrating the diﬀerence in the objective function alternatives we have presented for the ISO.35 38.

77 2410.5).88 Max ﬂow (MF) 222. In the ISF formulation the capacity in the network is therefore more restricted than in the WF formulation (the reason for including the ISF formulation is that it is a common approach for economic analysis in gas networks). In the MSS and the MVF formulation.2).. it has increased to the maximum possible ﬂow in the network in Table 5.17 3085. Case 2: ISF versus WF formulation In this example we look at the diﬀerence between using the WF formulation and the ISF formulation (see Section 5.92 763. Max social surplus (MSS) 258.39 Max value (MVF) 194.73 638.72 88.3. In Table 5.3 for a discussion of how the ISF capacities are determined).32 11668. WF formulation.15 7162.49 704.3 we see the results from changing the currency 1 from 100 NOK to EUR (this is done by changing the weighting of the ﬂow for the ISO. the change of currency will not aﬀect the solutions. By using a 1 currency of 100 NOK we put the emphasis on the large producers. so the units are comparable with the results in Table 5. If we change the currency (corresponds to changing the weighting of the ﬂow for the ISO). interpret since the units are diﬀerent in the objective functions for the ISO and the large producers.33 2282. While the ﬂow in the MF formulation was the lowest among the three alternatives in Table 5.35 11207..17 2595. we get a solution close to the benchmark. We see the same pattern in these results as we saw for the WF formulation: the MSS formulation 168 . and since the social surplus corresponds to producer surplus in our models.79 704. We use the same parameters as in the previous example (Section 5.89 2435. Every ﬂow pattern obtained with the ISF formulation is feasible within the WF formulation. The results from this optimization is shown in Table 5.4.35 678.39 76.51 6540.Chapter 5 Capacity booking in a Transportation Network. the equilibrium also changes.20 Competitive fringe g1 (NOK ) Competitive fringe g2 (NOK ) Producer 1 (NOK ) Producer 2 (NOK ) ISO proﬁt (NOK ) Social surplus (NOK ) Flow (Sm3 ) Value of ﬂow (NOK ) Table 5.5 7105.2: Results from the diﬀerent ISO objective functions.85 12870.2.05 80.22 704.17 3129. and in the remaining examples we will therefore focus on the MSS and the MVF formulations.

06 591. High correlation may result in less diﬀerence between the ISF and the WF formulation (since the ﬂexibility in the network is less important in this case).17 2328.80 782.85% for the MVF formulation. gives the highest social surplus in the network.93 Max value 174.10 13192.18 2599.91 91. the total surplus is reduced with 2.55 2098. ISF formulation.5.5 Numerical examples Max ﬂow (MF) Competitive fringe g1 (NOK ) Competitive fringe g2 (NOK ) Producer 1 (NOK ) Producer 2 (NOK ) ISO proﬁt (NOK ) Social surplus (NOK ) Flow (Sm3 ) Value of ﬂow (NOK ) 174.09 13191.42 707. The correlation between prices is also important.86 6090.97 10705.18 2329.82 2904. the uncertainty in prices and the volume uncertainty in the TOP-contracts. 169 .39 91. Compared with the WF formulation.19 71.96 6092.4: Results from the diﬀerent ISO objective functions. The importance of using the WF formulation depends on the network structure.40 2097.78 785.20 Table 5.18 704. Max social surplus 174. Large ﬂuctuations (as is common in natural gas prices) give more value to ﬂexibility and therefore the WF formulation will improve the eﬃciency in the network.71 6977.28 Competitive fringe g1 (NOK ) Competitive fringe g2 (NOK ) Producer 1 (NOK ) Producer 2 (NOK ) ISO proﬁt (NOK ) Social surplus (NOK ) Flow (Sm3 ) Value of ﬂow (NOK ) Table 5.42 707.58% for the MSS formulation and 6.3: Results from the MF formulation with a larger weight on the ISO objective function.

c22 = 7 . we assume that 170 . When the optimal booking in the ﬁrst stage is similar in all scenarios. That is. Node/pipeline g1 g2 j1 m1 m2 g1 -j1 g2 -j1 j1 -m1 j1 -m2 R 190 185 170 130 130 R 170 170 130 100 90 Kij 0.. The eﬀects of stochasticity are largest when the price is volatile. the 15 scenarios are solved independently and we then ﬁnd the expected value over the 15 scenarios.68% and 5.59.4 0.6 shows the results from the optimization. or negative. If volatility is low. the capacity allocation in the primary market is done such that the marginal unit goes to the player that has the largest expected marginal value.5 0. the optimal booking in the ﬁrst stage varies less between the scenarios. When prices are very volatile.35 Table 5. In the wait-and-see solution (WSS). The benchmark solution in this case is 9008. c12 = 4. the eﬀect of stochasticity is reduced. c2 = 9. In a stochastic setting. and the fact that all booked capacity must be used. We have chosen to use negative correlation and uniformly distributed prices between 75 and 225. and for the competitive fringe 2 in ﬁeld g: c1 = 9. this means that the large producers in some scenarios have more capacity than they ideally would have wanted to have. respectively.5: The design parameters for the network Case 3: The eﬀect of stochasticity In this example we look at the eﬀect of stochasticity in our model. The reason for these results is the capacity allocation we have chosen (focus on the VI solution). We see that the total expected social surplus in the network has been reduced with 3. We use the network parameters in Table 5.. Table 5.92% for the MSS and MV formulation. c21 = 4.6 0. The tariﬀs in the primary market are put at 10 for all ﬁeld-market combinations. and the following cost parameters for the large producers clg : c11 = 3. compared to the benchmark solution. or correlation is very high. We have also looked at the wait-and-see solution (Madansky 1960) and expected result of using the expected value solution (Birge & Loveaux 1997). and the correlation between the market prices is low.Chapter 5 Capacity booking in a Transportation Network.5.

Since the price of capacity is based only on one producer’s marginal cost (the competitive fringe).52 2985. and columns 4-5 shows the wait-and-see solution with unlimited booking.73 3275.19 726.28 8909.52 1155.37 15040. we found equilibria with a large distance 171 . The situation without a primary market We have also tested the model without a primary market (booking limits equal to zero).6) shows that the total surplus in the network has increased drastically in the WSS solution.5. we ﬁrst solve a deterministic problem where the stochastic variables are represented with their expected values (EVP). and found that the pricing mechanism in the secondary market was ineﬃcient in this case.10 3124.84 765.98 581. Columns 2-3 shows the result with unlimited booking for each producer. The results from this test (columns 4-5 in Table 5.82 3180.6: Results from the model with stochasticity.52 786.84 14522. and each ﬁeld-market combination.29 14830.62 3462.2% for the MVF formulation. The results from the EEV formulation is shown in Table 5. We then use the booking decisions from the EVP in the stochastic problem. The diﬀerence between the WSS solution and the solution from the stochastic model is the expected value of perfect information (EVPI).1% lower than the benchmark solution for the MSS formulation.99 84.50 637.77% higher than the EEV solution. In order to ﬁnd the the expected result of using the expected value solution (EEV).25 Competitive fringe g1 Competitive fringe g2 Producer 1 (NOK ) Producer 2 (NOK ) ISO proﬁt (NOK ) Social surplus (NOK ) Flow (Sm3 ) Value of ﬂow (NOK ) Table 5.28 758.93 96. the large producers somehow get perfect information of the future before they make their decisions in the ﬁrst stage.5 Numerical examples Booking limit = +∞ Max social Max value surplus 441.02 Wait-and-see solution Max social Max value surplus 453. and 3. EVPI tells us how much each player would have been willing to pay for knowing the outcome in the second stage.63 621.30 2927.43 99.30 94.71 3666.89 13660.7. The diﬀerences are small for the MVF formulation.06 8719. For the MSS formulation. The total expected social surplus is now only 1.91 420.92 1505.18 8475.55 8677.53 3303. we see that the stochastic solution is 1.

a diﬀerent market clearing mechanism in the secondary market is needed.11 1235.36 2994. Max social surplus 420..7: Results form the EEV formulation. For each of the large producers. In order to represent a situation without a primary market.. to the benchmark solution. In the deterministic setting we found a diﬀerence of 0. We have compared the results from our model with a benchmark model where a central planner with full information maximizes social surplus in the network.44 99. The design and tests of new clearing mechanisms is an interesting topic for future research. a decision to increase production will lead to an increase in production cost in addition to an increase in price of transportation capacity (when hlgms is increased.55 Max value 420.8% between the benchmark solution and the MSS solution. The formulation requires that the system operator has full information regarding the 172 . We found that the MSS formulation for the ISO lead to a higher total social surplus in the network than the alternatives.73 613.56 8463.84 14587.Chapter 5 Capacity booking in a Transportation Network.84 15040.18 95.63 696.6 Conclusions We have presented a stochastic MCP model based on Generalized Nash Equilibrium for analyzing a capacity distribution system with two stages: a primary market where only privileged players can participate and an open secondary market. 5.77 3216. This system is based on the existing capacity distribution system in the North-Sea.84 3262.02 Competitive fringe g1 (NOK ) Competitive fringe g2 (NOK ) Producer 1 (NOK ) Producer 2 (NOK ) ISO proﬁt (NOK ) Social surplus (NOK ) Flow (Sm3 ) Value of ﬂow (NOK ) Table 5. It may therefore be beneﬁcial for the large producer to decrease the production even if the marginal production cost is lower than the marginal revenue. We have shown that there exists at least one equilibrium solution (the VI solution) to our models. As illustrated in the numerical examples in this section. the price of capacity increase).63 1236.85 2892. the market clearing mechanism we have chosen works well in the presence of a primary market.14 8526.

the relatively low percentage diﬀerences we have shown in this paper still amounts to a substantial amount of money. we found that stochasticity is important for our results. The results from the WF formulation were highly dependent on the chosen weighting in the objective functions. Secondly. In the deterministic case. the distance to the benchmark solution was 9.41% for the MVF formulation. In this case we only need to assume that the ISO knows the market prices of natural gas.5. An alternative that we have considered in this paper is to maximize value of ﬂow to the market nodes. 173 . We still found that the ﬂexibility in the WF formulation was valuable.6 Conclusions cost structure of the producers in the ﬁelds.65 % for the MSS formulation. and 2.77% for the MSS formulation). Given that the value of the ﬂow in the pipelines in the North-Sea in 2006 was approximately 130 billion NOK. the possibility for the large producers to hold back capacity in the secondary market and strategic behavior in the primary market. The WSS solution indicated a high value of perfect information (social surplus increased with 2.3% lower in the stochastic case.67% for the MSS formulation). inclusion of elastic demand functions in the spot markets for natural gas. The booking rights lead to suboptimal solutions in some of the scenarios when prices are uncertain. The EEV solution illustrated that there was a value of solving the stochastic problem (social surplus increased with 1. In this article we have set the ﬁxed capacities such that the total throughput of the system is maximized. Possible future extensions of the model are other market clearing mechanisms in the secondary market. Finally we found that modelling the pressure constraints in the network is important.6% lower than the social surplus in the MSS formulation for the deterministic case. The social surplus for the MVF formulation was 8. we found that the WF formulation gave an increase of 2. In our example.

A The equilibrium conditions In this section we give the equilibrium conditions for our model.. The matching of shadow prices with constraints can also be seen from the KKT-conditions. The complementarity condition states that either G(x) − a or must be equal to zero. The large producers The KKT-conditions for producer l is found through the Lagrangian function: 174 . Appendix 5. Shadow prices for constraints are introduced directly in the Lagrangian function.. we use the following notation for the complementarity condition with the belonging constraint: G(x) − a ≤ 0 ⊥ ≥ 0.Chapter 5 Capacity booking in a Transportation Network. For the shadow prices that are restricted in sign. We distinguish two types of shadow prices: those that are unrestricted in sign (URS) and those that are restricted in sign.

5.A The equilibrium conditions

Ll = −

g∈G m∈M

**Tgm blgm + γlgm (Blgm − blgm ) φs
**

s∈S

+ +

s∈S

**(pms qlms + Plm vlms )
**

m∈M

hlgms pms − cg

m ∈M

φs φs

s∈S g∈G g∈G m∈M

zgm s +

l ∈L

hl gm s

−

**Clg (dlgs ) (blgm − hlgms ) − dlgs
**

m∈M

+

s∈S

**φs µ1lgs φs µ2lms
**

s∈S g∈G

+ +

s∈S

(blgm − hlgms ) − qlms − vlms

**φs [αlgms (blgm − hlgms )] φs τgms
**

s∈S

+

zgms +

l∈L

hlgms −

j∈O( g)

m∈M

+

s∈S

φs ugs

l∈L

blgm + zgms

fgjs fjms

+

s∈S

φs ums

g∈G l∈L

blgm + zgms pm − cg

m ∈M

−

j∈I(m)

+

s∈S

φs χgms

zgm s +

l∈L

hlgm s

.

Finding the derivative of the Lagrangian function with respect to the decision variables we get the KKT-conditions for optimality:

175

Chapter 5 Capacity booking in a Transportation Network...

**∂Ll = −Tgm − γlgm ∂blgm +
**

s∈S

φs (µ1lgs + µ2lms + αlgms + ugs + ums ) ≤ 0 ⊥ blgm ≥ 0,

(5.37) (5.38) (5.39) (5.40)

∂Ll = Blgm − blgm ≥ 0 ⊥ γlgm ≥ 0, ∂γlgm ∂Ll = pms − µ2lms ≤ 0 ⊥ qlms ≥ 0, ∂qlms ∂Ll ∂Clg =− − µ1lgs ≤ 0 ⊥ dlgs ≥ 0, ∂dlgs ∂dlgs ∂Ll = pms − cg zgm s − cg ∂hlgms

m ∈M

hl gm s − cg

m ∈M

hlgm s (5.41) (5.42) (5.43) (5.44) (5.45) (5.46) (5.47) (5.48)

l ∈L m ∈M

− cg

m ∈M

**χgm s − µ1lgs − µ2lms − αlgms + τgms = 0, hlgms URS , (blgm − hlgms ) − dlgs = 0, µ1lgs URS ,
**

m∈M

∂Ll = ∂µ1lgs

∂Ll = ∂µ2lms

**(blgm − hlgms ) − qlms − vlms = 0, µ2lms URS ,
**

g∈G

**∂Ll = blgm − hlgms ≥ 0, ⊥ αlgms ≥ 0, ∂αlgms ∂Ll = zgms + hlgms ≥ 0, ⊥ τgms ≥ 0, ∂τgms
**

l∈L

∂L = ∂ugs

blgm + zgms

m∈M l∈L

−

j∈O(g)

fgjs = 0, ugs URS , fjms = 0, ums URS ,

∂L = ∂ums

blgm + zgms

g∈G l∈L

−

j∈I(m)

∂L = pm − cg ∂χgms

zgm s +

m ∈M l∈L

hlgm s

≥ 0, ⊥ χgms ≥ 0.

**The network operator
**

For the network operator, we present the KKT-conditions for the three diﬀerent objective function alternatives. First the maximize ﬂow objective.

176

5.A The equilibrium conditions Maximize ﬂow The Lagrangian function for the system operator can be formulated as 1 :

Ls =

s∈S

1 2 fims + ηnils Knil rns − Knil ris − fnis

φs

m∈M i∈I(m)

** blgm + zgms
**

l∈L

+

s∈S

φs ugs

m∈M

−

j∈O( g)

fgjs hlgm s

l∈L

+

s∈S

**φs χgms φs ujs
**

s∈S

pm − cg

m ∈M

zgm s + fgjs

g∈I(j)

+

fjms −

m∈O(j)

−

j∈I(m)

+

s∈S

φs ums

g∈G l∈L

blgm + zgms

fjms

+

s∈S

**φs ω1ns Rn − rns + ω2ns (rns − Rn ) φs τgms
**

s∈S

+

zgms +

l∈L

hlgms

.

KKT-conditions The KKT-conditions:

1 We

1 2 have simpliﬁed the Weymouth equation such that Knil and Knil represents the constants in the expression

177

∂L = ∂rms ∂L = ∂rjs + − j∈I(m) l∈L 2 ηjmls rms Kjml − ω1ms + ω2ms ≤ 0 ⊥ rms ≥ 0.49) (5. j∈O(g) (5. (5.50) (5. ∂fgjs ∂L = 1 − ηjms + ujs − ums ≤ 0 ⊥ fjms ≥ 0.58) (5.51) (5. ums URS . ∂ω1ns ∂L = rns − Rn ≥ 0 ⊥ ω2ns ≥ 0. ∂τgms l∈L 178 . ∂ηnis ∂Ll = zgms + hlgms ≥ 0.60) (5.57) (5.56) (5. ⊥τgms ≥ 0..54) (5.. ⊥ χgms ≥ 0 ∂L = Rn − rns ≥ 0 ⊥ ω1ns ≥ 0.52) ∂L = ∂rgs 1 ηgjls rgs Kgjl j∈O(g) l∈L − ω1gs + ω2gs ≤ 0 ⊥ rgs ≥ 0. ∂L = −ηgjs − ugs − ujs ≤ 0 ⊥ fgjs ≥ 0.Chapter 5 Capacity booking in a Transportation Network. ujs URS . fjms = 0. ∂ω2ns ∂L 2 2 = Kni rns − ris − fnis ≥ 0 ηnis ≥ 0.53) − g∈I(j) l∈L 1 ηgjls rjs Kgjl 2 ηjmls rjs Kjml m∈O(j) l∈L − ω1js + ω2js ≤ 0 ⊥ rjs ≥ 0. ∂zgms m inM (5. fgjs = 0. j∈I(m) ∂L = ∂ums blgm + zgms g∈G l∈L − ∂L = pm − cg ∂χgms zgm s + m ∈M l∈L hlgm s ≥ 0.62) ∂L = ∂ugs ∂L = ∂ujs blgm + zgms m∈M l∈L − fjms − m∈O(j) g∈I(j) fgjs = 0. ugs URS .61) (5.55) (5.59) (5. ∂fjms ∂L = −cg χgm s + ugs + ums + τgms ≤ 0 ⊥ zgms ≥ 0.

KKT-conditions The KKT-conditions: 179 .5.A The equilibrium conditions Maximize value The Lagrangian function for the system operator can be formulated as: L= s∈S pms fims − l∈L φs m∈M i∈I(m) vlms 1 2 + ηnils Knil rns − Knil ris − fnis − j∈O( g) + s∈S φs ugs m∈M l∈L blgm + zgms pm − cg m ∈M fgjs hlgm s l∈L + s∈S φs χgms φs ujs s∈S zgm s + fgjs g∈I(j) + fjms − m∈O(j) − j∈I(m) + s∈S φs ums g∈G l∈L blgm + zgms fjms + s∈S φs ω1ns Rn − rns + ω2ns (rns − Rn ) φs τgms s∈S + zgms + l∈L hlgms .

∂L = ∂rms ∂L = ∂rjs + − j∈I(m) l∈L 2 ηjmls rms Kjml − ω1ms + ω2ms ≤ 0 ⊥ rms ≥ 0.75) (5. j∈O(g) (5.. ⊥ rjs ≥ 0. ⊥ χgms ≥ 0 ∂L = Rn − rns ≥ 0 ⊥ ω1ns ≥ 0. j∈I(m) ∂L = ∂ums blgm + zgms g∈G l∈L − ∂L = pm − cg ∂χgms zgm s + m ∈M l∈L hlgm s ≥ 0. (5.70) (5.73) (5.Chapter 5 Capacity booking in a Transportation Network. fgjs = 0 ugs URS . ∂τgms l∈L 180 .72) (5. ⊥τgms ≥ 0. ∂L = −ηgjs − ugs − ujs ≤ 0 ⊥ fgjs ≥ 0.76) ∂L = ∂ugs ∂L = ∂ujs blgm + zgms m∈M l∈L − fjms − m∈O(j) g∈I(j) fgjs = 0. ∂fjms ∂L = −cg χgms + ugs + ums + τgms ≤ 0 ⊥ zgms ≥ 0. ∂ηnis ∂Ll = zgms + hlgms ≥ 0.67) − g∈I(j) l∈L 1 ηgjls rjs Kgjl 2 ηjmls rjs Kjml m∈O(j) l∈L − ω1js + ω2js ≤ 0. ∂ω2ns ∂L 2 2 = Kni rns − ris − fnis ≥ 0 ηnis ≥ 0.63) (5.66) ∂L = ∂rgs 1 ηgjls rgs Kgjl j∈O(g) l∈L − ω1gs + ω2gs ≤ 0 ⊥ rgs ≥ 0.69) (5. ∂fgjs ∂L = pms − ηjms + ujs − ums ≤ 0 ⊥ fjms ≥ 0. ∂ω1ns ∂L = rns − Rn ≥ 0 ⊥ ω2ns ≥ 0.64) (5.. fjms = 0 ums URS .74) (5.65) (5. ∂zgms m ∈M (5.68) (5. ujs URS .71) (5.

A The equilibrium conditions Maximize social surplus The Lagrangian function for the system operator can be formulated as: L= s∈S pms j∈O(g) φs m∈M i∈I(m) fims − l∈L vlms + m∈M l∈L Plm vlms − s∈S φs g∈G 1 MC g 2 2 fgjs + s∈S 1 2 φs ηnils Knil rns − Knil ris − fnis blgm + zgms l∈L − j∈O + s∈S φs ugs m∈M (g)fgjs hlgm s l∈L + s∈S φs χgms φs ujs s∈S pm − cg m ∈M zgm s + fgjs g∈I(j) + fjms − m∈O(j) − j∈I(m) + s∈S φs ums g∈G l∈L blgm + zgms fjms + s∈S φs ω1ns Rn − rns + ω2ns (rns − Rn ) φs τgms s∈S + zgms + l∈L hlgms .5. 181 .

∂ηnis ∂Ll = zgms + hlgms ≥ 0.85) (5. fgjs = 0 ugs URS . ⊥ rjs ≥ 0. ∂zgms m ∈M ∂L = ∂rgs 1 ηgjls rgs Kgjl j∈O(g) l∈L − ω1gs + ω2gs ≤ 0 ⊥ rgs ≥ 0.80) ∂L = pms − ηjms + ujs − ums + τgms ≤ 0 ⊥ fjms ≥ 0.79) (5.82) (5.84) (5. ujs URS .86) (5.. ⊥τgms ≥ 0. ⊥ χgms ≥ 0 ∂L = Rn − rns ≥ 0 ⊥ ω1ns ≥ 0. fjms = 0 ums URS .78) (5. ∂L = −MC g ∂fgjs fgjs + ηgjs − ugs − ujs ≤ 0 ⊥ fgjs ≥ 0. (5. ∂ω2ns ∂L 2 2 = Kni rns − ris − fnis ≥ 0 ηnis ≥ 0.89) (5.77) (5.87) (5.Chapter 5 Capacity booking in a Transportation Network. j∈I(m) ∂L = ∂ums blgm + zgms g∈G l∈L − ∂L = pm − cg ∂χgms zgm s + m ∈M l∈L hlgm s ≥ 0. j ∈O(g) (5. ∂τgms l∈L 182 . ∂L = ∂rms ∂L = ∂rjs + − j∈I(m) l∈L 2 ηjmls rms Kjml − ω1ms + ω2ms ≤ 0 ⊥ rms ≥ 0. j∈O(g) (5.81) − g∈I(j) l∈L 1 ηgjls rjs Kgjl 2 ηjmls rjs Kjml m∈O(j) l∈L − ω1js + ω2js ≤ 0.83) (5. ∂ω1ns ∂L = rns − Rn ≥ 0 ⊥ ω2ns ≥ 0..88) (5.90) ∂L = ∂ugs ∂L = ∂ujs blgm + zgms m∈M l∈L − fjms − m∈O(j) g∈I(j) fgjs = 0. ∂fjms ∂L = −cg χgms + ugs + ums ≤ 0 ⊥ zgms ≥ 0.

Debreu. J. D. Norway. G. (1974). Mathematics of Operations Research 7. & Kiet. A. R. 183 . C. 211–222. & Debreu. (1997). European Union (1998). C. S. G. (1982). SIAM Journal on Control and Optimization 12. ‘The PATH solver: A non-monotone stabilization scheme for mixed complementarity problems’. (1952). K. J. Bensoussan. Europa-programmet. in ‘Proceedings of the National Academy of Sciences’. ‘Points de nash dans le cas de fonctionnelles quadratiques et jeux diﬀdrentiels lindaires a N personnes’. S. ‘Directive 98/30/EC of the european parliament and of the council’.Bibliography Arrow. (1995). J. ‘A large-scale complementarity model of the North American natural gas market’. Finite-Dimensional Variational Inequalities and Complementarity Problems: Volume I. New York. Ferris. & Pang. M. Norwegian Natural Gas. O. European Union (2003). P. (2003). ‘Engineering and economic applications of complementarity problems’. A social equilibrium existence theorem. J. Optimization Methods and Software 5. J. Liberalization of the European Gas Market. (2005). Birge. (2003). 639– 665. S. S. A. G. Econometrica 22. F.. Oslo. ‘Directive 2003/55/EC of the european parliament and of the council’. F. Springer series of operations research. ‘The generalized quasi-variational inequality problem’. & Loveaux. Introduction to Stochastic Programming. Springer. J. M. ‘Existence of an equilibrium for a competitive economy’. & Pang. Chan. 669–713. Zhuang. Facchinei. V. Austvik. (1997). 265–291. SIREV 39. Energy Economics 27. & Pang. (1954). & Ferris. Dirkse. Gabriel. 123–156.

(1991). Ralph. ‘Variational inequalities’. T.. Operations Research 47(1). Z. P. Mathematical Programming 48. J. Academic Press. & Pang. & Stampacchia. Midthun. & Smeers. X. lchiishi. 102–112. Electricity generation with looped transmission networks: Bidding to an iso. (1967). W. A. Lions. B. G. 65. Hu. (2006). Norway. 197–204. P.. T. Ralph. 493–519. New York. C. NTNU. P. T. & Wallace. Submitted to international journal. T. J. (2005). E. S. (1983). M. 184 . Trondheim. (1960). (1990). T. P. Hobbs. J. CRC Press. Harker. S. (1999). Chichester. IEEE Transaction on Power Systems 16(2). Stochastic Programming. Luo. Kall. K. S. & Ralph. Mathematical programs with equilibrium constraints. 81–94. E. Bjørndal. & Ferris.-Q. ‘Generalized nash games and quasivariational inequalities’. 799–818. ‘Linear complementarity models of nash cournot competition in bilateral and poolco power markets’. 413–423. S. S. F. & Howson. (2005). Bardsley. Working paper. Harker.. K. & Tomasgard. Management Science 6. Operations Research 53(5). Jing-Yuan. 161–220. ‘Spatial oligopolistic electricity models with cournot generators and regulated transmission prices’. ‘Equilibrium points of bimatrix games’. (1964). W. Menon. Y. algorithms and applications’. J. CMI EP Working Paper No. Communications on Pure and Applied Mathematics 20.. John Wiley & Sons. J. A. (1996). & Zhuang. Pang. Game Theory for Economic Analysis. D. Kiet. (2004). Lemke.com/abstract=556809. D.Bibliography Gabriel. Journal of the Society for Industrial and Applied Mathematics 12. (2001). ‘A mixed complementarity-based equilibrium model of natural gas markets’. European Journal of Operational Research 54. Gas pipeline hydraulics. Modeling optimal economic dispatch and ﬂow externalities in natural gas networks. ‘Finite-dimensional variational inequality and nonlinear complementarity problems: a survey of theory. ‘Inequalities for stochastic linear programming problems’.-S. M. J. Boca Raton. (1994). C. E. Available at SSRN: http://ssrn.. Cambridge University Press. USA.. Madansky. Technical report.

. Pang. Modeling and computing two-settlement oligopolistic equilibrium in a congested electricity network. Quak. (2007). (2003b). Equilibrium points in n-person games. (1950).-A. Rosen. (2006). Oren. USA. Tomasgard. L.. F. Y. Working paper. Operations Research Letters 25. S. 520–534. Patriksson. & Fukushima. (2005). S. M. J. and multi-leader-follower games’. J. eds. Fodstad. 185 . Y. K. Smeers. S. Computing cournot equilibria in two settlement electricity markets with transmission constraints. 175–196. S. in G. Part II: The forward and real time markets’. Econometrica 33. (2003a). B. ‘Existence and uniqueness of equilibrium points for concave n-person games’. (1999). M. Lie & E. Numerical Simulation. Networks and Spatial Economics 3. M. 159–167.. Optimization models for the natural gas value chain. Springer. UCEI. 151–174. Adler. K.Bibliography Nash. Part I: The forward market’. J. (2006). University of California. & Oren.. (2004). & Midthun. (1965). Cournot Equilibrium in Two-Settlement Electricity Markets: Formulation and Computation. generalized nash equilibria. & Wynter. Networks and Spatial Economics 3. & Adler. Yao. ‘Geometric Modelling. J. Computational management science 2. Rømo. PhD thesis. I. Yao. ‘Market incompleteness in regional electricity transmission. I. J. Berkeley. Smeers. ‘Market incompleteness in regional electricity transmission. in ‘Proceedings of the national academy of sciences’. in ‘Proceedings of the 37th Hawaii International Conference on System Sciences’. ‘Quasi-variational inequalities. ‘Stochastic mathematical programs with equilibrium constraints’. A. Yao. chapter Optimization Models for the Natural Gas Value Chain. Hasle. and Optimization: Applied Mathematics at SINTEF’.

- Ordinary Differential Equations (Arnold)by peterherbert
- Basic Sedimentary and Petroleum Geologyby Esther Anabu Ezima
- Energy: Natural Gas: The Production and Use of Natural Gas, Natural Gas Imports and Exports, EPAct Project, Liquefied Natural Gas (LNG) Import Terminals and Infrastructure Security, Underground Working Gas Storage, Fischer-Tropsch Fuels from Coal, Natural Gas, and Biomassby TheCapitol.Net
- Jem Walsh Webby Jother Amutan

- All Natural Gas Slides
- 1.Petroleum Geology 1
- Israel Natural Gas Demand Forecast 2014-2040
- Petroleum Geology
- Petroleum Geology
- Basics of Petroleum Geology
- Evaluation of Pipeline Failure Causes
- 01-Geology of Petroleum Systems
- Ordinary Differential Equations (Arnold)
- Basic Sedimentary and Petroleum Geology
- Energy
- Jem Walsh Web
- National Emission Standards for Hazardous Air Pollutants for Source Categoriesn Standards for Hazardous Air Pollutants for Source Categories
- Tut Unit 3.1 q and A
- Jerin Project
- for use
- 2004_ECE_Q3
- Formation Rate of Natural Gas Hydrate (Reactors Experiments and Models).pdf
- cv
- City Gas Distribution Report
- Up in flames
- Oil and Gas Global Market Press Release Report
- Natural Gas Cleaning and Treatment
- 12 Williams
- Processing
- Natural Gas Pipeline Planning & Economics(1)
- 13-H2O CONT OF GASES_33s.pdf
- Natural_Gas_360_Press_Release copy.pdf
- Pages From Chapter 10-407d050ddc413b94f1d41e8f8cdb7b8e
- TheNaturalGasIndustry-102605-1
- Optimization Models for Liberalized Natural Gas Markets

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd