IT Licentiate theses 2001-005
Market and Resource Allocation Algorithms
with Application to Energy Control
PER CARLSSON
UPPSALA UNIVERSITY
Department of Information Technology
with Application to Energy Control
BY
PER CARLSSON
May 2001
DEPARTMENT OFCOMPUTING SCIENCE
INFORMATION TECHNOLOGY
UPPSALA UNIVERSITY
UPPSALA
SWEDEN
Dissertation for the degree of Licentiate of Philosophy in Computing Science at Uppsala University 2001
with Application to Energy Control
Per Carlsson
Per.Carlsson@cs.lth.se
Department of Computing Science Information Technology
Uppsala University Box 337 SE-751 05 Uppsala
Sweden
http://www.it.uu.se/
c
Per Carlsson 2001 ISSN 1404-5117
Printed by KFS i Lund AB, Sweden Printing sponsored by EnerSearch AB
The energy markets of todayare markets with ratherfew activeparticipants.
Theparticipantsare,withfewexceptions,largeproducersanddistributors. The
market mechanisms that are used are constructedwith this kind of amarket
situation in mind. With an automatic or semiautomatic approach, the mar-
ket mechanismwould be ableto incorporate a largernumber of participants.
Smaller producers,and evenconsumers, couldtakeanactivepartin themar-
ket. Thegainisinmoreecientmarkets,andduetosmalleructuations in
demand betterresourceusagefromanenvironmentalperspective.
Theenergy markets of the Nordiccountries (as wellasmany others)were
deregulated during the last few years. The change has been radical and the
situation is still rather new. We believe that the market can be made more
ecientwiththehelp ofthedynamicsofthesmallactors.
Theidealisedworldoftheory(ofeconomics)oftenreliesonassumptionssuch
as continuous demand and supply curves. These assumptions are useful, and
they do not introduce problems in the powermarket situation of today, with
relatively few, large, participants. When consumers and small producers are
introduced on themarket,the situation is dierent. Then it is adrawbackif
themarketmechanismcannothandlediscontinuoussupplyanddemand.
The growth in accessibility to computational powerand data communica-
tionsthatwehaveexperiencedinthelastyears(andareexperiencing)couldbe
utilised whenconstructingmechanismsfortheenergymarketsoftomorrow.
Inthis thesiswesuggestanewmarketmechanism,ConFAst, thatutilises
the technological progressto make it possible to incorporate a large number
of active participants on the market. The mechanism does not rely on the
assumptions above. Thegainis amore ecient market with lessuctuations
in demandovertheday.
Tomakethis possiblethere is aneedfor ecient algorithms, in particular
this mechanismrelies on anecientaggregationalgorithm. An algorithm for
aggregationofobjectivefunctionsispartofthisthesis. Thealgorithmhandles
maximisationwith non-concave,evennoisy,objectivefunctions. Experimental
results show that the approach, in practically relevant cases, is signicantly
faster thanthestandardalgorithm.
1 Introduction 1
1.1 TheProject . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 TheProblem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 ConFAst, anApproach to Ecient and yet Fast Power ControlMarkets . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.2 ResourceAllocationwithNoisyFunctions . . . . . . . . . 3
1.3.3 Implementation. . . . . . . . . . . . . . . . . . . . . . . . 4
2 An ApproachtoEcientand yetFastPowerControl Markets 7 2.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 TheStrainsonMarketMechanismsinExtreme Situations 7 2.1.2 TheDynamicsoftheSmallActors . . . . . . . . . . . . . 8
2.1.3 TheTechnologicalProgress . . . . . . . . . . . . . . . . . 10
2.1.4 Discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 MarketMechanismsTheoryandPractice . . . . . . . . . . . . 10
2.2.1 IncreasingCostandDecreasingUtility . . . . . . . . . . . 11
2.2.2 Information . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.3 TheMarket ofTodayandTomorrow . . . . . . . . . . . . 13
2.3 ConFAst,A NewMarketMechanism . . . . . . . . . . . . . . . 14
2.3.1 ConcaveandNon-ConcaveUtility . . . . . . . . . . . . . 15
2.3.2 OutlineoftheConFAstMechanism . . . . . . . . . . . . 16
2.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3 Resource Allocationwith NoisyFunctions 29 3.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.1.1 ResourceAllocation . . . . . . . . . . . . . . . . . . . . . 29
3.1.2 AggregatingObjectiveFunctions . . . . . . . . . . . . . . 30
3.1.3 Anoteonthecomplexityofthegeneralproblem . . . . . 32
3.2 MainResult . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 ExperimentalResults. . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3.1 Tree-structuredAggregation. . . . . . . . . . . . . . . . . 36
3.3.2 TheExperiments . . . . . . . . . . . . . . . . . . . . . . . 37
3.3.3 SummaryoftheExperiments . . . . . . . . . . . . . . . . 43
3.4 TechnicalDetails . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4.1 AlgorithmWithout NoiseFiltering . . . . . . . . . . . . . 44
3.4.2 AlgorithmBasedonNoiseFiltering . . . . . . . . . . . . 46
3.4.3 ConstructingSegmentsBasedonHullFunctions . . . . . 48
3.4.4 FindingallPossibleCandidatesinSegmentsWithConvex andConcaveHullFunctions. . . . . . . . . . . . . . . . . 50
3.4.5 ProofofTheorem3.4.1 . . . . . . . . . . . . . . . . . . . 63
3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Bibliography 65
Introduction
1.1 The Project
The thesis work started in a project, DiMALLOC 1
, that was nancedby the
Swedish National Board for Industrial and Technical Development, NUTEK.
As this project ended, the work hasbeencompleted within the IT in Energy
ProgrammeoftheSwedishResearchInstituteforInformationTechnology,SITI.
ThisprogrammeisoneoffourprogrammesofSITI.ThefocusofITinEnergyis
on(i)localcommunicationsoverthepowerlines,(ii)vulnerabilityandsecurity
in electronic commerce (with focus on power line communication), and (iii)
electronicenergymarkets(withfocusoneconomicalandcomputationalaspects
ofelectronicpowermarkets). Thethesispresentssomeoftheresultsinthelast
area.
A Swedish research company, EnerSearch AB, plays a central partin the
programme together with Blekinge Institute of Technology, Lund University,
andUppsalaUniversity.
The aim of EnerSearch AB that is owned by interests in the energy
sector is to initiate and coordinate research in areasthat in dierent ways
are related to the energy and business sectors. As in the case of the IT in
EnergyProgramme,theresearchisconductedincloserelationwithuniversities.
Current owners of EnerSearch AB are: ABB Automation Products, ECN -
Netherlands Energy Research Foundation, EDP - Electricidade de Portugal,
Iberdrola,IBMUtilityServices,E.ON-Energie,andSydkraft.
1.2 The Problem
Powermarketsaroundtheworldarederegulating. Howtodesignecientpower
marketsisanissuethatisdiscussedamongauthorities,energyutilities,andre-
searchers. Itisaninterdisciplinaryproblem,withmainaspectsineconomicsas
1
DistributedMarketalgorithmsandresourceALLOCation
wellastechnicalelds. Theproblemhasinterestingcomputationalaspectstoo,
andduetothegrowingaccesstocomputationalpoweranddatacommunica-
tioncapabilitiesthe computerscienceapproachisableto opennewmarket
possibilities.
A wayto enhancetheeciencyofpowermarkets istoincorporate alarger
numberof participantsonthe activemarket. Today theend consumeris not
presentonthemarket(withfewexceptions)and,asaconsequence,themarket
has to rely onmore orless accurate predictionsof consumption patterns and
volumes. Thepricesat e.g. theNordic spot market,Nordpool 2
, variessigni-
cantlyfromhourto hour,duetouctuations indemand. Ontheotherhanda
largeshareoftheloadshassuchpropertiesthatthetimeproleoftheirdemand
could be utilised to even outasignicantpartof the uctuations (if they are
incorporatedintothe active powermarket), and this canbedonewith no (or
minor)lossin comfortoreciency fortheconsumer.
The gains of introducing the small participants on the active market are
concerningmostoftheaspectsthatwecanthinkof;whentheyareintroduced
supportedbyapropermarketmechanismtheuctuationsindemand,and
hence in price, becomes smaller, the planning of production will be based on
moreaccuratenumbers,thereislessstressonthetransmissionanddistribution
gridonhighpeakhours,wegetabetterresourceusagefromanenvironmental
perspective,andsoon;awinwinsituation.
Whenconsumersandsmallscaleproducersareintroducedontheactivemar-
ket,specialcarehastobetakenconcerningdiscontinuoussupplyanddemand.
Market mechanisms oftoday, based onstandardtheory of economics, assume
that supplyand demand iscontinuous. Due to this assumptionit ishardand
risk prone to participate on the market for any actor with a non-continuous
curve. Howtoconstructamechanismthat handlesnon-continuoussupplyand
demandis nontrivial.
Thepossibilityofalargescaleintroductionoftheconsumptionsideonthe
activemarketisaconsequenceofthedevelopmentincomputationalpowerand
data communicationsthat hastakenplace during the last years. Theaccessi-
bility to the Internet is growingand can be utilised, great eorts are put on
developmentofinexpensivelocalareanetworks,andmoreandmoreequipment
isenhancedwith(whatisoftenreferredas)intelligence. Inourhomeswehave
(orsoonwearehaving)computationalcapacityinwashingmachinesandrefrig-
eratorsaswellasheating/coolingsystems. Ifitispossibletocommunicatewith
the equipment,there is apotentialto introduceit onthemarket,represented
bysmallpiecesofsoftware. Consumers,otherthanlargeones,willneverpartic-
ipateonthemarketiftheyhavetocalculatetheirdemandbyhandandexpress
itasmathematicalfunctions. Whenincorporatingconsumersonthemarketwe
think ofthem asrepresentedbysoftwareagents,i.e. small pieces ofcomputer
software,thatgivetheirinputto thesystem.
2
1.3 Contributions
This thesisconsists of twopapersandan implementation ofthe algorithmin-
troducedin oneofthem;therstpaperintroducesanovelmarketmechanism,
ConFAst, a (semi) automatic mechanism with properties that meet the de-
mands that we focused on in the previous section. The second paper is an
articlewhichintroducesanalgorithm foraggregationof averygeneralclassof
(separable)objectivefunctions.
Thecharacterofthetwopapersdierssignicantly. Itisnotonlythetopic,
but the perspectivetoo. The rstone, dealingwithpowercontrolmarkets, is
on aratherhigh level, whilethesecond one,focusingon theconstructionand
behaviourofanalgorithm,israthertechnical.
Thethird partof the thesisis an implementation of the aggregationalgo-
rithm,animplementationthatwasproducedforcomparisonbetweenouralgo-
rithm and the standard algorithm for the problem. The work on the market
mechanismincludedimplementationtoo,but itis notin aformthat is suited
forpublication.
1.3.1 ConFAst, an Approach to Ecient and yet Fast
Power Control Markets
The ConFAst mechanism is designed with large computational markets in
mind. It is well suitednotonly formarketswhere thenumber ofparticipants
countsintensorhundreds,butitscalestohundredsofthousandsormore,both
producers,distributors,andconsumers. Themarketcomputationisdistributed
overcomputersinthenetwork,i.e. itisnotnecessarytogatherallinformation
needed at acentral spot. It handles non-continuous supplyand demand, and
henceitreducestheriskofparticipatingwithsuchcurvessignicantly.
Thepaperhasbeenpresented ataseminaronInformation andCommuni-
cation Technology in the Energy Sector, held by the Nordic Energy Research
Scientic Program, in Trondheim, Norway, March 8 9, 2001. A conference
versionofthepaperisundersubmission.
TheConFAst mechanismdepends onthe aggregationof utility functions
functionsthatexpresstherelativevalueofconsumptionsbundles,i.e. akind
ofobjectivefunctions. Computationallythisisthemostdemandingpartofthe
workanditisessentialthat theaggregationisperformedinanecientway.
1.3.2 Resource Allocation with Noisy Functions
Alotofeortshavebeenputontheaggregationofseparableconcaveobjective
functions 3
. Thegeneralcase,whennothingcanbesaidabouttheobjectives,is
hard. Apracticallyrelevantclassoffunctionsisseparablenon-concaveobjective
functions. Standard algorithms for aggregation of this class of functions are
basically constructed onapairwise aggregationof functions and abrute force
3
Concaveinthemaximisationversionoftheproblem,convexobjectivesintheminimisation
testing of allcombinations of thetwofunctions. Moreecientaggregation of
separableobjectivefunctions thatmightnotbeconcaveisnontrivial.
Thearticlepresentsanalgorithmthat focuses onthefactthat whenaggre-
gating non-concavefunctions the resulting function very fast becomes almost
concave,but noisy. The algorithm utilisesregularitiesin the aggregatedfunc-
tionswhenthefunctionisalmostconcaveorisclosetoanothersmoothcurve
sothat thesearchspacecanbeprunedwhenaggregatingthefunctions.
The algorithmis generic, but developed withresourceallocationand com-
putationalmarketsin mind. Thealgorithmcanbeusedasasubroutineofthe
ConFAstmarketmechanism.
Thearticleisrathertechnical. For thereadernotinterestedinalltechnical
detailsitisprobablysucienttoreadtheintroductionthroughtheexperimental
results, and skip the technical description of the algorithm. Hopefully this
partgivesenoughto grasp themain ideasbehind thealgorithm, and givesan
impressionoftheperformance.
Thearticleissubmittedforjournal publication.
1.3.3 Implementation
Implementationsandtestsof boththemarket mechanismandtheaggregation
algorithmispartofthethesiswork.
TheJavaclasses(andtestdata)neededtorunsometestsontheaggregation
algorithm are available on the Internet and on the CD-ROM included in the
thesis. Theimplementationisdonetotestandcomparetheperformanceofthe
algorithmondierentinput(thecomparisonisdoneonthestandardalgorithm
foraggregationof non-concaveobjectivefunctions). Theresults,that arepart
of the article, showthat the algorithm indeed is competitive. This holds for
bothadversarydataand practicallyrelevantdata. Itshouldbenotedthat the
standardalgorithmthat weuseforcomparisoncanbeexpected tobecloseto
optimalduetoitssimplicity. Muchcouldprobablybedonetoimprovethisrst
testimplementationofouralgorithm,asitisrathercomplicated.
Thethreepartsofthethesisare,insomeways,ratherdisparate. Atthesame
timetheyareheld togetherbytheaim todevelopalgorithmsand mechanisms
for resourceallocation and markets. As saidearlier, theapplication areathat
wehavein mindisthepowermarkets.
Acknowledgements
The thesis work has been partof twodierent projects. It startedin theDi-
Malloc project, sponsored by the Swedish National Board for Industrial and
Technical Development, NUTEK 4
. The work wasnished within the Swedish
National Programme on IT in Energy, aprogramme of the Swedish Research
InstituteforInformation Technology,SITI 5
.
ThemasterobjectiveoftheprogrammeITinEnergyistodemonstratehow
informationtechnologycanimprovetheeciencyin futureenergysystems.
EnerSearch AB 6
, a Swedish research company owned by interests in the
energysector,hasplayedacentralpartintheprojectsascoordinatingpartner
ofUppsalaUniversityandothers.
IwanttothankmyadvisorsFredrikYggeandArneAnderssonforsupport,
ideas,discussions,andmuchmore,andThoreHusfeldt,thathasbeenavaluable
local support in Lund. Both the articles of this thesis are joint work with
AnderssonandYgge. FurthermoreIwanttothankmyfriendsattheComputer
Science Department,Lund University,whereI livemyeveryday(working)life.
EventhoughIamnotatmyhomedepartmentinUppsalatooften,Iappreciate
beingpartoftheteam.
Finally, Iwantto thankmyfamily.
4
Duetorestructuring,the partofNUTEKthatsponsoredtheprojectisnowpartofthe
SwedishAgencyforInnovationSystems,Vinnova,http://www.vinnova.se.
5
http://www.siti.se
6
An Approach to Ecient and
yet Fast Power Control
Markets
Powermarketsaroundtheworldarederegulating. Theissueofhowtoproperly
designpowermarketsisadelicateone,intensivelydiscussedamongauthorities,
researchers,energyutilities,etc. Inthispaperwefocusontwomajoraspectsof
powermarket design: (i)managementofdiscontinuousdemand/supplycurves,
and(ii)computationaldesignofmarketsofhugesizewithrelativelyshorttime
frames.
Weintroduce amechanism, ConFAst, for highly dynamic powermarkets
thatallowsforhugenumbersofactiveparticipants(alsoconsumers)inthemar-
ket,eveniftheyhavediscontinuousdemand/supplycurves. Thecomputational
eciencyoftheapproachenablesusagealsoinmarketswithshorttimeframes.
Wearguethat ConFAst signicantlycanimprovetheeciency ofenergy
systems.
2.1 Introduction
2.1.1 The Strains on Market Mechanisms inExtreme Sit-
uations
Around January 24, 2000, the weather in all of Sweden was so cold that the
demand for electricity wasextremelyhigh andthe pressureon the wholepower
systemwasequallyhard. Asaconsequencethepriceonthespotmarketroseto
levels where energy utilities with no or small production capacity of their own
riskedtolose allof their annualprotduring onesingleday.
Thiswasthehardesttestofthenewmarket systemsofar,sincetherecent
deregulationoftheenergymarketsofFinland,Norway,andSweden. However,
in the times of deregulation the energy system is pushed closer to its limits
andwewillsee moreof thesetypesofsituations in thefuture. Skeptics ofthe
deregulationseeanewdawnin thisphenomenon.
Inanextremesituationlikethisone,newproblemsariseorcomeintofocus.
Onesuch problem is how torelease thepressureon thesystem in awaythat
aects the society and theeconomyas littleaspossiblewhen the demand for
energyistoohigh.
Today,inSweden,thereisanongoingdiscussiononhowtobettercopewith
this type of situations. One of the main issuesis how to properly utilise the
dynamics of the energy consumers. This questionis an aspect of the issue of
howtoconstructmoreecientpowermarkets. Theuctuationsindemandand
pricesonthemarketsoftodayaresignicantevenundernormalconditions,see
Figure2.1.
Figure2.1: Sincedemandvariesverymuchfromhourtohourthemarketprices
ofaderegulatedmarketvarysignicantlyaswell. Thesegraphsshowthemarket
priceoftheNordicspotmarket(left,peakprice55Euro/MWh)andtheSwedish
regulatingmarket(right,peakpricecloseto190Euro/MWh)aweekinJanuary
2001.
2.1.2 The Dynamics of the Small Actors
Today theloadside of themarket, theconsumerdemand, is essentiallytaken
as exogenous and the suppliers and distributors have to relyon moreor less
accuratepredictions. Thisarticleexploitspossibilitiesfortheloadside,aswell
assmallscaleproduction,to takeanactivepartin themarket.
A large portion of the loads within the power grid are controllable (in a
controlsystemsense). Thetermcontrollableloadcanbeillustratedbyexamples
fromeverydaylife: A lightbulboraTVsetisnotacontrollableloadwhen
the userswitches iton it is supposed to turn on instantly, and theother way
aroundwhenitisswitchedo. Ontheotherhandtheheatersofanapartment,
or any equipment that supply a comfortable indoor temperature, are typical