Department of Telecommunications MASTER THESIS Nr. 608 MASTER THESIS Nr. 608 INTELLIGENT TRADING AGENT FOR POWER TRADING THROUGH WHOLESALE MARKET Ivo Buljević.

Department of Telecommunications

MASTER THESIS Nr. 608

INTELLIGENT TRADING AGENT FOR POWER TRADING THROUGH

WHOLESALE MARKET

Ivo Buljević

2012/2013

Zagreb, July 2013

Department of Telecommunications

Contents

¨ Introduction¨ Smart grid¨ Wholesale market¨ CrocodileAgent 2013¨ Conclusion

Zagreb, July 2013 2 of 12

Department of Telecommunications

Introduction

¨ Characteristics of the traditional energy market: Centralized Vertically integrated market structure No competition

¨ Liberalization and deregulation of the traditional energy market

¨ Increased number of renewable energy sources ¨ Progressive transformation of traditional power

systems into evolved systems called smart grids

Zagreb, July 2013 3 of 12

Department of Telecommunications

Smart grid

¨ A modernization concept of the electricity delivery system¨ Enables real-time banacing of energy supply and demand¨ A two-way flow of electricity and information

Zagreb, July 2013 4 of 12

¨ Multi-agent market models Entities are represented by

intelligent software agents Opportunity to test software

solutions in order to prevent market crashes (California 2001)

Department of Telecommunications

Wholesale market

¨ Result of liberalization and deregulation of the traditional energy market, enables energy trade between market entities

¨ Power exchanges and power pools¨ Day-ahead market¨ Examples of wholesale markets:

Chile Great Britain and Wales Nord Pool California

Zagreb, July 2013 5 of 12

Department of Telecommunications

Wholesale market (2)

¨ Energy load forecasting Statistical approach

Similar-day method Exponential smoothing Regression methods

Artifficial intelligence – based tecniques Reinforcement learning

¨ Energy price forecasting Spike preprocessing Time series models with exogenous variables Interval forecasts

Zagreb, July 2013 6 of 12

Department of Telecommunications

CrocodileAgent 2013

¨ Intelligent software agent developed at University of Zagreb

¨ Participant of PowerTAC 2013¨ Main emphasis:

Zagreb, July 2013 7 of 12

Development of wholesale bidding strategy which will minimize negative effects on the balancing market

Responsive and context-aware agent design

Department of Telecommunications

CrocodileAgent 2013Modular architecture

Zagreb, July 2013 8 of 12

CLEARINGINFORMATION

WHOLESALE MARKET

CUSTOMER MARKET

GENCOS OTHER BROKERS

C1 C2 C3

WEATHER

PAST ENERGY USAGE

ALL FORECASTED DATA

PAST CLEARING PRICES

BID/ASK TARIFFS

CONSUMPTION PRODUCTION INTERUPTABLE CONSUMPTION

Office complexVillage types

Centerville homes

Solar

Wind

Frosty storageHeat Pump

FORECAST MANAGER

TARIFF MANAGER MARKET MANAGER

MARKET REPOSITORY

TARIFF REPOSITORY

MAIN SERVICE (MESSAGE SENDER/

RECEIVER)

OTHER TARIFF SPECIFICATION, TRANSACTION

PUBLISH TARIFFS

PASTUSAGE

FUTURE ENERGY USAGE/PRICES

CURRENT WHOLESALESTATE

BIDS/ASKS

SEND TO SERVER

CrocodileAgent 2013

LEARNING MODULE

BIDDING STRATEGIES

GENERATEDORDERS

ENERGYPRICES

NEEDEDENERGY

Contribution of this master thesis

Department of Telecommunications

CrocodileAgent 2013Learning module

¨ Based on reinforcement learning Erev-Roth method specially adapted for PowerTAC

wholesale market¨ Enables broker to adapt to various market

conditions¨ Key features:

Zagreb, July 2013 9 of 12

Multiple strategies Advanced strategy

evaluation based on its efficiency

RL module Simulator

InitializationChoose strategy

ExecuteResults

Set rewards

Department of Telecommunications

CrocodileAgent 2013Learning module (2)

¨ Uses basic order as an input Generated by forecast module, based on past usage of

subscribers on the retail market Holt-Winters method

¨ Life cycle:

Zagreb, July 2013 10 of 12

Initialization Choose strategy Place order Set reward

¨ Strategies used to model amount of energy and unit price

Department of Telecommunications

CrocodileAgent 2013Results

¨ Broker progressively learns to adapt to current market conditions – manifestation of the learning period Minimization of balancing cost

¨ Broker buys an excessive amount of energy on the wholesale market

Zagreb, July 2013 11 of 12

Results from May trial indicates that broker buys 125% of energy needed on the retail market

A need to optimize basic order generation (energy load forecasting)

Department of Telecommunications

Conclusion

¨ Robustness of the CrocodileAgent’s wholesale module Broker is able to adapt to changes in competition

environment¨ Adapted Erev-Roth algorithm was proved to be

suitable for the PowerTAC wholesale market¨ Future work:

Improvement of energy load forecasting Improvement in unit price calculation Design of intelligent strategies

Zagreb, July 2013 12 of 12