Multi-agent systems: an investigative tool to study electricity markets

1. Multi-agent systems: an investigative tool to study electricity markets Toshiyuki Sueyoshi New Mexico Institute of Mining & Technology and National Cheng Kung University

2. Overview • Introduction • Objective • Multi-Agent Systems • Application • California Electricity Market • Analysis Results

3. Motivation • Summer 2000 • Electricity Prices in California were approx. 500% higher than those during the same months in 1998-1999 • Explanations available in Literature • Drought in northwest pacific region • Increase in the price of natural gas • “Unusually hot” summer • Increase in power demand due to economic growth of industry and due to opening of many startup companies

4. Questions • Can intelligent agents predict a price change in a competitive market with many constraints? • Can the agent structure provide a framework for analysis of a dynamic system? • Can the intelligent system explain the occurrence of the California Electricity crisis?

5. Approach • Build a multi-agent platform such that agents can interact with the environment • Implement learning algorithms in the agents • Represent the transmission constraints in the model • Compare with other approaches • Perform an analysis of the different parameters affecting the system

6. Important Features • Transmission constraints • Multiple Learning Rates • Partial Reinforcement Learning [Bereby-Meyer and Roth (2006)] • Trader cannot always win • Law of Effects [Erev and Roth (1998)] • “choices that have led to good outcomes in the past are more likely to be repeated in the future”

7. Important Features • Power Law of Practice [Erev and Roth (1998)] • “learning curve tended to be steep initially and then be flat” • Learning from Mistakes [Chialvo and Bak (1999) and Si and Wang (2001)]

8. Three Zones in CA ISO (Source: http://www.ucei.berkeley.edu/)

9. Market Clearing Scheme for Multiple Zones

11. Supply Demand z zone DA (Financial) the z-th zone market z zone HA (Physical) the z-th zone market

12. Supply Demand z zone DA (Financial) the z-th zone market z’ zone z zone HA (Physical) the z-th zone market z’ zone

13. Two Types of Learning • Type I – Partial Reinforcement Learning • Type II – Myopic Learning

14. Data description on CA market price • Before crisis period • SP-15 and NP-15 • 1st April, 1998 to 30th April, 2000. • ZP-26 • 1st February, 2000 to 30th April, 2000. • During crisis period • 1st May 2000 to 31st January 2001.

15. Data description on CA market price • DA markets, • a maximum price of $2499.58/MWH was observed at 7 PM on 21st January 2001. • HA markets • a maximum price of $750/MWH starting from 26th June 2000. • It was observed that prices started rising steadily from the summer of 2000.

16. Market Composition-Generators((http://www.energy.ca.gov/maps/electricity_market.html) ) • 964 generators • 343 are hydroelectric with 20% market capacity, • 44 are geothermal with 3% market capacity • 373 are oil/gas with 58% market capacity • 17 are coal with 6% market capacity • 94 are wind with 4% market capacity • 80 are WTE with 2% market capacity • 2 are nuclear with 7% market capacity • 11 are solar with 1% market capacity.

17. Market Composition-Wholesalers(http://www.energy.ca.gov/ electricity/electricity_consumption_utility.html ) • 48 wholesalers • Pacific Gas and Electric has 30% of the share • San Diego Gas & Electric has 7% of the share • Southern California Edison has 31% of the share • LA Department of Water and Power has 9% of the share. • Sacramento Municipal Utility District has 4% of the share, • California Dept. of Water Resources has 3% of the share, and other 41 utilities have a 12% share. • Self-generating agencies account for 4% of the share.

18. Estimation Accuracy with different combinations of traders

19. Estimation Accuracy Comparison (%) among four approaches

20. MAIS Price Estimate - SP-15 – Before Crisis

21. MAIS Price Estimate - SP-15 – During Crisis

22. Findings • Significant difference between observed market prices and MAIS estimates • May not be explained as a natural outcome of changes in market fundamentals • Price range before crisis • from $0/MWH to $90/MWH • Price range during crisis • from $25/MWH to $400/MWH

23. Learning Speed

24. Findings • The learning speeds of agents depend upon the dynamic change of market fundamentals. • Before the crisis, agents can adjust themselves to a market change within a shorter time period because the market is stable and predictable. In contrast, they need a long time for their learning during the crisis because a large market fluctuation occurs in the market.

25. Agent Behavior in crisis – Type I and Type II Composition

26. Analysis – Which of these market fundamentals caused the CA crisis? • Increase in Marginal Cost Joskow and Kahn (2002) and Borenstein et al (2002) • Increase in Real Demand (Lee 2004) • Greed of Traders • Electricity withholding by Generators Joskow and Kahn (2002, pp. 19-28) and Borenstein et al. (2002) • Capacity limit in Transmission Lines Joskow and Kahn (2002, p.8) • Competitive Rent Borenstein et al. (2002, p.1397) • A Combination of the above

27. Zero and One Market Fundamental

28. Two Market Fundamentals

31. Findings • Transmission capacity and competitive rent did not have a major influence • 40.46%[= (83.96%-73.06%) / (100%-73.06%)] of the price increase during the crisis was due to an increase in marginal cost • 17.85% [= (88.77%-83.96%) / (100%-73.06%)] to traders’ greediness

32. Findings • 5.27% [= (90.19%-88.77%) / (100%-73.06%)] to a real demand change • 3.56% [= (91.15%-90.19%) / (100%-73.06%)] to market power (withholding electricity). • The remaining 32.86% is from other unknown market components and an estimation error.

33. Findings • The price hike during the crisis occurred due to an increase in fuel prices and real demand at the level of 45.73% (= 40.46%+5.27%). • The responsibility of energy utility firms (their greediness and a use of market power) was 21.41% (= 17.85%+3.56%). • The increase of fuel price and real demand was twice more influential than the responsibility of energy utility firms in terms of the price hike and large fluctuation during the crisis.

34. Related Publications • T. Sueyoshi, G.R. Tadiparthi, A Wholesale Power Trading Simulator with Learning Capabilities, IEEE Transactions on Power Systems, 20(3), (2005), pp. 1330 – 1340 • T. Sueyoshi, G.R. Tadiparthi, An Electric Power Trading System: Network-based Framework and Simulator with Learning Capabilities, International Journal of Operations Research, 3(3), (2006), 193-203 • T. Sueyoshi, G.R. Tadiparthi, Agent-based Approach to Handle Business Complexity in US Wholesale Power Trading, IEEE Transactions on Power Systems, 22(2), (2007), 532-543

35. Related Publications • T. Sueyoshi, G.R. Tadiparthi, An Agent-based Decision Support System for Wholesale Electricity Market, Decision Support Systems, (2008) Vol. 44, No. 2, pp. 425-446. • T. Sueyoshi, G.R. Tadiparthi, Wholesale Power Price Dynamics under Transmission Line Limits: A Use of Agent-based Intelligent Simulator, IEEE Transactions on Systems, Man, and Cybernetics--Part C: Applications and Reviews , (2008), (Forthcoming) • T. Sueyoshi, G.R. Tadiparthi, Why did California Electricity Crisis Occur? A Numerical Analysis Using Multi-Agent Intelligent Simulator, IEEE Transactions on Systems, Man, and Cybernetics--Part C: Applications and Reviews, (2008), (Forthcoming)