LICENTIATE THESIS,

THIRD CYCLE
STOCKHOLM, SWEDEN 2022 kth royal institute
of technology

Weekly planning of
Licentiate Thesis in Electric Power and Energy Systems

Weekly planning ofin

hydropower hydropower in
systems with large volumes of
systems
varying powerwith large
generation
volumes of varying
CHARLOTTA AHLFORS

power generation

Licentiate thesis

Charlotta Ahlfors
Stockholm, Sweden 2022
Weekly planning of hydropower in
systems with large volumes of
varying power generation
CHARLOTTA AHLFORS

Academic Dissertation which, with due permission of the KTH Royal Institute of Technology,
is submitted for public defence for the Degree of Licentiate of Engineering on Tuesday the 11th
October 2022, at 10:00 a.m. in Sten Velander rum 3412, Teknikringen 33, Stockholm.

Licentiate Thesis in Electric Power and Energy Systems

KTH Royal Institute of Technology
Stockholm, Sweden 2022
© Charlotta Ahlfors

TRITA-EECS-AVL-2022:56
ISBN: 978-91-8040-345-0

Printed by: Universitetsservice US-AB, Sweden 2022

Abstract

Hydropower is the world’s largest source of renewable electricity generation.

Hydropower plants with reservoirs provide flexibility to the power systems. Efficient
planning techniques improve the flexibility of the power systems and reduce carbon
emissions, which is needed in power systems experiencing a rapid change in balance
between power production and consumption. This is due to increasing amount of
renewable energy sources, such as wind and solar power. Hydropower plants have low
operating costs and are used as base power. This thesis focuses on weekly planning
of hydropower in systems with large volumes and varying power generation and a
literature review and a maintenance scheduling method are presented.

The topic of hydropower planning is well investigated and various research questions
have been studied under many years in different countries. Some of the works
are summarized and discussed in literature reviews, which are presented in this
thesis. First, some reviews are presented, which covers several aspects of hydropower
planning. Literature reviews for long term, mid term and short term planning,
respectively, are described.

Maintenance scheduling in power systems consists of preventive and corrective

maintenance. Preventive maintenance is performed at predetermined intervals
according to a prescribed criteria. This type of maintenance is important for power
producers to avoid loss in electricity production and loss in income. The maintenance
scheduling for hydropower plants prevent these phenomena since spill in the reservoirs
and wear on the turbines can be avoided. Usually, the maintenance in hydropower
plants is performed on the turbines or at the reservoir intake. A deterministic and a
stochastic method to solve a mid term maintenance scheduling problem formulated as
a Mixed Integer Linear Programming using dynamic programming is presented. The
deterministic method works well in terms of computational time and accuracy. The

iii
stochastic method compared to the deterministic method yields a slightly better result
at the cost of a need for larger computational resources.

Keywords

Hydropower, Literature review, Weekly planning, Low carbon operation, Maintenance

scheduling, Preventive maintenance, Mixed Integer Linear Programming, Dynamic
programming, Cascaded river systems, value of stored water.

iv
Sammanfattning

Vattenkraft är världens största källa till förnyelsebar elproduktion. Vattenkraftverk

med magasin erbjuder flexibilitet till elkraftsystem. Effektiva planeringsmetoder för­
bättrar flexibiliteten hos kraftsystemen och minskar koldioxidutsläppen, vilket är nöd­
vändigt i kraftsystem som utsätts för snabb förändring med obalans mellan produk­
tion och konsumtion av effekt. Detta beror på ökad andel förnyelsebara energi­
källor, som vind­ och solkraft, i kraftsystemen. Vattenkraftverk har låga driftkostnader
och används som baskraft. Den här avhandlingen fokuserar på veckoplanering av
vattenkraft i kraftsystem med stora volymer och varierande kraftproduktion, samt en
litteraturstudie och en metod för underhållsplanering presenteras.

Ämnet vattenkraftplanering är väl undersökt och varierande forskningsfrågor har

studerats under många år i olika länder. En del av arbetena sammanfattas och
diskuteras i litteraturstudier, vilka presenteras i den här avhandlingen. Först
presenteras några litteraturstudier, som täcker flera aspekter av vattenkraftplanering.
Litteraturstudier, för långtids­, medeltidsplanering, respektive korttidsplanering
beskrivs.

Underhållsplanering i elkraftsystem består av förebyggande och korrigerande under­

håll. Förebyggande underhåll utförs vid förutbestämda intervall enligt förbestämda
kriterier. Denna typ av underhåll är viktig för att kraftprodcenter ska kunna
undvika förlorad elproduktion och förlorad inkomst. Underhållsplaneringen för
vattenkraftverk förebygger dessa fenomen, eftersom spill i magasinen och slitage
på turbinerna kan undvikas. Vanligen utförs underhållen i vattenkraftverken på
turbinerna eller vid intaget i magasinet. En deterministisk metod och en stokastisk
metod att lösa ett medeltidsplaneringsproblem, formulerat som ett blandat heltals­
programmeringsproblem presenteras. Den deterministiska metoden fungerar väl i
termer av beräkningstid och noggrannhet. Den stokastiska metoden jämfört med den

v
deterministiska metoden ger ett något bättre resultat dock till priset av ett behov av
större datorresurser.

Nyckelord

Vattenkraft, Litteraturstudie, Veckoplanering, Låg koldioxid drift, Underhålls­

planering,
Förebyggande underhåll, Blandad Heltalsprogrammering, Dynamisk programmering,
Kaskadälvsystem, vattenvärden.

vi
Acknowledgements

I want to express my gratefulness to my main supervisor Mikael Amelin and my second

supervisor Lennart Söder for contributing with theoretical knowledge throughout the
work. Furthermore, I want to say thank you to Jonas Funkquist, Joakim Näsström
and Lars Abrahamsson at Vattenfall for contributing with industrial knowledge within
hydropower planning. Last but not least, huge thanks to amazing colleagues at the
Division of Electric Power and Energy Systems ­ Evelin, Priyanka, Elis, Saeed, Manuel,
Abolfazl, Iasonas and many more.
I also would like to say thank you to my parents Birgitta and Lennart Ahlfors, my
brother Martin and Tommy.

Solna 2022­05­08
Charlotta Ahlfors

vii
Acronyms

DP Dynamic programming
MILP Mixed Integer Linear Programming
SDP Stochastic dynamic programming
SDDiP Stochastic dual dynamic integer programming
SDDP Stochastic dual dynamic programming
UC Unit Commitment

viii
Contents

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research question . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Overview of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Hydropower planning 6
2.1 Planning methods for power systems . . . . . . . . . . . . . . . . . . . 6
2.2 Mathematical optimization . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 The Nordic electricity market . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Hydropower production . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4.1 The Nordic river systems . . . . . . . . . . . . . . . . . . . . . . 10
2.5 Renewable electricity production . . . . . . . . . . . . . . . . . . . . . . 11

3 Literature review 13
3.1 Previous reviews on hydropower planning . . . . . . . . . . . . . . . . . 13
3.2 Long term planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Mid term planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4 Short term planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.5 Maintenance scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.5.1 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4 Preventive maintenance scheduling 26

4.1 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

ix
CONTENTS

4.1.1 Model limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

4.1.2 Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.1.3 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.1.4 Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1.5 Dynamic programming . . . . . . . . . . . . . . . . . . . . . . . 29
4.1.6 Value of stored water . . . . . . . . . . . . . . . . . . . . . . . . 30
4.1.7 Efficiency curves for the power production . . . . . . . . . . . . 31
4.1.8 MILP formulation of the full problem . . . . . . . . . . . . . . . . 32
4.1.9 Subproblem and relaxation . . . . . . . . . . . . . . . . . . . . . 33
4.1.10 Solution by dynamic programming . . . . . . . . . . . . . . . . . 35
4.1.11 A solution to the full problem . . . . . . . . . . . . . . . . . . . . 37
4.2 Validation of the method . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.2.1 Two solution methods: MILP­solver and dynamic
programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.2.2 Case study with maintenance . . . . . . . . . . . . . . . . . . . 40
4.2.3 A second case study . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 Empirical study full river . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.4 Results full river . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

5 Stochastic model 50
5.1 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.2 Results of stochastic computations . . . . . . . . . . . . . . . . . . . . . 55

6 Discussion 58
6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

A Tables 61

References 64

x
Chapter 1

Introduction

1.1 Background
Sustainability and environment is a global concern. The global established
sustainability goals [1] aim to ensure access to affordable, reliable, sustainable and
modern energy for all (goal number 7) and to take urgent action to combat climate
change and its impacts (goal number 13).

In 2018 the amount of total world energy production originating from fossil fuel energy
sources (oil, coal and gas) was 81% and only around 14% was from renewable energy
sources [2]. To reach the goal of the Paris agreement of at most 2 degrees increase
in global average temperature from pre­industrial levels the CO2 emissions need to be
reduced to net­zero emissions by the middle of the 21st century [3]. Many countries
have also introduced legislation to achieve net­zero emissions. This puts a strong
incentive to increase the amount of renewable energy sources.

Among the renewable energy sources, hydropower is responsible for around half of
the power production [2]. Hydropower production uses the potential and kinetic
energy of water to produce power using reservoirs and hydropower turbines. Besides
being renewable, hydropower is also a flexible energy source. Other renewable
energy sources are intermittent (mainly solar and wind), which means that their
production fluctuates over short time intervals due to changing weather conditions.
To compensate for the power fluctuations of these other power sources and increasing
fluctuations in consumption the hydropower production can fulfil the important role
to stabilise the power system, as there must always be a balance between power

1
CHAPTER 1. INTRODUCTION

production and consumption. Therefore, hydropower plays an important role in the

power system and will so even more in the future since renewable energy sources
increase and other non­renewable flexible power sources are phased out.

However, there are some drawbacks with hydropower that must be kept in mind
related to sustainability. A hydropower unit has a large local environmental effect
disrupting the fauna and wildlife surrounding the unit and displacing nearby residents
in towns and villages. There is also an effect on the whole ecosystem of the river where
the unit is located as it changes water ways and increases the risks of drying up parts
of the river during the dry season. The hydropower unit blocks fishes from swimming
past the reservoir and diminishes the population of important species. On a global scale
most of the suitable river systems already have hydropower installed and this leaves
small room for increasing the hydropower production in the future. This means that
hydropower must be utilised in an efficient manner and create sustainable planning
methods [4].

An important part when running a hydropower unit is the planning of the power
production. In this planning different time horizons must be considered. In the long­
term planning, stochastic seasonal variations in precipitation need to be considered as
well as even variations over longer time periods of multiple years (e.g., changes due to
climate change). These plans concentrates on the utilization of the largest reservoirs
which are usually located at the top of the river system. In the short term planning,
with a time horizon of just a few days, more immediate factors influence the discharge
plans, such as the current electricity price. Somewhere between these perspectives
weekly planning of hydropower is relevant.

The planning period of a few weeks serves as an important bridge between the
long term planning and short term planning. Weekly planning can use more
detailed information about reservoir levels, electricity prices, power consumption and
maintenance of turbines not considered in the long term planning. In contrast to the
short term planning, weekly planning spans over a longer time period and it is therefore
possible to account for changes within the planning that have a duration of a few days,
which is not possible in the short term planning. Moreover, since the forecasts for the
relevant information is more uncertain in 1­3 weeks horizon than in the next 24 hours
it is important to consider stochastic behaviour in the weekly planning.

2
CHAPTER 1. INTRODUCTION

1.2 Research question

In weekly planning it is of significance to retain a high level of details in the problem
formulation. This enables the study of various important questions, such as the
influence of reservoir sizes and delay time of water running between reservoirs, on
the ability of hydropower to function as a regulating energy source in the presence of a
large amount of renewable energy sources. If stochastic behaviors are modelled, such
as reservoir inflow, power production from other renewable energy sources (wind and
solar) and electricity prices, this bring further significance to the model. However, the
challenge with such detailed models that span over time periods of weeks is that the
computational burden to solve the planning problem increases dramatically compared
to solving the same problem for a short time period of a few hours to days. Therefore
various approximations need to be included in the model.

The research question in this thesis is to study how detailed models can be used for
weekly planning of hydropower. Especially, modeling turbines and keeping a high time
resolution in the model, are parts of the research question.

1.3 Objectives
Earlier research has paid attention to long term planning and short term planning
of hydropower. This project aims at developing efficient methods to plan the
coordination between wind power and hydropower for weekly planning of hydropower.
Weekly planning is needed when planning should be performed for one or two weeks
ahead in order for the power producer to make an early plan before bidding to the day
ahead market or to make specific maintenance plans. This type of planning makes it
possible to reduce loss in income, avoid failure in turbines and unplanned disruptions.
Stochastic behaviors can also be considered, such as inflow to the reservoirs and
electricity prices. The following specific research objectives have been the foundation
for the work:

• Develop efficient planning methods for power systems with a large share of
renewable power sources, such as wind power. The first model is a deterministic
maintenance scheduling model and the second is a stochastic maintenance
scheduling model. This is due to the fact that this has been identified as a research
gap and is relevant for the power producers. Relevant time resolution to be used

3
CHAPTER 1. INTRODUCTION

for these models should be considered. The models should be tested on two case
studies, where two subsets of a cascaded river are used as test systems.

• Study the connection between short term, mid term and long term planning of
hydropower.

1.4 Contribution
This section gives an account of the contributions in this thesis.

• The first contribution is a method based on dynamic programming that is used to

solve detailed hydropower models including several hydropower stations, several
turbines and detailed time resolution.

• The second contribution is to show that the method can be used to asses the costs
of turbine maintenance.

• The third contribution is an application of the dynamic programming method to

stochastic planning problems.

The contributions are presented in one conference paper to Power Tech 2021 and one
Journal paper.

• Weekly planning of hydropower in systems with large volumes and varying power
generation: A literature review [5],

• Preventive maintenance scheduling of hydropower plants using dynamic

programming [To be published].

1.5 Overview of the thesis

The thesis consists of six chapters. The main chapters are the literature review,
preventive maintenance scheduling and the stochastic model, which include scientific
contributions. A summary of each chapter is found below:

• Chapter 1: Provides an introductory to the thesis, explains the research

question, objectives, contribution and overview of the thesis.

• Chapter 2: Explains hydropower planning and related topics; different

methods, mathematical optimization, the nordic electricity market, hydropower

4
CHAPTER 1. INTRODUCTION

production and renewable electricity production.

• Chapter 3: Presents a literature review that has been carried out including
previous reviews on hydropower planning in different time perspectives and
maintenance scheduling.

• Chapter 4: Describes preventive maintenance scheduling and contribution 1

and 2.

• Chapter 5: Describes the stochastic model and contribution 3.

• Chapter 6: Gives a summary of the content and suggestions for future work.

5
Chapter 2

Hydropower planning

2.1 Planning methods for power systems

Planning is aimed to utilize resources in the power system as efficiently as possible.

It requires considering both technical and economical aspects. In the shortest time
frame, seconds and minutes, the technical aspects are most important. It is important
to keep a stable frequency and keep the balance between production and consumption
at all times. In a time frame of hours or days, maintaining the function of the system
is important but it also becomes important to consider economical decisions. When
the time perspective is weeks or months, economical decisions play a major role. The
available energy that can be utilized must be planned for the coming period. For
hydropower plants this means that the producer must plan on how to utilize the water
that is stored in the reservoirs. But also maintenance planning is important for this
time frame. In the longest time frame, for months and years, investment decisions are
crucial. Furthermore, in the case of hydropower plants, decisions are also taken on
how to utilize the water that is stored in the largest reservoirs for the coming year or
months [6].

For hydropower planning, specifically, the planning is divided in plans for the coming
years or seasons (long term planning), months (mid term planning) and plans for
short time frame of the next few days (short term planning). This can be seen in
figure 2.1.1.

6
CHAPTER 2. HYDROPOWER PLANNING

Short term
Long term planning Mid term planning planning

Years or several months Months or weeks Week or days

Figure 2.1.1: Different planning horizons

2.2 Mathematical optimization

The planning problems studied in this thesis are formulated as mathematical

optimization problems and what follows in this section is a short description of the
theory of optimization problems and a few examples of the most common types of
optimization problems.

A general formulation of an optimization problem can be written in just two lines

min f (x),
s.t. x ∈ U.

The variables of the problem are collected in a vector x ∈ Rn and under the constraint
that x ∈ U the objective is to find the minimal value of the function f (x) [7].

Probably the most common type of optimization problem are linear programming
problems (LP) and they can be formulated as

min cT x
s.t. Ax ≥ b,
x ≥ 0,

where A is an m×n matrix, b ∈ Rm and c, x ∈ Rn . In these types of problems the

objective function f (x) = cT x is a linear function and the set U is defined in terms of
linear inequalities. The mathematical theory for these types of problems is rich and an
extensive set of methods have been developed to solve them. Two prominent methods
are the Simplex Method and various interior point methods [7]. If some of the variables
are assumed to be integers then we have a mixed integer linear programming problem
(MILP). Some of the problems in this these are formulated as MILP problems.

For every linear programming problem there is a corresponding problem called the

7
CHAPTER 2. HYDROPOWER PLANNING

dual problem. For every variable in the original problem, denoted the primal problem,
there is a constraint in the dual problem and for every constraint in the primal problem
there is a variable in the dual. The dual problem has the form

max bT y
s.t. AT y ≤ c,
y ≥ 0,

where AT is the transpose of A and y ∈ Rm . An interpretation of the values of the

dual variables y is that the component yi measures how the optimal value of the primal
problem changes if constraint i is relaxed. The dual variables are therefore often called
shadow prices. This formulation is not directly used in the thesis but is important in
the methods of the solver that has been used.

A useful and interesting class of optimization problems is stochastic optimizations

problems. In these problems a subset of the variables are stochastic and the objective
function is to optimize the expected value of an expression. In the linear case such a
problem takes the form

min cT x + E (dsT ys )
s.t. Ax ≥ b,
Ts x + W ys = hs
x, ys ≥ 0

where s is a stochastic variable, and ds , Ts and hs are stochastic functions dependent

on s and ys is a variable in the problem that can depend on s. This type of problem is
used in this thesis when the planning problem is studied with multiple price scenarios.
However, due to the scenario based formulation these problems tends to be large and
needs solution algorithms such as dynamic programming to be solved, which is also
used in this thesis to handle this problem. Another option is to solve these type of
problems by sampling the outcomes and apply Monte Carlo simulation [8].

One way to handle certain large linear optimization problems is by applying dynamic
programming. For example linear programming problems including integer variables,
so called Mixed Integer Linear Programming problems, can in favourable situations
be solved by dynamic programming. A description of dynamic programming is given

8
CHAPTER 2. HYDROPOWER PLANNING

in chapter 4. Dynamic programming can also be applied to stochastic optimization

problems and is then called stochastic dynamic programming.

2.3 The Nordic electricity market

Usually, in everyday language, the market that is referred to as the electricity market is
the day­ahead market. In this market the power producers trade power for a number
of trading periods in advance. The Nordic electricity market Elspot, is a spot market
and day­ahed market where hourly contracts for the coming 24 hours are traded.
Bids for the next day can be submitted until 12:00 the day before and at 15:00 prices
and quantities for the different hours in the next 24 hours are released. The market
price is computed as a price cross between demand and supply curves for the different
hours [9].

When the day ahead market has closed, in the hours before the power should be
delivered, adjustments to the trading plan may have to be performed. This trading
is handled in a separate market. The Nordic intraday market is called Elbas and actors
can place or accept bids for a specific hour after the release of prices on Elspot, until
one hour before the delivery hour [9].

The system operator has the responsibility to keep power balance between production
and consumption at all times and this is done at balancing markets. In Europe the
nominal value for the frequency is 50 Hz and a portfolio of frequency reserves are used
to keep this value. The Swedish system operator uses the reserves FFR, FCR­N, FCR­D
up, FCR­D down, aFRR and mFRR [10]. The reserves are called support services and
operate either automatically or manually and at different frequencies. This can be seen
in table 2.3.1.

2.4 Hydropower production

Hydropower is a carbon emission free energy source that can be used to keep the
balance in the power system when use of intermittent renewable energy sources, e.g.
photo voltaic generation and wind power, is rapidly increasing. In Figure 2.4.1 a
typical hydropower plant is shown. Power is generated using the difference in potential
energy between two water surfaces. Water is released from the reservoir, through the

9
CHAPTER 2. HYDROPOWER PLANNING

Table 2.3.1: Support services for the power system used by the Swedish system
operator [10].

Type Activation Activation time

FFR Automatic Three options for 100%: 0.7 s
at frequency deviations at low at 49.5 Hz, 1.0 s at 49.6 Hz
levels of inertia & Automatic and 1.3 s at 49.7 Hz
at 49.9­50.10 Hz
FCR­N Automatic at 49.9­50.10 Hz 63% within 60 s and 100%
within 3 min
FCR­D (up) Automatic linear activation at 50% within 5 s and 100%
49.9­49.50 Hz within 30 s
FCR­D (down) Automatic linear activation at 50% within 5 s and 100%
50.1­50.5 Hz within 30 s
aFRR Automatic activation at 100% within 120 s
deviations from 50.00 Hz
mFRR Manually activated 100% within 15 min

intake into the penstock. After passing the penstock water passes a turbine, which
is connected to a generator, to the outlet and the river. The amount of power that
can be generated depends on discharge, head height and efficiency of the turbine. The
relation between discharge and power production in a power plant is a non­linear, non­
concave function. This causes problems when models are constructed, since a global
maximum cannot be guaranteed when a solution is reached and other assumptions are
required [8]. In section 4.1.7 this non­linear relation is modeled by a piecewise linear
function.

2.4.1 The Nordic river systems

Sweden has several regulated rivers. Most of the rivers are controlled by
Vattenregleringsföretagen, which has the responsibility to make sure that several
hydropower owners can operate in the same river [12]. Vattenregleringsföretagen
decides how much each hydropower owner can produce each our and keep track of the
production to guarantee a fair system. Typically the Swedish rivers have low heads, i.e.
less than 100m. The rivers are cascaded, which means that water that is released in the
upper reservoirs can be used in hydropower plants downstream. The most commonly
used tubines in Swedish hydropower plants are Kaplan and Francis. A map of the
Swedish hydropower plants can be seen in figure 2.5.1. The Norwegian rivers have
larger head, are less cascaded and Pelton turbines are commonly used.

10
CHAPTER 2. HYDROPOWER PLANNING

Figure 2.4.1: Principle sketch of a hydropower plant [11].

2.5 Renewable electricity production

Renewable energy is continuously recycled from natural resources. The use of
renewable energy leads to reduced carbon emission and less global ecological foot
print. Examples of renewable energy sources are wind power, solar power and
hydropower. Wind and solar power are intermittent power sources that cannot be used
as regulating power since they are weather dependent and their stochastic behaviours
create fluctuations in the power production. These fluctuations can be compensated
by introducing energy storage in the power system. Another option is to combine wind
and solar power with hydropower plants that can balance the fluctuations. This also
means that the system operator has to account for larger frequency deviations when the
amount of wind power and solar power is increased in the power system [8]. These
are challenges that must be handled since the electricity demand is globally increasing
and there is a limited possibility to construct new hydropower plants.

The scope in this thesis is the Nordic river system and the case studies in chapter 4 and
5 are performed in one single owned river.

11
CHAPTER 2. HYDROPOWER PLANNING

Figure 2.5.1: Map of the Swedish hydropower plants and their sizes [13].

12
Chapter 3

Literature review

3.1 Previous reviews on hydropower planning

The topic of hydropower planning is well investigated and various research questions
have been studied under many years in different countries. Some of the works are
summarized and discussed in literature reviews, which are presented in this chapter.
First, some reviews are presented, which cover several aspects of hydropower planning.
Literature reviews for long term, mid term and short term planning, respectively, are
then described. This chapter is based on the Power Tech 2021 conference paper Weekly
planning of hydropower in systems with large volumes and varying power generation:
A literature review.

In [14], a review of scheduling under uncertainty is presented. This study spans

over many fields of science and the different techniques used to model uncertainties.
Hydropower planning is one of the discussed topics. Some of these techniques are two­
stage stochastic programming, parametric programming, fuzzy programming, chance
constraint programming, robust optimization, conditional value­at­risk and other risk
reducing techniques. In [15], a review of operation of hydropower plants is presented.
It is concluded that the main parameters affecting the operation of a hydropower plant
are head, discharge, turbine and the generators. In [16], a review on hydropower plant
models and control is performed. It summarizes the research work from 2005 and
earlier of hydropower plant model development and its controller design.

In [17], from 2020, a comprehensive review is performed which spans over different
time frames of hydropower scheduling. The different time spans for planning are

13
CHAPTER 3. LITERATURE REVIEW

defined as follows: short term 2­14 days, mid term 3­18 months and long term 1­5
years. An important conclusion and suggestion for future work is that scheduling of
hydropower systems would be more valuable by considering other sustainable energy
sources, such as wind and solar power, included in the optimization. Some models
take this into account but it can be further developed.

A review of stochastic dual dynamic programming (SDDP) for long term hydropower
planning is presented in [18]. It is concluded that the main strength of the method is to
obtain reliable water values for all reservoirs and part of the computations can be done
using parallel processing, due to the dual formulation of the optimization problem. Its
weakness is the long computational times. In [19], a review is made of the application of
the SDDP method applied to long term planning in the Nordic countries. For example,
head variations and price modelling are discussed.

Several other methods for long term planning is reviewed in [20]. The methods are
categorized in four main classes: Implicit stochastic optimization, explicit stochastic
optimization, real­time control with forecasting and heuristic programming.

Recent hydropower optimization research and development using metaheuristic

approaches are discussed in [21]. For example, physics based algorithms, gravitational
search algorithms, and swarm based algorithms, such as particle swarm optimization,
are reviewed and suitable for long term planning.

A review of short term planning in power systems with a large amount of wind power in
the system is presented in [22]. The paper discusses deterministic model approaches
as well as stochastic models. The conclusion in this paper is mainly that introducing
several types of uncertainties lead to more realistic models but also models that are
more difficult to study and analyse.

Earlier literature reviews have paid attention to the focus areas which have been
mentioned. However, new research related to hydropower planning is of importance
and weekly planning has not been in focus in earlier reviews. The scope of this chapter
is to summarize the reviews until today within hydropower planning, present selected
representative studies within the different time frames of hydropower planning and
present papers focusing in particular on weekly planning.

14
CHAPTER 3. LITERATURE REVIEW

3.2 Long term planning

The aim of the long term planning is to make plans for one to several years in advance,
where many parameters are uncertain, such as inflow and electricity prices. There are
many examples of models where the stochastic behaviour is taken into account. The
output from the long term planning is often water values or a weekly discharge plan of
power generation. The output is used in mid term or short term planning directly.

In [23], water value methods that dates back to the 1960’s are used. In water value
methods aggregation of the reservoirs is done to prevent the curse of dimensionality
and make the computations efficient. In this article only one reservoir is considered to
overcome this problem. Stochastic dynamic programming (SDP) is used and compared
to results from deterministic dynamic programming (DP). A conclusion is that the
hydropower plants is operated more efficiently with DP rather than with SDP, since the
DDP method results in higher reservoir levels and higher efficiency. However, more
water spillage can be a consequence. The conclusions are drawn based on a case study
in Brazil.

Another paper where the SDP method is used is [24] where the costs are reduced
compared to when a standard Markovian model is used. Inflows are determined in
a stochastic process as Markov­chains.

SDDP was introduced in [25], which allows for modeling of multi­reservoir systems
without aggregation and leads to more realistic and detailed models.

One major factor affecting the outcome of the SDDP method is a correct statistical
model of the inflow and this is studied in [26] and [27]. In [26] a new level of detail is
presented with respect to the hydro reservoirs. The approach is tested as a large scale
market model. Typical SDDP models tend to consider smaller systems. The case study
is performed on the Norwegian power system. The method is not recommended to
involve individual decision variables for each reservoir.

Recent focus of the research on SDDP is on convergence properties, strategic bidding,

risk aversion, emission reduction and integration of wind power and pumped­storage
as outlined in [28]. A hybrid SDP/SDDP aproach is used. The model is tested on a
7 reservoir system in Norway with corresponding power stations with a scheduling
horizon of two years and weekly decision stages. Traditionally long/mid term models
only consider the energy market but a conclusion of the work is that a market for a

15
CHAPTER 3. LITERATURE REVIEW

reserve capacity can be introduced as an extension.

Another example of a significant paper for long term planning is [29], which compares
two models for hydro­thermal scheduling. Both models takes into account inter­
connected power markets and using detailed hydro modelling. The reference model
is the EMPS model and the other model is the SOVN­model. The SOVN­ model uses
formal optimization while the EMPS­model uses a heuristic approach. The EMPS
model uses the objective to minimize the expected cost in the whole system subject
to all constraints. The simulated system in the EMPS model can be the Nordic system
or Northern Europe. The planning horizon is up to 25 years and the time step is one
week.

In [30] a stochastic dynamic programming and successive two­stage optimization

problem is solved. The proposed method is tested in the hydropower system of
Yuanhe River in Hunan province China. The results show that the method has good
performance to optimal scheduling. In [31] flexible robust optimization dispatch
for hybrid wind/photovoltaic/hydro/thermal power system is used. The model and
the algorithm performs well. In terms of reliability and economy the results of the
optimization dispatch are good.

3.3 Mid term planning

Mid term planning models do not consider multi­year horizons but limit their attention
to months. The models for mid term planning include more details and output such as
water values.

In [32], a model for the weekly scheduling of a hydropower system is developed,

where bidding to both the day­ahead market and the reserve market are included. The
model is constructed to protect against the risk of shortage of water and shortage of
storage. It is shown that hydropower producers who participate in both markets have
a significantly higher expected income than those only taking part in the day­ahead
market. This method has the strength to account for both markets but at the same
time it requires a proper scenario tree representation of the stochastic variables.

The model in [33] adds forward contracts as an ingredient in the operation of a

hydropower plant, as well as pumped storage. It uses multistage stochastic quadratic
programming. The results are water values and optimal forward bids to support the

16
CHAPTER 3. LITERATURE REVIEW

daily operation. A case study is performed, which shows that the forward trading
contributes to a 20% increase in profit. Its main strength is that the model is both
realistic and tractable. On the other hand, it assumes a linear dependence between
the amount of commodities and market price and is also limited to a horizon of one
year.

In [34], a mid term and long term electricity contract decomposition method is
proposed that considers the cascade water quantity matching constraints. Mid term
and long term power quantity contracts are decomposed by calculating the matching
constraint of cascade water quantity. The method is tested in the Yunnan power spot
market in China. The results show that the model is applicable for the Yunnan power
spot market.

In [35], stochastic linear and non linear programming is used. The objective is to
maximize the expected revenue of a producer, and the decision variables are generation
and forward contracts in each period for each scenario. The parameters are limited by
constraints on reservoir balance of water, upper and lower limits of power generation,
bilateral contract and reservoir spillage.

In [36], a model is suggested of combined SDP and SDDP, where variables and
functional relationships are linear. The main purpose is to fulfil the hydropower
units operators’ demands to get steady operation for the power grid. A case study
is performed in Lysebotn in Norway. The study has resulted in a quantitative
valuation including sales of capacity. However, the results from the paper may be case
specific.

An extended version of the SDDP method is used in [37], stochastic dual dynamic
integer programming (SDDiP). The method is applied to mid term planning. A test
is performed on how to incorporate stagewise­dependent stochastic variables on the
right­hand side of the objective function. The method has proven convergence in the
nonconvex case. A case study is performed for a Norwegian hydropower producer.
The case study shows that it is possible to solve the mid­term hydropower scheduling
problem but time consuming to solve it to the optimal value.

In [38], the SDDiP method is applied in another Norwegian case study. The
paper highlights that the method can be used to include nonconvex environmental
constraints in mid term and long term planning. A case study is performed on a
multireservoir system, with three reservoirs and corresponding hydropower plants. A

17
CHAPTER 3. LITERATURE REVIEW

particular nonconvex environmental constraint is modelled and a scheduling horizon

of one year and weekly decision stages was applied.

3.4 Short term planning

The short term planning is performed for the nearest future and input parameters
depend on the planning in the longer time horizons. The outcome of the planning is
a production plan, usually with hourly resolution, of the amount of water that should
be used in each power plant of the planning period and submitted to the day­ahead
trading. Since the purpose is to find a detailed generation schedule, the models used
for this planning include detailed information of the hydropower plants, turbines, and
the topology in the river. Furthermore, the models include data for efficiency curves,
discharge, inflow, spillage and reservoirs sizes.

In [39], a mixed integer stochastic model is developed that optimizes the bidding in
a day­ahead market. The model treats the inflow and market price as stochastic.
Through simulation, the results of the model are compared to results from a
deterministic practice­based model, similar to what is used in the industry. The
conclusion of the simulations is that the stochastic model performs better by 0.7%.
This model has a number of positive features. For example, it is fast, can be used on
a daily basis and is designed for the scheduling hierarchy used by producers today.
The framework can be used to study other approaches to short term planning. The
study is, however, limited to one system and needs further work to reach a reliable
conclusion.

In [40], a stochastic short term planning model is proposed where day­ahead electricity
prices and water inflows are modeled as scenarios based parameters on time series
analysis. The model is constructed for a river with two cascaded hydropower plants.
The schedule is determined for 7 days.

A non­linear dynamic programming model of a hydropower plant is presented in [41],

where head and efficiency curves are treated as non­linear. The revenue of a producer
that sells electricity in a pool­based electricity market is maximized. The plan is made
for 1 day to 1 week with hourly resolution and the outcome is both the number of units
in operation and the generation dispatch. A case study is made where the model is
tested and an optimal weekly operation schedule has been obtained. The results of

18
CHAPTER 3. LITERATURE REVIEW

the model have been compared to results from a nonlinear programming model in 15
different daily cases. An increase in revenue is found with the dynamic programming
model, ranging from 0.9% to 3.3%.

In [42], the non­linearity effects of head and efficiency curves are included in a
stochastic model, where inflow and market price are considered uncertain. The non­
linearity is handled using successive linear programming. The model is tested in
three fictitious case studies and performs well. The paper investigates which factors
influence the result the most but reaches no general conclusion.

Non­linearity is studied in [43]. A deterministic non­linear mixed integer optimization

model is used to solve short term planning problems. The optimization problem is
solved using Lagrangian relaxation leading to separable dual subproblems that are
solved by the bundle method. A small hydropower system is used as a case study. The
conclusion is that this approach is feasible but needs to be further studied on larger
scale.

In [44], a mathematical model is proposed to simulate day­ahead markets of large­

scale multi energy systems with high penetration of renewable energy towards 2050.
Unit commitment (UC) is added to the model and the results are analysed. Including
unit commitment constraints with integer variables lead to a more realistic behavior of
the units at the cost of increasing the computational time. Relaxing integer variables
significantly reduces the computational time without compromising the accuracy of the
results. Hydro reservoirs were not modelled with UC in order to reduce the complexity
of the problem.

The novelty of [44] involves to develop a method for integration of renewable energy
sources, planned maintenance scheduling, influence of unit commitment, multi­
energy markets, full year hourly results, annual storage scheduling, several years of
scenarios and a large scale system. The optimisations and simulations, except for
variable renewable energy, are performed with the energy system model Balmorel
(an open source, energy system tool, deterministic with bottom­up approach). The
variable renewable energy hourly resolution time series data is simulated in CorRES
and used as input to Balmorel. Data is aggregated into the regions that are used in
Balmorel. A case study of the North Sea offshore Grid Denmark project was performed.
Countries involved are Germany, Great Britain, Norway, Denmark, Belgium and
Netherlands. The outcome of the tested scenario of electricity and heat show an

19
CHAPTER 3. LITERATURE REVIEW

increase in CO2 ­price: 6.06, 76.70 and 130.38 €/ton in 2020, 2030 and 2050,
respectively. The aggregated planned energy content along the year of hydro reservoirs
with seasonal inflow of the Norwegian hydro storage reservoirs is investigated. Results
from scenarios for 2020, 2030 and 2050 strengthen the importance of performing full
year optimisation towards 2050. A conclusion from the study is that the case study is
valid to other regions than the North Sea region, where high penetration of variable
renewable energy is expected. The penetration of wind and solar is likely to challenge
the need for balance in the system and the profitability of thermal units.

In [45], the integration of wind power into a power system with limits on transmission
capacity is studied. Stochastic linear programming is used and includes trading on
the spot and regulating market. It is concluded from the study that coordination of
hydropower production with wind power is beneficial for the utility of both power
sources, great reduction in wind power curtailment and a more efficient usage of the
transmission lines. Some model limitations are mentioned, such that the outcome of
stochastic variables are assumed to be known in planning for the next day. Future
work is suggested, for example, that a longer case study needs to be performed
and the impact of coordination of start­ups for hydropower plants needs to be
investigated.

In [46], a short term model, formulated as a mixed integer linear programming (MILP)
problem, is presented. Power output and the number of units committed at each
hydropower plant and hour of the day or week ahead are modeled. The aim of the
planning is to maximise the hydropower efficiency while reducing start up or shut down
costs and produce sufficient power to cover the load in the system. The method is to
relax the hydraulic constraints, simulate the solution and reintroduce constraints that
are violated in an iterative way. This leads to a computational efficient method that
can be used in a large hydropower system.

In [47], a non­linear model for hydropower planning and scheduling of maintenance

is proposed. The thermal generation complement is mimimized and the future value
of stored water is maximized. The maintenance scheduling is modeled as a continuous
variable. In [48], an algorithm is presented for the unit commitment problem that
includes head and efficiency curves. The optimization problem is formulated as
MILP problem. The optimization results in an accurate planning with reasonable
computational time, since aggregation is used in the model.

20
CHAPTER 3. LITERATURE REVIEW

A stochastic mixed integer model is developed in [49], that includes bidding to the day­
ahead and balancing market and with start/stop decisions. Three implementations
of the start/stop decisions (using none, real and binary variables) are tested. The
conclusion is that there is only a small additional value in participating in the balancing
market and there is a weak dependence of start/stop costs. The case study is performed
on a small hydropower system in Norway and the power prices do not include the effect
of a larger amount of intermittent power sources.

There are many other notable works on short term planning and a few are mentioned
in this paragraph. The commitment problem in power systems with intermittent and
fluctuating characteristics is studied in [50]. In [51], an overview is presented of
formulations and optimization methods for unit­based short term hydro scheduling. A
short term planning method is presented in [52] where Benders decomposition is used
to solve a hydropower maintenance scheduling problem. Moreover in [53] a stochastic
short­term hydropower optimization method which emphasizes inflow scenario tree is
presented. Results show that there is a gain in using a stochastic model for the short­
term hydropower optimization model with multiple scenarios compared to a median
scenario since it offers a more robust solution.

3.5 Maintenance scheduling

Maintenance scheduling in power systems consists of preventive and corrective
maintenance. Preventive maintenance is performed at predetermined intervals
according to a prescribed criteria. This type of maintenance is important for power
producers to prevent wear on turbines and to repair malfunctioning equipment, and
thus to avoid loss in electricity production and loss in income. Usually the maintenance
in hydropower plants is performed on the turbines or at the reservoir intake. The
maintenance periods usually last for a few hours to weeks and in a few cases for months.
During the maintenance a part of the power plant is out of service and in some cases
the whole power plant is affected and offline.

Many studies have been carried out on maintenance scheduling in different areas of
electrical applications. A literature review is done in [54], of maintenance scheduling
in regulated and deregulated electricity markets, and MILP models are discussed as
possible solution methods. For example, [55] is mentioned where outage planning
for thermal power plants is studied. The model presented is a stochastic MILP model

21
CHAPTER 3. LITERATURE REVIEW

and the maintenance scheduling is based on hourly price based unit commitment. The
random variables are modeled with scenarios and the problem is solved with the Monte
Carlo method. The maintenance is fixed within a time window for each thermal unit
and the duration of the maintenance is also specified. The objective function is the
best cost effective solution of the maintenance, thus the maintenance is a decision
variable. Furthermore, there are several constraints on the maintenance. A case study
is performed with 29 units for one year, time steps of one hour and the maintenance
duration for each unit is one day.

A model of maintenance for long term hydropower planning is presented in [56]. The
objective function is quadratic, i.e. the model is nonlinear. The duration of each
maintenance is fixed but the start time of the maintenance are decision variables. The
model is deterministic and the time steps are days. The objective is to decide the
most cost effective way to schedule the maintenance for each turbine in the power
plants.

In [57], a linear stochastic model is developed for long term planning of 1 year for
different types of power plants. Benders’ decomposition technique is used to solve the
resulting model. The maintenance periods for the hydropower plants are variable. The
objective of the planning is to minimize start­up costs of generators, production costs of
generators and maintenance cost to put a generator into maintenance. The conclusion
is that the method has interesting convergence characteristics. Several turbines in each
power plant are not considered nor the hydrological balance is not considered.

In [58], two models are developed. The first model is deterministic and formulated
as a MILP problem. The second model extends this approach to include the effect of
uncertain water inflows by formulating a two stage stochastic linear problem. This
problem is solved by using Benders decomposition in combination with acceleration
techniques. The conclusion is that the second model is superior to the first model.
The aim of the planning is to maximize the value of electricity production minus
maintenance costs under a number of constraints. The planning period is 30 days with
18 maintenance tasks. Maintenance is a decision variable. A case study is performed
on both models where four hydropower plants of Rio Tinto in Canada are used as test
system. Turbines are not modeled separately.

The importance of integrating the maintenance scheduling in the detailed hydropower

scheduling is highlighted in [59]. A method to solve these problems as one is developed

22
CHAPTER 3. LITERATURE REVIEW

Figure 3.5.1: World map showing countries (indicated by grey dots) where case studies
from papers in this chapter are performed.

by the use of Benders decomposition. The maintenance scheduling problem is solved

as a MILP problem in the first step providing a trial maintenance scheduling to be
considered in the hydropower scheduling. The hydropower scheduling is evaluated
using multi­stage Benders decomposition. Furthermore, the hydropower scheduling
part of the problem is evaluated using stochastic dynamic programming and stochastic
dual dynamic programming. Inflow to reservoirs, prices for energy and reserve
capacity are treated as stochastic variables. Maintenance is a decision variable and
the maintenance duration is weeks. Hydropower units are considered but turbines
are not modeled separately. The method has been tested in a case study on a
Norwegian watercourse with 7 hydropower reservoirs and associated power plants and
the conclusion is that the model works well on the tested system.

3.5.1 Tables

Three tables summarizing the review papers are presented in appendix A; Table A.0.1,
Table A.0.2 and Table A.0.3.

Figure 3.5.1 shows countries (indicated by grey dots) where case studies from papers
in this chapter are performed. Several studies are performed in Europe, but also in
Canada, Brazil, New Zealand and China.

23
CHAPTER 3. LITERATURE REVIEW

3.6 Summary

Different reviews have paid attention to, for example, review of reservoir based as well
as run of river hydropower plants, hydropower plants model and control, different
algorithms for different time frames, methods for long­term planning models and short
term planning models. Moreover, some reviews have paid attention to SDDP for long
term hydropower planning and its application in the Nordic countries.

Models for long term planning are performed in e.g. the Nordic countries, Brazil, New
Zealand and China. Mid term planning models have been developed in e.g. Spain,
Switzerland, China and Norway. Models for short term planning has been performed
in the Nordic countries, Spain, Brazil and Canada.

The long term planning models are aimed for planning for years or several months.
They cover both deterministic and stochastic models. The SDDP method is commonly
used in different models.

Mid term planning models do not consider multi­year horizons but limit their attention
to months. The models for mid term planning include more details and output such as
values of stored water. Earlier studies have paid attention to decomposition, stochastic
linear and nonlinear programming, combined SDP and SDDP and stochastic dual
dynamic programming. Short term planning is aimed for a week or days. Many studies
have been performed within this time frame and they include methods such as mixed
integer stochastic model, stochastic model handling non­linearities with time series
data, successive linear programming, non­linear Mixed Integer optimisation model,
unit commitment model and mixed integer linear programming model.

3.7 Conclusion

Power systems are rapidly changing into a large share of renewable power sources.
This requires development of hydropower models, which are efficient and suitable
for these power systems. The planning models used until today are denoted as
long term planning, mid term planning and short term planning. From this study
it can be concluded that the definitions of the different planning time frames are
ambiguous and vary. Moreover, within the studied works on maintenance scheduling
no work was found which focuses on weekly planning, where turbines are modeled

24
CHAPTER 3. LITERATURE REVIEW

separately and where dynamic programming is used as solution method. A possible

explanation to why this area has not been studied thoroughly is that weekly planning
is in between short term and mid term planning. This means that the methods used
in short term planning are difficult to extend to the longer time frame used in weekly
planning. At the same time, the approximate methods used in mid term planning are
not sufficiently accurate. A motivation for studying the planning problem within the
weekly perspective is that even a slight improvement of the outcome of a percentage
or a fraction of a percentage leads to a substantial increase in revenue in the order of
millions of euro for the power plant owner.

25
Chapter 4

Preventive maintenance scheduling

Preventive maintenance is used to keep hydropower plants in good condition and avoid
failures. The preventive maintenance usually lasts for one hour up to several days
and is performed on the hydropower turbines to prevent water spill, loss in electricity
production and wear on the turbines. Plans that take disruptions into account are
required to control the maintenance. These plans can be utilized by the power producer
in order to schedule the maintenance of turbines for a few hours up to the next fourteen
days. The net result of a successful maintenance planning can be a substantial financial
gain for the power producer.

In Sweden, there are rivers with several cascade connected power plants and the exact
configuration of the power plants varies between the different rivers. Each power
plant usually have several turbines and commonly a few turbines are disconnected
at the same time during the maintenance. A power plant is seldom shut down for
maintenance of all turbines at the same time. Therefore, the maintenance plans
require that the possible combinations to switch the turbines on and off during a
planning period are considered. However, the complexity of the problem is to avoid
a combinatorial explosion, i.e. a rapid growth of the complexity of a problem. The
combinatorial explosion may occur due to how the combinatorics of the problem is
affected by the input, constraints and bounds of the problem. A method that avoids
the combinatorial explosion is required to solve a maintenance scheduling problem.
Standard techniques, such as Mixed Integer Linear Planning (MILP), are not adequate
due to the size of the problem.

This chapter consists of the following parts. First the problem is formulated as a MILP

26
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

problem. Then a relaxation of the problem using value of stored water is performed.
The relaxed problem is solved by dynamic programming. The MILP problem is
solved by using the relaxed solution. To test the planning method a case study is
performed where a Swedish river with cascade coupled power plants has been used.
Finally, a stochastic method is presented, with three different setups for electricity
price scenarios, which is compared to the deterministic method.

4.1 Problem formulation

A method to solve preventive maintenance scheduling problems is developed in this

chapter, where planned outages in specific turbines are assumed. The purpose of this
chapter relates to the research objective: develop efficient planning methods for power
systems with a large share of renewable power sources, such as wind power. The
problem that should be solved is a scheduling problem for a mid term planning period
of two to three weeks. This problem is solved in two situations, in the first situation
the problem is deterministic and the second situation stochastic electricity prices are
taken into account. The first situation is handled in this chapter. This leads to a large
number of variables because of the number of hydropower units, reservoirs, segments
and time steps. Therefore, a solving strategy must be applied in several steps.

The problem is formulated as a Mixed Integer Linear Programming and solved by

dynamic programming. The final value of stored water, i.e. the total value of stored
water in all reservoirs, is used as target at the end of the planning period. The problem
is investigated from an economic perspective of a power producer with sole ownership
of the river in the context of a competitive market using time steps of one hour.

4.1.1 Model limitations

The model has some limitations. It is assumed that the power producer knows when
maintenance should be performed and on which turbine. The model is aimed to
compute the loss in income if the planned maintenance is executed. Moreover, the
model is constructed to be utilized by a single owner of the river system. Nonlinear
dependence of the power production in turbines is modeled as concave piecewise linear
functions. The value of stored water assumes discrete values and delay time of water
is not modeled.

27
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

4.1.2 Indices

We introduce the following indices,

i = unit,
j = reservoir,
k = segment,
t = time.

4.1.3 Parameters

We introduce the following parameters,



−1, if reservoir j releases water directly to unit i,
Auniti,j =

+1, if unit i releases water to reservoir j,


−1, if reservoir j spills water to j ′ ,
Aspillj ′ ,j =

+1, if reservoir j ′ spills water to j,

β = (value of stored water at the end of the planning period)

÷ (value of stored water at the start of the planning period),
costi = Start up cost in unit i,
Pmaxi = Maximal power output in unit i,
Pmini = Minimum power output in unit i,
pricet = Electricity price hour t,
Qmaxi,k = Allowed maximum
discharge level in unit i and
segment k,
Qmini,k = Allowed minimum
discharge level in unit i and
segment k,
ϱi,k = Slopes in efficiency curves
unit i and segment k,
θi = Constant term in the efficiency curve in unit i,

28
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

Vmaxj = Maximum volume in reservoir j,

Vminj = Minimum volume in reservoir j,
V0j = Start volume in reservoir j,
wj,t = Local inflow to reservoir j hour t.

4.1.4 Variables

We introduce the following variables,

pi,t = Power output in unit i hour t,

qi,t,k = Discharge in unit i segment k hour t,
sj,t = Spillage in reservoir j hour t,
ui,t = Binary variable, unit i is on or off hour t,
vj,t = Volume in reservoir j hour t,
yi,t = Binary variable, unit i is switched on or off hour t.

4.1.5 Dynamic programming

The full maintenance scheduling problem is formulated as a MILP problem and solved
with a dynamic programming strategy.

Dynamic programming is a method developed in the 1950’s by Richard Bellman

to solve large optimization problems that can be formulated in terms of smaller
subproblems in the same form.

An optimization problem that can be formulated as

∑
N −1
min ϕ(xN ) + f0 (k, xk , uk )
uk
k=0

subject to xk+1 = f (k, xk , uk ),

x0 given,
uk ∈ U (k, xk ),

for k = 0, . . . , N − 1 and where

29
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

• uk are optimization variables, also called control variables,

• xk are state variables controlled by the dynamics xk+1 = f (k, xk , uk ),

• ϕ and f are functions, and

• U (k, x) are feasible sets for the control variables uk .

can then be solved by Bellman’s equation

[ ( )]
J(n, x) = min f0 (n, x, u) + J n + 1, f (n, x, u)
u∈U (n,x)

working backwards from n = N , when J(N, x) = ϕ(x), down to n = 0 arriving at

the solution of the full problem. For further discussion of dynamic programming,
see [60].

4.1.6 Value of stored water

An important aspect of mid term planning is the value of stored water. The value of
stored water represents the total energy of the water stored in the reservoirs of the
river system. Water in a reservoir upstream has a higher energy value than water in
a downstream reservoir since the water can be used in several turbines downstream.
The value of stored water at time t can be written as

∑
αt = bj vj,t
j

where the coefficients bj are constants that are determined by how the reservoirs are
connected. The variables vj,t are the volumes of the reservoirs at time t.

Since the reservoirs have a known volume at the start of the planning period the value
of stored water is known in the beginning of the planning period.

An example of this can be illustrated with a three hydropower reservoir system as

in Figure 4.1.1. If the units (turbines) connected to the reservoirs are equal in
performance then the value of stored water in this case is αt = 3v1,t + 2v2,t + v3,t . Water
in reservoir 1 can flow through all downstream units of the three reservoirs, water in
reservoir 2 can flow through the units of the second and third reservoirs, while water
in reservoir 3 can only flow through the units connected to this reservoir.

30
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

v1,t

v2,t

v3,t

Figure 4.1.1: Example of a three reservoir hydropower system.

4.1.7 Efficiency curves for the power production

The power production has highly nonlinear dependence between discharge and
electricity production. In this chapter the power production in each unit is modeled as a
piecewise linear concave function with two segments as illustrated in Figure 4.1.2.

This dependence is formulated as

∑
pi,t = ϱi,k qi,t,k + θi ui,t (4.1)
k

where ϱi,k is the slope of each linear part and θi is a constant term. The discharge is
∑
written as a sum of parts k qi,t,k , where each term is the amount of discharge in the
corresponding segment of the piecewise linear curve.

Q
Figure 4.1.2: Example of a piecewise linear function of the power production. The
horizontal axis shows the discharge through the turbine and the vertical axis shows the
power output. Note that in this example there is a small amount of discharge through
the turbine before any power is produced.

31
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

4.1.8 MILP formulation of the full problem

The full problem is formulated as a MILP problem. The formulation is

max The income of sold electricity

− Start up costs for the turbines

subject to Hydrological constraints

Final value of stored water (Energy)
Power generation
Unit commitment
Variable limits
Maintenance periods

The objective function is expressed as

∑( )
max pricet pi,t − costi yi,t . (4.2)
i,t

The hydrological balance in the river system is formulated as the constraint

∑
vj,t+1 = vj,t + wj,t+1 + Auniti,j qi,t+1,k
i,k
∑
+ Aspillȷ′ ,j sȷ′ ,t+1 . (4.3)
ȷ′

At the start of the planning period the reservoirs are assumed to contain a known
volume of water,
vj,0 = V0j (4.4)

and at the end of the planning period the total value of the reservoir water should have
a predetermined percentage β of its value at the start of the planning period,

∑ ∑
bj vj,T = β bj vj,0 . (4.5)
j j

The reservoir levels are only allowed to vary between prescribed fixed values,

Vminj ≤ vj,t ≤ Vmaxj . (4.6)

32
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

The discharge from a unit i in segment k is limited by Qmini,k and Qmaxi,k unless the
unit is switched off,
Qmini,k ui,t ≤ qi,t,k ≤ Qmaxi,k ui,t . (4.7)

Note that Qmini,k = 0 if k > 1 since the discharge in segment k is equal to 0 as long as
segment k −1 is not fully utilised. As already discussed the dependence between power
production and discharge is formulated as

∑
pi,t = ϱi,k qi,t,k + θi ui,t . (4.8)
k

Limits for minimal and maximal power production are formulated as,

Pmini ui,t ≤ pi,t ≤ Pmaxi ui,t . (4.9)

If unit i is on during hour t then ui,t = 1 and if it is off then ui,t = 0. If unit i is switched
on hour t, i.e. it is off hour t − 1 and on hour t, then yi,t = 1, otherwise yi,t = 0. The
connection between ui,t and yi,t is

ui,t+1 − ui,t ≤ yi,t+1 . (4.10)

The maintenance of a turbine i during a time period [t1 , t2 ] is specified by introducing

the constraints
ui,t = 0 for t1 ≤ t ≤ t2 . (4.11)

4.1.9 Subproblem and relaxation

The full problem has been formulated in the previous subsection as an optimization
problem. This MILP problem cannot be solved for any longer periods of time than just
a few hours for a large system (such as for example in the test of a full river system
described in section 4.3). As the number of binary variables increase, it leads to an
exponential increase in time to solve the problem. To counter act this phenomena
a relaxation of the problem is performed before the problem is solved by dynamic
programming.

To start, the volumes of the reservoirs can be plotted over time as shown in Figure 4.1.3.
The planning period [0, T ] is divided into n intervals [0, ∆t], [∆t, 2∆t], …, [(n − 1)∆t, T ],
where ∆t = T /n. Now, the volumes of the reservoirs are studied at the end points

33
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

v1

v2
v3

t=0 t=1 t=∆t t=2∆t

Figure 4.1.3: Example of a three reservoir system with water volume in each reservoir
varying over time.

of these intervals, t = 0, t = ∆t, t = 2∆t and so on, as done in Figure 4.1.4. In the
Figure 4.1.4 two value of stored water scale bars are shown for each of these times.
The left scale bar at time t = ∆t corresponds to the time t = ∆t− , i.e. the instant
before t = ∆t, and the right corresponds to the time t = ∆t+ , i.e. the instant after
t = ∆t. The values of stored water are obviously equal at those two times. The full

v1,∆t− v1,∆t+
v1,0

v2,0 v2,∆t− v2,∆t+

v3,0 v3,∆t− v3,∆t+

t=0 t=∆t t=2∆t

Figure 4.1.4: Example of a three reservoir system with water volumes plotted at the
times 0, ∆t, 2∆t and so on.

problem is now relaxed by allowing the volumes of the reservoirs instantly before and
after each of the times t = ∆t, t = 2∆t, …, t = (n − 1)∆t to be unequal as shown
in Figure 4.1.5. However, the values of stored water αt of the system are required
to be equal at the times instantly before and after. Also the variables ui,t and yi,t are
decoupled at the times t = ∆t, t = 2∆t, …, t = (n − 1)∆t. This relaxation splits the full
problem into subproblems over the time intervals [0, ∆t], [∆t, 2∆t], …, [(n − 1)∆t, T ].
Each subproblem is connected to the next by the fact that the value of stored water at
the end of the subproblem is equal to the value of stored water at the start of the next.
This is illustrated with an example in Figure 4.1.6.

34
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

α∆t− = α∆t+ α2∆t− = α2∆t+

v1,∆t−
v1,0 v1,∆t+

v2,∆t− v2,∆t+
v2,0
v3,∆t+
v3,0 v3,∆t−

t=0 t=∆t t=2∆t

Figure 4.1.5: Example of a three reservoir system where the water volumes are allowed
to be different instantly before and after the times ∆t, 2∆t, …The values of stored water
instantly before and after these times are equal.

α∆t α2∆t α3∆t

Subproblem Subproblem Subproblem

0 ∆t ∆t 2∆t 2∆t 3∆t

Figure 4.1.6: The full problem is split into subproblems connected by the values of
stored water.

4.1.10 Solution by dynamic programming

Introduce the following functions

∆F (k, α, α′ ) = The optimal objective function

value for the subproblem
from t = (k − 1)∆t to t = k∆t
with α(k−1)∆t = α and αk∆t = α′ ,
F(k, α, α′ ) = The optimal objective function
value for the optimization
problem from t = (k − 1)∆t
to t = T with α(k−1)∆t = α
and αT = α′ .

35
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

According to Bellman’s equation the function F satisfies the recurrence equation

[ ]
F (k, α, α′ ) = max
′′
∆F (k, α, α′′ ) + F (k + 1, α′′ , α) (4.12)
α

for k = 1, . . . , n − 2. This equation is illustrated in Figure 4.1.7 and can be expressed

as: The optimal value of going from value of stored water α at time t = (k − 1)∆t to
value of stored water α′ at time t = T is equal to the optimal value of going one time
step to value of stored water α′′ added to the optimal value of going the rest of the time
steps from value of stored water α′′ to the value of stored water α′ at time t = T when
α′′ is chosen to maximize the expression.

00 ) F (k + 1,
,α α00 α 00, α 0)
∆F (k , α
α0
F (k, α, α )
0
α

t=(k − 1)∆t t=k∆t ......... t=T

Figure 4.1.7: Illustration of Bellman’s equation.

The recurrence relation can be used to calculate the values of F by working

backwards,

F (n − 1, α, α′ ) = ∆F (n − 1, α, α′ ),
[
F (n − 2, α, α′ ) = max
′′
∆F (n − 2, α, α′′ )
α
]
+ F (n − 1, α′′ , α′ ) , (4.13)
..
.
[ ]
F (1, α, α′ ) = max
′′
∆F (1, α, α′′ ) + F (2, α′′ , α′ ) .
α

To make it feasible to solve this recurrence relations it is also assumed that all values
of stored water assume a discrete set of values and values of ∆F are calculated for all
transitions between these levels as illustrated in figure 4.1.8. The value of F (1, α0 , αT ),
where α0 is the value of stored water at the start of the planning period t = 0 and αT
is the value of stored water at the end of the planning period, is then the optimal value

36
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

αT

α0

t=0 t=∆t t=2∆t ......... t=T −∆t t=T

Figure 4.1.8: The value of ∆F is calculated for all transitions between the discrete set
of value of stored water levels.

of the relaxed full planning problem. This value is calculated using the recurrence
relations (4.13) and can be seen as finding the optimal path from the start value of
stored water α0 to the final value of stored water αT in the graph of transitions as
illustrated in Figure 4.1.9.

αT

α0

t=0 t=∆t t=2∆t ......... t=T −∆t t=T

Figure 4.1.9: Illustration of solution to the full problem by determining values of stored
water at the intermediate times t = ∆t, t = 2∆t and so on.

4.1.11 A solution to the full problem

The solution derived for the relaxed problem determines the values of stored water at
the intermediate times ∆t, 2∆t, …, (n − 1)∆t. These values of stored water are used
in order to find a solution close to the optimal solution to the non­relaxed problem by
solving the full MILP problem stepwise:

• From t = 0 to t = ∆t with the constraint that the value of stored water is α∆t
at t = ∆t.

• From t = ∆t to t = 2∆t with the constraint that the value of stored water is α2∆t
at t = 2∆t.

37
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

• …

• From t = (n − 1)∆t to t = T with the constraint that the value of stored water
is αT at t = T .

The time interval length ∆t is short enough to make each of the subproblem in the list
above computationally tractable.

4.2 Validation of the method

The objective of the empirical studies is to reduce the system so that the dynamic
programming approach can be compared to a direct solution of the MILP­problem.
This is done by studying two subsets, A and B, of a full river system. The planning
problem is solved twice per case study, first using a MILP­solver in Gurobi and then
with the dynamic programming method explained in this chapter.

4.2.1 Two solution methods: MILP­solver and dynamic

programming

The two solution methods, using a MILP­solver and using dynamic programming, are
studied in this section. This is done in a case study that uses the river A subset of
the full river system. This subset of the river consists of 3 reservoirs and 4 turbines.
The objective of this case study is to evaluate how the dynamic programming method
performs by comparing the results with the optimal solution obtained by solving the
MILP problem.

In table 4.2.1 results from the case study is presented when the planning problem is
solved in Gurobi for one day with the two methods. The number of intermediate steps
in the dynamic programming were 4 and each step was 6 hours. The optimal values
for different target of the value of stored water at the end of the planning period is
presented as well as the running time.

When using the dynamic programming method the optimal values are slightly smaller
compared to solve the problem directly. This originates from the fact that in the
dynamic programming method first a relaxed problem is solved and then a constrained
problem is finally solved. However, since the difference is fairly small this shows that
this relaxation does not affect the original problem in any major way.

38
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

What is perhaps not readable from table 4.2.1 is that the DP method is sensitive to
how fine grained the value of stored water levels are chosen. A too coarse choice of
levels will lead to a solution that deviates from the true optimal solution and an optimal
values that is significantly worse. On the other hand, a too fine grained set of value of
stored water levels will increase the runtime. In this case study 10 levels are used. This
means that 10 × 10 = 100 subproblems need to be solved in each step of the dynamic
programming. By increasing this to 20 levels a total of 20×20 = 400 subproblems need
to be solved for each DP step. To find a suitable number of levels an iterative approach
needs to be considered.

To overcome this the DP method is implemented as one full optimization problem in

Gurobi where all the subproblems are grouped together and the intermediate values
of stored water joining the subproblems are introduced as additional optimization
variables. This means that the intermediate values of stored water can attain real
values and not just values from discrete set of values of stored water and are determined
by solving the optimization problem. This yields a fully satisfactory solution. The
running time of the DP method compared to solving the problem directly is roughly
equal. This approach is only used for this case study in contrast to the full river system
where this is not computationally tractable.

Table 4.2.1: Optimal values from empirical study subset of a river A. The first column
shows percentage of the final value of stored water compared to the initial value of
stored water. The second and fourth columns are the profits during the planning
period.

MILP problem Dynamic programming

Final v.w. Opt. value [kEUR] Time [s] Opt. value [kEUR] Time [s]
99.96% 276.801 2.0 271.533 1.7
99.97% 276.791 1.7 271.526 1.7
99.98% 276.780 2.2 271.518 2.2
99.99% 276.769 2.5 271.511 2.1
100.00% 276.759 2.3 271.504 1.7
100.01% 276.748 2.0 271.497 2.1
100.02% 276.738 2.6 271.489 1.9
100.03% 276.727 2.3 271.482 2.1
100.04% 276.716 2.2 271.475 1.8
100.05% 276.706 1.7 271.468 2.1
100.06% 276.695 1.7 271.460 1.5

In figure 4.2.1 the different values of stored water that are assumed during the planning
period are shown both when solving the MILP problem directly and by using dynamic

39
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

programming. It can be seen that the values of stored water are similar for the two
methods.
∆ [%]
dynamic programming
0.4 MILP problem

−0.4

Time [h]
4 8 12 16 20 24

Figure 4.2.1: The values of stored water of the system during the planning period when
the problem is solved with and without dynamic programming, and no maintenance.
The black curve is with dynamic programming and the blue curve is the MILP­problem.

A conclusion from this is that the dynamic programming method performs well
compared to solving the MILP problem directly since the optimal values only differ
1–2%. The intermediate values of stored water are comparable in the two solution
methods.

4.2.2 Case study with maintenance

The case study is repeated when maintenance of turbine 1 is required and it is forced
out of commission during 5 hours. In table 4.2.2 the corresponding results are
reported. The objective in this part of the case study is to evaluate how the introduction
of maintenance affects the dynamic programming method compared to the optimal
solution.

In figure 4.2.2 the discharge from a turbine with and without maintenance is plotted
when the dynamic programming method is applied. It can be seen from the figure that
when the maintenance period (hour 4 to 7) has stopped the two curves follow almost
the same pattern.

In figure 4.2.3 the discharge from a turbine when the dynamic programming method
is applied and the electricity prices is plotted. As expected the discharge is high when
the electricity prices are high during the planning period.

Both graphs in figure 4.2.2 and 4.2.3 show patterns which are expected compared to the

40
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

Table 4.2.2: Optimal values from empirical study subset of a river A when a turbine
is under maintenance. The first column shows percentage of the final value of stored
water compared to the initial value of stored water. The second and fourth columns
are the profits during the planning period.

MILP problem Dynamic programming

Final v.w Opt. value [kEUR] Time [s] Opt. value [kEUR] Time [s]
99.96% 272.001 1.4 266.774 0.8
99.97% 271.994 1.8 266.766 1.3
99.98% 271.987 1.4 266.759 1.0
99.99% 271.980 1.2 266.752 1.0
100.00% 271.972 1.4 266.744 1.5
100.01% 271.965 1.4 266.737 1.0
100.02% 271.958 1.4 266.730 0.9
100.03% 271.950 1.3 266.722 1.4
100.04% 271.943 2.3 266.715 1.1
100.05% 271.936 1.4 266.708 1.4
100.06% 271.929 1.7 266.700 0.9

Discharge [m3 /s]

maintenance
no maintenance

50
Time [h]
4 8 12 16 20 24

Figure 4.2.2: The discharge from a turbine in case study river A with and without
maintenance. Blue curve shows no maintenance and black curve with maintenance.

discharge and electricity prices during the planning period. When the electricity prices
are higher the values of stored water are decreasing compared to when the electricity
prices are lower.

Figure 4.2.4 corresponds to figure 4.2.1 but with maintenance. The figure shows the
values of stored water during the planning period.

A conclusion from introducing maintenance is that the dynamic programming method

still performs well in comparison with the optimal strategy.

41
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

Price [EUR/MWh]

50

Time [h]
4 8 12 16 20 24

Figure 4.2.3: The discharge from a turbine in case study river A with maintenance and
the electricity prices. Blue curve shows electricity prices and black curve discharge with
maintenance.

∆ [%]
dynamic programming
0.4 MILP problem

−0.4

Time [h]
4 8 12 16 20 24

Figure 4.2.4: The values of stored water of the system during the planning period when
the problem is solved with and without dynamic programming and maintenance. The
black curve is with dynamic programming and the blue curve is the MILP­problem.

4.2.3 A second case study

In this section a second case study is performed and a different subset of the full river
system is studied. This subset, river B, consists of three cascaded reservoirs and three
units. The objective this time is to make sure that the conclusions from studying river A
can be applied to a similar but different case.

In table 4.2.3 the optimal values from this case study is shown. The data that has been
used in this case study is the same as for the first case study. The planning period is
one day. The planning problem is solved as a MILP problem and with the dynamic
programming method. In the dynamic programming method a time step of 6 hours is
used.

The results in table 4.2.3 can be compared to the results obtained in table 4.2.4 where

42
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

Table 4.2.3: Optimal values from empirical study subset of a river B. No maintenance
is scheduled. The first column shows percentage of the final value of stored water
compared to the initial value of stored water. The second and fourth columns are the
profits during the planning period.

MILP problem Dynamic programming

Final v.w Opt. value [kEUR] Time [s] Opt. value [kEUR] Time [s]
99.96% 75.670 17.7 75.448 2.2
99.97% 75.668 16.9 75.445 2.3
99.98% 75.665 13.6 75.441 2.3
99.99% 75.662 18.2 75.437 2.3
100.00% 75.659 18.3 75.433 2.0
100.01% 75.656 25.1 75.429 1.7
100.02% 75.653 20.8 75.426 1.8
100.03% 75.651 18.9 75.422 2.1
100.04% 75.648 22.4 75.418 2.2
100.05% 75.645 21.3 75.414 2.2
100.06% 75.642 17.7 75.411 2.3

the problem is solved with maintenance scheduled for one unit during six hours. A
comparison shows that the optimal values are somewhat smaller in the DP case which
is explained by the fact that the dynamic programming approach can be seen as solving
a more restricted problem.

Table 4.2.4: Optimal values from empirical study subset of a river B. Maintenance of
turbine 2 is scheduled for hours 12 to 17. The first column shows percentage of the
final value of stored water compared to the initial value of stored water. The second
and fourth columns are the profits during the planning period.

MILP problem Dynamic programming

Final v.w Opt. value [kEUR] Time [s] Opt. value [kEUR] Time [s]
99.96% 74.495 2.6 72.911 3.6
99.97% 74.493 3.9 72.908 3.4
99.98% 74.490 3.2 72.905 3.2
99.99% 74.487 3.2 72.902 3.8
100.00% 74.484 3.8 72.899 4.3
100.01% 74.482 3.2 72.897 3.8
100.02% 74.479 3.7 72.894 3.8
100.03% 74.476 3.4 72.891 4.0
100.04% 74.473 3.6 72.888 4.0
100.05% 74.471 3.6 72.885 3.8
100.06% 74.468 3.5 72.882 3.9

In figure 4.2.5 the discharge from one turbine is plotted for the cases when
maintenance is not required and when it is required and the dynamic programming
method is applied. The maintenance period is between hour 12 to 17. There is a

43
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

significant difference in discharge between the two cases.

Discharge [m3 /s]
maintenance
no maintenance

50

Time [h]
4 8 12 16 20 24

Figure 4.2.5: The discharge from a turbine in case study river B with and without
maintenance. Blue curve shows no maintenance and black curve with maintenance.

In figure 4.2.6 the electricity price is plotted together with the discharge when the
dynamic programming method is applied. The discharge is high during peak electricity
prices.
Price [EUR/MWh]

50

Time [h]
4 8 12 16 20 24

Figure 4.2.6: The discharge from a turbine in case study river B with maintenance and
the electricity prices. Blue curve shows electricity prices and black curve discharge with
maintenance.

To conclude the comparison, the values of stored water of the system are plotted in
figure 4.2.7 and 4.2.8. As can be seen the difference between solving the problem with
and without dynamic programming is rather small in the no maintenance case. When
maintenance is performed on the second turbine there is no discharge during hours 12­
17 and this prevents the values of stored water to decrease as rapidly as in the optimal
solution when there is no maintenance.

The conclusion that can be drawn from this second case study is that the dynamic
programming method still performed in line with what was obtained from the first
case study.

44
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

∆ [%]
dynamic programming
0.4 MILP problem

−0.4

Time [h]
4 8 12 16 20 24

Figure 4.2.7: The values of stored water of the system during the planning period when
the problem is solved with and without dynamic programming, and no maintenance.
The black curve is with dynamic programming and the blue curve is without dynamic
programming.

∆ [%]
dynamic programming
0.4 MILP problem

−0.4

Time [h]
4 8 12 16 20 24

Figure 4.2.8: The values of stored water of the system during the planning period when
the problem is solved with and without dynamic programming, and maintenance.
The black curve is with dynamic programming and the blue curve is without dynamic
programming.

4.3 Empirical study full river

The model has been implemented in Python and tested on a Swedish cascaded river
system consisting of 15 hydropower plants, 34 turbines and 16 reservoirs. The objective
of the study is to show that the model works for a full river of more than 10 power
plants. The river consists of two branches with power plants and reservoirs which
merge into the same river downstream. The reservoirs located in the north part of the
river system are larger and can store water for long time periods while the reservoirs
located downstream are smaller and can store water for short time periods of a few
hours. The planning period is 11 days and based on historical data from public and
nonpublic sources. The number of maintenance are 8 and ranging from 2 to 160 hours,

45
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

see Figure 4.3.1. In the case study the index sets of the indices i, j, k and t are I =
{1, 2, . . . , 34}, J = {1, 2 . . . , 16}, K = {1, 2} and T = {1, 2, . . . , 264}, respectively.

Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 Day 8 Day 9 Day 10 Day 11
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

Figure 4.3.1: Diagram showing the planned maintenance (red color) of the 34
generators (vertical axis) during the planing period (horizontal axis).

The electricity prices during the planning period are spot prices from Nord Pool and
shown in Figure 4.3.2.

Price [EUR/MWh]

50

Time [h]
24 48 72 96 120 144 168 192 216 240 264

Figure 4.3.2: Electricity spot price during the planning period in the case study.

The local inflows are nonpublic data but are mostly constant for each reservoir during
the planning period.

46
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

4.4 Results full river

A table of optimal values with and without maintenance for different values of the
final value of stored water are shown in Table 4.4.1. An immediate observation is that
the optimal value is lower in the case of maintenance compared to the case without
maintenance. Typically the decrease is around 0.3%. The optimal value is lower when
maintenance is performed since water cannot be fully utilized at peak electricity prices.
Another aspect is that a larger final value of stored water target also means a lower
optimal value since water is then stored in the reservoirs that can be utilized in the
future.

The final values of stored water are chosen to deviate very little from the start
value of stored water since the planning period is during a part of the year
(November/December) when the water inflow to the reservoirs are low.

Table 4.4.1: Optimal values from empirical study full river.

Optimal value
Final v.w. Without m. [MEUR] With m. [MEUR]
99.996% 16.528756 16.248609
99.997% 16.528752 16.248604
99.998% 16.528747 16.248600
99.999% 16.528742 16.248595
100.000% 16.528737 16.248590
100.001% 16.528732 16.248585
100.002% 16.528728 16.248580
100.003% 16.528723 16.248575
100.004% 16.528718 16.248571
100.005% 16.528713 16.248565

The discharge in one particular turbine is plotted in Figure 4.4.1 under two situations.
In the first situation (plotted in blue) there is no scheduled maintenance and in the
second situation (plotted in black) the turbine is out of commission during the first
63 hours.

By combining Figure 4.3.2 and Figure 4.4.1 as done in Figure 4.4.2 it is observed that
the times the turbine is switched on corresponds strongly to the times the electricity
price is high.

The computational time of running the model in Python with and without
maintenance, respectively, is 6­7 hours. The solver used in Python is Gurobi. The
computer used an Intel Core i5­10400T CPU with a clock frequency of 2.00 GHz and

47
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

Discharge [m3 /s]

maintenance
no maintenance
218.77

103.36

Time [h]
24 48 72 96 120 144 168 192 216 240 264

Figure 4.4.1: The discharge from one turbine during the planning period with and
without maintenance scheduled.

a RAM­memory of 16.0 GB.

Price [EUR/MWh]

50

Time [h]
24 48 72 96 120 144 168 192 216 240 264

Figure 4.4.2: The discharge (black curve) from one turbine during maintenance and
the electricity price (blue curve).

In table 4.4.2 results are presented from a study were the final value of stored water
is chosen to differ up to two percent from the start value of stored water. The purpose
with this study is to show that the method works for larger differences in start and final
value of stored water than what has been presented earlier. This time the intermediate
values of stored water were chosen to be five. This change decreased the computational
time but the optimal values are different from if ten values would have been chosen.
It can be seen in 4.4.2 that the optimal values for 100% of the value of stored water
are higher compared to the optimal values at that level in 4.4.1. The reason to this is
that there is a larger span of possible values of stored water in 4.4.2. However, this
illustrates that the method also works for larger variations in start and final value of
stored water.

48
CHAPTER 4. PREVENTIVE MAINTENANCE SCHEDULING

Table 4.4.2: Optimal values from empirical study full river 2% difference in start and
final value of stored water.

Optimal value
Final v.w Without m. [MEUR] With m. [MEUR]
98% 16.750615 16.463418
98.5% 16.748299 16.461100
99.0% 16.743493 16.456244
99.5% 16.741017 16.453755
100.0% 16.738329 16.451115

4.5 Conclusion
A model for fixed preventive maintenance scheduling has been developed in this
chapter, which can be used by the industry. The case studies, river A and B, show
that the model works well. The computational time for the case studies is similar
when using dynamic programming compared to solving the MILP problem except in
one case. This case is when the MILP problem is solved for river B and there is no
maintenance. A possible explanation to the longer computational time compared to
when the problem is solved for river B with maintenance can be seen in figure 4.2.5.
From the figure it is clear that the optimal solutions for the two cases differ significantly
and this could explain why it might take different time to reach the solution even
though the only difference between the cases is the applied maintenance. However,
the solution strategy gives a solution close to the optimal one. This is established
by comparing the value of the objective function and values of stored water for the
obtained solution with and without using dynamic programming.

49
Chapter 5

Stochastic model

5.1 Problem formulation

In this part a stochastic model is presented and compared to the deterministic model
discussed previously. The aim is to investigate if and to what extent a stochastic
model for the electricity price performs better than the deterministic model. This is
determined by comparing the outcomes of the stochastic model and the deterministic
model by calculating the expected value of perfect information (EVPI) and the value of
stochastic solution (VSS). The data that has been used is assumed to be representative
of an average scenario for a year. The different scenarios are assumed to be equally
probable.

In the stochastic model the electricity price is viewed as a stochastic variable. Scenarios
for the electricity price are used to investigate a comparison of the stochastic model
with a deterministic model. These scenarios are not necessarily realistic real scenarios
but test scenarios. The prices from the last eight years (2013–2020) are used as the
foundation to create scenarios in the following manner:

1. Let priceR
y,t be the electricity price year 2012 + y hour t.

2. First, for a given year (fixed y), the time series {priceR
y,t } is normalized by defining

y,t = (pricey,t − m)/s, where m is the average of the time

a new series priceN R

series {priceR
y,t } and s is the standard deviation.

3. Then a scenario of the electricity price is defined as priceω,t = σ ′ priceN

y,t +z, where

z is the mean value and σ ′ is the standard deviation of the electricity price. The

50
CHAPTER 5. STOCHASTIC MODEL

mean value z is chosen as the outcome of a normal distributed random variable

with mean equal to the average value of the time series priceR
y,t and variance 20.

The standard deviation σ ′ will be allowed to vary to simulate different volatility

of the electricity price.

In Figure 5.1.1 the electricity price time series priceR

y,t are plotted. In Figure 5.1.2

the normalized electricity price time series priceN

y,t are plotted. In Figure 5.1.3 the

electricity price times series priceω,t used in the stochastic model are plotted.

Price [EUR/MWh]

50

Time [h]
24 48 72 96 120 144 168 192 216 240 264
Figure 5.1.1: Electricity price during the eleven days following the first Monday of the
years from 2013 to 2020.

Normalized price

−6
Time [h]
24 48 72 96 120 144 168 192 216 240 264
Figure 5.1.2: Electricity price time series normalized to have mean 0 and standard
deviation 1.

Another method is also used to construct the electricity price scenarios. These
scenarios are priceR
y,t + cy where suitable constants cy are added to the historical
electricity price to give the scenario a desirable mean value. The resulting scenarios are

51
CHAPTER 5. STOCHASTIC MODEL

Price [EUR/MWh]

50

Time [h]
24 48 72 96 120 144 168 192 216 240 264
Figure 5.1.3: Electricity price time series in the stochastic model (σ ′ = 40).

plotted in Figure 5.1.4. The difference from earlier constructed price scenarios is that
the variance of the scenario is the same as in the corresponding historical year.

Price [EUR/MWh]

50

Time [h]
24 48 72 96 120 144 168 192 216 240 264
Figure 5.1.4: Electricity price time series in the stochastic model, variance is
unchanged.

If we were to plan for each scenario then we would solve eight independent
deterministic optimization problems as described in section 4.1.8, one for each
scenario. In order to help to formulate the stochastic model these independent
optimization problems can be combined into one large joint optimization problem with
objective function

∑ (∑ )
max πω priceω,t pi,t,ω − costi yi,t,ω . (5.1)
ω i,t

52
CHAPTER 5. STOCHASTIC MODEL

where the variables are made scenario dependent

pi,t,ω , qi,t,k,ω , sj,t,ω , vj,t,ω , ui,t,ω and yi,t,ω .

The parameter πω is introduced as the weight for the scenarios and the index ω is used to
keep track of the scenarios. The constraints and variable limits from the deterministic
model are kept.

This can be concluded from the fact that in (5.1) the sum over ω is made up of terms
that are maximized independent of each other.

The stochastic model is a recourse problem where decisions are made in two stages.
The first stage decision is made at the start of the whole planning period when one
scheduling is done taking all eight scenarios into account. This planning is used during
a first time period 0 ≤ t ≤ T . At time t = T a schedule is made for the rest of the
planning period taking only one scenario into account, i.e. at time t = T it is revealed
which scenario occurs.

The stochastic planning model uses the large joint optimization problem with objective
function (5.1) by adding the following constraints

Stage 1 constraints (for 0 ≤ t ≤ T )

pi,t,ω are equal for all ω,
qi,t,k,ω are equal for all ω,
sj,t,ω are equal for all ω, (5.2)
vj,t,ω are equal for all ω,
ui,t,ω are equal for all ω,
yi,t,ω are equal for all ω.

This makes the optimization variables equal during stage 1 irrespective of the scenario.
(During stage 2 they can have different values depending on the scenario.) The
stochastic model is solved with a similar dynamic programming approach used for the
deterministic model and will be described now.

During stage 2, from time t = T to the end of the planning period t = Tend , the
scenarios are treated separately in the dynamic programming. If the following notation

53
CHAPTER 5. STOCHASTIC MODEL

is introduced

ω
Ft→T end
(α, αend ) = The optimal objective function value for the
optimization problem from time t to time Tend in
scenario ω with start value of stored water α and final
value of stored water αend ,
ω
∆Ft→t+∆t (α, α′ ) = The optimal objective function value for the
subproblem from time t to time t + ∆t in scenario ω
with start value of stored water α and end value of
stored water α′ .

then Bellman’s equation can be formulated as

[ ]
ω
Ft→T end
(α, αend ) = max
′
ω
∆Ft→t+∆t (α, α′ ) + Ft+∆t→T
ω
end
(α′ , αend ) (5.3)
α

and this recurrence relation can be used to calculate the optimal objective function
value for stage 2, i.e. from time T and onward for each scenario ω.

The dynamic programming steps in stage 1 starts by introducing the following quantity

∑
FT →Tend (α, αend ) = πω FTω→Tend (α, αend ) (5.4)
ω

which is the optimal expected value of the objective function over all scenarios from
time T to time Tend as expressed in (5.1). The calculation done in (5.4) is illustrated
in figure 5.1.5. The optimal expected value of the objective function Ft→Tend (α, αend ) is
calculated by using the following version of Bellman’s equation

[ ]
Ft→Tend (α, αend ) = max
′
∆Ft→t+∆t (α, α′ ) + Ft+∆t→Tend (α′ , αend ) (5.5)
α

for t = T − 1 down to t = 0. The quantity ∆Ft→t+∆t is the optimal expected value of

the subproblem from time t to t + ∆t over all scenarios. It is calculated by solving the
large joint stochastic optimization problem over all scenarios as outlined earlier in this
text, i.e. solving one eight times larger optimization problem than for the deterministic
problem.

By relaxing the constraints (5.2) the wait­and­see solution is obtained. This

corresponds to the situation where we have complete information and planning can

54
CHAPTER 5. STOCHASTIC MODEL

αend
ω=1

T +∆t T end
α
αend
ω=2
T

T +∆t T end
Figure 5.1.5: The optimal expected value of the objective function FT →Tend (α, αend ) is
calculated by averaging the optimal values for each of the scenarios. The figure only
includes two scenarios, but the full model contains eight scenarios.

be performed optimally, i.e. we know beforehand which scenario will occur and
can plan accordingly. From this the expected value of perfect information can be
computed
EVPI = WS − RP,

where WS is the expectation value of the optimal values when planning for each
scenario and RP is the optimal value of the stochastic problem.

The value of the stochastic solution can be calculated as

VSS = RP − DS,

where DS is the optimal value of the stochastic problem when the first­stage variables
are fixed to values provided by a deterministic problem using the average value for
electricity price during the stage 1 and scenario dependent electricity prices during
stage 2.

5.2 Results of stochastic computations

The case study is performed on river A and the total time period is one week, where
the first stage is 42 hours. To achieve reasonable computation time and memory
consumption the time step is increased from 1 hour to 7 hours. Using the smaller time

55
CHAPTER 5. STOCHASTIC MODEL

step results in hitting the memory limit of the computer after about 12 hours and the
computation is aborted. The electricity price scenarios used are those outlines above
by using the price at the start of each 7 hour time step for the whole 7 hours.

The results from the computations are reported in table 5.2.1 and the values of EVPI
and VSS are
σ ′ = 40: EVPI = 11.726 − 11.545 = 0.181,
VSS = 11.545 − 11.539 = 0.006,
′
σ = 80: EVPI = 11.913 − 11.366 = 0.548,
VSS = 11.366 − 11.339 = 0.026,
Shifted: EVPI = 11.712 − 11.365 = 0.347,
VSS = 11.365 − 11.337 = 0.028,

where the three different electricity price scenarios, as described at the start of this
section, have been used. The standard deviation of the shifted scenarios ranges from
30 to 120 with a mean of 63.

Table 5.2.1: Optimal values for calculating EVPI and VSS.

Scenario [Mkr] (σ ′ = 40)

Method 1 2 3 4 5 6 7 8
Perfect information 11.7 10.4 10.7 11.7 12.4 13.0 11.0 13.0
Average scenario 11.6 10.3 10.4 11.7 12.3 12.8 10.9 12.3
Stochastic 11.6 10.3 10.3 11.7 12.3 12.8 10.9 12.5

Scenario [Mkr] (σ ′ = 80)

Method 1 2 3 4 5 6 7 8
Perfect information 11.3 9.9 11.0 11.7 12.9 13.6 11.2 13.7
Average scenario 10.9 9.7 9.9 11.6 12.5 12.7 10.9 12.4
Stochastic 10.7 9.7 9.9 11.6 12.3 12.9 11.0 12.7

Scenario [Mkr] (shifted)

Method 1 2 3 4 5 6 7 8
Perfect information 10.9 11.4 11.2 12.6 12.5 11.9 11.0 12.2
Average scenario 10.5 11.3 10.9 12.3 12.3 11.2 10.9 11.4
Stochastic 10.5 11.1 11.0 12.3 12.3 11.2 10.9 11.7

From the calculation of EVPI a conclusion is that with increasing standard

deviation (σ ′ ) the gap between the result from the stochastic method and solving
the problem with perfect information increases. The VSS calculations show that the
stochastic method performs better than a purely deterministic method where the

56
CHAPTER 5. STOCHASTIC MODEL

expected values of the electricity price is used. In this comparison the difference
increases with increasing standard deviation (σ ′ ). With the introduction of a larger
share of intermittent power sources in the power systems the standard deviation is
expected to increase and the result in this study indicates that stochastic models will
perform better for planning purposes. Comparing EVPI and VSS show that the gap
between the stochastic method and solving the problem with perfect information is at
least an order of magnitude greater than the difference between the stochastic method
and the deterministic method.

The runtime of the calculations of EVPI and VSS are 10.6 minutes for the scenarios
with σ ′ = 40, 70.7 minutes for the scenarios with σ ′ = 80 and 27.8 minutes for the
scenarios that are only shifted (and not rescaled). In these calculations the optimality
gap was set to 0.5% instead of 0.01% in order have reasonable runtimes and memory
consumption, yet sufficient accuracy, when solving the subproblems in the dynamic
programming method. With the smaller optimality gap (0.01%) the runtime could
be up to 10–12 hours and in many cases the computation was aborted due to lack
of memory. In cases when the computation was completed with a smaller optimality
gap the results were not different from the results reported with the larger optimality
gap in table 5.2.1. The runtimes also increase significantly with increasing time period
(1 week) or shorter time resolution (7 hours).

All results in this study are affected by the model setup and data used. A conclusion
from the study is that adequate electricity price forecasts are important but the
runtimes of the stochastic models tend to be long and difficult to use. A general
conclusion is that building a stochastic model is time consuming and requires large
computer resources. The increase in the optimal value if fairly small when using the
stochastic model compared to the deterministic model. However, this can translate
to substantial increase in revenue over time. Some form of stochastic planning
might therefore be useful in practise, but it is needed to always keep computational
tractability in mind by, for example, using long time steps and sufficiently large
optimality gap.

57
Chapter 6

Discussion

6.1 Summary

One goal of the project was to study the connection between short term, mid term and
long term planning of hydropower. This is achieved in a literature review that has been
carried out with the focus of weekly planning of hydropower. Some earlier reviews
are presented followed by a presentation of works describing long term planning,
mid term planning and short term planning. The long term planning models cover
both deterministic and stochastic models. It could be seen that the SDDP method
is commonly used. The mid term planning models cover decomposition, stochastic
linear and nonlinear programming, combined SDP and SDDP and stochastic dual
dynamic programming. The short term models pay attention to methods such as mixed
integer stochastic model, stochastic model handling non­linearities with time series
data, successive linear programming, non­linear Mixed Integer optimisation model,
unit commitment model and mixed integer linear programming model.

Another goal of the project was to develop efficient planning methods for a power
system with a large share of renewable power sources. This goal has been fulfilled since
a model has been developed for weekly planning of hydropower where a relaxation is
performed before the optimization problem is solved by dynamic programming. The
model has been applied in an empirical study of a Swedish cascaded river system
consisting of 15 hydropower plants, 34 turbines and 16 reservoirs. To test the
model two case studies are performed where two subsets of the full river are used.
The test models have been developed since the size of the problem makes it hard

58
CHAPTER 6. DISCUSSION

to solve with standard techniques. The full optimization problem is to maximise

the income of sold electricity minus start up costs for the turbines due to a set of
constraints. The constraints are final value of stored water, power generation, unit
commitment, variable limits and maintenance periods. The method works well in
terms of computational time and monetary value, which are similar when dynamic
programming is used compared to when the solver Gurobi is used in the case of the
smaller test systems, river A and river B, except in one case. The solutions found with
the dynamic programming method are close to the optimal solutions.

Furthermore a stochastic method is developed. The stochastic model is compared

to the deterministic model. The stochastic model is a two stage model where the
electricity price is a stochastic variable and data from eight years are used to create
scenarios for the electricity price. The objective function is the expected value of profit.
In general, the stochastic method performs better than the deterministic method and
this is more evident when the price variations increase. However, accurate electricity
price forecasts are important and the number of scenarios needs to be kept down
in order to avoid long computational runtimes. In practice, it is not clear which of
stochastic or deterministic planning models perform better.

To conclude, this project has shown that it is possible to solve a detailed hydropower
planning problem with hourly resolution and a planning horizon of 2­3 weeks. The
deterministic method can be used to solve large hydropower planning problems.

6.2 Future Work

Below some directions of future work are given.

• One suggestion is to combine the planning between different markets (such as the
spot market, intraday trading and real time trading) and integrate maintenance
scheduling, i.e. detailed models for mid term planning.

• It would be interesting to include stochastic inflows in the model. One way of

approaching the problem could be with the SDDP method to circumvent the
problem with exponential growth of the required resources to solve the problem.

• A possible extension of the maintenance scheduling study is to model the

maintenance as a decision variable, which gives the power producer a larger

59
CHAPTER 6. DISCUSSION

flexibility to schedule the maintenance.

• It would be interesting to include nonlinear dependence of the power production.

• As the electricity market is developing, it would be interesting to study shorter

time steps in the model, e.g. 15 minutes instead of 1 hour. This could be
formulated as a continuous time problem leading to a formulation as an optimal
control problem and applying different numerical methods to solve the resulting
differential equations.

• Delay time of water could be further studied.

• Further studies of electricity price forecasts and stochastic models which are
adapted to planning periods of several weeks are needed.

• Use other structures and methods for stochastic planning where decision and
information arrive at different times.

• Study approaches to circumvent the problem with limited computer resources in

order to be able to solve larger problems by, for example, using more powerful
computer resources.

60
Appendix A

Tables

Table A.0.1: This table summarizes the review papers.

Literature reviews
Ref. Title Characteristics
[14] Planning and Review that spans over many fields of
Scheduling under Uncertainty: A review science and the different techniques used
Across Multiple Sectors to model uncertainties.
[15] Operation of hydro power plants­a review Review of operation of reservoir based as
well as run of river hydropower plants.
[16] Review on hydropower plant models and Review of hydropower plant models and
control control.
[17] A Review of Optimization Algorithms Review of optimization algorithms for
in Solving Hydro Generation Scheduling different time frames.
Problems
[18] Handbook of Power Systems Review of the SDDP method for long term
hydropower planning.
[19] Long­and Review
medium­term operations planning and of stochastic dual dynamic programming
stochastic modelling in hydro­dominated applied to hydropower scheduling in the
power systems based on stochastic dual Nordic countries.
dynamic programming
[20] Optimal Operation of Multireservoir Review of long term planning methods
Systems: State­of­the­Art­Review other than SDDP.
[21] Optimization of Review of methods useful for long­term
the hydropower energy generation using planning models.
Meta­Heuristic approaches: A review
[22] The State­of­the­Art of the Short Term Review of short term planning works,
Hydro Power Planning with large Amount deterministic and stochastic.
of Wind Power in the System

61
APPENDIX A. TABLES

Table A.0.2: This table summarizes the papers on long­ and mid term planning.

Long term planning

Ref. Main characteristics Case study Conclusions
[23] Deterministic versus stochastic Brazil Deterministic and stochastic dynamic
dynamic programming for long programming approaches in long term
term hydropower scheduling. hydropower scheduling.
[24] SDP, Markovian model for inflow Brazil Reduced costs compared to standard
aggregated over time scales. Markovian model.
[25] Optimal Brazil SDDP method
stochastic operations scheduling used for optimal operating strategies in
of large hydro­electric systems. a multi­reservoir hydroelectric system.
[26] SDDP, large system, model for Norway SDDP is preferred to
inflow. aggregate­disaggregate models.
[27] SDDP, New IFS model is an accurate representation
IFS model for inflow compatible Zealand of the inflow in two aspects.
with SDDP.
[28] Hybrid Norway Linearization error is pronounced in
SDP/SDDP, inflow, energy price sales of energy and reserve capacity.
and reserve capacity stochastic.
[29] Formal Fictious Both models give comparable results.
optimization model compared to
aggregated heuristic model.
[30] SDP, multi­stage China The method
problem is decomposed into two­ avoids the curse of dimensionality with
stage optimization problems. good performance.
[31] Flexible robust optimization China Flexible robust optimization with
dispatch for hybrid wind/solar/ adjustable uncertainty budget dispatch
hydro/thermal power system. model built for a hybrid power system.
Mid term planning
[32] Stochastic MILP, day­ahead and Spain Higher expected
reserve markets considered. income when participating in both day­
ahead and reserve markets than just
day­ahead.
[33] Stochastic multi­stage quadratic Switzerland Trading with forwards increase profit.
model, pumped storage and
forwards used
[34] Mid term and long term power China Tested model
quantity contracts are works well for the Yunnan power spot
decomposed. market.
[35] Medium term power planning Norway Stochastic
with bilateral contracts. linear and non­linear programming is
used.
[36] A case study on medium­term Norway SDP and SDDP are combined.
hydropower scheduling with
sales of capacity.
[37] Nonconvex Medium­ Norway The method works to solve to optimal
Term Hydropower Scheduling by value but is time consuming.
Stochastic Dual Dynamic Integer
Programming
[38] Nonconvex Environmental Norway A particular nonconvex environmental
Constraints in Hydropower constraint was modelled.
Scheduling 62
APPENDIX A. TABLES

Table A.0.3: This table summarizes the papers on short term planning.

Short term planning

Ref. Main characteristics Case study Conclusions
location
[39] Mixed Norway The model
integer stochastic model, inflow performs better than a deterministic
and stochastic market price. practice­based industrial model.
[40] Stochastic MILP, day­ahead Norway Profit and risk negatively correlated.
price and inflow stochastic.
[41] NLDP, head and efficiency curves Spain Model considers many non­linearites
non­linear. and gives useful results.
[42] Stochastic non­linear model Fictious The solution is close to optimal.
solved with successive LP. Stochastic model performs better than
deterministic.
[43] Deterministic non­linear mixed Brazil The approach is feasible but needs to
integer be further studied.
optimization model solved with
Lagrangian relaxation.
[44] Mathematical model to simulate North Sea The high penetration of wind and solar
Day­ahead markets of region is likely to challenge the need for
large­scale multi­energy systems balancing in the system. Flexibility
with large amount of renewable will be very important towards 2050.
energy towards 2050.
[45] Stochastic LP model Sweden Coordinating hydro and wind power is
with integrated wind power and beneficial.
trading.
[46] MILP model with power Brazil The model is efficient and generates a
output and unit commitment as smooth generation schedule.
variables.
[47] Non­linear model for power Brazil The model is compact and efficient but
output and unit commitment. not convex.
[48] MILP, unit commitment model. Brazil An accurate planning result within
reasonable computational work.
[49] Stochastic MIP Norway Small additional revenue
commitment model with bids to to participate in balancing market and
day­ahead and balancing market. weak dependence on start/stop costs.
[50] MILP commitment model Fictious A flexible model that ensures reliability
with large amount of wind power and security.
solved with robust optimization.
[51] Overview on formulations and Norway Fewer works
optimization methods for summarize hydro scheduling works on
the unit­based short­term hydro single turbine models compared to
scheduling problem. aggregated models.
[52] Acceleration of United The formulation approximates
Benders decomposition for two­ Kingdom the three dimensional nonlinearity of
stage stochastic maintenance hydropower production.
scheduling.
[53] Stochastic two­stage NLP and Canada Using multi­scenarios is more robust
MILP, inflow stochastic than using median scenarios.
[61] The model is a combination of Nordic The model is well suited to valuate the
simulation and stochastic two­ Countries flexibility of a power system.
stage LP. Inflow and electricity 63
price are stochastic.
Bibliography

[1] UN. THE 17 GOALS. https://sdgs.un.org/goals. Accessed: 06052021.

[2] IEA. Data and Statistics. https://www.iea.org/data-and-statistics/data-

tables?country=WORLD&energy=Balances&year=2018. Accessed: 07052022.

[3] UN. The Paris Agreement. https://unfccc.int/process-and-meetings/the-

paris-agreement/the-paris-agreement. Accessed: 07052022.

[4] American Rivers. How Dams Damage Rivers. https://www.americanrivers.

org / threats - solutions / restoring - damaged - rivers / how - dams - damage -
rivers/. Accessed: 08052022.

[5] Ahlfors, Charlotta and Amelin, Mikael. “Weekly planning of hydropower

in systems with large volumes and varying power generation: A literature
review”. In: 2021 IEEE Madrid PowerTech. 2021, pp. 1–7. DOI: 10 . 1109 /
PowerTech46648.2021.9495058.

[6] Söder, Lennart and Amelin, Mikael. “Efficient operation and planning of power
systems”. In: (2011).

[7] Griva Igor Nash Stephen, Sofer Ariela. Linear and Nonlinear Optimzation.
2009.

[8] Vardanyan, Yelena. “On stochastic optimization for short­term hydropower

planning”. PhD thesis. KTH Royal Institute of Technology, 2012.

[9] Olsson, Magnus. “Optimal regulating power market bidding strategies in

hydropower systems”. PhD thesis. KTH, 2005.

[10] kraftnät, Svenska. Översiktlig kravbild för stödtjänster. https://www.svk.se/

siteassets/4.aktorsportalen/systemdrift-o-elmarknad/information-om-
stodtjanster/marknader-for-stodtjanster.pdf. (Accessed: 25.01.2022).

64
BIBLIOGRAPHY

[11] Solapv. Hydropower A Reliable Renewable Energy Source. https://power.

solapv.com/hydropower-a-reliable-renewable-energy-source/. (Accessed:
09.06.2021).

[12] Verksamhet. https://www.vattenreglering.se/. Accessed: 2021­05­08.

[13] energimyndighet, Statens. Vad avgör ett vattenkraftverks betydelse för

elsystemet. https : / / www . energimyndigheten . se / globalassets / nyheter /
2014/vad- avgor- ett- vattenkraftverks- betydelse- for- elsystemet.pdf.
(Accessed: 25.01.2022).

[14] Verderame, Peter M, Elia, Josephine A, Li, Jie, and Floudas, Christodoulos A.
“Planning and scheduling under uncertainty: a review across multiple sectors”.
In: Industrial & engineering chemistry research 49.9 (2010), pp. 3993–4017.

[15] Singh, Vineet Kumar and Singal, SK. “Operation of hydro power plants­a
review”. In: Renewable and Sustainable Energy Reviews 69 (2017), pp. 610–
619.

[16] Kishor, Nand, Saini, RP, and Singh, SP. “A review on hydropower plant models
and control”. In: Renewable and Sustainable Energy Reviews 11.5 (2007),
pp. 776–796.

[17] Thaeer Hammid, Ali, Awad, Omar I, Sulaiman, Mohd Herwan, Gunasekaran,
Saraswathy Shamini, Mostafa, Salama A, Manoj Kumar, Nallapaneni, Khalaf,
Bashar Ahmad, Al­Jawhar, Yasir Amer, and Abdulhasan, Raed Abdulkareem.
“A Review of Optimization Algorithms in Solving Hydro Generation Scheduling
Problems”. In: Energies 13.11 (2020), p. 2787.

[18] Iliadis, Panos M. PardalosSteffen RebennackMario V. F. PereiraNiko A.

Handbook of Power Systems. Vol. 1. Springer Verlag, 2010. ISBN: 978­3­642­
02492­4.

[19] Gjelsvik, Anders, Mo, Birger, and Haugstad, Arne. “Long­and medium­term
operations planning and stochastic modelling in hydro­dominated power
systems based on stochastic dual dynamic programming”. In: Handbook of
power systems I. Springer, 2010, pp. 33–55.

[20] Labadie, John W. “Optimal operation of multireservoir systems: State­of­the­

art review”. In: Journal of water resources planning and management 130.2
(2004), pp. 93–111.

65
BIBLIOGRAPHY

[21] Azad, Abdus Samad, Rahaman, Md Shokor A, Watada, Junzo, Vasant, Pandian,
and Vintaned, Jose Antonio Gamez. “Optimization of the hydropower energy
generation using Meta­Heuristic approaches: A review”. In: Energy Reports 6
(2020), pp. 2230–2248.

[22] Vardanyan, Yelena and Amelin, Mikael. “The state­of­the­art of the short term
hydro power planning with large amount of wind power in the system”. In: 2011
8th International Conference on the European Energy Market (EEM). IEEE.
2011, pp. 448–454.

[23] Zambelli, Monica S, Soares, Secundino, and Silva, Donato da. “Deterministic
versus stochastic dynamic programming for long term hydropower scheduling”.
In: 2011 IEEE Trondheim PowerTech. IEEE. 2011, pp. 1–7.

[24] Scarcelli, ROC, Zambelli, MS, Soares, S, and Carneiro, AAFM. “Ensemble of
Markovian stochastic dynamic programming models in different time scales for
long term hydropower scheduling”. In: Electric Power Systems Research 150
(2017), pp. 129–136.

[25] Pereira, MVF. “Optimal stochastic operations scheduling of large hydroelectric

systems”. In: International Journal of Electrical Power & Energy Systems 11.3
(1989), pp. 161–169.

[26] Gjerden,
Knut Skogstrand, Helseth, Arild, Mo, Birger, and Warland, Geir. “Hydrothermal
scheduling in Norway using stochastic dual dynamic programming; a large­scale
case study”. In: 2015 IEEE Eindhoven PowerTech. IEEE. 2015, pp. 1–6.

[27] Pritchard, Geoffrey. “Stochastic inflow modeling for hydropower scheduling

problems”. In: European Journal of Operational Research 246.2 (2015),
pp. 496–504.

[28] Helseth, Arild, Fodstad, Marte, and Mo, Birger. “Optimal medium­term
hydropower scheduling considering energy and reserve capacity markets”. In:
IEEE Transactions on Sustainable Energy 7.3 (2016), pp. 934–942.

[29] Warland, Geir and Mo, Birger. “Stochastic optimization model for detailed long­
term hydro thermal scheduling using scenario­tree simulation”. In: Energy
Procedia 87 (2016), pp. 165–172.

66
BIBLIOGRAPHY

[30] Xu, Huijun and Rao, Wei. “Long Term Generation Optimization Using
Stochastic Dynamic Programming”. In: Computer and Management (CAMAN),
2011 International Conference on. IEEE. 2011, pp. 1–4.

[31] Peng, Chunhua, Xie, Peng, Pan, Lei, and Yu, Rong. “Flexible robust optimization
dispatch for hybrid wind/photovoltaic/hydro/thermal power system”. In: IEEE
Transactions on Smart Grid 7.2 (2015), pp. 751–762.

[32] Chazarra, Manuel, García­González, Javier, Pérez­Díaz, Juan I, and Arteseros,

Montserrat. “Stochastic optimization model for the weekly scheduling of a
hydropower system in day­ahead and secondary regulation reserve markets”.
In: Electric Power Systems Research 130 (2016), pp. 67–77.

[33] Abgottspon, Hubert and Andersson, Göran. “Stochastic scheduling for a price­
maker hydro producer considering forward trading”. In: 2013 IEEE Grenoble
Conference. IEEE. 2013, pp. 1–6.

[34] Yuhui, XING, ZHANG, Maolin, Bangcan, WANG, Wenjiao, DING, Siyu, GONG,
and ZHANG, Wenjing. “Midium and Long Term Contract Decomposition
Method Considering Cascade Water Quantity Matching Constraint”. In: 2019
IEEE Sustainable Power and Energy Conference (iSPEC). IEEE. 2019, pp. 176–
186.

[35] Shrestha, GB, Pokharel, Bikal Kumar, Lie, Tek Tjing, and Fleten, S­E. “Medium
term power planning with bilateral contracts”. In: IEEE Transactions on Power
Systems 20.2 (2005), pp. 627–633.

[36] Hjelmeland, Martin N, Helseth, Arild, and Korpås, Magnus. “A case study on
medium­term hydropower scheduling with sales of capacity”. In: (2016).

[37] Hjelmeland, M. N., Zou, J., Helseth, A., and Ahmed, S. “Nonconvex
Medium­Term Hydropower Scheduling by Stochastic Dual Dynamic Integer
Programming”. In: IEEE Transactions on Sustainable Energy 10.1 (2019),
pp. 481–490. DOI: 10.1109/TSTE.2018.2805164.

[38] Helseth, A., Mo, B., and Hågenvik, H. O. “Nonconvex Environmental

Constraints in Hydropower Scheduling”. In: 2020 International Conference on
Probabilistic Methods Applied to Power Systems (PMAPS). 2020, pp. 1–6. DOI:
10.1109/PMAPS47429.2020.9183590.

67
BIBLIOGRAPHY

[39] Aasgård, Ellen K, Andersen, Gørild S, Fleten, Stein­Erik, and Haugstvedt,

Daniel. “Evaluating a stochastic­programming­based bidding model for a
multireservoir system”. In: IEEE Transactions on Power Systems 29.4 (2014),
pp. 1748–1757.

[40] Fleten, Stein­Erik and Kristoffersen, Trine Krogh. Short­term hydropower

production planning by stochastic programming. Stochastic Programming E­
Print Series. 2006.

[41] Pérez­Díaz, Juan I, Wilhelmi, José R, and Arévalo, Luis A. “Optimal short­term
operation schedule of a hydropower plant in a competitive electricity market”.
In: Energy Conversion and Management 51.12 (2010), pp. 2955–2966.

[42] Belsnes, Michael Martin, Wolfgang, Ove, Follestad, Turid, and Aasgård, Ellen
Krohn. “Applying successive linear programming for stochastic short­term
hydropower optimization”. In: Electric Power Systems Research 130 (2016),
pp. 167–180.

[43] Takigawa, Fabrício YK, Finardi, Erlon C, and Silva, Edson L da. “A
decomposition strategy to solve the short­term hydrothermal scheduling based
on lagrangian relaxation”. In: 2010 IEEE/PES Transmission and Distribution
Conference and Exposition: Latin America (T&D­LA). IEEE. 2010, pp. 681–
688.

[44] Gea­Bermúdez, Juan, Das, Kaushik, Koduvere, Hardi, and Koivisto, Matti
Juhani. “Day­ahead market modelling of large­scale highly­renewable multi­
energy systems: analysis of the North Sea region towards 2050”. In: Energies
14.1 (2021), p. 88.

[45] Matevosyan, Julija, Olsson, Magnus, and Söder, Lennart. “Hydropower

planning coordinated with wind power in areas with congestion problems for
trading on the spot and the regulating market”. In: Electric Power Systems
Research 79.1 (2009), pp. 39–48.

[46] Kadowaki, Makoto, Ohishi, Takaaki, Martins, Leonardo SA, and Soares,
Secundino. “Short­term hydropower scheduling via an optimization­simulation
decomposition approach”. In: 2009 IEEE Bucharest PowerTech. IEEE. 2009,
pp. 1–7.

68
BIBLIOGRAPHY

[47] Guedes, LSM, Vieira, DAG, Lisboa,

AC, and Saldanha, RR. “A continuous compact model for cascaded hydro­power
generation and preventive maintenance scheduling”. In: International Journal
of Electrical Power & Energy Systems 73 (2015), pp. 702–710.

[48] Guedes, Lucas SM, Mendonça Maia, Pedro de, Lisboa, Adriano Chaves,
Vieira, Douglas Alexandre Gomes, and Saldanha, Rodney Rezende. “A unit
commitment algorithm and a compact MILP model for short­term hydro­power
generation scheduling”. In: IEEE Transactions on Power Systems 32.5 (2017),
pp. 3381–3390.

[49] Fodstad, Marte, Helseth, Arild, and Henden, Arild Lote. “Modeling start/stop in
short­term multi­market hydropower scheduling”. In: 2016 13th International
Conference on the European Energy Market (EEM). IEEE. 2016, pp. 1–5.

[50] Chen, Yue, Liu, Feng, Wei, Wei, Mei, Shengwei, and Chang, Naichao.
“Robust unit commitment for large­scale wind generation and run­off­river
hydropower”. In: CSEE journal of power and energy systems 2.4 (2016),
pp. 66–75.

[51] Kong, Jiehong, Skjelbred, Hans Ivar, and Fosso, Olav Bjarte. “An overview
on formulations and optimization methods for the unit­based short­term
hydro scheduling problem”. In: Electric Power Systems Research 178 (2020),
p. 106027.

[52] Rodríguez, Jesús A, Anjos, Miguel F, Côté, Pascal, and Desaulniers, Guy.
“Accelerating Benders decomposition for short­term hydropower maintenance
scheduling”. In: European Journal of Operational Research 289.1 (2021),
pp. 240–253.

[53] Séguin, Sara, Fleten, Stein­Erik, Côté, Pascal, Pichler, Alois, and Audet, Charles.
“Stochastic short­term hydropower planning with inflow scenario trees”. In:
European Journal of Operational Research 259.3 (2017), pp. 1156–1168.

[54] Froger, Aurélien, Gendreau, Michel, Mendoza, Jorge E, Pinson, Eric, and
Rousseau, Louis­Martin. “Maintenance scheduling in the electricity industry: A
literature review”. In: European Journal of Operational Research 251.3 (2016),
pp. 695–706.

69
BIBLIOGRAPHY

[55] Wu, Lei, Shahidehpour, Mohammad, and Li, Tao. “GENCO’s risk­based
maintenance outage scheduling”. In: IEEE Transactions on Power Systems 23.1
(2008), pp. 127–136.

[56] Foong, Wai Kuan, Simpson, Angus R, Maier, Holger R, and Stolp, Stephen. “Ant
colony optimization for power plant maintenance scheduling optimization—
a five­station hydropower system”. In: Annals of Operations Research 159.1
(2008), pp. 433–450.

[57] Canto, Salvador Perez. “Application of Benders’ decomposition to power plant

preventive maintenance scheduling”. In: European journal of operational
research 184.2 (2008), pp. 759–777.

[58] Sarasty, Rodriguez and Andrés, Jesus. “Mixed­Integer Programming

Approaches for Hydropower Generator Maintenance Scheduling”. PhD thesis.
École Polytechnique de Montréal, 2018.

[59] Helseth, Arild, Fodstad, Marte, and Mo, Birger. “Optimal hydropower
maintenance scheduling in liberalized markets”. In: IEEE Transactions on
Power Systems 33.6 (2018), pp. 6989–6998.

[60] Jönsson, Ulf. Optimal Control. SE­100 44 Stockholm Sweden: Royal Institute
of technology, 2010.

[61] Helseth, A, Mo, B, Henden, AL, and Warland, G. “Detailed long­term hydro­
thermal scheduling for expansion planning in the Nordic power system”.
English. In: Iet Generation Transmission & Distribution 12.2 (2018), pp. 441–
447. ISSN: 1751­8687.

70