Vous êtes sur la page 1sur 168

INTEGRATED MODELING OF ELECTRIC POWER SYSTEM

OPERATIONS AND ELECTRICITY MARKET RISKS WITH


APPLICATIONS

A Dissertation
Presented to
The Academic Faculty

by

Haibin Sun

In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy in the
School of Industrial and Systems Engineering

Georgia Institute of Technology


December, 2006
INTEGRATED MODELING OF ELECTRIC POWER SYSTEM
OPERATIONS AND ELECTRICITY MARKET RISKS WITH
APPLICATIONS

Approved by:

Dr. Shi-Jie Deng, Advisor Dr. Shabbir Ahmed


School of Industrial and Systems School of Industrial and Systems
Engineering Engineering
Georgia Institute of Technology Georgia Institute of Technology

Dr. A.P. Sakis Meliopoulos, Co-advisor Dr. Haizheng Li


School of Electrical and Computer School of Economics
Engineering Georgia Institute of Technology
Georgia Institute of Technology

Dr. Paul Griffin


School of Industrial and Systems
Engineering
Georgia Institute of Technology

Date Approved: October 3, 2006

ii
ACKNOWLEDGMENTS

I would like to express my gratitude to my advisors, Dr. Shi-Jie Deng and Dr. Sakis

Meliopoulos for their invaluable advice and support, especially for the inspiration they

provided me throughout my PhD study. I thank the rest of my thesis committee: Dr. Paul

Griffin, Dr. Shabbir Ahme, and Dr. Haizheng Li for their comments on my dissertation,

especially for Dr. Ahmes insightful comments on an early draft of this dissertation and

on my presentation.

I am grateful to Dr. Jim Dai and Dr. Houcai Shen for introducing me to this PhD

program. Its been a great opportunity to study and work at Georgia Tech together with

our wonderful professors and fellow students, while everyone pursues his dreams. I

would like to thank my friends in School of Industrial and Systems Engineering, School

of Electrical and Computer Engineering, and Quantitative and Computational Finance

program at Georgia Tech for their friendship. Thanks also go to Aram, Adrian, and Larry

at Constellation Energy, Mark, Yuan, and Edward at GE Energy, Gerald at Adsystech,

and Sharon and Jim at SunTrust, for their support and encouragement during my

internships. I enjoyed those wonderful summers with these great colleagues and fun

friends.

I am also grateful to Power System Engineering Research Center and National

Science Foundation for the financial support throughout my PhD study. The

opportunities to meet and learn from the enthusiastic researchers from academia and

industry helped me identify research problems and shape my career goals.

Finally but above all, I thank my wife, my parents, my sister, my brother-in-law, and

my nephew for their love, encouragement, and their faith in me all along.

iii
TABLE OF CONTENTS

ACKNOWLEDGEMENTS --------------------------------------------------------------------- III

LIST OF TABLES ------------------------------------------------------------------------------- VII

LIST OF FIGURES ----------------------------------------------------------------------------- VIII

SUMMARY----------------------------------------------------------------------------------------- X

CHAPTER 1 OVERVIEW-------------------------------------------------------------------------1

1.1 Electricity Power System And Market Risks ----------------------------------------2


1.2 Contributions Of The Research Work ------------------------------------------------5

CHAPTER 2 MARKET SIGNALS FOR TRANSMISSION ADEQUACY


INVESTMENT --------------------------------------------------------------------------------------7

2.1 Introduction ---------------------------------------------------------------------------- 10


2.1.1 Transmission Inadequacy---------------------------------------------------------- 11
2.1.2 Transmission Pricing--------------------------------------------------------------- 13
2.1.3 Transmission Investment ---------------------------------------------------------- 17
2.2 Market Dispatch ----------------------------------------------------------------------- 22
2.2.1 Linearization of Market Dispatching Problems -------------------------------- 26
2.2.2 Incorporation of Reliability Constraints ----------------------------------------- 30
2.3 Electricity Market Simulation-------------------------------------------------------- 35
2.3.1 Comparison of DC and AC Models ---------------------------------------------- 38
2.3.2 Evaluation of Reliability Constraints -------------------------------------------- 43
2.4 Stochastic Electricity Price Models ------------------------------------------------- 47
2.4.1 Quantile based GARCH Model--------------------------------------------------- 48
2.4.2 Parameter Calibration -------------------------------------------------------------- 50
2.5 Conclusions and Discussions -------------------------------------------------------- 56

CHAPTER 3 HYDROELECTRIC GENERATION SCHEDULING IN MULTI-


MARKETS WITH MARKET UNCERTAINTIES------------------------------------------- 57

3.1 Introduction ---------------------------------------------------------------------------- 60

iv
3.2 Literature Review---------------------------------------------------------------------- 63
3.3 Deterministic Scheduling Model ---------------------------------------------------- 65
3.4 Stochastic Model with Ancillary Service Uncertainty---------------------------- 71
3.5 Stochastic Model with Market Price Uncertainty --------------------------------- 73
3.6 Stochastic Model with Both Uncertainties ----------------------------------------- 76
3.7 Case Study------------------------------------------------------------------------------ 78
3.8 Conclusions and Discussions -------------------------------------------------------- 85

CHAPTER 4 DAY-AHEAD FORWARD RISK PREMIUM IN ELECTRICITY


MARKETS WITH TRANSMISSION NETWORK CONSTRAINTS--------------------- 87

4.1 Introduction ---------------------------------------------------------------------------- 90


4.2 Decision Problems of The Market Participants------------------------------------ 97
4.2.1 System Operator -------------------------------------------------------------------- 98
4.2.2 Generators and Load Serving Entities ------------------------------------------105
4.3 Market Equilibrium Prices ----------------------------------------------------------113
4.3.1 Theoretical Formulations ---------------------------------------------------------113
4.3.2 Numerical Examples --------------------------------------------------------------116
4.4 Day-Ahead Forward Risk Premium------------------------------------------------120
4.4.1 Theoretical Formulation ----------------------------------------------------------120
4.4.2 Empirical Evidences --------------------------------------------------------------122
4.5 Conclusions and Discussions -------------------------------------------------------143

CHAPTER 5 CONCLUSIONS AND FUTURE RESEARCH-----------------------------145

5.1 Market Signals for Transmission Adequacy Investment ------------------------145


5.2 Hydroelectric Generation Scheduling----------------------------------------------146
5.3 Electricity Trading with Forward Hedging----------------------------------------147

APPENDIX A PROCEDURES OF SENSITIVITY COEFFICIENT CALCULATION


------------------------------------------------------------------------------------------------------149

APPENDIX B SOLVING FOR EQUILIBRIUM FORWARD AND SPOT PRICES --151

APPENDIX C SOLVING FOR FORWARD RISK PREMIUM --------------------------152

REFERENCES -----------------------------------------------------------------------------------153

v
LIST OF TABLES

Table 2. 1 Spatial volatility of LMPs calculated by AC and DC power flow based market
dispatch ($/MWh) --------------------------------------------------------------------------- 42

Table 2. 2 Averaged spatial volatility of LMPs over the sample year when imposing
different TRM requirements in the market dispatch ($/MWh) ------------------------ 46

Table 2. 3 Estimates of model parameters using LMPs at buses 10 and 18 determined by


market dispatch without TRM requirement ---------------------------------------------- 53

Table 2. 4 Estimates of model parameters using LMPs at buses 10 and 18 determined by


market dispatch with 10% TRM requirement-------------------------------------------- 54

Table 3. 1 Parameters of Price Distributions (in $/MWh)------------------------------------ 79

Table 3. 2 Estimates of Normal Distribution Parameters (in Million $) ------------------- 84

Table 4. 1 Parameters of demand distributions (in $/MWh) --------------------------------117

Table 4. 2 Scenario A Spot market prices (in $/MWh)------------------------------------118

Table 4. 3 Scenario A Day-ahead forward market prices (in $/MWh) ------------------118

Table 4. 4 Scenario A Day-ahead forward premium (in $/MWh) -----------------------118

Table 4. 5 Scenario B Spot market prices (in $/MWh) -----------------------------------118

Table 4. 6 Scenario B Day-ahead forward market prices (in $/MWh) -----------------119

Table 4. 7 Scenario B Day-ahead forward premium (in $/MWh)-----------------------119

Table 4. 8 Scenario C Spot market prices (in $/MWh) -----------------------------------119

Table 4. 9 Scenario C Day-ahead forward market prices (in $/MWh) -----------------119

Table 4. 10 Scenario C Day-ahead forward premium (in $/MWh) ---------------------119

Table 4. 11 Scenario D Spot market prices (in $/MWh) ---------------------------------119

Table 4. 12 Scenario D Day-ahead forward market prices (in $/MWh)----------------119

vi
Table 4. 13 Scenario D Day-ahead forward premium (in $/MWh) ---------------------120

Table 4. 14 Statistics of forward and spot LBMPs (in $/MWh) ---------------------------126

Table 4. 15 Statistics of forward and spot shadow prices of the flowgates (in $/MWh)129

Table 4. 16 Estimation of PTDFs--------------------------------------------------------------130

Table 4. 17 Statistics of regression ------------------------------------------------------------131

Table 4. 18 s (bk ) value for the estimation of PTDFs ---------------------------------------132

Table 4. 19 t-Value for the estimation of PTDFs --------------------------------------------133

Table 4. 20 Confident Interval of PTDFs estimation----------------------------------------134

Table 4. 21 Estimation of regression coefficients -------------------------------------------139

Table 4. 22 Statistics of regression coefficients ---------------------------------------------139

Table 4. 23 t value of regression coefficients ------------------------------------------------140

Table 4. 24 Estimates of forecast error term Parameters ($/MWh) -----------------------142

Table 4. 25 Estimates of forecast error term parameters by excluding price spikes


($/MWh) -------------------------------------------------------------------------------------142

vii
LIST OF FIGURES

Figure 2. 1 Statistics of level 2+ TLR from July 1997 to December 2005----------------- 12

Figure 2. 2 Deadweight loss due to transmission congestion -------------------------------- 19

Figure 2. 3 The IEEE RTS24 system----------------------------------------------------------- 36

Figure 2. 4 Comparison of LMPs at all buses determined by AC and DC power flow


based market dispatches when the system load is 40% of peak level ---------------- 39

Figure 2. 5 Comparison of LMPs at all buses determined by AC and DC power flow


based market dispatches when the system load is 90% of peak level ---------------- 40

Figure 2. 6 Comparison of LMPs at all buses determined by adjusted AC and DC power


flow based market dispatches when the system load is 90% of peak level ---------- 41

Figure 2. 7 System averaged LMP with respect to transmission capacity using market
dispatch without TRM requirement when the system load is 80% of peak level --- 43

Figure 2. 8 Generation and load bus LMPs with respect to transmission capacity using
market dispatch without TRM requirement when the system load is 80% of peak
level ------------------------------------------------------------------------------------------- 44

Figure 2. 9 Probability distributions of LMPs at bus 10 over the sample year with
different TRMs requirements -------------------------------------------------------------- 45

Figure 2. 10 Probability distributions of LMPs at bus 18 over the sample year with
different TRMs requirements -------------------------------------------------------------- 46

Figure 2. 11 Q-Q plot for fitted distribution of hourly log returns of LMP at bus 10 (left
panel) and 18 (right panel) determined by market dispatch without TRM
requirement----------------------------------------------------------------------------------- 54

Figure 2. 12 Q-Q plot for fitted distribution of hourly log returns of LMP at bus 10 (left
panel) and 18 (right panel) determined by market dispatch with 10% TRM
requirement----------------------------------------------------------------------------------- 55

Figure 3. 1 Scenario tree with both uncertainties---------------------------------------------- 77

Figure 3. 2 Sensitivity of expected revenues with respect to U and D ------------------ 81

viii
Figure 3. 3 Comparison of simulated revenues with regulation service call uncertainty 82

Figure 3. 4 Comparison of simulated revenues with market price uncertainty ----------- 82

Figure 3. 5 Comparison of simulated revenues with regulation service call and market
price uncertainties --------------------------------------------------------------------------- 83

Figure 3. 6 Comparison of revenues PDF of deterministic and stochastic solutions----- 83

Figure 4. 1 Hourly day-ahead forward and spot prices on the reference bus in NYEM-- 91

Figure 4. 2 Day-ahead forward and spot price differences between zone CENTRL and
the reference bus in NYEM ---------------------------------------------------------------- 95

Figure 4. 3 A 3-bus example system ----------------------------------------------------------116

Figure 4. 4 NYISO control area load zones --------------------------------------------------123

Figure 4. 5 Hourly shadow prices on FG1 in forward and spot markets -----------------129

Figure 4. 6 Forward premiums on the reference bus at each hour-------------------------135

Figure 4. 7 Load level in the CENTRL zone at each hour ---------------------------------136

Figure 4. 8 Forward premium of the shadow price on FG1 at each hour-----------------136

Figure 4. 9 Historical and forecasted spot price at the reference bus in August 2006 --141

Figure 4. 10 Historical day-ahead shadow prices of flowgate FG1 in August 2006 ----142

ix
SUMMARY

Because of the physical nature of electricity and the complexity of system operations,

many challenging problems have arisen from the restructuring of electric power industry.

Tremendous uncertainties put market participants in the midst of unprecedented risks

when making the planning, operating, and trading decisions. This situation has motivated

the reported research work on modeling, evaluating, and constructing strategies to hedge

against the market risks involved. Through integrated modeling of power system

operations and market risks, this thesis addresses a variety of important issues on market

signals modeling, generation capacity scheduling, and electricity forward trading.

The level of investment in electricity transmission networks has been lagging behind

those in the generation and distribution sectors amid the industry restructuring. It affects

economic efficiency and impairs system reliability. The first part of the thesis addresses

a central problem of transmission investment which is to model market signals for

transmission adequacy. The proposed system simulation framework, combined with the

stochastic price model, provides a powerful tool for capturing the characteristics of

market prices dynamics and evaluating transmission investment. Numerical experiments

with the IEEE RTS24 system yield interesting insights. In contrast with the common

practice of using DC power flow formulations for market dispatch, we advocate the use

of an AC power flow formulations instead since it allocates transmission losses correctly

and reveals the economic incentives of voltage requirements. By incorporating reliability

constraints in the market dispatch, the resulting market prices yield incentives for market

participants to invest in additional transmission capacity.

In electricity markets, generators seek to maximize their profits by simultaneously

participating in multiple markets simultaneously. The uncertainties in market prices and

x
different extents of market participations make generation capacity allocation a

challenging task. The second part of the thesis presents a co-optimization modeling

framework that incorporates market participation and market price uncertainties into the

capacity allocation decision-making problem through a stochastic programming

formulation. Optimal scenario-dependent generation scheduling strategies are obtained.

The advantages of the proposed model are illustrated through a computational study with

realistic data provided by a hydroelectric producer.

The third part of the thesis is devoted to analyzing the risk premium present in the

electricity day-ahead forward price over the real-time spot price. This study establishes a

quantitative model for incorporating transmission congestion into the analysis of

electricity day-ahead forward risk premium. Through simulations with a three-bus study-

system, it is illustrated that the more frequently transmission congestion happens, the

higher the forward prices get at the load buses. Evidences from empirical studies with

the New York electricity market data confirm the significant statistical relationship

between the day-ahead forward risk premium and the shadow price premiums on

transmission flowgates.

xi
CHAPTER 1

OVERVIEW

Since electricity is an essential resource in the national economy and our daily lives,

the electric power industry has always been under the close watch of the general public.

In order to prompt technology innovation and improve economic efficiency, deregulation

was undertaken in the previously regulated industries of telecommunication, natural gas,

and airlines. Following this trend, the electric power industry has been undergoing

restructuring from a regime operated with vertically integrated monopolies to one with

competitive markets starting in the early 1990s. Competition has been introduced with

structural changes involving generation, transmission and distribution sectors. The new

business environment consists of primary wholesale markets for bulk energy trading,

open transmission systems for physical electricity delivery, and ancillary service markets

for reliable system. Instead of following regulatory orders, market based mechanisms

have gradually been developed which create incentives for market participants to take

advantage of market opportunities. However, due to the unique physical natures of

electricity and the complexities of power system operations, many problems have

occurred and made the restructuring process very complex and challenging. Tremendous

market uncertainties, such as fluctuating demands, unforeseen facility outages, and

extremely volatile electricity market prices have occurred, and these uncertainties pose

unparalleled market risks to market participants. Through integrated modeling of power

system operation and market risks, the research presented in this thesis addresses several

issues important to market participants, including the topics of market signals modeling,

generation capacity scheduling, and electricity forward trading.

1
1.1 Electricity Power System and Market Risks

Electricity is produced at 10-25kV voltage levels from various fuel sources, such as

coal, oil, natural gas, hydroelectric sources, wind power, and nuclear plants. The

ownerships of generation units usually belong to independent power producers or electric

utilities with franchised service areas. Electricity is injected into transmission networks

which operate at 230-765kV voltages to allow economical bulk transportation over long

distances by reducing the conductor heat loss. Individual transmission lines can be taken

out of the network in case of contingencies or for maintenances. At the load centers,

electricity is transformed down to lower voltages (11.5-12kV for large industrial and

commercial customers and 120-240V for most residential customers) for distribution to

end consumers. Alternating current (AC) systems, as opposed to direct current (DC)

systems, are adopted predominantly in bulk power systems for the flexibility of voltage-

conversion.

For decades, the electric power industry had been viewed as a collection of natural

monopolies viable for vertical integration operating under government regulations.

However, the consumers pursuit of economic efficiency and the innovations in

generation technology combined with the observation of deregulation successes in other

traditionally monopolized industries such as telecommunication, natural gas, and airline,

triggered the restructuring of the electric power industry. Competition and the forming of

open market platforms occurred. This process is also described as deregulation and re-

regulation since some aspects of the industry remain regulated. Regardless of how it is

labeled, though, electric power systems have been unbundled to horizontally separated

segments of generation, transmission, and distribution to facilitate free competition.

Market-based mechanisms are expected to drive the system operations through price

signals and economic incentives.

2
The electricity spot pricing scheme, as initially proposed by Schweppe et al (1980,

1985, 1988), Caramanis et al (1982), and Bohn et al (1984), and later extended with the

introduction of financial transmission rights (FTRs) for transmission congestion hedging

as illustrated by Hogan (1992), provides an economic basis for market transaction

settlements and establishes a platform for revealing market signals for generation

resource allocation and transmission capacity investment. Spot pricing is a natural

extension of the classical market equilibrium theory. A central market agent (the power

system operator) collects bids from market participants and determines winning bids by

solving optimal power flow (OPF) problems. The market-clearing locational marginal

prices (LMPs) can be obtained through this process and used to settle the electricity

transactions. Accordingly, the LMP-based transmission charges fluctuate subject to

market conditions, especially when transmission capacity is tight. In order to provide

cost certainty for transmission customers, FTRs are structured to entitle (and obligate)

their holders (possibly negative) revenues associated with specific congestions, which

offset their congestion charges. FTRs are proposed to be awarded to merchant

transmission investors for their capital recovery.

However, the inherent physical nature of electricity determines the unique

characteristics of the evolving electricity market and imposes unprecedented complexities

on its management. First, the fact that electricity travels at close to the speed of light and

is not economically storable in bulk requires the instantaneous balancing of production,

consumption, and delivery. Deprived of the buffering benefit of holding inventory,

electricity markets are subject to temporary demand-supply imbalance which induces

volatile LMPs. Second, without the widespread use of expensive control devices over the

3
transmission network, electricity flows in accordance with Ohms law1 and Kirchoffs

laws2 instead of on any contracted path. As a consequence, it impedes the free movement

of electricity as envisioned by economic theory. The power transactions at various

locations in the system are therefore physically interdependent. Third, various

restrictions such as equipment capacity, voltage bounds, frequency requirement and

stability concerns should be adhered to for continuous production and delivery of

electricity. Consequently, marginal generation cost, congestion rent, reliability and

security cost altogether measure the market value of generation and demand, which

further diversifies the LMPs. Generally, the dynamics of LMPs are in line with the ever-

changing system conditions and present market opportunities accordingly. It provides the

basis for a market mechanism to guide the allocation of scarce generation resources and

addition of new capacities. The same argument leads to the research work of evaluating

financial transmission rights which reflect the transmission service conditions and

provide market incentives to induce capital investment in the transmission sector.

Similar to LMPs in the energy market and FTRs in the transmission market, the values of

operating reserve services should be discerned accordingly to guide transactions in the

ancillary service markets as well.

Due to the physical constraints of electricity listed above and the imbedded

uncertainties of a competitive market, market dynamics cannot be foreseen with certain.

Risks haunt market participants in their planning, operating, and trading decisions. Risk

management is often a high priority for participants in deregulated electricity markets due

to the substantial price and volume risks that the markets can exhibit. This situation has

1
Ohms law describes the relationship between the electric potential difference across an ideal conductor
and the current through it.
2
Kirchoffs law of voltage reveals the meaning of potential - the voltages around a closed path in a circuit
must sum to zero; Kirchoffs law of current is the current conservation in Ohms law - the sum of currents
entering a node must equal the sum of currents exiting the node.

4
motivated the research work to model, evaluate, and construct strategies to hedge against

the market risks.

1.2 Contributions of the Research Work

The research presented here has been conducted along three lines.

The first line focuses on market signals modeling for transmission adequacy

investment. The generation sector has attracted a significant amount of attention since

the inception of the electricity market restructuring. The divesture of generation assets

from utility companies induced the development of various generating facilities valuation

models. However, while it is generally agreed that a highly reliable transmission system

is a necessity for a workable power market, the important issue of valuing transmission

asset has not been adequately addressed. An AC power flow based market dispatch

model with reliability constraints is proposed in chapter two. It allocates transmission

losses correctly, reveals the economic incentives of voltage bound, and provides

incentives for transmission capacity addition to market participants. Combining this

system simulation framework with the stochastic model adopted in section 2.3, it

provides a powerful tool to capture the characteristics of market prices movements and

evaluate transmission investments.

The second line of the research is about generation capacity scheduling. The third

chapter looks at the operation of a hydroelectric generation unit with market uncertainties.

If a profit-pursuing hydroelectric producer simultaneously participates in the electricity

energy spot market and ancillary services markets as a price taker, the producer needs

need to balance the market opportunities on both sides with respect to limited water

resource over a pre-specified time horizon. This research presents a co-optimization

modeling framework that accommodates market participation and market price

uncertainties based on stochastic programming techniques; this method gives rise to an

5
optimal scenario-dependent scheduling strategy instead of a fixed trajectory as in the case

of deterministic optimization. The proposed methodology can also be straightforwardly

adapted to incorporate additional uncertainties for determining optimal capacity

scheduling strategy.

The third line of inquiry addresses the issue of electricity forward trading. Market

participants trading strategies in day-ahead forward and spot electricity wholesale

markets are modeled, and the market equilibriums are addressed in chapter 4. Optimal

decision problems of generation producers as the electricity supplier and load serving

entities as the electricity buyer on forward and spot electricity wholesale markets are

modeled. Their respective optimal positions and the market equilibrium prices are

revealed by solving the respective risk-adjusted profit maximization decision problems

with respect to the market equilibrium conditions and transmission network constraints.

The study establishes a quantitative model for incorporating transmission congestion into

the analysis of electricity day-ahead forward price premium. The impact of transmission

congestions on the forward risk premium is illustrated through simulations with a three-

bus study-system. Evidences from empirical studies with the New York electricity

market data confirm the implication of the model.

The thesis concludes in chapter five and provides directions for future research.

6
CHAPTER 2

MARKET SIGNALS FOR TRANSMISSION ADEQUACY

INVESTMENT

The on-going power industry restructuring process with the subsequent evolution of

the market environment and regulatory rules have drastically affected the system

planning process. Since the restructuring started with the introducing competition into

the generation sector, active research has been undertaken to understand the impact of

competitive markets on the generation sector. However, the lingering regulation over the

transmission sector leaves important issues about how well market mechanisms can work

with transmission networks still unclear. In light of the unprecedented generation build-

up, statistics show a clear and increasing lag between transmission construction and

generation development in recent years. The market environment that encourages

generation owners to respond to opportunities by transferring larger quantities of power

more frequently and over longer distances further threaten transmission capacity. The

inconsistency and the mounting problems this causes call into question the adequacy of

transmission and create a need for investment incentives and effective cost recovery

mechanisms to alleviate the tight transmission capacity conditions. By identifying

incentives and obstacles to transmission investment from various perspectives, this

chapter addresses the central problems of modeling market signals to allow credits for

economic efficiency improvement, system reliability enhancement, and financial risk

mitigation. The proposed system simulation framework, combined with the stochastic

7
model, provides a powerful tool for capturing the dynamic market signals and evaluating

transmission investments.

The following notations apply to this chapter:

Acronyms

LMP Locational marginal price

{AC, DC} {Alternating, Direct} current

FERC Federal energy regulatory commission

FTR Financial transmission right

ISO Independent system operator

MW Mega watts, unit for real power, the rate of energy consumption

MWh Mega watts hour, unit for energy consumption

PJM Pennsylvania, New Jersey, and Maryland regional electricity market

PTDF Power transfer distribution factor

RTO Regional transmission organization

Indices and sets

i Index of suppliers in the system, i = 1, L , I .

j Index of demands in the system, j = 1,L, J .

k Index of postulated contingency event for market dispatch, k = 1, L , K .

l Index of transmission flowgates monitored for market dispatch, l = 1,L, L .

n Index of buses in the system, n = 1, L N .

Market dispatch models

C The total generation procurement cost in the system.

F Eq () Equality constraint set for market dispatch.

F Ineq () Inequality constraint set for market dispatch.

8
Pmn The real power flow over transmission line m n .

Qmn The reactive power flow over transmission line m n .

{ S
, T } Reserve margins of {generation, transmission} capacity in percentage of their

respective normal capacity.

Parameters of market participants

Ci () Generation procurement cost of supplier i .

d A vector of demand quantity dispatched, d = d j .[ ]


D Peak The year peak load of the system.

s A vector of supply quantity dispatched, s = [si ] .

p A vector of locational marginal prices, p = [ pn ] .

Parameters and functions of the transmission network

g() A set of power flow functions of system state and control variables.

g Q () A set of quadratic power flow functions of system state and control variables.

h() A set of branch loading functions of system state and control variables.

{Gmn , Bmn } Conductance and suseptance of transmission line m n .


{Rmn , X mn } Resistance and reactance of transmission line m n .
Vector of transmission capacity of flowgates, T = [Tl ] .
T
T

{x, u} The vectors of system {state, control} variables in normal scenario.

{x ( ) , u ( ) } The vectors of system {state, control} variables in contingency k


k k
scenario.

xQ The vector of the state variables for in quadratic power flow formulations.

v The vector of bus voltage magnitudes, v = [vn ] .

{v , v }
l u
[ ] [ ]
The {lower, upper} bounds of bus voltage magnitudes, v l = vnl , v u = vnu .

9
The vector of phase angles of bus voltages, = [ n ] .

n The set of buses with direct transmission connected to bus n .

2.1 Introduction

In the pre-deregulation era of the electric power industry, economies of scale and

scope, as well as the desire to avoid duplication of the infrastructure, led to the formation

of regulated monopolies. The industry adopted cost of service regulation (COSR) to

achieve a pre-approved rate of return. Although this structure was successful in

balancing development of generation and transmission and other operational aspects,

COSR incurred retarded innovation, obscure cost-reduction incentives, and improper

decision risk allocation. The perceived flaws of the monopoly structure ultimately led

regulators to gradually introduce competition to the industry. The enactment of the

public utility regulatory policies act in 1978 encouraged non-utility generation owners to

supply power to the existing utilities. The movement toward more competitive wholesale

power transactions was accelerated by the Federal Energy Regulatory Commission

(FERC)s Energy Policy Act of 1992, which opened the door of the previously monopoly

franchised generation market to independent power producers. The FERCs orders 888

and 889, in 1996, represent another major milestone by requiring non-discriminatory

access to the transmission network. The open access to transmission network and the

electricity wholesale market regardless of ownership of the generation unit triggered

another building boom of power plants. In addition, the frequent price spikes and

extreme volatility in late 1990s created profit and risk hedging incentives for independent

and utility-associated power marketers to participate in generation development actively.

Additional state and federal regulatory policies promoted the formation of independent

system operators (ISOs), and, in some cases, the divestiture of generation assets. ISOs,

10
where they were formed, separated operational control from ownership of the

transmission and generation assets to increase efficiency and inhibit detrimental activities

from conflicting interests. FERC Order 2000 continued this trend by promoting the

voluntary formation of regional transmission organizations (RTOs) and transmission

pricing reform.

2.1.1 Transmission Inadequacy

Regulatory changes in the transmission system were accompanied by nationwide

demand increases at an annual rate of 2-3% and substantial changes in the generation

sector. New generation technologies, particularly gas-fired combined-cycle turbines,

allowed electricity to be produced in more modular and flexible quantities with higher

efficiency. A building boom ensued added over 200GW of new generation between the

years of 1999 and 2004 (NERC, 2004). In many cases, these units were located

convenient to construction or fuel resource accesses while taking adequate transmission

connection capacity for granted. With the building boom reaching its end and the

evolution toward competitive markets well advanced, the transmission system is

becoming increasingly vital. The importance of its new role of supporting market

transactions is far beyond what is indicated by the relatively small capital cost it

represents in the electric power industry.

Compared with the steady increase of demand and generation, however, transmission

investment declined over the same time period. In 1972 approximately 30GW generation

was added, supported by $7.4billion (in year 2004 dollars) in transmission investment. In

2001, 40.6GW generation was added with only $4.6billion in transmission. By the year

2003, the numbers further diverged to having 52.4GW of new generation versus

$3.9billion invested in transmission. Normalized transmission capacity measured in

MW-miles/MW-demand and MW/MW-demand is declining at annual rates of 1.5% and

11
1.6%, respectively (see Hirst, 2004). The market environment strains the system further

because merchant power plants competing for short and long-term contracts with

multiple buyers are encouraged to transfer larger quantities of electricity over longer

distances more frequently to capture interregional market opportunities, raising power

flow patterns significantly different from the projected scenarios in system planning. As

a result, transmission loading relief (TLR) procedures, which dictate a certain percentage

of the power transactions which cause the monitored transmission overloading to be

curtailed, have been called frequently for managing transmission utilization to prevent

overload situations that put the system at risk. Statistics by North American Electric

Reliability Council (NERC) on the exercises of level 2 or higher TLR from July 1997 to

December 2005, shown in figure 2.1, illustrates the increasing frequency of transmission

system challenges in recent years. Note that load seasonality is the primary factor for the

TLR log fluctuations within a year.

320
Monthly Logs

240
12 Month Rolling Averages
TLR Times

160

80

Date
0
Jul-97 Nov-98 Apr-00 Sep-01 Jan-03 Jun-04 Nov-05

Figure 2. 1 Statistics of level 2+ TLR from July 1997 to December 2005

It is recognized nowadays that, while considerable effort has been devoted to

electricity market, development of the transmission sector has been largely overlooked.

The lack of widely accepted regulation rules and absence of effective market mechanisms

lead to vague signals for market participants. Potential transmission investors face

12
physical nature, organization structure, and market risk related obstacles that complicate

the projection of capital recovery. These risks include:

a. Free riding -- As a public good with non-excludable benefit sharing, typical

transmission investments are haunted with the free-riding problem due to the difficulty of

isolating the benefits to the investor. Market participants would choose to be free riders,

expecting positive externalities induced by others investments.

b. Market risks -- Given the ever changing electricity demands, generation portfolio,

network topology, and market rules, it is impossible to predict with accuracy the future

economic payoffs accrue from a transmission investment.

c. Regulation uncertainties -- One key element for the success of a transmission

project is regulatory approval. In many cases, however, alignment of federal, state, and

local regulations results in project delays, and rejection is a very real possibility.

Consequently, funds expended early in the project may be at substantial risk.

d. Lumpiness -- Transmission investments typically appear as additions of large

blocks of capacity. The obscured linkage between expected benefit and marginal cost

complicates the investment decisions. Joskow and Tirole (2003) illustrate that lumpiness

leads to transmission network underinvestment.

Note that these obstacles are confounded by fragmented ownership that easily leads to

sub-optimal solutions.

2.1.2 Transmission Pricing

Because it physically connects geographically dispersed regions and transport energy

among them, a transmission network is critical for supporting electric power trading and

making real the benefit of competitiveness and the economy of scale in the generation

sector. To provide incentives for adequate investment, systems need to set up an

effective pricing mechanism and value the transmission services correctly. Among

13
various approaches proposed along the years, the embedded cost and the marginal cost

based methods are two most notable ones. The primitive embedded cost method

allocates all transmission-incurred costs to transmission service customers as a

transmission tariff. Its variants include:

a. The postage stamp method -- The MW quantity and duration of a power

transaction determine the transmission charge irrespective of the actual transferring

distance and source/sink locations involved.

b. The contract path method -- This method assumes that the transaction incurred

power flows follow a contract path and the transmission cost allocation is confined to

transmission facilities along the contract path while ignoring the impact on others.

c. The boundary flow method -- The transmission charge is calculated based on

power flow changes, on a line or net interchange basis, incurred by the power transaction.

d. The MW-Mile method -- The MW quantity and mile transferring distance of each

power transaction is calculated to measure the transmission network usage. The system

transmission costs are allocated in proportion to the MW-Mile usage.

In summary, the embedded cost methods are based on approximated power flow

patterns that do not reflect actual system dispatching and do not follow up with the ever-

changing system conditions. Because Kirchoffs laws govern the flow of electricity

instead of economic expectations, the usage of transmission network is difficult to trace

as simplified in methods listed above. Transmission congestion and system externalities

are not fully reflected in individual prices but smeared over all network users. In contrast,

the marginal cost method measures the market value of delivering an additional unit of

electric power from the source to the sink bus. It is an extension of electricity locational

pricing theory. Since a power transfer is physically equivalent to an power injection and

a power withdrawal and the two locations involved, the market value of transmission

14
usage can be immediately revealed as the difference between the locational electricity

prices at the sink and source points. The opportunity cost of transmission service

provides economic incentives and promotes allocation of the scarce transmission capacity

to economic efficient users.

However, since LMPs are volatile and cannot be foreseen with accuracy in advance,

transmission usages are subject to fluctuating congestion charges. Such exposures to

market price risks create strong demands for congestion-hedging among risk-averse

transmission service customers. In PJM for example, the system operators identified

several groups of electrically neighboring buses with active power transactions and

defined them as trading hubs. The hub settlement LMPs are calculated as the average of

the group of bus to provide market participants with more stable prices since. Market

participants can also choose to withhold transactions in case of transmission congestion

to avoid congestion charges. Financial transmission rights (FTRs) also came up as

sophisticated tradable market instruments to meet market participants risk hedging

demands. A FTR entitles (or obligates) its holder to collect a stream of revenues

determined by LMP differences between the two underlying locations over a contractual

time period specified ex ante.

Before the structuring of FTRs, physical transmission rights (PTR) were proposed at

the conception of market restructuring. A PTR gives its holder the priority to access the

underlying transmission facility. In the PTR-based congestion hedging mechanism, to a

great extend, the allocation of PTRs through bilateral contracts or private auction markets

determines the usage of scarce transmission capacity and determines the system dispatch

as well. Therefore, PTR holders have market power to withhold transmission access and

hamper competition. The scheduling priority of PTR holders creates perverse incentives

which conflict with the bid-offer matching mechanism. Furthermore, the exclusion of the

15
system operators control from the withheld transmission capacity compromises the

system reliability. Its been argued that the physical interpretation of transmission rights

was the principal pitfall that buried the FERCs original capacity reservation tariff. In the

contrast, by entitling financial congestion rents instead of priority of transmission access

to FTR holders, the FTR-based congestion hedging mechanism keeps the centralized

market dispatch paradigm intact. It has been adopted into the PJM system since 1998, in

New York since 1999, and in New England since 2003.

Since the underlying LMPs are determined by the market dispatch model which takes

system operating constraints in both normal and contingent scenarios into consideration,

An FTR provides a perfect price hedge against transmission congestion for power

transactions between the underlying source and sink. Assume a market participant

schedules a transaction of q AB MW power from source A to sink B, where the fluctuating

LMPs are ~
p A and ~
p B during time period T , respectively. The congestion rent involved

will be q AB ( ~ p A ) . To hedge against the scenario that ~


pB ~ pB ~
p A is driven very high by

the ever-changing system operating conditions, the market participant can procure a FTR

of q AB MW from source A to sink B over time period T at a cost of C FTR . It entitles him

a revenue stream of q AB ( ~ p A ) . Therefore, the total cost of the transmission service


pB ~

C Total for the power transaction is the net of cost and revenue,

C Total = C FTR + q AB ( ~ p A ) q AB ( ~
pB ~ p A ) = C FTR
pB ~ (2.1)

Considering a random deviate q AB of the transaction quantity in real time market from
~
the projected q AB , the total cost of uncertain transmission service C Total is,

C Total = C FTR + q AB ( ~ pA )
~
pB ~ (2.2)

As long as the volume deviation q AB is trivial, the transmission service can be secured

at around the pre-determined FTR procurement cost.

16
Since a FTR is typically connected with an existing or projected power transaction

between two locations, it represents a power injection and withdrawal accordingly. To

conform to the system capability and ensure the revenue solvency of the system operator,

the setup of FTRs is subject to a simultaneous feasibility test (SFT) defined as follows

Definition 2.1 (Simultaneous feasibility test) Given a set of FTRs defined over a

common time period, the system operator needs to test if the transmission network can

accommodate the corresponding pairs of power injections and withdrawals (P I , PW ) ,

assuming they occur simultaneously.

Using DC power flow, assuming F and F' represent the system PTDF matrix in the

normal and a contingent scenario, respectively, correspondingly, the vector of

transmission capacity limits are T and T ' . The SFT problem can be described as follows,

F(P I PW ) T (2.3a)

F ' (P I PW ) T ' (2.3b)

Note that constraints (2.3a) and (2.3b) represent the feasibility criteria in normal and

contingency scenarios, respectively.

As a summary, the marginal cost based transmission pricing provides market

incentives for transmission capacity allocation, and the FTR mechanism provides

transmission service customers a transmission congestion risk hedging tool,

2.1.3 Transmission Investment

With the understanding of the valuation of transmission services, an efficient

mechanism is needed to indicate adequate transmission investment, to lower the

transmission congestion costs, and to facilitate the market based competitive electricity

trading. Due to the obstacles listed in section 2.1.1, investment cost allocation is a

complicated problem. The benefits of improving a transmission network as identified

17
bellow are essential to understanding the problem and finding incentives for transmission

investment,

a. Economic efficiency improvement -- Transmission expansion provides access to

alternative power sources and additional options to meet consumption at the lowest

possible social cost.

b. System reliability enhancement -- Higher transmission reliability margin and

diversified fuel accesses improve interconnection by accommodating more fluctuating

transactions, facility unavailability, and sudden disturbances.

c. Financial volatility reduction -- Transmission upgrades reduce system congestion

and alleviate market participants risk-hedging pressure created by volatile market prices.

d. Market power mitigation -- The electricity market is vulnerable to the exercise of

market power which intentionally creates scarcity and manipulates prices. This power

can be mitigated by proper transmission expansions which facilitate competition.

e. Environmental impact alleviation -- Although people are concerned about

transmission facilities right-of-ways, given the more severe environmental impact of

power plants, especially in highly-populated regions, transmission connection to remote

generation sources may provide a more environmentally-sound alternative to meet

demand (Bloyd et al., 2002).

Corresponding to the benefits listed above, various incentives drive investment

decisions or back up transmission investment cost allocations for capital recovery.

Different transmission investment forms can be envisioned as follows:

a. System-wide reliability enhancing and economic efficiency improving

transmission upgrades.

For a system with tight transmission capacity, transmission congestion happens

frequently and out-of-merit generation units have to be called to serve demands. The

18
inefficiency involved represents the social cost of the inadequate transmission capacity.

Figure 2.2 illustrates the net deadweight loss due to transmission congestion,

Figure 2. 2 Deadweight loss due to transmission congestion

In addition, system reliability is impaired when a critical transmission flowgate is

congested, since a subsequent contingency event may interrupt or compromise the quality

of bulk power supply.

System operators usually prepare regional transmission expansion plans which

consolidate reliability or improve system-wide economic efficiency. Since these projects

bring widespread benefits to most market participants in the electric vicinity, it is hard to

draw the beneficiary boundaries, and a regulatory process is involved to allocate costs

incurred to a large group of consumers through an added service charge. A request for

proposals (RFP) process is preferable since the RFP process promotes minimal cost while

assigning the projects risks to the winning respondent instead of to the end consumers.

Investments in this category should be limited to those projects the direct economic

benefits of which are non-significant to any market participant. A potential refinement of

this approach is to allocate investment costs through a cost-benefit analysis among sub-

regions so that a non-uniform portion of the cost can be allocated accordingly.

b. Voluntary transmission investment

Projects in this category include generation interconnection requests to increase

electricity delivery to the market, load connection requests to get access to desired

19
resources, and capacity expansions that reduce congestion energy cost for consumers in a

load pocket like, for example, New York City. The free-riders problem is less

bothersome here since the economic benefits of the projects to potential investors are

more exclusive, although they usually introduce more competition to mitigate the existing

market power of others. A good quantitative assessment of future market opportunities is

essential for the projection of economic cost-benefits analysis. Voluntary transmission

investment projects are to be sponsored by those who would gain the benefits. It is

necessary, however, to assure that the projects not degrade system reliability nor create

additional opportunities for the exertion of market power.

c. Merchant transmission projects

Merchant transmission investors rely on the LMP-based transmission service pricing

mechanism and seek financial transmission rights as payoffs. Three components are

essential to this category of market-motivated transmission investments. First, the

spatial-differentiated electricity price signals indicate where and how much to invest;

second, FTRs hedge against congestion-risks or, as tradable financial instruments, entitle

investors to the right to collect revenues in the future; third, an efficient financial

transmission rights identification and allocation mechanism is expected to assign the right

amount of rights to the right investors, which include eliminating the opportunities of

free-riding public good benefits. Note that the identification of incremental FTRs created

by merchant transmission investment is guided by the SFT problem (2.3a-b). With the

revenues entitled to the incremental FTRs, at least part of the cost can be recovered

without resorting to the traditional regulatory charges. Bushnell and Stoft (1996) show

that the transmission rights entitled revenues provide market incentives for transmission

investment. The merchant investment mechanism relies on free entry and unfettered

20
competition in the market. It places the risks of investment inefficiencies and cost

overruns on investment decision-makers instead of on the involuntary end consumers.

It is notable that, due to the network externality of power flows, a transmission

upgrade may affect the transfer capacity between two locations in another part of the

network. As a crucial issue regarding trade-offs between cooperation and competition in

the long run, the network externality may degrade system reliability and undermine open

competition. Should a transmission upgrade impair existing FTRs, the theory of public

economics suggests that the investor should buy back the disabled FTRs. Another option

for the system operator is to retain some transmission rights and avoid jeopardizing FTRs

held by market participants when a transmission upgrade causes negative impacts.

However, this induces additional overheads to be socialized to all consumers.

Nevertheless, regardless of which category it belongs to, a transmission investment

requires the assessment of market conditions in the future. The projection of locational

market prices as the market signals is one the most essential tasks involved. Two

competing approaches are available for market price modeling: a fundamental approach

that relies on system simulation; and a technical approach that models the stochasticity

directly. While the fundamental approach provides more realistic representations under

specific scenarios, it is computationally prohibitive due to the large number of scenarios

to be considered. The method to be presented in this chapter is to combine the strengths

of the two approaches by calibrating the stochastic price process models using

probabilistic system simulation results.

The rest of the chapter is organized as follows: as the core problem of market

simulation, the market dispatch is formulated as an optimal power flow (OPF) problem in

Section 2.2. The quadratic power flow (QPF) and costate based linearization approach

are used to solve the OPF problems efficiently. The impact of reliability requirement on

21
market prices is evaluated by incorporating the corresponding operational constraints. In

Section 2.3, the proposed market simulation model is applied to the IEEE RTS24 system.

The results according to different market dispatch models are compared and

incorporation of transmission reliability margin requirements is evaluated. In Section 2.4,

a quantile-based GARCH model is adopted to illustrate how to combine the strength of

market simulation and stochastic modeling to capture market dynamics. Market prices

obtained through system simulation are used to calibrate the stochastic model parameters.

Finally, Section 2.5 concludes with observations and presents discussion of additional

issues related to the modeling of market signals for transmission investment.

2.2 Market Dispatch

In a pool-based electricity market, market dispatch is one of the most essential

functions of the system operator (called ISO or RTO), which is an establishment that

coordinates the movement of wholesale electricity, acting neutrally and independently,

operating the competitive wholesale electricity market and ensures the reliability in

managing the regional transmission system and the wholesale electricity market. By

collecting electricity supply bids and demand requests, the system operator determines

the set of winning supply bids to meet the demands while observing all the system

operating constraints. This centralized market dispatch can be formulated as a network

constrained optimal power flow (OPF) problem, which solves a set of linear or nonlinear

equations and inequations to obtain the intended optimal operations. Alternative

objectives of an OPF problem include maximizing social welfare, minimizing customer

expense, minimizing system status deviation, minimizing transmission loss etc. The

rational for choosing different objectives and the respective implications on market

participants are discussed by Alonso et al. (1999). A primitive formulation of the market

dispatch OPF which minimize the total generation procurement cost is defined as follow,

22
Definition 2.2: (The market dispatch OPF problem) By collecting I supply bids and

J demand requests, the system operator conducts the market dispatch accordingly to

minimize the total generation procurement cost while accommodating all power flow

balance and transmission feasibility constraints.


I
Min C = C i (si ) (2.4a)
i =1

St. F Eq (s, d, x, u ) = 0 (2.4b)

F Ineq (s, d, x, u ) 0 (2.4c)

By minimization of the total generation procurement cost as the objective, the market

dispatch model actually determines the locational marginal prices. A LMP measures the

incremental system generation procurement cost for a unit of incremental demand at the

specific location. In addition, the settlement prices of transmission service market are

implicitly determined through this market clearing decision.

The system status variables x consist of magnitudes and phase angles of bus voltages,

and so on, the system control variables u includes real and reactive loads and generations,

voltage settings and bounds, transformer tap settings, and so on. For example, a

prescheduled multilateral transaction can be modeled as a set of power injections and

withdrawals at the corresponding source and sink buses. In a system with interface with

neighboring systems and power interchanges, the interface MW limits are usually treated

as required power injections or withdrawals in u , depending on the controlled power

flow directions.

The equality constraints F Eq (s, d, x, u ) = 0 are always binding at least to within a user

specified tolerance. They consist of generation bus voltage setting and the power balance

equations,

23
g (s, d, x, u ) = 0 (2.4d)

Specifically, according to the traditional power flow formulation, considering both

real and reactive power flows, the power balance equation at any bus n which matches

the locational generation and load is,

s n d n = v n v m [G nm cos( n m ) B nm sin ( n m )] (2.4e)


m n

+ jvn v [G
m n
m nm sin ( n m ) + Bnm cos( n m )]

Note that explicit load-flow equations are listed as essential constraints in the formal

establishment of the optimal power flow problems. Instead of approximating the losses

as a polynomial function of the power output of each unit (Wood and Wollenberg, 1996)

and calculate a penalty factor for each generation unit, the transmission losses are

accounted implicitly in the power flow equations and their market costs are imbedded in

the electricity locational marginal prices.

The inequality constraints F Ineq (s, d, x, u ) 0 consist of system operating limit and

bound constraints. In the primitive formulation of the market dispatch OPF, generation

capacity bounds

s s Max (2.4f)

and the transmission loading thermal limits,

h(s, d, x, u ) T (2.4g)

are usually considered.

Note that the transmission loading thermal limits apply to not only single

transmission lines, but to sets of transmission facilities with certain capacity limit,

defined as transmission flowgates, which constraint the power flow through the interface

involved.

The Lagrange function associated with OPF problem (2.4) can be defined as,

24
I
L (s , , , ) = Ci (si ) + g() + (s s max ) + (h() Tk ) (2.5)
i =1

where, [ , , ] are vectors of Lagarangian multipliers associated with constraints (2.4d),

(2.4f) and (2.4g), respectively.

( )
In order for a point s* , * , * , * to be optimal, in addition to (2.4b) and (2.4c), to

guarantee the gradient of C is normal to g () , h() T , and s s Max , it requires the

gradient of the objective function C be linear combination of the gradient vectors of

the active constraints g() , (h() T ) , and (s s Max ) ,

L
(s, , , ) = 0 (2.6a)
s s =s*

plus the complementary slackness condition,

* [h() T] = 0 , * 0 (2.6b)

( )
* s* s Max = 0 , * 0 (2.6c)

In the economic sense, the Lagrange multiplier n associated with the real power

balance at the bus n can be interpreted as the locational marginal price of energy because

it quantifies the cost (or value, from demand side) for supplying (or consuming) an

additional MW at the bus n of the network. On the other hand, the Lagrange multiplier

k associated with the power flow limit of the k th transmission flowgate is interpreted as

the variation in social generation procurement cost if the transmission capacity is relaxed,

called flowgate shadow price or congestion multiplier. And the Lagrange multiplier i

reveals the market opportunity cost associated with the scarcity of supplier i s generation

capacity.

By solving (2.4), LMPs can be read off the Lagrange multipliers associated with the

corresponding constraints, which measure the cost to serve the next MW of load at a

25
specific location, using the lowest production cost of all available generation, while

observing all operating constraints. In the absence of any binding constraints, all of

LMPs are identical. The FTR values can be readily derived from the price differences.

The non-zero shadow prices of binding transmission constraints are major factors which

diversifies the LMPs across the system. Market uncertainties due to fluctuating system

loads, varying generation bid function, and unexpected transmission circuit outages can

[
~ ~ ~
]
be incorporated by using random variables T, d, Ci carrying the distribution properties

of the underlying coefficients, see (Sun et al. 2005). With a parametric optimal power

flow formulation, the sensitivity of the optimal operating conditions with respect to any

parameter under interest corresponding to a market participants input or the system

operators control can be investigated and the corresponding economic values can be

interpreted.

2.2.1 Linearization of Market Dispatching Problems

Traditionally, to obtain the present system operating conditions for further analysis,

the power flow of the system can be described using the voltage magnitude and phase

angle at each bus, the transformer tap, and so on. These traditional power flow (TPF)

equations are formulated in terms of g (x, u ) = 0 based on the fact that the sum of the

injected power flows at a bus is zero. However, due to the high nonlinearity of some

equations, the converging to the solution is usually slow. For market dispatch, a common

practice in industry is to avoid the time-consuming iteration process of solving nonlinear

AC power flow equations and adopt DC formulation instead. DC power flow prompts

the simplicity and fast solution of the market dispatch problem. In the characterization of

transmission branches, DC model usually ignores conductance and reactive power flows.

It also assumes the voltage magnitudes at all buses are unit valued and the phase angles

26
are close to each other. Specifically, the power flow equations can be simplified as

follow in per-unit measures, not that only real power flows are modeled.

sn dn = B (
m n
nm n m ) (2.7)

However, accuracy is sacrificed unless sophisticated transformations of all limits into

energy or flow limits are made (see Alvarado, 2003).

Given the limitations of TPE and DC power flow modes, quadratic power flow (QPF)

model (Meliopoulos, 2001; Kang, 2001) is proposed based on modeling any power

system component as a set of linear or quadratic equations (which can be achieved with

the introduction of additional state variables). In formulating the QPF model, the system

states are expressed in the format of Cartesian coordinates instead of the polar

coordinates used in the traditional power flow model. This makes it possible to formulate

the power flow equations in linear or quadratic forms in terms of system state variables.

Unlike the TPF model, which consists of the power balance equations at each bus of the

system, the QPF model consists of writing the Kirchoffs current law at each bus of the

system. In general, a bus may be connected with generation, loads, circuit, shunt devices,

etc. While the circuits and shunt devices are linear elements, the loads and generation

may operate in such a way that imposes nonlinearities. The models of loads and

generators are expressed in terms of their terminal currents and additional equations in

additional state variables that define their operating modes. The additional equations may

be nonlinear but of order no higher than two. The resulting set of equations is consistent,

i.e., the number of equations equals number of unknowns. The convergence characteristic

of the QPF model is superior to that of TPF model since Newtons method is ideally

suitable to quadratic equations.

Application of connectivity constraints (Kirchoffs current law) at each bus yields the

quadratized power flow equations:

27
g Q (x Q , u) = [x Q , u ] [x Q , u ] + [x Q , u ] + = 0
T
(2.8)

where and are non-variable matrices and is a non-variable vector. Their values

are determined by the system parameters.

The solution to the quadratic equations can be obtained by the Newton-Raphson

method and iteration terminates when the norm of the QPF equations is less than certain

tolerance,

xQ
k +1 k
= xQ xQ ( )k 1
( )
g Q xQ
k
(2.9)

where, k the step of iterations

(x Q ) the Jacobian matrix of the set of power flow equations at state vector x Q

Although for a large-scale system the number of equations is large and so is the

dimension of the Jacobian matrix. However, the Jacobian matrix is highly sparse and the

computational burden can be significantly alleviated employing the sparsity technique.

An extension of the costate method to the QPF introduced in (Meliopoulos, 1988) and

applied in (Bakirtzis, 1991; Meliopoulos, 1994) can be applied to analyze the sensitivity

of a performance index f I (x Q , u ) (such as circuit loading and transmission loss) to

control variable u as follows,

Differentiation of the QPF equations (2.8) gives,

g Q (x Q , u ) g Q (x Q , u ) dx Q
+ =0
u x Q du

which leads to

g (x , u ) g (x , u )
1
dx Q
= Q Q Q Q (2.10)
du x Q u

The derivative of the system performance index function with respect to u c is given by:

28
df I (x Q , u ) f I (x Q , u )
+ x Q (x Q )
T
= (2.11)
du u

where x Q is the costate vector of the power flow equations, and

f (x , u ) g (x , u )
1

= I Q Q Q
T
x Q
x Q x Q

With the quadratic power flow model and costate method based linearization

approach, given a current system operating condition (s = s 0 , d = d 0 , x = x 0 ) , the solving

of a market dispatching problem with respect to changes in demand d can be

accelerated by linearizing the active and close-to-active constraints as follows and

converting market dispatch into a linear programming problem,


I
Min C = Ci (si + si ) (2.12a)
i =1

I J
St. si d j = 0
i =1 j =1
(2.12b)

I J

alh,i,s si + alh,,jd d j Tl Tl 0 , l
i =1 j =1
(2.12c)

s sMax s (2.12d)

where Tl 0 = hl (s, d, x ) are transmission loading at current operating conditions,


s =s 0 ,d =d 0 ,x = x 0

d d
alh,i,s = hl (s, d, x ) and alh,,jd = hl (s, d, x )
d ( si ) s =s
d (d j )
0 ,d =d 0 , x = x 0 s =s0 ,d =d 0 ,x = x0

are model coefficients obtained at the current system operating conditions, they measure

the sensitivity of transmission loading with respect to unit additional power injection by

supplier i and withdrawal at demand j , respectively.

Note that for any bus n , we have

29
alh,n,s = alh,n,d ,

which denotes the impacts on the power flow on flowgate l from one unit of incremental

power injection and withdrawal at bus n are in the opposite directions.

Since reactive power flows and voltage variations are accounted in the AC power

flow equations, alh,n,s and alh,n,d reflect the system operating conditions more accurately

compared with the PTDF coefficients derived by DC power flow equations. The detailed

computational procedures are referred to Appendix A.

To incorporate the changes in transmission loss due to the changes of demand and

supply and increase the accuracy of the linearization model, the system power balance

constraint (2.12b) can be extended as follow,

(1 a )s (1 + a )d
I J
l ,s
i i
l ,d
n j =0 (2.12f)
i =1 j =1

d d
where ail ,s = l (s, d, x ) , a lj,d = l (s, d, x ) .
d (si ) s =s
d (d n )
0 ,d =d 0 , x = x 0 s =s 0 ,d =d 0 ,x = x 0

anl ,s and anl ,d can be calculated in a similar procedure as for alh,n,s and alh,n,d . For n ,

anl ,s = anl ,d . They measure the sensitivity of system transmission loss with respect to unit

additional power injection by supplier i and withdrawal at demand j , respectively.

2.2.2 Incorporation of Reliability Constraints

Reliability of an electric power system is crucial to keep continuous supply of

electricity at required quality. Although there has been substantial research activity on

enhancing system reliability, the economic impact on market participants by imposing

transmission reliability constraints in a competitive electricity market has not yet been

adequately addressed. Without knowledge of the resulting market price dynamics, it is

hard for system operators to estimate the economic consequences and for market

30
participants to take proper market positions and manage risks involved. In retrospection,

the unprecedented volatile California market electricity prices in 2000-2001 led to rolling

blackouts and dramatic economic trauma to load serving entities. In addition to

fundamental reasons such as abnormal hydro resource and high demands, the inadequacy

of transmission capacity made the system vulnerable to the exercise of market power.

Suppliers intentionally leveraged the generation scarcity in the dysfunctional wholesale

market, since transmission bottlenecks held back the otherwise reachable alternative

generation sources to the buyers. The lessons call for imposing transmission adequacy

requirement to maintain a reliable environment for the energy trading and to support open

competition.

In an electricity market, as the competitions are getting more intensive and drive

market participants to chase market opportunities, more generation units and transmission

facilities are operated close to the edge for economic efficiency. This leads to higher

risks of equipment contingency status and puts the failure stakes higher than ever before.

A contingency event such as the forced outage of a transmission line or a generator unit

may jeopardize the entire system if there lacks back up transmission capacity or

generation source to support the power flows anticipated by the normal operation of the

system. To prevent that a single contingency events could trigger the occurrence of

cascading outages throughout the network, capacity reserves in generation and

transmission should be procured and contingency analysis should be conducted.

As mentioned earlier in section 2.1.1, reliability is deemed as public good with non-

excludable benefit sharing among market participants. Traditionally, the reliability

enhancement and the LMP-determining market dispatch are implemented as

economically separated activities. As a consequence, there lacks a market mechanism to

value reliability and the costs involved are smeared among all consumers. We propose

31
that economic values of reliability can, at least partially, be discovered with a properly

designed market dispatch mechanism, and a reliability-differentiated pricing scheme can

be constructed correspondingly. Market dispatch OPF formulation (2.4) can be

augmented to a security and adequacy constrained optimal power flow (SACOPF)

problem by incorporating extra security and adequacy related constraints to address

system reliability concerns.

The generation and transmission capacity reserve requirements are imposed for

system adequacy concerns to ensure the existence of sufficient facilities to satisfy system

disturbances. They can be imposed by allowing lower generation and transmission

capacity for market dispatch. For example, replacing the constraint (2.4f) with

( )
s 1 S s Max (2.13a)

is equivalent imposing a S 100% generation reserve margin. Similarly, a T 100%

transmission reserve margin can be imposed by replacing the constraint (2.4g) with

( )
h(s, d, x, u ) 1 T T (2.13b)

The unused capacity, namely, the generation reserve margin ss Max and the

transmission reserve margins T T can be called by contingency operating procedures

when contingency events such as load surges or equipment outages happen. With such

reserve margins, the systems can absorb the dynamics caused by the disturbances and

remain stable. The technical problems of determining spare capacity in each generation

unit and transmission facility to keep the system operation safety have been proposed by

Bobo et al. (1994) and McCalley et al. (1991) respectively. By incorporating the

generation and transmission reserve margin requirements as adequacy constraints into the

market dispatch, their economic incentives can be revealed and passed to the

32
corresponding beneficiaries. Therefore, an investment cost recovery mechanism can be

established accordingly.

The contingency test constraints are imposed for system security, which is related to

the ability of the system to respond to disturbances arising within the system. In the

power system operations, the N-1 criteria are widely adopted in industry for postulated

contingency tests. It means that given a normal operating condition [s, d, x, u ] , in case a

single postulated contingency event k happens, the demands d and the market dispatch

determined supplies s can still be accommodated with no pressure for any immediate

adjustment. However, the system control variables u will be changed to u (k ) and

deviated system states x (k ) will be reached. The system can stand the contingency states

x (k ) for a short period of time. However, to keep the safety of the system, the x (k ) should

not deviate out of the feasible regions and the system operator will go through certain

operational procedures to put system states from the edge of feasible regions back to

normal conditions. To reflect the N-1 contingency-proof criteria, for each postulated

contingency event k , the following constraints should apply,

( )
g s, d , x ( k ) , u ( k ) = 0 (2.13c)

h(s, d, x ( ) , u ( ) ) T ( )
k k k
(2.13d)

Note that under contingency event k , the transmission capacity vector may change to

T (k ) correspondingly.

It is straightforward to foresee that imposing the generation and transmission capacity

reserve requirements and the contingency tests leads to a more conservative market

dispatch at the cost of higher generation procurement cost. The incremental cost of

generation procurement incurred can be reflected by augmented market prices and passed

to the corresponding market participants accordingly.

33
Additional practices to promote a reliable market dispatch include imposing upper

and lower bounds for voltage magnitudes to keep the quality of power supply and the

safety of electric equipments,

vl v vu (2.13e)

and imposing upper and lower bounds for reactive power outputs which reflect the

corresponding generation unit feasible operation region,

Ql im(S ) Qu (2.13f)

where im(S ) denotes the reactive power outputs of generation units.

When solving the market dispatch OPF problem, as components of LMPs, marginal

costs associated with various reliability constraints can be read off the corresponding

Lagrange multipliers (Alvarado 2003). As more reliability constraints become active, the

LMPs become further differentiated between locations across the system. By taking

more conservative values for model parameters [ S , T , v l , v u , Ql , Qu ] , certain reliability

criteria can be imbedded accordingly. Different model parameters can apply to different

market participants according to their respective requirements. For example, for

consumers expecting lower loss of load probability, higher S and T values can be

determined to their generation suppliers and the corresponding transmission facilities.

One the other hand, for consumers expecting a relatively stable voltage level, a higher v l

or a lower v u should apply to the corresponding bus. Note that such parameter-settings

usually affect not individual but a group of end consumers, the decision should be based

on aggregated requirements. Due to the impact on the resulting LMPs, the reliability-

requirement differentiated pricing mechanism can be established.

By incorporating the reliability constraints explicitly into the market dispatch model,

the openness of the market pricing mechanism can be improved since market values

34
instead of regulation rules determine the service values and allocate the costs involved

accordingly. As the restructuring continuous, further details can be incorporated into

market dispatch model with the evolving market mechanism.

2.3 Electricity Market Simulation

As mentioned at the end of section 2.1, in the evaluation of a transmission investment,

one the most essential tasks involved is the modeling of locational market prices. As the

market signals, LMPs indicate the values of transmission services and the benefits of the

transmission investment. Therefore, capturing the characteristics of the LMPs correctly

is critical to the decision making regarding transmission investments. Section 2.2

presents a fundamental approach for market price modeling by representing the LMP-

determining market dispatch decision as an optimal power flow problem. Given the high

dimension of a practical electric power system, analytical evaluation of the probabilistic

properties of the market price behaviors requires very complicated numerical methods

and procedures. A natural alternative is to resort to system simulation, through which the

fundamental uncertainty factor of the system can be represented as random variable with

corresponding distributions.

Numerical experiments with the IEEE RTS-24 system are presented in this section to

empirically illustrate the fundamental modeling of market signals. The structure of the

IEEE RTS24 system is illustrated in figure 2.3,

35
Figure 2. 3 The IEEE RTS24 system

The system consists of 24 buses connected with 38 transmission lines. There are load

demands at 17 of the buses, and 32 generation units connected to 10 of buses. Note that

multi generation units of different capacities can reside at one bus and some buses are

connected by double transmission lines. Bus 15 is the system slack bus. Most generation

units reside at the upper part of the system, which consists of buses 11 - 24 and is

operated at 230 kilovolt (kV). The lower part of the system is operated at the 138 kV

and is connected with the upper system through voltage transformers. The system has a

total installed generation capacity of 3561MW. The yearly peak load D Peak is 2850 MW.

And the weekly, daily and hourly load peaks are given in percentage of D Peak .

36
Generation unit parameters [s max , Ql , Qu ] , transmission capacities T and T ( k ) in

contingency k , transmission parameters [Rmn ,X mn ,Gmn ,Bmn ] for each line m n , the

outage rates and the averaged failure-repair cycles of all types of generation and

transmission equipments are referred to (Billinton and Li, 1994). The function generation

procurement cost from each generation unit i is assumed to be a quadratic function of the

real power output si , C (si ) = ai + bi si + ci si where the coefficients [ai , bi , ci ], i are


2

referred to (Meliopoulos et al., 1990). To reflect the increase fuel cost over the years, the

values of these coefficients are doubled. We assume that the generation cost function as

used in the objective function of the market dispatch include the components of a profit

margin beyond the operational costs. In this case, the short-run marginal cost as used for

the first best pricing can provide enough market incentives for generation suppliers.

The original IEEE RTS24 system was configured for the test of system generation

adequacy related reliability analysis only, high transmission capacity was assumed in

(Billinton and Li, 1994). Since we are more interested in the modeling of transmission

adequacy related market signals, we create transmission congestions intentionally by

setting transmission line capacities T and T ( k ) , in normal and contingency k scenarios

respectively, to be 60% of the limits given in (Billinton and Li, 1994). Correspondingly,

the system peak load level is reduced to be 60% as well.

The rest of the section is organized as follows: In Section 2.3.1, we compare the

differences of LMPs determined by AC and DC power flow based market dispatch

models. The causes and the impact of the differences are identified. To investigate the

effect of incorporating reliability constraints into the market dispatch, comparative

experiments are conducted in Section 2.3.2 by using different formulations of the market

dispatch OPF problem. Preliminary observations on the changes of LMPs at different

types of buses are provided accordingly.

37
2.3.1 Comparison of DC and AC Models

In the proposed market dispatch model in Section 2.2, we advocate using AC power

flow formulation for equation (2.4d) instead of using the DC power flow formulation.

Since the transmission line resistance is ignored in the DC power flow equations, the

system is assumed to be lossless. Also assumed in the DC power flow equations is that

reactive power is perfectly distributed over the network and the voltage magnitudes at all

buses are exactly at the nominal valued. However, for a practical system, given the

available reactive power resources at certain but not all buses, controls over the reactive

power flow are exercised to prevent dramatic voltage fluctuates.

We compare the differences of LMPs determined by AC and DC power flow based

market dispatch models through numerical examples in this section to advocate using AC

power flow. For the N-1 contingency test described in (2.13c) and (2.13d), we consider

the outage of transmission lines 5-10 and 18-21 respectively. The voltage magnitude

bounds v l , v u are set to be 97% and 103% of the corresponding nominal values,

respectively. We also assume the installment of VAR compensation and voltage

regulation equipments and the voltage magnitude bounds under contingency are relaxed

to be 95% and 105% of the corresponding nominal values, respectively, as considered in

(Milano, 2003).

We first look at a scenario when the system load is low (40% of peak load). Figure

2.4 illustrates the LMPs determined by the AC and DC power flow based market

dispatches at all buses.

38
54

LMP ($/MWh)
AC Power Flow Based
Market Dispatch
46
DC Power Flow Based
Market Dispatch
38

30

22
Bus
1 6 11 16 21

Figure 2. 4 Comparison of LMPs at all buses determined by AC and DC power flow


based market dispatches when the system load is 40% of peak level

In this low-load scenario, no transmission line is congested, no generation unit

reaches its capacity bounds, and no bus voltage magnitude bound constraint is active.

The LMPs according to the DC model are the same for all buses. However, due to the

transmission losses, the LMPs on all buses according to the AC model show certain

variety across the system. The differences between the two models are most notable at

pure load buses 4-14. Note that according to the AC model, the total generation real

power output is 709 MW, while the total demand is 684 MW only. This indicates a

transmission loss of 25MW over the network, which accounts for 3.65% of the total

power withdrawal. Actually, by setting transmission line resistance to be zero in the AC

model, an uniform LMP of 26.09 $/MWh can be reached for all buses, which is

equivalent to the DC model result as shown in figure 2.4.

Overbye et al. (2004) argue that the ignorance of DC model can be compensated by

uplifting the system load by a certain percentage to account for the transmission loss as

reveal by the AC model. Following their suggestion, we calculate the LMPs using the

DC model while uplifting the loads at all buses by 3.65%. The resulted LMPs at all

buses are 26.34 $/MWh. Although it is closer to the AC model results shown in figure

2.4, differences remain on most buses. This is not a surprising result since the uplift of

39
loads at a same percentage is equivalent to the allocation of transmission losses to all

consumers in proportion to their respective loads regardless of their location in the

network. In a real system however, the incremental transmission losses caused by unit

power withdrawals at different locations are generally not the same. The impact of this

difference is more significant if we look at a high-load scenario when the system load is

90% of the peak level. Figure 2.5 illustrates the LMPs determined by the AC and DC

power flow based market dispatches at all buses.


LMP ($/MWh)

54
AC Power Flow Based
Market Dispatch
46
DC Power Flow Based
Market Dispatch
38

30

22
1 6 11 16 21 Bus

Figure 2. 5 Comparison of LMPs at all buses determined by AC and DC power flow


based market dispatches when the system load is 90% of peak level

In this scenario, power transfers in larger quantities cause transmission congestions

and lead to the differentiation of LMPs over the network. By uplifting the loads at each

bus to compensate for transmission losses in the DC model and setting the transmission

line resistance to be zero in the AC model, the LMPs determined by the respective

adjusted market dispatch models are illustrated in the following figure 2.6,

40
54

LMP ($/MWh)
DC Power Flow Based
Market Dispatch with
46 Load Uplift for
Transmission Losses

38 AC Power Flow Based


Market Dispatch with
Transmission Losses
30 Ignored

22
1 6 11 16 21 Bus

Figure 2. 6 Comparison of LMPs at all buses determined by adjusted AC and DC power


flow based market dispatches when the system load is 90% of peak level

When comparing figure 2.6 with figure 2.5, we note that by ignoring transmission

losses in the AC model, the LMPs determined are closer to the DC model results.

Similarly, uplifting the loads in DC model to compensate for transmission losses brings

the LMPs resulted closer to those determined by the AC model. However, the remaining

differences are more significant than in the low-load scenario.

Actually, another immediate observation from figure 2.5 is the significant difference

between AC and DC model determined LMPs at buses 7 and 8. It reveals the impact of

voltage bound constraints (2.13e) in the AC market dispatch model, which are absent

from the DC counterpart. Given power flow Pmn + jQmn over a transmission line m n ,

the voltage decrease Vmn along the line is

Vmn = (Rmn + jX mn )
(Pmn + jQmn )e j m

(2.14)
Vm

In the high-load scenario illustrated in figure 2.5, heavy power flow over the transmission

line 10-8 leads to significant voltage decrease along the transmission line and causes the

binding voltage lower bound at bus 8. This limits the output of low-cost generation at

41
bus 13 to meet load at bus 8 and incur the high cost generation output at bus 7 instead.

Consequently, LMPs at buses 7 and 8 are brought up.

It is foreseeable that if the transmission line impedance is high, the voltage decrease

incurred is significant and the voltage bound constraints tend to become active and cause

the redistribution of power flows over the network. Consequently, the LMP values

reflect the economic incentive of voltage regulation.

By reading figures 2.4 and 2.5 again, another difference between AC and DC power

flow based market dispatch can be observed: LMPs determined by the AC model show

more significant differences between buses in the system. To quantify this difference, we

define a spatial volatility measure of the LMPs as follow,

(P P ) (N 1)
N
2
SV = n (2.15)
n =1

N
1
where P =
N
P
n =1
n .

The spatial volatility of LMPs at the low and high load scenarios according to the AC

and DC power flow based models are listed in table 2.1 as follow,

Table 2. 1 Spatial volatility of LMPs calculated by AC and DC power flow based market
dispatch ($/MWh)
DC Model AC Model
Low-Load Scenario 0 1.57
High-Load Scenario 2.37 5.34

As shown in table 2.1 above, with more constraints explicitly incorporated, the AC

power flow based market dispatch model determines higher spatial volatility of LMPs in

the system. This observation coincides with intuitive reasoning.

Due to the inaccuracy of DC power flow approximation, for the rest of the numerical

examples in this chapter, we resort to the AC power flow based market dispatch models.

42
2.3.2 Evaluation of Reliability Constraints

In section 2.2.2, we proposed incorporating reliability constraints into the market

dispatch model to reveal the economic incentives of reliability enhancement though

augmented LMPs and allocate the costs involved to the beneficiaries. Among various

methods to reach higher reliability through market dispatch, we advocated the imposing

of generation reserve margin ss Max and the transmission reserve margin (TRM) T T as

indicated by constraints (2.13a) and (2.13b). Since we try to tackle the modeling of

market signals for transmission adequacy in this chapter, we focus on the TRM T T and

present numerical examples in this section to show the impact of imposing constraints

(2.13b).

Before jumping onto the experiments results, we start with a look at the sensitivity of

LMP with respect to transmission capacity using the market dispatch model without

TRM requirements. The following figure 2.7 plots the changes of system averaged LMP

(equal weight for LMPs at all buses) with respect to the assumed transmission capacity

limit expressed in percentage of their nominal values.


System Averaged LMP ($/MWh

42

40

38

36

34
60% 80% 100% 120%
Transmission Capacity in Percentage of Nominal Value

Figure 2. 7 System averaged LMP with respect to transmission capacity using market
dispatch without TRM requirement when the system load is 80% of peak level

43
From figure 2.7, we observe that the system averaged LMP decreases with the

increase of available transmission capacity. When transmission capacity reaches a

threshold point of adequacy when no transmission congestion appears throughout the

system, the LMP stays constant even when more transmission capacity is invested. This

is because the generation resources are dispatched in merit-order already and the increase

of transmission capacity does not change the profile of winning generation supply bids.

However, no more conclusions can be drawn from figure 2.7 unless that the transmission

expansion increases economic efficiency in the sense of using less-costly generation to

meet system demands. In order to find out the impact on distinct market participants, we

pick any 2 buses (5 and 10) with pure load demands and 2 buses (18 and 22) with high

installed low cost generation capacity and plot the changes of their LMPs as the available

transmission capacity changes in the following figure 2.8,

55
Bus 5
LMP ($/MWh)

Bus 10
48
Bus 18
Bus 22
41

34

27

20
60% 70% 80% 90% 100% 110% 120%
Transmission Capacity in Percentage of Nominal Value

Figure 2. 8 Generation and load bus LMPs with respect to transmission capacity using
market dispatch without TRM requirement when the system load is 80% of peak level

As the transmission capacity increases, the changes of LMPs at the 2 load buses 5 and

10 show quite similar tendency: decreases to a low value until transmission capacity

reaches a certain adequacy threshold and stays the same afterwards. It corresponds to the

44
observation from figure 2.6 above. For generation buses 18 and 22, the direction of the

changes is the opposite: increases to a low value until transmission capacity reaches a

certain adequacy threshold and stays the same afterwards. It is an interesting observation

since it reveals that low-cost generation suppliers can sell more energy at higher prices

given adequate transmission to support their power transactions.

Actually, similar observations are observed when applying different TRMs T T .

Higher T values actually reduce the available transmission capacity for market dispatch

in normal system operating conditions and reserve more capacity for contingency

operating procedures. According to figure 2.8, we expect different changes of LMPs at

load and low-cost generation buses when a TRM requirement is imposed. Using the

given hourly load over one sample year (52 weeks), we calculate the hourly LMPs at

buses 10 and 18 sequentially and plot their probability distributions in the following

figures 2.9 and 2.10, respectively.

T
=0%
0.12 T
=10%
T
=20%
Probability Density

0.1

0.08

0.06

0.04

0.02

30 35 40 45
LMP at Bus 10

Figure 2. 9 Probability distributions of LMPs at bus 10 over the sample year with
different TRMs requirements

45
T
=0%
0.45 T
=10%
T
=20%

Probability Density
0.35

0.25

0.15

0.05

22 23 24 25 26 27 28
LMP at Bus 18

Figure 2. 10 Probability distributions of LMPs at bus 18 over the sample year with
different TRMs requirements

Observations from figures 2.9 and 2.10 coincide with the indication from figure 2.8:

higher TRMs increase the LMP at the load bus due to transmission congestion charges

and lower the LMP at the generation bus due to lower output level, while lower TRMs

grant more transmission capacity for market dispatch, reduce transmission congestions

and determine less differentiated LMPs between buses. The quantitative measures of

averaged spatial volatility of LMPs over the sample year at different TRMs are listed in

table 2.2 as follow,

Table 2. 2 Averaged spatial volatility of LMPs over the sample year when imposing
different TRM requirements in the market dispatch ($/MWh)
0% TRM 10% TRM 20% TRM
4.02 4.67 5.12

Through simulation studies, this section compares alternative market dispatch models

which differ from each other in DC or AC power flow formulations and in the set of

reliability constraints considered. Their impacts on market prices are investigated and

interesting observations are made accordingly. In the simulation study, we calculate the

46
market prices at the interval of 1 hour. In a practical system, for example in PJM and

NYISO, LMPs are usually every 5 minutes. The difference between these different

market dispatch models would be more significant if applied in such practical systems.

In addition, as is one of the goals for the restructuring of electricity market, to establish

economic incentive based market mechanisms for ancillary services, market dispatch

models should be enhanced accordingly to increase the transparency of market pricing

and cost allocation, as is proposed in our model for reliability enhancement. More

information regarding the market dispatch should be open to market participants for

better market assessments. Due to complexity of the problem, the observations are

preliminary and system-specific. However, a framework is constructed and the

importance of the problem is revealed, more analytic results are expected in our further

research.

2.4 Stochastic Electricity Price Models

As mentioned earlier in section 2.1, two approaches are widely employed in studying

the electricity price behaviors, namely, a fundamental approach and a technical approach.

As the core of the fundamental approach, the market dispatch model is studies in section

2.2 and applied in system simulation study in section 2.3. Along the line of technical

approach, there is a growing literature on electricity price modeling utilizing reduced-

form stochastic processes (Deng, 1999), fussy regression combined with neural network,

and Fourier Hartley transform-based techniques (Clewlow and Strickland, 2000).

However, it is challenging for these models to reveal the impacts of bulk power system

constraints on price volatility. Moreover, the constantly changing market rules make the

technical models difficult to calibrate using market data. This section tackles the

technical approach by combining it with the fundamental approach and using the system

simulation data to calibrate the technical model parameters.

47
2.4.1 Quantile based GARCH Model

Due to the unique characteristics of electricity such as non-storability, transmission

capacity constraining, and inelasticity of electricity consumption, unparalleled volatilities

exhibit in electricity prices. Other observations of empirical electricity prices including

volatility clustering and the heavy tail have been revealed in literatures. Merited with the

flexibility of quantile-based probabilistic models in capturing stylized features of

empirical data, the quantile-based probabilistic model advocated in (Deng and Jiang,

2004) is implemented for modeling the unconditional distribution and to fit the marginal

distributions of hourly log-return of electricity prices. The electricity prices obtained by

simulation are used to calibrate the model parameters to make the model applicable to the

specific system setup. A non-Gaussian Generalized Autoregressive Conditional

Heteroskedasticity (GARCH) model using a quantile specific distribution is constructed.

In contrast to the traditional time series analysis which focus on modeling the conditional

first moment, GARCH model incorporates the conditional second moments explicitly to

better interpret the dynamics. The versatile capability of the quantile based distributions

is illustrated in (Jiang, 2000).

To characterize a probabilistic distribution of a univariate random variable X , the

monotonically increasing quantitle function q ( y ) , y (0,1) as opposed to the probability

distribution function F (x ) can be defined as the generalized inverse

q( y ) = F 1 ( y ) = inf {x : F ( x ) y} (2.16)

where the infimum is taken over the range of X .

The following quantile based distribution Q( , , , ) is constructed in Jiang (2000)

described by the following quantile function

48
1


1
x
q( x; , , , ) = ln
+ (2.17)
1 x

where , , R + , R and the superscript ( ) for > 0 represents the operation:

x x>0

x ( ) = 0 x=0
( x ) x<0

And the probability density function of distribution Q( , , , ) is,


(x ) ( ) exp 1 (x )( )

(x )
f ( x; , , , ) = 1
for x ( , ) ( ,+ ) (2.18)
1+
1
(x )( )

1 + exp

The intuitive interpretations for parameters of the quantile based distribution

Q( , , , ) are: parameter indicates the tail order, the smaller the value of , the

fatter the tail of the distribution; depicts the balance of distribution tails with = 1

meaning a balanced tail and < (> )1 meaning a fatter left (right) tail compared with the

other; scales the point probability density ; and determines the central location of

the distribution.

The time series of the hourly log returns of electricity prices at bus n can be

calculated as:

pn , h
n ,h = log = log( pn , h ) log( pn , h 1 ) (2.19)
p
n , h 1

where h is the index of hour, h = 1,L, H and pn ,h denotes the simulated LMP on bus n

at hour h .

49
Given the distributions defined by the quantile function, we turn to the time series

models based on the quantile based distribution Q( , , , ) defined above. In

particular, a non-Gaussian GARCH(1,1) model with Q( , , , ) -distributed error terms

specified in equation (2.20) as follow is adopted to model the hourly log returns of

electricity prices at bus n ,

n, h = n, h n, h (2.20)

n2, h = 0 + 1 n2, h 1 + 2 n2, h 1 , 0 0, 1 0, 2 0

where, 0 , 1 , 2 are non-negative constants, n,h s are identical and independent

{ }
Q( , , , ) random variables and are independent of n , h k , k 1 for h .

This quantile GARCH model describes the distribution of n ,h as a scale-transform

of the distribution of n, h where the scaling factor n, h depends on the past of the

n , process. Compared with GARCH models with normal distributed error terms, the

model (2.20) above has the following merits,

a. The closed-form quantile function of n ,h leads to the explicit likelihood function

and makes the maximum likelihood estimation an applicable approach for the inference

of model parameters [ , , , ] .

b. The diversity of tail patterns of the conditional distribution of n ,h with respect to

the model parameters and gives flexibility in accommodating various tail behaviors

imbedded in the empirical data.

c. The GARCH model provides inherent ability in capturing the volatility clustering

characteristic of electricity price processes.

2.4.2 Parameter Calibration

50
For the estimation of the model parameters, a quantile match approach was used in

(Jiang, 2000) which minimize the Ln norm distance between the theoretical quantile as a

function of model parameters and the empirical quantile. The explicit quantile and

probability density functions of the distribution Q( , , , ) make the likelihood

estimation (MLE) another viable approach for parameter inference in the quantile based

GARCH model.

The time series of n ,h modeled in (2.20) can be rewritten as follow,

n , h = 0 + 1 n2, h 1 + 2 n2, h 1 n , h (2.21)

or, n ,h = n ,h n ,h , where n , h = 0 + 1 n2, h 1 + 2 n2, h 1

According to distribution Q( , , , ) s property (see Deng and Jiang, 2004),

n ,h ~ f ( n ,h ; , , n,h , n ,h ) (2.22)

The joint probability distribution of the hourly log returns of LMP at bus n over a time

period h = 1,L, H is,

( ) ( )
H
f n ,1 , n , 2 ,L , n ,H = f n ,h | n ,h1 (2.23)
h =1

The log-likelihood function can be expressed as,


= ln ( f ( n ,1 , n , 2 ,L, n ,H )) = ln f ( n ,h | n ,h1 ) = ln ( n ,h | n ,h1 )
H H
(2.24)
h=1 h =1

( ( ))
H
= ln f n ,h ; , , n,h , n ,h
h =1

To estimate the parameters [ 0 , 1 , 2 ] for the GARCH model and the parametric

probability distribution parameters [ , , , ] , the optimization problem of maximizing

( )
the log-likelihood function { n ,h }h =1 ; , , , , 0 , 1 , 2 can be written as follow,
H

51
Max (
{ n ,h }h=1 ; , , , , 0 , 1 , 2
H
) (2.25a)

St. 0 = 0 , n2,h = 0 + 1 n2,h1 + 2 n2,h1 , h = 1, L, H (2.25b)

0 , 1 , 2 0 , , , 0 (2.25c)

1 + 2 < 1 (2.25d)

where constraint (2.25d) is intended to guarantee the stationary of the stochastic process

with E [ n2,h ] < (see Peng, 2005).

Due to the high nonlinearity of the problem (2-25), the convergence to optimum is

very slow and difficult to achieve. Since it is proved that assuming normality for error

terms n ,h does not affect the inference of GARCH model parameters [ 0 , 1 , 2 ] , the

problem (2-25) can be decomposed into two steps as follows. In the first step, estimate

the GARCH model parameters [ 0 , 1 , 2 ] by assuming the errors follow a normal

distribution N (~, ~ ) . This particular distribution has two parameters ~ and ~ only

instead of the four in the Q( , , , ) distribution, the estimation for [ 0 , 1 , 2 ] can be

obtained relatively easier by replacing (2.25a) with the following instead,


H
( n ,h ~ ) 2
(( ))
H
Max ln f n,h ; n,h , n,h = ln ( n,h ) +
~ 2 ~
~
2 ~2
+C (2.25f)
h =1 h =1 2 n ,h

where, the constant term C can be ignored without affect the optimality searching.

In the second step, obtain estimates for the Q( , , , ) distribution parameters

[ , , , ] using the values of [ 0 , 1 , 2 ] obtained in the first step. The corresponding

optimization problem is,

Max ( H
)
{ n ,h }h=1 ; , , , = ln ( f ( n ,h ; , , n,h , n,h ))
H

h =1
(2.26)

52

( n ,h n ,h )( ) exp 1 (

n ,h )( )

H ( n , h n ,h )

n ,h

=
n ,h
1
1+
h =1
( )



1
(
n ,h 1 + exp n ,h n ,h )

n ,h

St. , , 0

where n,h s are calculated by (2.25b).

The quantile based GARCH model (2.20) is applied to model the simulated IEEE

RTS24 system electricity price processes obtained in the previous section. First, the

LMPs at buses 10 and 18 determined by the AC power flow based market dispatch

without TRM requirement are selected for parameter calibrations. Following the quantile

function based parameter estimation procedures described above, the model parameters

for the quantile distribution that fit the simulated RTS24 electricity prices are estimated

and reported in table 2.3,

Table 2. 3 Estimates of model parameters using LMPs at buses 10 and 18 determined by


market dispatch without TRM requirement
0 1 2
Bus 10 0.5843 0.8316 0.0689 0.0148 0.0780 0.1092 0.3701
Bus 18 0.8971 0.9261 0.1223 -0.0045 0.1206 0.0804 0.2923

Figure 2.11 illustrates the quantile-quantile (Q-Q) plots of the empirical quantile

versus the respective theoretical quantiles. The theoretical quantiles are plotted on the x-

axis and the empirical quantiles on the y-axis. In the two QQ plots in figure 2.11 and

2.12, the mid part of the plot forms a relatively straight line, which indicates a good fit

between the empirical quantile function and the theoretical quantile function, except for

several outliers.

53
Bus 10 Bus 18
0.45 0.35

Emperical Quatiles

Emperical Quatiles
0.2 0.15

-0.05 -0.05

-0.3 -0.25
-0.3 -0.05 0.2 0.45 -0.25 -0.05 0.15 0.35
Theoretical Quatiles Theoretical Quatiles

Figure 2. 11 Q-Q plot for fitted distribution of hourly log returns of LMP at bus 10 (left
panel) and 18 (right panel) determined by market dispatch without TRM requirement

According to the estimations of [ , , , ] , the tails of the distribution of LMPs at

bus 10 is fatter than the tails of LMPs at bus 18. Nevertheless, for both distributions,

< 1 indicate that both distributions have fatter left tails than the right tails.

Next, the LMPs at buses 10 and 18 determined by the AC power flow based market

dispatch with 10% TRM requirement are selected for parameter calibrations. The model

parameters for the class I distribution that fit the simulated RTS24 electricity prices are

estimated in table 2.4,

Table 2. 4 Estimates of model parameters using LMPs at buses 10 and 18 determined by


market dispatch with 10% TRM requirement
0 1 2
Bus 10 0.5032 0.7963 0.0293 0.0201 0.0909 0.0988 0.3691
Bus 18 0.9204 0.8482 0.2840 -0.0012 0.0886 0.1911 0.2002

54
Bus 10 Bus 18
0.5 0.3

0.3 0.18

Emperical Quatiles

Emperical Quatiles
0.1 0.06

-0.1 -0.06

-0.3 -0.18
-0.3 -0.1 0.1 0.3 0.5 -0.18 -0.06 0.06 0.18 0.3
Theoretical Quatiles Theoretical Quatiles

Figure 2. 12 Q-Q plot for fitted distribution of hourly log returns of LMP at bus 10 (left
panel) and 18 (right panel) determined by market dispatch with 10% TRM requirement

By comparing the estimates of model parameters in table 2.4 with those in table 2.3,

we find that the LMP at bus 10 determined by market dispatch with the TRM

requirement shows fatter tail than without TRM requirement. However, the opposite is

found for bus 18. This observation matches the observations from figures 2.9 and 2.10

illustrated in the previous section.

This section illustrates how the parameters of the stochastic models can be calibrated,

and how the market simulation based fundamental approach can be combined with the

stochastic model based technical approach to model the market signals. With the

parameters well calibrated, these technique models can be employed to value economic

incentives and transmission services, which provide market signals for making the

investment decisions on where and how much to invest in generation and transmission.

Investment opportunities on transmission capacity such as maintenance, upgrade and

construction of new lines can be evaluated based on the projection of future market

conditions using the reduced-form electricity price models calibrated through the power

system simulation model.

55
2.5 Conclusions and Discussions

This chapter addresses the problem of modeling market signals for transmission

adequacy investment. An AC power flow based market dispatch model which takes into

account various system operating constraints as well as the system reliability constraints

is proposed. Simulation experiments with the IEEE RTS24 system illustrate the

advantages of this model as compared with others. The AC model captures the economic

incentives of adding the voltage bound constraints and therefore the allocation of reactive

power flows. Compared with ignoring transmission losses or simply allocate an

estimated transmission loss in proportion to the load consumption as indicated in the DC

model, the AC model allocate transmission loss correctly. The economic impact of

reliability enhance through market dispatch is evaluated for different market participants.

The cost of imposing reliability constraints can be recovered from consumers with

limited transmission accesses to generation resources. It indicates that a market based

mechanism for transmission reliability investment can be established accordingly. To

combine the strength of the market simulation based fundamental approach and the

stochastic model based technical approach, market prices obtained through system

simulations are used to calibrate the parameters of a quantile-based GARCH model. The

changes of system conditions can be captured by the stochastic model parameters. This

research work builds a power tool for projecting the dynamics of future market signals,

which is essential for the evaluation of transmission adequacy investments.

Future work can be carried out to consider the strategic interactions between

transmission investors, and the coordination of generation and transmission investments.

Finding a reduced form hub-and-spoke representation of a complex system to improve

trading liquidity, the adequacy of market incentives are intended to be investigated in the

near future as well.

56
CHAPTER 3

HYDROELECTRIC GENERATION SCHEDULING IN MULTI-

MARKETS WITH MARKET UNCERTAINTIES

This chapter proposes a stochastic programming framework for solving the optimal

scheduling problem faced by a hydroelectric power producer that simultaneously

participates in multiple markets. Specifically, the hydro-generator participates in both the

electricity spot market and the ancillary services market as a price taker. It seeks to

maximize its profit by jointly optimizing its energy/capacity sales and scheduling into all

markets subject to market uncertainties and operational constraints. The impact of

market uncertainties on the co-optimization problem over a pre-specified time horizon is

analyzed through the stochastic programming formulation incorporating both ancillary

service and market price uncertainties. Numerical studies on the advantages of the

proposed stochastic co-optimization strategy for a hydro-generator for hedging market

uncertainties are carried out in a realistic case study. The proposed model can also be

directly adapted for determining the optimal scheduling and bidding strategy for a power

producer facing additional sources of market and operational uncertainties.

The following notations apply to this chapter.

Indices and Sets

HA Set of planning time horizon hours, H A = {h = 1,2, L , H } .

HP Set of on-peak hours, H P = {h H A 13 mod(h,24 ) 20}.

Hn Set of hours at stage n , H n = {h = I n 1 + 1, I n 1 + 2,L, I n } , I 0 = 0 , I N = H .

j Scenario of market price, j = 1,2,L, J .

57
k Scenario of ancillary service, k = 1,2,L, K .

Lh Indicator of market price scenario in hour h .

n Stage index, n = 1,2,L, N .

Sn Indicator of market participation scenario at stage n .

Parameters and Coefficients

ph Vector of market settlement prices in hour h .

Tn{l ,u} Lower/upper MW-equivalent water release target at stage n , TNl = TNu = TN .

n{l ,u } Penalizing coefficients for violating lower/upper water target at stage n .

n Given tolerance of missing the water target at stage n , 0 n < 1 .

A coefficient matrix converting vector h to vector P h .

h [
Vector of phE , pUh , phD , phS , phN , phB ]
T
with the components denoting the prices of

energy market, regulation-up, regulation-down, spin reserve, non-spin reserve

service markets and balance market in hour h , respectively

State transition matrix of a market price level Markovian process.

Limiting probability vector of Markovian states.

h A constant coefficient vector, h = [1,1,0,1,1] .

h A constant coefficient vector, h = [0,1,0,1,1] .

[
A constant coefficient vector, = 1, u , d ,0,0 ]
{U , D , S , N } Expected call probabilities of {regulation-up, regulation-down, spin reserve,

non-spin reserve} services.

~h{U , D , S , N } Bernoulli random variables with expected values being {U , D , S , N } .

j ,j
1 2 One-step state transition probability from market price level scenario j1 to j2 .

58
Scenario dependent Variables

(k ) / (k ) Scenario k -dependent constant coefficients.

({ku),d } Scenario k -dependent regulation-up/down service call probabilities.

nk Probability of S n = k at stage n .

hj Probability of Lh = j in hour h .

nS Ex-post information of market participation at stage n .

hL Ex-post information of market prices in hour h .

Decision Variables

qh Vector of [qhE , qhU , qhD , qhS , qhN ]T with the components denoting generation

scheduled (GS) in energy market, regulation-up, regulation-down, spin reserve,

and non-spin reserve services in hour h , respectively.

qhR Total GS quantity in reservation services in hour h .

qhC Total GS capacity in hour h .

Decision Dependent Variables

Rh [ ]
Vector of RhE , RhU , RhD , RhS , RhN with the components denoting the expected

revenue from energy market, regulation-up, regulation-down, spin reserve, and

non-spin reserve services in hour h , respectively.


~
y h / yh Random number/expected value of MW generated in hour h .

Yn Accumulated MW generated at stage n .

() Revenue function over the whole planning time horizon or stage n if

augmented with subscript n .

Vector of Lagrange multipliers associated with the water target constraints.

59
Note that when augmented with the superscript ( k1 , k 2 ,L, k n 1 ) , ( j h ) , or both, the

above variables represent their values given scenario S i = k i , (i = 1,2, L , n 1) , Lh = j ,

or both have occurred. When augmented with subscript of a set of hours, the above

variables represent a vector of the variables indexed with each h in the set.

3.1 Introduction

In electricity market designs such as the one implemented in California, there often

exist multiple markets for selling and buying electric power. These markets typically

include forward energy markets, ancillary services (AS) markets, and real-time energy

markets. While the overall transaction volume of the ancillary services markets is

smaller than that of the energy markets, the revenue from selling ancillary services can

yield significant profit potential or cost reduction. A mature pool-based electricity

market offers participants the choice (and certain obligation) to participate in the ancillary

services markets besides the energy markets. To determine the best portfolio strategy for

selling electricity into these markets, market participants need to jointly optimize (or, co-

optimize) the allocation of electricity generation capacity dedicated to each market,

incorporating both price and operational uncertainties in the energy and ancillary services

markets. There has been a large amount of research on the co-optimization problem of

selling electricity into multiple markets given deterministic price forecasts. However,

much less literature is available on such a problem subject to uncertain price forecasts as

well as random service requests on the committed ancillary services capacity. This

chapter attempts to address this disparity by providing a stochastic co-optimization

framework for hydro-electric generation scheduling. The framework can be extended to

other applications such as optimal bidding of generation capacity in multiple electricity

markets subject to price and operational uncertainties.

60
While the electricity forward, ancillary services and real-time markets are similar in

the sense that they are all conducted through auctions, they differ in the types of products

offered for trading. Unlike the forward and real-time (instantaneous delivery) energy

markets which are for firm energy delivery, the ancillary services market is a forward

market for capacity with obligation to deliver energy only when called in real-time. The

need for ancillary services as a form of reserve capacity comes from the fact that it is

impossible to forecast system demand exactly and then purchase electricity for all the

customers ahead of time. Although the real-time energy market is also available for

load/generation balancing due to errors in forecasts (or strategic bidding), it is very

difficult to predict how much capacity would show up in real-time since there are no

forward obligations by the participants. In order to lock-in some reserves ahead of

time and thus ensure that at least some additional capacity is available in real-time to

balance the potential discrepancy between load and generation, the independent system

operator (ISO) pays a reservation price to all the participants who clear the AS market to

keep their capacity idle and available for generation increase (or load decrease) if needed

in real-time. In addition to the capacity reservation price, participants in the AS markets

also receive payments for energy generated in real-time at the real-time market clearing

price, which is determined by the aggregate energy bids in the real-time market.

Normally, market clearing prices in the energy and AS markets are formed from the

respective bids received. There are four major ancillary services (namely, regulation-up,

regulation-down, spinning reserve, and non-spinning reserve), which are differentiated by

the flexibility and response-time of service offered. Using the system frequency as the

control signal, regulation services are devoted to the continuous balancing of generation

resource and load to assist in maintaining normal system frequency. They are

accomplished by online synchronized generation capacity ready to respond to automatic

61
generation control signals by acting directly on the stream inlet valve. Regulation-up and

regulation-down services are procured separately. Spin and non-spin reserves (namely,

operating reserves) are prepared for purposes such as peak load shaving and security

maintenance in case of plant or transmission outages. Regulation up, spin and non-spin

all come from generation units operating at less than full capacity. Regulation-up is

replaced gradually by operating reserves if the system balancing requirements persist.

When simultaneously participating in the energy and AS markets, a critical issue that

market participants face on a daily basis is how to allocate their capacity between energy

and AS markets so as to maximize their profits or minimize their overall costs. In order

to be able to deal with this problem, the market participants need to be capable of

forecasting prices (and level of their volatility) in all of the energy and ancillary services

markets. While there has been significant research done in the area of price forecasting,

accurate energy price forecasting is still daunting at best. This suggests the need of a

rigorous yet practical stochastic co-optimization framework.

The ancillary service uncertainty is another important type of uncertainty borne by a

market participant. Ancillary service capacity is procured as insurance against an

unpredictable real-time imbalance between demand and supply. The regulation capacity

actually called in real-time is highly variable. However, we are not aware of any research

explicitly addressing the ancillary service uncertainty.

This chapter tackles the problem of obtaining a profit-maximizing capacity allocation

policy in the presence of energy and AS market uncertainties. Specifically, we consider a

profit-pursuing hydro-power producer participating in energy and AS markets as a price

taker. The co-optimization model of capacity allocation in multiple markets for this

hydro producer is formulated and solved utilizing stochastic programming, which gives

rise to an optimal scenario-dependent scheduling strategy instead of just a fixed trajectory

62
as is the case of deterministic optimization. We employ stochastic programming

techniques for hedging potential revenue loss against market uncertainties by explicitly

taking into account the energy price volatility and the AS request variability in the form

of regulation-up and down service call probabilities. The advantages of the stochastic

programming solutions over their deterministic counterparts have been demonstrated

over a wide spectrum of applications in (Birge, 1997).

The remainder of the chapter is organized as follows. Section 3.2 provides a

literature review. A deterministic multi-market hydro-generation scheduling model is

presented in Section 3.3. Sections 3.4 and 3.5 discuss the modeling of ancillary service

and market price uncertainties, respectively. In Section 3.6, a complete stochastic

programming model for the co-optimization problem incorporating both types of

uncertainties is presented. A case study with realistic data is carried out to illustrate the

advantage of the proposed stochastic formulation in Section 3.7 with output graphs and

tables. Section 3.8 concludes and outlines some future research directions.

3.2 Literature Review

Literature on stochastic co-optimization for hydroelectric generation scheduling is

scant. Most of the early works on the co-optimization problem adopt a deterministic

formulation and employ point estimates for the random variables.

A deterministic framework of scheduling in an integrated energy-reserve market is

proposed in (Chen et al., 2003). Li and Shahidehpour (2005) introduce a security-

constrained unit commitment model to achieve simultaneous optimization of energy and

ancillary services markets given market demand information. In the spot market, power

producers take on the responsibility of estimating the tradeoff between participating in

alternative markets and determine the optimal capacity allocation to each market in

response to time-varying market conditions over a given time horizon. Arroyo and

63
Conejo (2000) address the optimal response of a thermal unit to energy and spin reserve

spot markets assuming accurate market price forecasts. However, the applicability of the

results hinges on the accuracy of the price estimates. Profit opportunities can be lost if

the estimates deviate from the realized values significantly.

An overview of the application of stochastic programming techniques in energy

markets is provided in (Wallace and Fleten, 2003). The increased awareness of market

uncertainties calls for rigorous risk management of exposures to both ancillary service

and market price uncertainties. For a power producer, risk exposures are controlled by

putting proper financial instruments, such as futures contracts and flexible-load contracts,

into its portfolio. Mo et al. (2001) present a risk management approach that integrates

futures contract hedging and hydro generation scheduling in the energy markets. The risk

level can be controlled by setting prudent value-at-risk targets. Moreover, the operation

scheduling strategy can be expanded via more involved mathematical programming

techniques for hedging against uncertainties. Early applications of stochastic

programming techniques to unit commitment problems can be found in (Carpentier et al.,

1996; Jacobs et al., 1995; and Takriti et al., 1996). Carpentier et al. (1995) model the

uncertainty of demand and unit failures. Jacobs et al. (1995) introduces PG&E's optimal

scheduling system SOCRATES which coordinates hydro-generation with other energy

sources with respect to stream-flow forecasting models and other hydrological

information. Takriti et al. report a multi-stage stochastic programming model for unit

commitment considering demand uncertainty. Dantzig and Infanger (1997) present an

optimal intelligent control of powerhouse-flows and spill-flows in a hydro power system

with a network of river basins considering uncertainty in exogenous water availability

and electricity demand. Philpott et al. (2000) consider the problem of scheduling the

turbines in a chain of stations down a river valley subject to uncertain demand. Ozturk et

64
al. (2004) use chance constrained programming to ensure a high probability of satisfying

load. Ferrero et al. (1998) present a dynamic programming two-stage algorithm approach

to the long-term hydrothermal scheduling of multi-reservoir systems where stochasticity

of inflows is treated by considering a large finite sample of hydrological sequences. Only

in recent years has price uncertainty come to the attention of researchers. For example,

Plazas et al. (2005) study optimal bidding strategies for the day-ahead energy market and

automatic generation control (AGC) markets in the presence of price uncertainties.

Although the literature on thermal unit commitment is plentiful, the literature on hydro-

electric units is relatively limited. Conejo et al. (2002) study self-scheduling of cascaded

hydro generating plants along a river basin which sell energy in the day-ahead market.

However, the ancillary service uncertainty in the ancillary services markets, namely, the

real-time service call randomness, is ignored.

3.3 Deterministic Scheduling Model

In this section, we present a deterministic model. Similar versions of the model are

widely used in industry. We consider a hydro-electric power producer participating in an

electricity energy market and four ancillary services markets, namely, regulation-up,

regulation-down, spin reserve, and non-spin reserve. The energy market and the ancillary

markets clear simultaneously. The producer forecasts the hourly prices of each market

and determines the hourly capacity allocations correspondingly. Assume we have an


~
estimation of the hourly price process for the time horizon of interest, where

h = [ p hE , p hU , p hD , p hS , p hN , p hB ] , h H A
T

The revenue from each market is the product of the corresponding price and

committed capacity. Assuming equal incremental and decremental market prices, R h

can be determined as follows,

65
RhE = phE qhE

(
RhU = pUh + U phB qUh)
(
RhD = phD D phB qhD)
( )
RhS = phS + S phB qhS phS qhS

( )
RhN = phN + N phB qhS phS qhS

where S , N , U , and D denote transpired values of real-time call probabilities of spin

reserve, non-spin reserve, regulation-up, and regulation-down services, respectively.

Note that the revenues from ancillary services consist of both capacity payment and real-

time balancing energy payment. However, due to the fact that S and N , are typically

very small (less than 1% on average), the energy payment terms in spin and non-spin

reserve revenues are ignored in this chapter.

The objective of the optimization problem is to maximize the expected total revenues

defined in the following compact form,

(q H ) = p h T q h (3.1)
hH

where, p h = h (3.2)

1 0 0 0 0 0
0 1 0 0 0 U

= 0 0 1 0 0 -D (3.3)

0 0 0 1 0 0
0 0 0 0 1 0

The components of the vector p h correspond to settlement prices of the five

aforementioned markets. The operation and maintenance costs are ignored in the

objective function since they are relatively fixed and low compared with market revenues.

66
For scheduling of hydro units, the water supply is a major concern. We simplify the

problem by limiting the modeling to hydroelectric plants that are not hydraulically

coupled. The total MW generation potential within a time period depends on the

elevation difference between headwater and tailwater, reservoir inflows and outflows,

and/or contracted water flow limits. In the operational model of a hydro-electric

producer, we can envisage specification of weekly or monthly target values for water

flows, which are normally determined by some long-term hydro planning model. Such

water flow constraints are termed water target constraints herein. Because of small S
S
and N values, we can neglect q h and qhN in the expected MW generated in hour h and

define

yh = qhE + ~U qhU ~ D qhD , and


~

y h = E [y~h ] = q hE + U q hU D q hD = q h , h H A (3.4)

For each stage n , the water target constraint is imposed as

Tnl
{
y } Th
h H 1 ,L, H n
n
u
(3.5)

It ensures that certain reservoir levels or contract specifications be maintained during

the corresponding time period. Note that the Lagrange multiplier of the target

constraint actually reveals the marginal value (namely, shadow price) of the

corresponding water resource. Indeed, is one of the most important inputs for setting

optimal bid prices.

In light of the market rule against gaming behavior of double booking capacity, we

impose a cap of operating generation capacity on the regulation-down capacity as follow,

q hD q hE , h H A (3.6)

The total committed capacity in hour h is

67
qhC = qhE + qhU + qhS + qhN = h q h

out of which the capacity committed to reserve is

qhR = qUh + qhS + qhN = hq h

Note that qhD denotes downward commitment of capacity, it is not included in qhC nor

q hR above.

We also impose bound constraints on qhR , qhC , and q h to reflect operational

characteristics and leave capacity margin for both energy and operating reserve markets

as follows.

qhR qhR , qhC qhC , and 0 q h q h , h H A (3.7)

Without meaningful bounds, a producer may tend to dedicate most capacity to one of the

markets while leaving others under-attended. Such an allocation would not be favorably

considered as it could be viewed as manipulative from the market monitoring point of

view. A long-term market participant would let go potential market opportunities in

respecting these implied regulatory constraints. Note that q h can be quite different for

h H P and h H / H P due to different market conditions. For example, during off-peak

hours, most generation units are operated at low output levels, the upper bound of qhD is

set low since little or no regulation-down service capacity can be committed.

As in most other scheduling problems, uncertainty is a key ingredient of the

generation co-optimization problem. We do not have perfect information on all the

parameters such as those corresponding to ancillary services (~U , ~U ) and market prices
~
. The decisions based on (3.1) are questionable if the range of these random variables

is too large. A generalized stochastic formulation of (3.1) is as follow,

max E [ (q H , )] (3.8)
q H

68
where R N denotes the set of all feasible decisions which is defined by the constraints
~
(3.5-7), and denotes set of possible realization of , ~U , ~ D . [ ]
The deterministic formulation (3.1-3) is actually an expected value scenario of (3.8)

defined as follows,

( ) [[]
max q H , , where = E , E [~U ], E [~ D ]
q H
~
]
Usually, deterministic optimization leads to an inferior solution since embedded

randomness is disregarded altogether. According to Jensens Inequality, for any x and


( )
-concave function G(q H , ) , we have G q H , E [G (q H , )] , which leads to

( )
max q H , max E [ (q H , )] . Thus, the optimal value of the deterministic
q H q H

optimization is biased upward relative to the optimal value of the stochastic optimization,

due to the fact that the variability of is ignored.

The generation scheduling is a discrete-time control problem with fixed decisions

occurring at different points in time. The essence of resorting to stochastic programming

is to incorporate market risks into decision making. The scheduling problem is

essentially to strike a careful balance between using water now and using it at a future

point. While deterministic multi-period optimization yields decisions for all periods, a

stochastic approach yields policies and strategies. In stochastic programming with

recourse, recourse actions are taken after some uncertainty has been resolved. The set of

decisions is divided into periods before and after the realization of uncertainty. In our

problem, the sequence of events and decisions, when considering ancillary service

uncertainty at stage n only, is q H nS q{H


n n +1 ,LH N }. The corresponding optimization

problem corresponding to stage n is


Max
T
[ [
p H n q H n + E Max p{H n+1 ,LH N } ( ) q{H n+1 ,LH N } ( )
T
]] (3.9)

69
St. AQ H n = b

T ( )q H n + Wq{H n+1 ,LH N } ( ) = h( )

q H n 0 , q H n+1 ( ) 0

Where E denotes mathematical expectation with respect to , A , b , h , T , and W are

(scenario-dependent) matrices or vectors, specifically, W is called the recourse matrix

(Birge, 1997).

To solve (3.9) numerically, assuming a discrete distribution of random data with a

finite number (say K ) of possible realizations ( k ) with probabilities p(k ) , the stochastic

program can be written in the following extensive form to describe explicitly the decision

variables for each scenario,


K
p H n q H n + p(k )p{(kH)n+1 ,LH N } q{(kH)n+1 ,LH N }
T T
Max (3.10)
k =1

St. Aq H n = b

T (k )q H n + Wq {(kH)n+1 ,LH N } = h (k )

q H n 0 , q {(kH)n+1 ,LH N } 0 , k = 1,2,L, K

The recourse model transforms the randomness contained in a stochastic program into

some specific parameters according to some random distributions. Modeling details such

as possible set of outcomes or scenarios and the coarseness of the period structure must

be specified with great care so as to achieve an optimal trade-off between realism of the

optimization model, which affects the usefulness and quality of the obtained decisions,

and tractability of the problem, which is necessary for practical implementation.

The approach proposed in this chapter is based on probabilistic models of ancillary

service and price uncertainties. The objective function and constraints of the

70
corresponding mathematical programming models are defined based on market scenarios

of interest.

3.4 Stochastic Model with Ancillary Service Uncertainty

In hour h , the regulation services can be either called or not. According to (3.4), ~
yh
has the following possible values depending on which regulation service is called,

qhE + qhU + qhD if ~hU = 1, ~hD = 1


E
~ q + q h
U
if ~hU = 1, ~hD = 0
yh = hE (3.11)
q h q h
D
if ~hU = 0, ~ D = 1
h
qhE if ~hU = 0, ~ = 0
D
h

Note that even though it can happen that regulation-up and regulation-down services are

called in the same hour, it does not happen often. Therefore, we exclude this scenario

here. This mild assumption can be easily relaxed, albeit at the cost of higher dimensions

of the search space. By discretizing ~


y h according to (3.11) and following formulation
(3.9), we get the recourse actions corresponding to various ancillary service scenarios in

every hour. However, if we model all possible scenarios for each hour h , the number of

total scenarios throughout the whole planning horizon would be 4 H 1 , which is

astronomical even for a moderate value of H and, thus, impractical to handle. An

approximation is in order.

Corresponding to each stage n , we set


Yn ~y
hH n
h = (q
hH n
E
h + ~hU qUh ~hD qhD ) (3.12)

Assume that ~hU s and ~hD s are independent. For a realization of [q hE , q hU , q hD ] , h H n ,

given the large number of hours within stage n , we have, approximately,

Yn ~ Norm , 2 ( )

71
where, = E [Yn ] = (q
hH n
E
h + U qhU D qhD = ) q
hH n
h

2 = Var (Yn ) = U (1 U ) (qUh ) + D (1 D ) (qhD )


2 2

hH n hH n

Discretize Yn to K = 3 realizations which represent low, medium and high scenarios of

accumulated participation, respectively,

Yn(k ) = + (k )

where we set (k ) = k 2 to match the first two moments of Yn given

nk = Prob(Yn = Yn(k ) ) = 1 K . We assume Prob(S n = k nS1 ) = Prob(S n = k 1S , 2S ,L, nS1 )

= nk . According to this scheme, we have K n possible realizations of total capacity

commitment scenarios by the end of any intermediate stage n . Given the discretization

of Yn , in order to represent the expected regulation service revenues at stage n in

scenario k , which corresponds to discrete value Y n = Y n( k ) , we make ex-post expected

values of ~hU and ~hD for h H n as follows,

[ ] [ ]
E ~hU = U (k ) = U (k ) , and E ~hD = D (k ) = D (k )

where (k ) = Yn(k ) .

The ex-post expected value of market price vector p (hk ) corresponding to the realization

of scenario S n = k becomes

p (hk ) = ( k ) h (3.13)

1 0 0 0 0 0
0 1 0 0 0 U (k )

where, (k ) = 0 0 1 0 0 - D (k )

0 0 0 1 0 0
0 0 0 0 1 0

72
The scheduling strategy can be easily determined since the deterministic equivalent

formulation of the stochastic programming model is amenable to any standard

optimization solver. The overall formulation is as follow,

1 2 3 4
[(
Max E S , S , S , S q H1 , q(Hk12 ) , L , q(Hk1N,L, k N 1 ) )] (3.14)

K K
p (k )q ( k ) + L + K k p ( k )q ( k ,Lk )
= 1k p (Hk )q H + 2k
1 1 2

H H
2 1
N H H N N 1 N 1


N N

1 1 2 2
k =1
1 k =1 2 k =1 4

K
[ ] ( [ ] K
) ( [ ] )
K K
= E S p H 1 q H 1 + 1k1 E S p H 2 q (Hk12 ) + L + 1k1 2k 2 L Nk N11 E S p H N q (Hk1N,L, k N 1 )
1
k1 =1
2
k1 =1 k 2 =1 k N 1 =1 N

St. (h, n, k1 , k2 , L , kn 1 ) ,

n 1
Tnl Yi (k1i1 ) + y( h
k1 ,L, k n1 )
Tnu
i =1 hH n

qhD (k1 ,L, k n1 ) qhE (k1 ,L, k n1 )

qhC (k1 ,L, k n1 ) = hq(hk1 ,L, k i1 ) qhC

qhR (k1 ,L, k n1 ) = hq(hk1 ,L, k i1 ) qhR

0 q(hk1 ,L, k n1 ) q h

where, E S denotes mathematical expectation with respect to realization of nS , Y0 = 0 ,


n

and yh(k1 ,L, k n1 ) = q(hk1 ,L, k n1 ) .

Objective function (3.14) is readily extendible to the E [ ()] Var [ ()] form to

incorporate the mean-variance tradeoff, where > 0 is the risk-aversion coefficient. We

do not examine the mean-variance model in this chapter. It is reserved for future research.

3.5 Stochastic Model with Market Price Uncertainty

73
Significant price volatilities have been observed since the market de-regulation. In

this section, we focus on hedging against fluctuating market prices. We assume the

producer has a projected hourly market price process and, according to (3.2), the

market price process vector p in place. For example, in Southern California Edison, a

team of researchers is dedicated to price forecasting. A variety of commercial price

forecasting services are employed for validation purposes. Instead of developing a full-

fledged stochastic model for the price evolving processes, we model the price uncertainty

by considering a set of discrete price levels which are expressed in proportions of the

forecast and match the means and projected volatilities. The uncertainty is compensated

for by stochastic programming with recourse. For each hour h H A , we assume J

market settlement price scenarios with

p (hj ) = ( j )p h (3.15)
J
(
and corresponding probability hj with hj = 1 . Also assume that Prob Lh = j hL1
j =1
)
( )
= Prob Lh = j 1L , 2L , L, hL1 . The decision strategy is that for each realization of energy

price level scenario Lh 1 = j of hour h 1 , decision Q (h j ) is selected for hour h . A

Markovian process is employed for modeling the evolution of price level process. While

this Markovian assumption simplifies the complex energy price behavior, we concentrate

our effort on dealing with price fluctuations. To construct the price model, we choose
J J
( j ) in a way to keep j p (hj ) = j ( j )p h = p h and keep the correlation matrix of
j =1 j =1

market settlement prices intact. To simplify the formulation, we assume that the

decision strategy applies from h = 2 onward, meaning that the solution for the first hour

is not scenario-dependent. The objective function (3.1) of the optimization problem

becomes

74
p1( j1 )q1

p (2j2 )q (2j1 )
J

j1 J L
+ P(L2 = j2 L1 = j1 ) J
(3.16)
j1 =1
+
j2 =1

+ P(LH = jH LH 1 = jH 1 ) p (HjH )q (HjH 1 ) ( )


jH =1

Assuming the Markovian state transition matrix of the price level process being

[ ]
= j1 , j2 = Prob(L2 = j2 L1 = j1 ) , j{1, 2} = 1,2, L, J

with limiting probability vector

[ ]
J
= j , j = j ' j ', j
j '=1

the objective function (3.16) can be rewritten as follow,

J J J
j p1( j )q1 + j , j p (2j )q(2j ) + L + j
1 1 2 1
H 1 , j H
p (HjH )q (HjH 1 ) (3.17)

1 2
j1 =1 j2 =1 j H =1

or, equivalently,

H J J
q1 E [p1 ] + jh1 q (hjh1 ) jh1 , jh p (hjh ) (3.18)
jh1 =1
h=2 jh =1

Corresponding to the water target constraint (3.5), we impose

z hMin q (hj ) z hMax , j

z h
Min
h{H 1 ,L, H n }
Tnl , z h
Max
h{H 1 ,L, H n }
Tnu , n

where the introduction of z hMin and z hMax is to avoid the exponential growth of the number

of water target constraints with the number of time periods. The following constraints

corresponding to (3.6-7) also apply j ,

qhD ( j ) qhE ( j )

qhC ( j ) = hq (hj ) qhC

75
qhR ( j ) = hq (hj ) qhR

0 q (hj ) q h

3.6 Stochastic Model with Both Uncertainties

A complete stochastic model taking into account both ancillary service and market

price uncertainties is presented in this section. Corresponding to realizations of S n and

Lh , we define market price vector p (h j )(k ) = ( j ) (k ) h by combining (3.13) and (3.15).

Corresponding to the finite discretization of ancillary service and market prices

uncertainties, the probability assigned to a scenario node (S n = k , Lh = j ) is the product

of probabilities nk and hj . We have,

( ) ( )
K J K J
Prob S n = k , Lh = j L
h 1 = P Lh = j L
h 1 P (S n = k ) = k
j =1
n j h , j h
=1

k =1
1
k , jh k =1 jh =1 h

The following figure 3.1 illustrates the decision tree for a 4-stage complete model

considering both ancillary service and market price uncertainties with discretization

parameters K = 3 and J = 3 .

76
Figure 3. 1 Scenario tree with both uncertainties

The model aims to maximize

E L
HA
,1S , 2S , 3S , 4S
[ (q( j h 1 )
H1 , q(Hj2h1 )(k1 ) , L , q(Hj hN1 )(k1 ,L, k N 1 ) )] (3.19)


( ) ( ) ( )
K K
= 1 q(Hj1h1 ) + 1k1 2 q(Hjh21 )(k1 ) + L + Nk N11 N q (Hj hN1 )(k1 ,L, k N 1 )
k1 =1 k N =1


( )
J J
where, 1 q(hjhH11) = q1E [p1 ] + j h 1
q(hjh1 ) j h1 , j h q(hj h )

hH 1 j h 1 =1 j h =1

n (q(hj H )(k ,L, k ) =
J J
h 1 1 n 1 ) j h1
q(hj h1 )(k1 ,L, k n1 ) j h1 , jh p(hjh ) , n > 1
j h1 =1
n
hH n j h =1

St. (h, n, j , k1 , k 2 , L, k n 1 ) ,

( k1 ,L, k n1 ) ( k1 ,L, k n1 )
zhMin
H n q(hj hH1 n)(k1 ,L, k n1 ) zhMax
H n

n 1
Min ( k1 ,L, k n 1 )
Y (
i =1
k i 1 )
i 1 + z
hH n
h Tnl

77
n 1
Max ( k1 ,L, k n1 )
Y (
i =1
k i 1 )
i 1 + z
hH n
h Tnu

( j )( k1 ,L, k n1 ) ( j )( k1 ,L, k n1 )
qhD qhE

( j )( k1 ,L, k n1 )
qhC = hq(hj )(k1 ,L, k n1 ) qhC

qhR ( j )(k1 ,L, k n1 ) = hq (hj )(k1 ,L, k n1 ) qhR

0 q(hj )(k1 ,L, k n1 ) q h

3.7 Case Study

We consider a hydro-electric power producer controlling capacity up to 900MW

doing hourly multi-market generation scheduling for 1 month (30 days, 24 hours per day).

Corresponding to weekly and monthly targets, the whole time horizon is divided into

N = 4 stages with the first 3 weeks being the first 3 stages and the rest being the 4 stage.
th

Therefore, we have I 1 = 168 , I 2 = 336 , I 3 = 504 , I 4 = H = 720 . The MW-equivalent

water inflow target for the month is T N = 200000 , and for targets of intermediate stage

n , we set,

[T n
d
]
, Tnu = [1 n ,1 + n ]
In
IN
TN

where 1 = 10% , 2 = 5% , 3 = 2% , and 4 = 0 .

The rationale for the decreasing tolerance levels for target deviations is that the

monthly target is normally a hard target and there is more leeway for adjustment in the

beginning of the month than toward the end of the month. Actually, in reality, small

deviation (1% to 2%, and up to 5% in emergency cases) is allowed for monthly targets.

However, for scheduling purposes, monthly targets are normally treated as hard targets,

which explains why we have 4 = 0 .

78
Using Southern California Edisons historical (June - August 2005) market prices

data, we calibrate the joint distribution of market prices as follows. The price vector p h

has a linear dependence structure measured by the following correlation matrix

1.0000 0.4397 0.7768 0.6451 0.6413


0.6099
0.4397 1.0000 0.2641 0.5133 0.2261
0.4888

0.7768 0.2641 1.0000 0.3905 0.3619 0.5341
=
0.6451 0.5133 0.3905 1.0000 0.7250 0.4311
0.6099 0.4888 0.3619 0.7250 1.0000 0.4052

0.6413 0.2261 0.5341 0.4311 0.4052 1.0000

with triangular marginal distributions Tri ( T , T ) where T and T denote the mean

and width of the support, respectively. The parameters are listed in the table 3.1,

Table 3. 1 Parameters of Price Distributions (in $/MWh)


On-Peak (h H )P Off-Peak (h H / H P )
T T T T
p hE 78.42 28.29 44.99 25.65
p hU 25.30 13.71 17.80 9.62
p hD 7.67 4.52 23.09 16.26
phS 29.63 19.96 14.79 7.49
p hN 24.72 15.92 7.81 2.22
p hB 57.73 26.95 39.48 24.99

The expected call probabilities of regulation services are assumed to be

U = D = 0.3 , and the ad hoc capacity constraints and the regulatory constraints are:
U U D D
qHR P = 200, qHR / H P = 450, q H P = 150, q H / H P = 300, q H P = 200, q H / H P = 50

To address the uncertainty of regulation service calls, we set K = 3 and discretize Yn

as described in section 3.4. To capture the uncertainty of market prices, we set J = 7

and consider a set of discrete price levels of phE

LE = [0.85 0.90 0.95 1.00 1.05 1.10 1.15]


T

79
Considering the co-movement of ancillary service prices, the following discrete price

level matrix is calculated with each row corresponds to each component of p h

LE 0.850 0.900 0.950 1.000 1.050 1.100 1.150


LU 0.815 0.971 1.073 1.011 1.169 0.979 0.982

LD 1.120 1.030 1.112 0.939 0.965 1.026 0.809
L= S=
L 0.816 0.952 1.132 0.982 0.963 1.039 1.115
LN 0.786 1.075 1.023 0.954 0.994 1.113 1.055
B
L 0.945 0.974 0.975 0.837 1.071 1.017 1.183

Correspondingly, diagonal matrix ( j ) can be determined with its diagonal components

being the j th column of L .

Note that the co-movement follows the correlation matrix and the standard

deviation of each market price is assumed to be of the same percentage of the mean.

We assume the Markovian price level process to be mean reverting and set the state

transition matrix to be

1
14 (j
1 3)(j2 3) > 0
1
= [ j , j ], j , j
1 2 1 2
= (j
1 3)(j2 3) = 0
7
3
14
(j
1 3)(j2 3) < 0

with the limiting probability vector being = [ j ], j = , j = 1,2,L, J


1
7

In the presence of uncertainties, it is impossible to find a solution that is ideal under

all circumstances. Using the above parameters and probability measures, we model and

solve the co-optimized hydro scheduling problem in stochastic formulas as described in

sections 3.4, 3.5, and 3.6 (referred to as models S1, S2, and S3 hereinafter) and compare

them to the deterministic model given in section IV (referred to as model D).

With the price process parameters given above, we simulate a projected price process
~
h for h H . We solve the models D, S1, S2, and S3 using GAMS and obtain their

80
solutions. An immediate comparison between the models D and S1 is to plot their

respective expected revenues with respect to U and D . Since model S1

accommodates the uncertainty of regulation service calls, as shown in figure 3.2, its

expected revenue derived is less sensitive to that of the model D.

Figure 3. 2 Sensitivity of expected revenues with respect to U and D

To show the advantages of stochastic programming models which imply the hedging

strategies with respect to uncertainties, we resort to Monte Carlo simulation as follows.

By generating Bernoulli random variables ~hU and ~hD for h H A to simulate the

regulation service calls. The realized revenue according to the generation scheduling

solutions of models D and SI can be calculated. Note that the accumulation of ~


y h in
each stage determines the choice of scenario-dependent S1 solutions in the next stage.

Repeating the simulation for 200 times, the realized revenues are plotted in figure 3.3.

Note that the S1 solutions lead to higher revenues under most if not all circumstances.

81
21 Deterministic Model
Stochastic Model
20.5

Revenue ($ in Millions)
20

19.5

19

18.5

18

17.5
20 40 60 80 100 120 140 160 180 200
Simulation Path ID

Figure 3. 3 Comparison of simulated revenues with regulation service call uncertainty

Similarly, by simulating the market price level transition process with random starting

level at h = 1 , the realized revenue according to models D and S2 solutions are compared

and shown in figure 3.4,

20.5 Deterministic Model


Stochastic Model

20
Revenue ($ in Millions)

19.5

19

18.5

18

20 40 60 80 100 120 140 160 180 200


Simulation ID

Figure 3. 4 Comparison of simulated revenues with market price uncertainty

It is observed that the model S2 solution consistently outperforms the model D

solution.

To mimic the contemporary uncertainties of both regulation service call and market

prices, we simulate the regulation service call process and market price level transition

82
simultaneously and plot the realized revenues according the model D and S3 solutions in

figure 3.5 as bellow.

20.5 Deterministic Model


Stochastic Model

20

Revenue ($ in Millions)
19.5

19

18.5

18

20 40 60 80 100 120 140 160 180 200


Simulation ID

Figure 3. 5 Comparison of simulated revenues with regulation service call and market
price uncertainties

The advantage of our stochastic programming approach is clearly demonstrated by

the comparison of the revenue distributions resulting from the stochastic programming

approach and the deterministic optimization approach as shown in figure 3.6,

Deterministic Model
3
Stochastic Model

2.5
Probability Density

1.5

0.5

0
18 18.5 19 19.5 20
Revenue ($ in Millions)

Figure 3. 6 Comparison of revenues PDF of deterministic and stochastic solutions

83
The minimum variance unbiased estimates of the parameters of the normal

distribution N ( , ) and the 95% confidence intervals (CI) of the simulated revenue by

the deterministic and the stochastic models are displayed in the following table 3.2,

Table 3. 2 Estimates of Normal Distribution Parameters (in Million $)


Model 95% CI of 95% CI of
Deterministic 18.24 [18.22, 18.27] 0.18 [0.17, 0.20]
Stochastic 19.61 [19.59, 19.63] 0.17 [0.15, 0.19]

The comparison reveals a high than 7% improvement in the realized revenue on

average. To a real market participator, it may translate to a significant multi-million

revenue gain due to the large economic value associated.

To study the effect of considering more price levels, we set J = 13 and set the

discrete price levels of phE being

LE = [0.834 0.861 0.889 0.917 0.945 0.972 1 1.028 1.055 1.083 1.111 1.139 1.166] .
T

Note that more intermediate levels are considered compared with setting J = 7 , and the

price levels are adjusted to keep the hourly price volatility level unchanged. Considering

the co-movement of ancillary service prices, the following discrete price level matrix is

calculated with each row corresponds to each component of p h

LE 0.834 0.861 0.889 0.917 0.945 0.972 1 1.028 1.055 1.083 1.111 1.139 1.166
LU 0.915 1.048 0.691 1.002 1.036 1.089 1.048 1.028 1.012 0.928 1.076 1.020 1.107

LD 1.119 1.064 1.027 1.061 1.071 1.043 1.020 1.006 1.073 0.925 0.981 0.908 0.700
L= S =
L 0.837 1.051 0.864 0.978 0.932 0.948 0.981 0.992 0.997 1.061 1.277 1.044 1.039
LN 0.987 0.856 0.857 1.021 0.972 0.967 0.991 0.992 0.985 0.970 1.294 1.016 1.092
B
L 0.756 0.901 1.035 1.050 0.924 0.936 1.048 1.081 1.100 0.923 1.084 1.002 1.160

The Markovian price level state transition matrix is changed, correspondingly, to be,

1
26 (j
1 6 )(j2 6 ) > 0
1
= [ j , j ], j , j
1 2 1 2
= (j
1 6 )(j2 6 ) = 0
13
3
26
(j
1 6 )(j2 6) < 0

84
with the limiting probability vector being = [ j ], j = 1 , j = 1,2,L,13 .
13

Apply the parameters above to the stochastic formulation S2 which considers market

price uncertainty, and formulation S3 which considers both regulation service and market

price uncertainties. And repeat the procedures above to simulate regulation service calls

and market price levels and calculate the realized revenues accordingly. Assuming

normality of the revenues calculated, the statistics are illustrated in table 3.3 as follows.

Table 3. 3 Estimates of Normal Distribution Parameters (in Million $)


S2 S3
Price Level

J =7 19.22 0.079 19.61 0.17
J = 13 19.34 0.087 19.72 0.18

Note that the intensified modeling of market price levels can increase both expected

values and volatilities of generated revenues.

3.8 Conclusions and Discussions

This chapter presents a co-optimization modeling framework that incorporates market

participation and market price uncertainties into the hydroelectric generation capacity

allocation decision-making problem through a stochastic programming formulation.

Compared with a deterministic model, this approach yields scenario-dependent

scheduling strategies which effectively hedge the potential revenue loss cause by

unfavorable market price changes and the ancillary service requests. The computational

study with realistic data provided by a hydroelectric producer illustrates the advantages of

the proposed approach. Since the effectiveness of the scheduling strategies depends on

the assumed distributions of the uncertain variables, with the evolving of market

conditions, these distribution parameters should be updated continuously in application.

The proposed formulation can be adapted for determining optimal strategies to

85
accommodate other sources of uncertainties such as unplanned unit maintenance and

water inflow.

Several potential extensions of this model deserve further investigation: To

incorporate the market price uncertainty without a projected hourly market price process,

an improvement can be made to augment the model described in section 3.5 by using

correlated continuous-time stochastic processes to model energy and ancillary service

market prices. Impact of a producers decisions on the market prices, especially on the

ancillary service prices, can be modeled to make the model more realistic. Another

fruitful direction is to extend the model to address the co-optimization problem for a

power producer who owns a portfolio of hydro, thermal and other types of generation

units.

86
CHAPTER 4

DAY-AHEAD FORWARD RISK PREMIUM IN ELECTRICITY

MARKETS WITH TRANSMISSION NETWORK CONSTRAINTS

This chapter tackles the issues of optimal hedging and equilibrium price discovering

for market participants in the electricity day-ahead forward and spot wholesale markets

considering transmission capacity constraints. The equilibrium prices in day-ahead

forward and spot markets are derived by formulating the individual market participants

decision problems and solving them with respect to the market clearing conditions.

Consequently, the locational forward risk premium, defined as the difference between the

day-ahead forward price and the spot price, is expressed as a function of network

topology, shadow prices of transmission flowgates, and other economic measures of the

market participants. The implications of the model are illustrated through a series of

numerical experiments with a three bus study-system. Empirical studies with New York

electricity market data indicate that the forward risk premium determining model

captures key economic features based on the hypothesis that the market prices are

determined by rational risk-averse market participants.

The following notations apply to this chapter.

Acronyms

{AC, DC} {Alternating, Direct} current

GEN Generator

ISO Independent system operator

LBMP Locational based marginal price

87
LSE Load serving entity

MW Mega watts, unit for real power, the rate of energy consumption

MWh Mega watts hour, unit for energy consumption

NYISO New York independent system operator

NYEM New York electricity market

PTDF Power transfer distribution factor

Parameters of the transmission network

F Matrix of PTDFs, F = [ f n ,k ] , where f n , k denotes the proportion of power flow

on flowgate k for one unit of power injection into bus n paired with one unit

of power ejection out of slack bus.

gi Bus index of where the i th GENs units are connected. We assume that

different GENs do not compete at the same location, i.e. if i i ' , then g i g i ' .

lj Bus index of where the j th LSE is located. We assume that all LSEs have

non-overlapping franchised service area, i.e. if j j ' , then l j l j ' .

Vector of transmission capacity of flowgates, T = [Tk ] , k K .


T
T

Set of buses, = {n}, n = 1, L N .

K Set of flowgates, K = {k }, k = 1, L K .

Parameters, variables, and functions associated with market participants

{A i
G
, A jL } Risk aversion coefficients of the i th GEN or the j th LSE, assume that

AiG = AG > 0 , i , and A jL = AL > 0 , j .

C () Generation cost function.

Vector of bus-aggregated demand of end consumers, D = [d n ] , n .


T
D

Fi Fixed cost of the i th GENs generation cost function.

88
Vector of forward/spot positions taken by GENs, G {F , S } = [Gi{F , S } ] , i = 1,L, I .
T
G {F ,S }

G {F ,S }* Vector of optimal forward/spot positions of GENs,

G {F ,S }* = [G i{F ,S }* ] , i = 1,L, I .
T

Vector of total positions by GENs, G = [Gi ] , i = 1,L, I , and G = G F + G S .


T
G

i Index of GENs, i = 1,L, I .

j Index of LSEs, j = 1,L, J .

L{F , S } [ ]
Vector of forward/spot positions taken by LSEs, L{F , S } = L{jF , S } , j = 1,L, J .
T

Vector of optimal forward/spot positions of LSEs, L{F ,S }* = [L{jF ,S }* ] , j = 1,L, J .


T
L{F ,S }*

L [ ]T
Vector of total positions by LSEs, L = L j , j = 1,L, J , and L = LF + LS .

U {G ,L } () Utility function of a GEN/LSE.

{ , }
i
G L
j GEN i or LSE j s profit without forward hedging.

{{GF,,LS}} () Profit function of a GEN/LSE in forward/spot markets.


iG Parameter of GEN i s generation cost function, iG > 0 , assume iG = , i .

iG Parameter of GEN i s generation cost function, iG > 0 .

A random variable representing the market uncertain which leads to load

randomness and market price fluctuation.

Electricity market prices and shadow prices associated with transmission flowgates

Vector of the LMPs in forward/spot markets, P {F ,S } = [Pn{F ,S } ] , n .


T
P {F , S }

Vector of electricity retail price for end consumer, P R = [PnR ] , n .


T
PR

Pref{F , S } LMP at the reference bus in forward/spot markets.

{F , S } Vector of flowgate shadow prices in forward/spot markets, {F,S } = {kF , S } . [ ]T

89
4.1 Introduction

The deregulation of the electric power industry separates the production and delivery

of electricity as different market making functions from the previously vertically

integrated monopoly. Wholesale electricity markets and retail electricity markets are

established to allow trading between bulk suppliers, retailers and other financial

intermediaries. As an essential commodity for national economics and household uses,

electricity is traded in large quantities between these market participants. With an annual

consumption of over 3.6 billion MWh, the US power market is becoming one of the most

active commodity markets in the world. In the first few years following the conception

of deregulation, the complexity of electricity trading and the market risks were more or

less underestimated due to continuations of the power transaction contracts

predetermined in the regulated regime. Until the late 1990s, electricity marketers had

enjoyed substantial latitude in capital markets. However, the turmoil of the California

power market in 2000-2001 and the collapse of the energy giant Enron in late 2001

created unease and reminded market participants of the harsh realism in the rapidly

developing industry. The electricity trading business experienced dramatic impact

although it was not the exclusive driver to such crisis. In 2002, stock depreciation and

credit downgrades haunted the industry and increased the cost of capital access. The

market perceptions of electricity commodities and derivatives trading associated risks

came to the close attention of market participants. Stakeholders and potential investors

began to appreciate the related research.

As indicated in chapter 1, the imbedded physical nature of electricity determines the

extreme volatility of its price and the illiquid of its trading. The following figure 4.1

illustrates the day-ahead (left panel) and spot (right panel) market hourly electricity prices

90
on the reference bus in the New York power pool over the period from February 2005

through August 2006.


Forward Market Spot Market
800 800

600 600

400 400
Price ($/MWh)

Price ($/MWh)
200 200

0 0

-200 -200

Mar2005 Oct2005May2006 Mar2005 Oct2005May2006


Date Date

Figure 4. 1 Hourly day-ahead forward and spot prices on the reference bus in NYEM

From figure 4.1, electricity price spikes are observed frequently especially in the spot

market. In contrast to the general observation that in most commodity markets derivative

prices are more volatile than the fundamentals, figure 4.1 shows the opposite: electricity

prices in the day-ahead forward market are generally less volatile than in the underlying

in the spot market. Actually, the standard deviation of percentage changes of the hourly

electricity prices as displayed in figure 4.1 is 10.7% in the day-ahead forward market (the

left panel) and 121.7% in the spot market (the right panel), respectively. Compared with

the stock market, the standard deviation of daily returns on the S&P 500 index during the

most volatile single month in recent decades, October 1987, was 5.7% (Bessembinder

and Lemmon, 2002). The extreme price volatilities put market participants at significant

risk exposure in the spot market, which may lead to serious financial difficulties. It lead

to the future trading of bulk electricity. Soon after the conception of the electric power

industrys deregulation, the market for electricity futures emerged in the New York

mercantile exchange in March 1996. Financially settled monthly futures contracts for on-

peak and off-peak electricity transactions provide a channel for risk transferring between

91
market agents who have different market projections and risk preferences. Nowadays,

standardized futures contract trading for electricity delivery at specific locations provides

sophisticated bulk electricity suppliers and buyer opportunities to hedge against

unfavorable price fluctuations. Options on the monthly futures have been structured and

offer market agents additional trading opportunities and risk management tools.

Compared with electricity spot prices which clear the transactions in the spot market,

electricity forward prices clear the forward transactions and reflect the pricing of the

corresponding forward contracts. There usually exist forward risk premiums which

measure the difference between the forward and expected spot prices. Economic theories

suggest that forward premiums represent market compensation to financial market

participants for bearing the systematic risk. Although classical literatures suggest that a

typical forward premium is negative due to the systematic hedging-pressure effect, recent

literature (e.g., Longstaff and Wang, 2004) provide examples of positive forward

premiums. Routledge et al. (2000) develop an equilibrium model of term structure of

forward prices for storable commodities. In their model, a competitive rational

expectations model of storage is employed to study the impact of the embedded timing

option of spot commodity which is absent in the future contracts. However, the results

do not necessarily extend to electricity futures. The cost-of-carry relationship links spot

and forward prices by the no arbitrage condition. A forward contract can be synthesized

by taking a long position in the underlying asset and holding it till the maturity of the

contract. Since there lacks efficient technology to store electricity economically, the

buffering benefit of holding inventory to smooth temporary demand/supply imbalance

does not exist. Consequently, in addition to the resulting extreme intertemporal and

interspatial price volatility, the standard arbitrage-based approach cannot be applied as it

is to value electricity derivatives. Therefore, electricity forward prices do not necessarily

92
conform to the cost-of-carry relationship (Eydeland and Geman, 1999; Pirrong and

Jermakyan, 1999), and the no-arbitrage approach does not apply to power derivatives

valuing directly. Pirrong and Jermakyan (1999) note that electricity forward prices differ

from expected delivery date spot prices due to an endogenous market price of power

demand risk. However, no attempt to explicitly model the determinants of the forward

market price risks is made. Eydeland and Geman (1999) focus on electricity options and

note that as a consequence of non-storability, delta hedging involving the underlying

asset cannot be implemented and using the spot price evolution models for pricing power

options is not helpful. Their option pricing model relies on assumptions regarding the

evolution of forward power prices instead. A recent study of forward price premiums in

the natural gas market (Borenstein et al., 2006) argues that since gas local distribution

companies (LDCs) choose to pay to guarantee access to gas at times when gas

transportation capacity is expected to be constrained, positive forward price premiums

exist. However, this conclusion is back up by the fact that LDCs face a much higher

penalty for inadequate gas input than for storing excessive gas. Inspired by the fact that

storable natural gas can be converted into electricity in industrialized size, Routledge et al.

(2001) tackle the equilibrium pricing of electricity contracts by studying the spark

spreads between the natural gas and electricity markets. Their research indicates mean

reversion and positive skewness of electricity prices. Also revealed is the unstable

correlation between electricity and natural gas prices. These literatures suggest relying

on equilibrium-based models to derive how electricity forward prices are related to spot

market conditions. According to the general equilibrium model of electricity forward

prices proposed by Bessembinder and Lemmon (2002), there exist a functional

relationship between electricity forward prices and the expected underlying spot price

and load demand. The forward price is a biased forecast of the future spot price, with the

93
forward premium being a decreasing function of the expected variance of the wholesale

spot price and an increasing function of the expected skewness of wholesale spot prices.

More specifically, the forward premium is negative (termed normal backwardation) when

price volatility over-dominates the positive skewness effect, and positive (termed

contango) otherwise. Generally, it shows that fundamental economic factors and market

participants risk attitude determine the forward premium. Following the equilibrium

model proposed by Bessembinder and Lemmon (2002), Longstaff and Wang (2004)

implement empirical studies using high frequent electricity price data from the PJM

(Pennsylvania, New Jersey, and Maryland) electricity market. Their study confirms the

relationship between forward risk premium and the moment statistics of spot prices and

finds that the forward premiums are higher in peak hours compared with non-peak hours.

In addition to the markets for long-maturity futures, in many U.S. power pools, a two-

settlement mechanism has been established which involves the market clearing in a day-

ahead forward market and a real-time spot-market. Assisted with market monitoring

which deters market speculation behaviors, the creation of the day-ahead forward market

allows the system operators to collect information and make more accurate projections of

the next-days demand and available generation supply. The opening of the day-ahead

forward markets also moderates the impact of the otherwise more volatile spot market

prices and provides market participants a channel to hedge the market risks involved.

Since the establishment of the two-settlement mechanism, the day-ahead forward markets

have grown rapidly. The price dynamics in the day-ahead forward markets and the

relation between day-ahead forward and spot prices have become one of the major

concerns of market participants.

Although the model proposed by Bessembinder and Lemmon (2002) provides an

economic characterization of forward risk premium for monthly settled futures contracts,

94
it implicitly assumes single forward and spot prices apply to every market participant

within the broad delivery locations of their transactions with the transmission congestion

related price dynamics being smeared over the model parameters. However, if applied

directly to the day-ahead forward market, the same model would over-simplify the

realism since limited transmission capacity lead to different LBMPs at different locations.

In addition to the supply-demand balances, market prices are vulnerable to network

constraints. Since physics laws determine the transmission of electricity power flows,

given the limited transmission capacity of transmission network, diversified prices

exhibit at different locations even within a market. The following figure 4.2 illustrates

the day-ahead forward (left panel) and spot (right panel) market hourly LBMP

differences between the zone CENTRL and the reference bus in NYEM over February

2005 through August 2006,


Forward Market Spot Market
30 30

20 20

10 10
Price ($/MWh)

Price ($/MWh)

0 0

-10 -10

-20 -20

-30 -30
Mar2005 Oct2005May2006 Mar2005 Oct2005May2006
Date Date

Figure 4. 2 Day-ahead forward and spot price differences between zone CENTRL and
the reference bus in NYEM

It illustrates the effect of network congestion resulting from limited transmission

capacity. Note that spatial price differences make significant percentages of the market

prices, especially in the spot market. Actually, the spatial price differences between the

zone N.Y.C and the reference bus are more dramatic due to the tighter transmission

95
capacity connection. Consequently, to understand the relationship between the day-ahead

forward LBMPs and the spot LBMPs in a practical system, the model presented in

(Bessembinder and Lemmon, 2002) needs to be expanded to incorporate transmission

network constraints and model the interspatial price diversity.

The research work presented in this chapter intents to model the pricing of electricity

forward contracts in the day-ahead market while considering the impact of transmission

congestion. We assume that in both the day-ahead and the spot markets, market clearing

prices are determined by industry participants rather than outside speculators. We also

assume that electricity bulk suppliers and retailers are risk averse and pursue risk-

adjusted profits. An equilibrium model is presented which leads to a regression model

which implies that the forward risk premium can be expressed as a function of network

topology, shadow prices of transmission flowgates, and other economic measures of the

market participants. We conduct preliminary empirical analysis using historical

electricity prices and transmission congestion information form the New York electricity

wholesale market. The data used in this study consist of hourly day-ahead and spot

electricity prices, hourly day-ahead and spot flowgate shadow prices, and hourly load

level for the period from February 2005 to August 2006. The rational for using hourly

data instead of daily-averaged data, as mostly adopted in other literatures, is that the

occurrences of transmission congestions usually last no more than a few hours instead of

days. The congestion effect may become smeared if it is scaled into daily averaged data.

The observations from the empirical study are in general supportive of the relation

indicated in the proposed model. They confirm that the regression model for day-ahead

forward risk premium in power prices captures many of the key economic features based

on the hypothesis that the market prices are determined by rational risk-averse market

participants. This research contributes to the rapidly growing literature on electricity

96
market risks and is valuable for leading to broader insights regarding market perceptions.

Although the equilibrium model considers electricity markets exclusive characteristics,

the modeling framework and the insights obtained apply to other non-storable commodity

markets as well.

The rest of the chapter is organized as follows. Section 4.2 describes the general

setup of the day-ahead forward and spot markets for electricity wholesale. The decision

problems of each market participant in the day-ahead and spot markets are formulated.

Section 4.3 discusses the market clearing conditions and solves for economic

determinants of market clearing prices. By maximizing each market participants risk-

adjusted profit with respect to market clearing conditions, the model gives closed form

solutions for the equilibrium forward prices and optimal forward positions. The

implications of the model and impacts of various system parameters on the forward risk

premium are illustrated through a series of simulation studies with a 3-bus study system.

A regression model for day-ahead forward risk premium, which incorporates

transmission network constraints and various economic measures, is derived in section

4.4. Empirical statistical studies with New York electricity market data are conducted.

The observations indicate that the proposed model captures many of the key economic

features which determine the forward risk premium. The model is then used for spot

market price prediction and tested against out-of-sample historical data. Finally,

conclusions and directions for future research are presented in section 4.5.

4.2 Decision Problems of The Market Participants

In an open electricity pool market, the transactions of electric power are primarily

driven by market participants business incentives. End consumers activities determine

the total demand. Generators and load serving entities choose to sell or procure on either

day-ahead forward or spot markets and at what quantities. They interact continuously

97
through projected and real-time electricity transactions. There exist an independent

system operator t monitors the transmission network status and dispatches the system to

minimize the social energy acquisition cost. The winning electricity selling and buying

bids determine the market clearing prices consequently. However, the objective of each

market participant is to maximize the individual risk adjusted profit. At the market

equilibrium, each market participant cannot gain profit by making unilateral moves.

Since the cost-of-carry models cannot be readily applied to price electricity

derivatives, an equilibrium approach is adopted instead to assess equilibrium prices and

optimal hedging positions for market participants in the day-ahead forward and spot

markets. To model the two-settlement mechanism, we consider a single forward time

period for electricity wholesale transactions when market participants choose their

forward market positions by estimating the spot market conditions. When it comes close

to the settlement of spot market, electricity load, the fundamental factor of the market

uncertainty, can be estimated with substantial precision in the immediate future. A same

assumption in (Bessembinder and Lemmon, 2002) is made that market participants can

make decisions regarding spot market positions at precisely projected spot prices.

This section presents the decision problems of the major market participants.

Contingency events such as outages of generation units and transmission facilities are

ignored here and reserved for future researches.

4.2.1 System Operator

Assume an electric power system with the transmission network consisting of a set

of buses and a set K of monitored transmission flowgates. In our analysis, three

markets are considered as follows:

a. A day-ahead forward wholesale market

98
In the day-ahead wholesale forward market, electricity power suppliers (i.e., GENs)

and buyers (i.e., LSEs) choose their forward positions G F and LF , respectively. The

market clears at the day-ahead locational marginal prices P F .

b. A real-time spot wholesale market

In the wholesale spot market, the suppliers and buyers chose their spot positions G S

and LS , respectively, while the market clears at the spot prices P S .

c. A regulated retail market

The load serving entities act as electricity buyers in the wholesale markets and sell

electricity procured from the wholesale markets to the end consumers in a retail market.

Although LSEs have to buy bulk electricity at time-varying competitive market clearing

prices P F or P S in the day-ahead forward or spot market, respectively, the retail prices

to end consumers are under regulation and set fixed at P R . Usually, the regulated retail

price P R to end consumers is set based on the expected spot price P S plus a certain retail

profit margin.

Suppose there exist I generators acting as power suppliers selling electricity into the

competitive wholesale markets with J load serving entities being the wholesale buyers

and suppliers in the retail market. To ensure the economical efficiency of the wholesale

market, prevent discriminatory access to the transmission network, and encourage open

markets competitions, an independent system operator with central dispatch authority

coordinates the transmission of electricity to facilitate the transactions while monitoring

against the violation of transmission network capacity. Locational marginal energy

prices P {F ,S } and shadow prices {F , S } associated with the congested flowgates are

determined by the system dispatch and informed to the public. The aggregated electricity

demands in the LSEs franchised service territories are denoted by a vector of exogenous

random variable D , which appear as firm obligations to load serving entities.

99
The cost function of the i th generator is assumed to be a quadratic function of the

generation output Gi ,

iG
C (Gi ) = Fi + Gi
2
(4.1)
2 iG

In the model proposed in (Bessembinder and Lemmon, 2002), to account for the missed

market price dynamics do to ignored transmission congestion, the term of Gi in the cost

function is assumed to be higher than 2 in general. In the contrast, the transmission

network constraints are modeled explicitly in the proposed model, we use quadratic

functions to represent supply curves, which are, according to (Wood and Wollenberg,

1996), more realistic characterizations of the generation production.

The market participants generally optimize market decisions expecting to reach high

profit on average. Suppose that the information can be viewed as a random variable

with support and known probability distributions, for example, with cumulative

distribution function F ( ) , the profit on average can be defined as

E [ ( )] = ( )dF ( ) . However, for a particular realization of the system conditions,


the profit resulted could be very different from the corresponding expected value. The

deviation may be quite considerable if the underlying randomness has a large variance.

Relying on the corporate risk hedging literatures which argue that market agents can

benefit from hedging market risks by avoiding suboptimal decisions, and given the

extreme volatility of the electricity wholesale prices, electricity market participants are

modeled as risk averse and willing to seek hedges in forward market against adverse real

time price changes, that is, they take both expected profits and the imbedded volatilities

into consideration. A utility function linear in expected value and variance of profit is

chosen as follow to reflect their risk aversion,

100
A{G , L}
E [U {G , L} ] = E [ {G , L} ( )] Var ( {G , L} ( )) (4.2)
2

where A{G , L} 0 represents the weight given to the conservative part of the decision.

To simplify the problem, the forward and spot wholesale markets are modeled as

closed systems where market participation is limited to generators and load serving

entities with the prices determined by their physical delivery backed electricity trades.

The rational for this assumption is that currently the activities of market agents, who take

financial positions in the forward market and offset their positions prior to the next day

delivery, are under strict regulation. Actually, due to the lack of electricity price indices

(Ong, 1996), the speculative power transactions only account for a very limited

percentage compared with the overall volume. Although no explicit model is build to

represent market speculation, it is straightforward to envision that the unlimited costless

participation of risk-neutral speculators would drive forward price converge to the

expected spot price. Nevertheless, the existence of nonzero electricity forward premiums

as observed from the empirical data provides incentives for financial intermediaries to

create instruments, for example, power-indexed bonds, to allow outside speculators to

include power positions in their portfolios (Bessembinder and Lemmon, 2002). The

presence of outside speculators would be essential to market prices.

As mentioned in previous chapters, a major difference of electricity from other

commodities is that electricity flow over the transmission network following the law of

physics instead of flowing freely as envisioned from an economic perspective. Since

electricity is priced at each distinct location, the existence of network constraints presents

complications to the electricity markets. Since we are concerned with the real power

only, we use the DC instead of AC power flow equations based system dispatch model in

our analysis. The DC model identifies thermal limits constraints and eases the

101
understanding of economical issues in congestion management. We have the ISOs

system dispatch problems with network constraints as follows, note that the market prices

can be determined based on the dispatch decisions.

Definition 4.1: (ISOs system dispatch problem in the day-ahead forward wholesale

market) The ISO collects the load serving entities electricity demands in the day-ahead

forward market and dispatches the available generation resources accordingly to meet the

forward demand while minimizing the forward electricity energy acquirement cost. By

clearing the supply and demand at each location with respect to the transmission capacity

constraints on each monitored flowgates, the market clearing day-ahead forward location

marginal prices and flowgate shadow prices are determined by solving the following

optimization problem,

( ) ( )
I
Min C G F = C GiF (4.3a)
i =1

I J
St. GiF = LFj
i =1 j =1
(4.3b)

F T (G F LF ) T (4.3c)

GF 0 (4.3d)

Definition 4.2: (ISOs system dispatch problem in the spot wholesale market) The

ISO collects the load serving entities projected electricity demands in the spot market,

based on the scheduled generation resources as determined in the day-ahead forward

market, the ISO adjusts the available generation resources accordingly to meet the spot

demand while minimizing the total electricity energy acquirement cost. By clearing the

supply and demand at each location with respect to the transmission capacity constraints

on each monitored flowgates, the market clearing forward location marginal prices and

flowgate shadow prices are determined by solving the following optimization problem,

102
I
Min C (G ) = C (Gi ) (4.4a)
i =1

I J
St. G = L
i =1
i
j =1
j (4.4b)

FT (G L ) T (4.4c)

G0 (4.4d)

Locational marginal pricing is the marginal cost of supplying the next increment of

power demand at a specific location on the network, taking into account the marginal cost

of generation and the physical aspects of the transmission system. Without the network

constraints, electricity would be traded at a unique price wherever the physical delivery

were located in the system. The demands would be met by the merit-order generations,

i.e., the cheapest group of generation units, to incur the minimum social cost. However,

the limited transmission network capacity may limit the transactions as congestion occurs,

preventing the access to the cheapest generation resource while resorting to out of merit-

order generations instead with respect to the specific location of the demands. As a

consequence, the market clearing prices are differentiated across spatial locations. The

optimality conditions of (4.3) and (4.4) characterize the relation between location prices

and the reference bus price.

The problem (4.3), with a convex objective function and linear constraints, is a

convex programming problem. The associated Lagrange function can be defined as

I
( ) ( ) ( ) (F (G ) )
J
L G F , F , F = C G F + F G iF LFj + F
T T F
LF T
i =1 j =1

where F is the Lagrange multiplier associated with constraints (4.3b) and F is a vector

of Lagrange multipliers dual variables associated with constraints (4.3c).

103
In order for a solution (G F * , F , F ) to be optimal, in addition to constraints (4.3b-c),

to guarantee that the gradient of C (G F ) is normal to L Fj G iF and


J I

j =1 i =1

(F (G
T F
LF ) T ) , it requires C (G F ) be normal to G iF L Fj
I

i =1
J

j =1
and

(FT (G F LF ) T ) , that is, be linearly dependent vectors, the gradient of the objective

function should be a linear combination of the gradients of the active constraints.

L
G F
(
G F* , F , F = 0)
plus the complementary slackness condition,

( ) (F (G
F T T F*
LF * ) T ) = 0 , F 0

Due to the convexity of (4.3a-c), the second-order optimality conditions are satisfied.

In the economical sense, the Lagrange multiplier F * associated with the system-wise

power supply and demand balance can be interpreted as the locational marginal price at

the reference since it quantifies the electricity procurement cost for an additional unit of

load at the system reference bus. On the other hand, the Lagrange multiplier vector F *

associated with the power flow limit of the monitored transmission flowgate are

interpreted as the variation in electricity procurement cost if the transmission capacity is

relaxed, called flowgate shadow price hereinbelow. Therefore, the necessary conditions

for optimal system dispatch gives

P F = PrefF F F (4.5a)

where for each congested transmission flowgate k , kF > 0 , otherwise, kF = 0 .

Similarly, from problem (4.4), we can get

P S = PrefS F S (4.5b)

104
Note that the day-ahead forward prices P F re used to settle the day-ahead forward

market transactions, and the settlements of spot market position deviations from the day-

ahead schedules are based on the spot prices P S .

4.2.2 Generators and Load Serving Entities

With the consideration of transmission constraints, power suppliers cannot be deemed

as homogeneous as in (Bessembinder and Lemmon, 2002). Even if they share the same

production technology, their trading activities vary since they may experience diversified

market prices due to their respective spatial locations. Similarly, same arguments require

the LSEs been treated heterogeneously in their respective franchised service areas. With

the unbundling of generation and retail services, market agents pursue their individual

goals simultaneously. The market equilibrium under the multi-agent perspective can be

modeled by a set of optimization problems associated with different market participants

and clears according to the market coordination conditions. A variety of models have

been proposed to model market participants behaviors in a competitive market focusing

on choosing different variables to compete against their rivals. Bertrand, Cournot, and

Stackelberg models have been employed in Hogan (1997), Borestein et al. (1995), and

Otero-Novas et al. (2000) to study market participants activities. The Bertrand model

takes price decision as the move of each supplier while conjecturing that the rivals will

not react to its actions by changing their prices. In the Cournot model, the quantity is the

driver for market equilibrium. It is also termed as Nash equilibrium and can be reached

when no agent can improve individual profit by unilaterally changing equilibrium

strategy. A leader is assumed in the Stackelberg model to take the first move while

leaving other agents behaving as in the Cournot set up. In this section, a two-period

Cournot model is employed with the DC power flow approximation of network

constraints to show the interaction of strategic behaviors between different market

105
participants in diverse locations. Specifically, we evaluate GENs and LSEs forward

hedging and spot positions and obtain closed form solutions for the equilibrium

electricity prices based on market clearing conditions. This model is extendable to

include strategic behaviors in the generation reserve market.

We start with assessing the real-time spot wholesale market while taking into account

the day-ahead forward positions selected beforehand. Then step back to the day-ahead

forward market, the optimal decisions can be determined assuming that each GEN and

LSE would behave optimally in the spot market the next day.

Spot Market

According to the aforementioned assumption of market uncertainty being cleared in

the real-time spot market, market participants can make decisions regarding spot market

positions at precisely projected load demand and spot prices. LSEs can procure exactly

the amount of power as needed and meanwhile, GENs are able to decide spot positions

contracts in the spot wholesale market at known prices.

Definition 4.3: (Generators profit pursuing problem in the spot wholesale market)
The ex post profit of generator is the sum of revenues from electricity transactions in

forward and spot markets minus the total generation cost. Therefore, given the day-ahead

forward market position GiF * , GEN i s decision-making problem is to maximize the

profit as follow,

2
Max GS (GiS ) = PgF *GiF * + PgS ( )GiS Fi + G (GiF * + GiS ) (4.6)
i i
2 i

The first order necessary condition

GS (GiS )
= PgSi ( ) G (GiF * + GiS ) = 0
Gi S
i

106
gives the profit-maximizing spot market position by GEN i being,

iG S
GiS * = Pg ( ) GiF * (4.7)
i

And the second order sufficient condition satisfies

( )=
2 GS GiS
<0
GiS
2
iG

Note that as implied in the generation cost function (4.1), the locational marginal

price at the bus g i defined as the marginal cost of supplying the next increment of power

demand at that specific location of the network can be determined as:

dC (Gi )
Pg i = = G Gi
dGi i

which increases with Gi and implies that the supplier can set the market price at which it

sells the power. However, due to the correlations between market prices at various

locations of the same system as revealed by (4.5), this price-setting cannot be decided

independent of other market participants decisions. Instead, the conformation to the

market clearing conditions presented in section 4.3 combined with price relation (4.5)

provides the decision environment for generators to rely on.

Definition 4.4: (Load serving entities obligation fulfillment problems in the spot
wholesale market) For LSEs, by regulation they need to fulfill the real time demand of

end consumers in their franchised territory. Take LSE j as an example,

L j dl j
(4.8)

Therefore, LSE j has little control over LSj and simply buy in the shortage of

electricity power resulted from the forward contracted quantity LFj * ,

LSj * = dl j LFj * (4.9)

107
Therefore, the ex post profit of LSE j can be determined as,

j j j j
(
LS (LSj ) = Pl R d l Pl F * LFj * Pl S ( ) d l LFj *
j
) (4.10)

( ) (
= Pl Rj Pl Sj ( ) d l j + Pl Sj ( ) Pl Fj * )L
F*
j

Note that the profit gain (loss) from taking day-ahead forward position LFj * is

multiplied by Pl Sj ( ) Pl Fj * . And a reduced portion of load obligation d l j LFj * of LSE j

is subject to the spot market price surging risk. On the other side, the regulated retail

price Pl R should be high enough to create a profit margin and keep LSE j in the market
j

for service.

We next step back in time to formulate the day-ahead forward market decisions of

market participants.

Forward Market

In the day-ahead forward market, GENs and LSEs face uncertain load demands to be

revealed in the next day retail market. Given the uncertain P S ( ) and unrevealed

optimal decisions in the spot market G S and LS , the GENs and LSEs have to rely on the

projections of the spot market conditions and decide their forward market positions.

Given the subsequent optimal spot market positions G S * as reveal in (4.7) to be taken

by GEN i , his total profit with forward market position GiF can be determined as follow,


GF (GiF ) = PgF GiF + PgS * ( )GiS * ( ) (GiF + GiS * ( ))
2
(4.11)
i i
2 i
G

To facilitate the model derivation in following sections, we define GEN i s un-

hedged profit by assuming no position in forward market is taken, that is, GiF = 0 , that is,

the profit without forward hedging as:

108
iG S
iG ( ) = PgS ( )GiS * ( )
i
2 i
G
(Gi
S*
( ))2
=
2
(Pg ( ))
i
2
(4.12)

where, according to equation (4.5), PgS ( ) = PRefS ( ) f g ,k Sk ( ) , assuming k is the only


i i

transmission flowgate which has tight capacity and non-zero associated shadow price in

the system.

Taking the GiS * revealed in (4.7) and the iG ( ) defined in (4.12) into consideration,

equation (4.11) can be transformed into the following neat form,

GF (GiF )= iG ( ) + GiF (PgF PgS ( ))


i i
(4.13)

Definition 4.5: (GENs risk-adjusted profit pursuing problem in the day-ahead


forward wholesale market) With the projected overall profit (4.13), the risk-averse

generator i s utility maximization problem in the form of (4.2) in the day-ahead forward

market is as follow,

Max [ ( )] [ ( )]
E U GF GiF = E GF GiF
AG
2
( ( ))
Var GF GiF (4.14)

Given the rational that the un-hedged profit iG ( ) is calculated when no forward

market position is taken, we can assume that E [ iG ( )] and Var ( iG ( )) are independent

of GiF , which leads to explicit form of (4.14) as follow,

E [U GF (GiF )] = E [ iG ( )] + GiF (PgF E [PgS ( )])


i i


AG
2
[ ( ) ( ) ( )
Var iG ( ) + GiF Var PgSi ( ) 2GiF Cov iG ( ), PgSi ( )
2
( )]
The first order necessary condition for optimality

[
E U GF (GiF ) ] [ ] [ ( ) ( )]
= PgFi E PgSi * ( ) AG GiF Var PgSi * ( ) Cov iG ( ), PgSi * ( ) = 0
Gi F

gives the expected utility maximizing GEN i s forward market position being,

109
GiF * =
[
PgFi E PgSi * ( ) ] + Cov( ( ), P ( )) G
i
S*
gi

A Var (P ( )) Var (P ( ))
S* S*
(4.15)
G gi gi

where, according to (4.12), Cov iG ( ), PgSi * ( ) = ( ) iG


2
(
Cov (PgS * ( )) , PgS * ( )
i
2
i
)
And the second order sufficient condition of optimality satisfies,

2 E [U GF (GiF )]
= AGVar (PgS * ( )) < 0
GiF
2 i

As described in (4.15), the optimal forward position consists of two components with

the first reflects the position taken in response to the bias in the forward price as

compared to the expected spot prices. Note that since iG ( ) is positively correlated

with PgS * ( ) for i , the second term of GEN i s forward market position accounts for
i

the effort to reduce the total profits risk exposure to the spot market prices as compared

to without forward hedging. Note that in (4.15), the E [PgS * ( )] denotes the conditional i

expectation of PgS * ( ) , similarly, Var (PgS * ( )) and Cov( iG ( ), PgS * ( )) are both
i i i

conditional measures.

Taking (4.15) back into (4.7), the optimal position to take in the wholesale spot

market by GEN i can be expressed as,

iG S PgF E [PgS * ( )] Cov( iG ( ), PgS * ( ))


G =S*
Pg ( ) +
AGVar (PgS * ( )) ( )
i i i
(4.16)
i

i
Var Pg
S*
(i
)
i

Similarly, load serving entity j s profit function is to be determined as,

LS (LSj ) = Pl R L j ( ) Pl F LFj Pl S * ( )(LRj ( ) LFj )


j j j
(4.17)

( )
= Pl Rj Pl Sj * ( ) L j ( ) + Pl Sj * ( ) Pl Fj LFj ( )
By assuming no position to be taken in forward market, that is, LFj = 0 , we define

LSE j s profit without forward hedging as:

110
( )
Lj ( ) = Pl R Pl S * ( ) L j ( )
j j

where, Pl Sj ( ) = PRef
S
( ) fl j , k kS ( )
Note that LSE j s profit without forward hedging is exposed to risks of electricity

acquiring cost in wholesale spot market and revenue in the retail market. The retail

revenue co-varies positively with the electricity wholesale spot price since electricity

price is positively correlated with the locational demand in general. And this correlation

can be amplified by the regulated retail price level. However, the cost to acquire the

required electricity from wholesale market offsets the revenue.

Therefore, take in the LSj* revealed in (4.9) and the Lj ( ) defined in (4.17) into

consideration, (4.17) can be transformed into the following,

(
LF (LFj ) = Lj ( ) LFj Pl F Pl S * ( )
j j
) (4.18)

Definition 4.6: (LSEs risk-adjusted profit pursuing problem in the forward


wholesale market) With the projected overall profit (4.18), the risk-averse load serving

entity j s utility maximization problem in the form of (4.2) in the day-ahead forward

market is as follow,

AL
Max E[U LF (LFj )] = E[ LF (LFj )] Var ( LF (LFj )) (4.19)
2

By assume E[ Lj ( )] and Var ( Lj ( )) being independent of LFj , which is reasonable

given that un-hedged profit Lj ( ) considers zero LFj , we get,

[ ( )] [ ]
E U LF LFj = E Lj ( ) LFj Pl Fj + LFj E Pl Sj * ( ) [ ]

AL
2
( ( 2
) ( ) ( ) ( ))
Var Lj ( ) + LFj Var Pl Sj * ( ) + 2 LFj Cov Lj ( ), Pl Sj * ( )

The following first order necessary condition

111
[ ( )] = E [P
E U LF LFj S*
( )] Pl F AL (LFjVar (Pl S * ( ))+ Cov ( Lj ( ), Pl S * ( ))) = 0
lj
LFj j j j

gives the expected utility maximizing LSE j s forward market position being,

[ ]
E Pl Sj * ( ) Pl Fj (
Cov Lj ( ), Pl Sj * ( ) )
L =
( )
( )
F*
(4.20)
ALVar Pl j ( ) Var Pl j ( )
j S* S*

( ) ((
where, Cov Lj ( ), Pl Sj * ( ) = Cov Pl Rj Pl Sj * ( ) L j ( ), Pl Sj * ( ) ) )
= Pl R Cov (L j ( ), Pl S * ( )) Cov (Pl S * ( )L j ( ), Pl S * ( ))
j j j j

And the second order sufficient condition satisfies,

[ ( )] = A Var (P
2 E U LF LFj S*
( )) < 0
F2 L lj
L j

Similar to what is revealed in (4.15) for generators, the optimal forward position for

LSEs consists of two components with the first reflects the position taken in response to

the bias in the forward price as compared to the expected spot prices. Since Lj ( ) is

negatively correlated with Pl S * ( ) for l j , the second term of LSE j s forward market
j

position accounts for the effort to reduce the total profits risk exposure to the spot

market prices without forward hedging.

Therefore, taking (4.20) back to (4.9), the optimal position to take in the wholesale

spot market by LSE j can be expressed as,

[ ]
E Pl Sj * ( ) Pl Fj Cov Lj ( ), Pl Sj * ( )
L = L j ( )
( )

( ) ( )
S*
(4.21)
j
ALVar Pl S * ( ) Var Pl Sj * ( )
j

Since GENs and LSEs take balanced forward and spot market positions in day-ahead

and spot markets according to projected demand, the market risks imbedded are shared

by the market participants. The retail price regulation prevents LSEs from transporting

112
all of their market risk to end consumers who are vulnerable to price fluctuations due to

the inelasticity of electricity consumption.

4.3 Market Equilibrium Prices

The equilibrium of the market can be reached when, in addition to every market

participant choosing the optimal decision from each individual optimization problem

given others behavior strategies, the market reaches its coordination. In our case, it

requires the balance of supply and demand with respect to transmission capability.

4.3.1 Theoretical Formulations

In the forward wholesale market, the equilibrium requires,


I J

G
i =1
i
F
= LFj
j =1
(4.22)

In addition, according to the characterization of PTDF matrix, for the congesting

flowgate k , we have,

F, k (G F LF ) = Tk
T
(4.23)

We get (see appendix B),

2 ( 5 7 + Tk ) 3 ( 4 6 )
F
PRef = (4.24a)
2 2 1 3

1 ( 5 7 + Tk ) 2 ( 4 6 )
kF = (4.24b)
2 2 1 3

PnF =
( 2 f n ,k 1 )( 5 7 Tk ) ( 3 f n ,k 2 )( 4 6 )
(4.24c)
2 2 1 3
I J
1 1
where, 1 =
i =1 (S*
+
) (
AGVar Pg i ( ) j =1 ALVar Pl Sj * ( ) )

113
I f g ,k J f l ,k
2 = +
F F
i j

i =1 AGVar (PgS * ( )) j =1 ALVar (Pl S * ( ))


i j

3 =
I (f ) 2
J (f ) 2

) + A Var (P ( ))
gi ,k F l j ,k F

i =1 (
AGVar PgSi * ( ) j =1 L
S*
lj

4 =
I [
E PgSi * ( ) ] +
J [
E PLjS * ( ) ]
i =1 (
AGVar PgSi * ( ) ) j =1 ALVar PLjS * ( ) ( )
I [f g ,k F E PgSi * ( ) ] f E[P J
S*
( )]
=
( )) A Var (P
l j ,k F lj
5
l ( ))
+
A Var (P
i
S* S*
i =1 G gi j =1 L j

6 =
I (
Cov iG ( ), PgSi * ( ) )+ J (
Cov Lj ( ), Pl Sj * ( ) )
i =1 Var P ( ) ( S*
) j =1 Var Pl j ( )( S*
)
gi

I (
f g ,k F Cov iG ( ), PgSi * ( ) ) J fl ( )
Cov Lj ( ), Pl Sj * ( )
7 = +
F
j ,k

( ) ( )
i

i =1 Var PgSi * ( ) j =1 Var Pl Sj * ( )

And the optimal day-ahead forward market positions GiF * and LFj * can be derived by

combining (4.24c) with (4.15) and (4.20), respectively. The optimal forward positions

derived are useful in evaluating which market participants have comparative advantage in

participating as compared to being absent from forward transactions.

In the spot market, similar to (4.22), the coordination of the market requires the

balance of supply and demand with respect to transmission capability as follow,


I J

i =1
GiS * = LSj*
j =1
(4.25)

taking (4.16) and (4.21) into (4.25), it leads to,


F
I 1 J
1

( )
PRef +
S*
(
i =1 AGVar Pg i ( ) j =1 ALVar Pl j ( )
S*
)

I f g i ,k J f l j ,k
kF
i =1 AGVar PgS * ( ) ( )
+
k
i
(
j =1 ALVar Pl j ( )
S*
)

114
=
I [
E PgSi * ( ) ] +
J [
E Pl Sj * ( ) ]
i =1 AGVar PgSi * ( )( ) j =1 (
ALVar Pl Sj * ( ) )

I (
Cov iG ( ), PgSi * ( ) ) J (
Cov Lj ( ), Pl Sj * ( ) )+ I
iG S J
Pg L j ( )
i =1 (
Var PgSi * ( ) ) j =1 (
Var Pl Sj * ( ) )
i =1
i
j =1

Combining (4.22) and (4.25), since the total generation should equal to the total

demand, we have,
I
iG S J

i =1
Pg = L j = d l
j =1 l
i j
(4.26a)
j

which equates the total generation production to total demand, we get,

iG S
(PRef f g ,k Sk ) = d n
I


i =1 n
i
(4.26b)

I I
S
or, PRef iG Sk f gi ,k gi = L j
i =1 i =1
(4.26c)

Also, according to the characterization of PTDF matrix and for the congesting flowgate

k , corresponding to (4.23) in the day-ahead forward market, we have,

F, k
T
((G F
) (
+ G S LF + LS = Tk )) (4.27a)

That is,

( )
I I J
S
f gi ,k iG Sk f gi ,k iG = Tk + f l j ,k L j
2
PRef (4.27b)
i =1 i =1 j =1

We get,

2 (Tk + 4 ) 3 d n
P S
Ref = n
(4.28a)
2 1 3
2

1 (Tk + 4 ) 2 d n
=
S
k
n
(4.28b)
2 2 1 3

115
( 2 f n ,k 1 ) (Tk + 4 ) ( 3 f n ,k 2 ) d n
P =
n
S n
(4.28c)
2 2 1 3
I
where, 1 = iG
i =1

I
2 = iG f g ,k i
i =1

I
3 = iG f g ,k
2
i
i =1

J
4 = f l ,k d l j j
j =1

And the optimal spot market positions GiS * and LSj* can be derived by combining

(4.28c) with (4.16) and (4.21), respectively.

4.3.2 Numerical Examples

To illustrate the implications of economic determinants of the equilibrium forward

prices as expressed in (4.24c), a series controlled experiments are designed and

conducted on a 3-bus study system illustrated in figure 4.3.

Figure 4. 3 A 3-bus example system

The admittances of all transmission lines are assumed to be equal. Bus2 is the

reference bus. There exist I = 2 generators (GEN1 and GEN2) located at bus 1 and 2,

116
respectively, and J = 2 load serving entities (LSE 2 and LSE3) with respective

franchised service territory whose demands are aggregated at bus 2 and 3, respectively.

The ingoing and outgoing arrows illustrate power injections and ejections represented by

GENs and LSEs, respectively. Transmission flowgate 1-3 is monitored for transmission

congestion. We assume generation cost function parameters = 10 and 1G = 2G = 40 .

All market participants have the same risk aversion. We also assume the regulated retail

price by LSEs are 1.2 times the average spot market price.

Since the bus aggregated demands of end consumers are the most fundamental drivers

for market price uncertainty and transmission flowgate congestions, we study the

following scenarios by assuming different levels of aggregated demands at load buses 2

and 3. For simplicity, the demands are assumed to follow normal distribution N ( , ) .

The postulated parameters are as follows,

Scenario A: low demand

Scenario B: medium demand

Scenario C: high demand

In addition to the demand level, we study the impact of demand volatility as well by

adding another scenario

Scenario D: medium demands with high volatility

The parameter of demands assumptions at bus 2 and 3 at each scenario are listed in

table 4.1,

Table 4. 1 Parameters of demand distributions (in $/MWh)


Load Bus 2 Bus 3
Parameter
Scenario A 250 250 9 200 200 9
Scenario B 275 275 9 225 225 9
Scenario C 325 325 9 250 250 9
Scenario D 275 275 6 225 225 6

117
When control over the explanatory variables is exercised through random

assignments, since the randomization tends to balance out the effects of other variables

that might affect the response variable, the resulting experimental data provide strong

information about the cause-and-effect relationships than observational data. For each of

the following experiment setup, the results are based on system dispatching conditions

based on 100 load simulations.

The spot market prices, the day-ahead forward market prices, and the forward

premiums in each scenarios are listed in tables 4.2-13 are listed as bellows,

Scenario A:

Table 4. 2 Scenario A Spot market prices (in $/MWh)


Bus 1 Bus 2 Bus 3 Flow-gate Shadow Price
Mean 54.72 57.29 59.85 6.64
Variance 3.51 7.13 13.01 17.23
Table 4. 3 Scenario A Day-ahead forward market prices (in $/MWh)
Risk-aversion Coefficient Bus 1 Bus 2 Bus 3 FG Shadow Price
0.05 54.87 56.73 58.60 5.60
0.005 54.79 57.08 58.74 5.18
0.0005 54.76 57.20 58.79 5.02
Table 4. 4 Scenario A Day-ahead forward premium (in $/MWh)
Risk-aversion Coefficient Bus 1 Bus 2 Bus 3 FG Shadow Price
0.05 0.15 -0.56 -1.25 -1.04
0.005 0.07 -0.21 -1.11 -1.46
0.0005 0.04 -0.09 -1.06 -1.62
Scenario B:

Table 4. 5 Scenario B Spot market prices (in $/MWh)


Bus 1 Bus 2 Bus 3 FG Shadow Price
Mean 54.78 69.67 84.56 23.22
Variance 20.05 160.78 812.10 1440.00

118
Table 4. 6 Scenario B Day-ahead forward market prices (in $/MWh)
Risk-aversion Coefficient Bus 1 Bus 2 Bus 3 FG Shadow Price
0.05 52.57 55.11 57.65 44.30
0.005 54.56 68.21 81.87 40.97
0.0005 54.75 69.52 84.29 7.62
Table 4. 7 Scenario B Day-ahead forward premium (in $/MWh)
Risk-aversion Coefficient Bus 1 Bus 2 Bus 3 FG Shadow Price
0.05 -2.21 -14.56 -26.91 21.08
0.005 -0.22 -1.46 -2.69 17.75
0.0005 -0.02 -0.15 -0.27 -15.60

Scenario C:

Table 4. 8 Scenario C Spot market prices (in $/MWh)


Bus 1 Bus 2 Bus 3 FG Shadow Price
Mean 50.60 92.54 134.47 125.82
Variance 47.60 298.40 1650.00 4960.00
Table 4. 9 Scenario C Day-ahead forward market prices (in $/MWh)
Risk-aversion Coefficient Bus 1 Bus 2 Bus 3 FG Shadow Price
0.05 44.33 104.42 164.50 180.25
0.005 47.47 98.48 149.49 153.03
0.0005 49.97 93.72 137.48 131.26
Table 4. 10 Scenario C Day-ahead forward premium (in $/MWh)
Risk-aversion Coefficient Bus 1 Bus 2 Bus 3 FG Shadow Price
0.05 -6.26 11.88 30.02 54.43
0.005 -3.13 5.94 15.01 27.22
0.0005 -0.63 1.19 3.00 5.44
Scenario D:

Table 4. 11 Scenario D Spot market prices (in $/MWh)


Bus 1 Bus 2 Bus 3 FG Shadow Price
Mean 52.93 71.23 89.52 48.04
Variance 39.20 322.50 1570.00 4300.00
Table 4. 12 Scenario D Day-ahead forward market prices (in $/MWh)
Risk-aversion Coefficient Bus 1 Bus 2 Bus 3 FG Shadow Price
0.05 44.31 113.23 182.15 206.77
0.005 52.07 75.43 98.79 70.07
0.0005 52.85 71.65 90.45 56.40

119
Table 4. 13 Scenario D Day-ahead forward premium (in $/MWh)
Risk-aversion Coefficient Bus 1 Bus 2 Bus 3 FG Shadow Price
0.05 -8.63 42.00 92.63 158.72
0.005 -0.86 4.20 9.26 22.03
0.0005 -0.09 0.42 0.93 8.36

According to experimental results illustrated above, the following observations can be

made,

a. According to tables 4.4, 4.7, 4.10, and 4.13, the lower the risk aversion coefficients

of market participants are, that is, the more risk-neutral the market participants become,

the closer the forward prices, including on the reference bus, are to the expected spot

prices

b. By comparing tables 4.2, 4.5, 4.8, and 4.11, the higher the demands, the more

frequently the flow-gate gets congested in the spot market, and the higher the market

prices at load buses and the lower the market prices at the generation buses are. Also, by

comparing tables 4.3, 4.6, 4.9, and 4.12, similar results can be observed from the day-

ahead forward market as well.

c. By comparing scenario B and scenario D results, for the same expected load level,

the more volatile the load gets, the larger the forward prices at load buses and the lower

the forward prices at the generation buses are.

4.4 Day-Ahead Forward Risk Premium

4.4.1 Theoretical Formulation

By applying the relation of locational marginal prices as revealed by (4.5) to the day-

ahead forward and spot markets, we get the following (4.29a) and (4.29b), respectively,

PnF = PRef
F
f n ,k kF (4.29a)

[ ] [
E PnS ( ) = E PRef
S
] [ ]
( ) f n,k E Sk S ( ) (4.29b)

120
Note that they reflect the effect of transmission capacity constraints on market

clearing conditions. Following the forward market coordination (4.22), with the utility

maximizing forward market positions of GENs and LSEs given in (4.15) and (4.20),

respectively, we get

I [
PgFi E PgSi * ( ) ] + P E [P
J
F S*
( )]
A Var (P ( )) A Var (P
lj lj

l ( ))
S* S*
(4.30a)
i =1 G gi j =1 L j

=
I (
Cov iG ( ), PgSi * ( ) ) J (
Cov Lj ( ), Pl Sj * ( ) )
i =1 (
Var PgSi * ( ) )
j =1 (
Var Pl Sj * ( ) )
which leads to (see appendix C),

I
(P [ ])
J
1 1

E PRef +
( )
F S
Ref

S*
i
(
i =1 AGVar Pg ( ) j =1 ALVar Pl S * ( )
j
)

(4.30b)

I f l j ,k
= (kF E Sk ) [ ] f gi ,k J

+ (
i =1 AGVar PgS * ( ) j =1 ALVar Pl S * ( ) ) ( )
i j



(
I Cov iG ( ), PgS * ( )
+
L
j )
J Cov ( ), P ( )
S*
lj ( )
( ) ( )
i

i =1
Var Pg
S*
i
( ) j =1 Var Pl Sj * ( )

Equation (4.30b) connects the forward premium on the reference bus with the forward

premium of transmission flowgate shadow prices (kF E [Sk ]) k , the covariance

between the un-hedged profits of generators and load serving entities with their

respective spot prices in the real-time spot market.

Model (4.30b) can be extended to the following more general form when the effect of

multi transmission flowgates are taken into account,

I
(P E[PRefS ])
1 J 1
F
Ref
i =1 A Var(P S * ( ))
+
j =1 A Var(P S * ( ))
(4.31)
G g i L l j

121
I f g ,k f l ,k
= (kF E [Sk ])
J
i
+ j

i =1 A Var (P S * ( )) j =1 A Var (P S * ( ))
kK
G g i L l j

I Cov( iG ( ), PgS * ( )) J Cov ( Lj ( ), Pl S * ( ))


i
+ j

i =1 Var ( Pg
S*
i
( ) ) j =1 Var (Pl
S*
(
j
) )

Note that the economic incentives of transmission network constraints as revealed by

{F , S } play an important role and consist a non-negligible component of the forward risk

premium on the reference bus. Similarly, the effect on the forward risk premium on other

buses can be derived by combining (4.5) and (4.31).

4.4.2 Empirical Evidences

We start with an overview of the structure and functions performed by the New York

independent system operator (NYISO). The NYISO is an outgrowth of the NYPOOL

(New York power pool), which was created by New York's eight largest electric utilities

to coordinate the statewide interconnection following the big Northeast blackout of 1965.

In responding to the Federal Energy Regulatory Commission (FERC)s acts and related

state policies issued in mid 90s intended to create more competition in the nations

wholesale electricity markets, the transmission system gradually became open and

provide nondiscriminatory access to newly expanded non-utility generation. As the

restructuring of the electric power industry evolved, NYPOOL was dissolved in 1998 and

replaced with the not-for-profit NYISO. The NYISOs mission is to ensure the reliable,

secure and efficient operation of the interconnected transmission system and to create and

administer an open, competitive and nondiscriminatory electricity wholesale market, in

which power is traded on the basis of competitive bidding. It enables the transactions to

be settled at competitive prices, rather than regulated rates. Currently, the NYISO system

oversees over 160,000 GWh energy transactions each year with a recorded summer load

122
peak of over 32,075MW (July 26, 2005) and winter load peak of 25,540MW (December

20, 2004), which represent about $8 billion market value. Specifically, NYISO

establishes and enforces market trading protocols, coordinate the transmission of

electricity power from generation source to consumption centers, and clears the markets

at the settlement prices.

Figure 4. 4 NYISO control area load zones

Compared with many other ISOs, where only partial, if any, markets have been

created, the NYISO adopts a full, two-settlement system with comprehensive markets for

a variety of energy-related contracts, including the day-ahead and real-time energy

markets (balancing market).

In the day-ahead markets (DAM), a set of forward prices are determined on an hourly

basis for each zone within the control area (and the neighboring areas) at a pre-specified

time (11AM) by matching generation and energy transaction bids offered in advance to

the ISO. The ISO runs a security constrained unit commitment and determine the amount

of energy needed for each day. Transmission losses, congestion, shift factors, penalty

factors and other system mathematical quantities are calculated against a reference bus

(physically located at the Marcy 345 kV substation in Marcy, New York). The DAM

123
zonal locational based marginal prices (LBMPs) are obtained by adding the marginal

costs of energy, losses and congestion and used as the basis for settlements. Generating

units that can most economically satisfy the energy needed to supply customers' demand

and allow a sufficient reserve for contingencies are scheduled. Typically, most (>90%)

energy transactions processed by the NYISO occur in the DAM since market participants

can hedge against the next day price risks.

In the real-time market (RTM), the ISO runs a security constrained dispatch (SCD)

determines the amount of energy needed on a continual basis. SCD makes adjustments to

previous schedules to regulate generating units that can most economically satisfy the

energy needed to supply customers' real-time demand and allow a sufficient reserve for

contingencies. LBMPs are calculated at five-minute intervals throughout the day.

Typically less (<10%) energy transactions processed by ISO occur in the RTM. The

settlements based on RTM market prices is intended to credit or charge market

participants for energy transactions, due to variations in a power suppliers real-time

dispatch quantity from what is pre-determined in the DAM. The total settlement is

determined by the sum of balancing energy, incremental occurring losses, and

created/eliminated congestion. There is actually a rolling hour-ahead market (HAM)

intended to balance the most updated system situations with pre-schedules (sell excess

and buy shortage). Since it serves the same function as the RTM as far as the proposed

model is concerned, they are not distinguished here.

Note that the market settlements are based on LBMPs, which differ from the market-

clearing price determined at New York mercantile exchange in that the spatial location of

a power transaction matters even within a market.

We are interested in the following market participants, which can be categorized as

follows based on their primary business functions,

124
a. Power producers that own generation units and sell to the wholesale markets

b. Load serving entities that buy from the wholesale markets and sell electricity at

certain retail prices to end consumers via the distribution networks

c. End consumers that have access to electricity through load serving entities. They

are generally immune to the market price risks since the prices are generally stable.

Although in reality most market participants are not exclusively electricity sellers or

buyers but tend to appear on both sides of the market as the needs to fulfill contractual

commitments or as the opportunities to generate extra profit come up, the primary

business function of each market participant generally dominates its market activities and

in our model, such subsidiary functions can be ignored without missing the major

determinant factors for forward risk premium discovery.

One problem with empirical tests is the availability of market data. Since the market

functions changed gradually following the conception of the deregulation, the historical

data do not necessarily reflect the same market structure. Besides, market participants do

not behave consistently due to the ever-evolving of the market design. Therefore, the

historical market equilibrium prices over a long period of time bear discrete biases caused

by regulatory adjustment. Taking the concerns listed above into consideration, we

choose the time-horizon from February 1st 2005 to August 31st 2006 when the NYISO

market structure was relatively stable. The data used for this study consist of hourly

LBMPs and flowgate shadow prices in day-ahead and spot markets, and houly zonal

loads as well. These data are acquired from NYISO website. The Zonal LBMPs are

load-weighted LBMPs at all buses within the zone.

A summary of statistics for the electricity forward and spot prices on the system

reference bus and in each zone is reported in table 4.14 as bellow.

125
Table 4. 14 Statistics of forward and spot LBMPs (in $/MWh)
Forward Market Spot Market
Mean Median Volatility Mean Median Volatility
Ref. Bus 66.99 63.21 21.95 63.67 58.56 40.72
CAPITL 74.67 70.42 25.11 69.82 64.32 45.43
CENTRL 63.72 59.80 21.95 61.46 56.22 41.29
DUNWOD 79.78 75.60 27.26 73.78 68.13 48.06
GENESE 60.48 56.33 21.30 58.89 53.04 41.64
HUD VL 78.38 74.37 26.64 73.17 67.52 47.87
LONGIL 82.76 78.33 27.88 76.26 70.87 48.76
MHK VL 68.19 64.40 22.67 65.79 60.49 43.35
MILLWD 79.79 75.73 27.35 73.58 67.93 48.03
N.Y.C. 83.42 78.68 29.25 76.24 70.19 49.85
NORTH 63.72 59.70 20.64 61.72 56.57 39.41
WEST 54.42 49.53 20.38 53.64 47.77 38.81

Table 4.14 shows that there is considerable forward premium for the reference bus

and each individual zone. It also indicates the right-skewness of the electricity prices,

which is consistent with the implication of the model presented in (Routledge, 2001).

The forward premiums reflect the market buyers willingness to pay to secure the

procurement of electricity.

Equation (4.31) describes a linear functional relationship between the forward

premium on the reference bus PRefF E [PRefS ] and the forward premium of shadow prices

on the flowgates kF E [Sk ] . The PTDF matrix F determines the effect of kF E [Sk ] on

PRefF E [PRefS ] . However, since the topology of the NYISO transmission network is not

accessible to the public, we can not derive a regression model between PRefF E [PRefS ] and

kF E [Sk ] to test the empirical data. To estimate the PTDF structure of the system, we

resort to relationship between locational marginal prices and the reference bus price as

revealed in (4.5). Given empirical data of hourly locational based marginal prices on the

system reference and each zone, and shadow prices associated with transmission

126
flowgates over the time horizon of h = 1,L, H , the following multiple regression model

can be constructed for all n accordingly,


K
Yh ,n = X h ,k k ,n + h ,n with X i 0 1 (4.32a)
k =0

where, for, k = 1, L , K , and n = 1,L, N

Yn ,h price difference between zone n and the system reference bus at hour h .

X k ,h shadow price of transmission flowgate k at hour h .

n ,k regression coefficients representing the negative of PTDF f n ,k .

n ,h independent error terms of the regression model with expectation 0 and

variance 2

or, in the equivalent matrix form for each zone n ,

Yn = X + n (4.32b)
H 1 H K K 1 H 1

We expect the observational Yn to distribute around the regression surface (a hyper-

plane in this case since it is more than 2 dimensions) X . To set up interval estimates

and make tests, the distributions of the residual terms n are assumed to be normal. We

have n ~ iid .N (0, n2 )

The normality assumption for the error terms is justifiable since the error terms

represent the effect of random flowgate congestions omitted from the model (4.32) that

affect the response to some extend and that vary at random without reference to the

modeled explanatory variables X . Besides, as long as the distributions of error terms do

not depart significantly from normality, the t-distribution based tests results are reliable.

In equation (4.32b), components of are the partial regression coefficients since they

reflect the partial effect of one explanatory variable when other explanatory variables are

127
held constant. Since the explanatory variables have additive effects, the effect of X k on
1

the mean response does not depend on the level of X k , k1 k 2 .


2

In a large electric power system such the New York power pool, as indicated by the

empirical date, the system situation can be very complicated. To reduce the reality to a

manageable proportion for the regression model (4.32), we choose an incomplete list of

flowgates as the explanatory variables for the spatial price differences. However, the

flowgates enlisted are the most prominent ones for both day-ahead market and the real-

time spot market and can make good sense for the purpose of the analysis. The major

consideration for their choice is the extent to which the chosen flowgate contributes to

reducing the remaining variation in the dependent spatial price difference after allowance

is made for the contributions of other flowgate shadow prices as predictor variables that

have tentatively been included in the regression model. Accordingly, in the formulation

of the regression model (4.32), we restrict the coverage of the empirical data

corresponding time intervals when the selected predictor variables are non-trivial. The

shape of the regression function substantially outside this range should be treated

differently.

Three prominent flowgates which experience transmission congestion most

frequently in both the day-ahead and the real-time spot markets are picked. The

corresponding IDs in the NYISO network are 25091, 23330, and 25546. They are

denoted as FG1, FG2, and FG3 respectively hereinafter. The statistics of the shadow

prices associated with the capacity constraints represented by them in the market dispatch

model are reported in Table 4.15 as follow,

128
Table 4. 15 Statistics of forward and spot shadow prices of the flowgates (in $/MWh)
Forward Market Spot Market
Mean Median Volatility Mean Median Volatility
FG1 16.20 11.96 23.12 23.79 10.47 48.61
FG2 1.10 0.00 5.39 5.53 0.00 143.52
FG3 3.02 0.00 17.19 3.48 0.00 54.30

As shown in table 4.15, FG1 is the most active transmission flowgate of the three.

The time series of shadow prices associated with FG1 in the day-ahead and spot markets

are illustrated in figure 4.5,


Forward Market Spot Market

1000 1000
900 900
800 800
Shadow Price ($/MWh)

Shadow Price ($/MWh)

700 700
600 600
500 500
400 400
300 300
200 200
100 100
0 0
Mar2005 Oct2005 May2006 Mar2005 Oct2005 May2006
Date Date

Figure 4. 5 Hourly shadow prices on FG1 in forward and spot markets

According to table 4.10 and figure 4.5, it is observed that shadow prices of flowgates

are generally less volatile in the forward market compared in the spot market, which

conforms to the observation of electricity prices as revealed in figure 4.1. Besides,

flowgate shadow prices take less extreme values in the day-ahead forward market and the

averages are lower as compared with in the spot market.

The seasonality of electric load demands lead to similar pattern changes in market

prices and system conditions. For instance, high market prices and frequent transmission

congestions tend to be observed in summer and winter season. Transmission network

129
topology may be changed by equipment maintenances which are usually scheduled in

sprint and fall when load demands are relatively low. Since later in the section, the

inferred PTDF coefficients are used for the projection of out-of-sample (August 2006)

market prices using model (4.31), we limit the historical data to be taken from June to

August in 2005 and June to July in 2006 when the system structure is relatively stable

and the explanatory variables variation ranges are comparable to their counterparts in

August 2006. Preprocessing of the data also include getting ride of price spikes (we use

150$/MWh as the criteria for spikes) in locational market prices and shadow prices

associated with selected flowgates.

According to the subset of active flowgates selected for the study of the imbedded

statistical relation with the spatial price difference of electricity prices as modeled above,

the unbiased maximum likelihood squares estimates of the set of PTDF coefficients ,

denoted as b can be calculated and displayed in table 4.16.

Table 4. 16 Estimation of PTDFs


FG1 FG2 FG3
Zone-CAPITL -0.1229 -0.1078 0.0045
Zone-CENTRL -0.0376 0.0200 0.0782
Zone-DUNWOD -0.1431 -0.2073 0.0276
Zone-GENESE -0.0605 0.0615 0.2512
Zone-HUD VL -0.1409 -0.1722 0.0170
Zone-LONGIL -0.1493 -0.2172 0.0326
Zone-MHK VL -0.0215 -0.0266 0.0331
Zone-MILLWD -0.1513 -0.1775 0.0316
Zone-N.Y.C. -0.1800 -0.3314 0.0183
Zone-NORTH 0.0791 0.0652 -0.0341
Zone-WEST -0.0793 0.1264 0.2810
In the regression model (4.32), the variance of the error terms n indicate the

variability of the probability distributions of Y . Its unbiased estimate error mean square

(MSE) is calculated and displayed in the second column of table 4.17. The adjusted

coefficients of multiple determination shown in the third column of table 4.17 measure

the proportionate reduction of total variation in Y associated with the use of the set of

130
X variables, and in contrast to un-adjusted coefficient of multiple determination, it takes
the associated degrees of freedom of each sum of squares into consideration.

To test whether there is a regression relation between the response variable Y and

the set of X variables [X k ], k = 1,2, L , K , i.e., to test whether or not k = 0 for k , an F

test can also be conducted to determine between the alternatives:

H 0 : 1 = 2 = L = K = 0 , H : not all k = 0, k = 1, L , K

We use the test statistic:

MSR
F=
MSE

with the decision rule to control the type I error at as follows:

If F * F (1 ; K , H (K + 1)) , conclude H 0 , otherwise, conclude H .

Apply regression model (4.32a) to all NYISO zones, the statistical measures

including MSE, R 2 , F, and p values, together with the serial correlation ( ) of error

terms of the regression model are shown in table 4.17:

Table 4. 17 Statistics of regression


MSE R2 F p
Zone-CAPITL 8.7446 0.7051 4.3573 0.0055 0.0772
Zone-CENTRL 2.5466 0.7903 4.5583 0.0042 -0.0361
Zone-DUNWOD 12.7076 0.6564 3.7473 0.0121 0.0819
Zone-GENESE 12.8970 0.4678 3.2998 0.0217 0.0921
Zone-HUD VL 8.4994 0.6637 3.7299 0.0124 -0.0687
Zone-LONGIL 6.0739 0.6494 4.1753 0.0070 -0.0299
Zone-MHK VL 1.5187 0. 5413 3.2727 0.0225 0.0564
Zone-MILLWD 11.7480 0.7056 4.0583 0.0081 0.0478
Zone-N.Y.C. 15.6489 0.6728 2.9149 0.0358 -0.1141
Zone-NORTH 2.6171 0.573 3.8538 0.0106 0.0802
Zone-WEST 14.3732 0. 7393 3.1193 0.0275 -0.0744

Given the wide support of explanatory variable X , together with moderate MSE

values, high R 2 values imply good inference power. The p-value reported in table 4.17

is the probability of observing the given sample result under the assumption that the H 0

131
hypothesis of the F test is true, depend on assumptions about the independence and

normality assumptions of the random disturbances i in the model equation (4.32). Since

all but one of the p-values are less than = 0.05 in general, This is a quite significant

and strong indication that the null hypothesis can be rejected. F statistic values as

extreme as the observed ones in table 4.17 would rarely occur if the differences between

zonal prices and the reference bus prices are independent of the flowgate shadow prices.

To test whether there is a regression relation between the response variable Y and

each of the X variables [ X k ], k = 1,2, L , K , i.e., the tests whether or not k = 0 for each

k , t tests can also be conducted to determine between the alternatives:

H 0 : k = 0 , H : k 0

bk
We use the statistics t * =
s{bk }

with the decision rule:


If t * t 1 ; H (K + 1) , conclude H 0 , otherwise, conclude H .
2

where s (bk ) are square root of the diagonal components of s 2 {b} = MSE (X' X )1 and
( K +1)( K +1)

calculated and listed in table 4.18 as bellow,

Table 4. 18 s (bk ) value for the estimation of PTDFs


FG1 FG2 FG3
Zone-CAPITL 0.0344 0.0573 0.0038
Zone-CENTRL 0.0369 0.0082 0.0409
Zone-DUNWOD 0.0539 0.0839 0.0258
Zone-GENESE 0.0375 0.0244 0.1093
Zone-HUD VL 0.0449 0.0772 0.0113
Zone-LONGIL 0.0537 0.0850 0.0200
Zone-MHK VL 0.0070 0.0369 0.0292
Zone-MILLWD 0.0470 0.0746 0.0277
Zone-N.Y.C. 0.0644 0.1173 0.0831
Zone-NORTH 0.0341 0.0499 0.0421
Zone-WEST 0.0483 0.0768 0.1538

132
The following table 4.19 shows the t-values of the PTDF regression estimations,

Table 4. 19 t-Value for the estimation of PTDFs


FG1 FG2 FG3
Zone-CAPITL -3.5749 -1.8810 1.1886
Zone-CENTRL -1.0178 2.4509 1.9125
Zone-DUNWOD -2.6562 -2.4714 1.0695
Zone-GENESE -1.6152 2.5175 2.2975
Zone-HUD VL -3.1412 -2.2309 1.5074
Zone-LONGIL -2.7794 -2.5545 1.6295
Zone-MHK VL -3.0777 -0.7206 1.1319
Zone-MILLWD -3.2180 -2.3803 1.1392
Zone-N.Y.C. -2.7932 -2.8246 0.2201
Zone-NORTH 2.3206 1.3079 -0.8106
Zone-WEST -1.6402 1.6466 1.8274

For most of the t-values associated with FG1 and FG2, the corresponding null

hypothesis H 0 s can be rejected at confidence level higher than 95%. Even for FG3, the

H s are more favorable than the corresponding H 0 s.

For the normal error regression model (4.32), we also have

bk k
~ t (H (K + 1)) , k = 0,1,L, K
s (bk )

Note that if only the probability distributions of Y do not depart significantly from

normality, the sampling distributions of estimates b are approximately normal and the t-

distribution based tests can provide approximately the specified confidence coefficient.

Even if Y depart seriously from normality, the estimators b have the property of

asymptotic normality and approach normality under the general conditions as the sample

size increases. Therefore, the confidence intervals and the conclusions still apply. The

confidence limits for k with 1 confidence coefficient are


bk t 1 , H (K + 1) s (bk ) , k = 0,1,L, K
2

Table 4.20 shows the confidence interval of PTDF estimates with 95% confidence level.

133
Table 4. 20 Confident Interval of PTDFs estimation
FG1 FG2 FG3
Zone-CAPITL [-0.1908, -0.0550] [-0.2209, 0.0053] [-0.0030, 0.0120]
Zone-CENTRL [-0.1105, 0.0353] [0.0039, 0.0361] [-0.0025, 0.1589]
Zone-DUNWOD [-0.2494, -0.0368] [-0.3729, -0.0417] [-0.0233, 0.0785]
Zone-GENESE [-0.1344, 0.0134] [0.0133, 0.1097] [0.0354, 0.4670]
Zone-HUD VL [-0.2294, -0.0524] [-0.3245, 0.0199] [-0.0053, 0.0393]
Zone-LONGIL [-0.2553, -0.0433] [-0.3850, 0.0494] [-0.0069, 0.0721]
Zone-MHK VL [-0.0353, -0.0077] [-0.0995, 0.0463] [-0.0246, 0.0908]
Zone-MILLWD [-0.2441, -0.0585] [-0.3247, 0.0303] [-0.0231, 0.0863]
Zone-N.Y.C. [-0.3072, -0.0528] [-0.5630, 0.0998] [-0.1458, 0.1824]
Zone-NORTH [0.0118, 0.1464] [-0.0332, 0.1636] [-0.1171, 0.0489]
Zone-WEST [-0.1747, 0.0161] [-0.0251, 02779] [-0.0225, 0.5845]

Since the regression model (4.32) assumes that the congestion shadow prices are

known constants, the confidence coefficient and risks of errors are interpreted with

respect to taking repeated samples in which the congestion conditions are kept at the

comparable levels as in the observed sample.

As it is illustrated in tables 4.11-12, although there exist variations across individual

zones, the statistical measures provide evidence for a significant relation between the

flowgate shadow prices and electricity price spatial differences. We conclude that the

PTDF coefficients are estimated with satisfactory statistical significance. Note that the

PTDF matrix as inferred above is different from what PTDF is defined in the traditional

way since there does not exist a specific bus in a zone into which the power injection

cause a portion of power flow indicated as the PTDF coefficient. Nevertheless, the PTDF

matrix indicates the impact of aggregated power injection changes in one zone on a

specific transmission flowgate. This is of special meaning when the locational based

marginal prices are calculated for each zone of the NYISO system.

Previous research find that the forward risk premiums in electricity forward prices

vary systematically throughout the day and the forward premium manifests most

significant during on-peak hours when spot market prices are most volatile. In (Longstaff

134
and Wang, 2004), hour-by hour forward premiums in PJM market are tested showing that

forward premium tend to be positive in on-peak hours but negative in off-peak hours.

By examining the empirical data with NYISO, we find the similar patter in forward

risk premium on the reference bus as plotted as bellow,

Mean
Forward Premium ($/MWh) 7 Media

1
5 10 15 20
Hour

Figure 4. 6 Forward premiums on the reference bus at each hour

Bessembinder and Lemmon (2002) note that the existence of large forward premiums

reflect the lack of risk-sharing in electricity markets with market risk being bear by a

subset of participants. The higher forward premiums during on-peak hours represent the

time period within a day when the market participants face the greatest economic risks.

Note that the period of high forward risk premium corresponds to the high load period.

Taking the CENTRL zone as an example (and each zonal load has a similar daily pattern),

the hour-by-hour loads are plotted in figure 4.7 as follow,

135
Mean
2100 Media

Load Level (MW)


2000

1900

1800

1700

5 10 15 20
Hour

Figure 4. 7 Load level in the CENTRL zone at each hour

One interesting observation is that the forward risk premium of the shadow prices on

the transmission flowgate FG1 as plotted in figure 4.8 shows the opposite direction: the

premium is high during off-peak hours while low during on-peak hours.

2 Mean
Media
Forward Premium ($/MWh)

-2

-4

-6

-8

-10

-12

5 10 15 20
Hour

Figure 4. 8 Forward premium of the shadow price on FG1 at each hour

This is not surprising if we look back at the equation (4.31) and the estimated PTDF

matrix illustrated in table 4.16. Since the power transfer distribution factors of most

zones on the flowgate FG1 are negative numbers, the forward premium on the system

136
reference bus is negatively correlated to the forward premium of flowgate shadow prices.

This observation provides more insights to the forward premium in the electricity

wholesale contracts. Actually, part of the premium can be attributed to the transmission

capacity constraints effect.

To evaluate the forward risk premium determining model shown by (4.31), the same

set of empirical data as used for system PTDF coefficient inference is used to test the

following regression model,


5
Z h = h , p p + h with h 0 1
p =0

where, for h = 1,L, H

Zh system reference bus locational based marginal price forward premium term

I
(P [ ])
J
1 1
F
E PRef
S
+ at hour h .
Ref
Gi
*
(
i =1 Var P ( ) j =1 Var P ( )
S S
Lj
*
) ( )
h ,k for k = 1,2,3 , transmission flowgate shadow price forward premium term

I
( [ ]) f Gi ,k f Lj ,k
J
F
E Sk + at hour h .
k
(
Gi )
i =1 Var P S * ( ) j =1 Var P S * ( )
Lj ( )

h,4 the summation of covariance between GENs un-hedged profit and the spot
I
Cov Gi
*
( )
( ), PGiS * ( ) at hour h .
market price
i =1 (
Var PGiS * ( ) )
h,5 the summation of covariance between LSEs un-hedged profit and the spot

J (
Cov Lj* ( ), PLjS * ( ) ) at hour h .
market price j =1 (
Var P ( ) S*
)
Lj

the regression coefficient vector.

h independent error terms of the regression model with expectation 0 and

variance 2

137
or, in the matrix form

Z = H6 61 + H1
H 1
(4.33)

We expect the observational Z to distribute around the regression surface (a

hyperplane in this case since it is more than 2 dimensions) . Since the model (4.33)
does not take all flowgates into consideration, beside, the empirical prices are subject to

some exogenous factors which are not modeled in our model, we take normality

assumption for the error terms h to represent the effect of random factors that affect the

response to some extend and that vary at random without reference to the modeled

explanatory variables . We have h ~ iid .N (0, 2 )

We used rolling average of corresponding empirical data at the specific hour in the
S
previous 14 days for the conditional expectation measures E PRef [ ] and E[ ] . S
k For

( )
conditional variance and covariance measures, Var PGiS * ( ) , Cov Gi
*
( )
( ), PGiS * ( ) , and
Cov ( Lj* ( ), PLjS * ( )) at hour h , we use the corresponding empirical data at the specific

hour in the previous 14 days also. The regulated retail price at each zone is assumed to

be 1.2 times the average of the spot market price throughout the interested time period.

To distinguish the difference between on-peak and off-peak hours, we assume on-peak

and off-peak retail prices and the averages are taken over on-peak and off-peak hours,

respectively.

Since the parameters AG and AL in equation (4.31) are not directly observable based

on the empirical data, we do not incorporate them explicitly in the regression function but

keeping in mind that they are positive numbers for risk-averse market participants as in

the assumptions indicated above. Given that AG and AL should be comparable for the

market participants modeled, the estimation of regression parameters 4 and 5 based on

138
the (4.33) as compared with (4.31) would reveal, approximately, the values of parameters

AG and AL .
The estimates for the regression model coefficients are given in table 4.21,

Table 4. 21 Estimation of regression coefficients


1 2 3 4 5
0.4783 0.3658 0.2943 -0.0248 -0.0093

Based on the normality assumption of the error terms, to test whether there is a

regression relation between the response variable Z and the set of variables, i.e., to

test whether or not for each of its component, an F test can also be conducted to

determine between the alternatives:

H 0 : 1 = 2 = L = 6 = 0 , H : not all p = 0, p = 1, L ,6

with the decision rule to control the type I error at :

If F * F (1 ; K , H (K + 1)) , conclude H 0 , otherwise, conclude H .

The statistical measures including the coefficient of multiple determination R 2 , the F

value, and p-value, together with the serial correlation ( ) of error terms of the

regression model, are displayed in table 4.22 as follow,

Table 4. 22 Statistics of regression coefficients


R2 F p
0.5021 3.5583 0.0053 0.1201

Since the p-values shows that a F statistic as extreme as the observed F would occur

by chance close to once in 200 times if the forward premium term is independent of the

flowgate shadow prices forward premium and the covariance between market

participants un-hedged profit and the spot market price, We conclude that H 0 can be

rejected with significant confidence level.

139
Corresponding to the tests of whether there exists regression relation between the

response variable Z and each of the variables, i.e., the tests whether or not p = 0 for

each p , p = 1,L,5 ,

H 0 : p = 0 , H : p 0

the t values corresponding to each partial regression coefficient p are calculated and

listed in table 4.23,

Table 4. 23 t value of regression coefficients


1 2 3 4 5
t-Value 2.4352 1.9620 1.9703 1.4326 2.3801

The null hypothesis H 0 s can be rejected with confidence as high as 95% for all p s

expect for 4 . Given the small sample size and the difficulty of estimating the

conditional measures, however, we think the results support the model in general.

Although the statistical significance varies according to the individual regression

coefficient, the statistical measures do suggest strong relations exist between the forward

risk premium and the economic factors modeled. The estimates of s are different from

1 as suggested by the model (4.31) can be attributed to the scale difference between AG

and AL and to the fact that, as mentioned in earlier part of the section, the flowgate list as

constructed is not a complete list of flowgates which affect the forward risk premium.

The values of 4 and 5 also indicate that the risk-aversion coefficients AG and AL are

of the magnitudes of 10 3 ~ 10 2 , which are comparable to the values we adopted for the

simulation study in section 4.3.1.

To further test the implications of the forward risk premium model, using the

regression model based estimates of PTDF coefficients listed in table 4.19 and the

140
reference bus forward premium regression coefficients listed in table 4.21, the equation

(4.33) is used to forecast spot market prices.

The out-of-sample spot price data and their projections are plotted in figure 4.9,

800
Historical Price
Forecasted Price
600
Price ($/MWh)

400

200

-200

08/01 08/06 08/11 08/16 08/21 08/26 08/31


Date

Figure 4. 9 Historical and forecasted spot price at the reference bus in August 2006

Note that except for a few spikes of the spot market price and sags of the forecasted

price, the rest sections of the curves fit quite closely. By looking at the shadow price on

the flowgate FG1 as illustrated in figure 4.10, we understand the sags come from the

shadow price spikes.

141
500

400

Price ($/MWh)
300

200

100

0
08/01 08/06 08/11 08/16 08/21 08/26 08/31
Date

Figure 4. 10 Historical day-ahead shadow prices of flowgate FG1 in August 2006

Assuming the normality of the forecast error term (defined as the absolute deviation

of forecasted prices from the realized spot prices), its distribution parameters and their

confidence intervals (CI) are estimated in the table 4.24,

Table 4. 24 Estimates of forecast error term Parameters ($/MWh)


95% CI of 95% CI of
10.22 [8.97, 11.47] 16.67 [15.85, 17.58] 10.22

By excluding the LBMP and flowgate shadow price (FGP) spikes (defined at certain

quantiles of their corresponding distributions), the forecast error terms distribution

parameters and their confidence intervals (CI) would be as follow,

Table 4. 25 Estimates of forecast error term parameters by excluding LBMP and FGP
spikes defined at different quantiles of their distributions ($/MWh)
Quantile LBMP Spike FGP Spike 95% CI of 95% CI of
95.00% 98.44 76.41 6.98 [5.92, 8.04] 12.96 [12.28, 13.70]
97.50% 120.17 131.45 7.03 [5.93, 8.13] 13.37 [12.67, 14.13]
99.00% 199.79 266.67 11.93 [10.65,13.21] 17.04 [16.20, 17.97]

142
The empirical study with the NYISO data confirms the relations between the forward

premium and the factors revealed in model (4.31). The functional relation is insightful

for the understanding of how the forward premium can be affected by the market

participants as well as the transmission network constraints. Although the statistics in

tables 4.24 and 4.25 not indicate a price forecasting with high accuracy as compared with

some existing models (for example, see Conejo et al., 2005), we expect the predictive

power of the forward premium model can be greatly improved when we have more

information about the transmission network topology. In that case, the PTDF coefficients

can be calculated accurately without using the regression model for inference. At the

same time, more transmission flowgates can be taken into account to determine the

forward premium. With more accurate model coefficients and more explanatory

variables, the model (4.31) can be relied on for practical uses.

4.5 Conclusions and Discussions

This chapter studies the risk premium present in the electricity day-ahead forward

price over the real-time spot price. This study establishes a quantitative model for

incorporating transmission congestion into the analysis of electricity day-ahead forward

risk premium. Through simulations with a three-bus study-system, it is illustrated that

the more frequently transmission congestion happens, the higher the forward prices get at

the load buses. Consistent with the implications of the model, evidences from empirical

studies with the New York electricity market data confirm the significant statistical

relationship between the day-ahead forward risk premium and the shadow price

premiums on transmission flowgates. When applied for the forecasting of next day spot

prices, the accuracy of this model is yet to be improved. One important factor is that the

PTDF coefficients are inferred using historical market price data without the knowledge

of the transmission network structure. However, for a long-term market participant in a

143
specific power pool, who is more familiar with the system transmission conditions, these

PTDF coefficients can be determines analytically. In addition, he can identify more

transmission flowgates and use their shadow price premiums as explanatory variables in

the regression model as well. With such changes, it is foreseeable that the forecasting

accuracy of the model will be improved.

Despite the empirical evidences from the New York electricity market, caution should

be exercised in generalizing the results to other markets where factors, such as market

power and regulators price intervention, play important roles in market clearing.

Future expansion of the model can be made to take into account the generation

reserve markets in addition to the energy wholesale markets since market participants

decisions in these markets are interdependent. In addition, since the existence of non-

zero forward risk premiums provide incentives for market speculators to take positions in

the electricity derivatives markets, their behaviors and the impact on forward premiums

need to be studies as well.

144
CHAPTER 5

CONCLUSIONS AND FUTURE RESEARCH

Through integrated modeling of power system operations and market risks, this

dissertation addresses a variety of important issues faced by electricity market

participants on market signals modeling, generation capacity scheduling, and electricity

forward trading. The findings of the research and the directions for future research are

summarized below.

5.1 Market Signals for Transmission Adequacy Investment

Transmission investment in the competitive market environment remains an open

question which enjoys on-going debate among academic and industry researchers. By

addressing the central problem of modeling market signals, chapter two investigates the

possibility of establishing a LMP-based transmission pricing mechanism which values

transmission capacity scarcity and reliability enhancements. In contrast with the common

practice of using DC power flow formulations for market dispatch, we advocate the use

of an AC power flow formulations instead since it allocates transmission losses correctly

and reveals the economic incentives of voltage requirements. By incorporating reliability

constraints in the market dispatch, the resulting market prices yield incentives for market

participants to invest in additional transmission capacity and support competition in the

electricity market. To combine the strength of the market simulation based fundamental

approach and the stochastic model based technical approach, market prices obtained

through system simulations are used to calibrate the parameters of a quantile-based

GARCH model. The changes of system conditions can be captured by the stochastic

145
model parameters. This research work builds a power tool for projecting the dynamics of

future market signals, which is essential for the evaluation of transmission adequacy

investments.

Future research topics include identification of price distortion, coordination of

generation and transmission investments, and strategic interactions between transmission

investors. Finding a reduced form hub-and-spoke representation of a complex system to

improve trading liquidity, the adequacy of market incentives are intended to be

investigated in the near future as well.

5.2 Hydroelectric Generation Scheduling

Put in a volatile environment with various uncertainties imbedded, market

participants in the electricity markets need to find operational strategies to hedge against

the market risks imbedded. The operation of a hydroelectric unit considering market

participation and market price uncertainties is studied in chapter three. A profit-pursuing

hydroelectric producer, who participates in the electricity energy spot market and

ancillary services markets as a price taker, need to balance the market opportunities on

both sides with respect to limited water resource over a pre-specified time horizon. A co-

optimization modeling framework that incorporates market participation and market price

uncertainties into the capacity allocation decision-making problem through a stochastic

programming formulation is presented. Compared with a deterministic model, this

approach yields scenario-dependent scheduling strategies which effectively hedge the

potential revenue loss cause by unfavorable market price changes and the ancillary

service requests. The advantages of the proposed model are illustrated through a

computational study with realistic data provided by a hydroelectric producer. This

formulation can be adapted for determining optimal strategies to accommodate other

sources of uncertainties such as unplanned unit maintenance and water inflow.

146
Extensions of this model deserve further research include: To incorporate the market

price uncertainty without a projected hourly market price process, an improvement can be

made by using correlated continuous-time stochastic processes to model energy and

ancillary service market prices. Impact of a producers decisions on the market prices,

especially on the ancillary service prices, can be modeled to make the model more

realistic. Another fruitful direction is to extend the model to address the co-optimization

problem for a power producer who owns a portfolio of hydro, thermal and other types of

generation units.

5.3 Electricity Trading with Forward Hedging

By modeling market participants trading strategies in the day-ahead forward and spot

electricity wholesale markets, chapter four of the dissertation analyzes the risk premium

present in the electricity day-ahead forward price over the real-time spot price. The

decision problems of generators and load serving entities are formulated to maximize

their respective risk-adjusted profits. These problems are solved simultaneously with

respect to the market equilibrium conditions and transmission network constraints. This

study establishes a quantitative model for incorporating transmission congestion into the

analysis of electricity day-ahead forward risk premium. Through simulations with a

three-bus study-system, it is illustrated that the more frequently transmission congestion

happens, the higher the forward prices get at the load buses. Evidences from empirical

studies with the New York electricity market data confirm the significant statistical

relationship between the day-ahead forward risk premium and the shadow price

premiums on transmission flowgates.

Future expansion of the model can be made to take into account the generation

reserve markets in addition to the energy wholesale markets since market participants

decisions in these markets are interdependent. In addition, since the existence of non-

147
zero forward risk premiums provide incentives for market speculators to take positions in

the electricity derivatives markets, their behaviors and the impact on forward premiums

need to be studies as well.

148
APPENDIX A

PROCEDURES OF SENSITIVITY COEFFICIENT CALCULATION

According to equation (2.11), we have

d h
hk = k + x Q (x Q )
T
akh,,ns = (A.1)
d (s n ) sn

Assuming flowgate k is a transmission line m n , we have

hk = (Pmn )2 + (Qmn )2 (A.2)

where: Pmn = Gmn (Vmr2 + Vmi2 VmrVnr VmiVni ) + Bmn (VmrVni VmiVnr ) + Gsmn (Vmr2 + Vmi2 )

Qmn = Gmn (VmrVni VmiVnr ) + Bmn (VmrVnr + VmiVmi Vmr2 Vmi2 ) Bsmn (Vmr2 + Vmi2 )

Vmr = vm cos( m ) , Vmi = vm sin ( m )

hk
Since hk is not an explicit function of sn , we have =0.
sn

Therefore, equation (A-1) becomes

g (x , u )
1
d h
hk = k Q Q (x Q ) (A.3)
d (s n ) x Q x Q

g (x , u )
1

=
1 Pmn Pmn + Qmn Qmn Q Q (x Q )
hk x Q x Q x Q

where the non-zero components of the partial differential vector corresponds to the

[Vkr , Vki , Vmr , Vmi ] in the state vector x Q .

149
0 0
M M

Gmn (2Vmr Vnr ) + BmnVni + 2GsmnVmr GmnVni + Bmn (Vnr 2Vmr ) 2 BsmnVmr

Gmn (2Vmi Vni ) BmnVnr + 2GsmnVmi GmnVnr + Bmn (Vni 2Vmi ) 2 BsmnVmi
Pmn , Qmn = G V + B V
= GmnVmr BmnVmi
x Q x Q mn mi mn mr

GmnVmi + BmnVmr GmnVmr + BmnVmi
0 0

M M

0 0

150
APPENDIX B

SOLVING FOR EQUILIBRIUM FORWARD AND SPOT PRICES

According to the characterization of PTDF matrix in the day-ahead forward

wholesale market, take equations (4.15) and (4.20) into (4.23), we get the following,
I PRef
F
[ ]
f Gi ,k kF E PGiS * ( ) Cov Gi
*
(
( ), PGiS * ( ) )
f Gi ,k
(
AGVar PGiS * ( ) ) +
Var PGiS * ( ) ( )
i =1


J
f Lj ,k
[ ]
E PLjS * ( ) PRef
F
(
+ f Lj ,k kF Cov Lj* ( ), PLjS * ( )

) = T
j =1

A LVar (
P S*
Lj ( ) ) Var PLjS * ( ) ( )

k (B.1)

Taking (4.5a) into (B.1), we have,

I f Gi ,k J f Lj ,k
F
PRef (+ )
i =1 A Var P S * ( ) j =1 A Var P S * ( ) ( )


G Gi L Lj

I

( f Gi ,k )
2 J ( f Lj ,k )
2

i =1 AGVar PGiS * ( )
F
+

k
( )
j =1 ALVar PLj ( )
S*
( )

I f E PGiS * ( )
= Gi ,k
[ J ]
f Lj ,k E PLjS * ( ) [ ] + T
i =1 A Var P S * ( )
+
G Gi ( )
j =1 ALVar PLj ( )
S*
( ) kF

I f Gi ,k Cov Gi


*
( +
) (
( ), PGiS * ( ) J f Lj ,k Cov Lj* ( ), PLjS * ( ) )
i =1
Var PGiS * ( ) ( )
j =1 Var PLjS * ( ) ( )

(B.2)

Combined with market balancing condition (4.22), we have (4.24a-c).

151
APPENDIX C

SOLVING FOR FORWARD RISK PREMIUM

According to the market balancing conditions in the day-ahead forward markets


I J

Q
i =1
F
Gi = QLjF
j =1
(C.1)

Taking (4.15) and (4.20) into (C.1)


I
[
PGiF E PGiS * ( ) Cov Gi
*
( ]
), PGiS * ( ) ( )


i =1 AG Var PGi ( )
S*
+
(
Var PGiS * ( ) ) ( )

J
=
[ ]
E PLjS * ( ) PLjF Cov Lj* ( ), PLjS * ( )

( )
j =1 ALj Var PLj ( )
S*
(
Var PLjS * ( ) ) ( )

(C.2)

Based on

[
E PnS ( ) = E PRef
S
] [
( ) f n,k S E FGkSS ( ) ] [ ] (C.3)

(C.2) can be rewritten as,


I (PF
) ([
f Gi ,k kF E PRef
S
]
( ) f Gi ,k E Sk ( ) [ ]) + I
(
Cov Gi
*
( ), PGiS * ( ) )

i =1
Ref

AGVar P ( ) ( S*
)
i =1 (
Var PGiS * ( ) )
Gi

=
I (E [P ( )] fS
Ref Lj , k [
E Sk ( ) ) (PRef
F
] f Lj ,k kF )

J Cov ( Lj* ( ), PLjS * ( ))
ALVar (PLjS * ( )) Var (PLjS * ( ))
(C.4)
i =1 j =1

Reorganize the items of (C.4) leads to (4.30b)

152
REFERENCES

Alonso, J., Trias, A., Gaitn, V., and Alba, J. (1999), Thermal plant bids and market
clearing in an electricity pool: minimization of costs vs. minimization of consumer
payments, IEEE Trans. Power Syst., 14(4): 1327-1334.

Alvarado, F. (2003), Converting system limits to market signals, IEEE Trans. Power
Syst., 18(2): 422-427.

Arroyo, J. and Conejo, A. (2000), Optimal response of a thermal unit to an electricity


spot market, IEEE Trans., Power Syst., 15(3): 1098-1104.

Bessembinder, H. and Lemmon, M. (2002), Equilibrium pricing and optimal hedging in


electricity forward markets, The Journal of Finance, LVII(3): 1347-1382.

Billinton, R. and Li, W. (1994), Reliability assessment of electric power systems using
Monte Carlo methods, Plenum, New York.

Birge, J. and Louveaux, F. (1997), Introduction to Stochastic Programming, Springer,


New York.

Bloyd, C., Bharvirkar, R., and Burtraw, D. (2002), Investment in electricity transmission
and ancillary environmental benefits, Resources for the future, Discussion Paper.

Bobo, D., Mauzy, D., and Trefny, F. (1994), Economic generation dispatch with
responsive spinning reserve constraints, IEEE Trans. Power Syst., 9(1): 555-559.

Bohn, R., Caramanis, M., and Schweppe, F. (1984), Optimal pricing in electrical
networks over space and time, Rand J. Economics, 15(3): 360-376.

Borenstein, S., Busnhell, J., Kanh, E., and Stoft, S. (1995), Market power in California
electricity markets, Utilities Policy, 5(3-4): 219236.

Borenstein, S., Busses, M., and Kellogg, R. (2006), Security of Supply and forward price
premia: evidence from natural gas pipeline rates, working paper, University of
California.

Caramanis, M., Bohn, R., and Schweppe, F. (1982), Optimal spot pricing: practice and
theory, IEEE Trans. Power Apparatus and Systems, PAS-101(9): 3234-3245.

153
Carpentier, P., Cohen, G., Culioli, J., and Renaud, A. (1996), Stochastic optimization of
unit commitment: A new decomposition framework, IEEE Trans. Power Syst., 11(2):
1067-1073.

Chen, J., Thorp, J., Thomas, R., and Mount, T. (2003), Locational pricing and scheduling
for an integrated energy-reserve market, Proceedings of the 36th Hawaii
International Conference on System Sciences, Hawaii, US.

Conejo, A., Arroyo, J., Contreras, J. and Villamor, F. (2002), Self-scheduling of a hydro
producer in a pool-based electricity market, IEEE Trans. Power Syst., 17(4): 1265-
1272.

Conejo, A., Plazas, M., Espnola, R., and Molina, A. (2005), Day-ahead electricity price
forecasting using the wavelet transform and ARIMA models, IEEE Trans., Power
Syst., 20(2), 1035-1042.

Dantig, G. and Infanger, G. (1993), Approaches to stochastic programming with


application to electric power systems, in: Frauendorfer, K., Glavitsch, H., and Bacher,
R. (eds), Optimization in planning and operation of electric power systems, Physica-
Verlag, Heidelberg, 125-138.

Dantzig, G. and Infanger, G. (1997), Intelligent control and optimization under


uncertainty with application to hydro power, European Journal of Operational
Research, 97(2):396-407.

Deng, S. (1999), Financial methods in competitive electricity markets, Ph.D dissertation,


University of California at Berkeley.

Deng, S., Jiang, W. (2004), Quantile-based probabilistic models with an application in


modeling electricity prices, in: Bunn, D. (eds.), Modeling prices in competitive
electricity markets, John Wiley & Sons, 161-176.

Eydeland, A., and Geman, H. (1998), Pricing power derivatives, Risk, October, 7174.

Ferrero, R., Rivera, J., and Shahidehpour, M. (1998), A dynamic programming two-stage
algorithm for long-term hydro thermal scheduling of multi-reservoir systems, IEEE
Trans., Power Syst., 13(4): 1534-1540.

Hirst, E. (2004), U.S. Transmission capacity: present status and future prospects.
Technical report prepared for Edison Electric Institute and US Department of Energy.

Hogan, W. (1992), Contract networks for electric power transmission, Journal of


Regulatory Economics, 4(3): 211242.

154
Hogan, W. (1997), A market power model with strategic interaction in electricity
networks, The Energy Journal, 18(4): 107141.

Jacobs, J., Freeman, G., Grygier, J., Morton, D., Schultz, G., Staschus, K., and Stedinger,
J. (1995), Socrates: A system for scheduling hydroelectric generation under
uncertainty, Annals of Operations Research, 59: 99-133.

Jiang W. (2000), Some simulation-based models towards mathematical finance, Ph.D


Dissertation, University of Aarhus.

Joskow, P. and Tirole, J. (2003), Merchant transmission investment, National Bureau of


Economic Research working paper series, Working Paper 9534.

Kang, S. (2001), A new approach for power transaction evaluation and transfer capability
analysis, Ph.D dissertation, Georgia Institute of Technology.

Li, Z. and Shahidehpour, M. (2005), Security-constrained unit commitment for


simultaneous clearing of energy and ancillary services markets, IEEE Trans., Power
Syst., 20(2): 1079-1088.

Longstaff, F., and Wang, A (2004), Electricity forward prices: a high-frequency empirical
analysis, The Journal of Finance, LIX(4): 1877-1900.

McCalley, J., Dorsey, J., Qu, Z., Luini, J., and Filippi, J. (1991), A new methodology for
determining transmission capacity margin in electric power systems, IEEE Trans.
Power Syst., 6(3): 944-951.

Meliopoulos, A. (2001), Power system modeling, analysis, and control, Class Notes for
ECE 6320, Georgia Institute of Technology (to be published by Marcel Dekker).

Meliopoulos, A., Cokkinides, G., and Chao, X. (1990), A new probabilistic power flow
analysis method, IEEE Trans. Power Syst., 5(1): 182-190.

Meliopoulos, A., Cheng, C., and Xia, F. (1994), Performance evaluation of static security
analysis methods, IEEE Trans. Power Syst., 9(3),

Meliopoulos, A., Kovacs, R., Reppen, N., Contaxis, G., and Balu, N. (1988), Power
system remedial action methodology, IEEE Trans. Power Syst., 3(2): 500-509.

Mo, B., Gjelsvik, A., and Grundt, A. (2001), Integrated risk management of hydro power
scheduling and contract management, IEEE Trans. Power Syst., 16(2): 216-221.

North American Electric Reliability Council (2004), Longterm reliability assessment.

Ong, J. (1996), New paradigms for electric power, Derivatives Strategy, Dec./Jan. 1996.

155
Otero-Novas, I., Meseguer, C., Battle, C., and Alba, J. (2000), A simulation model for a
competitive generation market, IEEE Trans. Power Syst., 15(1): 250256.

Overbye, T., Cheng, X., and Sun, Y. (2004), A comparison of the AC and DC power
flow models for LMP calculations, Proc. the 37th Annual Hawaii International
Conference on System Sciences, 47-55.

Ozturk, U., Mzaumdar, M., and Norman, B. (2004), A solution to the stochastic unit
commitment problem using chance constrained programming, IEEE Trans. Power
Syst., 19(3): 1589-1598.

Peng, L., (2005), ARCH and GARCH models. Class Notes for Math 8813, Georgia
Institute of Technology.

Philpott, A., Craddock, M., and Waterer, H. (2000), Hydro-electric unit commitment
subject to uncertain demand, European J.ournal of Operational Research, 125: 410-
424.

Pirrong C. and Jermakyan, M. (2005), Valuing power and weather derivatives on a mesh
using finite difference methods, in: Kaminski, V. (eds.) Energy modeling: advances
in the management of uncertainty, Risk Books, London.

Plazas, M., Conejo, A., and Prieto, F. (2005), Multimarket optimal bidding for a power
producer, IEEE Trans. Power Syst., 20(4): 2041-2050.

Routledge, B., Seppi D., and Spatt, C. (2000), Equilibrium forward curves for
commodities, Journal of Finance, 55, 12971338.

Routledge, B., Seppi D., and Spatt, C. (2001), The spark spread: an equilibrium model
of cross-commodity price relationships in electricity, Working paper, Carnegie
Mellon University.

Schweppe, F., Caramanis, M., Tabors, R., and Bohn, R. (1988), Spot pricing of electricity,
Boston, MA: Kluwer Academic Publishers.

Schweppe, F., Caramanis, M., and Tabors, R. (1985), Evaluation of SPOT price based
electricity rates, IEEE Trans. Power Syst., PAS-104(7): 1644-1655.

Schweppe, F., Tabors, R., Kirtley, J., Outhred, Jr.H., Pickel, F., and Cox, A. (1980),
Homeo-static utility Control, IEEE Trans. Power Syst., PAS-99(3): 1151-1163.

Sun, H., Deng, S., and Meliopoulos, A. (2005), Impact of market uncertainty on
congestion revenue right valuation, Journal of Energy Engineering, 131(2): 139-156.

Takriti, S., Birge J., and Long, E. (1996), A stochastic model for the unit commitment
problem, IEEE Trans. Power Syst., 11(3): 1497-1508.

156
Wallace, S. and Fleten, S. (2003), Stochastic programming models in energy, in:
Ruszczynski, A. and Shapiro, A. (eds.), Stochastic programming, handbooks in
operations research and management science, 10: 637-677. ELSEVIER, Great
Britain.

Wood, A. and Wollenberg, B. (1996), Power generation, operation, and control, John
Wiley & Sons, New York.

157

Vous aimerez peut-être aussi