Sunteți pe pagina 1din 143

2/29/2020 Untitled

Page 1

Page 2

Foreword to the 2nd edition

Typing errors have been corrected in this new edition. The layout was
slightly revised. In addition, on pages 38, 53, 99 and 108 respectively
corrected the symbolism.

Except for the changes mentioned, the 1st and 2nd editions are identical in content
table.

Duisburg, in 2011
Sebastian Kuhn

https://translate.googleusercontent.com/translate_f 1/143
2/29/2020 Untitled

Page 3

Optimizing the operation of electrical


Power generation plants and
Transmission systems for incomplete

information

- 2nd Edition -

The dissertation published in 2008 was


from the Department of Mathematics
the University of Duisburg-Essen
(Duisburg campus)

to attain the academic degree of


https://translate.googleusercontent.com/translate_f 2/143
2/29/2020 Untitled

Doctors of science
- Dr. rer. nat. -
approved.

Page 4

Bibliographic i information i the i German National Library


i
The i Deutsche i Nationalbibliothek i lists i this i publication i in i der i Deutsche
:
Nationalbibliografie; i detailed i bibliographic i data i are i on i internet i uber
http://dnb.d-nb.de i available.
:
Second
i ed. i - i Gottingen: i Cuvillier, i 2011
Approved: i Duisburg-Essen, i Univ., I Diss., I 2008

978-3-86727-775-4

https://translate.googleusercontent.com/translate_f 3/143
2/29/2020 Untitled

:
© CUVILLIER i VERLAG, i Gottingen i 2011
:
Nonnenstieg i 8, i 37075 i Gottingen
Telephone: i 0551-54724-0
Fax : i 0551-54724-21
www.cuvillier.de

:
All i rights i reserved. i Without i express i approval
of the i publisher i is i it i not i permitted i the i book i or i parts

from it i on i photomechanical i way i (photocopy, i microcopy)


to multiply i . :
Second
i edition, i 20 11
:
Printed i on i acid-free i paper

978-3-86727-775-4

Page 5

thanksgiving

The dissertation published in 2008 is part of my work as

scientific assistant of the working group Discrete math and op


https://translate.googleusercontent.com/translate_f 4/143
2/29/2020 Untitled

timierung ”was created at the University of Duisburg-Essen. My"very special one


Thanks to Prof. Dr. rer. nat. Rüdiger Schultz, who is not just my collaboration
on an extremely interesting project, but on its support
I could count at any time. Last but not least, his suggestions helped
this work to success.

Many thanks to Dr. rer. nat. Ralf Gollmer, who often gives me advice and
Tat stood by the side. I would also like to contact Dr.-Ing. Hendrik Neumann
thank you for sharing your engineering knowledge with me
supported my work. I also thank Dipl.-Wirt.-Ing. Daniel
Waniek, active at the Chair for Energy Systems and Energy Management

University of Dortmund, and Prof. Dr.-Ing. Edmund Handschin as well


Prof. Dr.-Ing. Christian Rehtanz, also working there, for the common
Cooperation.

Prof. Dr. Marc C. Steinbach I would like to take over the


thank rats. Finally, I would like to join the Federal Ministry of Education
and research for the financial support of the project.

iii

Page 7
6

https://translate.googleusercontent.com/translate_f 5/143
2/29/2020 Untitled

Table of Contents

Summary / Abstract 1

1 Introduction 3

2 Mathematical optimization models 9

2.1 Ignoring the energy transmission system. , , , , , , , , , , , , 10

2.2 Integration of the energy transmission system. , , , , , , , , , , , , 12

2.3 Extension to a period. , , , , , , , , , , , , , , , , , , , 21

2.4 Model relations. , , , , , , , , , , , , , , , , , , , , , , , , , , , 24

2.5 Retrofitting electrical energy transmission systems. , , , , , , , 30

3 solution methods 33

3.1 A classic procedure. , , , , , , , , , , , , , , , , , , , , , , , 34

3.2 An external approximation method. , , , , , , , , , , , , , , , , 36

4 Stochastic optimization 47

4.1 Two-stage stochastic mixed-integer linear models. , , 48

4.2 Two-stage stochastic mixed-integer convex models. , 55

4.3 An outlook on multi-level stochastic models. , , , , , , , , 56

4.4 Applications. , , , , , , , , , , , , , , , , , , , , , , , , , , , , , 58

5 decomposition processes 61

5.1 Mixed-integer linear model classes. , , , , , , , , , , , , 62

5.2 Mixed-integer convex model classes. , , , , , , , , , , , , 66

Page 8

https://translate.googleusercontent.com/translate_f 6/143
2/29/2020 Untitled

6 A heuristic for the AC model 71


6.1 Quality assessments. , , , , , , , , , , , , , , , , , , , , , , , , , , 72

6.2 A transformation approach. , , , , , , , , , , , , , , , , , , , , , , 75


6.3 Heuristic Approaches. , , , , , , , , , , , , , , , , , , , , , , , , , 78

7 Numerical examples 81
7.1 Deterministic preliminary considerations. , , , , , , , , , , , , , , , , , 83

7.2 Stochastic optimization: motivating examples. , , , , , , , , 90


7.3 Efficiency of the decomposition process. , , , , , , , , , , , , , , , 92
7.4 Retrofitting a transmission system. , , , , , , , , , , , , , , 98
7.5 A case study from the energy industry. , , , , , , , , , , , , , , 100

Summary 105

A List of symbols 107

B Data situation: 6-node network 111

List of Tables 114

List of Figures 115

bibliography 117

vi

https://translate.googleusercontent.com/translate_f 7/143
2/29/2020 Untitled

Page 9

short version
Mathematical modeling and analysis of electrical energy transmission
systems with decentralized generation is due to the increasing importance
on renewable energy sources, such as wind or solar energy. One use
Such systems require not only the control of the power generation plants
also that of energy flows, taking into account uncertainties. It
in mathematical terms, combinations are combinatorial
and stochastic effects including networks that are too highly

gradual non-linear structures and suitable approximation techniques


require. Different interpretations of the risk, from risk neutral to
towards risk-averse modeling using suitable statistical parameters
meters, are addressed. Decomposition algorithms for two-stage stochastic

Tables mixed-integer models are presented, with the help of which it is possible
is, both linear and nonlinear approximations of the energy flows
adequate modeling of randomness. Numerous sample calculations
are shown.

Abstract
Increased importance of renewable energy such as wind or solar energy has moti-

vated mathematical modeling and analysis of electricity networks with dispersed


generation. The operation of these systems requires the simultaneous handling of
unit commitment and powerflow considering uncertainties. Mathematically, this
leads to a combination of combinatorial, network, and stochastic phenomena. search
models involving power flow lead to highly nonlinear structures and require ade-
quate model simplifications. Attention is paid to different perceptions of risk, from
risk neutral to risk averse models incorporating different statistical parameters
for risk measurement. Algorithms for two-stage stochastic mixed-integer programs

are presented to handle linear and nonlinear approximations of powerflow cons-


traints with adequate modeling of uncertainty. Numerous numerical examples are
presented.
https://translate.googleusercontent.com/translate_f 8/143
2/29/2020 Untitled

Page 11
10

Chapter 1

introduction

Through the increased use of renewable energy sources to generate electricity


have the uncertainties regarding the expected feed-in and purchase
increased enormously in electrical energy transmission systems. So is subject to

the wind feed stochastic fluctuations, which can only be


limited accuracy can be predicted. Decentralized power generation plants, such as
Combined heat and power plants or fuel cells are primarily heat-managed today
hazards. As a result, their electricity generation is primarily based on that

uncertain weather forecast. Should a generating plant operate at optimal cost


optimization does not only require information about the
expected development in terms of electricity and heat requirements, but also
Information about primary energy prices. These are also based on
Forecasts and are therefore fraught with uncertainties. This is reinforced
Electricity trading situation with limited predictable effects
on feed-in and acceptance in the energy transmission system. Thus it results
the question of how to control the decentralized generation plants quickly
https://translate.googleusercontent.com/translate_f 9/143
2/29/2020 Untitled

are, so that on the one hand a cost-optimal operation can be maintained and
on the other hand, security of supply even if the acceptance is uncertain
is guaranteed. A crucial importance comes in this context
the energy transmission system. This transports the required energy

from producer to consumer and creates interdependencies between


possible productions of the production plants connected to it, so that

Page 12

4 CHAPTER 1. INTRODUCTION

a single decentralized generation plant is not considered an autonomous unit


can be. But an energy transmission system raises new questions
on. A cost-minimal production decision of the generating plants that
the demanded load can surely result in an energy flow
who do not overload the lines of the energy transmission system.

The aim of this work is the development of mathematical models and processes
decentralized for stochastic power flow and operational optimization
Generating plants in the context above. The stochas-

table influences. Mathematical optimization as a sub-area


applied mathematics is a useful tool for numerous practical problems.
ready instruments. It finds its area of application, for example
also in production planning, telecommunications or logistics. So
can help develop appropriate mathematical models and also
their solution to a wide range of modeling techniques and common
Software packages can be used. Also numerous literature sources can be found

are mentioned, which deal with the feed optimization with safe input
deal with data, such as [13, 25, 28, 36, 71].

The work is structured as follows. In Chapter 2, suitable mathematic


presented optimization models according to the concerns outlined above,
https://translate.googleusercontent.com/translate_f 10/143
2/29/2020 Untitled

the different modeling techniques, starting from a complete one


Ignoring the electrical power transmission system, going to Appro-
Ximation possibilities of such up to an exact reproduction, use

and also examined in more detail with regard to relationships that apply to each other
become. Furthermore, an optimization model is presented, with the help of which
an inexpensive retrofit of an electrical energy transmission system
lets determine. Such retrofitting can help in the long term

Increasing security of supply and reducing production costs. The


Access selected in this chapter is purely deterministic.
wise, ie all input data are assumed to be known. Chapter 3
represents solution methods that apply to most of the model classes obtained

Page 13

are applicable. Chapter 4 deals intensively with the integration of with

Data records with uncertainties and provides suitable stochastic opti


Modeling models focusing on the two-stage stochastic approach.
Both risk-neutral and risk-averse approaches with the possibility
The use of various statistical parameters is used. An exact
Modeling the unsafe input data can, however, take the computing time required
enormously when using the algorithms presented in Chapter 3. Specific
Algorithms that take advantage of decomposition techniques and thus become an essential one
Reducing computing times are addressed in Chapter 5. chapter

6 provides in addition to a method used in engineering


Load flow calculation also heuristic approaches for the developed in Chapter 2
and difficult to treat optimization model with an exact modeling
the physical relationships of the energy transmission system, ie without

Use of approximation possibilities, where certain requirements


are to be fulfilled. The work concludes with numerous shown in Chapter 7

https://translate.googleusercontent.com/translate_f 11/143
2/29/2020 Untitled
numerical test calculations.
Below is a brief look at relevant basics of electrical engineering

Given energy supply. For any specializations in this area


at this point reference is also made to [4, 28, 29]. First, the concept of
electrical power transmission explained in more detail in the Internet dictionary
Know-Library [42] can be found. To provide consumers with electrical energy

supply, it is necessary to connect lines from the power generation plants to the
To lay consumers. In Germany, the energy is
three-phase alternating current with a frequency of 50 Hz and a voltage of
transmit up to 400kV. Only shortly before the consumer does he become familiar

Low voltage of 230V single-phase alternating current or 400V three-phase alternating


selstrom transformed. In other countries there are different tensions or
Frequencies possible. The transmission with high voltage has come from various
reasons because high voltages are technically easier to control

are lower than high currents, a high transmission performance is guaranteed


Transmission losses occur and large distances are bridged

can. Depending on the voltage applied, extra high voltage networks (220kV

Page 14

6 CHAPTER 1. INTRODUCTION

or 380kV), high-voltage networks (50kV to 150kV), medium-voltage networks


(10kV to 30kV) and low-voltage networks (230V or 400V)

become. Fig. 1.1 shows an example of an electrical energy transmission system.


Such a system consists of busbars (nodes), transmission lines

energy generation
purification system

https://translate.googleusercontent.com/translate_f 12/143
2/29/2020 Untitled

bus

b rragnsleitung
Practice
load

Figure 1.1: An electrical power transmission system

(Edges) and transformers. Operators to be stepped on the busbars


Loads when connecting a generating plant can also be an infeed
electrical energy. Usually exist with larger networks

several continuously adjustable generation plants, at least theoretically, infinitely


many permissible feeds into the energy transmission system that need to be operated
Covering loads, but leading to different production costs. This
The situation can, for example, be based on the transmission system from Fig. 1.2

be checked if the feed-in limits of the generators are


are given. From an economic point of view, the determination of a permissible
Production decision with minimal costs (operational optimization), but should
often transmission losses are minimized

Page 15

(Power flow optimization). These can be different types of tasks


act, as the following example shows:

example

Consider an electrical power transmission system with two pure feeders


https://translate.googleusercontent.com/translate_f 13/143
2/29/2020 Untitled

solution points and a load node. The line from node 1 to node 2

D2

Figure 1.2: A 3-node system

be equipped in such a way that with an electrical energy flow marginal losses

occur, whereas a transmission over the line from node 3 to


Node 2 leads to significant losses. One further presupposes that both
Generation plant 1 with the feed variable P G as well as generation plant
1

3 with the feed variable P G 3 alone would be able to handle the load P D 2 to

serve. In addition, the feed causes P G 1 Production costs, one


Feed with P G 3 at no cost, as is the case with wind turbines
the case is (maintenance costs are ignored here). An optimal feed
with minimal network losses, the load P D 2 certainly completely with

PG 1 serve. A cost-minimal feed could only be with sole


Operation by P G available.
3 is

In addition to minimizing transmission losses,

flow optimization also the improvement of voltage profiles and the reduction
of reactive power transport, but the last two tasks are in this
Work not treated. Voltages and currents in an electrical energy

Transmission systems are subject to a continuous-time process. In the


However, models to be presented will be a stationary system representation, ie

Page 16

8th CHAPTER 1. INTRODUCTION

https://translate.googleusercontent.com/translate_f 14/143
2/29/2020 Untitled

a representation at a fixed point in time. In this representation


is, as can be seen from [28, 29], the description of the symmetrically

driven three-phase system possible through a single-phase AC system.


The following models can also be optimized using one
Period, but such is always considered discretized
provided. The following considerations therefore relate exclusively

to a single-phase representation of the current. It is also stated that


here the energy transmission system is viewed as an island network and so
there is no possibility of exporting or importing electrical energy because
only the basic modeling aspects are discussed. In addition

Transformers only implicitly via the improved transmission


performance capabilities of the cables are taken into account. The following chapter presents suitable ones
mathematical optimization models that perform the aforementioned tasks,
initially in a purely deterministic view, ie all input data

will be assumed to be known as fair.

https://translate.googleusercontent.com/translate_f 15/143
2/29/2020 Untitled

Page 17

Chapter 2

mathematical
optimization models

In accordance with the problem outlined in the previous chapter, it now arises
the task of developing suitable mathematical optimization models that
meet the stated goals. When creating such models

the underlying problem can be further specified. The mo


dent accuracy, ie the fidelity of the physical relationships,
is essentially reflected in the complexity of the mathematical description.
This, in turn, can have a significant impact on solution processes

which are actually optimally operated, be it from a minimum cost or loss


Point of view, can be calculated. Therefore, it is not only the model which
reproduces the physical relationships 1: 1, treated. Rather be
also reduced model approaches addressed. The chapter is structured as follows.

First, models that are tasked with determining an optimal


Accept operation, presented. In section 2.1 is a in its presentation
very compact model ignoring the electrical energy transmission

systems and in section 2.2 three models, which take into account the
Enable energy transmission system. The fundamental modeling approaches
Here are the AC representation, an exact, but mathematically demanding
Representation of the physical relationships and approximated representa-

the AC model, which is expressed in the DC and DCOL modeling


https://translate.googleusercontent.com/translate_f 16/143
2/29/2020 Untitled

Page 18

10 CHAPTER 2. MATHEMATICAL OPTIMIZATION MODELS

find, but through their presentation the application of efficient solution methods
allow. The development of these models is initially based on a

It is only in section 2.3 that the expansion takes place over a period of time. section
2.4 is intensely concerned with establishing relations in relation to the
featured models. On the one hand, it is intended to provide a better understanding of the
delle are accommodated and checked on the other hand whether by

the approximate modeling techniques quality estimates for the math


table model difficult to treat - as will be shown - possible
are. Furthermore, a model is developed in section 2.5, with the help of which
cost-effective retrofitting of an electrical energy transmission system

that can contribute to long-term security of supply


increase and lower production costs.

2.1 Ignoring the energy transmission system

A strongly simplifying model is certainly the complete one


Ignoring the electrical power transmission system and an imaginary one
Concentration of feed and load in one network node. Approaches this
Art can also be found in [13, 19, 23, 28, 32, 36, 51, 52, 66, 68]. Part-

Overproduction of electrical energy is also permitted there, if for example


Exports of unnecessary energy exist. Since the electrical energy
The transmission system here as a closed system, ie as an island network
is considered without the possibility of exporting or importing electrical energy,

can the accounting equation, the equation that ensures that


energy is produced as follows:

https://translate.googleusercontent.com/translate_f 17/143
2/29/2020 Untitled
Σ PG -PD=0
k (2.1)
k∈V

The variable sizes P G ∈ R + denote the active power feeds


k

the power generation plant located at a node k ∈ V. The amount


V ⊆ V stands for the feed-in nodes, ie nodes at which a generation
is connected, whereas if there are n n N nodes the

Page 19

2.1. IGNORING THE ENERGY TRANSMISSION SYSTEM 11

Set V = {1, ..., n} the node numbers of the energy transmission system
holds. It is further assumed that every energy transmission system with
a so-called slack knot, on which an extremely powerful
capable generation plant is connected. This has the node number here

1 ∈ V. The use of a slack knot is common and can be


[8, 24, 36, 28, 71]. The total active power load P D is as
Sum of the active power loads P D ∈ R + given at all nodes k ∈ V, ie it
k

applies:

PD= Σ PD k

k∈V

An energy generation plant can even be differentiated from an entire system.


cher units for power generation exist. For example, a block

heating power plant next to the boilers for the production of thermal energy
several electrical energy generating components, such as gas
turbines or internal combustion engines. It should be noted that the generation
is considered here as a unit. For example, is it required that this one

should generate predetermined electrical power, so take the following


Models does not address the task of how the individual components of the generation
are to be controlled. The models are satisfied with the information whether
such a system is capable of this. So there are further restrictions

to be held up by the technical conditions of the power generation plants

https://translate.googleusercontent.com/translate_f 18/143
2/29/2020 Untitled
are set. The active power feed-in P G k ∈ R + may only be within
fixed limits P min G , Pk
max
G

k
Move R + . Together with the current switching

state s k ∈ {0,1} of the generating plant k ∈ V, which provides information as to whether

the respective power generation plant is switched on or off, the restrictions


tions
P Gmin · S k ≤ P G ≤ P max
k
k
Gk
·Sk ∀k∈V (2.2)

to be observed, which regulate the possible feed-in values. The optimization criteria
With this modeling approach, only the minimization of product
measured since the network is ignored. A power flow optimization
from a minimum loss perspective (losses occur when transporting energy
Transmission lines to) is not possible. The optimization criterion can be so

about

Page 20

12 CHAPTER 2. MATHEMATICAL OPTIMIZATION MODELS

Σ C k (P G , s k ) → min
k (2.3)
k∈V

can be scheduled. The functions C k for k ∈ V describe the combustion


material costs for a production for which a linear representation is assumed
becomes, ie C k (P G , s k ) = α k P G + β k s k with suitable constants α k , β k ∈ R + .
k k

These are to be adapted to the duration of the time step to be optimized. often

the cost functions have nonlinear, but convex curves. For


In this case, linearization techniques can also be used that the
Approximate the convex shape using a piecewise linear function. such
Modeling techniques can also be found in [38]. You can also

Start-up costs of the power generation plants, fixed costs that arise,
when a generating plant from the state of non-production to the state
the production changes when considering a period. such
Modeling is done in section 2.3.

The disadvantage of the modeling approach described here is obvious. By


https://translate.googleusercontent.com/translate_f 19/143
2/29/2020 Untitled

Concentration of load and feed in a node can neither be maximum


permissible transmission capacities or transmission costs are taken into account.

The test calculations presented in Chapter 7 also show that the


Type of modeling obtained optimal feeds the lines of the
massively overload electrical power transmission system. Furthermore,
the model cannot be used to optimize power flow. Still delivers

this approach first estimates of the optimal solution with regard to a


paint feed and is due to the cheap mixed-integer linear
Model description widely used. The model presented here is as follows
denoted by IG model (ignoring the energy transmission system).

2.2 Integration of the energy transmission system

An integration of the energy transmission system enables the consideration of the


actually existing energy flows on the respective lines. hereby

these can be limited and thus also the maximum possible transmission

Page 21

2.2. INTEGRATION OF THE ENERGY TRANSMISSION SYSTEM 13

Pay attention to the capacities, but also possibly existing transmission costs
involved in the optimization phase. One initially associates with the

Energy transfer system an undirected graph G = (V, E). As before


V denotes the amount of network nodes and V ⊆ V the amount of feed-in
node. The set E ⊂ V × V contains the edges (lines) of the graph
G. The set E ⊂ E provides information on which edges on one

Energy transportation incur additional costs. It should be noted that


for an edge kl ∈ E also lk ∈ E applies and this is also valid for E. This in
Basically redundant description enables an extremely compact notation of the
physical restrictions. Furthermore, N k ⊂ V denotes the set of k ∈ V

https://translate.googleusercontent.com/translate_f 20/143
2/29/2020 Untitled
adjacent nodes, ie there is N k = {l ∈ V: ∃ k ∈ V with kl ∈ E}.

An integration of the energy transmission system can be done via AC modeling


(alternating current), DC modeling (direct current) or DCOL

Modeling (direct current ohmic losses) take place. Before making a distinction
these approaches, as made in the following considerations
the basic equations should be given, which is a quick one
Enable the realization of these models (cf. [4]). Look at everyone

Network node k ∈ V existing complex power S k = P k + jQ k . Inscribed


P k ∈ R the active power and Q k ∈ R the reactive power at the node k ∈ V. The
complex accomplishments can in turn use the two equations

Pk = Σ p kl ∀ k ∈ V, (2.4)
l∈N k

Qk = Σ q kl ∀ k ∈ V (2.5)
l∈N k

be expressed. Here p kl ∈ R denotes the active power flow and q kl ∈ R


the reactive power flow on one edge kl ∈ E. The complex power S k can

now as the difference from complex feed S G k = P G + jQ G


k k and more complex
Load S D = P D + jQ D can be expressed. Thus one obtains from (2.4), (2.5) the
k k k

equations:

Page 22

14 CHAPTER 2. MATHEMATICAL OPTIMIZATION MODELS

PG -PD
k k = Σ p small ∀ k ∈ V (2.6)
l∈N k

QG -QD
k k = Σ q small ∀ k ∈ V (2.7)
l∈N k

−P D k = Σ p small ∀ k ∈ V \ V (2.8)
l∈N k

https://translate.googleusercontent.com/translate_f 21/143
2/29/2020 Untitled

−Q D k = Σ q small ∀ k ∈ V \ V (2.9)
l∈N k

The type of physical representation of the energy flows p kl and q kl now results
in the AC, DC and DCOL models. The restrictions should be listed beforehand

that all approaches have in common. For the purpose of displaying the energy flows
the variable voltage angle θ k ∈ R is required at each node k ∈ V. At the
Slack node 1 ∈ V becomes additional

θ1=0 (2.10)

required (see [4, 8, 28]). To the minimum and maximum possible feed
To comply with the limits of the power generation plants, the restrictions can
(2.2) are adopted. Three subsections are considered below,
in which according to the type of physical modeling of the energy flows

will be divorced.

AC modeling

AC modeling or alternating current modeling enables an exact

te representation of the physical relationships in an electrical energy

transmission system, but also results in a


full mathematical wording. Similar modeling approaches can be found
also in the works of [1, 40, 41, 49]. The representation of the energy
flow is mainly based on [4]. To the energy flows in AC
To be able to express modeling along the edges kl ∈ E are first
at all nodes k ∈ V the complex voltages U k e jθ k ∈ C with the Varia

blen U k ∈ R + , which describe the amounts of the complex tensions, and the

Page 23

2.2. INTEGRATION OF THE ENERGY TRANSMISSION SYSTEM 15

https://translate.googleusercontent.com/translate_f 22/143
2/29/2020 Untitled

Variables θ k ∈ R, which express the stress angle , were introduced. Further designated
the variables θ kl ∈ R denote the voltage angle differences between the nodes
k, l ∈ V, ie θ kl = θ k - θ l . Without going into an exact derivation here
wanting, the energy flows can be left with these newly introduced variables

p small = U 2 ∀ kl ∈ E, (2.11)
k g kl - U k U l g kl cosθ kl - U k U l b kl sinθ kl
q small = U k U l b kl cosθ kl - U k U l g kl sinθ kl - U 2 (2.12)
k (b kl + b 0 kl ) ∀ kl ∈ E

express. The specified conductances for all edges kl ∈ E


g kl ∈ R + , susceptances b kl ∈ R - and shunt resistances b 0
kl ∈ R+
give the transferability of an edge kl ∈ E. Here too

the transformers are taken into account implicitly, since their presence means that the
sustainability is improved and accordingly the aforementioned
parameters can be adjusted. Of course, an edge must not be of any thickness
be loaded so that the power flow with the maximum possible transfer

capacity d max kl
∈ R + over

p2 )2 ∀ kl ∈ E (2.13)
kl + q 2 kl ≤ (d max kl

is limited. In addition to an active power feed-in, the energy


systems with an AC model also a reactive power feed-in
solution. The reactive power feed Q G ∈ R for k ∈ V also applies
k

limit, it can here, however, to be independent of the switching state s k mo-


be dellated, ie it is constant from zero to its feed-in limits
controllable. Numerous literature sources support this modeling approach, such as
about [4, 8, 28]. It is also pointed out that reactive power feed-in
causes no production costs. So the restrictions apply

Q Gmin ≤ Q G ≤ Q max G
k
k
k
∀k∈V (2.14)

observed. The voltage amounts U k may also be applied to the generator and

Load nodes only move in certain areas and are over

U kmin ≤ U k ≤ U maxk ∀ k ∈ V \ {1} (2.15)

https://translate.googleusercontent.com/translate_f 23/143
2/29/2020 Untitled
Page 24

16 CHAPTER 2. MATHEMATICAL OPTIMIZATION MODELS

to limit. The amount of tension is fixed at the reference node itself, like this
can also be taken from [4, 28], and as a rule fixed at U 1 = 1.

The fundamental physical relationships in an energy transmission


systems are modeled by the considerations above. Certainly,
dell formations turn out to be even more comprehensive by adding further technical aspects,
such as transformers with variable gear ratios

the. The modeling shown here aims at the mathematical representation of the
basic properties of an electrical energy transmission system.
If operational optimization is selected as the optimization criterion, then
(2.3) can be used. In addition, due to the integration of the energy

transmission system, the possibility of considering transmission costs.


The optimization task for an operational optimization (now: minimization of
Production and transmission costs) can then be assessed as follows:

Σ C k (P G , s k ) + Σ
k E AC (2.16)
kl (p kl , q kl ) → min
k∈V kl∈E

The functions describe E AC kl the transmission costs incurred


an energy transport that depends on the amount of complex power
are flow, ie they are E AC
(p kl , q kl ) = δ kl √p 2
kl kl + q 2 kl with suitable constants

δ kl ∈ R + for kl ∈ E. On the other hand, should power flow optimization be based on loss
nimal view, the task arises according to [28]:

( )2 ( )2
Σ PG -PD
k + Σ QG -QD k → min (2.17)
k∈V k∈V

The AC model lies in a very demanding


thematic description. In addition to integer requirements, tri-
gonometric expressions and variable products in the restrictions. On
the problems that arise with regard to solution methods for models in such
Chapter 3 is dealt with separately. The demanding mo
dellbeschreibung also suggests model reduction approaches, which are in the two
the following sections.

https://translate.googleusercontent.com/translate_f 24/143
2/29/2020 Untitled

Page 25

2.2. INTEGRATION OF THE ENERGY TRANSMISSION SYSTEM 17

DC modeling

The DC modeling or also DC modeling represents a linearized one


Form of the AC load flows and is also used in [10, 18, 24]. Here
the following assumptions made in high-voltage networks:

(i) θ kl ≈ 0 ∀ ∈ E kl

(ii) U k = 1 ∀ k ∈ V

(iii) Neglecting reactive power components

With the assumption (i), the linearizations are cosθ kl ≈ 1 and sinθ kl ≈ θ kl for all
kl ∈ E possible. Together with assumption (ii), the active power flows p kl are over

p kl = b kl (θ l - θ k ) ∀ kl ∈ E (2.18)

given. These are now available in a much cheaper linear representation.


Here b kl = - 1 x small
, where x kl ∈ R + for the longitudinal reactance on the line kl ∈
E stands. Based on assumption (iii) the reactive power flows and thus
Equations (2.7) and (2.9) are also neglected. The relevant restrictions
in DC modeling, (2.2), (2.6), (2.8), (2.10) and (2.18) are. The
Transmission limits on the lines kl ∈ E can, again with the

maximum possible capacity d max kl


∈ R + , about the restrictions

−d max
kl ≤ p kl ≤ d maxkl ∀ kl ∈ E (2.19)

will be realized. In order to also be able to consider transmission costs, the


Determine the amount of the active power flow so that a mixed
integer linear model description can be maintained. According to this
The binary variable ˜p kl ∈ {0,1} is required for each line kl ∈ E.
Then p kl : = | p kl | with a suitable constant M> 2 · max kl∈E d max kl

https://translate.googleusercontent.com/translate_f 25/143
2/29/2020 Untitled
formulate as follows:
p small ≥ p small (2.20)

p small ≥ −p small (2.21)

p kl - p kl ≤ M · klp kl (2.22)

p kl + p kl ≤ M · (1 - ˜p kl ) (2.23)

Page 26

18 CHAPTER 2. MATHEMATICAL OPTIMIZATION MODELS

With the help of these restrictions, the amount can be calculated in mixed integers
linear description can be modeled. Due to the Gross-M modeling is for
kl ∈ E, depending on the state of the binary variable ˜p kl , only one of the restrictions (2.22),
(2.23) is actually a requirement. Together with the simultaneous fulfillment
From (2.20) and (2.21), p kl = max {p kl , −p kl } is modeled. With the
Transmission cost function E DC
kl (p kl ) = δ kl · p kl can now optimize the

rium from an operationally optimal perspective on

Σ C k (P G , s k ) + Σ
k E DC (2.24)
kl (p kl ) → min
k∈V kl∈E

can be scheduled. A disadvantage of DC modeling is the following


Considerations evident. Apparently, the complex transmission losses
of an electrical power transmission system can be calculated by, if
the selected feed is physically permissible, the difference from the total
generated power and the total existing load. This statement
can also be found in [28]. The complex network losses S L = P L + jQ L
are thus over
SL= Σ Sk= Σ (P k + jQ k )
k∈V k∈V

given. Based on assumption (iii) and the relationship b kl = b lk for all kl ∈ E


now applies:

S L = P L + jQ L = P L = Σ Pk= Σ Σ p kl = Σ p kl + p lk = 0

https://translate.googleusercontent.com/translate_f 26/143
2/29/2020 Untitled
k∈V k∈V l∈N k kl∈E: k <l

Thus, in DC modeling based on the assumptions chosen there


the transmission loss ignored. Now, however, are involved in high-voltage networks
geographically smaller dimensions the losses are rather small and only become
geographically larger networks again relevant, so their influence in terms
to an actually optimal operational or power flow condition
may be of concern. This is how the resulting transmission losses are calculated
for a transmission line with resistance R, voltage U and a transmission
active power from P to:

P2·R
PL=
U2

Page 27

2.2. INTEGRATION OF THE ENERGY TRANSMISSION SYSTEM 19

The active dissipation power P L therefore takes


table with the tension. Further can [69] under the keyword High Voltage
"
networks “that the transmission losses to about 6% per 100 km
for a 110 kV line and an 800 kV extra-high voltage line to about

0.5% per 100 km can be reduced. Since the DC model ignores such,
it is also not used to optimize power flow. The cheap
Representation of the network equations allows a mixed-integer
linear model description, so that efficient solution methods with regard to a

optimal operational condition remain applicable. Chapter 7 describes this


extensive test calculations presented. The previous considerations
but also suggest the construction of models, which on the one hand in terms of
have favorable representation on solution processes and on the other hand are able

are to be considered transmission losses. This concern is expressed in


of DCOL modeling.

Dcol modeling
https://translate.googleusercontent.com/translate_f 27/143
2/29/2020 Untitled

In the following, the DC modeling approach is used as a starting point.


The approximate ohmic transmission losses become again afterwards
into expected. Approaches of this kind can also be found in [60, 61]. Around
to be able to consider the ohmic transmission losses afterwards
the balance equations (2.6), (2.8) by the one occurring at each node k ∈ V
Loss size P L ∈ R + added and are then over
k

1 Σ
PG -PD -
k k PL k = p small ∀ k ∈ V, (2.25)
2
l∈N k

1 Σ
−P D - k PL k = p small ∀ k ∈ V \ V (2.26)
2
l∈N k

given. Again the power flow is over

p kl = b kl (θ l - θ k ) ∀ kl ∈ E

given. The ohmic transmission losses fall along an edge kl ∈ E

ν kl = g kl (U 2
i + U 2 j ) - 2g kl U i U j cos (θ k - θ l ) ∀ kl ∈ E

Page 28

20 CHAPTER 2. MATHEMATICAL OPTIMIZATION MODELS

on. With regard to solution methods as developed in Chapter 3 is


a description of the loss ν kl : R → R is an advantage. Therefore, the

Variables θ kl ∈ R for all kl ∈ E together with the restrictions

θ kl = θ k - θ l ∀ kl ∈ E (2.27)

introduced. Due to the approximation U i , U j ≈ used in DC modeling


1 gives approximate losses of

ν kl = 2g kl (1 - cosθ kl ) ∀ kl ∈ E. (2.28)

Now the approximation cos θ kl ≈ 1 is not used because it closes again


https://translate.googleusercontent.com/translate_f 28/143
2/29/2020 Untitled

ν kl ≈ 0 would lead. The transmission losses P L determined in this way are transferred to the
k

Nodes k, l ∈ V are evenly divided and the loss quantities P L are over k

PL = k
Σ ν kl ∀ k ∈ V (2.29)
l∈N k

stated. If you assume a fixed switching assignment for the modeling


binary variables involved, (2.28) in turn turns a non-convex
Optimization model evoked. In Section 3.2 an outer approximation
mation process addressed, which solves the present opti
model by repeatedly solving approximations of the DCOL
Models in mixed-integer linear representation allowed. As an optimization
task can again with cost minimization

Σ C k (P G , s k ) + Σ
k E DC
kl (p kl ) → min
k∈V kl∈E

be used. On the other hand, should the power flow be optimized with regard to
If transmission losses are made, then the optimization task is
given as follows:

Σ P L → min
k

k∈V

The last-mentioned objective function, however, reduces the ohmic


transmission losses overestimated, because the relationship applies

Page 29

2.3. EXPANSION IN A PERIOD 21

1 Σ Σ = Σ Σ
(T) (T)
PL ν
2
k
kl ,
t∈T k∈V t∈T kl∈E: k <l

(T) (T)
since for 1 ≤ t ≤ t max and k, l ∈ V the variable ν Lk and
in the calculation of P
kl
(T) (T)
the variable ν Ll flows. The relationship still applies
lk into that of P

https://translate.googleusercontent.com/translate_f 29/143
2/29/2020 Untitled
(T) (T)
ν =ν
kl
lk ∀ kl ∈ E ∀ t ∈ T,

since the conductances fulfill the property g kl = g lk for all kl ∈ E. The

Factor 1
can thus be added to the target function. It affects the
2

Optimal value, but not the optimal solution of the DCOL model.

2.3 Extension to a period

The models developed so far can be used to determine an operational or


flow-optimized state over a time step, depending on the model choice
become. In the following, the models are to be expanded so that with
they can also be optimized over a period of time,

in this consists of discretized time steps. There will be discretization in


t max + 1 equidistant time intervals and the amount of discretized
Time steps are set with T = {0, ..., t max }. All of the previous sections 2.1
and 2.2 required variables are also given a time index (t). All

The restrictions mentioned occur at every time step and are for all times
t ∈ T to be fulfilled. Now that a period is considered, there are more technical ones
Boundary conditions such as performance gradients and minimum running or minimum
downtimes of the power generation plants. Model extensions-

This type can also be found in [32, 43]. Because the power feed
may not be changed at will from one to the next time step
To meet restrictions that require compliance with the change in performance in the
Limits of the maximum possible increase ∆ k ∈ R + or maximum possible

Guarantee reduction ∆ k ∈ R + . With T: = T \ {0}, restrictions of the


kind

Page 30

22 CHAPTER 2. MATHEMATICAL OPTIMIZATION MODELS

https://translate.googleusercontent.com/translate_f 30/143
2/29/2020 Untitled
(T) (T-1) (T-1) (T-1)
P Gk -P Gk ≤Δk·s k + Γ k (1 - s k ) ∀ t ∈ T ∀ k ∈ V, (2.30)
(T) (T) (T) (T-1)
Γ k (s k
≤P Gk
-P Gk
∀t∈T∀k∈V (2.31)
k - 1) - Δ k · s

this fact. Here the parameters Γ k ∈ R + and Γ k ∈ R + also give


output values, of which switching on or off the
generation plant k ∈ V can be forced. The introduction of
(T) (T)
Indicators and ∈ {0,1}, which signal whether the generating plant k ∈ V im
,v k k

Time step t ∈ T was switched on or off. Art

(T) (T-1) (T) (T)


s k =u k
∀ t ∈ T ∀ k ∈ V, (2.32)
k -s k -v
(T) (T)
u k
≤1 ∀t∈T∀k∈V (2:33)
k +v

are required for this. With those caused by these restrictions


(T) (T)
Switch assignments of the variables u k and V k can easily start up the
Pay attention to power generation plants, ie fixed costs that arise when a
generation plant from the state of non-production to the state of production

passes, but also minimum running times and minimum downtimes of the same model
lose. Denotes Λ k ∈ N the minimum runtime and Λ k ∈ N the minimum downtime
of the power generation plant k ∈ V, this fact can, if Λ k , Λ k ≥ 2
applies about

(T + n) (T)
sk ≥u k
∀ t ∈ {1, ..., t max + 1 - Λ k } (2.34)

∀ n ∈ {1, ..., Λ k - 1} ∀ k ∈ V,
(T + n) (T)
sk ≤1-v k
∀ t ∈ {1, ... ,, t max + 1 - Λ k } (2.35)

∀ n ∈ {1, ..., Λ k - 1} ∀ k ∈ V

be modeled. For the time steps t max + 2 − Λ k ≤ t <t max or t max + 2 − Λ k ≤


If Λ k , Λ k > 2, t <t max must also have restrictions similar to
above type are included. It may also be necessary to start values P init Gk

and s initk with regard to the infeed values and the switching states of the energy
systems k ∈ V at the beginning of the optimization period. This can

by adding restrictions

https://translate.googleusercontent.com/translate_f 31/143
2/29/2020 Untitled
Page 31

2.3. EXPANSION IN A PERIOD 23

(0)
sk = s init k ∀ k ∈ V, (2:36)
(0)
PG k
= P init G k
∀k∈V (2.37)

(T)
happen. With the shift change variables u k can also easily start
costs γ k ∈ R + of the power generation plants are taken into account. You set this up
(T) (T)
for all k ∈ V, the functions D k D k (u on. Further it is assumed
) = γku k k
(0) (0)
gone that u k =v k = 0 applies to all k ∈ V, with different assumptions
are of course possible. The for an optimization period in an operational
The minimization task required for optimization can then, with a DC,
DCOL or AC modeling may still consider the transmission cost share

needs to be about

Σ Σ ( (T) (T) (T) ) → min


C k (P G , s k
k ) + D k (u k )
t∈T k∈V

can be scheduled. If a power flow optimization with regard to minimal


If transmission losses are to be made, the optimization criterion is

illustrated using the example of DCOL modeling, above

(T)
Σ Σ PL → min
k

t∈T k∈V

given. Table 2.1 shows the respective model approaches with regard to their
compared to each other and also records the associated problem class.
In the following section 2.4, relationships between the optimal

model IG DC Dcol AC
operational optimization + + + +

network integration - + + +
Power flow optimization - - + +
Problem class MILP MILP MINLP MINLP

Table 2.1: Advantages and disadvantages of the modeling approaches

solutions of the respective modeling approaches developed. In conclusion, that will


Considered the DC model fully formulated over a period of time.

https://translate.googleusercontent.com/translate_f 32/143
2/29/2020 Untitled

Page 32

24 CHAPTER 2. MATHEMATICAL OPTIMIZATION MODELS

[ ]
Σ Σ ( (T) (T) (T) ) (T)
(DC) C k (P G k , s k ) + D k (u k ) + Σ E DC kl ) → min
kl (p
t∈T k∈V kl∈E

Constraints:

(T) (T) (T) (T) (T)


PGk - P Dk = Σ p kl , −P Di = Σ p il
l∈N k l∈N i

(T) (T) (T) (T) (T)


p kl = b kl (θ l -θ k ), −d max
kl ≤p kl ≤ d max
kl ,θ1 =0
(T) (T) (T)
P Gmink · Sk ≤P Gk
≤ P max
Gk · Sk
(T) (T-1) (T-1) (T-1)
PGk - P Gk
≤Δk·s k + Γ k (1 - s k )
(T) (T) (T) (T-1)
Γ k (s k - 1) - Δ k · s k
≤P Gk -P Gk

(T) (T-1) (T) (T) (T) (T)


sk - seek =u k -v k ,u k +v k ≤1
(j 1 (T) (j 2 (T) (T) (T) (T) (T)
k+ n1 k) k+ n2 k)
sk ≥u k ,s k ≤1-v k ,p from ≥ p from, p from ≥ −p from

(T) (T) (T) (T) (T) (T)


pfrom - p from ≤ M · ˜p from, p from + p from ≤ M · (1 - ˜p from)

(0) (0) (T) (T) (T) (T)


sk = s init
k ,P Gk = P initG k , P k ,u k ,vk ∈ {0.1}
G k∈ R, s
∀j1
k∈ {1, ..., t max + 1 - Λ k }, ∀ n 1 k∈ {1, ..., Λ k - 1}
∀j2
k∈ {1, ..., t max + 1 - Λ k }, ∀ n 2 k∈ {1, ..., Λ k - 1}
∀ t ∈ T, ∀ t ∈ T, ∀ kl ∈ E ∀ ab ∈ E, ∀ k ∈ V, ∀ i ∈ V \ V

2.4 Model relations


In the following it should be verified whether between the optimal values of the four
Models, ie the IG, DC, DCOL and AC model, asserted relations

can be. The considerations here are limited to the use of the
corresponding target functions to minimize costs. First, the term
relaxation, which is necessary for the following considerations.

Definition (relaxation)

https://translate.googleusercontent.com/translate_f 33/143
2/29/2020 Untitled
Given the functions f: X ⊆ R n → R and g: Y ⊆ R n → R.
Optimization model min {g (x): x ∈ Y} means relaxation of min {f (x): x ∈ X},
if g (x) ≤ f (x) ∀ x ∈ X and X ⊆ Y.

Page 33

2.4. MODEL RELATIONS 25

Apparently the optimal value of a relaxation is a lower bound for the


Optimal value of the initial problem. The formation of a model relaxation
can be useful if this simplifies the problem class and thus
the achievement of an optimal solution or lower barrier more easily accessible
is as in the initial problem. With the optimal value of the relaxed problem or
a lower bound can be the quality of permissible points in the initial problem
can be estimated. This strategy in relation to the models presented here
Pursuing should, on the one hand, help to better understand the models
come and on the other hand also possibilities of a quality assessment of
permissible points in the AC model are discussed. Previously, the following one
Notation rule recorded.

Notation rule: Given a model M with associated optimal solving


solution x = (x 1 , ..., x l ) (M permissible, l ∈ N suitable). The i-th component of the
Optimal solution x is denoted by [x i ] M. The admissibility of the M model
be denoted by Z M.

First of all, the optimal value of the IG model represents a lower bound for
that of the DC model. This can be demonstrated as follows:

Theorem 2.1

https://translate.googleusercontent.com/translate_f 34/143
2/29/2020 Untitled
The optimal value z IG of the IG model represents a lower bound for the DC model
represents, ie z IG ≤ z DC .

Proof: The models (IG) and (DC) are permissible. DC modeling requires
the restrictions:

(T) (T) (T)


PG -P = Σ p kl ∀k∈V∀t∈T (2:38)
k Dk
l∈N k

(T) (T)
-P = Σ p kl ∀k∈V\V∀t∈T (2:39)
Dk
l∈N k

(T) (T) (T)


p kl = b kl (θ l
-θ (2:40)
k ) ∀ kl ∈ E ∀ t ∈ T

Page 34

26 CHAPTER 2. MATHEMATICAL OPTIMIZATION MODELS

If one adds up (2.38) and (2.39) for all k ∈ V and for k ∈ V \ V, the result is
the restrictions:

(T) (T) (T)


Σ PG -P Σ Σ p kl ∀t∈T
k
D =
k∈V k∈V l∈N k

There further
(T) (T) (T)
Σ Σ p Σ p ∀t∈T
lk
kl = kl +p
k∈V l∈N k kl∈E: k <l

(T) (T)
applies and because of (2.40) p
+p kl lk = 0 for all kl ∈ E and for all t ∈ T (man
note: b kl = b lk ) are the restrictions

(T) (T)
Σ PG -Pk
D =0∀t∈T
k∈V

(T)
for all P Gk
∈ Z DC fulfilled. Adding these equations to the DC model

therefore does not change the admissibility. Now remove the equations
(2.38) - (2.40) and the restrictions (2.10), (2.19) - (2.23) from the model description
exercise, this results in a relaxation of the DC model. The restriction area this
This relaxation is right, as can be easily verified, with that of the IG model

https://translate.googleusercontent.com/translate_f 35/143
2/29/2020 Untitled

match. There further


Σ Σ E DC
(T)

kl (p kl )≥0
t∈T kl∈E

(T)
for all p kl
∈ Z DC applies, (IG) is a relaxation of (DC) and therefore
z IG ≤ z DC . If (DC) is not permitted, then z IG ≤ z DC = + ∞ certainly applies . Falls (IG)
impermissible, trivially also (DC) impermissible. Overall, z IG ≤ z DC 2

An analog result can also be found between IG and AC modeling.


did derive. For the proof, the RIG model belonging to the IG model is used
built up. The RIG model agrees with the IG model except for the balance sheet
equations at (2.1), considering a time period. This will
by
(T) (T)
Σ PG -P (2:41)
k
D ≥0∀t∈T
k∈V

replaced. Current accounts of this kind can also be found in [23, 51]. Apparently
then (RIG) a relaxation of (IG). The following sentence can be drawn up.

Page 35

2.4. MODEL RELATIONS 27

Theorem 2.2
The optimal value z RIG of the RIG model represents a lower bound for the AC

Model, ie z RIG ≤ z AC .

Proof: The proof is initially analogous to the previous one. The optimization
Modeling (RIG) and (AC) are allowed. The addition of the active power track
for generator and load nodes leads to the restrictions:

(T) (T) (T) (T)


Σ PG -P Σ p ∀t∈T
k
D = kl +p lk
k∈V kl∈E: k <l

For AC modeling, the following applies to kl ∈ E:

(T) (T) [ (T) (T) (T) (T) (T) ] ∀t∈T


p (U l cos kl
kl +p lk = g kl k ) 2 + (U l ) 2 - 2U k U
https://translate.googleusercontent.com/translate_f 36/143
2/29/2020 Untitled

So the restrictions are

Σ (T)
-P
(T)
Σ [ (T) (T) (T) (T) (T) ] ∀t∈T
PG g small(U l cos kl
k
D = k ) 2 + (U l ) 2 - 2U k U
k∈V kl∈E: k <l

} {{
=: P L }

for all permissible points of the AC model. Since g kl ≥ 0 and together with
(T) (T)
the cosine rate p
kl +p lk ≥ 0 for all kl ∈ E, also P L ≥ 0. Thus is
(T) (T) (T) (T)
Σ PG -P Σ PG -P
k
D ≥ k
D - PL= 0 ∀ t ∈ T
k∈V k∈V

(T)
for all P Gk
AC Z AC always met. Now add the restrictions

(T) (T)
Σ PG -P
k
D ≥0∀t∈T
k∈V

to (AC) and removes (2.6) - (2.15), the relaxation of the


AC model that matches the RIG model when the transmis-

is removed from the target function. Hence z RIG ≤ z AC . If


(AC) inadmissible, z RIG ≤ z AC = + ∞ certainly applies . If (RIG) is not allowed
trivially also (AC) inadmissible. Overall, z RIG ≤ z AC 2

If the optimal solution of the RIG model meets the restrictions at (2.41)
Equality is fulfilled, then z IG = z RIG ≤ z AC . Especially with regard to the hard to

Page 36

28 CHAPTER 2. MATHEMATICAL OPTIMIZATION MODELS

problem class of the AC model, the question arises of simplified


models that allow good barrier formation. So far, though
that z RIG represents a lower bound, but must already be here
it should be noted that their quality is due to complete network ignoring and
the associated disregard of maximum transmission capacities
and transmission costs will be more than meager. Would be desirable

https://translate.googleusercontent.com/translate_f 37/143
2/29/2020 Untitled
hence the connection between the DC and AC models.
Intuitively, one surely suspects the DC model due to the Igno
of transmission losses that creates a lower bound with it
can, but it has generally not been possible to prove this presumption.

When using specially designed cost functions, however, another can be used

lower bound for the AC model, more precisely for the AC model
additional requirements for the switching assignments of the power generation plants,
be generated. Chapter 6 addresses this issue again. Also between
the RDC model, which is subsequently built, and the DCOL model

a relation can be established. Apparently a relaxation of the


DC model if the restrictions (2.6) and (2.8), when considering a
Period, to be replaced by:

(T) (T) (T)


PG - P ≥ Σ p kl ∀k∈V∀t∈T (2:42)
k Dk
l∈N k

(T) (T)
-P ≥ Σ p kl ∀k∈V\V∀t∈T (2.43)
Dk
l∈N k

The resulting model is called RDC. Then you can do that


prove the following result.

Theorem 2.3
The optimal value z RDC of the RDC model represents a lower bound for that
DCOL model, ie z RDC ≤ z DCOL .

Page 37

2.4. MODEL RELATIONS 29

Proof: Consider the DCOL model. Since P L ≥ 0 for all k ∈ V, the


k

https://translate.googleusercontent.com/translate_f 38/143
2/29/2020 Untitled

estimates:

(T) (T) (T) (T) 1 (T)


Σ (T)
PG -P k Dk
≥P Gk
-P Dk
- PL k
= p kl ∀k∈V
2
l∈N k

(T) (T) 1 (T) Σ (T)


-P Dk
≥ −P Dk
- PL k
= p kl ∀k∈V\V
2
l∈N k

If we now replace (2.25) and (2.26) with (2.42) and (2.43), we get a re-
laxation of the DCOL model. If you further remove the restrictions (2.27) - (2.29),
this results in a relaxation that corresponds to the RDC model, and that

Claim is proven 2

If further the optimal solution of the RDC model the restrictions (2.42),
(2.43) with equality, then z DC = z RDC ≤ z DCOL . Anticipating the

Test calculations in Chapter 7 are intended to provide some optimal values with regard to that in the introduction
6-node network shown.

example
Consider the 6-knot electrical energy transfer shown in Figure 1.1.
transmission system over a time. The data of the energy transmission system
and technical characteristics of the power generation plants as well as the load profiles

can be found in Appendix B. The optimization spanned


a time step without considering transmission costs. The
Optimal values z IG = z RIG = 287.64, z DC = z RDC = 455.81, z DCOL = 456.96 and
~Z F AC = 457.27 determined.

Here ˜z F AC denotes a by a heuristic according to Chapter 6


the permissible point of the AC model. The optimal value z RIG represents one
demonstrable lower bound for the AC model.

Page 38
https://translate.googleusercontent.com/translate_f 39/143
2/29/2020 Untitled

30 CHAPTER 2. MATHEMATICAL OPTIMIZATION MODELS

2.5 Retrofitting electrical energy transmission


supply systems

In Chapter 7 extensive test calculations using the developed


Optimization models with regard to an optimal feed-in introduced.
As shown in the previous example, the type of underlying

Modeling clearly influence the optimal value. While paying attention to


Network losses, at least in the previous example, only minor additional ones
Causes costs, leads to the addition of electrical energy transmission
systems with capacity-limited cables for a significant increase in costs

compared to the IG model. The direct comparison of z IG and z DC shows a slope


almost 60% reduction in production costs. So the question arises how
electrical power transmission system is to be retrofitted, be it through the
and the associated increase in the maximum possible transmission capacity
capacities as well as the completely new construction of lines to
cost to move in the direction of the optimal value z IG , in the mathematical ideal case
even achieve. An upgrade should certainly take place as inexpensively as possible.

Furthermore, situations can occur in which the IG model is permissible, but that
DC model due to insufficient line capacities or too few cables
is inadmissible. Again, the question arises of an inexpensive one
Network expansion that ensures admissibility for the DC model. This results again
the task of developing an appropriate mathematical optimization model.
The following modeling is done using the DC model. prin-
However, DCOL modeling can also be used. Can be found in [10]
also optimization models using the DC modeling approach
take on the task of preventing electrical energy transmission systems
on failures in the event of unforeseen or very unlikely events
suitable to expand. In the following considerations,
existing transmission costs are ignored, since only the basic egg
properties are to be modeled. However, extensions can be made if necessary
that allow consideration of these costs. The set ˜E ⊆ E gives

https://translate.googleusercontent.com/translate_f 40/143
2/29/2020 Untitled

Page 39

¨
2.5. RETROFITTING ELECTRICAL ENERGY TRANSFER SYSTEMS 31 iii

information about which existing lines for a possible connection


strengthening will be released. Furthermore E ⊂ V × V \ E gives the permissible quantity of
New construction of lines on and N k = {l ∈ V: kl ∈ E} which
ten k ∈ V in the case of a new line construction, additional adjacent nodes. The amount
T = {0, ..., t max } again contains the discretized points in time. The network track
The DC model changes in consideration of a time period to:

(T) (T) (T) (T)


PG -P = Σ p Σ p kl ∀k∈V∀t∈T
Dk
k
kl +
l∈N k l∈N k

(T) (T) (T)


-P = Σ p Σ p kl ∀k∈V\V∀t∈T
Dk
kl +
l∈N k l∈N k

(T)
The additionally required variables p
denote kl ∈ R for kl ∈ E and t ∈ T
the possible flow of power along a new line. The performance
flow itself is with a suitable constant M> 4π max kl∈E | b kl | about the
menhang

(T) (T) (T)


p kl = b kl (θ l
-θ ∀ kl ∈ E ∀ t ∈ T,
k )
(T) (T) (T)
p kl ≤ b kl (θ l

k ) + M · (1 - n kl ) ∀ kl ∈ E ∀ t ∈ T,
(T) (T) (T)
p kl ≥ b kl (θ l

k ) - M · (1 - n kl ) ∀ kl ∈ E ∀ t ∈ T,
(T)
p kl ≤ M · n small ∀ kl ∈ E ∀ t ∈ T,
(T)
p kl ≥ −M · n small ∀ kl ∈ E ∀ t ∈ T

given. When choosing the constant M, note that the load angle differences
limits θ kl for kl ∈ E∪E can amount to a maximum of 2π. The binary variables
n kl for kl ∈ E provide information about newly created edges. The limitation
the load flows with the variables d add kl
∈ R + for kl ∈ E~ ∪ E to record the
Capacity expansion of the lines to:

(T)
−d add
kl
≤p kl
≤ d add
kl
∀ kl ∈ E ∀ t ∈ T

https://translate.googleusercontent.com/translate_f 41/143
2/29/2020 Untitled
(T)
−d max
kl - d addkl ≤p kl ≤ d max
kl + d add
kl ∀ kl ∈ ˜E ∀ t ∈ T

Restrictions with regard to the observance of minimum


run times, minimum downtimes, performance gradients and any start conditions

Page 40

32 CHAPTER 2. MATHEMATICAL OPTIMIZATION MODELS

conditions, ie (2.30) - (2.37) must be observed. The optimization criterion can


measured by the minimization of expansion and operating costs, which then over

Σ Σ Σ Σ ( (T) (T) (T) )


F1 F2 C k (P G , s → min
kl (d add
kl ) + kl (n kl ) + k
k ) + D k (u k )
kl∈ ˜E∪E kl∈E t∈T k∈V

is scheduled. The functions F 1 describe here kl


· D kladd For
(d add
kl ) = c add kl

c kladd ∈ R + the costs incurred through the expansion of transmission capacities

and F 2 kl
∈ R + additionally through a new building
(n kl ) = c ins kl · n kl for c ins
kl

costs incurred by a line. The entire model for network expansion lies in this way
in a mixed-integer linear description. If the DCOL modeling
can be used for retrofitting the network expansion
flow optimization with regard to minimal transmission losses. The
The optimization criterion is then over with a suitable weighting factor ρ> 0

(T)
Σ F1 Σ F2 Σ Σ PL → min
kl (d add
kl ) + kl (n kl ) + ρ · k

kl∈ ˜E∪E kl∈E t∈T k∈V

stated. In the case of a fixed budget b ∈ R + , which excludes


Lich used for retrofitting the electrical power transmission system
The restriction can of course also be used as an alternative

Σ F1 Σ F2
kl (d add
kl ) + kl (n kl ) ≤ b
kl∈ ˜E∪E kl∈E

be included in the model and instead the optimization task with

sole minimization of production costs or transmission losses


https://translate.googleusercontent.com/translate_f 42/143
2/29/2020 Untitled

be set. In the following chapter, solution methods for the developed


presented models.

Page 41

Chapter 3

solution process

The development of appropriate mathematical models in Chapter 2


le to determine a cost-minimal feed or a minimum loss
Transmission certainly took some effort, but is more of an

to view more problematic tasks. It is now important to examine whether


Mathematical procedures are available with which the optimal solutions
of these models can be determined. This certainly depends to a large extent on

https://translate.googleusercontent.com/translate_f 43/143
2/29/2020 Untitled
from what description the models are in. Table 2.1 shows the
Problem classes of the respective models already shown, but should still be
are called:

(i) IG / DC: Both models are in a mixed-integer linear description


before (MILP).

(ii) DCOL / AC: Both models require a mixed-integer nonlinear


Representation (MINLP). When the integer requirements are relaxed
the model description is also non-convex.

Ideally, optimization methods are able to determine the global optimum


agree or based on the type of optimization model or one
high computational ability to calculate lower bounds with which
a quality assessment of the permissible
points. While for mixed-integer linear optics
powerful algorithms are available, the Si

33

Page 42

34 CHAPTER 3. PROCEDURE FOR SOLUTION

tuation with mixed-integer nonlinear optimization models, depending


whether there is convexity after relaxation of the integer requirements,
look less favorable. Section 3.1 describes a classic procedure for mixed
integer linear optimization models, such as with IG or DC modeling of the
Case, presented and its applicability also to the mixed-integer
nonlinear model class checked. Section 3.2 sets out a general

Description held method for mixed-integer convex model class


which can be applied to a relaxed form of the DCOL model.

3.1 A classic procedure

The mixed-integer linear model class can be


https://translate.googleusercontent.com/translate_f 44/143
2/29/2020 Untitled

driving [70] can be tackled. The basic idea of this is shown below
posed, also for their applicability to mixed-integer nonlinear models
dell classes to check. Consider for the vectors b ∈ R s , c ∈ R n , d ∈ R m and
Matrices A ∈ Mat (s, n, R), B ∈ Mat (s, m, R) the mixed-integer linear
Optimization problem:

min {cx + dy: (x, y) ∈ M}


x, y

M = {(x, y) ∈ R n × Z m : Ax + By ≤ b}

The strategy in the basic version of the Branch & Bound process consists of
Subdivision of the allowed amount of the optimization model and subsequent
Mastery of the sub-models by means of suitable barrier formation. In principle is
this method also applies to mixed-integer nonlinear optimization models
applicable, but points precisely to the mixed-integer linear class
favorable properties. The theoretical basics are in the two
summarized the following lemmas.

Theorem 3.1 (branch)


N
Let M = ⋃ k=1
M k and ϕ = min x, y {c T x + d T y: (x, y) ∈ M} and ϕ k =
min x, y {c T x + d T y: (x, y) ∈ M k }. Then ϕ = min k∈ {1, ..., N} ϕ k .

Page 43

3.1. A CLASSIC PROCEDURE 35

Theorem 3.2 (bound)


Be ϕ k
and ϕ k a lower or upper bound for ϕ k . Then:

ϕ: = min φ k ≤ ϕ ≤ min ϕk:=ϕ


k∈ {1, ..., N} k∈ {1, ..., N}

Both proofs are trivial. A lower bound ϕ can be used for ϕ k k


by

https://translate.googleusercontent.com/translate_f 45/143
2/29/2020 Untitled
Solving the LP relaxation of the corresponding sub-problem, ie a recycling
tion of the integer requirements, win and an upper bound ϕ k through
Determine permissible points in the corresponding sub-problem. The effectiveness of the

Branch & Bound procedure now offers the possibility of implicit


meration (cutting of knots). If there is inferiority, ie for a subproject
blem is the condition ϕ ≥ ϕ is fulfilled, the corresponding node in the
k

Branch & bound tree to be cut off. The complete algorithm man-
ma can again be taken [70].
When AC or DCOL modeling is used, even with a
laxation of the 0-1 requirements regarding the binary variables due to the restriction
(2.11) and (2.12) and (2.28) evoked a non-convex optimization model.
This is one of the main reasons for the practical efficiency of the Branch & Bound process
lies in the possibility of clipping nodes due to inferiority,
however, a lower barrier is required. In the case of mixed

integer linear models can be solved by solving the LP relaxation


and observing Lemma 3.2. Even with mixed-integer non-
linear models which, after relaxation of the integer requirements, convex
this methodology is applicable. If the relaxed is not convex

Model description can not be guaranteed, the global optimal


to reach point. There is therefore no possibility of obtaining lower ones
Barriers and the associated fact of removing knots
due to inferiority. This has an enormous impact on the effectiveness

Branch & Bound process, as all possible combinations


would have to be checked and a complete enumeration would be carried out.
In the following considerations, the AC model should therefore initially be ignored
and instead give priority to developing a suitable one

Page 44

36 CHAPTER 3. PROCEDURE FOR SOLUTION

Process to solve the DCOL model, which is

https://translate.googleusercontent.com/translate_f 46/143
2/29/2020 Untitled
belongs to the same model class, but for which a favorable relaxation is indicated
can be. The stochastic extensions made in Chapter 4
optimization models as well as the algorithms described in Chapter 5
for the treatment of stochastic optimization models are only for the IG and DC
Modeling or applicable to a relaxed form of the DCOL model. In Ka

Chapter 6 nevertheless presents heuristic approaches for the AC model based on


aim to determine good permissible points in order to
the accuracy of the approximate models
to be able to check the AC model.

3.2 An external approximation method


Due to the restrictions, the DCOL model is (2.28) and more available
Binary variables in a sophisticated description. It has an extension
of the DC model, especially for high and extra high voltage transmissions
Validity. In the case of such transmissions, it can in turn be used that
the load angle differences θ kl ∈ R on the lines kl ∈ E are small and in

Application of the condition θ kl ∈ [- π 2



2 ] are sufficient. Numerous

giving approximation options can also be found in [60]. Around


the number of binary variables in the DCOL model or in a variant with appro
The aim is not to further increase the maximized ohmic transmission losses

Loss function cannot be approximated by a piecewise linear function,


Rather, an approach that suitably relaxes the DCOL model is presented
become. Now replace the loss equation (2.28) with

(T) (T)
ν (3.1)
kl ≥ 2g kl (1 - cosθ kl ) ∀ kl ∈ E ∀ t ∈ T

and also demands that the load angle differences meet the conditions

π (T) π
- ≤θ ∀ kl ∈ E ∀ t ∈ T (3.2)
2 kl ≤ 2
(T)
are sufficient, the functions 2g kl (1 - cosθ ,
] for all kl ∈ E and t ∈ T
π
) | [- kl π 2 2

convex, since g kl ≥ 0 for kl ∈ E. The resulting model is created with RDCOL

Page 45

https://translate.googleusercontent.com/translate_f 47/143
2/29/2020 Untitled

3.2. AN EXTERNAL APPROXIMATION PROCEDURE 37

denotes and is available in a mixed-integer convex description. A


This is the case if, after relaxation of the integer requirements of the original
jump model, a convex optimization model remains. Between the RDC,
RDCOL and DCOL record the following relationships, where
production cost minimization as an optimization criterion for all models
underlying.

comment
If the RDCOL model is allowed and the load angle differences
the condition for the optimal solutions of the RDC and DCOL model
(T) (T)
[θ , π2 ] for all kl ∈ E and t ∈ T, so is
] RDC , [θ kl ] DCOL ∈ [- π
kl 2

z RDC ≤ z RDCOL ≤ z DCOL .

Certainly, the losses actually occurring in the RDCOL model can

be overestimated. Nevertheless, it should be noted that it is a


Minimization problem and an overestimation of additional costs
leads. However, it cannot be excluded that in certain cases a

Optimal solution exists for overestimated losses. Below is an exterior


Approximation method for the solution of the RDCOL model presented. Generally
is the mixed-integer convex model class considered here about

{
min cx + dy: Ax + By ≤ b
x, y

(P) f i (x) ≤ 0, i = 1, ..., m

x∈Rn,y∈Y⊂Zs

given. Here f i : R n → R are convex and differentiable. From the crowd


Y ⊂ Z s is a limitation. The matrices A, B and vectors
b, c, d are equipped with suitable dimensions. The model (P) is
thus a mixed-integer convex optimization model. Now it turns

Question how an optimal solution of (P) can be determined. In [17, 21] can be found
Cutting plane methods specially adapted to the model class in question.
Non-linear master problems must also be minimized. Below is
presented a method which is able to find an optimal point of (P)

https://translate.googleusercontent.com/translate_f 48/143
2/29/2020 Untitled

Page 46

38 CHAPTER 3. PROCEDURE FOR SOLUTION

or a permissible point with a verifiable quality certificate by solving


Master problems that exist in a mixed-integer linear description,
to investigate. This makes such a method particularly suitable
use stochastic models in conjunction with decomposition processes.
Decomposition methods such as scenario decomposition [12] require mixed
integer linear structures of the master problems. Therefore, in this work
the model class (P) does not use approximation methods of the type described in [17, 21]
are treated. The algorithm developed below builds on
the Kelley method [39] and its extension (see section 5.2)
also on the scenario composition. It is now the next sentence,

known from nonlinear optimization [3].

Theorem 3.3
Let g: C → R be differentiable on the open and convex set C. Then g is
convex on C if and only if g (x) ≥ g (x) + ∇g (x) (x - x) for all x, x ∈ C.

For the purpose of convergence statements of the algorithm presented below,


meet the two requirements:

(V1) There is a limited polytope


Z M ⊃ {x ∈ R n : f i (x) ≤ 0, i = 1, ..., m}.
0

(V2) There is a K ∈ R with ∇f i (x) ≤ K for all x ∈ Z M 0 and i = 1, ..., m.

According to (V1), Z M is and thus also limited to Z P. Now consider how


0

follows defined master problem in one step k ∈ N 0 :

(M k ) min {cx + dy: Ax + By ≤ b, x ∈ Z M , y ∈ Y ⊂ Z s }


k

x, y

The set Z M k is now defined. The optimal solution of (M k ) becomes if


Z M = ∅ (Z M again stands for the admissibility range), with (x ∗
k k
k , y ∗k ).

https://translate.googleusercontent.com/translate_f 49/143
2/29/2020 Untitled
Further, a (M k ) feasible point (~x k , ~y k ) considered to be predetermined
Tolerance τ k ≥ 0 and the bound ϕ valid for (M k ) k
the condition

c ˜x k + d ˜y k - ϕ k
≤τk (3.3)
φk

Page 47

3.2. AN EXTERNAL APPROXIMATION PROCEDURE 39

Fulfills. The set Z M with k ∈ N is then related


k

ZM k+1 =ZM k ∩{x∈Rn : f i (˜x k ) + ∇f i (˜x k ) (x - ˜x k ) ≤ 0

∀ i ∈ {j ∈ {1, ..., m}: f j (˜x k )> 0}

given. The outer approximation method is then as follows:

Algorithm 3.1
(S0) initialization
Find a polytope Z M , choose τ 0 ≥ 0 and set k = 0.
0

(S1) optimization
If possible, determine one that is admissible for the master problem (M k )
Point that satisfies (3.3).

(S2) termination
If f i (˜x k ) ≤ 0 for all i = 1, ..., m or Z M = ∅, then terminate.
k

(S3) updating
Choose 0 ≤ τ k + 1 ≤ τ k , set k: = k + 1 and go to (S1).

For the purpose of convergence statements, the following sentence is considered, which
for the convex model class, ie without the existence of integer

claims in addition to [39] and [53], but here on the


mixed-integer convex model class was adapted. The optimal solution
of (P), if Z P = ∅, is denoted by (x ∗ , y ∗ ). If Z P = ∅, then
cx ∗ + dy ∗ : = + ∞ defined.

https://translate.googleusercontent.com/translate_f 50/143
2/29/2020 Untitled

Theorem 3.4

(i) Z P ⊂ Z M for all k ∈ N 0 .


k

(ii) A point generated by algorithm 3.1 in a step k ∈ N 0


(˜x k , ˜y k ) ∈ Z M fulfills the condition:
k

c ˜x k + d ˜y k - cx ∗ - dy ∗
≤τk (3.4)
cx ∗ + dy ∗

(iii) If (˜x k , ˜y k ) / ∈ Z P , then (˜x k , ˜y k ) ∈ Z M \ Z M k k+1 .

Page 48

40 CHAPTER 3. PROCEDURE FOR SOLUTION

Proof:

(i) According to prerequisite (V1), Z P ⊂ Z M results . Let k ∈ N for a step


0

fulfills the condition Z P ⊂ Z M k−1 . The following applies to all permissible (x, y) ∈ Z P
Theorem 3.3 then:

0 ≥ f i (x) ≥ f i (˜x k ) + ∇f i (˜x k ) (x - ˜x k ) ∀ i = 1, ..., m

Since (x, y) ∈ Z P is arbitrary, Z P ⊂ Z M follows .


k

(ii) The point (˜x k , ˜y k ) ∈ Z M satisfies (3.3) by construction. With (i)


k

Lich:

c ˜x k + d ˜y k ≤ϕ (1 + τ k )
k

≤ (cx ∗
k + dy ∗ k ) (1 + τ k )
≤ (cx ∗ + dy ∗ ) (1 + τ k )

(iii) If (˜x k , ˜y k ) / ∈ Z P , then there is an index i ∈ {1, ..., m} with f i (˜x k )> 0. Then
but is the inequality

https://translate.googleusercontent.com/translate_f 51/143
2/29/2020 Untitled
f i (˜x k ) + ∇f i (˜x k ) (x - ˜x k ) ≤ 0

for x = ˜x k violated and thus (˜x k , ˜y k ) ∈ Z M , but it is (˜x k , ˜y k ) / ∈ Z M


k k+12

When algorithm 3.1 runs, the following three cases are now possible:

(i) In a step k ∈ N 0 , Z M = ∅. According to Theorem 3.4 (i) is then also


k

Z P = ∅ and (P) has no solution.

(ii) In a step k ∈ N 0 is (~x k , ~y k ) ∈ Z P . According to Theorem 3.4 (ii) then (3.4)

Fulfills.

(iii) An infinite sequence {˜x k , ˜y k } is generated.

If the infinite sequence {˜x k , ˜y k } is generated, each cluster point is one


Solution of (P) if {τ k } → 0 holds. The consequence must be at least one
because the compactness of the sequence terms of {˜x k , ˜y k }

Page 49

3.2. AN EXTERNAL APPROXIMATION PROCEDURE 41

from Z M 0 ⊇ZM k for all k ∈ N are in a compact set. The proof


is again based on [39, 53], with adjustments for the treatment of

mixed-integer convex model class.

Theorem 3.5
Algorithm 3.1 does not stop after a finite number of steps and let {τ k } → 0.
Then each cluster point of the sequence {˜x k , ˜y k } is an optimal solution of (P).

Proof:
Let {˜x k , ˜y k } be a partial sequence of {˜x k , ˜y k } that is against the cluster point (˜x, ˜y)
j j

converges. In addition, the consequences {x ∗ kj , y k∗ } and {τ k } → 0 are considered. It applies


j
j

with Theorem 3.4 (ii) the connection:

c ˜x k + d ˜y k
j j ≤ (cx ∗ + dy ∗ ) (1 + τ k ) j

https://translate.googleusercontent.com/translate_f 52/143
2/29/2020 Untitled
≤ (cx + dy) (1 + τ k ) ∀ x, y ∈ Z P
j

In the limit, the condition c + d ~x ~y ≤ cx + dy ∀ x, y ∈ Z is then P ER


crowded. It remains to show that (~x, ~y) ∈ Z P is. Suppose it is (~x, ~y) / ∈ Z P .

Thus an index exists i ∈ {1, ..., m} with f i (˜x) = η> 0. Da (˜x, ˜y) cluster
point and f i is continuous, there exists a k ∈ N with ˜x k - ˜x < η 2K
and f i (˜x k )> η 2
,
According to (V2) the condition ∇f i (x) ≤ K for all (x, y) ∈ Z M 0 Fulfills.
Then:
η η
f i (˜x k ) + ∇f i (˜x k ) (˜x - ˜x k )> -K =0
2 2K
Hence (˜x, ˜y) / ∈ Z M k+1 . Since Z M k + 1 is complete, there is a ϵ> 0 such that

for the sphere B = B ((˜x, ˜y), ϵ) now B ∩ Z M k+1 = nun . But since (˜x l , ˜y l ) ∈ Z M k+1 for
all indices l ≥ k + 1 must apply, this is a contradiction 2

By design, Z M is k+1 ⊆ZM k for all k ∈ N 0 . So is one

for the master problem (M k ) generated lower bound also for the master
valid (M k + 1 ).

Application of the procedure to the RDCOL model The presented external


Outer approximation methods can be applied to the RDCOL model. In the following
Refinements are made in the algorithm. In addition to the

Page 50

42 CHAPTER 3. PROCEDURE FOR SOLUTION

the exact description of the linear set to be approximated


Cuts to approximate the same are shown. Furthermore, upper

ken for the variables ν kl with kl ∈ E, which are the ohmic transmission losses
describe, called. Also the termination criterion shown in algorithm 3.1
will be adapted. The exact ohmic transmission losses can be compared to those for
kl ∈ E defined functions

https://translate.googleusercontent.com/translate_f 53/143
2/29/2020 Untitled
f kl : R → R
θ kl ↦ → 2g kl (1 - cosθ kl )

to be discribed. The permissible amount of loss on a line kl ∈ E, subsequent


designated with M kl , can then about the connection

{ π π }
M kl : = (ν kl , θ kl ) ∈ R × [- . ]: f kl (θ kl ) - ν kl ≤ 0
2 2

can be expressed and is illustrated in Fig.3.1. Since g kl = g lk for all kl ∈


E, is also f kl = f lk and thus also M kl = M lk for all kl ∈ E. So far

1 / g small

2
M small

2
0 2 kl

Figure 3.1: Graphical illustration of the permissible losses

the quantities M kl are unlimited, so that requirement (V1) is not met


can. Now take the functions f kl | [- π
their maximum at θ kl = ± π
2

2 ] 2
on.
There the function value is 2g kl . Seen overall - when looking at one
Period - can approximate the convex set to be approximated

Page 51

3.2. AN EXTERNAL APPROXIMATION PROCEDURE 43

{⊗ ⊗
(T) (T) (T) (T)
M= t∈T kl∈E: k <l
(ν , π2 ]) | T | | E2| : f kl (θ
kl ,θ kl ) ∈ (R × [- π 2 kl )-ν kl ≤0

https://translate.googleusercontent.com/translate_f 54/143
2/29/2020 Untitled
ν kl(T)≤ 2g kl
∀ kl ∈ E: k <l
∀t∈T

can be scheduled. In addition,

(T) (T)
ν lk
∀ kl ∈ E with k> l ∀ t ∈ T
kl =ν
(T)
required. There are even more upper bounds for the variable ν kl

specify. Of course, the ohmic transmission losses on a line


kl ∈ E never exceed the maximum possible flow. This leads to the
restrictions:
(T)
ν ∀ kl ∈ E ∀ t ∈ T
kl ≤ d max kl
(T) (T)
This can be made even sharper with the help of the previously introduced variable p
kl =|p kl |
be formulated because the following applies:

(T) (T)
ν kl
∀ kl ∈ E ∀ t ∈ T
kl ≤p

However, it should be noted that binary variables are required to calculate the amount
those who can significantly increase the computing effort. If you pay attention to

certain lines transmission costs anyway, can be on the sharper barrier


can be used. The support hyperplanes are shown below
the functions f kl at a point θ kl ∈ [- π 2
, π2 ] specified. With Theorem 3.3 these are

about

π π
fθ kl : [- . ]→R
2 2
θ kl ↦ → 2g kl [1 - cosθ kl + sinθ kl (θ kl - θ kl )]

given. It is calculated that this is due to the restrictions

f θ (θ kl ) ≤ ν kl
kl

⇔ 2g kl sinθ kl θ kl - ν kl ≤ 2g kl [sinθ kl θ kl + cosθ kl - 1]

Page 52

https://translate.googleusercontent.com/translate_f 55/143
2/29/2020 Untitled

44 CHAPTER 3. PROCEDURE FOR SOLUTION

leads. In the outer approximation process itself, minor adjustments


compared to algorithm 3.1. There will be an early termination

initiation if in a step k ∈ N 0 to a predetermined ϵ> 0 the conditions


supply
(T) (T)
f ij (˜θ (3.5)
ij, k ) - ˜ν ij, k <ϵ ∀ ij ∈ E: i <j ∀ t ∈ T
(T) (T)
is satisfied. Here denote ˜θ
and ˜ν ij, k for ij ∈ E and t ∈ T one through
ij, k

the algorithm determined in a step k ∈ N 0 permissible point that the


Tolerance condition (3.3) fulfilled. An underestimation of the network losses around the

Value ϵ is thus permitted.

Overestimation of network losses Situations can arise in which

the optimal solution or a permissible point of the RDCOL model, which (3.4)

fulfilled, overestimated network losses. Assumed on exactly one edge


from ∈ E this occurs at a time t ∈ T for the optimal solution, ie it is:

[ν t
ab ] RDCOL > 2g ab (1 - [cosθ t from ] RDCOL )

An overestimation of losses has immediate consequences due to restriction (2.29)


effects on sizes [P t La ] RDCOL and [P t Lb ] RDCOL . Thus, in the knobs
ten a, b ∈ V assumed fictitious burdens

rators are covered. Without these additional burdens, production,


if cost optimization was chosen as the optimization goal, either more expensive
or inadmissible. This can be done with the minimum feed limits and
different cost functions of the power generation plants as well as the maximum

possible transmission capacities of the lines are justified. It can


it makes sense to have technical devices ready that have electrical energy or
store appropriately converted or leave unused. If, however, a
Estimation of transmission losses should be prevented explicitly, can

an internal approximation of the network losses is carried out in addition to the external one
become. The idea is illustrated in Fig.3.2. In addition to an exemplary
internal outer approximation. With the inner
Approximation is a piece-by-piece linear function that is based on the

https://translate.googleusercontent.com/translate_f 56/143
2/29/2020 Untitled

Page 53

3.2. AN EXTERNAL APPROXIMATION PROCEDURE 45

1 / g small

2
M small

2
0 2 kl

Figure 3.2: Exemplary approximation of the losses

gray marked bases. The exact modeling

technology can be found, for example, [38]. With a sufficient number of


The actual loss function can of course be arbitrarily accurate
be approximated. But it becomes a mathematical description of one
piecewise-linear function requires a binary variable for each of its segments.

This increases when larger energy transmission systems and an optimization


period, the computational effort to solve the RDCOL model
enormously. This approach is not pursued further below and is instead based on
[47] referred. There, in addition to the last-mentioned modeling technique

also pursues so-called SOS approaches.

https://translate.googleusercontent.com/translate_f 57/143
2/29/2020 Untitled

Page 55
54

Chapter 4

Stochastic optimization

The previous considerations were based on a complete knowledge of the inputs


output data, but these are often subject to uncertainties. So
In the case of the models presented here, active and reactive power loads can

be sure, but also the maximum possible transmission capacities of the transmission
transmission lines depend on atmospheric conditions. With wind energy
or solar energy systems are the maximum possible feed-in values from the
current weather, ie the radiation intensity of the light and the intensity

dependent on the wind and therefore, even if good forecasts for these
sizes are available, to some extent also subject to chance. Contrary to
Modeling according to Chapter 2 is therefore part of the input
data associated with the previous explanations are uncertain.

Overall, this leads to random or stochastic optimization models.

https://translate.googleusercontent.com/translate_f 58/143
2/29/2020 Untitled
Not just for a cost-optimal feed, much more for a feed
the energy generation plants that must guarantee security of supply,
these random parameters can have a high impact. Below is supposed to

but at least assume that for these random parameters


good forecasts are available, ie one is able to assess the probability
that a random variable that specifies the random need for the nodes that
maximum possible transmission of the transmission lines or the maximum possible

The feed-in of certain energy generation plants describes certain values


assumes to specify. Now the question arises as to how these exist

47

Page 56

48 CHAPTER 4. STOCHASTIC OPTIMIZATION

Bring information into the process of optimal decision making


gen. This is the task of stochastic optimization. In from-

Section 4.1 first becomes two-stage stochastic mixed-integer linear


Models are presented that use their application in IG and DC modeling
find random input data. Section 4.2 describes two-stage
Chastic mixed-integer convex models addressed when in use

of the RDCOL model become relevant. Section 4.3 also gives an insight
in the multi-stage stochastic optimization, which is a very accurate modeling
that allows uncertainties. The accesses selected here are generally valid
eng description and so applicable to numerous problems in practice.

Section 4.4 deals separately with the task of how the subsequent
stochastic modeling techniques listed below to those in Chapter 2
presented optimization models can be transferred.

4.1 Two-stage stochastic mixed-integer


League linear models

Disregarding a possible level separation, a motivation for this


https://translate.googleusercontent.com/translate_f 59/143
2/29/2020 Untitled

is still to be done, a general stochastic is first mixed-


considered integer linear optimization model. Such a model will
also presented in [9, 37, 59]. This is subsequently used for the mixed integer

linear model class defined.

Definition 4.1
The optimization model

min {c (ω) x: T (ω) x = h (ω), x ∈ X} (4.1)


x

is called stochastic mixed-integer linear model. Here, X ⊆ R m a


Polyhedron with integer requirements, c: Ω → R m , T: Ω → Mat (s, m, R) and
h: Ω → R s random vectors on a probability space (Ω, A, P).

A polyhedron with integer requirements is not a


Polyhedra in the traditional sense. Apart from marginal cases, it also shows no

Page 57

4.1. TWO STAGE STOCHASTIC MIXED-INTEGRATED


LINEAR MODELS 49

quality characteristic. If X ⊆ R m is a polyhedron with integer requirements,


there is a polyhedron P ⊆ R m and an index set I ⊆ {1, ..., m} with

X = {x ∈ P: x i ∈ Z ∀i ∈ I}. The convex hull of X again corresponds to P.

The problem to be clarified below is minimization


of model (4.1). It is additionally assumed that the

The admissible range of the model (4.1) is limited for each ω ∈ Ω and not empty
is. Consider the high-quality optimal solution function

ϕ: Ω → P (R m )

ω ↦ → arg min {c (ω) x: T (ω) x = h (ω), x ∈ X}.


x

It should now be assumed that a function ψ: P (R m ) → R m for

https://translate.googleusercontent.com/translate_f 60/143
2/29/2020 Untitled
Is available, which uniquely contains an element from the set ϕ (ω) for each ω ∈ Ω
can choose. Now consider the function ϕ ∗ = ψ ◦ ϕ. Apparently, ϕ ∗ is one
Random variable, which for an ω ∈ Ω leads to an optimal decision for (4.1) ϕ ∗ (ω)
maps. It is important to make a decision before the chance is realized
meet that as cheaply as possible, ideally after realization of chance

is. Often one uses oneself if only the cost vector c is random and the
Restriction area itself is not subject to any uncertainties, one of the two
subsequent approaches.

(E1) Replace the random vectors with their expected value and solve the result
deterministic optimization problem.

(E2) Use the decision ∫ Ω


ϕ ∗ (ω) P (dω).

According to the optimization models from Chapter 2, the secondary


conditions, the balance equations, transfer limitations and power

Model meals, random. If approaches (E1), (E2)


used, both have in common that the resulting solution after implementation
may be inadmissible by chance. Consider the following example.

Page 58

50 CHAPTER 4. STOCHASTIC OPTIMIZATION

Example 4.1
The stochastic optimization model is given

{ 1
Max x1+ x2 : x 1 + x 2 ≤ h (ω)
x1,x2 2 ,
x 1 ≥ 0, x 2 ≥ 0

Consider the probability space (Ω, A, P) and the discretely distributed assignment
if variable h (ω) with realizations 0.8, 1 and 1.2. P (h = 0.8) = 0.2,
https://translate.googleusercontent.com/translate_f 61/143
2/29/2020 Untitled

P (h = 1.0) = 0.5 and P (h = 1.2) = 0.3. It's easy to calculate that the two
the previously shown approaches (E1) and (E2) the point (x 1 , x 2 ) = (1.02.0)
deliver. This is therefore not permitted when Realizations 0.8 and 1 occur.
When implementation 1.2 occurs, it is permissible, but not optimal, since the
Then the optimal solution would be given by (x 1 , x 2 ) = (1.2.0).

A possible inadmissibility of those obtained by (E1) and (E2)


Decision presents both approaches for the models discussed in this work
in question. Especially in the electrical power supply must be permissible
Guaranteed production decisions with regard to security of supply

that is, a production is to be determined which, no matter how the realization


the random variable will fail, all necessary constraints of the
100% complies with the models presented. A specialization of (4.1)
can be achieved through the two-step approach that works through

[7, 15] justified. Here, the optimization model can be divided into one level
separate it securely and into one level with insecure information. considered
one can again the general stochastic optimization model (4.1), so
in the two-stage case, the objective function vector is divided into two individual vectors

become. One of these two has secure information. Accordingly


are variables that have safe input data in the target function
Connected, differentiated from those that are associated with insecure.
A two-step stochastic mixed-integer linear optimization model

can be defined as follows:

Page 59

4.1. TWO STAGE STOCHASTIC MIXED-INTEGRATED


LINEAR MODELS 51

Definition 4.2
The optimization model

https://translate.googleusercontent.com/translate_f 62/143
2/29/2020 Untitled
{
min cx + d (ω) y: T (ω) x + W (ω) y = h (ω)
x, y
(4.2)
x ∈ X, y ∈ Y

is called a two-stage stochastic mixed-integer linear model. Are there


X ⊆ R m and Y ⊆ R m polyhedra with integer requirements and c ∈ R m .
1 2 1

The random vectors d: Ω → R m , T: Ω → Mat (s, m 1 , R), W: Ω → Mat (s, m 2 , R)


2

and h: Ω → R s are defined on the probability space (Ω, A, P).

The equivalent description can be given for model (4.2)

min {cx + φ (x, ω): x ∈ X}


x

with the function

φ: X × Ω → R

x × ω ↦ → min {d (ω) y: W (ω) y = h (ω) - T (ω) x, y ∈ Y}


y

build up. The two-stage stochastic optimization model can be integrated into one
Random internal and an external minimization problem can be broken down.
Accordingly, variables x ∈ X and y ∈ Y are used as first-stage and second-stage
called variable. A decision x ∈ X is non-anticipatory,
ie the following sequence of decisions and observations applies:

decide x → observe (d, T, W, h) (ω) → decide y = y (x, ω)

It is always assumed that the distribution of the random vector (d, T, W, h)


is independent of a decision x ∈ X. Certainly a second stage
decision y ∈ Y both from the first stage decision x ∈ X and from

Chance itself, since φ: X × Ω → R applies. Structural properties that under


know the requirements of real value, semi-continuity and piecewise li
show the proximity of the function φ can be found in [11]. How can you do that now?
an optimal first stage decision x ∈ X with random influence in the second stage

Page 60

https://translate.googleusercontent.com/translate_f 63/143
2/29/2020 Untitled

52 CHAPTER 4. STOCHASTIC OPTIMIZATION

to meet? Apparently φ (x,.) For fixed x ∈ X is one on the probability space


(Ω, A, P) defined random variable. The task of stochastic optimization is
thus finding an optimal random variable cx ∗ + φ (x ∗ ,.) for (4.2 ),
which is caused by the decision x ∗ ∈ X. According to this task
risk-neutral and risk-averse approaches are used.

Risk-neutral models

As standard works for the introduction and deepening in the field of the two-step
stochastic optimization should in addition to the literature source already mentioned [37]
also be held. For x ∈ X, consider the
space (Ω, A, P) defined random variables:

Fx:Ω→R

ω ↦ → cx + φ (x, ω)

Let Z = {F x : x ∈ X} the set of all random varieties allowed for (4.2)


ble. The determination of a best non-anticipatory decision x ∈ X is
synonymous with determining a best random variable F ∗ ∈ Z. At the
Risk-neutral modeling is based on the expected value-based approach
sets. This aims to keep the average cost long
Minimize view. If you consider the expected value functional

QE:X→R

x↦→ F x (ω) P (dω),
Ω

so follows

min {Q E (x): x ∈ X} (4.3)


x

as an optimization problem. An optimal decision for (4.3) x ∈ X minimized

hence, in the long run, the average cost. Find in [63]


are structural statements about the Q E functional, which means that the
Ensure functionality and even continuity under additional conditions. With
assuming that the distribution of the random vector (d, T, W, h) is discrete and that

https://translate.googleusercontent.com/translate_f 64/143
2/29/2020 Untitled

Page 61

4.1. TWO STAGE STOCHASTIC MIXED-INTEGRATED


LINEAR MODELS 53

N realizations or scenarios (d i , T i , W i , h i ) with the probabilities


π i for i = 1, ..., N, an equivalent representation of

(4.3) in a mixed-integer linear description. This is with


y: = (y 1 , ..., y N ):
N
{ Σ
min cx + πjdjyj :Tjx+Wjyj=hj
x, y
j=1
(4.4)
x ∈ X, y j ∈ Y

j = 1, ..., N

With this representation (with discrete distribution) you have the possibility
created to find a decision for (4.2) that would in the long term
minimized average costs. On algorithmic difficulties, in particular
in the case of a high number of scenarios N, Chapter 5 is discussed in more detail. There will be

also algorithms adapted to the problem are presented.

Risk-averse models

(4.3) minimizes the average costs in the long run,


after the realization of the coincidence, the actual costs can vary greatly
Expected value deviate. Under certain circumstances, an excessive deviation can occur

pose too great a risk. Another functional is used to record the risk

QR:X→R

needed. With a suitable weighting factor ρ> 0 this is the result


Expected value risk model:

min {Q E (x) + ρQ R (x): x ∈ X} (4.5)


x

Modeling approaches of the same type and other risk measures can also [48, 50, 67]
be removed. With the help of ρ> 0 it can be weighted how strong the risk
should be taken into account in the optimization. There are two typical risk measures

given level η ∈ R + the excess probability


https://translate.googleusercontent.com/translate_f 65/143
2/29/2020 Untitled

Q EP : X → R

x ↦ → P ({ω ∈ Ω: F x (ω)> η})

Page 62

54 CHAPTER 4. STOCHASTIC OPTIMIZATION

and the expected excess

Q EE : X → R

x↦→ max {F x (ω) - η, 0} P (dω).
Ω

Apparently, Q EP measures the probability of exceeding η while


Q EE also measures the expected level of exceeding η. More in the
Risk measures frequently encountered in stochastic optimization are the value in
Risk or the semi-deviation. As literature sources are [46, 67] that
go into detail about the structural properties of the risk measures Q EP and Q EE .
If the random vector (d, T, W, h) is distributed discretely, then again
(4.5) State reformulations in a mixed-integer linear description.
For the risk measure Q EP the reformulation is with a suitable constant
B> 0 and the variable vector Θ: = (θ 1 , ..., θ N ):

N N
{ Σ Σ
min cx + πjdjyj+ρ πjθj :Tjx+Wjyj=hj
x, y, Θ
j=1 j=1

cx + d j y j - η ≤ B · θ j (4.6)

x ∈ X, y j ∈ Y, θ j ∈ {0.1}

j = 1, ..., N

The additional variables θ j ∈ {0,1} required for j = 1, ..., N measure for


the optimal solution x ∗ ∈ X, whether scenario j ∈ {1, ..., N} occurs
Total costs η are exceeded. Denote θ ∗
j ∈ {0,1} and y ∗ j ∈ Y for
j = 1, ..., N an optimal assignment, the following applies:
https://translate.googleusercontent.com/translate_f 66/143
2/29/2020 Untitled

cx ∗ + d j y ∗ j
∀ j ∈ {1, ..., N}
j - η ≤ Bθ ∗

If the level η ∈ R + is exceeded for scenario j ∈ {1, ..., N} must


θ∗
j = 1 apply. From the constant B> 0, the existence for restricted X

are shown (cf. [64]). If the level η for a scenario j ∈ {1, ..., N} is not
*
is exceeded, then in the optimal solution θ assumed j = 0. In the
Use of the risk measure Q EE can be used for a discrete distribution of the random
vector (d, T, W, h) is also an equivalent representation in mixed integer

Page 63

4.2. TWO STAGE STOCHASTIC MIXED-INTEGRATED


CONVEX MODELS 55

linear description can be given with the additional variable vector


v: = (v 1 , ..., v N ) is given as follows:

N N
{ Σ Σ
min cx + πjdjyj+ρ πjvj :Tjx+Wjyj=hj
x, y, v
j=1 j=1

cx + d j y j - η ≤ v j (4.7)

x ∈ X, y j ∈ Y, v j ∈ R +

j = 1, ..., N

The solutions v ∗ achieved at the optimal point for j = 1, ..., N j


∈ Measure R +
for the optimal first stage decision x ∗ ∈ X, how strong in a scenario

j ∈ {1, ..., N} with optimal second stage decision y ∗


the level η ∈ R + j

is exceeded. In the optimal solution the following then applies to j ∈ {1, ..., N}:

{ cx ∗ + d j y ∗
v∗ j - η: cx ∗ + d j y ∗ j >η
j=
0 : cx ∗ + d j y ∗
j ≤η

4.2 Two-stage stochastic mixed-integer


League convex models
https://translate.googleusercontent.com/translate_f 67/143
2/29/2020 Untitled

With regard to the inclusion of stochastic data sets when using the
RDCOL model, in this subsection the associated stochastic
Model class are presented, but again in a general description
being held. The considerations are limited to stochastic
Models of this type, which are given by the following definition.

Definition 4.3
The optimization model

{
min cx + d (ω) y: T (ω) x + W (ω) y = h (ω)
x, y
(4.8)
x ∈ X ∩ E, y ∈ Y ∩ F

is called two-stage stochastic mixed-integer convex model. Are there


X⊆Rm 1 and Y ⊆ R m 2 Polyhedra with integer requirements and E ⊆

Page 64

56 CHAPTER 4. STOCHASTIC OPTIMIZATION

m1
R and F ⊆ R m 2 convex sets. The random vectors d: Ω → R m , T:2

Ω → Mat (s, m 1 , R), W: Ω → Mat (s, m 2 , R) and h: Ω → R s are on the

Probability space (Ω, A, P) defined.

The sets A ⊆ R m and B ⊆ R m are only assumed to be convex and must be


1 2

consequently have no polyhedral shape. The two-stage stochastic mixed

integer linear model (4.2) agrees with (4.8) for E = R m 1 and F = R m 2

match. The reformulation can be done again

min {cx + φ (x, ω): x ∈ X ∩ E}


x

With

φ: X × Ω → R

x × ω ↦ → min {d (ω) y: W (ω) y = h (ω) - T (ω) x, y ∈ Y ∩ F}


y

https://translate.googleusercontent.com/translate_f 68/143
2/29/2020 Untitled

use. The purely expectation-based approach leads to the optimization model:

min {Q E (x): x ∈ X ∩ E}
x

If one again assumes a discrete distribution of the random vector (d, T, W, h),
the equivalent representation is:
N
{ Σ
min cx + πjdjyj :Tjx+Wjyj=hj
x, y
j=1
(4.9)
x ∈ X ∩ E, y j ∈ Y ∩ F

j = 1, ..., N

Analogous to the mixed-integer linear case, the equivalent


Construct representations for the expected value risk models if the

Random vector (d, T, W, h) is distributed discretely.

4.3 A look at multi-level stochastic


models
In the two-stage stochastic optimization models defined above,
de from a complete knowledge of the random information about

Page 65

4.3. A VIEW OF MULTI-STAGE STOCHASTIC MODELS 57

at a fixed time. In practice and especially in the chapter


2 presented optimization models, which carry out a feed optimization

this information is often only obtained over time. Multi-level sto-


chastic optimization models are able to map this property.
Below is just an insight into the modeling techniques of the multi-stage
mixed-integer stochastic optimization models are given. For

for further information, see [2, 9, 57, 65]. The following will also appear
[33] cited. Consider one of the stages to be observed in T ∈ N
https://translate.googleusercontent.com/translate_f 69/143
2/29/2020 Untitled
discrete system. A multi-level mixed-integer model can be used with the
Vectors c t , b t and matrices A tτ , which are to be understood as data sets, and the

Polyhedra with integer requirements X t for τ = 1, ..., t and t = 1, ..., T as


can be applied as follows:

{ TΣ t
min ctxt :∑ τ=1 A tτ ≤ b t
x 1 , ..., x T
t=1

x t ∈ X t , t = 1, ..., T

All the sizes mentioned above are dimensioned appropriately.

allowed. Assuming that the above records are random


a stochastic process defined on the probability space (Ω, F, P)
{ξ t } T
considered t = 1 . You get a multi-stage stochastic mixed

integer model. This can be done if the purely expected value-based

approach is presented as follows:


T
{∫ Σ ∑ t τ = 1 A tτ (ξ t ) ≤ b t (ξ t )
min c t (ξ t ) x t P (dω):
x 1 , ..., x T
Ω
t=1

x t ∈ X t , t = 1, ..., T

It is also assumed that due to their randomness


now available functions c t , b t and A tτ for τ = 1, ..., t and t = 1, ..., T affin-
are linear. So far, no restrictions have been made to what extent

random events of the individual levels may influence. In this way two
if variables ξ t and ξ t with t, t ∈ {1, ..., T} and t = t are quite different from each other
be dependent. A dependence of the random events on the decisions
As in the two-stage case, gen is not permitted. With one to be met

Page 66

58 CHAPTER 4. STOCHASTIC OPTIMIZATION

Up to this point, the decision in a stage 1 ≤ t ≤ T can be made by the entire group
Information from the stochastic process can be used so that the decision

determination is moved to the last stage T. A sequence should now not be


https://translate.googleusercontent.com/translate_f 70/143
2/29/2020 Untitled

anticipatory decisions are made, assuming oBdA


that the decision in the first stage can be made randomly,
ie:

decide x 1 → observe ξ 2 → decide x 2 (x 1 , ξ 2 )

→ ...

→ observe ξ T → decide x T (x 1 , ..., x T − 1 , ξ 2 , ..., ξ T )

If we consider the σ-algebras F t : = σ (ξ 1 , ..., ξ t ) for 1 ≤ t ≤ T, then F t ⊆ F t + 1 applies


for all t = 1, ..., T −1, since the available information increases over time
and refine the σ-algebras. According to the previous

set is F 1 = {Ω, ∅}. F T = F should also apply. To actually get the one you want
It must be possible to keep to the order of decisions and observations
at the time t ∈ {1, ..., T} the decision variable x t F t be measurable. The-
This measurability condition can be represented mathematically, but is on
this point on their presentation due to the additional required for this
Considerations not discussed in detail and instead referred to [6, 33, 34]
sen. The measurability conditions lead to additional linear restrictions. INS
Overall, you get the desired multi-level mixed-integer
stochastic optimization model, which the order of decisions
and observes. The inclusion of risk-averse modeling
approaches are much more complex in the multi-stage case. This topic
is dealt with in more detail in [5, 20, 34, 33, 58].

4.4 Applications

The numerical examples shown in Chapter 7 concentrate on the


Application of the two-step approach. Its applicability to that described in Chapter 2
The optimization models presented should now be checked.

Page 67

https://translate.googleusercontent.com/translate_f 71/143
2/29/2020 Untitled

4.4. APPLICATIONS 59

Feed-in optimization It is assumed that for the period T =


{0, ..., t max } up to the point in time t ∗ ∈ T the input data are safe and down

uncertainties occur at the time t ∗ + 1. Thus there is a breakdown of the


Period in the stages T 1 = {0, ..., t ∗ } and T 2 = {t ∗ +1, ..., t max } before. given
are then those defined on the discrete probability space (Ω, A, P)
random variables

(T)
PD k
:Ω→R+∀k∈V∀t∈T2,
max, (t)
PG k
:Ω→R+∀k∈V∀t∈T2,
max, (t)
d kl : Ω → R + ∀ kl ∈ E ∀ t ∈ T 2 ,

model the random input data of the second stage. The maximum
possible feeds of the generators and maximum possible transmission
capabilities of the lines are here, in contrast to the original model
lation, considered as time-dependent in order to provide a more precise illustration of the

allow. Furthermore, the production and transmission costs in the second


Level can be regarded as random. When using IG modeling
(T) (T)
the connection P k∈V PD be exploited.
D =∑ k

Optimal expansion of an energy transmission system


Model for the expansion of the electrical energy transmission system is currently
lem then a practical application when load data and production costs as
be adopted stochastically. The step separation is fundamentally different here
opted for pure stochastic feed-in optimization. Here are all
Variables that describe network expansion, first stage variables, whereas all

other second stage variables are. The network expansion takes place completely in the first
Stage instead. The electrical load is covered in the second stage.
men. For the second stage there is usually a long period, the decades
simulated, scheduled.

Remarks Apparently has a production decision in the first stage


effects on the second only if the time stages are linked

https://translate.googleusercontent.com/translate_f 72/143
2/29/2020 Untitled

Page 68

60 CHAPTER 4. STOCHASTIC OPTIMIZATION

are otherwise a separate minimization can take place over each individual level
te. Couplings of the time stages occur when optimizing the feed-in

the existence of minimum terms and minimum downtimes, but also


due to the need to consider performance gradients. at
determining the optimal expansion of an electrical energy transmission
systems, the coupling is even stronger because of the way the electrical

Load in the second stage on the one hand from the load values, but above all also
from the available electrical energy transmission system
hangs. With pure stochastic feed-in optimization, coincidence can
can be modeled more simply using the multi-level approach, in which here only one

brief insight was given. But was also during the emergence of [33]
the progress of the present work is noted. The DC modeling approach
was also used in [33] in some numerical test calculations, which
Excerpts of their results are shown in Chapter 7.

https://translate.googleusercontent.com/translate_f 73/143
2/29/2020 Untitled

Page 69

Chapter 5

Decomposition

As the number of scenarios increases, the solving of the equivalent representation


lungs for the two-stage stochastic optimization models, which were
chapters were presented, increasingly difficult. For one thing, the
Solution effort for the relaxed problems due to the number of scenarios
increasing number of restrictions too and secondly the growing
Number of integer variables enormously affects the size of the branch & bound trees
influence. Overall, this can have a significant impact on computing time,
which is required to determine a permissible point of acceptable quality. in the
In the following, options are addressed, such as
number of such models remain treatable. For this purpose, in Section 5.1, a de-
composition method for the two-stage stochastic mixed-integer linear
https://translate.googleusercontent.com/translate_f 74/143
2/29/2020 Untitled

Model class presented. Section 5.2 also shows a decomposition procedure for
the two-stage stochastic mixed-integer convex model class. at
These algorithms become special structures of the equivalent representations
exploited. The determination of a solution can thus clearly

be designed more effectively than when using standard procedures of


Would be the case.

61

Page 70

62 CHAPTER 5. DECOMPOSITION PROCEDURE

5.1 Mixed-integer linear model classes

The linear model class for the two-stage stochastic mixed-integer


Applicable scenario decomposition can be found, for example, in [12, 45]
become. Since this is also in section 5.2 when developing an algorithm
for the two-stage stochastic mixed-integer convex model class one
plays a crucial role, it is presented below. In principle, this shows
a similar structure to the branch & bound
Procedure, but it also uses heuristic approaches. Below will be
some names for the formulation of the complete algorithm
leads. With P the list of open problems becomes, with ϕ LB (P) a lower bound
for the problem P ∈ P and with ¯ϕ the currently best upper bound for that
general expected value risk model (4.5), on which the decomposition process
driving to be explained. If the problem P ∈ P is inadmissible
the lower bound is defined with ϕ LB (P): = + ∞. Inscribed further
τ ≥ 0 the relative tolerance parameter, with the help of which the quality of the
Algorithm is evaluated, and Z P the permissible
Scope of a problem P ∈ P. The scenario decomposition is with the
lower bound ϕ ∈ {R, + ∞} initialized, choosing + ∞ if none

https://translate.googleusercontent.com/translate_f 75/143
2/29/2020 Untitled
lower bound can be specified. The algorithm is as follows:

Algorithm 5.1 (scenario decomposition)


(S0) initialization
Put P: = {(4.5)} and ϕ = + ∞.

(S1) termination
φ-φ
If ϕ = + ∞ and φ
≤ τ or P = ∅, then terminate with ¯x, for

that ¯ϕ = Q E (¯x) + ρ · Q R (¯x) or determine the inadmissibility of (4.5).

(S2) barrier formation


Choose and remove problem P ∈ P. Compute lower bound ϕ LB (P),

set φ: = min {φ LB (P), φ}, and determine a point x ∈ Z P . If not


permissible point ¯x ∈ Z P can be found, then go to

Page 71

5.1. MIXED-INTEGRAL LINEAR MODEL CLASSES 63

(i) (S1) if ϕ LB (P) = + ∞.

(ii) (S4) if ϕ LB (P) = + ∞.

(S3) cut off


If ϕ LB (P) = Q E (¯x) + ρ · Q R (¯x) (optimality) or ϕ LB (P)> ¯ϕ
(Inferiority), then go to (S1). If Q E (¯x) + ρ · Q R (¯x) <¯ϕ, then
¯ϕ: = Q E (¯x) + ρ · Q R (¯x).

(S4) branching

Create two new sub-problems by partitioning the allowable set


von P. Add these partial problems to P and go to (S1).

For a partitioning of the admissible amount of the first stage in (S4)

in order to be able to preserve mixed-integer linear sub-problems, only linear ones


Secondary conditions used. Could be a problem in (S2), for example
https://translate.googleusercontent.com/translate_f 76/143
2/29/2020 Untitled

P ∈ P no permissible point x ∈ Z P can be determined, then a component


i ∈ N of the vector x chosen (denoted by x (i) ) and two new sub-problems
generated by adding one of the following restrictions to P.
These have the description:

x (i) ≤ ⌊X (i) ⌋
xi∈Z⇒
x (i) ≥ ⌊x (i) ⌋ + 1

x (i) ≤ ⌊x (i) ⌋ - ϵ
xi∈R⇒
x (i) ≥ ⌊x (i) ⌋ + ϵ

In step (S2) an x is always generated, but this is not mandatory

x ∈ Z P must meet. The parameter ϵ> 0 can be specified appropriately and prevents
a possibly infinite branching on continuous components. The
The quality of the generated point ¯x ∈ Z P can be determined during the algorithm
generated barrier ϕ can be evaluated. To illustrate the barrier formation

in (S2) the expected value-risk model with the expected excess


considered as a measure of risk and the equivalent wording (4.7). The scenario
decomposition can also be purely analogous to the following illustrations
model based on the expected value or on the expected value risk model

Page 72

64 CHAPTER 5. DECOMPOSITION PROCEDURE

the risk measure excess probability are applied. First,


introduced the copies x j for j ∈ {1, ..., N} of x and reformulated (4.7) by

the additional condition x 1 = ... = x N is also included. It will now


N
assumed that this restriction over ∑ j=1 H j x j = 0 with suitable ones
Matrices H j ∈ Mat (l, m 1 , R) is shown. The result is:

N
{ NΣ Σ
min (cx j + π j d j y j ) + ρ πjvj :Tjxj+Wjyj=hj
x, y, v
j=1 j=1

cx j + d j y j - η ≤ v j
https://translate.googleusercontent.com/translate_f 77/143
2/29/2020 Untitled
(5.1)
∑ Nj = 1 H j x j = 0

x j ∈ X, y j ∈ Y, v j ∈ R +

j = 1, ..., N

N
Without the condition ∑ j=1
H j x j = 0, the coupling constraint, would be the re-
formulation (5.1) separable with regard to each individual scenario. This is the one already
At the beginning of the chapter, the special structure of the two-stage stochastic
tables mixed-integer linear models. By removing the coupling
constraint from (5.1) is the resulting relaxation, if the
this makes it possible to swap totals with the minimized
task is performed (decomposition effect) via

Σ Pj (5.2)
j=1

With

{
P j = min cx j + π j d j y j + ρ · π j v j :Tjxj+Wjyj=hj
xj,yj,vj

cx j + d j y j - η ≤ v j

x j ∈ X, y j ∈ Y, v j ∈ R +

given. By minimizing the relaxation above, a lower bound for


(5.1) and solutions x ∗
j ∈ X for j ∈ {1, ..., N}, which are certainly not necessary

dly meet the coupling constraint. Hence heuristics


N
applied, such as the weighted average x = ∑ πjx∗
j=1 j to admissibility

Page 73

5.1. MIXED-INTEGRAL LINEAR MODEL CLASSES 65

manufacture. In the case of integer requirements, the corresponding compo-


elements of the vector x rounded. To improve the barrier formation can also

the Lagrangian dual approach. With those defined for j = 1, ..., N

https://translate.googleusercontent.com/translate_f 78/143
2/29/2020 Untitled
features

L j (x j , y j , v j , λ): = π j (cx j + d j y j + ρv j ) + λ H j x j

is the Lagrangian function over


N

L (x, y, v, λ) = Σ L j (x j , y j , v j , λ)
j=1

given. The Lagrangian dual problem results in:

Max {D (λ): λ ∈ R l } (5.3)


λ

Here is

{ NΣ
D (λ) = min L j (x j , y j , v j , λ): T j x j + W j y j = h j
xj,yj,vj
j=1

cx j + d j y j - η ≤ v j ,
x j ∈ X, y j ∈ Y, v j ∈ R +

j = 1, ..., N

The above optimization problem can again be separated with respect to each individual scenario
N
become. Hence D (λ) = ∑ j=1
D j (λ) with

{
D j (λ) = min L j (x j , y j , v j , λ): T j x j + W j y j = h j
xj,yj,vj

cx j + d j y j - η ≤ v j

x j ∈ X, y j ∈ Y, v j ∈ R +

for j = 1, ..., N. It is easy to see that D: R l → R is a piecewise linear


is concave function. Since max {D (λ): λ ∈ R l } = −min {−D (λ): λ ∈ R l }, applies
to minimize the piecewise linear convex function −D. With help of a
Subgradient method can do this task, but much more effectively with
Solve the ConicBundle method [35].

Page 74

https://translate.googleusercontent.com/translate_f 79/143
2/29/2020 Untitled

66 CHAPTER 5. DECOMPOSITION PROCEDURE

Note 5.1
Algorithm 5.1 is finite if X ⊂ Z m 1 and is limited. However, X is
mixed-integer, you need an abort criterion in order to
prevent constant components, which if bei> 0

given is.

5.2 Mixed-integer convex model classes


While the scenario composition is mixed on two-stage stochastic

Integer linear problem classes can be used, as can be


IG or DC modeling with stochastic input data
in the following a decomposition procedure for two-stage stochastic mixed-
integer convex optimization models are addressed to the RDCOL
To be able to treat the model with stochastic input data. Is presented
using the external approximation method from Section 3.2
Problems according to definition 4.3 applicable algorithm. This problem class
is shown again here:

{
min cx + d (ω) y: T (ω) x + W (ω) y = h (ω)
x, y
(5.4)
x ∈ X ∩ E, y ∈ Y ∩ F

X⊆Rm ,Y⊆Rm 1 2 are polyhedra with integer requirements and E ⊆

R m , F m R m convex sets that do not have to be polyhedral. Further


1 2

it is assumed that convex and differentiable functions e k : R m → R 1

for k = 1, ..., p and f l : R m → R for l = 1, ..., q exist with E = {x ∈ R m


2 1 :
e k (x) ≤ 0, k = 1, ..., p} and F = {y ∈ R m 2 : f l (y) ≤ 0, l = 1, ..., q}. That for
Model (5.4) developed decomposition method is based on the

purely expected value-based approach explained. The model description is for


then discretely distributed random variables over

{ NΣ
min (cx + π j d j y j ): T j x + W j y j = h j
x, y
j=1
(5.5)
x ∈ X ∩ E, y j ∈ Y ∩ F

j = 1, ..., N

https://translate.googleusercontent.com/translate_f 80/143
2/29/2020 Untitled

Page 75

5.2. MIXED-INTEGRAL CONVEX MODEL CLASSES 67

given. Since the sets E and F are convex but not in polyhedron
position, model 5.5 can be used if the RDCOL

Modeling the case, in mixed-integer nonlinear, but probably convex


Description available. Again, according to section 3.2, adapted to the
The present model class (5.5) meets the following requirements:

(i) There is a polytope Z M 0 With

Z M ⊃ {(x, y) ∈ R m × R N · m
0 1 2 : e k (x) ≤ 0, k = 1, ..., p

f l (y j ) ≤ 0, l = 1, ..., q ,

j = 1, ..., N

(ii) There is a K ∈ R with ∇e k (x) ≤ K and ∇f l (y 1 ) ≤ K for all


(x, y) ∈ Z M 0 with 1 ≤ k ≤ p and 1 ≤ l ≤ q.

The master problem in a step ξ ∈ N 0 is over


N
{ Σ
min cx + πjdjyj :Tjx+Wjyj=hj
x, y
j=1

(M ξ ) x ∈ X, y j ∈ Y

(x, y 1 , ..., y N ) ∈ Z M ξ

j = 1, ..., N

stated. Again, a point permissible for (M ξ ) (˜x ξ , ˜y 1, ξ , ..., ˜y N, ξ ) is considered


tet, the given tolerance τ ξ ≥ 0 and the bound ϕ valid for (M ξ ) ξ

the tolerance condition


N
c ˜x ξ + ∑ j=1 π j d j ˜y j, ξ - ϕ ξ
≤τξ (5.6)
φξ

Fulfills. The sets Z M ξ are now analogous to the master problems from Al
algorithm 3.1 built. Here, however, it is important to check whether the non-linear ones
First stage restrictions are met by the first stage decision or
whether the non-linear restrictions of the second stage by all possible second

https://translate.googleusercontent.com/translate_f 81/143
2/29/2020 Untitled
step decisions are met. In a step ξ ∈ N the sets are then
about the context

Page 76

68 CHAPTER 5. DECOMPOSITION PROCEDURE

ZM ξ+1 = ZM ξ ∩
{
(x, y) ∈ R m × R N · m
1 2 : e v (˜x ξ ) + ∇e v (˜x ξ ) (x - ˜x ξ ) ≤ 0

f w (˜y j, ξ ) + ∇f w (˜y j, ξ ) (y j - ˜y j, ξ ) ≤ 0
j j

∀ v ∈ {k ∈ {1, ..., p}: e k (˜x ξ )> 0}

∀ w j ∈ {l j ∈ {1, ..., q}: f l (˜y j, ξ )> 0}


j

j = 1, ..., N

given. Since Z M is assumed to be a polytope, it is limited and therefore also


0

every master problem (M ξ ) given in a step ξ ∈ N 0 is limited. The

Master problems (M ξ ) are now dealt with using the scenario decomposition
to be able to take advantage of their favorable structural properties. The algorithm
then looks like this:

Algorithm 5.2
(S0) initialization
Find a polytope Z M , choose τ 0 ≥ 0, ϕ
0
0
: = + ∞ and set ξ = 0.

(S1) optimization
Initialize the scenario composition with the lower bound ϕ ξ
, the
Tolerance parameters τ ξ and apply them to the master problem (M ξ ).

(S2) termination
If e k (˜x ξ ) ≤ 0 for all k = 1, ..., p and f l (˜y j, ξ ) ≤ 0 for all l = 1, ..., q

and j = 1, ..., N or Z M = ∅, then terminate.


ξ

(S3) updating

https://translate.googleusercontent.com/translate_f 82/143
2/29/2020 Untitled
Choose 0 ≤ τ ξ + 1 ≤ τ ξ , set ϕ ξ+1 : = ϕ ξ , ξ: = ξ + 1 and go to (S1).

After construction of the algorithm, Clause 3.4 applies here analogously. If the procedure
stops after finally many steps, so a permissible point became more acceptable

Found goodness or found the inadmissibility of model (5.5). In the event of


non-termination is carried out taking into account the analogue to Theorem 3.5 and
Note 5.1 stated:

Page 77

5.2. MIXED-INTEGRAL CONVEX MODEL CLASSES 69

Note 5.2
If algorithm 5.2 generates the infinite sequence (˜x ξ , ˜y 1, ξ , ..., ˜y N, ξ ), {τ ξ } → 0
and X ⊂ Z m 1 is, each cluster point of the sequence is a solution of (5.5).
However, if X is a mixed integer, then for a step ξ ∈ N 0 in (S1)

Determination of a point with the tolerance parameter τ ξ condition (5.6)


fulfilled, not necessarily possible. This depends on the according to note 5.1 in
used in the scenario decomposition to prevent
infinite branching together on continuous components. Under the

additional requirement that for fixed tolerance parameters τ ξ : = τ for all


ξ ∈ N 0 in (S1) a point satisfying (5.6) can be determined is certain
any cluster point (˜x, ˜y 1 , ..., ˜y N ) of the sequence (˜x ξ , ˜y 1, ξ , ..., ˜y N, ξ ) allowed for (5.5)
N N
and fulfills the condition c ˜x + ∑ π j d j ˜y j ≤ (1 + τ) (cx ∗ + ∑ πjdjy∗
j=1 j=1 j ), where

(x ∗ , y ∗
1 , ..., y ∗ N ) denotes the optimal solution to (5.5).

https://translate.googleusercontent.com/translate_f 83/143
2/29/2020 Untitled

Page 79
78

Chapter 6

A heuristic for that


AC model

https://translate.googleusercontent.com/translate_f 84/143
2/29/2020 Untitled

In the two previous chapters, the model class of the AC model is

neglected due to their mixed-integer non-convex representation


the. But precisely because this type of modeling
the exact slope, it should be examined further to determine whether at least
least heuristic approaches can be developed to get good allowable points

to be able to determine. Using a few example calculations,


reduced model formulations, ie the IG, DC and DCOL modeling, the
AC model are compared. First, section 6.1 shows up again
the establishment of relationships between sections set out in section 2.4

the DC and AC model, so that in addition to the existing


to be able to construct another and better barrier. In section 6.2
In addition to the previous AC model, a transformed representation is presented,
the procedures for network state
determination builds up. Section 6.3 finally presents heuristic approaches for the
mixed-integer nonlinear AC model for both deterministic and
for stochastic data.

71

Page 80

72 CHAPTER 6. A HEURISTICS FOR THE AC MODULE

6.1 Quality assessments

Another lower bound can be used to represent the AC model in the


construct the binary variables with the help of the DC model.
Initially, transmission costs and start-up costs are ignored, ie it is E =
∅ and γ k = 0 ∀ k ∈ V. If under these conditions z DC a lower one
Represents the bound of z AC , the following equivalence chain ([.] Denotes

again the value of the relevant variable with regard to the optimal solution):

https://translate.googleusercontent.com/translate_f 85/143
2/29/2020 Untitled
z DC ≤ z AC
⇔ Σ Σ C k ([P (T)
] DC , [s
(T) Σ Σ (T)
C k ([P G ] AC , [s
(T)
Gk
k ] DC ) ≤ k
k ] AC )
t∈T k∈V t∈T k∈V
(T) (T) (T)
⇔ Σ Σ C k (P D + Σ [p
k
kl ] DC , [s k ] DC )
t∈T k∈V l∈N k


(T) (T) (T)
Σ Σ C k (P D + Σ [p
k
kl ] AC , [s k ] AC )
t∈T k∈V l∈N k

( )
(T) (T)
⇔ Σ Σ αk Σ [p
kl ] DC + β k [s k ] DC
t∈T k∈V l∈N k


( )
(T) (T)
Σ Σ αk Σ [p
kl ] AC + β k [s k ] AC
t∈T k∈V l∈N k

The chain of equivalences would be truthful if one considers the general


dingung

( ) ( )
Σ αk Σ [p
(T) (T)
≤ Σ αk Σ [p
(T) (T)
∀t∈T
kl ] DC + β k [s k ] DC kl ] AC + β k [s k ] AC
k∈V l∈N k k∈V l∈N k

(6.1)

could prove. Reminding of DC modeling is known for everyone


(T) (T)
kl ∈ E and t ∈ T the relationship [p
] DC + [p lk ] DC = 0 applies. To thiskl

Take advantage of the facts and without the DC and AC modeling in their way
Every pure load node can be changed from the set V \ V into one

Transform feed nodes. You do this at the knot, which is


are only load nodes, but now an imaginary generator is assumed

Page 81

6.1. G¨UTEABSCH¨ATZUNGEN 73

the associated minimum and maximum infeed values for both


Active power as well as reactive power feed to zero, so that consequently V = V

applies. Furthermore, the quantity V denotes the generators actually present. The

https://translate.googleusercontent.com/translate_f 86/143
2/29/2020 Untitled
Equivalent condition to (6.1) is:
( ) ( )
(T) (T) (T) (T)
Σ α k Σ [p ≤ Σ αk Σ [p ∀t∈T
kl ] DC + β k [s k ] DC kl ] AC + β k [s k ] AC
k∈V l∈N k k∈V l∈N k

It has not been possible to prove this condition, so that between DC and

AC model generally no relation could be established. For certain cases


le, in which the cost functions and switching states of the power generation
conditions, but can be a lower limit for the AC model
be constructed. Consider the following model:

S-model: creation of the barrier

(i) Use the DC model with E = ∅ and V = V.

(ii) For all k ∈ V use the cost functions C k (P G , s k ) = αP G + β k s k with


k k

α: = min {α k : k ∈ V}. Choose β k > 0 for all k ∈ V \ V.

(T) (T) (T) (T)


(iii) Add the restrictions s k = [s k = [s
k ] DC (see k ] RDCOL ) ∀ k ∈ V ∀ t ∈ T
added.

(iv) Set P min Gk = P max


Gk
= 0 for all k ∈ V \ V.

Depending on the choice of an optimal switching assignment in step (ii),


also spoken of the S (DC) or S (RDCOL) model. From proofing techniques

The following model is also constructed for reasons:

SAC model: necessary in terms of proof

(i) Use the AC model with E = ∅ and V = V.

(ii) For all k ∈ V use the cost functions C k (P G , s k ) = αP G + β k s k with


k k

α: = min {α k : k ∈ V}. Choose β k > 0 for all k ∈ V \ V.

Page 82

https://translate.googleusercontent.com/translate_f 87/143
2/29/2020 Untitled
74 CHAPTER 6. A HEURISTICS FOR THE AC MODULE

(T) (T) (T) (T)


(iii) Add the restrictions s k
= [s k
= [s
k ] DC (see k ] RDCOL ) for all k ∈ V
∀ t ∈ T added.

(iv) Set P min Gk = P max


Gk
= Q min
G k
= Q max
G k
= 0 for all k ∈ V \ V.

(T) (T)
From the assumptions (iv) it follows in particular that [s
] S = [s k k ] SAC = 0 for everyone
k ∈ V \ V and t ∈ T applies. Now the optimal value z S of the S model
a lower bound for the SAC model. This can be verified as follows
the optimal switching assignment of the DC model is used here:

z S ≤ z SAC
( )
(T) (T)
⇔ Σ Σ αk Σ [p
kl ] S + β k [s k ]S
t∈T k∈V l∈N k


( )
(T) (T)
Σ Σ αk Σ [p
kl ] SAC + β k [s k ] SAC
t∈T k∈V l∈N k

( )
⇐ Σ αk Σ [p
(T) (T)

kl ] S + β k [s k ]S
k∈V l∈N k


( )
(T) (T)
Σ αk Σ [p ∀t∈T
kl ] SAC + β k [s k ] SAC
k∈V l∈N k

⇔ Σ α Σ [p
(T) Σ β k [s
(T)

kl ]S+ k ] DC
k∈V l∈N k k∈V


(T) (T)
Σ α Σ [p Σ β k [s ∀t∈T
kl ] SAC + k ] DC
k∈V l∈N k k∈V

⇔0≤α Σ Σ [p
(T)
∀t∈T
kl ] SAC
k∈V l∈N k

} {{
≥0 }

If the RDCOL decision is based on the S model, the above considerations apply

analogous trends. Due to the shift assignment fixed for the S model, the
is also used in the SAC model, you can see immediately that z S ≤ z SAC also applies,
if start-up costs in the S and SAC models are observed. transmission costs
but must be ignored in the S model. For those presented in section 6.3

https://translate.googleusercontent.com/translate_f 88/143
2/29/2020 Untitled

Page 83

6.2. A TRANSFORMATION APPROACH 75

heuristic solution method, the FAC model defined as follows is


be sam. Generators intended to be used only for proof
of the above facts were waived.

FAC model

(i) Use the AC model.

(T) (T) (T) (T)


(ii) Add the restrictions s k = [s k = [s
k ] DC (see k ] RDCOL ) for all k ∈ V and
t ∈ T added.

Depending on the choice of the optimal switching assignment in step (ii), the
F AC (DC) or F AC (RDCOL) model spoken. The relatio-

z AC ≤ z F AC because the switching assignment selected in the FAC model is not optimal
for the AC model, and z SAC ≤ z F AC due to the

changed cost functions. If the S model is permissible, the


context

z S ≤ z SAC ≤ z F AC ,

so that z S alongside z RIG represents a lower bound for the FAC model.

6.2 A transformation approach

In addition to the AC model shown in section 2.2, a is shown below


additional assumptions equivalent reformulation addressed. This trans-

the AC model using a classic method for grid connection


status determination of electrical energy transmission systems. The type of network access
status determination comes in common software packages for load flow calculation,
such as the PowerWorldSimulator [56]. The transformed AC

But model is also shown here to show that in this


Work the methods and procedures developed in engineering
analyzed in detail and their applicability to the problem here
https://translate.googleusercontent.com/translate_f 89/143
2/29/2020 Untitled

positions was checked. It should also be considered whether this changed

Page 84

76 CHAPTER 6. A HEURISTICS FOR THE AC MODULE

Representation offers advantages over previous ones. The transformed Dar-


position initially uses a Cartesian representation of the complex voltages

so that a mathematical description of the load flow


without using trigonometric expressions. One is
further able to use the Newton method to specify a function that
from the active power feed-in and amount of voltage at the

generation plants (any generation plant except the one on the slack
nodes connected) to the reactive power supply of the voltage-controlled
Generation plants and on the complex feed of the generation plant on
Maps the slack node in such a way that the equations (2.6) - (2.12) are fulfilled. The exact one

The way to this function is due to the extensive


styles not explained in detail, but is described in detail in [4, 8, 28].
In principle, with the parameters

˜g kl + j˜b kl = −g kl - jb kl k, l = 1, ..., n with k = l,

˜g kl + j˜b kl = Σ (g kl + j (b kl + b 0
kl )) k, l = 1, ..., n with k = l
l∈N k

and the load flow equations


n

PG -PD -
k k
Σ (e k (˜g kl e l - ˜b kl f l ) + f k (˜g kl f l + ˜b kl e l )) = 0 ∀ k ∈ V, (6.2)
l=1
n

QG -QD -
k k
Σ (f k (˜g kl e l - ˜b kl f l ) - e k (˜g kl f l + ˜b kl e l )) = 0 ∀ k ∈ V, (6.3)
l=1

which are now in Cartesian representation and the conditions

U2
k- e2 k - f2 k =0∀k∈V

https://translate.googleusercontent.com/translate_f 90/143
2/29/2020 Untitled
set up a nonlinear system of equations using the Newton method
is solved, so that as a result the chip hitherto unknown in the nodes k ∈ V
U k = e k + jf k can be specified. Since the entire voltage

distribution is known, the still unknown quantities, such as the energy flows
and reactive power feeds can be determined quickly. So you get to
differentiable functions L P G1 ,LQ Gk for k ∈ V and L d 2
kl
for kl ∈ E that on

the set D = ( × i∈V \ {1} [P min Gi , P Gmax ]) × ( ×


i
i∈V \ {1} [U min i , U imax ]) are defined,

Page 85

6.2. A TRANSFORMATION APPROACH 77

and the active power feed-in of the energy generation plant at the Slackkno-
ten, the reactive power feeds of the power generation plants and on the

Map the squares of the load flows on the respective edges. That transformed
The AC model, hereinafter referred to as the TAC model, can be constructed in this way.
The restrictions must be observed

P Gmin k
·Sk≤PG k ≤ P max
G k
· S k ∀ k ∈ V \ {1},

U kmin ≤ U k ≤ U max
k ∀ k ∈ V \ {1},

the power generation plant at the Slack node now has no feed limit
is subjected to zen (cf. [8, 28]). Further with the vector P G , which

summarizes active power feed-in variables, and with the vector U G


the secondary condition in an analogous definition

LP G1 (P G , U G ) ≥ 0

to be fulfilled, since the active power supply of the slack generator in the transfor-
presentation does not represent an optimization variable and is only a result
from the selection of the other optimization variables. The limitation of

Reactive power feeds lists

Q Gmin ≤ L Q
k
Gk (P G , U G ) ≤ Q maxG k
∀ k ∈ V \ {1}.

https://translate.googleusercontent.com/translate_f 91/143
2/29/2020 Untitled

The load flows are over

L d (P G , U G ) ≤ (d maxkl
2 ) 2 ∀ kl ∈ E
kl

to limit. The optimization criterion equivalent to the AC model can be over

C 1 (L P G1 (P G , U G )) + Σ C k (P G ) +
k
Σ E Tkl AC (L d (P G , U G )) → min
2
kl

k∈V \ {1} kl∈E

can be scheduled. In order to be able to consider transmission costs,

E Tkl AC (d 2
kl ) = δ kl √d 2 kl scheduled. Note that for a small with

Ld 2
kl
(P G , U G ) = 0 the objective function cannot be differentiated. The Cartesian load
Flow equations are avoided as secondary conditions by means of the transformation
have been and are now implicitly observed in the objective function. The extension

Page 86

78 CHAPTER 6. A HEURISTICS FOR THE AC MODULE

for an optimization period is analogous to section 2.3, with the


power generation system connected to the Slack node neither minimum running

or minimum downtimes are still subject to performance gradients. The


sophisticated mixed-integer nonlinear model class remains
obtained using the transformation approach, since the functions L P G1 ,LQ Gk

for k ∈ V and L d 2 for kl ∈ E in a non-convex description. Between


kl

the AC and TAC models have the following relationship:

comment
Since the slack generator in the TAC model has no technical constraints

is subject to z T AC ≤ z AC . Are the records still like this


procure that P min G1 = 0, Q minG 1
= −∞, P max G1 = Q max
G 1
= + ∞ and the energy
generation facility at the Slack node no minimum run times, minimum downtimes
and subject to power gradients, then z T AC = z AC .

https://translate.googleusercontent.com/translate_f 92/143
2/29/2020 Untitled

Due to the mathematical description of the TAC model, the in section


3.1 Problems already mentioned when using a branch & bound

The procedure is also relevant here. Solutions were also considered, the
the variable products appearing in the load flow equations (6.2) and (6.3)
appropriately approximate to give a Cartesian representation
To convert AC modeling into an overall mixed-integer linear form

to lead. The number of required binary variables increases with the number of
Network nodes disproportionately strong. Only for very small networks under consideration
At one point in time permissible points could be determined. More single
units about such approximation options can also be found in [47]

become. In the following subsection, heuristic approaches for the


AC model presented.

6.3 Heuristic Approaches


The model class of the AC model leads to the algorithm-
difficulties. The following are intended to use heuristic approaches

are presented with which good permissible points can be determined.

Page 87

6.3. HEURISTIC APPROACHES 79

Deterministic data situation By first solving the approximate


DC or RDCOL models are supposed to be the integer requirements in the AC model
be eliminated. The heuristic is as follows:

Heuristics 6.1
(i) Solve the DC model (RDCOL model).

(ii) Replace the binary variables in the AC model with the optimal switching load
DC model (RDCOL model) ⇒ FAC model.

https://translate.googleusercontent.com/translate_f 93/143
2/29/2020 Untitled
(iii) Solve the FAC model.

An SQP- can be used to solve the FAC model in step (iii).


Processes, such as implemented in the SNOPT program package,
be applied. As further literature sources dealing with this,

nen [26, 27, 54, 62, 72]. Such procedures certainly will achieve
demonstrably only a stationary point and can check if necessary
is a locally optimal point. With the help of the beginning of this chapter
The quality of the barrier z S developed for the FAC model can be assessed.
Furthermore, the optimal value obtained by DC or RDCOL modeling
divorce can be used as the starting point for an SQP procedure.

A heuristic with two-stage stochastic modeling


Relation of random influences in the AC model is from a modeling point of view
View possible. As in Chapter 4 for the mixed-integer linear or

Convex case, a two-stage stochastic model can be


build up. Here, however, in a much more sophisticated description
under additional integer requirements. Since the basic functional
demonstrate the modeling of two-stage replacement models through the considerations

Chapter 4 has already clarified the resulting two-step model


not formulated here with AC modeling. The involvement of
Risk measures, such as when using the excess probability,
require additional binary variables that optimize the already com-
complex two-stage stochastic AC model even more difficult. following

Page 88

80 CHAPTER 6. A HEURISTICS FOR THE AC MODULE

becomes a heuristic for the two-stage, purely expectation-based AC model


presented, which in turn integer requirements by previous solving
of the approximated models.

https://translate.googleusercontent.com/translate_f 94/143
2/29/2020 Untitled
Heuristics 6.2
(i) Solve the two-step stochastic expectation-based DC model
(RDCOL model).

(ii) Replace the binary variables in the two-step stochastic expected


value-based AC model through the optimal switching assignment of the two-stage

stochastic expected value-based DC model (RDCOL model).

(iii) Solve the model from step (ii).

After using this heuristic, there are no statements about global optimality

of the solution point. A decision that, in the long run,


minimizing costs only makes sense if it is globally optimal or
at least their quality can be estimated. The considerations from section
6.1 can be applied analogously to the stochastic case, so that too

a certain estimate is made possible here. However, since


is the optimal value of the DC model as the lower bound for the AC model
Such estimates must be verified when considering a period
only very roughly possible. Therefore, the problem of a

Inclusion of stochastic effects in AC modeling


be directed. The determination of an inexpensive retrofit of an electrical
trical energy transmission system with the underlying AC modeling
equally problematic because additional binary variables for the purpose of recording possible ones

Rebuilds of edges are needed. The following chapter, in which


Numerous numerical test calculations for uncertain data situations were presented
due to these problems, it therefore focuses on deterministic
of all models presented on the stochastic DC

and RDCOL modeling.

Page 89

https://translate.googleusercontent.com/translate_f 95/143
2/29/2020 Untitled

Chapter 7

Numerical examples

The last chapter of this work shows numerous numerical example calculations
on. Section 7.1 contains some calculations with deterministic data
illustrates, also about the influence of the underlying modeling approach

to further clarify zes. Based on the calculations,


valid, but only for the examples selected here, the approximation
clarity of the DC and DCOL modeling compared to the AC model
become. The influences of observing the energy transmission system

additional factors that have become possible, such as load flow limitations or
Transmission cost shares in the target function are shown as examples.
Numerous test calculations for unsafe data sets are used
of the two-stage approach and also of the multi-stage approach, the

on the multi-stage case from [33]. In addition to the first motivating examples
in Section 7.2, in which risk-neutral and risk-averse modeling techniques
larger reference networks, the IEEE30-BUS and
IEEE118-BUS systems [55], their application. Section 7.3 shows the numerical

Efficiency of the decomposition methods presented in Chapter 5 for the two-stage


towards stochastic mixed-integer linear and convex model classes.
Section 7.4 shows test calculations to determine a cost-optimal follow-up
upgrading of an energy transmission system in order to

reduce costs and increase security of supply. The chapter shows


In conclusion, a concrete case study from the energy industry in which the

81

https://translate.googleusercontent.com/translate_f 96/143
2/29/2020 Untitled

Page 90

82 CHAPTER 7. NUMERICAL EXAMPLES

Energy transmission system of the Federal Republic of Germany is used. For


All of the test calculations presented here were used as an optimization criterion

cost-effective operation selected. Fig. 7.1 shows the 6-node reference network,
at which the basic properties of electrical
trical energy transmission systems have been explained. For many of the following
This system was used for the test calculations shown. In addition, Fig.

7.1 also the type of the nodes and the node numbers assigned to the nodes
again. The exact network parameters, the technical characteristics of the energy
Power generation plants and the underlying load data can be found in Appendix B
be removed. For all of the test calculations shown below

D D

D D

Figure 7.1: A 6-node network

was a Pentium IV system with a 3.2 Ghz clock and a 4 GB RAM


used. For the most part, the calculations were carried out on a Unix platform
a Windows platform for deterministic test calculations

https://translate.googleusercontent.com/translate_f 97/143
2/29/2020 Untitled
was used. Mixed-integer linear model classes, as with IG or

Page 91

7.1. DETERMINIST PREVIEWS 83

DC modeling available was handled with the software package CPLEX [14]
delt. The mixed-integer convex model used in RDCOL modeling

Dell class was solved with algorithm 3.1, implemented in C ++, whereby the
master problems to be solved were themselves dealt with CPLEX. Which he-
a promising point for the AC model was found
using the heuristic approaches and the SNOPT program package

[22], a powerful implementation of an SQP process. In the case of random


Depending on the model class, algorithms 5.1 and 5.2
used.

7.1 Deterministic preliminary considerations

In this section, only the 6-node network from Fig. 7.1 is used
det. In the following three subsections deterministic
shown calculations in which operational optimizations over a time without

Presence of transmission costs, over a point in time of existence


of transmission costs and over a period of time.

Feed-in optimization over one point in time

The following is based on the underlying model description, ie whether IG,


DC, RDCOL or AC modeling is available. In addition to the through the
The resulting production decisions will also result in the load flows
the total active power generated by the power generation plants is illustrated.

IG modeling: Using IG modeling results in product


https://translate.googleusercontent.com/translate_f 98/143
2/29/2020 Untitled

z IG = 287.64 units. The optimal production decision


is given via ([P G ] IG , ..., [P G ] IG ) = (7.61,0,0,0). The total generated
1 4

IG
Active power is P
= 7.61. If you put this production decision in the
G

Balance equation of the DC model results in a load flow, Table 7.1


can be removed. All entries marked with an asterisk

overload the lines. The actually maximum possible transfers

Page 92

84 CHAPTER 7. NUMERICAL EXAMPLES

1 2 3 4 5 6
* *
1 -3,208 -3,402
* *
2 3,208 -1,008
*
3 1588 0132
* *
4 1008 0814 -0,642
* * *
5 3,402 -1,588 -0,814

6 -0,132 0642

Table 7.1: Resulting load flow with IG modeling

can be found in Appendix B. The production decision made at


IG modeling is therefore not permitted for the DC model.

DC modeling: When using DC modeling, product

on cost of z DC = 455.81 units. The optimal production decision


is given by ([P G ] DC , ..., [P G ] DC ) = (0,3,3,1.61). The total generated
1 4

DC IG
Active power is P G = 7.61 and is with P G of course identical. The resultant
current load flow when using the balance equations of the DC model

Table 7.2. In contrast to the IG production decision

1 2 3 4 5 6

1 0743 0257
2 -0.743 -0,057

3 -0.827 -0.453
https://translate.googleusercontent.com/translate_f 99/143
2/29/2020 Untitled

4 0057 -0.430 -0.057

5 -0,257 0827 0430

6 0453 0057

Table 7.2: Resulting load flow with DC modeling

permissible load flows now result, since in the DC model the restrictions at (2.19),
which carry out a load flow limitation.

RDCOL modeling The tolerance parameter for the outer approximate

tion method according to algorithm 3.1 is set to ϵ = 10e-5. The


external approximation methods terminate with the desired tolerance and
Production costs of z RDCOL = 456,962 units. The optimal production

Page 93

7.1. DETERMINIST PREVIEWS 85

decision is made about ([P G ] RDCOL , ..., [P G ] RDCOL ) = (0,3,3,1.665).


1 4

RDCOL DC
The total active power generated is P G = 7,696 and is with P G

not identical since transmission losses occur. The resulting load flow
when using the balance equations of the DCOL model, Table 7.3 can be
be taken. Use of the RDCOL production decision in IG
or DC model is of course inadmissible because the RDCOL production decision
Compensates for transmission losses that are ignored in the IG and DC models.

1 2 3 4 5 6

1 0750 0259

2 -0.750 -0,040

3 -0.823 -0.443

4 0040 -0.451 -0.071

5 -0,259 0823 0451

6 0443 0071

Table 7.3: Resulting load flow with DCOL modeling

https://translate.googleusercontent.com/translate_f 100/143
2/29/2020 Untitled

FAC (RDCOL) modeling: When using the FAC model with the for
the RDCOL model optimal switching assignment result in production
vonz F AC = 457,273 units. The production decision is over
([˜P G ] F AC , ..., [˜P G ] F AC ) = (0,3,3,1,679). The determined solution and the
1 4

Objective function values are also marked with,, since they are
SNOPT determined solution is a stationary point. The total
F AC
generated active power is ˜P G = 7,679. The amount of the resulting
Load flow when using the balance equations of the AC model can be table

7.4 can be taken that has no symmetry. By using


Theorem 2.2 can determine the barrier z RIG = z IG = 287.64 for the FAC model
become. Furthermore, with the help of the S (RDCOL) model, the significantly better one
Calculate the barrier z S = 444.98. This is the quality assessment

˜z F AC - z S
100% ≈ 2.75%
zp

possible. Table 7.5 compares the results obtained after choosing the model. It
shows that the optimal, achieved by DC, RDCOL and AC modeling

Page 94

86 CHAPTER 7. NUMERICAL EXAMPLES

1 2 3 4 5 6

1 0744 0285

2 0.770 0001

3 0.831 0.442
4 0005 0.487 0.083

5 0.277 0866 0450

6 0448 0075

Table 7.4: Resulting load flow with FAC modeling

solutions differ only slightly, so that their optimal values


lie close together. The IG model, however, leads to a very different one
https://translate.googleusercontent.com/translate_f 101/143
2/29/2020 Untitled

Optimal solution and also compared to a significantly reduced optimal value


to the rest. The computing time was less than one second in all cases.

IG DC RDCOL FAC

Production costs 287.64 455.81 456,961 457,273

PG1 7.61 0 0 0
PG2 0 3 3 3

PG3 0 3 3 3

PG4 0 1.61 1665 1679

Table 7.5: Comparison of the determined optimal solutions / values

Influence of transmission costs

The influence of transmission costs on the previous


the sub-section test calculations examined. It was

assume that when transporting energy along the line 12.21 ∈ E Trans-
mission costs incurred. Hence E = {12,21} and that for the transmission
Cost functions required parameters are above δ 12 = δ 21 = 20 units
given. Table 7.6 shows the results achieved. In addition, the initial

transmission and total costs that would arise if


the share of transmission costs itself was not taken into account in the objective function.
that would be. It turns out that the presence of transmission costs is noticeable
Has an impact on the optimal solution. The circuit of the power generation

Page 95

7.1. DETERMINIST PREVIEWS 87

DC RDCOL FAC DC RDCOL FAC

δ 12 = δ 21 = 0 δ 12 = δ 21 = 20

production costs 455.81 456.96 457.27 467.64 470.03 471.38

transmission costs 29.72 30 30.28 00:32 1.84 00:08

total cost 485.53 486.96 487.55 467.96 471.87 471.46

https://translate.googleusercontent.com/translate_f 102/143
2/29/2020 Untitled
PG1 0 0 0 1.61 1.71 1696
PG2 3 3 3 3 3 3

PG3 3 3 3 3 3 3

PG4 1.61 1665 1679 0 0 0

Table 7.6: Comparison of solutions when adding transmission costs

locations varied. Due to the changed switching and adjustment of the active power

feed-in, the transmission costs incurred can be significantly reduced


the. Despite the additional production costs required for this are still
the total cost less than using the production decision for
δ 12 = δ 21 = 0. The achieved by DC, RDCOL and FAC (RDCOL) modeling

Optimal solutions / values differ only slightly.

Feed optimization over a period of time

The additional necessary for the feed optimization over a period of time

Data records, ie the load data, performance gradients, minimum running and minimum
downtimes can be found in Appendix B. Here too
uses the optimal switching assignment of the RDCOL model for the FAC model,
ie the FAC (RDCOL) model is minimized. The optimal

values z IG = 7856.53, z DC = 9226.52, z RDCOL = 9461.08 and ˜z F AC = 9467.36 input

units. The tolerance for the outer approximation method was ϵ =


10e-5 scheduled. With the help of the S (RDCOL) model, for the FAC (RDCOL) -
Model the barrier z S = 8979 units can be determined and the following gü-

Assessment is possible:

˜z F AC - z S
100% ≈ 5.44%
zp

Figures 7.2 - 7.5 show the active power feed-in achieved.

Page 96

88 CHAPTER 7. NUMERICAL EXAMPLES

https://translate.googleusercontent.com/translate_f 103/143
2/29/2020 Untitled

Figure 7.2: Optimized feed with IG modeling

Figure 7.3: Optimized infeed with DC modeling

https://translate.googleusercontent.com/translate_f 104/143
2/29/2020 Untitled

Page 97

7.1. DETERMINIST PREVIEWS 89

Figure 7.4: Optimized infeed with RDCOL modeling

Figure 7.5: Optimized infeed with FAC (RDCOL) modeling

https://translate.googleusercontent.com/translate_f 105/143
2/29/2020 Untitled

Page 98

90 CHAPTER 7. NUMERICAL EXAMPLES

7.2 Stochastic optimization: motivating


Examples

In the following, numerous test calculations for stochastic data


location, which continues the two-stage stochastic modeling approach
motivate and emphasize its practical relevance. All of these use as

Optimization criterion minimizing operating costs.

Introductory example: Consider the RDCOL model for the 6-node


Network over a period of time. The last deterministic point in time was with

t ∗ = 12. The probability space is here Ω = {ω 1 , ω 2 , ω 3 }


with (π 1 , π 2 , π 3 ) = (0.7,0.2,0.1). The data situation was like this
chosen that when the event ω 1 occurs exactly the same data situation
is present, as in the previous subsection of the purely deterministic one

Feed-in optimization was used over a period of time (see also Appendix
B). If the events ω 2 or ω 3 occur, however, a deviation of the
maximum possible transmission capacities on the lines 12.24 ∈ E for
all t ∗ + 1 ≤ t ≤ t max assumed as follows:

max, (t) max, (t) max, (t) max, (t)


d 12 (ω 2 ) = d 21 (ω 2 ) = d 24 (ω 2 ) = d 42 (ω 2 ) = 1.5
max, (t) max, (t) max, (t) max, (t)
d 12 (ω 3 ) = d 21 (ω 3 ) = d 24 (ω 3 ) = d 42 (ω 3 ) = 0

(T)
Furthermore, a possible drop in performance of P D2 (ω 2 ) =
(T)
P D (ω 3 ) = 1.5 for t ∗ + 1 ≤ t ≤ t max assumed. All other of the event
2

Sen ω 2 and ω 3 dependent data in turn agree with that of when it occurred
of the elementary event ω 1 . When using the purely expected value

https://translate.googleusercontent.com/translate_f 106/143
2/29/2020 Untitled
based on the two-stage approach, the expected mean production
cost at z RDCOL = 9331.84 units. Note that an overestimation
of the ohmic transmission losses on the lines 12.24 ∈ E
Event ω 3 occurs due to the restrictions

(T) max, (t)


ν kl (ω) ∀ t ∈ {t ∗ + 1, ..., t max } ∀ kl ∈ E ∀ ω ∈ Ω
kl ≤d

Page 99

7.2. STOCHASTIC OPTIMIZATION: MOTIVATING EXAMPLES 91

is prevented. Already in section 3.2 the physical validity


of these restrictions. Table 7.7 shows the extent to which the

random maximum transmission capacity on lines 12.24 ∈ E


can influence the optimal production. The optimal first stage
decision by solving the expected-value two-stage model
was won with the optimal decision for the 1-scenario model, whereby

the scenario ω 1 was used. It turns out that the accidental

Solution to scenario 1 Two-stage approach


t 10 11 12 10 11 12
(T)
PG 1
0 0 00:14 0 0.18 1.35
(T)
PG 2 3 3 3 3 2.7 2.1
(T)
PG 3 3 3 3 3 3 3
(T)
PG 4
2.69 2.88 2.94 2.69 3 2.6

Table 7.7: Impact of uncertainties on the optimal solution

Events here repercussions on the optimal production decision of the

have the first level, because on the one hand the levels of the two-level model with each other
are coupled and on the other hand the possible failure of the lines 12.24. E
must be compensated for before such an event occurs.

https://translate.googleusercontent.com/translate_f 107/143
2/29/2020 Untitled

Motivation of the risk approach: Using the IEEE30 BUS reference


netzes [55] is intended to further motivate the expected value risk approach in the following
become. The results presented here can, with the exception of minor adjustments,

gene, can also be found in [43]. The optimization period was twelve
Hours scheduled in one hour resolution. Furthermore, uncertainties have been removed
the tenth hour in terms of load and maximum transferability and 30
possible scenarios assumed. The purely expectation-based minimization

leads to Q E = 28368.09 units. After evaluating the resulting


The level with scenario-specific optimal values is η = 28800 units
recognized, which is exceeded with a probability of 51%.
If the expected value risk models are minimized, the result is

Page 100

92 CHAPTER 7. NUMERICAL EXAMPLES

Results shown in Table 7.8. X E , x EP and x EE identify


the determined solution vectors for the associated expected value or
Mean-risk models. Furthermore, the optimal production
positions shown at the last deterministic point in time. A clear one

(9) (9) (9)


QE Q EP Q EE PG 2
PG 5
PG 13

xE 28 368.09 0.51 25.85 10 13.8831 9.0169


x EP 28 398.00 0.27 24.99 9 13.9443 9.9557
x EE 28 400.65 0.27 24.93 9 13.9395 9.9605

Table 7.8: Influence of risk on optimal solution / value

Reduction of risk is on average with relatively low additional costs

possible. The weighting factor for these test calculations was ρ = 1000
stated. It also illustrates the influence of the weighting factor
at the risk of the optimal solution. Table 7.9 shows using the
Risk measure excess probability the influence of this factor.

https://translate.googleusercontent.com/translate_f 108/143
2/29/2020 Untitled

ρ QE Q EP
0 28 368.09 0.51
1 28 368.09 0.51
10 28 371.55 0.33
100 28 398.00 0.27
1000 28 398.00 0.27

Table 7.9: Influence of the weighting factor

7.3 Efficiency of the decomposition process


Using selected examples, the efficiency of the in Chapter

4 presented decomposition methods are illustrated. In particular


is the algorithmic behavior with increasing number of scenarios
testing. Larger energy transmission systems are considered here. Next
Results for the two-stage approach also become test calculations for the multi-stage

Page 101

7.3. EFFICIENCY OF DECOMPOSITION PROCEDURES 93

Approach cited. The scenario decomposition according to algorithm 5.1 was in


the program package ddsip [16] implemented and used here, whereas for

Algorithm 5.2 an own implementation in C ++ using ddsip and


CPLEX was done.

Two-step stochastic mixed-integer linear models: The following

presented calculations lay the 6-node network with 36 discrete time steps
based. Unsafe data (load, demand, maximum transmission capacity) occur
from the 7th time step on. The results listed here can also [43]
be removed. Tables 7.10 and 7.11 show the results achieved

https://translate.googleusercontent.com/translate_f 109/143
2/29/2020 Untitled
Use of the purely expectation-based approach as well as the expected value
Risk approach using the expected excess as a risk measure.
The necessary level η is used to minimize the expected value and risk
and ρ> 0 selected appropriately. Furthermore, a maximum computing time of one

Hour and a premature termination of the algorithms when falling below


a relative tolerance of 0.3% was initiated. Show the fields marked with "-"
a failure to find a permissible point in the maximum specified here

Computing time.

ILOG CPLEX Decomposition (ddsip)

N Time (sec.) QE GAP (%) time (sec.) QE GAP (%)

20 5 4 307.33 00:00 8th 4 318.69 00:28


100 3,600 4 266.95 0.76 227 4 278.06 00:29
500 3,600 - - 1 865 4 283.07 00:29
1000 3,600 - - 3,274 4 280.23 00:24

Table 7.10: Results with expected value minimization

Table 7.12 shows results using the expected value risk approach
using the risk measure excess probability. The maximal
Computing time was again set at one hour. An early termi-

If the previously set relative tolerance was undershot, no


initiated. It shows that there is an increasing number of scenarios
more difficult or no longer possible, with CPLEX in the maxi-
paint computing time to achieve solutions. Even for a large number of scenarios

Page 102

94 CHAPTER 7. NUMERICAL EXAMPLES

ILOG CPLEX Decomposition (ddsip)

N Time (sec.) QE Q EE GAP (%) time (sec.) QE Q EE GAP (%)

20 5 4 307.33 13.97 00:00 78 4 307.34 13.97 00:01


100 3,600 4 267.22 1.61 1:02 106 4 275.61 2:43 00:26
500 3,600 - - - 1 542 4 280.61 3.26 00:26
1000 3,600 - - - 3,600 4,283.37 3.61 00:34

https://translate.googleusercontent.com/translate_f 110/143
2/29/2020 Untitled

Table 7.11: Results with expected value-expected excess minimization

ILOG CPLEX Decomposition (ddsip)

N Time (sec.) QE Q EP GAP (%) time (sec.) QE Q EP GAP (%)

20 3,600 4 307.33 0.30 00:00 3,600 4 307.33 0.30 00:00


100 3,600 4,269.34 0.05 00:14 3,600 4,273.45 0.06 00:18
500 3,600 - - - 3,600 4,283.61 0.13 00:31
1000 3,600 - - - 3,600 4,283.83 0.15 00:34

Table 7.12: Results with expected excess probability minimization

the scenario composition is able to find solutions with acceptable quality


determine. Finally, Table 7.13 shows the resulting model dimensions
the deterministically equivalent formulations in 1000 scenarios in comparison
to the 1-scenario model (deterministic model).

N 1 1000 1000 1000

risk measure - - Surplus fortune. Expected surplus

Supplementary conditions 2,888 2,422,466 2,423,466 2,423,466

continuous variables 864 720 144 720 144 721 144


binary variables 432 360 072 361 072 360 072

Table 7.13: Model dimensions with stochastic optimization

Two-stage stochastic mixed-integer convex models: Again,


de the 6-node network, but used with changed input data.

The optimization period spanned 18 discretized time steps


uncertainties occurring from the 13th time step. The maximum computing time
was set at 3600 seconds and an early termination
Falling below a relative tolerance of 1% initiated. In addition, the

Page 103

7.3. EFFICIENCY OF DECOMPOSITION PROCEDURES 95

https://translate.googleusercontent.com/translate_f 111/143
2/29/2020 Untitled

Approximation accuracy according to (3.5), adapted to the two-step case with

ϵ = 0.01. This was done for the expected value risk minimization
necessary levels η and ρ> 0 are selected appropriately. The following scenario sets
were generated randomly, as before. Depending on that, the loosening
of the scenario-specific sub-problems require different computing efforts.
The computing time required does not necessarily have to increase with the number of scenarios,

but is certainly largely dependent on the underlying scenarios themselves


dependent. But also here it shows that an external approximation method in
Connection with a decomposition according to algorithm 5.2 much more efficient
is than this when the possibility of decomposition is neglected

Would be the case. The determination of a permissible point in the given here
maximum computing time is only possible through the decomposition. It will

Outer approximation Outer approximation + decomposition

N Time (sec.) QE GAP (%) time (sec.) QE GAP (%)

5 7 8 940.16 00:00 10 8 939.48 0.93


10 3,600 7 505.11 8.18 3,600 7 414.65 1:48
25 3,600 - - 3,600 7 455.14 2.74
100 3,600 - - 3,600 7 412.99 6.92

Table 7.14: Results with expected value minimization

Outer approximation Outer approximation + decomposition

N Time (sec.) QE Q EE GAP (%) time (sec.) QE Q EE GAP (%)

5 8th 8 940.22 4.51 00:00 46 8 939.35 00:52 0.76


10 3,600 - - - 3,600 7 419.61 0.66 12.80
25 3,600 - - - 3,600 7 472.43 0.64 5:46
100 3,600 - - - 3,600 7 512.71 79.21 92.17

Table 7.15: Results with expected value-expected excess minimization

emphatically points out that this is only the representation of the

numerical efficiency of the decomposition process. Often they show


Optimal values for the purely expected value-based model already minimal risks
on or there can be optimal solutions with expected value risk minimization
can be determined with a better resulting expected value than this with pure

https://translate.googleusercontent.com/translate_f 112/143
2/29/2020 Untitled
Page 104

96 CHAPTER 7. NUMERICAL EXAMPLES

Outer approximation Outer approximation + decomposition

N Time (sec.) QE Q EP GAP (%) time (sec.) QE Q EP GAP (%)

5 8th 8 940.16 0.40 00:00 124 8 939.33 0.20 00:19


10 3,600 - - - 3,600 7 416.36 0.20 9.23
25 3,600 - - - 3,600 7 472.30 0.12 16:57
100 3,600 - - - 3,600 7 498.81 0.58 48.87

Table 7.16: Results with expected excess probability minimization

expected-based minimization was the case. Results of this kind can


of course, with early termination of the procedures, if the

relative tolerance is not zero and actually better solutions


exist. Further test calculations can also be found in [44].

Optimization of larger energy transmission systems: The new

efficiency of the decomposition process using a large


Energy transmission system, the IEEE118-BUS system [55] clarified. The
The period was set at nine hours in half-hour resolution, by
where the first four showed no uncertainties. With a maximum computing

24 hours before the appointment was terminated


Relative tolerance of 1.5% initiated. Table 7.17 shows the results achieved
with DC and RDCOL modeling. The RDCOL modeling increases, here

DC modeling RDCOL modeling

N Time (sec.) QE GAP (%) time (sec.) QE GAP (%)

5 319 17 189.81 0.65 51 968 19 555.02 00:58


10 10 933 15 063.13 1.10 95 900 17 027.39 0.82
15 - - - 90 523 15 712.96 1:48

Table 7.17: Results for the IEEE118 node network

at least, the computing times considerably compared to those for DC modeling.


In the long term, the average costs will also increase noticeably
on. The consideration of transmission losses has significant effects here
to the optimal values. It should also be noted that for the 15 scenario case at
DC modeling no solution was found. This is the scenario set

https://translate.googleusercontent.com/translate_f 113/143
2/29/2020 Untitled

Page 105

7.3. EFFICIENCY OF DECOMPOSITION PROCEDURES 97

actually inadmissible for the DC model, but permissible for the RDCOL model.
Reminiscent of Theorem 2.3, which applies analogously to the 2-stage stochastic case

this case is entirely possible. By overestimating


network losses in the RDCOL model ( Energy destruction ”) may be admissibility
"
when critical scenarios occur.

Multi-stage stochastic optimization models: Some numerical test calculations


The multistage stochastic case is cited from [33]. there
only an outlook should be given. The aim is to show that the
multilevel stochastic optimization, which enables an even more precise modeling of

Random influences allowed to apply to the models developed in this work


The lets and procedures exist that efficiently solve such models
allow. The IEEE14 bus system was used as the basis, with results
for the purely expected value-based multi-stage model. The

maximum computing time was set at 1800 seconds. Furthermore, a


early termination of the bb-tree and bb-hash algorithms, special procedures
for the multi-level mixed-integer model class, when falling below one

relative tolerance of 0.1% initiated. The time at bb-tree and bb-hash,

t max N bb-tree (sec.) CPLEX (in%) bb-hash (sec.) CPLEX (in%)

3 50 66 3.5 135 3.2


3 100 358 9.5 1 376 8.1

5 50 160 6.1 315 5.3


5 100 404 6.0 684 5.3

Table 7.18: Results with multi-stage optimization

https://translate.googleusercontent.com/translate_f 114/143
2/29/2020 Untitled

which is less than 1800 seconds in all cases, signals the detection
a point with a relative tolerance of less than 0.1%. For comparison
became the available relative tolerance when using CPLEX in the same
Computing time shown.

Page 106

98 CHAPTER 7. NUMERICAL EXAMPLES

7.4 Retrofitting a transmission system


Using the model presented in Chapter 7, the example of the
6-node network, a cost-optimal retrofit of such was carried out. The

Production costs are to be reduced in the long term and also allowable
can be achieved when critical scenarios occur. The optimization
criterion chosen to minimize investment and production costs.
The budget available has no upper bound. basis

5 scenarios were set up, which show the load values occurring in the second
simulate fe over a longer period of time. Using the IG and the
DC modeling first determined the optimal values, which are
solve the second stage. Table 7.19 shows using

of these 5 scenario sets the minimum production cost that is in use


of the existing network, and those ignored
Network. Attention to the energy transmission system increases the production

N πi z DC z IG
1 0.3 93 887 308 55 664 544
2 0.1 80 965 468 48 924 600
3 0.2 112 274 558 65 086 800
4 0.25 - 68 245 656
5 0.15 110 115 938 62 937 096

https://translate.googleusercontent.com/translate_f 115/143
2/29/2020 Untitled

Table 7.19: Production costs before expansion

cost significantly. If scenario 4 occurs, DC- is also inadmissible.

Modeling before, although the total active power available for


Load coverage would suffice because z IG ∈ R. In order to reduce production costs
Most and establishing admissibility for the critical scenario, they were still
lines not available E = {14,41,34,43} released for new construction. The

Installation costs were c ins


= 2,000,000 units and reinforcement costs
kl

with c kladd = 500 units set for small ∈ E. Furthermore, with ˜E = E all
already existing lines for reinforcement released. The cost of one
Gain were added via c add kl = 2000 units for small ∈ E~ set. The 2-stage

Page 107

7.4. RETROFITTING A TRANSMISSION SYSTEM 99

based on a stochastic model based purely on expected values, was


Sen Disconnect the line 14.41 zusätzlich E and the other lines

sometimes significantly reinforced. Fig. 7.6 shows the extended 6-node network. The Andes
Values noted in the lines show the additionally provided transmission
capacities. This leads to expected costs of Q E = 69 454 489 units,

0.3
D
3 D

2.14
2.73
D
0.61 0.66 D

1.5

https://translate.googleusercontent.com/translate_f 116/143
2/29/2020 Untitled

Figure 7.6: The retrofitted 6-node network

which include the expansion costs of 4031874 units. The so developed


Netz enables productions, which are shown in Table 7.20. The expansion

N 1 2 3 4 5

z DC 55 664 544 48 924 600 72 437 900 77 545 856 66 378 316

Table 7.20: Production costs after retrofitting

establishes admissibility when scenario 4 occurs and reduces production costs


most significant, but can certainly never fall below the value z IG , since z IG ≤ z DC
must apply to all scenarios. It should also be noted that in order to accelerate
optimization of the branching strategy first to the one to rebuild

Page 108

100 CHAPTER 7. NUMERICAL EXAMPLES

Binary variables associated with edges have been changed. The model shown here
could be solved in less than a minute.

7.5 A case study from the energy industry

For the test calculations presented below, an energy transmission


system, which allows typical situations in the transmission
system of the Federal Republic of Germany. This is shown in Fig.7.7
anschaulicht. The results shown here were created in close cooperation
with the chair for energy systems and energy management of the technical university
university of Dortmund. The collaboration also resulted in [31]. The after-
https://translate.googleusercontent.com/translate_f 117/143
2/29/2020 Untitled

The following case study can be found in detail in this literature source
become. The parameters for the energy transmission system used, which in
[30] is presented in detail by the Chair of Energy Systems and
Energy management provided by the Technical University of Dortmund.
It is only mentioned here that the voltage level is 380 kV. The use
The network does not fully reflect the FRG's energy transmission system
in addition, such network data is not publicly available, let it
Despite a reduction to a total of 28 knots, the north
Identify south connections and investigate the influence of wind feed.

In particular, situations were


with network congestion, ie possible overloading of overload
transmission lines, can come.
Deviating from the previously presented optimization models are now also

Multiple lines between network nodes observed. There are also


Power generation plants from a whole park of power plants that stand out
Lignite, nuclear, hard coal and gas turbine power plants as well as wind
can assemble energy systems. Also the possibility of an import and

Exports of electrical energy (transit) with the associated costs


or winning is considered in this extended model (in Fig.7.7
Import / export options marked with the grid squares). Mo-
dent extensions of the type described above can again be found in [31].

Page 109

7.5. A CASE STUDY FROM THE ENERGY SECTOR 101

https://translate.googleusercontent.com/translate_f 118/143
2/29/2020 Untitled

Figure 7.7: A typified energy transmission system of the FRG

The chair for energy systems was also able to provide information on the power plant
and energy industry back to a database with detailed information
be gripped. Again, more information can be found [30, 31]

be removed.
The computing times required to solve the DC or RDCOL models
were only a few seconds. Table 7.21 shows the minimization
values of the DC and RDCOL models and other parameters, whereby

Page 110

102 CHAPTER 7. NUMERICAL EXAMPLES

https://translate.googleusercontent.com/translate_f 119/143
2/29/2020 Untitled
Roundings were made. The wind feed was 7200 MW.
This corresponds to exactly 40% of the maximum available power of the

Wind turbines, which amount to 18000 MW. A review of the

DC RDCOL

feedforward 1 181 596 1 200 164


total feed 59 018 59 442
WEP feed 7 200 7 200
import 5 882 5 882
export 3 125 3 125
load 61 775 61 775
losses 0 424

Table 7.21: Optimal values in euros and other parameters in MW at 40%


WEP feed

RDCOL solution regarding the approximate ohmic transmission losses resulted


also that these were not overestimated on any line. So underestimated the

Solution of the RDCOL model the transmission losses. Because the tolerance parameter

was assumed to be ϵ = 10 −5 , but was only slightly underestimated.


It also showed that the decision made by the RDCOL model

can be used as a starting point for a classic load flow calculation,


ie it converges the Newton method. More details
ten are still mentioned. In addition, test calculations were carried out,
where the wind feed was 80% or 100% of the installed capacity.

Table 7.22 shows the results obtained. Often there are high feeds
of the wind turbines is problematic because they overload the lines
to lead. Thanks to the DC or RDCOL modeling,
conditions are determined which do not overload the transmission lines. It showed

furthermore, that only with a 100% wind feed a slight


overestimation of lust on line 13.31 ∈ E was present. On all other lines
There was no overestimation of the network losses. From the determined RDCOL
Again, a convergence of the Newton method for the decisions

https://translate.googleusercontent.com/translate_f 120/143
2/29/2020 Untitled
Page 111

7.5. A CASE STUDY FROM THE ENERGY SECTOR 103

DC RDCOL DC RDCOL
80% wind turbine feed 100% wind turbine feed

feedforward 969 421 986 275 931 242 945 198


total feed 59 417 60 117 59 422 60 178
WEP feed 14,400 14,400 18,000 18,000
import 5,483 5,483 5 478 5,483
export 3 125 3 125 3 125 3 125
load 61 775 61 775 61 775 61 775
losses 0 700 0 762

Table 7.22: Optimal values in euros and other parameters in MW

classic load flow calculation can be initiated. Graduate industrial engineer. D. Waniek leading
te, in addition to providing the network and power plant data, verification

of the values determined by solving the RDCOL model. This validation


can be found in much more detailed form [31]. The were over
RDCOL minimization determined values as input data for the AC model
which then uses a classic method for load flow

invoice was solved. The commercial network calculation program Neplan was
de used here. Table 7.23 shows the data records that are referenced with Ref. (AC)
marked columns are determined by Neplan. Those with opt. (AC)
marked columns show the values determined by Neplan, if

the optimal decision of the RDCOL model is used as an input variable. It

Ref. Opt. Opt. Ref. Opt. Opt. Ref. Opt. Opt.

(AC) (RDCOL) (AC) (AC) (RDCOL) (AC) (AC) (RDCOL) (AC)

wind feed 40% 80% 100%

Optimal value [T €] 1231 1200 1201 971 986 987 858 945 945

Import [MW] 5347 5882 5882 5347 5483 5483 5437 5483 5483

Export [MW] 3472 3125 3125 3472 3125 3125 3472 3125 3125

Grid losses [MW] 444 424 434 1016 700 709 1468 762 768
Overload of

resources No No No Yes No No Yes No No

Table 7.23: Comparison of the modeling approaches

only the basic results were shown here. The literature source [30]

https://translate.googleusercontent.com/translate_f 121/143
2/29/2020 Untitled

Page 112

104 CHAPTER 7. NUMERICAL EXAMPLES

further details can be found. In particular, the


Articles much more detailed with the interpretation of the results. The following

Extract is cited from the source mentioned:

The comparison of the results of the presented case study shows that
There are high deviations in the power flows on individual lines
comes. These are largely due to the model itself and

differs from case to case. The goal of RDCOL modeling


could be achieved, however, because the optimized feeds of the
Power plants never overload equipment at the
Calculate the full AC load flow and the total

Network losses can be estimated very precisely. The results will be


favored by the network topology, since in the areas around the slack
and PU nodes where the greatest deviations occur, the
Network is underutilized. The heavily loaded lines are

on the other hand, represented very precisely by the RDCOL model.

In the future, you can also proceed to an entire period


simulate. Corresponding exemplary calculations have been made in this chapter
presented. Only in this way is it possible to use the two- or multi-stage approach

stochastic optimization unsafe data sets, such as wind strength,


to be able to consider in the optimization phase. In this work it was shown
how unsafe data records can be integrated. The featured
Test calculations, in which decomposition options were used, show

also that such models work in acceptable computing time, of course depending on
the modeling accuracy of the stochastic effects remain solvable.

https://translate.googleusercontent.com/translate_f 122/143
2/29/2020 Untitled

Page 113

Summary

The present work comes with a summary of what has been achieved and a
Outlook on possible further developments completed. First, in Ka

Chapter 1 explains the underlying problem and a first impression of the


given technical and mathematical aspects to be observed. According to this
In Section 2, the task was formulated in broad outline
position further specified and suitable mathematical optimization models

designed for deterministic data situations. In particular, it became a model


developed, which integrates the electrical energy transmission system,
power losses on lines are taken into account and the use of energy generation
were precisely recorded, ie in addition to the pure supply, the switching

status taken into account. In Chapter 3, solution procedures became more general
Description presented. An algorithm was developed that allows
the optimal solution of a relaxed form of the optimization model that
Considered ohmic transmission losses by solving master problems,

https://translate.googleusercontent.com/translate_f 123/143
2/29/2020 Untitled
which are only available in a mixed-integer linear description.
hern. Chapter 4 dealt intensively with the integration of insecure data
sets. This led to stochastic optimization models. Both
Risk-neutral as well as risk-averse two-stage approaches addressed, but also in

Basics of multi-stage models shown. The usability of the stochastic


Modeling approaches to the models mentioned in Chapter 2 were defined
provides. In Chapter 5, effective decomposition methods were presented, with which
stochastic optimization models even if the

Remain treatable at random. The for the two-stage mixed-integer conv


xe model class algorithm also allowed nonlinear influences

105

Page 114

with simultaneous use of decompositional properties that are clear


Efficiency increases in terms of computing time. In

Chapter 6 created lower bounds for the AC model and one


Method used in engineering to determine the state of the network
further examined. Heuristic approaches were presented here. Abschlie-
The sample calculations from Chapter 7 for purely deterministic data

positions that the determined solutions for the approximate direct current
delle close to that of the allowable obtained by the heuristic approach
Point of the AC model. There was also a focus on the
blematics when applying production decisions that are

if the electrical energy transmission system is disregarded in the


result in the massive overloading of the transmission line
were shown when using such decisions. Further was
the possible influence of transmission costs on the optimal solution is shown.

Numerous stochastic test calculations were illustrated and in addition to


tivative examples also presented the efficiency of the decomposition process.
A concrete case study from the energy industry concluded this chapter. overall
https://translate.googleusercontent.com/translate_f 124/143
2/29/2020 Untitled

according to the task, models and algorithms mentioned in Chapter 1


stochastic optimization of the operation of electrical power generation plants
can achieve this goal for the reduced model formulations as achieved
be considered. When using AC modeling, however, could

only heuristic approaches are developed. Research needs therefore remain with
like modeling techniques exist. It should be noted:

One could generally show that the DC model is a relaxation


of the AC model, you would be able to
to generate ken. You could possibly do the AC model

approach using a branch and bound process by using good lower bounds
can be obtained by solving the DC model and upper ones
Solve the AC model, which
using a globalized form of sequential quadratic
table programming is made possible.

106

Page 115

Appendix A

List of Symbols

amounts
https://translate.googleusercontent.com/translate_f 125/143
2/29/2020 Untitled

V node

V supply node
e edge
e Edges with transmission costs
e~ edges released for reinforcement

e Connections released for the creation of new edges


Nk to k ∈ V adjacent nodes
Nk to k ∈ V possible adjacent nodes due to new line construction
T discretized times

T T \ {0}

Data
PD total active power load

QD total reactive power load


PD k Active power load in k ∈ V
QD k Reactive power load in k ∈ V
g small Conductance of the line kl ∈ E

b small Susceptibility of the line kl ∈ E


b small Susceptibility of the line kl ∈ E for DC flow
x small Line reactance kl ∈ E
107

Page 116

bkl0 Shunt resistance acceptance of the line kl ∈ E


Klg small Real part of the element kl of the node-edge admittance matrix
~b small Imaginary part of the element kl of the node-edge admittance matrix

d klmax maximum transmission capacity of the line kl ∈ E


P Gmin k
minimum possible active power feed of the system k ∈ V
P Gmax k
maximum possible active power feed of the system k ∈ V
https://translate.googleusercontent.com/translate_f 126/143
2/29/2020 Untitled

Q Gmin k
minimum possible reactive power feed of the system k ∈ V

Q Gmax k
maximum possible reactive power feed of the system k ∈ V
U kmin minimum possible voltage at k ∈ V
U kmax maximum possible voltage at k ∈ V
P Ginit
k
Initialization of the active power feed from system k ∈ V

s kinit Initialization of the switching status of the system k ∈ V


∆k Max. possible increase in performance of the system k ∈ V
∆k Max. possible reduction in performance of the system k ∈ V
Γk Max. possible power value when switching on system k ∈ V

Γk Max. possible power value when switching off system k ∈ V


Λk Minimum term of plant k ∈ V
Λk Minimum downtime of plant k ∈ V
P Ginit
k
Active power feed-in of the system k at the beginning of the optimization

s kinit Switching status of the system k at the beginning of the optimization


αk linear fuel cost share of the plant k ∈ V
βk Fixed fuel cost share of the system k ∈ V
γk Start-up costs of the system k ∈ V

δ small linear share of transmission costs for line kl ∈ E


c kladd linear cost share with capacity expansion of the edge kl ∈ E ∪ e~
c klins Construction costs for new line kl ∈ E
M sufficiently large constant (multiple use)

n Cardinality of V
t max Cardinality of T
t∗ last deterministic point in time with two-stage modeling

108

Page 117

https://translate.googleusercontent.com/translate_f 127/143
2/29/2020 Untitled

variables
Sk ∈C complex power in k ∈ V
Pk ∈R Active power in k ∈ V
Qk ∈R Reactive power in k ∈ V
SL ∈C complex network losses
PL ∈R+ Active power losses
QL ∈R Reactive power losses
ek ∈R+ Real part of the complex voltage in k ∈ V
fk ∈R Imaginary part of the complex voltage in k ∈ V
PG k ∈R+ Active power feed of the system k ∈ V
QG k ∈R Reactive power feed of the system k ∈ V
PL k ∈R+ Auxiliary variable for the inclusion of transmission losses
sk ∈ {0.1} Switching status of system k ∈ V
uk ∈ {0.1} Activation detection of the system k ∈ V
vk ∈ {0.1} Switch-off detection of the system k ∈ V
n small ∈ {0.1} New building decision for small ∈ E
p small ∈R Active power flow along edge kl ∈ E
p small ∈R Active power flow along a possible edge kl ∈ E
d kladd ∈R additional transmission capacity for edge kl ∈ E
ν kl ∈R ohmic transmission losses along edge kl ∈ E
q small ∈R Reactive power flow along edge kl ∈ E

Klp small ∈ {0.1} Auxiliary variable for determining the amount of p kl ∈ E


p small ∈R+ Amount of active power flow p kl
Uk ∈R+ Voltage amount in k ∈ V
Uk ∈C complex voltage in k ∈ V

θk ∈R Voltage angle in k ∈ V
θ kl ∈R Voltage angle difference between k, l ∈ V

109

Page 118
https://translate.googleusercontent.com/translate_f 128/143
2/29/2020 Untitled

features

Ck Production costs of the system k ∈ V


Dk Start-up costs of the system k ∈ V

E AC
kl Transmission costs on line kl ∈ E (AC modeling)
E DC
kl Transmission costs on line kl ∈ E (DC modeling)
F kl1 Costs when expanding the transmission capacity of line kl ∈ E ∪ e~
F kl2 Costs for new line construction kl ∈ E

When considering a period, the variables listed here are partly


to be marked with an additional time index. Also refer
the designations given here refer only to the developed

dent according to Chapter 2.

https://translate.googleusercontent.com/translate_f 129/143
2/29/2020 Untitled
110

Page 119

Appendix B

Data situation: 6-node network

network data

V = {1, ..., 6}

V = {1, ..., 4}

E = {12,21,15,51,24,42,35,53,36,63,46,64,45,54}

g small = 5, b kl = −10, b 0
kl = 0 ∀ kl ∈ E
d klmax = 1 ∀ kl ∈ E \ {45.54}, d max 45 = d max
54 = 0.5

Generation plant data

(α 1 , α 2 , α 3 , α 4 ) = (24,20,18,21)

(β 1 , β 2 , β 3 , β 4 ) = (105,101,109.98)

(γ 1 , γ 2 , γ 3 , γ 4 ) = (60, 40, 20, 30)

(P min
G , P min
1 G , P Gmin , P Gmin ) = (0,0.5,0.2,0.5)
2 3 4

(P max
G 1
, P Gmax , P Gmax , P Gmax ) = (10,3,3,3)
2 3 4

https://translate.googleusercontent.com/translate_f 130/143
2/29/2020 Untitled

(Γ 1 , Γ 2 , Γ 3 , Γ 4 ) = (∆ 1 , ∆ 2 , ∆ 3 , ∆ 4 ) = (2,0.6,1.2,1.4)

(Γ 1 , Γ 2 , Γ 3 , Γ 4 ) = (∆ 1 , ∆ 2 , ∆ 3 , ∆ 4 ) = (2,0.6,1.2,1.4)

(Λ 1 , Λ 2 , Λ 3 , Λ 4 ) = (1,3,3,3)

(Λ 1 , Λ 2 , Λ 3 , Λ 4 ) = (1,3,3,3)

111

Page 120

load data
(T) (T) (T) (T) (T) (T)
tP D1 PD 2
PD 3
PD 4
PD 5
PD 6

1 1.00 2.20 1.72 1.18 1.00 0.51


2 1.00 2.20 1.72 1.18 0.80 0.51
3 1.10 2.42 1.83 1.21 0.60 0.55
4 1.10 2.42 1.83 1.21 0.30 0.55
5 1.15 2.53 1.89 1.23 0.20 0.57
6 1.20 2.64 1.94 1.24 0.15 0.58
7 1.30 2.86 2.04 1.26 0.10 0.61
8 1.35 2.97 2.09 1.27 0.06 0.63
9 1.35 2.97 2.09 1.27 0.06 0.63
10 1.40 3.08 2.13 1.27 0.06 0.64
11 1.44 3.17 2.16 1.28 0.06 0.65
12 1.49 3.28 2.20 1.28 0.06 0.66
13 1.51 3.32 2.22 1.28 0.06 0.67
14 1.55 3.41 2.25 1.29 0.06 0.67
15 1.60 3.52 2.28 1.29 0.07 0.68
16 1.58 3.48 2.27 1.29 0.07 0.68
17 1.56 3.43 2.25 1.29 0.07 0.68
18 1.60 3.52 2.28 1.29 0.07 0.68

(T) (T) (T) (T) (T) (T)


(Q D1 , Q D , Q D , Q D , Q D , Q D ) = (0.2,0.2,0.2,0.2,0.5,0.1) ∀ t ∈ T
2 3 4 5 6

https://translate.googleusercontent.com/translate_f 131/143
2/29/2020 Untitled

The load profiles for


based on t = 1.

112

Page 121

List of Tables

2.1 Advantages and disadvantages of modeling approaches. , , , , , , , , , , , , 23

7.1 Resulting load flow with IG modeling. , , , , , , , , , , , , 84

7.2 Resulting load flow with DC modeling. , , , , , , , , , , , 84


7.3 Resulting load flow with DCOL modeling. , , , , , , , , , 85
7.4 Resulting load flow with FAC modeling. , , , , , , , , , , , 86

7.5 Comparison of the determined optimal solutions / values. , , , , , , , , , 86

7.6 Comparison of solutions when adding transmission costs. , , , , 87

7.7 Impact of uncertainties on the optimal solution. , , , , , , , , , 91

7.8 Influence of risk on optimal solution / value. , , , , , , , , , , , , 92


https://translate.googleusercontent.com/translate_f 132/143
2/29/2020 Untitled

7.9 Influence of the weighting factor. , , , , , , , , , , , , , , , , , , , 92


7.10 Results with expected value minimization. , , , , , , , , , , , , 93

7.11 Results with expected value-expected excess minimization. , 94

7.12 Results with expected excess probability minimization. 94

7.13 Model dimensions with stochastic optimization. , , , , , , , , 94

7.14 Results with expected value minimization. , , , , , , , , , , , , 95


7.15 Results with expected value-expected excess minimization. , 95
7.16 Results with expected excess probability minimization. 96

7.17 Results on the IEEE118 node network. , , , , , , , , , , , , , , , 96

7.18 Results with multi-stage optimization. , , , , , , , , , , , , , , 97

7.19 Production costs before expansion. , , , , , , , , , , , , , , , , , , , 98

7.20 Production costs after retrofitting. , , , , , , , , , , , , , 99


7.21 Optimal values in euros and other parameters in MW at 40%

Wind turbine feed-in. , , , , , , , , , , , , , , , , , , , , , , , , , , 102

113

Page 122

7.22 Optimal values in euros and other parameters in MW. , , , , , , 103


7.23 Comparison of the modeling approaches. , , , , , , , , , , , , , , , , , 103

https://translate.googleusercontent.com/translate_f 133/143
2/29/2020 Untitled

114

Page 123

List of Figures
https://translate.googleusercontent.com/translate_f 134/143
2/29/2020 Untitled

1.1 An electrical power transmission system. , , , , , , , , , , , , 6


1.2 A 3-node system. , , , , , , , , , , , , , , , , , , , , , , , , , 7

3.1 Graphic illustration of the permissible losses. , , , , , , , 42


3.2 Exemplary approximation of losses. , , , , , , , , , , , , , 45

7.1 A 6-node network. , , , , , , , , , , , , , , , , , , , , , , , , , , 82


7.2 Optimized feed with IG modeling. , , , , , , , , , , , , 88
7.3 Optimized infeed with DC modeling. , , , , , , , , , , , 88
7.4 Optimized infeed with RDCOL modeling. , , , , , , , , , 89

7.5 Optimized feed with FAC (RDCOL) modeling. , , , , , 89


7.6 The retrofitted 6-node network. , , , , , , , , , , , , , , , , , , 99
7.7 A typified energy transmission system of the FRG. , , , , , , , 101

115

Page 125
124

https://translate.googleusercontent.com/translate_f 135/143
2/29/2020 Untitled

bibliography

[1] Aguado, JA; Rosehart, W .: Alternative Optimal Power Flow Formulations,

14h PSCC, Seville (2002)

[2] Ahmed, S .; King, AJ; Parija, G .: A Multi-stage Stochastic Integer Program-


ming Approach for Capacity Expansion under Uncertainty, Journal of Global

Optimization, Vol. 26, Issue 1, 3-24 (2003)

[3] Alt, W .: Nonlinear Optimization, Vieweg Studies (2002)

[4] Andersson, G .: Modeling and Analysis of Electric Power Systems, Lecture


227-0526-00, Power Systems Laboratory, ETH Zurich (2004)

[5] Artzner, P .; Delbaen, F .; Eber, J.-M .; Heath, D .: Coherent measures of risk,

Mathematical Finance, Vol. 9, No. 3, 203-228 (1999)

[6] Bauer, H .: Probability Theory, de Gryter (2002)

[7] Beale, EML: On minimizing a convex function subject to linear inequalities,


J. Royal Statistical Society, Series B 17, 173-184 (1955)

[8] Bergen, AR; Vittal, V .: Power Systems Analysis, Second Edition, Prentice-

Hall (2000)

[9] Birge, JR; Louveaux, FV: Introduction to Stochastic Programming,


Springer publishing house (1997)

[10] Bienstock, D .; Mattia, S .: Using mixed-integer programming to solve power


grid blackout problems, Discrete Optimization, Vol. 4, Issue 1, 115-141 (2007)

117

Page 126

https://translate.googleusercontent.com/translate_f 136/143
2/29/2020 Untitled

[11] Blair, CE; Jeroslow, RG: The value function of a mixed integer program,
Discrete Mathematics 19, 121-138 (1977)

[12] Caroe, CC; Schultz, R .: Dual decomposition in stochastic integer program

ming, Operations Research Letters 24, 37-45 (1999)

[13] Christensen, GS; El-Hawary, ME; Soliman, SA: Optimal Control Applica-

tion in Electric Power Systems, Wiley (1987)

[14] CPLEX Interactive Optimizer 9.1.3, ILOG (1997-2005)

[15] Dantzig, GB: Linear programming under uncertainty, Management Science

1, 197-206 (1955)

[16] ddsip, C package for the dual decomposition of stochastic programs with mixed

integer recourse, Department of Mathematics, University of Duisburg-Essen (2004)

[17] Duran, MA; Grossmann, E .: An outer-approximation algorithm for a class


of mixed-integer nonlinear programs, Mathematical Programming, Volume 36,
Issue 3, 307-339 (1986)

[18] Dyer, J .: Escort: DC optimization for production cost estimation, IEE Collo-
quium on Optimal Power Flow - Invaluable Tool or Expensive Toy? 8 / 1-8 / 4
(1997)

[19] Edelmann, H .; Theilsiefje, K .: Optimal network operation in electrical


Energy supply, Springer-Verlag (1974)

[20] Eichhorn, A .; Römisch, W .: Polyhedral risk measures in stochastic program-

ming, SIAM Journal on Optimization, Volume 16, Issue 1, 69-95 (2005)

[21] Fletcher, R .; Leyffer, S .: Solving mixed integer nonlinear programs by ou-

ter approximation, Mathematical Programming, Volume 66, Issue 3, 327-349


(1994)

[22] Gill, P .; Murray, W .: SNOPT - A software package for solving large-scale op-
timization problems (linear and nonlinear programs), Stanford Business Software
goods (1998-2006)

118
https://translate.googleusercontent.com/translate_f 137/143
2/29/2020 Untitled

Page 127

[23] Gollmer, R .; Nowak, MP; Roman, W .; Schultz, R .: Unit commitment in


power generation - a basic model and some extensions, annals of operations

Research, Vol. 96, Numbers 1-4, 167-189 (2000)

[24] Gönen, T .: Modern Power System Analysis, Wiley (1988)

[25] Grainger, JJ; Stevenson, WD: Power System Analysis, McGraw-Hill (1994)

[26] Han, SP: Superlinearly convergent variable metric algorithm for general non-

linear programming problems, Mathematical Programming, vol. 11, 263-282


(1976)

[27] Han, SP; Mangasarian, QL: Exact penalty function in nonlinear program
ming, Mathematical Programming, Vol. 17, 251-269 (1976)

[28] Handschin, E .: Electrical energy transmission systems, 2nd edition, Hüthig-

Publishing house (1987)

[29] Handschin, E .: Electrical energy technology, script, chair for energy


systems and energy management, University of Dortmund (2004)

[30] Handschin, E .; Häger, U .; Rehtanz, C .; Waniek, D .: Influences of Wind Energy

gy on the Operation of Transmission Systems, IEEE PES General Meeting,


Pittsburgh (2008)

[31] Handschin, E .; Kuhn, S .; Rehtanz, C .; Schultz, R .; Waniek, D .: Opti-


Painter power plant use in network congestion situations, manuscript: university

Duisburg-Essen / Dortmund University of Technology, (to be published) (2008)

[32] Handschin, E .; Neise, F .; Neumann, H .; Schultz, R .: Optimal operation of


dispersed generation under uncertainty using mathematical programming, in-
ternational Journal of Electrical Power & Energy Systems, Vol. 28, Issue 9,

618-626 (2005)
https://translate.googleusercontent.com/translate_f 138/143
2/29/2020 Untitled

[33] Heinze, T .: A process for the decomposition of multi-stage stochastic


Optimization problems with integer and risk aversion, dissertation,
Department of Mathematics, University of Duisburg-Essen (2007)

119

Page 128

[34] Heinze, T .; Schultz, R .: A branch-and-bound method for multistage stochastic


integer programs with risk objectives, Optimization (to appear) (2008)

[35] Helmberg, C .: Semidefinite Programming for Combinatorial Optimization,


ZIB report ZR-00-34, Konrad Zuse Center Berlin (2000)

[36] Herold, G .: Electrical energy supply I, J. Schlembach Fachverlag (2001)

[37] Kall, P .; Wallace, S .: Stochastic Programming, Wiley (1994)

[38] Kallrath, J .: Mixed-integer optimization: modeling in practice,


Vieweg (2002)

[39] Kelley, JE: The cutting-plane method for solving convex programs, Journal
of the Society for Industrial and Applied Mathematics, Vol. 8, No. 4, 703-712
(1960)

[40] Kim, SG; Outhred, H .: Voltage quality of supply incorporated spot market

model: a market efficiency perspective, AUPEC2003, New Zealand (2003)

[41] Kim, S .; Song, W .; Lee, B .: Generation Redispatch Model to Enhance Voltage


Security in Competitive Power Market Using Voltage Stability Constrained
Optimal Power Flow (VSCOPF), Korea University Seoul, School of Electrical
Engineering

[42] Know Library: Internet lexicon, http://know-library.net

https://translate.googleusercontent.com/translate_f 139/143
2/29/2020 Untitled

[43] Kuhn, S .; Schultz, R .: Risk Neutral and Risk Averse Power Optimization in
Electricity Networks with Dispersed Generation, Mathematical Methods of

Operations Research (to appear) (2008)

[44] Kuhn, S .; Schultz, R .: Optimization of load flow when information is incomplete,


VDI Reports 2018, Optimization in the Energy Industry, 237-248 (2007)

[45] Louveaux, FV; Schultz, R .: Stochastic integer programming, in: Handbooks


in Operations Research and Management Science, 10: Stochastic Program-
ming (A. Ruszczynski, A. Shapiro eds.), Elsevier, Amsterdam (2003)

120

Page 129

[46] Märkert, A .; Schultz, R .: On Deviation Measures in Stochastic Integer Pro


gramming, Operations Research Letters, vol. 33, no. 5, 441-449 (2005)

[47] Moritz, S .: A Mixed Integer Approach for the Transient Case of Gas Network
Optimization, dissertation, Department of Mathematics, Technical University
Darmstadt Verlag (2006)

[48] Müller, A .; Stoyan, D .: Comparison Methods for Stochastic Models and Risks,
Wiley (2002)

[49] Murillo-Sánchez, C .; Thomas, RJ: Thermal Unit Commitment Including Op-

timal AC Power Flow Constraints, Proceedings of the Thirty-First Hawaii


International Conference on System Sciences, Vol. 3, 81-88 (1998)

[50] Neise, F .; Schultz, R .: Algorithms for mean-risk stochastic integer programs in


energy, Revista Investigación Operacional, Vol. 28, No. 1, 4-16 (2007)

[51] Nowak, MP; Roman, W. Stochastic Lagrangian relaxation applied to power

scheduling in a hydro-thermal system under uncertainty, Annals of Operations

https://translate.googleusercontent.com/translate_f 140/143
2/29/2020 Untitled
Research, Vol. 100, 251-272 (2000)

[52] Nuremberg, R .; Römisch, W .: A two-stage planning model for power scheduling


in a hydro-thermal system under uncertainty, Optimization and Engineering,
Vol. 3, No. 4, 355-378 (2002)

[53] Overhagen, T .: Lecture Optimization I and II, University of Siegen (1999)

[54] Powell, MJD: Algorithms for nonlinear constraints that use Lagrangian func-

tions, Mathematical Programming, Vol. 14, 224-248 (1978)

[55] Power Systems Test Case Archive: Examples of Power Transmission Systems
me, http://www.ee.washington.edu/research/pstca, University of Washing-
ton, College of Engineering

[56] PowerWorld Simulator, PowerWorld Corporation, Illinois (2005)

121

Page 130

[57] Römisch, W .; Schultz, R .: Multistage stochastic integer programs: An intro


duction, Online Optimization of Large Scale Systems (M. Grötschel, SO

Krumke, J. Rambau, Eds.), Springer-Verlag (2001)

[58] Ruszczynski, A .; Shapiro, A .: Optimization of Convex Risk Functions, Math

matics of Operations Research, Vol. 31, Issue 3, 433-452 (2006)

[59] Ruszczynski, A .; Shapiro, A. (eds.): Stochastic Programming, Handbooks of


Operations Research and Management Science, Vol. 10, Elsevier (2003)

[60] Sánchez-Martin, P .; Ramos, A .: Modeling Transmission Ohmic Losses in a


Stochastic Bulk Production Cost Model, Instituto de Investigación Tecnológica,

Universidad Pontificia Comillas, Madrid (1997)

https://translate.googleusercontent.com/translate_f 141/143
2/29/2020 Untitled
[61] Sánchez-Martin, P .; Ramos, A .; Alonso, JF: Probabilistic Midterm Transmis-
sion planning in a liberalized market, IEEE transactions on power systems,

Vol. 20, No. 4, 2135-2142 (2005)

[62] Schittkowski, K .: On the convergence of a Sequential Quadratic Program-

ming Method with an Augmented Lagrangian Line Search Function Systems


Optimization Laboratory, Dept. of Operations Research, Stanford University,
California (1982)

[63] Schultz, R .: On structure and stability in stochastic programs with random


technology matrix and complete integer recourse, Mathematical Programming,
Vol. 70, Issue 1, 73-89 (1995)

[64] Schultz, R .; Tiedemann, S .: Risk Aversion via Excess Probabilities in Stocha-


stic Programs with Mixed-Integer Recourse, SIAM Journal on Optimization,
Vol. 14, Issue 1, 115-138 (2003)

[65] Shapiro, A .: On complexity of multistage stochastic programs, Operations Re-


search Letters, Vol. 34, Issue 1, 1-8 (2006)

[66] Takriti, S .; Birge, JR; Long, E .: A stochastic model for the unit commitment
problem, IEEE Transactions on Power Systems, Vol. 11, Issue 3, 1497-1508
(1996)

122

Page 131

[67] Tiedemann, S .: Risk Measures with Preselected Tolerance Levels in Two-Stage


Stochastic Mixed-Integer Programming, Dissertation, Cuvillier Verlag (2005)

[68] Wallace, SW; Fleten, S.-E .: Stochastic programming models in energy, in:
Handbooks in Operations Research and Management Science 10: Stochastic
Programming (A. Ruszczynski, A. Shapiro eds.), Elsevier (2003)

[69] Wikipedia: Wikipedia - The Free Encyclopedia, http://www.wikipedia.org


https://translate.googleusercontent.com/translate_f 142/143
2/29/2020 Untitled

[70] Wolsey, LA: Integer Programming, Wiley (1998)

[71] Wood, AJ; Wollenberg, BF: Power Generation, Operation, and Control,
2nd Edition, Wiley (1996)

[72] Zhou, G .: A Modified SQP Method and Its Global Convergence, Journal of
Global Optimization, Vol. 11, Issue 2, 193-205 (1997)

123

Page 134
132
133

https://translate.googleusercontent.com/translate_f 143/143

S-ar putea să vă placă și