BIDDING IN INTERRELATED DAY-AHEAD ELECTRICITY MARKETS: INSIGHTS FROM AN AGENT-BASED SIMULATION MODEL

Transcription

1 BIDDING IN INTERRELATED DAY-AHEAD ELECTRICITY MARKETS: INSIGHTS FROM AN AGENT-BASED SIMULATION MODEL Anke Wedlch and Danel Vet Insttute for Informaton Systems and Management Unversty of Karlsruhe (TH) Englerstr Karlsruhe Germany { a nk e. w e d l c h d a n e l. v e t w. u n - karlsruhe.de Abstract In ths paper we present results from an agent -based smulaton model of two seuentally cleared electrcty markets. Agents can bd on both a day-ahead market for physcal delvery contracts and a day-ahead balancng power market and learn from ther acheved results. Dfferent scenaros of the order of market clearng and prcng rules are tested and ther results a r e compared. We show that prces are lower n both markets when the day-a h e a d m a r k e t s c l e a r e d f r s t. W e a l s o s h o w t h a t p a y- a s -bd leads to lower resultng prces than a unform prce mechansm. Key words: agent - b a s e d s m u l a t o n a g e n t- based computatonal economcs (ACE) electrcty market modellng seuental market clearng 1. Introducton Durng the last decades electrcty markets have been restructured n many regons worldwde. One effort n ths ongong change n the regulatory framework s to separate transmsson system operatons from other parts of the electrcty system such a s generaton and retal servces. When transmsson and generaton are separated the transmsson system operator (TSO) needs to procure balancng power from the market. Ths leads to the emergence of markets for prmary and secondary reserve as well as for mnute reserve. In ths envronment a generator that dsposes of fast controllable generaton unts faces the problem of decdng whether to bd n a day-ahead market or to commt hs unts for balancng purposes. Ths ncreasng complexty of electr cty tradng and bddng decsons rases the necessty for methods and tools that allow modellng a varety of aspects. One mportant ssue n electrcty market modellng s strategc behavour. In compettve eulbrum models of electrcty markets t s assumed that no agent tres to game the market and all market partcpants are prce takers. In eulbrum all agents bd ther margnal generatng costs and the (unform) market clearng prce s eual to the margnal unt bd. In real electrcty markets however the market structure s olgopolstc and each agent can potentally nfluence market prces. As one example C ramton [2004] shows how agents ratonally bd n olgopol st c electrcty markets and argues that bddng above margnal cost s a ratonal behavour and must not be consdered as ant-compettve. Realstc electrcty market modellng must therefore account for proft-seekng agent behavour that nvolves bddng above margnal costs. A second mportant aspect that a realstc electrcty market model should account for s the factor o f d a l y repetton n tradng. Rothkopf [1999] argues that the repettve nature of electrcty tradng s a c r u c al factor n market desgn. He argues that the choce of an effcent market mecha nsm may dffer when movng from one -shot auctons to repeated auctons. A th r d a s p e c t that should be addressed by electrcty models s market nterrelatonshps. Electrcty s traded on dfferent tme scales e.g. day-ahead real -tme and also n form of dfferent products e.g. physcal delvery reserved () capacty. The outcome of one market may nfluence an

2 agent s bddng strategy on another market. R ealstc electrcty market mode llng should thus account for nterrelatons between dffe rent markets. Agent-based smulaton has the potental to meet the aforementoned reurements and s a very promsng approach for realstc electrcty market modellng. It allows representng bdders as proft-maxmsng adaptve agents that can learn from ther tradng results n daly repeated auctons. It s also possble to smulate learnng agents tradng on dfferent markets. The markets are then lnked together through the agents tradng decsons whch allows for studyng market nterrelatonshp aspects. Several studes have demonstrated the usefulnes s o f agent-based smulaton for the analyss o f energy systems e.g. [ Koesrndartoto et al. 2005] [Bunn Olvera 2003] or [Bagnall 2004]. Agent - based models consderng market nterrelatonsh ps nclude [Rupérez Mcola Estañol B u n n 2006] who studes vertcal ntegraton n a mult-ter energy market ncludng gas shppers electrcty generators and retalers; [ Vet et al. 2006] study the nterplay between a forward market and a spot market for electrcty contracts and ts effect on prce stablty. In the followng we report results from an agent -based smulaton that comprses a d a y-ahead market for physcal delvery contracts and a day-ahead balancng power market. Agents are endowed wt h learnng capabltes and develop strateges that maxmse ther profts. The remander o f ths paper s structured as follows: secton two ntroduces the smulaton model and the agents tradng problem. Secton three summarses the characterstcs of the smulaton scenaros and secton four reports and dscusses the results from these smulatons. Secton fve fnally concludes and gves an outlook on future research. 2. The model In our model we smulate seuenced tradng n both a day-ahead market for physcal delvery wth unform prce clearng or pay system margnal prce and a balancng power market where agents bd ther capacty to a transmsson system operator as mnute reserve. On the latter market both pay-a s -bd and unform prcng schemes have been mplemented for dfferent scenaros. The smulaton model has been mplemented wth Java usng Repast 1 a toolkt for agent-based modellng The day-ahead market A bd on the day-a h e a d market conssts of dfferent prce volume pars. A supply bd by generator takes the followng form: { } b = p p p h h 1 h 1 h s h s h S h S where p s the bd prce and s the uantty that the agent s wllng to sell/buy at ths bd prce. The ndex h denotes the hour for whch the bd s vald and s s an ndex for the dfferent prce volume pars 2. In our smulatons we assume that an agent submts the same bd for every hour so the ndex h can be omtted n the followng presentaton. In further extensons of the model however t s envsaged to let agents optmse ther bds ndvdually for each hour. The day-ahead market operator collects all supply and demand bds and orders them nto 24 hourly bd collectons. M a r k e t c l e a r n g s d o n e once per tradng day and separately for each hour; physcal delvery of the sold electrcty s effected on the followng day. The day-a h e a d market s desgned a s a sealed-bd double aucton. However n the current model actve demand bddng doesn t occur. Instead genera tors submt supply bds for sats fyng a fxed nelastc demand. 1 Infor maton about ths tool can be found on 2 The bd format s desgned n the style of the spot market concept used at the [European Energy Exchange 2005].

3 In order to determne the market outcome the market operator sums up all supply and demand bds. Supply bds are ordered from the lowe st to the hghest bd prce and from hghest to lowest bd volume n the case of eual bd prces. The market clearng prce and volume are determned by the ntersecton of the supply and demand curves. If necessary the margnal bd s only partally fulflled. If many market eulbra exst.e. when the ntersecton between the supply and demand curve s s n o t a s ngle pont but a vertcal lne segment the mdpont between the lower and the upper lmt of the ntersectng lne segment s taken as the market clearng prce. Fgure 1 shows an example of the summed supply and demand bds n a day-ahead electrcty aucton. Fgure 1: Summed supply and demand bds n a day-ahead market At the end of each tradng round every agent receves nformaton about the resultng market clearng prce and the ndvdual volume that t has been able to sell or buy n the aucton. Agents use ths nformaton n order to calculate ther profts and renforce ther chosen strateges The balancng power market In the balancng power market an agent b ds a capacty pr c e (cap) at whch t s wllng to provde ts capacty as postve mnute reserve to a transmsson system operator. If the TSO actually needs electrc work for stablsng the transmsson system he bases the decson on whch capacty t o deploy on the bass of the bd work prce among all accepted bds. A bd on the balancng power market thus conss ts of a capacty prce and a work prce together wth the bd uantty for each bddng bloc b: { } b = p p cap work b b b b The market clearng of the procurement aucton for mnute reserve s effected n two stages. In the frst stage whch takes place one day ahead of physcal delvery the TSO selects the power plants that are held n reserve. For smplcty we assume that there s only one bddng bloc coverng all 24 hours of one day. Hence the ndex b can be o m t t e d n the followng. Among several dfferent optons for choosng the best bds the algorthm appled n our smulatons s solely based on the capacty bd prces. Ths reflects the fact that n real -world balancng power markets for mnute reserve the actual mnute reserve deployment s only n few cases greater than zero 3 and thus work prces play a mnor role n the optmal bd allocaton. Bds are ordered from the lowest to the hghest capacty bd prce and from hghest to lowest bd volume n the case of eual bd prces. Bds are accepted up to the demand volume as put to tender by the TSO. 3 For data on mnute reserve deployment and resultng capacty and work prces n the German balancng power m a r k e t s the nter ested reader s referred to the webstes of the four German TSOs: -n etz.com and

4 In the second stage all accepted bds are sorted accordng to ther bd work prces (mert order). If on the followng day the TSO s n need of electrc work from mnute reserve he determnes the capacty to be started-up on the bass of the work prce mert order. In the case of unform prce clearng all successful generators are pad the unf orm capacty prce whch corresponds to the capacty bd prce of the last unt needed to satsfy the demand. Those generaton unts that act ually delvered electrc work wll be remunerated the unform work prce whch s agan determned by the last successful bd. In the pay-a s -bd case all successful generators receve ther ndvdual bd prce s for ther commtted capacty and f applcable ther bd work prce s for delvered electrc work Agent learnng Agents choose ther bds accordng to a probablty dstrbuton developed wthn a renforcement learnng algorthm. The learnng representaton a ppled n our smulaton takes the form of a Modfed Erev-Roth renforcement learnng algorthm as presented n [ N c olasen Petrov Tesfatson 2001]. It s a three parameter algorthm developed on the bass of the f ndngs n [ Erev and Roth 1998] and s descrbed n the followng. For each stuaton n whch learnng s appled an agent can choose from a set of possble actons M. If generator chooses hs k th acton at tme t and receves a renforcement of R ( x ) t updates ts ndvdual propenstes to choose acton j = (1 M ) accordng to the followng functon: ì Rx () (1- e) j = k sj( t + 1) =(1 - f) sj() t +í ï e ï sj() t otherwse ïî ( M - 1) Here s j s the propensty to choose a specfc acton n the next round f s the recency parameter and e the expermentaton parameter. The recency parameter reflects the agents tendency to forget past experence over tme and the expermentaton parameter d e f n e s the extent to whch agents engage n nformaton exploraton through tryng strateges that don t have the hghest propensty. It has the effect that agents do not lock nto one choce at a too early stage. The thrd parameter specfyng the presented renforcement learnng algorthm s the scalng parameter s( 0 ). I t defnes the ntal propenstes for all actons. Agents choose a n acton k accordng to the followng probablty: p k ( t + 1) = s k M å j j =1 ( t + 1) s ( t + 1) They thus renforce actons whch have resulted n a hgh payoff a n d c h o os e t h e s e a g a n w t h a hgher probablty n the future. Accordngly less successful actons a re weakened and chosen agan wth a lower probablty. As probabltes should not be negatve t must be ensured that the propenstes always have a postve value. If a gents bd a prce below ther margnal generaton costs they face the rsk to be called nto operaton at a prce at whch they make losses. For ensurng postve propenstes we follow the proposton of Erev and Roth [1998] who subtract the smallest possble payoff x mn f r o m all payoffs: R ( x ) = x x m n In order for the algorthm to represent agent learnng n a realstc manner the parameters specfyng the algorthm have to be set carefully. Erev and Roth [1998] state that a combnaton of f = 0.1 and e = 0.2 leads to the best predcton of the emprcal outcome for ther studed matrx games. The order of magntude of the scalng parameter s (0) depends on the magntude of the

5 possble renforcements. In an effort standardse ths parameter we restrct the renforcement to be n the nterval [01] by dvdng each renforcement by the hghest possble renforcement. In other smulatons that we have carred out we fnd that the value for the scalng factor doesn t nfluence t h e r e s u l t s sgnfcantly 4. We also found that the parameter combnaton stated by Erev and Roth as the best for predctng the results of ther studed human experments actually leads to a low varablty of smulaton outcomes over a seres of 100 smulatons w h c h means that the choce of a random number seed nfluences the results less than most other tested combnatons 5. These fndngs led us to the followng choce of a parameter combnaton for every Modfed Erev-Roth renforcement learnng algorthm: f = 0.1; e = 0.2; s(0) = The agents acton domans On the day-ahead market we model a n agent as tryng to optmse both t s bd prces and uanttes n order to maxmse ts ndvdual proft. Thus an agent can engage n wthholdng strateges f t fnds these proftable. The possble bd prces for an agent on the day-ahead market range from the mnmum to the maxmum admssble prce where p = 0 a n d mn p = A generator max c h o o s e s t h e b d u a n t t y a s a fracton of hs avalable capacty. The ntervals [ a b] of possble prces and capacty fractons are stratfed nto 21 dscrete values for the bddng prce and sx values for the bddng uantty. The acton doman for an agent bddng on the day-ahead m arket thus comprses M = 126 possble actons and summarses to the followng form: { } { } { } { } { } { } M = p = On the balancng power market a n agent always chooses ts bd uantty as eual to the net nstalled capacty reduced f applcable by the amount o f capac ty t has already commtted on the day-ahead market. Thus an agent only learns to choose ts bd prces and employs a fxed strategy for the bd volume. As a bd on the balancng power market comprses a capacty prce and a work prce ths agan leads to a two-dmensonal acton doman. Admssble prces range from cap cap work work p = 0 to mn p = 500 and max p = 0 to mn p = 100 stratfed nto 21 possble max c a p a cty pr ces and fve work prces whch results nt o the followng acton doman: { } { } { } { } { } { } cap w o r k M = p p = The agents strategy The problem facng the agent can be dvded nto the strategy to choo se for the day-ahead market and the strategy for th e balancng power market. The learnng task for each agent s conseuently separated nto two ndvdual learnng problems. On the level of mplementaton ths results n each agent employng two nstances of the learnng algorthm wth the same paramete r values but dfferent acton domans for both nstances. In our smulaton scenaro each agent owns one generaton unt u whch s charactersed by a lnear varable generaton cost functon.e. constant margnal generaton costs MC u. Each power nst plant has a net nstalled capacty Q a n d n o -load costs NLC u. Rampng costs or other commtment constrants are abandoned n ths smulaton for smplfcaton reasons. comm As some capacty o f a generaton unt may have been commtted n one electrcty market for a t aval perod of tme t the avalable capacty for ths unt s defned as t 4 The tested scalng parameters range f r o m s(0) = 0.5 t o s(0) = The tested parameter values range from f = 0.0 to f = 0.5 and e = 0.1 to e = 0.5

6 = Q aval nst comm t t for ths specfc perod. Other reasons for reduced power plant avalablty such as mantenance perods or unplanned outages are not ncluded n the current smulaton mplementaton. We assume that an agent bds one prce volume par for every generaton unt t owns; agents cannot bd more than ther avalable capacty nto the market: aval h h aval u b u b An agent tres to maxmse profts on both market s. It does so by favourng actons that have yelded hgher profts n the past tradng round through renforcement learnng. The renforcement of each chosen acton comprses the proft acheved on the market a n d also ncludes opportunty costs to a certan extent. On the day-ahead market the agent s proft s defned as follows: ( ( )) π = p MC u Here p u s the proft that agent acheves for ts generatng unt u p s the resultng unform p r c e o n t h e d a y-ahead market and u s the uantty of unt u that agent was able to sell on the day-ahead market. The opportunty cost oc dayah ead for agent o n t h e da y-a h e a d m a r k e t s t h e proft t could have acheved f t had bd ts capacty on the balancng power market. It s defned a s : ( ) oc = p N L C u The proft an agent acheves on the balancng power market s defned as f ollows : ( ) ( ( )) π = p N L C + p MC cap cap w o r k bal a n c e w o r k w o r k ( ) ( ) u Here the subscrpt (u) ndcates that the capacty and work prce can ether be a sngle unform prce or an ndvdual bd prce for each agent/unt dependng on the prcng rule employed. As the produced electrc wo rk s remunerated separately the cost that an agent faces for provdng mnute reserve capacty s only lmted to the no-load costs of ts unt. However an agent faces hgh opportunty costs because t could also use ts unt for producng electrcty that t can sell on the day-ahead market. Ths opportunty cost s calculated as: ( ( )) oc = p MC cap cap u In both markets the opportunty costs are subtracted from the proft that an agent receves. In order to prevent unntended effects we restrc t the opportunty costs to be postve through applyng a mn() functon. In addton we ensure that the ncluson of opportunty costs doesn t lead to negatve renforcements. 3. Smulaton scenaros Because of the probablstc nature of the appled renforcement learnng algorthm the outcome from our agent-based smulatons partly depends on the random number seed wth whch learnng n s t a n c es are ntalsed. We derve ndvdual random number seeds for each nstance of the learnng module so as to avo d unntended smlarty among agents whch would occur f all learners had the same seed. Every smulaton settng s run 100 tmes wth dfferent random number seeds at each run. All seeds employed n one set of 100 runs are stored and used agan for dfferent smulaton scenaros. By dong so we can best explot the advantages of agent -based smulaton: dfferent settngs e.g. dfferent market mechansms can be tested under exactly the same condtons. By ensurng that all agents n prncpal learn exactly n the same way n one settng as n another we can derve ualtatve conclusons about market effcency through c o m p a r n g the resultng outcome of both settngs.

7 In ths paper we explore four dfferent scenaros whch dffer n the order of market executon and the prcng mechansm on the balancng power market: DayAheadBalance_unform DayAheadBalance_payAsBd BalanceDayAhead_unform BalanceDayAhead_payAsBd DayAheadBalance refers to a scenaro where the day-ahead market s cleared frst and t hen results from ths market are publshed before the balancng power market clears. BalanceDayAhead swtches the order of market executon. The endngs _unform and _payasbd correspond to a unform prce and a pay-a s -bd prcng rule on the balancng power m a r k e t. O n t h e d a y-ahead marketnform prcng s employed n all cases. The partcpatng agents and ther generaton unt characterstcs are summarsed n Table 1. Agent name Capacty of Marg n a l No -l o a d generatng cost c o s t u n t [MW] [EUR/MWh] [EUR/h] Generator Generator Generator Generator Generator Generator Generator Generator Generator Generator Table 1: Characterstcs of the agent s generatng unts The statc demand throughout the smulaton s D = 1500 MW and D cap = 800 MW both constantly for every hour. The demand for electrc work from mnute reserve s D work = 160 MW whch occurs n one hour per day. 4. Smulaton outcome Each run wthn one smulaton scenaro has been smulated over at most 2000 tradng days. In many runs however a l l agents locked nto some preferred acton earler because the propenstes of all other actons were so low that they were hardly chosen. A smulaton run was stopped f the resultng market prce dd not change over a perod of 200 tra dng days. In all cases the mean bd/market prces and bd/resultng volumes over the last 200 tradng days was recorded as a result for each run. In the followng the smulaton results for each scenaro are reported. They comprse the average mnmum and maxmum resultng values over 100 runs wth dfferent random number seeds as well as the standard devatons (SD). Resultng prces on the day-ahead market are depcted n T a b l e 2. Prces attan a hgher level f the day-a h e ad market s cleared after the balancng power market. The ntuton behnd ths result s that competton s lower on the supply sde n these cases as some generators have already commtted a part of ther capacty on the balancng power market. When fe wer agents compete on the day-ahead market they can more successfully bd above margnal cost and thus acheve hgher prces.

8 S c e n a r o Scenaro Average Mnmum Maxmum SD 6 p p d a y A h e a d p d a y A h e a d p d a y A h e a d DayAheadBalance_unform DayAheadBalance_payAsBd BalanceDayAhead_unform BalanceDayAhead_payAsBd Table 2: Smulated market clearng prces on the day-ahead market Average Mnmum p b alancecap / p b a l a n c e c a p / p b a l a n c e w o r k p b a l a n c e w o r k Maxmum SD p b a l a n c e c a p / p b a l a n c e w o r k p b a l a n c e c a p / p b a l a n c e w o r k DayAheadBalance_unform DayAheadBalance_payAsBd BalanceDayAhead_unform BalanceDayAhead_payAsBd Table 3: Smulated capacty/work prces on the balancng power market T a b l e 3 represents the resultng capacty and work prces on the balancng power market. Here we observe the opposte structure as on the day-ahead market.e. prces tend to be lower when the market s cleare d second. Ths outcome cannot be explaned by the supply concentraton; another aspect nfluences the outcome n a stronger manner n ths case. Whereas prce nformaton on the s e c o n d m a r k e t p l a ys only a mnor role for bddng strateges on the day-a h e a d market on the balancng power market agents manly base ther bd decson on ther opportunty costs. Hgh p r c e s o n t h e d a y-ahead market mean hgh foregone profts for an agent that commts hs capacty for mnute reserve purposes. So when prces are hgh on the day-ahead market agents tend to bd hgher on the balancng power market as well. Agent Average d a y A h e a d p / Mnmum d a y A h e a d p / Maxmum d a y A h e a d p / SD d a y A h e a d p / Generator Generator Generator Generator Generator Generator Generator Generator Generator Generator Table 4: Indvdual agents bd prces and volumes on the day-ahead market (smulaton scenaro: DayAheadBalance_unform) The sngle agents bddng decsons on the day-ahead market are summarsed n T a b l e 4. I t c a n b e shown that agents wth low margnal costs rea lse ther strategc advantage through bddng at lower prces on average. Ths ensures that ther bds are accepted wth a hgher probablty so they are able to sell more electrcty. It can also be shown that agents wth hgh margnal costs tend to bd less capacty nto the market. Ths can be nterpreted as a wthholdng strategy whch s appled n order to rase the market prce. Accordng to the observed bddng strateges t can be stated that no-load costs don t play a sgnfcant role n the bddng decson on the day-ahead market; agents 6 SD = standard devaton

9 whose generaton unt characterstcs only dffer n no -load costs tend to apply very smlar bddng strateges. Table 5 shows the ndvdual bddng decsons on the balancng power market. Only capacty bd prces are dsplayed. It can be seen that on average most agents bd hgher prces n a market wth pay-a s -bd prcng as compared to a unform-prce market. However ths rse n bd prces does not lead to hgher market clearng prces n a pay-a s-b d s e t t n g (see results for the balancng power m a r k e t n Table 3). The average gan agents attan from recevng ther hghe r bd prces under pay-a s -bd does not outwegh the gan that nfra -margnal bdders have from recevng t h e u n f o r m prce whch corresponds to the hghest accepted bd. It s also nterestng to note that generators wth lower margnal cost tend to bd hgher prces n the balancng power market; ths too s due to the strong nfluence of opportunty costs on the bddng decson. cap cap Agent p BalancngDayAhead_unform p BalancngDayAhead_payAsBd Average Mn M a x SD Average Mn M a x SD Generator Generator Generator Generator Generator Generator Generator Generator Generator Generator Table 5: Indvdual agents bd prces and volumes on the balancng power market (smulaton scenaro: BalanceDayAhead_unform and BalanceDayAhead_payAsBd) The ueston whether pay -a s- bd or unform prce leads to lower market prces s controversally dscussed n the lterature (e.g. [Kahn et al. 2001] [Rassent Smth Wlson 2003] [Bower Bunn ] [Xong Okuma Fujta 2004]). Some researchers argue that pay-a s-bd leads agents to b d at hgher prces resultng n hgher average market prces. Others n contrast argue that t leads to lower overall prces because nfra -margnal agents reach lower prces as they would under unform - prcng. Our smulaton results suggest that pay-a s -bd does n fact result n hgher bd prces but the bd prce ncrease s not hgh enough to result n hgher overall prces. 5. Conclusons and outlook In ths paper we modelled dfferent scena ros wth seuentally c l e a r e d electrcty market s a n d two dfferent prcng mechansms. We apply agent-based smulaton for evaluatng market outcomes for the dfferent tested scenaro and for analysng market nterrelatonshps. Generators are modelled as adaptve agents that apply a renforcement learnng algorthm whch allows them to teratvely approxmate ther proft maxmzng strategy. W t h n the tested scenaros we dfferentate between bddng strateges n a day-ahead market wth physcal settlement and strateges n a d a y-ahead balancng power market. W e shft the order of market executon and vary the prcng mechansm from pay-a s -bd to unform prce. Smulatons for each settng are repeated over 100 runs n order to level out the nfluence of the random number seeds taken for each learnng nstance. We fnd that prces on the day-ahead market are hgher f ths market s cleared after the balancng power market. We argue that ths s due to the fact that competton s weaker n ths case a s s o m e agents have already commtted (part of) ther capacty on the balancng power market. The reduced suppler concentraton enables agents to successfully bd hgher mark-ups to ther margnal costs. Results on the balancng power market gve a d f ferent pcture : prces are lower f ths market s

10 cleared second. Here the effect of agents ntegratng ther opportunty costs nto ther result evaluaton leads to hgher prces when day-ahead prces are hgh and lower prces when day-ahead prces are low. As for the prcng rule on the balancng power market we fnd the followng result: average prces are hgher under unform prce than under pay-a s -bd although agents bd at hgher prces under pay-a s -bd. The ncrease n bd prces s outweghed by the effect of all nfra -margnal bdders recevng the margnal.e. hghest accepted bd under unform prcng. The fndngs presented n ths paper leave us confdent that agent-based smulaton can reproduce realstc market outcomes. It s therefore a sutable tool for market desgn uestons because dfferent market mechansms can be tested under exactly the same condton of agent learnng. In order to make agent -based smulaton a useful market desgn tool we wll nclude more realstc data cha ractersng the electrcty system nto our model. Our smulaton model s mplemented n a flexble way so that s also easly extensble to other forms of generaton or demand representaton. For example margnal costs can also take a lnear form nstead of beng constant or agents can own several generaton unts and engage n combned tradng strateges for all ther capacty. Hourly contracts and smaller bd blocs for the mnute reserve can also be modelled wth the current mplementaton. Usng the se features for analysng more settngs s subject to our future research. References B a g n a l l A. (2004) A Mult - Agent Model of the UK Market n Electrcty Generaton I n B u l l L. ( E d. ) : A p p l c a t o n s of Learnng Classfer Systems Seres: Studes n F uzzness and Soft Computng Vol. 150 pp Bower J.; D. Bunn (2001) Expermental analyss of the effcency of unform- Prce versus dscrmnatory auctons n the England and Wales electrcty market J o u r n a l o f E c o n o m c D y n a m c s a n d C o n t rol Vol. 25 pp March Bunn D.; F. Olvera (2003) Evaluatng ndvdual market power n electrcty markets va agent-b a s e d s m u l a t o n Annals of Operatons Research 2003 Vol. 121 p Cramton P. ( ) Compettve Bddng Beh a v o u r n U n f o r m-p rce Aucton Markets P r o c e e d n g s o f t h e 3 7 th Hawa Internatonal Conference on System Scences European Energy Exchange (2005) EEX Spot Market Concept Erev I. ; A. Roth (1998) Predctng How P eople Play Ga m e s : R e n f o r c e m e n t L e a r n n g n E x p e r m e n t a l G a m e s w t h Unue Mxed Strategy Eulbra Amercan Economc Revew 88(4):84881 September K a h n A. ; P. C r a m t o n R. P o r t e r R. T a b o r s ( ) U n f o r m P r c n g o r P a y-a s-b d Prcng: A Dlemma for Calfor n a and Beyond Electrcty Journal July Koesrndartoto D.; J. Sun L. Tesfatson (2005) A n a g e n t -based computatonal laboratory for testng the economc relablty of wholesale power market desgns I E E E Power Engneerng Socety General Meetng p p Ncolasen J. ; V. Petrov L. Tesfatson (2001) Market Power and Effcency n a Computatonal Electrcty Market W t h D s c r mnatory Double - Aucton Prcng ISU Economc Report No. 52 revsed August Rassent S.; V. Smth B. Wlson (2003) Dscrmnatory Prce Auctons n Electrcty Markets: Low Volatlty at the Expense of hgh Prce Levels Journal of Regulatory Economcs vol. 23 no. 2 pp March Rothkopf M. ( ) Daly Repetton: A Neglected Factor n the A nalyss of Electrcty Auctons The Electrcty Journal 11 pp A p r l Rupérez - Mcola A. A. Es tañol D. Bunn (2006) I n c e n t v e s a n d C o o r d n a t o n n V e r t c a l l y Related Energy Markets CIC Workng Papers No. SP II V e t D. ; A. W e d l c h J. Ya o S. O r e n ( ) S m u l a t n g t h e D y n a m c s n T w o- Settlement Electrcty Markets va an Agent- B a s e d A p p r o a c h Workng Paper ( ). Xong G.; S. Okuma H. Fujta (2004) Mult -a gent Based Experments on Unform Prce and Pay-a s- Bd Electrcty Aucton Markets IEEE Internatonal Conference on Electrc Utlty Deregulaton Restructurng and Power Technologes (DRF T2004) Hong Kong Aprl 2004.