Sei sulla pagina 1di 223

Title

Modeling electricity prices for generation investment and


scheduling analysis
Author(s) He, Yang; OU3
Citation
Issue Date 2010
URL http://hdl.handle.net/10722/130905
Rights
The author retains all proprietary rights, (such as patent
rights) and the right to use in future works.
Modeling Electricity Prices for
Generation Investment and
Scheduling Analysis
by
Yang HE
A thesis submitted in partial fulllment of the requirements for
the Degree of Doctor of Philosophy
at The University of Hong Kong.
February 2010
Abstract of thesis entitled
Modeling Electricity Prices for Generation
Investment and Scheduling Analysis
Submitted by Yang HE
for the degree of Doctor of Philosophy
at The University of Hong Kong
in February 2010
In the deregulated electric power industry, under the market environment,
electricity spot prices are highly volatile and uncertain. For private generation
companies, their prots are directly tied with and signicantly aected by
these uctuations of electricity spot prices. To make decisions on building new
power plants and scheduling the production thereafter, generation companies
desire an electricity spot price model to assist their making these decisions.
The system of electricity spot prices is of high complexity; it is driven by
various physical underlying forces that play in dierent timescales. In the
short time horizon of one week, the physical driving forces are intra-day and
intra-week variations of electricity load, generation forced outages, etc.; in
the mid-term of one year, it is the seasonal forces that are manifest, such
as seasonal weather and temperature, annual generation maintenance, etc.;
and in the time horizon of years and decades, the eecting physical forces
are economic development and economic cycles, generation investment and
retirement, uctuations of fuel prices, etc.
This work develops a Multi-granularity Framework to facilitate analyz-
ing electricity spot prices, which views electricity spot prices in three time-
perspectives, that is, multi-year yearly, intra-year weekly, and intra-week hourly.
In each time-perspective, how the various underlying physical forces give rise
to the very peculiar behaviors of electricity spot prices is carefully discussed.
Because the physical forces that underlie electricity spot prices are indepen-
dent to each other, play in dierent timescales, and aect electricity spot prices
in dierent time horizons, this work adopts the methodology Divide and Con-
quer to build the price model: it decomposes the historical electricity spot
price data into components that are driven by dierent and independent phys-
ical underlying forces, then models each price component respectively, and
nally constructs a complete electricity spot price model out of the resulting
sub-models.
The overall price model explicitly considers the various physical forces that
drive electricity spot prices, the model extends on a time horizon of multiple
years and has a time unit of one hour, and its nal result represents prices at
each hour by a probability density function of Lognormal distribution. The
model has been evaluated in the New-England and PJM electricity markets.
Upon proper revisions, the same analysis framework and modeling methodol-
ogy probably could be applied to many other electricity markets in the world.
The proposed price model is physically grounded, mathematically simple,
and computationally fast. It provides an analytical tool to generation compa-
nies for their making informed decisions in generation investment and schedul-
ing analysis. Besides generation companies, the model could also be widely
used by other players in electricity markets, like by power traders for pric-
ing and trading electricity contracts, futures, options, and other electricity
derivatives, and by power retailers and large power consumers for their power
purchase and risk management.
Declaration
I declare that the thesis and the research work thereof represents
my own work, except where due acknowledgment is made, and
that it has not been previously included in a thesis, dissertation or
report submitted to this University or to any other institution for
a degree, diploma or other qualications.
Signed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Yang HE
Acknowledgment
The saying Read Masters, Not Pupils, weighs heavily. I am fortunate in
the past four years having learned under a master. Prof. Wu is a master who
knows when to let the student loose, by allowing him to explore under the
guidance of his own interest; and when to tighten him back, by pointing to
him the direction that is the most worth pursuing. He knows the importance
of good literature, and he recommends to his student the best works he ever
knows. He knows with exact in which stage the student is: he is the strictest
critics of his students work, and he shows to him, by correcting the students
work, how mathematically rigid and practically sound the nal result could
be. He helps the student vision the ideal result of the work and helps him
setup the analysis framework, and asks him to work out a solution that is the
most mathematically concise and physically grounded. Prof. Wu has a good
habit in thinking, he starts with a problem, denes it clearly, and analyzes it
to a point by which he has got a thorough understanding. Prof. Wu is one
of the best lecturers I have ever seen. All these good examples Prof. Wu has
set, the pupil during the past four years has observed and mimicked, and the
student is determined to practice, in the rest of his life, these good fortunes he
has learned from his teacher.
Another person who makes this thesis work possible is the philosopher and
novelist Any Rand. At the end of the rst year of this Ph.D. study, the author
was in great diculty in seeking for the meaning of life. It was the novel The
Fountainhead that taught me that man is a heroic rational being who pursues
his own happiness, that one achieves his happiness through productive work,
by productive work it is meant that one works by using his own mind to think,
and that reason is the absolute tool for the guidance of thought. In her another
novel Atlas Shrugged I found the seven virtues she derived truly attractive, that
is, Rationality, Independence, Honesty, Integrity, Justice, Productiveness, and
Pride. These virtues have since then become the guidance of my action, every
day.
I grew up in a small town on the China eastern coast. My parents own
a small family business which buy, salt, dry, and sell sh. Their working
days start from 5 oclock in the morning till 6 oclock in the evening. They
work without holiday till the years ends. Since very early age I often went to
the factory watching them working. By observing my parents, I learned the
value of hard work and the responsibility of life: I learned that one has to be
responsible for his own work, and that one has to be diligent and persistent
in pursuing his own convinced goals. I developed my working morals learning
from my parents, and these principles have guided me and proved valuable
throughout my time as a student.
I am grateful to my former supervisor in Zhejiang University, Prof. De-
qiang Gan, it was his reference that helped me be admitted to this Ph.D.
program; I am in debt to Dr. Yunhe Hou, during the early and late stages of
this research project, it was his discussions and advices that help me keep the
work progressing; I am grateful to Dr. Jin Zhong, who generously gave me two
chances to go to conferences; I am thankful to our laboratory technician Mr.
Peter Tam, who kept my computer out of trouble during this four years; I am
grateful to Prof. Wus secretary Ms. Clara Chung, who expertly steered me
through all the administrative requirements during the study; I am thankful
to my colleagues in the Center for Electrical Energy Systems, whose atten-
dance to my presentations and their following-up discussions gave me valuable
perspectives to criticize my work; and I am fortunate during this four years
having a few very intelligent and considerate friends, who made this journey
much more pleasant than it otherwise could be, they are Dr. Yanhui Geng
and his wife Ms. Qiong Sun.
Contents
List of Figures xv
List of Tables xix
1 Introduction 1
1.1 Background and Motivations . . . . . . . . . . . . . . . . . . . . 1
1.2 Literature Review and the Gap to Fill . . . . . . . . . . . . . . 4
1.3 Elaboration of the Problem . . . . . . . . . . . . . . . . . . . . 7
1.3.1 The Requirements on the Price Model . . . . . . . . . . 7
1.3.2 The Analytical and Mathematical Tools . . . . . . . . . 8
1.4 The Analysis Framework . . . . . . . . . . . . . . . . . . . . . . 9
1.5 The Modeling Methodology . . . . . . . . . . . . . . . . . . . . 10
1.6 Organization of the Thesis . . . . . . . . . . . . . . . . . . . . . 11
2 A Survey of Electricity Price Models 13
2.1 Electricity Price Modeling Approaches . . . . . . . . . . . . . . 13
2.2 Time Series Models . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.1 The Autoregressive Model . . . . . . . . . . . . . . . . . 15
2.2.2 AR with Time-varying Mean . . . . . . . . . . . . . . . . 17
2.2.3 AR with Exogenous Variables . . . . . . . . . . . . . . . 18
2.2.4 The Autoregressive Moving Average Model . . . . . . . . 19
2.2.5 ARMA with Time-varying Mean . . . . . . . . . . . . . 20
2.2.6 ARMAX and Transfer Function . . . . . . . . . . . . . . 21
2.2.7 Periodic Autoregressive Models . . . . . . . . . . . . . . 21
2.2.8 ARIMA and its Extensions . . . . . . . . . . . . . . . . . 23
ix
CONTENTS
2.2.9 Regime-switching Models . . . . . . . . . . . . . . . . . . 25
2.2.10 Time-varying Volatility . . . . . . . . . . . . . . . . . . . 26
2.2.11 The GARCH Model . . . . . . . . . . . . . . . . . . . . 26
2.2.12 GARCH with Asymmetric Eect . . . . . . . . . . . . . 28
2.2.13 Time-varying Volatility and Model Parameter Calibration 29
2.3 Financial/Stochastic Process Models . . . . . . . . . . . . . . . 30
2.3.1 Modeling the Multi-seasonality . . . . . . . . . . . . . . 30
2.3.2 Modeling the Mean-reverting Price Nature . . . . . . . . 33
2.3.3 A Basic Electricity Price Model . . . . . . . . . . . . . . 35
2.3.4 Modeling the Time-varying Volatility . . . . . . . . . . . 36
2.3.5 Modeling the Multi Risk-factors . . . . . . . . . . . . . . 38
2.3.6 Modeling the Price Spikes . . . . . . . . . . . . . . . . . 39
2.4 Structural Models . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.5 The Eect of Fuel Prices . . . . . . . . . . . . . . . . . . . . . . 43
2.6 Decomposition Techniques . . . . . . . . . . . . . . . . . . . . . 44
2.6.1 Fourier Analysis . . . . . . . . . . . . . . . . . . . . . . . 44
2.6.2 Wavelet Analysis . . . . . . . . . . . . . . . . . . . . . . 45
2.6.3 Principal Component Analysis . . . . . . . . . . . . . . . 48
2.7 Concluding Discussions . . . . . . . . . . . . . . . . . . . . . . . 51
3 Microeconomics applied to the Electricity Markets 53
3.1 The Supply Curve of a Generation Company . . . . . . . . . . . 54
3.1.1 Heat Rate Curve of the Generation Company . . . . . . 54
3.1.2 Fuel Price Curve of the Generation Company . . . . . . 54
3.1.3 Marginal Cost Curve of the Generation Company . . . . 55
3.1.4 Supply Curve of the Generation Company . . . . . . . . 56
3.2 The Supply Curve of an Electricity Market . . . . . . . . . . . . 58
3.3 The Demand Curve of an Electricity Market . . . . . . . . . . . 58
3.4 The Spot Price of an Electricity Market . . . . . . . . . . . . . 59
3.5 The Eect of Fuel Prices on Spot Prices . . . . . . . . . . . . . 60
3.6 The Eect of Generator Outage on Spot Prices . . . . . . . . . 60
3.7 Intra-day Electricity Spot Prices . . . . . . . . . . . . . . . . . . 62
3.8 Time-varying Volatility of Intra-day Spot Prices . . . . . . . . . 63
x
CONTENTS
3.9 Electricity Spot Price Spikes . . . . . . . . . . . . . . . . . . . . 64
4 A Multi-granularity View of Electricity Prices 67
4.1 Intra-week Hourly Electricity Spot Prices . . . . . . . . . . . . . 67
4.2 Intra-year Weekly Prices . . . . . . . . . . . . . . . . . . . . . . 69
4.3 Multi-year Yearly Prices . . . . . . . . . . . . . . . . . . . . . . 71
4.4 The Eect of Fuel Prices on Yearly Prices . . . . . . . . . . . . 73
4.5 A Multi-granularity View of Electricity Prices . . . . . . . . . . 75
5 A Multi-granularity Electricity Spot Price Model 77
5.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.3 The New-England Market . . . . . . . . . . . . . . . . . . . . . 81
5.4 Decompose the Data to Find Orders . . . . . . . . . . . . . . . 82
5.4.1 Fuel-Price Eect on Electricity Spot Prices . . . . . . . . 82
5.4.2 Fuel-price Eect Adjustment . . . . . . . . . . . . . . . . 84
5.4.3 Frequency Spectrum of the Data . . . . . . . . . . . . . 85
5.4.3.1 The High-frequency Component . . . . . . . . . 85
5.4.3.2 The Mid-frequency Component . . . . . . . . . 87
5.4.3.3 The Low-frequency Component . . . . . . . . . 88
5.5 Dene the Orders by Models . . . . . . . . . . . . . . . . . . . . 89
5.5.1 The Intra-week Model . . . . . . . . . . . . . . . . . . . 89
5.5.2 The Intra-year Model . . . . . . . . . . . . . . . . . . . . 90
5.5.3 The Multi-year Model . . . . . . . . . . . . . . . . . . . 91
5.5.4 The Fuel Price Model . . . . . . . . . . . . . . . . . . . . 92
5.5.5 The Overall Model . . . . . . . . . . . . . . . . . . . . . 93
5.6 Calibrate the Models . . . . . . . . . . . . . . . . . . . . . . . . 94
5.6.1 The Intra-week Model Calibration . . . . . . . . . . . . . 94
5.6.2 The Intra-year Model Calibration . . . . . . . . . . . . . 97
5.6.3 The Multi-year Model Calibration . . . . . . . . . . . . . 100
5.6.4 The Sampling of the Data and the Time-Unit of Models 101
5.7 Simulate Electricity Spot Prices with the Models . . . . . . . . . 102
5.8 Use Price Simulations to Valuate Generator Prot . . . . . . . . 108
5.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
xi
CONTENTS
5.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6 The Multi-granularity Model applied to the PJM Market 113
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.2 The Fundamentals of the PJM Market . . . . . . . . . . . . . . 114
6.3 The Electricity Prices and Fuel Price Index . . . . . . . . . . . . 118
6.4 Data Decomposition . . . . . . . . . . . . . . . . . . . . . . . . 120
6.5 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.6 Model Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.6.1 The Intra-week Model Calibration . . . . . . . . . . . . . 124
6.6.2 The Intra-year Model Calibration . . . . . . . . . . . . . 124
6.6.3 The Multi-year Model Calibration . . . . . . . . . . . . . 128
6.7 Price Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.8 Generator Prot Valuation . . . . . . . . . . . . . . . . . . . . . 133
6.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
7 A Structural Electricity Spot Price Model 135
7.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
7.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
7.3 The Eect of Generator Forced Outage on Spot Prices . . . . . 138
7.4 The Structural Model . . . . . . . . . . . . . . . . . . . . . . . . 140
7.4.1 The Structural Model . . . . . . . . . . . . . . . . . . . . 140
7.4.2 The Available Generation Capacity . . . . . . . . . . . . 142
7.4.3 The Segment Slopes . . . . . . . . . . . . . . . . . . . . 142
7.4.4 The Expected Value of the Spot Prices . . . . . . . . . . 144
7.5 The Simplied Model . . . . . . . . . . . . . . . . . . . . . . . . 146
7.5.1 The Simplied Model . . . . . . . . . . . . . . . . . . . . 146
7.5.2 The Probability Density Function of the Spot Prices . . 147
7.6 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . 148
7.6.1 The Test Generation System . . . . . . . . . . . . . . . . 148
7.6.2 The Structural Model . . . . . . . . . . . . . . . . . . . . 148
7.6.3 The Simplied Model . . . . . . . . . . . . . . . . . . . . 151
7.6.4 The Computation Burden of the Structural Models . . . 151
7.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
xii
CONTENTS
7.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
8 Conclusion 157
8.1 Signicance of the Work . . . . . . . . . . . . . . . . . . . . . . 157
8.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
8.3 Future Work on the Price Model . . . . . . . . . . . . . . . . . . 158
8.4 Applications to Power System Operation and Planning . . . . . 160
Appendix 162
A A Simple Structural Electricity Spot Price Model 163
B The Test Generation System 167
C Basic Stochastic Processes 169
C.1 Wiener Process . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
C.2 Wiener Process with Drift . . . . . . . . . . . . . . . . . . . . . 170
C.3 Geometric Brownian Motion . . . . . . . . . . . . . . . . . . . . 170
C.4 Mean-reversion Process . . . . . . . . . . . . . . . . . . . . . . . 171
C.5 Geometric Mean-reversion with Constant Mean . . . . . . . . . 173
C.6 Geometric Mean-reversion with Time-varying Mean . . . . . . . 174
D The Price Filters and the Filtering 177
D.1 Design of the Price Filters . . . . . . . . . . . . . . . . . . . . . 177
D.2 Price Decomposition by Filtering . . . . . . . . . . . . . . . . . 179
E The Mean-reversion Process in Discrete Time 181
F Build the Fuel Price Index 183
F.1 The Fuel Price Index Model . . . . . . . . . . . . . . . . . . . . 183
F.2 Derivation of the Model for Parameter Estimation . . . . . . . . 184
F.3 Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . 186
References 189
xiii
CONTENTS
xiv
List of Figures
1.1 The Flow-chart of Generator Prot Valuation . . . . . . . . . . 2
1.2 Distribution of Literatures . . . . . . . . . . . . . . . . . . . . . 4
2.1 Index the Intra-day Hourly Spot Prices . . . . . . . . . . . . . . 22
2.2 A Two-regime Markov Chain . . . . . . . . . . . . . . . . . . . 25
2.3 Intra-day Electricity Price Pattern . . . . . . . . . . . . . . . . 31
2.4 Intra-week Electricity Price Pattern . . . . . . . . . . . . . . . . 32
2.5 Intra-year Seasonality of Electricity Prices . . . . . . . . . . . . 32
2.6 A Haar Wavelet . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.1 The Heat-rate Curve of the Generation Company . . . . . . . . 54
3.2 The Fuel Price Curve of the Generation Company . . . . . . . . 55
3.3 The Marginal Cost Curve of the Generation Company . . . . . . 56
3.4 The Supply Curve of the Generation Company . . . . . . . . . . 57
3.5 The Supply Curve of an Electricity Market . . . . . . . . . . . . 58
3.6 The Demand Curve of an Electricity Market . . . . . . . . . . . 59
3.7 The Spot Price of an Electricity Market . . . . . . . . . . . . . 60
3.8 The Eect of Fuel Prices on Spot Prices . . . . . . . . . . . . . 61
3.9 The Eect of Generator Outage on Spot Prices . . . . . . . . . 62
3.10 Intraday Electricity Load Demand . . . . . . . . . . . . . . . . . 62
3.11 Intraday Electricity Spot Prices . . . . . . . . . . . . . . . . . . 63
3.12 The Time-varying Volatility of Intra-day Electricity Spot Prices 64
3.13 The Marginal Cost Curve with an additional Third Segment . . 65
3.14 Intra-day Electricity Spot Price Spikes . . . . . . . . . . . . . . 66
4.1 Intra-week Hourly Electricity Spot Prices . . . . . . . . . . . . . 68
xv
LIST OF FIGURES
4.2 Intra-year Weekly Prices . . . . . . . . . . . . . . . . . . . . . . 69
4.3 Multi-year Yearly Prices . . . . . . . . . . . . . . . . . . . . . . 72
4.4 The Prices of Fossil Fuels across Many Years . . . . . . . . . . . 74
4.5 The Yearly Prices Considering Fuel Price Fluctuations . . . . . 74
4.6 A Multi-granularity View of Electricity Prices . . . . . . . . . . 75
5.1 New-England Market Fundamentals . . . . . . . . . . . . . . . . 81
5.2 Electricity Prices and Natural Gas Prices . . . . . . . . . . . . . 82
5.3 Fuel-price Eect on Electricity Spot Prices . . . . . . . . . . . . 83
5.4 Implied Marginal Generator Heat Rate Data . . . . . . . . . . . 84
5.5 The High-frequency Component Data . . . . . . . . . . . . . . . 86
5.6 The Mid-frequency Component Data . . . . . . . . . . . . . . . 88
5.7 The Low-frequency Component Data . . . . . . . . . . . . . . . 89
5.8 The Overall Electricity Spot Price Model . . . . . . . . . . . . . 93
5.9 The High-frequency Component and the Intra-week Pattern . . 96
5.10 The Mid-frequency Component and the Intra-year Pattern . . . 98
5.11 Seasonality of the New-England Market . . . . . . . . . . . . . . 99
5.12 Calibration of the Multi-year Model . . . . . . . . . . . . . . . . 101
5.13 Construct the Electricity Spot Prices in the Year 2005 . . . . . . 103
5.14 Day-ahead Electricity Spot Prices in the Year 2005 . . . . . . . 104
5.15 Empirical Probability Distributions of the Prices in Year 2005 . 105
5.16 The Mid-winter Week in 2005 . . . . . . . . . . . . . . . . . . . 106
5.17 The Mid-spring Week in 2005 . . . . . . . . . . . . . . . . . . . 107
5.18 The Mid-summer Week in 2005 . . . . . . . . . . . . . . . . . . 107
5.19 The Mid-fall Week in 2005 . . . . . . . . . . . . . . . . . . . . . 108
5.20 One-year Generator Prot of the Coal Unit and the Gas Unit . 110
6.1 PJM Market Fundamentals . . . . . . . . . . . . . . . . . . . . 115
6.2 Business-day Intra-day Pattern of Marginal Fuel Mixture . . . . 116
6.3 Intra-year Seasonal Pattern of Marginal Fuel Mixture . . . . . . 116
6.4 Multi-year Pattern of Marginal Fuel Mixture . . . . . . . . . . . 117
6.5 The Prices of the Three Marginal Fuels . . . . . . . . . . . . . . 118
6.6 PJM Market Electricity Prices and Fuel Prices . . . . . . . . . . 119
6.7 Fuel-price Eect on Electricity Spot Prices . . . . . . . . . . . . 119
xvi
LIST OF FIGURES
6.8 Implied Marginal Generator Heat Rate Data . . . . . . . . . . . 121
6.9 The High-frequency Component Data . . . . . . . . . . . . . . . 122
6.10 The Mid-frequency Component Data . . . . . . . . . . . . . . . 122
6.11 The Low-frequency Component Data . . . . . . . . . . . . . . . 123
6.12 The High-frequency Data and the Intra-week Pattern . . . . . . 125
6.13 The Mid-frequency Data and the Seasonal Pattern . . . . . . . . 126
6.14 Seasonality of the PJM Market . . . . . . . . . . . . . . . . . . 127
6.15 Calibration of the Multi-year Model . . . . . . . . . . . . . . . . 128
6.16 Day-ahead Hourly Electricity Spot Prices in Year 2008 . . . . . 130
6.17 Empirical Probability Distributions of the Prices in Year 2008 . 130
6.18 Mid-winter Week in 2008 . . . . . . . . . . . . . . . . . . . . . . 131
6.19 Mid-spring Week in 2008 . . . . . . . . . . . . . . . . . . . . . . 131
6.20 Mid-summer Week in 2008 . . . . . . . . . . . . . . . . . . . . . 132
6.21 Mid-fall Week in 2008 . . . . . . . . . . . . . . . . . . . . . . . 132
6.22 One-year Generator Prot of the Coal Unit and the Gas Unit . 134
7.1 The Generation Supply Stack of an Electricity Market . . . . . 139
7.2 The Eect of Generator Forced Outage on Spot Prices . . . . . 139
7.3 The Structural Electricity Spot Price Model . . . . . . . . . . . 140
7.4 The Integration Regions of the Three Cases . . . . . . . . . . . 145
7.5 The Generation Supply Stack of the Test Generation System . . 149
7.6 The Expected Value of the Spot Prices . . . . . . . . . . . . . . 150
7.7 The Standard Deviation and Skewness of the Spot Prices . . . . 150
7.8 The Errors of the Model . . . . . . . . . . . . . . . . . . . . . . 151
7.9 The Expected Value of the Spot Prices . . . . . . . . . . . . . . 152
7.10 The Standard Deviation and Skewness of the Spot Prices . . . . 152
7.11 The Errors of the Model . . . . . . . . . . . . . . . . . . . . . . 153
A.1 A Simple Structural Electricity Spot Price Model . . . . . . . . 165
F.1 Parameter Estimation of the Fuel Price Index Model . . . . . . 187
F.2 The Monthly Intra-year Seasonal Pattern . . . . . . . . . . . . . 187
xvii
LIST OF FIGURES
xviii
List of Tables
2.1 Separate the Intra-day Hourly Spot Prices into 24 Time Series . 22
5.1 Parameters of the Intra-week Model . . . . . . . . . . . . . . . . 95
5.2 Parameters of the Intra-year Model . . . . . . . . . . . . . . . . 97
5.3 Parameters of the Multi-year Model . . . . . . . . . . . . . . . . 100
5.4 Statistics of the Empirical Probability Distributions . . . . . . . 105
5.5 Characteristics of the Coal Unit and the Gas Unit . . . . . . . . 108
5.6 One-year Generator Prot of the Coal Unit and the Gas Unit . 109
6.1 Parameters of the Intra-week Model . . . . . . . . . . . . . . . . 124
6.2 Parameters of the Intra-year Model . . . . . . . . . . . . . . . . 127
6.3 Parameters of the Multi-year Model . . . . . . . . . . . . . . . . 129
6.4 Statistics of the Empirical Probability Distributions . . . . . . . 129
6.5 One-year Generator Prot of the Coal Unit and the Gas Unit . 133
7.1 The Computation Burden of the Structural Models . . . . . . . 153
B.1 Characteristics of the Generators in the Test System . . . . . . 168
D.1 The Price Filters . . . . . . . . . . . . . . . . . . . . . . . . . . 179
D.2 Electricity Spot Price Decomposition by Filtering . . . . . . . . 180
F.1 Parameters of the Fuel Price Index Model(2002-2009) . . . . . . 186
F.2 Parameters of the Fuel Price Index Model(2005-2009) . . . . . . 186
xix
LIST OF TABLES
xx
Chapter 1
Introduction
1.1 Background and Motivations
Large scale electric power systems have a history of less than one hundred
years. It was only in 1882 that Thomas A. Edison brought his power station
in Pearl Street on-line to supply electricity to the nancial district in New
York City, see Wasik (2006). Since early 20
th
century power systems gained
dramatic development rst in the Western world and thereafter spread to the
rest of the world. Up to early 1990s, partly due to historical reasons and partly
due to the nature of the electric power industry, most power systems around
the world operated under strict government regulation, in which government
granted the rights of building power plants to generation companies, xed
the electricity tari, and guaranteed a certain rate of return for the invested
capital.
In the regulated electric power industry, because government policies guar-
antee that generation companies are receiving a xed tari for generating a
certain amount of electricity, the concern of these generation companies thus
is how to build a particular power plant and then how to dispatch the gen-
erators in a manner so that to minimize the cost of generating that amount
of electricity. The cost of generating electricity is a function of electric load,
eciency and availability of generators, prices of fuel, and various operating
constraints of generators and transmission networks, etc. Power system re-
searchers have developed sophisticated methodologies and models to estimate
1
1. INTRODUCTION
the cost of generating electricity up to a time horizon of several decades, which
are conventionally categorized as Production Cost Models, refer to Wood &
Wollenberg (1996).
Since early 1990s, deregulating the electric power industry in the purpose of
introducing competition became a trend around the world. The deregulation
rst started on the generation side, generation assets were sold to a few private
generation companies. These generation companies then compete with each
other to supply electricity to electricity markets, and the generators with lower
bids have the priority to sell to the market. Electricity prices are no longer xed
by the government, but determined by the supply and demand of electricity
markets.
In the deregulated electric power industry, private generation companies, to
make their decisions on building power plants and scheduling their production,
concern the prot they are going to make if they build a particular power
plant and if they dispatch their generators in a certain manner. The prot
of a generator is a function of future electricity prices, the availability of the
generator, the cost of generating electricity, and the discount-rate for future
prot, refer to Hou & Wu (2008), and see Figure 1.1.
Electricity Spot Prices
0
t
1
T
2
T
t
t
1
T
2
T
Q2
Generator
Availability
( ) A t

Total
Profit
max ( ) ( ),0 S t C t

Hourly Profit
Q3
Present Value
Rt
e

Q4
Summation

Q5 Q1
( ) S t


Figure 1.1: The Flow-chart of Generator Prot Valuation
Namely, at the present time t
0
, the total prot for generating electricity
during a future time period [T
1
, T
2
] could be calculated as follows: at each
future hour t, if the generator is available and not in outage, namely A(t) =
1(otherwise A(t) = 0), the generator will watch the market and take chances
2
1.1 Background and Motivations
to generate electricity; if the electricity spot prices S(t) is greater than the cost
of generating electricity C(t), S(t) > C(t), the generate will produce and earn
a prot (t) = S(t) C(t); and if the electricity spot price is less than the
cost of generating electricity, S(t) < C(t), the generator does not operate and
earn a zero prot (t) = 0; the future prot at hour t, (t), is then discounted
to obtain its present value

(t) = e
Rt
(t); nally, the prots at each future
hour t is summed over the whole period [T
1
, T
2
] to arrive at the total prot
during the period, =

t[T
1
,T
2
]

(t).
Due to the limitations of human knowledge and the inherent uncertainties
of Nature, the forecasts on future prots are always uncertain: future electricity
prices, as they have been, will be highly volatile, generators could be forced to
outage due to equipment failure, the cost of generating electricity is subject to
uctuations of fuel prices, and the discount-rate for future prot is further a
function of interest rate and the extent of the uncertainty associated with the
prot in concern. Therefore, on the downside, generation companies very much
concern the possibility of not earning enough prot to cover their initial capital
investment; on the upside, these generation companies also care the possibility
of reaping unusual spectacular prot if the markets go in their favor.
Therefore, generation companies, in order to make informed decisions on
generation investment and scheduling, have to calibrate carefully the key fac-
tors which will signicantly aect their future earnings, that is, future electric-
ity prices, generation forced outage rate, future fuel prices, and future interest
rate. This thesis work will only focus on understanding, interpreting, and then
modeling the rst key factor, namely, electricity prices.
Even compared with the prices of other energy commodities like oil and
natural gas, which have been famously volatile, electricity, the most important
direct source of power for mankind, has prices that are notoriously volatile.
During a day, electricity spot prices are higher during the day-time and lower
in the night-time; during a week, electricity prices are higher during weekdays
and lower on weekends; during one year, electricity prices are high and volatile
in high demand seasons, and usually low and milder in low demand seasons;
across years, electricity prices are high when the economy is active and low
3
1. INTRODUCTION
when the economy falls to recessions. Meanwhile, short-life extreme-high price
spikes, which could be as high as tens of times of usual prices, are frequently
encountered in many electricity markets around the world.
Generation companies, whose prots are directly tied with and signicantly
aected by these movements of electricity prices, crave for a deeper understand-
ing on electricity prices and at best a physically well-grounded, simple, and
fast electricity price model to empower them to make informed decisions even
on a day-to-day basis.
1.2 Literature Review and the Gap to Fill
Modeling Approach
Statistical / Financial
Structural / Hybrid
Fundamental / Simulation
Time-horizon Long-term
/Multi-year
Mid-term
/Intra-year
Short-term
/Intra-week
Figure 1.2: Distribution of Literatures - according to their approaches and
the time-horizons in their concern, and the darkness of the cells represents the
number of publications
Many researchers have worked on electricity price modeling. In terms of
time-horizon, the price modeling problems can be categorized into short-term,
mid-term, and long-term, see Figure 1.2. In terms of modeling approach,
there are Statistical/Financial Approach, Structural/Hybrid Approach, and
Fundamental/Simulation Approach: many researchers have taken the statis-
tical/nancial approach to model either the short-term or mid-term prices,
using Time Series Models or Financial Models; a few authors have taken the
structural/hybrid approach to model either the short-term, mid-term, or the
long-term prices; and some researchers have taken the fundamental/simulation
4
1.2 Literature Review and the Gap to Fill
approach to study power markets. (A detailed discussion on the various mod-
eling methods could be found in Chapter 2.)
Looking at this landscape of literature, one observes that few people have
taken the statistical/nancial approach to model the long-term prices. The
reason for this observation is mostly due to the limitations of the mathemat-
ical tools that are available for the statistical/nancial approach, which are
Time-Series/Financial Models. According to the conventional wisdom, Time-
Series/Financial Models are only good for modeling short-term prices, and at
most for mid-term prices, but not for modeling long-term prices, because in the
long-run of many years, the fundamental system is undergoing so signicant a
change that previous calibrations of models from historical data are no longer
valid.
Another observation on this literature landscape is that, comparing with
the large number of publications on the statistical approach, only a few authors
have addressed the structural approach. The main reason for this observation
is due to the complexity of the structural approach. Structural approach mod-
els electricity spot prices indirectly: it rst models the constituent physical
underlying forces such as generation supply stack, electricity load, generation
forced outages, fuel prices, etc., and then constructs these constituent models
into an electricity spot price model. In order to build a structural price model
that is realistic and thus has its potential use in practice, the perquisite is that
one has to rst model the constituent physical forces satisfactorily. However,
the problems of modeling these constituent physical forces themselves are chal-
lenging and so far havent received denitive solutions. Further discussions on
the structural approach could be found in Section 2.4.
In this landscape of literature, we discuss the position of our work. Our
primary purpose is to build a long-term electricity spot price model, and this
model will be used by generation companies for generation investment and
scheduling analysis. The approach we take is the statistical/nancial approach.
The reason for taking the statistical/nancial approach is mostly because sta-
tistical/nancial models provide a nal result that is mathematically simple
and elegant, and this mathematically simple and elegant nal result the other
modeling approaches usually are not able to serve. This simple nal result is
5
1. INTRODUCTION
of particular importance for us because in generation investment and schedul-
ing analysis prices is only the input variable for much further computation.
Another reason for taking the statistical/nancial approach is that few people
have ever taken the statistical/nancial approach for modeling the long-term
prices, and there might be much space worthing further exploration.
Even though the modeling approach we will take is statistical/nancial, the
analysis approach we will take is both structural and fundamental. Namely, in
the process of this work, we will extensively employ structural approach and
fundamental approach to understand the fundamental operation of engineering
power systems and power markets, and based on this understanding, we then
move on to devise our modeling approach and design the price models.
Once we decided to pursue the statistical/nancial approach for modeling
the long-term prices, it immediately implies that we have to resolve its limita-
tion of vulnerability for system fundamental changes. In order to make statis-
tical models robust for the underlying changes of systems, statistical models
have to be designed to somehow contain these system changes:
a) Intuitively, one way to achieve that is to take the Hybrid Approach that
combines the Statistical Approach and the Structural Approach: namely,
one derives the parameters of the statistical models from the more fun-
damental structural models that are more capable of containing system
fundamental changes;
b) The other way is to decompose systematically the electricity spot price
data into a few components that are driven by dierent and independent
physical underlying forces, and then model each price component respec-
tively by one statistical model. Because each statistical model captures
only one type of driving forces and each type of physical forces behaves
consistently along its own time horizon, each statistical model proba-
bly is capable of capturing that one particular type of physical forces
satisfactorily. Finally a complete electricity spot price model might be
somehow constructed with the resulting statistical sub-models.
6
1.3 Elaboration of the Problem
Our research work thus has adventured into both paths, and the result
turned out that the second approach of decomposition is more feasible and
produces a desirable model. Therefore, a signicant portion of this thesis is
devoted to discussing the second approach of decomposition; and only the
very last chapter introduces an adventure of modeling the prices by Structural
Approach, which only achieves limited success.
1.3 Elaboration of the Problem
This section denes the problem by visioning the ideal result of the model and
enlists the analytical and mathematical tools that are available for solving the
problem.
1.3.1 The Requirements on the Price Model
Modeling the Physical Forces
Generation investment and scheduling analysis concerns future generator prot
from as short as a few months to as long as years and decades. Electricity spot
prices up to many years and even decades is a complex system, for that in this
long time-horizon electricity prices are driven by many dierent and indepen-
dent physical forces from both the supply and demand sides, and these physical
forces play in dierent timescales: some forces aect prices in short-term(intra-
week) and have ignorable eect beyond a week, such as regular generation
dispatches, generation forced outage, and intra-day and intra-week variations
of electricity load; the seasonal forces play at the mid-term(intra-year) level,
such as seasonal weather and temperature, seasonal hydro-generation capacity,
annual generation planned outage, etc., and these forces could inuence elec-
tricity spot prices up to a few weeks, but have negligible eect beyond a year;
in the longer run of many years and up to a few decades, the physical forces
that are manifest are economic development and economic cycles, generation
investment and retirement, and uctuations of fuel prices, etc. A model in-
tended for generation investment and scheduling analysis ought to dene and
7
1. INTRODUCTION
capture properly these various physical forces that underlie electricity spot
prices.
Time Horizon
The electricity spot price model for generation investment and scheduling anal-
ysis, therefore, ought to capture the dynamics of electricity spot prices in a
future time horizon from as short as a few months to as long as many years.
Time Unit
In terms of time unit, as between day-time and night-time electricity spot
prices change a lot, thus particularly for valuating the prot of peaking gen-
erators, which are usually only protable to operate during peak hours, an
hourly based model is necessary.
Final Result
Besides satisfying all the above requirements, the model ought to have a nal
result that is mathematically simple, for that electricity spot prices is the input
variable for generator prot valuation and further the generation investment
and scheduling analysis. Ideally, the prices at each hour ought to be captured
by a simple probability density function.
1.3.2 The Analytical and Mathematical Tools
Time-Series Models and Financial Models
The basic building blocks for the price model are borrowed from the estab-
lished eld of Time-Series Analysis and Financial Engineering. The reasons
for choosing the Time-Series/Financial Models as the basic tools are due to
that they are capable of modeling one very important feature of electricity
prices, that is, its mean-reverting nature, and that they provide a simple and
elegant nal representation for electricity spot prices as a continuous proba-
bility density function.
8
1.4 The Analysis Framework
Microeconomics Price Theory
After hundreds of years eorts of many generations of economists, the Price
Theory in Microeconomics has become a solid foundation for the science of
Economics. It provides the basic analytical tool for understanding and inter-
preting the outcomes of almost any markets that feature people who are free to
buy and to sell. Electricity markets, despite the complexity of its underlying
engineering systems and so far the inability for large-scale storage of electric-
ity, are no exception. As we will see, Microeconomics Price Theory provides a
powerful tool for analyzing and understanding electricity spot prices.
Electric Power Engineering
Underpinning any electricity market, there is a huge and complicated engineer-
ing power system, which consists of a large number of generators, a network
of transmission lines, substations, distribution networks, load centers, various
control and communication devices, and control centers. This system is oper-
ating at real-time, at any moment, power engineers have to ensure that power
generation and electricity load are always balanced. This power generation
and electricity load, if interpreted into the terms familiar to economists, power
generation corresponds to market supply and electricity load stands for mar-
ket demand. No surprise, power generation and electricity load have been the
subjects of study by power engineers for long time, and this rich expertise on
engineering power systems no doubt is a valuable source for understanding
electricity markets and hence electricity prices.
1.4 The Analysis Framework
The system of electricity prices up to a few decades is a complex system; there
are many dierent and independent physical forces under playing. To put these
various underlying physical forces in perspective, electricity spot prices will be
studied carefully in a Multi-granularity Framework, that is, electricity spot
prices will be analyzed into three time-perspectives of dierent granularity:
multi-year yearly, intra-year weekly, and intra-week hourly. In each of the
three time-perspectives, by applying the analytical tools like Microeconomics
9
1. INTRODUCTION
Price Theory and Knowledge on operations of fundamental engineering power
systems, how these physical underlying forces aect electricity spot prices will
be carefully investigated:
a) The multi-year perspective looks at the long-term trend of electricity
prices, which is under driven by the physical forces like economic devel-
opment and economic cycles, generation investment and retirement, and
uctuations of fuel prices.
b) The intra-year perspective examines the seasonal behavior of electricity
prices, which is driven by the mid-term forces such as seasonal variations
of electricity load due to seasonal weather and temperature, seasonal
hydro-generation capacity, annual generation planned outage, etc.
c) The intra-week perspective investigates the intra-day and weekday-weekend
variations of electricity spot prices, which are driven by short-term forces,
such as intra-day and weekday-weekend variations of electricity load, reg-
ular generation dispatches, generation forced outages, etc.
1.5 The Modeling Methodology
Electricity prices are driven by various physical underlying forces, and these
underlying forces are playing in dierent timescales. In other words, these
physical forces are of dierent frequencies, that is, the physical forces in the
multi-year perspective are of the lowest frequency, the intra-year forces are of
mid-frequency, and the intra-week forces the highest frequency; these physical
forces probably locate at dierent and separate bands in their frequency spec-
trum. Taking advantage of this insight, if somehow one could decompose the
electricity spot price data into a few components, of which each is driven by
a dierent and independent type of physical forces, one could thus divide the
original price modeling problem into a few sub-problems that are by nature
much simpler. This is the thought we went through when devising our plan
for solving the problem. This modeling methodology will be referred as Divide
10
1.6 Organization of the Thesis
and Conquer, which is a general strategy for solving complex problems of any
kind.
Therefore, the plan for our adventure goes as follows: we will begin with
historical electricity spot price data; the historical electricity spot price data
will then be decomposed into a few components that are driven by dierent
and independent physical underlying forces; then these price components will
be dened respectively by separate models; nally the resulting sub-models
will be constructed into a complete electricity spot price model.
1.6 Organization of the Thesis
The forthcoming chapters of this thesis are organized as follows. Chapter 2
surveys the existing literatures on electricity price modeling, namely, the par-
ticular price modeling problems, the usual approaches, popular mathematical
tools, and some special techniques. Chapter 3 applies the Microeconomics
Price Theory to electricity markets and develops the basic analytical tools for
analyzing electricity spot prices. Making use of the analytical tools that are
developed in Chapter 3, Chapter 4 introduces the Multi-granularity Framework
for analyzing electricity spot prices in three time-perspectives; in each perspec-
tive, it endeavors to understand how the operations of fundamental engineer-
ing power systems give rise to the very peculiar behaviors of electricity spot
prices. Taking advantage of the understanding on electricity spot prices that
has been gained in previous chapters, Chapter 5 proposes an electricity spot
price model that is based on the idea of decomposition, and demonstrates the
modeling methodology in the New-England electricity market. Upon some mi-
nor revisions on the model, Chapter 6 brings the same modeling methodology
to the Pennsylvania-New Jersey-Maryland(PJM) electricity market. Chap-
ter 7 records an early adventure of us trying to model electricity spot prices by
Structural Approach. Chapter 8 summarizes and concludes this thesis work.
11
1. INTRODUCTION
12
Chapter 2
A Survey of Electricity Price
Models
2.1 Electricity Price Modeling Approaches
From the point of view of a generation company, modeling and forecasting
electricity prices are either for formulating bidding strategies, scheduling gen-
eration production, pricing electricity derivatives, or generation investment
analysis. For formulating bidding strategies one concerns the day-ahead 24-
hour spot prices; for scheduling generation and pricing electricity derivatives
one cares the spot prices from as short as the next week to as long as the
next year; for generation investment analysis, one concerns the electricity spot
prices during the life time of a generator, that is, years or even decades. In
order to make these pricing, scheduling, and investment decisions, generation
companies need a thorough understanding on electricity spot prices, at best,
they want a physically well-grounded, simple, and fast electricity spot price
model to assist their making these decisions.
A price model suitable for a particular application has to at least satisfy
ve requirements: it has to capture properly the movements of electricity spot
prices, it must give the expected value of the foreseen spot prices, it must
quantify the uncertainty of the spot prices, it has to dene the prices of a time
unit that is suitable for a particular application, and it has to cover the time
horizon that a particular application concerns. There are a few mathematical
13
2. A SURVEY OF ELECTRICITY PRICE MODELS
tools, modeling approaches, or techniques that have been applied to analyze
and model electricity prices, i.e., Time Series Models, Financial/Stochastic
Process Models, Structural Models, Decomposition Techniques, etc.
In terms of Time Series Models, they start with the price data and dene
a model to capture the correlations between the prices at one time period
and the prices at the previous periods. Time series models are usually good
options for short-term applications, like forecasting next-day 24-hr and next
week 168-hr prices. We will, start with the simplest time series model and
end with some rather involved ones, step by step, discuss the nature of time
series models and their possible applications in electricity price modeling and
forecasting.
Financial/Stochastic Process Models have seen their successful applica-
tions in modeling the nancial, commodity, and energy markets. The nancial
models are dened by stochastic dierential equations. The nancial models,
though complicated at rst glance, give nal results that are mathematically
simple. These mathematically elegant nal results enable nancial models to
be used in applications that desire the least computation and the fastest speed,
such as pricing electricity futures, contracts, and other derivatives. We will
introduce, carefully, the development of nancial models in modeling electric-
ity prices, the problems that have been satisfactorily solved, and the problems
that are still remain.
Time series models and nancial models belong to Statistical Models. Both
of them use the historical data to induce, calibrate, and experiment the mod-
els. Structural Approach, on the other hand, models electricity spot prices in-
directly. It rst models the physical forces that underlie electricity spot prices
by a few constituent models, then constructs these constituent sub-models into
a complete electricity spot price model. For that Structural Approach utilizes
the more fundamental knowledge on electricity markets, it could model elec-
tricity spot prices of a longer time-horizon than that the statistical approach
could usually do. We will briey introduce the idea of structural models, their
advantages and disadvantages, and the literatures along to this line.
Besides the statistical approach and the structural approach in modeling
electricity prices, there is another method for studying electricity markets,
14
2.2 Time Series Models
which will be referred as Fundamental Approach. Compared with the struc-
tural approach, fundamental approach is of much further detail. It may model
each generator unit, behavior of each market player, seasonal generation main-
tenance and available generation capacity, the dynamics of electricity load
demand, the transmission networks, the rules of the electricity market, the
investment cycles of the electric power sector, etc. It then uses these detailed
models to construct electricity spot prices. Due to these further details, prob-
ably there is no analytical result for a fundamental model. In this survey,
we will not give any further discussions on the fundamental approach, inter-
ested readers please refer to the literatures, such as, Angelus (2001); Bastian
et al. (1999); Baughman & Lee (1992); Bessembinder & Lemmon (Jun. 2002);
Olsina et al. (2006); Ruibal & Mazumdar (2008); Wang & Mazumdar (2007).
Finally, decomposition techniques probably have huge potential in analyz-
ing and modeling electricity prices. Electricity prices has a complex nature:
intra-week hourly spot prices has an intra-day pattern and a weekday-weekend
pattern; intra-year weekly prices has a seasonal pattern, and in dierent sea-
sons prices behave distinctly; and short-life price spikes prevail in electricity
markets all over the world. This complex nature of electricity prices is due
to the various physical driving forces that underlie electricity spot prices. If
the electricity spot prices could somehow be decomposed into a few price com-
ponents that are driven by the dierent and independent physical underlying
forces, the original problem then is divided into several sub-problems that
are by nature much simpler and thus much easier to solve. We will introduce
briey the various decomposition techniques, such as Fourier Analysis, Wavelet
Analysis, and Principal Component Analysis, and discuss their possible appli-
cations in electricity price analysis and modeling.
2.2 Time Series Models
2.2.1 The Autoregressive Model
The discussions on the time series electricity price models are based on Franses
(1998); Hamilton (1994); Wei (2006); Weron (2006). We will start with the
15
2. A SURVEY OF ELECTRICITY PRICE MODELS
simplest Autoregressive model of order One, namely the AR(1) model. Denote
a time series as x
k
, k = 1, ...., T, where the tilde denes a random variable.
An AR(1) model denes x
k
as a function of its value in the previous time
period x
k1
, as
x
k
= c + x
k1
+
k
0 < < 1
in which c is a constant,
k
is a random variable of Normal Distribution with
standard deviation ,
k
N(0, 1). Given the initial value x
0
, x
k
could be
re-written as,
x
k
=
k
x
0
+c
_
1 + +... +
k1
_
+
k1

1
+
k2

2
+... +
k
=
k
x
0
+c(
1
k
1
) +
k1

1
+
k2

2
+... +
k
Thus given x
0
, the expected value and variance of x
k
are,
E{ x
k
} =
k
x
0
+c(
1
k
1
) VAR{ x
k
} =
1
2k
1

2
which implies that, when time goes to innite, k , the expected value
of x
k
reaches a constant, lim
k
E{ x
k
} =
c
1
, this constant will be referred as
the unconditional mean and denoted as

=
c
1
, and the variance of x
k
also
reaches a constant, lim
k
VAR{ x
k
} =

2
1
.
This simplest AR(1) model of time-lag one can be extended to have more
time lags, say, a AR(m) model, it is dened by x
k
and its m time lags, x
k1
,
x
k2
, . . . , and x
km
,
x
k
= c + (
1
x
k1
+
2
x
k2
+... +
m
x
km
) +
k
which is re-organized as,
x
k

1
x
k1

2
x
k2
...
m
x
km
= c +
k
(2.1)
More concisely it is written as:
a(L) x
k
= c +
k
16
2.2 Time Series Models
where a(L) = 1
1
L
1

2
L
2
...
m
L
m
. The unconditional mean of AR(m)
is found by taking the expectation on both sides of Equation (2.1),
E{ x
k

1
x
k1

2
x
k2
...
m
x
km
} = E{c +
k
}
E{ x
k
}
1
E{ x
k1
}
2
E{ x
k2
} ...
m
E{ x
km
} = c + E{
k
}

1

2
...
m
= c
(1
1

2
...
m
) = c
Thus,
=
c
1
1

2
...
m
The unconditional mean is a constant, namely, given the initial value x
0
,
when time goes to innite k , the expected value of x
k
is a constant.
AR and Electricity Prices
AR models assume a constant unconditional mean, while electricity spot prices
have a time-varying mean, and its mean varies by the time of day, week and
year. AR models alone thus are not suitable for modeling electricity prices,
but they are ideal candidates for capturing the mean-reverting stochastic com-
ponent of electricity prices, therefore, AR models, as we will see, will serve as
an vital component in building an electricity price model.
2.2.2 AR with Time-varying Mean
Lets introduce a time-varying function f(k) into the simplest AR(1) model
x
k
, the resulting model is named AR model with a Time-varying Mean(ARV),
denoted as y
k
,
y
k
= f(k) + x
k
Here x
k
is simplied to have the constant c = 0, thus it is x
k
= x
k1
+
k
.
Write y
k
into the form of AR(1),
y
k
= [f(k) f(k 1)] + y
k1
+
k
= g(k) + y
k1
+
k
where g(k)

= f(k) f(k 1).
17
2. A SURVEY OF ELECTRICITY PRICE MODELS
ARV and Electricity Prices
An ARV model is the sum of a deterministic time-varying function f(k) and
an AR process x
k
. Electricity prices are observed to have a time-varying
mean and a stochastic component that is mean-reverting. The deterministic
function f(k) could model the time-varying mean of electricity prices, and
the AR process x
k
could model the mean-reverting stochastic component of
electricity prices. The design of the deterministic function f(k) depends on
the specic modeling problems, it could include the intra-day and weekday-
weekend pattern of spot prices, and it may also include the annual seasonal
pattern. The specication of the AR process x
k
, say, how many time lags to
include in the model, may depend on data and particular applications. ARV
models have been empirically applied to model electricity prices by Bhanot
(2000); Escribano et al. (2002); Hamm & Borison (2006); Knittel & Roberts
(2005); Misiorek et al. (2006); Rambharat et al. (2005); Serna & Villaplana
(2007); Weron & Misiorek (2008).
2.2.3 AR with Exogenous Variables
Another way to model the time-varying function f(k) is to substitute f(k)
with a function of an exogenous variable. The exogenous variable is denoted
as z
k
. AR models with eXogenous variables are named ARX models. An ARX
model y
k
with the time lags of the exogenous variable z
k
is,
y
k
= (c +
0
z
k
+
1
z
k1
+... +
q
z
kq
)+(
1
y
k1
+
2
y
k2
+... +
m
y
km
)+
k
Move the z
k
terms to the left and write it in a more concise form,
y
k

1
y
k1

2
y
k2
...
m
y
km
= c +
0
z
k
+
1
z
k1
+... +
q
z
kq
+
k
_
1
1
L
1

2
L
2
...
m
L
m
_
y
k
= c +
_

0
+
1
L
1
+... +
q
L
q
_
z
k
+
k
a(L) y
k
= c +c(L) z
k
+
k
where a(L) = 1
1
L
1

2
L
2
...
m
L
m
, and c(L) =
0
+
1
L
1
+... +
q
L
q
.
18
2.2 Time Series Models
ARX and Electricity Prices
Driven by the intra-day, intra-week, and seasonal variations of electricity load
demand, electricity spot prices have a time-varying mean, and its mean varies
by the time of day, week, and year. Besides electricity load, electricity spot
prices are also aected by generation capacity, fuel prices, etc. Capable of in-
cluding these exogenous variables into consideration, ARX models are proba-
bly able to improve the accuracy of price forecasting, see Karakatsani & Bunn
(2004); Misiorek et al. (2006); Rambharat et al. (2005); Weron & Misiorek
(2008).
2.2.4 The Autoregressive Moving Average Model
The simplest Moving Average Model of order one is MA(1) Model,
x
k
= c +
0

k
+
1

k1
It says that x
k
equals to a constant c plus the weighted average of time-lagged
random noises
k
and
k1
. In other words, if x
k
is subtracted by the constant
c, the residuals r
k

= x
k
c =
0

k
+
1

k1
are highly correlated. A general
Moving Average Model of time lag n, MA(n), is,
x
k
= c + (
0

k
+
1

k1
+... +
n

kn
)
= c +b(L)
k
where b(L) =
0
+
1
L +... +
n
L
n
.
ARMA stands for Autoregressive-Moving Average. The purpose for ex-
tending an AR model to an ARMA model is to model the autocorrelation
of residuals. An ARMA model is a combination of an AR model and a MA
model. The simplest ARMA model is a combination of AR(1) and MA(1),
namely, ARMA(1,1),
(1
1
L) x
k
= c + (
0
+
1
L)
k
a(L) x
k
= c +b(L)
k
A general ARMA model, ARMA(m,n) is:
a(L) x
k
= c +b(L)
k
where a(L) = 1
1
L
1

2
L
2
...
m
L
m
, and b(L) =
0
+
1
L+... +
n
L
n
.
The unconditional mean of the an ARMA(m,n) model is =
c
1
1

2
...
m
.
19
2. A SURVEY OF ELECTRICITY PRICE MODELS
ARMA and Electricity Prices
The same as an AR model, an ARMA model assumes a constant uncondi-
tional mean. Thus it only models the mean-reverting stochastic component of
electricity prices. An ARMA model is of more complexity than an AR model.
Whether an AR model or an ARMA model is a better option for modeling the
mean-reverting stochastic component of electricity prices probably depends on
data and particular applications. Whether the extra complexity of an ARMA
model could be justied has to be carefully judged, for that a model is not at
all merely veried by its tting of the historical data, but more importantly it
ought to be measured by its ability in forecasting. A model of more complexity
is usually more vulnerable to fundamental changes of systems. Literatures on
modeling electricity prices by ARMA models are Garcia et al. (2005); Huurman
et al. (2008); Nogales & Conejo (April 2006); Swider & Weber (2007).
2.2.5 ARMA with Time-varying Mean
In analogy to the extension from AR model to an ARV model, an ARMA model
could be extended to an ARMA model with Time-varying Mean, denoted as
ARMAV model. An ARMAV model, denoted as y
k
, is the summation of a
deterministic time-varying function f(k) and an ARMA process a(L) x
k
=
c +b(L)
k
, as
y
k
= f(k) + x
k
Written in terms of y
k
, it is,
a(L) y
k
= g(k) +b(L)
k
ARMAV and Electricity Prices
The dierence between an ARMAV model and an ARV model is the way they
model the stochastic mean-reverting component of electricity prices. Whether
an ARMAV model or an ARV model is a better option for modeling electricity
prices probably depends on specic empirical cases. Empirical evidences from
Crespo Cuaresma et al. (2004); Weron & Misiorek (2005) show that ARMAV
models do not outperform the simpler ARV models. The moral lesson from
20
2.2 Time Series Models
these empirical studies is that, when a model of higher complexity is not well
justied by the particular modeling problem in concern, one must be cautious
to move beyond a simpler model and prefer to a model of higher complexity,
even though the model of higher complexity usually gives a better tting of
historical data.
2.2.6 ARMAX and Transfer Function
If one replaces the deterministic function g(k) by a function of exogenous vari-
ables, c+c(L) z
k
, the ARMAV model becomes an ARMA model with eXogenous
variables, referred as an ARMAX model,
a(L) y
k
= c +c(L) z
k
+b(L)
k
where c(L) =
0
+
1
L
1
+... +
q
L
q
. Divide a(L) from both sides, one arrives
at the ARMAX model written in the form of a Transform Function,
y
k
=
c
a(L)
+
c(L)
a(L)
z
k
+
b(L)
a(L)

k
ARMAX Model and Electricity Prices
The dierence between an ARMAX model and an ARX model is again the
way they model the mean-reverting stochastic component of electricity prices.
Though probably a ARMAX model of higher complexity ts the historical data
better, it does not necessarily outperform a simpler ARX model in forecasting.
Empirical studies on ARMAX models are Conejo et al. (2005a); Garcia et al.
(2005); Huurman et al. (2008); Knittel & Roberts (2005); Nogales & Conejo
(April 2006); Nogales et al. (2002); Swider & Weber (2007); Weron & Misiorek
(2005).
2.2.7 Periodic Autoregressive Models
In previous sections, we model the time-varying mean of electricity prices with
a deterministic function. In this section, we will introduce another way of mod-
eling the time-varying mean, which is Periodic Autoregressive Model(PAR).
21
2. A SURVEY OF ELECTRICITY PRICE MODELS
Lets look at the hourly electricity spot prices across several days, for ex-
ample, there are N days, each day has 24 hours, thus there are 24 N spot
prices. We are going to separate the hourly spot prices into 24 time series
according to their hour index, which is the order of a particular hour in the 24
hours in a day.
Hour 1 Hour 2 . . . Hour 24
Day 1 x
1,1
x
2,1
. . . x
24,1
Day 2 x
1,2
x
2,2
. . . x
24,2
.
.
. . . . . . . . . .
.
.
.
Day N x
1,N
x
2,N
. . . x
24,N
Table 2.1: Separate the Intra-day Hourly Spot Prices into 24 Time Series
Day 1 k Day k
Hour
th
i
Hour
th
1 i 1, i k
x

, i k
x
, 1 i k
x

1, 1 i k
x


Figure 2.1: Index the Intra-day Hourly Spot Prices
The spot prices are organized in Table 2.1: the columns expand the 24
hours, and the rows spread the days. Instead of considering the 24 N prices
as a whole, we group them according to their hour index, namely, the index
of the columns. Take the rst column for example, we took the price at the
1
st
hour of each day, order them by day 1, day 2, ..., till day N, we thus
obtained a new segment time series consisting of x
1,1
, x
1,2
, . . ., x
1,N
, and put
them in the rst column. Continue the same procedure for the rest of the 24
hours, we obtain 24 segment time series. This way of separating the original
22
2.2 Time Series Models
time series according to their hour index is the key concept underlying PAR
Models, Figure 2.1.
If one models each segment time series with one AR model, there will be
24 AR models. In the simplest case, each AR model is an AR(1) model, the
model for hour i, thus is,
x
i,k
= c
i
+
i
x
i,k1
+
i,k
i = 1, ...24
The above PAR model could be extended to include the spot price at the
immediate previous hour x
i1,k
, one then arrives at a more general PAR model,
x
i,k
= c
i
+
i
x
i,k1
+
i
x
i1,k
+
i,k
i = 1, ...24
A general PAR model could have more price lags. The purpose of PAR model
is to model prices at dierent hours by dierent AR models.
PAR and Electricity Prices
For hourly electricity spot prices, on-peak prices are more volatile than the
o-peak prices. It is reasonable to model on-peak prices and o-peak prices
by dierent AR models. Hourly spot prices have three orders of seasonality:
intra-day pattern, weekday-weekend pattern, and intra-year seasonality. A
PAR model that treats each hour respectively avoids modeling the intra-day
pattern. For daily average prices, prices on weekdays are usually more volatile
than these on weekends, it is reasonable to model weekday and weekend prices
with dierent AR models. Empirical Studies on PAR models could be found
in Crespo Cuaresma et al. (2004); Garcia-Martos et al. (2007); Guthrie &
Videbeck (2007); Huisman et al. (2007).
2.2.8 ARIMA and its Extensions
The simplest ARIMA model is an ARIMA(0,1,0), which is a Random Walk
model,
x
k
x
k1
= x
k
=
k
It describes a process whose increase from time k 1 to k is random and the
noise
k
is characterized by a random variable of Standard Normal Distribu-
tion. In some cases, the rst order dierence, namely the increase x
k
, may
23
2. A SURVEY OF ELECTRICITY PRICE MODELS
require more specications. One may describe it with an AR(1) Model. Lets
denote x
k

k
,

k
= c +

k1
+
k
Reorganize it as,
(1 L)

k
= c +
k
We call it an ARI(1,1) model. The ARI(1,1) could be extended into a more
complicated ARIMA(1,1,1) model, where the residuals are assumed to be cor-
related as a MA(1) process,
(1 L)

k
= c + (1 +L)
k
A general ARIMA(m,1,n) model thus is:
a(L)

k
= c +b(L)
k
In some cases one may need to describe the second order dierences,

2
x
k
=

k1
= x
k


x
k1
we then arrive at an ARIMA(m,2,n) model,
a(L)
2
x
k
= c +b(L)
k
Higher order of dierences is denoted as
d
x
k
, thus a general ARIMA(m,d,n)
model describing the d
th
order dierence is,
a(L)
d
x
k
= c +b(L)
k
ARIMA and Electricity Prices
The nature of ARIMA model is to model the price dierences. Based on our
understanding on electricity spot prices, which are the intersections of marginal
cost curve and electricity load demand, we think it is the original prices, rather
than the dierences of prices, that probably have a systematic structure. Em-
pirical studies concerning the ARIMA models could be found in Conejo et al.
(2005a,b); Contreras et al. (2003); Garcia et al. (2005); Huurman et al. (2008).
24
2.2 Time Series Models
2.2.9 Regime-switching Models
A regime-switching model assumes two regimes, R
1
and R
2
, the switching
between the two regimes is governed by a Markov Chain, Figure 2.2, and the
transient probability between the two regimes are,

11

= Pr
_

R(t) = R
1


R(t dt) = R
1
_

12

= Pr
_

R(t) = R
2


R(t dt) = R
1
_

21

= Pr
_

R(t) = R
1


R(t dt) = R
2
_

22

= Pr
_

R(t) = R
2


R(t dt) = R
2
_
12

11

22

21

2
R
1
R

Figure 2.2: A Two-regime Markov Chain
A regime-switching process p
t
has two sub-processes, p
1t
and p
2t
: when the
Markov Chain is in the rst regime

R(t) = R
1
, p
t
= p
1t
; and when it is in the
second regime

R(t) = R
2
, p
t
= p
2t
. Namely,
p
t
=
_
p
1t
if

R(t) = R
1
p
2t
if

R(t) = R
2
The two processes p
1t
and p
2t
each could be an AR process or extensions of an
AR process. The parameters of the model p
1t
and p
2t
are dierent, by this way
they are designed to capture the dierent price behaviors in the two dierent
regimes.
Regime-switching Models and Electricity Prices
During one day, on-peak prices are more volatile than the o-peak prices;
during one year, prices are usually higher and more volatile during the high
25
2. A SURVEY OF ELECTRICITY PRICE MODELS
demand seasons than these during the low demand seasons. Driven by some
extreme weather, when electricity load demand reaches some very high levels,
price spikes tend to occur. Overall, when prices are low, high, or in spike,
they behave dierently. Regime-switching price models are designed to cap-
ture these dierent behaviors of electricity prices in dierent time. Empirical
studies applying the regime switching models to electricity prices are Huisman
& Mahieu (2003); Mount et al. (2006); Swider & Weber (2007); Weron et al.
(2004).
2.2.10 Time-varying Volatility
Lets rst recall the AR(1) with Time-varying Mean model, see Section 2.2.2,
y
k
= g(k) + y
k1
+
k

k
N(0, 1)
where the volatility , measured by the standard deviation of the noise
k
, is
constant. Now the constant volatility is to be relaxed as time-varying, namely,
the standard deviation becomes a function of time k, as
k
. The ARV model
with time-varying volatility thus is,
y
k
= g(k) + y
k1
+
k

k

k
N(0, 1)
2.2.11 The GARCH Model
GARCH stands for Generalized Autoregressive Conditional Heteroscedasticity.
GARCH models are designed to capture the stochastic time-varying volatility.
The simplest GARCH(1,1) model is,

2
k
= c +
2
k1
+(
k1

k1
)
2
where + < 1, and
k
N(0, 1).
The term
2
k1
is an autoregressive model. It models the dependence of
the variance
2
k
on its lagged value
2
k1
. It means that the volatility tends
to persist: high volatility probably to be followed by high volatility and low
volatility is probably to be followed by low volatility; namely high volatility
periods tend to occur in cluster, so do the low volatility periods. The last term
26
2.2 Time Series Models
(
k1

k1
)
2
models the eect of the innovation noises: a large noise at time
k 1 will cause the variance
2
k
in the time k to be large, and a small noise at
time k 1 contributes little to the variance
2
k
.
Given the initial variance at time 0,
2
0
, the variance at time k,
2
k
, is derived
as,

2
k
=
2
0
k

i=1
_
+
2
ki
_
+c
_
1 +
k1

j=1
j

i=1
_
+
2
ki
_
_
The expected value of the variance
2
k
thus is,
E
_

2
k
_
=
2
0
k

i=1
_
+E
_

2
ki
__
+c
_
1 +
k1

j=1
j

i=1
_
+E
_

2
ki
__
_
=
2
0
( +)
k
+c
_
1 +
k1

j=1
( +)
j
_
=
2
0
( +)
k
+c
_
1 +
( +) + ( +)
k
1 ( +)
_
As time k goes to innite, k , its expected value is lim
k
E {
2
k
} =
c
_
1 +
(+)
1(+)
_
. Namely, the unconditional mean of
2
k
is a constant.
GARCH and Electricity Price Volatility
The volatility of electricity spot prices is time-varying, see Section 3.8, specif-
ically, the volatility is price-dependent: high prices are of high volatility; low
prices are of low volatility. The volatility of hourly electricity spot prices thus
is cyclic and highly predictable: price volatility during the day-time is higher
than that during the night-time, and the price volatility during the high de-
mand seasons is higher than that during the low demand seasons.
GARCH model, by its nature, is a stochastic process. The unconditional
mean of a GARCH process is a constant, it means that the variance has a
tendency to revert to a constant level. And in the GARCH model, that the
variance is time-varying is due to the random return noises. The GARCH
Model, therefore, is only capable of modeling the stochastic component of the
time-varying volatility of electricity prices, and it is not able to model the
27
2. A SURVEY OF ELECTRICITY PRICE MODELS
deterministic part of the time-varying volatility of electricity prices. The de-
terministic part of the time-varying volatility of electricity prices, probably, is
more signicant than the stochastic counterpart. Empirical studies on apply-
ing GARCH models to electricity prices modeling are Escribano et al. (2002);
Garcia et al. (2005); Hadsell & Shawky (2006); Hadsell et al. (2004); Higgs &
Worthington (2005); Karakatsani & Bunn (2004); Knittel & Roberts (2005);
Serna & Villaplana (2007); Swider & Weber (2007); Worthington et al. (2005).
And general discussions on the volatility of electricity prices are Li & Flynn
(2004b); Nakamura et al. (2006); Robinson & Baniak (2002); Simonsen (2005);
Zareipour et al. (2007).
2.2.12 GARCH with Asymmetric Eect
The standard GARCH model could be extended to model the asymmetric
eect of noises on the volatility. For example, the positive noises,
k1
> 0,
may have larger eect on the volatility than the negative ones,
k1
< 0.
The GARCH model with Asymmetric Eect is referred as Threshold-GARCH
model. The model is written as,

2
k
= c +
2
k1
+(
k1

k1
)
2
+ (
k1

k1
)
2
d
k1
where d
k1
=
_
1 if
k1
> 0
0 if
k1
< 0
.
The additional term (
k1

k1
)
2
d
k1
dierentiates the asymmetric eect
of the positive and negative noises. If > 0, the model implies that positive
noises,
k1
> 0, have more signicant eects on the volatility
k
than the
negative noises; if < 0 , the model indicates that the positive noises have less
eect on the volatility than the negative noises.
Threshold-GARCH and Electricity Price Volatility
The supply stack of an electricity market is usually convex, a positive shock of
electricity load could send load high and make it more probably to intersect the
supply stack at steeper segments, follows will be more volatile electricity prices;
on the other hand, a negative shock of electricity load, could send the load low
and make it more probably to intersect the supply stack at the at part,
28
2.2 Time Series Models
follows will be less volatile electricity prices. The TGARCH model, designed
for modeling the asymmetric eect of the noises, therefore, may match some
realities in electricity prices. One empirical study on the Threshold-GARCH
is Knittel & Roberts (2005). It observes that positive price shocks have a
larger eect on the volatility than the negative shocks.
2.2.13 Time-varying Volatility and Model Parameter Cal-
ibration
Lets use a simple linear least-square regression example to illustrate the eect
of time-varying volatility on parameter calibration. We will start with constant
volatility and than add time-varying volatility.
a) In the case of constant volatility, a regression model of three time periods
is,
_
_
y
1
y
2
y
3
_
_
=
_
_
x
1
x
2
x
3
_
_
+
_
_

3
_
_
Write it in vectors as, y = x + . The parameter calibration is to nd
the optimal that minimizes , as
min

f()=(y x)
T
(y x)
Take the rst order derivative and the optimal solution satises the
condition,
df()
d

=
= 0
one thus has x
T
y x
T
x = 0, thus the estimate is,
=
_
x
T
x
_
1
x
T
y =
_
3

i=1
x
2
i
_
1
3

i=1
x
i
y
i
b) If the volatility is relaxed to be time-varying, namely, the model becomes,
_
_
y
1
y
2
y
3
_
_
=
_
_
x
1
x
2
x
3
_
_
+
_
_

3
_
_
29
2. A SURVEY OF ELECTRICITY PRICE MODELS
where
1
,
2
, and
3
are dierent, written into vectors as, y = x+ .
The optimization problem becomes,
min

f()=(y / x /)
T
(y / x /)
denote y


= y /, and x


= x /, one has,
min

f()=(y

)
T
(y

)
Take derivative and let
df()
d

=
= 0, namely, x

T
y

T
x

= 0, thus
the estimate is,
=
_
x

T
x

_
1
x

T
y

=
_
3

i=1
x
2
i

2
i
_
1
3

i=1
x
i
y
i

2
i
The above equation tells that
i
determines the weight of the point (x
i
, y
i
).
If
i
,the eect of (x
i
, y
i
) on is canceled out; if
i
0, solely depends
on (x
i
, y
i
). If
1
=
2
=
3
, each point (x
i
, y
i
), i = 1, 2, 3 is weighed equally
in calculating . If
1
=
2
=
3
= 1, it becomes the example of constant
volatility.
Up to this point, it is natural to infer that if a time series is generated by a
process by nature of time-varying volatility, and then one calibrates it with a
model which assumes a constant volatility, surely the estimated parameters are
biased. The extent of the bias depends on the dierence between the assumed
volatility and the true volatility.
2.3 Financial/Stochastic Process Models
2.3.1 Modeling the Multi-seasonality
Electricity spot prices, driven by the temporal variations of electricity load,
have three orders of seasonality: during a day, due to the high electricity de-
mand in the day-time and the low demand in the night-time, prices are higher
during the day-time and lower during the night-time; during a week, caused by
30
2.3 Financial/Stochastic Process Models
the higher demand during the weekdays and lower demand during the week-
ends, prices during the weekdays are higher and prices during the weekends
are lower; during a year, driven by the seasonal variations of electricity load,
prices are usually higher during the high demand seasons and lower during the
low demand seasons.
24 Hrs
Figure 2.3: Intra-day Electricity Price Pattern
In terms of the intra-day pattern, Figure 2.3, in the literature, most authors
have avoided modeling it by only considering the daily average prices. A few
authors have analyzed statistically the intra-day pattern for dierent markets,
see Chan et al. (2008); Johnson & Barz (1999); Knittel & Roberts (2005); Li &
Flynn (2004a); Longsta & Wang (2004); Wilkinson & Winsen (2002); some
have modeled the intra-day pattern using typical intra-day proles of dierent
day-types, such as business days, non-business days, and days in dierent
seasons, refer to Culot et al. (2006); Skantze & Ilic (2001); Vollbrecht (2008);
and one particular work is Branger et al. (2009) who models the intra-day
patterns with Fourier Series up to the 4th order.
In terms of the weekday-weekend pattern, Figure 2.4, the following authors
compared the dierent intra-day patterns between weekdays and weekends,
see Chan et al. (2008); Johnson & Barz (1999); Knittel & Roberts (2005); Li &
Flynn (2004a); some authors avoided modeling the weekday-weekend pattern
31
2. A SURVEY OF ELECTRICITY PRICE MODELS
Wk
7 Days
Mon Tue Wed Thu Fri Sat Sun
Figure 2.4: Intra-week Electricity Price Pattern
by only considering the weekday or business day prices, see Mari (2008); Meyer-
Brandis & Tankov (2008); Skantze & Ilic (2001); among the authors who have
endeavored to model the weekday-weekend pattern, most of them choose to
model it by dummy variables, which was initiated by Lucia & Schwartz (2002)
and then followed by many others, see Bierbrauer et al. (2007); Borovkova &
Permana (2006); Branger et al. (2009); De Jong (2006); Escribano et al. (2002);
Huisman & Mahieu (2003); Kosater & Mosler (2006); Nomikos & Soldatos
(2008); Seifert & Uhrig-Homburg (2007); Truck et al. (2007).
1
Wks
2
52
1Yr
4-Wks
Figure 2.5: Intra-year Seasonality of Electricity Prices
32
2.3 Financial/Stochastic Process Models
The intra-year seasonal pattern, Figure 2.4, is relatively well studied. Most
authors model the seasonal pattern by Fourier Series, which is a linear com-
bination of sine and cosine functions of various frequencies. Modeling the
seasonal pattern by Fourier Series was started by Lucia & Schwartz (2002);
Pilipovic (1998), and then followed by the others, see Bierbrauer et al. (2007);
Borovkova & Permana (2006); Cartea & Figueroa (2005); Culot et al. (2006);
De Jong (2006); Escribano et al. (2002); Geman & Roncoroni (2006); Kosater &
Mosler (2006); Kresen & Husby (2000); Mari (2008); Meyer-Brandis & Tankov
(2008); Nomikos & Soldatos (2008); Seifert & Uhrig-Homburg (2007); Voll-
brecht (2008). Among these works, most of them use the Fourier Series up
to the second order, that is, the sine and cosine functions of period one year
and half a year. One exception is Cartea & Figueroa (2005), who models the
seasonal pattern of the England and Wales with a Fourier Series up to the 5th
order. An alternative to Fourier Series for modeling the seasonal pattern is
monthly or seasonally dummy variables, see Bierbrauer et al. (2007); Branger
et al. (2009); Knittel & Roberts (2005); Lucia & Schwartz (2002); Truck et al.
(2007).
For dierent markets, the intra-year seasonal patterns are probably dif-
ferent. The US and European electricity markets are relatively well stud-
ied. For modeling the US markets, see De Jong (2006); Geman & Ron-
coroni (2006); Knittel & Roberts (2005); Mari (2008); Meyer-Brandis & Tankov
(2008); Pilipovic (1998); Skantze & Ilic (2001), and for studying the Euro-
pean markets, refer to Bierbrauer et al. (2007); Borovkova & Permana (2006);
Branger et al. (2009); Cartea & Figueroa (2005); Culot et al. (2006); De Jong
(2006); Escribano et al. (2002); Kosater & Mosler (2006); Kresen & Husby
(2000); Lucia & Schwartz (2002); Mari (2008); Meyer-Brandis & Tankov (2008);
Nomikos & Soldatos (2008); Seifert & Uhrig-Homburg (2007); Truck et al.
(2007); Vollbrecht (2008).
2.3.2 Modeling the Mean-reverting Price Nature
Electricity spot prices are mean-reverting. By mean-reverting it is meant that
the prices, whenever it goes astray from its equilibrium levels, tend to be
33
2. A SURVEY OF ELECTRICITY PRICE MODELS
pulled back to the equilibrium levels. The reason for its being mean-reverting
is due to the balancing eect of the fundamental driving forces from both
the supply and demand sides. If we view the electricity spot prices within a
perspective of a week, the electricity spot prices follow an intra-day pattern
and a weekday-weekend pattern, and these intra-day and weekday-weekend
patterns reect the marginal cost of generating electricity in the short run. If
we view the prices with a perspective of a year, the prices swing around a highly
predictable seasonal pattern. Similarly with a perspective of many years, the
prices vary around a long-term trend that is determined by the marginal cost
of generating electricity in the long run. In short, either in the short-run of
hours, mid-term of weeks, or long-run of years, electricity spot prices reect
the marginal cost of supplying electricity, and whenever it deviates from the
equilibrium level of marginal cost, it tends to be pulled back to the equilibrium
levels, refer to Chapter 4.
To model this mean-reverting nature of electricity spot prices, the Mean-
reversion process, also named OU-process, is a natural option, refer to Ap-
pendix C.4. The simplest Mean-reversion process x
r
is,
d x = k xdt +d z k > 0
Look at the stochastic dierential equation, the random noise term d z is
constantly driving x
r
(t) away from the equilibrium level 0, and the reverting
term k x
r
dt is always pulling x
r
(t) back to the equilibrium level 0. If we take
the expected value on both sides of the stochastic dierential equation,
E {d x
r
} = E {k x
r
dt} +E {d z}
Re-denoted as,
d x
r
= k x
r
dt
Thus the derivative of x
r
(t) by time t is
d x
r
dt
= k x
r
, which implies that: when
x
r
(t) < 0, as k > 0 the derivative
d x
r
dt
= k x
r
> 0, physically it means that
x
r
(t) is being pulled up to 0; on the other hand, when x
r
(t) > 0, the derivative
d x
r
dt
= k x
r
< 0, it means that x
r
(t) is being pulled down to 0.
34
2.3 Financial/Stochastic Process Models
The physical phenomenon that the Mean-reversion process describes matches
the mean-reverting behavior of electricity spot prices well. In the litera-
ture, Mean-reversion process has been widely used in modeling the commod-
ity markets in general, see Schwartz (1997); Schwartz & Smith (2000). The
pioneers who brought the Mean-reversion process to the electricity markets
are Bhanot (2000); Deng (2000); Johnson & Barz (1999); Lucia & Schwartz
(2002); Pilipovic (1998). Mean-reversion process, thereafter, has been ad-
vocated by many other authors and became a standard tool for modeling
electricity spot prices, see Barlow et al. (2004); Benth et al. (2007); Bier-
brauer et al. (2004, 2007); Blanco & Soronow (2001a); Borovkova & Permana
(2006); Branger et al. (2009); Cartea & Figueroa (2005); Clewlow & Strick-
land (2000); Culot et al. (2006); De Jong (2006); Diko et al. (2006); Escrib-
ano et al. (2002); Geman & Roncoroni (2006); Hambly et al. (2009); Huis-
man & Mahieu (2003); Knittel & Roberts (2005); Kosater & Mosler (2006);
Mari (2008); Meyer-Brandis & Tankov (2008); Mount et al. (2006); Nomikos
& Soldatos (2008); Schmidt (2008); Seifert & Uhrig-Homburg (2007); Skantze
& Ilic (2001); Villaplana (2004).
2.3.3 A Basic Electricity Price Model
After understanding the two peculiar features of electricity prices, that is,
multi-seasonality and mean-reversion, as well as the mathematical tools that
are appropriate for modeling each of them, lets build a simplest electricity
spot price model that captures these two most basic behaviors of electricity
spot prices. This basic model is an additive model as the summation of a time-
varying deterministic function f(t) and a mean-reversion process x
r
(t), and in
the forthcoming analysis it will be referred as an Arithmetic Mean-reversion
Process with Time-varying Mean process,

S(t)

= f(t) + x
r
(t)
Written into the stochastic dierential equations, refer to Appendix C.4, it is,
d

S = k
_
g(t)

S
_
dt +d z
35
2. A SURVEY OF ELECTRICITY PRICE MODELS
This Arithmetic Mean-reversion Process with Time-varying Mean process, de-
spite its simplicity, captures the two most basic behaviors of electricity spot
prices. In the later discussions, when we move on to have a deeper understand-
ing on electricity prices, the model will turn out to be inadequate. Despite its
inadequacy, as we will see, it is to serve as an vital basic starting point for
further modication and experiment.
2.3.4 Modeling the Time-varying Volatility
The volatility of electricity prices is time-varying: look at the hourly prices
during a day, during the day-time when prices is high, the prices is usually
more volatile; look at the prices during a week, the prices during the week-
days is higher and more volatile than these during the weekends; look at the
prices during a year, the prices during the high demand seasons are higher and
more volatile than these during the low demand seasons. Observing this time-
varying volatility of prices carefully, one notices that the volatility of prices is
proportional to the level of the prices, namely, when prices are higher, it is usu-
ally more volatile; when prices are lower, it is usually milder. More specically,
this time-varying volatility will be referred as price-dependent volatility.
To build a price model that captures this price-dependent volatility of elec-
tricity spot prices, lets start with the Arithmetic Mean-reversion Process with
Time-varying Mean process,
d

S = k
_
g(t)

S
_
dt +d z
Even though the Arithmetic Mean-reversion Process with Time-varying Mean
process captures well the multi-seasonality and the mean-reverting nature of
prices, it does not model the price-dependent volatility, for that in the noise
term d w

= d z the volatility is a constant and not a function of price

S.
To model the price-dependent volatility of electricity spot prices, an al-
ternative to the Arithmetic Mean-reversion Process with Time-varying Mean
process is the Geometric Mean-reversion with Time-varying Mean process, Ap-
pendix C.6, which models the prices with the exponential of the Arithmetic
36
2.3 Financial/Stochastic Process Models
Mean-reversion process, namely,

S(t)

= e
f(t)+ x
r
(t)
= e
f(t)
e
x
r
(t)

= G(t)

X
r
(t)
To see the nature of the Geometric Mean-reversion with Time-varying Mean
process clearly, lets write it into the form of Stochastic Dierential Equation,
d

S(t) = k
_
g

(t) ln

S(t)
_

S(t)dt +

S(t)d z
in which the noise term d

W =

S(t)d z now is proportional to the price



S(t)
; also, the reverting ratio, k

= k

S(t) is proportional to price

S(t). In
other words, the Geometric Mean-reversion with Time-varying Mean process
characterizes the behavior of electricity prices price-dependently, both in terms
of volatility and the reverting ratio.
If the

S(t) term is divided from both sides, the stochastic dierential equa-
tion becomes,
d

S(t)

S(t)
= k
_
g

(t) ln

S(t)
_
dt +d z
Rather than the price dierentials, one notices that the Geometric Mean-
reversion with Time-varying Mean process denes the percentage change of
prices. The behavior of electricity spot prices is price-dependent: when prices
are low it behaves mildly, and when prices are high it is much volatile, therefore,
it is probably that the percentage change of prices, rather than the dierentials
of prices, behaves in a systematic manner.
Geometric Mean-reversion with Time-varying Mean process, therefore, is
a natural option for modeling electricity spot prices. Among the authors who
have adventured into modeling electricity spot prices with mean-reversion pro-
cesses, only a few early works explored using the Arithmetic Mean-reversion
Process with Time-varying Mean process, see Bhanot (2000); Escribano et al.
(2002); Knittel & Roberts (2005); Lucia & Schwartz (2002), most works have
turned to the Geometric Mean-reversion with Time-varying Mean process, e.g.,
Barlow et al. (2004); Benth et al. (2007); Bierbrauer et al. (2004, 2007); Blanco
& Soronow (2001b); Borovkova & Permana (2006); Branger et al. (2009);
Cartea & Figueroa (2005); Clewlow et al. (2001); Culot et al. (2006); De Jong
(2006); Diko et al. (2006); Geman & Roncoroni (2006); Hambly et al. (2009);
37
2. A SURVEY OF ELECTRICITY PRICE MODELS
Huisman & Mahieu (2003); Kosater & Mosler (2006); Mari (2008); Meyer-
Brandis & Tankov (2008); Mount et al. (2006); Nomikos & Soldatos (2008);
Schmidt (2008); Seifert & Uhrig-Homburg (2007); Skantze & Ilic (2001); Vil-
laplana (2004).
In literature, another line of research advocates using GARCH models to
model the time-varying volatility of electricity spot prices. The discussions on
the nature of the GARCH models and its possible applications in electricity
price modeling could be found in Section 2.2.11.
2.3.5 Modeling the Multi Risk-factors
Electricity spot prices is driven by various fundamental physical forces that
play in dierent timescales: look at hourly prices during a week, the prices
is driven by the intra-day and weekday-weekend variations of electricity load,
generation dispatches, and generation forced outages; look at prices during
a year, it is driven by seasonal weather and temperature, seasonally avail-
able generation capacity, and seasonal hydro-electric generation capacity; and
within the perspective of many years, the prices is driven by the long-term
forces such as economic development and economic cycles, generation invest-
ment and retirement, etc. These three types of forces that are physically
independent and play in dierent timescales, which will be referred as three
risk-factors, at best, ought to be dened respectively by separate models.
These independent physical forces that play in dierent timescales are not
unique in electricity prices. They exist in commodity and energy markets in
general: commodity prices have a long-term equilibrium level that is deter-
mined by the marginal cost of production in the long run, and the deviations
from this long-term equilibrium level are caused by short-term disequilibrium
between market supply and demand. For modeling both the variations of the
long-term equilibrium-level and the short-term deviations in commodity prices,
a 2-factor model has been proposed by Schwartz & Smith (2000). Multi-factor
models thereafter have been brought to the energy eld by Clewlow & Strick-
land (2000), applied to model the futures prices of natural gas by Manoliu &
38
2.3 Financial/Stochastic Process Models
Tompaidis (2002), and used to capture electricity futures prices by Diko et al.
(2006); Koekebakker & Ollmar (2005).
In the literature of modeling electricity spot prices with nancial models,
most authors used single-factor models, only a few works suggested the idea
of multi-factor models: Lucia & Schwartz (2002); Pilipovic (1998); Skantze
& Ilic (2001) proposed 2-factor models, and Culot et al. (2006); Kresen &
Husby (2000) explored the 3-factor models. Overall, the number of literatures
on modeling electricity prices by multi-factor models is very limited. The
multi-factor models, for that it captures the nature of electricity spot prices
particularly well, deserve much more attention and further eort. The mod-
eling methodology proposed in this thesis work belongs to the category of
multi-factor models, refer to Chapter 5 and Chapter 6
One particular diculty facing the multi-factor models is how to calibrate
the each factor model. The standard method is using Kalman-Filter to esti-
mate the parameters from both the futures prices and electricity spot prices.
Not to mention the diculty in obtaining the futures price data and the cred-
ibility of these data, whether the method is reliable or not is worth further
investigation. A practical study on using Kalman-Filter to calibrate a 2-factor
model could be found in Barlow et al. (2004). In this thesis work, a method of
data decomposition based on Fourier Analysis has been proposed to facilitate
calibrating each factor model, see Chapter 5 and Chapter 6.
2.3.6 Modeling the Price Spikes
Electricity price spikes of various magnitude are frequently encountered in
electricity markets all over the world. The causes for these price spikes are
complex. On the supply side, the supply stack of electricity is hockey-stick
like, which consists of a at part and a steep part, and the reason for the
supply stack being hockey-stick like is due to the inherent structure of gener-
ation systems and so far the inability for large-scale storage of electricity. On
the demand side, at this early stage of electricity markets, consumers dont
response actively to the price signal, thus the price elasticity of the electricity
demand is almost zero, namely, the demand curve of electricity is almost a
39
2. A SURVEY OF ELECTRICITY PRICE MODELS
vertical straight line. When the demand for electricity is high, and particu-
larly when the market is relatively tight in its supply(generation capacity), the
supply and demand frequently intersect at the steep part of the supply stack,
that is when price spikes occur. Other causes for price spikes may be conges-
tions of transmission networks, market power exercised by the market players,
insensibly designed market rules that invite market manipulations, etc. A few
literatures have provided careful discussions on the causes for price spikes, see,
e.g., Geman & Roncoroni (2006); Joskow & Kohn (2002).
Among the pioneers in modeling price spikes are Johnson & Barz (1999);
Kaminski (1997), they rst brought the Geometric Brownian Motion Jump
Diusion process, which has been widely used in nancial markets, to model
electricity prices. Geometric Brownian Motion Jump diusion process is not
appropriate for modeling electricity spit prices, for that it doesnt capture the
mean-reverting nature of electricity spot prices. Another model which has
gained limited success in modeling electricity prices with spikes is Geometric
Mean-reversion Jump Diusion process, with combines a Geometric Mean-
reversion process with a Poisson Jump process, see, e.g., Blanco & Soronow
(2001b); Cartea & Figueroa (2005); Clewlow et al. (2001); Villaplana (2004).
The disadvantage of the Geometric Mean-reversion Jump Diusion process is
that it relies on the reverting component of the mean-reversion process to pull
the price spikes back to the equilibrium level, and because electricity price
spikes usually decay very fast, this pulling-back eect turns out too slow.
Two methods have been proposed to remedy this shortcoming of the Ge-
ometric Mean-reversion Jump Diusion process, one is to add another down-
ward jump after the initial upward jump, see Deng (2000); another remedy
is to rely on a regime-switch model of three regimes to pull the initial jump
down, see Huisman & Mahieu (2003). Later on it was recognized that these
two methods does not allow the jumps to last for a second consecutive day,
while electricity price spikes sometimes do last a few consecutive days, then
decay fast to normal levels. To allow the price spikes to last a few days, the
regime-switch model of three regimes is simplied into a two-regime model,
see Bierbrauer et al. (2004); De Jong (2006). And later on the regime-switch
models have been extended to include further details: authors later observed
40
2.4 Structural Models
that the intensity of price spikes is time-varying, that is, during the high de-
mand seasons price spikes are more frequent than during the low demand sea-
sons, they therefore extended the regime-switch models by making the transi-
tion probability between regimes as functions of available generation capacity,
demand, weather, etc., see Cartea et al. (2008); Huisman (2008); Kosater &
Mosler (2006); Mount et al. (2006).
One emerging trend for modeling electricity price spikes is to model the
spikes as an independent stochastic process. Namely, the jumps are modeled by
another stochastic process that is decoupled from the mean-reversion process,
see Culot et al. (2006); Hambly et al. (2009); Nomikos & Soldatos (2008).
Among the works that model the jump with a decoupled jump process, one
stream of works models the jump process by a mean-reversion process that is
driven by non-Gaussian noises, particularly the Levy-process, refer to Benth
et al. (2007); Meyer-Brandis & Tankov (2008).
Another diculty in modeling electricity price spikes is how to dierentiate
between price spikes and normal prices, the pioneering works in this direction
are Bystrom (2005); Chan et al. (2008); Truck et al. (2007).
2.4 Structural Models
So far the last few sections have discussed the Time Series Models and the
Financial/Stochastic Process Models. Theses models belongs to the category
of Statistical Models: namely one begins with the price data, nds systematic
patterns in the data, and then brings in appropriate mathematical tools to
dene these patterns. Statistical methods usually do not venture into the
structure of electricity markets that give rise to all these price data. Structural
Approach, on the other hand, in order to understand the price data, looks into
the more fundamental structure of electricity markets.
By Structural Approach here it is meant that, in order to model electricity
spot prices, one rst models its constituent components, i.e., available gener-
ation capacity, marginal fuel prices, electricity load demand, the generation
supply stack, etc., then one constructs these constituent models into a struc-
ture that gives electricity spot prices. At best, the resulting model of electricity
41
2. A SURVEY OF ELECTRICITY PRICE MODELS
spot prices ought to be analytical and mathematically elegant. In order to have
an analytical nal model, one necessarily has to design the constituent models
carefully, because only when the constituent models are simple and concise,
the nal price model could possibly have an analytical and elegant expres-
sion. These simple and concise constituent models could only be achieved
by disregarding many irrelevant factors and negligible details and properly
approximating the important physical forces.
In the literature, a few authors have attempted the Structural Approach.
One pioneering work is Skantze & Ilic (2001) who modeled the supply curve
with an exponential function and dened the fundamental driving forces from
both the supply and demand sides with 2-factor stochastic processes. The
idea of modeling the supply curve with an exponential function is further
investigated by Barlow (2002) studying the Alberta market in Canada. Later
an idea of modeling the supply stack by a cubic spline function is proposed
by Burger et al. (2004, 2007) modeling the hourly prices for the European
Energy Exchange. One work that focuses particularly on getting the price
spikes right is Kanamura & Ohashi (2007). Another work that focuses on
investigating the seasonal risk-premium is Cartea & Villaplana (2008). Rather
than modeling the generation supply stack, the work Boogert & Dupont (2008)
models the hourly prices as a non-parametric function of reserve margin. A
recent development is the work Howison & Coulon (2009) who emphasizes the
eect of fuel prices on the movements of bidding stacks. Along this line of
structural models, there are some novel ideas. The work Davison et al. (2002)
studying the PJM market groups the daily prices into spikes and non-spikes,
each price regime modeled by a normal distribution, and the switching between
the two regimes is governed by a function as the ratio electricity load over
generation capacity. This work is further extended by Anderson & Davison
(2008) to include a more detailed generation capacity model that considers
generation outage.
Structural models provide a deeper understanding on electricity prices by
shedding light on the more fundamental structure of electricity markets. By
making use the fundamental structure of electricity markets structural models
probably are capable of modeling and thus forecasting electricity prices of
42
2.5 The Eect of Fuel Prices
longer time horizon than statistical models do, and thus it provides a tool for
the mid-term or long-term applications such as pricing contracts, generation
scheduling, and even generation investment analysis. The diculty faced by
the structural models, on the other hand, is to maintain the simplicity of the
overall model while at the same time providing all these benets.
2.5 The Eect of Fuel Prices
For that fuel prices profoundly inuence electricity spot prices, and that this
profound eect of fuel prices on electricity spot prices havent, in the literature
by large, received enough attention that is proportional to its importance, we
will thus devote this whole section to discuss it.
The reason for the profound eect of fuel prices on electricity spot price is as
follows. Electricity spot prices, in a competitive market, constantly reect the
marginal cost of generating electricity, and that the marginal cost of generating
electricity is a function of the heat rate of the marginal generator and the prices
of the fuel that is burned by the marginal generator. Any change of the price
of fuel that is burned by the marginal generator causes the marginal cost of
supplying electricity in the market to change. Electricity spot prices, therefore,
are profoundly aected by the prices of the fuel that is burned by the marginal
generators. Discussions of further detail could be found in Section 3.5, and
Section 4.4.
The earliest statistical studies on the eect of fuel prices, particularly nat-
ural gas prices, on electricity spot prices are based on the Californian markets,
see Emery & Liu (2002); Kaminski (1997); Kosecki (1999). Later works on
the PJM and New-England market could be found in Rose (2007), and on
the European Energy Exchange see Janssen & Wobben (2009). A discussion
taking a Microeconomics approach could be found in Kosecki (1999).
Among the large number of literatures on modeling electricity spot prices
with Time Series/Financial models, only a few works have considered fuel
prices in their model. The earliest work in record is Deng (2000), which mod-
els the correlation between electricity spot prices and natural gas prices by a
correlation coecient. Another work that considers fuel prices is Pirrong &
43
2. A SURVEY OF ELECTRICITY PRICE MODELS
Jermakyan (2008). Two works concerning the European markets are Zach-
mann (2007) studying the British and German markets and Aid et al. (June
2009) on the French markets. One recent work Howison & Coulon (2009) on
the PJM and New-England markets investigates the eect of uctuations of
fuel price on the movements of bidding stacks.
To summarize, fuel prices, particularly natural gas prices, have profound
eect on electricity spot prices, which has been both proved theoretically and
observed empirically. In the literature, however, there are only very limited
number of works that have seriously taken this eect of fuel prices into account.
In this thesis work, an electricity spot price model that explicitly considers the
eect of fuel prices on electricity spot prices will be proposed.
2.6 Decomposition Techniques
2.6.1 Fourier Analysis
Fouriers Theorem, refer to Maor (1998), states that a continuous time function
f(t), t = [0, T], can be reconstructed by a linear combination of sine and cosine
functions of dierent frequencies,
f(t) =
a
0
2
+

m=1
(a
m
cos(
f
mt) +b
m
sin(
f
mt))
where
f
=
2
T
, and
f
is referred as fundamental frequency. These sine and
cosine functions are orthogonal to each other, the parameters, a
m
and b
m
, thus
can be calculated by Eulers Formulas,
a
0
=
2
T
T
_
0
f(t)dt a
m
=
2
T
T
_
0
f(t) cos (
f
mt)dt
and
b
m
=
2
T
T
_
0
f(t) sin (
f
mt)dt
Fourier Analysis provides a superior tool for analyzing electricity spot
prices. Electricity prices are driven by underlying physical forces from both the
44
2.6 Decomposition Techniques
demand side and the supply side. These forces are of various frequencies and
independent to each other. The forces of the lowest frequency are the cycles of
economic development and generation investment, they aect electricity prices
in a time-horizon of many years; the forces of the mid-frequency are the sea-
sonal weather and the annual generation maintenance, they aect electricity
prices for weeks but not beyond a year; and the forces of the highest frequency
are the intra-day and intra-week variations of electricity load, they inuence
electricity price for hours but their eects are conned within a week. Due
to these cyclic forces from both the demand and supply sides, electricity spot
prices, accordingly, have periodic components of various frequencies. Fouriers
Theorem, on the other hand, analyzes the function f(t) into sine and cosine
functions of various frequencies, and these sine and cosine functions play con-
stantly and spread on the whole time range of the function t = [0, T]. The
idea of Fourier Analysis matches well the nature of electricity spot prices.
One could therefore apply Fourier Analysis to decompose electricity spot
prices into a few components of dierent frequencies that are driven by dier-
ent and independent underlying forces. These price components, thereafter,
could be studied and modeled respectively. Few literatures, to the best of
the authors knowledge, have applied the Fourier Analysis to study electricity
spot prices, one work Pedregal & Trapero (2007) applies a dynamic harmonic
regression model to hourly electricity prices. In this thesis work, we will use
Fourier Analysis to facilitate building an electricity spot price model.
2.6.2 Wavelet Analysis
Fourier Analysis decomposes the function f(t) into a series of sine and cosine
functions that spread on the whole time range t = [0, T]. Lets move on to see
how Wavelet Analysis works. We will rst introduce the basic idea of Wavelet
Analysis, and then analogize it with music tones, and at last discuss how
Wavelet Analysis might be applied to electricity price analysis and modeling.
The discussion on Wavelet Analysis is based on Strang & Nguyen (1996).
The simplest wavelet, a Haar wavelet, is an impulse wave function dened
on the time period t = [0, 1], Figure 2.6,
45
2. A SURVEY OF ELECTRICITY PRICE MODELS
w(t) =
_

_
1 0 t <
1
2
1
1
2
t < 1
1
-1
1
t
( ) w t
Figure 2.6: A Haar Wavelet
Scale w(t) to cover the range t = [0, T], the function becomes w(
t
T
); then
squeeze the function by half to cover the range t = [0,
T
2
], the wavelet function
becomes w(
2
T
t); if one squeeze the function by half for j times, the function
becomes w(
2
j
T
t), and it covers the range t = [0,
T
2
j
]. If one shifts the function
w(
2
j
T
t) by one of its range, it becomes w(
2
j
T
t 1), and if one shifts the function
w(
2
j
T
t) by k times, it becomes w(
2
j
T
tk), k = 0, ..., 2
j
1, which will be denoted
as,
w
jk
(t)

= w(
2
j
T
t k) k = 0, ..., 2
j
1
Obtained by squeezing and shifting, these smaller wavelets w
jk
(t) are orthog-
onal to each other, that is,
T
_
0
w
jk
(t)w
JK
(t)dt = 0
As the wavelets w
jk
(t) are orthogonal to each other, they form a basis for the
function f(t), namely,
f(t) =

j,k
b
jk
w
jk
(t)
46
2.6 Decomposition Techniques
The coecients b
jk
could thus be obtained by integration,
T
_
0
f(t)w
jk
(t)dt = b
jk
T
_
0
(w
jk
(t))
2
dt
then b
jk
could be calculated by, b
jk
=
T

0
f(t)w
jk
(t)dt
T

0
(
w
jk
(t)
)
2
dt
.
Wavelet Analysis reconstructs the function f(t) by the linear combination
of the wavelet functions, f(t) =

j,k
b
jk
w
jk
(t). These wavelet functions are
orthogonal to each other, and they are of dierent frequencies, length, and
positions. The one of the lowest frequency, w(
t
T
), covers the longest range
t = [0, T], and the one of the highest frequency, w(
2
j
t
T
), covers the shortest
range t = [0,
T
2
j
]. In other words, the lower frequency wavelets last longer
time and the higher frequency wavelets last shorter time. This idea of Wavelet
Analysis corresponds closely to the design of a musical note. In a musical
composition, each note species the frequency and the position of a tone, and
usually a low frequency tone lasts longer, and a high frequency tone lasts
shorter, refer to Maor (1998). The short high-frequency tone corresponds to a
short wavelet, and the long low-frequency tone corresponds to a long wavelet.
In electricity spot prices, the components of various frequencies, driven
by electricity load, are continuously playing along time, the timing of the
components of dierent frequencies are thus not important. Electricity Price
Spikes, on the other hand, is not a consecutive phenomenon, price spikes have
a short life. What we concern the most, is the timing, the length, and the
magnitude of price spikes, these attributes seem to be captured well by a
high frequency, short life, high magnitude, time-positioned wavelet function.
There have been some empirical studies applying Wavelets Analysis to model
electricity prices, refer to Conejo et al. (2005a,b); Kim et al. (2002). A recent
work that applies the Wavelet Analysis to dierentiate between price spikes
and non-spike prices is Truck et al. (2007).
47
2. A SURVEY OF ELECTRICITY PRICE MODELS
2.6.3 Principal Component Analysis
This section rst introduces the basic idea of Principal Component Analysis,
and then discusses the possible applications of Principal Component Analy-
sis in electricity price analysis and modeling. The discussions on Principal
Component Analysis is based on Lay (1997).
Given m highly correlated random variables, x
1
, x
2
, . . . , x
m
, because these
variables are highly correlated, there are much abundant information. The
variables x
1
, x
2
, . . . , x
m
, could be written into a column vector
x =
_
x
1
x
2
... x
m

T
By linearly recombining the variables x
1
, x
2
, . . . , x
m
, say
y
i
= v
T
i
x
in which v
i
=
_
v
i1
v
i2
v
im

T
and v
i
= 1, one arrive at some new
random variables y
1
, y
2
, . . . , y
m
, these m variables are uncorrelated, among
which y
1
explains the most variations of the original variables, mathematically
it is that the variance VAR{ y
1
} is maximized, y
2
explains the second most
variations, and till y
m
explains the least variations. For the k
th
random vari-
able y
k
, if k is large, its variance VAR{ y
k
} is so little that it may have little
useful information. The variables beyond k, y
k+1
, y
k+2
, . . ., y
m
, thus could be
discarded. This is the purpose of principal component analysis: it is to reduce
the dimension of highly correlated random variables, from m to k, and it is to
identify the most signicant common variations, y
1
, y
2
, . . . , from the original
random variables x
1
, x
2
, . . . , x
m
.
The realized values of the variable x
i
is written as, x
i
(j) = a
ij
, j = 1, ..., n,
they are organized into a column vector X
i
, as
X
i
=
_
a
i1
a
i2
a
in

T
The values of the m variables is organized into a matrix X, which is,
X =
_

_
X
T
1
X
T
2
.
.
.
X
T
m
_

_
=
_

_
a
11
a
12
a
1n
a
21
a
22
a
2n
.
.
.
.
.
.
.
.
.
.
.
.
a
m1
a
m2
a
mn
_

_
48
2.6 Decomposition Techniques
The realized values of the variable y
i
is written as, y
i
(j) = b
ij
, j = 1, ..., n,
organized into a column vector Y
i
, as
Y
i
=
_
b
i1
b
i2
b
in

T
The values of the m new variables y
i
is written in a matrix Y , which is,
Y =
_

_
Y
T
1
Y
T
2
.
.
.
Y
T
m
_

_
=
_

_
b
11
b
12
b
1n
b
21
b
22
b
2n
.
.
.
.
.
.
.
.
.
.
.
.
b
m1
b
m2
b
mn
_

_
a) To nd the rst vector v
1
, and thus the rst new random variables,
y
1
= v
T
1
x explains the most variations, namely VAR{ y
1
} is maximized.
In other words, in terms of the realized values of the random variables,
we are going to form the new vector Y
T
1
= v
T
1
X, and we want to nd
the optimal v
1
to maximize the function,
max
v
1
f (v
1
) = Y
T
1
Y
1
= v
T
1
XX
T
v
1
s.t. v
1
= 1
Since A

= XX
T
is a symmetric matrix, and usually of full rank m, it is
diagonalized by
A = PDP
1
in which P =
_
u
1
u
2
u
m

, and
D =
_

1
0
.
.
. 0
0
2
.
.
. 0

.
.
.
.
.
.
0 0
m
_

_
where u
i
is the unit eigenvectors corresponding to the eigenvalue
i
of
the matrix A. The eigenvalues are ordered as
1
>
2
> ... >
m
. The
optimization problem is thus transformed to
max
v
1
v
T
1
PDP
1
v
1
= v
T
1
PDP
T
v
1
=
_
P
T
v
1
_
T
D
_
P
T
v
1
_
s.t. v
1
= 1
49
2. A SURVEY OF ELECTRICITY PRICE MODELS
By changing variables, w = P
T
v
1
, the quadratic optimization is simpli-
ed as,
max
w
1
f(w
1
) = w
T
1
Dw
1
s.t. w
1
= 1
the optimal solution for this simple quadratic optimization problem is,
w

1
= e
1
, and the maximum of the function is f(w

1
) =
1
. The optimal
solution in terms of v
1
, thus is,
v

1
=
_
P
T
_
1
w

1
= Pw

1
= u
1
b) Now we move on to nd the second vector v
2
, and thus the second
random variable y
2
= v
T
2
x. y
1
and y
2
are uncorrelated, say E{ y
1
y
2
} = 0,
which is equivalent to Y
T
1
Y
2
= u
T
1
XX
T
v
2
= 0, and it is followed by
v
T
2
PDP
1
u
1
= 0, and thus v
T
2
u
1
= 0. The problem to nd the vector
v
2
is a quadratic optimization problem with the additional constraint
v
T
2
u
1
= 0,
max
v
2
f(v
2
) = v
T
2
PDP
1
v
2
s.t. v
2
= 1
u
T
1
v
2
= 0
The optimal solution, similarly by changing variables, is v

2
= u
2
, thus
the second new random variable y
2
= u
T
2
x. Similarly with the same
routine could be obtained the other vectors as v

i
= u
i
, and y
i
= u
T
i
x,
i = 3, ..., m.
In short, Principal Component Analysis transforms the highly correlated
random variables, x
1
, x
2
, . . . , x
m
to new uncorrelated random variables y
1
, y
2
,
. . . , y
m
, in which y
1
identies the most signicant common forces underlying
x
1
, x
2
, . . . , x
m
, y
2
the second most, and till y
m
the least.
Electricity spot prices are driven by a few physical underlying forces, such
as intra-day and intra-week variations of electricity load, temporal change of
weather and temperature, uctuations of fuel prices, seasonal variations of
load, etc. Lets take the electricity spot prices during a day for example, for
that they are driving by a few common underlying forces, these prices at dier-
ent hours are highly correlated with each other, thus one may apply Principle
50
2.7 Concluding Discussions
Component Analysis to identify and analyze these common underlying forces:
put it mathematically, if denote the prices at the rst hour p
1
(t), the second
hour p
2
(t), and till the last hour p
24
(t), these 24 variables, p
i
(t), i = 1, ..., 24,
are highly correlated, see, e.g., Guthrie & Videbeck (2007). With Principal
Component Analysis, one may nd a few random variables, q
1
(t), q
2
(t), . . . ,
which identify the common forces that underlie the hourly electricity spot
prices. Empirical studies on applying Principal Component Analysis to elec-
tricity prices are Burger et al. (2007); Diko et al. (2006); Skantze & Ilic (2001).
2.7 Concluding Discussions
If one surveys the landscape of the existing literatures on electricity price mod-
eling, Figure 1.2, in terms of the Statistical/Financial Approach, the space of
applying Time-Series Models to model short-term prices has been well occu-
pied; many authors have applied Financial Models to capture the electricity
spot prices within a time-horizon up to a year; and few researchers have applied
the Statistical/Financial Approach to model the long-term prices.
Most of the proposed Financial Models are single factor models, namely
it is designed to capture only one particular stochastic physical force that
underlies electricity prices. Electricity prices, however, by its nature, are driven
by a few dierent and independent physical underlying forces that play in
dierent timescales, and to model these multiple independent physical forces
a multi-factor model is probably more appropriate than a single-factor model.
Although this limitation of single-factor models have been recognized by a few
researchers, the number of published works on modeling electricity prices by
multi-factor models so far is very limited.
Prices of fuel such as coal and natural gas have signicant eect on electric-
ity spot prices. This signicant eect of fuel prices on electricity spot prices has
been recognized by many researchers, among whom a few have incorporates
the eect of fuel prices in their models. However, in the literature by large,
this signicant eect of fuel prices on electricity spot prices has not received
enough attention that is proportional to its importance. This lack of attention
from literature is probable due to the limitations of the mathematical tools,
51
2. A SURVEY OF ELECTRICITY PRICE MODELS
namely, the statistical models such as Time-Series/Financial Models allow lit-
tle extension to consider the eect of fuel prices. On the other hand, in the
structural approach, the eect of fuel prices might be included relatively more
easily.
The Structural Approach of modeling electricity spot prices, compared with
the large number of publications on the Statistical Approach, have only re-
ceived very limited attention. Structural Models incorporates the more fun-
damental structure of electricity markets, by which it provides a deeper un-
derstanding on electricity spot prices. The researchers in this direction have
achieved various degrees of success. The number of published works in this
direction, overall, is limited, and there is still much space inviting further ex-
ploration.
Various decomposition techniques, such as Fourier Analysis, Wavelet Anal-
ysis, and Principal Component Analysis, etc., have been applied to solve some
particular problems in electricity spot prices modeling. The system of electric-
ity spot prices is complex, if one could somehow use the decomposition tech-
niques to separate the components of electricity spot prices that are driven by
dierent physical underlying forces, it will not only help to understand better
the nature of electricity spot prices, but also tremendously simplify the original
price modeling problem. So far only a very limited number of published works
have applied the decomposition techniques to analyze and model electricity
spot prices.
52
Chapter 3
Microeconomics applied to the
Electricity Markets
This chapter applies Microeconomics Price Theory to electricity markets. Dis-
cussions on Price Theory in general could be found in Mankiw (2004); Samuel-
son (1980). We will, at this very beginning, postulate a generation company
that owns a few coal and gas generators. This generation company owns so
little generation capacity that it is not able to manipulate the spot prices in
the market. Given the spot prices of electricity in the market, at any single
hour, this generation company, in order to maximize its own prot, faces a
problem that how much should be its production, namely, how much elec-
tricity it should generate during that hour. Based on this micro-view on a
particular generation company, the outcomes concerning the whole market,
which consists of many generation companies, afterward, step by step, will be
derived.
53
3. MICROECONOMICS APPLIED TO THE ELECTRICITY
MARKETS
3.1 The Supply Curve of a Generation Com-
pany
3.1.1 Heat Rate Curve of the Generation Company
This generation company has a few coal and gas generators, the coal units are
assumed to be more ecient than the gas ones. The eciency of a generator is
measured by its heat rate(mmBtu/MWh), dened as the fuel (mmBtu) burned
for generating per (MWh) electricity. This heat rate is the average heat rate of
one generation unit. These generators, ordered according to their heat rates,
forms a climbing-up stairs, Figure 3.1a. This climbing-up stairs will be referred
as a heat rate curve. The coal units serve the base load and thus form the lower
part of the heat rate curve, and the gas units serve the peak load and thus
compose the upper part of the heat rate curve. The whole heat rate curve will
be modeled by a linear segment, Figure 3.1b.
HR(q) = a +b q 0 q C
where C is the overall generation capacity of the company.
Quantity
HR
Coal units Gas units
K
(a)
Q
HR
K
(b)
Figure 3.1: The Heat-rate Curve of the Generation Company
3.1.2 Fuel Price Curve of the Generation Company
During one hour, if this generation company is generating q MWh electric-
ity, the price of its marginal fuel is a function of the quantity q, denoted as
54
3.1 The Supply Curve of a Generation Company
P(q), Figure 3.2, in which marginal fuel is the fuel burned by the marginal
generator, and a marginal generator is the last generator that is dispatched to
supply electricity in this generation company, (usually this marginal generator
is the most expensive one among all the generators that are dispatched at this
moment in this generation company). The fuel price curve, dened as the
marginal fuel price P(q) plotted against the generating quantity q, 0 q C,
consists of two horizontal segments: the lower-left segment identies the fuel
price of coal, and the upper-right segment represents the fuel price of natural
gas. The two segments meet at the threshold K where the coal generation
capacity encounters the natural gas generation capacity.

Q
Fuel
Price
K
Figure 3.2: The Fuel Price Curve of the Generation Company
3.1.3 Marginal Cost Curve of the Generation Company
The fuel cost for generating the q
th
MWh electricity is the price of marginal
fuel P(q) US$/mmBtu times the heat rate HR(q) mmBtu/MWh, this fuel cost
is the marginal cost MC(q) of generating the total q MWh electricity, namely,
MC(q) = HR(q) P(q)
The marginal cost curve, which is the marginal cost MC(q) plotted against
the quantity q, 0 q C, by denition, is the product of the hear rate curve
HR(q) and the fuel price curve P(q). The marginal cost curve thus is a non-
continuous piecewise curve, consisting of one at segment and the other steep
segment, Figure 3.3. The at segment is composed of coal units, and the steep
one the natural gas units. Because coal is cheap and natural gas usually is
55
3. MICROECONOMICS APPLIED TO THE ELECTRICITY
MARKETS
relatively more expensive, the marginal cost of the coal units increases slowly,
and that of the natural gas units climbs rapidly.

Q
MC
K
Figure 3.3: The Marginal Cost Curve of the Generation Company
3.1.4 Supply Curve of the Generation Company
For this generation company, the overall fuel cost to generate the q (MWh)
electricity, at one hour, thus is
FuelCost(q) =
q
_
0
MC(x)dx
The total cost of generating electricity, includes fuel cost and x cost. Fix cost,
for this hour, includes the operation and maintenance cost, wage of employ-
ees, management compensation, interests, et al. This x cost, by large, does
not change with the production quantity q, namely that the x cost is not a
function of production q. Total cost for generating the q (MWh) electricity,
denoted as TC(q), therefore, is,
TC(q) = FixCost +FuelCost(q) = FixCost +
q
_
0
MC(x)dx
Given the market price of electricity at this hour p, and if the company is
generating q (MWh), the revenue R(q) of the company, during this hour, is
R(q) = p q
56
3.1 The Supply Curve of a Generation Company
The prot of the company during this hour, denoted as (q), is the revenue
R(q) subtracted by total cost TC(q), thus is
(q) = R(q) TC(q) = p q
_
_
FixCost +
q
_
0
MC(x)dx
_
_
Given the market price p, the company faces a decision that, in order to
maximize the prot (q) at this hour, how much should be the output q. The
problem, written as an optimization problem, is,
max
q
(q) = p q
_
_
FixCost +
q
_
0
MC(x)dx
_
_
s.t. q C
To maximize the prot (q), the company will generate the quantity q

, such
that the marginal cost of generating the q

(MWh) electricity equals to the


market price p, say, p = MC(q

). The marginal cost curve, by this way,


denes the supply curve of this generation company, Figure 3.4: given one
market price p, the company is going to generate, according to its marginal
cost curve, the quantity
q

= MC
1
(p)
Q
*
q

p
Spot
Price

Figure 3.4: The Supply Curve of the Generation Company
57
3. MICROECONOMICS APPLIED TO THE ELECTRICITY
MARKETS
3.2 The Supply Curve of an Electricity Mar-
ket
An electricity markets consists of many generation companies. Each generation
company has its own supply curve. The aggregation of the supply curves of
these generation companies composes the supply curve of the whole electricity
market. Lets take a simple market of two generation companies for example.
Company A has a marginal cost curve MC
1
(q), and company B has a marginal
cost curve MC
2
(q), given the market price p at one hour, company A, in
order to maximize its own prot, generates the quantity q
1
= MC
1
1
(p) ,
and company B, also to maximize its own prot, generates the quantity q
2
=
MC
1
2
(p), thus, during this hour, overall the market supplies the quantity
q = q
1
+q
2
= MC
1
1
(p) +MC
1
2
(p)

= MC
1
(p)
by this way, Figure 3.5, the supply curves of the company A and company B
constitute the supply curve of the market, as,
q = MC
1
(p)
Spot Price

p
1
q
Q
2
q

Spot Price

Q 1 2
q q +
Spot Price

p
Q
Figure 3.5: The Supply Curve of an Electricity Market
3.3 The Demand Curve of an Electricity Mar-
ket
At this early stage of electricity markets, during the time as short as one hour,
the responses from the demand side to the market prices are very limited. The
58
3.4 The Spot Price of an Electricity Market
market price of electricity by large does not aect the amount of electricity
load demand. Namely, the price elasticity of electricity load demand is little.
Mathematically, given any price p, the quantity of demand always equals to
the quantity of electricity load L. Thus, the demand curve of the market is a
vertical straight line, Figure 3.6, as,
q = L

Q
Spot Price

L
Figure 3.6: The Demand Curve of an Electricity Market
3.4 The Spot Price of an Electricity Market
Synthesizing the supply curve and the demand curve of the market, namely by
intersecting the supply and demand curves, one obtains the market spot price:
the supply curve is
q = MC
1
(p)
and the demand curve is
q = L
Thus the intersection of the two curves settles the market spot price S, Fig-
ure 3.7,
L = MC
1
(S)
Namely,
S = MC(L)
59
3. MICROECONOMICS APPLIED TO THE ELECTRICITY
MARKETS

Q
Spot Price

S
L
Figure 3.7: The Spot Price of an Electricity Market
3.5 The Eect of Fuel Prices on Spot Prices
Electricity spot prices are settled as the intersections of the market marginal
cost curve and the load demand, as S = MC(L). Recall that the marginal
cost curve is the multiplication of the heat rate curve and the fuel price curve,
as
S = MC(L) = HR(L) P(L)
Fluctuations of fuel prices drive the fuel price curve P(L) to shift, and there-
after cause the electricity spot prices to change. Lets look at the following two
cases for illustration: if the fuel price of natural gas, in the rst place, is P
1
,
load and the marginal cost curve settle the market price S
1
, Figure 3.8a; if the
fuel price of natural gas then increases to P
2
, the second segment of the fuel
price curve rises, consequently, the steep segment of the marginal cost curve
rises and becomes steeper, the marginal cost curve and the load demand then
settle a higher market spot price S
2
, Figure 3.8b.
3.6 The Eect of Generator Outage on Spot
Prices
Due to equipment failures, some generators might be forced to decrease their
output or even shut down. Lets see if one coal generator of very large ca-
pacity serving the base load is forced to withdraw from operation, what will
60
3.6 The Eect of Generator Outage on Spot Prices
Q
Heat Rate
K Q K
Fuel Price Spot Price
P
1
S
1

Q L

(a)
Q
Fuel Price
Spot Price
Q
Heat Rate
K
P
2

S
2

S
1

Q L

(b)
Figure 3.8: The Eect of Fuel Prices on Spot Prices
61
3. MICROECONOMICS APPLIED TO THE ELECTRICITY
MARKETS
happen for the marginal cost curve and thus the spot price. Simply, the with-
drawn generation capacity causes the marginal cost curve of the market to
shift leftward, the spot price will thus increase from S to S

, see Figure 3.9.



Q
Spot Price

S
L
S

Figure 3.9: The Eect of Generator Outage on Spot Prices
3.7 Intra-day Electricity Spot Prices
During a day between night-time and day-time electricity load varies a lot.
In the midnight the load is the lowest, in the early morning the load climbs
up rapidly, at noon it falls to a gentle valley, in the afternoon and the early
evening it rises to the highest, and in the late night it declines fast, Figure 3.10.
This typical pattern of intra-day load varies slightly its shape across seasons
and markets.
Time
Load
Night Day Night

Figure 3.10: Intraday Electricity Load Demand
Electricity spot price is a function of electricity load, as S = MC(L), Fig-
ure 3.11. When electricity load varies across the day, it causes the uctuations
62
3.8 Time-varying Volatility of Intra-day Spot Prices
Time
Spot Price

Time
Night
Day
Night
K Q
Spot Price

Figure 3.11: Intraday Electricity Spot Prices
of electricity spot prices. When load is low during the night, load intersects
the marginal cost curve at the at segment, and because the slope is little,
the prices rises slowly when load increases; when during the daytime load is
high, load intersects the marginal cost curve at the steep segment, thus prices
are much sensitive to the increase of load, so spot prices rise rapidly. Overall,
the intra-day pattern of electricity load, through the marginal cost curve, is
transformed to an intra-day pattern of electricity spot prices: namely, the load
pattern is rst cut at the threshold K, then the part below the threshold K is
squeezed down, and the part that is larger than the threshold K is pulled up.
3.8 Time-varying Volatility of Intra-day Spot
Prices
Electricity load during a day has a highly predictable pattern, it falls low in
the night-time, and climbs high in the daytime. In previous discussions, we
implicitly assumed that load is deterministic, however, electricity load is al-
ways inherent with a nature of uncertainty. Caused by temporal variations of
weather and temperature, electricity load may swing widely around its pre-
dicted mean. The shadowed area around the load pattern in Figure 3.12 indi-
63
3. MICROECONOMICS APPLIED TO THE ELECTRICITY
MARKETS
cates and denotes the uncertainty of intra-day electricity load.
Q
Spot Price

Time
Night
Day
Night
Time
Spot Price

Figure 3.12: The Time-varying Volatility of Intra-day Electricity Spot Prices
Electricity spot prices are the intersections of the marginal cost curve and
the load demand. The marginal cost curve, by its nature, is piecewise and non-
continuous. The at segment of the marginal cost curve squeezes the variations
of o-peak load; and the steep segment of the marginal cost curve amplies
the variations of on-peak load, Figure 3.12. Due to the squeezing eect, the
uncertainty of the o-peak prices is little; while due to the amplifying eect,
the uncertainty of on-peak prices is large. It implies that o-peak prices are
mild, and on-peak prices are volatile; namely, the volatility of the o-peak
prices is small, and the volatility of the on-peak prices is large.
3.9 Electricity Spot Price Spikes
In previous discussions, it has been implicitly assumed that in an electricity
market there are only large capacity generator that are driven by steam tur-
bines. These steam turbines are often fueled either by nuclear, coal or by
natural gas, and they are of large capacity and of high eciency. The heat
rate, (which is inversely proportional to eciency), of these large capacity
generators thus is low.
64
3.9 Electricity Spot Price Spikes
In an actual electricity market, there are generators of much smaller size
which are driven by combustion turbines, and these combustion turbines are
of small capacity and of much lower eciency. The heat rate of these small
capacity generators are thus much higher than the large capacity ones, and
these combustion turbines are usually fueled by more expensive natural gas
or diesel oil; the marginal cost of these small capacity generators, therefore,
is much higher. These small capacity units, therefore, form the third segment
of the marginal cost curve in an electricity market, and this third segment is
much steeper than the other two segments, Figure 3.13.

MC
K
1
K
2
Q

Figure 3.13: The Marginal Cost Curve with an additional Third Segment
At the time when the electricity load demand, relatively to the overall
generation capacity in the market, is high, and when the load exceeds the
threshold K
2
, due to the large slope of the third segment, the spot prices rise
rapidly, Figure 3.14; after a few hours, when the load decreases, spot prices
decline back to the normal level. The spot prices, by this way, spike during
the high load demand hours. Because the high load demand usually lasts only
for a few hours, the spike has a short life.
65
3. MICROECONOMICS APPLIED TO THE ELECTRICITY
MARKETS
Time
Q

Time
Spot
Price
Spot
Price
Figure 3.14: Intra-day Electricity Spot Price Spikes
66
Chapter 4
A Multi-granularity View of
Electricity Prices
The last chapter applies the Microeconomics Price Theory to electricity mar-
kets, it introduces how the supply and demand of electricity markets are
formed, and how the electricity spot prices are settled. In order to have a
deeper understanding on the behavior of electricity spot prices, this chapter
applies the analytical tools that were learned from the last chapter to look
at electricity spot prices as a function of time. Since electricity spot prices is
driven by a few independent physical underlying forces that play in dierent
timescales, thus, in the next few sections, electricity prices will be examined
carefully within three time-perspectives of dierent granularity, that is, intra-
week hourly prices, intra-year weekly prices, and multi-year yearly prices.
4.1 Intra-week Hourly Electricity Spot Prices
In most electricity markets, electricity spot prices are settled hour by hour.
Lets rst look at the hourly electricity spot prices during a week. Electricity
spot prices are settled as the intersections of the demand and supply of an
electricity market, in which the demand is the system electricity load and the
supply is composed of a large number of generators, see Figure 4.1:
a) System electricity load, from Monday to Sunday, has two cyclic patterns,
67
4. A MULTI-GRANULARITY VIEW OF ELECTRICITY
PRICES
Supply
Demand

Price
Time(Hour)
T W T F S S M
Figure 4.1: Intra-week Hourly Electricity Spot Prices
one intra-day pattern and the other weekday-weekend pattern: during
the day-time electricity load is high, and during the night-time load
is low; during the weekdays electricity load is higher, and during the
weekends load is lower.
b) During one week, the supply of the electricity market may change little,
for that in the short time of one week the overall available generation
capacity in the market varies little. A few generators, due to equip-
ment failures, might be forced to outage; the outage of these generators,
overall, probably does not aect the total available generation capacity
much, except when some of these generators in forced outage are of large
capacity.
c) Electricity spot prices, as the intersections of the supply and demand in
the market, naturally show a pattern similar to that of the system load
demand: high during the day-time, low during the night-time, higher
during the weekdays, and lower during the weekends.
d) The volatility of spot prices, on the other hand, due to the convexity of
the supply curve, is price-dependent, that is, high prices are accompanied
with high volatility and low prices are accompanied with low volatility;
68
4.2 Intra-year Weekly Prices
namely, during the day-time, prices are high and volatile, during the
night-time, prices are low and mild, and spot prices during the weekdays
are higher and more volatile than these during the weekends.
4.2 Intra-year Weekly Prices
Zooming out from the perspective of one week, now move on to look at a
longer time horizon of one year. In order to study clearly the prices in this one
year perspective, one has to compensate the details, for that too much detail
distracts our limited attention. The uctuations of hourly spot prices during a
week now in this perspective of one year is irrelevant, thus by taking the weekly
average of hourly spot prices, hourly prices are squeezed into weekly prices.
In this perspective of weekly prices during one year, spot prices are driven by
seasonal forces from both the supply and the demand sides, see Figure 4.2:

Supply
Demand
Price
Time(Week)
Figure 4.2: Intra-year Weekly Prices
a) On the demand side, it is the seasonal electricity load. Driven by seasonal
weather and temperature, weekly electricity load uctuates around a
highly predictable seasonal pattern. This seasonal pattern varies across
markets because seasonal weather depends on geography: e.g., in PJM
69
4. A MULTI-GRANULARITY VIEW OF ELECTRICITY
PRICES
market, it is cold in winter, hot in summer, and mild in spring and fall,
therefore, the weekly electricity load is high in winter and summer, and
low in spring and fall; in Nordic Pool, it is cold in winter, and temperate
in summer, thus the weekly electricity load is high in winter, low in
summer, and temperate in spirng and fall.
b) On the supply side, the amount of available generation capacity varies
across a year. Most generators require annual maintenance, which usu-
ally takes one to two months. The maintenance is usually scheduled
during low demand seasons. The overall available generation capacity
thus falls to the lowest point during the months when the load demand
is the lowest. Another source of seasonal variations of generation capac-
ity is from hydro-electric power plants, the generation capacity of hydro
plants depends on the water inow and the water level of the reservoirs,
and the water inow and water level of the reservoirs are seasonal due
to seasonal rain fall and snow melt.
c) Electricity spot prices, as the intersections of seasonal demand and sup-
ply of the market, naturally have a seasonal pattern. Electricity spot
prices during the high demand seasons are high; during the low demand
seasons, even though the demand reaches its annually lowest valley, due
to at the same time the reduction of available generation capacity caused
by planned annual generation maintenance, electricity spot prices may
reach its secondary annual peaks, which are usually of smaller magni-
tude than the ones in high demand seasons; during spring and fall, when
weather is temperate, electricity load demand is relatively low, and avail-
able generation capacity is plenty, electricity spot prices usually reach its
annually lowest levels. (Empirical evidences on this seasonal pattern
could be found in Figure 5.11 and Figure 6.14.)
d) In terms of the volatility of electricity prices, during the extreme sea-
sons(summer and winter), the prices are usually volatile, and during the
temperate seasons(spring and fall), the prices often behaves mildly. This
70
4.3 Multi-year Yearly Prices
is for that, in the extreme seasons, when the supply(relative to the de-
mand)is tight, the demand more often intersects the supply at the steep
part of the supply curve, as a result, any variations from both the sup-
ply and demand sides are amplied by the steep supply curve, therefore,
the electricity spot prices are much volatile; on the other hand, during
the temperate seasons(spring and fall), the supply relative to demand is
plenty, the demand more often intersects the supply at the at part of
the supply curve, as a result, any variations from both the supply and
demand sides are squeezed and thus reected little in electricity spot
prices.
e) Overall, electricity spot prices during the extreme seasons(summer and
winter) are usually high and volatile, and during the temperate sea-
sons(spring and fall) are usually low and mild. In other words, in the
one year perspective, the volatility of weekly electricity prices is propor-
tional to the level of prices, namely, the volatility of electricity spot prices
are price-dependent. As a result, the volatility of electricity spot prices
during a year also has a highly predictable seasonal pattern.
4.3 Multi-year Yearly Prices
Now zoom out further from the one year perspective to look at a time horizon
that is even longer, many years or even decades. In order to have a clear view
in this long-term perspective, the irrelevant details of seasonal variations will
be ignored. Namely, by taking the yearly average of electricity spot prices,
the time unit of prices is changed to yearly, and we are going to look at the
yearly electricity spot prices across many years, see Figure 4.3. Yearly prices is
driven by long-term forces that play in a grand scales, these forces determine
the trend of electricity spot prices in the long run:
a) On the demand side, in the long run of decades, driven by economic de-
velopment, electricity load by large has an upward long-term trend; and
in the relatively short run of a few years, due economic cycles(the econ-
omy sometimes is hyper-active, and sometimes falls to a deep recession),
71
4. A MULTI-GRANULARITY VIEW OF ELECTRICITY
PRICES

Heat-rate of marginal generators
Time(Year)
Supply
Demand
Figure 4.3: Multi-year Yearly Prices
the electricity load demand often deviates from this long-term upward
trend, namely, it varies slowly along the long-term upward trend.
b) On the supply side, in the long run generation investment by large fol-
lows the upward trend of electricity load demand, meanwhile new gen-
eration plants are being built and old generators are scheduled to retire.
The total amount of installed generation capacity, by large, matches the
increasing electricity load demand. In the short-run of a few years, gen-
eration investment also has its cycle of boom and burst, that is, in some
years, too many generation plants might be built and thus the installed
generation capacity are relatively abundant, while in some other years
too few generation capacity is added and the market may face a shortage
of generation supply.
c) Look at the supply and demand together, the ratio of the electricity load
over the amount of generation capacity, by large, is kept stable, though
it could vary year by year due to the interactions between cycles eco-
nomic development and generation investment. This stable ratio implies
that the types of marginal generatorsa marginal generator is the last
generator on the marginal cost curve that is being dispatched to meet
72
4.4 The Eect of Fuel Prices on Yearly Prices
the load demandkeep the same, e.g., coal and natural gas units.
d) This stable ratio further implies that the heat rate of marginal generators,
in the long run, by large, changes little, and this heat rate overall may
decrease, even though slowly, because more ecient new generators are
going to replace the less ecient old ones. Due to cycles of economic
development and generation investment, in the short run of a few years,
the overall level of the heat rate of marginal generators may somehow
deviate from and varies slowly along its long-term trend.
4.4 The Eect of Fuel Prices on Yearly Prices
In previous discussions, the eect of fuel prices on electricity spot prices has
been purposely avoided, for that fuel prices aect electricity spot prices so
signicantly that it is too distracting to consider it in earlier discussions. This
section will give it a careful treatment to which it is well deserved. Even though
fuel prices may aect electricity spot prices in a time-horizon that is shorter
than a year, we will discuss it in the perspective of multi-year yearly prices.
The principles learned from the following discussions are directly applicable to
fuel-price eects in any other shorter timescale.
Coal and Natural Gas are the dominant marginal fuels in most electricity
markets, for which the percentage share of time being at the margin may dier
across markets. In some markets, diesel oil serves as the marginal fuel during
the peak hours, and due to its relatively high prices, its percentage share of
time being at the margin overall is very limited. In terms of prices, the prices
of coal are mostly stable, natural gas prices are usually more volatile, and oil
prices are the most volatile. Fossil fuels such as coal, natural gas, and oil are
exhaustible natural resources, in the long run, due to the increasing demand
and the limited supply, the prices of fossil fuels will probably continue to climb,
unless the mankind nds an eective substitute for them, Figure 4.4.
Electricity spot prices, in an ideal competitive market, constantly reect the
marginal cost of generating electricity. The marginal cost is a direct function of
the fuel prices and the heat rate of marginal generators; namely, the marginal
73
4. A MULTI-GRANULARITY VIEW OF ELECTRICITY
PRICES
Time(Year)
Fuel Prices
Figure 4.4: The Prices of Fossil Fuels across Many Years

Long-run heat-rate of marginal generators
Time(Year)
Time(Year)
Fuel Prices
Time(Year)
Price

Figure 4.5: The Yearly Prices Considering Fuel Price Fluctuations


74
4.5 A Multi-granularity View of Electricity Prices
cost of generating electricity is the multiplication of the prices of marginal
fuels and the heat rate of marginal generators, refer to Section 3.5. For that
the average heat rate of marginal generators in the long run is mostly stable,
and that the prices of the marginal fuels have an upward trend, electricity
spot prices, which are the multiplication of the heat rate and the marginal fuel
prices, will necessarily have a long-term upward trend that is similar to that
of the marginal fuel prices, see Figure 4.5.
4.5 A Multi-granularity View of Electricity Prices
After studying the intra-week hourly spot prices, intra-year weekly prices, and
multi-year yearly prices, it is time to synthesize them into one big picture,
Figure 4.6:


Yearly Prices
Hourly Spot Prices
Weekly Prices
Figure 4.6: A Multi-granularity View of Electricity Prices
a) Look at the prices with the time-perspective of many years, it is the
yearly prices that lay the most signicant trend of electricity prices in
the long run of many years and further decades;
b) zoom in from the yearly prices to have the perspective of one year, it is
the weekly prices that demonstrate the seasonal variations of electricity
prices during a year;
75
4. A MULTI-GRANULARITY VIEW OF ELECTRICITY
PRICES
c) nally if zoom in further to look at the perspective of one week, it is
the hourly spot prices that illustrate the micro-structure of intra-day
and weekday-weekend hourly variations of electricity spot prices during
a week.
d) These yearly prices, weekly prices, and hourly prices, by this way, driven
by dierent and independent physical underlying forces, construct the
hourly electricity spot prices across years and decades.
76
Chapter 5
A Multi-granularity Electricity
Spot Price Model
5.1 Abstract
In the deregulated electric power industry, under the market environment,
electricity spot prices are highly volatile and uncertain. Private generation
companies, to make decisions on generation investment and scheduling, con-
cern future electricity spot prices. This work builds an electricity spot price
model that is suitable for generation investment and scheduling analysis. It
adopts a modeling methodology Divide and Conquer to build the price model;
the overall price model explicitly considers the various physical forces that un-
derlie electricity spot prices; the price model covers a time horizon of many
years and has a time unit of one hour; and the nal result of the model is
mathematically simple and elegant. Besides its applications in generation in-
vestment and scheduling analysis, the model could be widely used in pricing
and trading electricity options, futures, and other electricity derivatives.
5.2 Introduction
In the deregulated electric power industry, private generation companies, in
order to sell to the electricity markets, make their own decisions on building
77
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
new power plants and scheduling the production. To make these decisions, the
generators concern the prot that is to be earned and the risk of not earning
enough prot. The prot of a generator in a competitive electricity market
is primarily a function of electricity spot prices and the cost of generating
electricity, see Wu et al. (2008).
Electricity spot prices are usually settled hour by hour as the intersections
of the supply and demand of an electricity market. One day ahead of the
moment of delivering electricity, generators submit their bids, which specify
the quantity and price they are willing to sell, to the system operator; the
system operator receives the bids from the generators, stacks up the bids from
low to high, and forms an aggregate supply stack of the market; then the
system operator, according to the supply stack, decides the order of dispatching
generators; the bid of the marginal generator, which is the last generator that
is needed to generate electricity in order to meet the system load demand
forecast, sets the spot price of the market for that particular delivery moment.
As the supply and demand of an electricity market are constantly changing,
electricity spot prices are all the time uctuating. During a day, driven by ev-
eryday business activities, electricity spot prices are higher during the day-time
and lower in the night-time; during a week, driven by the weekday-weekend
variations of electricity load demand, electricity prices are higher during week-
days and lower on weekends; during one year, electricity prices, driven by the
seasonal electricity load demand, are high and volatile in high demand sea-
sons, and usually low and milder in low demand seasons; across years, due to
economy cycles, electricity prices are high when the economy is active and low
when the economy falls to recessions. Furthermore, in a competitive electricity
market, prices of marginal fuels have profound eect on electricity spot prices,
for that electricity spot prices constantly reect the marginal cost of generat-
ing electricity, and the marginal cost of a generator is a direct function of its
fuel prices, refer to Section 3.5.
Generation short-term scheduling concerns the future generator prot of
days and months, and generation investment planning concerns the future
generator prot of years and decades, an electricity spot price model for gen-
eration investment and scheduling, therefore, ought to capture the dynamics
78
5.2 Introduction
of electricity spot prices in a future time horizon of many years. In terms of
the time unit of the model, as between day-time and night-time electricity spot
prices change a lot, thus particularly for valuating the prot of peaking gener-
ators, an hourly based model is necessary. Furthermore, during one week, one
year, and across years, electricity spot prices are driven by dierent and inde-
pendent fundamental physical forces from both the supply and demand sides,
thus this electricity spot price model ought to capture these physical forces
properly. Besides satisfying all these requirements, the model ought to have a
nal result that is mathematically simple, for that electricity spot prices is the
input variable for generator prot valuation and further the generation invest-
ment and scheduling analysis. Ideally, the nal result of the model ought to
characterize the prices at each hour with a simple probability density function.
Researchers have modeled electricity prices of various time horizons for
dierent purposes: short-term next-day or next-week prices for formulating
bidding strategies, trading, and short-term generation scheduling; mid-term
next-month or next-year prices for pricing electricity derivatives and gener-
ation scheduling; long-term prices of many years for generation investment
planning. For modeling the short-term prices, time series methods are popu-
lar, see Conejo et al. (2005a); Contreras et al. (2003); Crespo Cuaresma et al.
(2004); Garcia et al. (2005); Nogales et al. (2002); Pedregal & Trapero (2007);
Weron (2006), as well as a few works on the structural approach, see Boogert
& Dupont (2008); Burger et al. (2007); for modeling the mid-term prices,
works on time series methods see Escribano et al. (2002); Knittel & Roberts
(2005), and on the structural approach are Anderson & Davison (2008); Barlow
(2002); Kanamura & Ohashi (2007); relatively few works have modeled long-
term prices, Hamm & Borison (2006) proposes a time series model, Cartea
& Villaplana (2008); Skantze & Ilic (2001) take the structural approach, and
Olsina et al. (2006) models the long-term system dynamics.
This work takes the Statistical Approach to model long-term electricity spot
prices: it begins with historical data, then nds orders in data, and nally de-
nes the orders with models. And this work adopted a modeling methodology
Divide and Conquer, based on the conviction that dierent physical forces that
underlie electricity spot prices ought to be modeled respectively. To build the
79
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
model, throughout this chapter, the New-England electricity market will be
taken for example: rst the historical electricity spot price data from the New-
England market are collected; then the data are decomposed into four compo-
nents that are driving by a few dierent and independent physical underlying
forces, which are a fuel-price component, a high-frequency component driven
by short-term physical forces, a mid-frequency component driven by mid-term
forces, and a low-frequency component driven by long-term forces; afterward
these four components are dened respectively by separate models; nally the
resulting four sub-models are aggregated to produce an overall electricity spot
price model. The proposed price model belongs to the category of Multi-factor
Models, which have been explored by Culot et al. (2006); Diko et al. (2006);
Koekebakker & Ollmar (2005); Kresen & Husby (2000); Lucia & Schwartz
(2002); Pilipovic (1998); Skantze & Ilic (2001).
This work has beneted extensively from the existing literatures. The
mathematical tools employed to build the models are learned from the es-
tablished eld of nancial engineering, refer to Dixit & Pindyck (1994); Hull
(2006), and the early adventurers who brought the nancial theories to electric-
ity markets, see Deng (1999); Johnson & Barz (1999); Kaminski (1997); Lucia
& Schwartz (2002); Pilipovic (1998). The understanding on the fundamental
physical forces that underlie electricity prices is learned from Skantze & Ilic
(2001); the eect of fuel prices on electricity spot prices learned from Howison
& Coulon (2009); Kosecki (1999); and the method of decomposing the data
by Fourier Analysis inspired by an excellent discussion on Fourier Theorem
by Maor (1998).
The rest of this chapter is organized as follows: Section 5.3 introduces the
New-England electricity market; Section 5.4 decomposes the electricity spot
price data into four components; Section 5.5 builds the mathematical model
for each price component; Section 5.6 calibrates the models by matching the
models with historical data; Section 5.7 evaluates the price model by con-
structing a sample path of hourly electricity spot prices throughout a year;
Section 5.8 applies price simulations to estimate the generator prot for a coal
unit and a gas unit; Section 5.9 discusses the limitations, extensions, and other
applications of the model; Section 5.10 summarizes the chapter and concludes.
80
5.3 The New-England Market
5.3 The New-England Market
New-England electricity market locates in northeastern United States. The
market adopts a Standard Market Design, and includes a day-ahead energy
market and a real-time energy market. From 2005 to 2009, the market has in-
stalled generation capacity around 30 GW, its system load ranges from 10 GW
to 25 GW, and annual generation planned outage during a year could reach
as high as 8 GW, see Figure 5.1. The system load data are collected from ISO
New England Inc. (2009b), and the generation capacity and planned outage
data are from ISO New England Inc. (2009a). Natural-gas fueled generators
play a signicant role in the New-England market: in 2007, natural-gas fueled
generators make up 50% of the installed capacity, and natural-gas fueled gen-
erators are the marginal generators setting the market prices in 74% of the
time in 2007, see (ISO New England Inc., 2007, P32 & P.42).
03 04 05 06 07 08 09 10
0
5
10
15
20
25
30
35
Time (hour)
G
W


installed capacity
systemload
planned outage
Figure 5.1: New-England Market Fundamentals - (top) seasonally
claimed generation capacity, (middle) electricity load, and (bottom) generation
planned outage
The historical electricity spot price data are from the New-England day-
81
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
ahead energy market, which are the hourly Locational Marginal Pricing data at
the Internal Hub, see Figure 5.2a. The price data are obtained from ISO New
England Inc. (2009b)). The electricity price data, to facilitate the forthcoming
analysis, will be denoted as S(t).
Natural gas prices in the New-England market region were also collected,
which are the monthly Massachusetts natural gas prices sold to the electric
power consumers, see Figure 5.2b. The data were collected from Energy In-
formation Administration (2009a). The natural gas price data will be denoted
as P(t).
03 04 05 06 07 08 09 10
0
100
200
300
400
500
600
Time(hour)
U
S
$
/
M
W
h
(a) New-England Market Day-ahead
Electricity Spot Prices
03 04 05 06 07 08 09 10
0
2
4
6
8
10
12
14
16
18
20
Time (hour)
U
S
$
/
m
m
B
t
u


(b) Massachusetts Natural Gas Prices
Figure 5.2: Electricity Prices and Natural Gas Prices
5.4 Decompose the Data to Find Orders
5.4.1 Fuel-Price Eect on Electricity Spot Prices
Given the historical electricity spot price data S(t), take the natural logarithm
on them, denoted as
s(t)

= ln S(t)
82
5.4 Decompose the Data to Find Orders
and plotted them in Figure 5.3a the upper graph(a). For the natural gas price
data P(t), also took the natural logarithm on them, denoted as
p(t)

= ln P(t)
and plotted in Figure 5.3a the lower graph(b).
04 05 06 07 08 09
3
4
5
6
7
a:
lo
g

U
S
$
/
M
W
h


04 05 06 07 08 09
1.5
2
2.5
3
b:
Time (hour)
lo
g

m
m
B
t
u
/
M
W
h


Electricity
Natural Gas
(a) Logarithm Prices - a: electricity
prices, b: natural gas prices
04 05 06 07 08 09
3.8
4
4.2
4.4
a:
lo
g

U
S
$
/
M
W
h


04 05 06 07 08 09
1.8
2
2.2
2.4
b:
Time (hour)
lo
g

U
S
$
/
m
m
B
t
u


Natural Gas
Electricity
(b) Low-frequency Components - a:
electricity prices, b: natural gas prices
Figure 5.3: Fuel-price Eect on Electricity Spot Prices
For that natural gas is the marginal fuel in the New-England market most
of the time, natural gas prices have signicant eect on the electricity spot
prices. To look at clearly this eect of natural gas prices on electricity spot
prices, the Logarithm Electricity Spot Prices s(t) and the Logarithm Natural
Gas Prices p(t) are ltered by a low-pass lter, which is a central moving
average of length one year(8760hrs), to get rid of the seasonal eect and other
irrelevant eects of higher frequency, to leave only the long-term dynamics.
The long-term dynamics of the electricity spot prices are plotted in Figure 5.3b
the upper graph(a), and the long-term dynamics of the natural gas prices are
plotted in Figure 5.3b the lower graph(b). Observing the two curves, the close
match of the electricity price curve and the natural gas price curve evidences
the profound eect of natural gas prices on electricity spot prices.
83
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
5.4.2 Fuel-price Eect Adjustment
To adjust this eect of natural gas prices on electricity spot prices, the loga-
rithm natural gas prices p(t) were simply subtracted from the logarithm elec-
tricity spot prices s(t), and obtain the fuel-price adjusted electricity spot prices
in the logarithm domain, which will be referred as, according to its physical
meaning, the Logarithm Implied Marginal Generator Heat Rate Data,
s

(t)

= s(t) p(t)
which are plotted in Figure 5.4a.
04 05 06 07 08 09
0.5
1
1.5
2
2.5
3
3.5
4
Time (hour)
H
e
a
t
r
a
t
e

(
lo
g

m
m
B
t
u
/
M
W
h
)
(a) Time Domain, s

(t)
0 1/168 2/168
0
500
1000
1500
Frequency (cycle per hour)
M
a
g
n
it
u
d
e
a:
0 1/8760 2/8760 3/8760 4/8760
0
500
1000
1500
Frequency (cycle per hour)
M
a
g
n
it
u
d
e
b:
(b) Frequency Domain, F()
Figure 5.4: Implied Marginal Generator Heat Rate Data - logarithm
hourly electricity spot price data after being adjusted the eect of natural gas
prices
The reason for adjusting the fuel-price eect on the electricity spot prices
this way is because that, in a competitive electricity market, spot prices con-
stantly reects the marginal cost of the marginal generators, and the marginal
cost of the marginal generator is mostly its heat rate, denoted as S

(t), times
its fuel price, denoted as P(t), namely,
S(t) = S

(t) P(t)
Take logarithm on both sides, it becomes
s(t) = s

(t) +p(t)
84
5.4 Decompose the Data to Find Orders
in which s

(t)

= ln S

(t), and move p(t) to the left, one obtains the implied
heat rate of the marginal generators that is in the logarithm domain,
s

(t)

= s(t) p(t)
5.4.3 Frequency Spectrum of the Data
The implied marginal generator heat rate data s

(t) could be further decom-


posed into three components that are driven by dierent and independent
physical forces. The process of decomposition will be demonstrated in the fre-
quency domain, take Fourier Transform on the marginal generator heat rate
data s

(t) to have the frequency spectrum F(),


F() =
T
_
0
s

(t)e
it
dt
which is plotted in Figure 5.4b.
Look at the frequency spectrum around the frequency
w
= 2
1
168
(radian
per hour), [because one week is 168hrs], the frequency spectrum spikes at the
frequencies = 2
1
168
, 2
2
168
, 2
3
168
and so on. It implies that, in the
heat rate data, cyclic components of periods one week, half a week, and so on,
are signicant, and these cyclic components are the periodic harmonics that
construct the intra-day and weekday-weekend patterns of the heat rate data.
Zoom in to look at the lower frequency range around
y
= 2
1
8760
(radian
per hour), [because one year is 8760hrs], the frequency spectrum spikes at the
frequencies = 2
1
8760
and 2
4
8760
. It implies that, in the heat rate data,
cyclic components of periods one year and a quarter of a year are signicant,
and these cyclic components are the periodic harmonics that construct the
seasonal pattern of the heat rate data.
5.4.3.1 The High-frequency Component
The high-frequency component is driven by short-term physical forces from
both the supply and demand sides, that is, intra-day and weekday-weekend
variations of electricity load, regular generation dispatches, and generation
85
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
forced outages. These short-term physical forces play in a short timescale
and have negligible eects even beyond a week. These forces are cyclic; they
repeat themselves day after day and week after week. Therefore, these short-
term forces are periodic and of periods equal or less than a week(168hrs), and
the high-frequency component thus locates at the frequency band
w
, in
which
w
= 2
1
168
.
The high-frequency component is thus obtained by applying a high-pass
lter H
h
() on the original frequency spectrum F(),
F
h
() = F() H
h
()
which is plotted in Figure 5.5b, and where the high-pass lter is specied as,
H
h
() =
_
0 <
w
1
w
The details on designing the lter H
h
(), and the procedures of ltering please
refer to Appendix D.
2 Feb 04 9 Feb 04
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
Time (hour)
H
e
a
t
r
a
t
e

(
lo
g

m
m
B
t
u
/
M
W
h
)
... ...
(a) Time Domain, s

h
(t)
0 1/168 2/168
0
500
1000
1500
Frequency (cycle per hour)
M
a
g
n
it
u
d
e
(b) Frequency Domain, F
h
()
Figure 5.5: The High-frequency Component Data - obtained by applying
the high-pass lter H
h
()
To have the high-frequency component in the time-domain, take Inverse
Fourier Transform on the frequency spectrum F
h
(),
s

h
(t) =
1
2

F
h
()e
it
d
86
5.4 Decompose the Data to Find Orders
which is plotted in Figure 5.5a, and two sample weeks from early 2004 are
highlighted. Observing the data: the intra-day and weekday-weekend patterns
are apparent, seasonal and long term trends are imperceptible.
5.4.3.2 The Mid-frequency Component
The mid-frequency component is driven by seasonal forces, that is, seasonal
electricity load demand due to seasonal temperature and weather, seasonal
available generation capacity due to annual generation maintenance, seasonal
hydro-electric generation capacity, etc. The seasonal forces by large aect
electricity spot prices in a timescale of a few weeks and have negligible eect
beyond a year. These fundamental forces are periodic and repeat themselves
year after year, namely, they are of periods equal or less than a year; on the
other hand, these forces, during a week, change little, thus they must be of
periods much longer than a week. The mid-frequency component, therefore,
locates at the frequency range
y
<
w
, in which
y
= 2
1
8760
, and

w
= 2
1
168
.
The mid-frequency component is thus obtained by applying a mid-pass
lter H
w
() on the original frequency spectrum F(),
F
w
() = F() H
w
()
which is plotted in Figure 5.6b, and where the mid-pass lter is,
H
w
() =
_
1
y
<
w
0 otherwise
To have the mid-frequency component in the time-domain, take Inverse
Fourier Transform on the frequency spectrum F
w
(),
s

w
(t) =
1
2

F
w
()e
it
d
which is plotted in Figure 5.6a, and two sample years 2005 and 2006 are
highlighted. Observing the data: the seasonal pattern is apparent, and the
values in summer are higher than these in the other seasons; the data overall
are volatile, and this volatility is due to the volatile electricity load, which is
further driven by temporal changes of weather and temperature.
87
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
05 06 07
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
Time (hour)
H
e
a
t
r
a
t
e

(
lo
g

m
m
B
t
u
/
M
W
h
)
...
(a) Time Domain, s

w
(t)
0 1/168 2/168
0
500
1000
1500
Frequency (cycle per hour)
M
a
g
n
it
u
d
e
(b) Frequency Domain, F
w
()
Figure 5.6: The Mid-frequency Component Data - obtained by applying
the mid-pass lter H
w
()
5.4.3.3 The Low-frequency Component
The low-frequency component is driven by long-term physical forces, that is,
slow variations of electricity load due to economic cycles, and changes of in-
stalled generation capacity due to generation investment and generation retire-
ment. These physical forces play in a time horizon of many years and change
little during a year, and thus they must be of periods much longer than a year,
the low-frequency component thus belong to the frequency range <
y
, in
which
y
= 2
1
8760
.
The low-frequency component is thus obtained by applying a low-pass lter
H
y
() on the original frequency spectrum F(),
F
y
() = F() H
y
()
which is plotted in Figure 5.7b, and where the low pass lter is,
H
y
() =
_
1 <
y
0
y
To have the low-frequency component in the time-domain, take Inverse
88
5.5 Dene the Orders by Models
04 05 06 07 08 09
1.5
1.6
1.7
1.8
1.9
2
2.1
2.2
2.3
2.4
2.5
Time (hour)
H
e
a
t
r
a
t
e

(
lo
g

m
m
B
t
u
/
M
W
h
)

(a) Time Domain, s

y
(t)
0 1/8760 2/8760 3/8760 4/8760 5/8760
0
500
1000
1500
Frequency(cycle per hour)
M
a
g
n
it
u
d
e
(b) Frequency Domain, F
y
()
Figure 5.7: The Low-frequency Component Data - obtained by applying
the low-pass lter H
y
()
Fourier Transform on the frequency spectrum F
y
(),
s

y
(t) =
1
2

F
y
()e
it
d
which is plotted in Figure 5.7a. Observing the data: the low-frequency com-
ponent changes slowly around an almost constant level, and this constant level
is determined by the heat rate of marginal generators in the long run.
5.5 Dene the Orders by Models
In the next few sections, we will move on to dene each price component by a
mathematical model.
5.5.1 The Intra-week Model
The high-frequency component data s

h
(t) are driven by short-term physical
underlying forces during a week, they will be dened by an Intra-week Model,
denoted as s

h
(t). Note that the tilde dierentiates the model from the
data.
89
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
The intra-week model s

h
(t) is of time unit one hour, and it has two com-
ponents: a deterministic function and a stochastic process,
s

h
(t) = f
h
(t) + x
hr
(t)
This deterministic function f
h
(t) is represented by a periodic function con-
sisting of a linear combination of sine and cosine functions of basic period
168hrs,
f
h
(t) =
0
+
84

m=1

m
cos
_
2
168
mt
_
+
83

m=1

m
sin
_
2
168
mt
_
The base period 168hrs is because one week has 168 hours. The deterministic
function f
h
(t) models the intra-day and weekday-weekend patterns of the high-
frequency data.
The stochastic process x
hr
(t) is modeled by a mean-reversion process,
d x
hr
= k
h
x
hr
dt +
h
d z
h
where d z
h
=
h
(t)

dt, and
h
(t) is a white noise of standard Normal distribu-
tion
h
(t) N(0, 1) and of no temporal correlation, E {
h
(t)
h
(t +dt)} = 0.
The mean-reversion process represents the stochastic component of the high-
frequency data that varies around the deterministic intra-day and weekday-
weekend patterns.
The solution of the intra-week model s

h
(t), given its initial value at time
0, s

h0
, at any future time t, refer to Appendix C.4, is
s

h
(t) = f
h
(t) + [s

h0
f
h
(0)] e
k
h
t
+


2
h
2k
h
(1 e
2k
h
t
)
h
(t)
The solution is a Normal distribution of mean value f
h
(t) +[ s

h0
f
h
(0)] e
k
h
t
and standard deviation
_

2
h
2k
h
(1 e
2k
h
t
).
5.5.2 The Intra-year Model
The mid-frequency data are driven by mid-term seasonal forces during a year,
they will be captured by an Intra-year Model, denoted as s

w
(t).
90
5.5 Dene the Orders by Models
The intra-year model s

w
(t) is of time unit one week, and it again has two
components: a deterministic function and a stochastic process,
s

w
(t) = f
w
(t) + x
wr
(t)
The deterministic function f
w
(t) is represented by a periodic function con-
sisting of a linear combination of sine and cosine functions of basic period 52
weeks and of harmonics up to the 4
th
order,
f
w
(t) = a
0
+

m=1,2,3,4
_
a
m
cos(
2
52
mt) +b
m
sin(
2
52
mt)
_
The base period of 52 weeks is because one year has 52 weeks, and the har-
monics up to the 4
th
order is due to the spike in the frequency spectrum at
the frequency = 2
4
8760
, see Figure 5.4b. The deterministic function f
w
(t)
models the seasonal pattern of the mid-frequency data.
The stochastic process x
wr
(t) is modeled by a mean-reversion process,
d x
wr
= k
w
x
wr
dt +
w
d z
w
It represents the stochastic component of the mid-frequency data that vibrates
around the deterministic seasonal pattern.
The solution of the intra-year model s

w
(t), given its initial value at time 0,
s

w0
, at any future time t, is
s

w
(t) = f
w
(t) + [s

w0
f
w
(0)] e
k
w
t
+


2
w
2k
w
(1 e
2k
w
t
)
w
(t)
which is again a Normal distribution.
5.5.3 The Multi-year Model
The low-frequency data are driven by long-term forces that play in a timescale
of many years, they will be characterized by a Multi-year Model, denoted as
s

y
(t).
The multi-year model s

y
(t) is designed on the time unit one month, and it
has two components: a deterministic function and a stochastic process
s

y
(t) = f
y
(t) + x
yr
(t)
91
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
The deterministic function f
y
(t) is simply represented by a linear trend,
f
y
(t) = a +b t
which models the long-term trend of the marginal generator heat rate.
The stochastic process x
yr
(t) is modeled by a mean-reversion process,
d x
yr
= k
y
x
yr
dt +
y
d z
y
It represents the slow variations of marginal generator heat rate around its
long-term trend.
The solution of the multi-year model s

y
(t), at any future time t, given its
initial value at time 0, s

y0
, is
s

y
(t) = f
y
(t) +
_
s

y0
f
y
(0)

e
k
y
t
+


2
y
2k
y
(1 e
2k
y
t
)
y
(t)
which is also a Normal distribution.
Here we would like to discuss briey on designing the multi-year model:
the multi-year model captures the long-term dynamics of the marginal gen-
erator heat rate, and this long-term dynamics is driven by grand forces like
economic cycles, and cycles of generation investment and retirement. Consid-
ering the complex nature of these fundamental forces, this multi-year model,
which incorporates only a long-term trend and a mean-reversion process, is a
simplied model; and it might only be valid when there are many historical
data and used for projecting a very long time-horizon such as a few decades.
If one wants to use the multi-year model for forecasting the time-horizon of
a few years, one probably should replace it with a more sophisticated econo-
metric model that incorporates the forecast information such as load growth,
generation investment and retirement, etc. in order to have a more accurate
forecast.
5.5.4 The Fuel Price Model
The fuel price model, denoted as p(t), for capturing dynamics of the monthly
natural gas prices, is left undened, because it is out of the scope of this work.
In the rest of this work, wherever the fuel price model p(t) is needed, the
historical monthly natural gas price data p(t) is used instead.
92
5.5 Dene the Orders by Models
5.5.5 The Overall Model
So far the natural gas prices, the high-frequency component, the mid-frequency
component, and the low-frequency component have been dened respectively
by a fuel price model p(t), an intra-week model s

h
(t), an intra-year model
s

w
(t), and a multi-year model s

y
(t). Now lets put the four models together
and view them in a big picture, see Figure 5.8.


Hours
Years
Weeks
I: ( )
y
s t
II: ( )
w
s t
III: ( )
h
s t
Figure 5.8: The Overall Electricity Spot Price Model - In the logarithm
domain, the yearly model s

y
(t) lays the big skeleton, on which the weekly model
s

w
(t) hangs the intra-year seasonal structure, and then the hourly model s

h
(t)
adds the intra-week variations
The overall electricity spot price model is the summation of the four sub-
models: rst the multi-year model s

y
(t) lays the long-term variations across
many years; then the intra-year model s

w
(t) hangs the seasonal structure on the
93
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
multi-year model; further the intra-week model s

h
(t) adds the hourly variations
on the intra-year model; and nally the fuel price model p(t) is constantly
adjusting the electricity spot prices.
s(t) = s

y
(t) + s

w
(t) + s

h
(t) + p(t)
Take exponential on the additive model s(t), it becomes an multiplicative
model

S(t),

S(t) = e
s

y
(t)
e
s

w
(t)
e
s

h
(t)
e
p(t)

=

S

y
(t)

S

w
(t)

S

h
(t)

P(t)
Namely, the overall electricity spot price model

S(t) is a multiplication of four
sub-models, refer to Appendix C.5.
5.6 Calibrate the Models
In the next few sections, we will use the data to calibrate the models.
5.6.1 The Intra-week Model Calibration
The intra-week model s

h
(t) was dened in continuous time, to calibrate the
model, it has to be written in discrete time, for that the historical data from
which the parameters is going to be estimated are in the discrete time of hourly
base.
The intra-week model s

h
(t) is the summation of a deterministic and a mean-
reversion process, s

h
(t) = f
h
(t) + x
hr
(t), in which the mean-reversion process
x
hr
(t), in discrete time, is an Autoregressive Model of order One, AR(1), refer
to Appendix E,
x
hr
(t + t) = e
k
h
t
x
hr
(t) +
_

2
h
2k
(1 e
2k
h
t
)
h
(t)
The discrete time version of the intra-week model s

h
(t), therefore, could be
derived as,
s

h
(t+t) = e
k
h
t
s

h
(t)+
_
f
h
(t + t) e
k
h
t
f
h
(t)


2
h
2k
h
(1 e
2k
h
t
)
h
(t)
(5.1)
94
5.6 Calibrate the Models
which is an autoregressive model with a time-varying deterministic mean func-
tion. Denoting,

= e
k
h
t
and


2
h
2k
h
(1 e
2k
h
t
)
the discrete time intra-week model is simplied as,
s

h
(t + t) =
h
s

h
(t) + [f
h
(t + t)
h
f
h
(t)] +

h

h
(t)
The parameters
h
and

h
, and the deterministic function f
h
(t) will be
estimated from the high-frequency component data using a non-linear least
square algorithm, refer to Lucia & Schwartz (2002). With the estimated pa-
rameters
h
and

h
, the parameters of the mean-reversion process k
h
and
h
will be calculated by,
k
h
t = ln
h
and

h
=

2k
h
(1 e
2k
h
t
)

h
Model Season k
h

h
x
hr
(t)
winter 0.1776 0.0667
summer 0.1484 0.0562
spring 0.2590 0.0550
fall 0.2341 0.0703
Table 5.1: Parameters of the Intra-week Model
The high-frequency data are divided according to seasons: winter, spring,
summer, and fall; one set of parameters is estimated for each season. In Fig-
ure 5.9(a), the dashed curves plot the high-frequency data of a few mid-winter
weeks, superimposing each week, and the solid curve plots the estimated deter-
ministic function f
h
(t); Figure 5.9(b) plots the case of summer; Figure 5.9(c)
spring; and Figure 5.9(d) fall. Table 5.1 records the estimated parameters of
the mean-reversion process for each season.
95
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
M T W T F S S
-1
-0.5
0
0.5
1
a:
M T W T F S S
-1
-0.5
0
0.5
1
b:
M T W T F S S
-1
-0.5
0
0.5
c:
M T W T F S S
-1
-0.5
0
0.5
1
d:
Figure 5.9: The High-frequency Component and the Intra-week Pat-
tern - The high-frequency component data s

h
(t) of four mid-season weeks su-
perimposed one by one, together with the intra-week deterministic pattern f
h
(t):
(a) winter, (b) summer, (c) spring, (d) fall
96
5.6 Calibrate the Models
5.6.2 The Intra-year Model Calibration
The intra-year model s

w
(t), in discrete time, similarly is an Autoregressive
Model with a time-varying deterministic mean function,
s

w
(t+t) = e
k
w
t
s

w
(t)+
_
f
w
(t + t) e
k
w
t
f
w
(t)


2
w
2k
w
(1 e
2k
w
t
)
w
(t)
(5.2)
The parameters k
w
and
w
, and the deterministic function f
w
(t) will be es-
timated from the mid-frequency data. The mid-frequency data originally are of
time unit one hour, the time-unit is changed to one week by taking the weekly
average of the hourly data, and then the parameters of the intra-year model
is estimated from the weekly data. In Figure 5.10, the blue solid curves plot
the mid-frequency data by superimposing each year, and the red dashed curve
plots the estimated deterministic seasonal pattern f
w
(t). Table 5.2 records the
estimated parameters of the intra-year model.
Model k
w

w
d x
wr
= k
w
x
wr
dt +
w
d z
w
1.1281 0.1607
Table 5.2: Parameters of the Intra-year Model
The seasonal pattern of the mid-frequency data f
w
(t) has four peaks, one for
each season, and the reason for these four peaks is due to the seasonal pattern of
electricity load and annual generation planned outage. In Figure 5.11, graph(a)
re-plots the seasonal pattern of marginal generator heat rate f
w
(t), graph(b)
plots the seasonal pattern of electricity load, and graph(c) the seasonal pattern
of generation planned outage. The seasonal pattern of electricity load has two
peaks one in winter and the other in summer; the winter peak is due to heating,
and the summer peak due to cooling. These two peaks of the seasonal pattern
of electricity load corresponds to the two peaks of marginal generator heat rate
during winter and summer. The seasonal pattern of the generation planned
outage has two peaks one in spring and the other in fall, for that generation
annual maintenance is mostly scheduled in spring and fall when electricity
97
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
Time (hour)
H
e
a
t
r
a
t
e

(
l
o
g

m
m
B
t
u
/
M
W
h
)
Figure 5.10: The Mid-frequency Component and the Intra-year Pat-
tern - the mid-frequency component data from 2005 to 2007 superimposing year
by year, together with the intra-year pattern f
w
(t)
98
5.6 Calibrate the Models
J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
-0.1
0
0.1
l
o
g

m
m
B
t
u
/
M
W
h
a:
J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
-2
0
2
G
W
b:
J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
0
5
10
G
W
c:
Figure 5.11: Seasonality of the New-England Market - (a) the seasonal
patterns of marginal generator heat rate, (b) system electricity load, and (c)
generation planned outage
99
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
load is low. These two peaks of generation planned outage correspond to the
other two peaks of marginal generator heat rate in spring and fall. Finally,
the resulting four peaks of the seasonal pattern of marginal generator heat
rate data explain the spike in the frequency spectrum F() at the frequency
= 2
4
8760
, see Figure 5.4b.
5.6.3 The Multi-year Model Calibration
The multi-year model s

y
(t), in discrete time, is again an Autoregressive Model
with a time-varying deterministic mean function,
s

y
(t+t) = e
k
y
t
s

y
(t)+
_
f
y
(t + t) e
k
y
t
f
y
(t)


2
y
2k
y
(1 e
2k
y
t
)
y
(t)
(5.3)
The parameters k
y
and
y
, and the deterministic function f
y
(t) will be
estimated from the low-frequency data. The low-frequency data originally are
of time unit one hour, the time unit is changed to one month by taking the
monthly average on the original data, and the parameters of the multi-year
model is then estimated from the monthly data. In terms of the deterministic
function f
y
(t), only the constant level a is estimated and the trend term b is not
estimated, because the long-term trend is not explicit in the data of only a few
years. Figure 5.12 evaluates the calibrated multi-year model by comparing the
actual data with a counterpart simulation, and Table 5.3 records the estimated
parameters of the multi-year model.
Model a b k
y

y
s

y
(t) 2.049 N/A 0.0945 0.0071
Table 5.3: Parameters of the Multi-year Model
100
5.6 Calibrate the Models
04 05 06 07 08
2
2.01
2.02
2.03
2.04
2.05
2.06
2.07
2.08
2.09
2.1
Time (month)
H
e
a
t
r
a
t
e

(
l
o
g

m
m
B
t
u
/
M
W
h
)


actual
simulation
long-term trend
Figure 5.12: Calibration of the Multi-year Model
5.6.4 The Sampling of the Data and the Time-Unit of
Models
The data used for calibrating the models, either the high-frequency, mid-
frequency, or the low-frequency data, are originally of time-unit one hour.
The intra-week model were calibrated from the original data of time-unit one
hour; for the mid-frequency data, they were taken the weekly average and then
used to calibrate the intra-year model; and for the low-frequency data, they
were taken the monthly average and then used for calibrating the multi-year
model. One may ask that how did we decide the sampling of the data and the
time-unit of the models.
The frequency of sampling is determined by the frequency range of the
original data. The high-frequency data were driven by intra-week short-term
forces, these short-term forces vary considerably hour by hour, and after exper-
imenting the time-unit of one hour, we concluded that the original time-unit of
one hour is appropriate for the intra-week model; for the mid-frequency data
that are driven by mid-term seasonal forces, these forces change little during
101
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
one week while vary considerably week by week, thus we took the weekly aver-
age of the data and chose the time-unit one week for the intra-year model; for
the low-frequency data that are driven by long-term forces, the time-unit of
one-month were decided after a few experiments: originally the time-unit was
set as one year, later we found that there exits signicant intra-year variations
and the time-unit of one year is too rough for the low-frequency data, and thus
nally set the time-unit of the multi-year model as one-month.
5.7 Simulate Electricity Spot Prices with the
Models
The electricity spot price model

S(t) is now complete: the structure of the
model has been specied and the parameters of the sub-models have been
estimated. In order to evaluate the overall model, a sample path of hourly
electricity spot prices throughout the whole year of 2005 will be constructed
step by step, and then the sample path will be compared with the actual
hourly electricity spot prices. The procedures are illustrated in Figure 5.13
and elaborated as follows:
a) A sample path of the low-frequency component is simulated: given the
initial value s

y0
, the subsequent values s

y
(t) are simulated according to
the discrete time multi-year model s

y
(t), Equ. 5.3. The low-frequency
component is constant during a month, changes across months, lasting
a whole year;
b) A sample path of the mid-frequency component is simulated: given the
initial value s

w0
, the subsequent values s

w
(t) are simulated according to
the discrete time intra-year model s

w
(t), Equ. 5.2. The mid-frequency
component is almost constant during a week, changes week by week,
crossing a whole year;
c) A sample path of the low-frequency component: given the initial value
s

h0
, the subsequent values s

h
(t) are simulated according to the discrete
102
5.7 Simulate Electricity Spot Prices with the Models
t
V:
( ) ( ) ( ) ( ) ( )
y w h
s t s t s t s t p t = + + +

IV:
( ) p t

III:
( )
h
s t

II:
( )
w
s t

I:
( )
y
s t



Figure 5.13: Construct the Electricity Spot Prices in the Year 2005 -
I. s

y
(t) low-frequency component; II. s

w
(t) mid-frequency component; III. s

h
(t)
high-frequency component; IV. p(t) natural gas prices; V. s(t) logarithm prices
time intra-week model s

h
(t), Equ. 5.1. The high-frequency component
lays the intra-week variations, hour by hour, crossing the whole year;
d) Take the historical Massachusetts natural gas price data of the year 2005,
p(t);
e) Finally, plus the four components together to have a sample path of
hourly electricity spot prices in the logarithm domain s(t).
s(t) = s

y
(t) +s

w
(t) +s

h
(t) +p(t)
f) Take exponential on the logarithm price s(t) to obtain the prices S(t),
S(t) = e
s(t)
One arrives at a sample path of hourly electricity spot prices through-
out the whole year 2005. In Figure 5.14 the lower graph(b) plots the
simulated sample path, and the upper graph(a) plots the actual prices.
103
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
0
50
100
150
200
250
U
S
$
/
M
W
h
a:


actual
J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
0
50
100
150
200
250
Time (hour)
U
S
$
/
M
W
h
b:


simulation
Figure 5.14: Day-ahead Electricity Spot Prices in the Year 2005 - the
upper graph plots the actual prices, and the lower graph plots a counterpart
simulation by the price model
104
5.7 Simulate Electricity Spot Prices with the Models
In Figure 5.15, the lower graph(b) compares a few simulated sample paths
with the actual price path in terms of the price histogram: the blue dashed
curves plot the histograms of a few simulated sample paths, and the red solid
curve the histogram of the actual price path; and the upper graph(a) compares
the simulated sample path with the actual price path in terms of the empirical
cumulative probability function. Table 5.4 records the statistics of the price
histograms.
0 50 100 150 200 250
0
0.2
0.4
0.6
0.8
1
C
D
F
a:


actual
simulation
0 50 100 150 200 250
0
0.05
0.1
0.15
0.2
Price (US$/MWh)
P
D
F
b:


actual
simulation
Figure 5.15: Empirical Probability Distributions of the Prices in Year
2005 - the upper graph plots the Cumulative Distribution Functions, and the
lower graph plots the price histograms.
Mean STD Skewness Kurtosis
Actual 78.69 26.13 0.78 3.16
Simulated Sample Path 79.62 29.15 0.94 3.65
Error 1.18% 11.56% 20.51% 15.51%
Table 5.4: Statistics of Empirical Probability Distributions
105
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
To further compare the simulated sample path with the actual price path,
lets zoom in from the one-year perspective to look at a few mid-season sam-
ple weeks. First the mid-winter week, which dates from 25 January 2005 to
31 January 2005, in Figure 5.16 the upper graph(a) plots the actual hourly
spot prices, the lower graph(b) plots a counterpart simulation, and the dashed
lines plot the expected mean value and delimit the 95% condence interval.
Similarly Figure 5.17 compares the case for the mid-spring week, Figure 5.18
the mid-summer week, and Figure 5.19 the mid-fall week.
24 J an 05 T W T F S S
50
100
150
200
P
r
i
c
e

(
U
S
$
/
M
W
h
)
a:


actual
mean
95%CI
24 J an 05 T W T F S S
50
100
150
200
Time (hour)
P
r
i
c
e

(
U
S
$
/
M
W
h
)
b:


simulation
Figure 5.16: The Mid-winter Week in 2005 - the upper graphs plot the
actual hourly spot prices, the lower graphs plot a counterpart simulation, and
the dashed lines plot the expected mean value and the 95% condence intervals.
The expected mean value and the condence interval are calculated from the
price model.
106
5.7 Simulate Electricity Spot Prices with the Models
25 Apr 05 T W T F S S
20
40
60
80
100
120
P
r
i
c
e

(
U
S
$
/
M
W
h
)
a:


actual
mean
95%CI
25 Apr 05 T W T F S S
20
40
60
80
100
120
Time (hour)
P
r
i
c
e

(
U
S
$
/
M
W
h
)
b:


simulation
Figure 5.17: The Mid-spring Week in 2005
25 J ul 05 T W T F S S
20
40
60
80
100
120
140
P
r
i
c
e

(
U
S
$
/
M
W
h
)
a:


actual
mean
95%CI
25 J ul 05 T W T F S S
20
40
60
80
100
120
140
Time (hour)
P
r
i
c
e

(
U
S
$
/
M
W
h
)
b:


simulation
Figure 5.18: The Mid-summer Week in 2005
107
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
24 Oct 05 T W T F S S
50
100
150
200
250
P
r
i
c
e

(
U
S
$
/
M
W
h
)
a:


actual
mean
95%CI
24 Oct 05 T W T F S S
50
100
150
200
250
Time (hour)
P
r
i
c
e

(
U
S
$
/
M
W
h
)
b:


simulation
Figure 5.19: The Mid-fall Week in 2005
5.8 Use Price Simulations to Valuate Genera-
tor Prot
In this section, the simulated price sample paths are used to estimate how
much a generator is going to earn during a year. Two sample generators will
be considered, one coal unit and the other gas unit, refer to Table 5.5: the
coal unit is of heat rate 10 (mmBtu/MWh) and fuel price 2 (US$/mmBtu),
and thus its marginal cost, calculated as the heat rate times fuel price, is 20
(US$/MWh); the gas unit is of heat rate 8 (mmBtu/MWh) and fuel prices the
monthly natural gas spot prices.
Gen. ID
Heat-rate Fuel Price Cost
(mmBtu/MWh) (US$/mmBtu) (US$/MWh)
Coal 10 2 20
Gas 8 Spot, Monthly range from 55 to 118
Table 5.5: Characteristics of the Coal Unit and the Gas Unit
108
5.8 Use Price Simulations to Valuate Generator Prot
The total one-year prot for one unit(1 MW) capacity is calculated as the
spread between the electricity spot price and the marginal cost summing over
one year,
=

T
max [S(t) C(t), 0]
which means that, at a particular hour t, if the electricity spot prices S(t) is
higher than the marginal cost C(t), the generator gets to generate electricity
and earns a prot (t) = S(t) C(t); if the electricity spot price is lower
than the cost, it doesnt get the chance to generate electricity and thus earns
the zero prot (t) = 0; the total one-year prot is the summation of hourly
prots throughout the whole year =

(t).
Lets take the coal unit for example, rst one thousand price sample paths
for the year 2005 are simulated; from each sample path a one-year generator
prot is calculated, thus there are 1000 simulated one-year generator prots;
from these 1000 simulated prots, the histogram of the generator prots is
calculated, and plotted in Figure 5.20a as the dark solid curve; from the prot
histogram one calculates the expected mean value of the generator prots,
plotted as the red dashed line; the expected mean value is then compared with
the actual one-year generator prot(plotted as the blue solid line), which is
calculated from the actual electricity spot prices. The expected prot turns out
close to the actual prot, with a deviation of 0.26%. Similarly, Figure 5.20b
plots the case of gas unit, and the expected prot deviates from the actual
prot by 2.48%. The above results are recorded in Table 5.6.
Prot(US$/MW)
Coal Gas
Actual 512,750 64,256
Expected 511,431 65,852
Error 0.26% 2.48%
Table 5.6: One-year Generator Prot of the Coal Unit and the Gas
Unit
109
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
0 100 200 300 400 500 600 700 800 900 1000
0
0.05
0.1
Profit (10
3
US$/MW )
P
r
o
b
a
b
i
l
i
t
y
a:


profit histogram
mean
actual
0 20 40 60 80 100 120
0
0.05
0.1
0.15
Profit (10
3
US$/MW )
P
r
o
b
a
b
i
l
i
t
y
b:


profit histogram
mean
actual
Figure 5.20: One-year Generator Prot of the Coal Unit and the
Gas Unit - (a) the coal unit, (b) the gas unit: the dark solid curve plots the
histogram of generator prots, the red dashed line plots the expected prot, and
the blue solid line plots the actual prot.
110
5.9 Discussion
5.9 Discussion
The New-England market has been taken for example to build the price model.
The reason for choosing the New-England market is due to its simple structure
of marginal fuel, that is, the marginal fuel in the New-England market is mostly
natural gas. The PJM market has a more complex mixture of marginal fuel,
both coal and natural gas account for a substantial proportion, as well as a
small percentage of diesel oil, thus in order to model the PJM Market, one has
to consider not only natural gas prices, but also coal and diesel oil prices. The
forthcoming Chapter 6 discusses modeling the PJM market.
In terms of modeling the Nordic Pool, because it has a large hydro-electric
generation capacity, the variations of reservoir water level during a year and
across years have profound eect on its system prices, thus in order to model
the Nordic Pool, an extension of the model to consider the reservoir level is
necessary. For modeling the other European markets, the prices of carbon
certicates have signicant eect on electricity spot prices, and it has to be
taken into account when building the models.
Electricity spot price spikes prevail in most electricity markets, see De Jong
(2006); Huisman & Mahieu (2003); Mount et al. (2006); Seifert & Uhrig-
Homburg (2007). Electricity price spikes are due to unexpected high electricity
load demand, shortage of generation capacity, generator operation constrains,
congestion of transmission networks, high fuel prices, market power, etc. De-
termined by the modeling methodology, the proposed model is probably not
able to model extreme price spikes that are caused either by the severe oper-
ation constraints of the engineering electric power systems, or by the manip-
ulations in an imperfect electricity market. The proposed model, however, is
probably able to model appropriately the price hikes that are caused either by
high electricity load demand or high fuel prices.
5.10 Conclusion
This work builds an electricity spot price model for generation investment
and scheduling analysis. The overall price model explicitly considers the vari-
111
5. A MULTI-GRANULARITY ELECTRICITY SPOT PRICE
MODEL
ous physical forces that underlie electricity spot prices, such as, intra-day and
weekday-weekend variations of electricity load, seasonal weather and tempera-
ture, annual generation planned outage, economic development and economic
cycles, generation investment and retirement, uctuations of fuel prices, etc.
The price model covers a time horizon of many years and has a time unit of
one hour. The nal result of the model is mathematically simple and elegant:
electricity spot prices at each hour are characterized by a Lognormal distri-
bution. This price model provides an essential analytical tool for generation
companies to make informed decisions on generation investment and schedul-
ing. Besides its application in generation investment and scheduling analysis,
the model could also be widely used by power trading companies in pricing
and trading electricity options, futures, and other electricity derivatives.
112
Chapter 6
The Multi-granularity Model
applied to the PJM Market
6.1 Introduction
In the last chapter, the New-England market was taken for example to il-
lustrate building the price model. The reason for choosing the New-England
market is due to its simple structure of marginal fuel, that is, its marginal fuel
is dominantly natural gas. This chapter applies a similar modeling methodol-
ogy to the Pennsylvania-New Jersey-Maryland(PJM) electricity market. The
general idea of decomposition applies well to the PJM market, while the di-
culty lies in that the PJM market has a more complex structure of marginal
fuels, that is, coal and natural gas are the marginal fuels most of the time, and
diesel oil accounts for a small portion of time. In order to model this complex
structure of marginal fuels in the PJM market, a Fuel Price Index of the three
marginal fuels is constructed and taken as the only proxy of fuel prices in the
market. The forthcoming sections elaborate with detail the modeling of the
PJM market.
113
6. THE MULTI-GRANULARITY MODEL APPLIED TO THE
PJM MARKET
6.2 The Fundamentals of the PJM Market
The Pennsylvania-New Jersey-Maryland(PJM) market started around 1998,
initially it covered mostly three states, Pennsylvania, New Jersey, and Mary-
land, PJM Interconnection (2009a). At the beginning, it only had a real-time
market, and the day-ahead market was initiated in June 2000. From 2002
to 2005, the market expanded to its west and south to include more control
areas. From 2005 to present(2009), the geographic footprint of the market
doesnt expand and keep the same. The market now includes 17 control zones
that locate in about 15 northeastern states. At present, the market has in-
stalled capacity around 160 GW; its system load ranges from 60 GW in spring
and fall, and 130 GW in summer and 110 GW in winter; its generation planned
outage during a year could reach as high as around 30 GW, see Figure 6.1. On
the source of these data, the installed generation capacity data are from PJM
Interconnection (2009f) and compiled from Monitoring Analytics (2009), the
system load data are collected from PJM Interconnection (2009d), and the
generation planned outage data from PJM Interconnection (2009c).
The PJM market has a more complex structure of marginal fuels than the
New-England market. In the New-England market, the marginal generators
are dominantly fueled by natural gas; while in PJM market, the marginal
generators are either fueled by coal, natural gas, or diesel oil. Lets look at the
mixture of marginal fuels during business days, Figure 6.2, during the o-peak
hours coal dominates the marginal fuel, and during the peak hours natural gas
and oil share the marginal fuel roughly equally with the coal; during one year
month by month, see Figure 6.3, coal is the marginal fuel in roughly about
70% of the time, and natural gas and diesel oil account for the rest, and in
high-demand seasons, e.g., in summer time, natural gas and diesel oil slightly
increase their share of time to about 40%; from 2004 to 2008, see Figure 6.4,
each year coal accounts for about 70% of the time as the marginal fuel, and this
percentage share of time increased slightly during these years. The data for
compiling the above statistics are the PJM Marginal Fuel Type Data, collected
from PJM Interconnection (2009e).
114
6.2 The Fundamentals of the PJM Market
02 03 04 05 06 07 08 09
0
20
40
60
80
100
120
140
160
180
Time(Hour)
G
W


capacity
load
planned outage
Figure 6.1: PJM Market Fundamentals - (top) the installed generation
capacity, (middle) system electricity load demand, and (bottom) the planned
generation outage
115
6. THE MULTI-GRANULARITY MODEL APPLIED TO THE
PJM MARKET

0%
20%
40%
60%
80%
100%
1 3 5 7 9 11 13 15 17 19 21 23
Oil
Gas
Coal
Figure 6.2: Business-day Intra-day Pattern of Marginal Fuel Mixture
- the percentage share of time of the three marginal fuels, hour by hour across
one day.

0%
20%
40%
60%
80%
100%
1 2 3 4 5 6 7 8 9 10 11 12
Oil
Gas
Coal
Figure 6.3: Intra-year Seasonal Pattern of Marginal Fuel Mixture -
the percentage share of time of three marginal fuels, month by month across a
year.
116
6.2 The Fundamentals of the PJM Market

0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
2004 2005 2006 2007 2008
oil
gas
coal
Figure 6.4: Multi-year Pattern of Marginal Fuel Mixture - the per-
centage share of time of the three marginal fuels, year by year from 2004 to
2008.
The marginal generators in the PJM market are mostly fueled either by
coal, natural gas, or diesel oil. For that the prices of these three fuels are sup-
posed to have signicant eect on electricity spot prices, it is necessary to look
at the prices of these three fuels. In Figure 6.5, the top curve plots the prices of
diesel oil from 2002 to 2008, the middle curve the prices of natural gas prices,
and the bottom curve that of coal. Look at these data, the natural gas and
diesel oil prices are volatile, and the coal prices stably follows a upward trend.
On the source of these fuel price data, the data of diesel prices are the No.
2 diesel spot prices at the New York harbor, collected from Energy Informa-
tion Administration (2009b), the data of natural gas prices are the natural gas
prices sold to the electric power sector in Pennsylvania, collected from Energy
Information Administration (2009c), and the coal prices are complied from the
annual reports of two large generation companies in the PJM region, which are
American Electric Power, refer to American Electric Power Co., Inc. (2009),
and Allegheny Energy, see Allegheny Energy Inc. (2009).
117
6. THE MULTI-GRANULARITY MODEL APPLIED TO THE
PJM MARKET
02 03 04 05 06 07 08 09
0
5
10
15
20
25
30
Time (month)
P
r
i
c
e

(
U
S
$
/
m
m
B
t
u
)


diesel
gas
coal
Figure 6.5: The Prices of the Three Marginal Fuels - (top) diesel oil,
(middle) natural gas, and (bottom) coal
6.3 The Electricity Prices and Fuel Price In-
dex
The historical electricity spot price data S(t) that will be used for this empirical
study are from the PJM day-ahead energy market, which are the hourly Load
Weighted Locational Marginal Pricing data. The price data were collected
from PJM Interconnection (2009b), and plotted in Figure 6.6a. The prices,
during 2002 to 2009, mostly range from around 50 to 150 US$/MWh, and the
infrequent price spikes could be as high as around 350 US$/MWh, e.g., in the
summer of 2006.
The marginal fuels of the PJM market are coal, natural gas, and diesel oil,
and the prices of these three fuels have signicant eect on electricity spot
prices in the PJM market. To consider this eect from all the three fuel prices,
a Fuel Price Index P(t), which is a function of these three fuel prices, is con-
structed and taken as the only proxy of the marginal fuel prices in the market.
The fuel price index is calculated as the weighted multiplication(product) of
118
6.3 The Electricity Prices and Fuel Price Index
02 03 04 05 06 07 08 09
0
50
100
150
200
250
300
350
Time (hour)
P
r
ic
e

(
U
S
$
/
M
W
h
)
(a) PJM Market Day-ahead Electricity
Spot Prices
02 03 04 05 06 07 08 09
2
3
4
5
6
7
8
9
10
11
Time (hour)
P
r
ic
e

(
U
S
$
/
m
m
B
t
u
)
(b) PJM Market Fuel Price Index
Figure 6.6: PJM Market Electricity Prices and Fuel Prices
02 03 04 05 06 07 08 09
0
1
2
3
4
5
6
a:
lo
g

U
S
$
/
M
W
h
02 03 04 05 06 07 08 09
1
1.5
2
2.5
b:
Time (hour)
lo
g

U
S
$
/
m
m
B
t
u
(a) Logarithm Prices - a: electricity
prices, b: fuel price index
03 04 05 06 07 08 09
3.25
3.5
3.75
4
4.25
a:
lo
g

U
S
$
/
M
W
h
03 04 05 06 07 08 09
1.25
1.5
1.75
2
2.25
b:
Time (hour)
lo
g

U
S
$
/
m
m
B
t
u
(b) Low-frequency Components - a:
electricity prices, b: fuel price index
Figure 6.7: Fuel-price Eect on Electricity Spot Prices
119
6. THE MULTI-GRANULARITY MODEL APPLIED TO THE
PJM MARKET
the three marginal fuel prices, and plotted in Figure 6.6b. The details on
building this fuel price index are elaborated in Appendix F.
With the electricity spot price data S(t) and the fuel price index data P(t)
ready, lets look at how the fuel prices aect the electricity spot prices. In Fig-
ure 6.7a the upper graph(a) plots the logarithm electricity spot prices s(t),and
the lower graph plots the logarithm fuel price index p(t). Then the logarithm
electricity spot prices and the logarithm fuel price index are ltered by a low
pass lter of one-year(8760hrs) central moving-average, to get rid of the less
relevant dynamics of higher frequencies, and leave only the low-frequency com-
ponents. The low-frequency component of the logarithm electricity spot prices
is plotted in Figure 6.7b the upper graph(a), and the low-frequency component
of the logarithm fuel price index is plotted in Figure 6.7b the lower graph(b).
Observing the two curves, their close match evidences the profound eect of
marginal fuel prices on electricity spot prices in the PJM market.
6.4 Data Decomposition
The data used for building the price model for the PJM market dates from
May 2005 to June 2009. The selection of this period is due to that from 2005
to 2009 the PJM market does not expand and its control areas remain the
same. Due to this change of time period in our concern, the Fuel Price Index
will be estimated again using the electricity price data and the fuel price data
from May 2005 to June 2009, see Appendix F.
Similarly as was done for modeling the New-England market, the electric-
ity spot price data s(t) will be decomposed into four components, that is, a
fuel price component p(t), a high-frequency component s

h
(t), a mid-frequency
component s

w
(t), and a low-frequency component s

y
(t).
a) The logarithm electricity spot prices s(t) are subtracted by the logarithm
fuel price index p(t), and one obtains the fuel-price adjusted electricity
spot prices in the logarithm domain s

(t): the data in the time domain


are plotted in Figure 6.8 the left graph, and the data in the frequency
spectrum F() is plotted in the right graph. For simplicity, the fuel-price
120
6.4 Data Decomposition
adjusted electricity spot price data s

(t), in the forthcoming analysis, will


again be referred according to its physical meaning Logarithm Implied
Marginal Generator Heat Rate Data.
b) The logarithm implied marginal generator heat rate data s

(t) is then
ltered by a high-pass lter H
h
() to get the high-frequency component
s

h
(t), see Figure 6.9 the left graph the high-frequency data in the time
domain, and the right graph its frequency spectrum F
h
().
c) Similarly a mid-frequency component s

w
(t) is obtained by applying a
mid-pass lter H
w
() on the logarithm heat rate data s

(t), see Fig-


ure 6.10 the left graph the data in the time domain, and the right graph
its frequency spectrum F
w
().
d) Finally a low-frequency component s

y
(t) is ltered out by applying a
low-pass lter H
y
(), see Figure 6.11 the left graph for the time domain
data, and the right graph the frequency spectrum F
y
().
The design of the lers, H
h
(), H
w
(), and H
y
(), and the procedures of
ltering, please refer to Appendix D.
06 07 08 09
-0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
Time (hour)
H
e
a
t
r
a
t
e

lo
g

U
S
$
/
M
W
h
(a) Time Domain, s

(t)
0 1/168 1/168
0
500
1000
1500
Frequency (cycle per hour)
M
a
g
n
it
u
d
e
a:
0 1/8760 2/8760 3/8760 4/8760 5/8760
0
500
1000
1500
Frequency (cycle per hour)
M
a
g
n
itu
d
e
b:
(b) Frequency Domain, F()
Figure 6.8: Implied Marginal Generator Heat Rate Data - logarithm
hourly electricity spot price data after being adjusted the eect of fuel prices
121
6. THE MULTI-GRANULARITY MODEL APPLIED TO THE
PJM MARKET
2 J an 06 9 J an 06
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Time (hour)
H
e
a
t
r
a
t
e

(
lo
g

m
m
B
t
u
/
M
W
h
)
... ...
(a) Time Domain, s

h
(t)
0 1/168 2/168
0
500
1000
1500
Frequency (cycle per hour)
M
a
g
n
it
u
d
e
(b) Frequency Domain, F
h
()
Figure 6.9: The High-frequency Component Data - obtained by applying
the high-pass lter H
h
()
06 07 08
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Time (hour)
H
e
a
t
r
a
t
e

(
lo
g

m
m
B
t
u
/
M
W
h
)
...
(a) Time Domain, s

w
(t)
0 1/168 2/168
0
500
1000
1500
Frequency (cycle per hour)
(b) Frequency Domain, F
w
()
Figure 6.10: The Mid-frequency Component Data - obtained by applying
the mid-pass lter H
w
()
122
6.5 Modeling
06 07 08
1.5
1.6
1.7
1.8
1.9
2
2.1
2.2
2.3
2.4
2.5
Time (hour)
H
e
a
t
r
a
t
e

(
lo
g

m
m
B
t
u
/
M
W
h
)

(a) Time Domain, s

y
(t)
0 1/8760 2/8760 3/8760 4/8760 5/8760
0
500
1000
1500
Frequency (cycle per hour)
M
a
g
n
it
u
d
e
(b) Frequency Domain, F
y
()
Figure 6.11: The Low-frequency Component Data - obtained by applying
the low-pass lter H
y
()
6.5 Modeling
So far the electricity spot price data s(t) have been decomposed into four
components that are driven by dierent and independent physical forces, that
is, a fuel price component p(t), a high-frequency component s

h
(t) driven by
short-term intra-week forces, a mid-frequency component s

w
(t) driven by sea-
sonal forces, and a low-frequency component s

y
(t) driven by long-term forces.
Similarly as for modeling the New-England market, the high-frequency com-
ponent s

h
(t) will be captured by an Intra-week Model s

h
(t), the mid-frequency
component an Intra-year Model s

w
(t), the low-frequency component a Multi-
year Model s

y
(t), and the fuel price index p(t) by a Fuel Price Model p(t).
The structure of the intra-week, intra-year, and multi-year model remains the
same as modeling the New-England market, and the Fuel Price Model p(t) is
again left unspecied. The design of the models please refer to Section 5.5.
6.6 Model Calibration
This section moves on to calibrate the three models, the intra-week model,
the intra-year model, and the multi-year model, similarly as has been done in
123
6. THE MULTI-GRANULARITY MODEL APPLIED TO THE
PJM MARKET
modeling the New-England market, by matching the historical data with the
models.
6.6.1 The Intra-week Model Calibration
The intra-week model s

h
(t) is calibrated from the high-frequency data s

h
(t),
the data are divided according to four seasons, winter, spring, summer, and
fall, and one set of parameters is estimated for each season. The estimated
deterministic function f
h
(t) that models the intra-day and intra-week pat-
terns, for each season, is plotted in Figure 6.12, together with the data of
a few mid-season sample weeks superimposing each other. The parameters
of the mean-reversion process x
hr
(t), k
h
and
h
, again for each season, are
recorded in Table 6.1. Comparing the estimated deterministic functions f
h
(t)
and the parameters of the mean-reversion process x
hr
(t) to those of the New-
England market, see Section 5.6.1, one notices the deterministic intra-day and
intra-week patterns, for each season, are to some extent similar, so are the
parameters of the mean-reversion process.
Model Season k
h

h
x
hr
(t)
winter 0.1132 0.0692
summer 0.1074 0.0606
spring 0.1342 0.0603
fall 0.1515 0.0715
Table 6.1: Parameters of the Intra-week Model
6.6.2 The Intra-year Model Calibration
The intra-year model s

w
(t) is calibrated from the mid-frequency data s

w
(t).
The deterministic function f
w
(t) that models the seasonal pattern is plotted
in Figure 6.13, together with three years of mid-frequency data superimposing
each year. Comparing this seasonal pattern of the PJM market with that of
the New-England market, refer to Figure 5.10, one notices the dierence that
124
6.6 Model Calibration
M T W T F S S
-1
-0.5
0
0.5
1
a:


M T W T F S S
-1.5
-1
-0.5
0
0.5
1
1.5
b:


M T W T F S S
-1
-0.5
0
0.5
1
Time (hour)
c:


M T W T F S S
-1
-0.5
0
0.5
1
Time (hour)
d:


Figure 6.12: The High-frequency Data and the Intra-week Pattern
- the high-frequency data s

h
(t) of three mid-season weeks superimposed one
by one, together with the estimated intra-week pattern f
h
(t): (a) winter, (b)
summer, (c) spring, (d) fall.
125
6. THE MULTI-GRANULARITY MODEL APPLIED TO THE
PJM MARKET
the seasonal pattern of the PJM market has three peaks, while that of the
New-England market has four peaks.
J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Time (Hour)
H
e
a
t
r
a
t
e

(
l
o
g

m
m
B
t
u
/
M
w
h
)


Figure 6.13: The Mid-frequency Data and the Seasonal Pattern - the
mid-frequency component data s

w
from 2006 to 2008 superimposed year by year,
together with the estimated seasonal pattern f
w
(t).
The reason for the four peaks in the New-England market is due to the high
electricity demand in winter and summer, and the annual generation planned
outage in spring and fall, see Figure 5.11. For the PJM market, the electricity
demand and generation planned outage have patterns, see Figure 6.14, that are
to some extent similar as these of the New-England market. There is, however,
a subtle dierence between the patterns of annual generation planned outage
of the two markets, look at the month of March: for the PJM market, the
amount of generation capacity in outage increases abruptly at the beginning of
March; while for the New-England market, the amount of generation capacity
in outage increases slowly across the whole month of March. This subtle
dierence may explain the missing peak of the seasonal pattern for the PJM
market, that is, the two peaks in the PJM market in winter and spring have
merged and become only one peak.
126
6.6 Model Calibration
J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
1.5
2
2.5
H
e
a
t

r
a
t
e

(
lo
g

M
M
B
t
u
/
M
W
h
) a:
J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
50
60
70
80
90
b:
L
o
a
d

(
G
W
)
J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
0
20
40
c:
Time(Month)
C
a
p
a
c
i
t
y

(
G
W
)
Figure 6.14: Seasonality of the PJM Market - (a) the seasonal patterns
of marginal generator heat rate, (b) system electricity load, (c) and generation
planned outage.
Model k
w

w
d x
wr
= k
w
x
wr
dt +
w
d z
w
1.475 0.2349
Table 6.2: Parameters of the Intra-year Model
127
6. THE MULTI-GRANULARITY MODEL APPLIED TO THE
PJM MARKET
The parameters of the mean-reversion process x
wr
(t), k
w
and
w
, of the
intra-year model are recorded in Table 6.2. These parameters of the mean-
reversion process are somehow close to those of the New-England market,
refer to Table 5.2.
6.6.3 The Multi-year Model Calibration
The multi-year model s

y
(t) is estimated from the low-frequency data s

y
(t),
and the estimated parameters are recorded in Table 6.3. The parameters of
the mean-reversion process x
yr
(t), k
y
and
y
, are again close to these of the
New-England market, refer to Table 5.3. To evaluate the calibrated multi-
year model, Figure 6.15 plots the monthly average of the actual low-frequency
data s

y
(t), as well as a counterpart simulation by the estimated model, and
in which the dashed curves plot the long-term trend and delimit the 95%
condence intervals.
06 07 08 09
2
2.05
2.1
2.15
Time (month)
H
e
a
t
r
a
t
e

(
l
o
g

m
m
B
t
u
/
M
W
h
)


actual
simulation
long-term trend
Figure 6.15: Calibration of the Multi-year Model
128
6.7 Price Simulations
Model a b k
y

y
s

y
(t) 2.095 N/A 0.0536 0.0096
Table 6.3: Parameters of the Multi-year Model
6.7 Price Simulations
After calibrating the intra-week model s

h
(t), the intra-year model s

w
(t), and
the multi-year model s

y
(t), and that the fuel price model p(t) has been left
undened, the overall electricity spot price model s(t) is now complete. In
order to evaluate the overall model, similarly as has been done for model-
ing the New-England market, a sample path of hourly electricity spot prices
throughout the whole year 2008 will be constructed, refer to Section 5.7.
The results of the simulation are summarized as follows: the lower graph
Figure 6.16b plots a simulated sample path of hourly electricity spot prices
throughout the whole year 2008, and the upper graph Figure 6.16a a counter-
part sample path of actual hourly electricity spot prices. Figure 6.17b plots the
histogram of the actual price path and the histograms of a few simulated sam-
ple paths; and the upper graph Figure 6.17a compares the actual and simulated
sample paths in terms of the empirical Cumulative Probability Distributions.
The statistics of these price distributions is summarized in Table 6.4.
Mean STD Skewness Kurtosis
Actual 66.58 30.90 1.75 8.54
Simulated Sample Path 67.52 31.50 1.24 5.06
Error 1.41% 1.95% 29.18% 40.75%
Table 6.4: Statistics of Empirical Probability Distributions
Again to have a micro-view on the actual and simulated price sample paths,
one zooms in from the one year perspective to look at a few mid-season sample
weeks. For the mid-winter week, in Figure 6.18 the upper graph(a) plots the
actual price sample path, and the lower graph(b) plots a simulated counter-
part. Similarly for the mid-spring week see Figure 6.19, the mid-summer week
Figure 6.20, and the mid-fall week refer to Figure 6.21.
129
6. THE MULTI-GRANULARITY MODEL APPLIED TO THE
PJM MARKET
J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
0
50
100
150
200
250
300
350
U
S
$
/
M
W
h
a:


J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
0
50
100
150
200
250
300
350
Time (hour)
U
S
$
/
M
W
h
b:


actual
simulation
Figure 6.16: Day-ahead Hourly Electricity Spot Prices in Year 2008
0 50 100 150 200 250
0
0.2
0.4
0.6
0.8
1
P
r
o
b
a
b
i
l
i
t
y
a:


0 50 100 150 200 250
0
0.05
0.1
0.15
0.2
Price (US$/MWh)
P
r
o
b
a
b
i
l
i
t
y
b:


actual
simulation
actual
simulation
Figure 6.17: Empirical Probability Distributions of the Prices in Year
2008
130
6.7 Price Simulations
14 J an 08 T W T F S S
0
50
100
150
200
P
r
i
c
e

(
U
S
$
/
M
W
h
)
a:


14 J an 08 T W T F S S
0
50
100
150
200
Time (hour)
P
r
i
c
e

(
U
S
$
/
M
W
h
)
b:


simulation
mean
actual
mean
95%CI
Figure 6.18: Mid-winter Week in 2008
14 Apr 08 T W T F S S
0
50
100
150
200
P
r
i
c
e

(
U
S
$
/
M
W
h
)
a:


14 Apr 08 T W T F S S
0
50
100
150
200
Time (hour)
P
r
i
c
e

(
U
S
$
/
M
W
h
)
b:


simulation
mean
actual
mean
95%CI
Figure 6.19: Mid-spring Week in 2008
131
6. THE MULTI-GRANULARITY MODEL APPLIED TO THE
PJM MARKET
14 J ul 08 T W T F S S
0
50
100
150
200
250
300
Time (hour)
P
r
i
c
e

(
U
S
$
/
M
W
h
)
a:


14 J ul 08 T W T F S S
0
50
100
150
200
250
300
P
r
i
c
e

(
U
S
$
/
M
W
h
)
b:


simulation
mean
actual
mean
95%CI
Figure 6.20: Mid-summer Week in 2008
13 Oct 08 T W T F S S
20
40
60
80
100
120
140
P
r
i
c
e

(
U
S
$
/
M
W
h
)
a:


13 Oct 08 T W T F S S
20
40
60
80
100
120
140
Time (hour)
P
r
i
c
e

(
U
S
$
/
M
W
h
)
b:


actual
mean
95%CI
simulation
mean
Figure 6.21: Mid-fall Week in 2008
132
6.8 Generator Prot Valuation
6.8 Generator Prot Valuation
After constructing a sample path of hourly electricity spot prices throughout
the whole year 2008, many more price sample paths will be simulated and then
used to estimate the whole-year total generator prot for a coal unit and a gas
unit. The characteristics of the two generator, that is, their heat rate, fuel
price, and marginal cost, could be found in Table 5.5, as well as the denition
on the generator prot, refer to Section 5.8.
Take the coal unit for example, one thousand sample paths of hourly prices
throughout the whole year 2008 are simulated, out of which one thousand one-
year generator prots are calculated, and with these one thousand simulated
generator prots a histogram of generator prots could be obtained, which is
plotted in Figure 6.22a as the dark solid curve, together with the expected
mean value(the red dashed line) of the simulated generator prots and the
actual generator prot(the blue solid line). Look at the error between the
expected generator prot and the actual generator prot, the deviation of the
expected prot from the actual prot is within 5%, Table 6.5. For the case
of the gas unit, the results are plotted in the lower graph Figure 6.22b and
recorded in Table 6.5.
Prot(US$/MW)
Coal Gas
Actual 384,119 38,959
Expected 401,651 52,755
Error 4.63% 3.38%
Table 6.5: One-year Generator Prot of the Coal Unit and the Gas
Unit
6.9 Conclusion
This chapter models the electricity spot prices for the PJM market. The mod-
eling methodology is mostly the same as that for modeling the New-England
133
6. THE MULTI-GRANULARITY MODEL APPLIED TO THE
PJM MARKET
0 100 200 300 400 500 600 700 800
0
0.05
0.1
Profit (10
3
US$/MW )
P
r
o
b
a
b
i
l
i
t
y
a:


0 10 20 30 40 50 60 70 80 90 100
0
0.05
0.1
0.15
Profit (10
3
US$/MW )
P
r
o
b
a
b
i
l
i
t
y
b:


PDF
mean
actual
PDF
mean
actual
Figure 6.22: One-year Generator Prot of the Coal Unit and the Gas
Unit
market, except that the PJM market has a more complex structure of marginal
fuels, and thus a Fuel Price Index is built as the only proxy for the prices of
the marginal fuels. The same models, that is, a fuel price model, an intra-week
model, an intra-year model, and a multi-year model, which have been used to
model the New-England market, are applied to the PJM market. The models
are then calibrated from the PJM market historical data, and the estimated
parameters of the models turn out somehow close to those of the New-England
market. To evaluate the overall model, a sample path of hourly electricity spot
prices through out the whole year 2008 is constructed and compared with the
actual price sample path. Afterward a large number of price sample paths
are simulated and then applied to estimate the one-year total generator prot
for a coal unit and a gas unit, and the estimates of the generator prots are
reasonably close to the actual generator prots. Overall, this case study evi-
dences the validity of the decomposition-based modeling methodology in the
PJM market.
134
Chapter 7
A Structural Electricity Spot
Price Model Considering
Generator Forced Outage
7.1 Abstract
The system of electricity spot prices is complex, for that there are many physi-
cal forces underlying electricity spot prices. Among these fundamental physical
forces, generation forced outage is one important stochastic source. This work
takes the structural approach to study the eect of generation forced outage
on electricity spot prices. To simplify the problem in the rst place, it focuses
on the time period of one hour, it starts from modeling each generation unit,
assembles these generator models into a supply stack, assumes a determinis-
tic electricity load, constructs electricity spot prices by the supply stack and
the electricity load, and nally derives an analytical expression for electricity
spot prices. The model is then evaluated by a numerical example based on a
test generation system. The analytical results by the model turn out close to
the results by Monte Carlo Simulations. This structural model demonstrates
explicitly the signicant eect of generation forced outage on electricity spot
prices.
135
7. A STRUCTURAL ELECTRICITY SPOT PRICE MODEL
7.2 Introduction
Electricity spot prices are settled by the interactions between the supply and
demand of electricity markets. On the supply side, generation companies sub-
mit their bids to the independent market operator, as well as on the demand
side, electricity retailers(buyers) also submit their bids to the market opera-
tor, the market operator then aggregates the suppliers bids into a supply stack
and the buyers bids into a demand stack, the intersection of the supply stack
and the demand stack sets the spot price for the market. This structure of
supply and demand of electricity markets provides the basic framework to un-
derstand, analyze, and then model electricity spot prices. In this framework,
from both the supply and the demand sides, the basic constituent components
that construct electricity spot prices are the shape of the generation supply
stack, the available generation capacity, the prices of fuels, system electricity
load, demand response, etc.
The Structural Approach models electricity spot prices indirectly; it rst
models the more fundamental driving forces, and then constructs the resulting
constituent models into an electricity spot price model. Ideally, the nal result
of the structural model ought to have an elegant analytical expression. To
satisfy this requirement, the constituent models for the fundamental driving
forces have to be as simple as possible, by which it is meant that the irrelevant
physical forces and the negligible details are excluded, and the signicant un-
derlying forces are approximated, simplied, and captured properly. On the
other hand, if one wants to study the eect of one particular physical underly-
ing force on electricity spot prices, one probably needs a detailed model on this
particular physical force, and in order to accommodate this extra complexity,
while at the same time maintaining a simple nal result, one necessarily has
to compensate the complexity of the other constituent models that are not the
immediate concern.
In the literature, a few authors have attempted the Structural Approach of
modeling electricity spot prices. One pioneering work is Skantze & Ilic (2001),
they modeled the supply curve with an exponential function, and the funda-
mental driving forces from both the supply and demand sides with 2-factor
136
7.2 Introduction
stochastic processes. The idea of modeling the supply curve with an expo-
nential function is further investigated by Barlow (2002); Cartea & Villaplana
(2008). Later the idea of modeling the supply stack by a cubic spline function
is proposed by Burger et al. (2004, 2007). The work Boogert & Dupont (2008)
models the hourly prices as a non-parametric function of reserve margin. A
recent development is Howison & Coulon (2009) who emphasizes the eect of
fuel prices on the movements of the bidding stacks.
Along this line of structural models, there are a few authors who have
somehow considered the eect of generation forced outage on electricity spot
prices. The work Davison et al. (2002) took a hybrid structural approach that
makes use of both the historical price data and the knowledge on engineering
power systems. It studied the PJM market and classied daily prices into price
spikes and non-spikes, it modeled each price regime by a normal distribution,
and the switching between the two regimes controlled by a function as the
ratio of electricity load over the available generation capacity. And the model
of available generation capacity has been further extended by Anderson &
Davison (2008) to consider generation planned outage and generation forced
outage. The dierence between their work and this work is as follows: they
took a hybrid approach by making use of both the historical price data and
the knowledge on engineering power systems, while our approach is bottom-up
and does not use the historical price data, we derive the electricity spot prices
using the fundamental power system models.
The focus of this work is on studying how the forced outage of individual
generators aects the shape of generation supply stack and thereafter the elec-
tricity spot prices. To simplify the problem in the rst place, only one time
period(one hour) will be considered, fuel prices is considered xed, and we will
start with modeling each generation unit, assemble these generator models into
a generation supply stack, assume a deterministic electricity load, and nally
constructs electricity spot prices out of the generation supply stack model and
the electricity load model.
One may notice that the resulting structural model looks somehow similar
as a hockey-stick model proposed by Kanamura & Ohashi (2007). The dier-
ence between our structural model and their hockey-stick model is two-fold:
137
7. A STRUCTURAL ELECTRICITY SPOT PRICE MODEL
rst Kanamura & Ohashi (2007) took a hybrid approach by using the histor-
ical electricity spot price data to estimate the hockey-stick like supply stack,
while our approach is bottom-up and does not use historical price data; second
the emphasis of Kanamura & Ohashi (2007) is to model the price spikes, while
the purpose of our model is to investigate how generation forced outage aects
electricity spot prices in general, and price spikes is only part of our concern.
The rest of this chapter is organized as follows: Section 7.3 introduces
how generation forced outage aects electricity spot prices; Section 7.4 builds
a structural electricity spot price model, derives the variables of the model,
and calculates the spot prices; Section 7.5 proposes a simplied model and
derives the probability density function of the spot prices; Section 7.6 con-
ducts numerical studies on a test generation system in order to evaluate the
model; Section 7.7 discusses the limitations of the work, and proposes further
extensions; Section 7.8 summarizes and concludes.
7.3 The Eect of Generator Forced Outage on
Spot Prices
Electricity spot prices, at any single hour, are settled as the intersections of
the generation supply stack and the electricity load demand. In a competi-
tive market, where each market player bids its marginal cost, the generation
supply stack is equivalent to the marginal cost curve of the electricity market,
Figure 7.1. The marginal cost curve consists of a large number of generators,
and these generators are ordered according to their marginal cost, from cheap
to expensive. Generators are classied into two types: base units that are
running throughout the day, and peaking units that only serve the daytime
peak hours. Base units are of large capacity size and low marginal cost, and
they are the hydro-electric, nuclear and coal power plants; peaking units are
of smaller capacity size and higher marginal cost, and they are usually fueled
by natural gas and oil. On the generation supply stack, base units locate at
the at bottom, and peaking units form the steeper shoulder.
138
7.3 The Eect of Generator Forced Outage on Spot Prices
Q
MC
G1 G2

Base Peak
Figure 7.1: The Generation Supply Stack of an Electricity Market
Generator Forced Outage is a major source of uncertainty on the supply
side. A generator may be forced to reduce its output or shut down due to
equipment failures. Generators, more costly, will line up to supply electricity.
While some generators are being forced to outage and some others are being
repaired and returning to operation, the supply-stack is constantly moving
rightward or leftward, and meanwhile adjusting its slopes.

Q
MC
S
S'
S
Prob.
G
i

L
Figure 7.2: The Eect of Generator Forced Outage on Spot Prices
For example, if given the demand curve, a vertical line representing the
amount of electricity load, it intersects the supply-stack originally at S, Fig-
ure 7.2, and then if a generator G
i
is forced to outage, the supply-stack moves
leftward, spot price becomes S

. Other generators, similar as the unit G


i
,
may also be forced to outage. The spot prices are thus randomly uctuating.
These uctuating spot prices could be modeled as a random variable

S. In
139
7. A STRUCTURAL ELECTRICITY SPOT PRICE MODEL
short, given a deterministic load L and the characteristics of the generators,
such as their capacity size a
i
, forced outage rate r
i
, and marginal cost c
i
, we
may derive the probability density function of the spot prices

S.
7.4 The Structural Model
7.4.1 The Structural Model
The supply stack is approximated by two segments, Figure 7.3, a base-segment
and a peaking-segment. The base-segment starts at the marginal cost of the
rst base unit c
1
, and ends at the marginal cost of the last(most expensive)
available base unit c
N
1
; its slope modeled by a random variable
1
; and with
the length of its basement, which represents the total available base generation
capacity, modeled by another random variable

A
1
; the peaking-segment starts
from the marginal cost of the rst peaking unit c

1
, and ends at the marginal
cost of the last available peaking unit c

N
2
; its slope is modeled by
2
; and the
length of its basement, which represents the total available peaking generation
capacity, is modeled by

A
2
. Electricity load L is given as deterministic.

Q
MC
2
A


L
S


1
A

Figure 7.3: The Structural Electricity Spot Price Model


The supply stack and the load L intersect and settle the spot prices

S.
140
7.4 The Structural Model
When load L exceeds the total available generation capacity

A
1
+

A
2
, spot
prices are set as the price cap J. This structural electricity spot price model
is summarized as follows,

S =
_

_
c
1
+
1
L L

A
1
c

1
+
2

_
L

A
1
_

A
1
<L

A
1
+

A
2
J L >

A
1
+

A
2
Total available base capacity

A
1
is the summation of the available capacity
of single base units, as

A
1
= a
1
+ a
2
+...+ a
N
1
, where a
i
is the available capacity
of a single base generator G
i
, which is,
a
i
=
_
a
i
Prob. 1 r
i
0 Prob. r
i
and where a
i
is the capacity size of generator G
i
and r
i
its forced outage rate.
The slope of the base segment is a function of the total available base ca-
pacity and the marginal cost of the available base units, which is approximated
by

1
=

C
1
c
1

A
1
where

C
1
is the marginal cost of the last available base unit, mathematically,

C
1
=
_
_
_
max
_
c
i
: a
i
= 0 in

A
1
_

A
1
= 0
c
0

A
1
= 0
and where c
0
is set as c
0
= J, c
1
is the marginal cost of the rst base unit,
and c
i
is the marginal cost of a base unit G
i
. And please note that these base
generators have been arranged according to their marginal cost from low to
high, namely, c
1
< c
2
< ... < c
N1
.
Total available peaking capacity is modeled by

A
2
, which is

A
2
= a

1
+ a

2
+
... + a

N
2
. The slope of the peaking segment
2
, similar as that of the base
segment, is modeled by
2
=

C
2
c

A
2
, where

C
2
is the marginal cost of the last
available peaking unit, as,

C
2
=
_
_
_
max
_
c

j
: a

j
= 0 in

A
2
_

A
2
= 0
c

A
2
= 0
141
7. A STRUCTURAL ELECTRICITY SPOT PRICE MODEL
and where c

0
is set as c

0
= J, c

1
is the marginal cost of the rst peaking
unit, and c

i
the marginal cost of the peaking unit G

i
. The peaking units
have also been arranged according to their marginal cost, from low to high,
c

1
< c

2
< ... < c

N2
.
7.4.2 The Available Generation Capacity
In a generation system of N generators, total available capacity

A satises the
equation,

A = a
1
+ a
2
+... + a
N
Denote the probability density function of

A as
p
N
(x)

= Pr
_

A = x
_
p
N
(x) could be derived by a recursive convolution algorithm, see Billinton
& Allan (1983): rst given the probability density function of the simplest
system, the rst generator, which is p
1
(x)

= Pr { a
1
= x}, and the forced outage
rate of the rest of the generators, r
2
, r
3
, . . . , r
N
, then p
N
(x) could be computed
by recursively calling the following convolution routine,
p
k
(x) = r
k
p
k1
(x) + (1 r
k
)p
k1
(x a
k
) k = 2, ..., N
7.4.3 The Segment Slopes
The slope of the segment has been dened as a function of the total available
generation capacity

A and the marginal cost of the last available generation
unit

C, as,
=

C c
1

A
Therefore, the probability of occurring the slope =
c
i
c
1
x
, in which the avail-
able generation capacity

A = x, and the marginal cost of the last available
generator

C = c
i
, is denoted as,
p
N
i
(x)

= Pr
_
=
c
i
c
1
x
_
= Pr
_

A = x;

C = c
i
_
i = 0, 1, ..., N
142
7.4 The Structural Model
It will be referred as the cost-indexed probability density function, and they
satisfy the equation, by denition,
p
N
(x) =
N

i=0
p
N
i
(x)
The algorithm for calculating p
N
i
(x), i = 0, 1, ..., N, could be derived from
the recursive convolution routine that has been introduced in the last section.
Given the probability density function of the available capacity of a genera-
tion system that has k 1, k = 2, ..., N, generators, p
k1
(x), it satises, by
denition,
p
k1
(x) =
k1

i=0
p
k1
i
(x)
According to the convolution routine, the probability density function of the
available capacity of a generation system that has k generators could be cal-
culated by,
p
k
(x) = r
k
p
k1
(x) + (1 r
k
)p
k1
(x a
k
)
Replace p
k1
(x) =
k1

i=0
p
k1
i
(x), then follows,
p
k
(x) = r
k
_
k1

i=0
p
k1
i
(x)
_
+ (1 r
k
)p
k1
(x a
k
)
= r
k
_
p
k1
0
(x) +p
k1
1
(x) +... +p
k1
k1
(x)

+ (1 r
k
)p
k1
(x a
k
)
=
_
r
k
p
k1
0
(x)

+
_
r
k
p
k1
1
(x)

+... +
_
r
k
p
k1
k1
(x)

+
_
(1 r
k
)p
k1
(x a
k
)

and by denition,
p
k
i
(x) =
_
r
k
p
k1
i
(x)

i = 0, 1, ..., k 1
and
p
k
k
(x) =
_
(1 r
k
) p
k1
(x a
k
)

one arrives at,


p
k
(x) = p
k
0
(x) +p
k
1
(x) +... +p
k
k1
(x) +p
k
k
(x)
143
7. A STRUCTURAL ELECTRICITY SPOT PRICE MODEL
Recursively calling the above routine, starting from k = 2, till k = N, one ob-
tains the cost-indexed probability density function of the available generation
capacity of a N generator system, p
N
i
(x), i = 0, 1, ..., N.
Lets demonstrate the simplest case, a generation system of two generators,
that is N = 2, to illustrate the algorithm. Given the rst generator G
1
, its
cost-indexed probability density functions, by denition, is
p
1
0
(x)

= Pr
_
a
1
= x,

C = c
0
_
p
1
1
(x)

= Pr
_
a
1
= x,

C = c
1
_
and they satisfy the equation,
p
1
(x) = p
1
0
(x) +p
1
1
(x)
The cost-indexed probability density function of the two generator system,
which consists of the rst and the second generator, p
2
0
(x), p
2
1
(x), and p
2
2
(x)
are thereafter calculated as follows,
p
2
(x) = r
2
p
1
(x) + (1 r
2
)p
1
(x a
2
)
= r
2
_
p
1
0
(x) +p
1
1
(x)

+ (1 r
2
)p
1
(x a
2
)
=
_
r
2
p
1
0
(x)

+
_
r
2
p
1
1
(x)

+
_
(1 r
2
)p
1
(x a
2
)

= p
2
0
(x) +p
2
1
(x) +p
2
2
(x)
7.4.4 The Expected Value of the Spot Prices
Given a deterministic electricity load L; the probability density functions of
the available generation capacity

A
1
and

A
2
, p
N
1
(x
1
) and p
N
2
(x
2
); and the
cost-indexed probability density functions of the slopes
1
and
2
, p
N
1
i
(x
1
)
and p
N
2
j
(x
2
); the algorithms for calculating the expected value of the spot
prices,

= E
_

S
_
, are derived as follows.
There are two major random variables involved,

A
1
and

A
2
, the expected
value is calculated by integrating over the joint distribution of the two random
variables, Figure 7.4: if condition

A
1
= x
1
, then 0 x
1
A
1
, where A
1
is
the total base capacity, A
1
= a
1
+ a
2
+ ... + a
N
1
; if condition

A
2
= x
2
, then
0 x
2
A
2
, where A
2
is the total peaking capacity, A
2
= a

1
+a

2
+... +a

N
2
.
Separate the three cases, when the load L intersects the generation supply
stack on dierent segments,
144
7.4 The Structural Model

2
A

1
A
L
L
2
x

1
S

2
S

LOLP
1
x

Figure 7.4: The Integration Regions of the Three Cases
a) L>

A
1
+

A
2
, when the given electricity load L exceeds the available gen-
eration supply

A
1
+

A
2
, spot prices

S equals to the price cap J. The
probability that load exceeds the available generation supply, namely
Pr
_
L >

A
1
+

A
2
_
, is familiar to power system engineers as Loss of Load
Probability(LOLP), refer to Billinton (1970). LOLP is calculated by
integrating over the range of

A
1
and

A
2
: if rst condition

A
1
= x
1
,
0 x
1
< L, then follows, by conditioning

A
2
= x
2
, 0 x
2
< L x
1
,
LOLP =
L
_
0
Lx
1
_
0
p
N
2
(x
2
) p
N
1
(x
1
)dx
2
dx
1
b)

A
1
<L

A
1
+

A
2
, when the given load L intersects the supply stack at
the peaking segment. The expected value of the spot prices is calculated
by integrating over the feasible region: if rst condition

A
1
= x
1
, then
0 x
1
< L, and then follows, by conditioning

A
2
= x
2
, Lx
1
x
2
A
2
,
S
1
=
L
_
0
A
2
_
Lx
1
N
2

j=0
_
c

1
+
c

j
c

1
x
2
(L x
1
)
_
p
N
2
j
(x
2
) p
N
1
(x
1
)dx
2
dx
1
c) L

A
1
, when the given load demand L intersects the supply stack at
the base-segment. The expected value of the spot prices is calculated by
integrating over the feasible region, by conditioning

A
1
= x
1
, L x
1

145
7. A STRUCTURAL ELECTRICITY SPOT PRICE MODEL
A
1
,
S
2
=
A
1
_
L
N
1

i=0
_
c
1
+
c
i
c
1
x
1
L
_
p
N
1
i
(x
1
)dx
1
The expected value of the spot prices, given the load L, thus is,
= S
1
+S
2
+J LOLP
Besides the expected value , other central movements of the spot prices S
could be calculated by a similar routine, e.g., the second central movement
standard deviation and the third central movement skewness. The standard
deviation is dened by,
=

E
_
_

S
_
2
_
and the skewness in this case is dened by,
=
3

E
_
_

S
_
3
_
7.5 The Simplied Model
7.5.1 The Simplied Model
Rather than settling with the analytical result of the expected value, standard
deviation, skewness of the spot prices, ideally, we seek for a probability density
function of the spot prices. The probability of occurring each spot price has
been computed by carefully multiplying the probability of occurring the right
slope and the right available generation capacity that give that spot price. The
most involved part of the model is the random slopes,
1
and
2
. If we could
make the slopes as deterministic, the model will be much simpler, and with
this simplied model a probability density function of the spot prices might be
obtained. By setting the slopes of the supply stack function deterministic, the
performance of the model, on the other hand, may have to be compensated.
If this compensation, however, is much less than the benet the probability
density function brings, the simplied model may be justiable.
146
7.5 The Simplied Model
If one sets the slopes
1
and
2
deterministic as their expected values,

1
=
A
1
_
0
N
1

i=0
_
c
i
c
1
x
1
_
p
N
1
i
(x
1
)dx
1
and

2
=
A
2
_
0
N
2

j=0
_
c

j
c

1
x
2
_
p
N
2
j
(x
2
)dx
2
The original segmental model is simplied as,

S =
_

_
c
1
+
1
L L

A
1
c

1
+
2

_
L

A
1
_

A
1
<L

A
1
+

A
2
J L >

A
1
+

A
2
7.5.2 The Probability Density Function of the Spot Prices
The probability density function of the spot prices could be derived by in-
tegrating over the joint distribution of the two random variables

A
1
and

A
2
:
by conditioning

A
1
= x
1
, thus 0 x
1
A
1
, and, by conditioning

A
2
= x
2
,
0 x
2
A
2
. Separate the three cases when the load L intersects the supply
stack on dierent segments.
a) L>

A
1
+

A
2
, when load L exceeds the available capacity, price equals
to the price cap J. The probability of occurring the price cap is LOLP.
LOLP is calculated by integrating over the region: if rst condition

A
1
=
x
1
, thus 0 x
1
< L, and then, by conditioning

A
2
= x
2
, 0 x
2
< Lx
1
,
LOLP = Pr
_

S = J
_
=
L
_
0
Lx
1
_
0
p
N
2
(x
2
) p
N
1
(x
1
)dx
2
dx
1
b)

A
1
<L

A
1
+

A
2
, when load L intersects the supply stack on the peaking
segment. Conditioning

A
1
= x
1
, thus 0 x
1
< L, the spot price is
S = c

1
+
2
(L x
1
)
147
7. A STRUCTURAL ELECTRICITY SPOT PRICE MODEL
its probability is obtained by integrating over the range, by conditioning

A
2
= x
2
, L x
1
x
2
A
2
,
Pr
_

S = c

1
+
2
(L x
1
)
_
=
A
2
_
Lx
1
p
N
2
(x
2
)dx
2
p
N
1
(x
1
)dx
1
c) L

A
1
, when load L intersects the supply stack on the base-segment.
The spot prices equal to
S = c
1
+
1
L
its probability is calculated by integrating over the range, by conditioning

A
1
= x
1
, L x
1
A
1
,
Pr
_

S = c
1
+
1
L
_
=
A
1
_
L
p
N
1
(x
1
)dx
1
7.6 Numerical Examples
7.6.1 The Test Generation System
A test generation system is built to conduct the numerical studies. This test
generation system, Figure 7.5, is designed to resemble an actual system. It
consists of base and peaking generators. Base units are of large capacity size,
they are fueled by coal; peaking units are of smaller capacity size, they are
fueled by natural gas. Marginal cost of the base units increases steadily, and
that of the peaking units climbs rapidly. Refer to the Appendix B for the
characteristics of these generators.
7.6.2 The Structural Model
The structural model will be evaluated by comparing it with Monte Carlo
Simulations. In Monte Carlo Simulations, spot prices are simulated as follows:
in each run for simulating one spot price, according to generator forced outage
rate, simulate the on-or-o state of each generator, these generators forms the
generation supply stack; given a deterministic load L, the intersections of the
148
7.6 Numerical Examples
0 1,000 1760 2650 3,000
0
10
20
30
40
50
60
70
80
90
100
Load(MW)
M
a
r
g
i
n
a
l

C
o
s
t
(
U
S
$
/
M
W
h
)
Peaking Generators Base Generators
Figure 7.5: The Generation Supply Stack of the Test Generation System
load L and the generation supply stack settles the spot prices. For any given
load L, overall 10000 spot prices are simulated. The expected value of the
spot prices is taken as the average of the simulated spot prices. This expected
value is compared with the analytical result that is calculated by the structural
model, Figure 7.6. Besides the expected value, the comparisons between the
model and Monte Carlo Simulations are also in terms of standard deviation in
Figure 7.7a, and skweness in Figure 7.7b.
Look at Figure 7.8a, which plots the errors of the model in terms of expected
value. When load L is low, L < 1000, the error, dened as the deviation of
the model from the simulation, reaches at most 12%; when load L ranges
1000 < L < 2650, the errors are mostly within 5%. In the cases of standard
deviation, Figure 7.8b and skewness, Figure 7.8c: when load is low, L < 1000,
the errors reach as much as 100%; when load is high, 1000 < L < 2650, the
errors decline fast to mostly within 10%. The large errors in the low-load range
do not undermine the validity of the model, because electricity load rarely goes
as low as that.
149
7. A STRUCTURAL ELECTRICITY SPOT PRICE MODEL
0 500 1000 1500 2000 2500 3000
0
20
40
60
80
100
120
140
J
160
Load(MW)
P
r
i
c
e

(
U
S
$
/
M
W
h
)


generation supply stack
simulation
model
Figure 7.6: The Expected Value of the Spot Prices as a Function of Load
0 500 1000 1500 2000 2500 3000
0
10
20
30
40
50
P
r
i
c
e

(
U
S
$
/
M
W
h
)
a:


0 500 1000 1500 2000 2500 3000
0
10
20
30
40
50
Load(MW)
P
r
i
c
e

(
U
S
$
/
M
W
h
)
b:


simulation
model
Figure 7.7: (a) The Standard Deviation and (b) Skewness of the Spot Prices
as Functions of Load
150
7.6 Numerical Examples
0 500 1000 1500 2000 2500 3000
-20
-10
0
10
20
%
a:


0 500 1000 1500 2000 2500 3000
-100
-50
0
50
100
b:
%
0 500 1000 1500 2000 2500 3000
-100
-50
0
50
100
c:
Load(MW)
%
Figure 7.8: The Errors of the Model: (a) Expected Value, (b) Standard Devi-
ation, (c) Skewness
7.6.3 The Simplied Model
The original structural model will be referred as Model I, and the simplied
model of deterministic slope will be referred as Model II. The results of Model
I and Model II are indistinguishable in terms of all the measures that are in
our concern, such as expected value, standard deviation, and skewness, refer
to Figure 7.9 and Figure 7.10. In terms of the percentage errors, Model II also
tracks closely the Model I, see Figure 7.11.
7.6.4 The Computation Burden of the Structural Mod-
els
The results of the Model I and Model II turned out close. The Model II is
simpler than Model I, thus, to compute a same task, Model II takes much less
time. Given the load L, to calculate the expected value, standard deviation,
and skewness of spot prices, Model II takes half of the time of the Model I;
and for the same task, Monte Carlo Simulation takes way longer time than the
models, Table 7.1. (In the Monte Carlo Simulation, it generates 10000 spot
151
7. A STRUCTURAL ELECTRICITY SPOT PRICE MODEL
0 500 1000 1500 2000 2500 3000
0
20
40
60
80
100
120
140
J
160
Load(MW)
P
r
i
c
e

(
U
S
$
/
M
W
h
)


generation supply stack
simulation
model I
model II
Figure 7.9: The Expected Value of the Spot Prices as a Function of Load
0 500 1000 1500 2000 2500 3000
0
10
20
30
40
50
P
r
i
c
e

(
U
S
$
/
M
W
h
)
a:


simulation
model I
model II
0 500 1000 1500 2000 2500 3000
0
10
20
30
40
50
Load(MW)
P
r
i
c
e

(
U
S
$
/
M
W
h
)
b:
Figure 7.10: (a) The Standard Deviation and (b) the Skewness of the Spot
Prices as Functions of Load
152
7.7 Discussion
0 500 1000 1500 2000 2500 3000
-20
-10
0
10
20
%
a:


model I
model II
0 500 1000 1500 2000 2500 3000
-100
-50
0
50
100
b:
%
0 500 1000 1500 2000 2500 3000
-100
-50
0
50
100
c:
Load(MW)
%
Figure 7.11: The Errors of the Model: (a) Expected Value, (b) Standard
Deviation, (c) Skewness
prices, each spot price is obtained by simulating the on-or-o state of each
generator in the generation system. All the computation is done on a HP PC
with Pentium 4 CPU(3.2GHZ) and 512M Ram memory.)
Model I Model II Monte Carlo Simulation
Time(s) 0.14 0.06 2.25
Table 7.1: The Computation Burden of the Structural Models
7.7 Discussion
We have evaluated the model by numerical examples on a test generation
system. By comparing the performance of the model with the result of Monte
Carlo Simulations, we intend to conclude that the structural model performs
satisfactorily. Due to the simplicity of the test generation system, however, one
may think that any proposition based on this over-simplied test generation
system is poorly grounded. Admittedly, an actual generation system is of
153
7. A STRUCTURAL ELECTRICITY SPOT PRICE MODEL
much more complexity: it consists of a large number of generators that are
of diverse types; for dierent markets around the world, the constitutions
of their generation systems are dierent. We defend these doubts by that,
the test generation system, despite its simplicity, reects the main features
of a typical generation system, that an actual generation supply stack may
be approximated by a segmental function that has more than two segments,
and that for each market, its generation supply stack are better modeled by
a segmental function that is tailor-made for that market. We admit, on the
other hand, to be condent in theses proposed structural models, extensive
empirical studies are necessary.
The limitations of this work are as follows. The model focused on the
time period of one hour. Generation investment analysis, however, concerns
the hourly spot prices across many years. The model, therefore, has to be
extended to have a much longer time perspective. When time starts to evolve,
both the generation supply stack and the electricity load demand begin to
change. As time passes, the constitution of a generation system adjusts slowly,
the division of the segments accordingly have to make change; fuel prices
uctuate along with time, marginal cost of the generators accordingly has to
make adjustment. On the demand side, electricity load varies between days
and nights, weekdays and weekends, and across seasons and years. The load
model has to be extended to include these features of electricity load demand.
When considering the time period of one hour, we assumed that load is
deterministic. Assuming that load is deterministic, however, is unrealistic for
the long-run concern of years and decades. A reasonable extension is to model
the load as a random variable. If load is modeled as a random variable, while
the generation supply stack remains modeled as a segmental non-continuous
function, an analytical results on the spot prices is no longer obtainable. To
have an analytical result when the load is modeled as a random variable, there-
fore, one has to simplify the segmental function, make it a simpler continuous
function, refer to the discussion on Structural Models in general, Section 2.4.
154
7.8 Conclusion
7.8 Conclusion
This work takes the structural approach to investigate the eect of generation
forced outage on electricity spot prices. It models each generator considering
its forced outage, assembles each generator model to form a generation supply
stack, assumes a deterministic electricity load demand, and constructs elec-
tricity spot prices out of the generation supply stack model and the electricity
load demand. It derives analytical results for the expected value, standard
deviation, skewness, and other central movements of electricity spot prices.
Upon a further simplied model, it derives the probability density function of
the electricity spot prices.
In order to evaluate the proposed structural models, numerical studies were
conducted on a test generation system. The results by the models are compared
with these by Monte Carlo Simulations: the values computed from the models
closely track the ones obtained from the Monte Carlo simulations, and the
errors between the results of models and these of the simulations, in percentage,
is mostly within 10%. Furthermove, the performance of the original structural
model and the simplied model is indistinguishably close.
This fundamental structural model provides a exible analytical tool for
investigating the eect of generation forced outage on electricity spot prices.
155
7. A STRUCTURAL ELECTRICITY SPOT PRICE MODEL
156
Chapter 8
Conclusion
8.1 Signicance of the Work
In the deregulated electric power industry, under the market environment,
electricity spot prices are highly volatile and uncertain. The various players
in electricity markets desire a simple, physically well-grounded, and fast elec-
tricity spot price model to facilitate their making informed decisions even on a
day-to-day basis. Generation companies need it for formulating bidding strate-
gies, scheduling production, and making decisions on generation investment;
power trading companies want it for pricing and trading various electricity
contracts, futures, options, and other derivatives; electricity retailers and large
industrial consumers use it to make decisions on purchasing power and man-
aging the inherent price risk. This work provides an analysis framework for
analyzing and understanding electricity spot prices, and develops an electric-
ity spot price model that is suitable for use by most of these applications in
electricity markets.
8.2 Contributions
This thesis work builds an electricity spot price model that is suitable for gen-
eration investment and scheduling analysis. The model takes into account the
various physical underlying forces that aect electricity spot prices in dierent
157
8. CONCLUSION
timescales, such as generation forced outages, temporal variations of electric
load, seasonal weather and temperature, annual generation planned outage,
economic development and cycles, generation investment and retirement, uc-
tuations of fuel prices, etc. The model covers a time horizon of many years
and has a time unit of one hour. The nal result of the model is mathemati-
cally elegant, namely, electricity spot prices at each hour are characterized by
a lognormal probability distribution.
This work develops a Multi-granularity Framework for analyzing and under-
standing electricity spot prices. Electricity spot prices are analyzed into three
time-perspectives of dierent granularity, multi-year, intra-year, and intra-
week; in these three perspectives, electricity spot prices are driven by dierent
and independent physical underlying forces, by applying Microeconomics Price
Theory to electricity markets, and making use of the knowledge on fundamen-
tal operations of electric power systems, it demonstrates how these dierent
and independent physical underlying forces give rise to the very peculiar be-
haviors of electricity spot prices.
This work employs the modeling methodology Divide and Conquer to build
the price model. Based on the understanding on the electricity spot prices pro-
vided by the Multi-granularity Framework, it applies Microeconomics Price
Theory and Fourier Analysis to decompose electricity spot prices into compo-
nents that are driven by dierent and independent physical underlying forces,
it then model each price component respectively by separate models, and -
nally constructs the resulting sub-models into a complete electricity spot price
model. It has demonstrated this modeling methodology in the New-England
and the PJM electricity markets.
8.3 Future Work on the Price Model
The proposed modeling methodology may be applicable to many other power
markets in the world. The proposed price model has been evaluated in the
New-England and the PJM electricity markets, so has it been applied to the
Nordic Pool. The result of the case study on the Nordic Pool so far is prelimi-
nary and thus has not been included in this thesis; nevertheless, the case study
158
8.3 Future Work on the Price Model
on the Nordic Pool evidences the validity of the methodology of decomposition.
Beyond these markets, the Multi-granularity Framework and the Methodology
of Decomposition could probably be applied to many other electricity markets
over the world, among which a few representative markets might be selected
for careful empirical studies in order to further evaluate, experiment, and thus
improve the modeling methodology.
The proposed price model is able to capture the modest price spikes but not
the extreme price spikes. Though the proposed model is capable of capturing
some usual price hikes that are driven by high electricity load demand, reduc-
tion of generation capacity due to planned outage, or high fuel prices, the model
is not able to capture the extreme-high price spikes that are probably due to
severe constraints of engineering power systems or the non-competitiveness of
electricity markets. These extreme-high price spikes deserve a separate study
by their own merit of importance.
How to consider the CO
2
prices in this modeling framework probably re-
quires further investigation. In these days, the possibility of CO
2
emission
causing climate change is catching much public attention. Since early 2005,
Europe has introduced the Emission Trading Scheme, and it has been observed
empirically that the prices of CO
2
certicates have signicant eect on elec-
tricity spot prices. The investigation of this CO
2
eect on electricity spot
prices, and thereafter incorporating it into the proposed price model, could be
one direction for future research.
The proposed modeling methodology has its relevance in future power mar-
kets which integrate a more uncertain supply side and a more active demand
side. On the supply side, there will be more distributed energy resources such
as Wind and Solar power, and on the demand side, there will be a much more
active demand side due to the participation of consumers to demand response
programs. This transformations of both the demand side and the supply side
will denitely aect the behavior of electricity prices. Because the modeling
methodology proposed in this work adopts a bottom-up analysis approach and
a statistical modeling approach, upon some minor revisions, it might be well
applicable to this emerging circumstance. Therefore, this application of the
159
8. CONCLUSION
modeling methodology to the changing industry is another direction for future
research.
8.4 Applications to Power System Operation
and Planning
As were kindly pointed out by the thesis examiners, the Multi-granularity
Analysis Framework and the Method of Decomposition are not only applicable
to power markets, but also might be utilized in power system operation and
planning.
One major concern for operating power systems is to ensure the balance
of power in any foreseeable future: immediate hour-ahead, day-ahead, month-
ahead, year-ahead, etc. Power engineers thus need to forecast the power gen-
eration and the load demand for dierent future time-horizons: the immediate
short-term for operation, the mid-term for unit-commitment and scheduling,
and the long-term for planning. This problem of forecasting the supply and
demand has been a major work facing power engineers for many decades. The
methodology developed in this thesis might be applicable to developing models
for better forecasting both the supply and the demand sides.
In the future, traditional power systems which are featured by centralized
generation and a passive demand side will undergo a fundamental transforma-
tion into integrating more distributed generation and a more active demand
side. On the supply side, there will be more intermittent renewable energy
resources like Wind and Solar power; and one the demand side, there will be
more active participation of consumers through demand side response. In this
emerging circumstance of a more uncertain supply side and a more active de-
mand side, the work of balancing power at any time any where will become
hugely more challenging.
In this forthcoming circumstance, the methodology that is developed in this
thesis work might be applicable to better understanding and then modeling
both the supply and demand sides. On the supply side, Wind and Solar
power are driven by both the seasonal and intra-day forces, these two forces
160
8.4 Applications to Power System Operation and Planning
belong to two dierent time-scales and probably behave dierently, and thus
one might decompose the Wind and Solar power into two dierent components
and then study them respectively; on the demand side, demand side response
will probably have intra-day, intra-week, and also seasonal patterns, these three
components might be obtained by decomposition and then studied respectively.
161
8. CONCLUSION
162
Appendix A
A Simple Structural Electricity
Spot Price Model
In an electricity market as a whole, at any single hour, the total cost to supply q
(MWh) electricity includes the x cost and the variable cost. Fix cost includes
maintenance, operation, and other cost. The variable cost is mainly fuel cost,
because in the time as short as one hour, the cost caused by increasing the
production is mostly from fuel. For an electricity market, the heat rates of the
marginal generators is a function of production q, denoted as HeatRate(q), in
which a marginal generator is the last(usually the most expensive) generator
that is dispatched to meet the load demand. For this market, the quantity of
the fuel burned for producing the q (MWh) electricity is a function of the heat
rates of the marginal generators HeatRate(q) as,
FuelQuantity(q) =
q
_
0
HeatRate(x)dx
Fuel cost is fuel quantity times fuel prices, thus it is,
FuelCost(q) =
q
_
0
HeatRate(x) FuelPrice(x)dx
163
A. A SIMPLE STRUCTURAL ELECTRICITY SPOT PRICE
MODEL
The total cost for supply the q (MWh) electricity, which is x cost plus fuel
cost, thus is,
TotalCost(q) = FixCost +FuelCost(q)
= FixCost +
q
_
0
HeatRate(x) FuelPrice(x)dx
For this market, the marginal cost of generating the q (MWh) electricity, by
denition, thus is
MarginalCost(q) =
dTotalCost(q)
dq
= HeatRate(q) FuelPrice(q)
In an ideal competitive market, where all market players have so little
generation capacity that no one is able to manipulate the market prices, given
a market price p, each market player is to maximize its own prot, as a result,
the market, overall, will supply a quantity q that is,
q = MargianlCost
1
(p)
This P-Q curve is the supply curve of the electricity market. The demand curve
of the market is a vertical line q = L dened by the amount of electricity load,
because in the time as short as an hour, the price doest not aect the amount
of electricity load demand L at all, namely, the price elasticity of the electricity
load demand is 0. The intersection of the two curves: q = MargianlCost
1
(p)
and q = L , settles the market spot price S as L = MarginalCost
1
(S),
namely, the intersection of the market marginal cost curve, MarginalCost (),
and the electricity load demand L settles the spot price, Figure A.1, by
S = MarginalCost (L) = HeatRate(L) FuelPrice(L)
To build a simple model of electricity spot prices S = MC (L), we will
rst dene the heat-rate of the marginal generators HeatRate(L) that is a
function of load L, re-denoted as HR(L). In an electricity market, large-
capacity generators that are of low heat rate serve the base load, and a large
number of small-capacity generators that are of high heat rate serve the peak
164



P
L
MC
S
Q
Figure A.1: A Simple Structural Electricity Spot Price Model
load. The heat-rate curve HR(L) by large is non-decreasing and convex. The
heat-rate curve HR(L) will be modeled by an exponential function,
HR(L) e
a+bL
Then we move on to dene the fuel prices of the marginal generators FuelPrice(L)
that is a function of load L. For simplicity, we will assume that the marginal
generators burn only one type of fuel, which is of price P. Finally, we arrive
at a simple structural model for electricity spot prices, as
S = MC(L) = HR(L) P
F
= e
a+bL
P
This simple structural electricity spot price model could serve as a basic frame-
work for understanding and analyzing electricity spot prices.
165
A. A SIMPLE STRUCTURAL ELECTRICITY SPOT PRICE
MODEL
166
Appendix B
The Test Generation System
This appendix records the test generation system: fuel types of the generators,
their capacity size, forced outage rate, heat rate, fuel prices, and marginal cost,
Table B.1. This test system is revised from the IEEE Reliability Test System,
Grigg et al. (1999). Only the large capacity coal and natural gas steam turbine
units are considered; small size combustion turbine units are excluded. The
capacity size of the units is adjusted slightly. The forced outage rate of the
units is taken from the Test System. The heat rate of the units is typical for
coal and natural gas generation units. The marginal cost of a unit is its heat
rate times its fuel price. All base units are fueled by coal, which is assumed
of price 2 (US$/mmBtu); all peaking units are fueled by natural gas, which is
of price 5 (US$/mmBtu). The marginal cost of the base coal units increases
steadily, and that of the peaking natural gas units climbs steeply, for that the
fuel price of natural gas is more than twice of that of coal.
167
B. THE TEST GENERATION SYSTEM
#
Size
FOR
Heat-rate Marginal Cost
(MW) (mmBtu/MWh) (US$/MWh)
1 400 0.08 8.0 16
2 200 0.05 9.3 18.6
Base 3 280 0.08 11.2 22.4
units 4 300 0.08 12.0 24
5 250 0.05 12.5 25
6 330 0.08 13.2 26.4
7 100 0.04 8.3 41.5
8 80 0.02 9.1 45.5
9 90 0.04 10.3 51.5
Peaking 10 110 0.04 11.0 55
units 11 115 0.04 11.5 57.5
12 120 0.04 12.0 60
13 150 0.04 12.9 64.5
14 125 0.04 13.9 69.5
Table B.1: Characteristics of the Generators in the Test System
168
Appendix C
Basic Stochastic Processes
C.1 Wiener Process
The Wiener process is written as,
d x = d z
where d z =
t

dt, and
t
N (0, 1), E {
t

t+dt
} = 0. Given the initial value
x(0) = x
0
, its solution at time t is
x(t) = x
0
+
t
_
0
d z
which is of mean E { x(t)} = x
0
, and variance
V ar { x(t)} = E
_
_
_
t
_
0
(d z)
2
_
_
_
=
t
_
0

2
dt =
2
t
Namely, x(t) is of a normal distribution with mean x
0
and standard deviation

t, as
x(t) N(x
0
,
2
t)
Wiener process is named after Norbert Wiener. Wiener process describes
the phenomenon Brownian Motion, discovered by Robert Brown in the 19
th
century, which is the random movements of pollen granules in hot water. The
169
C. BASIC STOCHASTIC PROCESSES
movements of the pollen granules are caused by the collisions of the surround-
ing water molecules. Water molecules are of much smaller size and of much
larger number than pollen granules. The above Wiener process describes the
Brownian Motion on one dimension. The normal distribution of the noise,

t
N (0, 1), is due to the averaging eect of large number of collisions in
that direction. That the averaging eect is of a normal distribution could be
derived by Central Limit Theorem. That the increment at time t,
t
, is inde-
pendent from that at the time t + dt,
t+dt
, namely, E {
t

t+dt
} = 0, follows
from the assumptions that each water molecule is independent, and that for
each molecule its movement at time t, denoted as

d
t
, is independent from that
at time t + dt,

d
t+dt
, that is, E
_

d
t


d
t+dt
_
= 0. The term

dt in d z is to
model that the variance of the accumulated increment is proportional to time
t, namely, E
_
t
_
0
(d z)
2
_
=
2
t.
C.2 Wiener Process with Drift
A Wiener process with drift rate v is written as
d x = vdt +d z
Given the initial value x(0) = x
0
, its solution at time t is
x(t) = x
0
+
t
_
0
vdt +d z = x
0
+v t +

t N(0, 1)
C.3 Geometric Brownian Motion
Geometric Brownian Motion describes a process

X(t), which is the exponential
of the Wiener process with drift,

X(t) = e
x
t
in which d x = vdt + d z. Write

X(t) in the form of stochastic dierential
equation, by applying Itos Lemma, is
d

X =
_
v +

2
2
_

Xdt +

Xd z
170
C.4 Mean-reversion Process
Denoting

= v +

2
2
, it becomes,
d

X =

Xdt +

Xd z
Observe the process, it implies that the drift rate,

X, is proportional to

X(t),
and the volatility of d

X,

X, is also proportional to

X(t). Given the initial
value

X(0) = X
0
, its solution in terms of x(t) is x(t) = ln X
0
+vt+

tN(0, 1),
thus the solution in terms of

X(t) is

X(t) = X
0
e
vt+

tN(0,1)
C.4 Mean-reversion Process
Mean-reversion process is known as Ornstein-Uhlenbeck process, refer to Dixit
& Pindyck (1994). OU process is named after two physicists G.E. Uhlenbeck
and L.S. Ornstein, for crediting their work in improving Einsteins models on
Brownian Motion.
a) The simplest OU process is
d x = k xdt +d z k > 0
Given x(0) = x
0
, its solution at time t is
x(t) = x
0
e
kt
+e
kt
_
t
0
e
ks
d z
s
which has mean and variance as,
E{ x(t)} = x
0
e
kt
Var { x(t)} =

2
2k
_
1 e
2kt
_
Write the solution in the form of a normal distribution, it is
x(t) = x
0
e
kt
+
_

2
2k
(1 e
2kt
)N(0, 1)
As time goes to innite t , the mean reaches 0, lim
t
E{ x(t)} = 0,
and the variance reaches a constant lim
t
Var { x(t)} =

2
2k
; namely this
simplest OU process has a tendency to reverting to the constant mean
0, and reaching a constant variance

2
2k
.
171
C. BASIC STOCHASTIC PROCESSES
b) The OU process with a constant mean m is derived as follows. An OU
process with a constant mean, denoted as y(t), equals to the summation
of the constant m and the OU process with zero-mean x(t), namely as,
y(t) = m+ x(t)
Take the dierential on both sides,
d y = d x
Replace d x = k xdt +d z, and then x = y m, one arrives,
d y = k (m y) dt +d z
Given the initial value of the mean-reversion process y(t), y(0) = y
0
,
because x(0) = y
0
m, thus the solution in terms of x(t) is
x(t) = (y
0
m) e
kt
+
_

2
2k
(1 e
2kt
)N(0, 1)
As y(t) = m+ x(t), thus the solution in terms of y(t), explicitly, is,
y(t) = m+ (y
0
m) e
kt
+
_

2
2k
(1 e
2kt
)N(0, 1)
c) An OU process with a time-varying mean is the summation of the time-
varying function f(t) and the OU process of zero mean x(t), as
y(t) = f(t) + x(t)
Take the dierential on both sides,
d y = df(t) +d x
Replace df(t) = f

(t)dt, d x = k xdt +d z, and then x(t) = y(t) f(t),


one arrives,
d y = k
__
f

(t)
k
+f(t)
_
y
_
dt +d z
172
C.5 Geometric Mean-reversion with Constant Mean
denoting g(t)

=
f

(t)
k
+f(t), one obtains a simpler expression,
d y = k [g(t) y] dt +d z
Again given the initial value of the mean-reversion process with time-
varying mean y(t), y(0) = y
0
, as x(0) = y
0
f(0), the solution in terms
of x(t) is,
x(t) = [y
0
f(0)] e
kt
+
_

2
2k
(1 e
2kt
)N(0, 1)
As y(t) = f(t) + x(t), thus the solution in terms of y(t) is,
y(t) = f(t) + [y
0
f(0)] e
kt
+
_

2
2k
(1 e
2kt
)N(0, 1)
C.5 Geometric Mean-reversion with Constant
Mean
Think about the process

X(t), which is the exponential of the simplest OU
process x(t), d x = kxdt +d z,

X(t) = e
x(t)
It says that the process x(t) moves around a zero mean, and then x(t) is
transformed by an exponential function. The eect of the exponential function
is that it amplies the values that is positive, x (0, ), to the range X
(1, ); and squeeze the values that are negative, x (, 0), to the range of
X (0, 1). One may imagine that

X(t) moves around 1, due to the amplifying
and squeezing eect, when it is less than 1 it behaves mildly; when it bypasses
1 it becomes more volatile.

X(t) in the form of a stochastic dierential equation is derived as follows.


As

X(t) = e
x(t)
, in which x(t) is the OU process, d x = kxdt + d z. Apply
the Ito Lemma, refer to Hull (2006), one arrives,
d

X = k
_

2
2k
ln

X
_

Xdt +

Xd z
173
C. BASIC STOCHASTIC PROCESSES
If given the initial value X(0) = X
0
, because x(0) = lnX(0) = lnX
0
, the
solution in terms of x(t) is,
x(t) = ln X
0
e
kt
+
_

2
2k
(1 e
2kt
)N(0, 1)
Thus the solution in the form of

X(t) is,

X(t) = e
x(t)
= e
ln X
0
e
kt
+


2
2k
(
1e
2kt
)
N(0,1)
As time t goes to innite, t , lim
t

X(t) = e


2
2k
N(0,1)
, its expected value
reaches the constant
lim
t
E
_

X(t)
_
= e

2
4k
C.6 Geometric Mean-reversion with Time-varying
Mean
The Geometric Mean-reversion with constant mean process

X(t) could be
extended to a Geometric Mean-reversion with time-varying mean process

S(t)
by multiplying a deterministic time-varying function G(t), as,

S(t) = G(t)

X(t)
Take natural logarithm on both sides, it becomes, ln

S(t) = ln G(t) +

X(t),
denote s(t)

= ln

S(t), and f(t)

= ln G(t), it becomes,
s(t) = f(t) + x(t)
Take dierential on both sides,
d s(t) = df(t) +d x(t)
Replaces df(t) = f

(t)dt, d x = k xdt +d z, and x(t) = s(t) f(t) one arrives


at,
d s = k
__
f

(t)
k
+f(t)
_
s
_
dt +d z
174
C.6 Geometric Mean-reversion with Time-varying Mean
As

S(t) = e
s(t)
, applying Itos Lemma, write the stochastic dierential equation
in terms of

S(t), one arrives,
d

S(t) = k
__
f

(t)
k
+f(t) +

2
2k
_
ln

S(t)
_

S(t)dt +

S(t)d z
Replace f

(t) =
dG(t)
G(t)dt
and f(t) = ln G(t), one has,
d

S(t) = k
__
dG(t)
G(t)kdt
+ ln G(t) +

2
2k
_
ln

S(t)
_

S(t)dt +

S(t)d z
Denoting g

(t)

=
dG(t)
G(t)kdt
+ln G(t) +

2
2k
, one obtains a more compact expression,
d

S(t) = k
_
g

(t) ln

S(t)
_

S(t)dt +

S(t)d z
If given X(0) = X
0
, the solution of

X(t) is,

X(t) = e
ln X
0
e
kt
+


2
2k
(
1e
2kt
)
N(0,1)
then the solution for

S(t) is

S(t) = G(t) e
ln X(0)e
kt
+


2
2k
(
1e
2kt
)
N(0,1)
As ln S(0) = ln G(0)+ln X(0), and thus by replacing X(0) = ln S(0)ln G(0),
the solution of

S(t), if given the initial value S(0) = S
0
, thus is,

S(t) = G(t) e
[ln S
0
ln G(0)]e
kt
+


2
2k
(
1e
2kt
)
N(0,1)
175
C. BASIC STOCHASTIC PROCESSES
176
Appendix D
The Price Filters and the
Filtering
D.1 Design of the Price Filters
a) The high-pass lter is,
H
h
() =
_
0 <
w
1
w
b) The mid-band-pass lter is,
H
w
() =
_
1
y
<
w
0 otherwise
c) The low-pass lter is,
H
y
() =
_
1 <
y
0
y
where
w
=
2
168
, and
y
=
2
8760
. These lters, H
h
(), H
w
(), and H
y
(), will
be constructed by two simple low-pass lters,
H
1
() =
_
1 <
y
0
y
H
2
() =
_
1 <
w
0
w
177
D. THE PRICE FILTERS AND THE FILTERING
as H
h
() = 1 H
2
(), H
w
() = H
2
() H
1
(), and H
y
() = H
1
(). The
lters H
1
() and H
2
() are ideal brick-wall lters of which the transition
between 0 and 1 is abrupt; the ideal lters are not realizable because they are
of innite length in the time domain. Approximate lters H

1
() and H

2
()
are used instead. The following two paragraphs elaborate the design of the
approximate lters H

1
() and H

2
().
i) H

1
() is a lter of moving average: in the time domain it is h

1
(t) =
1
8760
,
t = K
1
, ..., K
1
, and K
1
= 4380. The length of the lter h

1
(t) thus is
N
1
= 2K
1
+ 1 = 8761. Transform h

1
(t) to the frequency domain, it is,
H

1
() =

t=
h
1
(t)e
it
=
K
1

t=K
1
1
8760
e
it
=
1
8760
_
1 + 2
K
1

k=1
cos (k)
_
ii) H

2
() is an truncated ideal lter. The ideal lter H
2
() =
_
1 <
w
0
w
by Inverse Fourier Transform it becomes
h
2
(t) =
1
2

H
2
()e
it
d =
1
2

w
_

w
e
it
d =
sin (
w
t)
t
in which t = , ..., , note that the length is innite. The approximate
lter, by truncating the ideal lter, is
h

2
(t) =
sin (

w
t)
t
where t = K
2
, ...., K
2
, and K
2
= 168
5
2
, the length of the lter thus is
N
2
= 2K
2
+ 1 = 5 168 + 1; and where

w
=
2
192
, (note that it is not

w
=
2
168
, the reason for that will become clear soon). The lter h

2
(t),
by Fourier Transform, becomes
H

2
() =

h
2
(n)e
in
=
K
2

n=K
2
sin n

w
n
e
in
=

+2
K
2

k=1
sin 2k

w
k
cos k
in which

w
=
2
192
, and it is determined by subjecting to H

2
()|
=
w
0.
178
D.2 Price Decomposition by Filtering
If otherwise we set

w
=
w
=
2
168
, and the lter becomes
H

2
() =

+ 2
K
2

k=1
sin 2k
w
k
cos k
then follows H

2
()|
=
w
0.5, which is undesirable for ap-
proximating the ideal lter which requires that H
2
()|
=
w
=
0. The phenomenon that H

2
()|
=
w
0.5 is referred as
Gibbs Phenomenon (Strang & Nguyen, 1996, P. 46).
By the two approximate lters, H

1
() and H

2
(), the ideal lters H
h
(),
H
w
(), and H
y
() are approximated as H

h
(), H

w
(), and H

y
(), as summa-
rized in the Table D.1,
Frequency Domain Time Domain
H

h
() = 1 H

2
() h

h
(t) = 1 h

2
(t)
H

w
() = H

2
() H

1
() h

w
(t) = h

2
(t) h

1
(t)
H

y
() = H

1
() h

y
(t) = h

1
(t)
Table D.1: The Realized Approximate Price Filters
D.2 Price Decomposition by Filtering
Given logarithm marginal generator heat rate data s

(t), we will decompose it


into a high-frequency component s

h
(t), a mid-frequency component s

w
(t), and
a low-frequency component s

y
(t). The algorithm is developed as follows:
Lets take applying the high-pass lter h

h
(t) on the original data s

(t) to
obtain the high-frequency component s

h
(t) for example. First we take the
Fourier Transform of the original data s

(t),
F() =
+

t=
s

(t)e
it
then apply the high-pass lter in frequency domain H

h
() on F() to obtain
the frequency spectrum of the high-frequency component s

h
(t), F
h
(),
F
h
() = F() H

h
()
179
D. THE PRICE FILTERS AND THE FILTERING
then take the Inverse Fourier Transform of F
h
() to obtain s

h
(t), as
s

h
(t) =
1
2

F
h
()e
it
d
=
1
2

(F() H

h
()) e
it
d
=

k=
s

(t k)h

h
(k)

= s

(t) h

y
(t)
Therefore, as summarized in Table D.2, the algorithm to obtain the high-
frequency component s

h
(t) by working directly in the time domain is simply
an algorithm of convolution,
s

h
(t) = s

(t) h

h
(t)
Similarly for obtaining the mid-frequency component s

w
(t) the algorithm is,
s

w
(t) = s

(t) h

w
(t)
and for obtaining the low-frequency component s

y
(t) the algorithm is,
s

y
(t) = s

(t) h

y
(t)
Original Data Price Components after Decomposition Time Span
s

(t) t [0, T]
s

h
(t) = s

(t) h

y
(t) = s

(t) [1 h

2
(t)] t [K
2
, T K
2
]
s

w
(t) = s

(t) h

w
(t) = s

(t) [h

2
(t) h

1
(t)] t [K
1
, T K
1
]
s

y
(t) = s

(t) h

h
(t) = s

(t) h

1
(t) t [K
1
, T K
1
]
Table D.2: Electricity Spot Price Decomposition by Filtering
180
Appendix E
The Mean-reversion Process in
Discrete Time
Here we derive the discrete time version of the mean-reversion process x
r
,
d x
r
= k x
r
dt +d z
Move k x
r
dt to the left,
d x
r
+k x
r
dt = d z
Multiply both sides with e
kt
,
e
kt
d x
r
+k x
r
e
kt
dt = e
kt
d z
Apply the product rule of dierentiation dxy = ydx + xdy, the left hand side
becomes,
d x
r
e
kt
= e
kt
d x
r
+k x
r
e
kt
dt
thus the equation becomes,
d x
r
e
kt
= e
kt
d z
Integration over [t, t + t],
t+t
_
t
d x
r
e
kt
=
t+t
_
t
e
kt
d z
181
E. THE MEAN-REVERSION PROCESS IN DISCRETE TIME
The left hand side is
t+t
_
t
d x
r
e
kt
= x
r
(t + t)e
k(t+t)
x
r
(t)e
kt
And the right hand side,
E
_
_
_
t+t
_
t
e
kt
d z
_
_
_
= E
_
_
_
t+t
_
t
e
kt
(t)

dt
_
_
_
= 0
V ar
_
_
_
t+t
_
t
e
kt
d z
_
_
_
= E
_
_
_
t+t
_
t

2
e
2kt
( (t))
2
dt
_
_
_
=
t+t
_
t

2
e
2kt
dt
=

2
2k
t+t
_
t
de
2kt
=

2
2k
_
e
2k(t+t)
e
2kt
_
Thus the equation becomes,
x
r
(t + t)e
k(t+t)
x
r
(t)e
kt
=
_

2
2k
(e
2k(t+t)
e
2kt
) (t)
Divided the e
k(t+t)
from both sides,
x
r
(t + t) e
kt
x
r
(t) =
_

2
2k
(1 e
2kt
) (t)
Move e
kt
x
r
(t) to the right,
x
r
(t + t) = e
kt
x
r
(t) +
_

2
2k
(1 e
2kt
) (t)
182
Appendix F
Build the Fuel Price Index
F.1 The Fuel Price Index Model
In the PJM Market, the marginal fuels are either coal, natural gas, or diesel
oil, lets denote the prices of coal as

P
coal
(t), that of natural gas as

P
gas
(t), and
that of diesel oil as

P
diesel
(t). The Fuel Price Index

P(t) is then dened as the
weighted multiplication of the three fuel prices, as,

P(t) =

P
coal
(t)


P
gas
(t)


P
diesel
(t)

Take logarithm on both sides, denoting p(t)



= ln

P(t), p
coal
(t) = ln

P
coal
(t),
p
gas
(t) = ln

P
gas
(t), and p
diesel
(t) = ln

P
diesel
(t), the Fuel Price Index Model in
the logarithm domain thus is,
p(t) = p
coal
(t) + p
gas
(t) + p
diesel
(t)
Namely, in the logarithm domain, the Fuel Price Index is the linear combi-
nation of the prices of the three marginal fuels. The weighting parameters
, , and will be estimated statistically, and the model in discrete time for
estimating these parameters will be derived in the next section.
183
F. BUILD THE FUEL PRICE INDEX
F.2 Derivation of the Model for Parameter Es-
timation
Begin with the complete electricity spot price model that is derived in Sec-
tion 5.5.5, which is a multiplication of four sub-models, an intra-week model

h
(t), an intra-year model

S

w
(t), a multi-year model

S

y
(t), and a fuel price
model

P(t),

S(t) =

S

y
(t)

S

w
(t)

S

h
(t)

P(t)
Take logarithm on both sides, the model becomes an additive model,
s(t) = s

y
(t) + s

w
(t) + s

h
(t) + p(t)
where s

y
(t)

= ln

S

y
(t), s

w
(t) = ln

S

w
(t), s

h
(t)

= ln

S

h
(t), and p(t)

= ln

P(t),
and where now the fuel price model is replaced by the fuel price index model,
as p(t) = p
coal
(t) + p
gas
(t) + p
diesel
(t).
As the data of the fuel prices that will be used for parameter estimation
are of the time unit monthly, we thus take the monthly average on both sides
of the model,
T
k

t
k
s(t) =
T
k

t
k
s

y
(t) +
T
k

t
k
s

w
(t) +
T
k

t
k
s

h
(t) +
T
k

t
k
p(t)
Approximating the monthly average of the multi-year model by its long-term
constant level, namely,
T
k

t
k
s

y
(t) a; denoting the monthly average of the
intra-year model as, s

w
(k) =
T
k

t
k
s

w
(t); the monthly average of the intra-week
model is negligible, namely,
T
k

t
k
s

h
(t) 0; denoting the monthly average of the
fuel price index model as, p(k) =
T
k

t
k
p(t); denoting the monthly average of
the overall price model as, s(k) =
T
k

t
k
s(t); and one nally obtains the monthly
version of the price model, as,
s(k) = a + s

w
(k) + p(k)
184
F.2 Derivation of the Model for Parameter Estimation
in which the monthly fuel price index is again the linear combination of the
prices of three marginal fuels, as,
p(k) = p
coal
(k) + p
gas
(k) + p
diesel
(k)
where, p
coal
(k) =
T
k

t
k
p
coal
(t), p
gas
(k) =
T
k

t
k
p
gas
(t), p
diesel
(k) =
T
k

t
k
p
diesel
(t).
For that the monthly intra-year model is the summation of a deterministic
function f
w
(k) that models the seasonal pattern and an Autoregressive(AR)
model x
wr
(k) that models the stochastic variations, namely,
s
w
(k) = x
wr
(k) +f
w
(k)
where f
w
(k) =

m=1,2

m
cos
_
2
12
m k
_
+

m=1,2

m
sin
_
2
12
m k
_
. The monthly
model price model thus becomes,
s(k) = x
wr
(k) +a +f
w
(k) + p(k)
Denoting g(k)

= a +f
w
(k) + p(k), the model is simplied as,
s(k) = x
wr
(k) +g(k)
As the AR model of order one x
wr
(k) is expressed as,
x
wr
(k) = x
wr
(k 1) +
w
(k)
The monthly price model, written in the form an AR(1) model with a
time-varying mean, nally, is,
s(k) = s(k 1) + [g(k) g(k 1)] +
w
(k)
where g(k) is a simple linear function, as g(k)

= a + f
w
(k) + p(k), in which
f
w
(k) =

m=1,2

m
cos
_
2
12
m k
_
+

m=1,2

m
sin
_
2
12
m k
_
, and p(k) = p
coal
(k)+
p
gas
(k) + p
diesel
(k).
To summarize, it is these coecients of the linear function g(k) and the pa-
rameters and of the AR(1) process that one needs to estimate statistically.
And the estimation will be carried out in the next section.
185
F. BUILD THE FUEL PRICE INDEX
F.3 Parameter Estimation
The data used for estimating the coecients and the parameters are the elec-
tricity spot price data, denoted as S(t), coal price data P
coal
(t), natural gas
price data P
gas
(t), and diesel oil price data P
diesel
(t), see Section 6.2. All the
data range from the year 2002 to 2009, the electricity spot price data are orig-
inally of time unit one hour, and the fuel prices are originally of time unit one
month, see Section 6.3; all the data are rst transformed into the logarithm
domain and then changed to be of time unit one month, and then denoted
respectively as, s(k), p
coal
(k), p
gas
(k), and p
diesel
(k).
The result of the model calibration is shown in Figure F.1, where the price
data s(k) is plotted together with the estimated linear function g(k) , as well
as the 95% condence intervals. Figure F.2 plots the estimated monthly sea-
sonal pattern f

w
(k) = a + f
w
(k). The estimated parameters are recorded in
Table F.1.
Parameter a
Values 0.1542 0.5882 0.1992 1.999 0.2067 0.1984
t-Stat 1.229 7.768 2.707 21.52 1.797
Table F.1: Parameters of the Fuel Price Index Model - and the param-
eters of the intra-year model. The data used from calibration dates from 2002
to 2009.
When building the price model for the PJM Market, Section 6.4, the data
used for calibrating the models are from 2005 to 2009. The Fuel Price Index
model are thus calibrated from the data ranging from 2005 to 2009. The
parameters are recorded in Table F.2.
Parameter a
Values 0.1903 0.6574 0.1111 2.07 0.18 0.1044
t-Stat 1.33 7.33 1.1 9.97 1.07
Table F.2: Parameters of the Fuel Price Index Model - The data used
from calibration dates from 2005 to 2009.
186
F.3 Parameter Estimation
02 03 04 05 06 07 08 09
2.5
3
3.5
4
4.5
5
Time (month)
l
o
g

U
S
$
/
M
W
h


actual
estimate
95%CI
Figure F.1: Parameter Estimation of the Fuel Price Index Model
J an Feb Mar Apr May J un J ul Aug Sep Oct Nov Dec
1.8
1.85
1.9
1.95
2
2.05
2.1
2.15
2.2
Time (month)
H
e
a
t

R
a
t
e

(
l
o
g

m
m
B
t
u
/
M
W
h
)


a
a +f
w
(k)
Figure F.2: The Monthly Intra-year Seasonal Pattern - the estimated
deterministic seasonal pattern f

w
(t).
187
F. BUILD THE FUEL PRICE INDEX
188
References
Aid, R., Campi, L., Huu, A.N. & Touzi, N. (June 2009). A structural
risk-neutral model of electricity prices. 44
Allegheny Energy Inc. (2009). Annual reports.
http://www.alleghenyenergy.com. 117
American Electric Power Co., Inc. (2009). Annual reports.
http://www.aep.com/. 117
Anderson, C.L. & Davison, M. (2008). A hybrid system-econometric
model for electricity spot prices: Considering spike sensitivity to forced
outage distributions. Power Systems, IEEE Transactions on, 23, 927937,
0885-8950. 42, 79, 137
Angelus, A. (2001). Electricity price forecasting in deregulated markets.
The Electricity Journal , 14, 3241, 1040-6190 doi: DOI: 10.1016/S1040-
6190(01)00184-1. 15
Barlow, M.T. (2002). A diusion model for electricity prices. Mathematical
Finance, 287298. 42, 79, 137
Barlow, M.T., Gusev, Y. & Lai, M. (2004). Calibration of multifac-
tor models in electricity markets. International Journal of Theoretical and
Applied Finance, 7, 101120. 35, 37, 39
Bastian, J., Zhu, J., Banunarayanan, V. & Mukerji, R. (1999). Fore-
casting energy prices in a competitive market. Computer Applications in
Power, IEEE, 12, 4045. 15
189
REFERENCES
Baughman, M.L. & Lee, W.W. (1992). A monte carlo model for calculating
spot market prices of electricity. Power Systems, IEEE Transactions on, 7,
584590, 0885-8950. 15
Benth, F.E., Kallsen, J. & Meyer-Brandis, T. (2007). A non-gaussian
ornstein-uhlenbeck process for electricity spot price modeling and derivatives
pricing. Applied Mathematical Finance, 14, 153 169. 35, 37, 41
Bessembinder, H. & Lemmon, M.L. (Jun. 2002). Equilibrium pricing and
optimal hedging in electricity forward markets. The Journal of Finance, 57,
13471382. 15
Bhanot, K. (2000). Behavior of power prices: Implications for the valuation
and hedging of nancial contracts. Journal of Risk, 2(2000), 4362. 18, 35,
37
Bierbrauer, M., Truck, S. & Weron, R. (2004). Modeling electricity
prices with regime switching models. In Computational Science - ICCS 2004,
859867, Springer Berlin / Heidelberg. 35, 37, 40
Bierbrauer, M., Menn, C., Rachev, S.T. & Truck, S. (2007). Spot and
derivative pricing in the eex power market. Journal of Banking & Finance,
31, 34623485, 0378-4266 doi: DOI: 10.1016/j.jbankn.2007.04.011. 32, 33,
35, 37
Billinton, R. (1970). Power System Reliability Evaluation. Gordon and
Breach, Science Publishers, New York. 145
Billinton, R. & Allan, R.N. (1983). Reliability Evaluation of Engineering
Systems: Concepts and Techniquesp. Longman Scientic & Technical. 142
Blanco, C. & Soronow, D. (2001a). Jump diusion processes: Energy
price processes used for derivatives pricing & risk management. Commodities
Now. 35
Blanco, C. & Soronow, D. (2001b). Mean reverting processes: Energy
price processes used for derivatives pricing & risk management. Commodities
Now. 37, 40
190
REFERENCES
Boogert, A. & Dupont, D. (2008). When supply meets demand: The case
of hourly spot electricity prices. Power Systems, IEEE Transactions on, 23,
389398, 0885-8950. 42, 79, 137
Borovkova, S. & Permana, F. (2006). Modelling electricity prices by the
potential jump-diusion. In Stochastic Finance, 239263, Springer US. 32,
33, 35, 37
Branger, N., Reichmann, O. & Wobben, M. (2009). Pricing electricity
derivatives. 31, 32, 33, 35, 37
Burger, M., Klar, B., Muller, A. & Schindlmayr, G. (2004). A spot
market model for pricing derivatives in electricity markets. 42, 137
Burger, M., Graeber, B. & Schindlmayr, G. (2007). Managing Energy
Risk. John Wiley & Sons Ltd, West Sussex, England. 42, 51, 79, 137
Bystrom, H.N.E. (2005). Extreme value theory and extremely large electric-
ity price changes. International Review of Economics & Finance, 14, 4155.
41
Cartea, A. & Figueroa, M.G. (2005). Pricing in electricity markets: A
mean reverting jump diusion model with seasonality. 33, 35, 37, 40
Cartea, A. & Villaplana, P. (2008). Spot price modeling and the val-
uation of electricity forward contracts: The role of demand and capac-
ity. Journal of Banking & Finance, 32, 25022519, 0378-4266 doi: DOI:
10.1016/j.jbankn.2008.04.006. 42, 79, 137
Cartea, A., Figueroa, M.G. & Geman, H. (2008). Modelling electric-
ity prices with forward looking capacity constraints. Applied Mathematical
Finance, forthcoming. 41
Chan, K.F., Gray, P. & van Campen, B. (2008). A new ap-
proach to characterizing and forecasting electricity price volatility. In-
ternational Journal of Forecasting, 24, 728743, 0169-2070 doi: DOI:
10.1016/j.ijforecast.2008.08.002. 31, 41
191
REFERENCES
Clewlow, L. & Strickland, C. (2000). Energy Derivatives: Pricing and
Risk Management. Lacima Publications. 35, 38
Clewlow, L., Strickland, C. & Kaminski, V. (2001). Extending mean-
reversion jump diusion. Energy Power Risk Management. 37, 40
Conejo, A.J., Contreras, J., Espinola, R. & Plazas, M.A. (2005a).
Forecasting electricity prices for a day-ahead pool-based electric energy mar-
ket. International Journal of Forecasting, 21, 435. 21, 24, 47, 79
Conejo, A.J., Plazas, M.A., Espinola, R. & Molina, A.B. (2005b).
Day-ahead electricity price forecasting using the wavelet transform and
arima models. Power Systems, IEEE Transactions on, 20, 10351042, 0885-
8950. 24, 47
Contreras, J., Espinola, R., Nogales, F.J. & Conejo, A.J. (2003).
Arima models to predict next-day electricity prices. Power Systems, IEEE
Transactions on, 18, 10141020, 0885-8950. 24, 79
Crespo Cuaresma, J., Hlouskova, J., Kossmeier, S. & Obersteiner,
M. (2004). Forecasting electricity spot-prices using linear univariate time-
series models. Applied Energy, 77, 87106. 20, 23, 79
Culot, M., Goffin, V., Lawford, S., Menten, S.d. & Smeers, Y.
(2006). An ane jump diusion model for electricity. 31, 33, 35, 37, 39, 41,
80
Davison, M., Anderson, C.L., Marcus, B. & Anderson, K. (2002).
Development of a hybrid model for electrical power spot prices. Power Sys-
tems, IEEE Transactions on, 17, 257264, 0885-8950. 42, 137
De Jong, C. (2006). The nature of power spikes: A regime-switch approach.
Studies in Nonlinear Dynamics & Econometrics, 10. 32, 33, 35, 37, 40, 111
Deng, S. (1999). Financial Methods in Competitive Electricity Markets. Ph.D.
thesis, University of California at Berkeley, ph.D. Dissertation. 80
192
REFERENCES
Deng, S. (2000). Stochastic models of energy commodity prices and their
applications: Mean-reversion with jumps and spikes. (POWER) Working
Paper. 35, 40, 43
Diko, P., Lawford, S. & Limpens, V. (2006). Risk premia in electricity
forward prices. Studies in Nonlinear Dynamics & Econometrics, 10, Article
7. 35, 37, 39, 51, 80
Dixit, A.K. & Pindyck, R.S. (1994). Investment under Uncertainty.
Princeton University Press, Princeton, New Jersey. 80, 171
Emery, G.W. & Liu, Q. (2002). An analysis of the relationship between
electricity and natural-gas futures prices. Journal of Futures Markets, 22,
95122, (Wilson). 43
Energy Information Administration (2009a). Mas-
sachusetts natural gas price sold to electric power consumers.
http://tonto.eia.doe.gov/dnav/ng/hist/n3045ma3m.htm. 82
Energy Information Administration (2009b). New
york harbor no 2 diesel low sulfur spot price fob.
http://tonto.eia.doe.gov/dnav/pet/hist/rdlnyhm.htm. 117
Energy Information Administration (2009c). Pennsyl-
vania natural gas price sold to electric power consumers.
http://tonto.eia.doe.gov/dnav/ng/hist/n3045pa3m.htm. 117
Escribano, A., Pena, J.I. & Villaplana, P. (2002). Modeling electricity
prices: International evidence. 18, 28, 32, 33, 35, 37, 79
Franses, P.H. (1998). Time series models for business and economic fore-
casting. Cambridge University Press, Cambridge. 15
Garcia, R.C., Contreras, J., van Akkeren, M. & Garcia, J.B.C.
(2005). A garch forecasting model to predict day-ahead electricity prices.
Power Systems, IEEE Transactions on, 20, 867874, 0885-8950. 20, 21, 24,
28, 79
193
REFERENCES
Garcia-Martos, C., Rodriguez, J. & Sanchez, M.J. (2007). Mixed
models for short-run forecasting of electricity prices: Application for the
spanish market. Power Systems, IEEE Transactions on, 22, 544552, 0885-
8950. 23
Geman, H. & Roncoroni, A. (2006). Understanding the ne structure of
electricity prices. Journal of Business, 79, 12251261. 33, 35, 37, 40
Grigg, C., Wong, P., Albrecht, P., Allan, R., Bhavaraju, M.,
Billinton, R., Chen, Q., Fong, C., Haddad, S., Kuruganty, S.,
Li, W., Mukerji, R., Patton, D., Rau, N., Reppen, D., Schneider,
A., Shahidehpour, M. & Singh, C. (1999). The ieee reliability test
system-1996. a report prepared by the reliability test system task force of
the application of probability methods subcommittee. Power Systems, IEEE
Transactions on, 14, 10101020, 0885-8950. 167
Guthrie, G. & Videbeck, S. (2007). Electricity spot price dynamics: Be-
yond nancial models. Energy Policy, 35, 56145621, 0301-4215 doi: DOI:
10.1016/j.enpol.2007.05.032. 23, 51
Hadsell, L. & Shawky, H.A. (2006). Electricity price volatility and the
marginal cost of congestion: An empirical study of peak hours on the nyiso
market, 2001-2004. The Energy Journal , 27, 157179. 28
Hadsell, L., Marathe, A. & Shawky, H.A. (2004). Estimating the
volatility of wholesale electricity spot prices in the us. The Energy Jour-
nal , 25, 2340. 28
Hambly, B., Howison, S. & Kluge, T. (2009). Modelling spikes and
pricing swing options in electricity markets. 35, 37, 41
Hamilton, J.D. (1994). Time Series Analysis. Princeton University Press,
Princeton, New Jersey. 15
Hamm, G. & Borison, A. (2006). Forecasting long-run electric-
ity prices. The Electricity Journal , 19, 4757, 1040-6190 doi: DOI:
10.1016/j.tej.2006.07.003. 18, 79
194
REFERENCES
Higgs, H. & Worthington, A.C. (2005). Systematic features of high-
frequency volatility in australian electricity markets: Intraday patterns, in-
formation arrival and calendar eects. The Energy Journal , 26, 2341. 28
Hou, Y. & Wu, F.F. (2008). Valuation of generator prot from spot market:
Analytical approach. submitted to IEEE Trans. on Power Systems. 2
Howison, S. & Coulon, M.C. (2009). Stochastic behaviour of the electric-
ity bid stack: From fundamental drivers to power prices. The Journal of
Energy Markets, 2. 42, 44, 80, 137
Huisman, R. (2008). The inuence of temperature on spike probability in
day-ahead power prices. Energy Economics, 30, 26972704, 0140-9883 doi:
DOI: 10.1016/j.eneco.2008.05.007. 41
Huisman, R. & Mahieu, R. (2003). Regime jumps in electricity prices.
Energy Economics, 25, 425434. 26, 32, 35, 38, 40, 111
Huisman, R., Huurman, C. & Mahieu, R. (2007). Hourly electricity prices
in day-ahead markets. Energy Economics, 29, 240248, 0140-9883 doi: DOI:
10.1016/j.eneco.2006.08.005. 23
Hull, J.C. (2006). Options, Futures, and Other Derivatives. Prentice Hall,
Upper Saddle River, New Jersey, 6th edn. 80, 173
Huurman, C., Ravazzolo, F. & Zhou, C. (2008). The power of weather.
some empirical evidence on predicting day-ahead power prices through
weather forecasts. 20, 21, 24
ISO New England Inc. (2007). Annual markets report. http://www.iso-
ne.com/markets/mkt anlys rpts/annl mkt rpts/2007/amr07 nal 20080606.pdf.
81
ISO New England Inc. (2009a). Annual maintenance schedule.
http://www.iso-ne.com/genrtion resrcs/ann mnt sched/index.html. 81
ISO New England Inc. (2009b). Hourly zonal information. http://www.iso-
ne.com/markets/hstdata/znl info/hourly/index.html. 81, 82
195
REFERENCES
Janssen, M. & Wobben, M. (2009). Electricity and natural gas pricing. 43
Johnson, B. & Barz, G. (1999). Selecting stochastic process for modelling
electricity prices. In Energy Modelling and the Management of Uncertainty,
RiskPublications. 31, 35, 40, 80
Joskow, P. & Kohn, E. (2002). A quantitative analysis of pricing behavior
in californias wholesale electricity market during summer 2000. The Energy
Journal , 23, 136. 40
Kaminski, V. (1997). The challenge of pricing and risk managing electricity
derivatives. In The US Power Market, 149171, Risk Publications. 40, 43,
80
Kanamura, T. & Ohashi, K. (2007). A structural model for electricity
prices with spikes: Measurement of spike risk and optimal policies for hy-
dropower plant operation. Energy Economics, 29, 10101032, 0140-9883 doi:
DOI: 10.1016/j.eneco.2006.05.012. 42, 79, 137, 138
Karakatsani, N.V. & Bunn, D.W. (2004). Modelling stochastic volatility
in high-frequency spot electricity prices. Tech. rep., London Business School.
19, 28
Kim, C.i., Yu, I.K. & Song, Y.H. (2002). Prediction of system marginal
price of electricity using wavelet transform analysis. Energy Conversion
and Management, 43, 18391851, 0196-8904 doi: DOI: 10.1016/S0196-
8904(01)00127-3. 47
Knittel, C.R. & Roberts, M.R. (2005). An empirical examination of
restructured electricity prices. Energy Economics, 27, 791817. 18, 21, 28,
29, 31, 33, 35, 37, 79
Koekebakker, S. & Ollmar, F. (2005). Forward curve dynamics in the
nordic electricity market. Managerial Finance, 31, 73 94. 39, 80
196
REFERENCES
Kosater, P. & Mosler, K. (2006). Can markov regime-switching
models improve power-price forecasts? evidence from german daily
power prices. Applied Energy, 83, 943958, 0306-2619 doi: DOI:
10.1016/j.apenergy.2005.10.007. 32, 33, 35, 38, 41
Kosecki, R. (1999). Fuel-based power price modeling. In Energy Modeling
and the Management of Uncertainty, Risk Publications. 43, 80
Kresen, K.F. & Husby, E. (2000). A joint state-space model for electricity
spot and futures prices. Tech. Rep. Report No. 965, Norwegian Computing
Center. 33, 39, 80
Lay, D.C. (1997). Linear algebra and its applications. Addison-Wesley, 2nd
edn. 48
Li, Y. & Flynn, P.C. (2004a). Deregulated power prices: comparison
of diurnal patterns. Energy Policy, 32, 657672, 0301-4215 doi: DOI:
10.1016/S0301-4215(02)00331-2. 31
Li, Y. & Flynn, P.C. (2004b). Deregulated power prices: compari-
son of volatility. Energy Policy, 32, 15911601, 0301-4215 doi: DOI:
10.1016/S0301-4215(03)00130-7. 28
Longstaff, F.A. & Wang, A.W. (2004). Electricity forward prices: A
high-frequency empirical analysis. The Journal of Finance, 59, 18771900.
31
Lucia, J.J. & Schwartz, E.S. (2002). Electricity prices and power deriva-
tives: Evidence from the nordic power exchange. Review of Derivatives Re-
search, 5, 550, 10.1023/A:1013846631785. 32, 33, 35, 37, 39, 80, 95
Mankiw, N.G. (2004). Principles of Economics. Thomson South-Western,
third edition edn. 53
Manoliu, M. & Tompaidis, S. (2002). Energy futures prices: term structure
models with kalman lter estimation. Applied Mathematical Finance, 9. 38
197
REFERENCES
Maor, E. (1998). Fouriers theorem. In Trigonometric Delights, Princeton
University Press, Princeton. 44, 47, 80
Mari, C. (2008). Modeling power prices in competitive markets. Nonlinear
Phenomena in Complex Systems, 11, 215 224. 32, 33, 35, 38
Meyer-Brandis, T. & Tankov, P. (2008). Multi-factor jump-diusion
models of electricity prices. International Journal of Theoretical and Ap-
plied Finance, 11, 503528. 32, 33, 35, 38, 41
Misiorek, A., Trueck, S. & Weron, R. (2006). Point and interval fore-
casting of spot electricity prices: Linear vs. non-linear time series models.
Studies in Nonlinear Dynamics & Econometrics, 10. 18, 19
Monitoring Analytics (2009). Pjm state of the market.
http://www.monitoringanalytics.com/reports/PJM State of the Market/2009.shtml.
114
Mount, T.D., Ning, Y. & Cai, X. (2006). Predicting price spikes in elec-
tricity markets using a regime-switching model with time-varying parame-
ters. Energy Economics, 28, 6280. 26, 35, 38, 41, 111
Nakamura, M., Nakashima, T. & Niimura, T. (2006). Electric-
ity markets volatility: estimates, regularities and risk management
applications. Energy Policy, 34, 17361749, 0301-4215 doi: DOI:
10.1016/j.enpol.2004.12.019. 28
Nogales, F.J. & Conejo, A.J. (April 2006). Electricity price forecasting
through transfer function models. Journal of the Operational Research So-
ciety, 57, 350356(7). 20, 21
Nogales, F.J., Contreras, J., Conejo, A.J. & Espinola, R. (2002).
Forecasting next-day electricity prices by time series models. Power Systems,
IEEE Transactions on, 17, 342348, 0885-8950. 21, 79
Nomikos, N.K. & Soldatos, O. (2008). Using ane jump diusion models
for modelling and pricing electricity derivatives. 32, 33, 35, 38, 41
198
REFERENCES
Olsina, F., Garces, F. & Haubrich, H.J. (2006). Modeling long-term
dynamics of electricity markets. Energy Policy, 34, 14111433, 0301-4215
doi: DOI: 10.1016/j.enpol.2004.11.003. 15, 79
Pedregal, D.J. & Trapero, J.R. (2007). Electricity prices fore-
casting by automatic dynamic harmonic regression models. Energy
Conversion and Management, 48, 17101719, 0196-8904 doi: DOI:
10.1016/j.enconman.2006.11.004. 45, 79
Pilipovic, D. (1998). Energy risk: valuing and managing energy derivatives.
McGraw-Hill. 33, 35, 39, 80
Pirrong, C. & Jermakyan, M. (2008). The price of power: The valuation
of power and weather derivatives. Journal of Banking & Finance, 32, 2520
2529, 0378-4266 doi: DOI: 10.1016/j.jbankn.2008.04.007. 43
PJM Interconnection (2009a). http://www.pjm.com. 114
PJM Interconnection (2009b). Day-ahead lmp data.
http://www.pjm.com/markets-and-operations/energy/day-
ahead/lmpda.aspx. 118
PJM Interconnection (2009c). Historical pjm forecasted generation outage
postings. http://www.pjm.com/markets-and-operations/ops-analysis.aspx.
114
PJM Interconnection (2009d). Hourly load data.
http://www.pjm.com/markets-and-operations/energy/real-
time/loadhryr.aspx. 114
PJM Interconnection (2009e). Marginal fuel type data.
http://www.pjm.com/markets-and-operations/energy/real-time/historical-
bid-data/marg-fuel-type-data.aspx. 114
PJM Interconnection (2009f). Pjm daily generation capacity.
http://www.pjm.com/markets-and-operations/energy/real-time/historical-
bid-data/gen-unavail.aspx. 114
199
REFERENCES
Rambharat, B.R., Anthony, E.B. & Duane, J.S. (2005). A threshold
autoregressive model for wholesale electricity prices. 18, 19
Robinson, T. & Baniak, A. (2002). The volatility of prices in the english
and welsh electricity pool. Applied Economics, 34, 14871495. 28
Rose, K. (2007). The impact of fuel costs on electric power prices. Consul-
tancy report under contract with american public power association, Insti-
tute of Public Utilities, Michigan State University. 43
Ruibal, C.M. & Mazumdar, M. (2008). Forecasting the mean and the
variance of electricity prices in deregulated markets. Power Systems, IEEE
Transactions on, 23, 2532, 0885-8950. 15
Samuelson, P.A. (1980). Economics. McGraw-Hill Book Company, eleventh
edition edn. 53
Schmidt, T. (2008). Modelling energy markets with extreme spikes. In Math-
ematical Control Theory and Finance, 359375, Springer Berlin Heidelberg.
35, 38
Schwartz, E.S. (1997). The stochastic behavior of commodity prices: Im-
plications for valuation and hedging. The Journal of Finance, 52, 923973.
35
Schwartz, E.S. & Smith, J.E. (2000). Short-term variations and long-term
dynamics in commodity prices. 35, 38
Seifert, J. & Uhrig-Homburg, M. (2007). Modelling jumps in electricity
prices: theory and empirical evidence. Review of Derivatives Research, 10,
5985, 10.1007/s11147-007-9011-9. 32, 33, 35, 38, 111
Serna, G. & Villaplana, P. (2007). Modelling higher moments of electric-
ity prices. 18, 28
Simonsen, I. (2005). Volatility of power markets. Physica A: Statisti-
cal Mechanics and its Applications, 355, 1020, 0378-4371 doi: DOI:
10.1016/j.physa.2005.02.062. 28
200
REFERENCES
Skantze, P.L. & Ilic, M.D. (2001). Valuation, Hedging and Speculation
in Competitive Electricity Markets. Kluwers Power Electronics and Power
Systems Series, Kluwer Academic Publishers. 31, 32, 33, 35, 38, 39, 42, 51,
79, 80, 136
Strang, G. & Nguyen, T. (1996). Wavelets and Filter Banks. Wellesley-
Cambridge Press, Wellesley, Massachusetts. 45, 179
Swider, D.J. & Weber, C. (2007). Extended arma models for estimating
price developments on day-ahead electricity markets. Electric Power Systems
Research, 77, 583593, 0378-7796 doi: DOI: 10.1016/j.epsr.2006.05.013. 20,
21, 26, 28
Truck, S., Weron, R. & Wolff, R. (2007). Outlier treatment and robust
approaches for modeling electricity spot prices. 32, 33, 41, 47
Villaplana, P. (2004). Pricing power derivatives: A two-factor jump-
diusion approach. 35, 38, 40
Vollbrecht, H. (2008). An intraday spotmarket-price model based on clus-
tering. 31, 33
Wang, L. & Mazumdar, M. (2007). Using a system model to decompose the
eects of inuential factors on locational marginal prices. Power Systems,
IEEE Transactions on, 22, 14561465, 0885-8950. 15
Wasik, J.F. (2006). The Merchant of Power. Palgrave Macmillan, New York.
1
Wei, W.W. (2006). Time Series Analysis: Univariate and Multivariate Meth-
ods. Pearson Education, 2nd edn. 15
Weron, R. (2006). Modeling and Forecasting Electricity Loads and Prices.
John Wiley & Sons, Chichester, England. 15, 79
Weron, R. & Misiorek, A. (2005). Forecasting spot electricity prices with
time series models. Econometrics. 20, 21
201
REFERENCES
Weron, R. & Misiorek, A. (2008). Forecasting spot electricity prices:
A comparison of parametric and semiparametric time series models. In-
ternational Journal of Forecasting, 24, 744763, 0169-2070 doi: DOI:
10.1016/j.ijforecast.2008.08.004. 18, 19
Weron, R., Bierbrauer, M. & Truck, S. (2004). Modeling electricity
prices: jump diusion and regime switching. Physica A, 336, 3948. 26
Wilkinson, L. & Winsen, J. (2002). What we can learn from a statistical
analysis of electricity prices in new south wales. The Electricity Journal , 15,
6069, 1040-6190 doi: DOI: 10.1016/S1040-6190(02)00276-2. 31
Wood, A.J. & Wollenberg, B.F. (1996). Power Generation, Operation
and Control . John Wiley & Sons, Inc, New York. 2
Worthington, A., Kay-Spratley, A. & Higgs, H. (2005). Transmis-
sion of prices and price volatility in australian electricity spot markets: a
multivariate garch analysis. Energy Economics, 27, 337350, 0140-9883 doi:
DOI: 10.1016/j.eneco.2003.11.002. 28
Wu, F.F., Su, J., Zhou, H. & Hou, Y. (2008). Valuation of generator
prot in spot market: Simulation approach. submitted to IEEE Trans. on
Power Systems. 78
Zachmann, G. (2007). A markov switching model of the merit order to com-
pare british and german price formation. 44
Zareipour, H., Bhattacharya, K. & Canizares, C.A. (2007). Electric-
ity market price volatility: The case of ontario. Energy Policy, 35, 4739
4748, 0301-4215 doi: DOI: 10.1016/j.enpol.2007.04.006. 28
202

Potrebbero piacerti anche