A Real Options Valuation of Renewable Energy Projects PDF

A REAL OPTIONS VALUATION OF RENEWABLE ENERGY PROJECTS
(Thesis format: Monograph)
by
Natasha Burke
Graduate Program in Applied Mathematics
A thesis submitted in partial fulfillment

of the requirements for the degree of
Doctor of Philosophy
The School of Graduate and Postdoctoral Studies

The University of Western Ontario
London, Ontario, Canada
c Natasha Burke 2012
THE UNIVERSITY OF WESTERN ONTARIO

School of Graduate and Postdoctoral Studies
CERTIFICATE OF EXAMINATION
Examiners:
Supervisor:
.....................
Dr. Michel Denault
.....................
Dr. Matt Davison
.....................
Dr. David Bellhouse
Co-Supervisor:
.....................
Dr. David Jeffrey
.....................
Dr. Lindsay Anderson
.....................
Dr. Mark Reesor
The thesis by
Natasha Burke
entitled:
A Real Options Valuation of Renewable Energy Projects
is accepted in partial fulfillment of the
requirements for the degree of
Doctor of Philosophy
...............
Date
..............................
Chair of the Thesis Examination Board
ii
Abstract
Due to climate change concerns, high oil prices and nuclear dangers there is increasing support
for renewable energy. At the forefront of the debate for government support of renewable
energy are wind energy and biofuels. Used primarily for power generation and transportation,
respectively, there have been many debates surrounding the reliability and efficiency of these
resources. These debates often address the uncertainty in the economic value of the resource
through time, however it is often difficult to quantify this uncertainty, which stems from the
random behavior of prices and the unpredictable nature of the resource itself.
In this thesis we use well developed theory taken from quantitative finance, more specifically real options theory, as well as various mathematical and statistical techniques and models
used in option pricing to determine the economic value of these resources. Market design and
policy are key considerations throughout the analysis. A simplified model of a corn ethanol
plant is analyzed using a simple Margrabe exchange option as well as numerical techniques
such as bootstrapping and finite difference methods for solving partial differential equations. It
is determined that, as correlation between corn price and gasoline price increases, the value of
the ethanol plant decreases. The level of decrease is substantial, and the economic and political
consequences are discussed.
A simplified wind-storage model is also developed and analyzed as a dynamic program,
allowing for analytic solutions. This allows for the determination of an optimal bidding strategy
when the penalty for failing to meet a committed generation level is defined. The market
consequences of these results are discussed and compared with numerical results obtained
from a more complex wind-storage model where analytical solutions are not available. These
results are obtained using a modified block bootstrapping method, and the optimal storage size
is determined for a specific penalty structure.
Keywords: Renewable Energy, Corn Ethanol, Wind Power, Energy Storage, Real Options,
Optimal Control, Dynamic Programming, Bootstrapping, Partial Differential Equations
iii
Contents
Certificate of Examination
ii
Abstract
iii
List of Figures
vii
List of Tables
ix
.
.
.
.
1
1
3
4
5
Introduction
1.1 Corn Ethanol as Renewable Energy
1.2 Wind as a Renewable Energy Source
1.3 Modeling Choices . . . . . . . . . .
1.4 Thesis Road Map . . . . . . . . . .
Real Options
2.1 Margrabe Options . . . . . . .
2.2 Options vs. Real Options . . .
2.3 Mathematical Background . .
2.4 A Simple Real Options PDE .
2.5 Real Options and Renewables
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6
7
8
9
10
11
Corn Ethanol Plant Valuation

3.1 Spark Spread Options . . . . . . .
3.2 Retrospective Analysis . . . . . .
3.3 Corn and Gasoline Price Dynamics
3.4 Bootstrap Analysis . . . . . . . .
3.4.1 Adding Correlation . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
12
12
14
14
22
30
.
.
.
.
.
.
33
33
34
36
42
42
53
.
.
.
.
.
Real Options PDEs

4.1 The Exchange Option PDE . . . .
4.1.1 Two Dimensional Problem
4.1.2 ADI . . . . . . . . . . . .
4.1.3 Simulating GBM . . . . .
4.2 Transforming to a 1D Problem . .
4.3 Summary . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
iv
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5 Wind and Storage

5.1 Dynamic Programming Approach to Optimal Control Problem
5.1.1 Dynamic Programming . . . . . . . . . . . . . . . . .
5.1.2 Assumptions . . . . . . . . . . . . . . . . . . . . . .
5.1.3 One Period Model . . . . . . . . . . . . . . . . . . .
5.1.4 Extending to N time periods . . . . . . . . . . . . . .
5.2 Discussion and Summary of Results . . . . . . . . . . . . . .
6 Wind and Storage - An In Depth Analysis
6.1 Part One: Wind . . . . . . . . . . . . . .
6.2 Statistical Properties of Wind Speed Data
6.2.1 Historical Data . . . . . . . . . .
6.2.2 Distribution of data . . . . . . . .
6.2.3 Autocorrelation . . . . . . . . . .
6.3 Missing Data . . . . . . . . . . . . . . .
6.4 Bootstrapping . . . . . . . . . . . . . . .
6.5 Summary . . . . . . . . . . . . . . . . .
6.6 Part Two: Storage . . . . . . . . . . . . .
6.6.1 Storage Options . . . . . . . . . .
6.6.2 Battery Storage . . . . . . . . . .
6.7 Modeling a Battery used for Storage . . .
6.7.1 Battery Size . . . . . . . . . . . .
6.8 Summary . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7 Wind/Storage System Valuation

7.1 The Simple Model Revisited . . . . . . . .
7.2 A More Realistic Model . . . . . . . . . .
7.3 Forecasting Wind Speed . . . . . . . . . .
7.4 Power Prices . . . . . . . . . . . . . . . .
7.5 Battery Charge/Discharge Heuristics . . . .
7.5.1 Strategy 1 - Naive Strategy . . . . .
7.5.2 Strategy 2 - Utilizing Capacity . . .
7.6 Strategy Comparison as Parameters Change
7.7 Next Steps . . . . . . . . . . . . . . . . . .
8 Market Design
8.1 Capacity Markets . . . . . . . . . . . . . .
8.2 Penalties for Forecasting Error . . . . . . .
8.2.1 Penalty functions . . . . . . . . . .
8.2.2 Current design in European Markets
8.2.3 A simple penalty function . . . . .
8.3 Summary and Discussion . . . . . . . . . .
v
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
54
55
55
55
56
58
70
.
.
.
.
.
.
.
.
.
.
.
.
.
.
72
72
73
76
76
77
82
84
86
86
87
88
89
91
91
.
.
.
.
.
.
.
.
.
92
92
93
97
97
99
99
100
100
106
.
.
.
.
.
.
107
107
108
108
109
109
120
Discussion and Conclusions

123
9.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
9.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
A Finite Difference Methods

A.0.1 A Review of Finite Difference Methods
A.0.2 Implicit vs. Explicit . . . . . . . . . .
A.1 Crank Nicolson . . . . . . . . . . . . . . . . .
A.2 ADI . . . . . . . . . . . . . . . . . . . . . . .
A.2.1 Transforming the Real Options PDE . .
vi
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
132
133
133
133
134
135
List of Figures
2.1
The Margrabe parameter, , and the option value, M , as a function of . . . .
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
Corn and Gasoline prices, January 1976 December 2006 . . . . . . . . .

Historical Corn and Gas Prices with Resultant Plant Operation Pattern . . .
Fourier Analysis - Corn (January 1997-December 2006 data) . . . . . . . .
Fourier Analysis - Gasoline (January 1997-December 2006 data) . . . . . .
R(i) vs. P (i) (Corn), 1997-2006 monthly data . . . . . . . . . . . . . . . .
R(i) vs. P (i) (Gasoline), 1997-2006 monthly data . . . . . . . . . . . . . .
Scatter Plot of Corn Returns using 1997-2006 price data . . . . . . . . . .
Scatter Plot of Gasoline Returns using 1997-2006 price data . . . . . . . .
Bootstrapped Corn and Gas Prices with Resultant Plant Operation Pattern
acceptable simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bootstrapped Corn and Gas Prices with Resultant Plant Operation Pattern
unacceptable simulation . . . . . . . . . . . . . . . . . . . . . . . . . .
Average Plant values, 1997-2006 monthly data used for bootstrapping . . .
Frequency of Plant Losses, 1997-2006 monthly data used for bootstrapping
Average Plant Value with Increasing Correlation . . . . . . . . . . . . . . .
Number of Months Plant is Running with Increasing Correlation . . . . . .
3.10
3.11
3.12
3.13
3.14
4.1
4.2
4.3
4.4
4.5
4.6
4.7
4.8
4.9
4.10
4.11
4.12
4.13
4.14
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8
15
16
18
19
20
21
23
24
. 26
.
.
.
.
.
27
28
29
31
32
Histogram of Log Returns - Corn Data . . . . . . . . . . . . . . . . . . . . . .

Histogram of Log Returns - Gasoline Data . . . . . . . . . . . . . . . . . . . .
QQ plot - Corn Data, i.e. log(Ck+1 /Ck ) against the standard normal distribution
QQ plot - Gasoline Data, i.e. log(Gk+1 /Gk ) against the standard normal distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2D PDE surface plot corresponding to = 0 . . . . . . . . . . . . . . . . . . .
Lease Value with increasing for C0 = 3,G0 = 2 . . . . . . . . . . . . . . . .
Plant Value with increasing for C0 = 2,G0 = 3 . . . . . . . . . . . . . . . .
QQ plot - Difference of HG(t) C(t) time series, i.e. H(Gk+1 Gk ) (Ck+1
Ck ) against the standard normal distribution . . . . . . . . . . . . . . . . . . .
1D PDE surface plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Surface plot cross section, V (Q, 1000) . . . . . . . . . . . . . . . . . . . . . .
of the Spark Spread Time Series as increases . . . . . . . . . . . . . . . .
Ethanol plant value with varying . . . . . . . . . . . . . . . . . . . . . . . .
37
38
39
vii
40
41
43
44
45
46
47
49
50
51
52
6.1
6.2
6.3
6.4
6.5
6.6
6.7
6.8
6.9
6.10
6.11
Power Curve - Vestas 1.8 MW Turbine [38] . . . . . . . . . . . . . . . . . . .

Hourly Wind Speed Data - Nantucket Sound, 2003 [77] . . . . . . . . . . . . .
Hourly Wind Speed Data - The Netherlands, 2003 [43] . . . . . . . . . . . . .
Frequency of wind speed and Rayleigh pdf - Dutch Data, 2003 [43] . . . . . .
QQ plot of 2003 Dutch data (X) vs. a Rayleigh distribution (Y ) [43] . . . . . .
Frequency of wind speed and Rayleigh pdf - US Data, 2003 [77] . . . . . . . .
QQ plot of 2003 Nantucket Sound data (X) vs. a Rayleigh distribution (Y ) [77]
Autocorrelation vs Hourly Time Shift - Dutch Data, 2003 [43] . . . . . . . . .
Autocorrelation vs Hourly Time Shift - US data, 2003 [77] . . . . . . . . . . .
Nantucket Wind Speed Data, 2003 (location of missing data shown in red) . . .
System Ratings, Installed systems as of November 2008 [29] . . . . . . . . . .
73
74
75
78
79
80
81
82
83
85
88
7.1
7.2
7.3
7.4
7.5
7.6
7.7
V (F, k) and V (E, k) as k increases: p = 0.39, = 0.2, M = 0.04 . . . . . .

V (F, k) and V (E, k) for 0 x p: p = 0.39, = 0.2, M = 0.04, k = 100
V (F, k) and V (E, k) for p x p/(1 p): p = 0.39, = 0.2, M = 0.04
Mean Electricity Price at Each Hour of the Day (hour zero is 12am) [77] . . . .
Battery Cycle Life vs. Depth of Discharge . . . . . . . . . . . . . . . . . . . .
Battery Cycle Life vs. Depth of Discharge when DOD 50% . . . . . . . . . .
Strategy 1: Losses as and increase: = 0.1, = 1000, = 1, = min =
0.1, min = 0, = $200/kWh, = $5.50/kWh . . . . . . . . . . . . . . . . .
94
95
96
98
101
102
8.1
8.2
8.3
8.4
8.5
Total Penalty vs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Battery Cost vs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Annual Battery Cost vs . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Revenue Ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Net Gain/Loss as a function of for various penalty levels x: = 0.1, =
1000, = 1, = min = 0.1, min = 0, = $200/kWh, = $5.50/kWh . .
8.6 Net Gain/Loss as a function of for various penalty levels x: = 0.1, =
1000, = 1, = min = 0.1, min = 0, = $200/kWh, = $5.50/kWh . .
8.7 lower vs x: = 0.1, = 1000, = 1, = min = 0.1, min = 0, =
$200/kWh, = $5.50/kWh . . . . . . . . . . . . . . . . . . . . . . . . . .
8.8 upper vs x: = 0.1, = 1000, = 1, = min = 0.1, min = 0, =
$200/kWh, = $5.50/kWh . . . . . . . . . . . . . . . . . . . . . . . . . .
8.9 max vs x: = 0.1, = 1000, = 1, = min = 0.1, min = 0, =
$200/kWh, = $5.50/kWh . . . . . . . . . . . . . . . . . . . . . . . . . .
8.10 max vs x: = 0.1, = 1000, = 1, = min = 0.1, min = 0, =
$200/kWh, = $5.50/kWh . . . . . . . . . . . . . . . . . . . . . . . . . .
8.11 Net Cash as a function of : = 0.1, = 1000, = 1, = min = 0.1,
min = 0, = $200/kWh, = $5.50/kWh . . . . . . . . . . . . . . . . . . .
viii
.
.
.
.
105
111
112
113
114
. 115
. 116
. 117
. 118
. 119
. 121
. 122
List of Tables
3.1
3.2
3.3
Average Annual Spark Spread . . . . . . . . . . . . . . . . . . . . . . . . . . 17

Probability of k highs assuming a random draw . . . . . . . . . . . . . . . . . 17
Frequency of high/low prices . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.1
Historical GBM Parameter Estimates . . . . . . . . . . . . . . . . . . . . . . . 34
5.1
5.2
Benefit Matrix: Empty Battery . . . . . . . . . . . . . . . . . . . . . . . . . . 56

Benefit Matrix: Full Battery . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.1
6.2
Wind Variables and Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 89

Battery Model Variables and Parameters . . . . . . . . . . . . . . . . . . . . . 90
7.1
Number of Windy Days and p estimate under various windy day assumptions
(2003 Nantucket Sound data) . . . . . . . . . . . . . . . . . . . . . . . . . .
Cycles to Failure Data (Yang et. al) . . . . . . . . . . . . . . . . . . . . . .
Strategy 1: Loss Attribution - Charging ( = 10000), Historical Analysis . . .
Strategy 1: Loss Attribution - Discharging ( = 10000), Historical Analysis .
Strategy 1: Loss Attribution - Discharging ( = 1000), Historical Analysis . .
Strategy 2: Loss Attribution - Discharging ( = 10000), Historical Analysis .
7.2
7.3
7.4
7.5
7.6
7.7
7.8
ix
.
.
.
.
.
.
.
.
93
103
103
104
104
104
105
106
Chapter 1
Introduction
A great deal of effort is being put into researching and developing renewable energy technologies. Renewable energy can be generated from wind, biomass, sunlight, tides and flowing
water. The primary reason for this effort stems from the desire to reduce greenhouse gas emissions from non-renewable energy sources such as fossil fuels. Other environmental impacts of
fossil fuels are also a factor, such as nitrogen and sulfur oxides from coal (NOx and SOx) as
well as oil spills similar to the recent major spill in the Gulf of Mexico. Developing renewable
technologies will also serve to increase energy security by reducing North Americas reliability
on foreign imports of fossil fuels. Nuclear technology relies on a radioactive source such as
uranium. The potential risks associated with nuclear energy continue to be a topic of debate,
especially following the tragic nuclear disaster at the Fukushima Daiichi power plant in Japan
in 2011.
Due to their reliance on the sun, renewable energy sources are often variable. The resulting uncertainty in predicting the power availability of a given source at a given time causes
scheduling difficulties. There is also uncertainty in the economic value of the energy source
at a given time; this uncertainty stems from the random behavior of prices associated with the
energy source. Incorporating these uncertainties into economic models of renewable energy
sources undoubtedly leads to better decision making and planning.
In this thesis, we study two renewable energy sources: corn ethanol and wind power coupled with a storage technology. The sources of uncertainty are different for each resource,
however at the highest level of abstraction the underlying problems are the same; in the face
of uncertainty, how can we accurately determine the economic value of these resources? To
answer this question we rely on well developed theory used in quantitative finance as well as
various mathematical and statistical techniques and models used in options pricing. Before
we attempt to answer this and other related questions, a brief overview of each resource is
provided.
1.1 Corn Ethanol as Renewable Energy

In recent years it has become both politically and, to a lesser extent, environmentally popular
to produce fuel ethanol from corn. In 20022003, ethanol accounted for 9% of corn usage
in the U.S, and in four years this number increased to 18%, or 54% of nonfeed usage [19].
1
C HAPTER 1. I NTRODUCTION
In the same year the third largest corn crop ever was seen. It is predicted that by 2016 corn
ethanol will use over 35% of the total corn produced in the US [31]. It is important to note
that corn ethanol is different from sugar cane ethanol which is used in Brazil (for information
on sugar cane ethanol, see [15]). While the production costs of corn ethanol exceed those of
sugar cane ethanol, when transportation costs are considered the cost benefit is often negligible
or even nonexistent. However due to recent US mandates to produce advanced biofuel (i.e.
non-starch biofuel), sugar cane might be a viable option [20].
The process of converting corn into ethanol has received some criticism. One criticism that
has been advanced is that the process is energy negative that more energy is used to grow
the corn and convert it into ethanol than is produced in the process (see [41, 57, 59, 70]). We
do not enter that particular debate here except insofar as to note that corn ethanol production
must receive government subsidies to be economically attractive. Another criticism is that
converting corn production from food or feed use to fuel use is responsible for an increase in
food prices with concomitant ill effects for the worlds poor [59].
We do not suppose a deterministic relationship between corn ethanol production and food
prices in this study, but rather consider a corn ethanol plant which faces uncertainty in corn
(input) and gasoline (output) prices. Moreover, we consider the possibility that widespread
ethanol production will cause the price of gasoline and the price of corn, which historically
have been nearly uncorrelated, to become more highly correlated in the future. Our goal is
to develop a framework for determining the economic value of a corn ethanol plant given the
uncertainty in prices and future price correlations.
The framework we choose is that of real options. Real options allow us to capture the value
of operational flexibility based on the uncertainty we face in corn and gas prices. Real options
is the focus of Chapter 2. The model is developed in Chapter 3, and results and conclusions
based on statistical methods are presented. In Chapter 4 a PDE approach is discussed and the
resulting PDEs are solved using numerical methods.
The idea of using real options analysis to better understand the effect of changing economic conditions on investment decisions is certainly not new; in [61] some simple models
are reviewed which illustrate two important characteristics of investment expenditures: they
are largely irreversible, and they can be delayed in order to obtain new market information.
It also touches on the importance of public policy implications, which is clearly crucial here
given the current corn ethanol debate. One of the most common pieces of literature regarding
real options analysis is [27], a motivator for the PDE portion of this thesis. The idea of real
options valuation applied to natural resource investments was introduced in [13], and in particular the application to offshore petroleum leases is discussed in [55]. Most recently, in [68],
the influence of ethanol policy in the U.S. on plant investment decisions was discussed using a
real options approach. However, in this paper revenue and costs were the stochastic variables
modeled rather then the prices themselves. Closed form solutions are sought, and optimal exit
and entry triggers are determined. The model presented in Chapter 3, and discussed in [42], is
actually listed in this paper and is hence an example of related work which was motivated by
our study.
The idea of corn-gasoline correlation, in particular the impact of this correlation on plant
valuation and policy, has not been not explored in the previous real options paper on ethanol
production. In fact the spread between corn and gasoline prices, as well as the conversion
factor which allows for the analysis of this spread, is not common in the literature on ethanol
1.2. W IND AS A R ENEWABLE E NERGY S OURCE
plant valuation in general. This conversion factor, which is crucial to the model and analysis,
is discussed and calculated in Chapter 3.
1.2
Wind as a Renewable Energy Source
Although wind power is a popular renewable energy source, it is intermittent; when the wind
doesnt blow, or if it blows too hard, power cannot be produced [11]. Since perfect forecasts
for wind are currently unattainable, there is uncertainty in predicting the power availability
from wind at a given time, which causes scheduling difficulties leading to reliability problems
in power systems [3]. In addition, in many geoclimatic regions peak wind speeds tend to occur
during periods of low power demand.
Intermittency is mostly an issue only when the system is near capacity; when the system is
well below capacity there is no intrinsic problem with an intermittent resource, as long as the
resource is predictable. Hydro is an example of an intermittent but possibly quite predictable
resource [36], for instance spring freshet and dry season in Alberta. Wind resources are mostly
unpredictably intermittent. Capacity issues especially become important as wind penetration
rises ([16, 74]). The unpredictability issue is generally more severe day to day as the identical
amount of power must be generated as is required by consumers. Shortfalls and, to a lesser
extent, over-delivery of power is therefore a significant problem.
Adaptable power sources which are able to rapidly respond to central commands are one
way to deal with the unpredictable resource problem. Historically this has been provided by
coal plants. Due to anti greenhouse gas initiatives, as well as anti N Ox and SOx initiatives, it
is suggested that these will be decommissioned [71], at least in Ontario, Canada and possibly
in other jurisdictions. Gas plants provide the necessary quick response, but at a cost and with
emissions. In addition, as wind penetration increases, the demand for these adaptable power
sources will increase and eventually saturate the current capacity of natural gas plants. This
will also not solve the problem of over generation (wasting wind energy by spilling it).
To improve the usability of wind, something must be done about this. One idea is to improve wind forecasts [3]. To a limited extent this is possible, but it is costly, and to take advantage of improved forecasts requires some tools such as storage, flexible imports, or responsive
demand resources. In the absence of improved forecasts the ability to store energy would also
provide significant assistance in the management of variability. In a regulated market the regulator can assign costs to various participants as it sees fit. In a deregulated market, who will pay
for the enhanced forecasts and/or storage? Who will operate the storage? Some might argue
that the wind producers themselves should bear the costs of their resources unpredictability.
How should market rules be designed to encourage this? What different technologies and technological advances in storage and wind forecasting best facilitate this? Clearly all of these
questions are interconnected.
Punitive market rules for over- or under- delivering scheduled wind power will encourage
wind producers to invest in better forecasts or more storage, while lenient tariff agreements will
force the market regulator to find other ways to solve the problem of wind variability. Beyond
this rather obvious statement, the exact details of market fees and penalties will encourage a
rational response by wind producers. This rational response may or may not promote better
predictability. For example, if producers only get penalized for under delivery they might sim-
C HAPTER 1. I NTRODUCTION
ply consistently under promise, which imposes the cost of dealing with surplus power on the
regulator. The relative effectiveness and cost of improvements in forecasts vs. the technical
storage technology details and cost will, for a given regulatory structure, dictate the relative
investment in the two areas.
In Chapter 5 we introduce a simplified system of an energy storage unit with a wind generator. The solution to the optimal operating strategy is analyzed as a dynamic program and
its simplicity allows us to gain valuable insight in determining the added value of storage. In
Chapter 6 we perform an in depth analysis of wind and storage, carefully laying out the many
parameters necessary to properly model a storage facility combined with a wind turbine. In
Chapter 7 we consider a wind producer who has invested in battery storage technology. We
present a model for a sodium sulfur battery, the most economically feasible battery in current
use. Given a set of battery parameters, we explore a variety of operating strategies and compare
them in terms of the level of predictability they are able to provide. In Chapter 8 we explore
the level of penalty which can be imposed on wind producers before they are discouraged from
participating in the market. In a similar way to which the ethanol section accounts for uncertain
prices, our battery storage model accounts for uncertainty in wind speed. Many of the techniques used in Chapter 3 will be revisited and modified to suit the model discussed in Chapter
8.
1.3 Modeling Choices

The complexity of the above problems is extremely high; this is not only true from a mathematical standpoint, but also in terms of the physical systems and even more so the underlying
policy issues. Mathematically, we proceed as follows: the first step is to perform a thorough
statistical analysis on the relevant data. In the corn ethanol piece of this study, the relevant data
is historical corn and gasoline prices. In the wind portion of this study it is wind speed and, to
a lesser extent, electricity prices. This analysis is crucial in developing modeling assumptions,
as well as in understanding the relevance of the results arising from the data.
Without first understanding the physical system driving the data it is impossible to develop
reasonable models to value the overall projects. This is especially true when modeling wind
power coupled with storage. The data analysis is useful in understanding the properties of
wind, but further research must be done to understand the storage technology. In the case of a
battery, the vast number of technologies, coupled with their underlying properties, limitations
and costs, must be understood in order to properly choose and model the battery.
As is the case with renewable energy in general, the public policy issues surrounding these
projects are unavoidable. In the case of corn ethanol, the major policy issue to consider is
subsidies associated with ethanol production. In the case of wind, the policy pieces are market
factors such as penalties associated with an unpredictable energy source. The impact of these
issues are evident as they are integrated into the models discussed in future chapters.
1.4. T HESIS ROAD M AP
1.4
Thesis Road Map
Since this thesis is essentially two self-contained units the reader may wish to begin with, or
focus on, one of the two units. Chapter 2 provides a brief financial and mathematical background on real options which encompasses the entire document. Chapters 3 and 4 focus on the
ethanol study only. If the reader would prefer to begin with the wind portion of the thesis, it
would be appropriate to proceed directly to Chapter 5 from Chapter 2.
Chapter 2
Real Options
A real option is the right, but not the obligation, to undertake, expand, extend or abandon a
certain project. Real options based analysis captures flexibility in future operations where the
value of the operation is uncertain, and allows for decision making under this uncertainty [27].
Traditional valuation techniques such as the method of discounted cash flows do not capture
this flexibility or uncertainty in future values; all future cash flows are simply estimated and
their present value computed. If we subtract the initial investment from these cash flows we
have calculated the net present value (NPV). If the net present value is positive the decision is
to go ahead with the project; if it is negative the decision is to reject the project. When using
this approach to value a project, the only modeling variable that we have at our disposal is
the discount rate. When valuing the same project using a real options framework, decisions
are frequently updated based on the state of the uncertain variables (such as prices) at various
points throughout the life of the project. At some points the project may realize little or no
value, at which point the decision to temporarily suspend the project may be made. However,
the project retains some value due to its potential to operate at a profit at a later date. Because
of this, the real options value will be larger than the NPV due to the lack of future flexibility
present in the NPV approach.
A simple example of a real option is a spark spread option [25]. The name is derived from
electricity generation where one would spark up the natural gas input and look at the spread
between the gas price and the electricity price to determine whether or not to run a particular
generation asset at a particular time. The option has payoff
Vt = max(Et
HGt , 0),
where Et is the electricity price in $/MWh at time t, Gt is the gas price in $/MMBtu at t and
H is the heat rate, which is defined as the amount of energy, in MMBtu, required to generate
one megawatt hour (MWh) of electricity. For a fixed time t the real option looks similar to a
European call option with pay-off
V = max(S
K, 0)
where S is the stock price and K is the strike price.

The related concept of the crack spread is used in oil markets; in this case the spread
between the price of crude oil and the petroleum products extracted from it is examined. This
allows the refinery to determine the expected profit by cracking crude oil (i.e. breaking down
its hydrocarbons) to make refined products such as gasoline [34].
6
2.1. M ARGRABE O PTIONS
2.1
Margrabe Options
A Margrabe exchange option is a contract which gives the owner the right, but not the obligation, to exchange b units of a traded asset S1 for a units of another traded asset S2 at maturity
T [49]. The payoff is therefore
(aS1 (T ) bS2 (T ))+ .
(2.1.1)
The underlying assumptions of the pricing model are as follows: assume S1 and S2 follow
geometric Brownian motion, i.e.,
dS1 = 1 S1 dt + 1 S1 dW1
dS2 = 1 S2 dt + 2 S2 dW2 ,
(2.1.2)
(2.1.3)
where W1 and W2 are correlated Wiener processes with correlation , i.e. dW1 dW2 = dt.
The price M of this option satisfies the following PDE:
M
1
2M
1 2 2 2M
M
M
+ S2 S2
+ S2 1 S12
+
S2 S2
+ S1 S1
2
2
t
2
S1
2
S2
S1
S2
2
M
+S1 S2 S1 S2
rM = 0
S1 S2
(2.1.4)
with final condition

M (S1 , S2 , T ) = max(aS1 (T )
bS2 (T ), 0).
Consider a situation where we want to continuously make the decision of whether or not
to run a plant. We can consider this problem a strip of exchange options, each having a payoff
similar to (2.1.1). The advantage of this consideration is that there is a formula for valuing
exchange options; the formula is obtained from a dynamic hedging, no arbitrage type argument.
The formula is given below (assume a = b = 1).
M = S1 e(1 r) N (d1 )
S2 e(2 r) N (d2 )
(2.1.5)
where
d1 =
ln (S1 /S2 ) + (1 2 + 2 /2)
d2 = d1 ,
p
= 12 + 22 21 2
and is the time until expiry. It is assumed that the interest rate r is constant; the same is true
for 1 and 2 , the drifts corresponding to S1 and S2 , respectively. The volatilities of S1 and S2
are given by 1 and 2 and N is the standard normal cumulative distribution function.
This formula can be used to build intuition regarding the effect of the correlation : it is
clear from the above equation for (the volatility of S1 /S2 ) that as increases, decreases.
If we look at the corresponding effect on equation (2.1.5) it is easy to see that as increases,
C HAPTER 2. R EAL O PTIONS
M decreases. This is shown in Figure 2.1 for an initial fixed price pair, in particular a corn
and gasoline price pair. Parameter values are as follows: 1 =0.1803, 2 =0.3252, 1 =-0.0265,
2 =0.0125, r = 0.04, and S1 (0) = S2 (0) = 2.
Since the exchange happens just once (i.e. at time T ), this valuation is not sufficient for
the problems we encounter in this thesis. At each time t we wish to examine the difference
between the price of two assets to make a decision. In particular for the ethanol problem,
we want to examine the difference between corn and gasoline prices. Thus, we rely on other
existing tools to value a corn ethanol plant which allow us to continuously make decisions, but
use this Margrabe result to build intuition about the end result.
0.55
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
1
0.5
0.5
Figure 2.1: The Margrabe parameter, , and the option value, M , as a function of
2.2 Options vs. Real Options

At the beginning of the chapter we gave a definition of a real option which looks similar to
the definition of a financial option, i.e. the right, but not the obligation, to buy or sell an asset
for some agreed strike price on or before the expiration date. We also saw from the spark
spread and Margrabe illustrations that the payoff of a real option resembles the payoff of a
financial option. These similarities exist beyond the simple purchase and sale of assets in many
real life situations involving decision making given an unknown future. Decisions should be
2.3. M ATHEMATICAL BACKGROUND
made optimally, and while financial options theory tells us how to optimally invest in a certain
financial product, real options extends this idea to other life decisions and investments.
Investment decisions exhibiting the following three factors can benefit from real options
analysis:
there is uncertainty in the future, and therefore uncertainty in the outcome of an investment decision
there is a degree of irreversibility associated with an investment; that is, if you change
your mind after the decision has been made, you cannot recover all of your initial investment
there is some flexibility regarding the timing of the investment; it may be better to wait
until more is known about the future before making the investment [27].
The combination of these characteristics can be difficult to consider using more traditional
investment decision making tools, and hence there is potentially large option value lost when
making decisions using such tools.
The mathematics used in real options theory is similar to the mathematics used in financial
options theory, but there are differences which will be seen in the following section.
2.3 Mathematical Background

Many real options partial differential equations (PDEs) are very similar to the Black Scholes
PDE used in valuing financial options,
V
1 2 2 2V
V
+ S
+
rS
rV = 0,
t
2
s2
S
where the underlying asset price S follows geometric Brownian motion,
dS = Sdt + SdW .
Here, dW is an increment of a Wiener process, i.e. a process which satisfies the following
conditions:
W (0) = 0
the process has independent increments
Wt
Ws N (0, t s), t > s
the sample paths are continuous

The model contains two parameters, which is the drift and which is the volatility.
In finance, geometric Brownian motion is a continuous time stochastic process where the
logarithm of the process follows a Brownian motion. Since a Brownian motion has independent
increments which are normally distributed, and a common assumption in finance is that the log
of stock returns are normally distributed, it is a very suitable assumption in financial modeling.
10
C HAPTER 2. R EAL O PTIONS
One major difference between the Black Scholes PDE and a real options PDE is that the
Black Scholes PDE does not contain the growth parameter . The reason for this lies within
the assumptions governing the PDE; the Black Scholes PDE uses a risk neutrality argument
which computes the expected option payoff under the risk neutral measure rather than the real
world measure. This results from the assumption that a hedged portfolio can be created which
is composed of one option and delta shares of the underlying asset. The risk characteristics
are unchanged throughout the life of the option, hence a risk free rate is applied as the discount
rate at each point in time, yielding a risk neutral valuation. In real options the underlying asset
and the option may not be tradeable, and even if the underlying asset is tradeable, its owner
(or operator) might not want to hedge it, but rather extract value from > r. Moreover, the
actions of the decision makers can cause changes in the risk of the project throughout its life,
eliminating the risk neutral assumption.
2.4 A Simple Real Options PDE

Assume Q evolves randomly according to a Brownian motion with drift, i.e.,
dQ(t) = dt + dW.
(2.4.1)
In many real options problems Q could denote the instantaneous profit at any time t. Therefore,
Q can be negative. The parameters and are problem dependent, and can be constant (with
being positive or negative) or functions of Q and t. It makes sense to restrict > 0, although
since dX and dX are simply reflected processes this is not essential in theory.
A common real options problem deals with the ability to costlessly start and stop production
at a plant, depending on market conditions. Let Q denote profit at time t, where Q evolves
according to (2.4.1); if Q > 0 we run the plant and if Q < 0 we shut down. The value of the
plant, V (Q(t), t), can be viewed as the expectation (conditional on information known up to
time t) of V (Q(t) + dQ(t), t + dt) + Q(t)H (Q(t)), where H (Q(t)) is the Heaviside function,
discounted at some rate r. We use the first two terms of the Taylor expansion of erdt as the
discount factor. In other words,
V (Q(t), t) = (1 rdt)Et [V (Q(t) + dQ(t), t + dt) + Q(t)H (Q(t))].
We can use Itos Lemma, the stochastic calculus analogue to the chain rule,
1 2 2V
V
V
V
+
+
dX,
dV =
dt
+
t
2 Q2
Q
Q
(2.4.2)
(2.4.3)
and the fact that E[dX] = 0 to obtain

V
1 2V
V
+ 2 2 +
t
2 Q
Q
rV + QH (Q) = 0.
(2.4.4)
This is similar to the PDE obtained in [76]. Final and boundary conditions are problem dependent and will be discussed for a model ethanol plant in a later chapter.
2.5. R EAL O PTIONS AND R ENEWABLES
2.5
11
Real Options and Renewables
It was stated earlier that when valuing renewable energy projects, there is uncertainty stemming
from the random behavior of prices associated with the energy source itself. When considering
corn ethanol as a renewable energy source, the sources of uncertainty are future corn and
gasoline prices. Therefore, it may not always be optimal to run the plant at a loss, but rather
to use the flexibility of being able to shut down and restart the plant. Using real options theory
we can determine when it is optimal to run the plant as well as the value of the plant given a
set of ethanol plant parameters.
When considering wind as a renewable energy source, the sources of uncertainty are future
wind speeds as well as future electricity prices. When a storage system is considered the
uncertainty remains the same; although the storage system is in place to make the energy source
more predictable, there is still uncertainty in how much wind we will see at a given hour, as
well as uncertainty in the price of electricity at a given hour. We can use real options in this
setting to determine how to best operate our storage so that we maximize the predictability of
wind in the face of uncertain wind speeds and prices.
Chapter 3
Real Options Valuation of a Corn Ethanol
plant
As discussed in Chapter 1, we consider the possibility that widespread ethanol production will
cause the price of gasoline and the price of corn, which historically have been nearly uncorrelated, to become more highly correlated in the future. We present calculations which suggest
that possible future correlations between corn and gas prices may impact the level of public
subsidy required in order to prevent the value of ethanol plants from being destroyed. To show
this, we model an ethanol plant as a real option on the appropriate weighted spread between
gasoline (output) and corn (input) prices. We use the real options framework because recent
historical experience suggests that ethanol plants may not always be economically advantageous to operate even with a subsidy; however in this case the plant itself still retains value
because of the ability it represents to profitably run in the future. The resulting model may be
computed using a variety of techniques including a bootstrap-based simulation. With the aid
of this model we are able to quantify the value destroyed, for a realistic but simplified model
ethanol plant, by a given level of future corn-gasoline correlation. We draw public policy
conclusions from this result.
In Chapter 2 a simple Margrabe option valuation showed that ethanol plant value loss is
inevitable. Having considered this simplified approach, we introduce the idea of a spark spread
option and derive the formula for valuing an ethanol plant using this framework. Initially we
use historical spot price data to apply the formulation. However, to be more precise and obtain
more accurate results we explore corn and gasoline price dynamics in order to generate price
sequences for both commodities using a bootstrapping technique. These price sequences are
generated with a specified level of correlation between the commodity prices, and the effects
of increasing correlation are reported. The results of this Chapter are also given in [42].
3.1 Spark Spread Options

Spread options are extremely valuable decision making tools in situations where the difference
between the prices of two assets determines the economic value of an operation. As mentioned
in Chapter 2, spark spread options are used in electricity generation where the spread between
the price of fuel input and electricity output is examined in order to decide whether or not to
12
3.1. S PARK S PREAD O PTIONS
13
run a particular generation asset.

Ethanol producers can use spark spreads in a similar way. They can examine the appropriate
spread between corn (input) and gasoline (output) to determine whether or not they will be
profitable at a given time. If the spread becomes negative, they have a decision to make:
should the plant run at a loss, or should the plant temporarily shut down and resume production
when the spread becomes positive again? If the latter approach is taken, the situation is similar
to that of an option where the owner of a contract chooses whether or not to exercise their right
to buy or sell at a given time. We therefore value the operation as a calllike option where, at
each time t, the payoff is
t = (HGt Ct )+ .
(3.1.1)
Here, is the profit per unit of gasoline, Gt is the spot price of gasoline (dollars/gallon), Ct
is the spot price of corn (dollars/bushel) and H is the conversion factor from gasoline to corn
units (gallons/bushel). Thus, the plant is run when
H gasoline price > corn price,
otherwise it is unprofitable to do so and production should cease. This valuation method differs
greatly from the method of discounted cash flows, which is likely to undervalue the operation
since it does not capture its flexibility. Unlike a call option, both Gt and Ct are unknown for
t > 0.
We must compute a suitable range of H values to use in our simulations. A range of values
exist since the energy content of gasoline and corn is not constant. For instance, gasoline
blends may differ with batch, season and from one refinery to another. The energy content of
one bushel of shelled corn is given by = 448, 000 476, 000 British thermal units (Btus)
while the energy content of one gallon of gasoline is given by = 115, 000 125, 000 Btus.
One bushel of corn gives approximately = 0.8 bushels of shelled corn. Thus, we can compute
H to be in the range
H =
0.8
2.87
448, 000 476, 000

,
125, 000 115, 000
3.31.
Our range of H values are consistent with values used in [6, 72] and in our simulations we use
the midpoint, H = 3.09.
To add to the realism of this model, we also consider the fact that ethanol production is
highly subsidized. Thus, we add a constant term (in dollars/gallon) to the gasoline price, which
accounts for the volumetric tax credit given to ethanol producers. In other words, for each
gallon of ethanol produced at a given facility in the US, the facility receives a subsidy in the
form of a tax credit. This subsidy is around s = $0.60/gallon of ethanol produced [45]. We
must also account for the cost of running the plant. Some plants in the US use coal and even
methane from cow manure to run their plants, and one US ethanol plant uses excess steam
from a nearby power plant. However, most plants rely on natural gas, and the associated cost
is approximately p = $0.52/gallon of ethanol produced [6].
14
C HAPTER 3. C ORN E THANOL P LANT VALUATION

The model, which represents the annual value of the plant, is therefore given as
12
X
max(H(Gt + s p)
Ct , 0),
(3.1.2)
t=1
where is the number of bushels of corn consumed by the plant per month on average. Monthly
units are used since we are using average monthly corn and gasoline price data. The source for
this data is the CRB Commodity Yearbook (see [19]).
3.2 Retrospective Analysis

We first look at historical corn and gas prices as shown in Figure 3.1. Although Figure 1 depicts
31 years of historical data, we will concentrate on data between January 1997 and December
2006 since it more indicative of the current markets, especially in the case of gasoline.
Figure 3.2 uses (3.1.2) to show when the plant was profitable and when it operated at a loss
(top panel), on a monthly basis, between January 1997 and December 2006. A value of one
on the yaxis represents a profit during a particular month, and a value of zero represents a
loss. Corresponding monthly corn and gasoline spot price data, respectively, are shown in the
bottom panels. All prices are in 2006 dollars.
From this figure it is evident that during the last three years of data, under the assumptions
given in Section II, an ethanol plant should be operating at full capacity every month without
incurring any losses. This is due to a major increase in gasoline prices during those years without a corresponding large increase in corn prices. Assuming our future correlation assumption
holds, this situation will be highly unlikely as corn and gasoline markets will tend to go up and
down together. In fact, since 2006 corn prices have increased substantially so these last three
years of data may not be indicative of the current value of an ethanol plant. We look at how to
deal with this in the next section.
Table 3.1 shows the average annual value of the plant from 1997 until 2006. The overall
average annual value and the total number of months that the plant operated without incurring
loses, assuming (3.1.2) holds, are also given.
3.3 Corn and Gasoline Price Dynamics

Since past data is not indicative of the current situation, and may not be indicative of the
future situation, we can only consider it as one possible scenario. In order to examine multiple
scenarios, we must better understand the dynamics of corn and gasoline prices. To do this, we
examine the data for seasonality, mean reversion and serial autocorrelation.
We first wish to remove any seasonality which exists in the commodity prices. In order to
do this, we perform a Fourier Analysis on both data sets and look for a sine wave of period 12
that best fits the data (see [35] for details on this). From Figures 3.3 and 3.4 it is easily shown
that the best fit sine waves are not representative of the commodity data. To examine this even
further, we determine in which month corn prices are highest and lowest, and compute the
probabilities of k highs (or lows) in a given month assuming a completely random draw. The
results are given in Tables 3.2 and 3.3.
3.3. C ORN AND G ASOLINE P RICE DYNAMICS
15
4.5
Corn price
Gasoline price
4
3.5
Price
3
2.5
2
1.5
1
0.5
0
50
100
150
200
Month
250
300
350
Figure 3.1: Corn and Gasoline prices, January 1976 December 2006
Plant status
16
1
0.5
0
0
20
40
60
80
100
120
20
40
60
80
100
120
20
40
60
Month
80
100
120
Corn price
4
3
2
Gasoline price
1
3
2
1
0
Figure 3.2: Historical Corn and Gas Prices with Resultant Plant Operation Pattern

Year
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
Average Annual Plant Value
Number of Months plant is running
17
Plant Value ($ million)
$0
$0
$2.94
$13.48
$7.71
$2.95
$7.64
$17.67
$41.57
$47.75
$14.17
85 out of 120
Table 3.1: Average Annual Spark Spread
k
0
1
2
3
4
P(k highs)
0.4189
0.3808
0.1558
0.0378
0.006
Table 3.2: Probability of k highs assuming a random draw
Month
January
February
March
April
May
June
July
August
September
October
November
December
Number of highs
4
0
0
2
2
0
0
0
1
0
0
1
Number of lows
1
0
0
1
0
1
1
1
1
2
1
1
Table 3.3: Frequency of high/low prices
18
3.6
3.4
Corn Price (dollars/bushel)
3.2
3
2.8
2.6
2.4
2.2
2
1.8
1.6
20
40
60
Month
80
100
120
Figure 3.3: Fourier Analysis - Corn (January 1997-December 2006 data)
19
2.2
2
Gasoline Price (dollars/gallon)
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
20
40
60
Month
80
100
120
Figure 3.4: Fourier Analysis - Gasoline (January 1997-December 2006 data)
20
Although obtaining four highs in January is extremely unlikely (it is 0.6% likely to occur at
random), there is also a 7% chance that one of the 12 months will show up as an outlier. Given
this, combined with the apparent randomness of annual lows and the fact that the amplitude of
the sine wave is not reflective of the data range, we can conclude that the effects of seasonality
are at best weak in the corn data and are nonexistent in the gas data; in this study we ignore
seasonality entirely.
We also consider the possibility of mean reversion in the data. Figures 3.5 and 3.6 show
returns versus prices for both corn and gasoline. If there was significant mean reversion we
would expect the returns to be higher than average when prices are very low and lower than
average when prices are very high. This is difficult to detect from these figures, and determining
a reliable mean reversion parameter would be an even more difficult task, so we also neglect
mean reversion in this study.
0.15
0.1
return(i)
0.05
0.05
0.1
0.15
0.2
1.6
1.8
2.2
2.4
2.6 2.8
price(i)
3.2
3.4
Figure 3.5: R(i) vs. P (i) (Corn), 1997-2006 monthly data
21
0.3
0.2
return(i)
0.1
0.1
0.2
0.3
0.4
0.5
1.5
price(i)
Figure 3.6: R(i) vs. P (i) (Gasoline), 1997-2006 monthly data
2.5
22
The final check is for serial autocorrelation in the data. To inspect this, we plot log return
i + 1 versus log return i, shown in Figure 3.7. We find the line of best fit through the data is
y = 0.3271x + 0.0007157
(3.3.1)
with R2 = 0.1074. This value is an indication of how well the regression line approximates
the real data points. If we consider only the data from the past 10 years, the line of best fit is
y = 0.4311x + 0.0001564
(3.3.2)
with R2 = 0.1845. In both cases, this is a strong signal, suggesting that there is a linear
correlation in the corn data. Thus, we compute the correlation coefficient and Fisher z statistic
to confirm this. In general, the correlation coefficient between two data sets X and Y is defined
as
Cov(X, Y )
= p
.
(3.3.3)
V ar(X)V ar(Y )
Since in this case X and Y represent the same data, V ar(X) = V ar(Y ) and we compute
the correlation coefficient to be c = 0.4295. The Fisher z statistic transforms the correlation
coefficient into a value which, if the data is actually uncorrelated, is normally distributed with
mean zero and standard deviation one, and is given is
N 3
1+r
zF =
ln
.
(3.3.4)
2
1 r
This statistic is found to be zF = 4.9679, which means the probability that the log return data is
uncorrelated is extremely small and confirms that the corn data has significant autocorrelation.
We redo this procedure using the gasoline spot price data as well. Figure 3.8 shows return
i + 1 versus return i, and the line of best fit passing through this data is
y = 0.001115x + 0.0005547
(3.3.5)
with R2 = 1.676 106 . If we consider only the data from the past 10 years, the line of best
fit is
y = 0.03148x + 0.006919
(3.3.6)
with R2 = 0.0009948. In both cases the regression line is a poor approximation to the actual
spot data, and so it appears that the data is random. However, to confirm this we compute
the correlation coefficient and Fisher z statistic. We obtain g = 0.0315 and zF = 0.3408,
confirming that the data does not have significant autocorrelation.
3.4 Bootstrap Analysis

Now that we better understand the dynamics of corn and gasoline prices, we can generate
new sequences of price data, or bootstrap. We do this using the log returns from the past ten
years. There are many bootstrapping techniques available, however it is important to choose the
technique which bests suits the data. In this case the data does not exhibit obvious seasonality
3.4. B OOTSTRAP A NALYSIS
23
0.15
0.1
R(i+1)
0.05
0.05
0.1
0.15
0.2
0.2
0.15
0.1
0.05
0.05
0.1
0.15
R(i)
Figure 3.7: Scatter Plot of Corn Returns using 1997-2006 price data
24
0.3
0.2
R(i+1)
0.1
0.1
0.2
0.3
0.4
0.4
0.3
0.2
0.1
0.1
0.2
0.3
R(i)
Figure 3.8: Scatter Plot of Gasoline Returns using 1997-2006 price data
25
or mean reversion, but one of the two data sets exhibits a strong one period autocorrelation.
The procedure differs for each commodity due to the autocorrelation in the corn data. For
gasoline data, we sample (with replacement) from the log returns to generate a new sequence
of log returns. Using the average gasoline price from January 1997 as our initial price value,
we can obtain a new price sequence via
Gt+1 = Gt eRt ,
(3.4.1)
where Rt is determined by sampling from the original sequence of log returns, Rt0 = log(Gt+1 /Gt ).
For corn data we generate an innovation sequence, It , using
It+1 =
0
rt+1
rt0
1
(3.4.2)
where rt0 = log(Ct+1 /Ct ) and = c = 0.4295. From this sequence we can generate a new
sequence, Et , by sampling It with replacement. We can obtain a new price sequence using
rt+1 = rt + (1
)Et
(3.4.3)
and
Ct+1
= Ct ert ,
(3.4.4)
where r1 = E1 and C1 is the average corn price from January 1997.

The results are somewhat surprising. The average value of the plant is found to be approximately $42 million, and the plant is expected to run profitably 64 months out of 120. This
means that although the option value is zero for several realizations, other realizations give
such large option values that, on average, the plant is extremely profitable. However, a closer
look at the simulation values shows that, in several cases, corn and gas prices are unrealistically
high or low, and hence the average option value is not a good indication of actual plant value.
Thus, we place restrictions on the simulated price values: if corn or gasoline prices exceed a
certain value, or fall below a certain value, the simulation is not used. Note that this pruning of
unrealistic paths is in some ways similar to the assumption of mean reversion in the return data
mean reversion that was not observable in the data. Mean reversion is notoriously difficult to
measure in data sets which, like the ones used here, never move very far from their mean value.
Depending on the size of the restriction a large percentage of the simulations may be omitted.
For instance, when we use the restrictions 1 < Ct < 15 and 0.2 < Gt < 10, about 30% of the
simulations are excluded. Figures 3.9 and 3.10 show examples of acceptable bootstrapped
time series and unacceptable bootstrapped time series, respectively. Using the acceptable
time series, the plant runs profitably 72 months out of 120; using the unacceptable time series,
the plant runs profitably 14 months out of 120. This is since initially the corn price is relatively
low in comparison to the gas price; this allows the plant to run at a significant profit. However,
eventually the corn price increases above 15, at the same time as the gasoline price decreases,
and in this case the plant is never able to run profitably.
Using these ranges the average option value is approximately $25 million and the plant is
expected to run at a profit 61 months out of 120. Figures 3.11 and 3.12 are histograms showing
the range and frequency of option values for all included simulations as well as the frequency
of losses over the 120 month period.
Plant status
26
1
0.5
0
0
20
40
60
80
100
120
20
40
60
80
100
120
20
40
60
Month
80
100
120
Corn price
6
4
2
Gasoline price
0
6
4
2
0
Figure 3.9: Bootstrapped Corn and Gas Prices with Resultant Plant Operation Pattern - acceptable simulation
Plant status
27
1
0.5
0
0
20
40
60
80
100
120
20
40
60
80
100
120
20
40
60
Month
80
100
120
Corn price
30
20
10
Gasoline price
0
1.5
1
0.5
0
Figure 3.10: Bootstrapped Corn and Gas Prices with Resultant Plant Operation Pattern - unacceptable simulation
28
3.5
x 10
2.5
1.5
0.5
50
100
150
200
Figure 3.11: Average Plant values, 1997-2006 monthly data used for bootstrapping
29
600
500
400
300
200
100
20
40
60
80
100
120
Figure 3.12: Frequency of Plant Losses, 1997-2006 monthly data used for bootstrapping
30
3.4.1 Adding Correlation

The above technique does not include the possibility of future correlation between commodity
prices. To do this, suppose a bootstrapped sequence of corn and gasoline returns are denoted
by rt and Rt respectively. We induce correlation between these sequences by creating a new
gasoline return sequence Rt1 as follows:
Rt1 = [rt + (1
Rt
)Rt0 ] q
2 rt2
+ (1
)2 Rt2
(3.4.5)
Here Rt and rt are the average values of the bootstrapped log return sequences, and thus change
with each simulation. This means that future gasoline returns do not only depend on past
gasoline returns, but they also depend on the corn return in the same period. The amount by
which gasoline returns depend on corn returns is the correlation . We can use (3.4.1) and
(3.4.5) to obtain new price sequences where the gasoline price at time t depends on both the
corn and gasoline log returns from the past. Since we have no way of accurately determining
the future correlation of corn and gas prices, we use a variety of values to examine the effect
on the number of months that the plant is down, and the ultimate impact on the value of the
plant. The results are given in Figure 3.13. It is evident from this figure that even if corn
and gasoline prices become moderately correlated, the value of the plant is substantially lower
than when correlation is neglected. For instance, a 50% correlation would mean that profits
would be 10% less than if the correlation was zero. Moreover, as correlation increases, the
value monotonically decreases, and even more financial losses are expected as the correlation
approaches one. If the prices are perfectly correlated, profits would be 30% less than the
expected profit given a zero correlation assumption.
We also look at the number of months the plant is shut down (Figure 3.14) and notice
that it is generally increasing with increasing correlation. This means that, although the plant
is running more, when it is running it is making less profit. As correlation increases, corn
and gas prices will tend upward and downward together more often. Situations where corn
prices decrease substantially and gas prices increase substantially at the same time are now
very unlikely. Thus, the opportunity to take advantage of this large positive spread has been
eliminated, and ethanol producers see smaller positive spreads more often.
31
27
26
Value
25
24
23
22
21
20
1
0.5
0.5
Figure 3.13: Average Plant Value with Increasing Correlation
32
90
85
Months running
80
75
70
65
60
55
1
0.8 0.6 0.4 0.2
0.2
0.4
0.6
0.8
Figure 3.14: Number of Months Plant is Running with Increasing Correlation
Chapter 4
Valuing Ethanol Plants using Real
Options PDEs
It might seem reasonable to end the analysis given in Chapter 3 and simply draw conclusions
based on the given results. However, it is appropriate to look at the problem in a different way;
in particular, a way which does not so heavily rely on past data. More specifically, in Chapter
3 we studied historical corn and gasoline price data; some of this data dates back to 1997 when
corn ethanol was a new idea. We then induced correlation between the historical data sets,
creating one new data set based on the other. Using this new historical data, we simulated
our model in order to draw conclusions based on correlation. While there is certainly value in
this study, there is also value in the ability to solve problems analytically, or using well known
numerics. The value of analytic solutions was evident in referring to the Margrabe exchange
option, which allowed for a simple way to gain intuition on correlation effects without having
to run any simulations. Of course, the underlying model assumptions cannot be ignored, and
these will be further discussed in the sections that follow.
In this chapter we look at solving the problem by deriving partial differential equations to
satisfy assumptions regarding price paths and correlation. Finite difference methods are used
to solve the PDE, and the results provide an additional layer of confidence that our data driven
results are sound. A more basic review of finite difference methods is given in Appendix A.
4.1 The Exchange Option PDE

In Chapter 2 we introduced the well known Margrabe result for valuing an option to exchange
one asset for another, given in (2.1.5). In order to arrive at this result it was assumed that the
underlying asset prices S1 and S2 followed geometric Brownian motion, i.e.,
dS1 = 1 S1 dt + 1 S1 dW1
dS2 = 1 S2 dt + 2 S2 dW2 ,
(4.1.1)
(4.1.2)
where W1 and W2 are correlated Wiener processes with correlation . The PDE including final
condition was outlined.
The simplest proof of the Margrabe result relies on reducing the problem to one in which
the Black-Scholes formula can be applied; in other words, a simple change in numeraire is
33
34
C HAPTER 4. R EAL O PTIONS PDE S
performed which considers both assets to be priced in terms of S1 (S2 ) so that the original
option becomes a call (put) option with strike equal to one.
As noted in Chapter 2, the Margrabe formula is not sufficient for many real options problems since the exchange happens only once, at time T . Therefore, we must write down a more
complex 2D PDE which captures the ability to switch on and off the plant between t = 0 and
t = T.
4.1.1 Two Dimensional Problem

Combining the knowledge outlined in the previous section with the 1D real options PDE derived in Chapter 2 (given by equation (2.4.4)) we obtain the desired equation. The underlying
asset prices C(t) and G(t) follow correlated Geometric Brownian motion as in the Margrabe
model, i.e.,
dC = c Cdt + c CdW1
dG = g Gdt + g GdW2
(4.1.3)
(4.1.4)
and dW1 dW2 = dt. The values of c , g , c and g can be estimated using historical data.
The results are given in Table 4.1.
Parameter
c
g
c
g
Value
-0.0265
0.0125
0.1803
0.3252
Table 4.1: Historical GBM Parameter Estimates

The PDE which results is a two-dimensional FokkerPlanck equation with final and boundary conditions given as follows:
V
1 2 2 2V
1 2 2 2V
V
V
2V
+ c C
+ G
+ c C
+ g G
+ c g CG
t
2
C 2 2 g G2
C
G
CG
rV + (HG C)H (HG C) = 0,
(4.1.5)
V (C, G, T ) = 0
V (C, 0, t) = 0
lim V (C, G, t) = 0
HG (g r)(T t)
e
g r
HG (g r)(T t)
e
lim V (C, G, t) =
G
g r
V (0, G, t) =
C (c r)(T t)
e
c r
1 .
4.1. T HE E XCHANGE O PTION PDE
35
Final and boundary conditions depend upon assumptions. For instance, if we assume that
we are operating a leased facility, then at time T the value of the plant is zero since the facility is returned to its owner and no longer has value to the lease holder. The final condition,
V (C, G, T ) = 0 has been chosen to reflect this. We are not restricted to this condition, however
this is a simple yet realistic choice which can easily be checked.
The first two boundary conditions can be understood by examining the spread (HGt Ct )+ :
when Gt = 0, the spread will be identically zero for any Ct . Therefore, the plant will have zero
value in this case. Similarly, as Ct (where this limit can be interpreted as Ct large) the
spread will be zero for finite Gt . This boundary condition does not hold when HGt > Ct
which means that if a numerical solution is sought, the grid of corn and gasoline prices over
which the problem is solved must be considered.
The final two boundary conditions can best be explained by computing the expected value
of the plant when Ct = 0 as well as in the limiting case when Gt . Again, this limit should
be interpreted as what happens when Gt is large relative to Ct ; of course as Gt approaches
infinity so does HG C and hence so does V . These conditions require us to take the time t
conditional expectation of simple integrals as follows:
Z
V (0, G, t) = Et
r(st)
e
Z
t
T
= H
Z
HG(s)ds
er(st) Et [G(s)] ds
t
T
= H
e(g r)(st) G(t)ds
HG(t) (g r)(T t)
=
e
g r
and
Z
lim V (C, G, t) = Et
= H
T
= H
t
e
(HG(s) C(s))ds
Z T
r(st)
e
Et [G(s)] ds
er(st) Et [C(s)] ds
t
Z T
e(c r)(st) C(s)ds
e(g r)(st) G(s)ds
r(st)
HG(t) (g r)(T t)
=
e
1
g r
C(t) (c r)(T t)
e
c r
1 .
Equipped with this PDE we must answer two questions: are the underlying assumptions
accurate, and how can we solve it? Neither of these questions have simple answers. In Chapter
3 we examined corn and gasoline price dynamics. In order to generate new price paths for both
commodities we computed the log returns for both data sets and sampled from these returns
to obtain many new series of returns. Rather then sample from the log returns it would be
ideal to determine an approximate distribution for the log returns and benefit from the known
properties of this distribution. Whereas in Chapter 3 we only required serial independence of
36
the log returns (which was not the case, however it was possible to remove this dependence
and incorporate it back into the new time series), in this chapter we determine whether or not
the log returns are normal. We can check the validity of the normality of the log returns, and
hence the GBM assumption, by comparing a histogram of log returns for each data set to the
probability distribution function (pdf) for the normal distribution, given by
f (x; , ) =
1
2 2
(x )2
2 2 .
e
Here and are the mean and standard deviation of the appropriate data set. This is shown in
Figures 4.1 and 4.2 for corn and gasoline, respectively. From these figures alone it appears as
if neither of the data sets are normally distributed. Figures 4.3 and 4.4 provide us with another
method of determining whether or not the normal distribution is suitable for our data. These
figures, commonly called QQ plots, are quantile-quantile plots of the sample quantiles of our
return data versus theoretical quantiles from a Normal distribution. If the distribution of our
data is indeed normal then the plot will be close to linear. It is clear from these figures that a
normal distribution does not perfectly describe our data, but is it an acceptable approximation?
Perhaps it is best to look at solution methods to the PDE in order to decide if something
slightly simpler yet equally valid will suffice. A possible numerical solution to the PDE can be
determined using an Alternating Direction Implicit (ADI) method [58]. This method has the
desirable stability features of the Crank Nicolson method, but proceeds in two steps. The first
half step is taken implicitly in one space variable and explicitly in the other, while the second
half step reverses the explicit and implicit variables. Thus, we are able to reduce the problem
to solving two matrix equations (rather than one as in Crank Nicolson) where the systems are
tridiagonal.
4.1.2 ADI
Consider the 2-D heat equation,
ut = uxx + uyy .
In stage one of the ADI method we treat the discretization in the x-direction implicitly and take
a half step, i.e.
k+1/2
ui,j
k+1/2
uki,j
ui+1,j
=
t
k+1/2
2ui,j
h2
k+1/2
+ ui1,j
uki,j+1
2uki,j + uki,j1
h2
(4.1.6)
where i is the counter in x direction, j is the counter in the y direction, k is the time counter,
t is half a time step size and h is the x and y step size. The tridiagonal system resulting from
(4.1.6) is given in Appendix A. In stage two we take a half step implicitly in the y direction
and explicitly in the x direction, i.e.
uk+1
i,j
k+1/2
k+1/2
ui,j
t
k+1/2
ui,j+1 2ui,j
h2
k+1/2
+ ui,j1
uk+1
i+1,j
The resulting tridiagonal system is also given in Appendix A.
k+1
2uk+1
i,j + ui1,j
.
h2
(4.1.7)
37
35
30
25
20
15
10
0
0.2
0.15
0.1
0.05
0.05
0.1
Figure 4.1: Histogram of Log Returns - Corn Data
0.15
38
25
20
15
10
0
0.4
0.3
0.2
0.1
0.1
0.2
0.3
Figure 4.2: Histogram of Log Returns - Gasoline Data
0.4
39
QQ Plot of Sample Data versus Standard Normal

0.15
Quantiles of Input Sample
0.1
0.05
0.05
0.1
0.15
0.2
3
1
0
1
Standard Normal Quantiles
Figure 4.3: QQ plot - Corn Data, i.e. log(Ck+1 /Ck ) against the standard normal distribution
40

0.3
0.2
0.1
0.1
0.2
0.3
0.4
3
1
0
1
Figure 4.4: QQ plot - Gasoline Data, i.e. log(Gk+1 /Gk ) against the standard normal distribution
41
If we consider the simple case when = 0, (4.1.6) simplifies nicely to

1
1
V
V
V
2V
2V
+ c2 C 2 2 + g2 G2 2 + c C
+ g G
t
2
C
2
G
C
G
rV + (hG C)H (hG C) = 0.
(4.1.8)
This case can be solved fairly easily via ADI, and the solution is given in Figure 4.5.
25
V(C,G,t)
20
15
10
5
0
30
25
20
20
15
10
G(t)
10
0
5
0
C(t)
Figure 4.5: 2D PDE surface plot corresponding to = 0

This figure clearly highlights the boundary conditions; when the gasoline price is zero, the
value of the plant is zero regardless of the corn price. When the corn price is zero however, the
plant value increases linearly with increasing gasoline price. In addition this figure shows that
for a fixed non-zero corn price plant value resembles that of a call option; for a fixed non-zero
gas price plant value resembles that of a put option.
The PDE given in (4.1.6) is significantly more complex then the 2D heat equation, or even
(4.1.8). We are left with two choices: transform the PDE into something simpler (which will
42
eliminate the cross derivative term), or approximate the cross derivative term. The former
option is discussed in Appendix A, however neither solution leaves us with a well behaved
solution for 6
= 0. This leaves us exploring simpler solutions.
4.1.3 Simulating GBM

One potential solution is to, for a variety of initial corn and gasoline prices, simply use the
GBM assumptions and simulate price paths for corn and gasoline with embedded correlation
. Using these price paths we could compute the spread using (3.1.2).
To simulate price paths under the GBM assumption, we discretize as follows:
St+dt = St e(
2 /2)dt+
dtA(t)
where the A(t) are random samples from independent and identically distributed (iid) random
variables which are N (0, 1). To simulate correlated paths we generate a second random sample,
A2 (t), in the same way as A(t). We can then obtain
p
Ac (t) = A(t) + 1 2 A2 (t),
where A(t) and Ac (t) have the desired correlation . Using 100,000 simulations we compute
the average plant value for a range of values for a 10 year period. Figures 4.6-4.9 show how
plant value changes as increases for varying initial conditions.
4.2 Transforming to a 1D Problem

We could also assume that the difference between corn and gasoline prices (adjusted using H)
is a Brownian motion with drift, or arithmetic Brownian motion (ABM). This is not equivalent to assuming that the difference of two GBM processes is ABM, but rather that since the
GBM assumption is weak, an ABM difference process might be reasonable. In other words,
Q(t) = HG(t) C(t) where Q(t) evolves according to (2.4.1). Like the GBM assumption,
this assumption is not perfect. This is illustrated by a QQ plot shown in Figure 4.10. However, this assumption is in line with the 1D PDE introduced in Chapter 2, which can easily be
solved. Therefore, the assumption provides us with a simpler, less computationally intensive
alternative, despite also being a weak assumption.
The PDE with final and boundary conditions, under the leased facility assumption previously discussed, is given by
1 2V
V
V
+ 2 2 +
t
2 Q
Q
rV + QH (Q) = 0,
subject to
V (Q, T ) = 0
lim V (Q, t) = 0
lim V (Q, t) = a(t)Q + b(t).
(4.2.1)
4.2. T RANSFORMING TO A 1D P ROBLEM
43
530
520
Value
510
500
490
480
470
1
0.5
0
Correlation
0.5
Figure 4.6: Lease Value with increasing for C0 = 3,G0 = 2
44
985
980
Value
975
970
965
960
1
0.5
0
Correlation
0.5
Figure 4.7: Plant Value with increasing for C0 = 2,G0 = 3
45
420
400
380
Value
360
340
320
300
280
260
1
0.5
0
Correlation
0.5
46
1744
1742
1740
Value
1738
1736
1734
1732
1730
1
0.5
0
Correlation
0.5
47

2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
4
1
0
1
Figure 4.10: QQ plot - Difference of HG(t) C(t) time series, i.e. H(Gk+1 Gk ) (Ck+1 Ck )
against the standard normal distribution
48
The boundary conditions are straightforward: as the spread between corn and gasoline becomes
very negative, V (Q, t) approaches zero since it is very unlikely that the spread will recover
and hence the plant will always lose money if it operates. Thus, the decision is to shut down
permanently, making no money for the remaining life of the plant. As the spread between corn
and gasoline becomes highly positive (i.e. as Q gets large, represented by a limit to infinity as
in the 2D case), it is very unlikely that the spread Q will decrease to the point where it becomes
negative. Thus, the plant is always making money and its value is the sum of expected cash
flows, discounted at rate r. This can be captured by the following integral:
Z T t
rs
lim V (Q, t) = E
e Qs ds
Q

Q
=
1 er(T t) +
1
r
r
= a(t)Q + b(t).
er(T t) ter(T t)
We can write a MATLAB code to numerically solve our PDE using the Crank Nicolson
method [21]. This method is also discussed in detail in Appendix A. When implementing this
method we used a time step of 0.001. The space grid was kept relatively small in order to get
a clear picture of what is happening around zero: it ranged from -10 to 10 with a step size of
0.1. Figure 4.11 shows V (Q, t) for the range of Q and t values specified by our grid, called
a surface plot. A surface plot can be used to identify numerical glitches, often indicated by
bumpy sections rather then a smooth plot. It can also be used to capture the big picture. In
this case it is clear that for a given t, plant value looks similar to the payoff of a call option,
which is expected.
Figure 4.12 shows a cross section of the surface plot at time step 1000 (i.e. 1000 time steps
before expiry, which in this case is the scrap date). This plot is used as a reality check: when
the spread is highly negative (<-4) the probability that the option value will be nonnegative is
extremely low, and hence has zero value. As the spread becomes less negative the option has
a small positive value since, although the option is out of the money at the present time (i.e.
the plant is operating at a loss), there is still time for the spread to become positive, in which
case the plant would be making a profit. It is thus capturing potential value in the future. As Q
increases past zero, the option value increases rapidly. This is since once the spread between
corn and gasoline prices becomes high, the probability that it will become negative in the future
is very small. Thus, the plant is always operating at a profit, with the highest profit occuring
when the spread attains its maximum value.
From each pair of corn and gasoline price simulations we could estimate the best Brownian
motion with drift to the spark spread time series and estimate the drift and volatility of the
Brownian motion by computing
Sk = H(Gk+1
Gk )
(Ck+1
Ck )
for k = 1...N 1 where N is the number of corn (or gasoline) prices per simulation. The drift
for each simulation is estimated as the mean of S1 , ..., SN 1 and the volatility is the standard
deviation of S1 , ..., SN 1 . We predict that as correlation increases, the volatility of our spark
spread time series will decrease. This is since an increased correlation implies prices move in
the same direction; if prices move together the spread will have less variability from one time
Figure 4.11: 1D PDE surface plot
49
50
80
70
60
50
40
30
20
10
0
10
0
Q
Figure 4.12: Surface plot cross section, V (Q, 1000)
10
51
period to the next. On the other hand, a zero correlation means that prices can move in any
direction relative to the other, and the spread can become quite large and quite small in a short
time period. In other words, the spread is more volatile. To check this assumption we plot our
volatility estimate against varying correlation values from 0 to 1. The result is given in Figure
4.13. From this figure we see that the result is indeed as expected; as correlation increases
from zero to one, the volatility decreases by approximately 60%. We do not include negative
correlation values here as the underlying assumption is that correlation, which is historically
very small but positive, will increase.
To test the assumption that increased correlation implies decreased plant value we produce
ethanol plant real option values for varying corngasoline correlation using the same spark
spread time series. The result is given in Figure 4.14. It is clear from this figure that the
assumption indeed holds under the modeling assumptions described above.
0.34
0.32
0.3
0.28
sigma
0.26
0.24
0.22
0.2
0.18
0.16
0
0.2
0.4
0.6
0.8
rho
Figure 4.13: of the Spark Spread Time Series as increases
52
1.8
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
1.6
1.4
Value
1.2
1
0.8
0.6
0.4
0.2
0
20
40
60
Time
80
100
Figure 4.14: Ethanol plant value with varying
120
4.3. S UMMARY
4.3
53
Summary
When moving away from a data driven analysis, the modeling assumptions which are made
are crucial in obtaining meaningful results. It is also important to find stable solutions which
are valid in all cases and for all parameter choices. In this study, both the 1D and 2D models
faced modeling challenges, however the 2D model faced additional challenges in finding stable
solutions. However, the results do agree with the data driven results presented in Chapter 3,
adding confidence that the ethanol plant valuation is representative of ethanol plants in current
use today. The most important consequence of this analysis is that possible future correlations
between corn and gas prices may impact the level of public subsidy required in order to prevent
the value of ethanol plants from being destroyed.
Chapter 5
Wind and Storage as a Renewable Energy
Source
Biofuels are just one of many types of green energy technologies in current use. An equally
controversial green energy source is wind power. Wind power refers to the process of converting wind energy into electricity via wind turbines. Wind power is currently replacing a small
fraction of the power produced by burning fossil fuels such as coal. In the US in 2010, wind
power accounted for approximately 2% of the total electricity supply; this fraction is predicted
to reach 20% by 2030 [23].
Major criticisms of wind power are its inherent variability as well as its uncertainty; any
power source must be reliable in order to meet demand, especially when power demand is significantly higher than average. One way of achieving reliability is to couple wind turbines with
storage systems which have the ability to provide power when wind speeds are at undesirable
levels. See [17] for more information on the coupling of wind turbines with storage systems.
In order for storage to be a viable solution to winds intermittency, a great deal of research must
be carried out. This research should focus on three major areas, outlined below.
1. A statistical characterization of wind speed is required. This can be done by characterizing wind speed at a given time via probability distribution functions, as well as in terms
of times series analysis describing the patterns in wind time series. Reliable historical
wind speed data is essential in this piece of the research process. It is also important to
understand how wind speed is translated into power output.
2. A variety of storage technologies must be examined. These technologies can be compared based on cost, location and responsiveness. Once a technology is chosen, the key
parameters involved in using that technology for energy storage must also be examined.
3. Indicators for control strategies must be chosen. This requires consideration of wind
forecasting, how power is bid into the market, power prices as well as the fill level of the
storage unit.
Each of the areas described above is crucial in understanding the role that combining wind
and storage technologies can play in replacing fossil fuels for power generation. In order to
proceed, we present in the next section an exactly solvable model for a wind storage system
54
5.1. DYNAMIC P ROGRAMMING A PPROACH
TO
O PTIMAL C ONTROL P ROBLEM
55
and obtain optimal control results for this simple model. These results will provide guidance
in the remaining chapters when a more realistic, and hence more complex, model is developed.
Wind power is renewable but not predictable, as will be seen later in Chapter 6, which
combines a review of literature on wind modeling with an exploration of real wind data. One
solution for this is storage. There are a variety of storage technologies available, as will be
surveyed later in the same Chapter 6. All of these technologies share the characteristic that
there is cost associated with operating them: in physical terms, there is significant energy
loss. They are also expensive to construct. These two features make it desirable for wind
producers to ignore them; to avoid this, market rules must be created to encourage their use.
A great deal of work has been done on optimal control of wind energy storage solutions; this
has mainly been done without regulatory emphasis however. This work is reviewed in 6.6.1.
In the following sections we take an approach motivated by the way macro economists think
about this kind of problem, as reviewed for instance in [67]. We create a random model for
wind generation, a simple model for storage losses, and a simple model for market rules. This
is embedded into a dynamic program, which we can solve exactly. The result of this chapter
is used to motivate more detailed, though less tractable, models of wind storage systems in
Chapters 7. We return to power market design questions in Chapter 8.
5.1
Dynamic Programming Approach to Optimal Control

Problem
Before the model is described, a brief introduction to dynamic programming is given.
5.1.1
Dynamic Programming
Dynamic programming is a way of optimizing a problem by breaking it down into a series of

decisions over time. Given a value sequence V1 , V2 , ...Vk , the value at an earlier (or later) time is
found by working backwards (or forwards) via a recursion relation. In some cases it is possible
to solve the dynamic program exactly, allowing for meaningful results to come easily; the goal
in this chapter is to create a simple dynamic program which allows for analytic solutions. The
model is described and analyzed in the following sections.
5.1.2
Assumptions
Assume an extremely simple model for wind where, at a given time, the wind is either blowing
or it is not. If the wind is blowing, wind power can be produced regardless of the wind speed.
At each time period a wind producer must decide whether or not to bid electricity; if a decision
to bid is made and the wind is blowing the producer receives M , the price of electricity. On the
other hand, if the decision is to bid and the wind is not blowing the producer faces a penalty of
xM . To slightly increase the complexity of this problem, assume that the wind producer has
rented one unit of storage. At a given time period it may either be full or empty; in other words,
one time period of wind power equals the storage capacity of the storage. The wind producer
pays M for using the unit, however it can be used to help avoid the potentially hefty penalty
imposed in a case where a bid decision has been made but the wind doesnt blow.
56
C HAPTER 5. W IND AND S TORAGE
wind
no wind
bid
M
xM
dont bid
(1 )M
0
Table 5.1: Benefit Matrix: Empty Battery
wind
no wind
bid
M + (1 )M
(1 )M
dont bid
(1 )M
(1 )M
Table 5.2: Benefit Matrix: Full Battery
5.1.3 One Period Model

Consider an even simpler one period model. At the end of the period the storage facility is
returned and the producer receives (1 )M for returning it full and zero for returning it
empty, i.e.
V (E, 0) = 0
V (F, 0) = (1
)M.
(5.1.1)
(5.1.2)
Note that in this case, these initial conditions are really terminal conditions; in other words,
this is a backward dynamic program. It is common in dynamic programming problems to begin
with a time reversal transformation in the formulation of the problem, by always referring to
the number of periods remaining and by defining that as our time variable. Tables 5.1 and 5.2
show the benefit matrices associated with an initially full as well as an initially empty battery.
The wind blows with probability p. In other words, the probability that wind speed is
positive is p, and the probability that wind speed is zero is 1 p. Let B denote the decision to
bid, and N denote the decision not to bid. E represents an empty battery whereas F represents
a full battery. V (X, Y, t) is used to describe the value of the wind/storage system at time t
given X belongs to {E, F } and Y belongs to {B, N }. If we begin with an empty battery we
can write the value of each decision as
V (E, N, 1) = pV (F, 0) + (1 p)V (E, 0)
V (E, B, 1) = p[M + V (E, 0)] + (1 p)[ xM + V (E, 0)].
In other words, if the decision is not to bid we either fill our battery when the wind blows (p
percent of the time) or we do nothing (1 p of the time) if it doesnt. The motivation for
filling the battery instead of doing nothing is that we receive V (F, 0) for returning it full. If
the decision is to bid, we receive M for the wind and return an empty battery worth V (E, 0) (p
percent of the time) or we pay a penalty of xM and return an empty battery worth V (E, 0) (1 p
percent of the time). Given this, is it optimal to always bid, never bid, or some combination of
bidding and not bidding? Clearly, if the decision is to always bid then we never fill our storage.
Under what condition is it optimal to never bid when empty, or in mathematical terms, when
is V (E, N, 1) > V (E, B, 1)? Simple algebra tells us that the condition for never bidding is as
follows:

x>
TO
57
p
.
1 p
Unless this condition holds, we will never fill our storage. This condition makes a great deal
of sense given the meaning of and x. Assume for a moment that p = 1 p = 0.5, and thus
p/(1 p) = 1. If the imposed penalty for failing to make a bid commitment is smaller than the
cost of using battery storage then battery storage will have little to no value. However, if the
imposed penalty is larger than the cost of the storage, the storage becomes valuable. If p 6
= 0.5
then in order for the storage to be valuable, not only does it have to compete with the cost of
storage, it must also compete with nature (i.e. the ratio of windy to calm days).
Although other cases are analyzed in full, x > p/(1 p), or equivalently p < (1 p)x, is
the standing assumption used throughout this section. This assumption captures the fact that,
in order to promote wind producers to use storage to increase the reliability of wind power,
penalties must be high enough to beat the effects of storage costs and nature. In addition,
the final condition V (F, 0) = (1 )M provides incentive for producers to keep filling their
battery since once it is returned they receive payment for returning it full.
To optimize total value we will clearly choose max[V (E, N, 1), V (E, B, 1)], denoted by
V (E, 1), we conclude that
V (E, 1) = pV (F, 0) + (1
p)V (E, 0) = p(1
)M .
We next look at V (F, N, 1) and V (F, B, 1), the case when we begin with a full battery. We
can write
V (F, N, 1) = pV (F, 0) + (1 p)V (F, 0) = V (F, 0)
V (F, B, 1) = p[M + V (F, 0)] +
(1 p) max[ xM + V (F, 0), (1 )M + V (E, 0)].
If the decision is to not bid, then we simply do nothing; if there is wind we cannot store it
since the battery is already full. However we do receive V (F, 0) for returning a full battery.
If the decision is to bid then it is not quite as simple, if there is wind we use it and remain
full (collecting V (F, 0) for keeping the battery full). The max() expression within V (F, B, 1)
indicates that there are two choices if there is no wind; we either pay a penalty in order to keep
our battery full, or we use our battery in order to avoid paying a penalty. We can rewrite these
expressions as follows:
xM + V (F, 0) = (1 )M xM = V (F, 0)
(1 )M + V (E, 0) = (1 )M = V (F, 0).
xM
Clearly the latter is always greater, meaning it is always better to use our storage, and hence
V (F, B, 1) = p[M + V (F, 0)] + (1 p)V (F, 0).
Now we can write V (F, 1) = max[V (F, N, 1), V (F, B, 1)] and since pM +V (F, 0) > V (F, 0)
it is clear that it is optimal to bid, giving us
V (F, 1) = p[M + V (F, 0)] + (1 p)V (F, 0).
58
Finally, for the one period model we can combine this result and state that
V (E, 1) =
=
V (F, 1) =
=
=
pV (F, 0) + (1 p)V (E, 0)

p(1 )M
p[M + V (F, 0)] + (1 p)V (F, 0)
V (F, 0) + pM
(p + 1 )M.
5.1.4 Extending to N time periods

Restricting the model to only one time period is not sufficient in determining an optimal control
strategy. However, with the one period framework in place it is relatively simple to extend the
model to any number of time periods. The details, beginning with a two period model, are
given in this section. To skip these details and move directly to the general model for any
number of time periods the reader can proceed to Theorem 5.1.2.
For a moment assume t = 2. Using what we already know, we can write
V (E, N, 2) = pV (F, 1) + (1 p)V (E, 1)
V (E, B, 2) = p[M + V (E, 1)] + (1 p)[ xM + V (E, 1)].
Similar to the one period case, given an empty battery with two time periods remaining and a
decision to not bid, we either fill our battery when the wind blows (a fraction p of the time)
or we do nothing when it doesnt (1 p of the time). There is still incentive to fill the battery
when the wind does blow since there is potential to use this wind in the next time period or
receive payment for it at the end. The optimal decision, bid or not bid, is determined from
the expression V (E, 2) = max[V (E, N, 2), V (E, B, 2)]. To evaluate this expression, let us
revisit the base assumption which tells us not to bid in the one period case. V (E, N, 2) and
V (E, B, 2) can be rewritten as follows:
V (E, N, 2) = p2 M + p(1 )M + (1 p)V (E, 1)
V (E, B, 2) = pM + p2 (1 )M (1 p)xM + (1 p)V (E, 1).
Therefore, V (E, 2) can be determined from
max[p2 + p(1 ), p + p2 (1 ) (1
= p2 + p + max{ p, p2 (1 p)x}
= p(1 + p) min[p, p2 + (1 p)x].
p)x]
From our standing assumption it is clear that

V (E, 2) = p[M + V (E, 1)] + (1 p)[ xM + V (E, 1)]
= pV (F, 1) + (1 p)V (E, 1).
In other words the optimal strategy is to not bid as in the one period case.
TO
59
A similar calculation can be carried out under the assumption that at t = 2 the battery is
full. We can write
V (F, N, 2) = pV (F, 1) + (1 p)V (F, 1)
V (F, B, 2) = p[M + V (F, 1)]
+ (1 p) max[ xM + V (F, 1), (1
)M + V (E, 1)].
The max() expression within V (F, B, 2) once again indicates that there are two choices; we
either keep our storage full at the cost of paying a penalty, or we use our storage to avoid paying
a penalty. We can rewrite these expressions as follows:
xM + V (F, 1) = pM + (1 )M xM
(1 )M + V (E, 1) = (1 )M + p(1 )M.
The problem reduces to max[pM + (1 )M xM, (1 )M + p(1 )M ] = (1 +
p)M M min[x, p)]. Once again our standing assumption, p < (1 p)x implies that p < x
always holds, indicating that it is better to use our storage. Therefore
V (F, B, 2) = p[M + V (F, 1)] + (1
p)[(1 )M + V (E, 1)].
Next we want to show that V (F, B, 2) >= V (F, N, 2), or in other words we should always bid
when full with two periods remaining. In that case V (F, 2) = max[V (F, N, 2), V (F, B, 2)] =
V (F, B, 2). Now
V (F, B, 2) = p[M + V (F, 1)] + (1 p)[(1
V (F, N, 2) = pV (F, 1) + (1 p)V (F, 1),
)M + V (E, 1)]
so
V (F, B, 2)
But V (F, 1)
V (F, B, 2)
V (F, N, 2) = pM + (1 p)(1 )M (1 p)[V (F, 1) V (E, 1)]

= pM + (1 p)V (F, 0) (1 p)[V (F, 1) V (E, 1)].
V (E, 1) = V (F, 0) + pM so
V (F, N, 2) = pM + (1 p)V (F, 0) (1
= pM [1 (1 p)] > 0.
p)V (F, 0)
p(1
p)M
Thus
V (F, 2) = V (F, B, 2) = p[M + V (F, 1)] + (1 p)[(1 )M + V (E, 1)]
= p[M + V (F, 1)] + (1 p)[V (F, 0) + V (E, 1)]
as required, meaning that it is always optimal to bid.
After working through the calculations for two time periods we are ready to generalize the
results to more time periods. We begin by considering V (F, B, k):
V (F, B, k) = p[M +V (F, k 1)]+(1 p) max[(1 )M +V (E, k 1), xM +V (F, k 1)].
60
In particular we consider max[(1 )M + V (E, k 1), xM + V (F, k 1)]. Is it better to

empty your storage to avoid paying penalties, or are there situations where it is more valuable to
keep the storage facility full for future use, paying a penalty to do so? If it indeed makes sense
to empty your storage to avoid paying penalties, then it must be true that V (F, k) V (E, k)
(1 )M + xM . With our standing assumption that x(1 p) > p we will show that this is
indeed the case, and we will in fact later prove that this is true in general. To move directly to
this general case, see Theorem 5.1.2. Otherwise, we begin with the following theorem where
our standing assumption holds.
Theorem 5.1.1. Consider the recursive system:
V (E, N, k)
V (E, B, k)
V (E, k)
V (F, N, k)
V (F, B, k)
V (F, k)
=
=
=
=
=
=
pV (F, k 1) + (1 p)V (E, k 1)

p[M + V (E, k 1)] + (1 p)[ xM + V (E, k 1)]
max[V (E, N, k), V (E, B, k)]
pV (F, k 1) + (1 p)V (F, k 1) = V (F, k 1)
p[M + V (F, k 1)] + (1 p)[(1 )M + V (E, k 1)]
max[V (F, N, k), V (F, B, k)]
with final conditions

V (E, 0) = 0
V (F, 0) = (1 )M.
If in addition x(1
1. (1
p) > p then the following statements hold:
)M V (F, k) V (E, k) (1
)M + pM , k.
2. V (E, N, k) V (E, B, k), k

3. V (F, B, k) V (F, N, k), k.
Proof. Lets begin with V (F, k) V (E, k). When k = 0 we obtain V (F, 0) V (E, 0) =
(1 )M (1 )M + pM as well as (1 )M V (F, 0) V (E, 0). Now assume that
statement 1 holds for some non-negative integer k 1. Now examine V (E, N, k) V (E, B, k).
We know from the recursive system above that
V (E, N, k)
V (E, B, k) = p[V (F, k
1)
V (E, k 1)] + [(1 p)x
From our assumption we can say that V (F, k
1)
V (E, k 1) (1
V (E, N, k)
V (E, B, k) p(1 )M + (1 p)xM

= [(1 p)x p)]M
p]M .
)M so
pM
Since x > p/(1 p) we know this is positive, and hence V (E, N, k) V (E, B, k). Therefore
if statement 1 holds for k 1 then statement 2 holds for k.
From the assumption that statement 1 holds for k 1 it follows that

V (F, B, k)
V (F, N, k) = pM + (1
p)(1
From our assumption we can say that V (F, k

V (F, B, k)
TO
)M
1)
V (F, N, k) pM + (1 p)(1
= pM pM
= p(1 )M.
(1
p)[V (F, k
V (E, k 1) (1
)M
(1
1)
V (E, k
61
1)].
)M + pM so
p)[(1 )M + pM ]
Clearly this is positive, and hence V (F, B, k) V (F, N, k). Therefore if statement 1 holds
for k 1 then statement 3 holds for k.
Finally, we examine V (F, k) V (E, k). Since item 2 holds for k we can write V (E, k) =
V (E, N, k). Since item 3 holds for k we can also write V (F, k) = V (F, B, k). Therefore,
V (F, k) V (E, k) =
=
=
=
V (F, B, k) V (E, N, k)
pM + (1 p)(1 )M
p(1 )M + pM + (1 p)(1
(1 )M + pM.
)M
Therefore statement 3 also holds by induction and the proof is complete.
We also formally state (in words) the optimal control strategy which is given in the above
theorem.
Corollary 5.1.1. Under the assumption that
p
,
1 p
i.e. for sufficiently large penalty, it is always optimal to bid when the storage facility is full and
not bid when the storage facility is empty.
x>
Recalling an earlier assumption (when p = 0.5) that x > must hold in order for storage
to have value, the penalty inequality is re-written as x > max(1, p/(1 p)). This says that
the less efficient the storage facility the larger the penalty must be to incent producers to use it.
However if windy days are very common then extremely large penalties are needed to incent
producers not to bid when their storage is empty. If we consider an empty facility, the expected
penalty paid per turn when bidding will be (1 p)x, while the expected storage loss paid by
not bidding and getting filled will be p. We need to make the penalty loss higher than the
storage loss and hence x(1 p) > p, or x > p/(1 p).
For the system described in Theorem 5.1.1, an analytic solution can be found. This is
summarized in the following corollary.
Corollary 5.1.2. For the recursive system defined in Theorem 5.1.1, with condition x > p/(1
p), we can find an analytic solution to the value function, given below:
V (E, 0)
V (E, 1)
V (E, k)
V (F, 0)
V (F, 1)
V (F, k)
=
=
=
=
=
=
0
p(1 )M
kp(1 )M + (k 1)p2 M, k > 1
(1 )M
(1 )M + pM
(1 + kp)(1 )M + (1 + (k 1)p)pM, k > 1.
62
Proof. In Theorem 5.1.1 we showed that, k > 0, V (E, k) = V (E, N, k) and V (F, k) =
V (F, B, k). This allows us to rewrite the recursive system as
V (E, k)
V (F, k)
V (E, 0)
V (F, 0)
=
=
=
=
pV (F, k 1) + (1 p)V (E, k 1)

p[M + V (F, k 1)] + (1 p)[(1 )M + V (E, k
0
(1 )M.
1)]
In addition, we showed in proving that item 1 of Theorem 5.1.1 holds for k given that it holds
for k 1 that for k > 0 the following holds:
V (F, k) = V (E, k) + (1
)M + pM .
We can use this result to obtain a single first order difference equation initial (final) value
problem (note that the k > 0 condition above is necessary since it does not replicate the
final conditions given in (5.1.1) and (5.1.2) when k = 0). We can account for this by using
V (E, 1) = p(1 )M and V (F, 1) = p[M + (1 )M ] + (1 p)(1 )M = pM + (1 )M ,
detailed in the previous section, and using these as the new final conditions to manipulate. With
this in mind we can combine the expressions for V (E, k) to obtain
V (E, k) = p[V (E, k 1) + (1
)M + pM ] + (1
p)V (E, k
1),
which simplifies to the very simple linear difference equation

V (E, k) = p[(1
with final condition V (E, 1) = p(1
) + p]M + V (E, k
1)
)M . Therefore,
V (E, k) = kp(1
)M + (k
1)p2 M , k > 0.
Using the expressions for V (F, k) we obtain the similar result

V (F, k) = (1 + kp)(1
)M + (1 + (k
1)p)pM .
It is worthy to note that if the final condition changes the exact form of the above Corollary
will have to be modified. We have shown, provided x > p/(1 p), that for k = 1, 2, 3,
V (F, k) V (E, k) = (1 )M + pM . We can think of this difference as being the intrinsic
value of being full, (1 )M , plus some option value of being full, pM . Another choice for
the final conditions might be
V (E, 0) = 0
V (F, 0) = (1
)M + pM.
As before there is no value when empty, however such a final full condition would recognize
the option value of being full as well as the intrinsic value. This would make sense if the
facility is leased (in fact, if the facility was being abandoned then presumably V (F, 0) = 0
would apply).
Given these new final conditions, we are able to write
TO
63
V (F, k) = V (E, k) + (1 )M + pM , k
instead of restricting ourselves to k > 0. Therefore,
V (E, k) = p[(1 ) + p]M + V (E, k
1)
with final condition V (E, 0) = 0 yields

V (E, k) = kpM [(1
) + p].
As a result, we obtain
V (F, k) = M (kp + 1)[(1 ) + p].
What happens if we try to generalize this result, given our original final conditions, beyond
the case when x > max(1, p/(1 p))? It is clear that understanding the difference V (F, k)
V (E, k) is crucial in proving the results obtained thus far. In fact, the following theorem
regarding this difference allows us to obtain the optimal operating strategy for any x. In other
words we relax our standing assumption and state a wide range of results for the unrestricted
model.
Theorem 5.1.2. For the system of equations given below,
V (E, N, k)
V (E, B, k)
V (E, k)
V (F, N, k)
V (F, B, k)
=
=
=
=
=
pV (F, k 1) + (1 p)V (E, k 1)

p[M + V (E, k 1)] + (1 p)[ xM + V (E, k 1)]
max[V (E, N, k), V (E, B, k)]
V (F, k 1)
p[M + V (F, k 1)] + (1 p) max[(1 )M + V (E, k
xM + V (F, k 1)]
V (F, k) = max[V (F, N, k), V (F, B, k)]
1),
with final conditions

V (E, 0) = 0
V (F, 0) = (1 )M
the following result holds:
V (F, k)
V (E, k) = (1
)M + min(pM, (1 pk )xM ), k 0.
In order to prove the above theorem, we first prove the following lemma.
Lemma 5.1.1. Let V (F, k) V (E, k) = M m(k) + (1
)M . Then
m(k) = m(k 1) + min{x(1 p), p pm(k
min{p, x(1 p), (1 p)m(k 1)]}

m(0) = 0.
1)}
(5.1.3)
64
Proof. From the most general system defined in Theorem 5.1.2, we obtain
V (F, k)
V (E, k) = max[V (F, B, k), V (F, N, k)] max[V (E, B, k), V (E, N, k)]
= max{p[M + V (F, k 1)] + (1 p) max[ xM + V (F, k 1),
(1 )M + V (E, k 1)], V (F, k 1)}
max{p[M + V (E, k 1)] + (1 p)( xM + V (E, k 1)),

pV (F, k 1) + (1 p)V (E, k 1)}.
Let n(k) = [V (F, k) V (E, k)]/M . Then,

n(k) =
n(k 1) + max{ (1 p)x, (1 p)[n(k 1)

max{ x(1 p), pn(k 1) p}
n(k 1) + min{x(1 p), p pn(k 1)}
min{x(1 p), (1 p)[n(k 1) (1 )], p}.
Now let m(k) = n(k)
(1
(1
)], p}
). With this we easily obtain
m(k) = m(k 1) + min{x(1 p), p pm(k 1)}
min{x(1 p), (1 p)m(k 1), p}.
With this in place, it is clear that if we are able to show that m(k) = min{p, (1 pk )x}
satisfies the equation of Lemma 5.1.1 then the proof of Theorem 5.1.2 is complete. To show
this we use induction.
Proof. First we show that it holds for k = 1, i.e.
m(1) = m(0) + min{x(1 p), p} min{x(1 p), (1
= 0 + min{x(1 p), p} min{x(1 p), 0, p}
= min{x(1 p), p}.
p)m(0), p}
which is clearly satisfied by m(1) = min{p, (1 p1 )x}. Next we assume it holds for k 1
and show it holds for k + 1. We know that either x(1 pk ) p or x(1 pk ) > p. Let us
look at each case separately.
1. Case 1: x(1
pk ) p, so m(k) = x(1
pk ).
We have
m(k + 1) =
Now, we can write
m(k) + min{x(1 p), p pm(k)}

min{x(1 p), (1 p)m(k), p}
x(1 pk ) + min{x(1 p), p px(1
min{x(1 p), (1 p)x(1 pk ), p}.
pk )}

min{x(1
TO
pk )} = px(1
p), p px(1
pk ) + min{x(1 p) + px(1
65
pk ), p}
and
(1
min{x(1 p), (1 p)x(1 pk ), p} =

p)x(1 pk ) + min{x(1 p) (1 p)x(1 pk ), 0, p (1
p)x(1
pk )}.
Therefore,
m(k + 1) =
+
x(1 pk )
min{x(1
min{x(1
min{x(1
min{x(1
px(1 pk ) (1 p)(1 pk )
p) + px(1 pk ), p}
p) (1 p)x(1 pk ), 0, p (1 p)x(1
pk+1 ), p}
p) (1 p)x(1 pk ), 0, p (1 p)x(1
pk )}
pk )}.
If we can show that min{x(1 p) (1 p)x(1 pk ), 0, p (1 p)x(1 pk )} = 0 then

we are done. Starting with the first term we simplify as follows:
x(1
p) (1
p)x(1 pk ) = x(1 p)[1 (1 pk )]

= x(1 p)pk
> 0.
Moving on to the third term we note that since we have assumed that x(1
then
x(1
(1 p)x(1
p (1 p)x(1
pk ) p
pk ) (1 p)p
pk ) p (1 p)p
= p[1 (1 p)]
= p2
> 0.
Therefore,
min{x(1 p)
(1
p)x(1 pk ), 0, p
(1 p)x(1
and we finally arrive at

m(k + 1) = min{x(1 pk+1 ), p} 0
= min{x(1 pk+1 ), p}.
2. Case 2: x(1
pk ) > p, so m(k) = p.
pk )} = 0
pk ) p
66

We have
m(k + 1) = m(k) + min{x(1 p), p pm(k)}
min{x(1 p), (1 p)m(k), p}

= p + min{x(1 p), p p2 } min{x(1
Since p > (1
p)p, we know that min{x(1
m(k + 1) = p + min{x(1
= p.
The final questions is, is min{x(1
pk ), p} = p. But we know that
p), (1 p)p}
p)p, p}.
p)p, p} 6
= p. Therefore,
min{x(1
p), (1 p)p}
pk+1 ), p} = p? By our assumption, min{x(1
pk+1
pk+1
1 pk+1
x(1 pk+1 )
Therefore, min{x(1
p), (1
p), (1
<
>
>
>
pk
pk
1 pk
x(1 pk ).
pk+1 ), p} = p and we are done.
It is interesting to note that V (F, k) V (E, k) = (1 )M + min{pM, (1 pk )xM } is

non-decreasing in time remaining, like an option value. Furthermore, if x(1 pk ) < p it is
clear that V (F, k) V (E, k) = (1 )M + (1 pk )xM. As k , pk 0, hence we can
state
limk [V (F, k) V (E, k)] = (1 + x)M .
If x(1
pk ) > p then V (F, k) V (E, k) simplifies to

V (F, k)
V (E, k) = (1
= (1
)M + pM
(1 p))M,
which is independent of k, or more simply, a constant. Can we determine a k = k at which

point V (F, k) V (E, k) levels off? Consider x(1 pk ) < p. We can write
x(1 pk ) < p
p
1 pk <
x
p
k
p > 1
x p
k ln p > ln 1
x
ln(x p) ln x
k <
.
ln p
TO
67
The inequality flip in the last step is necessary since ln p < 0. Therefore, V (F, k) V (E, k)
increases until k = k where k is the largest integer satisfying k < [ln(x p) ln x]/[ln p],
then it reaches its limit of (1 + x)M .
Equipped with this Theorem we can easily state several other results, as well as state the
optimal control. One important result is given in the following Corollary:
Corollary 5.1.3.
xM + V (F, k
1) (1
)M + V (E, k
1) k,
(5.1.4)
or in other words, we can rewrite

V (F, B, k) = p[M + V (F, k 1)] + (1 p) max[(1
xM + V (F, k 1)]
)M + V (E, k
1),
as
V (F, B, k) = p[M + V (F, k
1)] + (1 p)[(1
)M + V (E, k
1)].
Proof. This result is a simple extension of Theorem 5.1.2 using the fact that xM min{pM, (1
pk )xM }. Therefore we can write
V (F, k)
V (E, k) = (1 )M + min(pM, (1 pk )xM )

(1 )M + xM,
or equivalently
xM + V (F, k
1) (1
)M + V (E, k
1) k.
This result implies that, if you bid and the wind doesnt blow, it is always better to use
storage then pay a non-negative penalty, no matter how small.
The next result allows us to simplify V (F, k) = max{V (F, B, k), V (F, N, k)} to V (F, k) =
V (F, B, k); in other words, when full, a decision to bid should always be made.
Corollary 5.1.4. V (F, B, k) V (F, N, k), k.
Proof. Given the system outlined in Theorem 5.1.2, as well as the previous Corollary, we can
write
V (F, B, K)
V (F, N, k) = p[M + V (F, k 1)] + (1 p)[(1 )M

+V (E, k 1)] V (F, k 1)
= pM + (1 p)[(1 )M [V (F, k 1) V (E, k
= pM (1 p) min{pM, (1 pk1 )xM }
> 0,
giving the desired result V (F, B, k) > V (F, N, k).
1)]]
68
In other words, when full it is always optimal to bid, regardless of x. Thus one half of
the optimal operating strategy has been discovered. What about when we are empty? Should
we never bid, always bid, or some combination of bidding and not bidding? The following
corollary answers this question.
pk ).
Corollary 5.1.5. When empty, it is better to bid if and only if p > x(1
Proof. Again, given the system outlined in Theorem (1.1.2) we can write
V (E, B, k)
V (E, N, k) =
=
=
+
=
=
=
=
If x < p then p(p
p[M + V (E, k 1)] + (1 p)[ xM + V (E, k 1)]

pV (F, k 1) (1 p)V (E, k 1)
pM p[V (F, k 1) V (E, k 1)] (1 p)xM
M [p x(1 p)] p[(1 )M
min{pM, (1 pk1 )xM }
M [p x(1 p)] pM min{p, (1 pk1 )x}
M [p x(1 p)] + pM max{ p, x(1 pk1 )}
M max{p x(1 p) p2 , p x(1 pk )}
M max{(p x) p(p x), p x + xpk )}.
x) < 0. Therefore,
V (E, B, k) V (E, N, k) = M (p x) + M xpk

= M [p x(1 pk )].
If p < x then we can rewrite as
V (E, B, k)
V (E, N, k) = M max{(1
p)(p
x), p
x + xpk )}.
The first term is certainly negative since x > p, and the second term is positive if p >
x(1 pk ).
In both cases, if p > x(1 pk ) then it is better to bid, as desired.
Given this result, let us consider the case where penalties are large. That is, consider
x max{1, p/(1
p)}.
Since x(1 pk ) > x(1 p), when p > 1/2 we also have x(1 p) > p. If p < 1/2 we
have that x(1 p) > xp > x. Hence V (E, B, k) V (E, N, k) < 0 p, meaning that it
is optimal to not bid. This is not surprising given a large penalty. Thus we can state that the
optimal control strategy when penalties are large is to bid when full and not bid when empty.
This way you never have to pay the large penalties, and you are able to refill the first time it is
windy after a calm day. In this case we can state that
V (F, k)
V (E, k) = M (1
(1
p)).
TO
69
We can also easily solve the difference equations, starting with V (E, k):
V (E, k) =
=
=
=
V (E, N, k)
pV (F, k 1) + (1 p)V (E, k 1)
V (E, k 1) + p[V (F, k 1) V (E, k
V (E, k 1) + pM (1 (1 p)).
1)]
Therefore,
V (E, k) = kpM (1
(1
p)), k 0
and hence
V (F, k) = V (E, k) + M (1 (1 p))
= M (kp + 1)(1 (1 p)), k 1.
When penalties are small it turns out to be always optimal to bid. For instance, consider
x p.
We can write V (E, B, k) V (E, N, k) as
V (E, B, k)
V (E, N, k) = M (p
x(1 pk )).
Since x > x(1 pk ), or x(1 pk ) < x < p, we know that V (E, B, k) V (E, N, k) > 0 and
hence it is always optimal to bid. Thus, whether empty or full, it is always optimal to bid when
penalties are small. Penalties are too small to encourage the use of storage; if you start empty
you will never fill the storage, and if you start full the storage is used exactly once - to empty
it. In this case we can state that
V (F, k)
V (E, k) =
=
=
=
p[V (F, k 1) V (E, k 1)] + M (1 p)[1 + x]

p[(1 )M + xM (1 pk1 )] + M (1 p)[1 + x]
M [px(1 pk1 ) px + 1 + x]
M (1 + x(1 pk )).
Solving the difference equations is fairly straightforward. We can write

V (E, k) = p[M + V (E, k 1)] + (1 p)[ xM + V (E, k
= V (E, k 1) + pM x(1 p)M
= V (E, k 1) + M (p x(1 p)).
1)]
Therefore,
V (E, k) = kM (p
x(1 p)), k 0
and hence
V (F, k) = kM (p x(1 p)) + (1
)M + xM (1
pk ), k 1.
70

Finally, let us consider medium penalties, in other words
p < x < max{1, p/(1
p)}.
When k < k
V (E, B, k) V (E, N, k) = M (p x(1
> 0,
pk ))
V (E, B, k) V (E, N, k) = M (p x(1

< 0.
pk ))
and when k k
In other words, far from maturity it is optimal to not bid when empty, but closer to maturity it is
optimal to bid when empty. If we consider that x(1 pk ) is the expected proportional penalty
paid with k time periods remaining, and that this is an increasing function with respect to k,
then it makes sense that with a large number of time periods remaining the decision is to not
bid. If k is close to maturity then the expected proportional penalty paid out is smaller and so it
is possible to get away with paying a penalty. When k < k solving the difference equation
is identical to the case when penalties are small, i.e.
V (E, k) = kM (p
x(1
p)), k < k
and
V (F, k) = kM (p x(1 p)) + (1 )M + xM (1 pk ), k < k .
When k > k we can use the initial condition V (E, k ) = k M [p
x(1
p)] and write
V (E, k) = k M [p x(1 p)] + (k k )pM [1 (1 p)]

= kpM k M x(1 p) (k k )M p(1 p), k k
and hence
V (F, k) = V (E, k) + M (1 (1 p))
= kpM k M x(1 p) (k k )M p(1
p) + M [1
(1
p)].
5.2 Discussion and Summary of Results

In this chapter we were able to show that, under simplified conditions, it is always optimal to
bid when the storage facility is full. Moreover, when penalties are small it is optimal to bid
when the storage facility is empty and when penalties are large is it optimal to not bid when
the storage facility is empty. When penalties are moderate, it is optimal to not bid far from
maturity and optimal to bid closer to maturity. The point at which the decision switch occurs
can be determined.
5.2. D ISCUSSION AND S UMMARY OF R ESULTS
71
Now that we understand how to operate a storage facility if one were provided, how much
value is being added by the storage? Let us consider a simple wind turbine with no storage
unit; the value of the turbine with k periods remaining is
W (B, k) =
=
W (N, k) =
=
W (k) =
=
=
p[M + W (k 1)] + (1 p)[ xM + W (k 1)]

W (k 1) + (p x(1 p))M
pW (k 1) + (1 p)W (k 1)
W (k 1)
max{W (B, k), W (N, k)}
W (k 1) + M max{p x(1 p), 0}
kM max{p x(1 p), 0}
since W (0) = 0. From this it is clear that if x < p/(1 p) then the optimal strategy is to
always bid, yielding W (k) = kM (p x(1 p)). If x > p/(1 p) then it is optimal to never
bid, with W (k) = 0. In other words, large penalties destroy the value of wind in the absence
of storage. Looking back at the solution of the difference equation when penalties are large,
we are able to state that for a k-period storage facility, the added value from storage is
kpM [1
(1
p)]
under the assumption that the storage is initially empty, and

(k + 1)pM [1
(1
p)]
under the assumption that the storage is initially full. Both equations imply that more wind is
better and a more efficient storage facility is better (captured by ). A storage facility which
will last longer is also better; this especially makes sense if the purchasing cost of such a facility
is considered.
It is already obvious that factors such as storage life and efficiency are crucial in understanding and considering storage as a solution to winds intermittency and unreliability. In the
next chapter we will talk more about available storage options, and determine the key parameters involved in properly modeling a wind turbine coupled with a storage facility. In addition
we give an overview of wind, as well as the process of converting wind to power via a turbine. Once we have looked at real data we will return to the analytic solutions described in this
chapter to better understand the results presented thus far.
Chapter 6
Wind and Storage - An In Depth Analysis
As mentioned in Chapter 2, the wind doesnt always blow when we want it to. Even when it
does blow, we cannot always produce power from it. This is a more complicated assumption
than the one presented in Chapter 5. Wind power is the conversion of wind energy into a more
useful form, such as electricity via wind turbines. There are many different types of wind
turbines currently in use, each having different characteristics such as height, blade length,
efficiency and range of wind speeds at which they can operate. The latter will be discussed in
more detail in the next section.
In this Chapter we will examine hourly wind speed data for two different regions, Nantucket
Sound in the United States as well as Valkenburg in the Netherlands. We will investigate the
properties of wind speed as well as how wind is converted into power. We will then provide an
overview of current storage technologies, and provide a rationale for choosing battery storage
throughout the rest of the study.
6.1 Part One: Wind

Wind speeds are typically measured and recorded in m/s, and wind power is typically given in
megawatts (MW). The range of wind speeds for which we can produce power varies slightly
from one wind turbine to another, but the approximate cut in wind speed for a wind turbine is 45 m/s. When the wind speed is less than the cut-in speed , typically about 4-5 m/s, the turbine
cannot produce power. Similarly, there is a maximum wind speed above which power cannot
be produced. This cut-out speed is usually around 25 m/s. Above the cut-out speed, there is
a risk of turbine damage if power production continues; all modern turbines are equipped with
automatic shut down features to avoid this. The maximum output of an average turbine occurs
at 13-14 m/s; this means that wind speeds of 14-25 m/s are not more desirable then a wind
speed of 13 m/s since the same amount of power is produced [1]. In this study we consider
a Vestas 1.8 MW turbine, meaning that in one hour the turbine can produce a maximum of
1.8 MWh of electricity. This is called the rated capacity of the turbine. Figure 6.1 shows the
power curve for this particular turbine. All power curves have the same basic shape; when the
wind speed is very low, the curve is at zero, indicating no electricity being produced. When
the wind speed reaches the cut in speed the graph increases monotonically. Once maximum
output speed is reached, the graph remains flat until the cut out speed. If wind speeds beyond
72
6.2. S TATISTICAL P ROPERTIES OF W IND S PEED DATA
73
this speed are reached, the turbine shuts off and we can no longer produce power, hence the
sharp decrease to zero.
2000
1800
1600
Power (kW)
1400
maximum output
1200
1000
800
steep power law increase
600
400
cut in speed
cut out speed
200
0
10
15
Wind Speed (m/s)
20
25
Figure 6.1: Power Curve - Vestas 1.8 MW Turbine [38]
6.2
Statistical Properties of Wind Speed Data
There are 8760 hours in a typical year, 8784 hours in a leap year. Figure 6.2 shows one year
(8760 hours) of wind speed data for Nantucket Sound in the United States (measured in m/s).
Figure 6.3 shows the same amount of data, over the same period, in the Netherlands. Both
data sets appear completely random, and it is difficult to gain insight from these figures alone.
However, from these figures it is certainly clear that wind speeds can vary a great deal from day
to day, and even hour to hour. The range of wind speeds shown in both figures is approximately
0 - 20 m/s.
74
C HAPTER 6. W IND AND S TORAGE - A N I N D EPTH A NALYSIS
25
Wind Speed (m/s)
20
15
10
1000
2000
3000
4000 5000
Hour
6000
7000
8000
9000
Figure 6.2: Hourly Wind Speed Data - Nantucket Sound, 2003 [77]
75
20
18
16
Wind Speed (m/s)
14
12
10
8
6
4
2
0
1000
2000
3000
4000 5000
Hour
6000
7000
8000
9000
Figure 6.3: Hourly Wind Speed Data - The Netherlands, 2003 [43]
76
6.2.1 Historical Data

Before we perform a statistical analysis of the data it is important to note that wind speed data
was historically not collected for power needs. Historically wind speed data has been used
primarily for aviation and marine needs. For example, some major meteorological centers may
have recorded wind speed data in knots rather then meters per second. Recording wind speed
as a whole number in knots may have been sufficient for aviation and marine purposes in many
areas, however when converting this whole number to m/s the data loses a great deal of the
statistical properties and is not useful to power producers. An example of this can be found
in [52]; the wind speed data is observed and recorded in knots, expressed as a whole number.
It is then converted to km/h, again as a whole number, and then archived. Due to the unit
conversion and rounding that occurs throughout the data collection process, the appearance of
certain whole number km/h values is impossible and is clearly not indicative of the true data.
The height of the measuring device, called an anemometer, may also influence the relevance
of the data to wind producers though approximate conversions exist. Therefore, it is essential
to evaluate the usefulness of the historical data before it is used to make important decisions
about the future.
A third issue with historical wind speed data is that it may be incomplete; missing data
points are not uncommon. The Nantucket Sound data is an example of incomplete data. The
reason for missing data points is sometimes unknown, but may be caused by turbine anemometer failure [77]. This failure could result from icing issues or poor tolerance to gusting winds.
In cases where there are multiple readings at a given time, missing data could result from obtaining two very different results at a given location, hence questioning the reliability of the
readings. Other issues surrounding historical data will be discussed throughout this chapter as
they are encountered.
6.2.2 Distribution of data

If instead of the time series we examine a histogram of the data, we observe some additional
information which is quite meaningful. Wind speeds are often found to fit a Weibull distribution, which is commonly used in weather forecasting (see [32] for a study of fitting wind speed
distributions). The probability distribution function for a Weibull distribution is given by
!k
x
k x k1
f (x; k, ) =
e
, x > 0.

Here, > 0 is a scaling parameter and k > 0 is a shape parameter. In particular, wind speeds
are often found to fit a special case of the Weibull distribution which occurs when the shaping
parameter k is 2. This particular distribution has its own name: the Rayleigh distribution. The
pdf of the Rayleigh distribution is given by
f (x; k, ) =
2x
2
2
x
p
Using the fact that the mean of the Rayleigh distribution with parameter is /2 we can
use our data to obtain an approximation for . Using Dutch data from 2001 to 2007 we obtain
77
4.12. We can now use this to plot the pdf, as well as the frequency of wind speed. This is
shown in Figure 6.4. From this figure it appears that the data fits the distribution quite well. We
can in fact gain a better sense of how well the data matches the distribution by producing a Q-Q
plot. A Q-Q plot displays a quantile-quantile plot of the sample quantiles of our data versus
theoretical quantiles from a Rayleigh distribution. If the distribution of our data is indeed the
Rayleigh distribution the plot will be close to linear. Figure 6.5 shows this plot and suggests
that the data is appropriately modeled using a Rayleigh distribution. However although the
Rayleigh distribution does appear suitable for a wide range of wind speeds, we do not obtain a
close fit for high wind speeds. This is clearly shown in the Q-Q plot.
Repeating this procedure using the US data gives a slightly different picture. The parameter
is estimated to be 5.94, and plotting the pdf and frequency of wind speed on the same
axis (Figure 6.6) it is clear that while the shape of the frequency plot appears to match the pdf,
it does not look as impressive as the Dutch data. This could certainly mean that the Rayleigh
distribution is simply a better fit to the Dutch data, however it is important to point out that there
was a considerable amount of missing US data. Therefore it is difficult to make a definitive
conclusion regarding the suitability of the Rayleigh distribution based on this figure alone. To
further investigate the Q-Q plot is also shown in Figure 6.7. As in Figure 6.5 this plot suggests
that for low wind speeds the Rayleigh distribution is a nice fit to the data. This is not the
case for high wind speeds. Therefore we also revisit a statistical tool used earlier in the thesis:
bootstrapping. Bootstrapping, as well as the impact of missing data, will be discussed later in
this chapter.
6.2.3
Autocorrelation
Wind speed data also exhibits large autocorrelations; this means that the data is highly correlated with a time shifted version of itself. This is not surprising since, although wind speeds
fluctuate, the wind speed at a given hour is a good prediction of the wind speed in the next hour.
Autocorrelation is an important aspect in understanding wind speed data since, when plotted,
wind speed data looks rather noisy. Figures 6.8 and 6.9 show the autocorrelation as a function
of the time shift for the data sets explored earlier. For short time shifts, the autocorrelation is
close to 1, indicating nearly perfect correlation. As the time shift increases, the autocorrelation
decreases. It does not monotonically decrease however, but rather oscillates. In the case of
the Dutch data, the autocorrelation oscillates with a period of approximately 24 hours. This is
since, and this is true for many locations where wind speeds are of interest, wind speeds are
fairly consistent from day to day. In many locations wind speeds tend to be higher at night, and
so the wind speed at midnight on a given day is often a good indication of the wind speed at
the same time the following day. However, due to seasonal changes in wind, etc, it may not be
a good indication of the wind speed at the same time several months from now. Therefore, we
see the oscillating behaviour, but the overall trend is decreasing. In the case of the U.S data, the
oscillating behaviour is not so clear. Although the trend is the same - monotonically decreasing
as the time shift increases followed by oscillating behaviour - the oscillatory behaviour does
not appear to have a constant period and is not as smooth as the Dutch data. There are many
possible explanations for this, the most obvious being the fact that each data set corresponds
to a completely different part of the world, and the data will certainly not be identical. The
Netherlands also has a more established wind power industry and because of this wind speed
78
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
10
15
20
25
Figure 6.4: Frequency of wind speed and Rayleigh pdf - Dutch Data, 2003 [43]
79
25
Y Quantiles
20
15
10
10
15
20
25
X Quantiles
Figure 6.5: QQ plot of 2003 Dutch data (X) vs. a Rayleigh distribution (Y ) [43]
80
0.14
0.12
0.1
0.08
0.06
0.04
0.02
10
15
20
25
30
Figure 6.6: Frequency of wind speed and Rayleigh pdf - US Data, 2003 [77]
81
35
30
25
Y Quantiles
20
15
10
5
0
5
10
15
X Quantiles
20
25
30
Figure 6.7: QQ plot of 2003 Nantucket Sound data (X) vs. a Rayleigh distribution (Y ) [77]
82
data collection is very important. Wind conditions support the development of a large number
as well as large scale wind farms, both onshore and offshore [53]. Therefore, the data may simply be more reliable than the U.S data which contains many missing data points, as previously
mentioned. Missing data is the next topic of this chapter.
1.2
Autocorrelation
0.8
0.6
0.4
0.2
0.2
50
100
150
200
250
Time shift
Figure 6.8: Autocorrelation vs Hourly Time Shift - Dutch Data, 2003 [43]
6.3 Missing Data

Handling missing data is an important and diverse field of statistical research (see [48] for more
details) since it is a common occurrence. Since missing data can potentially lead to misleading
and inaccurate results, a variety of statistical techniques have been developed to deal with
this important problem. To obtain a robust analysis in spite of missing data, it is sometimes
possible to delete data which may cause misleading results (individual data points or a series of
6.3. M ISSING DATA
83
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
50
100
150
200
250
Figure 6.9: Autocorrelation vs Hourly Time Shift - US data, 2003 [77]
84
data points depending on the properties of the data). It is also possible to fill in the missing
data using a simple or complicated set of rules or statistical procedure. For instance, a missing
data point could be computed by a simple interpolation.
In this thesis, missing data is handled in two ways: if only one (or a small number) of data
points is missing, the data from the exact time on the previous day is used. It is important
to note that the raw data for the U.S is 10 minute data, and the hourly data is determined by
taking the average of the 10 minute data. Therefore, a missing data point corresponds to a 10
minute time interval rather than a one hour interval. Replacing a missing data point with the
corresponding data from the previous day ensures the proper diurnal pattern for the data, but
may affect the autocorrelation, especially in the event of a large number of missing points. If
a large number of data points are missing (for instance, several days or weeks of data) the data
is omitted. Since the wind speed data shows significant autocorrelation from hour to hour and
day to day, it is difficult to delete a few days of data while keeping the autocorrelation trend
consistent. Therefore if we are able to choose a range of data for which the missing data points
are few and scattered, we can avoid deleting data and minimize the effects that replacing data
has on the autocorrelation.
Fortunately data obtained in 2003 is a good example of this; there are a total of 52560 data
points in a given year (in the case of a leap year the total increases to 52704) and the number
of missing data points is 972. This means that 1.85% of the data is missing, in comparison
to 2006 when 32% of the data is missing or 2007 when 8.25% of the data is missing. Figure
6.10 is a plot of the 2003 10-minute data, with the location of the missing data points indicated
in red below. It is clear from this figure that although some of the missing data points occur
nearby, they are generally scattered throughout the time series. For these reasons, the 2003
data is used in our bootstrapping calculations; the bootstrapping methodology is the final topic
discussed in this chapter.
6.4 Bootstrapping
To move beyond a historical analysis of wind data and begin an analysis of our wind-storage
system, we rely on re-sampling methods. In particular we use a block bootstrap method [18].
The idea of a block bootstrap is similar to that of a standard bootstrap, however instead of
sampling (with replacement) one data point, we sample a block of consecutive data points.
The advantage of the block bootstrap is that, if the proper block size is chosen, it captures
the dependence between neighbouring data points. If we consider our historical wind speed
data, we see that the autocorrelation is very high for small time lags. In the case of the Dutch
data the autocorrelation oscillates with a period of approximately 24 hours, however it is still
fairly strong beyond this point, decreasing substantially between 24 and 48 hours. In the case
of the U.S data the autocorrelation decreases to nearly zero more drastically at a time lag of
approximately 48 hours. This is shown in Figures 6.8 and 6.9. Therefore, choosing blocks of
length 48 will help ensure that our bootstrapped data has the same autocorrelation properties
as the data sets explored in this Chapter.
After choosing a block size, we ask ourselves two additional questions: should the blocks
overlap, and how fast is the stochastic structure of the data changing? The first question is
relatively easy to answer; in order for the autocorrelation to be consistent at the joining of two
6.4. B OOTSTRAPPING
85
25
Wind Speed (m/s)
20
15
10
3
Hour
6
4
x 10
Figure 6.10: Nantucket Wind Speed Data, 2003 (location of missing data shown in red)
86
blocks, the blocks should not be able to overlap. If overlapping is permitted, there may exist
jumps in the autocorrelation at the joining of each block. While the autocorrelation is consistent
with the historical data when the time lag is small, it is not consistent when it exceeds the block
size.
Our initial wind speed data sets are quite large. Therefore, it may not be practical to assume
that the time series are stationary; a time series is stationary if its statistical properties do not
depend on time. It might be more realistic to assume the data exhibits a slowly changing
stochastic structure [63]. If we consider a year of hourly wind speed data, this assumption
makes sense; the data corresponding to the summer months looks quite different than the data
corresponding to the winter months. Therefore, rather than simply perform block bootstraps in
the remaining chapters, we choose local block bootstrapping [56]. This means that we sample
blocks which are relatively nearby rather than any block in the entire time series. We choose to
sample blocks which are within 20 blocks from the block in question. Using a block size of 48
hours, or two days, this means we are resampling data which is within 40 days of the current
data, giving us a sampling range of 80 days which is nearly 3 months, or one full season.
Increasing the block size a great deal beyond becomes more and more unrealistic, as well as
limits the effectiveness of the bootstrap since there are less blocks to bootstrap from.
The U.S data is examined in the chapters that follows, despite the patchiness of this data.
The similarities between the U.S and Dutch data, as seen in Figures 6.5 and 6.7, reassure us that
the U.S data is appropriate for our purposes, and in a sense the Dutch data has helped validate
the flawed U.S data. This is especially convenient given our interest in North American power
markets, as well as the fact that we have corresponding electricity price data.
6.5 Summary
In part one of this chapter an overview of wind, focusing on the statistical properties of wind
speed data, was given. It is noted that the frequency of wind speed data follows a Rayleigh
distribution. It is also determined that the data exhibits strong autocorrelation, especially when
the time shift is small. For larger time shifts the autocorrelation oscillates, and in the case of
the Dutch data the period of the oscillation is approximately 24 hours although both data sets
exhibit high autocorrelation out to 48 hours. This is not surprising given the consistency of
wind from day to day in many areas.
Part one of this chapter also touched on the importance of high quality data, which is often
difficult to find due to the historical usage of wind speed data. In particular, major challenges
exist due to data rounding and unit conversion, as well as incomplete time series.
Equipped with the relevant wind data and knowledge, we can now move forward to part
two: storage technology options.
6.6 Part Two: Storage

A solid understanding of wind speed data is essential in developing a data driven model to
address the problem of intermittent wind. However, of equal importance is understanding
the technology which will be used to achieve this. In part two of this chapter we provide
6.6. PART T WO : S TORAGE
87
an overview of a variety of storage technologies in current use. Once a technology has been
chosen, the modeling details related to the choice will be discussed.
6.6.1
Storage Options
There are several options available when considering a storage system for wind [29]. No matter
which technology is used, the goal is the same: to generate and store energy when the wind is
blowing (typically at off peak periods) and release it when there is no wind (during peak load
periods).
One storage option is pumped hydro. Wind power is used to pump water from a lower to
higher elevation, creating potential energy. The water is then released through a turbine at peak
load periods. See [22, 73, 79] for more details on optimal control strategies of pumped hydro.
A clear disadvantage of pumped hydro storage is that it is location dependent, and therefore it
may not be feasible to build a pumped hydro facility near a wind farm. However, in situations
where it is feasible to build wind turbines near existing pumped hydro facilities this option can
be both economically and environmentally attractive.
Many studies have been conducted on the feasibility of buffering wind power with pumped
hydro, including several North American studies [8, 9, 24, 39, 40] as well as several European
studies [14, 5, 17, 33]. A large number of these studies focus on engineering constraints, such
as transmission, which are not considered in our research. Penalties, where discussed, are
generally taken as an input from the system. Overall, the focus of these papers is not to create
a solvable model as was done in Chapter 5, but to develop an accurate picture of the system
being discussed. In some cases optimization software is used to solve the system. Therefore
the insights drawn are different then those seen in Chapter 5.
Another option is compressed air energy storage (CAES). These systems, which are typically underground tanks or abandoned mines, have advantages and disadvantages similar to
pumped hydro storage. An added disadvantage of CAES relative to pumped hydro is that gas
fuel is required [29]. Studies are currently being conducted to determine the feasibility of a
large CAES facility in southwestern Ontario. In addition to several depleted oil and gas wells
which could be used to store compressed air, the area also has strong wind resources [44].
Perhaps the most accommodating option is battery storage. There are many battery types
to explore, each having a different set of parameters related to efficiency, cost, size and so on.
The clear advantage of battery storage is that they are not location dependent. However, there
are concerns about the economic feasibility of some battery types.
In [46], dynamic programming is used to carry out the optimization of wind and storage
over a specified scheduling period (a 48 hour period in the detailed case study). Similar to several other models mentioned above, transmission constraints are considered however the model
is general enough to cover intermittent power sources other than wind. The paper focuses on
Nordic market rules, and simulation results are given. Analytic solutions are not possible.
Figure 6.11 shows system ratings for a variety of storage systems, including pumped hydro,
CAES and various battery storage technologies. From this figure it is clear that pumped hydro
has the highest power rating and discharge time. However, as previously mentioned pumped
hydro systems are not always feasible due to their dependence on location. Sodium sulfur
(NaS) and lead acid (L/A) batteries are comparable to CAES in terms of their power rating,
whereas lithium ion (Li-Ion) batteries have a lower rating. All three battery types have similar
88
possible discharge times, although the range for NaS batteries is much smaller. Various battery
types, including NaS, L/A and Li-Ion are discussed in the next section.
Figure 6.11: System Ratings, Installed systems as of November 2008 [29]
6.6.2 Battery Storage

Batteries are often compared in terms of their capacity, life cycle, and power rating. The
capacity of a battery is the maximum energy which can be extracted from the battery, measured
in kWh. Many batteries cannot be fully discharged due to possible damage to the battery; the
depth of discharge is the amount (usually expressed as a percentage) of the total capacity which
is withdrawn in a given cycle. The rate of discharge is called the power rating, measured in
kW. Assuming a partially charged battery, once some of the energy has been withdrawn and
then charged again to some other level, a full cycle of the battery is complete. The life cycle of
the battery is the number of charge-discharge cycles a battery can complete before replacement
6.7. M ODELING A BATTERY USED FOR S TORAGE

wt
ft
wtp
ftp
wte
wts
89
realized wind at time t (m/s)

forecasted wind at time t (m/s)
realized wind power at time t (kWh)
forecasted wind power at time t (kWh)
excess wind power at time t (if any) (kWh)
shortfall of wind power at time t (if any) (kWh)
Table 6.1: Wind Variables and Parameters
is needed. The efficiency of the battery is the amount of energy required to charge the battery
compared to the energy that can be extracted during discharging [7, 65].
Some of the most frequently used and discussed batteries for energy storage are lead acid,
nickel-cadmium, lithium ion and sodium sulfur batteries (see [7, 65, 66, 28] for a detailed
comparison of these and other battery types). Lead acid batteries are relatively inexpensive,
however they also have a low life cycle and cannot be discharged to less than 40% of their
total capacity. Nickel-cadmium batteries offer a high discharge rate and longer life cycle then
lead acid batteries, however due to their high cost they are often not feasible. Lithium ion
batteries are becoming very popular due to their durability and high energy density. Unlike
most rechargeable batteries, the life of the battery is not decreased by cycling at low depths of
discharge. However, their cycle life is lower then many other rechargeable batteries and they
are not as durable. The production cost of these batteries is currently very high as well. In this
study we consider sodium sulfur batteries, which are considered the most economically feasible
at the present time. They have a high life cycle and high efficiency and are also environmentally
attractive. In the future, lithium ion batteries may be a better alternative.
6.7
Modeling a Battery used for Storage
The first step in determining the value of a wind/storage system is to create a battery model
which incorporates the major parameters which influence the value. The model must take
into account battery efficiency, capacity, power rating, depth of discharge and various costs
such as purchase price as well as operations and maintenance. Tables 6.1 and 6.2 provide
a complete list of battery parameters, as well as other variables and parameters that will be
needed throughout Chapters 7 and 8.
From Table 6.2 it is clear that modeling a battery is complex; determining an optimal operating strategy for the battery given the complete list of parameters will be at the very least
challenging. It is therefore beneficial to carry out a preliminary study which is able to shed light
on which battery parameters are most important. It is also important to have an understanding
of what values are reasonable for the key parameters involved. For instance, NaS batteries have
a round-trip efficiency of approximately = 80%. A common assumption is that the
charging
efficiency and discharging efficiency are equal, implying a one-way efficiency of = 89%.
The battery costs approximately 150-300 USD/kWh and operations and maintenance costs are
approximately 5-10 USD/kWh. [28]. The number of cycles the battery can make before replacement is dependent upon the depth of discharge of each cycle, given in Table 6.2 by i .
90
st
stored power at time t (kWh)
et
electricity price at time t ($/kWh)
eh
mean electricity price at a given hour h
c
opportunity cost due to power rating constraint ($) - charging
c,
opportunity cost due to both power rating and size constraints ($) - charging
c
opportunity cost due to battery capacity constraint ($) - charging
ch
opportunity cost due to heuristic ($) - charging
d
opportunity cost due to power rating constraint ($) - discharging
d,
opportunity cost due to both power rating and size constraints ($) - discharging
d
opportunity cost due to battery capacity constraint ($) - discharging
dh
opportunity cost due to heuristic ($) - discharging
Rt
Revenue at time t ($)
battery cycle life (number of cycles)
battery efficiency (%)
battery capacity (kWh)

i
depth of discharge of the ith cycle (% of )
minimum state of charge (% of )

min
minimum fill level (% of )
min
minimum level of discharge (% of )
power rating (kW)

max
maximum power, equal to (kW)
max /
operations and maintenance cost (/kWh)
acquisition cost (/kWh)

N
number of battery cycles
ai
relative frequency of ith depth of discharge
g(i )
cycles to failure given depth of discharge of each cycle is i
pt
penalty function
Table 6.2: Battery Model Variables and Parameters
6.8. S UMMARY
91
For a battery with a primary purpose of avoiding under-delivery, the depth of discharge should
not be fixed but should vary for each cycle of the battery. Therefore may be dependent upon
the delivery requirements at a given point in time, which is dependent upon the accuracy of
the wind forecast at that time. Depth of discharge, as well as forecasting wind speed, will be
discussed in more detail in the next chapter when some possible operating strategies are laid
out.
6.7.1
Battery Size
A very obvious and important question to ask when considering any storage technology is how
big it should be. Determining a reasonable battery size, , is not an easy task. Clearly bigger
is better in the sense that using a big battery means that little or no wind will be spilled due
to insufficient storage space, however it also means a higher purchase price. It is important to
consider the advantages and disadvantages of a given battery size before deciding whether or
not it is reasonable.
The approach taken in this study is to avoid restricting to a single value, and instead
to implement the model for a variety of battery sizes. This allows us to determine a suitable
range of values for which we are able to improve the usability of wind without overspending
on storage, i.e. the goal is to make our delivery commitment most of the time without an
unrealistic price tag relative to the fraction of the battery actually being used on average.
6.8 Summary
Part two of this chapter provided an overview of possible storage technologies relevant to wind
power. Due to the geographical restrictions present in the placement of wind farms as well
as pumped hydro or compressed air energy storage, battery storage is the main focus. The
parameters involved in modeling a battery were outlined and, due to the large number, it is
obvious that determining an optimal operating strategy is not an easy task; in fact it might be
useful to simply pinpoint the key parameters and develop operating strategies which account
for these parameters. This is the goal of the next chapter; it combines the findings of our studies
on wind speed and battery storage and suggests possible operating strategies for a wind/storage
system with a carefully chosen set or range of key parameter values.
Chapter 7
Real Options Valuation of a Wind/Storage
System
In Chapter 5 a simple model of a wind-storage system was developed. The model consisted
of four parameters, namely the value of a unit of power, M , the fractional loss in storing the
power, , the probability of wind (or equivalently the probability of generating power), p, and
the penalty for bidding electricity into the market but not delivering, x. The model was a
nonlinear system of difference equations which could be solved in closed form. In Chapter
6 wind speed data and storage technologies were explored. An analysis of wind data was
provided and a variety of battery technologies, including the parameters involved in modeling
a battery, were given.
In this chapter the key battery parameters discussed in Chapter 6, along with wind speed
data findings, are combined into a more realistic model of a wind-storage system. Two viable
operating strategies are introduced and explored in terms of their ability to mitigate the problem
of wind variability and uncertainty. Results are based on the bootstrapping approach outlined
in Chapter 6.
7.1 The Simple Model Revisited

In this section we run our simple model from Chapter 5 with numerical estimates of its parameters, mainly discussed in Chapter 6. Given the 2003 Nantucket Sound data used in chapter 6
we are able to obtain an estimate for p, the probability of wind. Using the approximation that
wind speeds of approximately 5 m/s to 25 m/s are suitable for power production, we assume
that wind speeds in this range indicate a windy day, and wind speeds outside this range indicate
a calm day. In this case p = 0.75. However Figure 6.1 indicates that when wind speeds are
near the cut-in speed little power can actually be produced, and hence this estimate is likely
quite high. It might be more reasonable to indicate a windy day using a smaller range of wind
speeds, for instance 10 m/s - 25 m/s. In this case p = 0.22. Table 7.1 shows the number
of windy days as well as the probability of wind under given windy day assumptions. To
illustrate our simple model we use p = 0.39 corresponding to the 8-25 m/s window.
From Chapter 6 we also have an estimate of for a sodium sulfur battery. We use the
round trip efficiency of = 80%, or in other words a fractional loss due to storage of = 0.2.
92
7.2. A M ORE R EALISTIC M ODEL

Windy Day Assumption
5-25 m/s
6-25 m/s
7-25 m/s
8-25 m/s
9-25 m/s
10-25 m/s
93
Number of Windy Days
6585
5543
4488
3446
2615
1938
p
0.75
0.63
0.51
0.39
0.30
0.22
Table 7.1: Number of Windy Days and p estimate under various windy day assumptions (2003
Nantucket Sound data)
Although power prices were not discussed in Chapter 6, they are discussed later in this chapter.
For this illustration we simply choose an average price of $0.04/kWh [77]. The only parameter
we cannot obtain an estimate for based on past data is x; in Canada x = 0 at present and in the
U.S. x > 0, however the actual penalty data is not available. Therefore, we explore the affect
of x on the value of the wind-storage system.
When penalties are large the value is unaffected by the actual value of x; this is since the
optimal strategy avoids these penalties by not bidding when the storage facility is empty. Figure
7.1 shows V (F, k) and V (E, k) as the number of periods remaining, k, increases. As expected
V (F, k) > V (E, k) k. When penalties are small the value is indeed affected by the actual
value of x. This is since the optimal strategy is to always bid and hence the penalty must be
paid on calm days when the storage facility is empty. Figure 7.2 shows V (F, k) and V (E, k),
k = 100, for 0 x p = 0.078. From this figure we see that as the penalty level increases,
V (E, k) and V (F, k) decrease. As expected V (F, k) > V (E, k) k. Finally, for medium
penalties, we consider k > k . Since k is dependent on x, we assume k = k + 10. For the
parameter set considered, medium penalties include 0.078 < x < 0.6393. This is shown in
Figure 7.3. When penalties are near p = 0.078, V (E, k) and V (F, k) decrease more rapidly
then when penalties are near p/(1 p). In fact, as penalty levels approach p/(1 p) values
are no longer affected by the level of penalty.
The results of this section show that the never bid when empty condition of Chapter 5,
x(1 p) > p, holds for penalty rates greater than about x = 0.13. In other words severe nondelivery penalties are not required to induce the use of storage. However, as seen in Chapter
6, a realistic wind storage model will be much more detailed than the simple four parameter
model of Chapter 5. In the next section we embark on the construction and analysis of a
much more realistic storage model, which we investigate using heuristic rules motivated by the
proven optimal rules on the simpler system.
7.2 A More Realistic Model

The analysis carried out in Chapter 6 clearly illustrated that 4 parameters were not sufficient
for a realistic wind-storage model, and that analytic solutions to such a realistic model would
likely not be found. In order to move forward in developing a realistic model for valuing a
wind-battery system some modeling choices still need to be made. For instance, there must be
94
C HAPTER 7. W IND /S TORAGE S YSTEM VALUATION
1.8
V(E,k)
V(F,k)
1.6
1.4
V(X,k), X = E,F
1.2
1
0.8
0.6
0.4
0.2
0
20
40
60
80
100
k
Figure 7.1: V (F, k) and V (E, k) as k increases: p = 0.39, = 0.2, M = 0.04
7.2. A M ORE R EALISTIC M ODEL
95
1.6
V(E,k)
V(F,k)
V(Y,k), Y={E,F}
1.55
1.5
1.45
1.4
1.35
0.01
0.02
0.03
0.04
x
0.05
0.06
0.07
0.08
Figure 7.2: V (F, k) and V (E, k) for 0 x p: p = 0.39, = 0.2, M = 0.04, k = 100
96
0.3
V(E,k)
V(F,k)
0.28
0.26
V(Y,k), Y={E,F}
0.24
0.22
0.2
0.18
0.16
0.14
0.12
0.1
0.2
0.3
0.4
0.5
0.6
0.7
x
Figure 7.3: V (F, k) and V (E, k) for p x p/(1
p): p = 0.39, = 0.2, M = 0.04
7.3. F ORECASTING W IND S PEED
97
a forecasting model in place which is used to determine how much wind is bid into the market;
our simple model assumed one unit of power regardless of the actual wind speed. There
must also be a model for electricity prices; again, our simple model assumed a constant price
M . Finally, we must be able to track the fill level of the battery at a given time. In Chapter
5 it was assumed that one unit of wind filled and emptied the storage, which is clearly not
sufficient. These indicators for control strategies will also be important in determining which
of the many parameters involved in the modeling process are most important. This will also
give a better indication of how reliable, if at all, our simple model is at predicting the outcome
of the more sophisticated and realistic model. We do not attempt to solve a full fledged optimal
control problem; instead we look at plausible control heuristics and get a better understanding
of which parameters have the largest influence on value. These operating strategies have the
necessary physical aspects, as well as public policy issues, built into them.
We begin this chapter by outlining our wind forecasting assumptions and discussing our
simple model of electricity prices. We then describe the operating strategies which will be
explored for a range of parameter values in order to pinpoint the key parameters of the model.
These strategies are compared in terms of their ability to increase the predictability of wind.
7.3 Forecasting Wind Speed

It is difficult to develop an accurate forecasting model for wind speed. For short term predictions (typically in the magnitude of a few hours) a persistence forecast, a forecast which
assumes that future conditions will be the same as the present conditions, is often used [3]. For
other short term forecasting models, see [51]. This approach is not suitable for longer term
forecasting, so a great deal has been invested in improving this simple strategy (see [3] for
details). For a comparison of various techniques applied to mean hourly wind speed data, see
[69].
In this study of wind and storage, we do not attempt to develop an improved forecasting
model for wind speed. Instead, we use a persistence model which assumes that tomorrows
hourly forecast is todays realized hourly wind speed. The reasons for using day ahead forecasting are straightforward; power producers offer their power on the day ahead market based
on what they believe they are able to provide for a certain period the following day. Since
autocorrelation in wind speed data is fairly strong 24 hours ahead, as seen in Chapter 6, the
day ahead persistence forecast is a reasonable baseline forecast which, while not perfect, is
extremely easy to implement.
7.4 Power Prices

Many complex models of electricity prices have been developed in recent years due to the
deregulation of electricity markets. Electricity prices are typically mean reverting as well as
very volatile; price spikes are not uncommon. Recent models typically account for both mean
reversion and price spikes (see [4, 26, 37]) for more details on some of these models).
In 2009, Ontario introduced a feed-in tariff (FIT) program, enabled by the Green Energy
and Green Economy Act [54]. This program provides a guaranteed pricing structure for the
98
production of renewable electricity, including onshore wind. Therefore, producers of renewable energy in Canada are not currently subject to the volatility of prices but instead receive
fixed prices for the electricity they produce. The purpose of this program is to help cover acquisition and operating costs, as well as provide a reasonable rate of return over a specified
contract term.
Because of this feed-in tariff it is not necessary to use a complex model for electricity
prices; instead we use historical prices in order to determine the average price of electricity at
a given hour. This is shown in Figure 7.4. It is important to note, however, that to accurately
assess the cost of this program a model of this type would be required.
0.06
0.055
Price ($/kWh)
0.05
0.045
0.04
0.035
0.03
0.025
0.02
10
15
20
25
Hour
Figure 7.4: Mean Electricity Price at Each Hour of the Day (hour zero is 12am) [77]
7.5. BATTERY C HARGE /D ISCHARGE H EURISTICS
7.5
99
Battery Charge/Discharge Heuristics
We discuss two heuristic control strategies here. It is important to note that we have not proven,
nor do we believe, that either of these heuristic strategies are optimal - but we do believe they
are reasonable guesses at what a decent approximate control strategy might be. In both cases,
we assume that the battery is never discharged to less than 10% of the total capacity [28].
Discharging below this level may cause damage to the battery. We also assume that when the
battery is purchased, its charge level is 50% of the total capacity . This eliminates the need to
use wind energy to initially fill our battery.
The base case is determined by examining the frequency of depths of discharge under a
given heuristic. A bigger battery results in a smaller depth of discharge (and a smaller probability of failing to meet the delivery requirement), but also a higher acquisition cost. Therefore,
if possible, our base case should allow us to make our delivery commitment most of the time
but doesnt have an unrealistic price tag relative to the fraction of the battery actually being
used on average. For a given we compute the number of charge-discharge cycles and record
the depth of discharge of each cycle. We also examine the total excess and total shortfall over
short time periods (e.g. 48 hour intervals) to ensure that allows us to absorb the excess/cover
the shortfall over this short time.
Given these criteria, it is determined that = 10000 kWh might be a reasonable battery
size. It is reasonable in the sense that it is large enough to avoid under-delivery most of the
time. We do not limit ourselves to this single value of , but use this capacity as a default
and compare other scenarios to this case. Given this value of the power rating is chosen
to be 10% of the capacity, or 1000 kW. Many battery types have a considerable amount of self
discharge and so this is an important factor when comparing the feasibility of different battery
types. Our model does not account for self discharge, as this is not an issue with NaS batteries.
We use a time series of historical hourly wind speed data, and assume that the hourly wind
forecast for the following day, ft , is the current days realized wind speed for each hour. The
realized wind speed at time t is wt . We commit to deliver ftp kWh of electricity at time t, or in
other words we commit the exact amount that we forecast for that hour. At a given hour, we
can at most charge the battery by its rating , without exceeding the capacity . This means
that wind energy spillage is inevitable some of the time due to a full battery or too much wind.
The situation is similar for discharging: at each hour, we can only discharge the battery by
, without falling below our minimum state of charge . This means that, even with battery
storage, under-delivery may occur due to not having enough in storage or not being able to take
enough to cover our shortfall. Because of this, batteries which have a higher power rating may
be more desirable in the future. This also highlights the need for a more accurate forecast for
wind; a better forecast means less commitment error at a given hour. If the difference between
what we commit and what we are actually able to provide is small, there is no need for a large
battery or a battery with a large power rating.
7.5.1
Strategy 1 - Naive Strategy
The table and figures below provide more insight into a proper battery operating strategy. Table
7.2 is taken from [78] and gives us the number of cycles to failure (g(i )) as the depth of
discharge (DOD) increases, for a NaS battery. Figure 7.5 displays this table, along with the
100
equivalent full cycles (Ce ) which is defined as

Ce = g(i ) i .
From Figure 7.5 we can see that it is better, in terms of optimizing the battery life, to use very
small cycles of the battery than moderate or large cycles of the battery. Therefore, a possible
strategy is to simply charge our battery when we under-forecast wind (when wt ft > 0)
and discharge the battery when we forecast surplus wind (when wt ft < 0). This of course
assumes that wt is large enough to generate power. This strategy is closest to the assumptions
outlined in Chapter 5. The parameters constraining this strategy are , , and . Economically speaking however, purchasing and maintaining a large battery which is frequently cycled
at a small fraction of its capacity is not financially realistic. Figure 7.6 zooms in on the region
of Figure 7.5 when DOD 50%. While Figure 7.5 clearly showed that very small battery
cycles are better for battery life, this figure shows that large battery cycles are actually better
than moderate ones. Therefore, we should also consider an alternative strategy which provides
a relatively long battery life (but shorter than that of Strategy 1) while consistently cycling
through the battery at a large fraction of its capacity.
7.5.2 Strategy 2 - Utilizing Capacity

To achieve a relatively long battery life while also utilizing the available battery capacity, a
possible strategy is to consistently charge the battery to a minimum level min and after this
level is reached the decision to keep charging or switch to discharging is made on an hourly
basis. Once the decision is made to switch to discharging (this decision is made if wt ft < 0 or
when the battery is full), we continue to discharge to a minimum level min . This strategy only
makes sense if min > 50% and min < 50% (realistically, min >> 50% and min << 50%)
so that the depth of discharge is consistently large. While this setup may appear to solve the
problem of a high frequency of small depths of discharge, it often does not allow us to achieve
our goal of providing a predictable power source as in Strategy 1. This strategy results in
a large amount of spilled wind energy, which from an economic standpoint is not desirable.
Since an increase in the predictability of wind would decrease the need for constant buffering
as in Strategy 1, significant forecast improvements would allow Strategy 2 to be more valuable,
and so we consider it for this reason.
7.6 Strategy Comparison as Parameters Change

To compare Strategies 1 and 2, while gaining a better understanding of which battery parameters are most important, the effects of increasing the power rating (Tables 7.3 and 7.4) and
increasing the battery size (Tables 7.5 and 7.6) are explored using our historical data. For each
parameter set we choose, we can observe the amount of wind energy spilled while charging
or the level of under-delivery while discharging for each Strategy. Using the calculated average electricity prices, as shown in Figure 7.4, we can convert this energy into a dollar amount
which represents the loss incurred by spilling energy and/or under-delivering. The base case
( = 1000 kW, = 10000 kWh) is shown in bold in Tables 7.3 and 7.4. In these tables,
7.6. S TRATEGY C OMPARISON
AS
PARAMETERS C HANGE
101
x 10
3.5
12
10
2.5
1.5
0.5
0
0
10
20
30
40
DOD (%)
Figure 7.5: Battery Cycle Life vs. Depth of Discharge
0
50
Equivalent Full Cycles (Dashed Line)
Cycles to Failure (Solid Line)
14
x 10
10000
5800
9000
5600
8000
5400
7000
5200
6000
5000
5000
50
60
70
80
90
4800
100
DOD (%)
Figure 7.6: Battery Cycle Life vs. Depth of Discharge when DOD 50%
Equivalent Full Cycles (Dashed Line)
Cycles to Failure (Solid Line)
102
AS
PARAMETERS C HANGE
i (%)
0
2.5
5
10
15
20
30
40
50
60
70
80
90
100
103
g(i )
1400000
1000000
621100
194700
83400
65700
39600
23000
9900
8700
7700
6900
6200
5700
Table 7.2: Cycles to Failure Data (Yang et. al)
= max /, where max corresponds to the new, and larger, power rating used in the simulation. Therefore, = 1.2 represents a 20% increase in the power rating, and so on. The
dollar losses are broken into 4 types for charging and discharging: spilled wind energy due
to the power rating constraint (c ), spilled wind energy due to a full battery (c ), spilled wind
energy due to both power rating and battery size constraints, i.e both constraints were unsatisfied as opposed to only one of the constraints being unsatisfied (c, ) and spilled wind energy
due to the heuristic (ch ). Similar losses occur during discharging: under-delivery caused by
the power rating constraint (d ), under-delivery due to a full battery (d ), under-delivery due
to both power rating and battery size constraints (d, ) and under-delivery due to the heuristic
(dh ). Note that ch = 0 and dh = 0 for Strategy 1 since the strategy is designed to switch
between charging and discharging as necessary and hence there are no added losses due to the
design of the strategy. For a given loss type l, Nl is the frequency of this loss type during the
simulation.
1
1.2
1.4
1.6
1.8
c
7.1 103
3.0 103
1.1 103
2.1 102
0
Nc
569
300
151
47
0
c,
3.9 104
3.3 104
2.5 104
1.5 104
1.3 103
Nc,
796
796
455
259
21
c
1.5 104
2.4 104
3.4 104
4.5 104
5.8 104
Nc
953
953
1434
1651
1887
Table 7.3: Strategy 1: Loss Attribution - Charging ( = 10000), Historical Analysis
By increasing the power rating by 20%, we decrease the total losses by less than 1%. However, d decreases by 55%. By increasing the power rating by 60% we decrease total losses by
1.64%, but decrease d by 97%. When = 1800 kW we eliminate the loss due to power rating
104
1
1.2
1.4
1.6
1.8
d
8.1 103
3.8 103
1.4 103
3.1 102
0
Nd
590
326
176
70
0
d,
3.7 104
3.3 104
2.6 104
1.7 104
0
Nd,
766
616
472
293
0
d
1.6 104
2.4 104
3.3 104
4.3 104
5.2 104
Nc
924
1163
1367
1570
1662
Table 7.4: Strategy 1: Loss Attribution - Discharging ( = 10000), Historical Analysis
constraints entirely; this is due to the turbine used in our simulations, a Vestas 1.8 MW turbine.
Overall, losses decrease by 8.2% by increasing the power rating by 80%.
5000
15000
30000
50000
c
2.4 103
1.2 104
1.8 104
1.9 104
Nc
247
853
1299
1361
c,
5.4 104
2.5 104
2.9 103
1.1 102
Nc,
1118
512
66
4
c
2.3 104
1.1 104
2.5 103
7.1 101
Nc
1374
745
226
5
5000
15000
30000
50000
d
2.5 103
1.2 104
1.9 104
2.0 104
Nd
219
859
1304
1354
d,
5.5 104
2.4 104
2.0 103
5.5 101
Nd,
1137
497
52
2
d
2.2 104
1.1 104
2.7 103
1.5 102
Nc
1278
650
199
11

By increasing by 50%, we decrease the total losses by 22%. By further increasing to
50000, we reduce the total losses by 68%. Although these losses appear more significant then
those seen by increasing , increasing corresponds to a large increase in the price you must
pay for the battery. For instance, assuming an acquisition cost of 300 USD/kWh, increasing
by 50% increases the battery cost by 1.5 million USD. We have also seen so far that a large battery is not optimal since, on average, we are only charging and discharging to a small fraction
of our capacity. Therefore, it may be beneficial for future efforts to not only focus on making
bigger batteries, but also on increasing the power rating of batteries, without compromising
safety.
Strategy 1 does not incur heuristic losses (ch = dh = 0); this is because we never choose
under-delivery in order to reach a certain charge level. We simply charge and discharge as
necessary, restricted only by and . Tables 7.7 and 7.8 show that, when Strategy 2 is used, ch
and dh are quite substantial.
As min and min decrease, losses also decrease. When min and min decrease from 1
to 0.9, total losses decrease by 2.39% (losses due to the heuristic decrease by 3.68%). When
10
1.24
15
AS
20
PARAMETERS C HANGE
25
(X 103)
30
35
105
40
45
50
13
12
11
10
1.2
9
8
1.18
7
1.16
Losses (X107) Dashed Line
Losses (X 10 ) Solid Line
1.22
1.14
1.1
1.2
1.3
1.4
1.5
1.6
1.7
3
1.8
Figure 7.7: Strategy 1: Losses as and increase: = 0.1, = 1000, = 1, = min = 0.1,
min = 0, = $200/kWh, = $5.50/kWh
min , min
1,0.666
1,1
0.9,0.9
0.8,0.8
0.7,0.7
0.6,0.6
c
6.4 103
5.6 103
6.5 103
7.1 103
8.0 103
8.3 103
Nc
504
451
512
548
609
638
c,
2.3 103
2.0 103
2.3 103
2.5 103
2.5 103
2.6 103
Nc,
69
60
67
73
74
73
c
4.3 102
4.3 102
3.5 102
2.9 102
2.8 102
2.9 102
Nc
38
39
32
26
25
23
ch
6.2 104
6.3 104
6.3 104
6.3 104
6.4 104
6.6 104
Nch
2093
2150
2160
2179
2218
2306
106

d
6.9 103
6.6 103
6.8 103
6.8 103
6.8 103
6.6 103
Nd
492
472
479
482
480
464
d,
1.7 103
1.7 103
1.8 103
1.7 103
1.7 103
1.6 103
Nd,
54
56
56
55
52
49
d
3.9 102
5.5 102
4.9 102
4.4 102
3.6 102
3.4 102
Nd
28
43
35
32
27
25
dh
6.0 104
6.4 104
5.9 104
5.5 104
5.1 104
4.8 104
Nd h
2079
2235
2110
1956
1778
1630
these parameters decrease to 0.6, total losses decrease by 7.52% (losses due to the heuristic
decrease by 10.73%). Therefore, the total losses using reasonable min and min values are
always larger than losses seen using Strategy 1.
7.7 Next Steps

In this chapter two battery charging and discharging strategies were introduced and compared
for a variety of battery parameters. This allowed us to isolate which battery parameters have
the largest impact on the level of predictability they add to the wind-battery system. The key
conclusions from this analysis is that the battery size as well as power rating have a great
impact on how much value the battery is able to add to the system. It is shown that it is
optimal to make as large as possible, and the key constraint on this parameter are safety
issues with ramping up and down too fast. This could potentially be an important area for
future research. On the other hand, it is not optimal to make as large as possible; if the
battery capacity is too small, it is not able to compensate for a shortage of wind over a long
period since the battery will empty quickly. However, the size is constrained by the battery
cost, which increases linearly with size. Therefore, for a given wind profile at a given location,
an appropriate battery size must be determined.
In the next chapter, we expand our model to include market design parameters which will
highly influence a wind producers operation strategy, and hence the success of a wind-battery
system. In particular, we examine penalty functions which are currently, or could be in the
future, imposed on wind producers to deal with their unpredictable energy source. The results
are then compared against the analytical results obtained using the simple model outlined in
Chapter 5.
Chapter 8
Market Design
In Chapter 7 we examined two possible operating strategies for a wind-storage system. The
feasibility of these strategies is highly dependent on the reliability of the wind forecast for
a given hour. Since a perfect forecast does not exist, without adequate storage in place the
electricity grid system is forced to keep a large number of fossil fuel burning plants on standby
in order to maintain reliability in the system. Clearly the solutions to this problem, as discussed
in Chapters 5 and 7, are improved forecasts or adding storage systems. How can the market be
designed to provide the correct incentives for this to happen?
In this chapter we answer this important, yet to date unsolved, issue by discussing two
important design considerations; penalties and capacity payments. With these tools in mind
the models developed in Chapters 5 and 7 are revisited. Using the framework developed in
Chapter 7 our modified bootstrapping approach suited for wind speed data is carried out. For
simple penalty rules these results are compared with the analytics outlined in Chapter 5.
8.1
Capacity Markets
Capacity markets have been designed to help encourage the introduction of new generation
into the market. They are particularly important to fill a niche for high cost, quick ramp rate
generation options which are required to minimize the probability of lost system load. This
is true despite the occurrence of price spikes; they occur such a small number of times during
each year that they do not allow the value of such facilities to be recovered. Therefore, capacity
payments received by these generators are an important revenue source which often cover
acquisition and operating costs above the cost of fuel used in producing the electricity, ensuring
that these generators are available when needed.
High delivery rate storage facilities would be another way to address lost system load.
Should developers of storage facilities receive capacity payments as well? At first thought this
sounds like an excellent idea. However, implementation of the idea faces challenges. A storage
facility provides the potential to meet load, but unless it is full this potential cannot be realized. Therefore the capacity of the facility depends on its operating characteristics (maximum
deliverability) as well as the strategy used to operate the facility.
If we presume that the facility is used to deliver power during price spikes and fill during
low prices then this may already be consistent with the goals of the capacity payment program.
107
108
C HAPTER 8. M ARKET D ESIGN
However in this case the need for quick ramping generation sources is still needed at times
when forecasted wind is greater then realized wind at a given hour. Moreover, high delivery rate
storage facilities such as pumped hydro are not always feasible, while batteries are limited in
how fast they can be emptied by their power rating. If the storage facility is instead intended to
manage the discrepancies between forecasted and true wind at a given hour, it is not achieving
the goals of the capacity payment program as high ramping generation is still required at peak
times. In this case, what incentives should be put in place to encourage the use of storage? In
the next section we discuss the possibility of imposing penalties on wind producers in the event
of over- or under-delivery.
8.2 Penalties for Forecasting Error

Penalties have already been discussed in Chapter 5, and were incorporated into our simple
model by a single parameter, x. The optimal control strategy for a single storage unit was
given in terms of this level of penalty. Penalties charged to wind producers to offset the cost of
the unpredictability of their energy source are undoubtedly a bigger consideration than capacity
payments. However, there is still the question of if wind producers should be treated differently (in addition to being part of the feed-in tariff in Ontario), or in markets where penalties
currently do not exist, should they impose penalties in the future? If so, what should the level
of penalty be and how does it affect the producer?
Looking back at our simplified model in Chapter 5, we were able to conclude that, for any
(constant) level of penalty it was always optimal to bid when the storage facility was full. When
penalties are small it is also optimal to bid when empty, but for large penalties it is optimal to
not bid. For moderate penalty levels it is better to bid closer to the end of the storage life,
but better not to bid when far from this end time or maturity. We have now built a more
complex model where storage size is not limited to a single unit, and factors such as battery
power rating, battery efficiency, wind forecasting and hourly power prices are considered.
In the next section we move even further away from our initial simple dynamic program
of a wind storage system and consider penalty functions which are dependent on the level of
over- or under-delivery of power.
8.2.1 Penalty functions

Instead of a single parameter penalty representation, what about a penalty function? The introduction of a penalty function is not a new idea in electricity markets; markets in many countries
already have rules in place.
Without a penalty for the over-delivery of power, wind producers could potentially always
commit less than the expected power available at a given hour. If the realized power is indeed
more than the committed amount, they simply store the excess or, if the storage facility cannot
accommodate it, deliver the excess. Therefore, it may be in the best interest of the regulator to
impose penalties for both the under- and over-delivery of power at a given time. Moreover, it
makes sense that the penalty charged be a function of the extent of under- or over-delivery of
power at that time, as well as a function of the power price at that time.
8.2. P ENALTIES FOR F ORECASTING E RROR
8.2.2
109
Current design in European Markets
In some European markets there is currently asymmetry in the penalties assigned to wind producers for up and down forecast errors (and hence over- and under-delivery of power). Therefore it is surprisingly often better for wind producers to commit more power than the forecast
suggests rather than risk over-delivery. This is not acceptable for the Independent Systems
Operator (ISO) who must decide whether or not to trust the wind producers predictions in order to ensure the safety and reliability of the system. It is stated in [51] that consumers will
suffer the costs of rewarding wind producers for cheating the system. These wind producers
are currently not equipped with storage facilities, however the addition of these facilities may
decrease the need to cheat.
The European Wind Energy Association states that when the market faces a deficit of power
at the same time that wind producers cannot provide the power they commit, wind producers
will receive a price lower than the spot price for the power it does provide [30]. The penalty
decreases as the difference between what they commit and actually provide increases. If they
are able to provide more power than previously committed, they receive the spot price for
the total power they provide (including the excess) since the excess power can be used to
decrease or eliminate the deficit. The presence of storage facilities could potentially decrease
or eliminate the penalties imposed on wind producers for under-delivery in times of power
deficit. When the market faces excess power at the same time that wind producers cannot
provide the power it commits, they receive the spot price for the power delivered. If they are
able to provide more power then previously committed they receive a price lower than the spot
price, and the received payment decreases as the amount of excess power increases. A storage
facility would also decrease or eliminate penalties imposed on wind producers for over-delivery
in times of excess power since wind producers would store rather than over-deliver whenever
possible.
The strategies discussed in Chapter 7 do not consider market design parameters such as
penalties for over- and under-delivery. It is obvious that, if under-delivery penalties exist (such
as a dollar amount per kWh under-delivered or a dollar amount for each occurrence), Strategy
2 would sometimes face severe penalties. While Strategy 1 would still face under-delivery
penalties, they would be less severe in most cases.
8.2.3
A simple penalty function
We begin by considering a simple penalty function for under-delivery: a linear function which
increases as the difference between the forecasted power and the realized power increases. The
function, pt , is given by the following:
pt = x eh max(ftp
wtp st1,t , 0).
Here st1,t is the amount taken from storage in order to decrease the penalty incurred at a given
hour. Note that if the sum of wtp and st1,t is less than ftp there is an under-delivery penalty.
Otherwise there is no under-delivery and hence no penalty. We do not consider a penalty for
over-delivery here. The operational cost of spilling wind is zero and modern wind turbines
have the technology to do this. We do not consider the political cost of spilling a substantial
fraction of the available wind here.
110
An obvious question follows from the formulation of the penalty function described above:
what is a reasonable value for the parameter x? To answer this question we rerun our simulation
from Chapter 7 with the penalty function included. We begin by choosing x = 0.5 in order
to determine the total penalty over a one year period for various battery sizes (Figure 8.1).
Clearly, as increases the total penalty decreases. The relationship is not linear since there is a
damping effect as becomes large: once reaches a certain level you are still restricted by
other battery parameters such as the power rating , so the penalty paid is no longer related to
the battery size. This reinforces our earlier observation that a smaller battery may be sufficient
in reducing the penalties imposed on wind producers, in turn reducing the cost of purchasing a
battery. This is shown in Figures 8.2 and 8.3 which illustrates the linear relationship between
battery size and battery cost.
To further illustrate the importance of choosing a reasonable battery size, we plot the ratio
of the revenue generated by the use of our battery to the total revenue from the sale of wind
power as a function of , shown in Figure 8.4. Clearly, as increases the ratio should increase,
as seen in the figure. As in the penalty vs. plot, the relationship is not linear; the addition of a
battery certainly generates revenue, but as the battery size becomes extremely large the added
revenue stabilizes while the battery cost continues to increase.
Since market rules have not yet been set, the parameter x is very important in understanding
the possible impacts of penalties. For instance, assuming small or no penalties, wind-battery
systems may appear more cost effective (or less cost ineffective) then they would actually be if
used in the future. Therefore for a fixed x we can calculate our annual net gain (or loss) as a
function of . This is done in Figure 8.5 for various x values. We define our net gain as
net gain = annual revenue - penalties incurred throughout the year - annual battery cost opportunity cost incurred throughout the year.
By opportunity cost we mean the dollar amount we could have earned if our battery acted as a
constant buffer for wind, hence achieving zero spilled wind as well as zero time periods where
we are unable to make our day ahead commitment.
From Figure 8.5 we observe that when x 1 the net gain is positive for some range of
values. However, when is very small, or very large, we observe a negative gain. When
is small, opportunity costs are high. This is since our battery is not large enough to store
a large amount of spilled wind, and similarly not large enough to provide enough storage
to compensate for forecasting too much wind. Even when a small penalty is imposed wind
producers will be operating at a loss. Given moderate battery sizes, , as well as moderate
penalties, governed by x, it makes sense that there should be some range of these parameters
for which wind producers are able to make a profit. This is shown to be the case from this
figure. As becomes large the high cost of large capacity overpowers the revenues accrued
from the capacity, even when penalty levels are relatively low. It is clear from this figure that
there certainly must be a balance between capacity and penalty size.
In Figure 8.6 we zoom in on x values near x = 1 to determine more precisely the range
of penalty levels for which we observe a positive net gain. From this figure we see that when
x > 1.04 penalties are severe enough to cause a negative gain no matter what battery size
we choose. Below this level we observe a positive net gain for some range of capacities. For
instance, when x = 0.99 we see a positive net gain when is approximately between 800kWh
and 7300kWh. This can be seen more clearly in Figures 8.7 and 8.8. These figures provide
111
upper and lower bounds for so that, for a given x, we can determine the range of battery sizes
which results in a positive net gain. Since the goal is to maximize profits, it is important to
know exactly where this maximum occurs. Hence we determine the maximum (max ) value
which allows a positive net gain for a given penalty parameter x. This is shown in Figure 8.9.
From this figure it is obvious that this occurs at = 3500kWh, when x = 1.
4
x 10
4.8
4.6
Total Penalty ($)
4.4
4.2
4
3.8
3.6
3.4
3.2
3
1000
2000
3000
4000
5000
6000
7000
8000
9000 10000
Figure 8.1: Total Penalty vs
For comparison as well as completeness, we revisit the net gain equation previously defined. This equation contains a term accounting for opportunity costs, or the cost of having
insufficient forecasts, insufficient battery capacity, etc. It is important to note that this definition
of net gain is not the same as the net cash flow. To examine the actual profit or loss incurred by
the wind producer, we must look at the related equation,
net cash = annual revenue - penalties incurred throughout the year - annual battery cost.
112
15
x 10
($)
10
0
1000
2000
3000
4000
5000 6000
(kWh)
7000
Figure 8.2: Battery Cost vs
8000
9000 10000
113
12
x 10
Annual Battery Cost ($)
10
0
1000
2000
3000
4000
5000 6000
(kWh)
7000
8000
Figure 8.3: Annual Battery Cost vs
9000 10000
114
Ratio of Revenue from Storage to Total Wind
0.2
0.18
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
2000
4000
6000
Figure 8.4: Revenue Ratio
8000
10000
115
x 10
0.5
0.6
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
Net Gain/Loss
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Figure 8.5: Net Gain/Loss as a function of for various penalty levels x: = 0.1, = 1000,
= 1, = min = 0.1, min = 0, = $200/kWh, = $5.50/kWh
116
Revenue(battery cost+losses+penalties)
5000
0
0.99
1
1.01
1.02
1.03
1.04
1.05
1.06
1.07
1.08
1.09
1.1
5000
10000
15000
20000
2000
4000
6000
8000
10000
Figure 8.6: Net Gain/Loss as a function of for various penalty levels x: = 0.1, = 1000,
= 1, = min = 0.1, min = 0, = $200/kWh, = $5.50/kWh
117
3000
2500
1500
lower
2000
1000
500
0
0.9
0.92
0.94
0.96
0.98
1.02
1.04
Figure 8.7: lower vs x: = 0.1, = 1000, = 1, = min = 0.1, min = 0, = $200/kWh,

= $5.50/kWh
118
9000
8500
8000
upper
7500
7000
6500
6000
5500
5000
4500
0.9
0.92
0.94
0.96
0.98
1.02
1.04
Figure 8.8: upper vs x: = 0.1, = 1000, = 1, = min = 0.1, min = 0, = $200/kWh,

= $5.50/kWh
119
3500
3000
2500
max
2000
1500
1000
500
0
0.5
1.5
Figure 8.9: max vs x: = 0.1, = 1000, = 1, = min = 0.1, min = 0, = $200/kWh,

= $5.50/kWh
120
This equation allows us to obtain similar insights into the balance between capacity and
penalty, and also gives us an accurate picture of whether or not the producer is making a
profit, as well as how much. Figure 8.10 shows max vs x for a large range of x values. As
expected, max increases as x increases until a certain x is reached; after this value max is zero.
Therefore, as the penalty level rises so does the battery capacity in which we see the largest
positive gain/largest net cash. At a certain point battery costs as well as the total penalty paid
out are so large that both values are always negative, resulting in max = 0. In other words,
there is no battery size for this level of penalty that will result in a positive gain/positive cash
value. The difference between Figures 8.9 and 8.10 is that the value of x for which this occurs
is larger. This results from the fact that we are omitting the somewhat fictitious opportunity
cost, a cost which we can measure but which is not actually paid out by the wind producer.
Instead, as previously mentioned, it measures the potential inflow of cash which would have
occured given a perfect battery. Therefore both profit measures are worth tracking. Figure 8.11
shows the net cash when x is near two. We choose values near two since from Figure 8.10 it
is clear that when x > 2 (approximately) max = 0 implying a negative net cash flow. Figure
8.11 confirms this and we can determine more accurately the value of x where this occurs. It
is clear from this Figure that when x 2.02, and moreover when x > 2.02, we see a negative
net cash flow no matter how big or small a battery we choose.
8.3 Summary and Discussion

In this chapter we looked at the impact that market design has on decisions made by wind
producers. Considering market design, specifically penalties which wind producers may face
for over- or under-delivery of power, gives us a better understanding of how big a battery
should be in order to be economically feasible. Incorporating a linear penalty function into the
Naive strategy shows that, for x < 1.04 there is a range of values for which our strategy
is economically feasible. Including the penalty function also provides insight for improved
operating strategies; with steep penalties in place, a wind producer may tend to make decisions
which reduce the total penalty paid out over a period of time.
Chapters 6 and 7 provide an in depth study of operational heuristics which include optimal
engineering design as well as market rules designed to incentize producers to make better use
of forecasts so as to mitigate system impacts. Retrospective and block bootstrap analysis were
used to explore this complete model. An optimal strategy for our storage system will depend
on progress made to the quality of wind forecasts. The goal of our work is not to develop better
forecasting techniques, however this would certainly be something to consider in the future.
This is briefly discussed in the final chapter.
8.3. S UMMARY AND D ISCUSSION
121
3500
3000
2000
max
2500
1500
1000
500
0
0.4
0.6
0.8
1.2
1.4
1.6
1.8
2.2
Figure 8.10: max vs x: = 0.1, = 1000, = 1, = min = 0.1, min = 0, = $200/kWh,

= $5.50/kWh
122
1.5
x 10
Net Gain/Loss
0.5
=1.9
=1.92
=1.94
=1.96
=1.98
=2.0
=2.02
=2.04
=2.06
=2.08
=2.1
0.5
1.5
2000
4000
6000
8000
10000
Figure 8.11: Net Cash as a function of : = 0.1, = 1000, = 1, = min = 0.1, min = 0,
= $200/kWh, = $5.50/kWh
Chapter 9
Discussion and Conclusions
This thesis discusses two important renewable energy sources, corn ethanol and wind power
coupled with battery storage. Both of these sources have been praised for their ability to help
replace fossil fuels, and criticized for their inefficiency and cost. An important factor which
rarely enters the debates surrounding these renewable resources is how to optimally run a facility which harnesses these natural resources into energy which can be used in place of fossil
fuels. With the degree of uncertainty we face in the underlying prices and environmental data
associated with each resource, ignoring this uncertainty may lead to highly inaccurate predictions of the value of a facility, and in turn the resource itself. This thesis addresses this
important issue in hopes of determining how to manage such a facility as well as accurately
extract the value of the operation.
9.1 Summary
The framework used in Chapters 3-4 is standard option pricing theory. However, as shown in
this thesis, this theory extends well beyond financial options pricing and can be used in making
a variety of decisions where the underlying variables are uncertain. Under this framework, as
well as the assumption that corn-gasoline price correlation will increase with increased usage
of corn in ethanol production, we are able to show that the value of an ethanol plant will
decrease with increasing correlation. A wide range of mathematics can be used to show this,
including a historical analysis, a bootstrap based analysis, partial differential equations with
appropriate boundary conditions and Monte Carlo simulations. Each of these methods have
advantages as well as drawbacks, but the combination of methods with the same end result
gives us confidence in our assumptions and results.
These results could potentially benefit an ethanol producers decision to build, or continue
to run, an ethanol plant in the future. Decisions based on calculating the net present value of
the plant do not include relationships between corn and gasoline commodity prices, hence the
value extracted from the plant may be exaggerated. At best, it provides an upper bound on the
potential value over the life of the plant. The real options based approach allows us to consider
the real possibility that corn and gasoline prices, which historically do not move up and down
together, may do so in the future so that the planning and scheduling of the facility can be
carried out appropriately.
123
124
C HAPTER 9. D ISCUSSION AND C ONCLUSIONS
Chapters 5-8 discuss a different, but related, renewable energy source: wind power. The
sources of uncertainty are different here, in particular the uncertainty stems from limitations in
predicting future wind speed as well as uncertainty in power prices. Given this uncertainty, as
well as the option to couple wind farms with battery storage technology, we develop a simple
dynamic program to determine the optimal control of the wind-storage facility in a simplified
setting. The ability to obtain analytic solutions allows us to gain intuition which is not obvious
in a more complex, yet also more realistic, model. With added complexity we examine possible
operating strategies in order to increase the predictability of wind as a renewable energy source.
These strategies are examined based on historical wind speed and power price data, as well as
a sodium sulfur battery model. It is determined that the strategy which best increases the
predictability of wind is one where the battery is used for constant buffering rather then one
where preserving the life of the battery is the major focus. After further study of this particular
strategy we determine the range of battery sizes for which wind producers see a profit, given
that they may be penalized for the unpredictability of their energy source. Looking at a wide
range of penalty levels as well as a large range of battery sizes allows us to see the tradeoff
between the two; in particular a low battery capacity results in large penalties, hence if the
penalty is large it will be difficult for wind producers to make money. On the other hand a large
battery capacity results in much lower penalties, however since battery cost increases linearly
with capacity there is a limit on how large the battery can become before its cost exceeds the
benefits realized from lower penalties. For a given penalty level, we calculate the maximum
battery size for which wind producers see a profit, as well as upper and lower bounds for for
which they see a profit.
Similarly to the corn ethanol portion of the thesis, we use a bootstrapping technique to
obtain key results so that we are not only drawing conclusions from one realization (i.e. the
one already seen in the past) of possible future outcomes. The bootstrapping must be performed
in a way which preserves the strong autocorrelation in the data, hence a local block bootstrap
is chosen.
These results could impact a wind producers decision about how to proceed with the addition of storage to a current facility, or the creation of a new wind/storage facility. Without
considering multiple operating strategies, as well as optimal battery parameters and even possible penalties which may be imposed, wind producers could face large opportunity costs while
realizing little or no profit.
9.2 Future Work

This research could be extended in a number of ways. The model described in Chapter 3, which
gives the annual value of an ethanol plant, includes cost and subsidy parameters. However, a
sensitivity analysis has not been completed around these parameters. Performing a sensitivity
analysis would give a better idea of how ethanol plant value would change if subsidies were
decreased, or the cost of running the plant increased. Clearly the value would decrease in these
cases, but how large a decrease is seen if subsidies decrease by a small amount or running
costs increase by a small amount? This information is clearly valuable to present and potential
ethanol producers.
Our analysis of a wind/storage system could be improved by solving a more complicated
9.2. F UTURE W ORK
125
optimal control problem. The dynamic program could be extended to handle forecasting techniques, a series of wind turbines and batteries, and natural extensions of the current analytics
could possibly be given. The current analysis of the more complex model is a heuristic approach, which is a valuable first step and is clearly beneficial to current and future wind producers. However, solving this optimal control problem would add more certainty to the results
obtained in this data based approach.
As mentioned in Chapter 8, the goal of our work is not to develop better wind forecasting
techniques. However, determining how increasing the quality of forecasts might impact battery
usage would be a valuable next step. It is also possible to assign a value to the ability to
perfectly forecast wind. Future work in this area might consist of choosing forecasts with
a known quality and rerunning our simulations to determine better strategies assuming this
quality. It makes sense that, if forecasts improve considerably, the Utilizing Capacity strategy
outlined in Chapter 7 could outperform the Naive strategy. This is since the Naive strategy
is designed for potentially large forecast errors at a given time step and is therefore flexible
to switch from charging to discharging as needed. A better forecast may lead to a strategy
requiring less switching flexibility, such as the Utilizing Capacity strategy, where the battery is
being used in a more uniform way. Better forecasts might also have an impact on the optimal
battery size required to adequately increase the predictability of wind.
Bibliography
[1] Ackermann, T., 2005. Wind Power in Power Systems, John Wiley & Sons Inc., England.
[2] Alexiadis, M.C., Dokopoulos, P.S., Sahsamanoglou, H.S., Manousaridis, I.M., 1998.
Short-Term Forecasting of Wind Speed and Related Electrical Power, Solar Energy 63[1],
61-68.
[3] And Now, the Electricity Forecast. The Economist Technology Quarterly, June 12th,
2010.
[4] Anderson, L., Davison, M., 2008. A Hybrid Simulation Model for Electricity Spot Price
in a Deregulated Market, IEEE Transactions on Power Systems 3, 927-937.
[5] Angarite, J.M., Usaola, J.G., 2007. Combining hydro-generation and wind energy: Biddings and operation on electricity spot markets, Electric Power Systems Research, 7,
393400.
[6] Babcock, Bruce A., 2007. The Long-Run Impact of CornBased Ethanol on the Grain,
Oilseed, and Livestock Sectors: A Preliminary Assessment, Presented at the Iowa Pork
Congress, Des Moines, Iowa.
[7] Baker, John, 2008. New Technology and Possible Advances in Energy Storage, Energy
Policy 36, 4368-4373.
[8] Blanger, C., Gagnon, L., 2002. Adding wind energy to hydropower. Energy Policy, 5,
12791284.
[9] Benitez, L.E., Benitez, P.C., van Kooten, G.C., 2008. The economics of wind power
energy storage, Energy Economics, 16, 19731989.
[10] Bose, B.K., 2000. Energy, Environment, and Advances in Power Electronics, IEEE Transactions on Power Electronics 15, 688-701.
[11] Boyle, Godfrey, 2007. Renewable Electricity and the Grid : The Challenge of Variability,
Earth Scan, London.
[12] Bjork, T. S., 2004. Arbitrage Theory in Continuous Time (2nd edn.), Oxford University
Press, New York.
[13] Brennan, M.J., Schwartz, E.S., 1985. Evaluating Natural Resource Investments, Journal
of Business 58[2], 135-157.
126
BIBLIOGRAPHY
127
[14] Bueno, C., Carta, J.A., 2006. Wind powered pumped hydro storage systems, a means
of increasing the penetration of renewable energy in the Canary Islands, Renewable and
Sustainable Energy Reviews, 28, 312340.
[15] Budny, Daniel, 2007. The Global Dynamics of Biofuels: Potential Supply and Demand
for Ethanol, Brazil Institute Special Report, April 2007, Issue 3. Woodrow Wilson International Center for Scholars.
[16] Cavallo, A.J., 1995. High-Capacity Factor Wind Energy Systems, Journal of Solar Energy
Engineering 117, 137-143
[17] Castronuovo, E., Lopes, J., 2004. Optimal operation and hydro storage sizing of a windhydro power plant. Electrical Power and Energy Systems 26, 771-778.
[18] Chernick, Michael, 2008. Bootstrap Methods: A Guide for Practitioners and Reseachers,
Second Edition, Wiley Series in Probability and Statistics.
[19] Commodity Research Bureau, 2007. The CRB Commodity Yearbook, John Wiley &
Sons, Inc.
[20] Crago, Christine L., Khanna, M., Barton, J., Giuliani, E., Amaral, W., 2010. Competitiveness of Brazilian sugarcane ethanol compared to US corn ethanol, Energy Policy 38,
74047415.
[21] Crank J. and Nicolson P., 1947. A Practical Method for Numerical Evaluation of Solutions of Partial Differential Equations of the Heat Conduction Type, Proceedings of the
Cambridge Philosophical Society 43, 5064.
[22] Davison, M., Zhao, G., 2012. Optimal Control of a 2-dam Hydro Facility, Systems Engineering Procedia 3, 1-12.
[23] DeCesaro, J.,Porter, K.,Milligan, M., 2009. Wind Energy and Power System Operations:
A Review of Wind Integration Studies to Date, Electricity Journal 22[10], 3443.
[24] Denault, M., Dupuis, D., Couture-Cardinal, S., 2009. Complementarity of Hydro and
Wind Power: Improving the risk profile of energy inflows, Energy Policy 37[12], 53765384.
[25] Deng, Shijie, Johnson, Blake, Sogomonian, Aram, 1999. Spark Spread Options and the
Valuation of Electricity Generation Assets, Proceedings of the 32nd Hawaii International
Conference on System Sciences.
[26] Deng, S., Jiang, W., 2005. Levy Process-driven Mean-reverting Electricity Price Model:
The Marginal Distribution Analysis, Decision Support Systems 40[3-4], 483-494.
[27] Dixit, Avinash K., Pindyck, Robert S., 1994. Investment under Uncertainty, Princeton
University Press.
128
BIBLIOGRAPHY
[28] Dufo-Lopez, R., Bernal-Agustin, J.L., Dominguez-Navarro, J.A., 2009. Generation Managament Using Batteries in Wind Farms: Economical and Technical Analysis for Spain,
Energy Policy 37, 126-139.
[29] Economic and Market Barriers to the Adoption of Energy Storage Projects for Renewable Energy Integration (2010), Power Advisory LLC, Prepared for Natural Resources
Canada.
[30] The European Wind Energy Association. Website: www.ewae.org, accessed in 2010.
[31] ERS (United States Department of Agriculture Economic Research Service), 2010. Feed
Grains Database. Website: www.ers.usda.gov, accessed in 2012.
[32] Garcia, A., Torres, J.L., Prieto, E., de Francisco, A. (1998). Fitting Wind Speed Distributions: A Case Study, Solar Energy, 62[2], 139-144.
[33] Garcia-Gonzalez, J., 2008. Stochastic Joint Optimization of Wind Generation and
Pumped-Storage Units in an Electricity Market, IEEE Trans Power Systems 23[2], 460468.
[34] Geman, Helyette, (2005). Commodities and Commodity Derivatives, John Wiley & Sons
Inc., England.
[35] Haberman, Richard, 2004. Applied Partial Differential Equations with Fourier Series and
Boundary Value Problems, Fourth Edition, Pearson, New Jersey.
[36] Hamlet, Alan F., Huppert, Daniel, Lettenmaier, Dennis P., 2002. Economic Value of
Long-Lead Streamflow Forecasts for Columbia River Hydropower, Journal of Water Resources Planning and Management 28, 91-101.
[37] Huisman, R., Mahieu, R., 2003. Regime Jumps in Electricity Prices, Energy Economics
25[5], 425-434.
[38] Idaho National Laboratory, Biofuels and Renewable Energy, Power Curve Files. Website:
https://inlportal.inl.gov/portal/server.pt/community/
wind power/424/software/3993, accessed in 2009.
[39] Kaldellis, J.K., 2002. Parametrical investigation of the wind-hydro electricity production
solution for Aegean Archipelago, Energy Conversion and Management, 16, 20972113.
[40] Kaldellis, J.K., Kavadias, K.A., 2001. Optimal wind-hydro solution for Aegean Sea islands electricity-demand fulfilment, Applied Energy, 21, 333354.
[41] Kim, Seungdo, Dale, Bruce E., 2005. Environmental Aspects of Ethanol Derived from
No-Tilled Corn Grain: Nonrenewable Energy Consumption and Greenhouse Gas Emissions, Biomass and Energy 28, 475-489.
[42] Kirby, Natasha, Davison, Matt, 2010. Using a Spark-Spread Valuation to Investigate the
Impact of Corn-Gasoline Correlation on Ethanol Plant Valuation, Energy Economics,
32[6], 1221-1227.
BIBLIOGRAPHY
129
[43] KNMI Climate data and Services, Station 210 Valkenburg, the Netherlands. Website:
http://www.knmi.nl/klimatologie/onderzoeksgegevens/
potentiele wind, accessed in Summer 2008.
[44] Konrad, J., Carriveau, R., Davison, M., Simpson, F., Ting, D., 2012. Geological Compressed Air Energy Storage as an enabling technology for renewable energy in Ontario,
Canada, The International Journal of Environmental Studies 69[2], 350-359.
[45] Koplow, Doug, 2006. Biofuels At What Cost? Government Support for Ethanol and
Biodiesel in the United States, International Institute for Sustainable Development, Manitoba.
[46] Korpaas, M., Holen, A.T., Hildrum, R., 2003. Operation and sizing of energy storage for
wind power plants in a market system, Electric Power and Energy Systems, 7, 599606.
[47] Levy, H., Lessman, F., (1961). Finite Difference Equations, Macmillan Company, New
York.
[48] Little, Roderick, J.A., Rubin, Donald B., 2002. Statistical Analysis with Missing Data,
Second Edition, Wiley & Sons, Inc.
[49] Margrabe, William, 1978. The Value of an Option to Exchange One Asset for Another,
The Journal of Finance 33, 177-186.
[50] McDonald, Robert, Siegel, Daniel, 1986. The Value of Waiting to Invest, THe Quarterly
Journal of Economics 101[4], 707-728.
[51] Miranda, V., Bessa, R., Gama, J., Conzelmann, G., Botterud, A., 2009. New Concepts in
Wind Power Forecasting Models, Argonne National Laboratory.
[52] National Climate Data and Information Archive, Environment Canada. Website:
http://climate.weatheroffice.gc.ca/climateData/canada e.html,
accessed in Summer 2008.
[53] Offshore wind farm opens off the coast of the Netherlands. The
New York Times 18 Apr. 2007:
n. pag. Web. 14 Oct. 2009.
http://www.nytimes.com/2007/04/18/business/worldbusiness
/18iht-wind.4.5338427.htm
[54] Ontario Power Authority, Feed in Tariff Program, An Overview. Website:
http://fit.powerauthority.on.ca/
Page.asp?PageID=1115&SiteNodeID=1052, accessed in 2010.
[55] Paddock, James L., Siegel, Daniel R., Smith, James L., 1988. Option Valuation of Claims
on Real Assets: The Case of Offshore Petroleum Leases, The Quarterly Journal of Economics 103[3], 479-508.
[56] Paparoditis, E., Politis, D.N., 2002. Local Block Boostrap, Comptes Rendus Mathematique 335[11], 959-962.
130
BIBLIOGRAPHY
[57] Patzek, Tad W., Anti, S-M., Campos, R., Ha, K.W., Lee, J., Li, B., Padnick, J., Yee, SA., 2005. Ethanol From Corn: Clean Renewable Fuel For The Future, Or Drain On Our
Resources And Pockets?, Environment, Development and Sustainability 7, 319336.
[58] Peaceman, D. W. and Rachford, H. H. Jr., 1955. The Numerical Solution of Parabolic and
Elliptic Differential Equations, Journal of the Society for Industrial and Applied Mathematics 3, 28-41.
[59] Pimentel, David, 2003. Ethanol Fuels: Energy Balance, Economics, and Environmental
Impacts are Negative, Natural Resources Research 12, 127-134.
[60] Pimentel, David, 2005. Ethanol fuels: Energy security, economics, and the environment,
Journal of Agricultural and Environmental Ethics 4, 1-13.
[61] Pindyck, R., 1991. Irreversibility, Uncertainty, and Investment, Journal of Economic Literature 29, 1110-1148.
[62] Price, Anthony, 2005. Electrical Energy Storage - a Review of Technology Options, Proceedings of ICE, Civil Engineering 158, 52-58.
[63] Priestley, M.B., 1988. Non-linear and non-stationary time series analysis, Academic
Press, London.
[64] Rask, Kevin N., 1998. Clean Air and Renewable Fuels: The Market for Fuel Ethanol in
the US from 1984 to 1993, Energy Economics 20, 325-345.
[65] Rydh, J.C., Sanden, B.A., 2005. Energy Analysis of Batteries in Photovoltaic Systems.
Part I: Performance and Energy Requirements, Energy Conversion and Management 46,
1957-1979.
[66] Rydh, J.C., Sanden, B.A., 2005. Energy Analysis of Batteries in Photovoltaic Systems.
Part II: Energy Return Factors and Overall Battery Efficiencies, Energy Conversion and
Management 46, 1980-2000.
[67] Sargent, Thomas J., 1987. Dynamic Macroeconomic Models, Harvard University Press,
Massachusetts.
[68] Schmit, T.M., Luo, J, Conrad, J.M., 2011. Estimating the influence of U.S. Ethanol Policy
on Plant Investment Decisions: A Real Options Analysis with Two Stochastic Variables,
Energy Economics 33[6], 1194-1205.
[69] Sfetsos, A., 2000. A comparison of various forecasting techniques applied to mean hourly
wind speed time series, Renewable Energy 21[1], 23-35.
[70] Shapouri, Hosein, Duffield, James A., Wang, Michael, 2002. The Energy Balance of Corn
Ethanol: An Update, United States Department of Agriculture, Agricultural Economic
Report Number 813.
BIBLIOGRAPHY
131
[71] Smitherman, The Hon G., Minister of Energy and Infrastructure, 2009. Bill 150, An Act
to enact the Green Energy Act, 2009 and to build a green economy, to repeal the Energy
Conservation Leadership Act, 2006 and the Energy Efficiency Act and to amend other
statutes, Legislative Assembly of Ontario.
[72] Tiffany, Douglas G., 2006. Fossil Fuels and Ethanol Plant Economics, Managing High
Energy Costs, Redwood Falls, MN.
[73] Thompson, M., Davison, M., Rasmussen, H., 2004. Valuation and Optimal Control of
Electrical Power Plants in Deregulated Markets, Operations Research 52[4], 546-562.
[74] Weisser, D., Garcia, R.S., 2005. Instantaneous Wind Energy Penetration in Isolated Electricity Grids: Concepts and Review, Renewable Energy 30, 1299-1308
[75] Wen, Zhaoyin, 2006. Study on Energy Storge Technology of Sodium Sulfur Battery and
its Application in Power System, IEEE International Conference on Power System Technology (PowerCon 2006), 1-4.
[76] Wilmott, Paul, 2000. Quantitative Finance, Volume 2, John Wiley & Sons, Ltd., New
York.
[77] Wind
Energy
Center,
University
of
Massachusetts.
http://www.umass.edu/windenergy/
resourcedata.Bishop and Clerks.php, accessed in 2009.
Website:
[78] Yang B., Makarov, Y., Desteese J., Viswanathan, V., Nyeng, P., McManus, B., Pease, J.
(2008). On the Use of Energy Storage Technologies for Regulation Services in Electric
Power Systems with Significant Penetration of Wind Energy, EEM 2008, 1-6.
[79] Zhao, G., Davison, M., 2009. Valuing Hydrological Forecasts for a Pump Storage Facility,
Journal of Hydrology 373, 453-462.
Appendix A
Finite Difference Methods
Finite difference methods refer to numerical methods for finding approximate solutions to differential equations. This is done by replacing derivative terms with difference quotients which
are nearly equivalent. The use of difference quotients is obvious if we consider the definition
of a derivative,
f 0 (x) = limh0
f (x + h) f (x)
.
h
To determine the proper difference quotient, we use Taylor series approximations. The
forward difference quotient can be derived from
f (x + h) = f (x) + hf 0 (x) +
h2 f 00 (x)
+ ....
2!
Considering only the first order term we get

f 0 (x) =
f (x + h) f (x)
h
(A.0.1)
which looks very similar to the derivative definition given above. Similarly, the backward
difference is given as
f (x) f (x h)
f 0 (x) =
.
(A.0.2)
h
When solving differential equations in both space and time, it is common convention to
use a forward difference for time and a central difference for the space variable. The central
difference comes from the average of the forward and backward difference, or
f 0 (x) =
f (x + h) f (x
2h
h)
(A.0.3)
This can be extended to higher order derivatives as well. Using the second order Taylor series
approximation we can easily obtain
f 00 (x) =
f (x + h)
2f (x) + f (x h)
.
h2
See [47] for more details on finite differences.

132
(A.0.4)
A.1. C RANK N ICOLSON
A.0.1
133
A Review of Finite Difference Methods
Using the finite differences outlined in the previous section, we are ready to solve simple partial
differential equations. We limit our discussion to those methods needed in Chapter 5.
A.0.2 Implicit vs. Explicit

Let us consider a simple and popular PDE, the heat or diffusion equation, given below:
ut = uxx .
Before we write down the finite difference we must outline the notation. We assume a uniform
mesh in time, tn where n = 1, 2, ...N as well as a uniform mesh in our space variable, xi where
i = 1, 2, ..., I. We let t denote the time step and x denote the space step. Therefore, instead
of writing u(x, t) we can write u(ix, nt) and let u(ix, nt) = uni .
Using a forward difference in time and a second order central difference in x we obtain
un 2uni + uni1
un+1
uni
i
.
= i+1
t
(x)2
Rearranging this equation we obtain
un+1
i
t n
u + 1
=
(x)2 i1
t
(x)2
uni +
t n
u .
(x)2 i+1
(A.0.5)
This is called an explicit method; we can explicitly write u at time step n + 1 in terms of u at
time step n. Using a backward difference in time we obtain
un+1 2un+1
+ un+1
un+1
uni
i
i1
i
= i+1
,
t
(x)2
(A.0.6)
an implicit method due to the inability to write un+1 in terms of un . Each method comes with
advantages and disadvantages; the explicit method is numerically stable only when t/(x)2
1/2 but is easy to compute while the implicit method is unconditionally stable but is more computationally expensive.
A.1 Crank Nicolson

To get the best of both worlds, we take the average of both methods. Mathematically this gives
us
n+1
+ un+1
uni+1 2uni + uni1
1 un+1
un+1
uni
i+1 2ui
i1
i
=
+
.
t
2
(x)2
(x)2
Rearranging we obtain a convenient tridiagonal structure which is both numerically stable and
computationally inexpensive:
t
t
t
t
t
t
n+1
n+1
n+1
n
n
+
1
+
u
+
1
u
=
u
u
un .
i1
i
i1
i
i+1
2(x)2
(x)2
2(x)2
2(x)2
(x)2
2(x)2 i+1
(A.1.1)
134
C HAPTER A. F INITE D IFFERENCE M ETHODS
This method, known as the Crank Nicolson method, is also more accurate than the explicit and
implicit methods [21].
Of course, to solve a differential equation we must also be given a set of initial and boundary
conditions. Since these conditions are problem specific, we discuss their role in Chapter 5 when
we introduce the PDEs related to ethanol plant valuation.
A.2 ADI
The benefit of the Crank Nicolson scheme is that it is unconditionally stable (i.e. there is no
restriction on the maximal time step that can be used) and it involves linear systems with a
tridiagonal matrix. Tridiagonal systems can be solved very quickly in Matlab using the left
division operator. The ADI method can be thought of as the 2-D version of this scheme: it is
unconditionally stable and involves linear systems with tridiagonal matrices.
Rearranging (4.1.6) we obtain the tridiagonal systems seen in equation A.2.1. This system
can be solved for U k+1/2 (:, j) using the left division operator in Matlab and the result is then
used in stage two.
t
2t
0
1+ 2
h
h2
t
2t
t
..
1
+
h2
h2
h2
k+1/2
...
...
...
U
(:, j)
0
0
2t
t
t
..
.
1+ 2
2
h
h
h2
t
2t
0
0
1
+
h2
h2
2t
t
1
0
h2
h2
2t t
t
..
h2
h2
h2
k
...
...
...
U (i, :)
0
0
t
2t
t
..
.
1
h2
h2
h2
2t
t
1
0
0
h2
h2
Rearranging (4.1.7) we obtain the tridiagonal system seen in equation (A.2.2).
(A.2.1)
A.2. ADI
135
t
2t
h2
h2
t
2t
t
..
.
1
+
h2
h2
h2
k+1/2
..
..
..
U
(:, j)
.
.
.
0
0
t
2t
t
..
1+ 2
2
2
h
h
h
t
2t
0
0
1+ 2
h2
h
2t
t
1
0
h2
h2
t
2t t
..
h2
h2
h2
k
...
...
...
U (i, :)
0
0
t
2t
t
..
.
1
h2
h2
h2
2t
t
1
0
0
h2
h2
1+
(A.2.2)
This system is solved in the same way as the first and so, in two steps, both involving
solving a simple tridiagonal system, we are able to advance a complete time step without
having to restrict the size of this time step.
A.2.1
Transforming the Real Options PDE
In this section, the transformations involved in simplifying (4.1.6) are outined. These transformations allow us to create a constant coefficient tridiagonal system without mixed derivative
terms. The first transformation is as follows:
=T
t.
(A.2.3)
In many finance applications, it makes sense to write down a final condition rather than an
initial condition to accompany the PDE. This is since we know what happens at maturity,
t = T , but not initially, at t = 0. This transformation allows for the final condition to become
the more standard initial condition, and moves the time derivative term to the other side of the
equation since /t = 1.
The second transformation, a log transformation, is commonly used in PDEs with variable
coefficients. This transformation is also used in simplifying the Black Scholes PDE. The 2-D
version is
x = ln C
y = ln G,
(A.2.4)
(A.2.5)
eliminating the variable coefficients preceding the first, second and mixed order derivative
136
C HAPTER A. F INITE D IFFERENCE M ETHODS
terms. After the time and log transformations we are left with
V
1 2V
1 2V
V
V
= c2 2 + g2 2 + c
+ g
t
2 x
2 y
x
y
2
V
+c g
rV + (hey ex )H (hey ex ) = 0,
xy
a constant coefficient PDE. It is important to note that, with this log transformation in place, the
finite difference grid must be carefully constructed so that interior points are not too clustered;
in addition, the boundary must also be properly transformed to retain accurate solutions.
Next, we write
u = x
v = x + y.
In order to eliminate the term containing mixed partial derivatives, we must set
=
c
.
g
To eliminate the first order derivative terms we use the following transformations:
z = u (c a)
s = v (c a b)
= t,
where
a=
c2
2
and
b=
g c
c g
.
2
The final transformation is

W = V er .
(A.2.6)
This transformation eliminates the rV term, and we are left with

W
1 2W
1 2 2W
W
= c2
+
g
+ (c 0.5c2 )
+ (g
2
2
t
2 x
2 y
x
2W
+c g
+ ert (hey ex )H (hey ex ) = 0.
xy
0.5g2 )
W
y
(A.2.7)
or, incorporating the final two transformations discussed,

W
2W
a
+a 2 +( 2
a)
2W
+ (hf
y 2
j)H (hf j) = 0
(A.2.8)
A.2. ADI
137
where
(s
z b )g
c
f = e
j = ez+(c a) .
Of course the boundary conditions must also be transformed via the above transformations.
This equation resembles the 2-D heat equation, with the addition of the Heaviside function.
Due to the geometry of the transformations, as well as the existence of the Heaviside function,
solving this simplified equation does not lead to well behaved solutions for non-zero .
Curriculum Vitae
Name:
Natasha Burke
Post-Secondary
Education and
Degrees:
Memorial University
St. Johns, NL, Canada
2000-2004 B.Sc.
Memorial University
St. Johns, NL, Canada
2004 - 2006 M.Sc.
University of Western Ontario
London, ON, Canada
2006 - 2012 Ph.D.
Related Work
Experience:
Quantitative Analyst
RBC, Toronto
2010 - present
Teaching Assistant
The University of Western Ontario
2006 - 2010
Lecturer
Memorial University
Fall 2005
Honours and
Awards:
(1) NSERC PGS D, 2007-2010

(2) Western Graduate Thesis Research Award, 2008
(3) Student Poster Competition (2nd place),
Second Canada-France Congress, 2008
(4) Western Graduate Research Scholarship, 2006-2010
(5) Fellow of the School of Graduate Studies, 2006
(6) George Albert Hatcher Memorial Scholarship, 2005-2006
(7) F.A. Aldrich Graduate Award,2005-2006
(8) Lou Visentin Award, 2004
(9) Deans List Award, 2000-2004
(10) North Atlantic Refining Ltd. Entrance Scholarship, 2000
(11) Millenium Scholarship, 2000
Publications:
(1) Foster, A., Kirby, N., 2011. Analysis of a Heterogeneous Trader Model for Asset Price
Dynamics, Discrete Dynamics in Nature and Society, 2011, 1-12.
(2) Kirby, Natasha, Davison, Matt, 2010. Using a Spark-Spread Valuation to Investigate the
Impact of Corn-Gasoline Correlation on Ethanol Plant Valuation, Energy Economics, 32[6],
1221-1227.
(3) Kirby, N., Foster, A.,2009. Effects of Contrarian Investor Type in Asset Price Dynamics,
International Journal of Bifurcation and Chaos 19[8], 2463-2472.
(4) Kirby, N., Foster, A., 2006. A Dynamical Systems Approach to Asset Pricing, ASAC 2006
refereed conference proceedings, 57-71.
Presentations:
(1) Kirby, N., Davison, M., Anderson, L., A Real Options Approach to Valuing Wind as a
Renewable Energy Source, CORS 2009.
(2) Kirby, N., Foster, A., Asset Pricing Models: A Dynamical Systems Approach, CAIMS
2009.
(3) Kirby, N., Davison, M., Using Real Options to Value an Ethanol Plant, CMS 2009, St.
Johns.
(4) Kirby, N., Davison, M., Options Pricing in Two Dimensions Using Numerical PDEs, MITACS 2009.
(5) Kirby, N., Davison, M., A Real Options Valuation of Ethanol Production (poster), Second
Canada-France congress, June 2008.
(6) Kirby, N., Davison, M., Optionality for Ethanol Producers, CORS, May 2008.
(7) Kirby, N., Davison, M., Spark Spread Options and Ethanol Production, SWORD 2007.
(8) Kirby, N., Davison, M., Ethanol as an Alternative to Petroleum Fuel, CORS, May 2007.

A Real Options Valuation of Renewable Energy Projects PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Real Options Valuation of Renewable Energy Projects PDF

Uploaded by

Copyright:

Available Formats

A REAL OPTIONS VALUATION OF RENEWABLE ENERGY PROJECTS

(Thesis format: Monograph)

Graduate Program in Applied Mathematics

A thesis submitted in partial fulfillment

The School of Graduate and Postdoctoral Studies

c Natasha Burke 2012

THE UNIVERSITY OF WESTERN ONTARIO

Corn Ethanol Plant Valuation

Real Options PDEs

5 Wind and Storage

7 Wind/Storage System Valuation

Discussion and Conclusions

A Finite Difference Methods

The Margrabe parameter, , and the option value, M , as a function of . . . .

Corn and Gasoline prices, January 1976 December 2006 . . . . . . . . .

Histogram of Log Returns - Corn Data . . . . . . . . . . . . . . . . . . . . . .

Power Curve - Vestas 1.8 MW Turbine [38] . . . . . . . . . . . . . . . . . . .

V (F, k) and V (E, k) as k increases: p = 0.39, = 0.2, M = 0.04 . . . . . .

Average Annual Spark Spread . . . . . . . . . . . . . . . . . . . . . . . . . . 17

Historical GBM Parameter Estimates . . . . . . . . . . . . . . . . . . . . . . . 34

Benefit Matrix: Empty Battery . . . . . . . . . . . . . . . . . . . . . . . . . . 56

Wind Variables and Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 89

1.1 Corn Ethanol as Renewable Energy

1.2. W IND AS A R ENEWABLE E NERGY S OURCE

Wind as a Renewable Energy Source

1.3 Modeling Choices

1.4. T HESIS ROAD M AP

Thesis Road Map

where S is the stock price and K is the strike price.

2.1. M ARGRABE O PTIONS

with final condition

ln (S1 /S2 ) + (1 2 + 2 /2)

C HAPTER 2. R EAL O PTIONS

2.2 Options vs. Real Options

2.3. M ATHEMATICAL BACKGROUND

2.3 Mathematical Background

Ws N (0, t s), t > s

the sample paths are continuous

C HAPTER 2. R EAL O PTIONS

2.4 A Simple Real Options PDE

and the fact that E[dX] = 0 to obtain

2.5. R EAL O PTIONS AND R ENEWABLES

Real Options and Renewables

3.1 Spark Spread Options

3.1. S PARK S PREAD O PTIONS

run a particular generation asset.

448, 000 476, 000

C HAPTER 3. C ORN E THANOL P LANT VALUATION

3.2 Retrospective Analysis

3.3 Corn and Gasoline Price Dynamics

3.3. C ORN AND G ASOLINE P RICE DYNAMICS

C HAPTER 3. C ORN E THANOL P LANT VALUATION

3.3. C ORN AND G ASOLINE P RICE DYNAMICS

Table 3.1: Average Annual Spark Spread

Table 3.2: Probability of k highs assuming a random draw

Table 3.3: Frequency of high/low prices

C HAPTER 3. C ORN E THANOL P LANT VALUATION

Corn Price (dollars/bushel)

Figure 3.3: Fourier Analysis - Corn (January 1997-December 2006 data)

3.3. C ORN AND G ASOLINE P RICE DYNAMICS

Gasoline Price (dollars/gallon)

Figure 3.4: Fourier Analysis - Gasoline (January 1997-December 2006 data)

C HAPTER 3. C ORN E THANOL P LANT VALUATION