A Fundamental Approach: Valuation, Hedging and Speculation in Competitive Electricity Markets
A Fundamental Approach: Valuation, Hedging and Speculation in Competitive Electricity Markets
A Fundamental Approach: Valuation, Hedging and Speculation in Competitive Electricity Markets
SPECULATION IN COMPETITIVE
ELECTRICITY MARKETS
A Fundamental Approach
THE KLUWER INTERNATIONAL SERIES
IN ENGINEERING AND COMPUTER SCIENCE
by
Petter L. Skantze
Caminus Corporation
Marija D. Ilie
Massachusetts Institute ofTechnology
P.L.s.
To my husband Jeff,
who continues to believe in what I do.
M.D.I.
Table of Contents
PREFACE ........................................................................ xii
1 INTRODUCTION ................................................................................. 1
APPENDIX B ....................................................................201
Introduction
The purpose of this book is to build a framework to solve physical and
financial commitment decisions contingent on electricity, in a deregulated
market environment. The set of problems include valuing investment
opportunities in physical assets, valuing and hedging obligations to serve
customers, and the pricing of electricity dependent derivative contracts.
Electricity markets suffer from a severe case of over-dimensionality. Due to
the lack of economic storage of the commodity, each time interval of delivery
can be considered a separate product. Furthermore, the scarcity and
complexity of the transmission system leads to significant locational
variations in price. The combination of the temporal and spatial properties of
electricity poses significant barriers to market liquidity, and makes it
exceedingly hard for market participants to solve valuation and risk
management related optimization problems.
Using a model based on supply and demand states, one can explicitly
incorporate the network flow constraints. The resulting probability
distribution of the price spread is in stark contrast to what is implied by
traditional approaches. This has tremendous implications for the pricing of
transmission rights and locational price derivatives. The model is also
extended to allow for market based analysis of hypothetical expansions to the
transmission grid, allowing transmission companies to estimate the value of
new investments.
The final section of the book covers the dynamics of new investment in
generation capacity. It is shown how the time delay in information from the
spot market to the investment decision, and further in the installation process
of new power plants, can lead to periods of over and under capacity. We
further analyze the relationship between price trends and the physical
6
reliability of the market. Specifically we address the dangers related to
excessive government intervention in order to curb price spikes.
Chapter 2
f
T
where E{.} is the expected value operator, which is necessary since future
payoffs are generally uncertain.
The NPV rule states that if the net present value is positive, then the firm
should enter into the commitment, assuming there are no other commitment
options available. If there are multiple options, the firm should choose the
one with the highest NPV. To use the NPV criteria, one must overcome two
challenges. The first is estimating the expected value of future payoffs. The
second challenge lies in determining the appropriate discount rate o. The
discount rate contains information about the time value of money, or the
opportunity cost of any capital which is tied up in the commitment and cannot
be spent elsewhere. This cost would generally be set equal to the firms cost
of raising new capital, for example by issuing bonds. For the remainder of the
book, unless stated otherwise, we are going to assume that this component of
the discount rate is equal to zero. As long as the cost of capital is
deterministic, this assumption can be made without loss of generality.
The discount rate must also reflect the level of risk in the investment.
Assuming the firm is risk averse, the discount rate will increase with the level
of uncertainty associated with future payoffs. A commitment option with
lower expected cash flows, may be preferable to an option with higher
expected cash flows if the risk level is significantly lower. Intuitively this
argument is easy to understand. Quantifying a firm's risk preference is more
difficult. One possibility is to equate the level of risk with the variance of the
cash flow. This puts a greater requirement on the modeling of future cash
flows, since the firm needs to estimate variances in addition to expected
9
values. Furthermore, a firm needs to consider the risk associated with the
cumulative cash flow, not just the instantaneous variance at each point in
time. To calculate the variance of the sum of the cash flows, the firm needs to
estimate the covariance matrix of all the future cash flows.
T
\jItot =L \jIt
t=to
T T
var(\jItot)= LLcov(\jIj\jlj)
i=to j=to
The firm can then define its risk preference by stating its utility (U) in
terms of the tradeoff between the expectation and variance of the return on a
commitment. An example of this is the mean variance utility function
Cash Flow ..
--,.
Firm's Utility .
--,.
Optimal
Model Function Decision Rule
So far we have assumed that the market assigns a value to the traded asset
at any given point in time. This value, known as the spot price, is the cost of
purchasing the asset for immediate delivery. The spot market, however, is
only one of several venues available for a firm to trade in a given asset.
Mature markets generally trade forward contracts as well as numerous options
contracts in addition to the spot. Forward contracts specify the delivery of a
given quantity (q) of the asset for a fixed price (F) at a specific future time
(T), known as the maturity of the contract. The seller of the contract
generally has the option of paying the prevailing spot price at maturity rather
than delivering the actual asset. The payoff for the buyer, or long position, of
the forward contract at maturity can therefore be written as
The payoff, and therefore the value, of the forward and call option
contracts are functions of the spot price of the underlying commodity. We
refer to the general category of traded contracts with this property as
derivatives of the underlying asset. It is important to note that the
commitment for the individual firm, as defined in the previous section, is
effectively a derivatives contract. The payoff from the commitment is a
function of future levels of the spot price. All commitment opportunities
discussed in this book are effectively derivatives of the electricity market, or
the price of affiliated commodities, some with multiple underlying assets. A
key distinction is that some derivatives, such as forward contracts, are
publicly traded, while others are available to only a limited set of investors, at
a given point in time. In this section we address the relationship between
traded and non traded contracts. Particular emphasis is put on the information
value of price signals from traded contracts, and to which extent they can be
used to imply the value of non-traded contracts. A distinction has to be made
between the value of a contract to an individual firm, and the market valuation
of the contract. The value of a contract to a firm is contingent on the firm's
risk preference. The market value of a traded contract on the other hand is
uniquely defined by its current price.
14
By studying the historical behavior of the spot market, one might arrive at
a reasonable model of the stochastic properties of future spot prices. Assume
that, at time t, a firm uses all available historical information about the spot
market to estimate the distribution of the spot price at some future time T.
Next the firm observes a forward price F(t,T). We assume that participants in
the forward market are rational. That is, they will purchase or sell a forward
contract only if it increases their total utility. For the sake of discussion, we
assume that all participants have risk preferences which can be characterized
by the mean variance formulation, so that each individual firm j, has a utility
Ui defined by the properties of the total cash flow ~.
15
Furthermore, assume that all cash flows are the result of trades on the
forward market, and all cash flows occur at time T. If the utility function of
all participants were independent of risk, that is, the r parameter were zero for
all firms, then the utility from owning a long position in the forward contract
is given by
F(t, T) = E t {ST }.
ul =Et{ST}-F(t,T)-rl var{ST} ~O
U 2 =F(t,T)-Et{ST}-rl var{ST}~O
fl ,f2 ~ O.
There is no set of risk premium ri which will satisfy this criteria. Does this
mean that forward markets cannot exist if all market participants are risk
averse? What we have neglected is that when the variance of the return is
incorporated into the utility measure, the incremental change in the utility is
no longer independent of the firm's other trades. Specifically, if a firm has an
existing obligation with payoffs which are negatively correlated to the spot
price at time T, then entering into the long position in a forward contract will
tend to reduce the overall variance of the firm's returns. The forward contract
then takes on the function of a hedge, offsetting the risk from an underlying
17
position. In commodity markets, a major function of forward markets is to
hedge the positions of producers and consumers of the commodity. Producers
have physical assets which make them naturally long in the commodity (with
payoffs positively correlated to the spot price), while consumers are naturally
short (with payoffs negatively correlated to the spot price). When a producer
sells a forward contract to a consumer, both parties decrease the overall
variance of their future cash flows, and it is therefore possible to find a set of
forward prices for which both sides increase their expected utility.
L
n
TI(t) = Wi (t)X i (t)
i=l
TI(to) =0
and for some t > to
Prob(l1(tj < 0)= 0
Prob(l1(tj > 0» 0
This means that we can construct a portfolio with zero cost, which has zero
probability of decreasing in value and a strictly positive probability of
increasing in value. Since the portfolio has zero initial cost, any market
participant can purchase an unlimited amount of the portfolio, and enjoy a risk
free guaranteed profit. The theory is that as arbiters start to take advantage of
this opportunity, they will create an upward price pressure on assets with
positive weights in the arbitrage portfolio, and downward price pressure on
assets with negative weights. Prices will then reach a new equilibrium where
the arbitrage opportunity no longer exists.
19
2.2.5 Application of Arbitrage Pricing Theory in Valuing
Forward Contracts
Assume the current price of the stock, which pays no dividends, is St and
the risk free interest rate is r, continuously compounded. The price of a
forward contract on the stock (F(t,T» with delivery date T must then be ef(T-
t)St. To see why this is true consider the following cases:
1. If F(t,T» ef(T-t)Sr. the investor should sell one forward contract, borrow St
dollars at the risk free rate (assuming this is possible), and buy one unit of
stock. The net cash flow at time t is zero. At time T, the investor delivers
the stock against the forward contract, receives F(t,T) dollars as payment
for the forward, and er(T-t)St dollars to payoff his debt. The net cash flow
at time T is F(t,T)-ef(T-t)St>O. This is a pure arbitrage opportunity, which
cannot be sustained in an efficient market, and therefore sets the upper
limit for the forward price.
2. If F(t,T)< ef(T-t)Sr. the investor should buy a forward contract, short-sell one
stock, and lend St at the risk free rate. The net cash flow at time t is zero.
At time T, the investor pays F(t,T) and receives delivery of the stock from
the forward contract. He uses this stock to repay his short-selling
obligation. He also recovers ef(T-tJS t from the money loan. The net cash
flow is er(T-t)Sr-F(t,T»O . This is again a pure arbitrage opportunity,
setting the lower limit for the forward price.
In this case the upper and lower limits for the forward price are identical
and, therefore, in an efficient market where participants can borrow and lend
at the risk free rate, the forward price must be given by: F(t,T)=er(T-t)St. This
20
illustrates two important points. First, under no-arbitrage conditions, the
forward price of a stock is a deterministic function of the spot price and the
time to maturity (T-t). Second, there is a smooth convergence of the spot and
forward prices at maturity.
Assume the current unit price of the commodity is Sb the present value of
the total cost of storage incurred during the length of the futures contract is U,
and the risk free interest rate is r. The lower bound on the futures price for
delivery at time T is F(t,T»(St+U)er(T-t). If this does not hold, an investor can
receive a risk-free profit by borrowing St+U at the risk free rate, purchase the
commodity and payoff the storage cost, and short a forward contract in the
commodity. The cash-flow at time t is zero, and the cash-flow at time T is
F(t,T)-(St+U)er(T-t»O. This is known as cash and carry arbitrage.
Payoff at each time step from cash and carry arbitrage:
t T
Buy commodity to be -St 0
delivered against forward
contract.
Sell forward contract 0 F(t,T)
Pay storage cost -U 0
Borrow now, repay at maturity St+U _(St+U)er(T-t)
Total Cash Flow 0 F(t,T) _(St+U)er(T.t»O
Cash and carry arbitrage establishes an upper bound for the forward price
of the commodity. The bound converges to the spot price as we reach
maturity (T=t), and hence if the forward price is lower than the spot price then
the two prices must converge at maturity.
21
Much of the work in finance in the last thirty years is a direct outgrowth of
a set of seminal papers published by Fisher Black, Myron Scholes and Robert
Merton in the early seventies (see [2],[4]). In their work, Black and Scholes
proposed that stock prices (S) could be modeled as a simple stochastic process
known as geometric Brownian motion:
dS = IlSdt + crSdz
Next they recognized that since the same random input (dz) was driving
both stochastic processes, an investor could hold a combination of the stock
and derivative so that the random components cancel each other out perfectly.
Specifically, the risk neutral portfolio is formed by purchasing one derivative
contract, and short selling aflaS contracts of the stock. The process
describing the change in value (or return) of the portfolio (n) is then given by
This equation does not contain any random terms. If the portfolio return is
deterministic, Black and Scholes argued, the return must be equal to that of a
risk free loan, i.e. the risk free interest rate. Equating the return of the
portfolio to that from lending money at the risk free rate,
22
dII = rTIdt
iT = max(Sr - X) .
L1 = ai
f as
While APT provides a convincing argument for why physical and financial
forward prices must be equal at all times, actual observations in the market
place show that the two markets can diverge at times. The reasons for this
inconsistency can be found in the assumptions underlying the arbitrage
argument. The following points illustrate how market realities deviate from
the theory:
The effects of cash and carry arbitrage can also be interpreted as a dynamic
relationship between spot and forward prices. Assume that at time t we
observe a forward price F(t,T), which violates the upper bound imposed by
APT. We would expect the following behavior in the market.
25
Now consider the reverse condition, when forward prices drop below spot
market levels. In this case, no pure arbitrage strategy is present, since it may
not be possible to short sell a physical commodity on the spot market.
However, consider the position of a market participant who is currently
holding an inventory of the commodity. For this person, the optimal strategy
will be to sell the inventory today, and purchase cheap forward contracts
which can be used to restore the inventory at a later date. If there is significant
inventory in the market, this will put downward pressure on the spot price,
and upward pressure on the forward price.
One can question whether the bounds set by APT are valid under realistic
market conditions. This is especially true for commodities with thin forward
markets and high transaction cost. However, whether or not the bounds are
quantitatively accurate, the qualitative interaction between spot and forward
prices can certainly be observed.
The need to model the dynamic relationship between the spot and forward
prices in storable commodities has led to the notion of convenience yield (y)
(see [2], [5], and [6]), which is defined as:
The convenience yield represents the premium the market is willing to pay
in order to physically hold the commodity today, rather than a promise for
delivery at time T. We can model y as a deterministic parameter, or a
stochastic state of the system depending on the market.
F(t,T) = E t {Sr} .
Mean reversion dictates that if there is a spike in today's spot price, the
effect of that spike on the expectation of future spot prices will decrease with
time (T). Equivalently, the effect of a spike in the spot price on the forward
price F(t,T) is a decreasing function of the time to maturity (T-t).
Term structure is easiest to understand in the context of a specific model.
One of the simplest, and most popular, is the one-factor, lognormal, mean
reverting model
dF =ae-a(T-t)dZ.
F
df.1 = Kdt+aLdZL.
f.1
29
The market entities described so far are generally for profit companies,
looking to take advantage of the opportunities provided by the competitive
marketplace. Electricity markets however, are delicate physical networks,
which can easily break down if pushed beyond their operating limits. To
ensure the physical safety of the grid, regulators have encouraged the
formation of independent system operators (ISOs). An ISO is a non profit
entity, which acts as a supervisor of the physical transactions registered
between power suppliers and customers. The two main functions of the
system operator are to balance power, and manage congestion on the grid.
The power balancing problem is an inherent result of the non-storability of
electricity, forcing the system operator to maintain a constant stand-by reserve
of spinning power capacity, so that a sudden loss of generation on the system
does not lead to a drastic drop in frequency. The congestion management
problem results from a nonlinear relationship between power injections and
flows, forcing the system operator to implement complex pricing systems to
prevent the overloading of transmission lines (see chapter on multi market
modeling).
34
3.2 ELECTRICITY MARKETS
There are three fundamental markets available for trading electricity: the
spot market (day ahead), the physical forward or bilateral market, and the
financial futures market. In addition, there are a number of standard as well as
over the counter options contracts traded, either through exchanges or on a
purely bilateral basis. Before attempting to develop models that describe the
pricing of these various contracts, we need to understand the manner in which
electricity is traded.
A producer wishing to sell power submits a bid curve to the exchange. The
bid curve describes the willingness of the producer to deliver power as a
function of market price. For example a producer may be willing to supply a
total of 50MW if the price is $20IMW, and may offer to supply a total of
lOOMW if the price increases to $30IMW. Bid curves are generally supplied
on a day-ahead basis, and a different bid curve may be specified for each of
the 24 operating hours.
The exchange gathers all the bids from power producers, and similar bids
from consumers. The bids are used to compile aggregate supply and demand
curves for each hour. The intersection of the demand and supply curves
determines the market clearing price (Mep). All supply bids with a price less
35
than the MCP are accepted, and are paid the clearing price. Similarly all
demand bids with a price higher than the MCP are accepted, and are charged
the clearing price. This ensures that demand and supply commitments match
perfectly, and also that the exchange remains revenue neutral.
I
I
MCP
r---------,
I I
I I I
I I
________ 1
I
I
I I
L _________ J
Hour 1 Hour 2 I
I
Hour 3 I
I
Time
g janOO (t)
G(t) = g febOO (t)
gmarOl (t)
where ST is the spot price at maturity T, and F(t,T) is the price of the futures
contract at the time t it was entered into. The problem which occurs with
electricity is that the delivery period for the futures contract is one month,
while the underlying spot process is updated on a day-ahead basis. As a
result, when the futures contract matures on the 4th business day prior to the
1st day of the delivery period, the spot prices for hours in the delivery month
are not yet known. Hence the contract cannot be settled financially at this
time. To circumvent this problem, exchanges have taken on two different
approaches: ex-post settling and ex-ante settling.
The total cash flow for the long position over the duration of the delivery
month is given by
38
n
CFex - anre =Lq(G(T,T) - F)
;=1
Both the day-ahead spot and physical forward are based on the same
commodity: electricity delivered at a specific grid location. However there is
no simple mapping between the ex-post average spot price and the ex-ante
physical forward price. This is a very crucial point to understand in electricity
markets. While the settling procedure differs from market to market, the
dominant trend seems to be in the direction of ex-post settling, as seen in
California and Nordpool. Unless otherwise specified we will from now on
assume that financial forwards settle ex-post.
39
3.2.4 The Derivatives Markets
Without the ability to execute an arbitrage between the spot and forward
markets, APT is useless in predicting the relationship between the two
markets. Instead we have to address the forces underlying the supply and
demand in forward markets. One approach is to assume that the market as a
whole is liquid enough that every participant holds a small fraction of the total
risk. As a result the market effectively behaves in a risk neutral manner, even
if the individual participants are risk averse, allowing us to pose the
relationship
F(t, T) = E {ST}.
f
Risk neutral formulation is the basis for most risk management and option
pricing theories in commodities markets. The problem with this assumption is
that electricity markets are relatively illiquid, with a small number of
participants. In light of this we here propose a more general model allowing
for the existence of a risk premium in the market. We model the forward price
as a function of the spot price, the variance of the spot price, and a random
disturbance (ZF):
The exact structure of the forward risk premium is likely to vary from
market to market.
43
4.1 IS ELECTRICITY REALLY NON·STORABLE?
Consider the owner of a gas fired generator, who purchases the fuel for the
plant on the gas spot market, and sells the output on the electricity spot
market. We assume that the generator is unable to exercise market power: that
is, he takes fuel and electricity prices to be exogenous variables. In our setup,
there are two time periods, a current time t and a future time T. For these
periods we define the following variables:
Table 4-1
aS tg > set
Fe (t,T) > aFg (t,T)
Fg(t,T»S/ +U
The first condition states that it is uneconomical to run the plant in the first
period based on the gas price. The second condition states that' based on the
forward price of gas and electricity, it will be economical to run the plant in
the second period. Now consider the case of a generator with no gas storage
ability. Its cash flows for the two periods are given in the table below.
t T
Plant idle in period 0 0
t
Buy a*q gas 0 -aqFg (t,T)
forwards, for physical
delivery
Sell q electricity 0 qF e(t,T)
forwards, for physical
delivery
Produce electricity 0 0
for delivery against
forward
Total cash flow 0 q(Fe (t,T) - aFg (t - T»)
Next consider the cash flow from the same generator with the option to
store gas.
45
t T
Plant idle in period 0 0
t
Buy a*q Btu of gas -aqStg 0
on the spot at t.
Store gas -aqU 0
Sell q electricity 0 qF e(t,T)
forwards, for physical
delivery
Produce electricity 0 0
for delivery against
forward
Total cash flow -aq(Stg +U) qpe (t,T)
Assuming a zero discount rate, we have arrived at the following total cash
flows:
Under the third condition of our price levels, the revenue with storage will
be higher.
It seems that the owner of the plant has been able to carry out a temporal
arbitrage in the electricity market using gas storage. However, if we take a
closer look, the strategy is really nothing more than cash and carry arbitrage
on the gas market. The condition which must hold in order for the strategy
above to be successful,
46
is exactly the same constraint which we previously stated could not exist in an
arbitrage free gas market. Furthermore, the increase in profit with the storage
strategy is exactly the same as the profit from a separate gas arbitrage deal of
the same magnitude, and the optimal production strategy for the electricity
generator will be the same regardless of whether there is fuel storage
capability or not. Assuming that the fuel and electricity spot markets evolve at
the same rate, the owner of the plant will always have the option of
purchasing additional fuel, or reselling unused fuel on the spot market. There
are temporal aspects to operation of these generators, related to minimum and
maximum run times and maximum ramp rates. These lead to the so-called
unit commitment problem in the production decision problem [7]. It is not
possible, however, to exploit or circumvent these constraints through fuel
storage.
Hydroelectric plants are different from gas or oil fired in that their fuel is
not a traded commodity. The water flowing into the reservoir does not have
an explicit cost, but using the water does represent an opportunity cost to the
operator since there is only a limited supply. The operation of a hydroelectric
plant is therefore naturally a temporal resource allocation problem under
uncertainty. Solving this problem requires advanced dynamic programming
techniques.
Some hydro plants are equipped with the ability to act as loads and pump
water back up into the reservoir. This setup is known as a pump storage
device. It gives the operator the ability to purchase power when the price is
low, and produce when the price is high. This process is considered by many
to be the equivalent of storing electricity. The inefficiency, or loss of power,
in the pump cycle is the equivalent of storage cost. While it is true that the
opportunity to capture the difference between high and low price swings adds
a temporal component to the electricity production, it is important to
47
differentiate between pump storage and the pure storage of electricity. To
illustrate this difference, consider the following example. Note that we use a
number of unrealistic simplifications. The purpose is to illustrate the
qualitative difference between pump storage and pure storage, rather than
provide quantitative results.
In scenario I, the operator has at his disposal a pumping device which can
replace water in the reservoir. The speed of the pump is such that it can
replace enough water in one hour to allow the generator to run at full speed
for one hour. Furthermore we assume the pump is lossless, that is it will
require lOMWh of electricity to run the pump for one hour, exactly equal to
the amount which can be generated by this quantity of water. The electricity
used in running the pump can be purchased from the spot market. The optimal
strategy for operating the plant, and the associated cash flows in each hour,
are illustrated in the table below.
Hour hI h2 h3 Total
Activity run pump run run
generator generator
Cash Flow -$100 $200 $300 $400
Reservoir 20p.u. 10p.u. Op.u.
level at end
of hour.
Hour hI h2 h3 Total
Activity charge no run
battery activity generator,
discharge
battery
Cash Flow -$100 $0 $600 $500
Reservoir lOp.u. 10p.u. Op.u
level at end of
hour.
Battery lOMWh 10MWh OMWh
level at end of
hour
We see that the owner of the plant is able to make more money with a pure
storage device than with pump storage. This is directly due to the capacity
constraint of the generator. No matter how much water is stored in the
reservoir, the turbine can only produce power at a rate of lOMW. The pure
storage device, however, is able to discharge in parallel with the turbine, thus
capturing a larger share of the high price hour. This effect is predominant in
electricity markets. Prices are extremely sensitive to the ratio of
instantaneous demand and the instantaneous total generation capacity.
Therefore, while hydro storage introduces a new inter-temporal component to
the electricity price process, it does not add additional capacity, and therefore
does not offer the pure arbitrage opportunities available in storable
commodities.
49
4.2 ARBITRAGE AND THE RELATIONSHIP BETWEEN
PHYSICAL AND FINANCIAL CONTRACTS FOR
ELECTRICITY
t T
buy physical 0 NqF(t,T)
N
sell financial 0
NqF(t,T)- LqSj
;=1
N
sell spot 0
LqSj
;=1
t T
sell physical 0 -NqF(t,T)
N
buy financial 0
LqS; -NqF(t,T)
;=1
N
sell spot 0
LqS;
;=1
Based on this constraint we will from now on use physical and financial
forward contracts interchangeably.
Chapter 5
Several times in this book we have touched on the question of how best to
deal with the dimensionality of price dynamics in electricity markets. This
includes defining the dynamic and stochastic relationship between different
time periods, different locations in the network, different markets (spot,
forward, derivatives), and different commodities (electricity, oil, gas). We
found that on the basis of arbitrage pricing theory, we cannot impose
constraints on the temporal relationship of prices. This is a direct result of the
non-storability of electricity. Similarly, it is not possible to create pure
arbitrage strategies between electricity and fuel markets, or between the
electricity prices at different locations in the network, except in degenerate
cases. From a strictly theoretical point of view, we could therefore conclude
that each spot, forward and derivative contract, for each delivery period, at
every location in the network, should be modeled as a separate state of a
dynamic process. This modeling approach, however, is highly impractical.
The dimension of the state space would be so high that we would be unable to
solve even the simplest optimization problems, not to speak of defining
consistent methods of estimating model parameters. The focus in this section
of the book therefore is on finding a reasonable compromise: a model which
preserves the unique characteristics of electricity production, consumption
and transmission, while limiting complexity. To achieve this goal, we have
identified a set of fundamental drivers. These are external processes, physical
as well as financial, which have significant impact on the dynamics of
electricity prices, and whose effects on the supply of, and demand for,
electricity are reasonably well understood.
55
Figure 5-1 illustrates the full blown flow of physical and financial signals
in the model. As we proceed to model the dynamics of the system, we identify
how the interplay between the components can be broken down based on the
time scale at which the interaction occurs. This allows us to derive simpler
versions of the models for use in specific applications. For example, a user
interested in developing day-ahead bidding strategies or short term price
hedging strategies need not be concerned with load growth and investment
dynamics. It is crucial, however, to start with the full scale version of the
model in order to understand the inter-temporal dynamics, and thus have a
solid basis for any separation or model reduction arguments.
56
Fast load
weather'" fluctuations
Fuel Price
Rain Fall
Unit Outages ~
Emissions
Rights
..
250
200
:2
3: 150
~
..
~ 100 .
J
-;;
'" 50
'~
::E
. ~
..... .. •• ... ... ............ ."
o
5000 10000 15000 20000 25000 30000 35000 40000 45000 50! lOO
-50
Total Output (MWll)
where a is a fixed parameter characterizing the slope of the bid curve, qk is the
market clearing quantity in hour k, and b k denotes the position (or shift) of the
curve. Next we add the constraint that demand bids are inelastic. The market
clearing quantity qk must then always be equal to the system load L k • We can
now write the market clearing price in terms of our two fundamental drivers,
load and supply:
The following section will outline the models used and the reason for
choosing that specific form. In later sections we present step-by-step
descriptions on how model parameters were calibrated based on historical
market data.
,
.. .. .. ~.
L d =Jlm+rd'
L L
)( 10 "
1-0
1.4
s:
E.- i,2
u
ro
o
-'
(LB
15
25
10
20
15
10
5
Month o 0
HOUf
Fig, 6-5: Average monthly patterns of daily load, /-lmL, New England
L d=J1m+wd
L L L
v m'
where
Figure 6-6 Reversion of the load weights wdL to the long-term mean bm \ year
1998
Figure 6-7 Load weights wl, and long-term mean bm\ New England
72
6.S STOCHASTIC SUPPLY PROCESS
Recall our underlying price model as a function of load and supply states
Lk and bk:
This implies that the aggregate supply bid curve is an exponential function
of fixed shape (given by a), which shifts over time.
Let us consider the input drivers, which could cause the supply curve to
shift:
I. Fuel price: An increase in fuel prices would force suppliers to increase
their bids into the spot market in order to remain profitable. An increase in the
fuel price would therefore be accompanied by a positive shift in bk.
2. Unit Outages and Scheduled Maintenance: The withdrawal of a
generation unit from the market, whether through an unexpected failure or a
scheduled maintenance, causes a significant shift in the supply bid function.
The size and duration of such a shift, as well as the frequency of their
occurrence, is technology dependent.
3. Gaming and Strategic Bidding: It has been shown that generators with
significant market share may increase their profits by unexpectedly removing
part of their generation assets from the market, forcing up price and increasing
the payoff for the remaining units [36]. Such an event can be characterized by
a positive shift in bk •
4. Unit Commitment Decisions: While generators are often modeled as
having well behaved quadratic cost functions, in reality there are significant
non-standard costs and constraints associated with starting up and shutting
down a generator. Translating such constraints into bids will cause generators,
even though they may have no market power, to deviate from marginal cost-
bidding schemes.
73
We now attempt to translate the impact of these drivers into a stochastic
process for the supply process. As with the load we characterize supply by a
[24xl] daily vector b d containing hourly supply levels. This daily vector is
then decomposed into its deterministic and random components:
b d =J1 bm +rdb
Although less pronounced than the load, the supply process does exhibit
seasonality over multiple time scales. The most pronounced are monthly and
intra-day seasonality.
1. On a monthly time scale, we see the scheduling of maintenance. In a
practice that has carried over from the regulated industry, units are
regularly scheduled for maintenance during the off-peak seasons (mainly
fall and spring), when demand spikes are unlikely. From the modeling
perspective this creates a repeating twelve-month pattern of supply bid
shifts.
The fuel markets feeding the generators also experience seasonality on
this time scale, mainly due to seasonal demand for oil and gas. Seasonal
fuel prices therefore create a second pattern of supply shifts. The aggregate
effect of these repeating yearly patterns is captured by the deterministic
shifts in the monthly parameter )..tm.
Regions with significant amounts of hydro generation may exhibit a
different type of seasonality, since the level of water in the reservoirs is
linked to precipitation patterns, and the melting of snow caps.
2. The second time scale in which seasonality is observed in the supply
process is intra-day, where we observe repeating 24-hour patterns of
supply curve shifts. This type of seasonality is mainly attributed to unit
commitment decisions made by the dispatchers. The operator of the unit
will estimate a day ahead of time the hours during which it will be
profitable to run the unit, based on the startup/shutdown constraint of the
74
generator. Once this decision is made he may choose not to submit bids
for the remaining hours, so as not to risk being scheduled and incurring a
substantial startup cost. The result is a repeated pattern of hourly shifts in
the aggregate supply bid curve. This behavior is captured by the daily
shape of the vector )..lmb.
The process defining the evolution of the weights is similar to that used for
the load process:
where
.$ ............................
-.~ .. -.-~-~-~-~-~-
So far our supply model has included smooth changes in the behavior of
the supply bid curve, which can be characterized by an Ito process. However,
there exists a set of high impact, low probability events that cannot be
approximated through random walk type models. One such event is the
unexpected failure of a major generator in the market. There are a number of
unknowns associated with this event:
200 r ------ -.
• •
I
• • ....
JIll"""
0 6 12 18 24
hours
Fig. 6-9 Daily shape \jI for a 400 MW base load plant and a 200 MW
peaking plant
2. The impact of the outage on the market clearing price will depend on
the capacity of the unit and its characteristic operating schedule. An
outage in a plant which is scheduled to deliver at full capacity results
in a positive shift in b d , equal to the capacity of the plant. If, however,
the plant was not scheduled to deliver (i.e. bid in above market
clearing price) then there is no effect on the price. The probability of
a plant being selected to produce in a given hour generally depends
on its cost structure, and therefore on its technology. We incorporate
this effect by assigning a [24xl] vector \jJmi to each technology i. The
vector denotes the capacity of the unit as well as the likelihood of the
unit being scheduled in a given hour. Figure 6-9 denotes the daily
shape of \jI for two types of generation technologies, a 400 MW base
load plant and a 200 MW peaking plant.
3. The outage duration is modeled as a deterministic minimum outage
time plus a stochastic Bernoulli component. By combining this
78
process with the random arrival time of the outage, we can
characterize the process for the state 1tid as a Markov chain, as
illustrated in Figure 6-10. Here the numbers next to the arrows
designate the probability of a state transition for a given day. The
probability of going from normal operation to an outage for each day
is given by "'out. The probability of the unit returning on-line after the
minimum outage period is given by "'in' For the case shown the
minimum outage time is four days.
1 1 1
Figure 6-10: Modeling of outage duration for the state 1tid as a Markov
chain
L L L L L L
e d +! - e d = - a e d +amz d
8~+! - 8~ = x: L+ aU z;t5 ,
where,
L L;;:L
e d =Wd -U d ·
Supply Model:
b b b ~ i i
b d = Jim + Wd V m + £,. 1rdlf/m •
where,
b
ed = Wdb - ;;:b
Ud .
and ltd is Markov process with parameters leOU! and lein as described in the
previous section.
•
market clearing prices (Sh) market clearing prices (Sh)
t
Time series of
supply curve shifts (b h)
i ~,
• •
seasonal mean shape seasonal mean shape
+
Time series of
+
Time series of
daily weights (WL) daily weights (wb)
~
Estimation of state space Estimation of state space
parameters, supply parameters, load
The parameters describing the load state dynamics can estimated directly
from the time history of the market demand. The supply state bh, on the other
hand, represents the hourly position of the aggregate supply bid curve, and as
such is not directly observable. To overcome this problem we use the
historical load and price data to calculate the implied time history of the
supply state, given by the equation
After the time series of historical supply states has been generated, the rest
of the calibration process is virtually identical for the supply and demand
components of the model. The next step is to transform the hourly time
history of the load and supply states into a history of daily weights. First,
however, we need to extract the deterministic seasonal components from the
data. For each hour of the day, the mean load and supply level are calculated,
creating the [1 *24] vectors ~L and ~b. The process is carried out separately
for each calendar month, thus creating a [12*24] mean value surface for the
load and supply processes, as shown in figures 6-12 and 6-13.
1)3
1.6
1/
D8
25
iO
20
5 '10
Figure 6-12 Average monthly patterns of daily load, ~mL for, New England
83
Daily shift of supply curve, MU b
.' ..
--
2.5
., "
"
.'
"';
"
~ 2
£:.
v.>
W .,
c.
:;)
'-' 1,5
.. ..
'"
"£,i
Q.
:;
W
{5 ,
¢ ,
.':! '"
-- ..
V)
0,5 .'
15 '--.
'.
.. , . 25
--
'
10 "
20
1&
5 10
G
Q 0
MonU1 Hour
Figure 6-13. Monthly mean vector for the supply state; Ilmb, New England
Month 1 2 3 4 5 6
Var (%) 92.78 95.12 94.60 93.37 94.13 96.33
Month 7 8 9 10 11 12 Avg.
Var (%) 96.75 96.46 93.91 94.27 93.06 93.36 94.51
Tab. 6-2: Variance of load explained by the first PC for different months
C.35
~. 0.25
~
a;o 0.2
d
0.1
45
25
20
15
5 10
5
Month G (1
-.:...
""0'
E 0.5
~
"
'-' 0
'"
(i
Q
"
'""'0 ·05
<lJ
.':!
(f) ~:..-~.
-1
15
25
10 20
15
5 10
5
MenU) o 0
Hour
To calibrate the supply side of the BSM using the New England supply
data, the problem was that we only had 14 months of hourly b available. To
use a full-size PC Analysis, the number of instances in data (in our case
workdays in a month) should be at least equal to the number of original
variables, in our case the number of hours analyzed. Since on average there
are only about 22 workdays in a month, we would require at least two
instances for each month, raising the required number of months to 24.
To extend the available amount of data, three approaches were
investigated.
1. Duplicating the missing months to obtain 24 months worth of data. Since
the data was the result of two distinct stochastic processes, this would
significantly alter the data beyond usability, introducing a deterministic
pattern.
Treating the entire year as composed of 12 equal months, thus introducing
a single set of j principal components. Since the PC analysis is used to
86
model deviation from the monthly daily load pattern 11m, this approach
would have adverse effects on the amount of information retained by the
model.
First two Principal Components (24) of Deviations from Daily Supply Shape, NUb
1~------;-------:-------;-------~======~
-~ ~
------- -1- --------- -----------i- ---------- ~IL-'_:!'_'_._:_g_~---,I
1\
0.8
t I I I
I \ :I :I :I :I
+I \'\
I
0.6 ------7--r~----------,-----------r----------r----------
I
,I ,I ,I
\ I I I I
c 0.4 \' I I
-----~----\~----------'-----------r----------r----------
I
I, , , ,
til
Q) I
----1- -/~. -~ ---------~ -----------!- ----------~ ----------
\1 I I I
E
0.2
E
0
J::: +---i /,,:~, , ,
.G/.\,.~:::;;;::!~-~~~r~~~.
c
0 0
~
'5
Q)
0 -0.2
!
\: I _______ ~ ___________
\ I
-0.4 - - - - - - - - - -\ I -/- - I
L __________
I
~I ______ _
1\ I /
\: .I I I I
Figure 6-16 First two pes of deviation from 11mb for 24 hours for
December
Tab. 6-3 Variance of supply (in %) explained by the first four PCs for
different months
Month 1 2 3 4 5 6 7 8 9 10 11 12 Avg
PC
1 50.12 36.08 64.98 63.50 49.18 45.68 64.29 51.57 52.10 86.92 59.47 53.99 52.22
2 19.58 19.90 21.37 11.96 42.25 28.22 14.71 24.98 13.21 8.00 12.13 23.16 18.57
3 12.30 18.66 4.30 9.32 4.09 11.44 9.17 13.41 1l.75 l.74 10.65 9.56 9.18
4 6.99 8.72 3.31 8.34 1.32 4.74 5.64 3.41 7.21 1.31 5.04 4.07 4.93
Here, the choice of only one PC is less obvious than in the load data. An
average amount of variance explained by each of 12 original variables is 8.33
%, so according to the guidelines (Appendix A) we should have in some
months considered using two or even three PCs. For the rest of this chapter,
however, we will assume that a first order approximation, using a single
principal component vector, gives sufficient accuracy.
The Bid-Based Stochastic Model can be expressed in the state space as:
The problem of joint estimation of system and noise parameters has been
solved in the literature for simpler problems; see [8],[24],[25] and [26].
However, there are significant differences between our problem and others.
Some of the approaches were using a simpler two-factor model, or they
assumed risk-neutrality, which does not hold true in electricity markets. In the
context of financial markets, the parameters of the spot process are often
88
calibrated using a time series of historical forward prices; see [8]. The
forward curve provides a richer set of data, since it indicates the market
expectation of future spot price levels for a series of maturity dates (at every
time step we observe a price curve rather than a single spot price). The
additional data provided by the forward markets simplifies the estimation
process and eliminates the stability problem encountered when estimating
from historical spot and load data (see discussion below).
The standard estimation techniques usually assume known covanance
matrices of the stochastic processes in the model, i.e. process noise covariance
matrix Q and measurement noise covariance matrix R. Alternatively, other
techniques for estimation of noise covariances require a complete knowledge
of the other system parameters. Since among the unknown BSM parameters
there were also the stochastic process variances cr and cro, the elements of the
matrices Q and R, the standard estimation techniques failed to converge.
The parameters had to be estimated separately in several consecutive steps,
in which the parameters were estimated independently. Since the load and the
supply processes are described in a similar way in the model, their parameters
a\ KL, cr L, cr LO and exb, K b, crb, cr bo can be estimated separately and in the same
way.
Estimation of the Bid-Based Stochastic Model parameters can therefore be
summarized in three successive phases.
1. Long-term drift of the mean K is estimated using a linear least
squares fit. After K is determined, the data is de-trended, i.e. the long-
term drift is eliminated.
2. Calculation of mean reversion factor: Factor ex, determining the
mean reversion speed of the weight process, can be estimated using
linear regression over de-trended data.
3. Estimation of process volatilities: Using the estimated ex, the
remaining parameters of the model in state space from cr and crO can
be estimated using the adaptive Kalman Filter and the technique for
identification of the variance-covariance matrices of the process and
measurement noise Q and R [30].
89
6.7.4.1 Calculation of the mean reversion factor
After assuming the initial values of (<>0 = 0) and the linear trend already
eliminated from the data in the previous step (K = 0), the following sequence
of equations unfolds.
There is, however, a significant problem with this approach. The variance
associated with each incremental new observation grows linearly with the
total number of observations, and is therefore unbounded. This violates a
basic premise of the linear regression technique, which requires that the
90
variance of the estimation error be bounded. We illustrate problem by
performing a series of regression estimates on the a parameter of the load and
supply process. For each consecutive estimation run we add an additional
observation, using the result form the previous run as the initial guess on the
parameter. A seen in figures 6-17 and 6-18, this process becomes unstable.
An interesting result was that with a limited number of samples, under
1,500 hours (about two months), the estimation technique gave reasonably
stable results. The estimates for this sample region are presented in table 6-4.
This property seems to indicate that there is a natural separation in the
timescales at which the reverting and non reverting states evolve. In the
following section we will exploit this property by proposing a time scale
separated version of the Bid-based model.
I
R
M
i
.. J..
I .... J..
I
. .... .,..
I
..'......... ····f
,
I
I
.~ .i __ .l~,~~~ __ _
1()<JD
- - Starting data
I _ Ending data
I I I
0.95 - - - - - - I - - - - - - -1- - - - - - - T - - - - - - -I - - - - - - - ,- - - -
I I
!
0.9 - - - - - - 'I - - - - - - -1- - - - - - - i - - - - - - ""1 - - - - - - - r- - - -
I I I
- - - - - - -+ - - - - - - -1- - - - - - - +- - - - - - - ---j - - - - - - - l-
0::
I I I
!
_______ ..J ________ 1_ _ _ _ _ _ _ L _ _ _ _ _ _ _ I _ _ _ _ _ _ _ _ _ _ _
\ I
0.75
I _II___ /__ ~~ _,.-'~l~~+~~-_-_~ ___: ____ _
, I 1\,,' ~/ I I I
R f\ I I
M 0.7 - - - - -1':-0 "c~I- - - - -1- - - - - - -1- --
i 0.65
I
I
- - - - - - I - - - - - - -1- - - - - - - T - - - - - -
I I I
-I - - - - - - -
I
I
ill
-l t- -
I 06 - - - - - -
I
I
"i -
I
- - - - -
I
I
-1- - - - - - -
I
I
T - - - - - - -I - - - -
I I
I
- - -
I
r- - -1- ,
I
I
I
I I
I
I
I I I I I I I I
0.55 - - - - - - --+ - - - - - - -1- - - - - - - +- - - - - - - -1- - - - - - - I-- - 1- r- ---1 - -
I I I I{ I
I, I
0.5 0L ---~50-----C-'OO~---''':'50:--------:C200~-----C2'':'50:--L..J-lL--:c'300
Number of observations considered
where
and
The benefit of the time-scale separated model is that it makes the 'mean'
process directly observable from the historical data. If the separation is
genuine, then we can accurately approximate the monthly 8 of supply or
demand as the mean of the weights during this month,
2. For every year and for every month within the current year, a mean of
Wd, 8 m , was calculated. D is the total number of days in the data, while
The mean of the stochastic process is zero and is not affected by the
process volatility measure, crm • Coefficient ex could therefore be
determined using linear regression to satisfy the least-squares
criterion [27].
95
y'=a+bx
f(Xd -~~d - y)
b = ..::d_=l'---::,---_ _ __
f(Xd -~)
d=!
X=[Xd]=(Wd+!-W d)
Y = [y d]= (()m - W d)
a=b
5. The difference between the estimated W'd+l and Wd+l was the
contribution of the stochastic component of the process, (jmZd. It was
therefore possible to calculate the monthly volatility measure (jm of
the process by subtracting the estimated values of W'd+l from the
actual values Wd+l and calculating standard deviation of the parts of
the time series vectors, belonging to a particular month:
- I (Xi--IXj J
2
1 DID
StDev(x) =
D-I i=! D j=l
was calculated:
aL KL (JmOL
a ab Kb (JOb
TBSM BSM
0.0204 0.0318
75 -0.0537
685 0.2411
2023 0.413 t
a 1. 13484e-4
98
6.9 SIMULATIONS
The dynamics of the load and supply states are described by a set of
equations of the form:
Based on historical load and spot price data, parameters for both the
supply and demand processes are estimated.
• The long-term mean on the other hand develops much faster, and its
volatility should therefore be much smaller, divided by a square root
of the time constant factor. Assuming that there are about 25 working
days in a month, the daily volatility should be about 5-times smaller,
a'
DVM
= °TSVM
JT' T == 25 days
The monthly average spot price in the actual data to which the model was
calibrated is shown in figure 6-19 There are relatively big differences among
certain months, describing a year with unusually high summer prices. The
summer of 1999 was very hot and the prices were higher than the historical
levels. Since only a limited amount (14 months) of price data was available,
the influence of a single month in calibration of supply process was stronger
than in load process calibration, where almost 20 years of data were available
and the influence of excessive months were less prominent.
101
45
40 - - -
35
30
25
20
15 - - ..
January February March April May June July August September October November December
Month
In Figure 6-20, the seasonal evolution of the average simulated daily spot
prices are shown. The price as generated by the model exhibits similar
properties as the actual average monthly price in Figure 6-19. The main
difference could be observed during the summer months, where the influence
of load process dampens the excessive shift in supply curve toward higher
prices, as dictated by supply process.
102
45 ··············f·····
40
35
20
30 60 90 120 150 180 210 240 270 300 330 360 390
Days
The stochastic properties of the model on the other hand can be illustrated
without the interference of monthly mean values by examining the daily
weight processes w L and wb • They are driven by four stochastic processes -
ZkL, ZkLO , Zkb and Zkbll - and governed by the daily process parameters a, K, cr m
and cro. The interplay of the short-term w process variances, cr m Land cr mb , and
the long-term 8 process variances cr L and crb in the model is shown in figure 6-
21.
The short-term variances of the mean reverting process, which are
bounded, dominate in the short run. As time progresses, however, the long
term variances will gradually become the dominant source of uncertainty.
103
Similar conclusions can be drawn from Figure 6-22, where the evolution
of the mean value of w L and its volatility boundaries are shown. The standard
deviation of the process is not uniform over the months, what is the
consequence of interplay between two volatility measures, cr mL and aU>. At the
same time, the volatilities of electricity price differ from one month to
another. During the periods of peak load, prices tend to be much more volatile
than in spring or fall, which is reflected in the model output.
The mean grows steadily according to the long-term growth parameter 0.
The weights were simulated for a two-year period with a 10.000 simulation
runs.
The mean of the supply process weight, w b, and its volatility boundaries
are shown in figure 6-22.The mean slowly drifts downwards, and the monthly
shapes in standard deviation are more pronounced than in load process in
Figure 6-23.
2.5 ~-----,--------,---------.-------,.-----~
1.5
0.5
o ~
-0.5
-1
-1.5
-2
-05
Figure 6-22: Daily weights wrnL : mean value and standard deviation
105
-2
-3
-4
Using the BSM it is possible to generate hourly spot price ST and its
volatility. Because the intra-day dynamics that can be found both in hourly
development of load and hourly clearing of market in supply, it is important
to have the model that is able to capture the hourly price dynamics. On the
other hand, it is sometimes also necessary to neglect the hourly dynamics and
deal with daily prices, as it is the case in certain applications such as forward
contracts.
In the Stochastic Model the price evolves as a sequence of daily, 24-hour
vectors of prices.
106
The daily price would therefore be calculated as a daily average of the vector
Sd.
_ 1 24
Sd =-LSdh
24 h=!
Days
"'"
::;:
E
o
<=
Days
With the deregulation of the electric utilities came a significant increase in the
financial risk to the companies, known as load serving entities (LSE) or
energy service providers (ESP), serving the end users. The risk faced by the
energy service provider can be traced back to the physical and economic
interactions between the ESP and its customers. The physical configuration of
the distribution network does not generally allow for any differentiation in the
service provided to different customers, in terms of power quality or
reliability. Economically, the ESP is limited by its contracts to serve retail
customers, typically known as standard offer contracts, which are structured
to shield the customers from any fluctuation from the wholesale price of
electricity. The financial risk from the wholesale market must therefore be
absorbed by the ESP. In this chapter we break down the sources of this risk,
and propose a methodology for the ESP to hedge its customer portfolio using
a dynamic futures trading strategy.
In the deregulated environment, the idea that the customer has a 'right' to
electricity does not correspond to the objectives of the energy service
providers. Energy service providers are profit driven entities, with an
obligation to their share holders to extract the maximum profit possible from
the provision of electricity to retail customers. In some instances, this may
entail not serving a customer if the cost of providing the service exceeds the
willingness of the customer to pay. This is where the historical role of
utilities catches up with the deregulation process. Since customers were
traditionally considered to have a right to service, and furthermore all
customers had the right to the same service (at least at the retail level), the
distribution infrastructure was not built to accommodate customer segregation
of any type. The two major deficiencies in the current distribution system are
the metering and interruptability of customers.
7.4 MODELING
Optimal
sd
Price • Cash Flow Risk Strategy
•
Model • Model Preference
i
PI,T ~
Ld
Demand Id
Model
Figure 7-1
The variables and parameters of the model are defined in Table 7-1.
118
The time period during which the ESP has committed to serving its
customers is broken down into N delivery periods, each spanning one month.
This is done in order to accommodate the structure of the futures market.
Recall that each traded futures contract requires delivery for one month. We
will furthermore make the simplification of using a single load and price
variable per day. In reality there are 24 hourly spot prices in a given day. At
the end of the chapter we discuss how the modeling approach can be
extended, using principal component theory, to account for intra-day
variations
We begin by modeling the cash flow for the LSE before any purchases in
the forward market, which we call the unhedged cash flow ('I'u). The total
'I'u is given by
119
L
N M(m+l)
tpV = LId (R -Sd)
m=!d=Mm+!
For simplicity we will here consider the case where the hedging period is a
single month. The cash flow function then becomes:
LId (R -Sd)
T+M
tpV =
d=T
The index t represents the time at which the forward contracts are
purchased or sold. In contrast to the spot market, which clears at discrete daily
intervals, the forward market trades in real time. For computational purposes,
however, we will restrict trading to discrete intervals tj. Each tj represents a
hedging interval. Changing the number of futures contracts held from one
interval to the next is known as rebalancing the portfolio, or rolling over the
hedge. Since a forward contract Ft,T cannot be purchased after the starting
date of the delivery month (T), we include the constraint tj:ST. The timeline of
the hedging process is shown in Figure 7-2.
T+M
I I I I I I I I I I
hedging period deli very period
Figure 7-2: Time line for dynamic hedging problem
120
Note that the forward contracts have no cash flow prior to the delivery period
(they are not pre-paid), so that all cash flow occurs during the delivery period.
\f is therefore equal to the net profit of the ESP with respect to the load
serving contract.
The load level, as well as the spot and forward prices at any future date,
are random variables in the cash flow formulation. To characterize the
distribution of these variables we apply the bid-based stochastic mode
developed in the previous chapter. Since future values of the forward price
occur as explicit terms in the cash flow, we must extend the price model to
incorporate the dynamics of the forward price.
The model for the spot price is identical to the one developed in Chapter 6,
except that we are applying a daily rather than an hourly model, and are
ignoring the unit outage component.
Load Model:
where
Supply Model:
where
We model the forward price as the expectation of the average spot price
over the delivery period of the contracts. In addition we allow for the
presence of a seasonal risk premium A. The risk premium can be positive or
negative depending on the market, and furthermore is allowed to be seasonal
in the maturity date of the contract.
122
We model the risk premium of the forward risk premium so that there it
grows linearly with the log of the average spot price. This makes sense since
the log of the spot price corresponds to the sum of the load and supply states.
Recall from our discussion of the properties of the bid-based model that the
long term component of the load and supply states (corresponding to the ()
processes) has a variance which grows linearly in time. In essence the model
proposes that the market assign a premium to the forward contract in
proportion to the variance of the underlying asset. While this makes intuitive
sense, there is not yet sufficient data available to test this assumption
properly. A more general model may assign a stochastic process to the risk
premium.
The model can be characterized III state space format, where the state
vector is given by
a vector of disturbances
Z liL,b
m
Zl
d '
J
The state dynamics are linear, while the output variables are a nonlinear
function of the states. We can write the model in compact form
So far we have defined the cash flow function and the dynamic constraints
on the underlying stochastic variables. Now we must define the firm's risk
preference, in order to arrive at an objective function for the optimization
problem. The mean-variance formulation defines a firm's utility in terms of a
linear tradeoff between the expected value and variance of the payoff:
The firm needs to develop a decision rule, or policy, for the quantity of
forward contracts to purchase at each hedging interval (qtj), in order to
maximize its objective function:
or in expanded form,
Next, the mean variance formulation is superimposed upon the new value
function. VtO represents the initial expected return of the un hedged portfolio.
The objective of the hedging strategy is to maximize the expected value of the
126
portfolio while minimizing the risk (or variance) of the return. We define the
mean variance objective function as,
The new objective function differs from the old in that it does not optimize
over the variance of the actual cash flow. Instead we attempt to maximize the
risk adjusted value of the expected cash flow up to the end of the hedging
period.
With the reformulated cost function, the problem would seem to fit into a
dynamic programming framework. This however is not the case. The penalty
on the variance of the value makes the problem impossible to solve through
backward iteration. Specifically, it can be shown that the Bellman Equation is
violated [38], thus invalidating the dynamic programming approach. There
127
have been results, however, illustrating that variance-penalized Markov
decision problems can be solved through nonlinear programming. Methods
presented in [37],[38],[39] and [42] present such techniques for a variety of
interpretations of variability. The discussion is usually in terms of infinite
horizon problems, but can be generalized to the finite horizon case.
Under the objective function defined above, the firm optimally manages
the change in the value function Vt up to the final stage in the hedging period
(t=T). In reality, however, the firm is concerned with the uncertainty of the
cash flow occurring over the delivery interval [T, T +N]. Assume that the firm
has executed an optimal policy over the hedging interval. At time T, the firm
then holds a portfolio of forward contracts defined by q=[qtO qtl ... qT-d.
Conditional on the q vector, the ESP can then solve a single stage decision
problem with respect to the original optimization problem to find the optimum
number of forward contracts to buy or sell in the last stage:
The end state problem has several features which make it easier to solve
than the original problem.
2.5~
I
'.5-~
I
If·
0.5 ~
In determining the proper structure of the hedging period, the user has to
consider two questions. How far in advance does one need to start
hedging the cash flow at delivery, and how often must the portfolio be
rebalanced? The first question is limited by the forward market, which
currently only trades fifteen months into the future. The second part is
constrained by transaction costs, as well as by the computational
complexity of the DP problem. Given a fixed number of times the user
can rebalance the portfolio over a long hedging period, it may useful to
adopt a nonlinear function for the length of each rebalancing period. For a
mean reverting process, the further away in time the user is projecting the
133
uncertainty, the less new information will enter over a given time interval.
Consequently, the further away from the delivery period, the less likely it
is that the portfolio has deviated from its optimal value. Further research
is needed to find an optimal function for the length of the hedging interval
as a function of time to delivery.
Chapter 8
8.1 INTRODUCTION
The price model used is a simplified version of the bid-based price model
introduced in [19]. We define a daily [24* 1] price vector P; , whose elements
are the 24 hourly electricity prices. Next we define the log of the price vector
137
to be the sum of a deterministic and stochastic component. The deterministic
component is composed of a monthly vector )lm' which captures the seasonal
characteristics of the electricity spot price. The stochastic component is
modeled as the product of the principal component vector v m and a daily
stochastic scalar weight Wd. The principal component captures the shape of
daily price variations from the seasonal mean )lm while the weight describes
the magnitude of the deviation as well as its correlation over time. The log of
the vector of spot market prices can be written as
Next we model the process describing the evolution of the weights w~.
where,
The form of the price process postulates that hourly spot prices will be log-
normally distributed. Furthermore prices inside of a day are perfectly
correlated, since they are a function of a single daily random variable Wd. This
reduction in complexity is made possible by choosing the principal
component in an intelligent manner.
24-1
CFd (xo) = max E(L P;,kq -C k (q)}
1[ k=O
23
CFd (xo) = Jo(x o) = maxE{L Uk (P;kqk -C k(qk ,P1,k)-
1[ k=O
Reward-to-go in hour k:
_ {max(1,x k + 1)
X k +1 -
min(-1,xk -1)
ICTRUE') = 1
139
I('FALSE') = 0
The optimal policy is applied to obtain the maximum expected profits. The
above problem is a full-blown version in which there are multiple sources of
uncertainties. The first one is from electricity spot prices, and the second one
is from fuel prices.
This effectively states that in the very short term (day-ahead) we can
ignore the mean reversion as well as the long term volatility. It should be
noted that we only use this assumption to arrive at a bidding strategy. When
simulating future spot price for valuation purposes we use the full-blown
version of the price model.
Therefore,
and,
We have shown how to calculate cash flow and a onloff policy for a
generator in a given day d. This cash flow is an expected cash flow, given
Wd = Wd,i' Wd,i is a sample value of Wd ' which is continuously normally
distributed with (w d-J + Ke ) mean and standard deviation a:. To create a
lookup table mapping a pair (w d,j' Wd-J) to a cash flow, we generate Wd,j
given Wd-J,i ; and apply the optimal policy n(w d-J,i) to determine the cash
flow in the d period with Wd,j , or P;,j . Therefore, in period d, we have
23
CFd(w d-Li' Wd ) = ~) Uk (w d-J)(P;,k (w d)' q - c(q) - I(x k < O)S)
k=O
+(l-Uk(Wd_JJ),(c f +I(xk <O)T)},\tj
Repeating this process using Wd-J,i for other i, one obtains a cash flow
matrix which an element (i,j) is a cash flow associated with both Wd,j and
141
W d-J,i' This matrix captures possible cash flows in a given month m with a
simplified price process for each day d.
CFm =
For each month m, a cash flow matrix can be calculated by using the same
method. Note that each cash flow matrix is obtained by assuming a constant
fuel price P;,k for all k and d within each month m.
Note that we do not need to apply the principal component approach since
the price is a scalar. The stochastic component is described by a two factor
mean reverting model.
To apply this model we first need to expand the lookup table to include gas
price as a third dimension. Note that the assumption that unit commitment
takes the day-ahead gas price as deterministic allows us to add only one rather
than two dimensions to the lookup table.
We now generate simulated paths for future w's for electricity and gas.
The lookup table converts them into paths of future cash flows.
Once the lookup table has been created, we can use the full-blown price
model to generate simulated weights. The lookup table is then used to
generate a path of cash flows from a path of weights. The simulation time is
linear in the length of the valuation period. Furthermore we are not restricted
to the proposed model for generating weights. The lookup table can be linked
to any stochastic model which produces weights for the principal components.
D
V = E{I(r)d . CFd(w~)},
d~O
[::J
144
=L
D
Vi (r) d • CF~ (w ~ , w ~ )
d=O
V di = CFi(
d W ej
d'-J ' W ei)
i+1 + (1 -
d' + r ( p. CFd+l ) CFi-l)
p. d+l ,J, = {'I'
1- , 1+ I}
145
d d+l d+2
For a case with more than one sources of uncertainty, two additional
branches will be added to capture each additional source of uncertainty. This
will make the problem become more complex since nodes grow exponentially
with time. A Monte Carlo approach might be more applicable to deal with
more than one source of uncertainty.
8.6 FIGURES
A Principal Com ponent Representation of a Typical Daily
Price of Electricity
45
40
~5
30
""
:;:: 25
~ 20
1S
1 {)
25
20
:='" 15
~ 10
o
:;::
H 0 u r
Figure 8-3. A Typical 24 Hour Basis for Representing the Daily Price
Vector for Electricity (Hour 1)
2 6
2 0
1 :.:,
.."'
£:
::;
j 0
H O'IJ f
Figure 8-4 A Typical 24 Hour Basis for Representing the Daily Price
Vector for Electricity (Hour 23).
148
Average Spot Price and Spot Prices of Electricity
50
c::::J S pot P ric e s
45 1
-+-Average Price
40
35
30
.
'"::;~ 25
20
15
10
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
H 0 Ur
9.1 INTRODUCTION
In this chapter we will address the question of how market participants can
quantify and hedge locational price risk. The work draws on results from the
finance, economics and engineering community, attempting to find a middle
ground that allows us to solve the unique problems facing the electricity
industry. This includes developing methods for valuing newly emerging
transmission dependent derivative contracts. Furthermore we examine the
relationship of these new contracts with existing forward and option contracts
on locational spot prices. We extend this analysis to include the valuation of
investment opportunities in transmission assets, thus allowing a for-profit
transmission provider to arrive at a market based valuation of a potential
investment, based on observed forward and derivative prices.
PTR, Ownership of
Physical
transmission line
Financial Locational spread option FTR
In markets for transmission, the line between physical and financial rights
IS necessarily blurred. To better understand why this is true, consider the
following example, where our world consists of two electricity markets (or
two zones of the same market) connected by a transmission line. 1 supplier in
market 1 has entered into a contract to supply physical power to a consumer
in market 2. To enable this transaction, the supplier has also purchased a
physical transmission right from 1 to 2. The transmission right, however, may
be curtailed by the system operator under certain circumstances, such as a
physical failure of the transmission line. Would such a curtailment represent a
physical or financial risk to the parties in the bilateral contract?
Market 1 Market 2
Inter-tie
A curtailment of the transmission right does not absolve the supplier from
his obligation to serve the customer, but it makes it impossible for him to
transmit his locally generated power to the load. This problem can be
circumvented if the supplier sells his power in market 1 and purchases an
equivalent amount of power in market 2. The physical risk of curtailment has
then been transformed into a financial risk from the price spread between the
two markets. There is also an inherent assumption that there is power
available to be bought in market 2 at any price. This may not be the case,
especially if the region is heavily dependent on imports coming through the
downed transmission line. If power is not available, the load in the bilateral
154
contract cannot be served and the transmission right curtailment represents a
physical risk.
In this paper we will attempt to value three different types of assets: fixed
transmission rights, flexible transmission rights and the ownership of a
transmission line. Each of these assets can be thought of as a derivative of the
locational spot price at the end nodes. Consider the setup described in
figureO. We proceed to calculate the value of each asset at maturity, i.e. at the
actual time of use of the transmission asset, financially or physically. We find
that in each case, this value is a function of the underlying spread between the
spot prices Sl and S2. In each case it is assumed that the transmission right
held is in the direction 2 to I, and the quantity is q MW.
The value of the fixed transmission right is simply the difference in the
locational spot values at the time of maturity, and can thus be positive or
negative:
155
The owner of the flexible transmission right will only exercise the contract
if the price differential is positive. The payoff is therefore given by
The analysis above focused only on the value of the transmission rights at
maturity. In order to use these contracts as a part of hedging or speculation
portfolios, however, the investor needs to be able to project the future values
of the contract, and understand the dynamics of contract prices over time. To
address this issue we need to postulate stochastic models for the underlying
spot prices. We begin by reviewing the current state of the art modeling
techniques.
156
Value Fixed Transmission Right
$/MW
Figure 9-2
157
9.4 OVERVIEW OF EXISTING PRICE MODELS
The set of models used in pricing derivatives and managing financial risk
are commonly referred to as volatility models. The purpose of these models
are twofold: to characterize the probability distribution of future spot prices,
and to estimate the correlation between future prices at different points in
time. The most well-known application of volatility modes is the Black-
Sholes option valuation formula, originally derived to value derivatives on
equity. The basis for Black-Sholes is the assumption that the price of a stock,
S, can be characterized by a random walk process known as Geometric
Brownian Motion (GBM),
dS = J1Sdt + aSdz ,
This approach was taken by Deng, Johnson and Sogomonian [44] in order
to price locational spread options in electricity markets.
The net effect on price of the two actions is equivalent. Without loss of
generality we decide to interpret all flows as the effect of load cross-bidding.
To incorporate this behavior into the model, we introduce a new variable qd i,
representing the actual quantity bid into market i at time d. The variable Ldi is
interpreted as the native load of the market, which is physically located inside
the market's borders. Price in market i is a function of the total load and
supply bid into this market:
Sdi = ea'q'+b'
'
.=
I
I .. 2
q~ = L~ + FJ2
qJ=L~-FJ2
1. The prices equalize, thus removing any incentive for further cross-bidding.
2. The transmission line becomes congested, preventing the native loads from
being supplied from the other market beyond a certain level.
s~ = s3
a I qdI + bld = a 2 qd2 + b2d
The flow necessary to reach price equality ftJ2, as a function of native load
and supply states, is given by
, 12 1 rr 2L2 b2) {I I b I \1
Fd = I 2 l\a d + d - \a Ld + d}J·
a +a
The actual flow between the markets Pd 12 , accounting for the limits, can
therefore be written as
161
The prIces in two markets, Sd 1 and si, are always equal, until the
transmission flow reaches the maximum capacity. At this point prices will
diverge, and the dynamics of the two markets will decouple.
a K a
Load 0.3 50 500 2000
0.00015
Supply 0.3 0.05 0.0005 0.25
162
Tab. 4: Monthly BSM parameters, used in simulations
J.lm
Market Market
1 2
Load 13000 20000
Supply 1.7 1.7
We simulated the behavior of the two market model for various values of the
maximum transmission capacity, varying from 0 to 3500. For each scenario,
the model was then run 10,000 times for a 31-day period. The plots show the
correlation between locational prices and the value of the flexible
transmission right on the 31 st day.
163
Daily Pl and P2, lL =0 Daily P 1 and P2, lL = 1750
140
120 ' I I
II ,"- , I~ Mar1<et 1
Mar1<et2
I 100
90 I
' I'
\ I I f
r\
\
1\
11'\
I I I
1',
/ 1'1
I,
I I, \ 1'\1
I \1 \ I I' \'
100
I
I ,
I I , 80
V
I
I'
) )
~
I 1/
,I I
W II W I \
II
~ 80
~ 70 v
l 60
l
40~
20
0 10 20 30 40 10 20 30 40
Day Day
) )
W W
~ ~
! !
10 20 30 40 10 20 30 40
Day Day
Figure 9-3: Simulated joint spot price dynamics for two markets joined by
a transmission line of varying capacity.
164
Histogram of Spread option value, TL = 0 MW Histogram of Spread option value, TL = 1750 MW
10oo,-------------, 1000,-------------,
800 800
VI
1;l '"
1;l
~ 600 ~ 600
i;
-g 400 :u 400
z"
~
z
200 200
o
40 60 80 100 120 20 40 60 80
Value ($/MWh) Value ($/MWh)
Histogram of Spread option value, TL = 3500 MW Histogram of Spread option value, TL = 5250 MW
3000,-------------, 5 0 0 0 , - - - - - - -___- - - - - ,
2500 4000
'"~ 2000 VI
1;l
fJ 3000
~ 1500 '0
1l 1i 2000
~ 1000 ~
z z
1000
~L5---~1~0----5r--~0L--~-~10
10 20 30
Value ($/MWh) Value ($/MWh)
Figure 9-4
\\
.... ..
\
'~
"
Figure 9-5 Spread option value e l2 and price correlation coefficient in
various TL capacities
165
As the size of the transmission line increases, the correlation between the
locational prices increases. Consequently the probability of the prices
diverging is reduced, so that the value of the spread option, and the
transmission right, decreases.
A simulation based approach allows the user to estimate the expected cash
flow from a contract, and the associated risk, or variance, of the cash flow. In
addition, traders also need to understand the relationship between the contract
price and the current value of the underlying state variables. This knowledge
allows the user to cancel out risk between a variety of contracts, as long as the
risk is derived from a limited number of underlying sources. A well-known
example is the case of delta hedging. As shown in section 9-4, the Black
Scholes model assumes that all uncertainty affecting the price of a stock can
be modeled as originating from a single Wiener process. By Ito's lemma, all
derivatives of the stock price will follow Ito processes driven by the same
Wiener process as the stock price. This result allows the trader to hold a
combination of the derivative and underlying stock so that the uncertainty
cancels itself out exactly. The ratio of the stock (S) to the derivative (g)
required to eliminate the uncertainty is known as the delta of the derivative,
given by
L1 = agl
g as.·.1
The delta hedge will only cancel out the uncertainty for a given stock price
S. Since the stock price evolves continually, the value of delta will also
constantly change. To perfectly eliminate risk, a trader therefore needs to
rebalance his portfolio constantly. The process of rebalancing in response to
changes in the underlying asset price is known as dynamic hedging. The
effectiveness of a dynamic hedging strategy depends in large part on the
166
accuracy of the underlying model. Furthermore, due to transaction costs and
other real market constraints, a continuous replication strategy cannot be
implemented. Instead, traders must rely on approximate strategies where the
portfolio is rebalanced at discrete time steps. In this case, one must analyze
the robustness of the linearization, as spot prices diverge from the initial
operating point.
50
ot-----~ ........~----~
8 20000 30000 40900
~ ~ i
.100 ..--.-.-.-.---.-..-.------.--.-.----.----.-----... -.~.-J
Market 1 Native Load
140·· 0.00025
120··
. 100 0.0002
(\I
(Ii
.,.. 80
if) 60 0.00015
'tI 40
!!!
!! 20
a. 0.0001
(Ii 0
$
g
'0::
c.. 0.00005
-60
-80 I 11111111111 II III 0
0 8000 16000 24000
Market 1 Native load
Figure 9-7: Price spread between two markets as a function of native load
in market 1, with superimposed load distributions.
Figure 9-7 shows two normal distributions superimposed on the plot of the
sensitivity of price spread to changes in L]. The two distributions, high load
and low load, represent the projected distributions of load on a high demand
and low demand respectively. The means of the distributions are 13,000 MW
169
and 23,000MW, and the standard deviation in both cases is 2,OOOMW. The
figure illustrates an interesting condition. Each load distribution covers two
regions of the graph, but the probability of ending up in the third region is
nearly zero. In other words, a load change during a high demand period could
cause the transmission line to move from uncongested to congested in the
1~2 direction. However, it is extremely unlikely that a load change would be
large enough to cause congestion in the 2~ 1 direction. Conversely, during a
low demand period, the transmission line will be either uncongested or
congested in the 2~ 1 direction, but is unlikely to be congested in the 1~2
direction. Combining this result with the link between the direction of
congestion and the sign of the price spread gives an interesting result outlined
in the table below.
If one attempt to find a set of parameter for the traditional spot price
model, which produces a similar set of probabilities for the price spread, one
would find this impossible. The bid based dictates a qualitatively different
behavior of the price spread.
The nature of the price spread lends itself nicely to this problem. Consider
three states of the transmission line.
1. State 1. The line is congested in the direction 1-72. In this case S2>S I.
2. State 2. The line is uncongested. In this case S I=S2.
3. State 3. The line is congested in the direction 2-71. In this case SI>S2.
Now consider the payoff at maturity from the call option, as well as the
two replicating portfolios, under each of the three transmission line states,
(assuming the forward contracts are prepaid).
171
State 1 State 2 State 3
Call option Max(O, Sl_ Max(O, Sl_ Max(O, Sl_
S2)=O S2)=O S2)= SI_S2>O
Portfolio 1 SI-S2<O SI_S2=O SI-S2>O
Portfolio 2 0 0 0
The table shows that the call option has the same cash flow as portfolio 1
in states 2 and 3. The option has the same cash flow as portfolio 2 in states 1
and 2. Combining this with the probability distributions displayed in Figure
9-7, we find that in many cases a very simple replication strategy will do. For
the high demand period, the transmission line is very likely to be in state 2 or
3, so portfolio 1 provides a good replication. For the low demand period, the
transmission line is likely to be in state 1 or 2, and therefore the payoff from
the spread option is zero and the replicating portfolio is empty.
qdi = T'd"
L +"
3
£... Fdij ,
j=i
}• -r-
-1-'1, 1• = 1.. 3
while the power balance equation postulates that the amount of native load
equals the amount of bid-in load.
;=1 ;=1
The algorithm for calculation of native generation and prices in each of the
three markets for a 31-day period is shown in Figure 9-9. Using the respective
daily loads, the DC OPF is used to compute native generation Gi , i = 1.. 3,
according to the cost function J and transmission constraints Fmaxij • From
native generation Gi , load Li and supply curve shift bi , market i price pi is then
calculated. This algorithm is ran in a loop where the transmission capacity
Fmaxij of all three lines is gradually increased in 6 steps from 0 MW to 3,750
MW. The plots in show loads and prices for the period of 30 days.
As the transmission capacity is gradually increased the flows on the lines
increase as well, as shown in Figure 9-10, yet the effects of congestion are
eminent on the first four graphs. At the same time, correlation in prices among
markets increases as the lines become less congested, see Figure 9-10.
174
k=k+l
DCOPF
=I. (a i qi + b i )
3
- Compute cost function J
i=1
Figure 9-9
175
Tl= F",
3000 4000
~3OO0
~2oo0 ::;:
~
pOOO
It 1000
0 0
0 10 20 3Q 40 0 10
Day
10.1 INTRODUCTION
where ~mL captures the seasonal behavior of the load. The state OmL
represents the long term uncertainty in load, which grows stochastically with
drift K and volatility cr.
179
10.3 MODELING INVESTMENT DYNAMICS
where
Figure 10-1
Table 10-1
182
Sp 300~--------------------'----------~
ot. 250
PrJ
ce 200 - n o delay
($/ 150 ---'~d--'A--~V--4__1~+_tH . _. delay
M 100
VV 50+-----------------~~~~-----1
h)
21 41 61 81
time in months
Figure 10-2
estimator 1-----,
Figure 10-3
The challenge in the estimation problem lies in the fact that it requires the
user to model the decision process of all other investors. The problem may be
tractable in the case where there is sufficient historical data available to
estimate the cumulative investment rate in response to market price (the G
parameter in our model). However in the early stages of a market, such as the
current situation in the United States, one would be forced to arrive at this
parameter by deriving likely competitor strategies. This would be a very
complex game theoretic problem, where the outcome would depend on how
sophisticated market participants are in their decision process.
When capital investment fails to keep up with load growth, there are two
measurable effects in the market. The first is an increase in the spot price, as
discussed in the previous section. The second effect is a reduction in the
186
available generation reserve R, defined as the amount of unused generation
available in the market as a fraction of the total load,
To illustrate the link between reliability and spot price dynamics, the
model is further amended. Starting with a total capacity equal to the initial
load, plus a reserve margin X,
In periods of high price levels, consumer advocates can put pressure on the
government to impose price caps on the market. The argument is that
suppliers are taking advantage of the generation shortage in order to drive up
prices, either by withholding their generation or bidding it in at inflated price
levels. The issue of 'fair' pricing of electricity will not be addressed here.
Instead, we will try to answer the question of whether price caps are an
effective means of reducing price levels in the long term. To do this the
market is simulated under two conditions. The first is without a price cap, as
shown above. In the second case, a price cap is introduced, leading to the
condition
Sk = min(cap,eGL,-b,)
From the simulation it is clear that while the cap eliminates periods of high
prices, it also raises price levels during the low price cycles. This result is
easy to understand if one goes back and examines the signal which drives new
investment:
S;"-I.
By reducing price levels when supply is scarce, the regulator reduces the
rate of new investment into generation. As a result, prices drop off at a slower
188
rate, causing higher future spot pnces. In the case described, the average
power price is higher in the case where price caps are imposed.
300
250
::c
~
200
.,
"'-" 150
~
8. 100
'"
50
0
21 41 61 81
Time (months)
Figure 10-4
Reserve Margin
0.15
c
.~ 0.1
~ - - n o cap
~ 0.05
~ - .. - - with cap
~
[/J
~ 0
~
-0.05
21 41 61 81
Time (months)
Figure 10-5
189
G(SCnliCal -I) = aK .
If the price cap is set below SCritical, then investment cannot keep up with
load growth, and the system is invariably headed towards blackouts.
The implications of the results in this paper must not be interpreted as
rejecting all forms of regulatory intervention in general, and price caps in
particular. There may be instances where it is necessary for the government to
190
set temporary limits to the price in a market to prohibit suppliers from
exploiting shortages. What the model illustrates is that the regulator must be
very careful in setting these limits. Price caps must be set higher rather than
lower, to ensure that the economic feedback is not blocked, and that market
forces are allowed to bring the system back to stable price levels. Once price
caps have been put in place at a too low level, they become increasingly
difficult to remove as the generation shortage worsens.
This chapter addresses the interplay between spot price levels and
investment into new generation capacity in competitive electricity markets.
The problem was addressed from the viewpoint of economic efficiency as
well as the physical reliability of the system. Special emphasis was placed on
the dynamic properties of the investment process. It was shown that delays
caused by backward looking investment, as well as by the licensing and
construction time of the asset, lead to periods of over and under investment.
This in turn leads to a cyclical long term price behavior, driven by a stochastic
growth in demand, which does not settle to an equilibrium level. The structure
of the problem indicates that the presence of liquid forward markets could
reduce the information delay, and help stabilize the system. This assumes
however that forward markets contain information which is not reflected in
historical spot prices, or that is otherwise part of the public knowledge.
Further research into the effect of forward markets on information flow could
involve simulations of bottom up, agent-based models, to determine the extent
to which locally held information is reflected in the forward price. While it
may not be possible to accurately calibrate such models to the market, they
would provide important qualitative insights into optimal decision rules for
investors, as well as intelligent market designs for the deregulated electricity
industry.
In the final part of the paper, the dynamics of spot price and new
investment were linked to the physical reliability of the system. Periods of
191
under-investment not only lead to higher price levels, but also reduce the
reserve margin of available generation, and can lead to generation deficiency
and blackouts. The first reaction of regulators to periods of high prices is
often to try to force price levels back down through the use of price caps.
Price caps, however, inhibit the economic feedback which would allow the
market to readjust itself. Imposing the caps reduces the rate of new
investment, leading to a slower recovery from the price hike. If the regulator
continues to force the issue by reducing the cap levels, the lack of new
investment will eventually lead to an erosion of the reserve margin, leading to
load curtailments and blackouts in the system. The results presented in this
paper indicate that regulators have to be cautious in the use of price caps.
They must respect the unique characteristics of electricity as a commodity:
non-storability, inelasticity of demand, and a highly constrained transmission
system. These characteristics lead to an uncommonly strong link between
market price signals and physical stability. Any attempt to block the true
economic signals from the market could therefore prove disastrous.
193
Chapter 11
Conclusion
The models introduced in this book are mainly a tool for communicating
the advantage of an alternative set of state variables in the price modeling
process for non-storable commodities. Due to the limited amount of available
data, rigorous testing against a general set of model formulations was not
possible. As the markets mature, they should provide more information
regarding optimal choices of model structure. In addition, the emergence of
new technologies for load management on a retail level is likely to generate
more elasticity in the overall market demand for electricity. As a result,
further development of the model is required to cope with alternate shapes of
the supply and demand curves.
Though not part of the original scope of this book, we found that the
modeling presented lent some insight into the impact of market structure and
government intervention on the physical and economic prosperity of the
system. As shown in the chapter on long term price dynamics, the futures
market serves a role, not only as a tool for hedgers and speculators, but as a
medium for the transfer of information. The analysis suggests that the
presence of a transparent futures market may be crucial for the physical and
financial stability of the system by preventing unnecessary delays in future
investment. A comparative study of the reliability levels in regions with and
without futures exchanges could provide some interesting insights into the
validity of this claim.
Appendix A
I I
n n
var(v i ) = var(Yi)
i=1
2. Repeat the process for the subsequent PCs, until the number of PCs =
rank(C).
3. Determine how many PCs are necessary to describe the process
adequately. Form the reduced-order principal component [j x n] matrix v*,
where only the first m PCs are retained. A detailed description of the routine
can be found in [22].
The eigenvalue ~, associated with the i-th PC corresponds to the
equivalent number of variables this PC represents. A PC with an eigenvalue
of ~ = 3.9 describes on average as much variance as 3.9 original variables.
Dividing the eigenvalue by the total number of PCs, j, we can obtain a total
percentage on variance explained by each pc.
When all n PC have been determined, it is necessary to determine j: how
many PCs are necessary to describe the data accurately enough. The three
most common measures are:
1. Retain all PCs that represent more variance than original variables on
average (its Ai < 1).
2. The Scree plot. The incremental plot of variance accounted for by every
PC is called the scree plot. The number of points before leveling-off of the
curve is the number of PCs retained.
3. Total variance of the data accounted for by the retained PCs. Some
authors propose to retain as many PCs as to account for about 90% of the
199
variance [23], while others propose less stringent criteria, depending on the
reasons for performing the peA [21].
Appendix B
where A is the system matrix, relating the system state XI at time t to the next
state xI+lat time t+ 1 in the absence of the controlling input. B is the matrix
relating the input 0 1 to the state XI' The measured noisy system output at time t
YI IS
The random variables '111 and £1 represent process and measurement noise.
They are assumed to be independent of each other and with normal
probability distributions.
202
'It z N(O,Q)
ct z N(O,R)
To write the Bid-based model in state space form, the system state Xb the
process noise llb the input U t and output Yt signals, the system matrices A, B,
C and r, and the noise covariance matrices Q and R take the following
values.
Q~[l'O] R=l
0,1
Let's define xt+l/t as our a-priori estimate of the state vector at step HI,
and xt+I/t+! our a-posteriori estimate of the state vector at HI, given
measurement Zt+!. We can then define a-priori and a-posteriori estimate errors
et+ lit and etft as
et+l/t = x, - Xt+I/1
e,l, = x, - X'I'
Pt+l/t = E~t+l/le;+l/t J
Ptll = Ele,/le~t J
203
The Kalman Filter algorithm computes the optimal a-posteriori estimate
Xt+l/t+l as a linear combination of the a-priori estimate xt+l/t and a weighted
difference between an actual measurement Zt and predicted measurement
C xt+l/ t • When they agree completely, the residual Yr+l/ t is zero.
The factor K in the equation is called the Kalman Gain and is chosen in
such a way as to minimize the a-posteriori covariance Pllt.
11.2 KF ALGORITHM
Kalman gain,
204
... ~
Measurement Ul!date {"Correct"}
r
Time Ul!date {"Predict"} l. Compute the Kalman Gain
l. Project the state ahead Kt+l = Pt+l/tC Tt+1~Ct+1Pt+l/tCt+l
T
I
Fig. B.l Kalman Filter operation flowchart.
After constructing the model representation in the state space and setting
up the KF procedure, we construct a vector of unknown parameters e that
contains the unknown parameters of the model.
o= [a K (Y (Yo y
Using the covariance of the innovation process Nt+1/t> obtained by the
Kalman Filter,
Nr+l/r = CPr+l/r CT
we can construct a log likelihood function J.
J = log L = -~
2 r~l
I fy:l/tN~~l/t
Yr+l/t + log (det (N r+l/r ))]
Kalman Filter
- obtain J
Unconstrained optimization
- obtain M, aJ/ae
[47] Ilie, M., Zaborszky, J., Dynamics and Control of Large Electric
Power Systems, Wiley & Sons, 2000, ISBN 0-471-29858-1.
211
[48] Gubina F., Grgie D., Banie I., "A Method for Determining the
Generators' Share in a Consumer Load ", to appear in IEEE T-
PWRS in 2001.
[49] Joskow, P.L., Tirole, J., "Transmission Rights and Market Power on
Electric Power Networks", Rand Journal of Economics, Vol. 31,
No.3, 2000, p. 450-487.
[50] California ISO Corp., "FERC Electric Tariff", First Rep. Vol., No.
I, October 13,2000.
[51] Hogan, W., "Contract Networks for Electric Power transmission:
Technical Reference", Harvard University, September 1990.
[52] Ilie, M., Hyman, L., Allen, E., Younes, Z., "Transmission Scarcity:
Who Pays?", The Electricity Journal, July 1997, p. 38-49.
[53] Oren, S., Spiller, P., Varaiya, P., Wu, F., "Nodal Prices and
Transmission Rights: a Critical Appraisal", University of
California, Berkeley, December 1994.
[54] Yoon, Y.T., Arce, J.R., Collison, K.K., IIie, M., "Implementation of
Cluster-based Congestion Management Systems", ICPSOP 2000,
"Restructuring the Power Industry for the Year 2000 and Beyond",
2000.
[55] Yu, C-N., IIic, M.D., "Congestion Cluster-Based Markets for
Transmission Management", IEEE Winter Meeting, January 1999,
pp 821-832.
[56] Ott, A., "Fixed Transmission Right Auction", IEEE 1999
[57] PJM, "FTR Auction Training", April 1999,
http://www.pjm.comitraining/training index.html
[58] Chicago Mercantile Exchange, Futures and Options Listing
Schedule, http://www.cme.com/clearingllistings/index.html
[59] Baker, M. P., Mayfield, E. S., Parsons, J. E.: "Alternative Models
of Uncertain Commodity Prices for Use with Modem Asset Pricing
Methods", The Energy Journal, Vol. 19, No. 1.
[60] Mount, T.: "Market Power and Price Volatility in Restructured
Markets for Electricity", Proceedings of the Hawaii International
Conference on System Sciences, January 1999, Maui, Hawaii.
212
[61] Backerman, S. R., Denton, M. J., Rassenti, S. J., Smith, V. L.,
"Market Power in a Deregulated Electrical Industry: An
Experimental Study", Economic Science Laboratory, University of
Arizona, Tuscon, AR.
[62] Borenstein S., Bushnell J., Wolak, F., "Diagnosing Market Power in
California's Restructured Wholesale Electricity Market", August
2000.
[63] Moerch von der Fehr N.-H. , Harbord D., "Spot Market
Competition in the UK Electricity Industry", Economic Journal,
No. 103, May 1993, pp. 531-546. Blackwell Publishers,
Cambridge, MA.
[64] Mount, T.: "Market Power and Price Volatility in Restructured
Markets for Electricity", Proceedings of the Hawaii International
Conference on System Sciences, January 1999, Maui, Hawaii.
Index
agent-based modeling, vii, 58 generation assets, 31, 53, 72, 135,
arbitrage 136, 145, 150, 179, 186
cash and carry, 20, 24, 41, 43, Geometric Brownian Motion
45 (GBM),21
definition of, 18 hedging
battery, 48 dynamic, 114, 119, 165, 169
bid-based model, 80, 111, 113, static, 114
114, 122, 132, 158, 166, 168, industrial customers, 32
194, 195 investment
bidding strategy, 139, 146 dynamics of, 55, 180, 182, 184
Black Scholes model, 21 Kirchhoff's laws, 2, 5
commodities least squares estimation, 81, 88
non-storable, xi, 194 load serving entity (LSE), 113
storable, 3, 19,26,48,55 locational pricing, 5
congestion clusters, 150 lookup table, 136, 140, 142, 143,
congestion management, 33, 150 144, 146
cross bidding, 160 Monte Carlo techniques, 143, 144,
double auction, 61 145
dynamic programming, 46, 53, net present value (NPV), 7, 8
125, 126, 135 New York Mercantile Exchange
dynamic replication, 40, 169 (NYMEX),36
economic equilibrium models, vii, options, xi, 5, 8, 12, 32, 34, 39,
58 152,158, 168
economic feedback, 190, 191 over the counter (OTC), 34, 35,
end state problem, 129 195
energy service provider (ESP), 32, swing, 39
113, 115, 116, 117, 131, 152 physical risk, 153, 154
experimental models, vii, 59 price caps, 187, 188, 189, 191
forward markets pncmg
information content, 185 arbitrage pricing theory (APT),
fundamental modeling, vii, 59 xi, 7,17,19,54
futures markets, 183, 184 risk neutral, 22
214
principal component analysis supply bid curve, 72, 74, 81
(peA), 74,83, 111 temperature, xii, 4, 39, 55, 60,
production based models, vii 122, 193, 195
reliability, 6, 113, 116, 154, 177, term one, 29
179, 186, 190, 194, 195 time scale separation, 55, 70
dynamic notion of, x, 185 transmission assets
retail customers, 113, 115 valuation, 149
risk measures transmission rights,S, 39, 151,
mean-variance formulation, 16, 152, 153, 154, 155, 158, 161,
114, 124 162, 165, 170, 173
value at risk (VAR), 9, 10, 129, fixed, 151, 154
131 flexible, 151, 154, 155
risk neutral pricing, 22 unit commitment, 46, 73, 135,
risk preference, 8, 9, 10, 13, 14, 136, 137, 139, 141, 142, 145
17, 114, 117, 124 valuation, vi, ix, xi, xii, 1, 7, 12,
scheduled maintenance, 62 13, 16, 69, 135, 136, 139, 143,
spark spread, 39 144, 146, 149, 151, 157, 161
storage market based, 149
cost of, 3, 20, 46 of generation assets, 136
gas, 44, 45 weather, 2,4,24,39,61,68, 195
hydro-electric, vii, 46 Wiener process,S, 21, 27, 157,
pump storage facilities, 46, 48 165
strategies, vii, 43 zonal pricing, 150