Quantifying the system balancing cost

when wind energy is incorporated into
electricity generation system
Natalia Issaeva
Contents
Notation 1
1 Introduction 4
1.1 Benefits of wind energy . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Challenges of wind energy . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Balancing of the electricity system . . . . . . . . . . . . . . . . . . 6
1.4 Timescale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Aim of the research . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.6 Structure of thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Wind Speed Analysis 13
2.1 Generating wind energy . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Generating wind speed data . . . . . . . . . . . . . . . . . . . . . 14
2.2.1 One-minute wind speed data . . . . . . . . . . . . . . . . . 15
2.2.2 Methods of generating missing data . . . . . . . . . . . . . 18
2.2.3 Description of Gibbs sampling . . . . . . . . . . . . . . . . 19
2.2.4 Transformation to normal variable . . . . . . . . . . . . . 23
2.2.5 Suggested modifications to Gibbs sampler . . . . . . . . . 27
2.3 Implementation of modified Gibbs sampler . . . . . . . . . . . . . 30
2.3.1 Multivariate normal distribution . . . . . . . . . . . . . . . 31
2.3.2 Two-variable sampling . . . . . . . . . . . . . . . . . . . . 35
2.3.3 Multivariate sampling . . . . . . . . . . . . . . . . . . . . 38
2.3.4 Testing the results of wind speed sampling . . . . . . . . . 43
2.4 Diversification of wind energy . . . . . . . . . . . . . . . . . . . . 46
3 Modelling Electricity Generation with Wind Energy 48
3.1 Elements of the British power system and simplifying assumptions 48
3.1.1 Electricity generation . . . . . . . . . . . . . . . . . . . . . 49
3.1.2 Cost of generation . . . . . . . . . . . . . . . . . . . . . . 51
3.1.3 Electricity demand . . . . . . . . . . . . . . . . . . . . . . 52
i
3.1.4 Cost of demand-and-supply mismatch . . . . . . . . . . . . 53
3.1.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.2 Deterministic model for the electricity generation with wind energy 55
3.2.1 Decision variables . . . . . . . . . . . . . . . . . . . . . . . 55
3.2.2 Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2.3 Objective function . . . . . . . . . . . . . . . . . . . . . . 59
3.2.4 Formulation of the LP deterministic optimization model . 60
3.3 Calculating cost of wind energy variability . . . . . . . . . . . . . 60
3.3.1 Case of constant wind energy that reduces the operational
cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.3.2 Case of variable wind energy that increases operational cost 63
4 Statistical analysis of operational cost 67
4.1 Formulation of the problem . . . . . . . . . . . . . . . . . . . . . 67
4.2 Case of marginal generator following a profile of the wind energy
curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.3 Case of wind energy curve having a zigzag profile . . . . . . . . . 70
4.3.1 Analytical calculation of the accumulated fuel cost . . . . 71
4.3.2 Improving the estimation of the accumulated fuel cost . . . 77
4.4 Case of the wind energy curve with changing standard deviation . 79
5 Stochastic Modelling 82
5.1 Uncertainty of the wind energy . . . . . . . . . . . . . . . . . . . 83
5.2 Uncertainty of lost generating capacity . . . . . . . . . . . . . . . 89
5.3 Solving a problem with the uncertainty . . . . . . . . . . . . . . . 91
5.3.1 Decision variables and constraints of the LP stochastic op-
timization model . . . . . . . . . . . . . . . . . . . . . . . 91
5.3.2 Formulation of the LP stochastic optimization model . . . 96
5.3.3 Output of the stochastic model with the uncertainty of the
wind power and the capacity loss . . . . . . . . . . . . . . 97
5.4 The system balancing cost when there is the uncertainty of the
wind power in the electricity generation system . . . . . . . . . . 101
6 Conclusion 103
Appendices 106
A. Definitions by the National Grid . . . . . . . . . . . . . . . . . 106
Bibliography 110
ii
List of Figures
1.1 Schematic description of the power system . . . . . . . . . . . . . 6
1.2 Process of balancing power system shown in terms of the Settle-
ment Period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Wind energy that would have been generated by a small wind farm
in Aberystwyth on the 5th of January,2001 . . . . . . . . . . . . . 10
2.1 Transformation function between the wind speed and the wind power 14
2.2 Autocorrelation of the original and the deseasonalised wind speed
data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Schematic description of Gibbs sampler as a Markov-chain process 21
2.4 Cumulative probability distribution of one-minute wind speed recorded
with MST Radar located at Frongoch farm . . . . . . . . . . . . . 23
2.5 Truncated and full plot of normal probability density function . . 24
2.6 Normal variable plot with respect to the deseasonalised wind speed 26
2.7 Fitted sum of exponential functions to approximate the autoco-
variance of the normalised wind . . . . . . . . . . . . . . . . . . . 32
2.8 Contours of Normal distribution and four linear functions used in
the modification to Gibbs algorithm . . . . . . . . . . . . . . . . . 36
2.9 Amount of the information carried by each of the ten variables . . 42
2.10 An example of the sampled and the original values of 1-minute
wind speed plotted along with the ten minute average wind speed 44
2.11 Autocovariance of the sampled and the original normalised 1-minute
wind speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.12 An example of wind speed sampled for one of the Utah sites . . . 46
2.13 An example of geographically distributed wind energy output . . 47
3.1 Half-an-hour average of the daily electricity demand in the United
Kingdom and its piecewise linear approximation with 1-minute
frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
iii
3.2 A mismatch between the electricity demand and the generation
with the cost attached . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3 Area formed by Run-Up and Run-Down constraints for thermal
generators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.4 Load mismatch costs: the difference between the curves gives the
penalty for generation deviating from demand. . . . . . . . . . . . 57
3.5 Electricity generation when 11% of wind energy with flat profile is
introduced to the system . . . . . . . . . . . . . . . . . . . . . . . 62
3.6 Electricity generation when 11% of variable wind energy is intro-
duced to the system . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.7 Electricity generation and load mismatch when 5% of variable wind
energy is introduced to the system . . . . . . . . . . . . . . . . . 65
3.8 The fuel cost of the wind energy variability shown for different
levels of it introduced into the electricity system . . . . . . . . . . 65
4.1 Electricity generation with only the marginal generator modifying
the power output . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Electricity generation with the wind energy profile as a zigzag . . 70
4.3 Electricity generation for a wind energy curve as one zigzag and
the thermal plants increasing and decreasing their output in a sym-
metric Λ-shape . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.4 An example of the solution with the thermal plants increasing and
decreasing their output in symmetric shapes . . . . . . . . . . . . 73
4.5 Seven cheapest thermal plants increasing and decreasing their out-
put in symmetric shapes . . . . . . . . . . . . . . . . . . . . . . . 73
4.6 Seven most expensive thermal plants increasing and decreasing
their output in symmetric shapes . . . . . . . . . . . . . . . . . . 73
4.7 Comparing the calculated and the LP-modelled marginal fuel costs 77
4.8 Fitting a quadratic term into the model that estimates the accu-
mulated fuel cost of electricity generation . . . . . . . . . . . . . . 78
4.9 Comparison of the error term for the LP-modelled and the calcu-
lated accumulated fuel cost with the model (4.2) for a set of 30
wind energy curves . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.10 Comparison of the error term for the LP-modelled and the calcu-
lated accumulated fuel cost with the model (4.7) for a set of 30
wind energy curves . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.11 Comparing an error of calculating the accumulated fuel cost of the
electricity generation with the models (4.2), (4.7) and (4.8) . . . . 81
iv
5.1 An example of a scenario tree . . . . . . . . . . . . . . . . . . . . 84
5.2 Tree of scenarios that represent possible changes in wind speed . . 85
5.3 A scenario tree of the changes in the electricity load equivalent to
the loss of the capacity . . . . . . . . . . . . . . . . . . . . . . . . 90
5.4 An example of the rolling horizon stochastic programming solution
with the uncertainty of the capacity loss when 11% of the wind
power is incorporated in the total generation . . . . . . . . . . . . 98
5.5 An example of the rolling horizon stochastic programming solution
with the uncertainty of the wind power output and the capacity loss
when 7% of the wind energy is incorporated in the total generation 100
5.6 The fuel cost of the uncertainty in wind power output and the
available thermal capacity, shown for different levels of the wind
introduced into the electricity system . . . . . . . . . . . . . . . . 101
v
List of Tables
2.1 Covariance matrix ϕ calculated for the normal variables trans-
formed from the minute-by-minute wind speed data . . . . . . . . 40
2.2 M-matrix such that MM
T
= ϕ . . . . . . . . . . . . . . . . . . . 40
2.3 Variance of the original and the sampled normalised wind speed . 45
3.1 Plant/Demand balance for SYS Background (2007/2008) . . . . . 50
3.2 An example of the Run-Up and Run-Down export rates for a typ-
ical coal plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.1 Accumulated fuel cost weighted by the generated power and cal-
culated for different branching combinations, £/(MW*day) . . . . 88
vi
Abstract
Incorporation of wind energy into the electricity generation system requires a
detailed analysis of wind speed in order to minimize system balancing cost and
avoid a significant mismatch between supply and demand. Power generation
and consumption in the electricity networks have to be balanced every minute,
therefore it is necessary to study wind speed on a one-minute time scale. In this
thesis, we examine the statistical characteristics of one-minute average values
of wind speed. One-minute wind speed is available from a single site in Great
Britain while there are records of ten-minute wind speed available. We apply a
modified Gibbs sampling algorithm to generate one-minute wind speed required
for optimization modelling from the available ten-minute wind speed.
System balancing costs are estimated through optimization modelling of the
short-term electricity generation with wind energy contributing to the total sup-
ply. Two main drivers of additional system cost caused by wind energy are vari-
ability and unpredictability of one-minute wind speed. Further, a linear math-
ematical optimization model for a problem of short-term electricity generation
is presented to calculate an additional balancing cost that appears as a result
of wind energy variability. It is then shown that this additional balancing cost
can be estimated using the statistical characteristics of wind energy present in
the system. The unpredictable characteristic of wind speed is analysed with the
techniques of stochastic programming. Uncertainty of the expected wind speed
is represented through scenario trees and stochastic linear optimization models
are used to calculate the extra cost due to uncertainty. Alternative optimization
models are compared by calculating the additional balancing cost and the extent
of imbalance between power generation and consumption in the system.
Notation
G a set of all generators
C a set of conventional generators
W a set of wind generators
T a number of time periods
T
Υ
a number of time periods in one year
T
S
a modelling horizon
T
n
a number of intervals associated with node
N a number of nodes
S a number of scenarios
V a number of intervals of mismatch between
the electricity supply and demand
α power law exponent
t, τ, ν, d, δ, m a single period of time
l, L a lag
i, j, k an index
g an index of generator
n a node
a
n
an ancestor of node
p
n
a path probability
κ a number of thermal generators
f
+
w
saving in fuel cost
f

w
additional fuel cost
S
i
, S
j
a set of indices
z, z
r
height where wind speed was recorded
θ, q, s
g
, k
1
, k
2
parameters
λ a parameter of Lagrangian function
β
±
percent of cost band
ǫ an error term
1
b
1
, b
2
, b
3
parameters of quadratic function
β
i
, ς
i
parameter of an exponential curve
a
1
, a
2
, ..., a
T
Υ
deseasonal scaling curve
r
±
1
, r
±
2
, r
±
3
Run-Up and Run-Down export rates
e
±
1
, e
±
2
“elbows” of Run-Up and Run-Down export rates
D
1
, D
2
, ..., D
T
the electricity demand
c
g
, c
coal
, c
gas
fuel costs
c
m
fuel cost of the marginal generator
ζ
±
1
, ζ
±
2
, ..., ζ
±
V
cost of imbalance between the electricity supply and demand
k
dm
a weight in calculation of deseasonal scaling curve

D
a percent mismatch allowed between
the electricity supply and demand
h
w
height of wind energy curve
γ
w
the average of the absolute value of gradients of wind energy curve
γ
g
the average of the absolute value of gradients of thermal plants
γ
g
maximum of the absolute gradients of thermal plants
γ

sum of the absolute value of gradients of thermal plants
σ
x
standard deviation of wind energy
σ
γ
standard deviation of the absolute value of gradients
µ mean value
σ standard deviation
ϕ covariance matrix
ψ an inverse to covariance matrix
M a matrix such that MM
T
= ϕ
1 vector consisting of scalers 1 only
Y
1
, Y
2
, ..., Y
k
a vector of normal variable (normally includes 10 values)
U
1
, U
2
, ..., U
k
a vector of a standard normal variable (normally includes 10 values)
X
1
, X
2
, ..., X
k
a vector of decision variable (normally includes 10 values)
W
s
1
, W
s
2
, ..., W
s
k
a vector of wind speed (normally includes 10 values)
W
d
1
, W
d
2
, ..., W
d
k
a vector of deseasonalised wind speed (normally includes 10 values)
W
e
1
, W
e
2
, ..., W
e
k
a vector of wind energy (normally includes 10 values)
y
1
, y
2
, ..., y
k
one-minute normal variable
u
1
, u
2
, ..., u
k
one-minute standard normal variable
x
1
, x
2
, ..., x
k
one-minute decision variable
w
s
1
, w
s
2
, ..., w
s
k
one-minute wind speed
w
d
1
, w
d
2
, ..., w
d
k
one-minute deseasonalised wind speed
w
e
1
, w
e
2
, ..., w
e
k
one-minute wind energy
2
d
±
1
, d
±
2
, ..., d
±
V
mismatch between the electricity supply and demand
W
s
, w
s
the average of wind speed over a period of time
x
g
, x
g
maximal and minimal capacity limits
d
±
, d
±
upper and lower limits of mismatch intervals
Y
A
, Y
B
, Y
C
, Y
D
vectors
ˆ y a value of normal variable
f(∗), h(∗) a function
P(∗) a probability function
φ(∗) probability density function
Φ(∗) cumulative distribution function
Θ
d
(W
s
, a
ν
) function transferring original wind speed into deseasonal
Θ
n
(W
d
) function transferring deseasonalised wind speed into normal variable
F
0
, F
flat
w
, F
var
w
accumulated fuel cost of the deterministic optimization model
F
(a)
stoch
, F
(b)
stoch
accumulated fuel cost of the stochastic optimization model
F
A
actual accumulated fuel cost
F
m
marginal fuel cost
[Y
1
, Y
2
], [Y
1
|Y
2
], [Y ] joint, conditional and marginal densities
3
Chapter 1
Introduction
How expensive is the operating of the electricity generation system when wind
energy is introduced into it? This issue is becoming important as the amount
of renewables in the system increases in order to meet the target that the Gov-
ernment of United Kingdom has set. In 2002 the Government introduced the
Renewable Obligation (RO) to encourage development of renewable sources of
energy[1, 2] and aimed to reach 15.4% of renewables in the overall electricity
supply by 2015. Generators receive an RO Certificate (ROC) for each 1MWh of
renewable electricity they generate. These are sold to electricity suppliers (de-
fined in Appendix A) that pay the fixed penalty for each MWh it falls short of
its obligation[1].
Various renewable technologies can contribute to the goal set by the Gov-
ernment but wind resource takes a leading competitive position among other
renewable sources of energy (solar, tidal, wave, hydro) by the availability and ef-
ficiency. However, using wind for the electricity generation introduces a problem
of intermittency and unpredictability of wind energy that will be studied in this
work. The problem of meeting a target set by the Government of United King-
dom was investigated by the researchers of the Environmental Change Institute
at Oxford University. In the report “Wind power and the UK wind resource”[10],
Sinden states that in order to integrate wind power into electricity networks it is
essential to understand the characteristics of wind resources. He believes variabil-
ity of the wind energy is one of the obstacles to its extensive use by an electricity
supplier and suggests a diversified renewable energy portfolio that includes wave
and tidal steam combined with wind power. Such a combination reduces effect of
variability in wind speed and consequently the cost of balancing electricity supply
and demand.
4
1.1 Benefits of wind energy
In the 21st century the use of modern wind turbines for electricity generation
has been growing very rapidly in Europe and this growth is expected to continue.
Integration of wind generation in the operation and development of the electricity
system is associated with both benefits and costs. The following reasons create
plenty of opportunities to realise the technical potential of renewable energy,
particularly wind power:
• Wind turbines produce green energy unlike conventional generators that
produce CO
2
emissions;
• Renewable sources of energy do not have fuel costs that contribute the most
to the expenses of the conventional generators;
• In spite of high percentages of natural gas and coal in British electricity
generation, these resources are limited and their long-term supply can be
disrupted. Meanwhile wind, tides, solar are known as renewable sources of
energy.
1.2 Challenges of wind energy
However desirable the incorporation of wind energy into the electricity system
there are a number of issues associated with its installation and management of
the system. Detailed estimation of additional costs associated with the presence
of the wind energy in the electricity generation system was performed by a few
researchers since 2002 ([8, 9, 10, 11, 12]). Strbac et al. wrote a detailed paper
on “Quantifying the system costs of additional renewables by 2020”[8] that ex-
amined the issue of challenges that incorporation of additional renewable sources
faces. Besides the high investment cost of renewable sources of electricity, there
is additional cost that occurs when balancing intermittent and variable wind en-
ergy. The work of Strbac et al. shows that in the current system wind energy is
not able to replace fully the capacity and flexibility of conventional generating
plants.
This thesis concentrates on the system costs associated with the balancing
of supply and demand, however, there are also system costs appearing in trans-
mission and distribution of electricity. Strbac et al.[8, 9] state different system
cost drivers, such as location of the generation sources in relation to the source of
demand, lack of flexibility in generation as well as intermittency and variability
5
of supply. Moreover, wind turbines are generally not able to provide the range of
system support services (eg. voltage and frequency regulation) that are provided
by conventional thermal and hydro plant. At relatively low levels of penetration
this can usually be tolerated, but at the higher levels indicated by the Renew-
ables Obligation target, it will require systematic solutions in order to maintain
stability and integrity of the transmission system. These issues reduce the value
of wind generation that displaces conventional generation.
1.3 Balancing of the electricity system
Electricity generation and delivery to the customers is a complex process. It is de-
scribed in detail by the National Grid and Balancing and Settlement Codes[3, 4]
that we refer to while formulating a problem of scheduling the electricity gener-
ation.
The main groups included into the electricity generation and distribution sys-
tem are presented on Figure 1.1.
Figure 1.1: Schematic description of the power system
Most of the electricity in the power system is contracted. Some of the contracts
are signed directly between the Generating companies and large industrial con-
sumers. Residential consumers pay for their electricity through the Suppliers,
companies like EDF, Scottish power or npower. Suppliers sign agreements to
deliver contracted amounts of power to the Consumers. At the same time the
Generators have an obligation to provide this power to the transmission system.
6
This section explains general terms of the system functioning while formal defi-
nitions of what these participants represent are given in Appendix A.
The interaction process of different participants in the electricity generation,
transmission and consumption is complex and contains many procedures. Let us
give a general description of this process in terms of a Settlement Period, a
half-hour period of time when real-time balancing of the electricity supply and
demand takes place (Figure 1.2).
Figure 1.2: Process of balancing power system shown in terms of the Settlement
Period
In order to manage the electricity transmission system efficiently National Grid
Electricity Transmission (NGET), company that stands for the System Operator,
collects the necessary data from the appointed parties. Pre Gate closure pro-
cedures are intended for the submission of Balancing Mechanism (BM) Unit Data
(as defined in Appendix A) which includes characteristics of each generator such
as maximum and minimum power output, ramp up rates, availability and other.
This information would enable NGET to assess which BM Units are expected to
be operating in order that NGET can ensure (so far as possible) the integrity of
the Great Britain Transmission System, and the security and quality of supply
[4]. From the information provided by BM Units the System Operator has an
insight into how much it can vary their generation above or below their submitted
profile and what price must be paid for every possible amount of deviation from
the planned generation.
Gate Closure takes place one hour before the beginning of the Settlement
Period. Data for any Operational Day may be submitted to NGET up to several
7
days in advance of the day to which it applies, as provided in the Data Vali-
dation, Consistency and Defaulting Rules. The data to be used by NGET for
operational planning will be determined from the most recent data that has been
received by NGET by 11:00 hours on the day before the Operational Day to
which the data applies, or if no data has been submitted by 11:00 hours on that
day by the default data values. The individual data items that Generators and
Suppliers submit include Physical Notifications, Quiescent Physical Notifications,
Export and Import Limits, Bid-Offer Data and Dynamic Parameters (all defined
in Appendix A).
NGET as the System Operator has the responsibility of providing the informa-
tion on the Demand Estimates. On the basis of historic trends in the availability
of generating capacity and the data submitted by the generating units NGET also
decides on the necessary Contingency and Operating Reserve, as well as System
Margins.
Electricity transmission and distribution system of Great Britain operates
with a certain Target Frequency. The frequency is allowed to vary 1% up
or down from 50Hz. As long as the amount of produced electricity equals the
demand, frequency in the transmission system stays at 50Hz. If the the gener-
ation exceeds demand frequency would rise above 50Hz and if demand exceeds
generation then frequency in the system drops. The System Operator balances
the supply and the demand in the system every moment of the real time through
different procedures and is responsible for keeping the system frequency in the
target range.
In order to deal with unpredicted variations in demand and generation, the
System Operator requires an appropriate automatic Response, to neutralise rapid
variations from a few seconds to a few minutes, and operating Reserve to deal
with slow variations over time horizons from several minutes to several hours
[24]. Response is defined as an automatic change in active power output by the
large generating units that effects the frequency in the system. Response is used
as an earliest available instrument of the balancing mechanism. It supports the
system from the first second up to 30 minutes after the imbalance was recorded.
By that time the operating Reserve is ready to change the Export (Import) level
(as defined in Appendix A). Some of the generating units acting as the Reserve
are operating at part-time load, some are off-line but able to start within a short
time. Pumped storage, for example, can respond very rapidly to counteract any
loss of generation or surge in demand. Gas turbines are able to provide generation
on timescale of few minutes while steam plant would require a few hours to warm
8
up. The participants need to ensure that each of their BM units, applying Good
Industry Practise, follows the submitted notification of export or import level.
Balancing costs appear when the System Operator uses or keeps on stand-
by flexible but expensive generating units in order to deal with an unexpected
mismatch between the electricity supply and demand or operates generators at
non-optimal generation levels so as to provide flexibility for rapid change in out-
put level. The levels of reserve required at any given time depend partly on
uncertainties in the predictions of demand, but also on the need to deal with the
sudden loss of substantial amounts of generation, either due to a power station
faults or the loss of transmission circuits. In England and Wales, for example,
key criteria are possible loss of one circuit of the cross-Channel link (1000 MW),
or of Sizewell B nuclear power station (1320 MW). The value of possible loss of
generating capacity can significantly increase after wind energy is introduced into
the power system.
1.4 Timescale
Electricity supply and demand in the transmission and distribution power systems
need to be balanced every single moment of time. On the scale of seconds it is
managed through the automated process of frequency response. At the same time,
scheduling of electricity generation reserve on hourly and daily basis has been ex-
tensively studied by various researchers, for example in [29, 30, 31, 32, 33]. Hourly
and daily production planning of power systems involves start-up and mainte-
nance of the thermal plants. This leads to dynamic, mixed-integer programming
problems, and methods of solving these models were analysed by R¨ omisch, M¨ oller,
Dentcheva, Feltenmark and other researchers ([29, 30, 31]) in the 1990s. Later,
mixed-integer programming problems were extended by capturing the uncertainty
in thermal power systems and the stochastic mixed-integer programming was ap-
plied to solve complex models ([32, 33]).
We wish to study scheduling of the electricity generation at time intervals
that were not well-researched before but have an important application in the
industry. This work focuses on short-term variability of the wind energy and the
electricity generation with one-minute frequency. At this short scale there can be
no decision made on the start-up or maintenance of thermal plants, therefore, we
avoid the complexity of mixed-integer programming problems.
When scheduling the operating Reserve, the System Operator takes into ac-
count the uncertainties in demand and generation on various timescales. Uncer-
9
tainty increases with the time horizon, but broadly speaking, the costs of the
appropriate Reserve decrease as the time scale over which they have to respond
increases. Operating Reserve (as defined in Appendix A), for example, may cost
around £1/MWh, but fast Response plant may cost up to £5/MWh, or more[24].
Figure 1.3 plots an example of the wind power that would have been generated
by a small wind farm in Aberystwyth based on the recorded wind speed. It varies
significantly from minute to minute
1
which makes variability of wind energy one of
the strongest drivers of the balancing cost. One possible solution for this problem
involves diversifying locations of wind farms. Another solution suggested by
Sinden in his work on diversified renewable energy portfolios[12], was to combine
wind energy with more predictable tidal and wave energy.
1
2
3
4
5
6
7
8
100 120 140 160 180 200
e
n
e
r
g
y
,
M
W
time,min
Wind energy that one wind farm would generate at Aberystwyth in 2001,Januray,5,01:40-03:20
wind energy curve
Figure 1.3: Wind energy that would have been generated by a small wind farm
in Aberystwyth on the 5th of January,2001
1.5 Aim of the research
Different participants of the electricity industry (Figure 1.1) can benefit from the
efficient operation and planning of electric power generation systems. On the
one hand, suppliers of energy reduce the cost of mismatch between the reported
and actual electricity load. On the other hand consumers receive uninterrupted
electricity flow in their houses and offices. This thesis is aiming at the vertically
integrated utilities (like Scottish Power or EDF) that are capable of combining
1
The MST Radar Facility at Aberystwyth is funded by the UK Natural Environment Re-
search Council and the data are provided through the British Atmospheric Data Centre
10
different energy portfolios and, thus, minimize their costs of meeting the con-
tracted demand.
NGET acting as the System Operator manages the electricity generation sys-
tem of the United Kingdom. Its profit is independent of the operational cost
of the electricity generation, however the System Operator has to ensure the in-
tegrity of the Great Britain Transmission System while minimizing the balancing
costs.
1.6 Structure of thesis
In Chapter 2 we investigate how the wind speed varies and how this affects
the wind power generation. There is only one publicly available site of wind
speed one-minute averages. In order to obtain a wind power output from the
geographically diverse sources we will develop a Gibbs sampling algorithm to
generate one-minute average wind speed from available average ten-minute values
of wind speed at geographically dispersed locations.
Obtained in Chapter 2, wind power output is used later in the thesis when
testing the optimization model of the electricity generation. Chapter 3 describes
the power system of the United Kingdom and how it can be modelled using
linear programming. By setting different parameters of the deterministic model,
the system balancing cost of wind power variability is calculated. In deterministic
models we assume that wind can be predicted exactly, so the costs of incorporating
the wind power into the electricity generation system originate only in variability
of wind speed.
The additional system balancing cost that appears after the wind power is
introduced into the electricity generation system can be modelled as a function
depending on the statistical parameters of one-minute wind power output. There
is a number of tests performed in Chapter 4 that derives a model approximating
the operational cost of the electricity generation.
In practise it is not possible to predict the wind speed accurately even over
a short time period of 30 minutes. Chapter 5 investigates into the additional
system balancing cost that comes from coping with the unpredictable nature of
wind speed. The problem of electricity generation is then solved using stochastic
programming methods where the uncertainty is represented with a scenario tree.
The issue is complicated by a possible loss of the generating capacity. This
uncertainty can also be pictured on the scenario tree and resolved along with the
uncertainty of the wind power output using stochastic programming.
11
Deterministic and stochastic models formulated in this work are implemented
using a mathematical programming modelling language (AMPL)[47]. Data used
for the solving of the electricity scheduling problem are managed using JAVA
programming language.
12
Chapter 2
Wind Speed Analysis
The aim of this thesis is to study electricity generation on one-minute time scale.
Implementation of optimization models requires data with the appropriate reso-
lution, however, there is no one-minute wind speed available on the country level.
Least frequent data is provided by two sources: the Meteorological Office of the
United Kingdom that publishes wind speed averaged over one hour for different
sites in the UK and Utah Geological Survey (USA) [21] that gives an access to
ten-minute average wind speed with standard deviation known.
To obtain the necessary data, in this chapter we wish to study the only avail-
able source of one-minute wind speed recorded in the UK ([20]) and apply this
information to generating an unbiased sample conditional on a given average
value. Samples generated for different sites of a geographically diverse territory
are further transformed into power that can be used for optimization modelling
of the electricity generation system.
2.1 Generating wind energy
United Kingdom is situated on an island with a unique location that defines its
climate. Due to the proximity to the Atlantic Ocean and the North Sea, weather
on the British islands frequently changes producing high wind speeds. As a result
the United Kingdom has some of the best wind resources in Europe, if not the
world, in both onshore and offshore locations. This makes the country a very
attractive location for wind developments, as high average wind speeds and good
reliability result in more power output and lower costs[6].
Currently there are over 200 onshore and offshore wind farm projects oper-
ating in Great Britain with a total installed capacity of 3.3GW. Wind energy
contributes around 2.2% of the total electricity generation in the United King-
13
dom and this number is expected to grow as more wind farms are connected to
the grid. Rapidly developing technology introduces more efficient wind turbines
built with the towers up to 100m tall that can capture strong winds. The aver-
age turbine size in 2008 was considered as 2MW that produces energy with the
average output of 40% of the maximum possible capacity[6].
In order to estimate wind energy produced by a turbine, the available wind
speed is transformed into power using a wind-power curve. These curves are
uniquely determined for the different wind turbine type. Let us take NEG Micon
2750/92 as a typical wind turbine with the capacity 2.75MW, 3m/s cut in and
25m/s cut out wind speed. Its hub height is 70m and rotor diameter is 92m. The
wind-power curve for this wind turbine is plotted in Figure 2.1.
0
0.5
1
1.5
2
2.5
3
0 5 10 15 20 25 30
p
o
w
e
r
,

M
W
wind speed, m/s
Figure 2.1: Transformation function between the wind speed and the wind power
2.2 Generating wind speed data
This section starts with the analysis of the available source of one-minute wind
speed recorded in the United Kingdom ([20]) that is used to identify the probabil-
ity distribution required for sampling one-minute wind speed. Further, we wish
to develop a unique method that generates an unbiased sample from the wind
speed stochastic process conditional on each consecutive group of wind speeds
having a given average value. A new algorithm includes Monte Carlo Markov
Chain method, Gibbs sampling algorithm, originally adapted to a problem of
generating one-minute wind speed data.
14
2.2.1 One-minute wind speed data
The one-minute resolution wind speed stochastic process is based on the statistical
properties of one-minute wind speed measurements made at one location. The
one-minute wind speed measurements required for this work are provided by
Mesosphere-Stratosphere-Troposphere (MST) Radar located at Frongoch farm
near Aberystwyth (140 m above mean sea level, 52.42N, 4.05W; British National
Grid Reference SN 604826) funded by the Natural Environment Research Council
(NERC). Wind speed and direction are recorded at one-minute intervals using
vector instruments, an A100R anemometer and a W200P wind vane[20]. This
research uses 12 years of data for the period between 1996 and 2007.
Further adjustments are necessary. Wind speed read by the MST Radar is
registered at 5-meter height, however, a hub height of NEG Micon 2750/92 (a
typical wind turbine for this research) is 70 meters. Thus, it is necessary to
transfer available 1-minute wind speed to this height. The power law represents
a simple model for the vertical wind speed profile. Its basic form is provided by
Manwell, McGowan and Rogers in [22]:
W
s
(z)
W
s
(z
r
)
=
_
z
z
r
_
α
where W
s
(z) is the wind speed at height z, W
s
(z
r
) is the reference wind speed
at height z
r
, and α is the power law exponent.
In practise, the exponent α is a highly variable quantity. It has been found that
α varies with such parameters as height above the sea level, time of day, season,
nature of terrain, wind speed, temperature, and various thermal and mechanical
mixing parameters. Some researchers have developed methods for calculating α
from the parameters of the log law. Many researchers, however, feel that these
complicated approximations reduce the simplicity and applicability of the general
power law and that wind energy specialists should accept the empirical nature of
the power law and choose values of α that best fit available wind data. For flat
and open areas α equals
1
7
, which is what expected for a flow over a flat plane.
In what follows the wind speeds have been transformed to correspond to a
height of 70m above ground level.
Definition Let W
s
1
, ..., W
s
T
be a vector of wind speed corresponding to a height
of 70m above ground level. T is a number of minutes in 12 years of one-minute
records. We assume 360 days in a year so that T = 6220800.
Let us test the wind speed data for periodicity, as this can interfere with the
15
probability distribution of the wind speed data later. There are two different
types of periodicity affecting original records. There is a daily pattern which
may be caused by the interaction between warm and cold air of the land and the
sea consequently (geographical feature of the United Kingdom) or due to heating
during the day over land (even when there is no sea involved). There is also
a seasonal pattern which affects a strength of the wind speed during different
months of the year. We use a multiplicative model to evaluate the seasonal and
daily effect in order to preserve a non-negative property of the data.
Definition Let {a
ν
}
ν=1,...,T
Υ
be a deseasonal scaling curve that carries the in-
formation on daily and seasonal patterns of the wind speed. T
Υ
is a number of
minutes in a year. We assume 360 days in a year so that T
Υ
= 518400.
For a convenience of further calculations every minute ν of a year can be
described by a δ day of the year and a τ minute of the day:
ν = 1440(δ −1) + τ,
where δ = 1, ..., 360 and τ = 1, ..., 1440. Similarly, every minute t of the 12-year
data can be described by a δ day of an Υ year and a τ minute of the day:
t = 518400(Υ−1) + 1440(δ −1) +τ,
where Υ = 1, ..., 12, δ = 1, ..., 360 and τ = 1, ..., 1440.
Then a
ν
is obtained by calculating a weighted average of the wind speeds over
all 12 years for a day δ − 10 to δ + 10 and for each of these days over minutes
τ −20 to τ + 20.
a
ν
= a
1440·δ+τ
=
12

Υ=1
10

d=−10
20

m=−20
k
dm
W
s
Υ,(δ+d),(τ+m)
(2.1)
where k
dm
is a weight of the wind speed W
s
Υ,(δ+d),(τ+m)
in minute (τ +m) of day
(δ +d) in year Υ. In (2.1) ν = 1, ..., T
Υ
is a minute of a year and an average value
a
ν
is assumed to be effected by the preceding and the following 20 minutes of the
preceding and the following 10 days to the current.
Weight k
dm
varies depending on the distance between the current minute (δ, τ)
and a minute effecting it (δ + d, τ + m). If plotted in terms of day d or minute
m, weights k
dm
form a shape of an isosceles triangle with the middle vertex at
the current time moment of τ minute of δ day. Every weight k
dm
can be defined
16
as follows:
k
dm
=
(10 −|d| + 1)(20 −|m| + 1)
12 ∗ (20 + 1)
2
∗ (10 + 1)
2
(2.2)
In (2.2) values 10 and 20 correspond to intervals of 10 days and 20 minutes
assumed to contribute to an average value a
ν
. Note, that all 12 years are given
the same weight in the average.
As a result a deseasonal scaling curve a
ν
can be constructed, each value of
which is a weighted average over 12 years of data.
Definition Let us denote W
d
1
, ..., W
d
T
as deseasonalised wind speed, or the wind
speed with removed daily and seasonal components. T again is a number of
minutes in 12 years.
To estimate deseasonalised wind speed W
d
t
we remove daily and seasonal effects
from the wind speed W
s
t
dividing the latter by a corresponding value of the
deseasonal scaling curve. For a convenience of calculation we assume a linear
connection between a minute of a year ν and a minute t of the 12-yearly data:
t = 518400(Υ−1) + 1440(δ −1) +τ ⇒ν = 1440(δ −1) + τ, ∀Υ
Definition Let us denote Θ
d
(W
s
t
, a
ν
) as a function transferring original wind
speed W
s
t
into deseasonalised wind speed W
d
t
for ∀t = 1, ..., T.
By defining a time position of a wind speed t with a minute τ of a day δ in year
Υ the deseasonilised value W
d
t
can be found as follows:
W
d
t
= Θ
d
(W
s
t
, a
ν
) = Θ
d
(W
s
Υδτ
, a
δτ
) =
W
s
Υδτ
a
δτ
where a
δτ
= a
ν
is the deseasonal scaling curve.
Notice, that the hypothesis about daily periodicity of wind speed is supported
by the autocorrelation of the original data. The series of autocorrelation coeffi-
cients measure the correlation, if any, between observations at different distances
apart and provide useful descriptive information [43]. Given T observations of
wind speed W
s
1
, . . . , W
s
T
, on a time series, for every integer l we can form T − l
pairs of observations, namely, (W
s
1
, W
s
1+l
), . . . , (W
s
t
, W
s
t+l
), . . . , (W
s
T−l
, W
s
T
), where
each pair of observations is separated by a time lag l. Regarding the first obser-
vation in each pair as one variable, and the second observation in each pair as
a second variable, then, we can measure the autocorrelation coefficient between
17
adjacent observations, W
s
and W
s
l
, using the formula:
corr(W
s
, W
s
l
) =

T−l
t=1
(W
s
t
−W
s
)(W
s
t+l
−W
s
)

T
t=1
(W
s
t
−W
s
)
2
·
T
T −l
(2.3)
where W
s
is the mean value of the series W
s
1
, . . . , W
s
T
.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 2000 4000 6000 8000 10000 12000
a
u
t
o
c
o
r
r
e
l
a
t
i
o
n
time, min
without seasonal component
original data
Figure 2.2: Autocorrelation of the original and the deseasonalised wind speed
data
Figure 2.2 plots the autocorrelation of one-minute wind speed recorded with
MST Radar located near Aberystwyth and the autocorrelation of the same desea-
sonalised data. The fluctuations of the autocorrelation function for the original
data repeat every 1440 minutes, which supports the presence of a daily periodic
component in the wind speed, however this pattern is not visible in the desea-
sonalised data.
2.2.2 Methods of generating missing data
After wind energy is incorporated into the the electricity generation system its
optimization modelling requires geographically distributed samples of one-minute
wind speed data. US Geological Survey provides sample of ten-minute wind speed
recorded over a large territory of Utah state. In this section we wish to describe
an algorithm that would allow us to sample one-minute wind speed from a known
distribution when an average value over ten minutes is given.
18
To ensure a generated set of data values reflects the properties of one-minute
wind speed distribution, one-minute records from MST Radar located near Aberys-
twyth are analysed. However, it is hard to describe this distribution analytically.
A survey of existing techniques on data sampling identified Monte Carlo Markov
Chain (MCMC) methods as a group of methods that can be applied to generate
one-minute wind speed when the distribution is not fully formulated. One of the
MCMC algorithms, Metropolis-Hastings produces a sequence of random samples
from a probability distribution, however it requires a joint probability distribu-
tion of one-minute wind speed which is hard to find. Meanwhile, it is possible
to find a conditional distribution of wind speed at every minute depending on
the past realisations and, hence, to use a special case of Metropolis-Hastings al-
gorithm - Gibbs sampling algorithm. This method ensures that as a number of
iterations of sampling from known conditional distributions increases, the density
of a resulting set of variables converges to the required one.
Gibbs sampling algorithm was developed during the last 30 years. In the
1980s two important papers were published that discussed a question of comput-
ing an estimate of a posterior distribution. Tanner and Wong (1987)[15] estimated
a distribution of the set of missing and observed values by data-augmentation.
Shortly before this, in 1984 Geman and Geman[13] offered a new restoration al-
gorithm for computing the maximum a posteriori estimate of the original image
given a degraded image. In 1990, Gelfand and Smith[14] showed a close rela-
tionship between the Gibbs sampler introduced by Geman and Geman (1984)
and the data-augmentation algorithm proposed by Tanner and Wong. In their
paper “Sampling-based approaches to calculating marginal densities”, Gelfand
and Smith gave a description of a sampling technique that we shall extend for
the generation of one-minute wind speed when the average value over each ten
minutes is known.
Glasbey et al. in ([16, 17, 18, 19]) applied Gibbs sampling to a problem of
sampling hourly rainfall when a daily value is given. It worked better for dry days
as Glasbey required repeated sampling until the desired average value appeared.
We wish to apply Gibbs sampling to the sites with the highest wind speed, hence,
the algorithm is originally adapted to preserve good samples and reduce their
number.
2.2.3 Description of Gibbs sampling
Let us formulate an algorithm of Gibbs sampling for a problem of generating T
variables correlated with the variables in a certain “neighbourhood” [14]. In a
19
relation to a collection of random vectors, Y
1
, Y
2
, . . . , Y
T
, suppose that for i =
1, . . . , T, the conditional distributions [Y
i
|Y
1
, . . . , Y
i−1
, Y
i+1
, . . . , Y
T
] are available
in reduced forms, i.e. depend only on {Y
j
}
j∈S
i
, where S
i
⊂ {1, . . . , T}:
[Y
i
|Y
1
, . . . , Y
i−1
, Y
i+1
, . . . , Y
T
] = [Y
i
|Y
j
; j ∈ S
i
⊂ {1 . . . T}], i = 1, . . . , T (2.4)
In (2.4), brackets denote densities, so joint, conditional, and marginal forms, for
example, can be written as [Y
1
, Y
2
], [Y
1
|Y
2
], and [Y ]. We denote by S
i
a certain
“neighbourhood” subset of {1, . . . , T}.
Gibbs sampler is an iterative Markov Chain Monte Carlo method. The algo-
rithm uses the following representation of the desired posterior density:
[Y
1
] =
_
[Y
1
|Y
T
, . . . , Y
2
] ∗ [Y
T
|Y
T−1
, . . . , Y
2
] ∗ . . . ∗ [Y
3
|Y
2
] ∗ [Y
2
] (2.5)
[Y
2
] =
_
[Y
2
|Y
1
, Y
T
, . . . , Y
3
] ∗ [Y
1
|Y
T
, . . . , Y
3
] ∗ . . . ∗ [Y
4
|Y
3
] ∗ [Y
3
] (2.6)
[Y
3
] =
_
[Y
3
|Y
2
, Y
1
, Y
T
, . . . , Y
4
] ∗ [Y
2
|Y
1
, Y
T
, . . . , Y
4
] ∗ . . . ∗ [Y
5
|Y
4
] ∗ [Y
4
] (2.7)
· · · (2.8)
[Y
T−1
] =
_
[Y
T−1
|Y
T−2
, . . . , Y
1
, Y
T
] ∗ [Y
T−2
|Y
T−3
, . . . , Y
1
, Y
T
] ∗ . . . ∗ [Y
1
|Y
T
] ∗ [Y
T
]
(2.9)
[Y
T
] =
_
[Y
T
|Y
T−1
, Y
T−2
. . . , T
1
] ∗ [Y
T−1
|Y
T−2
, . . . , Y
1
] ∗ . . . ∗ [Y
2
|Y
1
] ∗ [Y
1
] (2.10)
In (2.5) − (2.10), ∗ denotes a multiplication of densities, for example [Y
1
, Y
2
] =
[Y
1
|Y
2
] ∗[Y
2
]. The process of marginalisation (i.e. integration) is denoted by forms
such as [Y
1
|Y
2
] =
_
[Y
1
|Y
2
, Y
3
, Y
4
] ∗[Y
3
|Y
4
, Y
2
] ∗[Y
4
|Y
2
], with the convention that all
variables appearing in the integrand but not in the resulting density have been
integrated out. Therefore, for this example the integration is with respect to Y
3
and Y
4
[14].
Substitution of (2.10) into (2.9) and further up into (2.5) produces a marginal
distribution for Y
1
[14]:
[Y
1
] =
_
h(Y
1
,
ˆ
Y
1
) ∗ [
ˆ
Y
1
]
where h(Y
1
,
ˆ
Y
1
) is an integral that was resulted from the substitutions with
ˆ
Y
1

Y
1
. By similar substitutions, marginal distributions of all the random vectors
20
Y
1
, Y
2
, . . . , Y
T
are found:
[Y
i
] =
_
h(Y
i
,
ˆ
Y
i
) ∗ [
ˆ
Y
i
] (2.11)
The method of successive substitution for solving (2.11) suggests an itera-
tive method for the estimation of marginal distributions [Y
i
], i ∈ 1, ..., T. Its
implementation by Tanner and Wong[15] requires, however, the availability of
all conditional distributions on the right-hand side of (2.5) − (2.10), which are
not known in our case. The full conditional distributions uniquely determine the
joint distribution [Y
1
, . . . , Y
T
] that we wish to estimate by marginal distributions
using Gibbs sampling algorithm.
Figure 2.3: Schematic description of Gibbs sampler as a Markov-chain process
Gibbs sampling can be described as the following sequence of actions schemat-
ically presented in Figure 2.3.
Step 1: Given an arbitrary starting set of values Y
(j)
1
, . . . , Y
(j)
T
we draw Y
(j+1)
1
∼ [Y
1
|Y
(j)
2
, . . . , Y
(j)
T
]
Step 2: After the value of Y
(j+1)
1
is found
we draw Y
(j+1)
2
∼ [Y
2
|Y
(j+1)
1
, Y
(j)
3
, . . . , Y
(j)
T
]
· · · : · · ·
21
Step T: After the values of Y
(j+1)
1
, . . . , Y
(j+1)
T−1
are found
we draw Y
(j+1)
T
∼ [Y
T
|Y
(j+1)
1
, Y
(j+1)
2
, . . . , Y
(j+1)
T−1
]
Thus, each variable is visited in the natural order and a cycle requires T random
variate generations. After i such iterations, results a set of variables (Y
(i)
1
, . . . , Y
(i)
T
).
Under mild conditions, Geman and Geman showed that the following results
hold[13, 14]:
Convergence. [Y
(i)
1
, . . . , Y
(i)
T
] → [Y
1
, . . . , Y
T
] as i → ∞, and so [Y
(i)
j
] → [Y
j
] for
j ∈ 1, . . . , T. In fact, a slightly stronger result is proved. Rather than requiring
that each variable is visited in repetitions of the natural order, convergence still
follows under any visiting scheme, provided that each variable is visited infinitely
often.
Rate. Using the sup norm, rather than L
1
norm, the joint density of (Y
(i)
1
, . . . , Y
(i)
T
)
converges to the true joint density at a geometric rate in i, under visiting in the
natural order. A minor adjustment to the rate is required for an arbitrary “in-
finitely often” visiting scheme.
For i sufficiently large, the resulting set of variables (Y
(i)
1
, . . . , Y
(i)
T
) is a suffi-
ciently accurate sequence of wind speed values. It does not reflect, however, our
property that the average value of wind speed over each group of 10 one-minute
values should equal a given ten-minute measured value. Let us denote {Y
j
}
j=1,...,T
as a vector consisting of 10 scalars, or in other words Y
j
= (y
j,1
, . . . , y
j,10
), j =
1, . . . , T. Then, we can write an additional condition required in the probability
density function to reflect the fact that we require 10 sampled values of one-minute
wind speed having the measured ten-minute wind speed average:
Y
j
∼ [Y
j
|Y
1
, Y
2
, . . . , Y
j−1
, Y
j+1
, . . . , Y
T
, f(Y
j
) = W
s
j
] (2.12)
where f(Y
j
) is a function of 10 components of Y
j
and W
s
j
is the ten-minute
measured value.
Glasbey and Allcroft in [16] performed spatio-temporal rainfall disaggregation
using Gibbs sampler for a truncated Gaussian Markov random field model. In
order to satisfy (2.12) they suggested repeating sampling from the conditional
distributions (2.4) until the right average value is reached within a specified tol-
erance. Dry blocks were easier to match but approximately 0.1% of blocks that
described intense rainfall needed more than 1000 attempts. For the purpose of
electricity generation using wind speed turbines it is in the interest of produc-
ers to locate wind farms in windy places. Therefore, we would like to suggest a
modification to Gibbs sampling of one minute wind speed that reduces a number
22
of iterations. Later in subsection 2.2.5, a technique is described that shifts every
generated point towards the required average wind speed value in a way that
does not bias the sampling. This reduces a number of samples required by using
samples that initially do not satisfy the condition f(Y
j
) = W
s
j
.
2.2.4 Transformation to normal variable
This section presents transformation of the deseasonalised one-minute wind speed
data so that the marginal distribution is normal. If the stochastic process de-
scribing the deseasonalised wind speed was fully known, then in theory we could
generate the samples from this. However, there is not enough data to decide on
a unique stochastic process for the deseasonalised wind speed so we proceed as
follows.
1. We do a non-linear transformation of the deseasonalised wind speed so
that the distribution of wind speeds at any time is normal and so that
when the normalised wind speed falls below a threshold, the corresponding
deseasonalised wind speed is zero. In the normalised wind speed variable
the distribution of wind speed at any one time is truncated normal.
2. We assume that the stochastic process in the normalised variables is mul-
tivariate normal with the same autocorrelation function as the measured
data.
Points 1 and 2 make the generation of data computationally tractable.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 1 2 3 4 5 6 7
c
u
m
u
l
a
t
i
v
e

p
r
o
b
a
b
i
l
i
t
y

d
e
n
s
i
t
y

f
u
n
c
t
i
o
n
wind speed without periodical component, m/s
Figure 2.4: Cumulative probability distribution of one-minute wind speed
recorded with MST Radar located at Frongoch farm
23
The nature of wind speed is such that there can be no negative values. Further,
typical measurements show a big number of values zero. This is due to the fact
that the measurements are recorded in discrete quantities. These features are
depicted with the distribution of 1-minute wind speed plotted in the Figure 2.4.
As a multiplicative model was applied to remove seasonal effect from the data,
the non-negative property of the original data is still present. Since 8% of the
wind speeds are zero we require:
P(W
d
≡ 0) = 0.08 (2.13)
where W
d
is a deseasonalised wind speed.
Definition Let us denote Y
1
, . . . , Y
T
as normalised wind speed, or the deseason-
alised wind speed transfered to the normal variable. T again is a number of
minutes in 12 years.
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
-8 -6 -4 -2 0 2 4 6 8
p
r
o
b
a
b
i
l
i
t
y

d
e
n
s
i
t
y

f
u
n
c
t
i
o
n
normal variable, y
truncated normal pdf
standard normal pdf
Figure 2.5: Truncated and full plot of normal probability density function
Let us consider standard normal distribution N(0, 1) with mean equals 0 and
standard deviation 1. Eight percent of wind speed data equals zero, hence we wish
to transform the standard normal distribution so that it is truncated from the
left (Figure 2.5). Single truncation means that certain outcomes are constrained,
in our case these are the wind speeds recorded at zero.
The corresponding lowest value of the normal variable can be found from the
equation:
P(Y ≤ ˆ y) ≡ P(W
d
≡ 0) = 0.08 (2.14)
24
From (2.14) we have ˆ y equals −1.1225.
There is a large body of literature on the subject of estimation of the pa-
rameters of the original data (µ, σ) based upon data from truncated samples.
For our work Schneider provides an excellent overview of parameter estimation
of truncated normal distributions in his detailed work “Truncated and Censored
samples from Normal Populations” [44]. Johnson and Thomopoulos applied most
of the analysis to a special case of left truncated normal distribution [45].
Definition Let us denote U
1
, . . . , U
T
as standard normal random variable with
mean equals 0 and standard deviation 1.
The probability density function (pdf) of U is given by
φ(U) =
1


e

U
2
2
−∞≤ U ≤ ∞
Let Φ(U) denote the cumulative distribution function (cdf) of U. The cdf of a
normal random variable Y with mean µ and variance σ
2
is Φ(
Y −µ
σ
). The random
variables Y and U are linearly related and the relationship is given by U =
Y −µ
σ
.
A random variable Y has a single truncation from the left normal distribution if
its probability density function is:
f(Y ; µ, σ, ˆ y) =
_
_
_
φ(
Y −µ
σ
)
σ(1−Φ(ˆ y))
ˆ y ≤
Y −µ
σ
0 otherwise
where ˆ y = −1.1225 is left (lower) truncation point and a degree of truncation is
Φ(ˆ y) = 0.08 for the considered dataset.
Let us define θ for a singly truncated from the left normal distribution:
θ =
φ(ˆ y)
1 −Φ(ˆ y)
Then the expected value and variance of Y can be obtained in the following way:
E(Y ) = µ + σθ (2.15)
V (Y ) = σ
2
(1 −θ
2
+ θˆ y) (2.16)
Every variable from the truncated normal distribution described by (2.15) and
(2.16) corresponds to a single value of the deseasonalised one-minute wind speed.
Y (W
d
) : P(Y < ¯ y) ≡ P(W
d
< ¯ w) (2.17)
25
where ¯ y and ¯ w are the corresponding values of the normal variable and the de-
seasonalised wind speed respectively.
Wind speed is normalised according to the cumulative density function of the
left truncated normal distribution. If we write now a normalised wind speed in
discrete form with a step equal to 0.01, it has a “one-to-one” relation with the
deseasonalised wind speed. Hence, it is possible to look for an analytic fit that
can simplify further calculations.
Figure 2.6 plots the original empirical relation between normalised and desea-
sonalised wind speed. It appears a simple function can be used as an analytic
monotonic fit to the transformation function between the normalised and desea-
sonalised wind speed (Monotonicity is a necessary condition for this transforma-
tion. If transformation function were not monotonic there would be a possibility
of multiple solutions, this would prevent the method working). In ([16, 17, 18, 19])
Glasbey et al. suggest using a quadratic as the function to transform the rainfall
to normal.
-2
-1
0
1
2
3
4
5
0 1 2 3 4 5 6 7
n
o
r
m
a
l

v
a
r
i
a
b
l
e
wind speed without seasonal component, m/s
original plot
-0.0854*x*x+1.3652*x-1.1225
Figure 2.6: Normal variable plot with respect to the deseasonalised wind speed
Definition Let us denote Θ
n
(W
d
t
) as a function transferring deseasonalised wind
speed W
d
t
into normalised wind speed Y
t
for ∀t = 1, . . . , T.
We wish to find an analytical fit to Θ
n
(W
d
t
) so that during sampling it is easy
to calculate its inverse W
d
t
= Θ
−1
n
(Y
t
). We assumed Θ
n
(W
d
t
) to be a logarith-
mic or quadratic function from a plot on figure 2.6 and tested a set of points
{(W
d
t
, Y
t
)}
t=1,...,T
for a best fit. During the test quadratic function gave a better
26
result, therefore, we write a function Θ
n
(W
d
t
) as a full square:
Y
t
= Θ
n
(W
d
t
) = b
1
(W
d
t
+ b
2
)
2
+ b
3
, ∀t = 1, . . . , T (2.18)
where W
d
t
denotes the deseasonalised one-minute wind speed and Y
t
denotes the
normalised wind speed, i.e. a variable from the normal distribution truncated to
the left. Using a quadratic function formulated as a full square it is straightfor-
ward to find the inverse to Θ
n
(W
d
t
):
W
d
t
= Θ
−1
n
(Y
t
) = ±
_
Y
t
−b
3
b
1
−b
2
, ∀t = 1, . . . , T (2.19)
Parameters of the equation (2.18), b
1
, b
2
and b
3
can be estimated by solving si-
multaneously three linear equations for the points from the original plot (W
d
t
, Y
t
),
giving b = (−0.0854, −7.993, 4.3335). With these parameter values, the function
is monotonic along all the period [0 : 6] (m/s) with a range for the normalised
wind speed [−1.1225 : 4.3335] and probability of the deseasonalised wind speed
exceeding 6m/s at a single site is only 1.6∗10
−5
, which is small enough to ignore.
Plus or minus in front of the square root in (2.19) is chosen so that a value of
the deaseasonalised wind speed W
d
t
is in a range [0 : 7.993] (m/s). The estimated
transformation function is plotted on Figure 2.6 along with the original plot.
2.2.5 Suggested modifications to Gibbs sampler
Let us now reformulate the Gibbs sampling algorithm for the normalised wind
speed. On page 19 of the current work we introduced Y
t
as a vector containing
10 normalised wind speeds or Y
t
= (y
t,1
, . . . , y
t,10
), t = 1, . . . , T. Then f(Y
t
)
represents the average of the wind speeds corresponding to the components of Y
t
,
i.e.
f(Y
t
) =
1
10
10

j=1
Θ
−1
d

−1
n
(y
t,j
)),
where y
t,j
is normalised wind speed that is transfered into one-minute wind speed
with inverse functions to Θ
n
and Θ
d
defined on page 23 and page 14 accordingly.
We wish to sample a collection of random normalised wind speed variables,
Y
1
, Y
2
, . . . , Y
T
from the available conditional distributions:
[Y
t
∼ [Y
t
|Y
1
, Y
2
, . . . , Y
t−1
, Y
t+1
, . . . , Y
T
, f(Y
t
) = W
s
t
], t = 1, . . . , T
To achieve this, first Y
t
is sampled from the distribution [Y
t
|Y
1
, Y
2
, . . . , Y
T
] and
27
then shifted to
˜
Y
t
in the space to satisfy the equation f(
˜
Y
t
) = W
s
t
.
Assume a vector Y
t
is generated. We wish to preserve a chance of this vector
to appear. So Y
t
is shifted along a vector of maximum probabilities for the desired
normal probability distribution function. Let us consider a matrix form of the
probability density function for the normalised wind speed.
φ(Y ) =
1


e

1
2
(Y
T
ϕ
−1
Y −2Y
T
ϕ
−1
µ−µ
T
ϕ
−1
µ)
(2.20)
where Y is the normalised wind speed with the covariance matrix ϕ and the mean
µ.
Then a point Y which maximizes φ(Y ) subject to satisfying the constraint
f(Y ) = W
s
can be found by solving an optimization problem:
arg max
_
1


e

1
2
(Y
T
ϕ
−1
Y −2Y
T
ϕ
−1
µ−µ
T
ϕ
−1
µ)
_
subject to:
f(Y ) = W
s
which occur at the same
arg min
_
1
2
(Y
T
ϕ
−1
Y −2Y
T
ϕ
−1
µ −µ
T
ϕ
−1
µ)
_
subject to:
f(Y ) = W
s
Write Lagrangian for this minimization problem to solve it:
L(Y, λ) =
1
2
(Y
T
ϕ
−1
Y −2Y
T
ϕ
−1
µ −µ
T
ϕ
−1
µ) + λ(f(Y ) −W
s
) (2.21)
Function f(Y ) in (2.21) is complex, every normalised wind speed is transformed
to one-minute wind speed using two inverse functions Θ
−1
n
and Θ
−1
d
. In order
to work in the space of the normalised wind speed we assume further that the
equation f(Y ) = W
s
can be replaced with its linear approximation 1
T
Y = Y ,
where 1 is a vector of the same size as vector Y and consists of scalars 1 only.
Value of Y is found so that 1
T
Y is situated closer to the center of the distribution
28
[Y ]:
If 1
T
Y ≤ 1
T
µ, 1
T
Y is the tangent to f(Y ) −W
s
at the point of
the shortest distance between f(Y ) −W
s
and µ, and
If 1
T
Y > 1
T
µ, we assume Y = b
3
,
where µ is mean value of the distribution [Y ] and b
3
is a parameter in a quadratic
function at (2.18).
When using a linear approximation of the equation f(Y ) = W
s
, the resulting
Lagrangian function obtained from (2.21) is easily differentiated with respect to
Y :

Y
L =
∂L
∂Y
=
∂[
1
2
(Y
T
ϕ
−1
Y −2Y
T
ϕ
−1
µ −µ
T
ϕ
−1
µ) + λ(1
T
Y −Y )]
∂Y
= ϕ
−1
Y −ϕ
−1
µ + λ1
T
First-order necessary condition for Y to be a minimum for the above optimization
problem stipulates that ∇
Y
L = 0. From here we can find an equation for Y :
ϕ
−1
(Y −µ) = −λ1
T
(2.22)
Y = −λ1
T
ϕ + µ (2.23)
In (2.23) ϕ and µ denote the covariance matrix and the mean vector, they de-
scribe the probability distribution of the normalised wind speed. The scalar λ
is a Lagrangian constant and defines a position of Y on a vector of maximum
probabilities. Parameter λ is found so that Y in (2.23) satisfies a linear equation
1
T
Y = Y .
In the context of the Gibbs sampling, a generated vector Y
t
is shifted along
the vector described by (2.23). This means that
˜
Y
t
acquires a value from the
vector
˜
Y
t
= −λ1
T
ϕ + Y
t
(2.24)
where Y
t
is the original generated variable and
˜
Y
t
is a shifted normalised value of
wind speed so that it satisfies now the condition 1
T
˜
Y
t
= Y .
A shifted variable
˜
Y
t
is found so that it satisfies f(Y ) = W
s
within a certain
tolerance as the equation f(Y ) = W
s
is close to its linear approximation 1
T
Y = Y
which
˜
Y
t
satisfies exactly. Assume, there exists a vector Y

t
such that it satisfies
both equations exactly, f(Y

t
) = W
s
and Y

t
= −λ1
T
ϕ+Y
t
. Then
˜
Y
t
is accepted
as a valid normalised wind speed if the probability of
˜
Y
t
is within a specified
29
tolerance of the probability of Y

t
. The values of the one-minute wind speed W
s
t
in this case are found using W
s
t
= Θ
d

n
(
˜
Y
t
)).
2.3 Implementation of modified Gibbs sampler
We wish to generate an unbiased sample from the wind stochastic process con-
ditional on each consecutive group of wind speeds having a given average value.
Assume the group size is q and we wish to generate T · q values:
w
s
1
, . . . , w
s
q
. ¸¸ .
, w
s
q+1
, . . . , w
s
2q
. ¸¸ .
, . . . , w
s
(T−1)q+1
, . . . , w
s
Tq
. ¸¸ .
,
t = 1 t = 2 . . . t = T
such that
1
q
(w
s
(t−1)q+1
+ w
s
(t−1)q+2
+ . . . + w
s
tq
) = W
s
t
where w
s
i
, i = 1, . . . , Tq are the one-minute wind speeds and W
s
t
is the measured
average wind speed for times i = (t − 1)q + 1, . . . , tq. For the real problems we
are dealing with q = 10, but we will also illustrate the case q = 2.
Instead of operating with wind speed variables w
s
i
directly, we work with the
transformed stochastic process when the periodicity (seasonal and daily effects)
has been removed so that their marginal distribution is a truncated normal. We
will do this in two steps, first the removal of periodic effects by transforming to
new variables w
d
i
= Θ
d
(w
s
i
) followed by a transformation to variables y
i
= Θ
n
(w
d
i
)
which have a truncated normal distribution.
Definition Let us denote Θ as a compound function, such that
y
i
= Θ
n

d
(w
s
i
)) := Θ(w
s
i
)
and since the transformation will be “one-to-one”,
w
s
i
= Θ
−1
d

−1
n
(y
i
)) := Θ
−1
(y
i
)
Transformation functions Θ
d
and Θ
n
were described earlier in this Chapter.
Section 2.2.1 dealt with daily and seasonal periodical components of the original
wind speed data. Section 2.2.4 discussed transformation to the normalised wind
speed. Given these functions are known and given µ and ϕ as described in Section
2.2.4, the algorithm of generating one-minute wind speed values is the following:
Step 1: Start with t = 1.
30
Step 2: Generate normalised wind speed Y
t
= (y
(t−1)q+1
, . . . , y
tq
)
from N
q
(µ, ϕ).
Step 3: Find w
s
i
= Θ
−1
(y
i
), i = [(t −1)q + 1 . . . tq].
Step 4: If
1
q
(w
s
(t−1)q+1
+ w
s
(t−1)q+2
+ . . . + w
s
tq
) = W
s
t
within a specified tolerance
Go to Step 8
Step 5: Shift (y
(t−1)q+1
, . . . , y
tq
) along the vector of maximum prob-
abilities to find (˜ y
(t−1)q+1
, . . . , ˜ y
tq
)
Step 6: If a probability of (˜ y
(t−1)q+1
, . . . , ˜ y
tq
) is within a specified tol-
erance of a probability of (y

(t−1)q+1
, . . . , y

tq
), which if trans-
ferred to the wind speed satisfies
1
q
(w
s
(t−1)q+1
+ w
s
(t−1)q+2
+ . . . + w
s
tq
) = W
s
t
exactly,
Go to Step 8
Step 7: Go to Step 2
Step 8: t = t + 1
Step 9: If t > T Stop
else Go to Step 2
This algorithm describes a single iteration of the Gibbs sampler. To generate
a sample with the required probability density function we wish to perform a
significantly big number of such iterations.
2.3.1 Multivariate normal distribution
Section 2.2 of this Chapter introduced the Gibbs sampling technique for a collec-
tion of normalised wind speed variables Y
1
, Y
2
, . . . , Y
T
. Every generated variable
Y
t
= (y
t,1
, . . . , y
t,10
), t = 1, . . . , T has to satisfy f(Y
t
) = W
s
t
, which is ten wind
speed values have to add up to the measured average ten-minute wind speed. The
fact that Y
t
includes 10 univariate variables defines a vector Y
t
as a multivariate
variable.
A stochastic process can be described with the moments of the process, par-
ticularly the first and second moments, mean and autocovariance function re-
spectively. Let us find mean and autocovariance of the univariate one-minute
normalised wind speed that can be used further to define a stochastic process for
multivariate normalised wind speed. First moment of the stochastic process of
31
wind speed generation is described earlier as the expected value of the truncated
normal distribution. It can be found using the formula (2.15).
Given Tq variables y
1
, . . . , y
Tq
, autocovariance coefficients can be found:
cov(y, y
l
) =
1
Tq −l
Tq−l

i=1
(y
i
−y)(y
i+l
−y) (2.25)
where l denotes a lag between the variables and y is the mean value of the series.
Figure 2.7 plots the autocovariance function in terms of a lag l calculated
for the normalised wind speed. We assume a lag up to 1440 minutes (1 day) is
significant for the one minute wind speed stochastic process.
In order to use the autocovariance in the process of wind speed generation
it is convenient to find a continuous function that would fit in the original plot.
We assumed cov(y, y
l
) to be a quadratic or exponential function from a plot on
figure 2.7 and tested a set of points {(l, cov(y, y
l
))}
l=1,...,L
for a best fit. During
the test the exponential function gave a better result, therefore, we estimate the
autocovariance function with a mixture of exponential curves:
cov(y, y
l
) ≡
m

i=1
β
i
e
−ς
i
|l|
m

i=1
β
i
= 1 (2.26)
According to Allcroft and Glasbey [16] a condition of all coefficients β
i
being
positive is sufficient for the function to be valid.
0
0.1
0.2
0.3
0.4
0.5
0.6
0 200 400 600 800 1000 1200 1400
a
u
t
o
c
o
v
a
r
i
a
n
c
e
lag, min
original normalised wind speed
fitted exponential function
Figure 2.7: Fitted sum of exponential functions to approximate the autocovari-
ance of the normalised wind
32
Figure 2.7 shows that the autocovariance of the normalised wind speed is
influenced by certain processes that effect its smoothness, for example, for l >
1000 a gradient to the curve changes and it is hard to find a simple function to fit
it perfectly. We found a good fit to the autocovariance function to be a mixture
of six exponentials with the following parameters:
β = (0.04, 0.026, 0.19, 0.204, 0.25, 0.29)
and
ς = (1.5, 0.13, 0.0023, 0.0018, 0.0005, 0.0002)
Figure 2.7 plots the original autocovariance for the normalised wind speed along
with the fitted exponential function.
Calculated autocovariance and mean describe a stochastic process for the
univariate normalised wind speed y
1
, . . . , y
Tq
. It can be used now to formulate a
multivariate stochastic process for the series Y
1
, . . . , Y
T
.
For the convenience of further calculations a covariance can be described in a
matrix form:
ϕ =
_
_
_
_
_
_
_
_
ϕ
11
· · · ϕ
1j
· · · ϕ
1L
· · · · · · · · · · · · · · ·
ϕ
i1
· · · ϕ
ij
· · · ϕ
iL
· · · · · · · · · · · · · · ·
ϕ
L1
· · · ϕ
Lj
· · · ϕ
LL
_
_
_
_
_
_
_
_
where L denotes a biggest lag of the autocovariance. Covariance matrix ϕ is
found based on the autocovariance function described by (2.26):
ϕ
ji
= ϕ
ij
= cov(y, y
i−j
), i ≥ j (2.27)
For a collection of multivariate normalised wind speeds Y
1
, . . . , Y
T
every vector
Y
t
follows the distribution N
q
(µ, ϕ). For the case of generating ten minutes of wind
speed, we take q = 10 << L. To find mean and covariance of a q-dimensional
variable let us partition an interval of length L. Suppose that Y ∼ N
L
(µ, ϕ),
partitioned into q and (L −q) components as
Y =
_
Y
q
Y
L−q
_
In our case Y
q
, of dimension q, is a set of ten normalised wind speeds for which we
wish to find a probability distribution. Variable Y
L−q
, of dimension (L −q), is a
set of variables that first q variables correlate with. Partition µ and ϕ analogously
33
as
µ =
_
µ
q
µ
L−q
_
and
ϕ =
_
ϕ
qq
ϕ
q(L−q)
ϕ
(L−q)q
ϕ
(L−q)(L−q)
_
with dimensions of the subvectors and submatrices as induced by the partition
of Y . Then Y
q
∼ N
q

q
, ϕ
qq
) and Y
L−q
∼ N
L−q

L−q
, ϕ
(L−q)(L−q)
).
Assume that ϕ
(L−q)(L−q)
is positive definite. Then the conditional distribution
of Y
q
given Y
L−q
= Y

L−q
, is q-variate normal with the following parameters:
µ
q|(L−q)
:= E[Y
q
|Y
L−q
= Y

L−q
] = µ
q
+ ϕ
q(L−q)
ϕ
−1
(L−q)(L−q)
(y
L−q
−µ
L−q
) (2.28)
and
ϕ
qq|(L−q)
:= Cov[Y
q
|Y
L−q
= Y

L−q
] = ϕ
qq
−ϕ
q(L−q)
ϕ
−1
(L−q)(L−q)
ϕ
(L−q)q
(2.29)
Mean and covariance stated in (2.28) and (2.29) represent conditional normal
probability distribution that describes the stochastic process of ten-minute wind
speed generation.
For any vector µ ∈ R
q
and positive semidefinite symmetric matrix ϕ of di-
mension (q ×q), there exists a unique multivariate normal distribution with mean
vector µ and covariance matrix ϕ [46] .
Further, to generate a multivariate normal variable with mean µ and covari-
ance ϕ we sample first a multivariate vector U ∼ N
q
(0, I
q
). For any vector µ ∈ R
q
and positive semidefinite, symmetric matrix ϕ of dimension (q×q), random vector
Y with a N
q
(µ, ϕ)-distribution can be generated as
Y = µ + MU (2.30)
where U ∼ N
q
(0, I
q
), and MM
T
= ϕ.
If ϕ is a positive semidefinite symmetric matrix of dimension q ×q, then there
exists a matrix M of dimension (q ×q) such that
ϕ = MM
T
, (2.31)
where M
T
is the transposition of matrix M. Matrix M can be found precisely
using different methods, for example Cholesky decomposition, since in our case
covariance matrix ϕ is a positive definite matrix. More specifically, M can be
34
computed as follows:
M :=
_
_
_
_
_
_
M
11
0 . . . 0
M
21
M
22
. . . 0
. . . . . . . . . . . .
M
q1
M
q2
. . . M
qq
_
_
_
_
_
_
where every element M
ij
, i ≥ j can be found with a sequence of calculations:
Step 1: M
11
=

φ
11
Step 2: M
i1
=
φ
i1
M
11
, i ≤ q
Step 3: Set j = 2
Step 4: M
jj
=
_
φ
jj

j−1
k=1
M
2
jk
Step 5: M
ij
=
φ
ij

j−1
k=1
M
ik
M
jk
M
jj
, j < i ≤ q
Step 6: Set j = j + 1
Step 7: If j ≤ q Go to Step 4
Else Finish algorithm
Matrix M is used now to generate a multivariate normal variable Y of the Gibbs
sampling algorithm.
2.3.2 Two-variable sampling
We wish to generate data for each minute t where each consecutive group of ten
variables has a given average value. To illustrate the method we will first derive a
case where the variables are divided into groups of two. Therefore, on the interval
{0, 1, . . . , 2T} we wish to generate points w
s
i
subject to a given average of two
adjacent values of wind speed W
s
t
:
w
s
i
+ w
s
i+1
2
= W
s
t
, i = 0, 1, . . . , 2T, t = 0, 1, . . . , T
This can be achieved through a sequence of steps described below in this section
and illustrated in Figure 2.8. Let us start with normalising a given average wind
speed by transforming it to the truncated normal variable:
Y
(0)
t
:= (y
(0)
i
, y
(0)
i+1
) := (Θ(W
s
t
), Θ(W
s
t
))
35
This gives us an initial set for Gibbs sampling.
The repeated subproblem when using Gibbs sampling algorithm to generate
samples with given means, is to sample from one group of variables conditional on
achieving a given average wind speed. Here are the main steps of j-th iteration:
1. Assume we wish to generate a sample Y
(j)
t
= (y
(j)
i
, y
(j)
i+1
). In the normalised
variables this will be a bivariate normal Y
(j)
t
∼ N
2
(µ, ϕ), where ϕ is the
same for all samples and is given in general by, for our example,
ϕ =
_
0.8 0.6
0.6 0.7
_
and µ is calculated using the neighbouring sample values
(Y
(j)
1
, Y
(j)
2
, . . . , Y
(j)
t−1
, Y
(j−1)
t+1
, . . . , Y
(j−1)
T
).
Assume, a point A := Y
∗(j)
t
:= (y
∗(j)
i
, y
∗(j)
i+1
) = (1.2595, −0.3102) is generated
(Figure 2.8) using N
2
(µ, ϕ).
Figure 2.8: Contours of Normal distribution and four linear functions used in the
modification to Gibbs algorithm
2. The generated normalised wind speeds are transformed further using inverse
36
function to Θ:
w
∗s
i
= Θ
−1
(y
∗(j)
i
) (2.32)
w
∗s
i+1
= Θ
−1
(y
∗(j)
i+1
) (2.33)
Calculated with (2.32) and (2.33) new wind speed values are W
s(j)
t
:=
(w
∗s
i
, w
∗s
i+1
) with their average equals
W
s(j)
t
:=
w
∗s
i
+ w
∗s
i+1
2
where W
s(j)
t
is the average wind speed corresponding to the generated values
(y
∗(j)
i
, y
∗(j)
i+1
).
The generated point (y
∗(j)
i
, y
∗(j)
i+1
) is tested. If W
s(j)
t
equals W
s
t
within a
specified tolerance, the point (w
∗s
i
, w
∗s
i+1
) is accepted, otherwise (y
∗(j)
i
, y
∗(j)
i+1
)
is shifted in order to satisfy W
s
t
as explained in steps 3 and 4 of this algo-
rithm.
3. We assumed λ to vary in the interval between −10 and 10, that are suffi-
ciently large to find Y in (2.23). By advancing λ from one border of the
interval to another with a step 0.001, we search along the line
Y
(j)
t
= −λ1
T
ϕ + Y
∗(j)
t
, (2.34)
where Y
∗(j)
t
is the point A sampled in step 1 of this algorithm, ϕ is a
covariance matrix of distribution [Y
t
] and 1 is a vector of the same size as
Y
t
and consists of scalars 1 only. Points Y
C
and Y
D
(C and D in Figure
2.8) are found during the search.
Definition Let us denote Y
B
(B in Figure 2.8) as the point on the curve
f(Y
t
) = W
s
t
that acquires the shortest distance to the center of the distri-
bution [Y ].
Definition Let us denote Y
C
as the intersection of the line (2.34) with the
tangent to the curve f(Y
t
) = W
s
t
at the point Y
B
, i.e. a line 1
T
Y
t
= Y
defined in the Section 2.2.5 of this work. A scalar Y is calculated as in the
Section 2.2.5:
If 1
T
Y
B
≤ 1
T
µ, Y = 1
T
Y
B
, and
if 1
T
Y
B
> 1
T
µ, we assume Y = b
3
,
37
where b
3
is a parameter of the quadratic function in (2.18). By modifying
the value of λ, Y
C
(and a corresponding λ
C
) are found so that 1
T
Y
C
and
Y match to small given tolerance.
Definition Let us denote Y
D
as the point on the line (2.34) that satisfies
the condition of two adjacent one-minute wind speeds being equal a given
average value f(Y
D
) = W
s
t
. By modifying the value of λ, Y
D
(and a
corresponding λ
D
) are found so that f(Y
D
) and W
s
t
match to small given
tolerance.
4. The probabilities of the points Y
D
and Y
C
are further compared:
If
φ(Y
D
)
φ(Y
C
)
> ρ accept Y
(j)
t
= Y
C
Else reject Y
C
and Go to step 1,
where φ(∗) is the probability distribution function for truncated normal
distribution used for sampling normalised wind speed. We assumed ρ to
be a fixed number 0.95. Glabey et al. in [17] suggested to take ρ as a
random number from the uniform [0, 1] distribution in order to add random
character to the test.
5. Take i = i + 1. If i ≤ T Go to step 1, Else Finish j-th iteration of Gibbs
sampling algorithm.
Steps 1 to 5 are repeated until j reaches a specified number of iterations j

.
We determine j

so that a stochastic process of the resulting sample is weakly
stationary.
2.3.3 Multivariate sampling
The two-variable algorithm was described in order to illustrate wind speed gen-
eration when the average value of adjacent values is given. We wish to apply the
demonstrated method to the sampling of ten minutes of wind speed when their
average value is given.
In [16], Glasbey et al. generate hourly rainfall with a given daily value by
repeated sampling from the multivariate normal distribution until the data total
fell within a certain margin of the target. To sample hourly rainfall values it
required an average of 170 simulations per day (sometimes considerably larger
38
numbers for wet days) for one iteration of the Gibbs sampling. This means that
many samples were rejected.
On the interval [0, 1, . . . , qT] we wish to generate wind speeds w
s
i
subject to
the measured ten-minute average wind speed W
s
t
:
1
q
(w
s
i
+ w
s
i+1
+ . . . + w
s
i+q−1
) = W
s
t
,
where q = 10 and t = 1, . . . , T. This can be achieved through a sequence of
steps similar to that described for two-variable sampling in Section 2.3.2. The
algorithm starts with normalising a given average wind speed by transforming it
to the truncated normal variable:
Y
(0)
t
:= (y
(0)
i
, y
(0)
i+1
, . . . , y
(0)
i+q−1
) := (Θ(W
s
t
), Θ(W
s
t
), . . . , Θ(W
s
t
))
This gives us an initial set of variables for Gibbs sampling. The repeated sub-
problem when using Gibbs sampling to generating samples with given means, is
to sample from one group of variables conditional on achieving a given average
wind speed. Here are the main steps of j-th iteration:
1. We wish to generate a multivariate Y
(j)
t
∼ N
q
(µ, φ), where
µ =
_
_
_
_
µ
1
.
.
.
µ
q
_
_
_
_
and
ϕ =
_
_
_
_
_
_
_
_
ϕ
11
. . . ϕ
1j
. . . ϕ
1q
. . . . . . . . . . . . . . .
ϕ
i1
. . . ϕ
ij
. . . ϕ
iq
. . . . . . . . . . . . . . .
ϕ
q1
. . . ϕ
qj
. . . ϕ
qq
_
_
_
_
_
_
_
_
are given (see below). We are interested in generating ten minutes of wind
speed at any single time, hence q = 10.
A covariance matrix for a set of one-minute wind speed data recorded with
MST Radar near Aberystwyth is first calculated for a lag L = 1440 minutes.
For sampling of a multivariate variable of size q = 10 we use a covariance
matrix of size (10 × 10) and mean of size (10 × 1). Both can be found as
a conditional covariance ϕ
qq|(L−q)
and a conditional mean µ
q|(L−q)
as was
shown in section 2.3.1. The calculated conditional covariance ϕ does not
39
depend on the values of adjacent normalised wind speeds and has the same
values for any sample.
Table 2.1: Covariance matrix ϕ calculated for the normal variables transformed
from the minute-by-minute wind speed data
0.076 0.031 0.017 0.011 0.009 0.008 0.007 0.006 0.005 0.003
0.031 0.089 0.038 0.021 0.015 0.012 0.010 0.009 0.007 0.005
0.017 0.038 0.092 0.040 0.023 0.016 0.013 0.011 0.009 0.006
0.011 0.021 0.040 0.094 0.041 0.023 0.016 0.013 0.010 0.007
0.009 0.015 0.023 0.041 0.094 0.041 0.023 0.016 0.012 0.008
0.008 0.012 0.016 0.023 0.041 0.094 0.041 0.023 0.015 0.009
0.007 0.010 0.013 0.016 0.023 0.041 0.094 0.040 0.021 0.011
0.006 0.009 0.011 0.013 0.016 0.023 0.040 0.092 0.038 0.017
0.005 0.007 0.009 0.010 0.012 0.015 0.021 0.038 0.089 0.031
0.003 0.005 0.006 0.007 0.008 0.009 0.011 0.017 0.031 0.076
Conditional covariance for our illustrative example is presented in the Table
2.1. The calculated conditional mean, however, is defined by the normalised
wind speed values of the partition (L−q). It is calculated for every sample
separately.
Assume a point U
∗(j)
t
= (u
∗(j)
i
, . . . , u
∗(j)
i+q−1
) ∼ N
q
(0, I
q
) is generated. It is
then transferred to the required conditional normal probability distribution
N
q
(µ, ϕ).
Y
∗(j)
t
= µ + MU
∗(j)
t
where µ is a conditional mean and a conditional covariance matrix ϕ =
MM
T
. Matrix M can be found using the algorithm demonstrated in Section
2.3.1. For the covariance matrix ϕ presented in Table 2.1, matrix M can
be found as in Table 2.2.
Table 2.2: M-matrix such that MM
T
= ϕ
0.276 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
0.113 0.276 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
0.061 0.112 0.276 0.000 0.000 0.000 0.000 0.000 0.000 0.000
0.041 0.060 0.111 0.276 0.000 0.000 0.000 0.000 0.000 0.000
0.033 0.040 0.059 0.111 0.275 0.000 0.000 0.000 0.000 0.000
0.028 0.032 0.039 0.057 0.110 0.275 0.000 0.000 0.000 0.000
0.025 0.027 0.030 0.038 0.056 0.108 0.274 0.000 0.000 0.000
0.022 0.023 0.025 0.028 0.036 0.054 0.106 0.273 0.000 0.000
0.018 0.019 0.020 0.022 0.025 0.032 0.050 0.102 0.270 0.000
0.012 0.013 0.014 0.015 0.016 0.019 0.025 0.041 0.088 0.255
40
2. Generated normalised wind speed is transformed further using the inverse
of function Θ:
W
s(j)
t
= Θ
−1
(Y
∗(j)
t
) (2.35)
Calculated with (2.35) new values of wind speed are W
s(j)
t
:= (w
∗s
i
, . . . , w
∗s
i+q−1
)
with ten-minute average wind speed equals
W
s(j)
t
=
1
q
(w
∗s
i
+ . . . + w
∗s
i+q−1
)
where W
s(j)
t
is an average wind speed corresponding to the generated earlier
values of the normalised wind speed (y
∗(j)
i
, . . . , y
∗(j)
i+q−1
).
The generated point Y
∗(j)
t
:= (y
∗(j)
i
, . . . , y
∗(j)
i+q−1
) is tested. If W
s(j)
t
matches
W
s
t
within a specified tolerance, the point (w
∗s
i
, . . . , w
∗s
i+q−1
) is accepted.
Otherwise (y
∗(j)
i
, . . . , y
∗(j)
i+q−1
) is shifted in order to satisfy W
s
t
as explained
in steps 3 and 4 of this algorithm.
3. As in the case of two variables a shifted point (y
∗(j)
i
, . . . , y
∗(j)
i+q−1
) acquires
a value from the vector of maximum probabilities of the truncated normal
probability distribution, so the following applies. We assumed again λ to
vary in the interval between −10 and 10, that are sufficiently large to find
Y in (2.23). By advancing λ from one border of the interval to another with
a step 0.001, we search along the line
Y
(j)
t
= −λ1
T
ϕ + Y
∗(j)
t
, (2.36)
where Y
∗(j)
t
is a point sampled in step 1 of this algorithm, ϕ is a covariance
matrix of distribution [Y
t
] and 1 is a vector of the same size as Y
t
and
consists of scalars 1 only. Points Y
C
and Y
D
are found during the search.
Direction in which point Y
∗(j)
t
is shifted is defined by the covariance matrix
ϕ and evaluates how much information each of ten variables contributes
to the average wind speed W
s
t
. Information carried by each of the ten
generated variables is calculated as a sum of elements in every column of
the covariance matrix:
η = (η
1
, η
2
, . . . , η
q
)
where
η
j
=
q

i=1
ϕ
ij
41
Figure 2.9 plots the amount of information carried by each of the ten gen-
erated points in the total sum.
0.1
0.11
0.12
0.13
0.14
0.15
0.16
0.17
0 1 2 3 4 5 6 7 8 9
i
n
f
o
r
m
a
t
i
o
n
order of variable
Figure 2.9: Amount of the information carried by each of the ten variables
Point Y
∗(j)
t
is shifted in the specified above direction η towards a curve
that presents a required sum of q wind speed values in terms of a normal
variable, i.e. f(Y
t
) = W
s
t
.
Definition Let us denote Y
B
as a vector that satisfies the equation f(Y
t
) =
W
s
t
and acquires the shortest distance to the center of the distribution Y
t
.
Definition Let us denote Y
C
as a vector on the intersection of the line
(2.36) and the tangent to the curve f(Y
t
) = W
s
t
, i.e. a line of the form
1
T
Y
t
= Y as defined in the Section 2.2.5 of this work. A scalar Y is
calculated as in the section 2.2.5:
If 1
T
Y
B
≤ 1
T
µ, Y = 1
T
Y
B
, and
if 1
T
Y
B
> 1
T
µ, we assume Y = b
3
,
where b
3
is a parameter of the quadratic function in (2.18). By modifying
the value of λ, Y
C
(and a corresponding λ
C
) are found so that 1
T
Y
C
and
Y match to a small given tolerance.
Definition Let us denote Y
D
as the point on the line (2.36) that satisfies
the condition of ten one-minute wind speeds being equal a given average
42
value f(Y
t
) = W
s
t
. By modifying the value of λ, Y
D
(and a corresponding
λ
D
) are found so that f(Y
D
) and and W
s
t
match within a small given
tolerance.
4. The probabilities of the points Y
D
and Y
C
are further compared:
If
φ(Y
D
)
φ(Y
C
)
> ρ accept Y
(j)
t
= Y
C
Else reject Y
C
and Go to step 1,
where φ(∗) is the probability distribution function for truncated normal
distribution used for sampling wind speed. As in the case of two-variable
sampling, we assumed ρ to be a fixed number 0.95.
5. Take i = i + 1. If i ≤ T Go to step 1, Else Finish j-th iteration of Gibbs
sampling algorithm.
Steps 1 to 5 are repeated until j reaches a specified number of iterations j

.
We determine j

so that a stochastic process of the resulting sample is weakly
stationary. For an interval of 5 years when the average wind speed is given, there
are j

= 12 Gibbs sampling iterations being performed.
Out of five years of ten-minute average wind speed data records there were
between 34 to 87 rejects reported for a single iteration. Therefore, shifting a
generated point is an improvement to the method used by Allcroft and Glasbey
in [17], even with high wind speeds it rejects very view trial points.
2.3.4 Testing the results of wind speed sampling
On Figure 2.10 there is one hour of generated wind speed plotted along with
original wind speed. However good the generated sample can look on Figure 2.10
it necessary to evaluate it statistically. Let us begin by defining a second-order
stationary process. A process is called second-order stationary (or weakly
stationary) if its mean is constant and its autocovariance function depends only
on the lag [43], so that
E[Y
(j

)
] = µ
and
Cov[Y
(j

)
, Y
(j

)
l
] = ϕ(l)
No requirements are placed on moments higher than second order. By letting
l = 0, we note that the form of a stationary autocovariance function implies that
43
the variance, as well as the mean, is constant.
0
5
10
15
20
25
100 120 140 160 180 200 220
w
i
n
d

s
p
e
e
d
,

m
/
s
time, min
original wind speed
sampled wind speed
10 min average of original wind speed
10 min average of sampled wind speed
Figure 2.10: An example of the sampled and the original values of 1-minute wind
speed plotted along with the ten minute average wind speed
The chains encountered in Monte Carlo Markov Chain settings satisfy a sta-
bility property, namely a stationary probability distribution that exists by con-
struction. This means, if there exists a distribution [Y ] such that Y
(j

)
i
∼ [Y ],
then Y
(j

)
i+1
∼ [Y ].
Let us analyse the first two moments of the probability distribution for the
original and sampled data sets. The first moment of the probability distribution
is calculated as an expected value for the sampled normalised wind speed, it
reflects the mean of truncated normal distribution described in the section 2.2.4:
E[Y
(j

)
] = 0.2412079 while µ = 0.2419591
where Y
(j

)
is the j

-th sample of the Gibbs sampling method.
The second moment of the probability distribution is calculated with the
autocovariance coefficients of the sampled normal variables. Let us take 5 years
of the recorded one-minute wind speed for the analysis, which is 2592000 points
in the data set. Figure 2.11 plots the autocovariance functions of the original and
sampled normalised wind speeds.
We see that the difference between the mean of the i-th sample and the original
data is insignificant, µ −E(Y
(j

)
) < 10
−3
. One day is chosen as a significant test
interval to compare the autocovariance of the sampled and the original datasets
where they are very close to each other.
44
0
0.1
0.2
0.3
0.4
0.5
0.6
0 200 400 600 800 1000 1200 1400
a
u
t
o
c
o
v
a
r
i
a
n
c
e
lag, min
original normalised wind speed
sampled normalised wind speed
Figure 2.11: Autocovariance of the sampled and the original normalised 1-minute
wind speed
For the next test we calculate variance of one-minute wind speeds inside
ten-minute time sections. The variance is calculated for all consecutive groups of
ten minute wind speeds.
var(Y
j

) =
1
Tq
Tq

t=1
(y
j

t
−Y
j

i
)
2
(2.37)
where y
j

t
are the sampled normalised wind speeds from the ten-minute interval i
with the average Y
j

i
. As data was sampled for ten-minute sections it is important
to ensure that there is no boundary effect appearing in the results. To test this,
variance (2.37) is calculated for shifted ten-minute intervals.
Table 2.3: Variance of the original and the sampled normalised wind speed
shift original data sampled data
0 0.027672 0.025126
1 0.027672 0.025126
2 0.027672 0.025126
3 0.027672 0.025126
4 0.027672 0.025126
5 0.027672 0.025126
6 0.027672 0.025126
7 0.027672 0.025126
8 0.027672 0.025126
9 0.027672 0.025126
45
Table 2.3 shows that independent of the start point for the grouping of the
interval 1 . . . Tq cumulative variance has the same value 0.025126. This supports
a statement that there are no boundary effect appearing in the sampled data.
2.4 Diversification of wind energy
As it was stated at the beginning of this chapter, we require a dataset of ge-
ographically distributed one-minute wind speed in order to model short-term
electricity generation. Utah Geological Survey publishes ten-minute wind speed
([21]) recorded in different sites of the Utah state of the United States of Amer-
ica. Section 2.2 of this work introduced Gibbs sampling algorithm and in Section
2.3 we originally adapted this algorithm so that one-minute wind speed is sam-
pled when the average value of ten-minute wind speed is given. We applied
further the modified Gibbs sampling to generation of one-minute wind speed us-
ing ten-minute wind speed provided by the Utah Geological Survey. Statistical
characteristics of one-minute wind speed were based on the data received with
the MST Radar located at the Frongoch farm in Wales ([20]) that allowed us to
study short-term wind electricity generation with wind energy in Great Britain.
0
2
4
6
8
10
12
14
16
0 20 40 60 80 100 120
w
i
n
d

s
p
e
e
d
,

m
/
s
time, min
wind speed
Figure 2.12: An example of wind speed sampled for one of the Utah sites
We chose 22 different sites from the Utah state where ten-minute wind speed
was recorded by Utah Geological Survey. They are distributed around the state
that introduces an effect of the geographical diversity into the accumulated wind
energy curve. For every site we generated a valid sample of one-minute wind
46
speed using the modified Gibbs sampling algorithm explained in Section 2.3.3,
so that a given value of ten-minute wind speed is achieved. Figure 2.12 plots an
example of wind speed sampled for one of the sites in Utah.
0
10
20
30
40
50
60
70
0 20 40 60 80 100 120
e
n
e
r
g
y
,

M
W
time, min
wind turbines at 22 sites
22 wind turbines at one site
Figure 2.13: An example of geographically distributed wind energy output
We sampled 22 sets of one-minute wind speed for different sites in Utah that
were further transfered into energy using the wind power curve plotted in Figure
2.1. Notice, variability of the combined wind energy from 22 sites in Utah varies
less than the same amount of energy produced at one site. To demonstrate this,
Figure 2.13 plots the combined wind energy that could have been generated from
22 sites in Utah against wind energy that could have been generated by 22 wind
turbines using a single wind speed sample plotted in Figure 2.12.
Considering a geographical diversity of wind farms in Great Britain we wish to
use the combined wind energy from 22 sites in the Utah state for the optimization
modelling of the electricity generation further in this work.
47
Chapter 3
Modelling Electricity Generation
with Wind Energy
Scheduling electricity generation is a complex engineering problem that takes into
account various parameters of generation, demand, transmission and distribution.
This chapter specifies the details of the electricity demand and generation in
the United Kingdom and formulates a mathematical programming problem of
scheduling the operating reserve so that the supply and demand match within
the target range. We use the formulated optimization model to investigate the
costs involved in incorporating different levels of wind energy into the electricity
generation system of Great Britain.
We assume in this chapter that wind speed is forecast with certainty. There-
fore, the formulated optimization model is deterministic and includes some of the
common constraints of power generation. The uniqueness of the model developed
in this chapter consists in the data representing the power system of the UK and
the application of the optimization model to estimating the system balancing cost
of the variable wind energy.
3.1 Elements of the British power system and
simplifying assumptions
Section 3.1 introduces the electricity generation system of the United Kingdom
and how it can be adapted for the optimization modelling. In Chapter 2 we devel-
oped an algorithm that would produce one-minute wind energy data, therefore,
this section focuses on the parameters of thermal generation and its fuel cost.
We assume that the electricity load only slowly changes, leaving the matter of
48
interaction between the demand and wind power for further research.
3.1.1 Electricity generation
Currently electricity generation in the United Kingdom includes a diversity of
generating units. The amount of generation provided by different types of gen-
erating units vary depending on the time of the year and the size of the load.
For the purpose of this work we use Plant/Demand Balance of the “Seven Year
Statement” published by the National Grid Electricity Transmission in 2008 [35].
Table 3.1 shows the percent of power from the various plant types in the total
generation.
According to the figures in Table 3.1 two types of plants that cover more than
70% of the total British electricity generation are Gas and Coal turbines. Various
nuclear generators contribute to another 14% of the total generating capacity.
Coal, gas and nuclear units have common characteristics with other units of the
same type that we wish to incorporate in the model. Parameters describing a
certain type of the generator are available through the Balancing Mechanism
(BM) reporting system ([5]). This website publishes on a daily basis Physical
and Dynamic parameters (as defined in Appendix A) of the units participating
in the balancing mechanism. We wish to incorporate some of the characteristics
into the mathematical programming model, in particular the available capacity
and the maximum rates of change in the power output.
The first characteristic that the System Operator takes into account while
planning an output of a thermal generator, like gas, coal or nuclear, is a capacity
limit. Capacity limit is specified in the BM reporting system as the Export Limits
parameter (defined in Appendix A) and describes a maximum level at which the
generating unit may be exporting power to the transmission system. The capacity
limit of any type of the generator depends on the size of the unit and can vary
within the same type of generators. However, for the modelling purpose we wish
to use a single value as a capacity limit for every type of a same plant. Let us use
a capacity limit for a typical gas generator as 500MW and 600MW for typical
coal and nuclear types of generating units.
The output levels of thermal plants are not constant and can vary following
the changes in the electricity load. The maximum rate at which units change their
production can be described by Run-Up and Run-Down export rates (defined
in Appendix A). These rates depend on the current output of the generator
and are used by the System Operator in scheduling the operating Reserve when
balancing the supply and demand in real-time.
49
Table 3.1: Plant/Demand balance for SYS Background (2007/2008)
hline Plant type Capacity (MW) %
biomass 45 0.06
CCGT 25532 32.58
CHP 1725 2.20
Hydro 1028 1.31
Interconnector 1988 2.54
Large coal 4413 5.63
Large coal+AGT 21462 27.39
Medium coal 1152 1.47
Medium coal+AGT 1102 1.41
Small coal 783 1.00
Nuclear AGR 8365 10.68
Nuclear Magnox 1450 1.85
Nuclear PWR 1190 1.52
OCGT 589 0.75
Offshore wind 140 0.18
Oil+AGT 3496 4.46
Pumped storage 2300 2.94
Wind 1597 2.04
The Run-Up and Run-Down export rates specified in the BM reporting system
are further adjusted to be used in the model. Maximum Run-Up and Run-Down
rates are published in the form of a table with three different values for each of
the rates that depend on a production level. The rates for different power output
can cause a non-convex constraint in the optimization model. Let us consider the
following example:
Table 3.2: An example of the Run-Up and Run-Down export rates for a typical
coal plant
Export rate r
1
e
1
r
2
e
2
r
3
Run-Up 5.0 177 0.2 180 5.0
Run-Down 5.0 240 0.4 234 58.5
Table 3.2 gives an example of the Run-Up and Run-Down export rates pub-
lished for a typical coal plant by the Balancing Mechanism reporting system.
There are usually three Run-Up and Run-Down export rates, r
1
, r
2
and r
3
sub-
mitted by the generating unit before the Gate Closure. Every rate is applied to
a different output level specified by the “elbows”, e
1
and e
2
. So, Run-Up rate
r
1
= 5.0 MW per minute is applied to the production level from 0 to 177 MW.
And a maximum Run-Down export rate r
3
= 58.5 MW per minute can be applied
when a coal generator cools down between 234 MW and 0.
50
The specified sequence of maximum Run-Down and Run-Up export rates
leads to the non-convex constraint describing the power output rate of change
between the current and the following minute. Hence, we wish to simplify the
rates presented in Table 3.2. First, the rate r
2
can be dropped as it is likely to
be an internal work taking place in the generating unit and is not important for
our work. Later in this chapter we modify the values of the remaining rates r
1
and r
3
so that they form a convex feasible set. The resulting constraints will be
formulated in Section 3.2.2.
Short time scale water reservoir limits are often not building and a hydro
generator can be modelled in the same way as a thermal generator except that
the “fuel” is water. Table 3.1 also shows that the hydro and pumped storage
contribute no more than 5% to the total electricity generation. Considering a
limit of water resources in United Kingdom and capital cost of pumped storage
generation we wish to avoid modelling hydro generation in this work.
According to the “Seven Year Statement” published in the beginning of 2008
wind contributes only 2% to the total electricity generation, however, the installed
wind energy capacity have doubled during the last year and is expected further
to increase. We wish to introduce the wind energy output into the electricity
generation system based on the wind speed data provided by the Utah Geological
Survey. Every site is assumed to contribute an equal number of wind turbines to
the total wind energy curve. This curve is further scaled so that wind contributes
a required amount of energy to the total electricity generation.
3.1.2 Cost of generation
Due to the fact that the electricity generation is scheduled minute by minute
during the real time, there is not enough time to make a decision about switching
additional thermal plants. For some coal generators it takes up to 2 hours to
be warm and ready to respond. Therefore, we assume that only generators that
are currently onstream can be used to meet the load. Short term scheduling
determines a cost of the electricity generation as the fuel cost for all types of
generators. Following this, generating cost of wind and nuclear energy is zero
and cost of gas and coal is defined by the fuel prices.
In the report “Keeping the lights on: Nuclear Renewables and Climate Change”
Milborrow [41] discusses current cost of the electricity generation and the way it
can be calculated. Also governmental organisation “Department for Business En-
terprise and Regulatory Reform” (BERR) regularly publishes the average prices
of fuels purchased by the major UK producers. We wish to use the quarterly
51
energy prices reported by BERR for year 2008 [7]. An average price of coal for
power stations was increasing since 2007 and amounted to 0.72 p/KWh. The
price of gas has increased by more than 50% up to 1.5 p/KWh.
In order to apply these prices to the electricity generation in one-minute fre-
quency we recalculate the given prices in £/MW*min.
c
coal
= 0.72p/KWh = 0.12£/MW*min
c
gas
= 1.5p/KWh = 0.25£/MW*min
Note, there is no capital cost included in the cost of the electricity generation,
we are interested in minimizing the fuel cost later in this work.
3.1.3 Electricity demand
We wish to model the electricity demand with the same time frequency as the
generation, i.e. on a minute by minute basis. However, a most frequent electric-
ity demand reported through the Balancing Mechanism reporting system is the
historic and real-time records of the national demand.
32000
34000
36000
38000
40000
42000
44000
46000
48000
50000
0 200 400 600 800 1000 1200 1400 1600
o
u
t
t
u
r
n

d
e
m
a
n
d
,

M
W
*
m
i
n
time, min
1-minute curve
half hour average
Figure 3.1: Half-an-hour average of the daily electricity demand in the United
Kingdom and its piecewise linear approximation with 1-minute frequency
As the electricity demand is available in half-an-hour average estimates we wish to
transfer it into a continuous piecewise linear function with one minute frequency.
Figure 3.1 shows an example of the winter electricity demand used later for the
deterministic and stochastic models.
52
3.1.4 Cost of demand-and-supply mismatch
In reality it is very hard to generate exactly the value of the national electricity
demand at any moment. Considering the uncertainty of consumer’s behaviour
and a probability of losing a generating capacity in the system, it is important
to have a flexibility of meeting the demand. As it was stated in Chapter 1 the
transmission system of the United Kingdom operates at the frequency 50Hz with
a range of 1% variation above and below 50Hz.
We wish to incorporate an option of the mismatch between the electricity
generation and demand in the model. The transformation function that calculates
the deviation in power, based on a given deviation in frequency is complex: it
depends on the system parameters and the load level. Transmission system is
not modelled in this work, therefore, a mismatch between the generation and the
electricity demand that corresponds to the variations in the frequency is assumed
to be a function of the electricity demand. Further, the generation is allowed to
vary 2% above or below the electricity load which corresponds to the 1% variation
in frequency.
Although it is possible for the generation to deviate from the demand it is
not a good practise to get close to the upper and lower limits. This can get the
transmission system into zones of too high or too low frequency. As a result of
modelling minimization of the electricity generating costs, the solution tends to
produce an output of the electricity generation 2% below the electricity demand.
To prevent this happening let us introduce a penalty of mismatch between the
demand and generation.
Figure 3.2: A mismatch between the electricity demand and the generation with
the cost attached
53
Let us divide the allowed mismatch between the electricity demand and generation
into several intervals so that the cost of a mismatch increases with the higher
deviation between the demand and generation.
Gas generating units are assumed to be a marginal type of the electricity
generation. Gas turbines have the highest fuel cost among the thermal plants
but the Run-Up and Run-Down rates are also higher than that of the coal and
nuclear generators. Determined as the marginal generator, gas turbines are used
first to respond to the mismatch between the electricity demand and generation,
so that a cost of the mismatch is estimated according to the fuel cost of the gas
generator. In Figure 3.2 the cost c
+
1
corresponding to the mismatch between the
electricity demand and the generation d
+
1
is marginally higher than the fuel cost
of the gas generator. If gas turbines are able to cope with the load fluctuations,
higher cost of mismatch encourages the generation to meet the demand exactly.
However, it is rather allowed for the generation to deviate from the demand
than to modify the power output of the coal or gas turbines. The cost c
+
3
is
significantly higher than the cost of the marginal generator so that the mismatch
between the electricity demand and generation exceeds d
+
1
+ d
+
2
only if no other
feasible solution exists.
3.1.5 Summary
Let us summarise the above points in order to formulate a deterministic problem
of the electricity generation. To estimate the system balancing cost when variable
wind energy is incorporated into the electricity generation system we assume:
a. there is no transmission cost;
b. there are three types of thermal generators (gas, coal and nuclear) included in
the model. They are characterised by the limited capacity and the Run-Up
and Run-Down rates;
c. there is only generating (fuel) cost considered in the problem. Time scale is
short so there is no start-up costs of thermal generation;
d. wind power generation is determined by the incoming data and is perfectly
forecast for the deterministic problem;
e. there is no hydro energy included in the model;
54
f. the electricity demand is deterministic;
g. there is a cost of mismatch between the electricity demand and generation.
3.2 Deterministic model for the electricity gen-
eration with wind energy
This section formulates a linear programming (LP) mathematical optimization
model for the electricity generation scheduling problem specified in Section 3.1.
3.2.1 Decision variables
Let T denote the number of time intervals obtained by discretising the planning
horizon. This discretisation may be chosen uniformly or non-uniformly. For the
current problem we use one-minute uniform discretization of the planning horizon.
Let G denote the set of all the generators and consist of the union of C, the
set of the conventional generators, and W, the set of the wind generators. The
set of conventional generators also consists of Coal, Nuclear and Gas generating
plants. The decision variables in the model correspond to the outputs of each
thermal unit, i.e. the electric power generated by the Coal, Gas and Nuclear
generators. The decision variables are denoted by
x
tg
, t = 1, 2, . . . , T; g ∈ C
where x
tg
is the production level of thermal unit g during time period t. We
assume a constant power output during 1 minute and measure it in MW*min.
Wind power generation is determined by the wind speed and equals to the
power w
e
t
at time t:
x
tg
= w
e
t
, t = 1, 2, . . . , T, g ∈ W (3.1)
3.2.2 Constraints
The electricity generation system is complex so it is possible to include different
constraints in its optimization model (capacity, emission limits, system security
and others). The model described in this section would contain only those con-
straints that allow us studying a problem of incorporating the wind energy into
the electricity generation system. Decision variables corresponding to the ther-
mal generation mentioned above have finite upper and lower bounds representing
55
unit capacity limits of the generation system:
x
g
≤ x
tg
≤ x
g
, t = 1, 2, . . . , T, g ∈ C (3.2)
The constants x
g
and x
g
denote minimal and maximal possible outputs of the
thermal units respectively.
Figure 3.3: Area formed by Run-Up and Run-Down constraints for thermal gen-
erators
Further single-unit constraints describe maximum Run-Up and Run-Down
export rates introduced in Section 3.1.1. An example of Run-Up and Run-Down
export rates that a generating plant provides to the System operator before the
“Gate closure” is given in Table 3.2. As it was stated then, these export rates
cause a non-convex constraint. Hence, let us modify the limits of maximum Run-
Up and Run-Down rates in a way that they form a convex set (x
tg
, x
(t+1)g
), t =
1, 2, . . . , (T −1), g ∈ C of the decision variables.
As mentioned in Section 3.1.1 we use only two Run-Up and two Run-Down
rates that are connected in a piecewise linear function and form a convex set.
The piecewise linear function can be chosen in a way that tighten or relax the
maximum export rate limits. Constraints (3.3) − (3.6) formulated below relax
the limits of Run-Up and Run-Down rates of generating plants. Let r
1+
and
r
2+
denote two Run Up rates separated by the “elbow” e
+
, while r
1−
and r
2−
- denote two Run Down rates separated by the “elbow” e

. The corresponding
constraints of the change in the power output are described by inequalities:
x
t+1,g

e
+
+ r
2+
−r
1+
e
+
x
t,g
+ r
1+
, t = 1, 2, . . . , T, g ∈ C (3.3)
56
x
t+1,g
≤ x
t,g
+ r
2+
, t = 1, 2, . . . , T, g ∈ C (3.4)
and
x
t+1,g

x
g
−e

+ r
2−
−r
1−
x
g
−e

x
tg

r
2−
x
g
−r
1−
e

x
g
−e

, t = 1, 2, . . . , T, g ∈ C
(3.5)
x
t+1,g
≥ x
t,g
−r
2−
, t = 1, 2, . . . , T, g ∈ C (3.6)
Figure 3.3 plots a convex set (x
tg
, x
(t+1)g
), t = 1, 2, . . . , (T −1), g ∈ C with new
Run-Up and Run-Down export rates in a form of the shaded area bordered by
two lines from the above and two lines from the below. Every line on the graph
states a maximum rate of change between the output at the current minute and
the following minute. The two lines from the above in Figure 3.3 intersect in a
point of the “elbow” e
+
and are described by the inequalities (3.3) −(3.4). The
two lines from the below intersect in a point of the “elbow” e

, their gradients
are r
1−
and r
2−
respectively. Run-Down rates are described by the inequalities
(3.5) −(3.6).
-40000
-20000
0
20000
40000
60000
80000
100000
120000
140000
-800 -600 -400 -200 0 200 400 600 800
c
o
s
t
,

G
B
P
/
M
W
*
m
i
n
deviation from demand, MW*min
excessive generation
insufficient generation
gas turbine cost
Figure 3.4: Load mismatch costs: the difference between the curves gives the
penalty for generation deviating from demand.
The loading constraint couples across all the generating units. It is essential
for the operation of the power system and means that the difference between
the total power generation and the load demand does not exceed the allowed
(by the transmission system) limit at any time. Denoting by D
t
the electricity
demand during a time period t and by ∆
D
a percent mismatch allowed between
the electricity load and the generation, the loading constraint is described by the
57
inequalities:
(1 −∆
D
)D
t

g∈G
x
tg
≤ (1 + ∆
D
)D
t
t = 1, 2, . . . , T (3.7)
In (3.7) a parameter ∆
D
is taken as 2% that corresponds to 1% variation in the
system frequency.
In Section 3.1.1 we discussed a cost of the mismatch between the electricity
supply and demand. Logically the generators should produce close to the demand
and as far away from the lower and upper bounds as possible. To achieve this
result in the model we introduce a piecewise linear cost for the load imbalance

+
v
; ζ
+
v
}, v = 1, 2, . . . , V where V denotes a number of bands in the penalty cost.
To avoid giving an incentive to violate the load constraint, this piecewise linear
load imbalance cost should be above the cost of the marginal generator at that
time. Figure 3.4 plots the fuel cost of the marginal generator (gas in our case)
and three different cost bands for the generation exceeding the demand and the
demand exceeding the generation. The higher mismatch between the electricity
demand and generation the higher the cost of coping with it.
Let us denote a mismatch between the electricity generation and the load by
a set of variables d
+
tv
and d

tv
, t = 1, 2, . . . , T, v = 1, 2, . . . , V such that

g∈G
x
tg
−D
t
=

v∈1..V
(d
+
tv
−d

tv
) t = 1, 2, . . . , T (3.8)
In (3.8) T denotes the planning horizon while V is a number of the cost bands
in the accepted mismatch between the supply and demand [(1 −∆
d
)D
t
; D
t
] and
[D
t
; ((1 + ∆
d
)D
t
)]. Size of every interval v = 1, 2, ; dots, V is measured in MW
and defined by the inequalities:
0 ≤ d
+
tv
≤ d
+
tv
(3.9)
0 ≤ d

tv
≤ d

tv
(3.10)
where d
+
tv
is an upper bound of interval v when the electricity demand exceeds
the generation at time t,

g∈G
x
tg
< D
t
. d

tv
is an upper bound of interval v
when the generation exceeds the demand at time t,

g∈G
x
tg
> D
t
.
Set of parameters [d
+
tv
, d

tv
] is time dependent and defined by the changing
value of demand. It can be calculated in the following way
d
+
tv
= ∆
d
∗ D
t
∗ β
+
v
(3.11)
58
where β
v
determines a part of the demand envelope ∆
d
∗ D
t
that corresponds to
the cost band v = 1, 2, . . . , V .
3.2.3 Objective function
The constraints described above form a feasible set of solutions for our optimiza-
tion problem of scheduling the electricity generation. To complete our LP model,
let us introduce the operational cost of the electricity generation for different
types of plant. Let c
g
denote the fuel cost for each thermal generator g ∈ C.
Fuel cost of the coal and gas generation was introduced in the section 3.1.2.
Then

g∈C
c
g

T
t=1
x
tg
is the operational cost paid by the Supplier to generate

g∈C

T
t=1
x
tg
amount of electricity.
In order to decrease a mismatch between the electricity generation and de-
mand we wish to include the cumulative cost of the mismatch into the objective.
If parameters ζ
+
v
and ζ

v
denote the cost of the excessive and insufficient gener-
ation respectively, we wish to minimize the total cost of the electricity demand
and generation mismatch:
V

v=1
ζ
+
v
T

t=1
(

g∈G
x
tg
−D
t
) +
V

v=1
ζ

v
T

t=1
(D
t

g∈G
x
tg
)
If we define d
+
tv
:=

T
t=1
(

g∈G
x
tg
− D
t
) and d

tv
:=

T
t=1
(D
t

g∈G
x
tg
) this
can equivalently be rewritten as:
V

v=1
ζ
+
v
T

t=1
d
+
tv
+
V

v=1
ζ

v
T

t=1
d

tv
Taking into consideration the fuel cost and the cost of the load imbalance, the
objective of the mathematical programming model is to minimize the cumulative
operational cost of the electricity for the planning horizon T:
min
x

g∈C
c
g
T

t=1
x
tg
+
V

v=1
ζ
+
v
T

t=1
d
+
tv
+
V

v=1
ζ

v
T

t=1
d

tv
(3.12)
In the objective function (3.12)

g∈C
c
g

T
t=1
x
tg
denotes the fuel cost of the
thermal generators g ∈ C and

V
v=1
ζ
+
v

T
t=1
d
+
tv
+

V
v=1
ζ

v

T
t=1
d

tv
denotes the
cost of mismatch between the electricity supply and demand.
59
3.2.4 Formulation of the LP deterministic optimization
model
The constraints and and the objective described in Section 3.2.2 and Section 3.2.3
are combined below:
Objective (minimize cost):
min
x

g∈C
c
g
T

t=1
x
tg
+ +
V

v=1
ζ
+
v
T

t=1
d
+
tv
+
V

v=1
ζ

v
T

t=1
d

tv
Subject to:
x
tg
= w
tg
∀g ∈ W
x
g
≤ x
tg
≤ x
g
∀g ∈ C
x
(t+1),g

e
+
+ r
2+
−r
1+
e
+
x
t,g
+ r
1+
∀g ∈ C
x
(t+1),g
≤ x
t,g
+ r
2+
∀g ∈ C
x
(t+1),g

x
g
−e

+ r
2−
−r
1−
x
g
−e

x
tg

r
2−
x
g
−r
1−
e

x
g
−e

∀g ∈ C
x
(t+1),g
≥ x
t,g
−r
2−
∀g ∈ C
(1 −∆
D
)D
t

g∈G
x
tg
≤ (1 + ∆
D
)D
t

g∈G
x
tg
−D
t
=
V

v=1
(d
+
tv
−d

tv
)
0 ≤ d
+
tv
≤ d
+
tv
∀v = 1, 2, . . . , V
0 ≤ d

tv
≤ d

tv
∀v = 1, 2, . . . , V
d
+
tv
= ∆
d
∗ D
t
∗ β
+
v
∀v = 1, 2, . . . , V
d

tv
= ∆
d
∗ D
t
∗ β

v
∀v = 1, 2, . . . , V
where t = 1, 2, . . . , T.
3.3 Calculating cost of wind energy variability
An output of the mathematical optimization model (3.1) − (3.12) formulated in
Section 3.2 represents a series of one-minute electricity generation level for every
thermal generator g ∈ C over a time period of T minutes. The optimal solution
is chosen according to the minimization function that includes fuel cost of the
60
thermal generators and the cost of a mismatch between the electricity supply and
demand.
The deterministic LP optimization model formulated in Section 3.2.4 includes
two sets of variables {x
tg
} and {d
tv
}, where g ∈ C, t = 1, 2, . . . , T and v =
1, 2, . . . , V . If we denote κ as a number of thermal generators in the set C, a total
number of decision variables in the model equals κ∗T +V ∗T. To reduce a number
of variables we wish to group thermal generators so that there are only three
generators included in the set C: nuclear, gas and coal. Each of these thermal
generators represent a big number of plants with similar fuel costs and capacity
limits. Then, taking V = 6 there are 12960 decision variables in the deterministic
optimization problem of daily (T=1440) electricity generation. Assuming further
one accumulated wind energy curve in the set of wind generators, a number of
constraints results in 60480 for a daily electricity generation problem.
Formulated deterministic optimization problem is linear so that it can be
efficiently solved with a CPLEX solver. We used version 10.00 of the CPLEX
solver that resulted in CPU time < 1s.
Formulated deterministic model can be used now to estimate the system bal-
ancing cost that appears as a result of wind energy variability. In order to estimate
this cost we compare the results of two different cases. The power system bene-
fits initially from the wind energy being present in the system as there is no fuel
cost associated with the generation of wind energy. But this advantage of wind
energy is challenged by its variability. If a significant amount of the wind energy
fluctuates, thermal generators must frequently modify their production level in
order to balance these fluctuations in the power system. Fluctuations of the wind
energy result not only in the additional fuel cost but also the cost of a mismatch
between the electricity generation and the load.
3.3.1 Case of constant wind energy that reduces the op-
erational cost
First let us consider a case when a wind energy profile if flat, i.e. there are no
fluctuations in the wind output. There are three types of thermal generators used
to meet the electricity load: nuclear, coal and gas generators. These generating
units can be ordered by the fuel cost and the output flexibility, i.e. a size of the
maximum Run-Up and Run-Down export rates.
61
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
20000
0 20 40 60 80 100 120
p
o
w
e
r
,

M
W
time, min
gas
coal
nuclear
wind
Figure 3.5: Electricity generation when 11% of wind energy with flat profile is
introduced to the system
Figure 3.5 plots an example of the optimization model output, which is a
generation level for the thermal plants as well as the available wind energy with
the flat profile. Nuclear generators are the cheapest, we price the fuel cost as
zero in the model. It is also the least flexible type of generation as it has the
lowest Run-Up and Run-Down rates. Gas turbines are the most expensive type of
generation in the model but they are the most flexible generators. As a result the
model output shows the nuclear and coal generation at the maximum available
capacity and gas generators produce the remaining electricity to meet the load.
Definition Let us denote a flexible generation producing electricity away from
its upper and lower boundaries as the marginal generation with the fuel cost
equals c
m
. Marginal generators are the first to react on changes in the electricity
load.
Wind energy is constant for this solution with the output of 5173 MW*min that
makes up to 11% of the load in two hours.
The cumulative fuel cost corresponding to this problem of scheduling the
generation is reduced by the fuel cost of the marginal generator that would be
used if there was no wind present in the system.
Definition Let us denote F
flat
w
as the cumulative fuel cost of the mathematical
optimization problem with flat wind energy profile and f
+
w
as the saving in fuel
cost by the Supplier of the electricity when it uses the wind energy in the total
generation.
62
Cumulative fuel cost F
flat
w
is related to the problem with no wind energy in the
system in the following way:
F
flat
w
= F
0
−f
+
w
(3.13)
In (3.13), F
0
is the fuel cost related to the optimization problem with no wind
energy available. The cost of the mismatch between the electricity demand and
the generation equals zero in the current case as the load can be met exactly.
3.3.2 Case of variable wind energy that increases opera-
tional cost
For the next case a profile of the wind energy curve is variable. It is balanced by
the group of thermal generators: nuclear, coal and gas. As in the previous case
they are ordered by the fuel cost and maximum Run-Up and Run-Down export
rates.
4000
6000
8000
10000
12000
14000
16000
18000
0 20 40 60 80 100 120
p
o
w
e
r

o
u
t
p
u
t
,

M
W
time, min
gas
coal
nuclear
wind
Figure 3.6: Electricity generation when 11% of variable wind energy is introduced
to the system
Figure 3.6 plots an example of the optimization model output, which is a
generation level for the thermal plants as well as the available wind energy. Unlike
the case presented in the Section 3.3.1, wind energy for the current problem
fluctuates. Let the fluctuations of the wind energy curve be proportional to
the amount introduced to the power system. Figure 3.6 plots 11% of the wind
energy in total electricity generation of 1 day. In order to balance the variable
63
wind energy the thermal generators have to modify their production profile. Gas
turbines have the highest Run-Up and Run-Down exports rates so it reflects most
of the variation in the wind energy. However, gas generation is not able to cover
some of the steep changes in wind energy from minute to minute. Then the
coal generators modify the power output so that the supply and the electricity
demand are within the target range. The nuclear plants have the lowest Run-Up
and Run-Down rates which results in the output profile being flat at almost all
the time t = 1, 2, . . . , T.
Figure 3.6 shows that it is hard to meet the electricity demand exactly when
variable wind energy is introduced into the power system. In the reality coal
generators can not modify the power output as often. Therefore, we wish to
allow the total generation to deviate from the electricity demand within a certain
range. This can be achieved by setting the penalty cost of the first band v = 1
marginally higher than the fuel cost of the gas turbine that acts most of the
time as the marginal generator. If the gas turbines are able to cope with the
variation of the wind energy it is cheaper for the total generation not to deviate
from the electricity demand. However, a small mismatch between the supply and
the electricity demand is preferable to the modification of the coal power output
as it is shown in Figure 3.7.
The cumulative fuel cost corresponding to the problem of scheduling the elec-
tricity generation with the variable wind energy is affected not only by the fuel
cost of the marginal generator but also by the fuel cost of the generators that
modify the output profiles in order to balance the fluctuations of the wind energy.
Definition Let us denote F
var
w
as the cumulative fuel cost of the optimization
problem with variable wind energy profile and f

w
as the additional fuel cost of the
generators that modify the output profiles in order to balance the fluctuations of
the wind energy.
Cumulative fuel cost F
var
w
is related to the problem with no wind energy in the
system in the following way:
F
var
w
= F
0
−f
+
w
+ f

w
F
flat
w
+ f

w
,
where F
0
is the fuel cumulative cost related to the optimization problem with no
wind energy available and F
flat
w
was calculated in (3.14) as the accumulated fuel
cost of the problem with the flat wind energy profile.
64
-800
-600
-400
-200
0
200
400
600
800
0 20 40 60 80 100 120
m
i
s
m
a
t
c
h

b
e
t
w
e
e
n

g
e
n
e
r
a
t
i
o
n

a
n
d

l
o
a
d
,

M
W
time, min
Figure 3.7: Electricity generation and load mismatch when 5% of variable wind
energy is introduced to the system
Figure 3.8 plots the cumulative fuel cost of a daily electricity generation
weighted by the total amount of the produced power. The highest cost is the
fuel cost calculated for a problem when there is no wind energy in the power
system. The fuel cost from the bottom reflects a problem with the variable wind
energy that can be balanced by the marginal generator. So the weighted fuel cost
is a piecewise linear function with the gradient of the marginal generator.
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14
o
p
e
r
a
t
i
o
n
a
l

c
o
s
t

w
e
i
g
h
t
e
d

b
y

t
o
t
a
l

g
e
n
e
r
a
t
i
o
n
,

p
o
u
n
d
s
/
M
W
*
m
i
n
percent of wind power in total generation, %
zero wind
flat wind profile
variable wind profile
Figure 3.8: The fuel cost of the wind energy variability shown for different levels
of it introduced into the electricity system
The system balancing cost that appears when variable wind energy is introduced
into the power system can be estimated as the difference between the fuel cost
65
calculated for the problem with variable wind energy that can be balanced with
a combination of different thermal generators and the fuel cost calculated for a
problem with variable wind energy that can be balanced by the marginal gener-
ator.
66
Chapter 4
Statistical analysis of operational
cost
In the previous chapter we introduced the deterministic optimization LP model
of the electricity generation system. This model was used to estimate the system
balancing cost that appears when variable wind energy was incorporated into
the generation system. We assume further that the saving in fuel cost and the
additional fuel cost (as defined in Section 3.3.1 and Section 3.3.2) can be deter-
mined as a function of the statistical parameters of the wind energy and develop
a unique model that estimates the parameters of such a model.
4.1 Formulation of the problem
To evaluate how well the statistical model estimates the system balancing cost,
the optimal value of the cumulative fuel cost is found using the deterministic
optimization model described in Chapter 3. Again we assume that the wind
speed can be forecast exactly over the planning horizon and there is no uncertainty
about the electricity load. For the purpose of this chapter we wish to make the
following modifications to the optimization model:
• The electricity demand is constant so that variability in wind energy is
clearly presented in the output. However, the general conclusions can be
applicable for the case of slowly changing load.
• The electricity load has to be met exactly by the generating units, which
means that there is no cost of mismatch between the electricity supply and
demand. The objective function of the optimization models in this chapter
is presented only by the accumulated fuel cost.
67
• There is hydro generation included in the optimization model, however, it
is small and does not affect the solution.
Let us consider a small electricity generation system where the set of the
conventional generators C consists of 15 thermal units. Thermal generators are
not differentiated by the available capacity and every plant has the following
upper and lower bounds:
x
g
= 0MW and
x
g
= 30MW, g ∈ C.
Hydro generation is small for all the models in this chapter and has a constant
output of 20 MW. Maximum Run-Up and Run-Down export rates (introduced
in Section 3.1.1) of thermal generating units are the same for all the range of
the electricity output, x
g
≤ x
t,g
≤ x
g
, therefore, inequalities (3.3) − (3.6) are
simplified as
x
t+1,g
≤ x
t,g
+ r
+
x
t+1,g
≥ x
t,g
−r

t = 1, 2, . . . , T, g ∈ C,
where r
+
equals 0.125 MW*min and r

equals 0.125 MW*min for all the models
in this chapter.
The thermal generators are differentiated by the fuel cost chosen in a certain
merit order so that cheaper thermal plants are used first. Let us take a merit
order of the fuel costs to be a piecewise linear function, so that the cost of the
first unit c
1
= 1£/MW*min increasing by 0.1£/MW*min for every next unit so
that the 15th thermal plant is priced at c
15
= 2.4£/MW*min. There is no cost
associated with hydro generation.
Definition Let us denote F
A
as the actual accumulated fuel cost calculated using
the deterministic optimization model.
Further in this chapter we will consider different statistical models of the ac-
cumulated fuel cost and compare these with the actual accumulated fuel cost.
Every next model is chosen so that the error between the actual and calculated
accumulated fuel cost decreases. We assume the accumulated fuel cost of the
electricity generation to be a piecewise linear function. The value of F
A
is calcu-
lated per unit time (per minute in the current case), therefore all the statistical
models in this chapter are independent of the planning horizon T.
68
4.2 Case of marginal generator following a pro-
file of the wind energy curve
Let us begin with a simplified situation of the electricity generation when the
wind energy is constant and equals zero. In this case the accumulated fuel cost
can be calculated accurately with an error ǫ = 0 by:
F
A
= F
0
(4.1)
where F
0
is the generated cost when there is no wind energy available and can
be found from the merit order fuel costs of the thermal generators.
The model described by (4.1) is not, however, suitable for calculating the
accumulated fuel cost if there is wind energy incorporated into the power system.
In this case error of the calculation ǫ = F
A
−F
0
can be significantly higher than
zero. To improve the model we take into account another simplified situation of
electricity generation, when a marginal thermal generator (as defined in Section
3.3.1) has the Run-Up and Run-Down export rates big enough to cope with
variability of wind energy curve.
0
5
10
15
20
25
30
0 20 40 60 80 100 120 140 160 180
e
n
e
r
g
y
,

K
W
time
Electricity generation with hydro, wind and thermal power
Th1
Th2
Th3
Th4
Th5
Th6
Th7
Th8
Th9
Th10
Th11
Th12
Th13
Th14
Th15
wind
hydro
Figure 4.1: Electricity generation with only the marginal generator modifying the
power output
Figure 4.1 plots an example of the solution for this simplified case of electricity
generation. On this graph thermal plant with the index “Th8” is the marginal
generator. It frequently changes the production plan in order to balance the
69
wind energy output and meet the electricity load exactly. The first seven gen-
erators have the lowest fuel cost so that they produce at full capacity. The last
seven generators with fuel cost higher than that of the marginal generator do not
produce.
This case improves the model that estimates the operational cost of the elec-
tricity generation:
F
A
= F
0
−c
m
x
w
−ǫ. (4.2)
In (4.2) c
m
denotes the fuel cost of the marginal thermal generator and x
w
is an
average wind energy over the modelled period of time T.
For both simplified cases considered in this section, the operational cost can be
calculated with (4.2) with an error ǫ = 0. However, for a situation when the rates
of change of the wind energy curve are higher than that of the marginal generator,
there can be an error calculating the operational cost from the right-hand side of
formula (4.2). The error term in (4.2) is non-zero and can be significant. Thus,
we wish to find a better model in order to improve accuracy of the operational
cost calculation.
4.3 Case of wind energy curve having a zigzag
profile
Assume, the other elements of the model that estimates the operational cost also
depend on the statistical parameters of the wind energy curve.
0
5
10
15
20
25
30
0 20 40 60 80 100 120 140 160 180
e
n
e
r
g
y
,
K
W
min
Generation with equal number of thermal generators on top and bottom
Th1
Th2
Th3
Th4
Th5
Th6
Th7
Th8
Th9
Th10
Th11
Th12
Th13
Th14
Th15
wind
hydro
Figure 4.2: Electricity generation with the wind energy profile as a zigzag
70
To evaluate them we construct another simplified case when the marginal gen-
erator has the Run-Up and Run-Down export rates smaller than those of the
wind energy curve, however, the wind energy itself is periodic and has a form of a
zigzag. Figure 4.2 shows that the solution is also periodic and makes it possible to
calculate the accumulated fuel cost exactly as it is found using the deterministic
optimization model.
Further in this section there is a number of tests performed with one V-shaped
drop in wind energy that is plotted in Figure 4.3. Because the solution shown in
Figure 4.2 is periodic, calculations for one V-shaped drop are applicable to the
whole zigzag curve.
4.3.1 Analytical calculation of the accumulated fuel cost
Definition Let us denote a height of wind energy curve as h
w
that represents
the difference between the highest and the lowest value of wind energy in the
time horizon 1, 2, . . . , T.
A first test investigates how the accumulated fuel cost changes depending on the
modification of the height of every V-shaped wind energy curve defined in Figure
4.3 while a length of the planning period of time T remains the same.
0
5
10
15
20
25
30
0 20 40 60 80 100 120 140 160 180
K
W
/
m
in
min
Electricity generation with hydro, wind and thermal power
Th1
Th2
Th3
Th4
Th5
Th6
Th7
Th8
Th9
Th10
wind
hydro
Figure 4.3: Electricity generation for a wind energy curve as one zigzag and the
thermal plants increasing and decreasing their output in a symmetric Λ-shape
Definition We denote γ
w
as the average of the absolute value of the gradients
calculated for the wind energy curve.
Modification of the wind energy height h
w
is performed in a way that the absolute
value of the gradient of the drop in the symmetric V-shaped wind energy curve
71
changes proportionally to the height while the average of wind energy x
w
remains
the same.
Definition Let γ
g
denote the average of the absolute value of the gradients of a
thermal plant g ∈ C; then γ
g
denotes the maximum of the absolute gradients of
the thermal plant and is defined by the Run-Up and Run-Down export rates.
In Section 4.1, γ
g
was assumed to be the same for all κ = 15 thermal generators.
Then, for every γ
w
such that γ
w
≤ κγ
g
there exist situations where the output
from thermal plants decreases or increases in a symmetric Λ-shape. Note, that
for the case where generation has to meet the electricity load exactly:
γ
w
=

g∈C
γ
g
(4.3)
i.e. the absolute value of the average gradient of the wind energy equals a sum
of the absolute values of the gradient of κ thermal plants.
We wish to find a formula that calculates the accumulated fuel cost of the
electricity generation where a V-shaped wind energy curve is given. To do this,
let us look into the output of the thermal generators for a series of problems with
the absolute value of gradient of the wind energy curve increasing. After finding
a way every thermal plant responds to the increase of the variability in wind
energy, we can calculate its total generated energy and, hence, the accumulated
fuel cost.
Besides the cases where the output of the thermal generators increases or
decreases in a symmetric Λ-shape, there are also transition cases between the two
adjacent Λ-shapes where the the output of the thermal generators increases or
decreases in a symmetric broken-Λ-shape. Figures 4.5 and 4.6 plot an example of
a symmetric broken-Λ-shape for the generators “Th6” and “Th10” respectively.
Definition Let us denote the active thermal generators as generators that run
away from the upper and lower capacity limits so that an increase or decrease
in their output has a broken-Λ-shape. For an odd number of the thermal plants
there always exists a pair of matching active thermal generators, one of which
increases the output and the other decreases it.
Figure 4.4 plots an example of the solution with the broken-Λ-shapes; Figure
4.5 emphasizes this solution for the thermal generators running close to the upper
capacity limit and Figure 4.6 - for the thermal generators running close to the
lower capacity limit. Thermal generators “Th6” and “Th10” are matching ther-
72
mal generators and change the gradient of the Λ-shape at the same time intervals
t = 30 and t = 150.
0
5
10
15
20
25
30
0 20 40 60 80 100 120 140 160 180
e
n
e
r
g
y
, M
W
*
m
in
time, min
Th1
Th2
Th3
Th4
Th5
Th6
Th7
Th8
Th9
Th10
Th11
Th12
Th13
Th14
Th15
Hydro
Wind
Figure 4.4: An example of the solution with the thermal plants increasing and
decreasing their output in symmetric shapes
27
27.5
28
28.5
29
29.5
30
30.5
0 20 40 60 80 100 120 140 160 180
e
n
e
r
g
y
, M
W
*
m
in
time, min
Th1
Th2
Th3
Th4
Th5
Th6
Th7
Figure 4.5: Seven cheapest thermal plants increasing and decreasing their output
in symmetric shapes
-0.5
0
0.5
1
1.5
2
2.5
3
0 20 40 60 80 100 120 140 160 180
e
n
e
r
g
y
, M
W
*
m
in
time, min
Th9
Th10
Th11
Th12
Th13
Th14
Th15
Wind
Figure 4.6: Seven most expensive thermal plants increasing and decreasing their
output in symmetric shapes
73
A point in time t where a pair of two matching active thermal generators
running at the maximum and minimum production levels modify a gradient of
the Λ-shape is determined by a proportion between their fuel costs and the fuel
cost of the marginal generator:
c
m+i
−c
m
c
m+i
−c
m−j
where c
m+i
is the fuel cost of a thermal generator, higher than that of the fuel cost
of the marginal generator, c
m
, and c
m−j
is the fuel cost of a matching thermal
generator, lower than that of the marginal generator.
Using information on the shapes of the thermal generation, increase or de-
crease in the produced energy can be calculated as the volume of an isosceles
triangle. Then, the accumulated fuel cost, F
A
, for the planning horizon T can be
estimated using Formula (4.4) below. In order to simplify visually further analy-
sis, the right-hand side of Formula (4.4) is divided into three parts: (I), (II) and
(III).
F
A

w
, T) =F
0
−c
m
x
w
+ (I)
c
m
γ
w
T
4
+ (II)
γ

T
κ−1

i=1

g∈C
_
(∆
i+1
g
−∆
i
g
)Tc
m
1 −s
g
4
+ ((∆
i+1
g
)
2
−(∆
i
g
)
2
)c
g
s
g
4
_
(III)
where

i
g
:=
_
¸
¸
_
¸
¸
_
˜ c
i
g
−c
m
˜ c
i
g
−c
i
g
T, c
g
< c
m
T, c
g
= c
m
T(1 −
c
i
g
−c
m
c
i
g
−˜ c
i
g
) =
c
m
−˜ c
i
g
c
i
g
−˜ c
i
g
T, c
g
> c
m
;
(4.4)
F
0
, c
m
and x
w
are defined as in the previous model (4.2);
T denotes a planning horizon with one-minute discrete intervals;
c
i
g
and ˜ c
i
g
are the fuel costs of the paired active thermal generators;
γ

denotes the sum of the absolute values of the average gradient of the thermal
generators g ∈ C as in (4.3);
κ represents a number of thermal generators;
s
g
is a parameter that equals 1 for the active generators g ∈ C running at the
minimum capacity, equals −1 for the active generators g ∈ C running at the
maximum capacity and equals 0 for the marginal generator.
The actual accumulated fuel cost F
A
is significantly close to that calculated
with the right-hand side of (4.4) and weighted by the planning horizon T with
74
respect to the gradient of a V-shaped drop of the wind energy curve. To find how
close the LP-modelled and the calculated accumulated fuel costs are, we wish to
estimate the marginal fuel cost for each of them.
Let us take a discrete interval of the absolute values of the gradient for the
wind energy curve with a discretisation step j.
Definition Let us denote the marginal fuel cost as F
m
. It is calculated as the
difference between the accumulated fuel cost for the adjacent values of the abso-
lute value of the gradient of the wind energy curve weighted by the difference in
the height of the wind energy h
w
.
Then, for every γ
w
, the marginal fuel cost can be estimated by (4.5):
F
m
=
F
A

w
, T) −F
T
A

w
−j, T)
h
w

w
) −h
w

w
−j)
(4.5)
where (γ
w
− j) is the average of the absolute value of the gradients of the wind
energy preceding the current value γ
w
.
Proposition 4.3.1 The marginal fuel cost does not depend on the planning hori-
zon T and is constant for every transition case where the active thermal generators
increase or decrease the output in the broken-Λ-shape.
Proof Let us consider, first, a simple case where γ
w
≤ γ
g
for the planning horizon
t = 1, 2, . . . , T, i.e. the variability of the wind energy can be balanced with the
marginal generator. We wish to calculate the marginal fuel cost by analysing
parts (I) −(III) of Formula 4.4. By construction of the case on the modification
of a height of wind energy curve (introduced at the beginning of Section 4.3.1), the
average wind energy x
w
remains the same through the series of tests. Therefore,
a part (I) of (4.4), calculated for F
A

w
, T) and F
A

w
−j, T) has the same value;
part (I) of (4.4) is set to zero in the numerator of (4.5). A part (III) of (4.4)
consists of the fuel costs for the active generators, however, there are no thermal
plants deviating from the capacity limits when γ
w
≤ γ
g
, thus, a part (III) of
(4.4) equals zero for F
A

w
, T) and F
A

w
−j, T). As parts (I) and (III) are not
used, the numerator of (4.6) is calculated as the difference in parts (II) of (4.4)
for F
T
A

w
) and F
T
A

w
− j). Taking into account a V-shape of the wind energy,
the marginal fuel cost F
m
equals the fuel cost of the marginal generator:
F
m
=
c
m
γ
w
T
4
−c
m

w
−j)
T
4
h
w

w
) −h
w

w
−j)
=
c
m
j
T
4
j
T
4
= c
m
(4.6)
75
Next, we wish to look at the general case of (4.5), again by analysing parts
(I)−(III) of Formula (4.4). As for the simple case, a part (I) of (4.4), calculated
for F
A

w
, T) and F
A

w
−j, T) has the same value, therefore, part (I) of (4.4) is
set to zero in the numerator of (4.5). For every transition case pairs of the active
generators are the same when calculating the accumulated fuel cost, thus, a part
(III) of (4.4) calculated for F
T
A

w
) and F
T
A

w
− j) has the same value. Then
applying the results of (4.6) the marginal fuel cost for a transition case can be
calculated in the following way:
F
m
= c
m
+
γ

T

κ−1
i=1

g∈C
_
(∆
i+1
g
−∆
i
g
)Tc
m
1−s
g
4
+ ((∆
i+1
g
)
2
−(∆
i
g
)
2
)c
g
s
g
4
_
h
w

w
) −h
w

w
−j)

γ

−j
T

κ−1
i=1

g∈C
_
(∆
i+1
g
−∆
i
g
)Tc
m
1−s
g
4
+ ((∆
i+1
g
)
2
−(∆
i
g
)
2
)c
g
s
g
4
_
h
w

w
) −h
w

w
−j)
= c
m
+
j
T

κ−1
i=1

g∈C
_
(∆
i+1
g
−∆
i
g
)Tc
m
1−s
g
4
+ ((∆
i+1
g
)
2
−(∆
i
g
)
2
)c
g
s
g
4
_
j
T
4
= c
m
+

κ−1
i=1

g∈C
_
(∆
i+1
g
−∆
i
g
)Tc
m
(1 −s
g
) + ((∆
i+1
g
)
2
−(∆
i
g
)
2
)c
g
s
g
_
T
2
Using the definition for ∆
i
g
in (4.4) the marginal operational cost is further cal-
culated as
F
m
= c
m
+

g∈C
_
(
c
i+1
g
−c
m
c
i+1
g
−˜ c
i+1
g
T −
c
i
g
−c
m
c
i
g
−˜ c
i
g
T)Tc
m
(1 −s
g
) + ((
c
m
−˜ c
i+1
g
c
i+1
g
−˜ c
i+1
g
T)
2
−(
c
m
−˜ c
i
g
c
i
g
−˜ c
i
g
T)
2
)c
g
s
g
_
T
2
= c
m
+

g∈C
_
(
c
i+1
g
−c
m
c
i+1
g
− ˜ c
i+1
g

c
i
g
−c
m
c
i
g
− ˜ c
i
g
)c
m
(1 −s
g
) + ((
c
m
− ˜ c
i+1
g
c
i+1
g
− ˜ c
i+1
g
)
2
−(
c
m
− ˜ c
i
g
c
i
g
− ˜ c
i
g
)
2
)c
g
s
g
_
where i = 1, 2, . . . , (κ −1). This shows that an additional unit of the calculated
accumulated fuel cost depends only on the combination of the fuel costs of the
thermal generators g ∈ C.
Figure 4.7 demonstrates that the marginal fuel cost is constant for every tran-
sition case where the active thermal generators increase or decrease the output
in the broken-Λ-shape. A length of the interval where the marginal fuel cost is
constant equals to the Maximum Run-Up and Run-Down Export Rates of the
thermal plants, r
+
= r

= 0.125 for our small electricity generation system. No-
tice, that for the first interval, 0 ≤ γ
w
≤ 0.125, marginal fuel cost equals to the
fuel cost of the marginal generator, c
m
= 1.7.
The calculated marginal fuel cost acquires the same value as the LP-modelled,
which is shown in Figure 4.7. This provides an insight into the next element of
76
the statistical model calculating the accumulated fuel cost that would reduce the
error ǫ. We assume further in Section 4.3.2 that another element of the statistical
model includes the height and the average of the absolute value of the gradient
of the wind energy curve and test it through a series of cases.
1.65
1.7
1.75
1.8
1.85
1.9
1.95
2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
m
a
r
g
in
a
l
c
o
s
t
,
p
/
K
W
m
in
average of absolute value of gradient
Comparing calculated and modeled marginal costs
calculated solution
modeled solution
Figure 4.7: Comparing the calculated and the LP-modelled marginal fuel costs
4.3.2 Improving the estimation of the accumulated fuel
cost
Analysis provided in the section 4.3.1 helps us improving the model (4.2) that
evaluates the accumulated fuel cost of the electricity generation. Section 4.3.1
showed that the accumulated fuel cost depends on the absolute value of the
gradient of the wind energy curve, hence, we wish to perform two tests with the
wind energy curves having various profiles similar to the tests performed for the
wind energy curve with a zigzag profile.
a. For a first test the average of the absolute value of gradients of the wind
energy curve γ
w
remains the same while the height and planning horizon
T are modified linearly. This can be achieved by the scaling of the wind
energy curve.
b. For another test the average of the absolute value of gradients of the wind
energy curve γ
w
and the height of the wind energy curve h
w
are changing
while the planning horizon T remains the same.
For the calculation later in this chapter we chose a set of 30 wind energy curves
so that they provide a variety of the statistical parameters such as the mean, the
absolute value of the gradient, the standard deviation and the height.
77
It was noticed that the height of the wind energy and the average of the
absolute value of gradients are linearly dependent and a remaining error of the
model (4.2) plotted against γ
w
has a quadratic form. Figure 4.8 demonstrates
the results of the tests a) and b) performed for the different values of h
w
. In the
test a) only the height of the wind energy drop changes while the average of the
absolute value of the gradients remains constant. So the error ǫ in the model
(4.2) is linear with respect to the changing height h
w
. Test b) is more complex as
two variables: the height and the average value of the absolute gradients of the
wind energy curve are changing. Error ǫ can be plotted as quadratic function as
long as the height and the absolute value of the gradient are modified linearly.
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0 5 10 15 20 25
d
i
f
f
e
r
e
n
c
e

i
n

c
o
s
t
height of wind energy curve, h
Ftot-Fo+1.7Xav
Ftot-Fo+1.7Xav
Ftot-Fo+1.7Xav
0.0155*Gav*h
0.0155*Gav*h
0.0155*Gav*h
Figure 4.8: Fitting a quadratic term into the model that estimates the accumu-
lated fuel cost of electricity generation
Tests a) and b) support a statement that an error in the model (4.2) depends
on a quadratic term containing the average of the absolute value of gradients γ
w
and the height of the wind energy curve h
w
. In Figure 4.8 three functions from the
below use the same quadratic 0.0155γ
w
h
w
to approximate another term included
in the model that estimates the accumulated fuel cost. Hence, an improved model
estimating the accumulated fuel cost of the electricity generation is formulated
further as (4.7).
F
A
= F
0
−c
m
x
w
+ k
1
γ
w
h
w
+ ǫ (4.7)
In (4.7) k
1
≈ 0.0155 is a parameter that can be estimated with the fuel costs of
the thermal plants, it is independent of the wind energy curve.
Let us find how good the model (4.7) is for the estimation of the accumulated
fuel cost of the electricity generation. Figure 4.10 plots the error term of the
78
model (4.2) and compares it to the quadratic term 0.0155γ
w
h
w
calculated for a
set of 30 wind energy curves.
0
1
2
3
4
5
6
7
0 2 4 6 8 10 12 14 16 18
c
o
s
t
,

p
/
u
n
i
t

t
i
m
e
square root of product h*Gav
Approximation of the difference between LP-modeled and calculated cost for zigzag wind curve
Ftot-Fo+1.7*Xav
0.0165*Gav*h
Figure 4.9: Comparison of the error term for the LP-modelled and the calculated
accumulated fuel cost with the model (4.2) for a set of 30 wind energy curves
Figure 4.9 shows that within a range of the wind energy curves with low variability
(represented by a low γ
w
) the model (4.7) can be considered a good estimate of
the accumulated fuel cost of the electricity generation. However, for the more
variable wind energy we wish further to improve the model (4.7).
4.4 Case of the wind energy curve with chang-
ing standard deviation
Let us estimate the error term in the model (4.7) to decide on the elements that
would improve it. We wish to calculate the error for two different cases: for all
the planning horizon T and an accumulated error if calculated for the ten-minute
intervals that a planning horizon T is divided into.
Figure 4.10 plots the error of estimating the accumulated fuel cost calculated
for all the planning horizon T and an accumulated error calculated for
T
10
intervals
of the planning horizon. For both curves in Figure 4.10 the error is relatively
high. However, an accumulated error for
T
10
intervals of the planning horizon
is convenient for further improvement as all the values are positive. An error
calculated for all the planning horizon T has positive and negative values for the
different wind energy curves so that it is hardly possible to improve the result.
79
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0 0.5 1 1.5 2 2.5 3
c
o
s
t

e
r
r
o
r
,

p
/
u
n
i
t

t
i
m
e
standard deviation of gradient
Difference between LP-modeled and calculated cost
whole:Ftot-Fo+1.7Xav-f(G,h)
18int:Ftot-Fo+1.7Xav-f(G,h)
Figure 4.10: Comparison of the error term for the LP-modelled and the calculated
accumulated fuel cost with the model (4.7) for a set of 30 wind energy curves
When plotted along the axis of the standard deviation of the absolute value
of the gradient of the wind energy curve, an accumulated error for
T
10
intervals
of the planning horizon has a trend of a power function. Thus, we wish to use a
standard deviation of the wind energy and a standard deviation of the absolute
value of the gradients of the wind energy to formulate another element in the
model estimating the accumulated fuel cost of the electricity generation. Again
we are looking at two different ways of calculating both values of the standard
deviation: over the whole planning horizon T and accumulated for
T
10
intervals of
the planning horizon T.
An improved model estimating the accumulated fuel cost of the electricity
generation is formulated as follows:
F
A
= F
0
−c
m
x
w
+ k
1
γ
w
h
w
+ k
2
σ
x
σ
α
γ
+ ǫ (4.8)
where σ
x
denotes the standard deviation of the wind energy and σ
γ
is the standard
deviation of absolute value of gradient in a power α = 1.28 for our case. Parameter
k
2
equals 0.05 for our case.
Figure 4.11 plots a comparison of the error term in three models estimating
the accumulated fuel cost of the electricity generation, (4.2), (4.7) and (4.8),
calculated for a set of 30 wind energy curves. The figure shows that the error,
found as the difference between the LP-modelled and the calculated accumulated
fuel cost, decreases with the additional statistical parameters being added to the
model. This is an expected result as the latter models are richer and include the
80
former when setting parameters k
1
:= 0 and/or k2 := 0.
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 5 10 15 20 25 30
e
r
r
o
r
,

p
curve number
Error calculating cost of electricity generation
18int:Ftot-Fo+1.7Xav
18int:Ftot-Fo+1.7Xav-F(h,G)
18int:Ftot-Fo+1.7Xav-F(h,G)-0.05Dx(Dg**1.28)
Figure 4.11: Comparing an error of calculating the accumulated fuel cost of the
electricity generation with the models (4.2), (4.7) and (4.8)
A small error of estimating the accumulated fuel cost when calculated with
the statistical model (4.8) shows that the saving in fuel cost and the additional
fuel cost (introduced in Section 3.3.1 and Section 3.3.2) are determined by the
statistical parameters of the wind energy curve. When wind energy is incorpo-
rated into the electricity generation system, the accumulated fuel cost is reduced
by the amount of wind energy present in the system, however, one-minute fluc-
tuations of the wind speed, described by the absolute value of gradients and the
standard deviation, cause an increase in the accumulated fuel cost.
81
Chapter 5
Stochastic Modelling
Chapters 3 and 4 use deterministic modelling for scheduling the electricity gen-
eration assuming the electricity load is known and all the generating units are
strictly following the operating plan. However, this is a strong assumption in
real-life situations, so this chapter introduces an uncertainty of the electricity
generation. We wish to use stochastic programming as a method of structuring
possible future scenarios and estimating the system balancing cost that appears
when wind energy is not forecast for certain.
The System Operator of the United Kingdom balances the electricity genera-
tion and consumption in the power system every moment of the real time and is
responsible for keeping the system frequency within a target range. In order to
cope with an unexpected mismatch between the electricity supply and demand
the System Operator uses the available Response and Reserve provided by the
generators. As it was specified in Chapter 1, Response and Reserve are being used
on different time scales. Response is an automatic process provided by the large
generating plants that keeps the power system within the target frequency in a
time range from 0 seconds to several minutes. By that time operating Reserve
is expected to increase or decrease the output and support the power system in
minutes or hours.
Response is an automatic process provided by the generators while a decision
about the amount of operating Reserve is made by the System Operator and
depends on possible losses or fluctuations of the load that can happen in the
future. For example, 1GW of Reserve means that some generators run away
from their maximum export limits so that they can increase the power output by
1GW if required by the System Operator. The aim of the stochastic programming
is to formulate an operating plan that keeps feasible possible scenarios of wind
energy output and minimizes the system balancing cost.
82
The stochastic optimization model is formulated using the deterministic model
introduced in Chapter 3 and includes common constraints of the electricity gen-
eration. Our original contribution consists in developing a scenario tree that
reflects two uncertainties: uncertainty in the available capacity of the thermal
plants and uncertainty in wind energy output. We use the formulated stochastic
optimization model to achieve the aim of this thesis by calculating the system
balancing cost when unpredictable wind energy is introduced into the electricity
generation system.
5.1 Uncertainty of the wind energy
A motivation to model the uncertainty of the wind energy comes from the fact
that wind speed is hard to forecast. Mathematical techniques of forecasting an
average wind speed for the next hour or half an hour have improved significantly
during last 30 years ([36, 37, 38, 39, 40]), however, one minute variations as well
as sub-hourly capacity of wind energy are still calculated with possible errors.
Until the short-term prediction of the wind speed is further improved, the avail-
able wind energy creates an additional uncertainty in the electricity generation
system. Currently the installed wind capacity does not exceed 4% of the total
electricity generation, so that the matter of wind speed low predictability can be
resolved without an increase in the operating Reserve. However, if the wind en-
ergy production amounts to 15% and higher, as it is planned by the Renewables
Obligation, then the system balancing cost would significantly increase.
Assuming the uncertainty of the wind energy output, planning the electricity
generation can be described as a sequence of random realisations and decisions of
the System Operator on the modification of the operating plan for the thermal
plants. In its simplest form the discrete stochastic process can be represented as
a scenario tree describing the unfolding of uncertainty over the planning horizon
[28].
A scenario tree consists of nodes and arcs as in the example pictured in Figure
5.1. For our case the nodes represent the states when the information about the
wind energy output is revealed and the operating plan of the thermal generators
is determined to match the electricity load. The arcs do not have a physical
meaning and are used to structure the nodes and form possible scenarios. For
every two nodes connected by an arc the preceding node is named a parent and
the following node is named a child. The current state is taken as the first node
of every scenario, it is named the root of the tree while the last node of a scenario
83
is named a leaf. A path, from the root to a leaf of the scenario tree, represents a
scenario. Thus, node 1 on Figure 5.1 is parent of three nodes-children, 2, 3 and
4, while node 2 is parent of only two nodes, 5 and 6 that are considered leaves
as they do not have any children. An example of a scenario is the set of nodes
{1, 3, 7}.
1.Root
2.Child
5.Leaf
Scenario 1
6.Leaf
Scenario 2
3.Child
7.Leaf
Scenario 3
8.Leaf
Scenario 4
4.Child
9.Leaf
Scenario 5
10.Leaf
Scenario 6
Figure 5.1: An example of a scenario tree
It is important to capture the nonanticipativity property with the structure
of the scenario tree. Nonanticipativity property can be explained in the following
way. At each state the System Operator makes a decision about the operating
Reserve under the uncertainty of the future realisations of the wind energy. This
uncertainty is gradually reduced, since the decision process is being accompanied
by a flow of information, however, at the moment of making a decision, the
System Operator does not have a preference for any of the scenarios so that the
decision is nonanticipative [34]. For example, node 1 of the tree presented in
Figure 5.1 corresponds to the first stage, and the associated decisions at this
node are identical for all six scenarios.
Let us specify further the parameters of the scenario tree. The System Oper-
ator balances the electricity generation and demand every moment of the time,
however, in Chapter 3 we used the discrete time with the smallest interval of 1
minute. The aim of the modelling is to schedule the operating Reserve of gas
and coal plants whose production level can not be modified too frequently. At
the same time, the Reserve has to respond to the changes in the electricity load
as soon as possible, hence one minute is considered the most appropriate fre-
quency for short-term scheduling. We wish to use the discrete time further for
the stochastic programming as well, and incorporate it in the scenario tree so
that a decision about the electricity generation is made every other minute. A
84
length of the Settlement Period (as it was defined in Chapter 1) is 30 minutes.
Denote further a length of the modelling horizon as T
S
, and for the current
problem T
S
= 30.
Let us define an event as an output of the available wind farms in MW every
minute t = 1, 2, . . . , T where T ≥ T
S
is a certain planning horizon. Assume
that wind speed and, therefore, the wind energy output are known at the current
state t = 1 when decision about the generating plan for the conventional plants
is made. It is unknown whether at the minute t +1, t +2, . . . , T
S
the wind speed
remains as it was forecast or a wind speed profile deviates from the forecast or
shifts in time. This means that a number of possible scenarios for the wind energy
profile can be significantly large.
Assume further multistage stochastic programming so that the wind energy
branches several times during the modelling horizon T
S
. Let us branch the sce-
nario tree every 10 minutes so that there are 4 stages in this multistage stochastic
program. Accumulated length of all the stages equals the length of the modelling
interval T
S
so that for every stage s ∈ S := {1, 2, 3, 4} we set:
T
1
= 1, T
2
= 9, T
3
= 10, T
4
= 10 :
4

s=1
T
s
= T
S
= 30
Branching the scenario tree every 10 minutes reflects the uncertainty of the wind
energy and keeps a limit on the number of nodes that can grow dramatically
when branching the scenario tree every minute of the modelling time period.
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
0 5 10 15 20 25 30
e
n
e
r
g
y
,

M
W
time, min
Figure 5.2: Tree of scenarios that represent possible changes in wind speed
85
Branching of the scenario tree is an individual process and depends on the
characteristics of the underlying events. Although variability and capacity of the
wind energy turbines depend on the geographical and weather characteristics of
the site, we wish to outline further some general directives for building a scenario
tree for the wind energy output. Assume there are four main elements of the
available wind energy taken into account when building a scenario tree, namely:
1. the current output from the wind energy turbines;
2. an average wind energy predicted for the next hour;
3. a heuristic distribution of the wind energy; and
4. the preceding values of the wind energy;
The transformation function that allows us calculating the wind power from the
wind speed, provides a useful information for building a scenario tree. Figure 2.1
plots a typical transformation function and shows that the wind energy output
has upper as well as lower bounds. Besides, when the wind speed exceeds 15m/s
the power output flattens at the maximum capacity. Hence, we wish to build a
scenario tree of the wind power rather than the wind speed as it can reduce a
number of nodes in the tree and increase the speed of solving a problem.
We wish a scenario tree to capture the unpredictable nature of the wind power,
however, it does not reflect the variability of the wind power that was estimated
in Chapter 3. Every possible scenario of the wind power output is a piecewise
linear function so that the slope of a linear segment of this function remains the
same for every node.
When calculating the probability distribution of the wind power, bullet points
3 and 4 of the above list can be combined so that the distribution of the wind
energy output is conditional to the preceding values.
P(w
e
t+1
) = P(w
e
t+1
|w
e
t
, w
e
t−1
) (5.1)
The conditional probability distribution (5.1) provides the information on the
variability of the wind power from a given value w
e
t
. While building a scenario
tree we consider three ways of branching, such as two extreme outcomes; two
extreme outcomes balanced by the mean wind power; and five or more branches.
• two extreme outcomes. In this case a wind power value of the parent
node branches into two child nodes so that two new nodes are represented
86
by the highest positive and negative deviation of the wind power from the
current value:
w
e+
t+1
: max(w
e
t+1
−w
e
t
) and w
e−
t+1
: min(w
e
t+1
−w
e
t
) (5.2)
where w
e
t+1
is found with the probability distribution (5.1).
The probability for each of the two child nodes is calculated so that the
stochastic process of the wind power output is preserved. In the case of
two nodes it requires the expectation of the probability distribution (5.1)
to remain the same. If w
e+
t+1
and w
e−
t+1
are new child nodes acquired by
respectively the highest positive and negative deviation of the wind power
then the probability of their outcome can be calculated as follows:
p(w
e−
t+1
) =
w
e+
t+1
−w
e
t+1
w
e+
t+1
−w
e−
t+1
p(w
e+
t+1
) = 1 −p(w
e−
t+1
)
where w
e
t+1
is the expected value of the conditional probability distribution
(5.1).
• two extreme outcomes balanced by the mean wind power. In this
case a wind power value of the parent node branches into three child nodes
two of which are represented by the highest positive and negative deviation
of the wind power from the current value (as in (5.2)) and the third child
node is the expected value of the conditional distribution (5.1).
The probability for each of the three child nodes is calculated so that the
stochastic process of the wind power output is preserved. In the case of three
nodes it requires that the expectation and the variance of the probability
distribution (5.1) remain the same. If σ
w
t+1
denotes the variance of the
probability distribution (5.1) then the probability of the outcomes w
e+
t+1
,
w
e−
t+1
and w
e
t+1
are as follows:
p(w
e+
t+1
) =
σ
w
t+1
(w
e+
t+1
−w
e
t+1
)
2
−(w
e+
t+1
−w
e
t+1
)(w
e−
t+1
−w
e
t+1
)
p(w
e−
t+1
) =
σ
w
t+1
(w
e−
t+1
−w
e
t+1
)
2
−(w
e+
t+1
−w
e
t+1
)(w
e−
t+1
−w
e
t+1
)
p(w
e
t+1
) = 1 −p(w
e+
t+1
) −p(w
e−
t+1
)
• five or more branches. In this case a wind power value of the parent node
87
branches into five or more child nodes two of which are represented by the
highest positive and negative deviation of the wind power from the current
value (as in (5.2)) and the remaining possible outcomes are the mean values
of the equally spaced segments.
A scenario tree can be built by combining different branching methods. For
example, a scenario tree in Figure 5.2 was built with the combination of (T
1
:
T
2
: T
3
: T
4
) := (1 : 5 : 3 : 2). This combination means that at stage s = 1
there is only one node that branches into 5 nodes at stage s = 2. Each of these 5
nodes branches into 3 nodes resulting in 15 nodes at stage s = 3, each of which
branches further into 2 nodes resulting in 30 nodes at stage s = 4.
We tested different combinations of branching methods for building a four-
stage scenario tree by finding a solution of the stochastic linear optimization
model, described further in Section 5.3.1. All the branching methods result in a
feasible solution, hence, we determined the best branching combination with the
accumulated fuel cost over the planning horizon of 5 days. Table 5.1 compares
the accumulated fuel cost weighted by the generated power and calculated for
different branching methods.
Table 5.1: Accumulated fuel cost weighted by the generated power and calculated
for different branching combinations, £/(MW*day)
(T
1
: T
2
: T
3
: T
4
) day 1 day 2 day 3 day 4 day 5 total
(1 : 2 : 2 : 2) 15.28 17.43 16.10 16.81 17.71 83.33
(1 : 3 : 2 : 2) 14.21 15.12 14.55 14.85 15.17 73.90
(1 : 3 : 3 : 2) 13.44 14.36 13.72 13.88 14.43 69.83
(1 : 3 : 3 : 3) 12.66 13.13 12.87 12.36 12.98 64.00
(1 : 5 : 2 : 2) 11.67 11.95 11.20 11.77 12.03 58.62
(1 : 5 : 3 : 2) 11.12 11.51 10.82 10.96 11.84 56.25
(1 : 5 : 5 : 2) 11.56 12.07 11.15 11.65 12.16 58.59
Table 5.1 shows that the accumulated fuel cost calculated for the branching
combination (1 : 5 : 3 : 2) is the lowest, therefore, we wish to use this branching
method further in this chapter for the stochastic optimization modelling. Figure
5.2 plots an example of scenario tree where every scenario represents a possible
output of the wind turbines. This scenario tree is used to solve a single stochas-
tic programming problem of scheduling the electricity generation at the current
minute t of the planning horizon 1, 2, . . . , T such that the resulting generation
prepares the system for a variety of wind power outcomes. After the solution
is obtained, the electricity generation of the root of the scenario tree is fixed at
its optimal value that establishes the final schedule for this minute of the time
88
horizon.
For the next minute t = 1, 2, . . . , T the modelling horizon 1, 2, . . . , T
S
is rolled
forward in time and a new scenario tree is built based on a new value of the
wind power w
e
t
. The next stochastic programming problem is solved using a new
scenario tree and the optimal electricity generation for the root is fixed for the
current minute t. This process of the rolling modelling horizon is repeated so
that the optimal values obtained from each iteration yield the final solution for
every minute of the planning time horizon 1, 2, . . . , (T −T
S
).
5.2 Uncertainty of lost generating capacity
Another uncertainty that the System Operator of the United Kingdom has to take
into account while balancing the electricity demand and generation involves a loss
of the generating capacity. An example of an event captured by this uncertainty
could be a conventional plant tripped and taken out of service. A probability
of such an event is low, however it can happen at any moment and effect the
target frequency with a significant mismatch between the electricity supply and
demand.
As it was specified in Chapter 1, the System Operator can use the automatic
Response and the planned operating Reserve to increase or decrease an active
power in the system and preserve the integrity of the GB Transmission System.
An impact of the generating capacity loss on the power system increases when
there is also the uncertainty of the wind energy output taken into consideration.
Therefore, we wish to build a scenario tree that describes a possible loss of the
capacity and estimate the system balancing cost when resolving both uncertain-
ties by the System operator, loss of the generating capacity and the wind energy
output.
A scenario tree that captures a possible loss of the capacity is built so that it
is compatible with the wind power scenario tree. Every node represents a state
when the information about the available capacity of the conventional plants
enters the system and the System Operator determines whether their operating
plan should be modified. The arcs connect the nodes forming a path from the
root to the leaves of the scenario tree. Modelling time horizon T
S
is discretized
for the scenario tree of the possible capacity loss and, similar to the wind power
scenario tree, equals 30 minutes.
In order to introduce a possible capacity loss of any type of conventional
plant in one scenario tree, we define an event as a size of the electricity load
89
in MW every minute t = 1, 2, . . . , T
S
. The fact that the electricity demand is a
deterministic value in this work allows us interpreting a loss of the capacity as
an increase in the electricity load. Assume that the available generating capacity
and, therefore, the electricity load are known at the current state t = 1 when a
decision is made whether to modify an operating plan of the conventional plants
or not. It is unknown whether the onstream generating capacity is available at
minutes t + 1, t + 2, . . . , T
S
or there is an importing or exporting capacity lost.
On the event tree it can be pictured as a linear change in the forecast demand
for a minute t + 1, t + 2, . . . , T
S
comparing to the previous state. Figure 5.3
plots a scenario tree of the electricity load values where every minute there is an
expected loss of the exporting or importing capacity. Every scenario suggests a
loss of 600MW in the generating capacity and a loss of 100MW in the importing
capacity at a certain minute t.
37800
38000
38200
38400
38600
38800
39000
0 5 10 15 20 25 30
l
o
a
d
,

M
W
*
m
i
n
time, min
Figure 5.3: A scenario tree of the changes in the electricity load equivalent to the
loss of the capacity
A scenario tree plotted on Figure 5.3 presents three possible outcomes every
minute t of the modelling horizon 2, 3, . . . , T
S
: all the planned generating capacity
is available; there is a loss of the generating capacity; there is a loss of the
importing capacity. For every scenario we assume that once the capacity is lost,
the remaining generating units are available until the end of the modelling horizon
T
S
.
We wish to capture the expectation of a loss in the generating capacity every
minute of the modelling horizon. In this case, however, the number of stages when
a new information becomes available is T
S
−1, which is higher than that for the
90
scenario tree of the changes in the wind power where S = 4. In order to formulate
a mathematical model for both uncertainties, a loss of the generating capacity and
the change in the wind power, we assume there are four stages in the scenario tree
of the changes in the electricity load rather than T
S
. This assumption leads to the
situations where parent and child belong to the same stage, they are formulated
in the next section so that the nonanticipativity property is preserved.
The optimal solution at the root of the scenario tree ensures that the electricity
generating system is prepared for a capacity loss at any minute t of the modelling
horizon. This solution is fixed and the modelling horizon is rolled forward so
that the optimal solution can be found for the next minute t = 1, 2, . . . , T. The
optimal solution for the root of the scenario tree which is now a new minute t, is
fixed again. The rolling horizon method is applied until the optimal solution
is found for every minute t of the planning horizon 1, 2, . . . , (T −T
S
).
5.3 Solving a problem with the uncertainty
The uncertainties presented in sections 5.1 and 5.2 are formulated in the stochastic
programming problem later in this section. The output of a number of tests
performed with the stochastic model result in the estimate of the system balancing
cost when the wind power and the available capacity are uncertain.
5.3.1 Decision variables and constraints of the LP stochas-
tic optimization model
To formulate the deterministic equivalent of the multistage stochastic program-
ming problem, let us combine the wind power and the electricity load scenario
trees described in sections 5.1 and 5.2, and enumerate all nodes of the combined
scenario tree. We use a breadth-first search order [28], i.e. start from a root node
corresponding to the first stage.
Definition Let n denote a node of the scenario tree and a
n
denote the direct
ancestor of a node n to capture the dynamics in the stochastic optimization
model.
The root of the scenario tree has index n = 1, so the stage 2 nodes start from index
2. The numbers of children for each node in the event tree may differ, as they
depend on a probability distribution of the appropriate stochastic process and a
choice of the branching. However, the “parent-child” structure of the scenario
tree preserves the nonanticipativity property of the stochastic problem.
91
Let N denote the last node at the stage 4 so that each scenario goes through a
certain number of nodes n ∈ {1, 2, . . . , N}. A number of time intervals associated
with every node n is denoted by T
n
> 0. For the root node, a number of intervals
T
1
= 1 shows that there is only the current minute information available at the
moment of the decision making. Depending on a choice of the branching method
(as described in Section 5.1) nodes of the scenario tree can be grouped by stages
of the scenario tree. Then, for the scenario tree that captures the uncertainty of
the wind power output, T
n
equals a length of the stage s corresponding to node
n. However, the scenario tree of the changes in the electricity load branches every
minute, so a number of intervals T
n
for a node n of this tree varies depending
on a minute when the generating capacity is lost, and ancestor a
n
sometimes
corresponds to the same stage as node n.
Definition Let the index τ denote a minute of node n, such that τ = 0, 1, . . . , T
n
.
If a certain event describes a loss of the generating capacity at time τ = τ

< T
s
of node a
n
then T
a
n
= τ

and T
n
= T
s
−τ

, where s is a stage of node a
n
.
Let G denote the set of all the generators and consist of the union of C, the set
of the conventional generators, with W, the set of the wind generators. The set
of the conventional generators also consists of Coal, Nuclear and Gas generating
plants. The decision variables in the model correspond to the output of each
generating unit in the power system. The decision variables are denoted by
x
nτg
∀n = 1, 2, . . . , N, τ = 0, 1, . . . , T
n
, g ∈ C (5.3)
where x
nτg
is a production level of the thermal unit g during time interval τ of
node n. The wind power generation is determined by the wind power scenario
tree.
x
nτg
= w
nτg
∀n = 1, 2, . . . , N, τ = 1, 2, . . . , T
n
, g ∈ W (5.4)
There are two groups of constraints that can be formulated for the stochastic
model: the constraints applied to every node n = 1, 2, . . . , N and the constraints
that reflect the stochastic properties of the model. First let us formulate the
constraints that connect the nodes into the scenarios and show the “parent-child”
relationship between the nodes.
x
n0g
= x
a
n
T
n
g
∀n = 2, 3, . . . , N, g ∈ G (5.5)
These inter-nodal constraints state that for every generator g the first value of an
output in a node n has to be the same as the last value of its ancestor a
n
. The
92
first value of a node n is described with an index 0, it is fixed by the ancestor
while decisions are made beginning with τ = 1. The last value of a node n is
determined by an index T
n
.
Further there are constraints that are applied to every node n, they are equiv-
alent to the constraints formulated for the deterministic model in Chapter 3.
Planning the production level of the electricity generation units is bounded by
the available resources and technology. Decision variables introduced above have
finite upper and lower bounds representing unit capacity limits of the generation
system.
x
g
≤ x
nτg
≤ x
g
∀n = 1, 2, . . . , N, τ = 1, 2, . . . , T
n
, g ∈ C (5.6)
The parameters x
g
and x
g
denote the lowest and the highest possible output of
unit g respectively.
Further, Run-Up and the Run-Down constraints also applied to the thermal
generation. Following the notation of Chapter 3, let r
1+
and r
2+
denote two
Run-Up rates separated by the “elbow” output level e
+
and r
1−
and r
2−
denote
two Run-Down rates separated by the “elbow” e

output level. Then the corre-
sponding constraints of the possible change in the output of the thermal plants
are formulated as the inequalities:
x
n(τ+1)g

e
+
+ r
2+
−r
1+
e
+
x
nτg
+ r
1+
(5.7)
x
n(τ+1)g
≤ x
nτg
+ r
2+
(5.8)
and
x
n(τ+1)g

x
g
−e

+ r
2−
−r
1−
x
g
−e

x
nτg

r
2−
x
g
−r
1−
e

x
g
−e

(5.9)
x
n(τ+1)g
≥ x
nτg
−r
2−
(5.10)
where n = 1, 2, . . . , N, τ = 0, 1, . . . , T
n
, g ∈ C.
The loading constraint combines different generating units and applied for
every node separately. Denoting by D

the load demand during interval τ of
node n and by ∆
D
a percent mismatch allowed between the electricity load and
the generation, the loading constraint is formulated with the inequalities:
D

(1 −∆
D
) ≤

g∈G
x
nτg
≤ D

(1 + ∆
D
) (5.11)
93
where n = 1, 2, . . . , N, τ = 1, 2, . . . , T
n
.
Similarly to the case of the deterministic problem, the minimization of the
accumulated fuel cost affects the stochastic solution by setting the generation of
the thermal units close to the lower bound of the allowed mismatch between the
electricity generation and the demand D
τn
(1 −∆
D
). This fact increases the risk
of operating at a frequency below the target level. To avoid this we reintroduce
a cost of mismatch between the electricity demand and the generation {ζ
+
v
; ζ

v
}
as it was described in Chapter 3.
Let us introduce further another set of variables:
{d
+
nτv
, d

nτv
} (5.12)
where n = 1, 2, . . . , N, τ = 1, 2, . . . , T
n
, v = 1, 2, . . . , V . These variables denote
the difference between the electricity demand and the generation formulated in
the following way:

g∈G
x
nτg
−D

=
V

v=1
(d
+
nτv
−d

nτv
) (5.13)
where n = 1, 2, . . . , N, τ = 1, 2, . . . , T
n
and V is a number of the penalty cost
bands in the allowed mismatch between the electricity demand and the generation
[(1 −∆
d
)D

; (1 + ∆
d
)D

].
The size of every penalty cost band v = 1, 2, . . . , V is formulated with the
inequalities:
0 ≤ d
+
nτv
≤ d
+
nτv
(5.14)
0 ≤ d

nτv
≤ d

nτv
(5.15)
where d
+
nτv
is an upper bound of the band v when the generation exceeds the
electricity demand and d

nτv
is an upper bound of the band v when the electricity
demand exceeds the generation.
The set of parameters [d
+
nτv
, d

nτv
] depends on the scenario tree of the electricity
load and can be calculated in the following way
d
+
nτv
= ∆
d
∗ D

∗ β
+
v
(5.16)
d

nτv
= ∆
d
∗ D

∗ β

v
(5.17)
where β
±
v
is the part of the allowed mismatch between the electricity demand and
the generation ∆
d
∗ D

that corresponds to the penalty band v = 1, 2, . . . , V .
94
As distinct from the deterministic case there are four penalty cost bands
for the stochastic programming problems. Because of the uncertainty that the
System Operator has to deal with, the amount of the wind power in the electricity
generation system that produced a feasible solution for the deterministic problem
maybe give an infeasible result for the stochastic model. First, three penalty cost
bands are associated with the costs {ζ
+
v
; ζ

v
} and the size {d
+
nτv
; d

nτv
} as in the
deterministic case. The fourth price band has a very high cost and big enough
size associated with it so that the variables (d
+
nτ4
; d

nτ4
) are nonzero only in the
exceptional cases where there is no other way to avoid infeasible solution.
The objective of the stochastic problem is to find a feasible solution {x
nτg
}
such that it satisfies all the constraints (5.2) −(5.17) at the minimal operational
cost, i.e.:
min
x
N

n=1
p
n
(

g∈C
c
g
T
n

τ=1
x
nτg
+
V

v=1
ζ
+
v
T
n

τ=1
d
+
nτv
+
V

v=1
ζ

v
T
n

τ=1
d

nτv
) (5.18)
In the objective function (5.18) the operational cost is calculated for every
node n and includes the fuel cost

g∈C
c
g

T
n
τ=1
x
nτg
and the cost of a mismatch
between the electricity generation and the load

V
v=1
ζ
+
v

T
n
τ=1
d
+
nτv
+

V
v=1
ζ

v

T
n
τ=1
d

nτv
.
It is uncertain if a node n = 1, 2, . . . , N will realise in the future, however, there
is a probability p
n
associated with node n. The probabilities in the objective
function are not those that were formulated for the scenario tree but path prob-
abilities
p
n
= P(n) ∗ P(a
n
) ∗ P(a
a
n
) ∗ · · · ∗ P(1)
where P(n) is a probability formulated for a node n in the scenario tree. This
probability is multiplied by the probabilities of its ancestors to get the path
probability p
n
. Therefore, the objective function (5.18) estimates the expected
operational cost of the electricity generation.
The stochastic LP optimization model formulated in this section includes two
sets of variables {x
nτg
} and {d
nτv
}, where g ∈ C, v = 1, 2, . . . , V n = 1, 2, . . . , N
and , τ = 1, 2, . . . , T
n
. We take κ again as a number of thermal generators
in the set C so that a total number of decision variables in the model equals
κ

N
n=1
T
n
+ V

N
n=1
T
n
. Similarly to the case of the deterministic optimization
model, we wish to reduce a number of variables by grouping thermal generators:
one nuclear, one gas and one coal generators in the set C. Number of nodes N
in the stochastic optimization problem depends on a chosen scenario tree. For
the problem formulated in this section we combine two scenario trees plotted in
95
Figure 5.2 and Figure 5.3. The aggregated tree results in N = 1000 nodes. Then,
taking V = 6 there are 72, 882 decision variables in the stochastic optimization
problem for a single t of the rolling horizon 1, 2, . . . , T. We assume again one
accumulated wind energy curve in the set of wind generators, thus, number of
constraints result in 344, 112 for a single t of the rolling horizon 1, 2, . . . , T.
5.3.2 Formulation of the LP stochastic optimization model
The constraints and the objective function described in Section 5.3.1 are combined
below:
Objective (minimize cost):
min
x
N

n=1
p
n
(

g∈C
c
g
T
n

τ=1
x
nτg
+
V

v=1
ζ
+
v
T
n

τ=1
d
+
nτv
+
V

v=1
ζ

v
T
n

τ=1
d

nτv
)
Subject to:
x
nτg
= w
nτg
∀g ∈ W
x
g
≤ x
nτg
≤ x
g
∀g ∈ C
x
n(τ+1)g

e
+
+ r
2+
−r
1+
e
+
x
nτg
+ r
1+
∀g ∈ C
x
n(τ+1)g
≤ x
nτg
+ r
2+
∀g ∈ C
x
n(τ+1)g

x
g
−e

+ r
2−
−r
1−
x
g
−e

x
nτg

r
2−
x
g
−r
1−
e

x
g
−e

∀g ∈ C
x
n(τ+1)g
≥ x
nτg
−r
2−
∀g ∈ C

g∈G
x
nτg
≤ D

(1 + ∆
D
)

g∈G
x
nτg
≥ D

(1 −∆
D
)

g∈G
x
nτg
−D

=
V

v=1
(d
+
nτv
−d

nτv
)
0 ≤ d
+
nτv
≤ d
+
nτv
∀v = 1, 2, . . . , V
0 ≤ d

nτv
≤ d

nτv
∀v = 1, 2, . . . , V
d
+
nτv
= ∆
d
∗ D

∗ β
+
v
∀v = 1, 2, . . . , V
d

nτv
= ∆
d
∗ D

∗ β

v
∀v = 1, 2, . . . , V
x
n0g
= x
a(n)T
n
g
∀n = 2, 3, . . . , N, g ∈ G
96
where n = 1, 2, . . . , N, τ = 1, 2, . . . , T
n
.
Formulated stochastic optimization problem is linear and, although a number
of decision variables and constraints is high, it can be efficiently solved with a
CPLEX solver. Using version 10.00 of the CPLEX solver CPU time of solving the
optimization problem was 0.54s for a single t of the rolling horizon 1, 2, . . . , T.
However, there is 2.36s of CPU time spent for calculating a new scenario tree
for the stochastic optimization problem. This adds-up to CPU time of 2.8s for a
single iteration or 67.2min for a daily electricity generation problem, or T = 1440.
5.3.3 Output of the stochastic model with the uncertainty
of the wind power and the capacity loss
The decision on the power output of the thermal plants at the current minute t is
based on the expected operational cost of the stochastic problem. The expected
cost depends on the heuristic probability distribution of the wind power illustrated
with the scenario tree. We wish the System Operator to be ready and respond to
the highest fluctuations recorded historically which means that the fluctuations
may be stronger than that happening in real time. To evaluate the operational
cost of the electricity generation during the real time rather than the expected
cost based on the chosen scenario tree we use the decision variables recorded at
the root of the scenario tree while solving the rolling horizon problems.
Let (x
1g
, x
2g
, . . . , x
T−T
S
,g
) be a vector of the real-time decisions after solving
a sequence of T stochastic problems:
x
tg
:= x
(t)
1,1,g
, t = 1, 2, . . . , (T −T
S
)
where x
(t)
1,1,g
is the power output at the root of the stochastic tree for the t-th
problem of the rolling horizon.
In a similar way a mismatch between the electricity generation and the load
is recorded.
d
+
tv
:= d
+(t)
1,1,v
,v = 1, 2, . . . , V
d

tv
:= d
−(t)
1,1,v
,v = 1, 2, . . . , V
where d
+(t)
1,1,v
and d
−(t)
1,1,v
are the positive and the negative mismatch between the
electricity generation and the load respectively at the root of the scenario tree
for the t-th problem of the rolling horizon.
Knowing the fuel cost c
g
of each generating plant g ∈ C and the cost of the
97
mismatch between the electricity generation and the load (ζ
+
v
, ζ

v
), the operational
cost can be calculated for the planning horizon t = 1, 2, . . . (T −T
S
) as follows.
F
stoch
=

g∈C
c
g
T−T
S

t=1
x
tg
+
V

v=1
ζ
+
v
T−T
S

t=1
d
+
tv
+
V

v=1
ζ

v
T−T
S

t=1
d

tv
(5.19)
where {x
tg
, d
+
tv
, d

tv
} is a set of real-time decision variables recorded for (T −T
S
)
stochastic problems.
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
20000
0 20 40 60 80 100 120
p
o
w
e
r

o
u
t
p
u
t
,

M
W
time, min
gas
coal
nuclear
wind
coal limit
nuclear limit
Figure 5.4: An example of the rolling horizon stochastic programming solution
with the uncertainty of the capacity loss when 11% of the wind power is incor-
porated in the total generation
To evaluate the system balancing cost that appears when the wind power is
incorporated into the electricity generation system let us find the operational cost
in (5.19) for two cases:
Case A: there is uncertainty in the available capacity of the thermal
plants but wind energy output is known for the planning
horizon T. A scenario tree for this stochastic problem is
illustrated in Figure 5.3.
Case B: there is uncertainty in the available capacity of the thermal
plants and uncertainty in wind energy output is represented
with the scenario tree. A scenario tree for this stochastic
problem is a combination of two scenario trees illustrated in
Figure 5.2 and Figure 5.3.
98
Figure 5.4 shows an example of the solution for Case A when there is un-
certainty of a possible capacity loss but the wind speed is forecast for certain.
Problem with 11% of wind energy in the total generation is feasible but the oper-
ational cost is high because the Coal and Nuclear generators are taken down from
the maximum capacity levels and the cheaper power is replaced by the more ex-
pensive gas power output. Such a scheduling of the thermal plants ensures that in
the event of capacity loss, the power system remains within the target frequency
until the end of the modelling horizon.
To calculate the operational cost for the Case A there are decision variables
recorded for the solution at the root of the scenario tree:
(x
(a)
1g
, x
(a)
2g
, . . . , x
(a)
T−T
S
,g
)
(d
+(a)
1v
, d
+(a)
2v
, . . . , x
+(a)
T−T
S
,v
)
(d
−(a)
1v
, d
−(a)
2v
, . . . , x
−(a)
T−T
S
,v
)
The above variables are used to calculate the operational cost by applying them
into the right-hand side of (5.19). The operational cost of the electricity gener-
ation when there is only uncertainty in the available thermal capacity but the
wind power output is known, is then the following:
F
(a)
stoch
=

g∈C
c
g
T−T
S

t=1
x
(a)
tg
+
V

v=1
ζ
+
v
T−T
S

t=1
d
+(a)
tv
+
V

v=1
ζ

v
T−T
S

t=1
d
−(a)
tv
If compared with the operational cost F
var
calculated in Section 3.3.2, the
cost of the stochastic problem F
(a)
stoch
is higher because of the thermal generation
running at the level that would allow them responding to a sudden increase or
decrease in load. The scheme of having a number of thermal generators away
from the output limits to secure feasibility of solution is currently being applied
in practise. It is also illustrated with the solution of the stochastic programming
problem. The system balancing cost that appears with the uncertainty of the
available capacity of the thermal generation can be estimated as follows:
f
(a)
= F
(a)
stoch
−F
var
We wish to compare further the results of Case A with that calculated for
Case B when there is also the uncertainty about wind power output added to the
uncertainty of the load. Similarly to Case A, there are decision variables recorded
99
for a rolling horizon stochastic problem of Case B:
(x
(b)
1g
, x
(b)
2g
, . . . , x
(b)
T−T
S
,g
)
(d
+(b)
1v
, d
+(b)
2v
, . . . , x
+(b)
T−T
S
,v
)
(d
−(b)
1v
, d
−(b)
2v
, . . . , x
−(b)
T−T
S
,v
)
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
20000
0 20 40 60 80 100 120
p
o
w
e
r

o
u
t
p
u
t
,

M
W
time, min
gas
coal
nuclear
wind
coal limit
nuclear limit
Figure 5.5: An example of the rolling horizon stochastic programming solution
with the uncertainty of the wind power output and the capacity loss when 7% of
the wind energy is incorporated in the total generation
The above variables are further used to calculate the operational cost of the
electricity generation when there are two uncertainties that the System Operator
faces: uncertainty of capacity loss and uncertainty of wind power output.
F
(b)
stoch
=

g∈C
c
g
T−T
S

t=1
x
(b)
tg
+
V

v=1
ζ
+
v
T−T
S

t=1
d
+(b)
tv
+
V

v=1
ζ

v
T−T
S

t=1
d
−(b)
tv
In Case B the System Operator does not have the information about the
variability of future wind power output as it is not reflected in the scenario tree.
The solution of the stochastic programming problem for Case B is found with
the expectation that wind power output is slowly changing over the modelling
horizon. Therefore, the cost of the uncertainty in the wind power is calculated
as follows:
f
(b)
= F
(b)
stoch
−F
(a)
stoch
+ f
+
w
(5.20)
where f
+
w
is the additional cost of wind power fluctuations.
100
Figure 5.5 plots an example of the rolling horizon stochastic programming
solution when 7% of the wind power is incorporated in the total generation. This
is the highest possible amount of wind energy that can be incorporated into the
electricity generation system. With wind power higher than 7% in the total
generation, the variables of the fourth penalty cost band (d
+
t4
; d

t4
) are given non-
zero values, which means that the mismatch between the electricity generation
and the load is so high that it affects the target frequency.
5.4 The system balancing cost when there is the
uncertainty of the wind power in the elec-
tricity generation system
In order to calculate the system balancing cost of the uncertainty in the wind
and thermal generation, the operational cost of Case A and Case B are estimated
for a day (or 1440 minutes) of electricity generation for different levels of wind
power in the total generation.
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14
o
p
e
r
a
t
i
o
n
a
l

c
o
s
t

w
e
i
g
h
t
e
d

b
y

t
o
t
a
l

g
e
n
e
r
a
t
i
o
n
,

p
o
u
n
d
s
/
M
W
percent of wind power in total generation
zero_wind
flat wind profile
variable wind profile
capacity loss uncertainty
wind capacity uncertainty
Figure 5.6: The fuel cost of the uncertainty in wind power output and the avail-
able thermal capacity, shown for different levels of the wind introduced into the
electricity system
Figure 5.6 plots the fuel costs of a daily electricity generation weighted by the
total amount of the produced power. For up to 13.3% of wind energy in the total
generation, the highest cost is the fuel cost calculated for a problem when there is
no wind energy in the power system. However, when the amount of wind exceeds
101
13.3% in the total generation, the balancing cost of the uncertainty in the wind
power output combined with the uncertainty of the available thermal capacity is
so high that the power system does not benefit from the incorporation of extra
wind into the electricity generation system.
Two optimization models that we developed in this work, deterministic and
stochastic, interact when estimating the system balancing cost of variable and
unpredictable wind energy incorporated into the electricity generation system.
Each of the models covers only one aspect of the wind speed: variability or
uncertainty. In case variable wind energy is also uncertain the system balancing
cost can be estimated as in (5.20). Even at 0% of wind energy integration the
balancing cost of variability is complemented by the cost of uncertainty in loss of
thermal generating capacity.
Even though it is possible to estimate the system balancing cost of variable
and unpredictable wind energy using the results of deterministic and stochastic
optimization models, we wish to emphasize the importance of developing a model
that captures both characteristics of wind. Building a scenario tree that includes
both, variability and uncertainty of wind energy can be suggested as a future
project.
102
Chapter 6
Conclusion
This thesis presented a study of the statistical characteristics of one-minute wind
speed and estimated the system balancing costs when the wind power is incorpo-
rated into the electricity generation system. The problem of the increase in the
system balancing cost arises as a result of variable and unpredictable nature of
the wind speed that transferred into the wind power output. National Grid has to
balance the electricity supply and demand every moment, which leads to different
problems formulated on time scales of a year, hour, one minute or one second. In
this work we were focusing on one-minute electricity generation when the System
Operator makes a decision on the output of the available thermal generators.
This area was not well-researched before but has an important application in the
industry.
We wish to use optimization modelling to estimate the system balancing cost
that requires the appropriate data with one-minute resolution. However, there is
no one-minute wind speed available on the country level. Least frequent data was
presented by ten-minute wind speed published by the Utah Geological Survey,
therefore, we wish to generate unbiased samples from the wind stochastic process
conditional on each consecutive group of wind speeds having a given average
value. To do this, we developed a unique algorithm based on one of the Monte
Carlo Markov Chain methods, Gibbs sampling algorithm. It ensures that as the
number of iterations of sampling from a known distribution increases, the density
of a resulting set of variables converges to the required one.
We originally adapted an algorithm that transfers one-minute wind speed
into the deseasonalised and normalised wind speed, first used by Glasbey et al.
([16, 17]), so that the modified Gibbs sampling algorithm is described as follows:
step A: generate normalised wind speed that further transfered into
original wind speed;
103
step B: if the average of the generated one-minute wind speed does
not equal a given average ten-minute wind speed, it is shifted
along the vector of the maximum probabilities of the prob-
ability distribution;
step C: a new value of the shifted normalised wind speed is found
within a specified tolerance interval of the probability space;
Iterations: steps A to C are repeated a big number of times that ensures
the probability distribution of the resulting set is second-
order stationary process.
Generated samples of one-minute wind speed were further used in determin-
istic and stochastic optimization modelling of the electricity generation system.
Both optimization models include common constraints of the electricity genera-
tion while our original contribution consists in reflecting the characteristics of the
UK power system and developing algorithms that estimate the system balancing
costs of variable and unpredictable wind energy.
The system balancing costs appear when the System Operator uses or keeps
on stand-by flexible but expensive generating units in order to deal with an un-
expected mismatch between the electricity supply and demand. The levels of the
operating reserve required at any given time depend partly on the uncertainties
in the electricity load but also in the available generating capacity.
Electricity supply and demand in the transmission and distribution power
systems need to be balanced every single moment of time. This thesis investigated
the electricity generation on one-minute basis and, therefore, the variability of
the wind energy and the electricity generation with one-minute frequency. The
system balancing cost caused by the fluctuations of the wind energy was estimated
through a series of cases. The power system benefits from the fact that wind power
does not include fuel costs but the operational cost is increased when the thermal
plants modify the power output in order to balance the electricity generation and
the load.
We assumed in this work that the saving in fuel cost and the additional fuel
cost appearing when wind energy is incorporated into the electricity generation
system, can be determined as a function of the statistical parameters of the
wind energy. To demonstrate this, we developed a unique model of the actual
accumulated fuel cost, depending on the mean, absolute value of gradients and
standard deviation of the wind energy curve.
104
In reality the System Operator does not have the information about the vari-
ability or the capacity of the future wind power output. The system balancing
cost of this uncertainty was evaluated using stochastic programming when the
flow of the information and a sequence of the decisions by the System Operator
were illustrated with a scenario tree. Our original contribution consisted in devel-
oping a scenario tree that reflected two uncertainties: uncertainty in the available
capacity and uncertainty in wind energy output. The fact that the determinis-
tic and the stochastic optimization models cover only one aspect of wind speed,
variability or uncertainty, means that they have to interact in order to estimate
the system balancing cost. We wish to suggest exploring different scenario trees
in future work that would allow us to estimate the system balancing cost of vari-
able and unpredictable wind speed in one model. We also leave the matter of
interaction between the electricity demand and wind speed for future research.
In summary, the statistical characteristics of the wind speed have been ex-
plored and the system balancing costs of wind speed variability and unpredictabil-
ity have been estimated. The results showed that depending on the statistical
parameters and the geographic distribution of the wind speed the system bal-
ancing cost of the electricity generation can be significant. However, the recent
developments in construction of wind turbines as well as introduction of smart
metering into the global energy market could open a new area of application for
the results of this work.
105
Appendices
Appendix A. Definitions by the National Grid
NGET is National Grid Electricity Transmission plc (NO: 2366977) whose reg-
istered office is at 1 −3 Strand, London, WC2N5EH.
Supplier is
• (a) A person supplying electricity under an Electricity Supply Licence; or
• (b) A person supplying electricity under exemption under the Act; in each
case acting in its capacity as a supplier of electricity to Customers in Great
Britain.
Customer is a person to whom electrical power is provided (whether or not he
is the same person as the person who provides the electrical power).
Generator is a person who generates electricity under licence or exemption un-
der the Act acting in its capacity as a generator in Great Britain.
Network operator is a person with a User System directly connected to the
GB Transmission System to which Customers and/or Power Stations (not form-
ing part of the User System) are connected, acting in its capacity as an operator
of the User System, but shall not include a person acting in the capacity of an
Externally Interconnected System Operator.
Settlement Period is a period of 30 minutes ending on the hour and half-hour
in each hour during a day.
Gate Closure means, in relation to a Settlement Period, the spot time 1 hour
before the spot time at the start of that Settlement Period.
Operational Day is the the period from 05 : 00 hours on one day to 05 : 00 on
the following day.
Balancing Mechanism (BM) is a period of time which allows the System Oper-
ator to call upon additional generation/consumption or reduce generation/consumption
in order to balance the System minute by minute. From July this period of time
will be one hour before each trading period.
106
BM Unit means a unit established and registered (or to be established and
registered) by a Party in accordance with Section K3 or, where the context so
requires, the Plant and/or Apparatus treated as comprised in or assigned to such
unit for the purposes of the Code.
Physical Notification means, in respect of a Settlement Period and a BM Unit,
a notification made by (or on behalf of) the Lead Party to the Transmission Com-
pany under the Grid Code as to the expected level of Export or Import, as at the
Transmission System Boundary, in the absence of any Acceptances, at all times
during that Settlement Period.
Quiescent Physical Notification is the data that describes the MW levels to
be deducted from the Physical Notification of a BM Unit to determine a resultant
operating level to which the Dynamic Parameters associated with that BM Unit
apply, and the associated times for such MW levels. The MW level of the QPN
must always be set to zero.
Export (Import) Limits is a series of MW figures and associated times, mak-
ing up a profile of the maximum level at which the BM Unit may be exporting
(importing) in MW to the GB Transmission System at the Grid Entry Point or
Grid Supply Point, as appropriate.
Dynamic Parameters comprise
• Up to three Run-Up Rate(s) and up to three Run-Down Rate(s), ex-
pressed in MW/minute and associated Run-Up Elbow(s) and Run-Down
Elbow(s), expressed in MW for output and the same for input. It should
be noted that Run-Up Rate(s) are applicable to a MW figure becoming
more positive;
• Notice to Deviate from Zero (NDZ) output or input, being the notifi-
cation time required for a BM Unit to start importing or exporting energy,
from a zero Physical Notification level as a result of a Bid-Offer Acceptance,
expressed in minutes;
• Notice to Deliver Offers (NTO) and Notice to Deliver Bids (NTB),
expressed in minutes, indicating the notification time required for a BM
Unit to start delivering Offers and Bids respectively from the time that the
Bid-Offer Acceptance is issued. In the case of a BM Unit comprising a
Genset, NTO and NTB will be set to a maximum period of two minutes;
• Minimum Zero Time (MZT), being either the minimum time that a
BM Unit which has been exporting must operate at zero or be importing,
107
before returning to exporting or the minimum time that a BM Unit which
has been importing must operate at zero or be exporting before returning
to importing, as a result of a Bid-Offer Acceptance, expressed in minutes;
• Minimum Non-Zero Time (MNZT), expressed in minutes, being the
minimum time that a BM Unit can operate at a non-zero level as a result
of a Bid-Offer Acceptance;
• Stable Export Limit (SEL) expressed in MW at the Grid Entry Point
or Grid Supply Point, as appropriate, being the minimum value at which
the BM Unit can, under stable conditions, export to the GB Transmission
System;
• Stable Import Limit (SIL) expressed in MW at the Grid Entry Point or
Grid Supply Point, as appropriate, being the minimum value at which the
BM Unit can, under stable conditions, import from the GB Transmission
System;
• Maximum Delivery Volume (MDV), expressed in MWh, being the max-
imum number of MWhr of Offer (or Bid if MDV is negative) that a partic-
ular BM Unit may deliver within the associated Maximum Delivery Period
(MDP), expressed in minutes, being the maximum period over which the
MDV applies.
Target Frequency is that Frequency determined by NGET, in its reasonable
opinion, as the desired operating Frequency of the Total System. This will nor-
mally be 50.00Hz plus or minus 0.05Hz, except in exceptional circumstances as de-
termined by NGET, in its reasonable opinion when this may be 49.90 or 50.10Hz.
An example of exceptional circumstances may be difficulties caused in operating
the System during disputes affecting fuel supplies.
High Frequency response is an automatic reduction in Active Power output
in response to an increase in System Frequency above the Target Frequency (or
such other level of Frequency as may have been agreed in an Ancillary Services
Agreement). This reduction in Active Power output must be in accordance with
the provisions of the relevant Ancillary Services Agreement which will provide
that it will be released increasingly with time over the period 0 to 10 seconds
from the time of the Frequency increase on the basis set out in the Ancillary Ser-
vices Agreement and fully achieved within 10 seconds of the time of the start of
the Frequency increase and it must be sustained at no lesser reduction thereafter.
Primary Response is the automatic increase in Active Power output of a
108
Genset or, as the case may be, the decrease in Active Power Demand in response
to a System Frequency fall. This increase in Active Power output or, as the case
may be, the decrease in Active Power Demand must be in accordance with the
provisions of the relevant Ancillary Services Agreement which will provide that
it will be released increasingly with time over the period 0 to 10 seconds from
the time of the start of the Frequency fall on the basis set out in the Ancillary
Services Agreement and fully available by the latter, and sustainable for at least
a further 20 seconds.
Secondary Response is the automatic increase in Active Power output of a
Genset or, as the case may be, the decrease in Active Power Demand in response
to a System Frequency fall. This increase in Active Power output or, as the case
may be, the decrease in Active Power Demand must be in accordance with the
provisions of the relevant Ancillary Services Agreement which will provide that it
will be fully available by 30 seconds from the time of the start of the Frequency
fall and be sustainable for at least a further 30 minutes.
Operating Reserve is the additional output from Large Power Stations or the
reduction in Demand, which must be realisable in real-time operation to respond
in order to contribute to containing and correcting any System Frequency fall to
an acceptable level in the event of a loss of generation or a loss of import from
an External Interconnection or mismatch between generation and Demand.
Good Industry Practice is the exercise of that degree of skill, diligence, pru-
dence and foresight which would reasonably and ordinarily be expected from a
skilled and experienced operator engaged in the same type of undertaking under
the same or similar circumstances.
109
Bibliography
[1] Government of United Kingdom. White paper on energy. UK, May 2007.
[2] Government of United Kingdom. Renewables Obligation. UK, 2002.
[3] National Grid Code http://www.nationalgrid.com/uk/Electricity Issue 3 Re-
vision 32, 8 December 2008
[4] Balancing and Settlement Code http://www.elexon.co.uk V1.1 Designated
21st March 2001
[5] Balancing Mechanism Reporting System http://www.bmreports.com/ sup-
ported by ELEXON Ltd.
[6] Wind Power in the UK. A guide to the key issues surrounding onshore power
development in the UK. Report of the Sustainable Development Comission
UK, 2005
[7] Quaterly energy prices: December 2008 Department for Business Enterprise
and Regulatory Reform (BERR) http://www.berr.gov.uk
[8] G. Strbac. Quantifying the system costs of additional renewablesin 2020. UK,
October 2002.
[9] G. Strbac. Challenges for market design to facilitate cost effective integration
of renewables: a UK perspective. presentation, Berlin, October 2007.
[10] G.Sinden. Wind power and the UK wind resource. Environmental Change
Institute, University of Oxford, UK, 2005.
[11] G.Sinden. Diversity in the electricity generating sector. Environmental
Change Institute, University of Oxford, UK, 2005.
[12] G.Sinden. Diversified renewable energy portfolios for the UK. Environmental
Change Institute, University of Oxford, UK, 2005.
110
[13] S.Geman and D.Geman. Stochastic relaxation, Gibbs distributions, and the
Bayesian restoration of images. IEEE, 1984.
[14] A.E.Gelfand and A.F.M.Smith. Sampling-based approaches to calculating
marginal densities. Journal of the American Statistical Association, Vol.85,
N410, 1990.
[15] M.A.Tanner and W.H.Wong. The calculation of posterior distributions by
data augmentation. Journal of the American Statistical Association, Vol.82,
N398, 1987.
[16] D.J.Allcroft and C.A.Glasbey. A latent Gaussian Markov random field model
for spatio-temporal rainfall disaggregation. Applied Statistics, 52, 487-498,
2003.
[17] D.J.Allcroft and C.A.Glasbey. Spatial disaggregation of rainfall using a la-
tent Gaussian Markov random field. Proceedings of the 17th International
Workshop on Statistical Modelling, 2002.
[18] M.Durban and C.A.Glasbey. Weather modelling using a multivariate latent
Gaussian Markov random field. Agricultural and Forest Meteorology, 109, 187-
201, 2001.
[19] D.J.Allcroft, M.Durban and C.A.Glasbey. Modelling weather data. SCRI
Annual Report, 192-195, 2001.
[20] British Atmospheric Data Center. Surface wind speed measurements funded
by Natural Environment Research Council. http://mst.nerc.ac.uk/
[21] Utah Geological Survey. Utah’s Anemometer Loan Program.
http://geology.utah.gov/SEP/wind/anemometerdata/index.htm
[22] J.F.Manwell, J.G.McGowan and A.L.Rogers. Wind Energy Explained. The-
ory, Design and Application. John Wiley and Sons, Inc., 2002.
[23] B.Everett, G.Boyle and J.Ramage, editors. Energy systems and sustainabil-
ity. The Open University, UK, 2003.
[24] G.Boyle. Renewable electricity and the grid: the challenge of variability.
Earthscan, UK, 2007.
[25] J.R.Birge and F.Louveaux. Introduction to stochastic programming. Sringer-
Verlag New York, USA, 1997.
111
[26] A.Pr` ekopa. Stochastic Programming. Springer, UK, 1995
[27] P.Kall and S.W.Wallace. Stochastic Programming. Wiley, Chichester etc.,
1994.
[28] S.W.Wallace and W.T.Ziemba. Applications of Stochastic Programming.
SIAM, USA, 2005
[29] D.Dentcheva, R.Gollmer, A. M¨ oller, W.R¨ omisch and R.Schultz. Solving the
Unit Commitment Problem in Power Generation by Primal and Dual Methods.
Preprint, Mathematik, 1996
[30] A.M¨ oller, W.R¨ omisch. A Dual Method for the Unit Commitment Problem.
Preprint Nr 95-1, Humboldt-Universitat Berlin, Institut fur Mathematik, 1995
[31] S.Feltenmark, K.C.Kiwiel, P.Lindberg. Solving Unit Commitment Problems
in Power Production Planning. Operations Research Proceedings, 1996
[32] C.C.Caroe and R.Schultz A two-Stage Stochastic Program for Unit Commit-
ment Under Uncertainty in Hydro-Thermal Power System. Preprint SC 98-11,
Konrad-Zuse-Zentrum fur Informationstechnik,1998
[33] D.Dentcheva and W.R¨ omisch Optimal Power Generation under Uncertainty
via Stochastic Programming. Stochastic Programming Methods and Technical
Applications, 1997
[34] S.D.Flam. Nonanticipativity in Stochastic Programming Journal of Opti-
mization theory and applications, Vol.46, No 1, 1985
[35] National grid electricity transmission. web-site www.nationalgrid.com, 2008.
[36] J.Hardenberg, M.S.Roulston, D.T.Kaplan and L.A.Smith. Using medium-
range weather forecasts to improve the value of wind energy production. Re-
newable Energy, 28:585–602, 2003.
[37] R.W.Katz and A.H.Murphy, editors. Economic value of weather and climate
forecasts. Cambridge University Press, UK, 2003.
[38] A.Costa, A.Crespo, J.Navarro, G.Lizcano, H.Madsen, E.Feitosa. A review
of the young history of the wind power short-term prediction Renewable and
Sustainable Energy Reviews, 12 (2008) 1725-1744
112
[39] P.R.J.Campbell and K.Adamson. A novel approach to wind forecasting in
the United Kingdom and Ireland International Journal of Simulation: Systems,
Science and Technology, Vol. 6, No 12-13, 2005
[40] U.Schlink and G.Tetzlaff. Wind speed forecasting from 1 to 30 minutes.
Theoretical and Applied Climatology, No 60, 1998
[41] Report from the Environmental Audit Committee Keeping the lights on:
Nuclear, Renewables and Climate Change Sixth report Session 2005-2006
[42] E.H.Huitema. The analysis of Covariance and Alternatives. John Wiley and
Sons, Inc., USA, 1980.
[43] C.Chatfield. The analysis of Time Series: an Introduction. 6th edn., Chap-
man and Hall, London, UK, 2003.
[44] H.Schneider. Truncated and Cencored samples from Normal Populations
Marcel Dekker, Inc., New York, USA, 1986.
[45] A.C.Johnson, N.T.Thomopoulos. Characteristics and tables of the left-
truncated normal distribution Midwest Decision Sciences Institute, 2004 An-
nual Conference Proceedings, Cleveland, April 2004
[46] B.Flury. A first course in Multivariate statistics Springer-Verlag New York,
USA, 1997.
[47] R.Fourer, D.M.Gay, B.W.Kernighan. AMPL: A modelling language for
mathematical programming Springer-Verlag New York, USA, 2004.
113

Contents
Notation 1 Introduction 1.1 Benefits of wind energy . . 1.2 Challenges of wind energy 1.3 Balancing of the electricity 1.4 Timescale . . . . . . . . . 1.5 Aim of the research . . . . 1.6 Structure of thesis . . . . 1 4 5 5 6 9 10 11 13 13 14 15 18 19 23 27 30 31 35 38 43 46

. . . . . . . . . . system . . . . . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

2 Wind Speed Analysis 2.1 Generating wind energy . . . . . . . . . . . . . . 2.2 Generating wind speed data . . . . . . . . . . . . 2.2.1 One-minute wind speed data . . . . . . . . 2.2.2 Methods of generating missing data . . . . 2.2.3 Description of Gibbs sampling . . . . . . . 2.2.4 Transformation to normal variable . . . . 2.2.5 Suggested modifications to Gibbs sampler 2.3 Implementation of modified Gibbs sampler . . . . 2.3.1 Multivariate normal distribution . . . . . . 2.3.2 Two-variable sampling . . . . . . . . . . . 2.3.3 Multivariate sampling . . . . . . . . . . . 2.3.4 Testing the results of wind speed sampling 2.4 Diversification of wind energy . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

3 Modelling Electricity Generation with Wind Energy 48 3.1 Elements of the British power system and simplifying assumptions 48 3.1.1 Electricity generation . . . . . . . . . . . . . . . . . . . . . 49 3.1.2 Cost of generation . . . . . . . . . . . . . . . . . . . . . . 51 3.1.3 Electricity demand . . . . . . . . . . . . . . . . . . . . . . 52 i

3.2

3.3

3.1.4 Cost of demand-and-supply mismatch . . . . . . . . . . . . 3.1.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . Deterministic model for the electricity generation with wind energy 3.2.1 Decision variables . . . . . . . . . . . . . . . . . . . . . . . 3.2.2 Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.3 Objective function . . . . . . . . . . . . . . . . . . . . . . 3.2.4 Formulation of the LP deterministic optimization model . Calculating cost of wind energy variability . . . . . . . . . . . . . 3.3.1 Case of constant wind energy that reduces the operational cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.2 Case of variable wind energy that increases operational cost

53 54 55 55 55 59 60 60 61 63 67 67 69 70 71 77 79

4 Statistical analysis of operational cost 4.1 Formulation of the problem . . . . . . . . . . . . . . . . . . . . . 4.2 Case of marginal generator following a profile of the wind energy curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Case of wind energy curve having a zigzag profile . . . . . . . . . 4.3.1 Analytical calculation of the accumulated fuel cost . . . . 4.3.2 Improving the estimation of the accumulated fuel cost . . . 4.4 Case of the wind energy curve with changing standard deviation .

5 Stochastic Modelling 82 5.1 Uncertainty of the wind energy . . . . . . . . . . . . . . . . . . . 83 5.2 Uncertainty of lost generating capacity . . . . . . . . . . . . . . . 89 5.3 Solving a problem with the uncertainty . . . . . . . . . . . . . . . 91 5.3.1 Decision variables and constraints of the LP stochastic optimization model . . . . . . . . . . . . . . . . . . . . . . . 91 5.3.2 Formulation of the LP stochastic optimization model . . . 96 5.3.3 Output of the stochastic model with the uncertainty of the wind power and the capacity loss . . . . . . . . . . . . . . 97 5.4 The system balancing cost when there is the uncertainty of the wind power in the electricity generation system . . . . . . . . . . 101 6 Conclusion 103

Appendices 106 A. Definitions by the National Grid . . . . . . . . . . . . . . . . . 106 Bibliography ii 110

. . . . . . . . . . . . .12 2. . . . . .2001 .1 Transformation function between the wind speed and the wind power 14 Autocorrelation of the original and the deseasonalised wind speed data . . . . . .8 2. 32 Contours of Normal distribution and four linear functions used in the modification to Gibbs algorithm . . .1 1. . . . . . . . . . . . .11 2. . . . . . . 47 Half-an-hour average of the daily electricity demand in the United Kingdom and its piecewise linear approximation with 1-minute frequency . . . . . . . . . . .4 2. 6 7 10 2. . . . .9 2. . . . . . . . . . 42 An example of the sampled and the original values of 1-minute wind speed plotted along with the ten minute average wind speed 44 Autocovariance of the sampled and the original normalised 1-minute wind speed . . .10 2. . . . . .List of Figures 1. . . . . . . . . . .5 2. . . . 36 Amount of the information carried by each of the ten variables . .6 2. . . . . . . . . . . . . . . . . . . .2 2. . . . . . .1 2. . . . .2 1. 23 Truncated and full plot of normal probability density function . . . . Process of balancing power system shown in terms of the Settlement Period . . .3 Schematic description of the power system . . .3 2. . . . . . . . . . . . . . 24 Normal variable plot with respect to the deseasonalised wind speed 26 Fitted sum of exponential functions to approximate the autocovariance of the normalised wind . . Wind energy that would have been generated by a small wind farm in Aberystwyth on the 5th of January. . . . . .7 2. . . .13 3. . . . . . . . . . . . . 46 An example of geographically distributed wind energy output . . . . . . . . . . 45 An example of wind speed sampled for one of the Utah sites . . . . . . . . . . . . . . . . . . . . . . 18 Schematic description of Gibbs sampler as a Markov-chain process 21 Cumulative probability distribution of one-minute wind speed recorded with MST Radar located at Frongoch farm . . . . . . . . . . 52 iii . .

. . . . . . . .4 3. . . . . 4. . . . . . . . . . . (4. . . . 53 56 57 62 63 65 65 Electricity generation with only the marginal generator modifying the power output . . Electricity generation and load mismatch when 5% of variable wind energy is introduced to the system . . . . . . . . . . . 4. Load mismatch costs: the difference between the curves gives the penalty for generation deviating from demand. . . . . . . . . . .3. . . . . . . . . .2). . . . . . . . . . . .7 3. . . . . . . . . . .5 Seven cheapest thermal plants increasing and decreasing their output in symmetric shapes .2 3. . . . . . . . . . . . . . . . . . . . . . . . . .11 Comparing an error of calculating the accumulated fuel cost of the electricity generation with the models (4. . . . .10 Comparison of the error term for the LP-modelled and the calculated accumulated fuel cost with the model (4. . . 4. Electricity generation when 11% of wind energy with flat profile is introduced to the system . . . . 4. . . . . . . . . .9 Comparison of the error term for the LP-modelled and the calculated accumulated fuel cost with the model (4. . . . . . . . . . . . .4 An example of the solution with the thermal plants increasing and decreasing their output in symmetric shapes . . . . . . . The fuel cost of the wind energy variability shown for different levels of it introduced into the electricity system . . . . Electricity generation when 11% of variable wind energy is introduced to the system . . . . . . 4. . . . . . . . . . . . . . . . 4. . . . . iv 4. . . . . . . . . . . . . . . .6 3. . . . . . . . . . . .1 69 70 71 73 73 73 77 78 79 80 81 . . . . . . . . . . . . . . . . . . . . . . . . 4. . . . .3 3.2 Electricity generation with the wind energy profile as a zigzag . . . . . . 4.7) and (4. . . . . . . . . . . . .7 Comparing the calculated and the LP-modelled marginal fuel costs 4. . . . . . . . . . . . .6 Seven most expensive thermal plants increasing and decreasing their output in symmetric shapes . . . . . . . . . Area formed by Run-Up and Run-Down constraints for thermal generators . . . . . . . . . . .8 Fitting a quadratic term into the model that estimates the accumulated fuel cost of electricity generation . . . . . . . . . 4. . . . . . . . . .8) . . .5 3. . .3 Electricity generation for a wind energy curve as one zigzag and the thermal plants increasing and decreasing their output in a symmetric Λ-shape . . . . . .7) for a set of 30 wind energy curves . .2) for a set of 30 wind energy curves . . . . . . . . . . . .8 A mismatch between the electricity demand and the generation with the cost attached . . . .

. . .2 5. A scenario tree of the changes in the electricity load equivalent to the loss of the capacity . . . . . . . . . . . Tree of scenarios that represent possible changes in wind speed . . . . . . . . . . . . . . . . . . . . . . . . An example of the rolling horizon stochastic programming solution with the uncertainty of the capacity loss when 11% of the wind power is incorporated in the total generation . . .3 5.4 5. .1 5. . 84 85 90 98 100 101 v . . . An example of the rolling horizon stochastic programming solution with the uncertainty of the wind power output and the capacity loss when 7% of the wind energy is incorporated in the total generation The fuel cost of the uncertainty in wind power output and the available thermal capacity. . . . . . . . . . . . . . . . . . . . shown for different levels of the wind introduced into the electricity system . . .5. .6 An example of a scenario tree .5 5. . . . . .

. . . . . .1 88 vi . . . . . £/(MW*day) . . . . An example of the Run-Up and Run-Down export rates for a typical coal plant . . . . .List of Tables 2. . .1 2. . . . . . Plant/Demand balance for SYS Background (2007/2008) . .2 Covariance matrix ϕ calculated for the normal variables transformed from the minute-by-minute wind speed data . . . . Accumulated fuel cost weighted by the generated power and calculated for different branching combinations. . . . . . . . . . . . . . . .2 2.3 3. . . . . . . . . 40 40 45 50 50 5. M-matrix such that M M T = ϕ . . . . . .1 3. Variance of the original and the sampled normalised wind speed . . . . .

a linear mathematical optimization model for a problem of short-term electricity generation is presented to calculate an additional balancing cost that appears as a result of wind energy variability. The unpredictable characteristic of wind speed is analysed with the techniques of stochastic programming. Uncertainty of the expected wind speed is represented through scenario trees and stochastic linear optimization models are used to calculate the extra cost due to uncertainty. It is then shown that this additional balancing cost can be estimated using the statistical characteristics of wind energy present in the system. therefore it is necessary to study wind speed on a one-minute time scale.Abstract Incorporation of wind energy into the electricity generation system requires a detailed analysis of wind speed in order to minimize system balancing cost and avoid a significant mismatch between supply and demand. We apply a modified Gibbs sampling algorithm to generate one-minute wind speed required for optimization modelling from the available ten-minute wind speed. we examine the statistical characteristics of one-minute average values of wind speed. In this thesis. Power generation and consumption in the electricity networks have to be balanced every minute. One-minute wind speed is available from a single site in Great Britain while there are records of ten-minute wind speed available. Further. . Alternative optimization models are compared by calculating the additional balancing cost and the extent of imbalance between power generation and consumption in the system. System balancing costs are estimated through optimization modelling of the short-term electricity generation with wind energy contributing to the total supply. Two main drivers of additional system cost caused by wind energy are variability and unpredictability of one-minute wind speed.

Notation G C W T TΥ TS Tn N S V α t. m l. d. τ. Sj z. ν. sg . L i. zr θ. k g n an pn κ + fw − fw Si . δ. j. k2 λ β± ǫ a set of all generators a set of conventional generators a set of wind generators a number of time periods a number of time periods in one year a modelling horizon a number of intervals associated with node a number of nodes a number of scenarios a number of intervals of mismatch between the electricity supply and demand power law exponent a single period of time a lag an index an index of generator a node an ancestor of node a path probability a number of thermal generators saving in fuel cost additional fuel cost a set of indices height where wind speed was recorded parameters a parameter of Lagrangian function percent of cost band an error term 1 . q. k1 .

Xk s s s W1 .... ....b1 . cgas cm ± ± ± ζ1 ...... Wk y1 .. Uk X1 . Y2 . uk x1 ..... W2 . . Yk U1 .. xk s s s w1 . W2 ... aTΥ ± ± ± r 1 .. . wk parameters of quadratic function parameter of an exponential curve deseasonal scaling curve Run-Up and Run-Down export rates “elbows” of Run-Up and Run-Down export rates the electricity demand fuel costs fuel cost of the marginal generator cost of imbalance between the electricity supply and demand a weight in calculation of deseasonal scaling curve a percent mismatch allowed between the electricity supply and demand height of wind energy curve the average of the absolute value of gradients of wind energy curve the average of the absolute value of gradients of thermal plants maximum of the absolute gradients of thermal plants sum of the absolute value of gradients of thermal plants standard deviation of wind energy standard deviation of the absolute value of gradients mean value standard deviation covariance matrix an inverse to covariance matrix a matrix such that M M T = ϕ vector consisting of scalers 1 only a vector of normal variable (normally includes 10 values) a vector of a standard normal variable (normally includes 10 values) a vector of decision variable (normally includes 10 values) a vector of wind speed (normally includes 10 values) a vector of deseasonalised wind speed (normally includes 10 values) a vector of wind energy (normally includes 10 values) one-minute normal variable one-minute standard normal variable one-minute decision variable one-minute wind speed one-minute deseasonalised wind speed one-minute wind energy 2 . . Wk e e e W1 .. ζ2 . x2 . U2 . .. . ... D2 . w2 . ςi a1 . ... . w2 .. W2 . ζV kdm ∆D hw γw γg γg γ∗ σx σγ µ σ ϕ ψ M 1 Y1 .. wk e e e w1 . ... . ccoal . ... w2 ..... a2 ... b2 .. y2 . . Wk d d d W1 . r2 . yk u1 .. b3 βi .. X2 . wk d d d w1 . r3 e± . u2 . e± 1 2 D1 .. . . DT cg ....

[Y ] joint.. Y D vectors y ˆ a value of normal variable f (∗). d± . aν ) function transferring original wind speed into deseasonal d Θn (W ) function transferring deseasonalised wind speed into normal variable f var F0 . d± 1 2 V W s ... ws xg . Fw accumulated fuel cost of the deterministic optimization model (a) (b) Fstoch . Y C .d± . h(∗) a function P (∗) a probability function φ(∗) probability density function Φ(∗) cumulative distribution function Θd (W s . Fstoch accumulated fuel cost of the stochastic optimization model FA actual accumulated fuel cost Fm marginal fuel cost [Y1 . xg mismatch between the electricity supply and demand the average of wind speed over a period of time maximal and minimal capacity limits d± . Fw lat . [Y1 |Y2 ]. d± upper and lower limits of mismatch intervals Y A . Y2 ]. . conditional and marginal densities 3 . Y B .

Chapter 1 Introduction
How expensive is the operating of the electricity generation system when wind energy is introduced into it? This issue is becoming important as the amount of renewables in the system increases in order to meet the target that the Government of United Kingdom has set. In 2002 the Government introduced the Renewable Obligation (RO) to encourage development of renewable sources of energy[1, 2] and aimed to reach 15.4% of renewables in the overall electricity supply by 2015. Generators receive an RO Certificate (ROC) for each 1MWh of renewable electricity they generate. These are sold to electricity suppliers (defined in Appendix A) that pay the fixed penalty for each MWh it falls short of its obligation[1]. Various renewable technologies can contribute to the goal set by the Government but wind resource takes a leading competitive position among other renewable sources of energy (solar, tidal, wave, hydro) by the availability and efficiency. However, using wind for the electricity generation introduces a problem of intermittency and unpredictability of wind energy that will be studied in this work. The problem of meeting a target set by the Government of United Kingdom was investigated by the researchers of the Environmental Change Institute at Oxford University. In the report “Wind power and the UK wind resource”[10], Sinden states that in order to integrate wind power into electricity networks it is essential to understand the characteristics of wind resources. He believes variability of the wind energy is one of the obstacles to its extensive use by an electricity supplier and suggests a diversified renewable energy portfolio that includes wave and tidal steam combined with wind power. Such a combination reduces effect of variability in wind speed and consequently the cost of balancing electricity supply and demand.

4

1.1

Benefits of wind energy

In the 21st century the use of modern wind turbines for electricity generation has been growing very rapidly in Europe and this growth is expected to continue. Integration of wind generation in the operation and development of the electricity system is associated with both benefits and costs. The following reasons create plenty of opportunities to realise the technical potential of renewable energy, particularly wind power: • Wind turbines produce green energy unlike conventional generators that produce CO2 emissions; • Renewable sources of energy do not have fuel costs that contribute the most to the expenses of the conventional generators; • In spite of high percentages of natural gas and coal in British electricity generation, these resources are limited and their long-term supply can be disrupted. Meanwhile wind, tides, solar are known as renewable sources of energy.

1.2

Challenges of wind energy

However desirable the incorporation of wind energy into the electricity system there are a number of issues associated with its installation and management of the system. Detailed estimation of additional costs associated with the presence of the wind energy in the electricity generation system was performed by a few researchers since 2002 ([8, 9, 10, 11, 12]). Strbac et al. wrote a detailed paper on “Quantifying the system costs of additional renewables by 2020”[8] that examined the issue of challenges that incorporation of additional renewable sources faces. Besides the high investment cost of renewable sources of electricity, there is additional cost that occurs when balancing intermittent and variable wind energy. The work of Strbac et al. shows that in the current system wind energy is not able to replace fully the capacity and flexibility of conventional generating plants. This thesis concentrates on the system costs associated with the balancing of supply and demand, however, there are also system costs appearing in transmission and distribution of electricity. Strbac et al.[8, 9] state different system cost drivers, such as location of the generation sources in relation to the source of demand, lack of flexibility in generation as well as intermittency and variability 5

of supply. Moreover, wind turbines are generally not able to provide the range of system support services (eg. voltage and frequency regulation) that are provided by conventional thermal and hydro plant. At relatively low levels of penetration this can usually be tolerated, but at the higher levels indicated by the Renewables Obligation target, it will require systematic solutions in order to maintain stability and integrity of the transmission system. These issues reduce the value of wind generation that displaces conventional generation.

1.3

Balancing of the electricity system

Electricity generation and delivery to the customers is a complex process. It is described in detail by the National Grid and Balancing and Settlement Codes[3, 4] that we refer to while formulating a problem of scheduling the electricity generation. The main groups included into the electricity generation and distribution system are presented on Figure 1.1.

Figure 1.1: Schematic description of the power system

Most of the electricity in the power system is contracted. Some of the contracts are signed directly between the Generating companies and large industrial consumers. Residential consumers pay for their electricity through the Suppliers, companies like EDF, Scottish power or npower. Suppliers sign agreements to deliver contracted amounts of power to the Consumers. At the same time the Generators have an obligation to provide this power to the transmission system. 6

availability and other. This information would enable NGET to assess which BM Units are expected to be operating in order that NGET can ensure (so far as possible) the integrity of the Great Britain Transmission System. Pre Gate closure procedures are intended for the submission of Balancing Mechanism (BM) Unit Data (as defined in Appendix A) which includes characteristics of each generator such as maximum and minimum power output.This section explains general terms of the system functioning while formal definitions of what these participants represent are given in Appendix A. The interaction process of different participants in the electricity generation.2). collects the necessary data from the appointed parties. and the security and quality of supply [4]. Gate Closure takes place one hour before the beginning of the Settlement Period.2: Process of balancing power system shown in terms of the Settlement Period In order to manage the electricity transmission system efficiently National Grid Electricity Transmission (NGET). a half-hour period of time when real-time balancing of the electricity supply and demand takes place (Figure 1. Let us give a general description of this process in terms of a Settlement Period. Figure 1. transmission and consumption is complex and contains many procedures. company that stands for the System Operator. Data for any Operational Day may be submitted to NGET up to several 7 . From the information provided by BM Units the System Operator has an insight into how much it can vary their generation above or below their submitted profile and what price must be paid for every possible amount of deviation from the planned generation. ramp up rates.

Gas turbines are able to provide generation on timescale of few minutes while steam plant would require a few hours to warm 8 . to neutralise rapid variations from a few seconds to a few minutes. The frequency is allowed to vary 1% up or down from 50Hz. or if no data has been submitted by 11:00 hours on that day by the default data values. The System Operator balances the supply and the demand in the system every moment of the real time through different procedures and is responsible for keeping the system frequency in the target range. The individual data items that Generators and Suppliers submit include Physical Notifications. Electricity transmission and distribution system of Great Britain operates with a certain Target Frequency.days in advance of the day to which it applies. and operating Reserve to deal with slow variations over time horizons from several minutes to several hours [24]. NGET as the System Operator has the responsibility of providing the information on the Demand Estimates. as well as System Margins. Quiescent Physical Notifications. By that time the operating Reserve is ready to change the Export (Import) level (as defined in Appendix A). Export and Import Limits. As long as the amount of produced electricity equals the demand. If the the generation exceeds demand frequency would rise above 50Hz and if demand exceeds generation then frequency in the system drops. It supports the system from the first second up to 30 minutes after the imbalance was recorded. Consistency and Defaulting Rules. as provided in the Data Validation. Pumped storage. the System Operator requires an appropriate automatic Response. Some of the generating units acting as the Reserve are operating at part-time load. On the basis of historic trends in the availability of generating capacity and the data submitted by the generating units NGET also decides on the necessary Contingency and Operating Reserve. In order to deal with unpredicted variations in demand and generation. for example. Response is defined as an automatic change in active power output by the large generating units that effects the frequency in the system. can respond very rapidly to counteract any loss of generation or surge in demand. some are off-line but able to start within a short time. frequency in the transmission system stays at 50Hz. Response is used as an earliest available instrument of the balancing mechanism. Bid-Offer Data and Dynamic Parameters (all defined in Appendix A). The data to be used by NGET for operational planning will be determined from the most recent data that has been received by NGET by 11:00 hours on the day before the Operational Day to which the data applies.

4 Timescale Electricity supply and demand in the transmission and distribution power systems need to be balanced every single moment of time. The participants need to ensure that each of their BM units. or of Sizewell B nuclear power station (1320 MW). follows the submitted notification of export or import level. At this short scale there can be no decision made on the start-up or maintenance of thermal plants. the System Operator takes into account the uncertainties in demand and generation on various timescales. 31. 32. The levels of reserve required at any given time depend partly on uncertainties in the predictions of demand. mixed-integer programming problems. scheduling of electricity generation reserve on hourly and daily basis has been extensively studied by various researchers. either due to a power station faults or the loss of transmission circuits. we avoid the complexity of mixed-integer programming problems. Uncer9 . 33]). 30. and methods of solving these models were analysed by R¨misch. 1. mixed-integer programming problems were extended by capturing the uncertainty in thermal power systems and the stochastic mixed-integer programming was applied to solve complex models ([32.up. Hourly and daily production planning of power systems involves start-up and maintenance of the thermal plants. The value of possible loss of generating capacity can significantly increase after wind energy is introduced into the power system. We wish to study scheduling of the electricity generation at time intervals that were not well-researched before but have an important application in the industry. therefore. At the same time. When scheduling the operating Reserve. Feltenmark and other researchers ([29. but also on the need to deal with the sudden loss of substantial amounts of generation. 31]) in the 1990s. for example. Balancing costs appear when the System Operator uses or keeps on standby flexible but expensive generating units in order to deal with an unexpected mismatch between the electricity supply and demand or operates generators at non-optimal generation levels so as to provide flexibility for rapid change in output level. key criteria are possible loss of one circuit of the cross-Channel link (1000 MW). On the scale of seconds it is managed through the automated process of frequency response. 30. This work focuses on short-term variability of the wind energy and the electricity generation with one-minute frequency. o o Dentcheva. This leads to dynamic. applying Good Industry Practise. In England and Wales. 33]. Later. for example in [29. M¨ller.

On the other hand consumers receive uninterrupted electricity flow in their houses and offices.3: Wind energy that would have been generated by a small wind farm in Aberystwyth on the 5th of January. but broadly speaking. for example. Another solution suggested by Sinden in his work on diversified renewable energy portfolios[12].MW 5 4 3 2 1 100 120 140 time.min 160 180 200 Figure 1.1) can benefit from the efficient operation and planning of electric power generation systems. the costs of the appropriate Reserve decrease as the time scale over which they have to respond increases. This thesis is aiming at the vertically integrated utilities (like Scottish Power or EDF) that are capable of combining 1 The MST Radar Facility at Aberystwyth is funded by the UK Natural Environment Research Council and the data are provided through the British Atmospheric Data Centre 10 . but fast Response plant may cost up to £5/MWh.Januray. On the one hand. was to combine wind energy with more predictable tidal and wave energy.5. or more[24]. One possible solution for this problem involves diversifying locations of wind farms. suppliers of energy reduce the cost of mismatch between the reported and actual electricity load.3 plots an example of the wind power that would have been generated by a small wind farm in Aberystwyth based on the recorded wind speed. Operating Reserve (as defined in Appendix A). Wind energy that one wind farm would generate at Aberystwyth in 2001.2001 1.5 Aim of the research Different participants of the electricity industry (Figure 1. It varies significantly from minute to minute1 which makes variability of wind energy one of the strongest drivers of the balancing cost.tainty increases with the time horizon. Figure 1. may cost around £1/MWh.01:40-03:20 8 wind energy curve 7 6 energy.

11 .6 Structure of thesis In Chapter 2 we investigate how the wind speed varies and how this affects the wind power generation. Obtained in Chapter 2. thus. By setting different parameters of the deterministic model. wind power output is used later in the thesis when testing the optimization model of the electricity generation. however the System Operator has to ensure the integrity of the Great Britain Transmission System while minimizing the balancing costs. Chapter 5 investigates into the additional system balancing cost that comes from coping with the unpredictable nature of wind speed. so the costs of incorporating the wind power into the electricity generation system originate only in variability of wind speed. NGET acting as the System Operator manages the electricity generation system of the United Kingdom. minimize their costs of meeting the contracted demand. The additional system balancing cost that appears after the wind power is introduced into the electricity generation system can be modelled as a function depending on the statistical parameters of one-minute wind power output. There is only one publicly available site of wind speed one-minute averages. There is a number of tests performed in Chapter 4 that derives a model approximating the operational cost of the electricity generation. The problem of electricity generation is then solved using stochastic programming methods where the uncertainty is represented with a scenario tree. In practise it is not possible to predict the wind speed accurately even over a short time period of 30 minutes. 1. In deterministic models we assume that wind can be predicted exactly.different energy portfolios and. Chapter 3 describes the power system of the United Kingdom and how it can be modelled using linear programming. This uncertainty can also be pictured on the scenario tree and resolved along with the uncertainty of the wind power output using stochastic programming. In order to obtain a wind power output from the geographically diverse sources we will develop a Gibbs sampling algorithm to generate one-minute average wind speed from available average ten-minute values of wind speed at geographically dispersed locations. The issue is complicated by a possible loss of the generating capacity. the system balancing cost of wind power variability is calculated. Its profit is independent of the operational cost of the electricity generation.

Data used for the solving of the electricity scheduling problem are managed using JAVA programming language.Deterministic and stochastic models formulated in this work are implemented using a mathematical programming modelling language (AMPL)[47]. 12 .

As a result the United Kingdom has some of the best wind resources in Europe. Wind energy contributes around 2. as high average wind speeds and good reliability result in more power output and lower costs[6].Chapter 2 Wind Speed Analysis The aim of this thesis is to study electricity generation on one-minute time scale. This makes the country a very attractive location for wind developments. however. Samples generated for different sites of a geographically diverse territory are further transformed into power that can be used for optimization modelling of the electricity generation system. To obtain the necessary data. Least frequent data is provided by two sources: the Meteorological Office of the United Kingdom that publishes wind speed averaged over one hour for different sites in the UK and Utah Geological Survey (USA) [21] that gives an access to ten-minute average wind speed with standard deviation known. there is no one-minute wind speed available on the country level. Implementation of optimization models requires data with the appropriate resolution. weather on the British islands frequently changes producing high wind speeds.2% of the total electricity generation in the United King13 . Currently there are over 200 onshore and offshore wind farm projects operating in Great Britain with a total installed capacity of 3. Due to the proximity to the Atlantic Ocean and the North Sea. in both onshore and offshore locations. in this chapter we wish to study the only available source of one-minute wind speed recorded in the UK ([20]) and apply this information to generating an unbiased sample conditional on a given average value.3GW. 2. if not the world.1 Generating wind energy United Kingdom is situated on an island with a unique location that defines its climate.

we wish to develop a unique method that generates an unbiased sample from the wind speed stochastic process conditional on each consecutive group of wind speeds having a given average value.5 2 power. MW 1. the available wind speed is transformed into power using a wind-power curve.5 0 0 5 10 15 wind speed. Gibbs sampling algorithm. originally adapted to a problem of generating one-minute wind speed data.75MW. Let us take NEG Micon 2750/92 as a typical wind turbine with the capacity 2. 3 2. Its hub height is 70m and rotor diameter is 92m.1: Transformation function between the wind speed and the wind power 2. 14 . The wind-power curve for this wind turbine is plotted in Figure 2.1. These curves are uniquely determined for the different wind turbine type. The average turbine size in 2008 was considered as 2MW that produces energy with the average output of 40% of the maximum possible capacity[6].5 1 0. 3m/s cut in and 25m/s cut out wind speed. m/s 20 25 30 Figure 2. A new algorithm includes Monte Carlo Markov Chain method.2 Generating wind speed data This section starts with the analysis of the available source of one-minute wind speed recorded in the United Kingdom ([20]) that is used to identify the probability distribution required for sampling one-minute wind speed. Further. Rapidly developing technology introduces more efficient wind turbines built with the towers up to 100m tall that can capture strong winds. In order to estimate wind energy produced by a turbine.dom and this number is expected to grow as more wind farms are connected to the grid.

however. WT be a vector of wind speed corresponding to a height of 70m above ground level. s s Definition Let W1 . Many researchers. season.2. Thus. wind speed. and α is the power law exponent. feel that these complicated approximations reduce the simplicity and applicability of the general power law and that wind energy specialists should accept the empirical nature of the power law and choose values of α that best fit available wind data. 52.42N. however.1 One-minute wind speed data The one-minute resolution wind speed stochastic process is based on the statistical properties of one-minute wind speed measurements made at one location. Some researchers have developed methods for calculating α from the parameters of the log law. This research uses 12 years of data for the period between 1996 and 2007.05W. 4. In what follows the wind speeds have been transformed to correspond to a height of 70m above ground level. which is what expected for a flow over a flat plane. British National Grid Reference SN 604826) funded by the Natural Environment Research Council (NERC). For flat 1 and open areas α equals 7 . Wind speed read by the MST Radar is registered at 5-meter height. the exponent α is a highly variable quantity. We assume 360 days in a year so that T = 6220800. nature of terrain. Further adjustments are necessary. It has been found that α varies with such parameters as height above the sea level. ..2. McGowan and Rogers in [22]: W s (z) = W s (zr ) z zr α where W s (z) is the wind speed at height z. Its basic form is provided by Manwell. In practise. and various thermal and mechanical mixing parameters. The one-minute wind speed measurements required for this work are provided by Mesosphere-Stratosphere-Troposphere (MST) Radar located at Frongoch farm near Aberystwyth (140 m above mean sea level. T is a number of minutes in 12 years of one-minute records. temperature. a hub height of NEG Micon 2750/92 (a typical wind turbine for this research) is 70 meters. an A100R anemometer and a W200P wind vane[20].. The power law represents a simple model for the vertical wind speed profile. time of day. Wind speed and direction are recorded at one-minute intervals using vector instruments. it is necessary to transfer available 1-minute wind speed to this height.. as this can interfere with the 15 . Let us test the wind speed data for periodicity. W s (zr ) is the reference wind speed at height zr .

. TΥ is a minute of a year and an average value aν is assumed to be effected by the preceding and the following 20 minutes of the preceding and the following 10 days to the current. In (2... Definition Let {aν }ν=1.(δ+d). 1440... where Υ = 1.(τ +m) Υ=1 d=−10 m=−20 s where kdm is a weight of the wind speed WΥ. There are two different types of periodicity affecting original records.. τ ) and a minute effecting it (δ + d.. There is also a seasonal pattern which affects a strength of the wind speed during different months of the year.TΥ be a deseasonal scaling curve that carries the information on daily and seasonal patterns of the wind speed. every minute t of the 12-year data can be described by a δ day of an Υ year and a τ minute of the day: t = 518400(Υ − 1) + 1440(δ − 1) + τ.. τ + m).. .... δ = 1. We use a multiplicative model to evaluate the seasonal and daily effect in order to preserve a non-negative property of the data.1) 16 . .. . 12 10 20 s kdm WΥ. We assume 360 days in a year so that TΥ = 518400. 1440. where δ = 1. 360 and τ = 1. . 12.probability distribution of the wind speed data later. weights kdm form a shape of an isosceles triangle with the middle vertex at the current time moment of τ minute of δ day.. Weight kdm varies depending on the distance between the current minute (δ. There is a daily pattern which may be caused by the interaction between warm and cold air of the land and the sea consequently (geographical feature of the United Kingdom) or due to heating during the day over land (even when there is no sea involved). TΥ is a number of minutes in a year. .. Then aν is obtained by calculating a weighted average of the wind speeds over all 12 years for a day δ − 10 to δ + 10 and for each of these days over minutes τ − 20 to τ + 20. 360 and τ = 1...(δ+d). ..(τ +m) in minute (τ + m) of day (δ + d) in year Υ.1) ν = 1.. Similarly.. If plotted in terms of day d or minute m... Every weight kdm can be defined aν = a1440·δ+τ = (2. For a convenience of further calculations every minute ν of a year can be described by a δ day of the year and a τ minute of the day: ν = 1440(δ − 1) + τ.

each value of which is a weighted average over 12 years of data. To estimate deseasonalised wind speed Wtd we remove daily and seasonal effects from the wind speed Wts dividing the latter by a corresponding value of the deseasonal scaling curve.. WT . ∀Υ Definition Let us denote Θd (Wts . W1+l ). T again is a number of minutes in 12 years. .. . . Notice. . T . Regarding the first observation in each pair as one variable.. if any. WT as deseasonalised wind speed.2) In (2. (WT −l . we can measure the autocorrelation coefficient between 17 . aν ) = Θd (WΥδτ . d d Definition Let us denote W1 . that all 12 years are given the same weight in the average. for every integer l we can form T − l s s s s s pairs of observations. WT ). (W1 . For a convenience of calculation we assume a linear connection between a minute of a year ν and a minute t of the 12-yearly data: t = 518400(Υ − 1) + 1440(δ − 1) + τ ⇒ ν = 1440(δ − 1) + τ. Given T observations of s s wind speed W1 .. aν ) as a function transferring original wind speed Wts into deseasonalised wind speed Wtd for ∀t = 1. .2) values 10 and 20 correspond to intervals of 10 days and 20 minutes assumed to contribute to an average value aν . . or the wind speed with removed daily and seasonal components. where each pair of observations is separated by a time lag l. The series of autocorrelation coefficients measure the correlation. . then. on a time series. . (Wts . Wt+l ). . . that the hypothesis about daily periodicity of wind speed is supported by the autocorrelation of the original data. . As a result a deseasonal scaling curve aν can be constructed.. . By defining a time position of a wind speed t with a minute τ of a day δ in year Υ the deseasonilised value Wtd can be found as follows: s Wtd = Θd (Wts . . namely.. Note.as follows: kdm = (10 − |d| + 1)(20 − |m| + 1) 12 ∗ (20 + 1)2 ∗ (10 + 1)2 (2. and the second observation in each pair as a second variable. . aδτ ) = s WΥδτ aδτ where aδτ = aν is the deseasonal scaling curve. between observations at different distances apart and provide useful descriptive information [43].

adjacent observations. . 1 without seasonal component original data 0. . . 18 . however this pattern is not visible in the deseasonalised data.5 0.8 0.7 0. W s and Wls .6 0.1 0 0 2000 4000 6000 time.2: Autocorrelation of the original and the deseasonalised wind speed data Figure 2. using the formula: corr(W s s .2 0.2.3) s s where W is the mean value of the series W1 . The fluctuations of the autocorrelation function for the original data repeat every 1440 minutes. 2. . Wls ) = s T −l s s t=1 (Wt − W )(Wt+l − s 2 T s t=1 (Wt − W ) W ) s · T T −l (2.9 0.3 0. In this section we wish to describe an algorithm that would allow us to sample one-minute wind speed from a known distribution when an average value over ten minutes is given.2 plots the autocorrelation of one-minute wind speed recorded with MST Radar located near Aberystwyth and the autocorrelation of the same deseasonalised data.4 0. min 8000 10000 12000 Figure 2. WT .2 autocorrelation Methods of generating missing data After wind energy is incorporated into the the electricity generation system its optimization modelling requires geographically distributed samples of one-minute wind speed data. US Geological Survey provides sample of ten-minute wind speed recorded over a large territory of Utah state. which supports the presence of a daily periodic component in the wind speed.

to use a special case of Metropolis-Hastings algorithm . Shortly before this. This method ensures that as a number of iterations of sampling from known conditional distributions increases. In a 19 .2. it is hard to describe this distribution analytically. the density of a resulting set of variables converges to the required one. Gelfand and Smith[14] showed a close relationship between the Gibbs sampler introduced by Geman and Geman (1984) and the data-augmentation algorithm proposed by Tanner and Wong. 18. Meanwhile. In their paper “Sampling-based approaches to calculating marginal densities”. One of the MCMC algorithms.To ensure a generated set of data values reflects the properties of one-minute wind speed distribution. one-minute records from MST Radar located near Aberystwyth are analysed. In the 1980s two important papers were published that discussed a question of computing an estimate of a posterior distribution.3 Description of Gibbs sampling Let us formulate an algorithm of Gibbs sampling for a problem of generating T variables correlated with the variables in a certain “neighbourhood” [14]. We wish to apply Gibbs sampling to the sites with the highest wind speed. Metropolis-Hastings produces a sequence of random samples from a probability distribution. 2. it is possible to find a conditional distribution of wind speed at every minute depending on the past realisations and. hence. however it requires a joint probability distribution of one-minute wind speed which is hard to find. 17. It worked better for dry days as Glasbey required repeated sampling until the desired average value appeared. Gibbs sampling algorithm was developed during the last 30 years. In 1990. 19]) applied Gibbs sampling to a problem of sampling hourly rainfall when a daily value is given. A survey of existing techniques on data sampling identified Monte Carlo Markov Chain (MCMC) methods as a group of methods that can be applied to generate one-minute wind speed when the distribution is not fully formulated. Glasbey et al. Gelfand and Smith gave a description of a sampling technique that we shall extend for the generation of one-minute wind speed when the average value over each ten minutes is known. Tanner and Wong (1987)[15] estimated a distribution of the set of missing and observed values by data-augmentation.Gibbs sampling algorithm. hence. However. in ([16. in 1984 Geman and Geman[13] offered a new restoration algorithm for computing the maximum a posteriori estimate of the original image given a degraded image. the algorithm is originally adapted to preserve good samples and reduce their number.

. . . . YT ] ∗ [YT −2 |YT −3 . . .4) In (2.5) produces a marginal distribution for Y1 [14]: ˆ ˆ [Y1 ] = h(Y1 . .10). Therefore.relation to a collection of random vectors. . . . Y4 ] ∗ [Y3 |Y4 . . The process of marginalisation (i. . . and marginal forms. The algorithm uses the following representation of the desired posterior density: [Y1 ] = [Y2 ] = [Y3 ] = ··· [YT −1 ] = [Y1 |YT . Yi+1 . ∗ denotes a multiplication of densities. . . . . . . Y1 . with the convention that all variables appearing in the integrand but not in the resulting density have been integrated out. YT . Y1 . YT ] = [Yi |Yj . . where Si ⊂ {1. i. . . . . . Yi+1 . . . . . can be written as [Y1 . ∗ [Y3 |Y2 ] ∗ [Y2 ] [Y2 |Y1 . the conditional distributions [Yi |Y1 . . Y3 . Y2 ]. Y2 ] = [Y1 |Y2 ]∗[Y2 ]. T }. j ∈ Si ⊂ {1 . . YT . . YT ] ∗ .9) [YT ] = [YT |YT −1 .8) [YT −1 |YT −2 . for example. suppose that for i = 1. for example [Y1 . Gibbs sampler is an iterative Markov Chain Monte Carlo method. . T1 ] ∗ [YT −1 |YT −2 . for this example the integration is with respect to Y3 and Y4 [14]. conditional. . Y3 ] ∗ . Yi−1 .5) − (2. . . . . . . .10) In (2.9) and further up into (2. . Y2 ] ∗ [Y4 |Y2 ].5) (2. . . . T . ∗ [Y4 |Y3 ] ∗ [Y3 ] [Y3 |Y2 . marginal distributions of all the random vectors 20 . . . T }: [Yi |Y1 . .7) (2. . . . We denote by Si a certain “neighbourhood” subset of {1. . Y1 ) is an integral that was resulted from the substitutions with Y1 ≡ Y1 . . . . T (2. . Y1 . YT . .e. . . so joint. . . Y2 ] ∗ [YT |YT −1 . . . [Y1 |Y2 ]. . .10) into (2. YT −2 .6) (2. .4). . YT ] are available in reduced forms. . brackets denote densities. and [Y ]. ∗ [Y2 |Y1 ] ∗ [Y1 ] (2. Y2 ] ∗ . Y4 ] ∗ . Y3 ] ∗ [Y1 |YT . T }]. By similar substitutions. . Y1 ] ∗ . i = 1. . ∗ [Y1 |YT ] ∗ [YT ] (2. ∗ [Y5 |Y4 ] ∗ [Y4 ] (2. . Y1 .e. . Y4 ] ∗ [Y2 |Y1 . . Yi−1 . . depend only on {Yj }j∈Si . . integration) is denoted by forms such as [Y1 |Y2 ] = [Y1 |Y2 . . . . Y2 . . . Y1 ) ∗ [Y1 ] ˆ ˆ where h(Y1 . . . . . . Substitution of (2. YT . . .

. T . . . . . YT ] that we wish to estimate by marginal distributions using Gibbs sampling algorithm. . Y3 . i ∈ 1. Y2 .11) The method of successive substitution for solving (2. which are not known in our case. YT (j+1) (j) (j) we draw Y1 ∼ [Y1 |Y2 . YT ] After the value of Y1 is found (j+1) (j+1) (j) (j) we draw Y2 ∼ [Y2 |Y1 .. .. YT are found: [Yi ] = ˆ ˆ h(Yi .10).3.Y1 . . YT ] ··· 21 (j+1) (j) (j) Step 2: ···: . The full conditional distributions uniquely determine the joint distribution [Y1 . . Step 1: Given an arbitrary starting set of values Y1 . . Its implementation by Tanner and Wong[15] requires. . .11) suggests an iterative method for the estimation of marginal distributions [Yi ]. . Yi ) ∗ [Yi ] (2. . . .5) − (2.. . however. . . the availability of all conditional distributions on the right-hand side of (2. Figure 2.3: Schematic description of Gibbs sampler as a Markov-chain process Gibbs sampling can be described as the following sequence of actions schematically presented in Figure 2. . .

4) until the right average value is reached within a specified tolerance. . . each variable is visited in the natural order and a cycle requires T random (i) (i) variate generations. T .10 ).12) where f (Yj ) is a function of 10 components of Yj and Wjs is the ten-minute measured value. . In fact. . under visiting in the natural order. rather than L1 norm. the resulting set of variables (Y1 .12) they suggested repeating sampling from the conditional distributions (2. YT ] → [Y1 . . . YT ) converges to the true joint density at a geometric rate in i. . . .Step T : After the values of Y1 . (i) (i) Rate. j = 1. or in other words Yj = (yj. . . we can write an additional condition required in the probability density function to reflect the fact that we require 10 sampled values of one-minute wind speed having the measured ten-minute wind speed average: Yj ∼ [Yj |Y1 . Y2 . [Y1 . Yj+1 . the joint density of (Y1 . Rather than requiring that each variable is visited in repetitions of the natural order.. convergence still follows under any visiting scheme. Under mild conditions. .. . . . . . .T as a vector consisting of 10 scalars. . yj. Y2 . Dry blocks were easier to match but approximately 0. . results a set of variables (Y1 .. . YT ] as i → ∞. . . . f (Yj ) = Wjs ] (2. Let us denote {Yj }j=1. . . a slightly stronger result is proved. . . . provided that each variable is visited infinitely often. Geman and Geman showed that the following results hold[13. . T . (i) (i) For i sufficiently large. A minor adjustment to the rate is required for an arbitrary “infinitely often” visiting scheme. YT ). . Yj−1 . . . YT −1 ] (j+1) (j+1) Thus. . In order to satisfy (2. . . Using the sup norm. however. . . For the purpose of electricity generation using wind speed turbines it is in the interest of producers to locate wind farms in windy places. It does not reflect.1% of blocks that described intense rainfall needed more than 1000 attempts. we would like to suggest a modification to Gibbs sampling of one minute wind speed that reduces a number 22 . .1 . YT . . . Then.. . . YT −1 are found (j+1) (j+1) (j+1) (j+1) we draw YT ∼ [YT |Y1 . . Glasbey and Allcroft in [16] performed spatio-temporal rainfall disaggregation using Gibbs sampler for a truncated Gaussian Markov random field model. . . After i such iterations. Therefore. 14]: (i) (i) (i) Convergence. . our property that the average value of wind speed over each group of 10 one-minute values should equal a given ten-minute measured value. and so [Yj ] → [Yj ] for j ∈ 1. . YT ) is a sufficiently accurate sequence of wind speed values.

5.4: Cumulative probability distribution of one-minute wind speed recorded with MST Radar located at Frongoch farm 23 . However.7 0.4 Transformation to normal variable This section presents transformation of the deseasonalised one-minute wind speed data so that the marginal distribution is normal. there is not enough data to decide on a unique stochastic process for the deseasonalised wind speed so we proceed as follows. Points 1 and 2 make the generation of data computationally tractable. a technique is described that shifts every generated point towards the required average wind speed value in a way that does not bias the sampling.1 0 0 1 2 3 4 5 wind speed without periodical component. 1. 1 0. In the normalised wind speed variable the distribution of wind speed at any one time is truncated normal. the corresponding deseasonalised wind speed is zero. If the stochastic process describing the deseasonalised wind speed was fully known. Later in subsection 2.2.4 0. 2.of iterations.8 cumulative probability density function 0.9 0. We assume that the stochastic process in the normalised variables is multivariate normal with the same autocorrelation function as the measured data. We do a non-linear transformation of the deseasonalised wind speed so that the distribution of wind speeds at any time is normal and so that when the normalised wind speed falls below a threshold.2 0.6 0. This reduces a number of samples required by using samples that initially do not satisfy the condition f (Yj ) = Wjs . m/s 6 7 Figure 2.3 0.2. then in theory we could generate the samples from this.5 0. 2.

05 0 -8 -6 -4 -2 0 normal variable. or the deseasonalised wind speed transfered to the normal variable. As a multiplicative model was applied to remove seasonal effect from the data.15 0. (2.3 0. .08 where W d is a deseasonalised wind speed.25 0. Eight percent of wind speed data equals zero.4 0. y 2 4 6 8 Figure 2. This is due to the fact that the measurements are recorded in discrete quantities.14) 24 .08 ˆ (2. Since 8% of the wind speeds are zero we require: P (W d ≡ 0) = 0.The nature of wind speed is such that there can be no negative values.4. .5). T again is a number of minutes in 12 years. .35 probability density function 0. in our case these are the wind speeds recorded at zero.13) 0. Definition Let us denote Y1 .1 0.5: Truncated and full plot of normal probability density function Let us consider standard normal distribution N (0. These features are depicted with the distribution of 1-minute wind speed plotted in the Figure 2. . Further. the non-negative property of the original data is still present. 1) with mean equals 0 and standard deviation 1. The corresponding lowest value of the normal variable can be found from the equation: P (Y ≤ y ) ≡ P (W d ≡ 0) = 0. hence we wish to transform the standard normal distribution so that it is truncated from the left (Figure 2.45 truncated normal pdf standard normal pdf 0. typical measurements show a big number of values zero.2 0. Single truncation means that certain outcomes are constrained. YT as normalised wind speed.

14) we have y equals −1. . UT as standard normal random variable with mean equals 0 and standard deviation 1. Definition Let us denote U1 .1225. Y (W d ) : P (Y < y ) ≡ P (W d < w) ¯ ¯ 25 (2. The probability density function (pdf) of U is given by U2 1 φ(U ) = √ e− 2 2π −∞≤U ≤∞ Let Φ(U ) denote the cumulative distribution function (cdf) of U . A random variable Y has a single truncation from the left normal distribution if its probability density function is:   −µ φ( Y σ ) σ(1−Φ(ˆ)) y f (Y . .15) (2. y Let us define θ for a singly truncated from the left normal distribution: θ= φ(ˆ) y 1 − Φ(ˆ) y 0 otherwise Then the expected value and variance of Y can be obtained in the following way: E(Y ) = µ + σθ V (Y ) = σ 2 (1 − θ2 + θˆ) y (2. The random −µ variables Y and U are linearly related and the relationship is given by U = Y σ . .1225 is left (lower) truncation point and a degree of truncation is ˆ Φ(ˆ) = 0. For our work Schneider provides an excellent overview of parameter estimation of truncated normal distributions in his detailed work “Truncated and Censored samples from Normal Populations” [44]. σ.17) .From (2. Johnson and Thomopoulos applied most of the analysis to a special case of left truncated normal distribution [45].15) and (2. ˆ There is a large body of literature on the subject of estimation of the parameters of the original data (µ. µ.08 for the considered dataset. σ) based upon data from truncated samples. .16) Every variable from the truncated normal distribution described by (2. The cdf of a −µ normal random variable Y with mean µ and variance σ 2 is Φ( Y σ ). y ) = ˆ y≤ ˆ Y −µ σ where y = −1.16) corresponds to a single value of the deseasonalised one-minute wind speed.

We wish to find an analytical fit to Θn (Wtd ) so that during sampling it is easy to calculate its inverse Wtd = Θ−1 (Yt ).3652*x-1. suggest using a quadratic as the function to transform the rainfall to normal. 17. .6 plots the original empirical relation between normalised and deseasonalised wind speed.1225 4 3 normal variable 2 1 0 -1 -2 0 1 2 3 4 5 wind speed without seasonal component. During the test quadratic function gave a better 26 . .where y and w are the corresponding values of the normal variable and the de¯ ¯ seasonalised wind speed respectively. 18.01. Hence. . it is possible to look for an analytic fit that can simplify further calculations. it has a “one-to-one” relation with the deseasonalised wind speed. 5 original plot -0.6 and tested a set of points {(Wtd . If transformation function were not monotonic there would be a possibility of multiple solutions. 19]) Glasbey et al. T . In ([16. m/s 6 7 Figure 2... this would prevent the method working). We assumed Θn (Wtd ) to be a logarithn mic or quadratic function from a plot on figure 2. Figure 2.T for a best fit. If we write now a normalised wind speed in discrete form with a step equal to 0..0854*x*x+1.. Yt )}t=1.6: Normal variable plot with respect to the deseasonalised wind speed Definition Let us denote Θn (Wtd ) as a function transferring deseasonalised wind speed Wtd into normalised wind speed Yt for ∀t = 1. . Wind speed is normalised according to the cumulative density function of the left truncated normal distribution. It appears a simple function can be used as an analytic monotonic fit to the transformation function between the normalised and deseasonalised wind speed (Monotonicity is a necessary condition for this transformation.

. .6 along with the original plot. With these parameter values. the function is monotonic along all the period [0 : 6] (m/s) with a range for the normalised wind speed [−1. Yt ). . T (2.1225 : 4. −7. . . . Using a quadratic function formulated as a full square it is straightforward to find the inverse to Θn (Wtd ): Wtd = Θ−1 (Yt ) = ± n Yt − b3 − b2 . yt.993] (m/s). .19) is chosen so that a value of the deaseasonalised wind speed Wtd is in a range [0 : 7.j )). . The estimated transformation function is plotted on Figure 2. . Y2 . Yt+1 . . . . Y2 . T .e. Plus or minus in front of the square root in (2. n 10 j=1 d where yt. Y1 . . .10 ). YT .5 Suggested modifications to Gibbs sampler Let us now reformulate the Gibbs sampling algorithm for the normalised wind speed. .3335] and probability of the deseasonalised wind speed exceeding 6m/s at a single site is only 1. Yt−1 . b1 ∀t = 1. i. Y2 . a variable from the normal distribution truncated to the left.1 .6 ∗ 10−5 .19) Parameters of the equation (2. .2. b2 and b3 can be estimated by solving simultaneously three linear equations for the points from the original plot (Wtd . 10 1 f (Yt ) = Θ−1 (Θ−1 (yt.993. we write a function Θn (Wtd ) as a full square: Yt = Θn (Wtd ) = b1 (Wtd + b2 )2 + b3 . . . ∀t = 1.18). which is small enough to ignore. . . . b1 . t = 1. On page 19 of the current work we introduced Yt as a vector containing 10 normalised wind speeds or Yt = (yt. . t = 1. 4. . .j is normalised wind speed that is transfered into one-minute wind speed with inverse functions to Θn and Θd defined on page 23 and page 14 accordingly. . f (Yt ) = Wts ]. 2.3335). . YT from the available conditional distributions: [Yt ∼ [Yt |Y1 . i. . giving b = (−0. . Then f (Yt ) represents the average of the wind speeds corresponding to the components of Yt .e. therefore.0854. T (2. . YT ] and 27 .18) where Wtd denotes the deseasonalised one-minute wind speed and Yt denotes the normalised wind speed. . We wish to sample a collection of random normalised wind speed variables. T To achieve this. .result. first Yt is sampled from the distribution [Yt |Y1 . . . . . .

1 1 T −1 T −1 T −1 φ(Y ) = √ e− 2 (Y ϕ Y −2Y ϕ µ−µ ϕ µ) 2π (2. In order n d to work in the space of the normalised wind speed we assume further that the equation f (Y ) = W s can be replaced with its linear approximation 1T Y = Y .21) is complex. So Yt is shifted along a vector of maximum probabilities for the desired normal probability distribution function. Value of Y is found so that 1T Y is situated closer to the center of the distribution 28 .21) 1 T −1 (Y ϕ Y − 2Y T ϕ−1 µ − µT ϕ−1 µ) 2 Function f (Y ) in (2. We wish to preserve a chance of this vector to appear. every normalised wind speed is transformed to one-minute wind speed using two inverse functions Θ−1 and Θ−1 . Assume a vector Yt is generated. Let us consider a matrix form of the probability density function for the normalised wind speed. where 1 is a vector of the same size as vector Y and consists of scalars 1 only. Then a point Y which maximizes φ(Y ) subject to satisfying the constraint f (Y ) = W s can be found by solving an optimization problem: 1 1 T −1 T −1 T −1 arg max √ e− 2 (Y ϕ Y −2Y ϕ µ−µ ϕ µ) 2π subject to: f (Y ) = W s which occur at the same arg min subject to: f (Y ) = W s Write Lagrangian for this minimization problem to solve it: 1 L(Y.˜ ˜ then shifted to Yt in the space to satisfy the equation f (Yt ) = Wts . λ) = (Y T ϕ−1 Y − 2Y T ϕ−1 µ − µT ϕ−1 µ) + λ(f (Y ) − W s ) 2 (2.20) where Y is the normalised wind speed with the covariance matrix ϕ and the mean µ.

24) ˜ where Yt is the original generated variable and Yt is a shifted normalised value of ˜ wind speed so that it satisfies now the condition 1T Yt = Y .21) is easily differentiated with respect to Y: ∇Y L = ∂[ 1 (Y T ϕ−1 Y − 2Y T ϕ−1 µ − µT ϕ−1 µ) + λ(1T Y − Y )] ∂L = 2 ∂Y ∂Y −1 −1 T = ϕ Y − ϕ µ + λ1 First-order necessary condition for Y to be a minimum for the above optimization problem stipulates that ∇Y L = 0. The scalar λ is a Lagrangian constant and defines a position of Y on a vector of maximum probabilities.23) satisfies a linear equation 1T Y = Y . and we assume Y = b3 . they describe the probability distribution of the normalised wind speed. there exists a vector Yt∗ such that it satisfies ˜ both equations exactly.18). the resulting Lagrangian function obtained from (2. Parameter λ is found so that Y in (2. When using a linear approximation of the equation f (Y ) = W s .[Y ]: If If 1T Y ≤ 1T µ. f (Yt∗ ) = W s and Yt∗ = −λ1T ϕ + Yt .23) Y = −λ1T ϕ + µ In (2. Assume.22) (2. From here we can find an equation for Y : ϕ−1 (Y − µ) = −λ1T (2. where µ is mean value of the distribution [Y ] and b3 is a parameter in a quadratic function at (2.23) ϕ and µ denote the covariance matrix and the mean vector. a generated vector Yt is shifted along ˜ the vector described by (2. 1T Y > 1T µ.23). In the context of the Gibbs sampling. This means that Yt acquires a value from the vector ˜ Yt = −λ1T ϕ + Yt (2. 1T Y is the tangent to f (Y ) − W s at the point of the shortest distance between f (Y ) − W s and µ. Then Yt is accepted ˜ as a valid normalised wind speed if the probability of Yt is within a specified 29 . ˜ A shifted variable Yt is found so that it satisfies f (Y ) = W s within a certain tolerance as the equation f (Y ) = W s is close to its linear approximation 1T Y = Y ˜ which Yt satisfies exactly.

T q are the one-minute wind speeds and Wts is the measured average wind speed for times i = (t − 1)q + 1. The values of the one-minute wind speed Wts ˜ in this case are found using Wts = Θd (Θn (Yt )).3 Implementation of modified Gibbs sampler We wish to generate an unbiased sample from the wind stochastic process conditional on each consecutive group of wind speeds having a given average value. tq. first the removal of periodic effects by transforming to d s d new variables wi = Θd (wi ) followed by a transformation to variables yi = Θn (wi ) which have a truncated normal distribution. .1 dealt with daily and seasonal periodical components of the original wind speed data. s Instead of operating with wind speed variables wi directly. . Section 2. . 30 . . Definition Let us denote Θ as a compound function. We will do this in two steps. .tolerance of the probability of Yt∗ . such that s s yi = Θn (Θd (wi )) := Θ(wi ) and since the transformation will be “one-to-one”. .4..2. wq .2. s s wq+1 . + wtq ) = Wts q (t−1)q+1 s where wi ... s wi = Θ−1 (Θ−1 (yi )) := Θ−1 (yi ) n d Transformation functions Θd and Θn were described earlier in this Chapter.4 discussed transformation to the normalised wind speed.2. . . Section 2. . For the real problems we are dealing with q = 10. i = 1. . 2.. . . Given these functions are known and given µ and ϕ as described in Section 2. wT q .. . . . . t=1 such that t=2 t=T 1 s s s (w + w(t−1)q+2 + . . . . but we will also illustrate the case q = 2. we work with the transformed stochastic process when the periodicity (seasonal and daily effects) has been removed so that their marginal distribution is a truncated normal. s s w(T −1)q+1 . the algorithm of generating one-minute wind speed values is the following: Step 1: Start with t = 1. . . . w2q . . . Assume the group size is q and we wish to generate T · q values: s s w1 .

A stochastic process can be described with the moments of the process. ytq ) along the vector of maximum probabilities to find (˜(t−1)q+1 . Every generated variable Yt = (yt. The fact that Yt includes 10 univariate variables defines a vector Yt as a multivariate variable. . .3. . ytq ) y ˜ If a probability of (˜(t−1)q+1 . i = [(t − 1)q + 1 .1 . First moment of the stochastic process of 31 .2 of this Chapter introduced the Gibbs sampling technique for a collection of normalised wind speed variables Y1 . Let us find mean and autocovariance of the univariate one-minute normalised wind speed that can be used further to define a stochastic process for multivariate normalised wind speed. . . .10 ). . tq]. which if transferred to the wind speed satisfies 1 s s s (w(t−1)q+1 + w(t−1)q+2 + . s Find wi = Θ−1 (yi ). . To generate a sample with the required probability density function we wish to perform a significantly big number of such iterations. . . . . . T has to satisfy f (Yt ) = Wts . Y2 . which is ten wind speed values have to add up to the measured average ten-minute wind speed. . . . . . . 2.Step 2: Generate normalised wind speed Yt = (y(t−1)q+1 . . . ytq ) is within a specified toly ˜ ∗ ∗ erance of a probability of (y(t−1)q+1 . . . . + wtq ) = Wts q within a specified tolerance Go to Step 8 Step 3: Step 4: Step 5: Shift (y(t−1)q+1 . . YT . . . + wtq ) = Wts exactly. q Go to Step 8 Go to Step 2 t=t+1 If t > T Stop else Go to Step 2 Step 6: Step 7: Step 8: Step 9: This algorithm describes a single iteration of the Gibbs sampler. .1 Multivariate normal distribution Section 2. s s s If 1 (w(t−1)q+1 + w(t−1)q+2 + . particularly the first and second moments. . . . . ytq ). . ϕ). . yt. . t = 1. mean and autocovariance function respectively. . ytq ) from Nq (µ. . .

0. .1 0 0 200 400 600 800 lag. we estimate the autocovariance function with a mixture of exponential curves: m m cov(y. During the test the exponential function gave a better result.. We assume a lag up to 1440 minutes (1 day) is significant for the one minute wind speed stochastic process. yl ))}l=1..6 original normalised wind speed fitted exponential function 0. yl ) = Tq − l T q−l i=1 (yi − y)(yi+l − y) (2. yl ) to be a quadratic or exponential function from a plot on figure 2. therefore.L for a best fit. .. . . In order to use the autocovariance in the process of wind speed generation it is convenient to find a continuous function that would fit in the original plot.4 autocovariance 0.7 plots the autocovariance function in terms of a lag l calculated for the normalised wind speed.26) According to Allcroft and Glasbey [16] a condition of all coefficients βi being positive is sufficient for the function to be valid.3 0.wind speed generation is described earlier as the expected value of the truncated normal distribution.7: Fitted sum of exponential functions to approximate the autocovariance of the normalised wind 32 .5 0. We assumed cov(y. Figure 2.25) where l denotes a lag between the variables and y is the mean value of the series. Given T q variables y1 . yl ) ≡ βi e i=1 −ςi |l| i=1 βi = 1 (2. autocovariance coefficients can be found: 1 cov(y. It can be found using the formula (2. cov(y.15)..7 and tested a set of points {(l. yT q .2 0. min 1000 1200 1400 Figure 2.

0002) Figure 2. 0. ϕ). Calculated autocovariance and mean describe a stochastic process for the univariate normalised wind speed y1 . for l > 1000 a gradient to the curve changes and it is hard to find a simple function to fit it perfectly.25. yT q . . 0. 0. . Variable YL−q . 0. . YT every vector Yt follows the distribution Nq (µ. for example. . ϕ). 0. Covariance matrix ϕ is found based on the autocovariance function described by (2. .Figure 2. we take q = 10 << L.7 plots the original autocovariance for the normalised wind speed along with the fitted exponential function. For the convenience of further calculations a covariance can be described in a matrix form:   ϕ11 ··· ϕ1j ··· ϕ1L    ··· ··· ··· ··· ···    ϕ =  ϕi1 ··· ϕij ··· ϕiL      ··· ··· ··· ···   ··· ϕL1 ··· ϕLj ··· ϕLL where L denotes a biggest lag of the autocovariance. We found a good fit to the autocovariance function to be a mixture of six exponentials with the following parameters: β = (0. of dimension (L − q). is a set of ten normalised wind speeds for which we wish to find a probability distribution. Partition µ and ϕ analogously 33 .204.04. YT .026. is a set of variables that first q variables correlate with. 0. . 0. It can be used now to formulate a multivariate stochastic process for the series Y1 .19. 0. of dimension q.7 shows that the autocovariance of the normalised wind speed is influenced by certain processes that effect its smoothness. partitioned into q and (L − q) components as Y = Yq YL−q In our case Yq . . i≥j (2.29) and ς = (1. Suppose that Y ∼ NL (µ.0023. . . 0. For the case of generating ten minutes of wind speed.0005. 0.13. To find mean and covariance of a q-dimensional variable let us partition an interval of length L.26): ϕji = ϕij = cov(y.27) For a collection of multivariate normalised wind speeds Y1 . .5. yi−j ). .0018. .

random vector Y with a Nq (µ. there exists a unique multivariate normal distribution with mean vector µ and covariance matrix ϕ [46] . symmetric matrix ϕ of dimension (q×q). ϕqq ) and YL−q ∼ NL−q (µL−q . M can be 34 .28) and ∗ ϕqq|(L−q) := Cov[Yq |YL−q = YL−q ] = ϕqq − ϕq(L−q) ϕ−1 (L−q)(L−q) ϕ(L−q)q (2. Assume that ϕ(L−q)(L−q) is positive definite. For any vector µ ∈ Rq and positive semidefinite symmetric matrix ϕ of dimension (q × q). Then Yq ∼ Nq (µq . for example Cholesky decomposition. to generate a multivariate normal variable with mean µ and covariance ϕ we sample first a multivariate vector U ∼ Nq (0. (2. If ϕ is a positive semidefinite symmetric matrix of dimension q × q.29) Mean and covariance stated in (2.as µ= and ϕ= ϕqq ϕq(L−q) ϕ(L−q)q ϕ(L−q)(L−q) µq µL−q with dimensions of the subvectors and submatrices as induced by the partition of Y .29) represent conditional normal probability distribution that describes the stochastic process of ten-minute wind speed generation. ϕ)-distribution can be generated as Y = µ + MU (2. Iq ). Then the conditional distribution ∗ of Yq given YL−q = YL−q . is q-variate normal with the following parameters: ∗ µq|(L−q) := E[Yq |YL−q = YL−q ] = µq + ϕq(L−q) ϕ−1 (L−q)(L−q) (yL−q − µL−q ) (2. since in our case covariance matrix ϕ is a positive definite matrix. Matrix M can be found precisely using different methods. Iq ).31) where M T is the transposition of matrix M . and M M T = ϕ. Further. More specifically.30) where U ∼ Nq (0. For any vector µ ∈ Rq and positive semidefinite.28) and (2. ϕ(L−q)(L−q) ). then there exists a matrix M of dimension (q × q) such that ϕ = MMT .

. ... yi+1 ) := (Θ(Wts ). . .. . Let us start with normalising a given average wind speed by transforming it to the truncated normal variable: Yt (0) := (yi .i M11 Set j = 2 Mjj = Mij = φjj − φij − j−1 k=1 j−1 k=1 2 Mjk Mik Mjk . 1. . . on the interval s {0.3. .8. . . Mqq         M :=    Step 1: Step 2: Step 3: Step 4: Step 5: Step 6: Step 7: M11 = Mi1 = √ where every element Mij . 1. To illustrate the method we will first derive a case where the variables are divided into groups of two.computed as follows:  M11 0 M21 M22 . 0 . Mq1 Mq2 . 0 . 2T } we wish to generate points wi subject to a given average of two adjacent values of wind speed Wts : s s wi + wi+1 = Wts . Therefore. Θ(Wts )) 35 (0) (0) .j Mjj <i≤q Set j = j + 1 If j ≤ q Go to Step 4 Else Finish algorithm Matrix M is used now to generate a multivariate normal variable Y of the Gibbs sampling algorithm. . .. 2. i ≥ j can be found with a sequence of calculations: φ11 ≤q φi1 . .. T This can be achieved through a sequence of steps described below in this section and illustrated in Figure 2. . .. 2T..2 Two-variable sampling We wish to generate data for each minute t where each consecutive group of ten variables has a given average value... . t = 0.. 1... 2 i = 0. .

6 0.6 0. Here are the main steps of j-th iteration: 1.8 0. In the normalised (j) variables this will be a bivariate normal Yt ∼ N2 (µ. ϕ= 0. ϕ). for our example.8: Contours of Normal distribution and four linear functions used in the modification to Gibbs algorithm 2. a point A := Yt (Figure 2.2595. . where ϕ is the same for all samples and is given in general by. is to sample from one group of variables conditional on achieving a given average wind speed.7 (j) (j) (j) and µ is calculated using the neighbouring sample values (j) (j) (j) (j−1) (j−1) (Y1 . The generated normalised wind speeds are transformed further using inverse 36 . . . Yt−1 . yi+1 ) = (1. yi+1 ). . . . . . ϕ). ∗(j) ∗(j) . := (yi Assume.3102) is generated ∗(j) Figure 2. Yt+1 . The repeated subproblem when using Gibbs sampling algorithm to generate samples with given means. Y2 . Assume we wish to generate a sample Yt = (yi .This gives us an initial set for Gibbs sampling.8) using N2 (µ. −0. YT ).

3.e. wi+1 ) with their average equals Wt s(j) s(j) := := ∗s ∗s wi + wi+1 2 where Wt is the average wind speed corresponding to the generated values ∗(j) ∗(j) (yi . yi+1 ). (2. Y = 1T Y B . If Wt equals Wts within a ∗(j) ∗(j) ∗s ∗s specified tolerance.function to Θ: ∗s wi = Θ−1 (yi ∗(j) ∗(j) ) (2. a line 1T Yt = Y defined in the Section 2. Definition Let us denote Y C as the intersection of the line (2. Definition Let us denote Y B (B in Figure 2. Points Y C and Y D (C and D in Figure 2.34) with the tangent to the curve f (Yt ) = Wts at the point Y B .8) as the point on the curve f (Yt ) = Wts that acquires the shortest distance to the center of the distribution [Y ]. i. The generated point (yi .33) new wind speed values are Wt ∗s ∗s (wi . By advancing λ from one border of the interval to another with a step 0. yi+1 ) is shifted in order to satisfy Wts as explained in steps 3 and 4 of this algorithm.34) is the point A sampled in step 1 of this algorithm. 37 if 1T Y B > 1T µ. wi+1 ) is accepted.2.2. and we assume Y = b3 . otherwise (yi . we search along the line Yt ∗(j) (j) ∗(j) ∗(j) s(j) = −λ1T ϕ + Yt ∗(j) .5: If 1T Y B ≤ 1T µ. ϕ is a where Yt covariance matrix of distribution [Yt ] and 1 is a vector of the same size as Yt and consists of scalars 1 only. We assumed λ to vary in the interval between −10 and 10. the point (wi .23).32) (2.001.33) s(j) ∗s wi+1 = Θ−1 (yi+1 ) Calculated with (2.32) and (2.8) are found during the search.5 of this work. that are sufficiently large to find Y in (2. yi+1 ) is tested. A scalar Y is calculated as in the Section 2. .

Steps 1 to 5 are repeated until j reaches a specified number of iterations j ∗ . Glasbey et al. To sample hourly rainfall values it required an average of 170 simulations per day (sometimes considerably larger 38 . In [16]. The probabilities of the points Y D and Y C are further compared: If φ(Y D ) >ρ φ(Y C ) accept Yt (j) =YC Else reject Y C and Go to step 1. We determine j ∗ so that a stochastic process of the resulting sample is weakly stationary. By modifying the value of λ. Y C (and a corresponding λC ) are found so that 1T Y C and Y match to small given tolerance. We wish to apply the demonstrated method to the sampling of ten minutes of wind speed when their average value is given. Y D (and a corresponding λD ) are found so that f (Y D ) and Wts match to small given tolerance. Else Finish j-th iteration of Gibbs sampling algorithm. 4.18). Definition Let us denote Y D as the point on the line (2. 2.where b3 is a parameter of the quadratic function in (2. generate hourly rainfall with a given daily value by repeated sampling from the multivariate normal distribution until the data total fell within a certain margin of the target. where φ(∗) is the probability distribution function for truncated normal distribution used for sampling normalised wind speed. in [17] suggested to take ρ as a random number from the uniform [0. By modifying the value of λ. 1] distribution in order to add random character to the test. 5. We assumed ρ to be a fixed number 0.95.3. If i ≤ T Go to step 1. Take i = i + 1.3 Multivariate sampling The two-variable algorithm was described in order to illustrate wind speed generation when the average value of adjacent values is given. Glabey et al.34) that satisfies the condition of two adjacent one-minute wind speeds being equal a given average value f (Y D ) = Wts .

ϕqj . . ϕ1j .. . is to sample from one group of variables conditional on achieving a given average wind speed.. . .   . The calculated conditional covariance ϕ does not 39    ϕ=    . .  µ= . . . . . ϕiq .3. . . . yi+q−1 ) := (Θ(Wts ). . The algorithm starts with normalising a given average wind speed by transforming it to the truncated normal variable: Yt (0) := (yi .. q where q = 10 and t = 1. 1.3.  µq and  ϕ11 . For sampling of a multivariate variable of size q = 10 we use a covariance matrix of size (10 × 10) and mean of size (10 × 1). ϕ1q .. ..2. T ... Θ(Wts )) (0) (0) (0) This gives us an initial set of variables for Gibbs sampling. . .. . We wish to generate a multivariate Yt  (j) ∼ Nq (µ. yi+1 . . qT ] we wish to generate wind speeds wi subject to the measured ten-minute average wind speed Wts : 1 s s s (wi + wi+1 + . .. . . . . ϕqq         are given (see below). ϕi1 . .. A covariance matrix for a set of one-minute wind speed data recorded with MST Radar near Aberystwyth is first calculated for a lag L = 1440 minutes. Here are the main steps of j-th iteration: 1. . Both can be found as a conditional covariance ϕqq|(L−q) and a conditional mean µq|(L−q) as was shown in section 2. ϕij . where  µ1  . φ). .numbers for wet days) for one iteration of the Gibbs sampling. + wi+q−1 ) = Wts . . hence q = 10. ϕq1 .1..... . This can be achieved through a sequence of steps similar to that described for two-variable sampling in Section 2. . Θ(Wts )... s On the interval [0. . The repeated subproblem when using Gibbs sampling to generating samples with given means. .. This means that many samples were rejected. .. . . ... We are interested in generating ten minutes of wind speed at any single time. . . .

038 0.094 0.014 2.011 0.000 0.017 0.007 0.023 0.000 0.111 0.000 0.000 0.008 0.255 40 .003 0.012 0.056 0.013 0.2.012 0.111 0.000 0. Iq ) is generated.092 0.088 0.009 0.006 0.040 0.015 0.021 0.041 0.025 0.041 0.018 0.000 0.000 0.013 0.000 0.041 0.013 0.021 0.275 0.021 0.016 such that M M T = ϕ 0.094 0.076 0.012 0.000 0.015 0.000 0.006 0.108 0.005 0.000 0.000 0.106 0.040 0.009 0.007 0.032 0.013 0.276 0. Matrix M can be found using the algorithm demonstrated in Section 2.041 ∗(j) ∗(j) ∗(j) 0.1.022 0.276 0. matrix M can be found as in Table 2.031 0.017 0.016 0.016 0.010 0.030 0.depend on the values of adjacent normalised wind speeds and has the same values for any sample.036 0. is defined by the normalised wind speed values of the partition (L − q). For the covariance matrix ϕ presented in Table 2.000 0.023 0.000 0.008 0.1.000 0.040 0.000 0. Assume a point Ut = (ui .000 0.000 0. .089 0.009 0.039 0.025 0.011 0.059 0.040 0.031 0.023 0. It is calculated for every sample separately.061 0.000 0.000 0.1.270 0.000 0.000 0.110 0.031 0. .009 0.000 0.028 0.010 0.000 0.015 0.006 0.023 0. ui+q−1 ) ∼ Nq (0.273 0.000 0.005 0.000 0.038 0.1: Covariance matrix ϕ calculated for the normal variables transformed from the minute-by-minute wind speed data 0.023 0.000 0.017 0.038 0.041 0. however.000 0.015 0.009 0.033 0.025 0.000 0.027 0.060 0.019 0.038 0.015 0.013 0. .000 0.007 0. ϕ).010 0.000 0.017 0.000 0.009 0.031 0.000 0.276 0.094 0.028 0.006 0.016 0.009 0.007 0.000 0.032 0.102 0.019 0.005 0.007 0.089 0.000 0.020 0.000 0.041 0.009 0.011 0.023 0.011 0.000 0. Table 0.000 0.021 0.2: M-matrix 0.041 0.3. It is then transferred to the required conditional normal probability distribution Nq (µ.003 0.008 0.038 0.025 0.113 0.000 0.000 0. Table 2.274 0.000 0.023 0.016 0.007 0. .000 0.008 0.011 0.275 0. The calculated conditional mean.000 0.012 0.054 0.092 0.012 0.050 0.057 0.023 0.000 0.005 0.094 0.041 0.112 0.000 0.016 0.022 0.011 0.040 0.016 0.276 0.076 Conditional covariance for our illustrative example is presented in the Table 2. ∗(j) ∗(j) = µ + M Ut Yt where µ is a conditional mean and a conditional covariance matrix ϕ = M M T .023 0.010 0.

. .35) Calculated with (2. yi+q−1 ). . By advancing λ from one border of the interval to another with a step 0. . . yi+q−1 ) acquires a value from the vector of maximum probabilities of the truncated normal probability distribution. . . . . . Information carried by each of the ten generated variables is calculated as a sum of elements in every column of the covariance matrix: η = (η1 . (2.2.23). . yi+q−1 ) is shifted in order to satisfy Wts as explained in steps 3 and 4 of this algorithm. 3. .001. . . Points Y C and Y D are found during the search. so the following applies. . ϕ is a covariance where Yt matrix of distribution [Yt ] and 1 is a vector of the same size as Yt and consists of scalars 1 only. If Wt matches ∗s ∗s s Wt within a specified tolerance. η2 . . . . . that are sufficiently large to find Y in (2. wi+q−1 ) 1 ∗s ∗s = (wi + . . wi+q−1 ) is accepted. . As in the case of two variables a shifted point (yi . .36) is a point sampled in step 1 of this algorithm. we search along the line Yt ∗(j) (j) ∗(j) ∗(j) ∗(j) ∗(j) ∗(j) s(j) = −λ1T ϕ + Yt ∗(j) . . . is shifted is defined by the covariance matrix Direction in which point Yt ϕ and evaluates how much information each of ten variables contributes to the average wind speed Wts . . yi+q−1 ) is tested. The generated point Yt := (yi . + wi+q−1 ) q where Wt is an average wind speed corresponding to the generated earlier ∗(j) ∗(j) values of the normalised wind speed (yi . . . . . ∗(j) ∗(j) Otherwise (yi .35) new values of wind speed are Wt with ten-minute average wind speed equals Wt s(j) s(j) s(j) ∗s ∗s := (wi . . We assumed again λ to vary in the interval between −10 and 10. Generated normalised wind speed is transformed further using the inverse of function Θ: s(j) ∗(j) Wt = Θ−1 (Yt ) (2. the point (wi . ηq ) where q ∗(j) ηj = i=1 ϕij 41 .

e. Y = 1T Y B . where b3 is a parameter of the quadratic function in (2.11 0.13 0. Definition Let us denote Y B as a vector that satisfies the equation f (Yt ) = Wts and acquires the shortest distance to the center of the distribution Yt . a line of the form 1T Yt = Y as defined in the Section 2. Y C (and a corresponding λC ) are found so that 1T Y C and Y match to a small given tolerance.2.9: Amount of the information carried by each of the ten variables Point Yt is shifted in the specified above direction η towards a curve that presents a required sum of q wind speed values in terms of a normal variable.15 information 0.1 0 1 2 3 4 5 order of variable 6 7 8 9 Figure 2.e. Definition Let us denote Y D as the point on the line (2. f (Yt ) = Wts .12 0.Figure 2. By modifying the value of λ. and we assume Y = b3 . ∗(j) if 1T Y B > 1T µ.17 0. Definition Let us denote Y C as a vector on the intersection of the line (2.2. A scalar Y is calculated as in the section 2.9 plots the amount of information carried by each of the ten generated points in the total sum. 0.14 0. i.18).36) that satisfies the condition of ten one-minute wind speeds being equal a given average 42 . i.5 of this work.5: If 1T Y B ≤ 1T µ.36) and the tangent to the curve f (Yt ) = Wts .16 0.

For an interval of 5 years when the average wind speed is given.10 it necessary to evaluate it statistically. there are j ∗ = 12 Gibbs sampling iterations being performed. Else Finish j-th iteration of Gibbs sampling algorithm. Out of five years of ten-minute average wind speed data records there were between 34 to 87 rejects reported for a single iteration. The probabilities of the points Y D and Y C are further compared: If φ(Y D ) >ρ φ(Y C ) accept Yt (j) =YC Else reject Y C and Go to step 1. 4.3.value f (Yt ) = Wts .4 Testing the results of wind speed sampling On Figure 2.10 there is one hour of generated wind speed plotted along with original wind speed. Y D (and a corresponding λD ) are found so that f (Y D ) and and Wts match within a small given tolerance. Yl ∗ (j ∗ ) ] = ϕ(l) No requirements are placed on moments higher than second order. shifting a generated point is an improvement to the method used by Allcroft and Glasbey in [17]. As in the case of two-variable sampling. even with high wind speeds it rejects very view trial points. Take i = i + 1. Let us begin by defining a second-order stationary process. we note that the form of a stationary autocovariance function implies that 43 . We determine j ∗ so that a stochastic process of the resulting sample is weakly stationary. so that ∗ E[Y (j ) ] = µ and Cov[Y (j ) . By letting l = 0. If i ≤ T Go to step 1. Therefore. A process is called second-order stationary (or weakly stationary) if its mean is constant and its autocovariance function depends only on the lag [43]. where φ(∗) is the probability distribution function for truncated normal distribution used for sampling wind speed. 5. we assumed ρ to be a fixed number 0. By modifying the value of λ. 2. Steps 1 to 5 are repeated until j reaches a specified number of iterations j ∗ . However good the generated sample can look on Figure 2.95.

2412079 ∗ ∗ while µ = 0. as well as the mean.2. it reflects the mean of truncated normal distribution described in the section 2. which is 2592000 points in the data set. One day is chosen as a significant test interval to compare the autocovariance of the sampled and the original datasets where they are very close to each other. 25 original sampled 10 min average of original 10 min average of sampled 20 wind wind wind wind speed speed speed speed wind speed. Let us analyse the first two moments of the probability distribution for the original and sampled data sets. Let us take 5 years of the recorded one-minute wind speed for the analysis. The first moment of the probability distribution is calculated as an expected value for the sampled normalised wind speed. Figure 2. µ − E(Y (j ) ) < 10−3 . The second moment of the probability distribution is calculated with the autocovariance coefficients of the sampled normal variables.2419591 where Y (j ) is the j ∗ -th sample of the Gibbs sampling method.4: E[Y (j ) ] = 0.10: An example of the sampled and the original values of 1-minute wind speed plotted along with the ten minute average wind speed The chains encountered in Monte Carlo Markov Chain settings satisfy a stability property. min 180 200 220 Figure 2.the variance. We see that the difference between the mean of the i-th sample and the original ∗ data is insignificant. (j ∗ ) then Yi+1 ∼ [Y ]. m/s 15 10 5 0 100 120 140 160 time. namely a stationary probability distribution that exists by con(j ∗ ) struction. 44 . if there exists a distribution [Y ] such that Yi ∼ [Y ]. This means.11 plots the autocovariance functions of the original and sampled normalised wind speeds. is constant.

1 0 0 200 400 600 800 lag.025126 0.025126 6 7 0.0. Table 2.027672 0.3 0.3: Variance of the original and the sampled normalised wind speed shift original data sampled data 0 0. min 1000 1200 1400 Figure 2.6 original normalised wind speed sampled normalised wind speed 0.025126 45 .027672 0.025126 8 0. As data was sampled for ten-minute sections it is important to ensure that there is no boundary effect appearing in the results.027672 0.025126 5 0.025126 1 2 0.025126 3 0.027672 0.2 0.027672 0. variance (2.37) is calculated for shifted ten-minute intervals.027672 0. To test this. 1 var(Y ) = Tq j∗ ∗ Tq j (yt − Yij )2 ∗ ∗ (2.37) t=1 j where yt are the sampled normalised wind speeds from the ten-minute interval i ∗ with the average Yij .027672 0.027672 0.025126 4 0. The variance is calculated for all consecutive groups of ten minute wind speeds.11: Autocovariance of the sampled and the original normalised 1-minute wind speed For the next test we calculate variance of one-minute wind speeds inside ten-minute time sections.027672 0.5 0.027672 0.4 autocovariance 0.025126 0.025126 9 0.

min 80 100 120 Figure 2. We applied further the modified Gibbs sampling to generation of one-minute wind speed using ten-minute wind speed provided by the Utah Geological Survey.12: An example of wind speed sampled for one of the Utah sites We chose 22 different sites from the Utah state where ten-minute wind speed was recorded by Utah Geological Survey. For every site we generated a valid sample of one-minute wind 46 . m/s 10 8 6 4 2 0 0 20 40 60 time. 2. we require a dataset of geographically distributed one-minute wind speed in order to model short-term electricity generation. They are distributed around the state that introduces an effect of the geographical diversity into the accumulated wind energy curve. Utah Geological Survey publishes ten-minute wind speed ([21]) recorded in different sites of the Utah state of the United States of America.2 of this work introduced Gibbs sampling algorithm and in Section 2. Statistical characteristics of one-minute wind speed were based on the data received with the MST Radar located at the Frongoch farm in Wales ([20]) that allowed us to study short-term wind electricity generation with wind energy in Great Britain. . This supports a statement that there are no boundary effect appearing in the sampled data.3 we originally adapted this algorithm so that one-minute wind speed is sampled when the average value of ten-minute wind speed is given. 16 wind speed 14 12 wind speed. T q cumulative variance has the same value 0. Section 2.3 shows that independent of the start point for the grouping of the interval 1 . .4 Diversification of wind energy As it was stated at the beginning of this chapter.025126.Table 2.

12 plots an example of wind speed sampled for one of the sites in Utah. Considering a geographical diversity of wind farms in Great Britain we wish to use the combined wind energy from 22 sites in the Utah state for the optimization modelling of the electricity generation further in this work.1. 70 wind turbines at 22 sites 22 wind turbines at one site 60 50 energy. 47 .3. Notice.speed using the modified Gibbs sampling algorithm explained in Section 2. so that a given value of ten-minute wind speed is achieved.13: An example of geographically distributed wind energy output We sampled 22 sets of one-minute wind speed for different sites in Utah that were further transfered into energy using the wind power curve plotted in Figure 2.12. To demonstrate this. MW 40 30 20 10 0 0 20 40 60 time. Figure 2. variability of the combined wind energy from 22 sites in Utah varies less than the same amount of energy produced at one site.13 plots the combined wind energy that could have been generated from 22 sites in Utah against wind energy that could have been generated by 22 wind turbines using a single wind speed sample plotted in Figure 2. Figure 2. min 80 100 120 Figure 2.3.

leaving the matter of 48 . therefore. This chapter specifies the details of the electricity demand and generation in the United Kingdom and formulates a mathematical programming problem of scheduling the operating reserve so that the supply and demand match within the target range. the formulated optimization model is deterministic and includes some of the common constraints of power generation. demand. Therefore. We assume that the electricity load only slowly changes.Chapter 3 Modelling Electricity Generation with Wind Energy Scheduling electricity generation is a complex engineering problem that takes into account various parameters of generation. In Chapter 2 we developed an algorithm that would produce one-minute wind energy data. this section focuses on the parameters of thermal generation and its fuel cost. We assume in this chapter that wind speed is forecast with certainty. transmission and distribution.1 introduces the electricity generation system of the United Kingdom and how it can be adapted for the optimization modelling. We use the formulated optimization model to investigate the costs involved in incorporating different levels of wind energy into the electricity generation system of Great Britain. 3. The uniqueness of the model developed in this chapter consists in the data representing the power system of the UK and the application of the optimization model to estimating the system balancing cost of the variable wind energy.1 Elements of the British power system and simplifying assumptions Section 3.

3. According to the figures in Table 3. like gas. Various nuclear generators contribute to another 14% of the total generating capacity.1 two types of plants that cover more than 70% of the total British electricity generation are Gas and Coal turbines.1 Electricity generation Currently electricity generation in the United Kingdom includes a diversity of generating units. Table 3. The capacity limit of any type of the generator depends on the size of the unit and can vary within the same type of generators. gas and nuclear units have common characteristics with other units of the same type that we wish to incorporate in the model. The amount of generation provided by different types of generating units vary depending on the time of the year and the size of the load. for the modelling purpose we wish to use a single value as a capacity limit for every type of a same plant. is a capacity limit. Capacity limit is specified in the BM reporting system as the Export Limits parameter (defined in Appendix A) and describes a maximum level at which the generating unit may be exporting power to the transmission system. coal or nuclear. 49 .interaction between the demand and wind power for further research. Let us use a capacity limit for a typical gas generator as 500MW and 600MW for typical coal and nuclear types of generating units. For the purpose of this work we use Plant/Demand Balance of the “Seven Year Statement” published by the National Grid Electricity Transmission in 2008 [35]. Coal.1. These rates depend on the current output of the generator and are used by the System Operator in scheduling the operating Reserve when balancing the supply and demand in real-time.1 shows the percent of power from the various plant types in the total generation. in particular the available capacity and the maximum rates of change in the power output. Parameters describing a certain type of the generator are available through the Balancing Mechanism (BM) reporting system ([5]). The output levels of thermal plants are not constant and can vary following the changes in the electricity load. However. The maximum rate at which units change their production can be described by Run-Up and Run-Down export rates (defined in Appendix A). We wish to incorporate some of the characteristics into the mathematical programming model. This website publishes on a daily basis Physical and Dynamic parameters (as defined in Appendix A) of the units participating in the balancing mechanism. The first characteristic that the System Operator takes into account while planning an output of a thermal generator.

Table 3.47 Medium coal+AGT 1102 1.41 Small coal 783 1.94 Wind 1597 2.68 Nuclear Magnox 1450 1.85 Nuclear PWR 1190 1. There are usually three Run-Up and Run-Down export rates. 50 .75 Offshore wind 140 0. And a maximum Run-Down export rate r3 = 58.5 MW per minute can be applied when a coal generator cools down between 234 MW and 0. Run-Up rate r1 = 5.04 The Run-Up and Run-Down export rates specified in the BM reporting system are further adjusted to be used in the model.52 OCGT 589 0.00 Nuclear AGR 8365 10.46 Pumped storage 2300 2.0 Run-Down 5. Every rate is applied to a different output level specified by the “elbows”.54 Large coal 4413 5.1: Plant/Demand balance for SYS Background (2007/2008) hline Plant type Capacity (MW) % biomass 45 0.31 Interconnector 1988 2. Let us consider the following example: Table 3.0 177 0.2: An example of the Run-Up and Run-Down export rates for a typical coal plant Export rate r1 e1 r2 e2 r3 Run-Up 5.4 234 58.0 MW per minute is applied to the production level from 0 to 177 MW. The rates for different power output can cause a non-convex constraint in the optimization model. r1 .58 CHP 1725 2. Maximum Run-Up and Run-Down rates are published in the form of a table with three different values for each of the rates that depend on a production level.5 Table 3.63 Large coal+AGT 21462 27. e1 and e2 . r2 and r3 submitted by the generating unit before the Gate Closure.2 gives an example of the Run-Up and Run-Down export rates published for a typical coal plant by the Balancing Mechanism reporting system. So.06 CCGT 25532 32.2 180 5.18 Oil+AGT 3496 4.20 Hydro 1028 1.39 Medium coal 1152 1.0 240 0.

Therefore.2. Later in this chapter we modify the values of the remaining rates r1 and r3 so that they form a convex feasible set.The specified sequence of maximum Run-Down and Run-Up export rates leads to the non-convex constraint describing the power output rate of change between the current and the following minute. Also governmental organisation “Department for Business Enterprise and Regulatory Reform” (BERR) regularly publishes the average prices of fuels purchased by the major UK producers.2.1. The resulting constraints will be formulated in Section 3. Considering a limit of water resources in United Kingdom and capital cost of pumped storage generation we wish to avoid modelling hydro generation in this work. we wish to simplify the rates presented in Table 3. however. there is not enough time to make a decision about switching additional thermal plants. Following this.2 Cost of generation Due to the fact that the electricity generation is scheduled minute by minute during the real time.2. the installed wind energy capacity have doubled during the last year and is expected further to increase. 3. the rate r2 can be dropped as it is likely to be an internal work taking place in the generating unit and is not important for our work. Every site is assumed to contribute an equal number of wind turbines to the total wind energy curve. Short time scale water reservoir limits are often not building and a hydro generator can be modelled in the same way as a thermal generator except that the “fuel” is water. According to the “Seven Year Statement” published in the beginning of 2008 wind contributes only 2% to the total electricity generation.1 also shows that the hydro and pumped storage contribute no more than 5% to the total electricity generation. generating cost of wind and nuclear energy is zero and cost of gas and coal is defined by the fuel prices. We wish to introduce the wind energy output into the electricity generation system based on the wind speed data provided by the Utah Geological Survey. Hence. First. Short term scheduling determines a cost of the electricity generation as the fuel cost for all types of generators. This curve is further scaled so that wind contributes a required amount of energy to the total electricity generation. In the report “Keeping the lights on: Nuclear Renewables and Climate Change” Milborrow [41] discusses current cost of the electricity generation and the way it can be calculated. For some coal generators it takes up to 2 hours to be warm and ready to respond. Table 3. We wish to use the quarterly 51 . we assume that only generators that are currently onstream can be used to meet the load.

25£/MW*min Note. However.1: Half-an-hour average of the daily electricity demand in the United Kingdom and its piecewise linear approximation with 1-minute frequency As the electricity demand is available in half-an-hour average estimates we wish to transfer it into a continuous piecewise linear function with one minute frequency.3 Electricity demand We wish to model the electricity demand with the same time frequency as the generation. on a minute by minute basis. MW*min 44000 42000 40000 38000 36000 34000 32000 0 200 400 600 800 time. a most frequent electricity demand reported through the Balancing Mechanism reporting system is the historic and real-time records of the national demand. 3.1.72p/KWh cgas = 1. min 1000 1200 1400 1600 Figure 3. 52 .5 p/KWh. i.1 shows an example of the winter electricity demand used later for the deterministic and stochastic models. there is no capital cost included in the cost of the electricity generation.12£/MW*min = 0.5p/KWh = 0. 50000 1-minute curve half hour average 48000 46000 outturn demand.72 p/KWh. An average price of coal for power stations was increasing since 2007 and amounted to 0.energy prices reported by BERR for year 2008 [7]. Figure 3.e. The price of gas has increased by more than 50% up to 1. In order to apply these prices to the electricity generation in one-minute frequency we recalculate the given prices in £/MW*min. we are interested in minimizing the fuel cost later in this work. ccoal = 0.

Further.1. Transmission system is not modelled in this work.3. the solution tends to produce an output of the electricity generation 2% below the electricity demand. This can get the transmission system into zones of too high or too low frequency. based on a given deviation in frequency is complex: it depends on the system parameters and the load level. Although it is possible for the generation to deviate from the demand it is not a good practise to get close to the upper and lower limits. To prevent this happening let us introduce a penalty of mismatch between the demand and generation. The transformation function that calculates the deviation in power. therefore. We wish to incorporate an option of the mismatch between the electricity generation and demand in the model. the generation is allowed to vary 2% above or below the electricity load which corresponds to the 1% variation in frequency. Considering the uncertainty of consumer’s behaviour and a probability of losing a generating capacity in the system. As a result of modelling minimization of the electricity generating costs. As it was stated in Chapter 1 the transmission system of the United Kingdom operates at the frequency 50Hz with a range of 1% variation above and below 50Hz.2: A mismatch between the electricity demand and the generation with the cost attached 53 . a mismatch between the generation and the electricity demand that corresponds to the variations in the frequency is assumed to be a function of the electricity demand. it is important to have a flexibility of meeting the demand. Figure 3.4 Cost of demand-and-supply mismatch In reality it is very hard to generate exactly the value of the national electricity demand at any moment.

coal and nuclear) included in the model. The cost c+ is 3 significantly higher than the cost of the marginal generator so that the mismatch between the electricity demand and generation exceeds d+ + d+ only if no other 1 2 feasible solution exists. there is no transmission cost. Time scale is short so there is no start-up costs of thermal generation. higher cost of mismatch encourages the generation to meet the demand exactly. To estimate the system balancing cost when variable wind energy is incorporated into the electricity generation system we assume: a.2 the cost c+ corresponding to the mismatch between the 1 electricity demand and the generation d+ is marginally higher than the fuel cost 1 of the gas generator.1. Gas turbines have the highest fuel cost among the thermal plants but the Run-Up and Run-Down rates are also higher than that of the coal and nuclear generators. wind power generation is determined by the incoming data and is perfectly forecast for the deterministic problem. If gas turbines are able to cope with the load fluctuations. They are characterised by the limited capacity and the Run-Up and Run-Down rates. 54 . However. d. c. gas turbines are used first to respond to the mismatch between the electricity demand and generation. there are three types of thermal generators (gas. b. so that a cost of the mismatch is estimated according to the fuel cost of the gas generator. 3. there is only generating (fuel) cost considered in the problem. In Figure 3.5 Summary Let us summarise the above points in order to formulate a deterministic problem of the electricity generation. Gas generating units are assumed to be a marginal type of the electricity generation. Determined as the marginal generator. there is no hydro energy included in the model. it is rather allowed for the generation to deviate from the demand than to modify the power output of the coal or gas turbines.Let us divide the allowed mismatch between the electricity demand and generation into several intervals so that the cost of a mismatch increases with the higher deviation between the demand and generation. e.

the set of the wind generators.1 Decision variables Let T denote the number of time intervals obtained by discretising the planning horizon. T . The decision variables in the model correspond to the outputs of each thermal unit.2. system security and others). The decision variables are denoted by xtg . We assume a constant power output during 1 minute and measure it in MW*min. g ∈ W (3.f. . the set of the conventional generators. g. This discretisation may be chosen uniformly or non-uniformly. Wind power generation is determined by the wind speed and equals to the e power wt at time t: e xtg = wt . 3. the electric power generated by the Coal. . Let G denote the set of all the generators and consist of the union of C. g ∈ C where xtg is the production level of thermal unit g during time period t. Nuclear and Gas generating plants.1. 2. . T.1) 3. The set of conventional generators also consists of Coal.2. . Decision variables corresponding to the thermal generation mentioned above have finite upper and lower bounds representing 55 . . t = 1. 2. For the current problem we use one-minute uniform discretization of the planning horizon. i. 3.2 Constraints The electricity generation system is complex so it is possible to include different constraints in its optimization model (capacity.2 Deterministic model for the electricity generation with wind energy This section formulates a linear programming (LP) mathematical optimization model for the electricity generation scheduling problem specified in Section 3. and W . emission limits. t = 1. .e. . . there is a cost of mismatch between the electricity demand and generation. Gas and Nuclear generators. The model described in this section would contain only those constraints that allow us studying a problem of incorporating the wind energy into the electricity generation system. the electricity demand is deterministic.

3) − (3.unit capacity limits of the generation system: xg ≤ xtg ≤ xg .1. while r1− and r2− .1 we use only two Run-Up and two Run-Down rates that are connected in a piecewise linear function and form a convex set. (T − 1). . Hence.1. . Figure 3.1. 2. 2. . As mentioned in Section 3. g ∈ C of the decision variables. e+ 56 t = 1. . . T. let us modify the limits of maximum RunUp and Run-Down rates in a way that they form a convex set (xtg . . An example of Run-Up and Run-Down export rates that a generating plant provides to the System operator before the “Gate closure” is given in Table 3. Constraints (3. As it was stated then. t = 1. .6) formulated below relax the limits of Run-Up and Run-Down rates of generating plants. g ∈ C (3.3) . x(t+1)g ). . t = 1. g ∈ C (3. The corresponding constraints of the change in the power output are described by inequalities: xt+1.2. 2. .denote two Run Down rates separated by the “elbow” e− .3: Area formed by Run-Up and Run-Down constraints for thermal generators Further single-unit constraints describe maximum Run-Up and Run-Down export rates introduced in Section 3.g + r1+ . . . .g ≤ e+ + r2+ − r1+ xt. The piecewise linear function can be chosen in a way that tighten or relax the maximum export rate limits.2) The constants xg and xg denote minimal and maximal possible outputs of the thermal units respectively. T. Let r1+ and r2+ denote two Run Up rates separated by the “elbow” e+ . these export rates cause a non-convex constraint.

.6) t = 1. 2. MW*min 400 600 800 Figure 3. xg − e− xg − e− xt+1. t = 1.g + r2+ .3 plots a convex set (xtg .5) − (3. 2. T. and xt+1.3 intersect in a point of the “elbow” e+ and are described by the inequalities (3. g ∈ C (3. Run-Down rates are described by the inequalities (3. . g ∈ C Figure 3.g − r . . . g ∈ C with new Run-Up and Run-Down export rates in a form of the shaded area bordered by two lines from the above and two lines from the below. 2− t = 1. 140000 excessive generation insufficient generation gas turbine cost 120000 100000 80000 cost. their gradients are r1− and r2− respectively. GBP/MW*min 60000 40000 20000 0 -20000 -40000 -800 -600 -400 -200 0 200 deviation from demand. Every line on the graph states a maximum rate of change between the output at the current minute and the following minute. It is essential for the operation of the power system and means that the difference between the total power generation and the load demand does not exceed the allowed (by the transmission system) limit at any time.4: Load mismatch costs: the difference between the curves gives the penalty for generation deviating from demand. . T.g ≥ t = 1. . . 2.3) − (3. . . .g ≤ xt. .4) r2− xg − r1− e− xg − e− + r2− − r1− xtg − .5) (3.6). . The two lines from the below intersect in a point of the “elbow” e− . The two lines from the above in Figure 3. T.xt+1. g ∈ C (3. x(t+1)g ).g ≥ xt. the loading constraint is described by the 57 . 2. The loading constraint couples across all the generating units. . . . Denoting by Dt the electricity demand during a time period t and by ∆D a percent mismatch allowed between the electricity load and the generation. (T −1). .4).

. . ((1 + ∆d )Dt )]. . . . dots. . ζv }.7) a parameter ∆D is taken as 2% that corresponds to 1% variation in the system frequency. 2.8) T denotes the planning horizon while V is a number of the cost bands in the accepted mismatch between the supply and demand [(1 − ∆d )Dt . . this piecewise linear load imbalance cost should be above the cost of the marginal generator at that time. . . The higher mismatch between the electricity demand and generation the higher the cost of coping with it.1 we discussed a cost of the mismatch between the electricity supply and demand. To achieve this result in the model we introduce a piecewise linear cost for the load imbalance + + {ζv . T.inequalities: (1 − ∆D )Dt ≤ xtg ≤ (1 + ∆D )Dt t = 1. Set of parameters [d+ . It can be calculated in the following way + d+ = ∆d ∗ Dt ∗ βv tv (3. To avoid giving an incentive to violate the load constraint. V where V denotes a number of bands in the penalty cost. Logically the generators should produce close to the demand and as far away from the lower and upper bounds as possible. v = 1.9) (3.8) g∈G v∈1. .1. . . . .11) 58 . . Dt ] and [Dt .7) g∈G In (3. g∈G xtg < Dt . T (3. Figure 3. T (3. 2. d− ] is time dependent and defined by the changing tv tv value of demand.10) where d+ is an upper bound of interval v when the electricity demand exceeds tv the generation at time t.4 plots the fuel cost of the marginal generator (gas in our case) and three different cost bands for the generation exceeding the demand and the demand exceeding the generation. In Section 3. . Let us denote a mismatch between the electricity generation and the load by − a set of variables d+ and dtv . 2. Size of every interval v = 1.V In (3. t = 1. .. 2. . 2. V such that tv xtg − Dt = (d+ − d− ) tv tv t = 1. 2. v = 1. V is measured in MW and defined by the inequalities: 0 ≤ d+ ≤ d+ tv tv 0 ≤ d− ≤ d− tv tv (3. d− is an upper bound of interval v tv when the generation exceeds the demand at time t. g∈G xtg > Dt . . . .

t=1 g∈C In order to decrease a mismatch between the electricity generation and demand we wish to include the cumulative cost of the mismatch into the objective.3 Objective function The constraints described above form a feasible set of solutions for our optimization problem of scheduling the electricity generation.where βv determines a part of the demand envelope ∆d ∗ Dt that corresponds to the cost band v = 1. the objective of the mathematical programming model is to minimize the cumulative operational cost of the electricity for the planning horizon T : T V T + ζv v=1 t=1 V T − ζv v=1 t=1 min x g∈C cg t=1 xtg + d+ tv + d− tv (3. 2. 3. . let us introduce the operational cost of the electricity generation for different types of plant.2. . .2. Fuel cost of the coal and gas generation was introduced in the section 3.12) t=1 xtg denotes the fuel cost of the g∈C cg V T V T + − + − thermal generators g ∈ C and v=1 ζv t=1 dtv + v=1 ζv t=1 dtv denotes the cost of mismatch between the electricity supply and demand.12) T In the objective function (3. . To complete our LP model.1. 59 . Then g∈C cg T xtg is the operational cost paid by the Supplier to generate t=1 T xtg amount of electricity. V . + − If parameters ζv and ζv denote the cost of the excessive and insufficient generation respectively. we wish to minimize the total cost of the electricity demand and generation mismatch: V + ζv v=1 T V T − ζv v=1 t=1 ( t=1 g∈G xtg − Dt ) + (Dt − − T − ζv xtg ) g∈G If we define d+ := T ( g∈G xtg − Dt ) and d− := tv tv t=1 can equivalently be rewritten as: V + ζv v=1 t=1 T T t=1 (Dt g∈G xtg ) this V d+ tv + v=1 d− tv t=1 Taking into consideration the fuel cost and the cost of the load imbalance. Let cg denote the fuel cost for each thermal generator g ∈ C.

4 Formulation of the LP deterministic optimization model The constraints and and the objective described in Section 3. V ∀v = 1. . . 2. . .g + r1+ e+ x(t+1). . .g + r2+ xg ≤ xtg ≤ xg ∀g ∈ W ∀g ∈ C ∀g ∈ C ∀g ∈ C ∀g ∈ C ∀g ∈ C x(t+1). . The optimal solution is chosen according to the minimization function that includes fuel cost of the 60 . 2.g − r2− xtg ≤ (1 + ∆D )Dt V g∈G g∈G xtg − Dt = v=1 (d+ − d− ) tv tv ∀v = 1. .2. V ∀v = 1. 2.2. V + d+ = ∆d ∗ Dt ∗ βv tv − d− = ∆d ∗ Dt ∗ βv tv 0 ≤ d− ≤ d− tv tv 0 ≤ d+ ≤ d+ tv tv where t = 1. T . . . V ∀v = 1. 2. .2 represents a series of one-minute electricity generation level for every thermal generator g ∈ C over a time period of T minutes.g ≥ xt.3 Calculating cost of wind energy variability An output of the mathematical optimization model (3.3.g ≥ xg − e− + r2− − r1− r2− xg − r1− e− xtg − xg − e− xg − e− (1 − ∆D )Dt ≤ x(t+1). . .2 and Section 3. .12) formulated in Section 3. . . . 2.3 are combined below: Objective (minimize cost): T V T + ζv v=1 t=1 V T − ζv v=1 t=1 min x g∈C cg t=1 xtg + + d+ tv + d− tv Subject to: xtg = wtg e+ + r2+ − r1+ xt.1) − (3.g ≤ x(t+1).2. . .g ≤ xt. 3. .

2.1 Case of constant wind energy that reduces the operational cost First let us consider a case when a wind energy profile if flat. . there are no fluctuations in the wind output. We used version 10. Then. i. .e. Assuming further one accumulated wind energy curve in the set of wind generators. thermal generators must frequently modify their production level in order to balance these fluctuations in the power system. Each of these thermal generators represent a big number of plants with similar fuel costs and capacity limits. i. a number of constraints results in 60480 for a daily electricity generation problem. There are three types of thermal generators used to meet the electricity load: nuclear.thermal generators and the cost of a mismatch between the electricity supply and demand. t = 1. But this advantage of wind energy is challenged by its variability. coal and gas generators. 3. These generating units can be ordered by the fuel cost and the output flexibility. . 61 . where g ∈ C. The deterministic LP optimization model formulated in Section 3. . Formulated deterministic model can be used now to estimate the system balancing cost that appears as a result of wind energy variability. 2. V . . a total number of decision variables in the model equals κ∗T +V ∗T . taking V = 6 there are 12960 decision variables in the deterministic optimization problem of daily (T=1440) electricity generation. If a significant amount of the wind energy fluctuates. gas and coal. Fluctuations of the wind energy result not only in the additional fuel cost but also the cost of a mismatch between the electricity generation and the load. In order to estimate this cost we compare the results of two different cases.e. T and v = 1.4 includes two sets of variables {xtg } and {dtv }. . To reduce a number of variables we wish to group thermal generators so that there are only three generators included in the set C: nuclear. a size of the maximum Run-Up and Run-Down export rates. Formulated deterministic optimization problem is linear so that it can be efficiently solved with a CPLEX solver. The power system benefits initially from the wind energy being present in the system as there is no fuel cost associated with the generation of wind energy.3. . If we denote κ as a number of thermal generators in the set C. 2. .00 of the CPLEX solver that resulted in CPU time < 1s.

20000 18000 16000 14000 12000

power, MW

10000 8000 6000 4000 2000 0 0 20 40 60 time, min 80 100 120 gas coal nuclear wind

Figure 3.5: Electricity generation when 11% of wind energy with flat profile is introduced to the system Figure 3.5 plots an example of the optimization model output, which is a generation level for the thermal plants as well as the available wind energy with the flat profile. Nuclear generators are the cheapest, we price the fuel cost as zero in the model. It is also the least flexible type of generation as it has the lowest Run-Up and Run-Down rates. Gas turbines are the most expensive type of generation in the model but they are the most flexible generators. As a result the model output shows the nuclear and coal generation at the maximum available capacity and gas generators produce the remaining electricity to meet the load. Definition Let us denote a flexible generation producing electricity away from its upper and lower boundaries as the marginal generation with the fuel cost equals cm . Marginal generators are the first to react on changes in the electricity load. Wind energy is constant for this solution with the output of 5173 MW*min that makes up to 11% of the load in two hours. The cumulative fuel cost corresponding to this problem of scheduling the generation is reduced by the fuel cost of the marginal generator that would be used if there was no wind present in the system.
f Definition Let us denote Fw lat as the cumulative fuel cost of the mathematical + optimization problem with flat wind energy profile and fw as the saving in fuel cost by the Supplier of the electricity when it uses the wind energy in the total generation.

62

f Cumulative fuel cost Fw lat is related to the problem with no wind energy in the system in the following way: f + Fw lat = F0 − fw

(3.13)

In (3.13), F0 is the fuel cost related to the optimization problem with no wind energy available. The cost of the mismatch between the electricity demand and the generation equals zero in the current case as the load can be met exactly.

3.3.2

Case of variable wind energy that increases operational cost

For the next case a profile of the wind energy curve is variable. It is balanced by the group of thermal generators: nuclear, coal and gas. As in the previous case they are ordered by the fuel cost and maximum Run-Up and Run-Down export rates.
18000 gas coal nuclear wind

16000

14000

power output, MW

12000

10000

8000

6000

4000 0 20 40 60 time, min 80 100 120

Figure 3.6: Electricity generation when 11% of variable wind energy is introduced to the system

Figure 3.6 plots an example of the optimization model output, which is a generation level for the thermal plants as well as the available wind energy. Unlike the case presented in the Section 3.3.1, wind energy for the current problem fluctuates. Let the fluctuations of the wind energy curve be proportional to the amount introduced to the power system. Figure 3.6 plots 11% of the wind energy in total electricity generation of 1 day. In order to balance the variable 63

wind energy the thermal generators have to modify their production profile. Gas turbines have the highest Run-Up and Run-Down exports rates so it reflects most of the variation in the wind energy. However, gas generation is not able to cover some of the steep changes in wind energy from minute to minute. Then the coal generators modify the power output so that the supply and the electricity demand are within the target range. The nuclear plants have the lowest Run-Up and Run-Down rates which results in the output profile being flat at almost all the time t = 1, 2, . . . , T . Figure 3.6 shows that it is hard to meet the electricity demand exactly when variable wind energy is introduced into the power system. In the reality coal generators can not modify the power output as often. Therefore, we wish to allow the total generation to deviate from the electricity demand within a certain range. This can be achieved by setting the penalty cost of the first band v = 1 marginally higher than the fuel cost of the gas turbine that acts most of the time as the marginal generator. If the gas turbines are able to cope with the variation of the wind energy it is cheaper for the total generation not to deviate from the electricity demand. However, a small mismatch between the supply and the electricity demand is preferable to the modification of the coal power output as it is shown in Figure 3.7. The cumulative fuel cost corresponding to the problem of scheduling the electricity generation with the variable wind energy is affected not only by the fuel cost of the marginal generator but also by the fuel cost of the generators that modify the output profiles in order to balance the fluctuations of the wind energy.
var Definition Let us denote Fw as the cumulative fuel cost of the optimization − problem with variable wind energy profile and fw as the additional fuel cost of the generators that modify the output profiles in order to balance the fluctuations of the wind energy. var Cumulative fuel cost Fw is related to the problem with no wind energy in the system in the following way: var Fw = − + F0 − fw + fw − f Fw lat + fw ,

where F0 is the fuel cumulative cost related to the optimization problem with no f wind energy available and Fw lat was calculated in (3.14) as the accumulated fuel cost of the problem with the flat wind energy profile.

64

06 0.8: The fuel cost of the wind energy variability shown for different levels of it introduced into the electricity system The system balancing cost that appears when variable wind energy is introduced into the power system can be estimated as the difference between the fuel cost 65 .12 0.08 0.04 0.14 0 Figure 3.16 0.14 0.08 0.800 600 mismatch between generation and load.06 0.12 0.1 percent of wind power in total generation.04 0. % 0. 0.02 zero wind flat wind profile variable wind profile 0 0. min 80 100 120 Figure 3.18 operational cost weighted by total generation. The highest cost is the fuel cost calculated for a problem when there is no wind energy in the power system. MW 400 200 0 -200 -400 -600 -800 0 20 40 60 time.02 0. pounds/MW*min 0. So the weighted fuel cost is a piecewise linear function with the gradient of the marginal generator. The fuel cost from the bottom reflects a problem with the variable wind energy that can be balanced by the marginal generator.8 plots the cumulative fuel cost of a daily electricity generation weighted by the total amount of the produced power.7: Electricity generation and load mismatch when 5% of variable wind energy is introduced to the system Figure 3.1 0.

66 .calculated for the problem with variable wind energy that can be balanced with a combination of different thermal generators and the fuel cost calculated for a problem with variable wind energy that can be balanced by the marginal generator.

3. the general conclusions can be applicable for the case of slowly changing load. The objective function of the optimization models in this chapter is presented only by the accumulated fuel cost.1 Formulation of the problem To evaluate how well the statistical model estimates the system balancing cost.3. 67 .2) can be determined as a function of the statistical parameters of the wind energy and develop a unique model that estimates the parameters of such a model. 4.1 and Section 3. which means that there is no cost of mismatch between the electricity supply and demand. We assume further that the saving in fuel cost and the additional fuel cost (as defined in Section 3.Chapter 4 Statistical analysis of operational cost In the previous chapter we introduced the deterministic optimization LP model of the electricity generation system. • The electricity load has to be met exactly by the generating units. However. the optimal value of the cumulative fuel cost is found using the deterministic optimization model described in Chapter 3. Again we assume that the wind speed can be forecast exactly over the planning horizon and there is no uncertainty about the electricity load. For the purpose of this chapter we wish to make the following modifications to the optimization model: • The electricity demand is constant so that variability in wind energy is clearly presented in the output. This model was used to estimate the system balancing cost that appears when variable wind energy was incorporated into the generation system.

Every next model is chosen so that the error between the actual and calculated accumulated fuel cost decreases. g ∈ C.4£/MW*min. Maximum Run-Up and Run-Down export rates (introduced in Section 3.• There is hydro generation included in the optimization model.125 MW*min and r− equals 0. . . T.1.g ≤ xt. and g ∈ C.125 MW*min for all the models in this chapter. We assume the accumulated fuel cost of the electricity generation to be a piecewise linear function. . therefore. so that the cost of the first unit c1 = 1£/MW*min increasing by 0. xg ≤ xt. it is small and does not affect the solution. . Hydro generation is small for all the models in this chapter and has a constant output of 20 MW.g ≤ xg .g + r+ xt+1. Let us take a merit order of the fuel costs to be a piecewise linear function. 68 . The value of FA is calculated per unit time (per minute in the current case).g ≥ xt. Thermal generators are not differentiated by the available capacity and every plant has the following upper and lower bounds: xg = 0M W xg = 30M W. inequalities (3. Definition Let us denote FA as the actual accumulated fuel cost calculated using the deterministic optimization model. where r+ equals 0. 2. The thermal generators are differentiated by the fuel cost chosen in a certain merit order so that cheaper thermal plants are used first. Let us consider a small electricity generation system where the set of the conventional generators C consists of 15 thermal units. Further in this chapter we will consider different statistical models of the accumulated fuel cost and compare these with the actual accumulated fuel cost. however. There is no cost associated with hydro generation.1) of thermal generating units are the same for all the range of the electricity output.1£/MW*min for every next unit so that the 15th thermal plant is priced at c15 = 2.g − r− t = 1.3) − (3. therefore all the statistical models in this chapter are independent of the planning horizon T .6) are simplified as xt+1.

It frequently changes the production plan in order to balance the 69 .3. when a marginal thermal generator (as defined in Section 3.1) is not. suitable for calculating the accumulated fuel cost if there is wind energy incorporated into the power system.1 plots an example of the solution for this simplified case of electricity generation. On this graph thermal plant with the index “Th8” is the marginal generator.2 Case of marginal generator following a profile of the wind energy curve Let us begin with a simplified situation of the electricity generation when the wind energy is constant and equals zero.1) where F0 is the generated cost when there is no wind energy available and can be found from the merit order fuel costs of the thermal generators. Electricity generation with hydro. KW 15 10 5 0 0 20 40 60 80 time 100 120 140 160 180 Figure 4.1: Electricity generation with only the marginal generator modifying the power output Figure 4. wind and thermal power 30 Th1 Th2 Th3 Th4 Th5 Th6 Th7 Th8 Th9 Th10 Th11 Th12 Th13 Th14 Th15 wind hydro 25 20 energy. To improve the model we take into account another simplified situation of electricity generation. In this case the accumulated fuel cost can be calculated accurately with an error ǫ = 0 by: FA = F0 (4.1) has the Run-Up and Run-Down export rates big enough to cope with variability of wind energy curve. In this case error of the calculation ǫ = FA − F0 can be significantly higher than zero. The model described by (4.4. however.

Thus.2). The first seven generators have the lowest fuel cost so that they produce at full capacity.wind energy output and meet the electricity load exactly. 4.2) In (4.KW 15 10 5 0 0 20 40 60 80 min 100 120 140 160 180 Figure 4. the operational cost can be calculated with (4.2: Electricity generation with the wind energy profile as a zigzag 70 . we wish to find a better model in order to improve accuracy of the operational cost calculation. Generation with equal number of thermal generators on top and bottom 30 Th1 Th2 Th3 Th4 Th5 Th6 Th7 Th8 Th9 Th10 Th11 Th12 Th13 Th14 Th15 wind hydro 25 20 energy. This case improves the model that estimates the operational cost of the electricity generation: FA = F0 − cm xw − ǫ. The error term in (4. the other elements of the model that estimates the operational cost also depend on the statistical parameters of the wind energy curve.2) with an error ǫ = 0. (4. For both simplified cases considered in this section. there can be an error calculating the operational cost from the right-hand side of formula (4. The last seven generators with fuel cost higher than that of the marginal generator do not produce.2) cm denotes the fuel cost of the marginal thermal generator and xw is an average wind energy over the modelled period of time T . for a situation when the rates of change of the wind energy curve are higher than that of the marginal generator.2) is non-zero and can be significant.3 Case of wind energy curve having a zigzag profile Assume. However.

To evaluate them we construct another simplified case when the marginal generator has the Run-Up and Run-Down export rates smaller than those of the wind energy curve.3: Electricity generation for a wind energy curve as one zigzag and the thermal plants increasing and decreasing their output in a symmetric Λ-shape Definition We denote γw as the average of the absolute value of the gradients calculated for the wind energy curve. Further in this section there is a number of tests performed with one V-shaped drop in wind energy that is plotted in Figure 4.2 shows that the solution is also periodic and makes it possible to calculate the accumulated fuel cost exactly as it is found using the deterministic optimization model. 2.3. 4.2 is periodic.3 while a length of the planning period of time T remains the same. . calculations for one V-shaped drop are applicable to the whole zigzag curve. .1 Analytical calculation of the accumulated fuel cost Definition Let us denote a height of wind energy curve as hw that represents the difference between the highest and the lowest value of wind energy in the time horizon 1. Electricity generation with hydro. A first test investigates how the accumulated fuel cost changes depending on the modification of the height of every V-shaped wind energy curve defined in Figure 4. the wind energy itself is periodic and has a form of a zigzag. Figure 4. Because the solution shown in Figure 4. . T . Modification of the wind energy height hw is performed in a way that the absolute value of the gradient of the drop in the symmetric V-shaped wind energy curve 71 . however. wind and thermal power 30 Th1 Th2 Th3 Th4 Th5 Th6 Th7 Th8 Th9 Th10 wind hydro 25 20 KW/min 15 10 5 0 0 20 40 60 80 min 100 120 140 160 180 Figure 4. .3.

let us look into the output of the thermal generators for a series of problems with the absolute value of gradient of the wind energy curve increasing.for the thermal generators running close to the lower capacity limit. Figure 4. We wish to find a formula that calculates the accumulated fuel cost of the electricity generation where a V-shaped wind energy curve is given. Then.4 plots an example of the solution with the broken-Λ-shapes. γg was assumed to be the same for all κ = 15 thermal generators. Thermal generators “Th6” and “Th10” are matching ther72 . Figure 4.changes proportionally to the height while the average of wind energy xw remains the same.5 and 4.3) i. hence.5 emphasizes this solution for the thermal generators running close to the upper capacity limit and Figure 4. then γg denotes the maximum of the absolute gradients of the thermal plant and is defined by the Run-Up and Run-Down export rates. Definition Let us denote the active thermal generators as generators that run away from the upper and lower capacity limits so that an increase or decrease in their output has a broken-Λ-shape.6 . the absolute value of the average gradient of the wind energy equals a sum of the absolute values of the gradient of κ thermal plants. In Section 4. Besides the cases where the output of the thermal generators increases or decreases in a symmetric Λ-shape. After finding a way every thermal plant responds to the increase of the variability in wind energy.e.1. for every γw such that γw ≤ κγg there exist situations where the output from thermal plants decreases or increases in a symmetric Λ-shape.6 plot an example of a symmetric broken-Λ-shape for the generators “Th6” and “Th10” respectively. Figures 4. Note. there are also transition cases between the two adjacent Λ-shapes where the the output of the thermal generators increases or decreases in a symmetric broken-Λ-shape. For an odd number of the thermal plants there always exists a pair of matching active thermal generators. one of which increases the output and the other decreases it. the accumulated fuel cost. that for the case where generation has to meet the electricity load exactly: γw = g∈C γg (4. To do this. we can calculate its total generated energy and. Definition Let γg denote the average of the absolute value of the gradients of a thermal plant g ∈ C.

5 27 0 20 40 60 80 100 120 140 160 180 time. 30 Th1 Th2 Th3 Th4 Th5 Th6 Th7 Th8 Th9 Th10 Th11 Th12 Th13 Th14 Th15 Hydro Wind 25 20 energy.5 Th9 Th10 Th11 Th12 Th13 Th14 Th15 Wind 2 energy. MW*min 1. MW*min 15 10 5 0 0 20 40 60 80 100 120 140 160 180 time. min Figure 4.5: Seven cheapest thermal plants increasing and decreasing their output in symmetric shapes 3 2.6: Seven most expensive thermal plants increasing and decreasing their output in symmetric shapes 73 .4: An example of the solution with the thermal plants increasing and decreasing their output in symmetric shapes 30.5 30 Th1 Th2 Th3 Th4 Th5 Th6 Th7 29.5 28 27.5 energy.5 1 0.mal generators and change the gradient of the Λ-shape at the same time intervals t = 30 and t = 150. min Figure 4. min Figure 4.5 0 -0. MW*min 29 28.5 0 20 40 60 80 100 120 140 160 180 time.

the accumulated fuel cost. In order to simplify visually further analysis. for the planning horizon T can be estimated using Formula (4. κ represents a number of thermal generators. and cm−j is the fuel cost of a matching thermal generator. (II) and (III). sg is a parameter that equals 1 for the active generators g ∈ C running at the minimum capacity. cm and xw are defined as in the previous model (4. lower than that of the marginal generator. Using information on the shapes of the thermal generation.4) = cm −˜i cg T. The actual accumulated fuel cost FA is significantly close to that calculated with the right-hand side of (4.4) is divided into three parts: (I). equals −1 for the active generators g ∈ C running at the maximum capacity and equals 0 for the marginal generator.4) below. F0 . ci −ci ˜g g c g < cm cg = cm (4.2).4) and weighted by the planning horizon T with 74 . T ) =F0 − cm xw + T c m γw + 4 ∗ κ−1 γ sg 1 − sg + ((∆i+1 )2 − (∆i )2 )cg (∆i+1 − ∆i )T cm g g g g T i=1 g∈C 4 4 where i ∆g (I) (II) (III) :=        ci −cm ˜g T. ˜g g ∗ γ denotes the sum of the absolute values of the average gradient of the thermal generators g ∈ C as in (4. increase or decrease in the produced energy can be calculated as the volume of an isosceles triangle. ci and ci are the fuel costs of the paired active thermal generators. Then. T (1 − ci −cm g ) ci −˜i g cg c g > cm . FA . ci −˜i g cg T.A point in time t where a pair of two matching active thermal generators running at the maximum and minimum production levels modify a gradient of the Λ-shape is determined by a proportion between their fuel costs and the fuel cost of the marginal generator: cm+i − cm cm+i − cm−j where cm+i is the fuel cost of a thermal generator. the right-hand side of Formula (4. higher than that of the fuel cost of the marginal generator.3). T denotes a planning horizon with one-minute discrete intervals. cm . FA (γw .

5): Fm = T FA (γw . the variability of the wind energy can be balanced with the marginal generator. T ) hw (γw ) − hw (γw − j) (4.6) is calculated as the difference in parts (II) of (4.5) where (γw − j) is the average of the absolute value of the gradients of the wind energy preceding the current value γw .respect to the gradient of a V-shaped drop of the wind energy curve. . T ) has the same value. As parts (I) and (III) are not used. By construction of the case on the modification of a height of wind energy curve (introduced at the beginning of Section 4. the marginal fuel cost Fm equals the fuel cost of the marginal generator: cm γw T − cm (γw − j) T cm j T 4 4 Fm = = T 4 = cm hw (γw ) − hw (γw − j) j4 75 (4. however. Taking into account a V-shape of the wind energy. the numerator of (4. It is calculated as the difference between the accumulated fuel cost for the adjacent values of the absolute value of the gradient of the wind energy curve weighted by the difference in the height of the wind energy hw .4) T T for FA (γw ) and FA (γw − j). a simple case where γw ≤ γg for the planning horizon t = 1. Therefore.1 The marginal fuel cost does not depend on the planning horizon T and is constant for every transition case where the active thermal generators increase or decrease the output in the broken-Λ-shape. . . calculated for FA (γw . 2.6) . Proof Let us consider.4) equals zero for FA (γw . . A part (III) of (4. Let us take a discrete interval of the absolute values of the gradient for the wind energy curve with a discretisation step j. T ) and FA (γw − j. there are no thermal plants deviating from the capacity limits when γw ≤ γg . we wish to estimate the marginal fuel cost for each of them. We wish to calculate the marginal fuel cost by analysing parts (I) − (III) of Formula 4. thus. a part (III) of (4.4. part (I) of (4. Proposition 4. i. T ). Then. the marginal fuel cost can be estimated by (4.e.3. T ) and FA (γw − j. Definition Let us denote the marginal fuel cost as Fm .4). To find how close the LP-modelled and the calculated accumulated fuel costs are.3.4) consists of the fuel costs for the active generators.1).4) is set to zero in the numerator of (4. a part (I) of (4. first. the average wind energy xw remains the same through the series of tests. T . for every γw .5). T ) − FA (γw − j.

4). The calculated marginal fuel cost acquires the same value as the LP-modelled. As for the simple case.7.7 demonstrates that the marginal fuel cost is constant for every transition case where the active thermal generators increase or decrease the output in the broken-Λ-shape. that for the first interval.4) the marginal operational cost is further calg culated as g∈C m g ( ci+1 −˜i+1 T − c g g ci+1 −c Fm = cm + = cm + ci −cm g T )T cm (1 ci −˜i g cg m g − sg ) + (( ci+1 −˜i+1 T )2 − ( cm−˜ig T )2 )cg sg i c c g g g g c −˜i+1 c c −˜i c T2 ( g∈C ci − cm cm − ci+1 2 ˜g cm − ci 2 ˜g ci+1 − cm g g − i )cm (1 − sg ) + (( i+1 ) −( i ) )cg sg i+1 − ci+1 i i+1 cg ˜g cg − cg ˜ cg − cg ˜ cg − ci ˜g where i = 1. calculated for FA (γw . This provides an insight into the next element of 76 . (κ − 1). 0 ≤ γw ≤ 0. a part T T (III) of (4.5).5).125 for our small electricity generation system. cm = 1.6) the marginal fuel cost for a transition case can be calculated in the following way: γ∗ T κ−1 i=1 g∈C i+1 (∆i+1 − ∆i )T cm 1−sg + ((∆g )2 − (∆i )2 )cg s4g g g g 4 Fm = cm + − = cm + = cm + γ ∗ −j T κ−1 i=1 g∈C (∆i+1 − ∆i )T cm 1−sg + ((∆i+1 )2 − (∆i )2 )cg s4g g g g g 4 hw (γw ) − hw (γw − j) jT 4 hw (γw ) − hw (γw − j) j T κ−1 i=1 g∈C (∆i+1 − ∆i )T cm 1−sg + ((∆i+1 )2 − (∆i )2 )cg s4g g g g g 4 κ−1 i=1 g∈C (∆i+1 − ∆i )T cm (1 − sg ) + ((∆i+1 )2 − (∆i )2 )cg sg g g g g T2 Using the definition for ∆i in (4. marginal fuel cost equals to the fuel cost of the marginal generator. . therefore. .125. Notice. Then applying the results of (4. .4) calculated for FA (γw ) and FA (γw − j) has the same value. For every transition case pairs of the active generators are the same when calculating the accumulated fuel cost. part (I) of (4. T ) has the same value. . we wish to look at the general case of (4. This shows that an additional unit of the calculated accumulated fuel cost depends only on the combination of the fuel costs of the thermal generators g ∈ C. which is shown in Figure 4. 2. T ) and FA (γw − j. again by analysing parts (I) − (III) of Formula (4.7. Figure 4. thus.Next.4) is set to zero in the numerator of (4. A length of the interval where the marginal fuel cost is constant equals to the Maximum Run-Up and Run-Down Export Rates of the thermal plants.4). a part (I) of (4. r+ = r− = 0.

75 1.the statistical model calculating the accumulated fuel cost that would reduce the error ǫ. For the calculation later in this chapter we chose a set of 30 wind energy curves so that they provide a variety of the statistical parameters such as the mean. hence. 77 .3.7 1.2 0. b. Comparing calculated and modeled marginal costs 2 calculated solution modeled solution 1.2 Improving the estimation of the accumulated fuel cost Analysis provided in the section 4.2 that another element of the statistical model includes the height and the average of the absolute value of the gradient of the wind energy curve and test it through a series of cases. This can be achieved by the scaling of the wind energy curve.3.5 0.2) that evaluates the accumulated fuel cost of the electricity generation.7: Comparing the calculated and the LP-modelled marginal fuel costs 4.9 marginal cost.8 0. Section 4. For another test the average of the absolute value of gradients of the wind energy curve γw and the height of the wind energy curve hw are changing while the planning horizon T remains the same. We assume further in Section 4. For a first test the average of the absolute value of gradients of the wind energy curve γw remains the same while the height and planning horizon T are modified linearly.8 1.3.3.7 average of absolute value of gradient 0.1 showed that the accumulated fuel cost depends on the absolute value of the gradient of the wind energy curve.6 0.3 0. the absolute value of the gradient.1 helps us improving the model (4.85 1. a. we wish to perform two tests with the wind energy curves having various profiles similar to the tests performed for the wind energy curve with a zigzag profile.4 0.65 0 0.95 1.p/KWmin 1. the standard deviation and the height.1 0.9 1 Figure 4.

0155*Gav*h 0.7) k1 ≈ 0. Hence.2) plotted against γw has a quadratic form.8: Fitting a quadratic term into the model that estimates the accumulated fuel cost of electricity generation Tests a) and b) support a statement that an error in the model (4.0155*Gav*h 0.5 0. Figure 4.0155 is a parameter that can be estimated with the fuel costs of the thermal plants.7) In (4.4 difference in cost 0. FA = F0 − cm xw + k1 γ w hw + ǫ (4. it is independent of the wind energy curve.7) is for the estimation of the accumulated fuel cost of the electricity generation. h 20 25 Figure 4. 0.2) is linear with respect to the changing height hw . In the test a) only the height of the wind energy drop changes while the average of the absolute value of the gradients remains constant. So the error ǫ in the model (4.3 0.1 0 5 10 15 height of wind energy curve.8 three functions from the below use the same quadratic 0.8 demonstrates the results of the tests a) and b) performed for the different values of hw . an improved model estimating the accumulated fuel cost of the electricity generation is formulated further as (4.1 0 -0.2 0.7).2) depends on a quadratic term containing the average of the absolute value of gradients γw and the height of the wind energy curve hw .0155γ w hw to approximate another term included in the model that estimates the accumulated fuel cost. In Figure 4.7Xav Ftot-Fo+1. Test b) is more complex as two variables: the height and the average value of the absolute gradients of the wind energy curve are changing. Figure 4.10 plots the error term of the 78 . Error ǫ can be plotted as quadratic function as long as the height and the absolute value of the gradient are modified linearly.7Xav Ftot-Fo+1.6 Ftot-Fo+1. Let us find how good the model (4.7Xav 0.It was noticed that the height of the wind energy and the average of the absolute value of gradients are linearly dependent and a remaining error of the model (4.0155*Gav*h 0.

7*Xav 0. For both curves in Figure 4.4 Case of the wind energy curve with changing standard deviation Let us estimate the error term in the model (4.7). Approximation of the difference between LP-modeled and calculated cost for zigzag wind curve 7 Ftot-Fo+1.10 the error is relatively T high. An error calculated for all the planning horizon T has positive and negative values for the different wind energy curves so that it is hardly possible to improve the result.model (4. for the more variable wind energy we wish further to improve the model (4.2) for a set of 30 wind energy curves Figure 4.7) can be considered a good estimate of the accumulated fuel cost of the electricity generation.7) to decide on the elements that would improve it. We wish to calculate the error for two different cases: for all the planning horizon T and an accumulated error if calculated for the ten-minute intervals that a planning horizon T is divided into.9: Comparison of the error term for the LP-modelled and the calculated accumulated fuel cost with the model (4. an accumulated error for 10 intervals of the planning horizon is convenient for further improvement as all the values are positive.10 plots the error of estimating the accumulated fuel cost calculated T for all the planning horizon T and an accumulated error calculated for 10 intervals of the planning horizon.0155γw hw calculated for a set of 30 wind energy curves.2) and compares it to the quadratic term 0. Figure 4. However. However. p/unit time 4 3 2 1 0 0 2 4 6 8 10 square root of product h*Gav 12 14 16 18 Figure 4.0165*Gav*h 6 5 cost.9 shows that within a range of the wind energy curves with low variability (represented by a low γw ) the model (4. 79 . 4.

Difference between LP-modeled and calculated cost 0.7) and (4. The figure shows that the error.28 for our case. Thus.h) 0. Again we are looking at two different ways of calculating both values of the standard T deviation: over the whole planning horizon T and accumulated for 10 intervals of the planning horizon T .1 0 0. Parameter k2 equals 0.5 0.7) for a set of 30 wind energy curves When plotted along the axis of the standard deviation of the absolute value T of the gradient of the wind energy curve. we wish to use a standard deviation of the wind energy and a standard deviation of the absolute value of the gradients of the wind energy to formulate another element in the model estimating the accumulated fuel cost of the electricity generation.5 standard deviation of gradient 2 2. Figure 4.7Xav-f(G. calculated for a set of 30 wind energy curves.h) 18int:Ftot-Fo+1.5 3 Figure 4.4 cost error.6 whole:Ftot-Fo+1.8).3 0. (4. An improved model estimating the accumulated fuel cost of the electricity generation is formulated as follows: α FA = F0 − cm xw + k1 γ w hw + k2 σx σγ + ǫ (4.1 0 -0.2). This is an expected result as the latter models are richer and include the 80 .5 1 1. decreases with the additional statistical parameters being added to the model.7Xav-f(G. p/unit time 0. an accumulated error for 10 intervals of the planning horizon has a trend of a power function. found as the difference between the LP-modelled and the calculated accumulated fuel cost. (4.2 0.8) where σx denotes the standard deviation of the wind energy and σγ is the standard deviation of absolute value of gradient in a power α = 1.11 plots a comparison of the error term in three models estimating the accumulated fuel cost of the electricity generation.05 for our case.10: Comparison of the error term for the LP-modelled and the calculated accumulated fuel cost with the model (4.

former when setting parameters k1 := 0 and/or k2 := 0.
Error calculating cost of electricity generation 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 -0.1 0 5 10 15 curve number 20 25 30 18int:Ftot-Fo+1.7Xav 18int:Ftot-Fo+1.7Xav-F(h,G) 18int:Ftot-Fo+1.7Xav-F(h,G)-0.05Dx(Dg**1.28)

Figure 4.11: Comparing an error of calculating the accumulated fuel cost of the electricity generation with the models (4.2), (4.7) and (4.8)

A small error of estimating the accumulated fuel cost when calculated with the statistical model (4.8) shows that the saving in fuel cost and the additional fuel cost (introduced in Section 3.3.1 and Section 3.3.2) are determined by the statistical parameters of the wind energy curve. When wind energy is incorporated into the electricity generation system, the accumulated fuel cost is reduced by the amount of wind energy present in the system, however, one-minute fluctuations of the wind speed, described by the absolute value of gradients and the standard deviation, cause an increase in the accumulated fuel cost.

error, p

81

Chapter 5 Stochastic Modelling
Chapters 3 and 4 use deterministic modelling for scheduling the electricity generation assuming the electricity load is known and all the generating units are strictly following the operating plan. However, this is a strong assumption in real-life situations, so this chapter introduces an uncertainty of the electricity generation. We wish to use stochastic programming as a method of structuring possible future scenarios and estimating the system balancing cost that appears when wind energy is not forecast for certain. The System Operator of the United Kingdom balances the electricity generation and consumption in the power system every moment of the real time and is responsible for keeping the system frequency within a target range. In order to cope with an unexpected mismatch between the electricity supply and demand the System Operator uses the available Response and Reserve provided by the generators. As it was specified in Chapter 1, Response and Reserve are being used on different time scales. Response is an automatic process provided by the large generating plants that keeps the power system within the target frequency in a time range from 0 seconds to several minutes. By that time operating Reserve is expected to increase or decrease the output and support the power system in minutes or hours. Response is an automatic process provided by the generators while a decision about the amount of operating Reserve is made by the System Operator and depends on possible losses or fluctuations of the load that can happen in the future. For example, 1GW of Reserve means that some generators run away from their maximum export limits so that they can increase the power output by 1GW if required by the System Operator. The aim of the stochastic programming is to formulate an operating plan that keeps feasible possible scenarios of wind energy output and minimizes the system balancing cost. 82

The stochastic optimization model is formulated using the deterministic model introduced in Chapter 3 and includes common constraints of the electricity generation. Our original contribution consists in developing a scenario tree that reflects two uncertainties: uncertainty in the available capacity of the thermal plants and uncertainty in wind energy output. We use the formulated stochastic optimization model to achieve the aim of this thesis by calculating the system balancing cost when unpredictable wind energy is introduced into the electricity generation system.

5.1

Uncertainty of the wind energy

A motivation to model the uncertainty of the wind energy comes from the fact that wind speed is hard to forecast. Mathematical techniques of forecasting an average wind speed for the next hour or half an hour have improved significantly during last 30 years ([36, 37, 38, 39, 40]), however, one minute variations as well as sub-hourly capacity of wind energy are still calculated with possible errors. Until the short-term prediction of the wind speed is further improved, the available wind energy creates an additional uncertainty in the electricity generation system. Currently the installed wind capacity does not exceed 4% of the total electricity generation, so that the matter of wind speed low predictability can be resolved without an increase in the operating Reserve. However, if the wind energy production amounts to 15% and higher, as it is planned by the Renewables Obligation, then the system balancing cost would significantly increase. Assuming the uncertainty of the wind energy output, planning the electricity generation can be described as a sequence of random realisations and decisions of the System Operator on the modification of the operating plan for the thermal plants. In its simplest form the discrete stochastic process can be represented as a scenario tree describing the unfolding of uncertainty over the planning horizon [28]. A scenario tree consists of nodes and arcs as in the example pictured in Figure 5.1. For our case the nodes represent the states when the information about the wind energy output is revealed and the operating plan of the thermal generators is determined to match the electricity load. The arcs do not have a physical meaning and are used to structure the nodes and form possible scenarios. For every two nodes connected by an arc the preceding node is named a parent and the following node is named a child. The current state is taken as the first node of every scenario, it is named the root of the tree while the last node of a scenario 83

An example of a scenario is the set of nodes {1. A path. 5 and 6 that are considered leaves as they do not have any children.Child 5. The aim of the modelling is to schedule the operating Reserve of gas and coal plants whose production level can not be modified too frequently. node 1 on Figure 5.Leaf 3. 3 and 4. in Chapter 3 we used the discrete time with the smallest interval of 1 minute. however.1 is parent of three nodes-children. at the moment of making a decision. 3.is named a leaf. represents a scenario.Leaf Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5 Scenario 6 Figure 5.Leaf 7. and incorporate it in the scenario tree so that a decision about the electricity generation is made every other minute. from the root to a leaf of the scenario tree. For example.1: An example of a scenario tree It is important to capture the nonanticipativity property with the structure of the scenario tree. This uncertainty is gradually reduced. At each state the System Operator makes a decision about the operating Reserve under the uncertainty of the future realisations of the wind energy. the Reserve has to respond to the changes in the electricity load as soon as possible. and the associated decisions at this node are identical for all six scenarios.Leaf 9. hence one minute is considered the most appropriate frequency for short-term scheduling. The System Operator balances the electricity generation and demand every moment of the time. We wish to use the discrete time further for the stochastic programming as well. Let us specify further the parameters of the scenario tree. the System Operator does not have a preference for any of the scenarios so that the decision is nonanticipative [34]. since the decision process is being accompanied by a flow of information. Nonanticipativity property can be explained in the following way.Leaf 4.Child 10.Root 2. 7}. while node 2 is parent of only two nodes. A 84 .Child 8. node 1 of the tree presented in Figure 5.1 corresponds to the first stage. 1. however. At the same time. 2. Thus.Leaf 6.

T = 9. 2. It is unknown whether at the minute t + 1. Let us define an event as an output of the available wind farms in MW every minute t = 1. . . and for the current problem T S = 30. Denote further a length of the modelling horizon as T S . T = 10 : s=1 1 2 3 4 T s = T S = 30 Branching the scenario tree every 10 minutes reflects the uncertainty of the wind energy and keeps a limit on the number of nodes that can grow dramatically when branching the scenario tree every minute of the modelling time period. . Assume that wind speed and. therefore. T where T ≥ T S is a certain planning horizon. . . . 2.length of the Settlement Period (as it was defined in Chapter 1) is 30 minutes.2: Tree of scenarios that represent possible changes in wind speed energy. 4} we set: 4 T = 1. Assume further multistage stochastic programming so that the wind energy branches several times during the modelling horizon T S . T S the wind speed remains as it was forecast or a wind speed profile deviates from the forecast or shifts in time. min 20 25 30 Figure 5. T = 10. the wind energy output are known at the current state t = 1 when decision about the generating plan for the conventional plants is made. MW 85 . This means that a number of possible scenarios for the wind energy profile can be significantly large. . Accumulated length of all the stages equals the length of the modelling interval T S so that for every stage s ∈ S := {1. t + 2. 5000 4500 4000 3500 3000 2500 2000 1500 1000 500 0 0 5 10 15 time. 3. . Let us branch the scenario tree every 10 minutes so that there are 4 stages in this multistage stochastic program.

the current output from the wind energy turbines. The transformation function that allows us calculating the wind power from the wind speed. In this case a wind power value of the parent node branches into two child nodes so that two new nodes are represented 86 . such as two extreme outcomes. namely: 1. We wish a scenario tree to capture the unpredictable nature of the wind power. wt−1 ) (5. While building a scenario tree we consider three ways of branching. it does not reflect the variability of the wind power that was estimated in Chapter 3. 2. Every possible scenario of the wind power output is a piecewise linear function so that the slope of a linear segment of this function remains the same for every node. e e e e P (wt+1 ) = P (wt+1 |wt . we wish to outline further some general directives for building a scenario tree for the wind energy output. a heuristic distribution of the wind energy. and 4.1) The conditional probability distribution (5.1) provides the information on the e variability of the wind power from a given value wt . Besides. When calculating the probability distribution of the wind power.1 plots a typical transformation function and shows that the wind energy output has upper as well as lower bounds. Figure 2. when the wind speed exceeds 15m/s the power output flattens at the maximum capacity. two extreme outcomes balanced by the mean wind power. Assume there are four main elements of the available wind energy taken into account when building a scenario tree.Branching of the scenario tree is an individual process and depends on the characteristics of the underlying events. • two extreme outcomes. provides a useful information for building a scenario tree. however. we wish to build a scenario tree of the wind power rather than the wind speed as it can reduce a number of nodes in the tree and increase the speed of solving a problem. the preceding values of the wind energy. Hence. and five or more branches. an average wind energy predicted for the next hour. Although variability and capacity of the wind energy turbines depend on the geographical and weather characteristics of the site. 3. bullet points 3 and 4 of the above list can be combined so that the distribution of the wind energy output is conditional to the preceding values.

(5.1) e+ e− to remain the same. If wt+1 and wt+1 are new child nodes acquired by respectively the highest positive and negative deviation of the wind power then the probability of their outcome can be calculated as follows: e− p(wt+1 ) e+ e wt+1 − wt+1 = e+ e− wt+1 − wt+1 e− e+ p(wt+1 ) = 1 − p(wt+1 ) e where wt+1 is the expected value of the conditional probability distribution (5.2) The probability for each of the two child nodes is calculated so that the stochastic process of the wind power output is preserved.1).1). In this case a wind power value of the parent node 87 . In this case a wind power value of the parent node branches into three child nodes two of which are represented by the highest positive and negative deviation of the wind power from the current value (as in (5. In the case of two nodes it requires the expectation of the probability distribution (5.1). The probability for each of the three child nodes is calculated so that the stochastic process of the wind power output is preserved. e− e wt+1 and wt+1 are as follows: e+ p(wt+1 ) e− p(wt+1 ) e p(wt+1 ) w σt+1 = e+ e+ e− e e e (wt+1 − wt+1 )2 − (wt+1 − wt+1 )(wt+1 − wt+1 ) w σt+1 = e− e+ e− e e e (wt+1 − wt+1 )2 − (wt+1 − wt+1 )(wt+1 − wt+1 ) e+ e− = 1 − p(wt+1 ) − p(wt+1 ) • five or more branches. • two extreme outcomes balanced by the mean wind power.2)) and the third child node is the expected value of the conditional distribution (5. In the case of three nodes it requires that the expectation and the variance of the probability w distribution (5. If σt+1 denotes the variance of the e+ probability distribution (5.1) remain the same.1) then the probability of the outcomes wt+1 .by the highest positive and negative deviation of the wind power from the current value: e+ e e e− e e wt+1 : max(wt+1 − wt ) and wt+1 : min(wt+1 − wt ) e where wt+1 is found with the probability distribution (5.

2 plots an example of scenario tree where every scenario represents a possible output of the wind turbines.20 11. therefore. 2. .55 14.17 73.36 13. T such that the resulting generation prepares the system for a variety of wind power outcomes. we determined the best branching combination with the accumulated fuel cost over the planning horizon of 5 days.65 12. each of which branches further into 2 nodes resulting in 30 nodes at stage s = 4. After the solution is obtained.1 compares the accumulated fuel cost weighted by the generated power and calculated for different branching methods.1 shows that the accumulated fuel cost calculated for the branching combination (1 : 5 : 3 : 2) is the lowest.1.56 12.13 12. Table 5.96 11.28 17.84 56.16 58.72 13. This combination means that at stage s = 1 there is only one node that branches into 5 nodes at stage s = 2. .66 13. hence.82 10. Each of these 5 nodes branches into 3 nodes resulting in 15 nodes at stage s = 3. A scenario tree can be built by combining different branching methods. £/(MW*day) (T 1 : T 2 : T 3 : T 4 ) day 1 day 2 day 3 day 4 day 5 total (1 : 2 : 2 : 2) 15. we wish to use this branching method further in this chapter for the stochastic optimization modelling. All the branching methods result in a feasible solution.44 14.83 (1 : 3 : 3 : 3) 12.43 16. .10 16. Figure 5.25 (1 : 5 : 5 : 2) 11. For example.2)) and the remaining possible outcomes are the mean values of the equally spaced segments.1: Accumulated fuel cost weighted by the generated power and calculated for different branching combinations.62 (1 : 5 : 3 : 2) 11.67 11. We tested different combinations of branching methods for building a fourstage scenario tree by finding a solution of the stochastic linear optimization model.59 Table 5.88 14.77 12.2 was built with the combination of (T 1 : T 2 : T 3 : T 4 ) := (1 : 5 : 3 : 2).07 11.36 12. Table 5. the electricity generation of the root of the scenario tree is fixed at its optimal value that establishes the final schedule for this minute of the time 88 . described further in Section 5.12 14.21 15.85 15.81 17.branches into five or more child nodes two of which are represented by the highest positive and negative deviation of the wind power from the current value (as in (5.12 11.90 (1 : 3 : 3 : 2) 13.3.43 69. a scenario tree in Figure 5.33 (1 : 3 : 2 : 2) 14.51 10.98 64.03 58.87 12.71 83.00 (1 : 5 : 2 : 2) 11. . This scenario tree is used to solve a single stochastic programming problem of scheduling the electricity generation at the current minute t of the planning horizon 1.15 11.95 11.

As it was specified in Chapter 1. . . A scenario tree that captures a possible loss of the capacity is built so that it is compatible with the wind power scenario tree. 5. . Therefore. . 2. similar to the wind power scenario tree. we wish to build a scenario tree that describes a possible loss of the capacity and estimate the system balancing cost when resolving both uncertainties by the System operator. Every node represents a state when the information about the available capacity of the conventional plants enters the system and the System Operator determines whether their operating plan should be modified.horizon. A probability of such an event is low. T S is rolled forward in time and a new scenario tree is built based on a new value of the e wind power wt . the System Operator can use the automatic Response and the planned operating Reserve to increase or decrease an active power in the system and preserve the integrity of the GB Transmission System. we define an event as a size of the electricity load 89 . (T − T S ). . equals 30 minutes. The arcs connect the nodes forming a path from the root to the leaves of the scenario tree. .2 Uncertainty of lost generating capacity Another uncertainty that the System Operator of the United Kingdom has to take into account while balancing the electricity demand and generation involves a loss of the generating capacity. 2. In order to introduce a possible capacity loss of any type of conventional plant in one scenario tree. This process of the rolling modelling horizon is repeated so that the optimal values obtained from each iteration yield the final solution for every minute of the planning time horizon 1. Modelling time horizon T S is discretized for the scenario tree of the possible capacity loss and. An impact of the generating capacity loss on the power system increases when there is also the uncertainty of the wind energy output taken into consideration. An example of an event captured by this uncertainty could be a conventional plant tripped and taken out of service. . . . . For the next minute t = 1. . . 2. loss of the generating capacity and the wind energy output. however it can happen at any moment and effect the target frequency with a significant mismatch between the electricity supply and demand. T the modelling horizon 1. The next stochastic programming problem is solved using a new scenario tree and the optimal electricity generation for the root is fixed for the current minute t.

. . 2. . Assume that the available generating capacity and. however. . . min 20 25 30 Figure 5. 3. . On the event tree it can be pictured as a linear change in the forecast demand for a minute t + 1. there is a loss of the importing capacity. therefore.3 plots a scenario tree of the electricity load values where every minute there is an expected loss of the exporting or importing capacity. . We wish to capture the expectation of a loss in the generating capacity every minute of the modelling horizon. Figure 5. . the electricity load are known at the current state t = 1 when a decision is made whether to modify an operating plan of the conventional plants or not. . T S : all the planned generating capacity is available. t + 2. The fact that the electricity demand is a deterministic value in this work allows us interpreting a loss of the capacity as an increase in the electricity load. . It is unknown whether the onstream generating capacity is available at minutes t + 1. .3: A scenario tree of the changes in the electricity load equivalent to the loss of the capacity A scenario tree plotted on Figure 5. . . MW*min 38400 38200 38000 37800 0 5 10 15 time. there is a loss of the generating capacity. In this case. T S . the remaining generating units are available until the end of the modelling horizon T S. T S or there is an importing or exporting capacity lost. t + 2. . .in MW every minute t = 1. the number of stages when a new information becomes available is T S − 1. 39000 38800 38600 load.3 presents three possible outcomes every minute t of the modelling horizon 2. which is higher than that for the 90 . For every scenario we assume that once the capacity is lost. Every scenario suggests a loss of 600MW in the generating capacity and a loss of 100MW in the importing capacity at a certain minute t. T S comparing to the previous state. .

. The optimal solution for the root of the scenario tree which is now a new minute t. let us combine the wind power and the electricity load scenario trees described in sections 5. The root of the scenario tree has index n = 1.2 are formulated in the stochastic programming problem later in this section. .3 Solving a problem with the uncertainty The uncertainties presented in sections 5. . This solution is fixed and the modelling horizon is rolled forward so that the optimal solution can be found for the next minute t = 1. 5.scenario tree of the changes in the wind power where S = 4.2. . as they depend on a probability distribution of the appropriate stochastic process and a choice of the branching. 5. 2. (T − T S ).e. and enumerate all nodes of the combined scenario tree.1 and 5. so the stage 2 nodes start from index 2. . Definition Let n denote a node of the scenario tree and an denote the direct ancestor of a node n to capture the dynamics in the stochastic optimization model. In order to formulate a mathematical model for both uncertainties. However. The rolling horizon method is applied until the optimal solution is found for every minute t of the planning horizon 1.1 Decision variables and constraints of the LP stochastic optimization model To formulate the deterministic equivalent of the multistage stochastic programming problem. . . they are formulated in the next section so that the nonanticipativity property is preserved.1 and 5. a loss of the generating capacity and the change in the wind power. i. is fixed again. 91 .3. . This assumption leads to the situations where parent and child belong to the same stage. The numbers of children for each node in the event tree may differ. The output of a number of tests performed with the stochastic model result in the estimate of the system balancing cost when the wind power and the available capacity are uncertain. The optimal solution at the root of the scenario tree ensures that the electricity generating system is prepared for a capacity loss at any minute t of the modelling horizon. We use a breadth-first search order [28]. 2. T . we assume there are four stages in the scenario tree of the changes in the electricity load rather than T S . start from a root node corresponding to the first stage. the “parent-child” structure of the scenario tree preserves the nonanticipativity property of the stochastic problem.

. . A number of time intervals associated with every node n is denoted by Tn > 0. τ = 1. The 92 . N. . . The wind power generation is determined by the wind power scenario tree. If a certain event describes a loss of the generating capacity at time τ = τ ∗ < T s of node an then Tan = τ ∗ and Tn = T s − τ ∗ .4) There are two groups of constraints that can be formulated for the stochastic model: the constraints applied to every node n = 1. However. . xnτ g = wnτ g ∀n = 1. 2.3) where xnτ g is a production level of the thermal unit g during time interval τ of node n. . Then. Depending on a choice of the branching method (as described in Section 5. 2. . the scenario tree of the changes in the electricity load branches every minute. . . Definition Let the index τ denote a minute of node n. . 1. Let G denote the set of all the generators and consist of the union of C. g ∈ C (5. for the scenario tree that captures the uncertainty of the wind power output. 2. N. . N and the constraints that reflect the stochastic properties of the model. 2. xn0g = xan Tn g ∀n = 2.Let N denote the last node at the stage 4 so that each scenario goes through a certain number of nodes n ∈ {1. For the root node.5) These inter-nodal constraints state that for every generator g the first value of an output in a node n has to be the same as the last value of its ancestor an . . g ∈ W (5. . 3. . Nuclear and Gas generating plants. . . . . g ∈ G (5. . Tn . τ = 0. . . . . N. Tn . .1) nodes of the scenario tree can be grouped by stages of the scenario tree. so a number of intervals Tn for a node n of this tree varies depending on a minute when the generating capacity is lost. the set of the conventional generators. where s is a stage of node an . First let us formulate the constraints that connect the nodes into the scenarios and show the “parent-child” relationship between the nodes. Tn . 1. The decision variables are denoted by xnτ g ∀n = 1. and ancestor an sometimes corresponds to the same stage as node n. . . the set of the wind generators. . 2. N }. . The decision variables in the model correspond to the output of each generating unit in the power system. . . Tn equals a length of the stage s corresponding to node n. . The set of the conventional generators also consists of Coal. . such that τ = 0. with W . a number of intervals T1 = 1 shows that there is only the current minute information available at the moment of the decision making.

Decision variables introduced above have finite upper and lower bounds representing unit capacity limits of the generation system. . xg ≤ xnτ g ≤ xg ∀n = 1.8) xn(τ +1)g ≥ xnτ g − r2− where n = 1.first value of a node n is described with an index 0. they are equivalent to the constraints formulated for the deterministic model in Chapter 3. . Run-Up and the Run-Down constraints also applied to the thermal generation. τ = 0. 2. . . τ = 1. Further. . The last value of a node n is determined by an index Tn . . . . . Following the notation of Chapter 3. N. Then the corresponding constraints of the possible change in the output of the thermal plants are formulated as the inequalities: xn(τ +1)g ≤ xn(τ +1)g and xn(τ +1)g ≥ r2− xg − r1− e− xg − e− + r2− − r1− xnτ g − xg − e− xg − e− (5. let r1+ and r2+ denote two Run-Up rates separated by the “elbow” output level e+ and r1− and r2− denote two Run-Down rates separated by the “elbow” e− output level. Planning the production level of the electricity generation units is bounded by the available resources and technology.9) (5. Tn . . the loading constraint is formulated with the inequalities: Dnτ (1 − ∆D ) ≤ xnτ g ≤ Dnτ (1 + ∆D ) (5. .7) (5.10) e+ + r2+ − r1+ xnτ g + r1+ e+ ≤ xnτ g + r2+ (5. . 1.11) g∈G 93 . Tn . Denoting by Dnτ the load demand during interval τ of node n and by ∆D a percent mismatch allowed between the electricity load and the generation. . it is fixed by the ancestor while decisions are made beginning with τ = 1. 2. g ∈ C. . 2.6) The parameters xg and xg denote the lowest and the highest possible output of unit g respectively. g ∈ C (5. . The loading constraint combines different generating units and applied for every node separately. . N. Further there are constraints that are applied to every node n.

d− v ] depends on the scenario tree of the electricity nτ nτ load and can be calculated in the following way + d+ v = ∆d ∗ Dnτ ∗ βv nτ − d− v = ∆d ∗ Dnτ ∗ βv nτ (5. .14) (5. (1 + ∆d )Dnτ ]. . V is formulated with the inequalities: 0 ≤ d+ v ≤ d+ v nτ nτ 0 ≤ d− v ≤ d− v nτ nτ (5. N.where n = 1. The set of parameters [d+ v . This fact increases the risk of operating at a frequency below the target level. . . 2. To avoid this we reintroduce + − a cost of mismatch between the electricity demand and the generation {ζv . . . Tn .16) (5. . 2. 94 . . V . d− v } nτ nτ (5. . . 2. . . . N. τ = 1. Let us introduce further another set of variables: {d+ v . . . Tn and V is a number of the penalty cost bands in the allowed mismatch between the electricity demand and the generation [(1 − ∆d )Dnτ . 2. . . ζv } as it was described in Chapter 3. . . The size of every penalty cost band v = 1. .13) where n = 1. . . . . . . Similarly to the case of the deterministic problem.12) where n = 1.15) where d+ v is an upper bound of the band v when the generation exceeds the nτ electricity demand and d− v is an upper bound of the band v when the electricity nτ demand exceeds the generation. . N. 2. Tn . 2. V . . . . . τ = 1. 2. . 2. . the minimization of the accumulated fuel cost affects the stochastic solution by setting the generation of the thermal units close to the lower bound of the allowed mismatch between the electricity generation and the demand Dτ n (1 − ∆D ).17) ± where βv is the part of the allowed mismatch between the electricity demand and the generation ∆d ∗ Dnτ that corresponds to the penalty band v = 1. . τ = 1. . 2. v = 1. . These variables denote the difference between the electricity demand and the generation formulated in the following way: V g∈G xnτ g − Dnτ = v=1 (d+ v − d− v ) nτ nτ (5.

v = 1. We take κ again as a number of thermal generators in the set C so that a total number of decision variables in the model equals κ N Tn + V N Tn . For the problem formulated in this section we combine two scenario trees plotted in 95 . . ζv } and the size {d+ v . It is uncertain if a node n = 1. however. 2. This probability is multiplied by the probabilities of its ancestors to get the path probability pn .2) − (5. 2. . . . . the amount of the wind power in the electricity generation system that produced a feasible solution for the deterministic problem maybe give an infeasible result for the stochastic model. Similarly to the case of the deterministic optimization n=1 n=1 model. The objective of the stochastic problem is to find a feasible solution {xnτ g } such that it satisfies all the constraints (5.: N Tn V Tn + ζv v=1 τ =1 + dnτ v V Tn − ζv v=1 τ =1 min x n=1 pn ( g∈C cg τ =1 xnτ g + + d− v ) nτ (5. Because of the uncertainty that the System Operator has to deal with.18) In the objective function (5. N will realise in the future.As distinct from the deterministic case there are four penalty cost bands for the stochastic programming problems. we wish to reduce a number of variables by grouping thermal generators: one nuclear. . The probabilities in the objective function are not those that were formulated for the scenario tree but path probabilities pn = P (n) ∗ P (an ) ∗ P (aan ) ∗ · · · ∗ P (1) where P (n) is a probability formulated for a node n in the scenario tree. . . 2. . First. . . N and . 2.e. Number of nodes N in the stochastic optimization problem depends on a chosen scenario tree. where g ∈ C. . τ = 1. V n = 1.17) at the minimal operational cost. i. . dnτ 4 ) are nonzero only in the nτ exceptional cases where there is no other way to avoid infeasible solution. .18) estimates the expected operational cost of the electricity generation. three penalty cost + − bands are associated with the costs {ζv . . there is a probability pn associated with node n.18) the operational cost is calculated for every n node n and includes the fuel cost g∈C cg T=1 xnτ g and the cost of a mismatch τ Tn V Tn + + − − between the electricity generation and the load V ζv v=1 τ =1 dnτ v + v=1 ζv τ =1 dnτ v . Tn . Therefore. d− v } as in the nτ nτ deterministic case. the objective function (5. The stochastic LP optimization model formulated in this section includes two sets of variables {xnτ g } and {dnτ v }. The fourth price band has a very high cost and big enough − size associated with it so that the variables (d+ 4 . one gas and one coal generators in the set C. .

2 Formulation of the LP stochastic optimization model The constraints and the objective function described in Section 5. . . The aggregated tree results in N = 1000 nodes. 112 for a single t of the rolling horizon 1. . . .3. . g ∈ G 96 . thus. taking V = 6 there are 72. .2 and Figure 5. 3. 2. . . . . . 2. . . . 882 decision variables in the stochastic optimization problem for a single t of the rolling horizon 1. . . T . . V nτ nτ 0 ≤ d+ v ≤ d+ v ∀v = 1. . V nτ 0 ≤ d− v ≤ d− v ∀v = 1. . Then.1 are combined below: Objective (minimize cost): N Tn V Tn + ζv v=1 τ =1 V Tn − ζv v=1 τ =1 min x n=1 pn ( g∈C cg τ =1 xnτ g + d+ v nτ + d− v ) nτ Subject to: xnτ g = wnτ g ∀g ∈ W xn(τ +1)g ≤ e+ + r2+ − r1+ xnτ g + r1+ ∀g ∈ C e+ xn(τ +1)g ≤ xnτ g + r2+ ∀g ∈ C xg ≤ xnτ g ≤ xg ∀g ∈ C xn(τ +1)g ≥ r2− xg − r1− e− xg − e− + r2− − r1− xnτ g − ∀g ∈ C xg − e− xg − e− xnτ g ≤ Dnτ (1 + ∆D ) xnτ g ≥ Dnτ (1 − ∆D ) V xn(τ +1)g ≥ xnτ g − r2− ∀g ∈ C g∈G g∈G g∈G xnτ g − Dnτ = v=1 (d+ v − d− v ) nτ nτ + d+ v = ∆d ∗ Dnτ ∗ βv ∀v = 1. .3. N. T . 2. .3. number of constraints result in 344. V nτ − d− v = ∆d ∗ Dnτ ∗ βv ∀v = 1. . . 2. .Figure 5. . . 2. 5. We assume again one accumulated wind energy curve in the set of wind generators. 2. . V nτ nτ xn0g = xa(n)Tn g ∀n = 2.

.54s for a single t of the rolling horizon 1.v = 1.v and d1. However.1. .v = 1. .3. t = 1. .1. 2.8s for a single iteration or 67. Let (x1g . N .g ) be a vector of the real-time decisions after solving a sequence of T stochastic problems: xtg := x1. To evaluate the operational cost of the electricity generation during the real time rather than the expected cost based on the chosen scenario tree we use the decision variables recorded at the root of the scenario tree while solving the rolling horizon problems. or T = 1440. 2.g . although a number of decision variables and constraints is high. . 2. 2. . .1. d+ := d1. . . (T − T S ) where x1.v . . x2g . 2.v . . Tn . In a similar way a mismatch between the electricity generation and the load is recorded.1. V tv where d1. . .g is the power output at the root of the stochastic tree for the t-th problem of the rolling horizon. .where n = 1. T . it can be efficiently solved with a CPLEX solver. . Using version 10. . .36s of CPU time spent for calculating a new scenario tree for the stochastic optimization problem. . .3 Output of the stochastic model with the uncertainty of the wind power and the capacity loss The decision on the power output of the thermal plants at the current minute t is based on the expected operational cost of the stochastic problem. 2.1. τ = 1.00 of the CPLEX solver CPU time of solving the optimization problem was 0. . . . . . Knowing the fuel cost cg of each generating plant g ∈ C and the cost of the 97 +(t) −(t) −(t) +(t) (t) (t) . . This adds-up to CPU time of 2. The expected cost depends on the heuristic probability distribution of the wind power illustrated with the scenario tree. 5. We wish the System Operator to be ready and respond to the highest fluctuations recorded historically which means that the fluctuations may be stronger than that happening in real time. Formulated stochastic optimization problem is linear and. . xT −T S . V tv d− := d1. . .1.2min for a daily electricity generation problem.v are the positive and the negative mismatch between the electricity generation and the load respectively at the root of the scenario tree for the t-th problem of the rolling horizon. there is 2.

T −T S V T −T S + ζv v=1 t=1 V T −T S − ζv v=1 t=1 Fstoch = g∈C cg t=1 xtg + d+ tv + d− tv (5.19) for two cases: Case A: there is uncertainty in the available capacity of the thermal plants but wind energy output is known for the planning horizon T . .3. . 2. . A scenario tree for this stochastic problem is illustrated in Figure 5. MW 12000 10000 8000 6000 4000 2000 0 0 20 40 60 time. the operational cost can be calculated for the planning horizon t = 1.3. (T − T S ) as follows. A scenario tree for this stochastic problem is a combination of two scenario trees illustrated in Figure 5. d+ . ζv ). 20000 18000 16000 14000 power output.19) where {xtg . there is uncertainty in the available capacity of the thermal plants and uncertainty in wind energy output is represented with the scenario tree.4: An example of the rolling horizon stochastic programming solution with the uncertainty of the capacity loss when 11% of the wind power is incorporated in the total generation To evaluate the system balancing cost that appears when the wind power is incorporated into the electricity generation system let us find the operational cost in (5. d− } is a set of real-time decision variables recorded for (T − T S ) tv tv stochastic problems.+ − mismatch between the electricity generation and the load (ζv .2 and Figure 5. min 80 100 120 gas coal nuclear wind coal limit nuclear limit Figure 5. 98 Case B: .

xT −T S .4 shows an example of the solution for Case A when there is uncertainty of a possible capacity loss but the wind speed is forecast for certain. To calculate the operational cost for the Case A there are decision variables recorded for the solution at the root of the scenario tree: (x1g . the (a) cost of the stochastic problem Fstoch is higher because of the thermal generation running at the level that would allow them responding to a sudden increase or decrease in load. d2v . Problem with 11% of wind energy in the total generation is feasible but the operational cost is high because the Coal and Nuclear generators are taken down from the maximum capacity levels and the cheaper power is replaced by the more expensive gas power output.v ) (d1v . Similarly to Case A. . there are decision variables recorded (a) 99 . .2.v ) The above variables are used to calculate the operational cost by applying them into the right-hand side of (5. is then the following: T −T S (a) Fstoch V (a) xtg t=1 T −T S + ζv v=1 t=1 +(a) dtv V T −T S − ζv v=1 t=1 −(a) −(a) −(a) +(a) +(a) +(a) (a) (a) (a) = g∈C cg + + dtv −(a) If compared with the operational cost Fvar calculated in Section 3. The scheme of having a number of thermal generators away from the output limits to secure feasibility of solution is currently being applied in practise. . . . . . The system balancing cost that appears with the uncertainty of the available capacity of the thermal generation can be estimated as follows: f (a) = Fstoch − Fvar We wish to compare further the results of Case A with that calculated for Case B when there is also the uncertainty about wind power output added to the uncertainty of the load. . . x2g . .3. It is also illustrated with the solution of the stochastic programming problem. .Figure 5. xT −T S .19). d2v . The operational cost of the electricity generation when there is only uncertainty in the available thermal capacity but the wind power output is known. the power system remains within the target frequency until the end of the modelling horizon. Such a scheduling of the thermal plants ensures that in the event of capacity loss.g ) (d1v . xT −T S . .

xT −T S . x2g .20) + where fw is the additional cost of wind power fluctuations. min 80 100 120 gas coal nuclear wind coal limit nuclear limit Figure 5. . . .v ) −(b) −(b) −(b) +(b) +(b) +(b) (b) (b) (b) 20000 18000 16000 14000 power output. . xT −T S .g ) (d1v . . The solution of the stochastic programming problem for Case B is found with the expectation that wind power output is slowly changing over the modelling horizon.v ) (d1v . T −T S (b) Fstoch V (b) xtg t=1 T −T S + ζv v=1 t=1 +(b) dtv V T −T S − ζv v=1 t=1 = g∈C cg + + dtv −(b) In Case B the System Operator does not have the information about the variability of future wind power output as it is not reflected in the scenario tree. . . Therefore. MW 12000 10000 8000 6000 4000 2000 0 0 20 40 60 time. 100 . . . . xT −T S . .5: An example of the rolling horizon stochastic programming solution with the uncertainty of the wind power output and the capacity loss when 7% of the wind energy is incorporated in the total generation The above variables are further used to calculate the operational cost of the electricity generation when there are two uncertainties that the System Operator faces: uncertainty of capacity loss and uncertainty of wind power output. .for a rolling horizon stochastic problem of Case B: (x1g . d2v . the cost of the uncertainty in the wind power is calculated as follows: (b) (a) + f (b) = Fstoch − Fstoch + fw (5. d2v .

18 operational cost weighted by total generation. However.1 percent of wind power in total generation 0. pounds/MW 0.5 plots an example of the rolling horizon stochastic programming solution when 7% of the wind power is incorporated in the total generation. 5. when the amount of wind exceeds 101 .04 0.1 0.16 0. shown for different levels of the wind introduced into the electricity system Figure 5.08 0. This is the highest possible amount of wind energy that can be incorporated into the electricity generation system. For up to 13. the highest cost is the fuel cost calculated for a problem when there is no wind energy in the power system.06 0.6: The fuel cost of the uncertainty in wind power output and the available thermal capacity.06 0.14 0. the operational cost of Case A and Case B are estimated for a day (or 1440 minutes) of electricity generation for different levels of wind power in the total generation.4 The system balancing cost when there is the uncertainty of the wind power in the electricity generation system In order to calculate the system balancing cost of the uncertainty in the wind and thermal generation. d− ) are given nont4 t4 zero values.14 0. which means that the mismatch between the electricity generation and the load is so high that it affects the target frequency.08 0.12 0.Figure 5. With wind power higher than 7% in the total generation. the variables of the fourth penalty cost band (d+ .02 0 Figure 5.02 0.6 plots the fuel costs of a daily electricity generation weighted by the total amount of the produced power.3% of wind energy in the total generation.12 0.04 zero_wind flat wind profile variable wind profile capacity loss uncertainty wind capacity uncertainty 0 0. 0.

Two optimization models that we developed in this work.20). 102 . Building a scenario tree that includes both. In case variable wind energy is also uncertain the system balancing cost can be estimated as in (5. the balancing cost of the uncertainty in the wind power output combined with the uncertainty of the available thermal capacity is so high that the power system does not benefit from the incorporation of extra wind into the electricity generation system.3% in the total generation. deterministic and stochastic. Even at 0% of wind energy integration the balancing cost of variability is complemented by the cost of uncertainty in loss of thermal generating capacity. Even though it is possible to estimate the system balancing cost of variable and unpredictable wind energy using the results of deterministic and stochastic optimization models. variability and uncertainty of wind energy can be suggested as a future project.13. we wish to emphasize the importance of developing a model that captures both characteristics of wind. Each of the models covers only one aspect of the wind speed: variability or uncertainty. interact when estimating the system balancing cost of variable and unpredictable wind energy incorporated into the electricity generation system.

Chapter 6 Conclusion This thesis presented a study of the statistical characteristics of one-minute wind speed and estimated the system balancing costs when the wind power is incorporated into the electricity generation system. In this work we were focusing on one-minute electricity generation when the System Operator makes a decision on the output of the available thermal generators. Gibbs sampling algorithm. It ensures that as the number of iterations of sampling from a known distribution increases. Least frequent data was presented by ten-minute wind speed published by the Utah Geological Survey. therefore. The problem of the increase in the system balancing cost arises as a result of variable and unpredictable nature of the wind speed that transferred into the wind power output. To do this. first used by Glasbey et al. We originally adapted an algorithm that transfers one-minute wind speed into the deseasonalised and normalised wind speed. there is no one-minute wind speed available on the country level. the density of a resulting set of variables converges to the required one. ([16. we developed a unique algorithm based on one of the Monte Carlo Markov Chain methods. which leads to different problems formulated on time scales of a year. However. one minute or one second. This area was not well-researched before but has an important application in the industry. so that the modified Gibbs sampling algorithm is described as follows: step A: generate normalised wind speed that further transfered into original wind speed. We wish to use optimization modelling to estimate the system balancing cost that requires the appropriate data with one-minute resolution. 17]). 103 . hour. we wish to generate unbiased samples from the wind stochastic process conditional on each consecutive group of wind speeds having a given average value. National Grid has to balance the electricity supply and demand every moment.

can be determined as a function of the statistical parameters of the wind energy. 104 . a new value of the shifted normalised wind speed is found within a specified tolerance interval of the probability space. we developed a unique model of the actual accumulated fuel cost. The levels of the operating reserve required at any given time depend partly on the uncertainties in the electricity load but also in the available generating capacity. The system balancing cost caused by the fluctuations of the wind energy was estimated through a series of cases. The power system benefits from the fact that wind power does not include fuel costs but the operational cost is increased when the thermal plants modify the power output in order to balance the electricity generation and the load. This thesis investigated the electricity generation on one-minute basis and. absolute value of gradients and standard deviation of the wind energy curve. step C: Iterations: Generated samples of one-minute wind speed were further used in deterministic and stochastic optimization modelling of the electricity generation system.step B: if the average of the generated one-minute wind speed does not equal a given average ten-minute wind speed. depending on the mean. the variability of the wind energy and the electricity generation with one-minute frequency. We assumed in this work that the saving in fuel cost and the additional fuel cost appearing when wind energy is incorporated into the electricity generation system. steps A to C are repeated a big number of times that ensures the probability distribution of the resulting set is secondorder stationary process. therefore. To demonstrate this. The system balancing costs appear when the System Operator uses or keeps on stand-by flexible but expensive generating units in order to deal with an unexpected mismatch between the electricity supply and demand. Electricity supply and demand in the transmission and distribution power systems need to be balanced every single moment of time. it is shifted along the vector of the maximum probabilities of the probability distribution. Both optimization models include common constraints of the electricity generation while our original contribution consists in reflecting the characteristics of the UK power system and developing algorithms that estimate the system balancing costs of variable and unpredictable wind energy.

We wish to suggest exploring different scenario trees in future work that would allow us to estimate the system balancing cost of variable and unpredictable wind speed in one model. 105 . The fact that the deterministic and the stochastic optimization models cover only one aspect of wind speed. means that they have to interact in order to estimate the system balancing cost. the statistical characteristics of the wind speed have been explored and the system balancing costs of wind speed variability and unpredictability have been estimated. In summary. However. We also leave the matter of interaction between the electricity demand and wind speed for future research.In reality the System Operator does not have the information about the variability or the capacity of the future wind power output. Our original contribution consisted in developing a scenario tree that reflected two uncertainties: uncertainty in the available capacity and uncertainty in wind energy output. The system balancing cost of this uncertainty was evaluated using stochastic programming when the flow of the information and a sequence of the decisions by the System Operator were illustrated with a scenario tree. the recent developments in construction of wind turbines as well as introduction of smart metering into the global energy market could open a new area of application for the results of this work. variability or uncertainty. The results showed that depending on the statistical parameters and the geographic distribution of the wind speed the system balancing cost of the electricity generation can be significant.

Network operator is a person with a User System directly connected to the GB Transmission System to which Customers and/or Power Stations (not forming part of the User System) are connected. 106 . Definitions by the National Grid NGET is National Grid Electricity Transmission plc (NO: 2366977) whose registered office is at 1 − 3 Strand. Gate Closure means. Generator is a person who generates electricity under licence or exemption under the Act acting in its capacity as a generator in Great Britain. London. Settlement Period is a period of 30 minutes ending on the hour and half-hour in each hour during a day.Appendices Appendix A. Supplier is • (a) A person supplying electricity under an Electricity Supply Licence. but shall not include a person acting in the capacity of an Externally Interconnected System Operator. acting in its capacity as an operator of the User System. Operational Day is the the period from 05 : 00 hours on one day to 05 : 00 on the following day. in relation to a Settlement Period. Balancing Mechanism (BM) is a period of time which allows the System Operator to call upon additional generation/consumption or reduce generation/consumption in order to balance the System minute by minute. or • (b) A person supplying electricity under exemption under the Act. Customer is a person to whom electrical power is provided (whether or not he is the same person as the person who provides the electrical power). the spot time 1 hour before the spot time at the start of that Settlement Period. From July this period of time will be one hour before each trading period. in each case acting in its capacity as a supplier of electricity to Customers in Great Britain. W C2N 5EH.

from a zero Physical Notification level as a result of a Bid-Offer Acceptance. • Minimum Zero Time (MZT). at all times during that Settlement Period. in the absence of any Acceptances. • Notice to Deviate from Zero (NDZ) output or input. expressed in MW/minute and associated Run-Up Elbow(s) and Run-Down Elbow(s). Export (Import) Limits is a series of MW figures and associated times. 107 . Physical Notification means. where the context so requires. expressed in MW for output and the same for input. a notification made by (or on behalf of) the Lead Party to the Transmission Company under the Grid Code as to the expected level of Export or Import. • Notice to Deliver Offers (NTO) and Notice to Deliver Bids (NTB). expressed in minutes. expressed in minutes. being either the minimum time that a BM Unit which has been exporting must operate at zero or be importing. In the case of a BM Unit comprising a Genset. Dynamic Parameters comprise • Up to three Run-Up Rate(s) and up to three Run-Down Rate(s). as at the Transmission System Boundary. in respect of a Settlement Period and a BM Unit. It should be noted that Run-Up Rate(s) are applicable to a MW figure becoming more positive. NTO and NTB will be set to a maximum period of two minutes. as appropriate. and the associated times for such MW levels. the Plant and/or Apparatus treated as comprised in or assigned to such unit for the purposes of the Code. Quiescent Physical Notification is the data that describes the MW levels to be deducted from the Physical Notification of a BM Unit to determine a resultant operating level to which the Dynamic Parameters associated with that BM Unit apply. The MW level of the QPN must always be set to zero. indicating the notification time required for a BM Unit to start delivering Offers and Bids respectively from the time that the Bid-Offer Acceptance is issued.BM Unit means a unit established and registered (or to be established and registered) by a Party in accordance with Section K3 or. making up a profile of the maximum level at which the BM Unit may be exporting (importing) in MW to the GB Transmission System at the Grid Entry Point or Grid Supply Point. being the notification time required for a BM Unit to start importing or exporting energy.

in its reasonable opinion. This will normally be 50. under stable conditions. expressed in minutes. except in exceptional circumstances as determined by NGET. Target Frequency is that Frequency determined by NGET. An example of exceptional circumstances may be difficulties caused in operating the System during disputes affecting fuel supplies. Primary Response is the automatic increase in Active Power output of a 108 . This reduction in Active Power output must be in accordance with the provisions of the relevant Ancillary Services Agreement which will provide that it will be released increasingly with time over the period 0 to 10 seconds from the time of the Frequency increase on the basis set out in the Ancillary Services Agreement and fully achieved within 10 seconds of the time of the start of the Frequency increase and it must be sustained at no lesser reduction thereafter. • Stable Export Limit (SEL) expressed in MW at the Grid Entry Point or Grid Supply Point. being the minimum value at which the BM Unit can. being the maximum period over which the MDV applies. as appropriate. • Minimum Non-Zero Time (MNZT). being the maximum number of MWhr of Offer (or Bid if MDV is negative) that a particular BM Unit may deliver within the associated Maximum Delivery Period (MDP). being the minimum time that a BM Unit can operate at a non-zero level as a result of a Bid-Offer Acceptance.before returning to exporting or the minimum time that a BM Unit which has been importing must operate at zero or be exporting before returning to importing. • Stable Import Limit (SIL) expressed in MW at the Grid Entry Point or Grid Supply Point.10Hz. under stable conditions. export to the GB Transmission System. being the minimum value at which the BM Unit can.00Hz plus or minus 0. as the desired operating Frequency of the Total System. High Frequency response is an automatic reduction in Active Power output in response to an increase in System Frequency above the Target Frequency (or such other level of Frequency as may have been agreed in an Ancillary Services Agreement). • Maximum Delivery Volume (MDV). expressed in MWh. as appropriate.90 or 50.05Hz. expressed in minutes. as a result of a Bid-Offer Acceptance. expressed in minutes. in its reasonable opinion when this may be 49. import from the GB Transmission System.

as the case may be. Operating Reserve is the additional output from Large Power Stations or the reduction in Demand. This increase in Active Power output or. Good Industry Practice is the exercise of that degree of skill. the decrease in Active Power Demand in response to a System Frequency fall. prudence and foresight which would reasonably and ordinarily be expected from a skilled and experienced operator engaged in the same type of undertaking under the same or similar circumstances. as the case may be. diligence. and sustainable for at least a further 20 seconds. the decrease in Active Power Demand in response to a System Frequency fall. the decrease in Active Power Demand must be in accordance with the provisions of the relevant Ancillary Services Agreement which will provide that it will be released increasingly with time over the period 0 to 10 seconds from the time of the start of the Frequency fall on the basis set out in the Ancillary Services Agreement and fully available by the latter. as the case may be. This increase in Active Power output or. Secondary Response is the automatic increase in Active Power output of a Genset or. the decrease in Active Power Demand must be in accordance with the provisions of the relevant Ancillary Services Agreement which will provide that it will be fully available by 30 seconds from the time of the start of the Frequency fall and be sustainable for at least a further 30 minutes. 109 .Genset or. which must be realisable in real-time operation to respond in order to contribute to containing and correcting any System Frequency fall to an acceptable level in the event of a loss of generation or a loss of import from an External Interconnection or mismatch between generation and Demand. as the case may be.

Diversified renewable energy portfolios for the UK.Sinden. UK. 2005 [7] Quaterly energy prices: December 2008 Department for Business Enterprise and Regulatory Reform (BERR) http://www.co. Quantifying the system costs of additional renewablesin 2020. Wind power and the UK wind resource.Bibliography [1] Government of United Kingdom. [2] Government of United Kingdom. 2005. 110 . Strbac.uk [8] G. UK. Berlin. [3] National Grid Code http://www.com/ supported by ELEXON Ltd. May 2007.1 Designated 21st March 2001 [5] Balancing Mechanism Reporting System http://www. 2005. Renewables Obligation. October 2002. UK. Report of the Sustainable Development Comission UK. A guide to the key issues surrounding onshore power development in the UK. presentation.berr.com/uk/Electricity Issue 3 Revision 32.Sinden. University of Oxford. 2002.bmreports. University of Oxford. White paper on energy. October 2007. Environmental Change Institute. Strbac. Challenges for market design to facilitate cost effective integration of renewables: a UK perspective.gov.uk V1. [10] G.Sinden. Diversity in the electricity generating sector. University of Oxford. [11] G. UK. Environmental Change Institute. [12] G. 2005. UK.elexon. Environmental Change Institute. UK.nationalgrid. [9] G. [6] Wind Power in the UK. 8 December 2008 [4] Balancing and Settlement Code http://www.

F. Design and Application. SCRI Annual Report.G.Allcroft and C.Glasbey. Weather modelling using a multivariate latent Gaussian Markov random field.Durban and C. 2001.82. Surface wind speed measurements funded by Natural Environment Research Council. Applied Statistics. 192-195. N410.M.Allcroft and C.L.utah.J. G. The Open University. M.Ramage. Stochastic relaxation.85.Birge and F. Theory. Earthscan. [14] A. [17] D.E.Glasbey. 2003.Durban and C. Wind Energy Explained.Geman. J. 2007.htm Program. 187201. [18] M. UK. Agricultural and Forest Meteorology.Gelfand and A.H.A.Boyle and J. Sampling-based approaches to calculating marginal densities. http://mst.Louveaux.. 1997.Wong. SringerVerlag New York. A latent Gaussian Markov random field model for spatio-temporal rainfall disaggregation. [20] British Atmospheric Data Center. Spatial disaggregation of rainfall using a latent Gaussian Markov random field.uk/ [21] Utah Geological Survey. Renewable electricity and the grid: the challenge of variability.gov/SEP/wind/anemometerdata/index.J. 52. Gibbs distributions. [22] J. IEEE. [16] D. 2003.A. [24] G. 1984. [25] J. The calculation of posterior distributions by data augmentation. Proceedings of the 17th International Workshop on Statistical Modelling. USA. Utah’s Anemometer Loan http://geology. Vol.Smith.Geman and D.ac.A. 111 .Everett. editors. 487-498. Journal of the American Statistical Association.Boyle. 2002.[13] S.Manwell.Glasbey.A. Introduction to stochastic programming. Energy systems and sustainability.Allcroft.Rogers. [15] M. 2001. [19] D.Tanner and W. Modelling weather data.R. UK.nerc.Glasbey.McGowan and A. 2002. 1990. Vol. Journal of the American Statistical Association. 109. [23] B. 1987.J. Inc.F. N398. John Wiley and Sons.A. and the Bayesian restoration of images.

Applications of Stochastic Programming.Flam.C. 1996 [32] C.Kall and S. USA. Preprint. Stochastic Programming Methods and Technical Applications. Institut fur Mathematik.Feitosa.com.H.W. 2003. Wiley.Hardenberg. Nonanticipativity in Stochastic Programming Journal of Optimization theory and applications. Renewable Energy. 28:585–602.T. Economic value of weather and climate forecasts.Lindberg. 1996 [30] A. M¨ller. H. Chichester etc.Ziemba. R. Stochastic Programming. Stochastic Programming.nationalgrid.R¨misch Optimal Power Generation under Uncertainty o via Stochastic Programming. Springer. E.Costa. Mathematik.Dentcheva and W. M. A. A.Smith. D.Roulston. 12 (2008) 1725-1744 112 .Lizcano. Operations Research Proceedings.S.T..Kaplan and L.Schultz. [36] J. 1995 e [27] P. Solving the o o Unit Commitment Problem in Power Generation by Primal and Dual Methods. UK. W. web-site www.C.Murphy. P.Navarro.Feltenmark. Konrad-Zuse-Zentrum fur Informationstechnik. 1994. 2003.Kiwiel. A Dual Method for the Unit Commitment Problem. [28] S. J.Schultz A two-Stage Stochastic Program for Unit Commitment Under Uncertainty in Hydro-Thermal Power System. Vol.Pr`kopa.R¨misch. Preprint SC 98-11. Using mediumrange weather forecasts to improve the value of wind energy production.46.M¨ller.Wallace and W. 2008.Dentcheva. 2005 [29] D.W. [37] R.Crespo. SIAM. UK. W.A. 1985 [35] National grid electricity transmission. A review of the young history of the wind power short-term prediction Renewable and Sustainable Energy Reviews.D. No 1.R¨misch and R. o o Preprint Nr 95-1.Gollmer. editors.Katz and A. G. 1997 [34] S. K.W.Wallace.[26] A.1998 [33] D.Caroe and R.Madsen. Cambridge University Press. Humboldt-Universitat Berlin. [38] A. 1995 [31] S. Solving Unit Commitment Problems in Power Production Planning.

. A novel approach to wind forecasting in the United Kingdom and Ireland International Journal of Simulation: Systems.Tetzlaff.R. 2004. John Wiley and Sons. Science and Technology... April 2004 [46] B. 6th edn.Thomopoulos. 2004 Annual Conference Proceedings. [43] C.Chatfield. Truncated and Cencored samples from Normal Populations Marcel Dekker.Fourer. USA. New York. The analysis of Covariance and Alternatives. D. Renewables and Climate Change Sixth report Session 2005-2006 [42] E. No 12-13. Cleveland.Campbell and K.Gay.J. No 60. [44] H. [45] A. The analysis of Time Series: an Introduction. Inc. Chapman and Hall. 113 .C.Johnson.T. 2003.W. 1997.Huitema. 1998 [41] Report from the Environmental Audit Committee Keeping the lights on: Nuclear. 6. Wind speed forecasting from 1 to 30 minutes. USA. N. [47] R. 1986.[39] P.H. USA.Schneider. USA. Inc.Kernighan. 1980. AMPL: A modelling language for mathematical programming Springer-Verlag New York. Vol.Flury. London. 2005 [40] U. B.Schlink and G. Theoretical and Applied Climatology. UK.M. A first course in Multivariate statistics Springer-Verlag New York.Adamson. Characteristics and tables of the lefttruncated normal distribution Midwest Decision Sciences Institute.