See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/309035528

Modelling Wind Power for Grid Integration Studies

Thesis · November 2016

CITATIONS READS

2 540

1 author:

Jon Olauson
Svenska Kraftnät (swedish TSO)
28 PUBLICATIONS   644 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Reanalyses View project

Wind Power Forecast Analysis View project

All content following this page was uploaded by Jon Olauson on 13 October 2016.

The user has requested enhancement of the downloaded file.

 Digital Comprehensive Summaries of Uppsala Dissertations
from the Faculty of Science and Technology 1428

Modelling Wind Power for Grid

Integration Studies

JON OLAUSON

ACTA
UNIVERSITATIS
UPSALIENSIS ISSN 1651-6214
ISBN 978-91-554-9690-6
UPPSALA urn:nbn:se:uu:diva-302837
2016
Dissertation presented at Uppsala University to be publicly examined in Polhemsalen,
Ångströmlaboratoriet, Lägerhyddsvägen 1, Uppsala, Friday, 4 November 2016 at 09:15 for
the degree of Doctor of Philosophy. The examination will be conducted in English. Faculty
examiner: Professor Thomas Hamacher (Technische Universität München).

Abstract
Olauson, J. 2016. Modelling Wind Power for Grid Integration Studies. Digital Comprehensive
Summaries of Uppsala Dissertations from the Faculty of Science and Technology 1428.
114 pp. Uppsala: Acta Universitatis Upsaliensis. ISBN 978-91-554-9690-6.

When wind power and other intermittent renewable energy (IRE) sources begin to supply a
significant part of the load, concerns are often raised about the inherent intermittency and
unpredictability of these sources. In order to study the impact from higher IRE penetration
levels on the power system, integration studies are regularly performed. The model package
presented and evaluated in Papers I–IV provides a comprehensive methodology for simulating
realistic time series of wind generation and forecasts for such studies. The most important
conclusion from these papers is that models based on coarse meteorological datasets give very
accurate results, especially in combination with statistical post-processing. Advantages with our
approach include a physical coupling to the weather and wind farm characteristics, over 30 year
long, 5-minute resolution time series, freely and globally available input data and computational
times in the order of minutes. In this thesis, I make the argument that our approach is generally
preferable to using purely statistical models or linear scaling of historical measurements.
In the variability studies in Papers V–VII, several IRE sources were considered. An important
conclusion is that these sources and the load have very different variability characteristics in
different frequency bands. Depending on the magnitudes and correlations of these fluctuation,
different time scales will become more or less challenging to balance. With a suitable mix of
renewables, there will be little or no increase in the needs for balancing on the seasonal and
diurnal timescales, even for a fully renewable Nordic power system. Fluctuations with periods
between a few days and a few months are dominant for wind power and net load fluctuations
of this type will increase strongly for high penetrations of IRE, no matter how the sources are
combined. According to our studies, higher capacity factors, more offshore wind power and
overproduction/curtailment would be beneficial for the power system.

Keywords: Wind power, Wind power modelling, Intermittent renewables, Variability,

Integration or renewables, Reanalysis data, Power system studies

Jon Olauson, Department of Engineering Sciences, Electricity, Box 534, Uppsala University,
SE-75121 Uppsala, Sweden.

© Jon Olauson 2016

ISSN 1651-6214
ISBN 978-91-554-9690-6
urn:nbn:se:uu:diva-302837 (http://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-302837)
Dedicated to the Royal Academic Orchestra
List of papers

This thesis is based on the following papers, which are referred to in the text
by their Roman numerals.

I J. Olauson and M. Bergkvist, “Modelling the Swedish wind power

production using MERRA reanalysis data”, Renewable Energy, vol.
76, pp. 717–725, 2015

II J. Olauson, H. Bergström and M. Bergkvist, “Restoring the missing

high-frequency fluctuations in a wind power model based on reanalysis
data”, Renewable Energy, vol. 96, pp. 784–791, 2016

III J. Olauson, M. Bergkvist and J. Rydén, “Simulating intra-hourly wind

power fluctuations on a power system level”, Resubmitted after
revisions to Wind Energy, August 2016

IV J. Olauson and M. Bergkvist, “A new approach to obtain synthetic

wind power forecasts for integration studies”, Resubmitted after
revisions to Energies, September 2016

V J. Widén, N. Carpman, V. Castellucci, D. Lingfors, J. Olauson,

F. Remouit, M. Bergkvist, M. Grabbe and R. Waters, “Variability
assessment and forecasting of renewables: A review for solar, wind,
wave and tidal resources”, Renewable & Sustainable Energy Reviews,
vol. 44, pp. 356–375, 2015

VI J. Olauson and M. Bergkvist, “Correlation between wind power

generation in the European countries”, Energy, vol. 114, pp. 663–670,
2016

VII J. Olauson, N. Ayob, M. Bergkvist, N. Carpman, V. Castellucci, A.

Goude, D. Lingfors, R. Waters and J. Widén, “Net load variability in
the Nordic countries with a highly or fully renewable power system”,
Resubmitted after revisions to Nature Energy, August 2016

Reprints were made with permission from the publishers.

The author has also contributed to the following work, not included in the
thesis.
A J. Olauson, A. Goude and M. Bergkvist, “Wind energy converters and
photovoltaics for generation of electricity after natural disasters”, Ge-
ografiska Annaler: Series A, Physical Geography, vol. 97, no. 1, pp.
9–23, 2015
B J. Olauson, J. Samuelsson, H. Bergström and M. Bergkvist, “Using the
MIUU model for prediction of mean wind speed at low height”, Wind
Engineering, vol. 39, no. 5, pp. 507–518, 2015.
C J. Olauson, D. Lingfors, M. Bergkvist and J. Widén, “Quantifying vari-
ability: A review of metrics and a case study of net load variability”, Pro-
ceedings for 13th wind integration workshop, Berlin, Germany, 2014.
D J. Olauson, H. Bergström and M. Bergkvist, “Scenarios and time se-
ries of future wind power production in Sweden”, Energiforsk report
2015:141, ISBN 978-91-7673-141-3, 2015.
Contents

Acknowledgements ............................................................................................ 9
Abbreviations ................................................................................................... 11
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.1 Wind integration challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.2 Research objectives and previous work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3 Guiding principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4 Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.5 Wind power research at UU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.6 Wind power and natural disasters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1 Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Meteorological models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2.1 Coarse reanalyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.2 Mesoscale models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.3 (Re)forecasts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3 Theory and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1 Wind speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2 Wind power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.1 Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.2 Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.3 Quantifying variability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2.4 Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2.5 Forecasts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3 Statistical methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3.1 Optimisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3.2 Transformation between distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.3.3 Splines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.3.4 Multivariate ARMA models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.3.5 Empirical orthogonal functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.4 Frequency domain methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.4.1 Power spectral density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.4.2 Synthesis of time series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.5 Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.6 Machine learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.6.1 Classification and regression trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
 3.6.2 Random forests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.6.3 Boosting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.1 Modelling wind power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.1.1 Basic model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.1.2 Improved model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1.3 Intra-hourly fluctuations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.1.4 Synthetic forecasts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2 Wind power in Sweden . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.2.1 Historical generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.2.2 Future generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.3 Variability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.3.1 Wind power correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.3.2 A fully renewable Nordic power system . . . . . . . . . . . . . . . . . . . . . . . . 92
5 Concluding discussion .............................................................................. 98
6 Future work .............................................................................................. 100
7 Summary of papers .................................................................................. 101
8 Svensk sammanfattning .......................................................................... 104
References ...................................................................................................... 106
Acknowledgements

And since dissertations can be written about everything under the sun, the
number of topics is infinite. Sheets of papers covered with words pile up in
archives sadder than cemeteries [...]

–M. Kundera, The unbearable lightness of being, 1984

Performing research can sometimes be a lonely task and, as Kundera kindly

reminds us, one can get the feeling that the output is often archived and for-
gotten. I however consider myself lucky in being recommended a research
topic which I find highly interesting and rewarding and with possibilities for
fruitful collaborations in and outside academia. For this, I am very thankful to
my main supervisor Mikael Bergkvist. I would also like to thank Mikael for
always finding time in his busy schedule for discussions and for contributing
to the good ambience at our division.
My co-supervisor Mats Leijon is a busy man with many irons in the fire.
Although Mats has not been involved in the details of my research, he is ac-
knowledged for obtaining funding for my work and for the confidence he puts
in us PhD students. MSB (Swedish civil contingencies agency) and Energi-
forsk are gratefully acknowledged for providing research funding. I hope you
find it money well spent.
Research is much more fun if you don’t do it yourself! I would like to
thank my co-authors Anders, Mikael, Joakim W, Nicole, Valeria, David, Flore,
Mårten, Rafael, Jonatan, Hans, Nasir, Jesper, Johan and Joakim L. I hope
we will have the chance to work more together in the near future. Thanks
also go to teaching colleagues and students, to the administrative personnel
Ingrid, Maria, Gunnel, Thomas and Anna for always being helpful, to Colin
for sharing his expertise on Belgian wind power, to Magnus and Tobias for
their help with filter design, to my office roommates Yue, Maria and Tobias,
to the organizers, senior researchers and PhD students at CNDS (Centre for
Natural Disaster Sciences) and to the people in the wind power group: Hans,
Sandra, Mikael, Anders, Per, Morgan, Stefan, Petter, Bahri, Marcus, Erik,
Senad and Victor.
I do not forget my former colleagues at Sweco: you taught me a lot. Thank
you Fredrik for being a very flexible manager and to Mattias, Olle and Pelle.

9
I would also like to thank Oskar Sämfors and Erik Böhlmark (Svenska Kraft-
nät), Daniel Gustafsson, Pierre-Julien Trombe, Roy Lilleberg, Linn Saarinen
and Tobias Nylander (Vattenfall), Lennart Söder and Richard Scharff (KTH),
Anton Steen (Svensk Vindenergi), Matthias Mohr, Jon Kjellin and Morgan
Rossander (Uppsala University), Åsa Elmqvist (Energiforsk) and Johanna
Laakso (Energimyndigheten) for valuable input to our work. The follow-
ing persons have had a particularly important impact on this thesis frame:
Dorothée, Anders, and Mikael who commented on the manuscript and Hannes
and Valny who made the beautiful cover (photographer back cover: Hans
Bernhoff).
Many persons have also helped with data access and interpretation. To
mention a few: Marcus Flarup and Sverker Hellström at SMHI, Mattias Won-
dollek at Svensk Vindenergi, Birger Fält at Svenska Kraftnät, Kari Fougman
at Sweco and Karin Salevid at Vattenfall. The hard work of many weather
services around the world have led to extremely important (and often freely
available) resources for research. Researchers at NASA and NOAA are ac-
knowledged for producing such datasets, which have been used in all papers
included in this thesis. Transmission system operators sharing their data are
also gratefully acknowledged.
A special thanks goes to my fellow musicians in the Royal Academic Or-
chestra and its committed leader Stefan Karpe. If it wasn’t for you my life
would have been so much duller and I would probably never have returned to
Uppsala for PhD studies. This thesis is therefore dedicated to you. Any or-
chestra would soon collapse without a strong viola section; great work Chizu,
Helga, Amanda, Tom, Anna, Samuel, Mari, Thomas, Ester, Maria, Johanna,
Johan, Kalle, Geraldine, Laura, Daniel and Alice. And thank you Julia, Ven-
dela, Emil, Matteo, Malin, Oskar, Lena, Erik, Anja, Moa, Andrea, Diego,
Gustav, Jonatan, Viktor, Anna, Klara, Linda, Viktor B, Pelle, Mattias, Tomas,
Beatriz, Jan, Björn, Bo, Anna-Maria etc. for our chamber music experiences.
Isak, Colin, Kalle, Frida, Jonatan, Mi-ae and Emil, thank you for wonderful
hiking, skiing, paddling and cycling trips. And to old friends whom I haven’t
seen as much as I would have liked; Berk, Tomas, Amanda, Jens, Marja, Olle,
Åsa and Jonas, I hope you can make it to the dissertation party. Finally I
would like to thank Dodo and my dear family Ingrid, Erland, Hannes, Viktor
and Linn for your love and support.

10
Abbreviations

ACF Autocorrelation function

ARMA Autoregressive moving average
CART Classification and regression tree
CF Capacity factor
EOF Empirical orthogonal function
FE Forecast error
FFT Fast Fourier transform
GB Gradient boosting
GEFS Global ensemble forecasting system
HVDC High-voltage direct current
IEA International Energy Agency
IFFT Inverse fast Fourier transform
IRE Intermittent renewable energy
LTC Long-term correction
MA Moving average
MERRA Modern-era retrospective analysis for
research and applications
MIUU Department of meteorology, Uppsala University
NWP Numerical weather prediction
PC Power curve
PSD Power spectral density
p.u. Per unit
PV Photovoltaics
RMS Root mean square
SD Standard deviation
SE Sweden (SE1–4 for the four different bidding zones)
SMHI Swedish Meteorological and Hydrological Institute
SNSP System non-synchronous penetration
TSO Transmission system operator
VI Variability index
WT Wind turbine

11
1. Introduction

If we combine the equations of Magnetic Force (B) with those of Electric Cur-
rents (C) [...]

–J.C. Maxwell, A Dynamical Theory of the Electromagnetic Field, 1865

In the early history of wind power, the wind turbines (WTs) were built
to supply local energy demands including e.g. water pumping and grinding.
When it became more common for WTs to be connected to large electric grids
in the 1970’s, a rapid increase in tower height, turbine diameter and generator
rating begun, which has continued ever since. [1] The main drivers of this
development have been the better wind conditions at higher hub heights and
economics of scale, which give lower costs per produced energy unit for larger
WTs. The dominating design of modern WTs is three-bladed horizontal axis,
but other concepts are also under consideration. Although the lion’s share
of todays installed capacity is grid-connected and large-scale, there is some
interest in small-scale wind power. In most areas of the world, photovoltaics
(PV), for which generation costs are much less dependent on scale, however
outperforms small-scale wind [2].
Starting from a global installed capacity of 5 GW at the end of 1995 (which
has now been surpassed by Sweden alone), the capacity reached 59 GW in year
2005 and 433 GW in 20151 . Currently, roughly 4% of the world’s electricity
demand is met by generation from WTs. In some countries and regions, wind
power supplies a more substantial part of the load. Two examples are Spain
and Denmark with 19% and 42% wind penetration2 . By including planned
and projected installations, several other regions will have high shares of wind
power in the near future. The Global Wind Energy Council (GWEC) has pub-
lished three long-term scenarios for wind deployment [3]. In the “moderate”
scenario, the capacity will supersede 700 GW by 2020 and 2000 GW by 2040.
Whether these projections will be realised or not is of course impossible to
tell. One can however note that the GWEC moderate scenario published in
2008 has so far been too conservative.
1 http://www.gwec.net/ Accessed: 2016-04-20.
2 Data from the system operators (Red Eléctrica and Energinet respectively) for year 2015.

12
Table 1.1. Levelised cost of electricity for selected generation technologies in Sweden
2014 (excluding policy instruments, 6% cost of capital). Results from [4] recalculated
with an exchange rate of 9.3 SEK/e. The rated capacities are the sizes of the (set of)
power plants for which costs have been calculated, e.g. a 150 MW onshore wind farm.
Technology Rated capacity Investment Energy cost
(MW) (e/W) (e/MWh)
Onshore wind 150 1.3 55
Offshore wind 600 2.5 81
Rooftop PV 0.005 1.7 183
PV farm 1 1.1 100
Nuclear 1600 4.3 58
Hydro power 90 2.2 49
Gas co-gen 150 0.75 63

According to a recent (year 2014) estimate for Sweden [4], onshore wind
power compares favourably to most other generation technologies in terms of
levelised cost of electricity. Table 1.1 shows a selection of the investment and
energy costs, assuming 6% cost of capital, an exchange rate of 9.3 SEK/Euro
and excluding policy instruments (taxes and subsidies). The costs include
those for capital, operation and maintenance and insurances but not integra-
tion costs induced by intermittency and unpredictability. If policy instruments
are taken into consideration, the competitiveness of wind power is increased
further.
Wind power in Sweden and the Nordic synchronous power system are the
main concerns of this thesis. Fig. 1.1 shows the eleven bidding zones3 in the
Nordic system and wind farms in operation or under construction as of April
2015. The data for Sweden are relatively accurate, but of somewhat poorer
quality for the other countries. The database, taken from Paper VII, comprises
800–1000 MW in eastern Denmark4 , Finland and Norway respectively and
6500 MW in Sweden. As can be seen in the figure, the farms in Denmark
and Sweden are well dispersed while deployment in Finland and Norway has
been concentrated to the coastal areas. Recently (early 2016), an investment
decision was taken for a 1000 MW project in bidding zone NO3. The six
farms in this project, not included in the map, are to be completed by 2020.
Maximum net transfer capacities5 in MW for import to Sweden are given in
red text in Fig. 1.1 and Swedish HVDC links are indicated by black arrows.
3 Because of bottlenecks in the transmission system, the Nordic system has been divided into
bidding zones. By allowing the electricity price to differ between the zones, actors are encour-
aged to act accordingly, e.g. install more capacity in zones with deficits. The current eleven
zones have been in place since 2011.
4 Western Denmark is not part of the Nordic synchronous system.
5 http://nordpoolspot.com/globalassets/download-center/tso/max-ntc.pdf Ac-

cessed: 2016-04-25.

13
 0 100 200 400 km

NO4

700 1100

SE1

250 3300

FI
600 SE2
NO3

2145 7300
NO5
NO1 1200

NO2 SE3

740 5300

SE4 700
1700

± DK2 615
600

Figure 1.1. The eleven bidding zones and wind farms in the Nordic synchronous
system (operating or under construction as of April 2015). The circle areas are pro-
portional to farm capacities. Offshore country borders are only illustrative. Maximum
net transfer capacities in MW for Sweden as of March 2016 are given in red text and
Swedish HVDC links are indicated by black arrows. For comparison, the average load
in Sweden is around 15,500 MW.

14
1.1 Wind integration challenges
Due to its variable and unpredictable nature, integration of wind power into
the power system poses several challenges and induces additional costs [5].
Depending on the characteristics of the power system, e.g. thermal or hydro
dominated, system size, load characteristics and transmission capacities, the
magnitudes of the challenges for similar wind penetration levels differ.
By definition, there is always a balance between generation and load. If the
load is high and the wind generation is low, sufficient capacity must be avail-
able from other generators, import and demand side flexibility. Assuming that
conventional power plants are dismantled when wind farms are constructed,
an important issue is therefore to make sure that the load can be supplied at all
times. The deployment of wind power can also increase the magnitudes of the
changes in net load6 , which may be an issue. Very short wind power fluctu-
ations due to turbulent gusts are close to uncorrelated between sites and thus
effectively dampened by aggregation. The ramps associated with passages of
large-scale weather systems are most critical according to [6]. In addition to
the inherent variability, wind power generation is difficult to predict, especially
for longer forecast horizons. The planning of the power system thus becomes
more challenging if the wind penetration is high.
The variability of wind power is not only a technical question, it also pro-
foundly affects the energy markets and the profitability of both conventional
generators and wind farms. By introducing intermittent generation with very
low operating costs, the electricity price is reduced but also becomes more
volatile. Systems with relatively high shares of intermittent generation, e.g.
Germany and Denmark, sometimes even experience periods with negative
prices. For operators of conventional plants, the introduction of more wind
power may lead to lower capacity factors (CFs)7 and more cycling. WT own-
ers, on the other hand, may find that the average revenue for their sold electric-
ity is lower than the time-averaged price [7]. A different generation portfolio
implies that different market structures may be necessary. As an example, ca-
pacity markets are already in place in some systems and under consideration
in others [8].
For several reasons, wind power requires investments in new and existing
transmission grids. Firstly, wind farms are often located in remote areas, and
thus requires construction of new lines. At least in Sweden, this expense is
covered by the farm owner, so it should not be considered an integration cost.
Investments in the existing grid can be necessary since wind farms have lower
CFs than conventional generation (so larger transmission capacity is needed
to transfer the same annual energy) and may be located far from load centres.
Again with Sweden as example, load is concentrated in the southern part of
the country, but wind farms are more evenly spread (see Fig. 1.1). Grid rein-
6 The net load is generally defined as load minus intermittent renewable generation.
7 The capacity factor is the mean generation normalised to the installed capacity.

15
forcements will become necessary in order to transfer wind power from north
to south. In order to cope with the increased net load variability, it may also
be rational to strengthen cross-border transmission capacity [9, 10]. Lastly,
if conventional generation is dismantled when more WTs are erected, there
may be deficits of reactive power in some regions. This leads to a less stable
voltage and reduces the transmission capacity of existing lines.
Last but not least, replacing synchronous generators with wind farms de-
coupled from the grid by power electronics reduces the inertia of the system.
The lower the inertia, the less damping is provided and the frequency thus
reacts faster to abrupt changes in load or generation. From the system oper-
ators’ point of view, the reduction of inertia is generally considered the main
challenge with integration of renewables [11]. The instantaneous wind pene-
tration levels (wind/load) can be considerably higher than the average. As an
example, in Sweden 2015 the mean penetration was 12% but the maximum
hourly penetration was 39%. In order to keep the inertia at acceptable levels,
limitations can be set on the instantaneous penetration. From a technical point
of view, it is not complicated to curtail wind power. If the average penetration
is not very high, relatively small amounts of energy have to be spilled at times
when prices can be expected to be low, see Section 4.2.2.
On the positive side, the dimensioning errors can be smaller for a system
without large centralised generators such as coal or nuclear. Wind power can
also enhance voltage and transient stability by providing reactive power [12],
have fault-ride-through capabilities [13] and provide synthetic inertia [11]. It
is likely that in a future with a significantly higher wind penetration, the farms
will (be required to) have these capabilities and assist in balancing the power
system. More ideas on how to tackle challenges with wind integration can be
found in two reports [14, 15] in Swedish.

1.2 Research objectives and previous work

Over the last decades, the levelized cost of electricity (LCOE) for wind power
has fallen steadily. According to [16], the global weighted LCOE for on-
shore wind was around 0.07 $/kWh year 2015, which can be compared to
0.20 $/kWh year 1995 (assuming a capital cost of 7.5%). For utility-scale
PV, the decline has been even faster and the cost is now around 0.12 $/kWh.
For year 2025, global weighted costs for onshore wind and PV are projected
to be 0.05–0.06 $/kWh. It is thus reasonable to believe that integration chal-
lenges rather than LCOEs will be the main barriers to overcome if intermittent
renewables should deserve a significant role in future power systems.
In order to achieve a smooth and effective transition to highly or fully
renewable power systems, integration studies are necessary. Realistic wind
power scenarios and corresponding time series, both generation and forecasts,
are key components in any such study. The main research objective of this the-

16
 Model package

Met. data
Simulation of Model of
Model of hourly
intra-hourly wind power Time series for
generation
fluctuations forecasts power system studies
Paper I and II
Wind farm Paper III Paper IV
info

Variability Correlation 100% renewable

review wind in Europe Nordic system?

Paper V Paper VI Paper VII

Figure 1.2. Scope of thesis. “Met. data” is short for meteorological data. Wind
farm information can be either on actual farms (when the objective is to evaluate the
models) or scenarios of future farms.

sis is to develop and improve methodologies for producing these time series.
Contrary to e.g. [17], which uses highly resolved meteorological models, we
rely on globally and freely available data from coarse, meteorological reanal-
yses8 . This approach has its obvious benefits, but also requires that particular
care is taken to ensure that e.g. high-frequency fluctuations are appropriately
represented. A secondary research objective is the study of variability of inter-
mittent renewables, possibly in combination with the electric load, and possi-
bilities for reducing this variability. It is important to stress that these studies
are statistical in nature and do not take into account transmission limitations,
costs, balancing capabilities of hydro power etc. Such issues will be the topic
for future work, see Chapter 6.
The scope of the papers included in this thesis is illustrated in Fig. 1.2.
Methods and results from these papers are presented in this thesis frame. Ad-
ditionally, selected results on wind power in Sweden are given. These are
partially new and partially taken from Paper D.
In the remainder of this subsection, some work that has been influential for
my thinking are credited. First, IEA Wind Task 25 [5, 18–20] on “Power sys-
tems with large amounts of wind power”, led by Hannele Holttinen, should
be acknowledged for their work on compiling results from wind integration
studies and providing guidelines for such studies including input data require-
ments.
Freely available meteorological reanalyses [21, 22] have become popular
for creating wind power time series [10, 23–28]. The complexity of the meth-
ods used and the extent of validations with measurements differ between the
8 Long meteorological time series produced with a consistent model and assimilation system,
see Section 2.2.

17
authors. Output from such models can be post-processed for obtaining appro-
priate high-frequency fluctuations or time series with higher temporal resolu-
tion [17, 29, 30]. In many types of wind integration studies, simulated (syn-
thetic) forecasts are necessary for determining the impact from wind power
uncertainty on the operation of power systems [18]. Our methodology for
generating such time series differed from earlier attempts, but the multivariate
ARMA method used by Söder [31] and others influenced our thinking.
Regarding statistical techniques for modelling and studying the variabil-
ity of wind power, power spectral density estimates [32, 33], machine learn-
ing [34] and empirical orthogonal functions [35] have all been used earlier.
Correlations of outputs between different farms or regions, which are impor-
tant for wind power smoothing, have been studied by e.g. [36–38].
Finally, the study by Staffell and Green [23] on deterioration of wind farm
performance and the work by Hirth et. al. [7,39] on the diminished revenues of
wind power for higher penetration rates are acknowledged for their originality
and high quality.

1.3 Guiding principles

In hindsight, a few guiding principles can be distinguished behind all papers
included in the thesis. In conclusion:
• Focus on wind power on the power system level.
• The methods and results should be directly useful (engineering approach)
and results should be shared.
• Use physical models as far as possible. When needed, statistical models
are subsequently employed.
• Always validate the models with measurements on the power system
level.
• Use as long time series as possible.
• Utilise, if possible, input and validation data that are publically and glob-
ally available.
Although large farms can have an effect on the regional grid [40], the largest
impact from wind power is on the power system scale. This scale has thus
been the focus of all papers included in the thesis. Better results are generally
obtained if generation and forecasts are first modelled for each farm. When
tuning the parameters for the models, measurements on the power system level
were however the most important training data. By using freely available in-
put data, demonstrating the methodology on a power system level (including
reasonable computational times) and sharing the results, the chance that the
methodologies are actually used by practitioners is increased.
A very important question is whether to use linear scaling of historical mea-
surements [6, 41, 42], physical models or statistical models [43, 44] for obtain-

18
 ing data for power system studies. Our take on this matter is to use physical
models as far as possible, employ statistical models for fine-tuning and only
use measurements for validating the results. Purely statistical models were
rejected since it is often crucial to have synchronous load and wind power
data [18, 45]. Moreover, physical models are more credible for modelling fu-
ture WTs and farms in new geographical areas (e.g. offshore).
Nobody questions that (quality controlled) measurements are valuable for
studying the past. As the following example shows, linearly scaled measure-
ments can however be very misleading for studying the future. For illustration,
a paper by Kiviluoma et. al. [6] is used. The paper was chosen partly because
it contains a readable analysis of wind power variability in several different
countries, but also because the authors have some influence in the wind inte-
gration community and express doubts whether modelling is appropriate. In
Section 3.2 in [6], net load ramps in future power systems with high wind
penetrations were studied. Historical time series were linearly scaled since
“geographical smoothing is already incorporated in the data”. According to
our analyses in Paper D, this is to some degree correct; if more wind farms
of the same type as already present are deployed in a mature power system,
the reduction in normalised variability is negligible. However, the WTs have
evolved over the years and can be expected to evolve in the future. This will
have a large impact on the variability. As an example (see Section 4.2.2 for
more results), let us consider the first percentiles of one hour step changes9 in
the Swedish net load. With no wind power, the first percentile for the years
2007–2014 was -1.14 GW. By adding 50 TWh/a of wind power, linearly scaled
from historical time series, this figure is reduced to -1.58 GW. With a time se-
ries from a scenario of future deployment (scenario C1 from Paper D), -1.34
GW results. The increase in extreme negative ramp magnitudes is thus less
than half of that obtained by using scaled measurements.10
Since the question of modelling versus measurements is so important, some
results from modelling Swedish wind generation are given already now. As
can be seen in Fig. 1.3, a physical model based on coarse reanalysis data and
some statistical fine-tuning can adequately predict the actual generation. As
for the whole thesis, different time periods were used for training the model
and evaluating the performance.
As can be anticipated from the discussion above, the normalised variabil-
ity of future wind power will most likely be significantly lower than today’s.
When studying the variability and related costs for wind power in the future,
9 The change in averaged power over a fixed number of time steps, e.g. change in hourly aver-
aged power over four hours.
10 For the sceptical reader it can be worth pointing out that modelled time series based on his-

torical farms yields in principle identical results as measurements (around 1% difference). The
difference is thus not due to modelling per se, but rather due to the expected increase in CFs for
future WTs. Scaling historical measurements implicitly assumes no change in WT characteris-
tics.

19
 b) 0.15
a) 0.7 -0.05

One hour step change [p.u.]

0.1
0.6 -0.1
Generation [p.u.]

0.5 0.05 -0.15

0.99 1
0.4 0
0.3 0.15
-0.05
0.2 Measurements
0.1
Model
-0.1
0.1 0.05
0 0.01
0 -0.15
0 1 2 3 0 0.2 0.4 0.6 0.8 1
Time [weeks] Fraction exceedance

Figure 1.3. Models can successfully predict wind generation: (a) time series example,
(b) step change duration curves with extremes as insets. Based on measurements and
model output for the whole of Sweden (results from Paper II).

especially if decisions are to be taken about desirable development paths, I be-

lieve that these must be compared to the energy production, not the installed
capacity. The former is much more informative regarding the benefits from
wind power. Ten GW of wind power with a high CF will be somewhat more
variable than ten GW of wind power with a low CF. This must not lead us to
the false conclusion that it would be better to install low CF farms; if we want
30 TWh/a of electricity, the variability becomes significantly lower if high CF
farms are deployed.
This section is concluded by a few remarks on hardware and software. A
standard desktop computer from year 2011 was used for all computations.
Since some of the datasets were quite large, an extra hard drive and 16 GB
RAM memory have been installed. Matlab has been used solely for the mod-
elling, except in Papers B and III were R was utilised to some extent. The
latter program is free of charge and the preferred choice by most statisticians,
so new methods generally appears first in R. Matlab, on the other hand, is
faster, more user friendly, better documented and has nice graphical quali-
ties (especially with the newly introduced colour schemes). No 3D-plots have
been used since these, in my opinion, are most often only confusing.

1.4 Nomenclature
As much as I would have liked it, the nomenclatures in the different papers
are not entirely consistent. Sometimes similar techniques have been used in
several papers, but with different variable notations. In the method section of

20
 this thesis frame, the nomenclature might thus be different from that in some
of the papers.
The original purpose of my PhD project was to design and construct a small
vertical axis wind turbine, see Section 1.6. In order to avoid confusion of the
rotor of the generator, the “aerodynamic rotor” (rotor blades and hub) and
the whole system (commonly denoted wind turbine), it was decided to denote
the latter wind energy converter (WEC). This denomination was used in the
earlier papers and in my licentiate thesis [2]. In the later papers and in this
thesis, wind turbine (WT) is used.
As for the energy generated by WTs, different terminologies and units have
been used. Average generation for certain fixed time periods has been denoted
“hourly energy”, “generation” or “(average) power” and units have been per
unit (p.u.)11 , Watt etc. If the term power was chosen, this should always be
interpreted as average power at sampling frequency.
When discussing a “wind power scenario”, this refers to a scenario of both
technical parameters of future WTs (capacity factor, specific rating12 etc.) and
the geographical distribution, i.e. coordinates of the farms. The term “aggre-
gated” is used exclusively for geographical aggregation, e.g. the combined
wind power generation in a country. Statistical metrics such as standard de-
viation and correlation are typically denoted by the symbol for the expected
value instead of the sample metric, i.e. σ and ρ instead of s and r. In the thesis
frame, variables and parameters are defined in their context. Abbreviations are
listed on page 11.

1.5 Wind power research at UU

At the Division of Electricity at Uppsala University, research on straight-
bladed, vertical axis WTs has been conducted for over ten years, focusing
on e.g. generators [46, 47], aerodynamics [48], noise [49] and the electrical
system [50]. Three WTs have been designed and constructed for research pur-
poses:
• Torsholm 200 kW [51] - Constructed by the spin-off company Vertical
Wind, but now owned by Uppsala University. A picture of Torsholm is
shown on the back cover of this thesis.
• Lucia 12 kW [52] - Research wind turbine.
• Birgit 10 kW [53] - A novel concept turbine developed to electrify tele-
communication towers.
Wind power studies are also conducted by meteorologists at the depart-
ment of earth sciences. Examples of research fields are the MIUU model (see
11 One per unit corresponds to the maximum theoretical generation, e.g. if the installed capacity
is 1000 MW and 300 MWh is generated a certain hour, this equals 0.3 p.u.
12 Rated power over swept rotor area.

21
Section 2.2.2), wind power in forests [54] and icing [55]. Paper II has a co-
author from the department of earth sciences. In 2013, Gotland University was
merged with Uppsala University. Parts of the island of Gotland have excellent
wind conditions. Many of the first WTs in Sweden were deployed here, most
notably the 3 MW, two-bladed “Matilda” in 1993. At Campus Gotland, the
focus of technical wind power research is on wake modelling, see e.g. [56].
Wind power researchers at (primarily) Uppsala University and the Royal
Institute of Technology (KTH) in Stockholm are organized in the network
“STandUP for Wind”. This centre, involving nearly 50 persons, is profiled
towards planning and establishing wind energy in the Swedish electrical net-
work.

1.6 Wind power and natural disasters

I am a member of the research school Centre for Natural Disaster Sciences
(CNDS)13 . My original research objective was to develop and construct a
portable WT, intended for generation of electricity in the immediate response
and for more long-term recovery work after a natural disaster. Based on the
results from [2] and Papers A and B, it was concluded that photovoltaics is
most often a better solution; the energy cost is generally lower and it is not so
common with extended periods of very low generation, thus a smaller battery
bank is needed. Exceptions are regions close to the poles, but these are not the
most prone to natural disasters and are sparsely inhabited.
The abovementioned results lead to the new research direction that is pre-
sented in this thesis. Although the link to natural disasters may seem weaker
than before, this is not the case. The whole electrical system is susceptible
to disasters such as storms, earthquakes, extreme icing events etc. One of the
eight key risks with a warmer climate pointed out by the IPCC [57] is sys-
temic risks due to extreme weather events leading to breakdown of infrastruc-
ture networks and critical services such as electricity, water supply and health
and emergency services. Moreover, the vulnerability of the power system to
natural disasters can be altered by the power mix; introducing large amounts
of highly variable production with little or no inertia jeopardise the stability
of the grid if proper measures are not taken. In conclusion, wind power and
natural disasters are linked in both directions and proper modelling of future
wind power will be crucial for securing a robust power system.

13 http://www.cnds.se Accessed: 2016-08-04.

22
2. Data

In this chapter, the data used in our studies are presented and some pitfalls
of using coarse reanalyses for modelling wind power are discussed. Some
datasets that were ultimately not utilized are also briefly mentioned. The data
fall into two main categories: measurements and output from meteorological
models.

2.1 Measurements
One of the guiding principles stated in Section 1.3 is that model outputs must
always be compared to actual measurements. If the aim is to generate wind
power time series for a power system, measurements for the whole system
should primarily be used. In some cases it can however be advantageous to
also utilise measurements for single farms, e.g. for determining correlations of
forecast errors.
Time series of hourly, aggregated wind generation for the four bidding
zones in Sweden are available at the Swedish transmission system operators
(TSO) web page [58]. The data were normalised to the installed capacity at
each time step, i.e. expressed in p.u. The definition of installed capacity is not
clear-cut. Swedish Wind Energy1 [59], includes a wind farm in their statis-
tics when it has the “possibility to deliver power to the grid”. The Swedish
Energy Agency [60], on the other hand, only includes turbines that has been
reported to the electricity certificate system. The two dates can be the same,
but the latter is often up to a few months later since the owner does not want
to report the turbine until the testing/trimming phase is completed. When the
installed capacity increases significantly each year, the two databases can thus
yield differences in the national capacity of a few percent. In Paper I, a con-
siderable amount of work was put into correcting and supplementing the data
from [60], thus producing a database of WTs installed year 2012 or earlier.
For the analyses in Section 4.2, this database was extended to include farms
deployed up until 2014. For the reason just mentioned, the installed capacity
in this database may be an underestimation. An average of this time series and
linearly interpolated end-of-year values from [59] was thus used2 .
1 Svensk Vindenergi; the trade association for companies working with wind power in Sweden.
2 Two other databases of wind farms are Vindstat and Vindbrukskollen, see http://vindstat.

com/ and http://www.vindlov.se/sv/vindbrukskollen1/karta/.

23
 In some of the papers, data from other power systems were needed in ad-
dition to the Swedish. Time series of wind power generation, with temporal
resolution ranging from five minutes to one hour, were thus retrieved from
the system operators’ web pages/servers or by mail. As for Sweden, the data
were transformed to p.u. by normalising the MW values to the installed ca-
pacity at each time step. The resolution of the time series of installed capacity
differed between the systems; for some, values for each day were available,
but in some cases interpolation had to be used since only monthly or yearly
values were available. All time series were quality-controlled, removing obvi-
ously erroneous data samples and, in some instances, filling the resulting gaps
with interpolated data. When needed, the time series were shifted in time to
get synchronous data. Most often, all data was converted to UTC+00 (coordi-
nated universal time) or to UTC+01 (Swedish winter time). Particular care had
to be taken to daylight saving time, which is handled differently in different
databases.
Measurements from a few met masts and farms were obtained from the
power company Vattenfall. Forecasted generation time series were shared by
the same company. Data on electric load in the Nordic countries were attained
from the respective TSO’s web page. The Swedish load dataset, used in Pa-
per VII and Section 4.2, is not entirely complete; electricity that is produced
and consumed locally (e.g. in pulp industries) is not included3 .

2.2 Meteorological models

Numerical Weather Prediction (NWP) models are used to forecast wind speed
and other variables for the present (“nowcasts”) up to several days ahead. The
same type of models can also be run with historical data from a constant data
assimilation system in order to produce long and relatively consistent4 datasets
for research and climate services. Several of these retrospective analyses (re-
analyses) are publicly available, e.g. ERA Interim [22], MERRA [21, 61] and
JRA-55 [62]. See [63–65] for reviews that also include earlier generation
reanalyses. This type of datasets has become popular for modelling wind
power [10, 23–28] and was used in all papers included in this thesis. In Pa-
pers V and VII, solar irradiation and wave data from NWP reanalyses were
also used. The following three subsections on coarse reanalyses, mesoscale
models and forecasts respectively are however only devoted to wind power
related variables.

3 The lacking energy amounts to around 4% of the total load.

4 With consistent we mean that the quality of the analysis/forecast does not change substantially

over the years.

24
2.2.1 Coarse reanalyses
Because of the relatively high correlations with measurements [63] and tem-
poral resolution, MERRA was the primary reanalysis choice in the papers
included in this thesis. In Paper II, some variables were also taken from the
ERA Interim dataset. CFSR data [66] were considered but ruled out since the
time series are not consistent in time; from 2011 and onwards, only outputs
from the operational model CFSv2 are available.
MERRA, Modern-Era Retrospective analysis for Research and Applica-
tions, is a NASA reanalysis for the satellite era, i.e. since 1979 [21]. Output
is available with a spatial resolution of 0.5 degrees in latitude and 0.67 de-
grees in longitude and one hour temporal resolution5 . The spatial resolution
of MERRA is quite typical for modern global reanalyses. Available variables
include e.g. northward and eastward wind speeds, temperature, humidity and
air pressure. Wind speeds are available at two and ten meters above the dis-
placement height (see Section 3.1) and at 50 m above ground. Recently, the
atmospheric general circulation model was updated, resulting in a second gen-
eration reanalysis: MERRA2 [61]. The production of MERRA thus stopped
in February 2016. MERRA2 has roughly the same resolution and spans the
same time period as its predecessor. Results from citestaffell2016 show that
the performances are similar in terms of modelling wind power generation.
Although a coarse reanalysis can be very useful for modelling the aggre-
gated generation for a region or country, there are analyses that such a model
should not be used for. A few examples are given in the numbered list below.
For readers not familiar with wind power computations, it is recommended to
read Sections 3.1 and 3.2 before proceeding.
1. Determine likely farm locations. As can be seen in Fig. 2.1, local (and
some regional) variations in the mean wind speed are not captured by a
coarse reanalysis. In [67], the geographical distribution of farms in Ger-
many and the Nordic countries was optimised in terms of a low com-
bined variability. Only sites with an estimated CF of 0.15 or higher
were considered. The NWP data (11 km resolution) thus ruled out most
regions in Sweden except an area in the south, the mountain region bor-
dering Norway and some coastal sites. When comparing to the actual
deployment and planned farms (see Fig. 4.15 on page 83), this picture is
drastically wrong6 .
2. Calculate generation time series without first properly scale the wind
speed. The actual mean wind speed at a site has little in common with

5 Some variables, not used in our work, have a coarser temporal resolution.
6 Note that [67] was chosen as an example since details on farms in Sweden were available for
comparison and that the difference between prediction and actual outcome is so striking. Other
authors [39, 68] have used reanalyses with an even coarser resolution for determining potential
sites for deployment and/or compute generation time series without first properly scale the wind
speed.

25
 that obtained from coarse reanalysis datasets, in particular in more com-
plex terrain. Because of the shape of the power curve, an error of 1 m/s
in the mean wind speed gives around 20–40% error in the annual en-
ergy yield. The mean wind speed should thus be determined either from
more highly resolved resource maps or from the anticipated annual en-
ergy yield of planned farms. Without a proper scaling, the variability in
generation is also misjudged. If the mean wind speed is underestimated,
the variability will be overestimated, see Fig. 2.2a for an example.
3. It is often desirable to perform a LTC (long-term correction) [69] of a
shorter measurement, e.g. for estimating the mean output of a future
farm or, as in [23], for quantifying deterioration in wind farm perfor-
mance. In the latter case, the LTC is done in order to separate possible
long-term trends in the wind climate from trends in farm performance.
If a trend in the wind climate exists and an erroneous mean wind speed
(e.g. taken directly from the coarse reanalysis) is assumed, an artificial
trend in the farm performance will result, see Fig. 2.2b. Even if no long-
term trend is present, the LTC will be more uncertain for shorter mea-
surements (noisy signals in Fig. 2.2b).

2.2.2 Mesoscale models

In order to get more refined results, the output from global, coarse NWPs can
be fed into regional models with higher resolution, a procedure referred to as
dynamical downscaling. Typically, these mesoscale models have a spatial res-
olution ranging from a few hundred meters to 10 km. Results from a mesoscale
model [70–72] developed at the former Department of Meteorology, Uppsala
University (the MIUU model) were used in Paper B and for generation of sce-
narios of wind farm deployment in Paper D. Freely available data7 on mean
wind speeds can be retrieved for different heights and with spatial resolution
0.5 km, see Fig. 2.1a.
The Weather Research and Forecasting (WRF) model [73] is a NWP sys-
tem intended for research and operational forecasts and is used by both major
weather services and many universities. In the beginning of my PhD studies,
we considered using WRF for producing more highly resolved data for mod-
elling of wind power. For various reasons this never happened; firstly there
is a significant learning threshold, secondly, as described in Section 1.3, we
wanted as many as possible to be able to use our methodologies and thirdly,
the fine results obtained (on the power system level) by using coarser models.
The Swedish Meteorological and Hydrological Institute (SMHI) uses meso-
scale models for regional forecasts of various variables. Recently, a large
amount of their data became publically available8 . The MESAN(HIRLAM)
7 http://www.s1106835.crystone.net/?q=en/node/55 Accessed: 2016-07-15.
8 http://opendata-catalog.smhi.se/explore/ Accessed: 2016-04-14.

26
Figure 2.1. Mean wind speeds 100 m above displacement height according to the
MIUU and MERRA models (see Sections 2.2.2 and 2.2.1 respectively). Horizontal
resolutions are 500 m and 0.5 × 0.67◦ respectively. Local, and even some regional,
variations are not captured by MERRA.

27
 2.5 u = 7 m/s (correct) 0.44
a) b)
u = 6 m/s

Capacity factor after LTC

Normalised generation 0.42
2 u = 8 m/s
0.4
1.5 0.38

1 0.36

0.34
0.5
0.32

0 0.3
20 40 60 80 100 120 20 40 60 80 100 120
Month Month
Figure 2.2. Example based on a MERRA time series for illustrating the effects of
assumed mean wind speed on variability and long-term corrected generation. For
a clear demonstration of the latter, a fictive linear trend was added to the time series,
which was subsequently linearly scaled to means of 6, 7 and 8 m/s respectively. In this
example it is assumed that 7 m/s is the correct mean wind speed. (a) The variability in
normalised monthly generation is over- and underestimated with mean wind speeds of
6 and 8 m/s respectively. (b) When the monthly outputs based on a mean wind speed
of 7 m/s are long-term corrected using time series with means of 6 and 8 m/s, artificial
trends in LTC output results (dotted lines).

is an analysis model, describing the current state of the weather every hour
in an eleven km grid. The MESAN dataset was found inconsistent, which is
not surprising since, in contrast to reanalyses, the model has evolved over the
years. As an example, the spatial resolution was not the same throughout the
period. Furthermore, data were only available since 1998, i.e. significantly
shorter than for MERRA. It was thus decided to not use MESAN-data.

2.2.3 (Re)forecasts
For creating synthetic wind power forecast errors with the methodology de-
scribed in Section 3.2.5, forecasted wind speeds and other variables are nec-
essary. Three sources of forecasts were considered and are described briefly
below: regional forecasts produced by SMHI, global forecasts available in
the TIGGE project and GEFS reforecasts. For readers not familiar with wind
power forecasts, it may be useful to read Section 3.2.5 before proceeding.
SMHI uses a limited-area model to generate regional forecasts covering
essentially the Nordic countries. SMHI models have a spatial resolution of
between 2.8 and 11 km and forecast horizons up to 240 h. As mentioned in
Section 2.2.2, much of the data is now publically available. Regarding fore-
casts, only analysis data (“nowcasts” for hour zero) can however be accessed

28
Table 2.1. Information on a selection of global, freely available wind forecast
datasets. Horizontal and temporal resolutions are given for forecasts up to one week
ahead.
GEFS ECMWF NCEP
Type Reforecast Forecast Forecast
Horizontal resolution 0.5◦ 0.28◦ 0.7◦
Time (from) Dec 1984 Oct 2006 Mar 2007
Horizon (days) 0–15 0–14 0–15
Temporal resolution 3–6 h 6h 6h
Perturbed members 10 50 20
Runs per day 1 2 4
Height 80 m 10 m 10 m

free of charge. Furthermore, forecasts are not stored, so even if the policy
would change, historical forecasts would not be available.
TIGGE, THORPEX Interactive Grand Global Ensemble, was a part of the
international research collaboration THORPEX (The Observing System Re-
search and Predictability Experiment). The aim with TIGGE was to provide
operational ensemble forecast data to the research community [74]. Although
the THORPEX project was finalised in year 2014, global forecasts from sev-
eral weather services can still be accessed from e.g. ECMWF’s web page9 .
NCEP’s GEFS (Global Ensemble Forecasting System) [75] is, to my knowl-
edge, the only long-term, global and freely available dataset of reforecasts.
In analogue to reanalyses being produced from historical data using approx-
imately consistent models and assimilation systems, reforecasts that are con-
sistent with an operating forecast system can be produced.
In Table 2.1, information on GEFS second generation reforecast and the
global forecasts from ECMWF and NCEP is given. The latter two were cho-
sen out of the TIGGE members since these were most promising for our needs;
the datasets should preferably span several decades, be consistent in time, have
a high temporal and spatial resolution, have long enough forecast horizon, be
issued several times each day, give wind forecasts at roughly the hub height
of modern WTs and have several members (if probabilistic forecasts are de-
sirable). As an overall judgement, GEFS was considered the most appropriate
choice.

9 http://apps.ecmwf.int/datasets/data/tigge/levtype=sfc/type=cf/ Accessed:
2016-04-14.

29
3. Theory and Methods

In this chapter, a theoretical background for the papers included in the thesis
is given. Most of the methods used in the papers are thoroughly described
in textbooks on e.g. wind engineering, statistics and signal processing and
references are given for further reading. Since Matlab has been the main tool
for the modelling, references are given to both built-in functions and code
available on Matlab’s file exchange area. For readers familiar with modelling
of wind power, parts of this chapter may seem trivial and can be omitted.
Based on feedback on our work, I nevertheless believe that it can be useful
to introduce the theoretical background and methods by starting from a more
basic level.

3.1 Wind speed

Large-scale wind patterns are driven by pressure differences and the rotation of
the earth (the Coriolis effect). On the meso- and microscale, phenomena such
as sea breeze, terrain etc. also affects the wind flow. In more complex terrain,
it can therefore be large differences in wind speeds on the scale of hundreds of
meters. In the surface layer (up to about 100 m above ground [76]), the wind
speed generally increases with height, i.e. there is a positive wind shear. The
shear depends on surface roughness and orography. In and near forests, the
wind speed is typically low near the ground, but increases sharply with height.
The shear also varies in time depending on the atmospheric stability and other
factors. The wind profile is sometimes described by the empirical power law:

z−d α
 
u(z) = ure f , (3.1)
zre f − d
where u is the wind speed, z is the height above ground, subscript ref denotes
a reference height where a measurement is available, d is the (zero-plane)
displacement height and α is the shear exponent. The displacement height
describes the elevation of the zero level of the wind in and near cities, forest
and other vegetation and is often around 0.6–1.0 of the canopy height. In
practice, the wind speed never reaches zero, but the profile at heights relevant
for wind power is better described if the displacement height is taken into
account. Wind speeds at a particular site are often considered to follow a
Weibull distribution,

30
 0.2
Wind speed

Probability or Energy share

Energy production
0.15 Weibull fit,
c=7.9, k=2.3

0.1

0.05

0
0 5 10 15 20
Wind speed [m/s]

Figure 3.1. Histogram and Weibull fit of wind speeds for a measurement taken 100
m above ground (mean wind speed is 7.0 m/s). Energy production shares (see Sec-
tion 3.2), calculated for Vestas V90 2MW, are shown in grey.

k  u k−1 −(u/A)k
p(u; k, A) = ·e , (3.2)
A A

where p is the probability density function of wind speed u, given the Weibull
scale factor A and shape factor k. The Weibull parameters A and k are linked to
the mean wind speed ū; given two of these parameters, the third can be calcu-
lated. If nothing more than mean wind speed is known, a Rayleigh distribution
is commonly assumed, i.e. a Weibull distribution with shape factor equal to
2. One must though be careful assuming Weibull/Rayleigh distributed wind;
some sites exhibit two peaks in the distribution and an assumed Rayleigh dis-
tribution can give severe errors in energy calculations. An example of the dis-
tribution of measured wind speeds and the corresponding Weibull fit is shown
in Fig. 3.1.
The temporal variation in the wind speed is significant. Seasonal and diur-
nal patterns are present in many regions. In Sweden, the generation is in aver-
age 74% higher in November–March than in June–August, see Section 4.2.1.
An example of wind speed variations for a measurement taken 96 m above
ground is shown in Fig. 3.2.
In order to determine the wind conditions for a potential wind farm, on-site
wind measurements are normally carried out during a one- or two-year period.
Because of the variability of wind speed and direction, a long-term correction
(LTC) of the measurement is necessary. Commonly, this is done with the
“MCP” (Measure Correlate Predict) method [69]. A simple LTC method is
described in Section 3.2.

31
Figure 3.2. A few weeks wind speed measurements (10-minute resolution, 96 m above
ground), showing the significant temporal variability.

3.2 Wind power

The available power Pw in the air mass flow is a function of the wind speed u,
air density ρ and area A perpendicular to the wind direction:

1
Pw = ρAu3 . (3.3)
2

It has been shown that the amount of wind energy convertible to mechanical
energy has a theoretical upper limit of 16/27 ≈ 59% [77]. A real WT has
a lower aerodynamic efficiency and by including mechanical and electrical
losses the overall efficiency is often peaking at 45–50% for large-scale modern
WTs. Because of component and grid connection costs, there is a trade-off in
the rating of the generator. For a given turbine area, a generator with a higher
rating will increase the electricity production but also the costs. The optimum
ratio of rated power to rotor area (specific rating) depends on the site-specific
wind conditions. The resulting power to wind speed relation is known as the
power curve (PC) and can readily be obtained for any commercial WT, at least
for standard air density (1.225 kg/m3 ). According to the IEC standard [78],
the PC is given as a function of 10-minute mean wind speeds. Some examples
of normalized PCs are shown in Fig. 3.3. Three wind speeds of importance
for the WT are the cut-in (where the WT begins to operate, often 3–4 m/s), the
rated (the lowest wind speed with full output, often 11–14 m/s) and cut-out
(where the turbine is shut down for protection, often 25 m/s). As can be seen
in the example in Fig. 3.1, the largest share of the energy is often produced at
relatively high, but less probable, wind speeds.
In reality, there are several factors influencing the power output of a WT.
The PC accounts for aerodynamic and electrical losses in the WT. Additional
reductions in the power occur due to aerodynamic degradation from dirt and

32
 1

Normalized Power [p.u.]

0.8

0.6 Medium wind WT,

ρ = 1.225 kg/m 3
0.4 High wind WT,
ρ = 1.225 kg/m 3
Medium wind WT,
0.2
ρ = 1.1 kg/m3

0
0 5 10 15 20 25 30
Wind speed [m/s]

Figure 3.3. Three selected power curves for illustrating the impacts from specific
rating and air density. “High wind” windturbine (WTs) have higher specific ratings
and thus requires higher wind speeds to reach the rated power.

icing, hysteresis losses1 , grid and transformer losses and stops due to mainte-
nance, failures or icing. Since the certification of the PC is performed under
certain standard conditions, operation of the WT in a different wind regime
might change the PC. As an example, a more turbulent flow gives higher power
at low wind speeds but lower power near rated wind speed.
As described in Section 3.1, a LTC is routinely used to determine the long-
term wind conditions at a site before the investment decision for a farm is
taken. A basic LTC method called wind index [79] was used to determine
the long-term corrected yearly averages of wind generation in Sweden, see
Section 4.2.1. For each farm, a 35-year generation time series was computed
from MERRA data which was first linearly scaled so that the average gener-
ation equalled that specified by the farm owner. The wind index (WI) for a
farm i and year y was subsequently computed as

Pi,y
WIi,y = , (3.4)
Pi

where Pi is the average power for the 35-year period and Pi,y the average
power for year y. The national WI was subsequently calculated as the capacity-
weighted average of the WIs of all farms in operation during the year in ques-
tion. By dividing the measured annual generation by the corresponding wind
index, the “normal year” generation is obtained.

1 After
a high wind stop the WT does not start until the wind speed falls below a certain value
(which is lower than the cut-out wind speed).

33
3.2.1 Modelling
Most papers included in this thesis rely heavily on the idea of modelling ag-
gregated wind power generation from reanalysis data. The basic model is
presented in Paper I and extensions in Papers II and III.
Parameters were tuned to give good agreement between model output and
measurements on the system scale. As the following example illustrates, this
approach can have a strong impact on optimal parameter values. For a sin-
gle farm, the output varies between zero and (very close to) rated capacity. If
parameters were to be tuned to measurements for separate farms, the model
would produce time series with a similar range. However, when such time
series are combined, one would get nation-wide generation with almost the
same range, which is clearly not correct; the yearly maximum of generation
for the whole of Sweden is often around 80% of the installed capacity. One
explanation to this phenomenon is that out of thousands of WTs in a country,
some will always be out of operation. Even more important is the underesti-
mation of wind speed diversity by coarse NWP models; we simply never have
situations were all WTs simultaneously experience wind speeds above rated.
Somewhat simplified, the methodology involves the steps in the numbered list
below. The first six steps are described in detail in Paper I while step 7 is
described in Paper II and step 8 in Paper III.
1. A database of WTs is created, containing e.g. coordinates, rated capaci-
ties and rotor diameters.
2. Time series of wind speed and direction for each WT are bilinearly in-
terpolated2 from MERRA reanalysis data.
3. The wind speed time series are scaled in order to get annual generation
in agreement with that specified by the farm owner. As described in
Section 2.2.1 this is, in my opinion, a better method than calculating the
wind speed directly from reanalysis wind speeds at different heights.
4. Hourly generation time series are computed for each WT, taking into
account farm specific PCs, losses of different kind, PC smoothing (see
Section 3.2.4), wind direction etc.
5. The output is aggregated for the studied area and subsequently multi-
plied with a variable less than one (“external” losses, see Paper I).
6. Corrections are performed, accounting for seasonal and diurnal bias as
well as the fact that the annual generation predicted by turbine owners
was systematically too high in the 90’s and early 00’s.
7. Fluctuations in the high-frequency range, f > (10h)−1 , were underesti-
mated by the model described in Paper I (points 1–6 above), see expla-
nation below. This leads to an underestimation of step changes, which
is not a good thing for power system studies. Thus a statistical model
2 Vector summation is employed in order to get correct interpolation of wind direction (the mean

of e.g. 359◦ and 1◦ should be 0◦ rather than 180◦ ).

34
Figure 3.4. Time series of measurements and model outputs for a farm, a region and
the whole of Sweden. It is clear that the two latter can adequately be modelled with
relatively coarse reanalysis data, but for a single farm the results are not satisfactory.

was developed in order to increase these fluctuations, see Paper II and

Section 4.1.2.
8. A statistical model, simulating intra-hourly variations, was developed
in Paper III. A short methodology description is given in Section 4.1.3.
This step is only executed if data with a higher resolution than hourly is
necessary.
When NWP reanalyses are used as input for generating wind power time
series, the high-frequency fluctuations are often not properly represented. For
one WT, the magnitudes of these fluctuations are often underestimated. An
explanation to this effect is the limited spatial and temporal resolution of the
NWP models; local features causing variations in the wind speed are not cap-
tured. A model with a certain resolution is only able to correctly represent
phenomena on significantly larger scales, see [33] for a more in-depth discus-
sion. The “effective” resolution of NWP models is often in the order of seven
times the nominal resolution. When the outputs from many farms in a power
system are combined, the results can change; in [45] it was found that the
variability in terms of step changes was severely overestimated by NWP mod-
els. This could be explained e.g. by too high correlations between the time
series for different farms. In most studies, the high-frequency variability is
nevertheless underestimated also for the aggregated generation, see Paper II.
A legitimate question is whether a relatively coarse reanalysis is appropri-
ate for modelling wind power. As the results in Section 4.1 show, the answer
to this question is yes. The following example however demonstrates that our
approach is only suitable for modelling generation aggregated over relatively
large areas, where errors for individual farms are cancelled out. Let us con-
sider the output from one farm, one region in Sweden (SE4, see Fig. 1.1 on
page 14) and the whole of Sweden. Fig. 3.4 clearly shows that the coarse NWP
model cannot resolve the local wind variations seen by a farm but works fine
for larger regions.

35
3.2.2 Scenarios
When modelling the output of future farms in a system, it is necessary to de-
velop scenarios of deployment, both in terms of geographical locations and
of technical specifications of the farms. Several scenarios of future Swedish
farms were developed in Paper D. As discussed in Section 2.2.1, coarse reanal-
yses are neither suitable for determining likely sites for deployment of wind
farms nor for estimating mean wind speeds. Even a more highly resolved
NWP can be misleading for determining likely sites since many other factors
than the mean wind speed have a strong influence. Distance to a suitable grid-
connection point, quality of the infrastructure, the risk of icing, conflicting
interests (natural values, the military, aviation, telecommunications etc.) and,
not the least, the attitude of the local municipality, both neighbours and the
local government, are often more important than the wind resource itself. In-
stead of trying to model all these factors, we started from a database of present
and planned farms, see Fig. 4.15 on page 83. Since the cost for obtaining
permits, measuring and analysing the wind etc. is generally several hundred
thousand euros for each farm, it is reasonable to assume that such database
gives a good estimate of the actual distribution of future WTs. The farms in
the database were assigned probability weights depending on e.g. the status of
the permitting process, mean wind speed according to the MIUU model (see
Section 2.2.2) and whether the farm is located in a designated area for wind
power or not. The technical characteristics of the farms, most importantly CFs,
were estimated from MIUU wind speeds, historical trends and interviews with
persons in the wind energy community.
In Papers VI and VII, scenarios of deployment in the Nordic synchronous
power system and Europe were developed. Because of the larger scopes of
these studies and since the authors do not have the same in-depth knowledge of
wind power in these countries, simpler methodologies were used. Still though,
the geographical locations were not based on coarse reanalysis wind speeds,
but current and planned farms and/or mesoscale wind maps. Also, the mean
wind speeds were not taken directly from reanalyses, but were rather computed
from the anticipated energy yields and WT characteristics.
The appearance of the aggregated generation profile of a region or country
naturally depends on both geographical distribution and technical specifica-
tions of the farms. In Paper D it was found that the dispersion of farms and
average CF were the two most influential factors and consequently needed
most attention for the development of scenarios. The CF for a farm is a func-
tion of e.g. mean wind speed, specific rating, wake losses and efficiency of
the WTs. The mean wind speed, in its turn, is determined by the wind condi-
tions at the site and the hub height. From a variability point of view, it is of
relatively little importance how the average CF is achieved. As an example,
farms with relatively low wind speeds and large rotor areas give more or less
the same results as farms with higher wind speeds and smaller rotors.

36
3.2.3 Quantifying variability
The output from a WT varies on all time scales, from sub-seconds to decades.
As mentioned in the introduction, the variability can be challenging for the
power system since the fluctuations need to be balanced in one way or an-
other. In order to quantify and visualise the variability, many different metrics
and figures have been used in the literature. In Paper C, a review on variabil-
ity metrics, these were classified as “Distribution”, “Change” or “Temporal
Characteristics”. The same categorisation is used in the following exposition,
where some of the metrics are presented and, when considered necessary, ex-
plained.
The distribution metrics naturally describe the distribution of the time se-
ries and are independent of observation order. The mean generation in its
normalised form is often denoted capacity or load factor. A 3 MW turbine that
produces in average 1.1 MW thus has a CF of 1.1/3 ≈ 0.37. The mean gen-
eration can also be expressed as equivalent full-load hours (or only full-load
hours): the number of hours in a year that the turbine would have to run at its
rated capacity in order to produce its annual energy. The abovementioned WT
thus generates 0.37 · 8760 ≈ 3200 full-load hours, which is typical for a good
onshore project in Sweden as of today.
The minimum generation for a single WT or farm is zero, but also for a
relatively large region (e.g. the Nordic synchronous area) the minimum is often
very close to zero. I therefore believe that the minimum generation level is not
a very informative metric; it is preferable to study low generation in terms
of quantiles/percentiles3 , perhaps conditioned on the electric load. Maximum
generation, in particular for larger regions, can however be useful to analyse.
The firm capacity is defined as the output exceeded a certain percentage p of
the time (e.g. 90%), and is thus the same as the 100 − p percentile.
A duration curve is a plot of all data sorted from highest to lowest and can
often be an important tool for studying variability. Of particular relevance is
the study of net load duration curves for different scenarios of wind deploy-
ment. An empirical cumulative distribution functions (ECDF) can be attained
by a 90 degrees clockwise rotation of the duration curve. These plots thus
hold the same information and it is only a matter of taste which to use. In
most wind studies, duration curves are the preferred option. The sample stan-
dard deviation (SD) or variance can quantify the variability of the generation
with one single number. Probability density function estimates in the form of
histograms give more information.
Of equal or higher importance than the analysis of the distribution of gener-
ation is that of changes in generation, which can be measured as step changes
or ramp rates. The former is defined as change of averaged power over a fixed
number of steps, e.g. change in hourly energy over four hours. The definition
3 Seehttp://se.mathworks.com/help/stats/quantiles-and-percentiles.html for
how these metrics are computed in Matlab.

37
of ramp rates, however, differs between authors. As put in [80] “depending on
the defining criteria of a ramp event, the starting and ending point and dura-
tion of a ramp can change substantially”. Some authors also define ramp rates
in the same manner as step changes [81]. In our papers, one and four hour step
changes in hourly generation are normally considered.
Most of the metrics used for quantifying the distribution of generation can
be directly transferred to the analysis of step changes. The SD of step changes
is possibly the most common of all variability metrics. Since the distribution
of step changes can be skewed and normally have fat tails, the higher moments
(skewness and kurtosis) are sometime given [37]. It can be mentioned that the
changes in power depend on the initial generation level. Sörensen et. al. [82],
among others, therefore plotted percentiles of ramp rates over initial power.
In Paper C, metrics in the “Temporal Characteristics” category, admittedly
a bit confusing, included:
• Metrics of temporal characteristics of the data (excluding the metrics of
change discussed earlier)
• Frequency domain metrics
The latter type of metrics is treated in Section 3.4. The autocorrelation func-
tion (ACF) is the linear correlation of a time series with itself, shifted in time.
If, for instance, the ACF is high for lag one, it implies that the generation is
not likely to change much from one sample point to the next. In some cases,
e.g. for wind power forecast errors, a high ACF is not desired since it implies
that the energy imbalance (actual - forecasted generation) can build up to sig-
nificant levels. The sample ACF at lag k of a stationary time series x1 , ..., xn
can be computed as
n
∑t=k+1 (xt − x̄)(xt−k − x̄))
ACFk = n . (3.5)
∑t=1 (xt − x̄)2
The volatility of wind power is not constant; sometimes the output is rela-
tively steady and sometimes it fluctuates heavily. A useful way of quantifying
the varying volatility is the moving variance, which is the variance of a sig-
nal in a moving window4 . The moving SD is defined analogously. When the
moving SD is applied on a filtered time series, only the fluctuations of certain
frequencies are considered. This metric, referred to in [34] as the “Variability
Index”, was used in Papers II and III.

3.2.4 Smoothing
When the outputs from several WTs are combined, be it in a farm or in a whole
synchronous power system, the resulting time series is smoother than that for
4A fast implementation in Matlab (movingvar) has been written by Aslak Grinsted, see
http://www.mathworks.com/matlabcentral/fileexchange/8252-moving-variance
Accessed: 2016-08-04.

38
 1
a) b) 1
One farm

Hourly generation [p.u.]

0.8 0.8 Three farms

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 2 4 6 8 10 12 14 0 0.2 0.4 0.6 0.8 1
Time [days] Fraction exceedance

Figure 3.5. Smoothing of output when combining several farms. (a) Two weeks
example of time series. The profile for three farms combined is smoother; ramp rates
are lower and the generation is more seldom very high or very low. (b) Duration
curves based on 19 months of measurements.

a single WT [6]. An example of outputs from one farm and three farms in
aggregation is given in Fig. 3.5. The three farms combined have lower ramp
rates and the generation is more seldom very high or very low.
If the correlations between generation time series are known, the reduction
in variability (in terms of variance) can be quantified. Let yt be a weighted
sum of n discrete time series xi,t
n
yt = ∑ wi xi,t . (3.6)
i=1
The variance of y can then be computed from the weights wi , the standard
deviation σi of each xi,t and the pair-wise, linear correlations ρi j between the
time series:
n n
Var(yt ) = ∑ ∑ wi w j σi σ j ρi j . (3.7)
i=1 j=1
As an example, let us consider the aggregation of two time series x1,t and x2,t ,
each with σ = 1 and w = 0.5. With ρ12 of -1, 0, and 1, the variance of the
combined outputs become 0, 0.5 and 1 respectively. Eq. (3.7) is, due to its low
computational cost, very powerful for optimising the w’s for lowest possible
variance of yt .
The smoothing effect makes it beneficial to connect many farms over a
wide area. The variability is reduced and hence integration of wind power be-
comes less of a challenge. As was demonstrated in Eq. (3.7), a low correlation
between outputs gives better smoothing. A commonly [36] used model for
describing the correlation between farms is given by

ρ(d) = α · e−d/D , (3.8)

39
where d is the distance between the farms and D is a parameter determining
how fast ρ decreases with distance. The α parameter is sometimes fixed to
unity and sometimes allowed to take other values as well. In Paper VI, the
correlation coefficients between wind generation in different countries in Eu-
rope were quantified and it was shown that Eq. (3.8) is appropriate also for
country-wise generation.
There is also smoothing within a farm or a smaller region. A “multi-
turbine” PC therefore gives a better representation of the output than the origi-
nal PC from the manufacturer. A methodology for calculating this is proposed
in [83]. In our database (see Section 2.1), detailed coordinates are not avail-
able for all WTs. Furthermore, no turbulence intensity data was available and
the temporal resolution of the wind speed time series was one hour instead of
the ten minutes which is the standard for PCs. Thus a simplified version of
the methodology given in [83] was employed, using only one smoothing pa-
rameter which was tuned to generation data for the whole of Sweden. Let us
assume there are several WTs of the same type distributed within one grid cell
in the meteorological model. At a certain hour, the mean wind speed is 6 m/s
which, according to the manufacturers PC, translates to a generation of 0.19
p.u. However, some WTs will experience lower or higher mean wind speeds
than 6 m/s and there will also be differences within the hour for each WT.
We therefore assume, as in [83], that the 10-minute wind speeds for all WTs
are normally distributed around 6 m/s. The smoothing parameter was defined
as the standard deviation of this distribution. In our example, a smoothing
parameter of 2 m/s results in an average generation of 0.24 p.u for the consid-
ered hour. An example of an original PC and its multi-turbine counterpart is
shown in Fig. 3.6. For low and above cut-out wind speeds, the output is higher
for the multi-turbine PC. For wind speeds between a few m/s lower than rated
and cut-out, the output is lower. In our implementation, the same smoothing
parameter was applied for all WTs and all wind conditions. Although this is a
quite rough simplification, a smoothing parameter of around 1.0 m/s improved
the model of aggregated Swedish wind generation considerably as compared
to using original PCs, see Fig. 11 in Paper I.

3.2.5 Forecasts
For secure operation of power systems and profitable operation of generators,
load and intermittent generation need to be predicted in advance. Forecasts for
different time horizons are of interest, e.g. short-term forecasting for intra-day
operation, day-ahead forecasts for energy trading and one-week forecasts for
hydropower planning. For more information, in particular on Sweden and the
Nordic system, see [84].
A real example of generation, day-ahead forecasts and forecast errors (FEs)
for a wind farm is given in Fig. 3.7. The forecast generally follows the gen-

40
 1

Generation [p.u.]
0.8

0.6

0.4

0.2 Original PC
Multi-turbine PC
0
0 5 10 15 20 25 30 35
Wind speed [m/s]

Figure 3.6. Examples of original and multi-turbine power curves (PCs). The latter
was calculated with a smoothing parameter of 2 m/s (see Section 3.2.4).

1
Generation
Generation and FE [p.u.]

Forecast
FE
0.5

-0.5
0 1 2 3 4 5 6 7 8
Time [days]

Figure 3.7. Example of actual generation, day-ahead point forecasts and forecast er-
rors (FE) for a farm in Sweden. Note the phase error at day seven, leading to first a
large negative error and then a large positive error.

eration quite well, but for some hours the mismatch is large. Two main types
of FEs exist: level and phase errors [85]. The latter (misjudgements of when
the generation ramps up or down) are short-lived but can be very large. The
FEs are autocorrelated; if the error is positive one hour it is likely that it is also
positive the following hours.
Traditionally, research focus has been on improving the performance of
point forecasts [86] but recently a lot of attention has been drawn to prob-
abilistic forecasts, providing the user with prediction intervals. Commonly,
probabilistic forecasts are obtained from ensemble NWP models where sev-
eral members are produced by perturbing the initial conditions. Different
NWP models can also be combined, both to improve point forecasts and to
get information on prediction intervals [87].
An example of an ensemble forecast from the GEFS dataset is shown in
Fig. 3.8. For this particular forecast, there are relatively small differences

41
 15
MERRA
Control member
Ensemble mean
Perturbed members
Wind speed [m/s] 10

0
0 1 2 3 4 5 6 7 8
Forecast horizon [days]

Figure 3.8. Example of an ensemble forecast from GEFS. The ten perturbed members
are obtained by perturbing the initial conditions of the control forecast.6

between the ensemble members for the first three days, but for longer horizons
the spread is larger. The control member, shown in blue, is the forecast for
unperturbed initial conditions.
For low penetrations of wind power, the uncertainty in net load is dominated
by load uncertainty. Wind power FEs however become increasingly important
for higher wind penetration levels and thus need to be accounted for in wind
integration studies. According to IEA wind [18], several years of “synthetic”
forecasts are preferable for simulation of unit commitment and economic dis-
patch including reserve requirements. As for generation time series, synthetic
forecasts can stem from (linear scaling of) historical data, NWP models, statis-
tical models or hybrids of the latter two. A review of wind integration studies
published between 2007 and 2016 (see Paper IV), revealed that NWP based
forecasts were rarely used but purely statistical approaches were common.
Actual FEs have several characteristics, some of which make them chal-
lenging to simulate with purely statistical methods: the errors increase with
terrain complexity [88], depend on the wind speed [89] and are often skewed
and heavy-tailed [90]. Furthermore, there exist level and phase errors and
consecutive forecasts are not independent7 .
Despite the challenges, many attempts have been made to simulate realistic
wind power forecasts using statistical methods. After some initial experiment-
ing, we decided to use a different strategy: to base the synthetic forecasts on
GEFS reforecasts (see Section 2.2.3) and some statistical fine-tuning. In or-
6 Note the butterfly effect; small changes in the initial conditions of a chaotic system eventually
leads to large differences. This is why weather is so difficult to predict for longer time horizons
(for all but Laplace’s Demon).
7 If, for example, the five day-ahead (D+5) forecast issued a certain day is too conservative

(forecast < generation), then it is likely that the D+4 forecast issued the following day is also
too conservative.

42
der to retain the benefits of NWP based forecasts (from a modelling perspec-
tive), e.g. the occurrence of phase errors and dependencies between consecu-
tive forecasts, we strived to keep the statistical post-processing to a minimum.
In short, the method involves the following steps, also illustrated in Fig. 3.9:
1. The “true” generation is calculated from reanalysis data with the method-
ology described in Section 3.2.1.
2. “Raw” wind speed forecasts for day 0 through 7 are interpolated from
GEFS reforecasts.
3. All or most of the bias in the forecasted mean wind speed is removed.
4. Some multivariate ARMA noise, described in Section 3.3.4, is added to
the forecast in order to decrease the correlation between FEs for different
farms and obtain an appropriate spectrum.
5. A fixed part of the errors for each hour is removed.
6. Generation forecasts are computed from wind speed forecasts.
7. A machine learning model (Gradient boosting, see Section 3.6.3) is op-
tionally used in order to improve the generation forecasts.
For trimming the model, measurements/forecasts from the whole of Belgium
and individual farms in Sweden were used. FE correlations were also com-
pared to results in the literature. Although no probabilistic forecasts were gen-
erated, this could be done in a straight-forward manner by using the perturbed
members of GEFS.

3.3 Statistical methods

This section comprises five subsections. First, different methods for optimi-
sation and the partitioning of data into training, validation and test sets are
reviewed. Secondly, the transformation between different distributions is de-
scribed briefly. The three last subsections are devoted to splines, ARMA mod-
els and empirical orthogonal functions respectively.

3.3.1 Optimisation
In the papers included in this thesis, several different optimisation methods
were employed for finding suitable parameter sets. In general, the main ob-
jective was not to infer “true” parameter values or find causal relationships,
but rather to build models suitable for prediction. As discussed in [91], there
are often multiple models (parameter sets) having similar performance in this
sense.
In most models there are two types of parameters that need to be optimised:
hyperparameters and model parameters. The former determines the structure
of the model and are set beforehand. As an example, in classification and

43
 a)
Wind speed [m/s]

MERRA
20 GEFS raw

10

0
b)
Wind speed [m/s]

20 Bias reduced

10

0
c)
Wind speed [m/s]

ARMA noise added

20

10

0
d)
Wind speed [m/s]

20 Errors reduced

10

0
1
Power [p.u.]

0.5

e)
0
0 1 2 3 4 5 6
Time [days]

Figure 3.9. Illustration of the method for generating synthetic forecasts. The steps
referred to in the following are described in the numbered list on page 43. (a) Step
1–2, (b) Step 3, (c) Step 4, (d) Step 5, (e) Step 6.

44
regression trees, the maximum number of splits is a hyperparameter set by
the user and the actual splits are determined by the data when the model is
trained. Different algorithms are often employed for optimising the two types
of parameters. E.g. a grid of hyperparameters can be searched and for each
set of hyperparameters, the splits that minimise a loss function are chosen. In
most of my papers, the function that should be minimised is denoted objective
function. An objective function can be of the type reward/utility function (if it
should be maximised) or loss function (if it should be minimised).
When training and evaluating the performance of a model, it is crucial to not
use the same data for these two purposes. If many different sets of hyperpa-
rameters are considered, it is useful to partition the data into three independent
sets:
1. Training data for determining optimal model parameters given certain
hyperparameters.
2. Validation data for determining optimal hyperparameters. For each set
of hyperparameters, a model is trained using the training data and the
performance is assessed with the validation data.
3. Test data to evaluate the performance of the chosen model.
We can illustrate the concepts of optimisation with a simple example. Let

Y (X) = 40 + 0.8X − 0.01X 2 + 0.00002X 3 + ε,

X ∼ U (0, 500) ε ∼ N (0, 1002 ), (3.9)

where U (0, 500) is the continous uniform distribution (equal probability in

the range 0–500 and zero probability outside this range), be an unknown func-
tion we try to fit with polynomials

Pn (X) = a0 + a1 X + a2 X 2 + ... + an X n . (3.10)

In this case n is a hyperparameter that is determined beforehand. All values of

n between one and nine were considered. Three datasets were generated with
N = 100 sample points each: training, validation and test sets. The model
parameters (coefficients) a0 , ..., an were obtained by minimising the squared
residuals of the training data. Fig. 3.10 show some important characteristics
of the optimisation:
• The data is somewhat scattered. In reality this can be due to e.g. mea-
surement uncertainty and neglected predictors.
• Model performance is generally best for the training data.
• For the training data, a more complex model (higher n is this case) gives
a better fit. For n ≥ N − 1, the polynomial would pass through all data
points.

45
 150
a) b)
600 Training data
140 Training set
Underlying distribution
Validation set
400 P2 (X) 130 Test set

P9 (X)
120

RMSE
200
Y

110
0

100
-200

90
-400
0 100 200 300 400 500 2 4 6 8
X n
Figure 3.10. (a) Training dataset, underlying distribution Y (X) = 40 + 0.8X −
0.01X 2 + 0.00002X 3 and two examples of polynomial fits. (b) RMS error for the
different sets depending on polynomial order n. The best polynomial orders for each
set are indicated with circles.

• Due to overfitting, a too complex model will give poorer results for the
validation and test data. The n ≥ N − 1 polynomial mentioned in the pre-
vious point would most likely give terribly bad results for the validation
and test sets.
• If the same dataset is used for determining optimal values of the hy-
perparameters and assessing model performance, overoptimistic results
may be achieved. In our example, n = 5 is optimal for the validation set
which results in a higher RMS error for the test set as compared to n = 3.
The simplest method for optimisation is probably the search over a regular
grid. For each parameter, allowed values are specified and subsequently the
objective function for all possible combinations of parameters is evaluated.
This method was employed e.g. in Paper II, where a coarse grid of a few
hyperparameters was evaluated. For high-dimensional problems and/or when
a fine resolution near local optima is desired, this method quickly leads to
impractically long computational times.
In Paper I, a random restart hill-climb algorithm was used for optimising
the model of hourly wind generation in Sweden. This algorithm starts with a
random set of parameter values and the performance of the model is evaluated.
Subsequently, a randomly selected parameter is increased or decreased. If this
leads to a better model, the change is accepted and a new randomly selected
parameter is altered. In the beginning of the search, relatively large changes of
the parameter values were used but subsequently smaller steps in the parameter
space were taken. When no further improvement is possible, the parameter set
is saved and a new hill-climb with a random start point is undertaken. In
Paper I, the model with the lowest objective function for the training set was

46
chosen. In hindsight, this led to a slightly overfitted model; somewhat better
results could have been attained if a validation set had also been used.
Matlab’s built-in fminsearch, utilised in several of our papers, employs the
Nelder-Mead simplex algorithm [92, 93] to find the minimum of a function.
A simplex in n dimensions has n + 1 vertices, e.g. a triangle in two dimen-
sions. The function is evaluated at the vertices. New, better simplexes are con-
structed by replacing one of the vertices, e.g. by reflection of the worst point
through the centroid of the remaining points. This unconstrained, derivative-
free algorithm often requires relatively few function evaluations, although it
can sometimes fail to find the true minimum [94]. A transformation can be
used to convert a bound constrained problem into an unconstrained problem.
This has been implemented in the fminsearchbnd script8 .
If the variance of a combined time series is to be minimised, e.g. optimising
the allocation of wind power capacity in several countries in order to get a
smooth output, fminsearchbnd with Eq. (3.7) as loss function can be used for
very fast computations (a few seconds). Naturally, zero is set as the lower
boundary for each weight, but additional constraints may also be needed. In
Papers VI and VII, the variance was e.g. minimised while keeping the total
energy production constant.
In two of the papers, phase angles of the frequency domain representation
of signals were optimised. The algorithm for finding appropriate phases is
described in Section 3.4.2.

3.3.2 Transformation between distributions

Transformations between distributions were used in Paper III for obtaining
normally distributed, stationary time series. The transformations were made
either from a t location-scale distribution (see Paper III) fitted to data to the
normal distribution. Since it is not possible to express the cumulative distri-
bution functions of these distributions in terms of elementary functions, the
transformations could not be written as closed-form expressions. Instead, the
Matlab cdf and icdf commands were used. For fitting a distribution to data at
hand, the Matlab fitdist function was used. An example of the fitting and trans-
formation procedure is shown in Fig. 3.11. First a t location-scale distribution
is fitted to the 1000 sample points. The dashed line illustrates how a sample
point is then transformed to the normal distribution (−0.005 → −1.18).

3.3.3 Splines
It is often desirable to smoothen data in one way or another. An option to low-
pass filtering (described in Section 3.5) is to use splines; piecewise polynomi-
8 Script
written by John D’Errico, http://www.mathworks.com/matlabcentral/
fileexchange/8277-fminsearchbnd--fminsearchcon Accessed: 2016-08-04.

47
 z (Normal distribution)
-3 -2 -1 0 1 2 3
1 1

Example:
0.8 0.8
-0.005 is
transformed
to -1.18 N(0,1)
0.6 0.6
F(z)

0.4 0.4
Fitted t location scale

0.2 Empirical 0.2

0 0
-0.02 -0.015 -0.01 -0.005 0 0.005 0.01 0.015 0.02
z (empirical and t distribution)

Figure 3.11. Illustration of the transformation from a fitted t location-scale to a normal

distribution. The vertical axis gives the cumulative distribution functions F(z). Note
that two different horizontal axes are used and that the t location-scale distribution
gives a good fit to the empirical data.

als with a continuous derivative. In Paper III, wind power fluctuations with
temporal resolution of 5–15 minutes were simulated and added to smoothened
hourly generation. Because of the limitation in article length, the smoothing
splines used, both Matlab’s built-in cubic splines and our own fourth order
splines, were described relatively briefly. More details are thus given in the
following paragraphs. Let us first define the following variables:

fs Sampling frequency (samples per hour)

Pt Mean generation with frequency fs
Ph Mean hourly generation
Ct Cubic spline fit of Ph , evaluated at resolution fs
xt High-frequency component (Pt − Ct for the training data)

Let also additional subscripts sim0, sim and obj denote preliminary simulated
time series, simulated time series and the objective power system respectively.
For generating Ct,obj from Ph,obj , Matlab fit command with fitType set to cu-
bicinterp was used. The algorithm fits cubic polynomials between each pair of
data points. Cubic interpolation is a special case of smoothing splines where
the fitted function passes through all data points. If desired, smoothing splines
can be made smoother, at the expense of not passing through all data points,
by adding a penalty for the integral of the squared second order derivative of
the splines.
The preliminary simulated generation was achieved as Pt,sim0 = Ct,obj +
xt,sim0 . Due to the stochastic nature of xt,sim0 , an adjustment was necessary

48
in order to make the mean of Pt,sim0 (t ∈ h) the same as Ph,obj . Let ∆Ph denote
the difference in hourly means for hour h. Simply adding ∆Ph to Pt,sim0 would
sometimes result in large step changes at new hours. For the fs samples in
hour h, a fourth order spline adjustment Sh (r) was therefore used:

xt,sim (t ∈ h) = xt,sim0 (t ∈ h) + Sh (n), (3.11)

where

Sh (n) = ah · n4 + bh · n3 + ch · n2 + dh · n + eh (3.12)
and n is the vector (0 1 ... fs -1). The coefficients ah , ..., eh in Eq. (3.12) were
chosen to satisfy

∆P +∆P


 Sh (0) = h−12 h
 ∆Ph +∆Ph+1
 Sh ( f s ) =

2

Sh (n) = ∆Ph (3.13)
0 0

Sh (0) = Sh−1 ( fs )




 MinimiseR fs (S00 )2 dr

0 h

The coefficients ah , ..., eh are uniquely defined by the constraints in Eq. (3.13).
The algebraic expressions for the coefficients can be derived in a straight-
forward fashion but are quite long, so they are not given here. Pt,sim was
subsequently calculated as

Pt,sim = Ct,obj + xt,sim . (3.14)

In an earlier version of the adjustment function, cubic splines were used and
the last condition in Eq. (3.13) was not included. This resulted in oscillations
with a one-hour period, clearly observable as a peak in the power spectral
density (see Section 3.4.1) of Pt,sim . Examples of the difference in hourly mean
generation ∆Ph (t) as well as the spline adjustment Sh (t) and the resulting Pt,sim
are shown in Fig. 3.12.

3.3.4 Multivariate ARMA models

In Papers II-III, IFFT method was utilized to generate noise with a desired
frequency domain spectrum, see Section 3.4.2. This is a powerful method in
that it allows synthesis of time series with an arbitrary spectrum. In Paper IV,
we desired noise with certain correlations between different sites. An Autore-
gressive Moving Average (ARMA) model was thus deemed more appropriate
since the correlations can then be controlled.
An ARMA process with AR order p and MA order q is notated ARMA(p,q)
and satisfies the equation

49
 ×10 -3
10
a) b)
∆P h (t) P t,sim0
0.75
S h (t) P t,sim
5
P h,obj
Power [p.u.]

Power [p.u.]
0.7

0 0.65

0.6
-5

0.55
0 1 2 3 4 5 6 0 1 2 3 4
Time [hour] Time [hour]

Figure 3.12. (a) Spline adjustment for correcting differences in mean hourly genera-
tion (∆Ph ). If ∆Ph would be added directly, abrupt step changes at new hours would
occur. (b) Resulting time series Pt,sim after adjustment. Note that the same hours are
not shown in the two sub-figures.

Yt = φ1Yt−1 + φ2Yt−2 + ... + φ pYt−p + εt + θ1 εt−1 + θ2 εt−2 + ... + θq εt−q ,

(3.15)
where φi are AR coefficients, θ j are MA coefficients and ε is Gaussian noise
with mean zero and variance σ 2 . In order to generate correlated Yt time series,
multivariate ε noise were generated with the Matlab mvnrnd function. The
entry for row n and column m in the covariance matrix, necessary as input to
mvnrnd, was computed as

cnm = σ 2 e−dnm /D , (3.16)

where dnm is the distance between wind farms n and m and D is a decay factor.
This technique resembles that used in e.g. [95]. As an example, let D = 50 km,
σ = 2 m/s and d1,2 = 17 km. This gives c1,2 = 2.85 and the ε time series for
farms 1 and 2 will thus have a correlation coefficient close to 2.85/22 = 0.71.
This correlation is preserved in the ARMA filtering. The generation of noise
from the ARMA model was implemented with a Matlab filter using a rational
transfer function determined by the ARMA parameters, which in turn were
fitted to give an appropriate spectrum.

3.3.5 Empirical orthogonal functions

Empirical Orthogonal Functions (EOFs) are described relatively informally in
this section, see e.g. [96, 97] for a more rigorous treatment. Meteorological
variables, e.g. temperature, wind speed and pressure, are normally available
as time series for each point in a grid. The detrended (mean removed) data for
a variable can be ordered in a matrix X:

50
  
x1 (1) x2 (1) · · · xm (1)
x1 (2) x2 (2) · · · xm (2)
X =  .. ..  , (3.17)
 
.. ..
 . . . . 
x1 (n) x2 (n) · · · xm (n)

with n time steps and m grid points. For a highly resolved model and/or large
geographical area, m quickly becomes large and it may be inappropriate or
impossible to use all m time series as regressors in a model. Fortunately, an
EOF decomposition allow us to reduce the dimensionality of the problem sig-
nificantly with very little loss of information. Informally, one finds the domi-
nating spatial patterns of the variable (the EOFs) and the corresponding EOF
time series. Often, most of the variance of the variable van be explained by
just a handful EOFs. Mathematically, a decomposition of a real-valued X into
three matrices is performed with singular value decomposition:

X = U · D · VT (3.18)

where V contains the EOFs and U · D are the EOF time series. In Papers II and
IV, EOFs were computed in Matlab using the pcatool9 . Note that different
terminology is sometimes used in the literature; EOFs can e.g. be denoted
principal components and the EOF time series denoted principal component
amplitudes. Apart from reducing the dimensionality, EOFs can also reveal
important characteristics of a system in terms of modes of variability.
We can illustrate the use of EOFs by looking at wind speeds from the
MERRA reanalysis dataset for Sweden (latitude 54−70◦ and longitude 10.0−
25.3◦ ). The horizontal resolution is 1/2◦ in latitude and 2/3◦ in longitude, so
m = 33 × 24 = 792. Fig. 3.13 shows the first three EOFs10 . In Fig. 3.14, the
actual wind speeds for an hour is compared to the sum of mean wind speeds
and the first nine EOFs for the same hour. As can be seen, the patterns are very
similar. For the whole dataset, the first nine EOFs explain 90% of the variance
and the first 20 EOFs explain 99%. The 792 original time series can thus, for
regression purposes, be replaced by e.g. nine or twenty time series. Note that
the number of EOFs needed can differ significantly; for some variables used
in Paper II, the first EOF alone explained over 90% of the variance and for
some, 14–15 EOFs were needed. Also note that in an EOF analysis, one has
to determine the geographic region to include, see e.g. [35].

9 Script written by Guillaume Maze, available at http://www.mathworks.com/

matlabcentral/fileexchange/17915-pcatool Accessed: 2016-08-04.
10 Note that each EOF corresponds to a column in V. For presentation, each column vector was

first reshaped to the native grid of MERRA.

51
Figure 3.13. The first three EOFs of wind speed. EOF number one is strictly negative
and strongly negatively correlated to the mean wind speed (-0.94). A positive contri-
bution of this EOF thus gives lower and more even wind speeds. The second and third
EOFs imply certain patterns predominantly in the offshore winds.

Figure 3.14. Illustration of the use of EOFs. (a) MERRA wind speeds for a certain
hour and (b) sum of mean wind speeds and the first nine EOFs for the same hour.
These EOFs explain 90% of the variance. For regression purposes, the 792 original
time series, one for each grid cell in (a), can be replaced with only e.g. nine.

52
3.4 Frequency domain methods
A discrete signal xt in the time domain can be decomposed into its frequency
domain components Xk by the fast Fourier transform (FFT). Each element in
Xk can expressed in rectangular notation (ak + bk i) or in polar notation (mag-
nitude and phase). The inverse procedure (IFFT) synthesises a time domain
signal from the frequency domain components.

3.4.1 Power spectral density

In many signal processing applications, an important aspect of a FFT analysis
is to isolate and study the dominating frequencies in a noisy signal. For wind
power, the frequencies corresponding to one year and one day (and their har-
monics) are often prominent. Wind power generation is however, to a large
degree, stochastic and all frequencies are of interest. A useful method to anal-
yse the strength of fluctuations with different frequencies is power spectral
density (PSD) estimates, which are closely related to the squared magnitudes
of the frequency-domain representation of the signal. The simplest PSD esti-
mate comes from the periodogram and is computed as

1
PSDk = |Xk |2 (3.19)
fs · N
The periodogram is often very noisy. In order to reduce the variance of
the PSD estimate, several different options are available. The Welch’s PSD
estimate [98] split the signal in several overlapping segments. Each segment
is windowed; samples near the ends are reduced in amplitude (in the Matlab
pwelch function, the Hamming window is the default). This is done in order to
reduce spectral leakage, see e.g. [99]. The final PSD estimate is calculated as
the average of periodograms of the windowed segments. Two disadvantages
with segmenting the data are that the frequency resolution becomes poorer
and that information on very low frequencies is lost. If very long time se-
ries are available, it is possible to achieve both high resolution and low noise.
An option is to compute the Welch PSD for high frequencies using multiple
segments and the PSD for low frequencies with one or a few segments.
Log-log plots of the Welch PSD estimates for Swedish wind power genera-
tion as well as output from one farm are shown in Fig. 3.15a. Some character-
istics, typical for wind power PSDs, are worth mentioning:
• The PSD is generally lower for higher frequencies (shorter periods). The
contribution from relatively high frequencies to the variance of the signal
can however be significant, see Fig. 3.15b.
• The PSD for the whole of Sweden is considerably lower than that for one
farm when the generation is given in p.u., indicating that the combined
output is smoother than that for one farm.

53
 a) b) 0.03

Cumulative sum of PSD·∆f

10 0 0.025
PSD [(p.u.) 2 · h]
0.02

10 -2 0.015

0.01
Sweden
One farm 0.005
10 -4

0
1w 1d 2h 1y 1m 1w 1d 2h
Period Period

Figure 3.15. Power spectral density (PSD) estimates of hourly wind generation. (a)
Welch estimates with 257 sample segments for the whole of Sweden and one farm,
(b) cumulative sum of PSD · ∆ f computed with periodogram for the whole of Sweden.
The contributions sum up to the variance of the time series (0.029 p.u.), see Eq. (3.20).

• The PSD slope is steeper for the whole power system, see also [100,
101]. This is a result of higher frequency components of different farms
being less correlated than lower ditto and thus more effectively damp-
ened by aggregation, see Paper VI.
• There is a smaller peak corresponding to diurnal wind variations. Be-
cause of the non-sinusoidal diurnal patterns, harmonics with frequency
(24h/2)−1 , (24h/3)−1 , ... are also present.
In Fig. 3.15b the cumulative sum of PSD · ∆ f for the whole of Sweden
is shown (∆ f is the frequency increment). An advantage with this type of
plot is that noise in the PSD becomes less visible, thus a periodogram of the
whole time series can be used and information on lower frequencies becomes
available. If the mean value of a signal xt is subtracted before the PSD is
computed, as is the case for Fig. 3.15, we furthermore have

∑ PSDk (xt ) · ∆ f = Var(xt ). (3.20)

k

Fig. 3.15b can thus be interpreted as the contribution to the overall variance
of the generation from different frequency bands; a steep curve in a certain
region implies that this is important in terms of variability. Very little of the
variance can be explained by variations with periods shorter than 24h (around
3%). Almost half (47%) of the variance can be attributed to fluctuations with
periods between two days and two weeks while around 9% comes from the
yearly cycle.
PSD analyses were used in Papers II, III, IV and VII. Generally, the fre-
quencies were expressed in terms of periods since this is easier to interpret,
e.g. f = (24h)−1 instead of f ≈ 0.417h−1 . The Hz unit was seldom used since

54
this becomes unpractical for the time scales considered. As mentioned above,
it can be advantageous to remove the mean from the signal before computing
the PSD. If this is not done, the relation in Eq. (3.20) does not hold and the
PSD for the lowest frequencies can be distorted. The mean was not always
removed in the papers if it was not important for the analysis.

3.4.2 Synthesis of time series

In Papers II and III, time series of wind generation were constructed by com-
bining low-frequency components from meteorological models and stochastic
high-frequency noise. This was done either because no high-frequency data
were available (Paper III) or because the high-frequency fluctuations from the
models were too smooth. Other authors have used statistical methods to gener-
ate purely synthetic time series of generation and forecasts [31, 44, 101–104],
but as discussed in Section 1.3, we believe that it is beneficial to use physical
models as far as possible.
For the following exposition, let Pt denote the wind generation in discrete
time and xt the high-pass filtered Pt . Let also an additional subscript sim indi-
cates simulated data and no additional subscript indicates measurements. The
volatility of xt vary considerably in time. In Papers II and III, xt was therefore
transformed to an approximately stationary time series yt , see Sections 3.3.2
and 3.5.
When synthesising yt,sim , there are often several criteria that need to be ful-
filled e.g. appropriate distribution, autocorrelation, spectrum and step change
distribution. Furthermore, the varying volatility of xt should preferably be
captured. The FFT-IFFT approach provides a very flexible method that, to-
gether with e.g. stationarity transformations and predictions of volatility, can
meet these demands. The analysis and synthesis (FFT and IFFT) themselves
are not complicated, the challenging part is to find appropriate magnitudes and
phases of Ysim , the frequency domain representation of yt,sim . The magnitudes
of Ysim were interpolated from the magnitudes of Yt . In an early version of the
methodology (Paper D), uniform random phase angles were used to synthesize
yt,sim . This gives a correct PSD, but yt,sim becomes normally distributed, which
was not always desired. Furthermore, the moving SD of yt,sim fluctuated quite
a bit which led to a deterioration of the models’ ability to predict the varying
volatility of xt . Inspired by [101], we therefore used an algorithm to search for
appropriate phases.
Let τt denote the moving SD of yt,sim and στ the SD of τt . The loss function
L to be minimized was chosen as

L = στ + c · |κsim − κdesired |, (3.21)

where κ is the kurtosis and c is a constant. The latter term in Eq. (3.21) was
included in order to impose appropriate heaviness of the tails on yt,sim . In

55
Figure 3.16. The volatility of τt (the moving standard deviation of yt,sim ) is reduced by
using a phase angle search algorithm before synthesising yt,sim . Based on data from
Paper III.

order to minimize L, it was found that an “elitist” algorithm with small but
many generations was robust and efficient:
1. Start with ten uniform random phase angle sets (ranging from −π to π).
2. Synthesize time series (a generation with ten individuals) from the mag-
nitudes and phases using IFFT and calculate L.
3. Choose the angles corresponding to the individual with lowest L. Con-
struct ten new angle sets by adding random noise from a U (−a, a) dis-
tribution.
Step 2 and 3 were repeated 2000 times with the a parameter gradually re-
duced from π/10 to π/300. The reduction in the volatility of τt for one case
in Paper III is illustrated in Fig. 3.16; στ was reduced from 0.104 to 0.041.
A further reduction would be possible, but 2000 iterations was considered a
reasonable trade-off between low στ , desired kurtosis and computational time.
The computational time for simulating a 300,000 sample point time series
was around ten minutes, which should be compared to only ten seconds with
random phases. The improved performance of the model must therefore be
weighted against the increased computational time.
Two comments can be made on the methodology outlined above. Firstly, it
is often the case that the length of the simulated time series is different from the
training data. As an example, one might have a few years of training data and
want to simulate high-frequency fluctuations for a 30+ year long time series
stemming from reanalysis data. p In this case, the magnitudes of Y must be
interpolated and scaled by length(yt,sim )/length(yt ) in order to get correct
magnitudes for Ysim . Secondly, if yt,sim should be real-valued, Ysim must be

56
made symmetrical: even symmetry for the magnitudes and odd symmetry for
the phases.

3.5 Filters
Filters of different kinds were used in almost all papers. For a very accessi-
ble introduction to filters and other signal processing techniques, see [99]. It
should be recognised that the author of this thesis is not an expert on filter
design; more knowledgeable persons at our division are gratefully acknowl-
edged for their help on choosing appropriate filters. The motives for filtering
data were e.g.
• Smoothing time series, e.g. with moving average filters.
• Separate the high- and low-frequency components of a signal, e.g. when
intra-hourly wind power fluctuations were to be simulated.
• Splitting a signal into several frequency bands in order to deepen the
analysis of variability and correlations.
Operations like ARMA simulations and calculations of step changes of
wind power generation can also be interpreted as filtering of data. The ker-
nel (impulse response) defines the filter in the time domain; the output from a
filter is the input signal convoluted with the kernel. An example of a moving
average (MA) filter is shown in Fig. 3.17. The kernel has the amplitude 1/10
for sample 5–14 and zero for all others. This means that e.g. sample 47 of the
1
output signal is achieved as 10 ∑14
i=5 x47−i , where xn is sample n of the input
signal. The length of the output from a convolution is the sum of the lengths
of the input and of the kernel minus one. In order to get an output that is the
same length as the input and without phase-shift, the following adjustment can
be done. First, the kernel is made symmetrical, e.g. for a 5 point MA filter the
kernel is defined as 1/5 for sample -2 through 2. Sample n in the output, yn , is
thus the average of samples n − 2 to n + 2 in the input. Secondly, the endpoints
of the input signal need some special treatment. In the Matlab MA implemen-
tation (the smooth function), this is done by letting y1 = x1 , y2 = (x1 + x2 )/2
etc. Zero-phase implementation and endpoints handling of other filters are
described below.
Filters can be classified as infinite impulse response (IIR) or finite impulse
response (FIR). For an IIR, the kernel has an infinite length, i.e. all samples in
the output are affected by all samples in the input. An example of IIR is the
low-pass sinc filter, which kernel is given by

hn = 2 fC sinc(2 fC n), (3.22)

where fC is the cutoff frequency and sinc(·) is the normalised sinc function
defined by

57
 Input signal Filter kernel Output signal
5 5
0.1
4
4 0.08
Amplitude

3
0.06
3 ∗ =
2
0.04
2
0.02 1

1 0 0
0 50 100 150 0 10 20 0 50 100 150
Sample Sample Sample

Figure 3.17. Moving average filtering of a noisy signal. The input signal is convoluted
with the filter kernel, resulting in a smoother output signal.

1 0.2
a) b) Sinc
0.8 0.15 Windowed
sinc
Amplitude

Amplitude

0.6 0.1

0.4 0.05

0.2 0

0 -0.05
-50 0 50 -50 0 50
Sample Sample

Figure 3.18. (a) Hamming window, (b) Kernels for a low-pass sinc and a 101 point
windowed sinc filters with fC = 0.1. The latter was windowed with the Hamming
window shown in (a).

(
1 x=0
sinc(x) = sin(πx) (3.23)
πx otherwise.

Ideally, all frequencies lower than fC are passed and all higher frequencies are
blocked. When the sinc filter is windowed (i.e. the kernel is multiplied by a
window function) it becomes a FIR filter called windowed sinc. An example
of such filter kernel is shown in Fig. 3.18.
In Paper VII, high-order Butterworth filters were used. For numerical sta-
bility, these were implemented as second order sections (serially cascaded bi-
quadratic sections). For both windowed sinc and Butterworth filters, two is-
sues possibly need to be dealt with. Firstly, applying a filter will give a phase
shift in the output. If the filtering is not performed in real-time, this issue can
be resolved by applying the filter in both forward and reverse directions, e.g.
using the Matlab filtfilt function. Secondly, the endpoints of the input time

58
series requires special attention, especially for low-pass filtering. In Paper VII
this was handled by padding the time series with 10,000 samples of the mean
of its first and last 1000 samples respectively. The resulting 20,000 extra sam-
ples of the output signal were subsequently removed. In Paper VI, no padding
was performed but the first and last year of the filtered time series were simply
removed.
Depending on the purpose of the filtering, different characteristics are more
or less important. In signal processing, it can e.g. be desired to filter a square
wave with noise. The time domain of the output is then in focus, e.g. an effec-
tive reduction of noise while keeping sharp edges of the square wave. For the
applications in my papers, it was often more important to have sharp divisions
in the frequency domain. The windowed sinc and high-order Butterworth fil-
ters both provide good separation of the frequency bands. Furthermore, when
separating a signal xt into frequency components y1,t , y2,t , ..., each with its dis-
tinct frequency content and together covering the whole frequency range, the
following two relations hold to a very close approximation:

xt = ∑ yi,t (3.24)
i

Var(xt ) = ∑ Var(yi,t ) (3.25)

i

The latter equation is a direct result from Eq. (3.7) given that the different
components are uncorrelated.

3.6 Machine learning

With the increased computational power, many different machine learning (or
statistical learning) methods have become accessible. These often perform
better and can handle more predictor variables than parametric procedures
such as linear regression. The ability to handle many predictors makes ma-
chine learning suitable to use in combination with EOFs of meteorological
variables, see Section 3.3.5. A common feature for machine learning tech-
niques is that the functional forms linking the predictors to the response are
derived inductively from the data and the techniques can thus be called non-
parametric [105]. In Papers II and IV, random forests and/or gradient boost-
ing were employed, both based on classification and regression trees (CART).
Other methods include e.g. neural networks and support vector machines. An
accessible introduction to machine learning with many references for further
reading can be found in [105].

59
3.6.1 Classification and regression trees
CART [106] are the building blocks for many of the more advanced machine
learning techniques, including random forests and gradient boosting presented
in the following two subsections. As the name suggests, CART can be used
either for classification of categorical response variables or for regression of
quantitative response variables, and is particularly useful for high-dimensional
problems. The focus of this and the forthcoming subsections will be on regres-
sion trees, since these were used in Papers II and IV.
CART works by recursively partitioning the data, i.e. earlier stages are not
revisited after results from later stages are known. Assuming a quadratic loss
function, the impurity i in node τ is defined by the within-node sum of squares:

i(τ) = ∑(y j − ȳ(τ))2 , (3.26)

where ȳ(τ) is the mean of all sample points in node τ. At each stage, the
split that maximises the impurity reduction is chosen. [105] The partitioning
process is illustrated in Figs. 3.19-3.20 with the example function

Y := sin(4 · X) + Z 2 + ε, X, Z ∼ U (0, 1) ε ∼ U (0, 4), (3.27)

i.e. Y is a non-linear function with two predictors X and Z and a stochastic

component, but no interaction effects. As can be seen in Fig. 3.19, the values
of Y , indicated by different colours, are generally lower when X is high. The
best first split, reducing the impurity mostly, is for X = 0.80 (solid line). The
second split (dashed line) was made at Z = 0.66 for X < 0.80. The partitioning
can continue until there are very little data, sometimes only one sample point,
in each terminal node. In order to avoid overfitting, it is often recommended
to limit the number of partitions. This can be done by using a penalty function
for tree complexity, limit the number of splits or to set a lower limit for the
number of samples in the terminal nodes. Fig. 3.20 shows a grown tree with
the maximum number of splits set to five. One can easily follow a path from
the root node to a terminal node; the sample mean of Y for X < 0.80, Z < 0.66
and X >= 0.19 is for instance 2.92.
Regression trees can be used for prediction in a straightforward manner:
a sample is dropped at the root node, and the predictors determine in which
terminal node the sample ends up. The forecasted value of Y is the mean of
the training data for that node. Although the CART methodology might seem
simple, it is surprisingly good at handling multiple predictors and complex
dependencies between these. Nevertheless, even better results can be accom-
plished by combining several trees, which will be the topic of the following
two subsections.

60
 1

0.8

0.6
Z

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
X

Figure 3.19. Partitioning of data with classification and regression trees (CART).
Different colours indicate the value of Y = f (X, Z); blue (low), yellow (high). The fist
split is marked with a solid line and the second split with a dashed line.

x < 0.80 x >= 0.80

z < 0.66 z >= 0.66 x < 0.91 x >= 0.91

x < 0.19 x >= 0.19 z < 0.76 z >= 0.76

3.49  1.65

2.44 2.92 2.09 2.81

Figure 3.20. Regression tree with five splits and thus six terminal nodes (see Fig. 3.19
for the dataset that has been partitioned). The sample mean of Y for a certain region
in the X − Z space can be found by following a path from the root node to a terminal
node.

61
 3.6.2 Random forests
The idea behind random forests [107] is to grow an ensemble, often hundreds
or thousands, of trees and combine the results in order to get more accurate and
robust predictions. For each tree, a bootstrap sample11 is taken from the data.
For each partition, only a random sample of the predictors are considered.
In regression mode, one third of the predictors are normally chosen for each
partition. Generally, deep trees (many splits) are allowed to be grown. The
data samples that were not sampled are called “out-of-bag” (OOB) and amount
to around 37% of all data. The OOB data is subsequently dropped down the
tree and mean values are computed for each terminal node. For predictions,
the average of the predictions for each individual tree is used. The advantage
of using OOB data is that the tree becomes robust to overfitting. One reason
that the random forest algorithm is so effective is that only a subset of the
predictors are used in each partition. This implies that very local features of
the data will be captured in some of the trees. As a contrast, using only one
CART with all regressors will only let the most dominant predictors influence
the splits. [105]
In order to determine the importance of each predictor, Breiman [107] sug-
gests to permute the predictors one by one and measure the increase in error,
but other approaches are also possible [105]. It can often be useful to study
the impact of each predictor. This can be done by partial dependence plots
constructed in the following manner [105] for predictor X:
1. Grow a forest.
2. For all unique values of X, make a new database where other variables
are left unchanged but X takes that value.
3. Plot the average predictions for each value of X.
Predictor importance and partial dependence plots for our earlier example
in Section 3.6.1 is shown in Fig. 3.21. Matlab TreeBagger with standard set-
tings was used to grow a forest with 100 trees (which is a relatively small
forest). The Matlab OOBPermutedVarDeltaError property was employed to
infer predictor importance. In the wind power field, random forests have been
used by e.g. [34, 35, 108] and in Paper II.

3.6.3 Boosting
Boosting is also an ensemble method that can be based on CART, but there
are some important differences to random forests. The latter, as we saw in
Section 3.6.2, grow deep trees that are as independent as possible by using
random samples of the data and the predictors. Boosting algorithms, on the
other hand, improve the model by building upon the predecessor model. In
11 A random sample with replacement of the same length as the data. A bootstrap sample of
[1, 2, 3, 4] can for instance be [3, 1, 3, 4].

62
 3 4
a) b)

2.5
3.5
Predictor importance

Response
3

1.5

2.5
1
X
Z
2
0.5 sin(4X) + c
Z2 + c
0 1.5
X Z 0 0.2 0.4 0.6 0.8 1
Predictor value

Figure 3.21. Predictor importance and partial dependence plots for a random forest
trained on the dataset presented in Section 3.6.1 (a) Predictor importance in terms of
increase in RMS error when variables are permuted (b) Partial dependence plots. In
this example, the partial dependence plots are in good agreement with the underlying
first-order contributions to Y (dashed lines).

contrast to random forests, there is generally no chance element involved in

boosting algorithms and only a few splits are allowed for each tree.
Gradient boosting [109] (GB) stage-wise improves the model by finding
and adding the tree that minimises the residuals from the previous tree:

Fm (x) = Fm−1 (x) + νρm h(x; am ), (3.28)

where Fm (x) is the model at iteration m and h(x; am ), which is a CART with
parameters am , corresponds to the steepest gradient descent in the loss func-
tion. The optimal length of the descent is ρm , but by multiplying ρm with a
learning rate parameter ν ∈ [0, 1], the learning process can be forced to be
slower, thereby increasing flexibility at the expense of an increased computa-
tional time.
When GB is used in regression mode minimising least square residuals,
Friedman [109] calls it LSBoost. This algorithm is implemented in the Mat-
lab fitensemble function. With random forests, there is a built-in protection
for overfitting since OOB data are used for determining the fitted values for
each terminal node. Boosting procedures generally does not have this built-in
protection, so it may be useful to set aside some of the data as a validation set
and use this to determine the optimum hyperparameters, e.g. number of itera-
tions, learning rate and maximum number of splits for each individual CART.
In the wind power field, boosting trees have been used by e.g. [35, 110] and in
Papers II and IV.

63
4. Results and discussion

This section is structured in the following manner. First, in Section 4.1, evalu-
ation results from modelling of aggregated wind generation and forecasts (Pa-
pers I–IV) are given, i.e. the model outputs are compared to measurements.
Section 4.2 is devoted to historical and future wind generation in Sweden and
contains results from Paper D and unpublished results based on datasets, mod-
els and methods described in this thesis. Section 4.3, finally, is based on Pa-
pers V-VII and contains results on variability and correlations of wind power
and other IRE sources in the Nordic synchronous system and Europe.

4.1 Modelling wind power

Improving the methods for modelling wind power for power system stud-
ies is the main objective of this thesis. This section therefore contains the
most important contributions. Firstly, results are given for the “basic” model
of hourly, aggregated wind power generation (Paper I). Subsequently, results
from Papers II and III are presented; improvements of the high-frequency part
of the basic model and simulations of intra-hourly fluctuations respectively.
Finally, the model for generating synthetic wind power forecasts is evaluated
(Paper IV). The focus here will be on comparing model outputs to actual mea-
surements and not so much on using the models for exploring the character-
istics of future wind power. The latter is however the main motivation for
developing the models, see Chapter 6 for some thoughts on future studies.

4.1.1 Basic model

The basic model, described in Paper I and briefly in Section 3.2.1, is based
on MERRA reanalysis data and information on WTs. Separate models were
trained for the whole of Sweden (SE) and bidding zones SE2, SE3 and SE4.
Three years of data were used for training the models (2007, 2009 and 2011).
The results given below are for the three evaluation years 2008, 2010 and
2012. No model was developed for SE1 since the data on operating farms
in this zone was flawed: the measured generation was sometimes significantly
higher than the installed capacity. During the studied time period, the installed
capacity in SE increased from around 600 to 3500 MW (see Fig. 2 in Paper I).
Especially in the first years, very few farms were operating in SE2 and the

64
Table 4.1. Performance of the “basic” wind power model when comparing to mea-
surements from the Swedish transmission system operator. Results are given for the
evaluation years 2008, 2010 and 2012 for Sweden (SE) and three regions therein
(SE2–4). All errors are given as percentages of the installed capacity.
SE SE2 SE3 SE4
Mean Absolute Error 2.9% 6.5% 3.7% 4.2%
RMS Error 3.8% 9.1% 5.0% 5.9%
Mean Error -0.1% -0.7% -0.5% 0.4%
Correlation 0.98 0.89 0.97 0.97

0.8 Measurement
Hourly energy [p.u.]

Model
0.6

0.4

0.2

0
0 1 2 3 4 5 6 7 8
Time [weeks]

Figure 4.1. Evaluation of the modelling of hourly wind generation in Sweden (SE).
Model output and validation data for eight weeks. The RMS error for the period is
0.039 p.u., i.e. almost the same as the average for all three years of validation.

installed capacity was below 100 MW. It is thus not surprising that the errors
are higher for this region.
Some key performance metrics and results are given in Table 4.1 and Figs.
4.1–4.2. Overall, the results are satisfactory for SE, SE3 and SE4, but less so
for SE2. For SE, the RMS error was 3.8% and the correlation to measure-
ments 0.98, which is good in comparison to earlier results in the literature.
As shown in Fig. 4.2, the distributions of hourly generation are also in good
agreement. Before the application of a season/diurnal correction term, there
was a systematic overprediction of wintertime generation by the model. A
plausible explanation is icing losses.
Several different hyperparameters were tuned to give the best performance
(low objective function). When the air density is higher, the available power
in the wind increases. However, including air density in the computations of
hourly energy did not improve the results. A possible explanation is the re-
lationship between higher air density and icing losses. It was also found that
constant wind shear coefficients for each farm gave as good results as using
hourly values computed from the MERRA wind speeds at different heights.
The inclusion of PC smoothing, seasonal/diurnal corrections and different
losses in different wind sectors, in contrast, improved the model considerably.

65
 1000
Measurement
800 Model

600

Count
400

200

0
0 0.2 0.4 0.6 0.8
Hourly energy [p.u.]

Figure 4.2. Histogram of hourly wind generation in Sweden (SE) calculated for bins
of width 0.01 p.u.

Figure 4.3. The basic model underestimated high-frequency fluctuations of wind gen-
eration. (a) Step change duration curves. The distribution of four hour step changes
are adequately captured by the model, but one hour step changes are underestimated
(see also Fig. 4.6). (b) Power spectral density (PSD) estimates. The PSDs starts to
deviate for frequencies above (10h)-1 .

The output from the basic model was overly smooth which led to an under-
estimation of step changes. As an example, the SD of one hour step changes
was understimated by 14% for SE and even more for the separate bidding
zones. It is possible to detect the difference in smoothness in the time series
plot in Fig. 4.1 and in one hour step change magnitudes in Fig. 4.3a. The dif-
ference is however much more obvious from PSD estimates, see Fig. 4.3b. The
distribution of step changes can be crucial for determining the challenges of
integrating more wind power into a power system. The lack of high-frequency
variability was the reason for developing the improved model described in the
next subsection.

66
4.1.2 Improved model
In this section, the model for simulating high-frequency variations of hourly,
aggregated wind power generation is described briefly and results from a com-
parison with measurements are given. The elements in the modelling, e.g.
filtering, FFT-IFFT transformations and machine learning algorithms, are de-
scribed in more details in Chapter 3. As in Paper II, the basic model described
in the previous subsection is denoted “the MERRA model”, measurements
are denoted “SvK” and subscript “sim” denotes simulated time series. The
methodology involves the following steps:
1. Separate hourly time series from SvK and the MERRA model (M) into
their low- and high-frequency components, e.g. SvKHF and MLF . A win-
dowed sinc filter with a 5000 samples long kernel and cutoff frequency
of (10h)-1 was used.
2. Transform SvKHF to an approximately stationary time series ySvK . This
was done by dividing ySvK by its variability index (VI), defined as the
24-hour moving SD of a signal.
3. Find the magnitudes of the frequency domain representation of ySvK us-
ing FFT (Fast Fourier Transform).
4. Generate ysim of the same length as the MERRA time series, using in-
verse FFT with interpolated and scaled magnitudes from 3) and appro-
priate phase angles. The algorithm for finding phase angles is described
in Section 3.4.2.
5. Transform ysim to MHF,sim , using predicted VI. Nearest neighbour, ran-
dom forest and GB models were trained and used for prediction. For the
former, only calendar month, hour of the day and time series extracted
from the MERRA model (e.g. the VI of MHF ) were used as predictors.
For the latter two, EOF time series for several meteorological variables
were optionally used in addition.
6. The time series for the improved model is achieved as MLF + MHF,sim .
The volatility of the high-frequency component varies in time. By sim-
ulating a stationary time series ysim which was subsequently transformed to
MHF,sim using the predicted VI, the varying volatility of measurements could
be reproduced in the simulations. If was found that a GB machine learning
model was somewhat better than k nearest neighbour and random forests for
predicting the VI. By using EOF time series as predictors in addition to calen-
dar month, hour of the day and time series extracted from the MERRA model,
the correlation between VIs of simulations and measurements was increased
slightly. A scatter plot of observed versus predicted VIs for the GB model
with EOFs as predictors is shown in Fig. 4.4.
Fig. 4.5 gives some results from the simulation. The histogram in Fig. 4.5a
clearly shows that the magnitudes of the high-frequency fluctuations from the
MERRA model are too small. For the improved model, however, these are in

67
Figure 4.4. Scatter plot of observed vs. predicted variability indices (VI, a metric of
the volatility of the high-frequency time series). The correlation is 0.67.

good agreement with measurements. Fig. 4.5b gives a representative exam-

ple of the varying magnitudes of high-frequency fluctuations. The improved
model can relatively well capture the magnitudes, although the match is not
perfect. Figs. 4.5c–d show time series excerpts and PSD estimates for mea-
surements, the MERRA model and the improved model. The fluctuation mag-
nitudes from the improved model are more similar to those seen in measure-
ments.
The underestimation of one hour step changes was the primary reason to
improve the basic model described in Section 4.1.1. As Fig. 4.6 shows, the
distributions of step changes for measurements and the improved model are
very similar. Table 4.2, finally, gives RMS errors, correlations of VIs and
some metrics describing one hour step changes. Overall, the improved model
is really an improvement. Due to the filtering and addition of stochastic noise,
the RMS error is slightly increased. The observant reader may have noted that
the metrics given for the MERRA model are not identical to those presented
in Section 4.1.1. The reason for this is that all six years of output (2007–
2012) from the MERRA model was used for training, validating and testing
the improved model.

4.1.3 Intra-hourly fluctuations

In many power system studies, wind data with a higher resolution than hourly
is preferable [18, 111]. Two independent studies [112, 113] of the Irish power
system compared the results from using hourly and sub-hourly (5–15 minutes)
data in the simulations. Markedly higher levels of generator cycling and ramp-
ing rates were seen in the highly resolved simulations. Furthermore, hourly
resolution models seem to underestimate the value of flexible generation and
energy storage.

68
 10 4
a) b)

10 3
0
Count

2
10

10 1 0

10 0
-0.05 0 0.05 0 1 2 3 4 5 6
High-frequency component [p.u.] Time [weeks]

0.6 10 0
c) The two lower curves d)
are displaced 0.1
Hourly Energy [p.u.]

0.4 and 0.2 p.u.

PSD [(p.u.) 2 · h] 10 -2
0.2

10 -4 Measurement
0
MERRA
Improved
-0.2
0 1 2 3 4 5 6 7 8 (24h) -1 (10h) -1 (2h) -1
Time [days] Frequency

Figure 4.5. Comparisons of measurements and outputs from the basic (MERRA) and
improved models. (a) Histogram of high-frequency components ( f > (10h)−1 ) using
bin width 0.001 p.u., (b) representative example (in terms of correlation of variability
indices) of high-frequency components, (c) example of hourly energy time series and
(d) power spectral density estimates.

0.06
Measurement
-0.05
0.04 MERRA
Improved
1h step change [p.u.]

-0.1
0.02
-0.15
0.99 1
0
0.15
-0.02
0.1

-0.04 0.05
0 0.01
-0.06
0 1
Exceedance probability

Figure 4.6. Duration curves for one hour step changes of the hourly generation time
series (with the tails as insets). With the improved model, the step changes are no
longer understimated.

69
Table 4.2. Various metrics comparing the performance of the original MERRA model
and the improved models. ∆P1h is short for one hour step change in hourly energy and
SD for standard deviation. All metrics except correlation are given in relation to the
installed capacity, i.e. 3.6% corresponds to 0.036 p.u.
Measure- MERRA Improved Improved model
ment model model (GB) (GB, no EOFs)
RMS error - 3.6% 3.8% 3.8%
Correlation VI - 0.58 0.67 0.65
SD ∆P1h 2.0% 1.7% 2.0% 2.0%
1st %-ile ∆P1h -5.3% -4.5% -5.2% -5.2%
99th %-ile ∆P1h 5.4% 4.9% 5.5% 5.5%

For many power systems, only hourly measurements are available. The me-
teorological models described in Section 2.2 are also often of hourly or poorer
resolution. In Paper III, a methodology for simulating intra-hourly fluctuations
(deviations from hourly mean values) was therefore developed. The ideas were
relatively similar to those used for simulating hourly fluctuations described in
Section 4.1.2; isolate the high-frequency part of actual measurements, simu-
late similar (stochastic) fluctuations using the IFFT algorithm with a search for
phase angles, taking into account the non-stationarity of the data (right volatil-
ity at the right time) and finally add the simulated noise to smoothened data
for the objective power system. Since no intra-hourly measurements were
available for the Swedish power system, models were trained and evaluated
using data from Denmark and Germany (5- and 15-minute temporal resolu-
tion respectively). The PSDs of these systems are similar to that for Sweden
for frequencies up to (2h)-1 and it can therefore be expected that the PSDs are
also similar for higher frequencies.
As compared to the methodology described in Section 4.1.2, the approach
differed in some aspects. Firstly, no EOFs were used for prediction of volatil-
ity since the EOFs are system specific. Only hour of the day, calendar month
and variables derived from the hourly energy production time series were used
as predictors. Secondly, instead of predicting the variability index, all obser-
vations were sorted into bins depending on the predictor values. Distributions
(t location-scale) were fitted for each bin of the training data. The normally
distributed and stationary simulated time series were transformed into simu-
lated high-frequency components using the inverse of these transformations.
Finally, a spline adjustment (see Section 3.3.3) was used to get correct hourly
mean generation. Models trained with data from Denmark and Germany were
used to simulate time series for both these systems. For our purposes, i.e. us-
ing these models for simulating Swedish intra-hourly fluctuations, the most
interesting results are for the Danish model evaluated on the German system
and vice versa.

70
 a) b)
10 4 w t,val 10 4
w t,sim
Count

10 2 10 2

10 0 10 0
-0.1 -0.05 0 0.05 0.1 -0.1 -0.05 0 0.05 0.1
wt [p.u.] wt [p.u.]
0.05

c) 0
0.05
-0.05
0

-0.05
0 1 2 3 4 5 6 7 8 9 10 11
Time [day]

Figure 4.7. Similarity of deviations from hourly means (wt ) for simulations and
validation data. (a) Histogram with Denmark as both training and objective system
(Denmark-Denmark), (b) histogram (Denmark-Germany) (c) representative time se-
ries example (Denmark-Germany).

Let us begin by having a look on the simulated and measured deviations

from hourly mean values (wt,sim and wt,val ). Fig. 4.7a–b give histograms of
wt for Denmark as both training and objective system (however using differ-
ent time periods) and for Denmark as training system and Germany as the
objective system. These combinations were chosen based on their, as an over-
all judgement, best and poorest performance respectively. In both cases, the
simulated and measured deviations have very similar distributions, also in the
tails. Fig. 4.7c shows a time series example of wt,sim and wt,val for Denmark
as training system and Germany as the objective system. When comparing
to the results in Section 4.1.2, the varying volatility of wt for simulation and
measurements are much higher correlated (0.94–0.95 as compared to 0.65–
0.67). One reason for this is that wt,val is strongly dependant on the slope of
the hourly energy generation Ph . When Ph is increasing, wt,val is likely to be
strongly negative in the beginning of the hour and strongly positive in the end
of the hour. When Ph is nearly constant, wt,val is often small in magnitude.
This behaviour is easy to mimic in the simulations.
Fig. 4.8a shows a comparison between the simulated output Pt,sim and mea-
surements Pt,val . By visual inspections these time series have similar charac-
teristics. The PSDs in Fig 4.8b–c strengthen the impression that the simula-
tions were successful. As can be seen in Table II in Paper III, the SDs of one
sample step changes are also very similar for Pt,sim and Pt,val .

71
 a)
0.4 Pt,val
0.35 Pt,sim
Power [p.u.]

0.3 Ph

0.25

0.2

0.15

0 2 4 6 8 10 12 14
Time [Hours]
b) c)
PSD [(p.u.) 2 /Hz]

0
10
10 0

10 -2
10 -2

(4h) -1 (1h) -1 (10min) -1 (4h) -1 (1h) -1 (0.5h)-1

Frequency Frequency

Figure 4.8. (a) Time series example of the simulated output Pt,sim and measurements
Pt,val (Denmark as both training and objective system). (b–c) Power spectral den-
sity estimates for (b) Denmark-Denmark and (c) Denmark-Germany as training and
objective systems respectively.

72
4.1.4 Synthetic forecasts
In Paper IV, a new method was developed for creating synthetic wind power
forecasts for wind integration studies, see also Section 3.2.5. The idea is to use
reforecasts from the GEFS dataset and some statistical processing for produc-
ing the synthetic forecasts. The modelled FEs can subsequently be computed
as

Modelled FE = Modelled generation − Synthetic forecast, (4.1)

where the modelled generation is based on MERRA data as described in Sec-
tion 3.2.1. Before evaluating the similarity of modelled and actual FEs, let us
have a look at the terrain dependence and temporal consistency of modelled
FEs.
Fig. 4.9a–b show how the SDs of D+1 and D+5 FEs vary over Sweden. Re-
sults are given for generation calculated from “raw” GEFS wind speed fore-
casts, i.e. before the addition of multivariate ARMA noise and removal of a
fixed share of the error each hour. When comparing to the complexity of the
terrain, quantified in Fig. 4.9c as the SD of ground heights within each grid
cell, it is obvious that the day-ahead (D+1) errors are strongly correlated to the
terrain complexity. FEs for longer horizons are, in contrast, not so terrain de-
pendent, but are larger for offshore farms due to higher CFs1 . Although we did
not have enough measurements to properly evaluate if the relative differences
in FE magnitudes were correct, the results seem reasonable when comparing
to an earlier study [88] and FEs for three farms in Sweden.
An advantage with reforecasts, i.e. forecasts produced from historical data
using approximately consistent models and assimilation systems, as compared
to historical, operational forecasts is that the quality does not change much
over time. In Paper IV, 30 years long generation and forecast time series were
computed for 50 random locations in Sweden. In general, monthly CFs agree
well except for year 2010–2012, which is likely due to a change in source for
GEFS initial conditions. The quality of the forecasts only improves slightly
over the years. A linear fit to SDs of hourly FEs computed for each year gives
a reduction from 0.179 p.u. year 1985 to 0.161 p.u. year 2014 (average for the
individual farms). The same analysis performed for aggregated forecasts gives
a reduction from 0.078 p.u. to 0.075 p.u.
Let us now turn to a comparison between modelled and actual FEs for the
whole of Belgium, see Fig. 4.10. An example of modelled generation and
synthetic D+1 and D+7 forecasts is first presented in Fig. 4.10a. Note that
towards the end of day two, there is a phase error in the D+1 forecast; the
generation ramps up slightly earlier than the forecast. This type of errors,
also present for operational forecasts, are generally not captured with purely
1 When the errors are expressed in relation to the energy production, and not in relation to
installed capacity as in Fig. 4.9, the opposite however holds.

73
Figure 4.9. (a–b) Standard deviations of forecast errors for D+1 and D+5 respec-
tively (“raw” forecasts). (c) Terrain complexity quantified by the standard deviation
of ground heights within each grid cell. The ground heights were obtained from a
database with 0.5 km horizontal resolution.

statistical methods for generating synthetic forecasts. Also note the abrupt
change in the D+7 forecast at the beginning of day one. Similar step changes
can be seen when actual forecasts are updated, especially for longer horizons.
Fig. 4.10b shows how the FEs for operational and synthetic forecasts de-
pend on horizon. Note that only D+1 and the D+7 forecasts were available
from the Belgian TSO. The good agreement should not come as a surprise
since the model parameters were tuned to give a good match. In Fig. 4.10c,
the impact from generation level on FE magnitude is displayed. For D+1,
model and measurements both give highest errors when the national genera-
tion is around 0.2–0.7 p.u. During these occasions it is likely that many WTs
are operating in the steepest part of the PC. For D+7, the curves are not agree-
ing as well, but both model and measurements give lower errors for very low
national generation levels. Quantile-quantile plots for measured and modelled
FEs are given in Fig. 4.10d–e. The distributions agree well, also in the tails.
This is beneficial since the extreme errors are important for dimensioning of
the reserves. The D+1 autocorrelations given in Fig. 4.10f have an excellent
match. For D+7, the autocorrelation is however somewhat overestimated by
the model. Fig. 4.10f, finally, shows PSD estimates for model and measure-
ments.
For Sweden, day-ahead forecasts were available for three individual farms.
The same ARMA model was used to generate noise for the synthetic fore-
casts, but slightly more of the error for each hour was subsequently removed
(higher e f ix parameter). Although somewhat better performance (i.e. similar-
ity to measurements) could be obtained by using separate ARMA models, we
believed it to be beneficial to show that one model works well for both coun-
tries and both for individual farms and for national output, thereby demonstrat-

74
 a) 1
Generation [p.u.] Generation (MERRA)
0.8 D+1 forecast
D+7 forecast
0.6

0.4

0.2

0
0 1 2 3 4 5 6
Time [days]
b)
0.25 c) 0.25
Forecast error (SD) [p.u.]

Forecast error (SD) [p.u.]

Measurements 0.2
0.2 Model
0.15
0.15 See f) for legend
0.1
0.1
0.05

0.05 0
D+1 D+3 D+5 D+7 0 0.2 0.4 0.6 0.8
Forecast horizon Generation [p.u.]

d) 0.5 e)
D+1 errors 0.5 D+7 errors
Model quantiles

Model quantiles

One-to-one One-to-one

0 0

-0.5

-0.5
-0.5 0 0.5 -0.5 0 0.5
Measurements quantiles Measurements quantiles

f) 1 g)
10 0
0.8 D+1
D+7
0.6 D+1 model
PSD
ACF

D+7 model D+1

10 -2
0.4 D+7
D+1 model
0.2
D+7 model
0 10 -4
0 10 20 30 40 50 10 -2 10 -1
Lag [hours] Frequency [h-1]

Figure 4.10. Evaluation of model performance for Belgium. (a) Time series ex-
ample of modelled generation and synthetic D+1 and D+7 forecasts. Sub-figures
(b)–(f) compare different properties of measured and modelled forecast errors (FEs).
(b) Standard deviation (SD) of FEs depending on horizon (measurements are only
available for D+1 and D+7), (c) Errors depending on generation level, (d–e) Q-Q
plots, (f) Autocorrelation and (g) Power spectral density.

75
Table 4.3. Impact from WT characteristics on modelled forecast errors. Results for
hypothetical farms in Sweden for horizon D+1 and D+7. S1 is the baseline sce-
nario, S2 has lower specific rating, S3 has higher assumed wind speed and S4 has
30% offshore capacity. All results are given as standard deviations of forecast errors,
normalised to annual energy yield.
D+1 D+7
S1 0.193 0.510
S2 0.177 0.465
S3 0.177 0.466
S4 0.158 0.448

ing the robustness of the method. The most important result for Sweden was
that when e f ix was tuned to give a correct magnitude of the errors for all farms
combined, the magnitude of the errors for individual farms are also appropri-
ate. Farm 1 is located in complex terrain and has the highest errors. Farm 2 is
an offshore farm with higher CF than the other two farms and relatively high
FEs. Farm 3, which has lowest errors both for measurements and model, is an
onshore farm in somewhat hilly terrain, bordering flat terrain near the sea in
the prevailing wind direction.
In Paper D, it was shown that the variability of wind power generation can
be reduced significantly by increasing the average CF, which can be accom-
plished by, for example, increasing the share of offshore farms, increasing
the hub heights or lowering the specific ratings. In Paper IV we investigated
whether the reduced variability also leads to lower FEs. Four different scenar-
ios were considered: S1 with a CF around 0.30 and S2–S4 with CFs around
0.36. S1–S3 consist of 50 randomly selected onshore farms in Sweden. The
same farms were used for all these scenarios. In S4, fifteen of the farms were
replaced with five offshore farms. As compared to the base scenario (S1),
the higher CFs in S2–S4 were obtained by reducing the specific rating (S2),
increasing the hub heights and thereby the mean wind speeds (S3) and by as-
suming 30% of the installed capacity as offshore farms (S4).
As can be seen in Table 4.3, increasing the CF not only leads to lower vari-
ability, but also to lower FEs2 . The lowest errors are obtained for S4, which
can be explained by longer average separation distance between the farms and
different weather patterns onshore and offshore. This leads to lower correla-
tion between the FEs of individual farms and consequently lower aggregated
errors. The smoothing effect is strongest for shorter horizons when FE corre-
lations are lower.

2 Boththese conclusions are valid when the results are normalised to the annual energy yield
which, in my opinion, is the preferable option when comparing different scenarios.

76
4.2 Wind power in Sweden
In Section 4.2.1, results are given for historical wind power generation and
trends of wind farm characteristics in Sweden. The analysis is based on hourly,
aggregated measurements from 2007 to 2014, see Section 2.1, and a database
of all operating wind farms. In Section 4.2.2, a selection of results from mod-
elling future wind power generation in Sweden are presented.

4.2.1 Historical generation

Beginning with trends3 , Fig. 4.11 displays how hub heights, specific rating
and CFs have evolved over time. The increase in hub heights has been more
or less linear, from below 40 m in 1990 to over 100 m today. Modern WTs
also have smaller installed capacity in relation to rotor areas; slightly above
300 W/m2 as compared to around 400 W/m2 in the 90s and early 00s. The
development is not linear as for hub heights, but rather it seem to be a change
of trend around year 2005. This is likely related to the increased deployment
of farms in forests at the same time. Increased hub heights, relatively larger
rotors and, to a smaller extent, improvements in WT efficiencies have led to
markedly higher CFs. When a farm is deployed, the estimated annual energy
yield is reported to the electricity certificate system [60]. Figure 4.11c shows
how the estimated CFs have increased from around 0.20 to 0.33. As shown in
Figure 4.11d, the measured national CF has increased from around 0.21 year
2007 to around 0.29 year 2013–2014 (long-term corrected data). The mea-
sured CF has increased faster than expected from the estimated CFs which
can be explained by the systematic overestimation of energy yields for farms
deployed in the 90s and early 00s. In year 2014, these old WTs only consti-
tuted a small fraction of the total capacity and consequently the overestimation
of the national CF was much smaller than in 2007.
Figure 4.12 and Table 4.4 give information on the distribution of hourly
generation. The generation very seldom, on average 14 hours per year, exceeds
0.8 p.u. The 90th percentile is only 0.51 p.u. If high wind penetration is an
issue, it is therefore not too costly to curtail wind power at times with low
load and high wind (especially since the price can then be expected to be low).
Some quantitative examples are given in Section 4.2.2.
There are significant seasonal differences in wind generation. Diurnal pat-
terns are also present, although not as strong as the seasonal variations. Fig.
4.13 illustrates the patterns. In Fig. 4.13b, the monthly means have been re-
moved in order to visualise the diurnal cycles each month more clearly. These
pattern, which are not properly captured by the MERRA reanalysis, seem to
be related to the rise and set of the sun.
The variability of intermittent energy sources is not a problem per se; if
the fluctuations are correlated to those of the electric load it can actually be
3 The values are weighted on the installed capacity of each farm.

77
 a) b)
100

Specific rating [W/m 2 ]

Hub height [m]

400
80

60 350

40
300

1990 1995 2000 2005 2010 2015 1990 1995 2000 2005 2010 2015
Year of deployment Year of deployment
c)
0.35 d) 0.3
Capacity factor (estimated)

Measurements

National capacity factor

0.28 LTC

0.3
0.26

0.25 0.24

0.22
0.2
0.2
1990 1995 2000 2005 2010 2015 2008 2010 2012 2014
Year of deployment Year

Figure 4.11. Capacity weighted trends for Swedish wind power, see Paper D for more
information on the dataset. (a) Hub heights, (b) specific ratings, (c) capacity factors
(estimated by farm owners before construction) and (d) measured national capacity
factors, both raw and long-term corrected (LTC) data.

Figure 4.12. Figures illustrating the distribution of Swedish hourly wind generation
2007–2014. (a) Histogram computed with bin width 0.01 p.u. and (b) duration curve.

78
Table 4.4. Distribution metrics for wind power in Sweden (hourly measurements
2007–2014). The time series has been normalised to the installed capacity at each
time step. The 10th percentile is considered “firm capacity” by the Swedish TSO.
Metric Value
Capacity factor 0.25
Standard deviation 0.17 p.u.
Minimum 0.003 p.u.
1st percentile 0.021 p.u.
10th percentile 0.066 p.u.
90th percentile 0.51 p.u.
99th percentile 0.71 p.u.
Maximum 0.90 p.u. (some years below 0.8 p.u.)

a) b)
2 2 0.02
0.3
4 4 0.01
Month

0.25 0
6 6
-0.01
8 8
0.2
-0.02
10 10
-0.03
12 0.15 12
5 10 15 20 5 10 15 20
Hour Hour

Figure 4.13. Seasonal/diurnal plots for wind power in Sweden (2007–2014). (a)
Average capacity factors (CFs). The generation is considerably higher in wintertime.
(b) Deviations from monthly CFs. The pattern seems to be related to the rise and set
of the sun.

79
advantageous. As can be seen in Fig. 4.14a, the seasonal variations of wind
power and load in Sweden are in good agreement. In the figure, both wind
generation and load have been normalised to their average values. Unfortu-
nately, a closer look at the wind-load dependence reveals that at times with
very high load, the wind generation is below its average. In Fig. 4.14b, the
mean and 10th percentile of wind generation are plotted as functions of the
load (grouped into 100 bins). As an example, the first load bin correspond to
the 1% of the time when the load is lowest (around 9 GW). The mean wind
generation is then around 70% of its average. Note that for the computation of
10th percentiles of wind generation, five load bins at the time were considered
in order to get more robust results. The 10th percentile is an important met-
ric since this is defined as ‘firm capacity” by the Swedish TSO. Recently, the
TSO’s estimated firm capacity of wind power was increased based on obser-
vations in winter-time (when the load is generally higher). According to my
analysis, this was a mistake; during the 5% of the time when the load is at its
maximum, the 10th percentile of wind power is actually very slightly below
the 10th percentile for the whole time period. A physical mechanism for this
is suggested in [114]: “...load extremes are often due to relatively infrequent
large-scale high-pressure weather systems that typically bring calm winds”.
Indeed, very high loads in Sweden generally coincides with high pressures
(national average pressure data from the MERRA reanalysis).
Fig. 4.14b illustrates the importance of using synchronous wind and load
data. If one insist of using purely statistical methods for producing wind power
time series, the relationship between load and wind generation must be taken
into account. The following example quantifies the error from neglecting the
relationship (but taking into account the seasonality). When randomly per-
muting4 the years of the wind power time series, the generation was in average
20% above mean for the 5% of time with largest loads. This can be explained
by the higher generation in wintertime. With synchronous data, the generation
is 7% below average.
The variability of Swedish wind power is relatively low in comparison to
that in other systems [6]. The main explanation for the low variability is the
large mean distance between farms. In Table 4.5, a few metrics of change (one
and four hour step changes) are given. The distributions of step changes are
given in Fig. 4.3 on page 66. Kurtosis is one way to quantify the occurrence
of relatively extreme values. Both one and four hour step changes have kur-
tosis around five, i.e. the distributions are significantly more heavy-tailed than
normal distributions.
The average revenue for wind farm owners does not necessarily equals the
time average of electricity prices. Firstly, wind energy may be produced at
4 Onethousand permutations were performed. Only the 382 permutations with no years in
common were considered. As an example, [2009, 2008, 2012, 2013, 2007, 2011, 2014, 2010]
would not be accepted since year 2008 is the second element both for the load data and for the
permuted wind power data.

80
a) 1.4 b) 1.4
Average wind/load (normalised)

1.2 Load 1.3

Normalised wind generation

Wind
1.2
1
1.1
0.8
1
0.6
0.9
0.4 Mean
0.8
10 th percentile
0.2 0.7

0 0.6
1 2 3 4 5 6 7 8 9 10 11 12 10 15 20 25
Month Load [GW]

Figure 4.14. Relationships between wind power generation and load in Sweden. The
data have been normalised to the means for the whole time period (2007–2014). (a)
Wind has a similar seasonality as the load. (b) Wind generation for 100 bins of load.
Each bin has the same amount of data, e.g. the first bin contains wind generation
corresponding to the lowest percent of load observations. The 10th percentiles of
wind generation are given for five load bins at the time.

Table 4.5. Metrics for one and four hour step changes for wind power in Sweden
(hourly measurements 2007–2014).
Metric Value 1h Value 4h
Standard deviation [p.u.] 0.019 0.061
Minimum [p.u.] -0.13 -0.33
1st percentile [p.u.] -0.051 -0.16
99th percentile [p.u.] 0.053 0.17
Maximum [p.u.] 0.12 0.41
Kurtosis 5.1 4.8

81
times with prices differing from the average, e.g. more wind energy is pro-
duced during winters. Secondly, if the penetration is significant, a high wind
power generation at certain hours will have a negative impact on the price.
The ratio of wind weighted price and time weighted price (i.e. a simple mean)
is called value factor (VF). Wind value factors reported in the literature is of-
ten around unity for low penetrations but declines when more wind power is
deployed. Hirth [7, 115] has showed that at a penetration level of 30%, wind
electricity can be worth 10–50% less than generation from a constant source.
The VF is system specific, e.g. depending on the relationship between wind
and load and the type of generators used for balancing wind power.
According to data from the Swedish TSO [58], the wind power generation
and load for year 2015 were 17 and 135 TWh respectively. Based on mar-
ket data from NordPool [116], the average prices weighted on time, load and
wind generation were 22.0, 23.3 and 21.3 e/MWh respectively. The VF for
wind power is thus only slightly below unity. A slower decline in the VF with
penetration rate can be expected for hydro dominated systems [115]. It can
also be noted that in 2015, prices were low in a historical context and dif-
fered little between the bidding zones. An interesting topic for future studies
(see Chapter 6) is how the VF in the Nordic system will change in a future
with considerably higher wind penetration and furthermore if a potential VF
reduction can be alleviated with proper measures.

4.2.2 Future generation

In this section, some results from modelling future wind power generation in
Sweden are presented. The results are given for scenarios developed in Pa-
per D and sometimes, for comparison, for linearly scaled historical generation
data. The database of potential farms, used for generating the scenarios, is il-
lustrated in Fig. 4.15. When comparing to the situation today (Fig. 4.15a), the
distributions of onshore farms with all permits (Fig. 4.15b) and farms in the
permitting process (Fig. 4.15c) are strikingly similar; most in bidding zones
SE2 and SE3, somewhat less in SE4 and the least in SE1. A large share of the
planned offshore farms are however located in SE4.
Two of the main scenarios from Paper D, A1 with 20 TWh/a and C1 with
50 TWh/a, are shown in Fig. 4.16. The A1 scenario, expected to be realised
around 2020, has a very small offshore capacity but for C1, the offshore energy
share is 26%.
As discussed in Section 1.1, duration curves of the net load can provide
insights into the challenges of integrating variable renewables such as wind
power. Fig. 4.17 shows duration curves for both hourly generation and one
hour step changes for the load and the net load with 50 TWh/a of wind power
(scenario C1). The mean net load is reduced by 5.6 GW, the maximum net
load by 2.3 GW and the minimum by 9.1 GW. Consequently, if wind power

82
 a) b) c)

0 100 200
±
400 km

Figure 4.15. Database of wind farms in Sweden from Paper D. Bidding zone borders
are indicated by thick lines (SE1 in north to SE4 in south). The circle areas are propor-
tional to capacity. Farms (a) in operation or under construction, (b) with all permits
in place, (c) in the permitting process.

replaces nuclear generation, some additional peak generating capacity might

be necessary. The distribution of net load step changes also differ from that
of the load, in particular for the tails. Ramping capabilities is however not the
main issue for the hydro-dominated Nordic power system.
In order to alleviate the impact from wind power on the power system it is
important to pinpoint the most influential parameters for the resulting variabil-
ity and quantify costs and benefits of different strategies to reduce the variabil-
ity. Some future studies on this topic are planned, see Chapter 6. In Fig. 4.18,
results from a parameter study in Paper D are given. As a starting point, a sce-
nario with 20 TWh/a onshore farms, average CF of 0.34 and a specific rating
of 280 W/m2 was used. The SD of one hour step changes was used as a proxy
for variability. All results were normalised to the reference case.
According to Fig. 4.18a, adding more farms of the same type has a rela-
tively small impact on the variability. The geographical smoothing effect is in
other words small when the number of farms is significant to start with and
no action is taken to steer the deployment of farms towards regions with little
capacity. However, as can be seen in Fig. 4.18d, if the placement of farms is
optimised, a considerably lower variability can be obtained with only a few
farms. Depending on whether only available onshore projects, all onshore lo-
cations or onshore and near offshore locations are allowed in the optimisation,

83
 Built New onshore
New onshore Repower
Repower New offshore

a) b)

65 ° N 65 ° N

60 ° N 60 ° N

20 MW 20 MW

100 MW 100 MW

500 MW 500 MW

55 ° N 55 ° N
10 ° E ° 10 ° E °
25 E
25 E 15° E °
20 E
15° E °
20 E

Figure 4.16. Two of the main scenarios from Paper D. “Built” indicates farms already
in operation or under construction, “Repower” indicates current wind turbines (WTs)
that have been replaced by larger, modern WTs. Bidding zone borders are indicated
by thick lines. (a) Scenario A1: around year 2020, 20 TWh/a, 7.5 GW, (b) scenario
C1: 50 TWh/a, 14.2 GW.

Hourly generation One hour step changes

30 4
Load -1
25 3
Net load -2
Step changes [GW]
Generation [GW]

20 2
-3
0.99 1
15 1

10 0
25 3
5 -1
2
0 20 -2
0 0.01 0 0.01
-5 -3
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Exceedance Exceedance

Figure 4.17. Duration curves of (a) hourly generation and (a) one hour step changes
for load and net load. The results are based on measured load for 2007–2014 and wind
power scenario C1 from Paper D. In this scenario, wind power supplies 50 TWh/a or
around 36% of the load.

84
a) 1.2 b) 1.2

Normalised SD ∆P 1.1 1.1

1 1

0.9 0.9

0.8 0.8

0.7 0.7
10 20 30 40 50 60 70 0.25 0.3 0.35 0.4 0.45
Annual generation [TWh] Capacity factor
d)
c) 1.2
Offshore share
1.2
Onshore + near offshore
Normalised SD ∆P

1.1 Share in north 1.1 Onshore

Onshore, available projects
1 1

0.9 0.9

0.8 0.8

0.7 0.7
0 0.2 0.4 0.6 0.8 1 5 10 15 20 25 30
Share Number of farms

Figure 4.18. Impact from several parameters on the standard deviation of one hour step
changes of wind generation. All results have been normalised to that for a reference
scenario. (a) Adding more farms with the same characteristics has a small impact
on variability, (b) variability is greatly reduced with higher average capacity factors,
(c) minimum variability is attained with shares of offshore farms and farms in north
around 40% and (d) optimising the geographical distribution of farms has the potential
of significantly reducing the step change magnitudes.

the SDs of step changes saturates at around 70–80% of that for the reference
case. Note that the same CF was assumed for all farms (including offshore) in
order to isolate the effect of optimal geographical dispersion.
Fig. 4.18b shows the impact from the average CF on variability. Increasing
the CF leads to a strong reduction of step change magnitudes. Fig. 4.18c
quantifies the variability depending on share of farms in northern Sweden (SE1
and SE2) and offshore energy production respectively. The share of capacity
in the north is not the most important factor, but lowest variability seem to
be obtained with around 40%. Increasing the share of offshore wind power
would reduce the variability. This can be explained both by an increase of the
mean distance between farms and a higher average CF (0.46 was assumed for
offshore farms). Too high shares of offshore wind is not desirable since the
generation is then concentrated to only a few locations. A final note is that the
specific rating has a very small impact on the step changes as long as the CF
is not altered, i.e. from a variability point of view it is not important whether a
certain CF is obtained by a low mean wind speed and a large rotor area or by
a higher wind speed and a smaller rotor.

85
 -1000
1h step changes in net load [MWh/h] 1 st percentiles

3000 -1500

99.9 th 0.1 th percentiles

percentiles
-2000
2500

Historical
-2500
Scenario A1
2000 Scenario C1
99 th percentiles
-3000
0 10 20 30 40 50 0 10 20 30 40 50
Wind penetration [%] Wind penetration [%]

Figure 4.19. Impact from wind penetration and from using different wind power
datasets on extremes of one hour step changes in the net load. The three wind power
time series were linearly scaled. For comparison, the maximum load was 26 GW.

In Section 1.3 it was stated that linear scaling of historical wind genera-
tion might be very misleading for studying the future. The following exer-
cise serves as an example. Let us begin by pointing out that the geographical
smoothing for Sweden is close to saturated; building more farms of the same
type that we already have in the same regions will reduce the normalised vari-
ability very little. It is thus reasonable to, as a first approximation, linearly
scale generation in order to study different penetration levels of similar WTs.
In Fig. 4.19, extreme values of net load step changes are given for different
penetration levels. Three sources of wind data were used: historical measure-
ments (as in [6]) and scenarios A1 and C1. The impact from using different
datasets is quite dramatic, especially when considering low percentiles of step
changes for high wind penetrations. Note that positive extreme ramps of the
load are considerably larger in magnitude than negative ramps. The latter are
however affected to a larger degree by increasing the wind penetration.
As in Section 1.3, it can be worth mentioning that the difference in step
change magnitudes are not due to differences between modelling and mea-
surements per se; modelling with the historical distribution and characteristics
of WTs gives almost identical results as using measurements. The main lesson
to be learned is that measurements should not be used for studying the future
if one does not have good reasons to assume that CFs etc. will not change for
future farms5 .
As shown in Table. 4.4 on page 79, it has historically been uncommon with
generation above 0.7 p.u. and relatively uncommon with generation above 0.5
p.u. If a high instantaneous wind penetration is considered a problem, curtail-
5 With that said, it is of course not easy to foresee the geographical distribution and characteris-
tics of future farms. Scaling historical measurements implicitly assumes no change at all; most
likely scenarios can do considerably better than that.

86
 A1, SNSP = 0.60
A1, SNSP = 0.75

Wind power curtailment

A1, SNSP = 0.90
10%
C1, SNSP = 0.60
C1, SNSP = 0.75
C1, SNSP = 0.90

1%

0.1%
0 0.1 0.2 0.3 0.4 0.5
Wind power penetration

Figure 4.20. Curtailed wind energy for different average penetration levels (after cur-
tailment). The amount of curtailed wind depends heavily on the allowed system non-
synchronous penetration (SNSP) and the average capacity factor (0.31 for scenario A1
and 0.40 for C1).

ment of wind power is therefore a realistic and potentially an economically

rational option. The system non-synchronous penetration (SNSP) is defined
as [117]

Wind − HVDC imports

SNSP = . (4.2)
Demand − HVDC exports
In the following analysis, the curtailed wind energies necessary to keep the
SNSP below certain values are calculated. The analysis was simplified by not
considering HVDC links and by only studying Sweden. Measured loads for
year 2007 to 2014 and wind power scaled from scenarios A1 and C1 from
Paper D were used. For Ireland, allowed SNSPs in the range 60–75% are
anticipated for year 2020 [117]. In other systems, e.g. Portugal and Denmark,
higher instantaneous wind penetration have been observed [14]. The amount
of curtailed wind energy for allowed SNSPs of 60%, 75% and 90% and wind
penetration levels up to 50% are given in Fig. 4.20. Note that the penetration
levels are after curtailment and that the curtailed energy is given in relation to
the uncurtailed generation. For up to 20% average wind penetration, negligible
amounts of wind would have to be curtailed, regardless of the allowed SNSP.
With 50% penetration, the curtailed energy increases to high levels if not a
SNSP of 90% is allowed. The average CF also has a relatively strong impact
of the curtailed energy. Scenario A1 has a CF of 0.31, only slightly higher
than today, while scenario C1 has a CF of 0.40.
Wind power curtailment deals with the issue of high wind and low load.
We not turn to the opposite problem of system adequacy; what happens when
the load is high and wind power generates little? Will future farms be better

87
Table 4.6. Metrics for quantifying low wind / high load for historical wind power
measurements as well as time series for two scenarios (A1 and C1). Firm capacity is
the 10th percentile in accordance with the Swedish TSO’s guidelines. Firm capacities
for high load is given for the top five percent of the load. The wind power time series
were linearly scaled to 30 TWh/a or in average 3.4 GW.
Historical A1 C1
Firm cap. 0.89 GW 1.0 GW 1.2 GW
Firm cap. high load 0.88 GW 1.0 GW 1.2 GW
Max load - max net load 1.0 GW 1.2 GW 1.5 GW

at handling such situations? The preferred method [114] for determining the
“capacity value” of wind power is to compute the constant increase in load
that gives the same loss of load expectation (LOLE) as before wind power
was added. These calculations require information on accepted LOLE, forced
outage rates for conventional generators and import capacities. Since these
data were not available, three other metrics were computed:
1. Firm capacity, defined as the 10th percentile in accordance with the
Swedish TSO’s guidelines.
2. Firm capacity for the 5% of the time with highest loads.
3. The difference between maximum load and maximum net load.
As before, the computations were performed for three different wind power
time series: historical measurements and scenarios A1 and C1 from Paper D.
The time series were linearly scaled to 30 TWh/a or a mean generation of
3.4 GW. The CFs were 0.25, 0.31 and 0.40 respectively. As can be seen in
Table 4.6, the firm capacity is highest for scenario C1. Firm capacities are
very similar for high loads and for all data. The differences between maximum
loads and maximum net loads are around 20% higher than the firm capacities.
The latter metric is very sensitive to the wind generation at hours with very
high load. For individual years this metric varies between 0.8–3.2 GW for
all datasets. This is consistent with the high variations in capacity values,
see [114] for a few examples. For robust assessment of the contribution from
wind power to system adequacy it is thus valuable with long time series, e.g.
35+ years from reanalysis data.

4.3 Variability
Besides the development of wind power models (Papers I–IV), one review ar-
ticle and two studies on variability/correlations are included in this thesis (Pa-
pers V–VII). Some results and conclusions from the review are given below.
Results on wind power correlations in Europe are presented in Section 4.3.1.

88
 Solar Wind Wave Tidal
Jan

>1

0.8

Dec 0.6

Jan 0.4

0.2

Dec
00:00 12:00 00:00 12:00 00:00 12:00 00:00 12:00

Figure 4.21. Site specific examples of variability at two sites per energy source (solar
irradiation, wind speed, significant wave height and tidal current speed). The temporal
resolutions differ from one minute to one hour. The measurements are normalised to
the 98th percentile measured for each site.

Section 4.3.2, finally, gives results on net load variability in a highly or fully
renewable Nordic power system.
As a background for future studies, a review on variability and forecasta-
bility of non-dispatchable renewable energy sources was conducted in collab-
oration with solar, wave and tidal power researchers (Paper V). The aim of the
paper was to compare the methodologies used for the different sources. The
foci were primarily on temporal variability and the effects of aggregation and
not so much on spatial variability.
Although both wind and wave power is driven by energy from the sun, the
temporal characteristics of these three sources are strikingly different. Solar
irradiation has a clear maximum defined by geographical location and time.
Cloud movements can introduce an almost binary pattern, where the point
irradiation switches between almost zero and maximum. Because of this, re-
search focus is primarily on short time scales. The wind speed can also have
seasonal and diurnal patterns, although the strength of these vary from site to
site. Stochastic variations of the wind are substantial and it has been shown
that wind variations on the time scale of 1–6 hours can be challenging for
the power system. Waves are induced by the wind, but the variations are
smoothened out. The diurnal variations are therefore small. Tidal currents
finally are driven by the gravity of the moon, and follows very regular pat-
terns. Commonly, tides are semi-diurnal with two high tides and two low per
day. Fig. 4.21 shows examples of variability of the different sources during
one year.

89
 Of great importance is also the correlation of production from spatially dis-
tributed sources. High correlations imply large variations for a distributed
fleet of generation units, while lower correlations give a smoothing of the
aggregated power. The correlation of wind power often exhibit an approx-
imately exponential decay with separation distance. Averaging over longer
time-periods gives higher correlation, while the short-term variations are more
independent. The solar irradiance has a substantially higher correlation than
wind power for similar separation distances.
Forecasting of intermittent renewables becomes increasingly important as
the penetration levels increase. A lot of research and practical experience is
available, not the least for wind power. Forecasting systems, often using phys-
ical NWP models and statistical post-processing, are operational in several
countries, and the results have improved considerably over the past years. A
conclusion from the review was that the accuracy of the forecasts are often
hard to compare since different metrics are used for the different fields.
A general conclusion in Paper V was that more studies on combinations of
the sources would be desirable. The disciplines could also learn from each
other and benefit from the use of more unified methods and metrics.
In Papers VI and VII, generation time series were separated into frequency
components. This gave us the possibility to study the variability characteristics
in more depth. The same cut-off frequencies were used:
• Long-term component (T > 4 months)
• Mid-term component (2 weeks < T < 4 months)
• Mid/short-term component (2 days < T < 2 weeks)
• Short-term component (T < 2 days)
An example of a filtered wind power time series is shown in Fig 4.22. Ap-
propriately filtered time series have two desirable properties: the sum of the
frequency components add up to the input signal and the sum of the variances
add up to the variance for the original (raw) time series (see Section 3.5). As
an example, the SD of Swedish wind generation 2007–2014 was 0.17 p.u. The
contributions were 0.054, 0.12, 0.083 and 0.073 p.u. from the short, mid/short,
mid and long-term components6 . Fluctuations with periods ranging from two
days to two weeks are thus contributing most to the SD of the raw signal. The
same analysis for one and four hour step changes shows that the short-term
component is clearly dominating. If one is interested in more details on which
frequencies contributes mostly to the variance, a cumulative PSD plot can be
informative, see Section 3.4.1.

6 Note that standard deviations (SD) does not sum in the same way as variances; the SD of the
raw signal is, to a close approximation, the square root of the sum of the squared SDs of the
constituent components

90
 0.8 Short Mid/Short Mid Long
Raw
T<2d 2d<T<2w 2w<T<4m T>4m

0.6

Hourly generation [p.u.]

0.4

0.2

-0.2

0 1 2 3 4
Time [weeks]

Figure 4.22. Filtering of a wind power time series. The raw signal is decomposed into
four frequency components.

4.3.1 Wind power correlations

The correlation between wind power generation in different countries is im-
portant since it helps us quantify the reduction in aggregated variability when
electrically interconnecting the countries. In Paper VI, hourly, country-wise
time series of wind power output were generated for all European countries (or
groups of smaller countries) and linear correlation coefficients ρ were studied.
In order to deepen the analysis, ρ’s were not only computed for these time se-
ries, but also for one hour step changes and for band-pass filtered data.
As compared to the methodology outlined in Sections 3.2.1 and 3.2.2, a
simplified approach was taken to develop scenarios for year 2020 and model
hourly output. The methodology is not given here, see Paper VI for details.
Overall, the results were however (surprisingly) good when comparing to mea-
surements and it was concluded that the simplified method is adequate for
studying correlations. In particular, there is a good match in ρ’s for hourly
energy, step changes as well as for all band-pass filtered time series.
It is well known that the outputs from wind farms are correlated and that, in
general, ρ decreases with longer separation distances [36]. Exponential mod-
els are commonly used to describe this relationship, see Eq. (3.8) on page 39.
Most often, correlations between outputs of individual farms or wind speed
measurements have been studied. The same analysis can be performed with
nationally or regionally aggregated wind power time series [37, 41].
Pair-wise correlations versus distance and an exponential fit to the data are
shown in Fig. 4.23a. Apparently, the exponential model is adequate (also)
for country-wise generation. Fig. 4.23b shows smoothing spline fits of cor-
relations versus distance for hourly energy, one hour step change and long,
mid, mid/short and short-term components respectively. The correlations be-
tween the long-term components are generally high. This is expected since the

91
Figure 4.23. Correlations between wind power generation in the European countries
depending on separation distance. (a) Pair-wise correlations and an exponential fit to
the data (hourly energy). (b) Smoothing spline fits of correlations for hourly energy,
one hour step change and long, mid, mid/short and short-term components respec-
tively. The dashed blue line shows the correlations between long-term components
when excluding Turkey and Portugal.

seasonal wind patterns in Europe are similar, with higher generation in winter-
time. The spread in the results is however very large; ρ’s for countries 2000–
2500 km apart, for instance, vary between 0.17 and 0.83. The largest part of
the spread can be attributed to the much lower correlations seen for Portugal
and Turkey, which have distinctly different seasonal patterns than the rest of
Europe. For fluctuations with shorter periods, ρ gradually decreases. In par-
ticular, the short-term variations are even less correlated than the step changes.
In comparison with the results for individual wind farms (see e.g. [41]), ρ is
markedly higher, especially for step changes.
For a single country, the correlations can be visualised in a map. Maps for
all countries are provided in Paper VI. Correlation maps for Sweden are shown
in Fig. 4.24.

4.3.2 A fully renewable Nordic power system

In Paper VII, the net load variability in the Nordic power system was stud-
ied. Different time scales and various shares of solar, wind, wave and tidal
power were considered. The hourly net load was calculated from historical
consumption for year 2010 through 2014 and meteorological resource data,
using scenarios for likely deployment across the Nordic countries. For wind
power, the methods described in Sections 3.2.1 and 3.2.2 were used to generate
scenarios and model hourly generation.
The principal research question was how the net load variability will in-
crease for a highly and fully renewable power system, assuming that new re-
newables replace fossil and nuclear generation on an equal energy basis. Fur-

92
Figure 4.24. Correlation between wind generation in Sweden and other European
countries. (a) Hourly energy, (b) one hour step change, (c) long-term component, (d)
mid-term component, (e) mid/short-term component, (f) short-term component. The
correlations are highest for the long-term components and lowest for the short-term
components. Similar maps for all countries are provided in Paper VI.

93
thermore, it was studied if a wise combination of the sources can reduce the
variability. As for the correlations described in Section 4.3.1, the analysis was
performed for hourly energy, one hour step changes and the four frequency
components specified on page 90. The net load (NL) was defined as

NL = load − IRE − thermal − nuclear = hydro + imports − exports, (4.3)

where IRE is short for intermittent renewable energy7 . In the highly renewable
scenario, 20% of the annual energy was obtained from IRE sources. All fossil
and 30% of the nuclear generation were assumed to be dismantled. In the
fully renewable scenario, all fossil and nuclear generation was replaced with
renewables, resulting in 36% energy from IRE sources.
As already seen in Fig. 4.21 on page 89, PV, wind, wave and tidal have
different generation patterns. Fig. 4.25 shows the SDs of the four frequency
components for the considered IRE sources and for the electric load. The dis-
tinctly different variability characteristics are now even more obvious. The
overall variability is lowest for the load, relatively similar for wind, wave
and tidal and largest for PV. Different components are dominating for all four
sources. For PV, the short-term component accounts for 80% of the SD of the
hourly generation time series (σraw ). For wind, the mid/short-term component
is dominating, for wave and load, long-term variability is most prominent and
for tidal, the mid-term component accounts for 54% of σraw .
As shown in Eq. (3.7) on page 39, the variability of the net load can be
explained by the weight and SD of each time series, and the correlations be-
tween them. In order to obtain a low net load variability, it is desirable with
high correlations to the load and low correlation between the sources. Tidal
generation is close to uncorrelated to the other sources and the load for all fre-
quency components. PV has a strong negative correlation (between -0.83 and
-0.75) to wind, wave and load for the long-term component. PV is, in contrast,
positively correlated (0.42) to the load for the short-term components. Adding
smaller amounts of PV will thus reduce the short-term component of the net
load and wind and wave deployment will reduce the long-term component.
However, after a certain point the net load variability is increased due to the
relatively large SDs of output from the IRE sources. As an example, there is a
short-term variability minimum when PV is supplying 2–3% of the load. For
high shares of PV, the variability can be very large. Wind, wave and load are
strongly correlated (0.74–0.93) to each other for the long-term components.
For each scenario (highly and fully renewable systems), both predefined
mixes and optimised shares of IRE were considered. Mix 1 was a reference
mix similar to today’s, only slightly more PV (90% of energy from wind
7 Note that this definition is different than that normally used.
The reason for subtracting nuclear
and thermal generation from the load was that we assumed the generation patterns from these
sources to be fixed.

94
 1.6
Contribution to
1.4 Long
σ raw = 80%
Mid

Standard deviation (normalised)

1.2 Mid/Short
Short
1

0.8
48%
54%
0.6 47%

0.4
72%

0.2

0
Load PV Wind Wave Tidal
σ raw = (0.21) (1.76) (0.60) (0.79) (0.67)

Figure 4.25. Standard deviations for the frequency components, normalised to mean
load/generation. For the IRE sources, results for the 3 TWh scenarios are given (see
Paper VII). The contributions to the raw σ from the dominating frequency components
are indicated in the figure. The normalised σ of the raw signals are given in the
horizontal axis labels.

and 10% from PV). Mix 2 was more futuristic with maximum tidal energy
(3 TWh/year) and with PV, wind and wave accounting for 40%, 40% and 20%
of the remaining respectively. The optimisations were carried out in order
to reduce the net load SDs for each frequency component (σLT , σMT , σMST
and σST ) as well as for the raw data (σraw ). As can be anticipated from the
SDs for and correlations between the IRE sources, the optimised mixes varied
heavily depending on the objective function (Fig. 2 in Paper VII). Most mixes
comprised the maximum allowed tidal generation. The mixes optimised for
low σraw , σLT and σST contained little or no PV. When σMT or σMST were
minimised, PV shares of around 50% were however obtained.
The resulting net load variability is shown in Fig. 4.26. The variability
is given in terms of SDs of raw data (hourly energy time series), the four fre-
quency components and one hour step changes. As an example, let us consider
the raw net load in the scenario with 20% IRE, i.e. the leftmost group of bars in
Fig. 4.26. With mix 1 (90% wind and 10% PV), the SD is 7.6 GW. When the
IRE shares were optimised to obtain a low σraw , the corresponding figure is
7.0 GW. Minimising σMT leads to a high share of PV, resulting in σraw = 11.4
GW. These figures should be compared to 5.9 GW for the present generation
portfolio (dashed line). An increased net load variability is thus inevitable,
no matter how the IRE sources are combined. The magnitude of the increase
depends heavily on the IRE mix.
Two important conclusions can be drawn. Firstly, no matter how the IRE
sources are combined, a large increase will be seen in σMT and σMST , i.e. fluc-
tuations with periods ranging from two days to four months. More studies are

95
 Mix optimised for low σ in
Mix 1 Mix 2 Raw Long Mid Mid/Short Short Today

16 16
Scenario 1: Highly renewable Scenario 2: Fully renewable
Net load standard deviation [GW]

14 14

12 12

10 10

8 8

6 6

4 4

2 2

0 0
σ raw σ LT σ MT σ MST σ ST σ ∆P σ raw σ LT σ MT σ MST σ ST σ ∆P

Figure 4.26. Net load standard deviations of different kind for scenario 1 and 2 com-
pared to today’s situation. The different colours indicate the mix of intermittent re-
newables (either predefined or optimised). Step changes are denoted ∆P.

needed in order to answer whether the remaining system, most importantly hy-
dro power, will be able to balance these fluctuations, taking into account both
technical and environmental constraints. Secondly, the improvements from
using optimised mixes over mix 1 are relatively small. A small overproduc-
tion and corresponding curtailment of mix 1 can give similar net load SDs as
for the optimised mixes. The system benefits from combining the sources are
thus limited and consequently the IRE sources will have to compete mainly
with their e/MWh costs. The “overproduction stategy” [118–120] can also be
employed to obtain an even lower net load SD, e.g. similar to that today, see
Fig. 4 in Paper VII.
Time series for the current situation and highly/fully renewable scenarios
for year 2014 are shown in Fig. 4.27. For the latter, the IRE mixes optimised
for low σraw were used. For the fully renewable scenario, the net load some-
times becomes very low and, seldomly, even negative. If export is not possible
at these occasions, curtailment of IRE might be necessary. The maximum net
loads are 6.5 and 9.0 GW higher than today’s 44 GW for the highly and fully
renewable scenarios respectively. For the reference mix (90% wind and 10%
PV), the same figures are 7.0 and 11.2 GW. In other words, the maximum ca-
pacity of hydro power, import-export and any new peak capacity such as gas
turbines needs to be significantly higher than the maximum utilised capacity
we have seen over the last five years. The energy that needs to be provided by
this additional (or currently unused) capacity is however below 1 TWh/year.

96
 Nuclear Thermal IRE Net load
60
Current situation
Power [GW]

40

20

0
60
Highly renewable
Power [GW]

40

20

0
60
Fully renewable
Power [GW]

40

20

0
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Figure 4.27. Contribution from nuclear, thermal, intermittent renewables and net load
(hydro + import - export) for the year 2014. For the highly and fully renewable sce-
narios, results for the mixes optimised to reduce σraw are shown.

97
5. Concluding discussion

This thesis contains seven papers with two main foci: Papers I–IV deal with
the modelling of wind power generation and forecasts and Papers V–VII with
the variability of intermittent renewables.
The most important conclusion from the former group of papers is that
coarse reanalyses are highly suitable for modelling, but only for generation on
the power system level. Statistical post-processing is also necessary, e.g. for
obtaining correct magnitudes of the high-frequency fluctuations and appropri-
ate correlations of forecast errors. Since physical models based on reanalysis
data perform so well (if used properly), linear scaling of historical generation
and purely statistical models should, in my opinion, be avoided for studies of
the future. If historical wind power measurements are used, a severe overesti-
mation of the challenges and costs of wind integration may result; future WTs
will have higher CFs than today’s and thus give lower variability. An impor-
tant reason for not using purely statistical models is that wind and load data
should be synchronous.
Nevertheless, a few potential issues with the “reanalysis approach” are
worth pointing out:
• Coarse reanalyses should not be used for determining mean wind speeds.
The time series should rather be scaled according to mean wind speeds
from highly resolved maps or estimated from energy yields of future
farms.
• Coarse reanalysis data are not suitable for determining possible or likely
sites for wind power deployment. Planned farms or highly resolved wind
maps are better options.
• The models should always be validated with measurements. If system-
level output is desired, system-level measurements should be used for
tuning the model parameters.
• Statistical post-processing of the output from reanalysis based models
might be necessary, e.g. if the time series are too smooth.
When developing scenarios of future farms, the most influential factors for
the combined variability are the average CF and the geographical distribution.
Offshore farms affect both these factors and thus require special attention. For
a given set of farm characteristics, the variability is saturated relatively quickly.
This implies that linear scaling of the time series are acceptable for studying
e.g. different penetration levels.
In Papers I–IV, a lot of details and extra features were considered. For prac-
tical purposes, the models can be simplified with little loss of performance. In

98
the basic model (Paper I), it is recommended to omit time-varying air density
and wind shear coefficients. In the improved model (Paper II), EOFs as predic-
tors are, for most purposes, not worth the effort and also make the model more
system specific. The search for phase angles in Papers II and III is however
recommended, since this improves the representation of the varying volatility
significantly.
An important conclusion from the variability studies in Papers V–VII is that
intermittent renewables and the load have very different variability characteris-
tics in different frequency bands. Wind power seasonality is highly correlated
between the European countries. Fortunately, the load has a similar seasonal
profile as wind power, at least for the northern countries. Even at high pen-
etration rates, there will thus be no, or only a small, increase in the need for
seasonal storage. Less fortunately, wind power in the Nordic system produces
slightly less than average during the peak-load hours, so additional peak ca-
pacity (e.g. gas turbines) and/or load flexibility is likely needed for a highly
renewable system.
Fluctuations with periods between a few days and a few months are domi-
nant for wind power. Since these fluctuations are not so highly correlated be-
tween countries, interconnection can reduce the variability significantly. For
the Nordic region, net load fluctuations of this type will increase strongly for
high penetrations of intermittent renewables, no matter how the sources are
combined. More studies, taking into account e.g. forecast uncertainties and en-
vironmental regulations, are necessary for determining whether hydro power
will be able to balance these fluctuations. High-frequency variability (e.g. di-
urnal fluctuations and hourly step changes) are lowly correlated and not very
prominent for wind power. With a suitable mix of renewables, there will be no
increase in net load fluctuations on this time scale, even for a fully renewable
power system.
Our work so far has been focused on modelling of renewables and studies
of variability. One must therefore be careful in drawing conclusions on the
impact of more wind power on the power system and possible/desirable levels
of wind penetration. It can however be noted that the increase in e.g. net load
step changes and required curtailments in order to not exceed certain instanta-
neous penetration levels are non-linear functions of the wind penetration. The
challenges related to e.g. a 30% wind penetration are therefore considerably
smaller than those for a 50% penetration.
According to our studies, higher CFs, more offshore wind power, more op-
timal geographical distribution and overproduction/curtailment would be ben-
eficial for the system in terms of lower variability and/or lower FEs. Especially
for higher wind penetrations, it might thus be rational to use policy instruments
to steer the development in this direction. Before such measures are recom-
mended, the socio-economic implications must be assessed, see suggestions
for future work in the subsequent chapter.

99
6. Future work

Now that the model package for generating time series of wind power pro-
duction and forecasts is completed, the next logical step is to focus on the
implications for the power system from an increased share of wind and other
intermittent renewables. The overarching goal is to answer whether a fully
renewable Nordic system is feasible and to investigate suitable strategies for
facilitating the integration of renewables. Two ongoing studies are however
planned to first be completed: one on the possible deterioration of wind farm
performance over time (similar to [23]) and one on comparing different re-
analyses for modelling wind power.
A highly interesting research topic is to evaluate different strategies for mit-
igating the variability of renewable energy. In [39], it was shown that “ad-
vanced” WTs with lower specific rating and higher hub heights are beneficial
from a system point of view and increase the revenues for wind farm owners.
In this thesis, it has been demonstrated that higher CFs leads to both lower
variability and lower FEs. Other possible strategies include wise combina-
tions of renewables (see Paper VII), more offshore wind power, overproduc-
tion/curtailment and a better geographic distribution of farms, e.g. by procur-
ing farms at strategic locations.
In order to fairly assess the abovementioned strategies from both a techni-
cal and socio-economical perspective, and to compare these to grid reinforce-
ments, investments in new peak generation etc., a lot of work will become nec-
essary. A transmission grid model of the studied area needs to be developed
as well as models for hydropower and the future electric load. Furthermore,
the electric market needs to be accounted for, perhaps through a collaboration
with energy economists.

100
7. Summary of papers

Paper I
Modelling the Swedish wind power production using MERRA reanalysis
data
Description and validation of a model for hourly, aggregated wind power gen-
eration in Sweden. The model uses MERRA reanalysis data and information
on WTs and takes into account e.g. power curve smoothing and direction de-
pendent losses. When comparing model output to measurements, the errors
were low, the correlation was high and the distributions were very similar.
The author collected the data, developed the model and wrote the paper.
Published in Renewable Energy, 76: 717–725, 2015.

Paper II
Restoring the missing high-frequency fluctuations in a wind power model
based on reanalysis data
Since the above-mentioned model (Paper I) understimated the high-frequency
fluctuations, a statistical supplement model was developed. By using fre-
quency domain and machine learning techniques, the spectrum and step change
distribution of the model output became very similar to measurements at the
expense of a small increase in the RMS error. Furthermore, the varying volatil-
ity of the high-frequency fluctuations was captured relatively well. The author
collected the data, developed the model and wrote the paper.
Published in Renewable Energy, 96: 784–791, 2016.

Paper III
Simulating intra-hourly wind power fluctuations on a power system level
For some power system studies, wind power data with higher resolution than
hourly are desirable. This paper describes and validates a model for simulation
of intra-hourly fluctuations, given that hourly data are available. It was shown
that a model trained on data from one power system can be used for simulat-
ing realistic fluctuations for another system, e.g. in terms of distribution and
varying volatility of the deviations from hourly means. The author collected
the data, developed the model in collaboration with the co-authors and wrote
the paper.
Resubmitted after revisions to Wind Energy, August 2016.

101
Paper IV
A new approach to obtain synthetic wind power forecasts for integration
studies
For higher penetration levels, the forecast errors of wind power can be sig-
nificant and must be accounted for in wind integration studies. This paper
describes a new approach for obtaining “synthetic” wind power forecasts with
a horizon of up to one week or more. Unlike earlier methodologies, freely
available, global reforecasts (from the GEFS dataset) were used. In order
to produce realistic forecasts, some statistical post-processing was employed.
Validation with aggregated observations from Belgium and three individual
farms in Sweden proved the adequacy of the approach. The author collected
some of the data, developed the model and wrote the paper.
Resubmitted after revisions to Energies September, 2016.

Paper V
Variability Assessment and Forecasting of Renewables: A Review for
Solar, Wind, Wave and Tidal Resources
Very large amounts of research has been conducted on the variability and fore-
castability of renewable energy sources, in particular for solar and wind power.
The questions studied and methods employed differs between the sources.
This review attempts to summarize and compare the research done so far, and
points out research gaps and areas of possible learning between the fields. The
author collected and read the references related to wind energy and wrote the
corresponding text. The author also contributed to merging the text and in dis-
cussing the scope and results.
Published in Renewable & Sustainable Energy Reviews, 44: 356-375, 2015.

Paper VI
Correlation between wind power generation in the European countries
The correlations of wind power generation in different countries are impor-
tant since they give information on the benefits of interconnection. Country-
wise time series were generated for a European scenario for year 2020. It
was shown that the correlation decreases in an exponential fashion with dis-
tance. In order to deepen the analysis, the time series were band-pass filtered.
Long-term components (seasonal variations) were highly correlated for most
countries, but fluctuations with shorter periods were less correlated. The au-
thor collected the data, performed the analyses and wrote the paper.
Published in Energy, 114: 663–670, 2016.

102
Paper VII
Net load variability in the Nordic countries with a highly or fully renew-
able power system
Scenarios for the Nordic power system were developed in which intermittent
renewables (PV, wind, wave and tidal) replaced fossil and nuclear generation.
Different mixes, predefined or optimised, of the renewables were considered.
The seasonal and diurnal fluctuations that hydropower needs to balance do
not necessarily increase, even for a fully renewable system. Fluctuations with
intermediate periods (days to months) will however inevitable increase sig-
nificantly, no matter how the renewables are combined. The author modelled
wind power generation, did the analyses of combinations of sources and wrote
most of the paper.
Resubmitted after revisions to Nature Energy, August 2016.

103
8. Svensk sammanfattning

Under det senaste årtiondet har stora mängder vindkraft installerats i många
länder. Denna utveckling har drivits av t.ex. lägre kostnader och miljöhänsyn
och förväntas fortsätta i framtiden. När vindkraft och andra intermittenta för-
nybara källor (IRE) börjar försörja en större andel av elförbrukningen ifråga-
sätts ibland om den inneboende variabiliteten och oförutsägbarheten hos dessa
kan hanteras. För att studera påverkan på kraftsystemet från en högre andel
IRE utförs integrationsstudier. Realistiska tidsserier av vindkraftsproduktion
och -prognoser är viktiga beståndsdelar i dessa studier. Tidsserierna ska helst
vara synkrona med förbrukningsdata, vara flera år långa och ha en tillräckligt
hög tidsupplösning.
Modellpaketet som beskrivs och utvärderas i uppsatserna I–IV ger en hel-
täckande metodik för att simulera vinddata för kraftsystemstudier. I uppsats I
beskrivs den grundläggande modellen för att generera tidsserier av vindkraft-
produktion utifrån information om vindkraftsparker samt grovupplösta mete-
orologiska data. Genom att jämföra den modellerade, timvisa produktionen
för Sverige med uppmätta data visar vi att metoden fungerar tillfredställande;
korrelationen är 0.98 och fördelningarna mycket likartade. Storleken på hög-
frekventa fluktuationer underskattas dock vilket motiverade oss att utveckla en
statistisk efterbehandlingsmodul vilken beskrivs och utvärderas i uppsats II.
För vissa typer av studier är det önskvärt med data av högre upplösning än en
timme. Eftersom inga sådana mätningar finns tillgängliga på aggregerad ni-
vå i Sverige samt att fritt tillgängliga meteorologiska modeller som bäst har
upplösningen en timme utvecklades en statistisk modell för att kunna simulera
stokastiska inomtimmen-variationer. Vi visar i uppsats III att modellen ger bra
resultat även om den tränas på data från ett kraftsystem och appliceras på ett
annat. Slutligen presenteras i uppsats IV en ny metodik för att generera “synte-
tiska” vindkraftsprognoser baserade på fritt tillgängliga data från grovupplösta
meteorologiska modeller.
Den viktigaste slutsatsen från ovanstående studier är att grovupplösta me-
teorologiska återanalyser är högst lämpliga, men bara för att modellera den
aggregerade produktionen i ett tillräckligt stort område, exempelvis ett land.
En viss statistisk efterbehandling är också nödvändigt för att t.ex. få en bra
återgivning av högfrekventa fluktuationer och lämpliga korrelationer av pro-
gnosfel. Fördelar med vår metodik inkluderar i) en fysisk koppling till väder-
data och vindkraftverkens egenskaper, ii) över 30 år tidsserier med upplösning
5–60 minuter, iii) fritt och globalt tillgängliga indata samt iv) beräkningstider
i storleksordningen minuter. I denna avhandling argumenterar jag för att vår

104
metodik generellt är att föredra framför att använda t.ex. rent statistiska mo-
deller eller linjär skalning av historiska mätningar. Ett antal fallgropar finns
dock vid användningen av återanalyser.
I variabilitetsstudierna som presenteras i uppsatserna V–VII studeras elpro-
duktion från sol, vind, havsvågor och tidvatten. En översikt av forskningsläget
(uppsats V) visar, föga förvånande, att forskningen kommit längst inom sol-
och vindkraft. En mängd studier finns av variabiliteten och prognostiserbar-
heten för dessa källor, både separat och i kombination. Fördelar finns av att
kombinera IRE och sprida produktionen geografiskt eftersom variabiliteten då
minskar, men mcket få studier analyserar alla fyra källorna tillsammans. Det
skulle kunna vara fördelaktigt för de olika forskningsfälten att lära av varandra
samt att använda mer konsistenta metoder och mått.
En viktig slutsats från uppsatserna VI och VII är att de olika IRE-källorna
samt elförbrukningen har väldigt olika karaktäristik i olika frekvensband. Be-
roende på storleken på och korrelationen av dessa fluktuationer kommer oli-
ka tidsskalor bli mer eller mindre utmanande att balansera i ett kraftsystem
med mycket förnybart. Vindkraften har liknande säsongsmönster i större de-
len av Europa vilket innebär att även en helt sammankopplad kontinent skulle
ha betydande variationer av produktionen på denna tidsskala. Korrelationerna
minskar dock ju kortare tidsskalor man studerar. Fluktuationerna med perioder
kortare än två dygn är t.ex. mycket lågt korrelerade även för närliggande län-
der och variabiliteten minskar följdaktligen effektivt genom sammankoppling.
I ett helt förnybart nordiskt kraftsystem kommer det bli ingen eller endast en
liten ökning av behovet för säsongslagring givet en lämplig mix av IRE. Fluk-
tuationer med perioder mellan några dagar och några månader är dominanta
för vindkraften och variationerna av denna typ som vattenkraften behöver ba-
lansera kommer att öka betydligt, oavsett hur de förnybara källorna kombi-
neras. Med en lämplig mix av förnybart kommer däremot variationerna inom
dygnet inte öka överhuvudtaget ens för ett helt förnybart system. En lovande
strategi för att minska variabiliteten är att bygga mer IRE och sedan strypa
produktionen vid tillfällen när den inte behövs.
Enligt våra studier leder högre kapacitetsfaktorer och mer havsbaserad vind-
kraft till lägre variabilitet och mindre prognosfel. I synnerhet för höga andelar
vindkraft skulle det därför kunna vara rationellt att använda styrmedel för att få
till stånd en sådan utveckling. En intressant inriktning för framtida forskning
är att jämföra olika strategier för att minska variabiliteten hos de intermittenta
källorna. För att kunna besvara om dessa är samhällsekonomiskt försvarbara
och hur väl de står sig i konkurrens med t.ex. ökad utbyggnad av balanskraft
och utlandsförbindelser krävs samarbeten med vattenkrafts- och elmarknads-
forskare.

105
References

[1] P. W. Carlin, A. S. Laxson, and E. B. Muljadi, “The history and state of the art
of variable-speed wind turbine technology,” Wind Energy, vol. 6, no. 2, 2003.
[2] J. Olauson, “Wind power and natural disasters,” Licentiate Thesis, Uppsala
University, 2014, ISSN 0349-8352; 337-14L.
[3] “Global wind energy outlook 2014,” Global Wind Energy Council, Tech. Rep.,
2014.
[4] I. Nohlgren, S. Herstad Sväd, M. Jansson, and J. Rodin, “El från nya och
framtida anläggningar 2014,” Elforsk, Tech. Rep. 14:40, 2014.
[5] H. Holttinen, P. Meibom, A. Orths, B. Lange, M. O’Malley, J. O. Tande,
A. Estanqueiro, E. Gomez, L. Söder, G. Strbac, J. C. Smith, and F. van Hulle,
“Impacts of large amounts of wind power on design and operation of power
systems, results of IEA collaboration,” Wind Energy, vol. 14, no. 2, pp.
179–192, 2011.
[6] J. Kiviluoma, H. Holttinen, D. Weir, R. Scharff, L. Söder, N. Menemenlis,
N. A. Cutululis, I. Danti Lopez, E. Lannoye, A. Estanqueiro,
E. Gomez-Lazaro, Q. Zhang, J. Bai, Y.-H. Wan, and M. Milligan, “Variability
in large-scale wind power generation,” Wind Energy, In print, DOI:
10.1002/we.1942.
[7] L. Hirth, “The market value of variable renewables: The effect of solar wind
power variability on their relative price,” Energy economics, vol. 38, pp.
218–236, Jul. 2013.
[8] P. C. Bhagwat, L. J. de Vries, and B. F. Hobbs, “Expert survey on capacity
markets in the US: Lessons for the EU,” Utilities Policy, vol. 38, pp. 11–17,
Feb. 2016.
[9] M. Huber, D. Dimkova, and T. Hamacher, “Integration of wind and solar
power in Europe: Assessment of flexibility requirements,” Energy, vol. 69, pp.
236–246, May 2014.
[10] K. Schaber, F. Steinke, and T. Hamacher, “Transmission grid extensions for the
integration of variable renewable energies in Europe: Who benefits where?”
Energy Policy, vol. 43, pp. 123–135, 2012.
[11] P. Tielens and D. Van Hertem, “The relevance of inertia in power systems,”
Renewable and Sustainable Energy Reviews, vol. 55, pp. 999–1009, Mar. 2016.
[12] A. K. Pathak, M. P. Sharma, and M. Bundele, “A critical review of voltage and
reactive power management of wind farms,” Renewable and Sustainable
Energy Reviews, vol. 51, pp. 460–471, Nov. 2015.
[13] J. Fortmann, R. Pfeiffer, E. Haesen, F. van Hulle, F. Martin, H. Urdal, and
S. Wachtel, “Fault-ride-through requirements for wind power plants in the
ENTSO-E network code on requirements for generators,” IET Renewable
Power Generation, vol. 9, no. 1, pp. 18–24, Jan. 2015.

106
[14] L. Söder, “På väg mot en elförsörjning baserad på enbart förnybar el i Sverige
- En studie om behov av reglerkraft och överföringskapacitet (Version 4.0),”
KTH, Tech. Rep., 2014.
[15] J. Bruce, S. Holmer, A. Badano, L. Söder, S. Larsson, N. Dahlbäck, J. Bladh,
J. Lönnberg, L. Göransson, B. Rydén, H. Sköldberg, T. Unger, and S. Montin,
“Reglering av kraftsystemet med ett stort inslag variabel produktion,” North
European Power Perspectives (NEPP), Tech. Rep., 2016.
[16] IRENA, “The power to change: Solar and wind cost reduction potential to
2025,” International Renewable Energy Agency, Tech. Rep. ISBN
978-92-95111-97-4 (PDF), 2016.
[17] C. Potter, D. Lew, J. McCaa, S. Cheng, S. Eichelberger, and E. Grimit,
“Creating the dataset for the western wind and solar integration study
(U.S.A.),” Wind Engineering, vol. 32, pp. 325–338, Jun. 2008.
[18] H. Holttinen, “Expert group report on recommended practices: 16. Wind
integration studies, 1st edition,” IEA Wind, Tech. Rep., Sep. 2013.
[19] H. Holttinen, M. O’Malley, J. Dillon, D. Flynn, A. Keane, H. Abildgaard, and
L. Söder, “Steps for a complete wind integration study,” in 46th Hawaii
International Conference on System Sciences, 2013.
[20] H. Holttinen, J. Kiviluoma, A. Robitaille, N. Cutululis, A. Orths, F. van Hulle,
I. Pineda, B. Lange, M. O’Malley, J. Dillon, E. Carlini, C. Vergine, J. Kondoh,
M. Gibescu, J. Tande, A. Estanqueiro, E. Gomez, L. Söder, J. Smith,
M. Milligan, and D. Lew, “Design and operation of power systems with large
amounts of wind power, Final summary report, IEA WIND Task 25, Phase
three 2012–2014,” VTT Technology 75, Espoo, Finland, Tech. Rep., 2013.
[21] M. M. Rienecker, M. J. Suarez, R. Gelaro, R. Todling, J. Bacmeister, E. Liu,
M. G. Bosilovich, S. D. Schubert, L. Takacs, G.-K. Kim, S. Bloom, J. Chen,
D. Collins, A. Conaty, A. da Silva, W. Gu, J. Joiner, R. D. Koster, R. Lucchesi,
A. Molod, T. Owens, S. Pawson, P. Pegion, C. R. Redder, R. Reichle, F. R.
Robertson, A. G. Ruddick, M. Sienkiewicz, and J. Woollen, “MERRA:
NASA’s modern-era retrospective analysis for research and applications,”
Journal of Climate, vol. 24, pp. 3624–3648, Jul. 2011.
[22] D. P. Dee, S. M. Uppala, A. J. Simmons, P. Berrisford, P. Poli, S. Kobayashi,
U. Andrae, M. A. Balmaseda, G. Balsamo, P. Bauer, P. Bechtold, A. C. M.
Beljaars, L. van de Berg, J. Bidlot, N. Bormann, C. Delsol, R. Dragani,
M. Fuentes, A. J. Geer, L. Haimberger, S. B. Healy, H. Hersbach, E. V. Hólm,
L. Isaksen, P. Kållberg, M. Köhler, M. Matricardi, A. P. McNally, B. M.
Monge-Sanz, J.-J. Morcrette, B.-K. Park, C. Peubey, P. de Rosnay, C. Tavolato,
J.-N. Thépaut, and F. Vitart, “The ERA-Interim reanalysis: configuration and
performance of the data assimilation system,” Quarterly Journal of the Royal
Meteorological Society, vol. 137, no. 656, 2011.
[23] I. Staffell and R. Green, “How does wind farm performance decline with age?”
Renewable Energy, vol. 66, pp. 775–786, Jun. 2014.
[24] M. L. Kubik, D. J. Brayshaw, P. J. Coker, and J. F. Barlow, “Exploring the role
of reanalysis data in simulating regional wind generation variability over
northern ireland,” Renewable energy, vol. 57, pp. 558–561, Sep. 2013.
[25] D. Heide, L. von Bremen, M. Greiner, C. Hoffmann, M. Speckmann, and
S. Bofinger, “Seasonal optimal mix of wind and solar power in a future, highly

107
 renewable Europe,” Renewable Energy, vol. 35, no. 11, pp. 2483–2489, Nov.
2010.
[26] G. B. Andresen, A. A. Søndergaard, and M. Greiner, “Validation of Danish
wind time series from a new global renewable energy atlas for energy system
analysis,” Energy, vol. 93, Part 1, pp. 1074–1088, Dec. 2015.
[27] D. R. Drew, D. J. Cannon, D. J. Brayshaw, J. F. Barlow, and P. J. Coker, “The
impact of future offshore wind farms on wind power generation in Great
Britain,” Resources, vol. 4, no. 1, pp. 155–171, Mar. 2015.
[28] I. Staffell and S. Pfenninger, “Using bias-corrected reanalysis to simulate
current and future wind power output,” Energy, vol. 114, pp. 1224–1239, Nov.
2016.
[29] F. van Hulle, “Integrating wind: Developing Europe’s power market for the
large-scale integration of wind power,” TradeWind, Tech. Rep., 2009.
[30] S. Rose and J. Apt, “Generating wind time series as a hybrid of measured and
simulated data,” Wind Energy, vol. 15, pp. 699–715, 2012.
[31] L. Söder, “Integration study of small amounts of wind power in the power
system,” KTH, Tech. Rep. Trita-EES-9401, 1994.
[32] J. Apt, “The spectrum of power from wind turbines,” Journal of Power
Sources, vol. 169, no. 2, pp. 369–374, Jun. 2007.
[33] W. C. Skamarock, “Evaluating mesoscale NWP models using kinetic energy
spectra,” Monthly Weather Review, vol. 132, pp. 3019–3032, 2004.
[34] R. J. Davy, M. J. Woods, C. J. Russell, and P. A. Coppin, “Statistical
downscaling of wind variability from meteorological fields,” Boundary-Layer
Meteorology, vol. 135, no. 1, pp. 161–175, Jan. 2010.
[35] N. Ellis, R. Davy, and A. Troccoli, “Predicting wind power variability events
using different statistical methods driven by regional atmospheric model
output,” Wind Energy, vol. 18, no. 9, pp. 1611–1628, Sep. 2015.
[36] C. M. St. Martin, J. K. Lundquist, and M. A. Handschy, “Variability of
interconnected wind plants: correlation length and its dependence on
variability time scale,” Environmental Research Letters, vol. 10, no. 4, p.
044004, 2015.
[37] H. Louie, “Correlation and statistical characteristics of aggregate wind power
in large transcontinental systems,” Wind Energy, vol. 17, no. 6, pp. 793–810,
2014.
[38] E. Fertig, J. Apt, P. Jaramillo, and W. Katzenstein, “The effect of long-distance
interconnection on wind power variability,” Environmental research letters,
vol. 7, pp. 1–6, 2012.
[39] L. Hirth and S. Müller, “System-friendly wind power: How advanced wind
turbine design can increase the economic value of electricity generated
through wind power,” Energy Economics, vol. 56, pp. 51–63, May 2016.
[40] P. Sørensen, N. A. Cutululis, A. Vigueras-Rodríguez, H. Madsen, P. Pinson,
L. E. Jensen, J. Hjerrild, and M. Donovan, “Modelling of power fluctuations
from large offshore wind farms,” Wind Energy, vol. 11, pp. 29–43, 2008.
[41] H. Holttinen, “Hourly wind power variations in the Nordic countries,” Wind
Energy, vol. 8, no. 2, pp. 173–195, 2005.
[42] J. Olsson, L. Skoglund, F. Carlsson, and L. Bertling, “Future wind power
production variations in the Swedish power system,” in Innovative Smart Grid

108
 Technologies Conference Europe (ISGT Europe), 2010 IEEE PES, 2010, pp.
1–7.
[43] R. Billinton, H. Chen, and R. Ghajar, “Time-series models for reliability
evaluation of power systems including wind energy,” Microelectronics
Reliability, vol. 36, no. 9, pp. 1253–1261, Sep. 1996.
[44] G. Papaefthymiou and B. Klöckl, “MCMC for wind power simulation,” IEEE
Transactions on Energy Conversion, vol. 23, pp. 234–240, 2008.
[45] M. Milligan, E. Ela, D. Lew, D. Corbus, Y.-h. Wan, and B. Hodge,
“Assessment of simulated wind data requirements for wind integration
studies,” IEEE Transactions on Sustainable Energy, vol. 3, pp. 620–626, Oct.
2012.
[46] S. Eriksson, “Direct driven generators for vertical axis wind turbines,” PhD
Thesis, Uppsala University, 2008, ISBN 978-91-554-7264-1.
[47] F. Bülow, “A generator perspective on vertical axis wind turbines,” PhD
Thesis, Uppsala University, 2013, ISBN 978-91-554-8642-6.
[48] A. Goude, “Fluid mechanics of vertical axis turbines: Simulations and model
development,” PhD Thesis, Uppsala University, 2012, ISBN
978-91-554-8539-9.
[49] E. Möllerström, F. Ottermo, J. Hylander, and H. Bernhoff, “Noise emission of
a 200 kW vertical axis wind turbine,” Energies, vol. 9, no. 1, 2016.
[50] J. Kjellin, “Vertical axis wind turbines: Electrical system and experimental
results,” PhD Thesis, Uppsala University, 2012, ISBN 978-91-554-8496-5.
[51] S. Eriksson, J. Kjellin, and H. Bernhoff, “Tip speed ratio control of a 200 kW
VAWT with synchronous generator and variable DC voltage,” Energy Science
& Engineering, vol. 1, no. 3, 2013.
[52] J. Kjellin, F. Bülow, S. Eriksson, P. Deglaire, M. Leijon, and H. Bernhoff,
“Power coefficient measurement on a 12 kW straight bladed vertical axis wind
turbine,” Renewable Energy, vol. 36, no. 11, pp. 3050–3053, 2011.
[53] S. Eriksson, H. Bernhoff, and M. Bergkvist, “Design of a unique direct driven
PM generator adapted for a telecom tower wind turbine,” Renewable Energy,
vol. 44, pp. 453–456, 2012.
[54] J. Arnqvist, “Mean wind and turbulence conditions in the boundary layer above
forests,” PhD Thesis, Uppsala University, 2015, ISBN 978-91-554-9123-9.
[55] P. Thorsson, S. Söderberg, and H. Bergström, “Modelling atmospheric icing:
A comparison between icing calculated with measured meteorological data
and NWP data,” Cold Regions Science and Technology, vol. 119, pp. 124–131,
Nov. 2015.
[56] K. Nilsson, S. Ivanell, K. S. Hansen, R. Mikkelsen, J. N. Sørensen, S.-P.
Breton, and D. Henningson, “Large-eddy simulations of the Lillgrund wind
farm,” Wind Energy, vol. 18, no. 3, pp. 449–467, Mar. 2015.
[57] C. B. Field, V. R. Barros, D. J. Dokken, K. J. Mach, M. D. Mastrandrea, T. E.
Bilir, M. Chatterjee, K. L. Ebi, Y. O. Estrada, R. C. Genova, B. Girma, E. S.
Kissel, A. N. Levy, S. MacCracken, P. R. Mastrandrea, and L. L. White,
“Summary for policymakers,” in Climate Change 2014: Impacts, Adaptation,
and Vulnerability. Part A: Global and Sectoral Aspects. Contribution of
Working Group II to the Fifth Assessment Report of the Intergovernmental
Panel on Climate Change. Cambridge, United Kingdom and New York, NY,

109
 USA: Cambridge University Press, 2014, pp. 1–32.
[58] “Svenska kraftnät, Statistik,” Available online:
http://www.svk.se/aktorsportalen/elmarknad/statistik/ (Accessed: 2016-04-13).
[59] “Svensk Vindenergi, Statistik om vindkraft,” Available online:
http://www.vindkraftsbranschen.se/statistik/ (Accessed: 2016-03-16).
[60] “Energimyndigheten, Marknadsstatistik,” Available online: http:
//www.energimyndigheten.se/fornybart/elcertifikatsystemet/marknadsstatistik/
(Accessed: 2016-03-16).
[61] A. Molod, L. Takacs, M. Suarez, and J. Bacmeister, “Development of the
GEOS-5 atmospheric general circulation model: evolution from merra to
merra2,” Geoscientific Model Development, vol. 8, no. 5, pp. 1339–1356, 2015.
[62] S. Kobayashi, Y. Ota, Y. Harada, A. Ebita, M. Moriya, H. Onoda, K. Onogi,
H. Kamahori, C. Kobayashi, H. Endo, K. Miyaoka, and K. Takahashi, “The
JRA-55 Reanalysis: General Specifications and Basic Characteristics,” Journal
of the Meteorological Society of Japan. Ser. II, vol. 93, no. 1, pp. 5–48, 2015.
[63] S. Liléo, E. Berge, O. Undheim, R. Klinkert, and R. E. Bredesen, “Long-term
correction of wind measurements - state-of-the-art, guidelines and future
work,” Elforsk report 13:18, Tech. Rep., Jan. 2013.
[64] E. Sharp, P. Dodds, M. Barrett, and C. Spataru, “Evaluating the accuracy of
CFSR reanalysis hourly wind speed forecasts for the UK, using in situ
measurements and geographical information,” Renewable Energy, vol. 77, pp.
527–538, May 2015.
[65] D. Carvalho, A. Rocha, M. Gómez-Gesteira, and C. Silva Santos,
“Comparison of reanalyzed, analyzed, satellite-retrieved and NWP modelled
winds with buoy data along the Iberian Peninsula coast,” Remote Sensing of
Environment, vol. 152, pp. 480–492, Sep. 2014.
[66] S. Saha, S. Moorthi, H.-L. Pan, X. Wu, J. Wang, S. Nadiga, P. Tripp,
R. Kistler, J. Woollen, D. Behringer, H. Liu, D. Stokes, R. Grumbine,
G. Gayno, J. Wang, Y.-T. Hou, H.-Y. Chuang, H.-M. H. Juang, J. Sela,
M. Iredell, R. Treadon, D. Kleist, P. Van Delst, D. Keyser, J. Derber, M. Ek,
J. Meng, H. Wei, R. Yang, S. Lord, H. Van Den Dool, A. Kumar, W. Wang,
C. Long, M. Chelliah, Y. Xue, B. Huang, J.-K. Schemm, W. Ebisuzaki, R. Lin,
P. Xie, M. Chen, S. Zhou, W. Higgins, C.-Z. Zou, Q. Liu, Y. Chen, Y. Han,
L. Cucurull, R. W. Reynolds, G. Rutledge, and M. Goldberg, “The NCEP
climate forecast system reanalysis,” Bulletin of the American Meteorological
Society, vol. 91, no. 8, pp. 1015–1057, Apr. 2010.
[67] L. Reichenberg, F. Johnsson, and M. Odenberger, “Dampening variations in
wind power generation - the effect of optimizing geographic location of
generating sites,” Wind Energy, vol. 17, no. 11, pp. 1631–1643, Nov. 2014.
[68] D. J. Cannon, D. J. Brayshaw, J. Methven, P. J. Coker, and D. Lenaghan,
“Using reanalysis data to quantify extreme wind power generation statistics: A
33 year case study in Great Britain,” Renewable Energy, vol. 75, pp. 767–778,
Mar. 2015.
[69] J. Carta, S. Velázquez, and P. Cabrera, “A review of measure-correlate-predict
(MCP) methods used to estimate long-term wind characteristics at a target
site,” Renewable and Sustainable Energy Reviews, vol. 27, pp. 362–400, Nov.
2013.

110
[70] H. Bergström, “Boundary-layer modelling for wind climate estimates,” Wind
engineering, vol. 25, no. 5, pp. 289–299, 2001.
[71] L. Enger, “Simulation of dispersion in a moderately complex terrain. Part A.
The fluid dynamic model,” Atmospheric Environment, vol. 24A, pp.
2431–2446, 1990.
[72] A. Andrén, “Evaluation of a turbulence closure scheme suitable for air
pollution applications,” Journal of Applied Meteorology, vol. 29, pp. 224–239,
1990.
[73] W. Skamarock, J. Klemp, J. Dudhia, D. Gill, M. Barker, K. Duda, Y. Huang,
W. Wang, and J. Powers, “A description of the advanced research WRF version
3,” National Center for Atmospheric Research, Tech. Rep., 2008.
[74] R. Swinbank, M. Kyouda, P. Buchanan, L. Froude, T. M. Hamill, T. D.
Hewson, J. H. Keller, M. Matsueda, J. Methven, F. Pappenberger,
M. Scheuerer, H. A. Titley, L. Wilson, and M. Yamaguchi, “The TIGGE
project and its achievements,” Bulletin of the American Meteorological
Society, vol. 97, pp. 49–67, 2016.
[75] T. M. Hamill, G. T. Bates, J. S. Whitaker, D. R. Murray, M. Fiorino, T. J.
Galarneau, Y. Zhu, and W. Lapenta, “NOAA’s second-generation global
medium-range ensemble reforecast dataset,” Bulletin of the American
Meteorological Society, vol. 94, no. 10, pp. 1553–1565, Feb. 2013.
[76] S. Emeis, Wind Energy Meteorology. Berlin, Germany: Springer, 2013.
[77] A. Betz, Windenergie und ihre ausnutzung durch windmüllen. Göttingen,
Germany: Vandenhoeck and Ruprecht, 1926.
[78] “Power performance measurements of electricity producing wind turbines,”
International Electrotechnical Commission, Geneva, Switzerland, Tech. Rep.
IEC 61400-12-1:2005(E), 2005.
[79] M. L. Thøgersen, M. Motta, T. Sørensen, and P. Nielsen,
“Measure-correlate-predict methods: Case studies and software
implementation,” in EWEA Conference Proceedings. EMD International
A/S, 2007.
[80] Y.-H. Wan, “Analysis of wind power ramping behavior in ERCOT,” National
Renewable Energy Laboratory, Tech. Rep. NREL/TP-5500-49218, 2011.
[81] E. D. Stoutenburg, N. Jenkins, and M. Z. Jacobson, “Power output variations
of co-located offshore wind turbines and wave energy converters in
California,” Renewable Energy, vol. 35, no. 12, pp. 2781–2791, Dec. 2010.
[82] P. Sorensen, N. Cutululis, A. Vigueras-Rodriguez, L. Jensen, J. Hjerrild,
M. Donovan, and H. Madsen, “Power fluctuations from large wind farms,”
IEEE Transactions on Power Systems, vol. 22, no. 3, pp. 958–965, 2007.
[83] P. Nørgaard and H. Holttinen, “A multi-turbine power curve approach,” in
Proceedings of Nordic Wind Power Conference, Gothenburg, Sweden, Mar.
2004.
[84] R. Scharff, “Design of electricity markets for efficient balancing of wind
power generation,” PhD Thesis, KTH, 2015, ISBN 978-91-7595-652-7.
[85] U. Focken, M. Lange, K. Mönnich, H.-P. Waldl, H. G. Beyer, and A. Luig,
“Short-term prediction of the aggregated power output of wind farms - a
statistical analysis of the reduction of the prediction error by spatial smoothing
effects,” Journal of Wind Engineering and Industrial Aerodynamics, vol. 90,

111
 no. 3, pp. 231–246, Mar. 2002.
[86] C. Wan, Z. Xu, P. Pinson, Z. Y. Dong, and K. P. Wong, “Probabilistic
forecasting of wind power generation using extreme learning machine,” IEEE
Transactions on Power Systems, vol. 29, no. 3, pp. 1033–1044, May 2014.
[87] A. M. Foley, P. G. Leahy, A. Marvuglia, and E. J. McKeogh, “Current methods
and advances in forecasting of wind power generation,” Renewable Energy,
vol. 37, no. 1, pp. 1–8, 2012.
[88] G. Kariniotakis, I. Marti, D. Casas, P. Pinson, T. S. Nielsen, H. Madsen,
G. Giebel, J. Usaola, I. Sanchez, A. M. Palomares, R. Brownsword, J. Tambke,
U. Focken, M. Lange, P. Pouka, G. Kallos, C. Lac, G. Sideratos, and
G. Descombes, “What performance can be expected by short-term wind power
prediction models depending on site characteristics?” in In CD-Rom
Proceedings, European Wind Energy Conference EWEC. European Wind
Energy Association (EWEA), 2005.
[89] A. Boone, “Simulation of short-term wind speed forecast errors using a
multi-variate ARMA(1,1) time-series model,” Master thesis, KTH, Stockholm,
Sweden, 2005.
[90] B.-M. Hodge, H. Holttinen, S. Sillanpää, E. Gómez-Lázaro, R. Scharff,
L. Söder, X. Larsén, G. Giebel, D. Flynn, J. Dobschinski, D. Lew, and
M. Milligan, “Wind power forecasting error distributions: An international
comparison,” in 11th Annual International Workshop on Large-Scale
Integration of Wind Power into Power Systems as well as on Transmission
Networks for Offshore Wind Power Plants Conference, Lisbon, Portugal, Nov.
2012.
[91] K. J. Beven, Environmental modelling: an uncertain future? Abingdon,
United Kingdom: Routledge, 2009.
[92] J. Nelder and R. Mead, “A simplex method for function minimization,”
Computer Journal, vol. 7, pp. 308–313, 1965.
[93] J. Lagarias, J. Reeds, M. Wright, and P. Wright, “Convergence properties of
the Nelder–Mead simplex method in low dimensions,” SIAM Journal on
Optimization, vol. 9, no. 1, pp. 112–147, Jan. 1998.
[94] K. McKinnon, “Convergence of the Nelder–Mead simplex method to a
nonstationary point,” SIAM Journal on Optimization, vol. 9, no. 1, pp.
148–158, Jan. 1998.
[95] L. Söder, “Simulation of wind speed forecast errors for operation planning of
multiarea power systems,” in 2004 International Conference on Probabilistic
Methods Applied to Power Systems, Sep. 2004, pp. 723–728.
[96] R. W. Preisendorfer and C. D. Mobley, Principal component analysis in
meteorology and oceanography. Amsterdam, the Netherlands: Elsevier,
1988.
[97] A. Navarra and V. Simoncini, A Guide to Empirical Orthogonal Functions for
Climate Data Analysis. Dordrecht, the Netherlands: Springer, 2010.
[98] P. D. Welch, “The use of fast fourier transform for the estimation of power
spectra: a method based on time averaging over short, modified periodograms,”
IEEE Trans. Audio and Electroacoust., vol. AU-15, pp. 70–73, June 1967.
[99] S. W. Smith, The scientist and engineer’s guide to digital signal processing,
2nd ed. San Diego, CA: California Technical Publishing, 1999, Available

112
 online: http://www.DSPguide.com.
[100] W. Katzenstein, E. Fertig, and J. Apt, “The variability of interconnected wind
plants,” Energy Policy, vol. 38, pp. 4400–4410, Aug. 2010.
[101] D. Lee and R. Baldick, “Future wind power scenario synthesis through power
spectral density analysis,” IEEE Transactions on Smart Grid, vol. 5, pp.
490–500, Jan. 2014.
[102] Y. Li, K. Xie, and B. Hu, “Copula-ARMA model for multivariate wind speed
and its applications in reliability assessment of generating systems,” Journal of
Electrical Engineering and Technology, vol. 8, no. 3, pp. 421–427, May 2013.
[103] A. Sturt and G. Strbac, “Time series modelling of power output for large-scale
wind fleets,” Wind Energy, vol. 14, no. 8, pp. 953–966, 2011.
[104] P. E. de Mello, N. Lu, and Y. Makarov, “An optimized autoregressive forecast
error generator for wind and load uncertainty study,” Wind Energy, vol. 14,
no. 8, pp. 967–976, Nov. 2011.
[105] R. A. Berk, Statistical learning from a regression perspective, 1st ed., ser.
Springer Series in Statistics. New York, N.Y: Springer, 2008.
[106] L. Breiman, J. Friedman, C. J. Stone, and R. A. Olshen, Classification and
regression trees. Monterey, CA: Wadsworth press, Jan. 1984.
[107] L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32,
Oct. 2001.
[108] V. Bulaevskaya, S. Wharton, A. Clifton, G. Qualley, and W. O. Miller, “Wind
power curve modeling in complex terrain using statistical models,” Journal of
Renewable and Sustainable Energy, vol. 7, no. 1, p. 013103, Jan. 2015.
[109] J. H. Friedman, “Greedy function approximation: A gradient boosting
machine,” The Annals of Statistics, vol. 29, no. 5, pp. 1189–1232, 2001.
[110] A. Kusiak and Z. Zhang, “Short-horizon prediction of windpower: A
data-driven approach,” IEEE Transactions on Energy Conversion, vol. 25,
no. 4, pp. 1112–1122, Dec. 2010.
[111] M. Olsson, M. Perninge, and L. Söder, “Modeling real-time balancing power
demands in wind power systems using stochastic differential equations,”
Electric Power Systems Research, vol. 80, pp. 966–974, Aug. 2010.
[112] J. P. Deane, G. Drayton, and B. P. Ó Gallachóir, “The impact of sub-hourly
modelling in power systems with significant levels of renewable generation,”
Applied Energy, vol. 113, pp. 152–158, Jan. 2014.
[113] N. Troy, D. Flynn, and M. O’Malley, “The importance of sub-hourly modeling
with a high penetration of wind generation,” in 2012 IEEE Power and Energy
Society General Meeting, Jul. 2012, pp. 1–6.
[114] A. Keane, M. Milligan, C. J. Dent, B. Hasche, C. D’Annunzio, K. Dragoon,
H. Holttinen, N. Samaan, L. Soder, and M. O’Malley, “Capacity value of wind
power,” IEEE Transactions on Power Systems, vol. 26, no. 2, pp. 564–572,
May 2011.
[115] L. Hirth, “The market value of wind energy - thermal versus hydro power
systems,” Energiforsk, Tech. Rep. 2016:276, 2016.
[116] “Nordpool, Historical market data,” Available online:
http://www.nordpoolspot.com/historical-market-data/ (Accessed:
2016-07-16).
[117] E. V. Mc Garrigle, J. P. Deane, and P. G. Leahy, “How much wind energy will

113
 be curtailed on the 2020 Irish power system?” Renewable Energy, vol. 55, pp.
544–553, Jul. 2013.
[118] D. Heide, M. Greiner, L. von Bremen, and C. Hoffmann, “Reduced storage
and balancing needs in a fully renewable European power system with excess
wind and solar power generation,” Renewable Energy, vol. 36, no. 9, pp.
2515–2523, Sep. 2011.
[119] C. Budischak, D. Sewell, H. Thomson, L. Mach, D. E. Veron, and
W. Kempton, “Cost-minimized combinations of wind power, solar power and
electrochemical storage, powering the grid up to 99.9% of the time,” Journal
of Power Sources, vol. 225, pp. 60–74, Mar. 2013.
[120] M. G. Rasmussen, G. B. Andresen, and M. Greiner, “Storage and balancing
synergies in a fully or highly renewable pan-European power system,” Energy
Policy, vol. 51, pp. 642–651, Dec. 2012.

114
Acta Universitatis Upsaliensis
Digital Comprehensive Summaries of Uppsala Dissertations
from the Faculty of Science and Technology 1428
Editor: The Dean of the Faculty of Science and Technology

A doctoral dissertation from the Faculty of Science and

Technology, Uppsala University, is usually a summary of a
number of papers. A few copies of the complete dissertation
are kept at major Swedish research libraries, while the
summary alone is distributed internationally through
the series Digital Comprehensive Summaries of Uppsala
Dissertations from the Faculty of Science and Technology.
(Prior to January, 2005, the series was published under the
title “Comprehensive Summaries of Uppsala Dissertations
from the Faculty of Science and Technology”.)

ACTA
UNIVERSITATIS
UPSALIENSIS
Distribution: publications.uu.se UPPSALA
urn:nbn:se:uu:diva-302837 2016

View publication stats