Agent-Based Energy 1-26-04

Evaluating Individual Market Power in Electricity Markets
Using
Agent-Based Simulation
Fernando Oliveira
[email protected]
A Presentation at the FERC

Washington
January 2004
WHY AGENT-BASED SIMULATION?
It can be used to represent:
• Discrete supply and demand functions

• The detailed internal structure of each firm
• The interaction processes in complex markets:
a) Analysing the relation between bilateral
markets and the balancing mechanism.
b) Studying the impact of forecasting errors on the
efficiency of the industry.
c) Evaluating the impact of the new market rules
on the performance of generators and suppliers.
• The learning processes created by repeated
interactions
WHY LEARNING?
• To reflect Bounded-rationality
• To analyse situations where players interact repeatedly
• Consistency with theory (organisations are adaptive

systems)
• To approximate solutions in non-linear problems
• To analyse problems in which no equilibrium solution

can be computed
a) There is no equilibrium
b) There are several equilibria
c) The problem is computationally too complex!!
Bunn and Day
• Analyse the impact of regulatory divestment, using a
computational model, based on the supply function
equilibrium.
• They simulate the supply function competition, using:

•Best response framework
•Assuming kinked, discontinuous supply functions
• They found no convergence to a unique supply function

equilibrium: a cyclic behaviour emerged instead
• They argue that the 1999 divesture of 40% joint capacity of

National Power and PowerGen would not be sufficient to
completely erode the possibility of market power.
Bower and Bunn
• Look at the switch from uniform to discriminatory
pricing.
• They model:
•Discrete supply functions
•Different marginal costs for each technology
•Interactions between players in a repeated game.
• The players use a rule of thumb that balances pricing and

generation quantities.
• It does not capture the interaction between the bilateral

trading and the balancing market.
• It does not take into account the supply side.

Bunn and Oliveira
• Extends the analysis of Bower and Bunn.
• The simulation platform developed here is a much more
detailed representation of how NETA was designed to work:
1. It actively models the demand side (suppliers).
2. It models the interactions between the bilateral market

and the balancing mechanism, and the settlement process.
3. It takes into accounts the daily dynamic constraints.
4. It uses learning to allow players to infer a pricing policy.

MOTIVATION AND RESEARCH OBJECTIVES
• In early 2000, Ofgem decided to modify the licenses of the

larger generators to include a condition prohibiting abuse of
market power.
• Six generators decided to sign the new licence but AES

and British Energy refused. The Competition
Commission had to resolve this dispute.
• Research question: Do these two apparently small

companies have enough “market power” to operate
against the public interest?
THE ENGLAND AND WALES GENERATION
INDUSTRY
Capacity of each Company (% of Total, 59 GW) in 2000
Total Nuclear Large Coal+CCGT Small Coal +OCGT + OIL + Pump. Storage
PG 16.5 17.8 24.9
NP 13.9 14.8 22.5
BE 12.4 58 7.1
Edison 10.6 9.1 30.7
TXU 9.7 10.5 14.7
AES 7.8 9.2 6.8

EDF 4.7 27.8 1.8
Magnox 3.9 5.1
Others 20.5 14.1 24.7 0.4
Total GW 59 7 45 7
THE ENGLAND AND WALES SUPPLY INDUSTRY
Peak Demand 50 GW in 2000

Market Share of each Company (%)
London 12.2 PG 10.2
Electricity
Eastern 11.2 Scot. & Southern 9.9
NP 10.2 Others 40.7
NETA’S STRUCTURE
1. Forward and Futures Market 3. Balancing Mechanism
Futures and forward contracts 3.5 hours ahead

Years, months, days ahead The SO accepts offers or bids from generators and suppliers
Generators and suppliers buy and sell Pay as bid auction
Future and forward prices
2. Day-Ahead Power Exchanges 4. Settlement Process
One day-ahead trading The SO charges the companies if

Generators and suppliers buy and sell contract positions <> metered volumes
Power exchange prices System sell price (SSP)
System buy price (SBP)
NETA MODEL
For the number of trading days specified
1. Suppliers predict demand for each hour
2. Generators define which plants can run
3. Generators offer in the PX
4. Suppliers bid in the PX
5. Trading in the PX and calculation of System Position in each one of the
hours
6. Generators and Suppliers offer (bid) into the BM
7. Trading in the BM and calculation of imbalance prices
8. Settlement Process: calculation of imbalances for each one of the
suppliers and generators (plant by plant)
9. Learning
SUPPLIERS AND GENERATORS
• Objectives: To maximise total daily profits and to attain their

objectives for the Balancing Mechanism exposure.
• Instruments: Define the Mark-ups for the Power

Exchange and for the Balancing Mechanism .
• Each agent uses reinforcement-learning.

GENERATORS’ BEHAVIOUR
For each plant in the portfolio:
1. Verify that the objectives for the BM exposure. If they were not
achieved, penalise the profit obtained.
2. Given the daily profits in the PX and in the BM revise, for the mark-up
used, the:
2.a) Expected profit in the PX and in the BM.
2.b) Expected acceptance rate in the PX and in the BM.
3. Define the new offering policy for the next day.
4. In the beginning of the next day. Define the quantities and prices offered
for each plant, after knowing which plants are available, making sure that
all the operational constraints are respected.

SUPPLIERS’ BEHAVIOUR
1. Verify that the objective for the BM exposure was achieved
If not, penalise the profit obtained in the PX.
2. Given the daily profits in the PX and in the BM revise the:
2.a) Expected profit in the PX and in the BM
2.b) Expected acceptance rate in the PX and in the BM
3. Define the new bidding policy for the next day.
4. In the beginning of the next day. After calculating predictions for
demand, define the quantities and prices bid.

SIMULATIONS
• 80 gensets, owned by 24 Generators

• 13 Suppliers – with 10% demand forecasting error
• 59 GW of capacity in the system
• A peak winter demand of 49 GW, 48 GW and 45 GW
• Summer maintenance outage of 14, 12 and 10 GW
50
45
Demand (GW)
W1
40
W2
35
W3
30 S
25
20
24 hours
A FIRST ANALYSIS
Average daily prices for the Winter (49 GW) Scenario
2 00
£ / MWh 1 50
SB P
1 00
P XP
50
SSP
0
2 0 0 d a y s (i te r a ti o n s )
• Less than 80% of capacity used: PXP Marginal costs.
• 80% - 85% of capacity used: PXP > Marginal costs.
• More than 85% of capacity used: PXP seem

unbounded
STABILITY AND CONVERGENCE
• A winter day, 10 Simulations, 400 iterations by

simulation.
• PXP behaviour (demand 45 GW)
80
70
60
Price (£/MWh)
average
50
maximum
40
minimum
30
difference
20
10
0
400 trading days
STABILITY AND CONVERGENCE - II
• PXP behaviour (demand 49 GW)
450
400
350
Price (£/MWh)
300 average
250 maximum
200 minimu m
150 difference
100
50
0
400 trading days
FIRST RESULTS
• The majority of trading (more than 98%) will take place

in the bilateral markets.
• Imbalances: much higher than the quantities traded.

How high depends on the prediction-errors by suppliers.
• Industry efficiency: will increase if the System Operator

publishes predictions more accurate .
FIRST RESULTS - II
• NETA will have a greater risk impact on suppliers

than on generators.
• Generation collusive behaviour is facilitated by low

Capacity Margins.
•Suppliers’ strategy: Demand forecasting is less

important than to predict the behaviour of the other
suppliers.
• Agent-based computational learning can provide

insights into pricing and strategic behaviour in electricity
markets, which was obtained in 2000 just before its
introduction in 2001.
AES – BE CASE STUDY: THE STRATEGIES
(ST1) AES and BE act as Bertrand Players with Capacity

constraints
(ST2) AES saves capacity for the BM (480 MW)
(ST3) BE saves capacity for the BM (490 MW)
(ST4) AES saves capacity for the BM (480 MW) and closes
a Drax genset (645 MW).
(ST5) BE saves capacity for the BM (490 MW) and closes an

Eggborough genset (480 MW).
(ST6) Strategies (4) and (5) combined.

AES – BE: RESULTS
• t-statistics for the PX price (PXP), the SSP and the

SBP
Strategies
ST2 ST3 ST4 ST5 ST6
PXP 1.12 2.21 2.61 2.37 3.55
SSP 0.23 0.83 1.92 0.26 1.7
SBP -0.15 0.95 0.75 1.98 1.5
• t-statistics for AES' s and BE' s profits
Strategies
ST2 ST3 ST4 ST5 ST6
AES -0.09 1.77 1.31 1.61 2.29
BE -1.03 0.32 0.58 1.57 3.04
CONCLUSIONS
• BE alone, and BE together with AES, may significantly

increase the Price in the Power Exchange and the
System Buy Price .
• AES alone cannot influence the wholesale market

prices.
• Only when acting together may the firms profit from

capacity and prices manipulation.
• The Competition Commission concluded that the

continuation without modification of the provisions of
AES and BE is not against the public interest.
DISADVANTAGES OF AGENT-BASED SIMULATION
• Lack of formalism.
• Unjustified assumptions of certain types of behaviour.
• Can easily be modelled to reflect the mental fears and
wishes of the modeller, instead of reflecting reality.
• Over-parameterisation and difficult parameterisation.
• Difficult to validate.
• “Impossible” to prove that a model is right.
• A model too complex can be perceived as a “black-
box”.
• Causal relationships maybe very difficult to justify,
given the decentralised decision process and the more
or less complex nature of the system modelled.
ADVANTAGES OF AGENT-BASED SIMULATION
• Lack of formalism.
• Modelling of non-linear systems.
• Detailed description of the systems analysed.
• Capture the learning and adaptation processes within a
company.
• Can mimic boundedly rational behaviour: players do
not have to optimise.
• Replicate mental models.
• Can look into the future, in a logic way, enabling
insights into the nature of complex systems, even in
the absence of data.
LANGUAGES AND SOFTWARE
• Visual C++ : Object Oriented, Very Flexible
• JAVA: Object Oriented, Internet Language
• MatLab: Mathematics oriented, fantastic for finance and

forecasting; efficiency problem!
• Excel Visual Basic

MODEL DEVELOPMENT
• Oliveira´s Masters in AI: Java

Development Time: a year
• Bower and Bunn: Excel Visual Basic

Development Time: a year
• Bunn and Oliveira: MatLab

Development time: 9 months
Design, validation and programming: 6 months
Verification and testing: 1 month
First experiments: 1 month
Data Analysis: 1 month
Running time: 50 iterations run in 2 hours, Pentium II
ABS LINKS
• Leigh Tesfatsion: Agent-based Computational Economics

http://www.econ.iastate.edu/tesfatsi/ace.htm
• Gerard Sheble
http://vulcan.ee.iastate.edu/~sheble/
• MIT ENERGY LAB: Marija Iic
• My website
http://phd.london.edu/foliveira/

Agent-Based Energy 1-26-04

Uploaded by

Copyright:

Available Formats

Agent-Based Energy 1-26-04

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Agent-Based Energy 1-26-04

Uploaded by

Copyright:

Available Formats

Evaluating Individual Market Power in Electricity Markets

A Presentation at the FERC

It can be used to represent:

• Discrete supply and demand functions

• To analyse situations where players interact repeatedly

• Consistency with theory (organisations are adaptive

• To approximate solutions in non-linear problems

• To analyse problems in which no equilibrium solution

• They simulate the supply function competition, using:

• They found no convergence to a unique supply function

• They argue that the 1999 divesture of 40% joint capacity of

• The players use a rule of thumb that balances pricing and

• It does not capture the interaction between the bilateral

• It does not take into account the supply side.

1. It actively models the demand side (suppliers).

2. It models the interactions between the bilateral market

3. It takes into accounts the daily dynamic constraints.

4. It uses learning to allow players to infer a pricing policy.

• In early 2000, Ofgem decided to modify the licenses of the

• Six generators decided to sign the new licence but AES

• Research question: Do these two apparently small

Capacity of each Company (% of Total, 59 GW) in 2000

AES 7.8 9.2 6.8

Peak Demand 50 GW in 2000

1. Forward and Futures Market 3. Balancing Mechanism

Futures and forward contracts 3.5 hours ahead

One day-ahead trading The SO charges the companies if

1. Suppliers predict demand for each hour

2. Generators define which plants can run

3. Generators offer in the PX

4. Suppliers bid in the PX

5. Trading in the PX and calculation of System Position in each one of the

6. Generators and Suppliers offer (bid) into the BM

7. Trading in the BM and calculation of imbalance prices

8. Settlement Process: calculation of imbalances for each one of the

suppliers and generators (plant by plant)

• Objectives: To maximise total daily profits and to attain their

• Instruments: Define the Mark-ups for the Power

• Each agent uses reinforcement-learning.

achieved, penalise the profit obtained.

2.a) Expected profit in the PX and in the BM.

2.b) Expected acceptance rate in the PX and in the BM.

3. Define the new offering policy for the next day.

all the operational constraints are respected.

If not, penalise the profit obtained in the PX.

2. Given the daily profits in the PX and in the BM revise the:

2.a) Expected profit in the PX and in the BM

2.b) Expected acceptance rate in the PX and in the BM

3. Define the new bidding policy for the next day.

4. In the beginning of the next day. After calculating predictions for

demand, define the quantities and prices bid.

• 80 gensets, owned by 24 Generators

• Less than 80% of capacity used: PXP Marginal costs.

• 80% - 85% of capacity used: PXP > Marginal costs.

• More than 85% of capacity used: PXP seem

• A winter day, 10 Simulations, 400 iterations by

• PXP behaviour (demand 49 GW)

• The majority of trading (more than 98%) will take place

• Imbalances: much higher than the quantities traded.