Using ICT-Controlled Plug-in Electric Vehicles to Supply Grid Regulation in California at Different Renewable Integration Levels

Transcription

1 1 Using ICT-Conrolled Plug-in Elecric Vehicles o Supply Grid Regulaion in California a Differen Renewable Inegraion Levels Chrisoph Goebel, Member, IEEE, and Duncan S. Callaway, Member, IEEE Absrac The purpose of his paper is o quanify he poenial for plug-in elecric vehicles (PEVs) o mee operaing reserve requiremens associaed wih increased deploymen of wind and solar generaion. The paper advances prior PEV esimaes in hree key ways. Firs, we employ easily implemenable scheduling sraegies wih very low cenralized compuing requiremens. Second, we esimae PEV availabiliy based on daa sampled from he Naional Household Travel Survey (NHTS). Third, we predic regulaion demand on a per minue basis using published models from he California ISO for 20% and 33% renewable elecriciy supply. Our key resuls are as follows: Firs, he amoun of regulaion up and regulaion down energy delivered by PEVs can be balanced by using a hybrid of wo scheduling sraegies. Second, he percenage of regulaion energy ha can be delivered wih PEVs is always significanly higher han he percenage of convenional regulaion power capaciy ha is deferred by PEVs. Third, regulaion up is harder o saisfy wih PEVs han regulaion down. Fourh, he scheduling sraegies we employ here have a limied impac on load following requiremens. Our resuls indicae ha 3 million PEVs could saisfy a significan porion - bu no all - of he regulaion energy and capaciy requiremens ha are anicipaed in California in Index Terms Plug-in Elecric Vehicles, renewable inegraion, regulaion, capaciy, load following, scheduling I. INTRODUCTION This paper examines he fuure poenial for plug-in elecric vehicles (PEVs) o balance variabiliy and uncerainy from wind and solar generaors. We focus specifically on approaches o leverage emerging Informaion and Communicaion Technologies (ICT, i.e. he smar grid ) o coordinae PEVs o provide ancillary services in he regulaion (secondary frequency conrol) ime frame. As a firs sep ino his direcion, we analyze in his paper how he charging demand of a large populaion of PEVs can be sculped such ha he demand for regulaion in California is reduced as much as possible. From an insiuional poin of view, he scenario we are invesigaing could be realized by concenraing he conrol over PEV charging a he sysem operaor. Since he sysem operaor already has full conrol over generaors providing load following and regulaion, we believe ha his is a realisic scenario. We model he relevan marke operaions of he California Independen Sysem Operaor (CAISO) and evaluae he performance of a basic greedy algorihm for load shifing. The work of D. S. Callaway was suppored by Rober Bosch LLC hrough is Bosch Energy Research Nework funding program, and he PSERC FuureGrid iniiaive. The approach in his paper requires he ICT-based communicaion beween he sysem operaor and he conrollable PEVs o be wo-way. Specifically, he sysem operaor (or an inermediae aggregaor) will collec informaion from he PEVs including: arrival and anicipaed deparure from he grid, baery sae of charge, maximum charge rae and baery capaciy. The sysem operaor hen sends he charging schedules o he PEVs. We will assume he ICT is capable of his informaion exchange a a minimum frequency of once per minue. The PEVs le he aggregaor know when hey connec o or disconnec from a charging saion. Upon connecion hey reveal heir baery s sae of charge. If he PEV owner does no specify he saring ime of he nex rip, he sysem operaor predics i. We do no address he deails of how his migh be done in his paper. In addiion o he sar ime of he nex rip and he connecion period, he sysem operaor would need o know he charging rae and baery capaciy of each PEV. Based on his informaion, i can devise a charging schedule and send i back o he PEV for execuion. Our evaluaion is based on one minue ime inervals, herefore he ICT infrasrucure should make i possible o send one schedule per minue o each conrollable PEV. We show how he regulaion poenial provided by our charging algorihm changes depending on a number of basic parameers such as PEV flee size and charging job scheduling mehod. The goal of his aricle is o evaluae he regulaion poenial of PEVs, no heir overall economic implemenabiliy. Neiher do we consider vehicle-o-grid (V2G), i.e. PEVs providing power o he grid, nor excepionally high charging raes. The charging algorihm we propose does no increase he number of charging cycles, he deph of discharge, or he power pu in or aken ou of he PEV baeries compared o unconrolled charging. Therefore is impac on baery life ime should no be significanly more negaive han he impac of unconrolled charging. II. RELATED WORK The auhors of [1] were among he firs o invesigae how he U.S. ligh vehicle flee could serve as grid resource. They argue ha, in he shor erm, PEVs should be apped for delivering high-value, ime criical services, in paricular regulaion and spinning reserve. The main difference beween [1] and our work is ha we are able o simulae he impac of conrolled PEVs charging based on acual grid and driving daa

2 2 over longer ime periods. This allows us o consider dynamics which saic, back-of-he-envelope ype of models as he one used in [1] canno. Recen effors o inegrae more inermien power from wind and solar have revived he ineres in demand response from a differen angle. The auhors of [2] explore he concepual requiremens and opporuniies of developing load conrol schemes ha are compeiive wih convenional generaionbased approaches. One of he major goals hey idenify is full responsiveness defined as enabling high-resoluion sysemlevel conrol across muliple ime scales. A second goal is non-disrupive conrol, meaning ha conrol should have an impercepible effec on end-use performance. The PEV charging conrol infrasrucure oulined in his aricle achieves boh goals. Similar o previous work in he area of demand response, e.g., [3] or [4], he goal of he coordinaion schemes invesigaed in [2] is o fill valleys of aggregaed elecriciy demand raher han providing regulaion as discussed in his paper. The auhors of [5] presen new mehods o model and conrol hermosaically conrolled elecric loads which can also be applied o deliver load following and regulaion services for he grid. The auhors of [6] use a basic economic model o invesigae he formaion of valley-filling Nash equilibria if charging conrol is decenralized and agens behave raionally minimizing heir operaion cos. The auhors of [7] describes a similar vision as pu forward in [2], namely fully responsive loads. They demonsrae he concep by providing deails on a prooype smar PEV charging sysem developed by Google. Their sysem also allows for dispaching PEV charging load o provide regulaion services for he grid. In conras o us, hey use a simplisic vehicle usage model. They focus on demonsraing he funcionaliy of heir prooype using non-represenaive daa of regulaion demand and PEV usage. The auhors of [8] compare he impac of conrolled and unconrolled PEV charging on emissions and generaion coss in he sae of Ohio. They use a uni commimen model o simulae he cos-minimal scheduling of fossil-fueled generaors assuming ha PEVs and generaion dispach are co-opimized. Our model has no ye been exended o capure emission and generaion cos. In conras o he auhors of [8], we focus on regulaion and load following and hus use a per minue insead of an hourly model. The auhors of [9] describe he oucome of a research projec which invesigaed he revenues ha could be generaed from providing regulaion services o he Californian grid. They used a prooypical PEV oufied wih all necessary upgrades for conrolling charging and discharging o analyze he impac of regulaion services provision on baery wear ou coss. In conras o us, hey used a hisorical grid regulaion signal as inpu o heir charging conrol algorihm and focus on comparing he benefi obained from selling regulaion capaciy o he cos caused by baery wear ou. The auhors of [10] presen he mos comprehensive modeling framework we could find in he lieraure. I consiss of groups of V2G-capable PEVs, hermal household loads, and combined hea and power (CHP) urbines. Their goal is also o evaluae he possible conribuion of deferrable loads o supplying regulaion, in heir case referred o as secondary conrol. Wihou doub, he auhors of [10] have done an excellen job inegraing differen models and considering as much deail as possible. However, i is a challenging ask o obain he daa required o paramerize hese models. For heir case sudy, hey use an invened 4 hub ransmission nework covering a residenial area wih 1,000 households. Insead of simulaing he demand for regulaion by modeling ISO operaions like us, hey use he swissgrid pre-qualificaion profile and an acual one day long regulaion signal in he range of up o 40 MW. In conras o all oher models presened so far, our model is capable of predicing he consequences of using PEVs for supplying regulaion over an enire year and in differen renewable inegraion scenarios. Considering longer ime horizons for his kind of analysis is imporan, since load (including he addiional one caused by PEV charging), as well as wind and solar energy supply, exhibi significan seasonaliies. The chosen level of absracion allows us o leverage acual daa made available by he CAISO via [11]. Thus, our resuls can be used for deriving concree policy recommendaions for he CAISO conrol area. Our mehod could also be applied for oher conrol areas, provided ha he relevan operaional characerisics are capured and corresponding daa is available. III. THE MODEL In he following Secion III-A, we presen a model used by he CAISO for deriving he demand for load following and regulaion [12], [13]. We modified his model o include he charging demand of PEVs as a separae variable. In Secion III-B, we explain our innovaive approach o derive PEV charging jobs from NHTS daa. Secion III-C oulines how we inegrae he CAISO model and he PEV model ino a simulaion ool ha can quanify he impac of coninuously arriving PEV charging jobs and load shifing on regulaion and load following. Fig. 1. Flow char of simulaion procedure. Figure 1 provides an overview of he complee simulaion procedure. I indicaes he cyclical relaionship of regulaion demand and generaion scheduling which our ool simulaes: A change of he PEV load a he curren ime sep influences fuure PEV demand and herefore he load forecas (cf., Equaions 3 and 9). The load forecas, in urn, influences he

3 3 regulaion demand and o he exen he demand is served by PEVs, i also influences fuure PEV charging schedules (cf., Equaion 2). The generaion schedules are updaed in regular ime inervals according o CAISO operaion procedures and hus do no reflec all changes of he PEV charging schedule a every given ime. A. Regulaion and Load Following Demand The CAISO is responsible for balancing supply of and demand for elecriciy wihin is conrol erriory. I also runs he corresponding wholesale markes on which paricipans can rade elecriciy. A he end of each rading inerval, he CAISO deermines a marke clearing price based on he submied bids and he consrains of he ransmission grid. The final aggregaed power generaion schedules for he hourly scheduling processes have 20 minue long ramps beween he oupu levels of subsequen operaing hours and are available 75 minues before he sar of he corresponding operaing hour [14], [15]. The generaion schedules for he real-ime scheduling process conain 5 minue ramps and are available 7.5 minues before he corresponding operaing inerval [16], [15]. Updaed forecass of demand and supply from renewable resources ha become available wihin he ime inerval beween dispach insrucions and energy delivery/consumpion canno be considered in he corresponding schedules. Demand and supply forecass can be inaccurae, especially he forecass of wind and solar oupu. This has a negaive impac on he economic efficiency of he menioned energy markes. In addiion o he markes for elecric energy, he CAISO also runs markes for capaciy producs or ancillary services (AS). We do no consider he marke for regulaion in furher deail in his work, alhough one could imagine ha PEV aggregaors will paricipae in he AS markes. Recen work in his direcion includes [9], [17] and [18]. Insead, we invesigae he heoreical siuaion in which he sysem operaor can direcly conrol he charging behavior of PEVs. Regulaion up and down is he mos demanding AS: I has o be available wihin one minue and is he mos expensive AS on a $/MW-hr basis [19]. In his work, we infer he demand for up and down regulaion using he model described in [15], [13]. We provide all relevan deails and he paramerizaion of our model in he following. Table I provides an overview of he differen variables used in he following. The modelling approach described in his secion akes advanage of he fac ha he acual power generaion mus be equal o he oal load minus he non-dispachable generaion, i.e., L W G SG, and ha he oal generaion commied hour-ahead can be obained from he differen hour-ahead forecass, i.e., LF HA, W GF HA, and SGF HA +L P EV, load following demand LF D, and regulaion demand RD. Equaion (1) summarizes his coherence. L + L P EV W G SG = LF HA W GF HA SGF HA + LF D + RD (1) Symbol Descripion (a minue ) LF HA Load according o hour-ahead forecas LF RT Load according o real-ime forecas LF D Load following demand L Acual non-pev load L P EV Acual PEV load RD Regulaion demand SG Acual solar oupu SGF HA Solar power oupu according o hour-ahead forecas SGF RT Solar power oupu according o real-ime forecas W G Acual wind oupu W GF HA Wind power oupu according o hour-ahead forecas W GF RT Wind power oupu according o real-ime forecas TABLE I VARIABLE DESCRIPTIONS. In he following, we make wo furher simplifying assumpions: The aggregaed hour-ahead and real-ime schedules are followed perfecly, i.e. generaors do no deviae from heir individual schedules. The ISO has sufficien load following resources a is disposal o mee he ramping requiremens. The demand for regulaion, RD, is hen equal o he difference beween he generaion according o he real-ime W GF RT SGF RT schedule, i.e., LF RT load minus he non-dispachable generaion, i.e., L L P EV, and he acual + W G SG. Equaion (2) provides a formal expression of he regulaion requiremen RD a ime based on he assumpions saed above. RD = L + L P EV W G SG LF RT + W GF RT + SGF RT (2) Hisorical values for he oal load as well as he oal oupu generaed by wind and solar power generaion resources are available for download from he CAISO websie [20]. The downloadable daa file conains projecions of expeced values in 2020 based on 2005 measuremens. The CAISO has provided echnical deails on he generaion of his daase in [13]. Regarding he non-pev elecriciy demand, we follow he CAISO s assumpion ha i grows exponenially according o he following equaion wih α = 1.5% [15], [13]: L 2005+i = (1 + α) i L 2005 The wind and solar races are scaled o reflec differen insalled capaciies. The capaciies we use in his aricle are shown in Table II. They are adoped from he CAISO s 20% and 33% renewable resources sudies (cf., [12], [21]). According o California s regulaory arges, 20% renewable elecriciy producion has o be reached in 2012 and 33% in As in [15], [13], we approximae he real ime marke s forecas of wind oupu W GF RT by he observed wind oupu 8 minues before, i.e. we assume ha W GF RT = W G 8. The compuaion of he demand resuling from PEV charging, L P EV, is described in Secion III-B. The compuaion

4 4 Renewable inegraion arge Resource Type Symbol 20% (2012) 33% (2020) Wind urbines W GC Y 6,688 9,194 Large solar PV SCY 1 1,076 3,527 Large solar hermal SCY 2 1,448 4,058 Disribued solar PV SCY 3 1,045 1,045 Cusomer solar PV SCY 4 1,749 1,749 TABLE II INSTALLED WIND AND SOLAR POWER GENERATION CAPACITY IN MW ACCORDING TO [22]. of he remaining values, i.e., LF RT and SGF RT will be described in he following. We obain he real-ime load forecas LF RT by adding forecas errors o 5 minue averages of he non-pev load L and connecing he resuling values wih 5 minue long ramps. The forecas errors are generaed using he following auo-regressive process: ɛ = γ s() ɛ 1 + x s() The random numbers xs() are drawn from a runcaed normal disribuion Nx wih zero mean and san- dard deviaion σs during season s. The disribuion N[ x is runcaed such ] ha ɛ lies in he inerval ( 3)σ s, (+3)σs. [15] is based on only one esimae of γ and σ for he enire year The mos recen source for he required parameer esimaes is [12]. The values in [12] are season-specific and hus more accurae, which is why we also use hese values. They are denoed by γs and σs. We repor he values in Table III o faciliae he reproducion of he resuls presened in his aricle. Parameer Symbol Winer Spring Summer Fall Auo- γ s correlaion γs HAload γs HAwind Sandard σ s deviaion σs HAload σs HAwind TABLE III AUTO-CORRELATION AND STANDARD DEVIATION OF LOAD AND WIND FORECAST ERRORS ACCORDING TO [22], [21]. Equaion (3) describes he mehod we use o derive he realime load forecas formally. The ramp generaion is denoed by he funcion AR 5min. LF RT = AR 5min ( 1 5 sop(i 5min()) i= sar(i 5min()) (Li + L P EV i ) + ɛ I5min()) (3) In Equaion (3), I 5min () reurns he 5 minue inerval ha minue belongs o, sar and sop reurn he sar and sop minue of a 5 minue inerval. The real-ime solar generaion forecas SGF RT is compued according o he mehod described in [13]. For each ype r of solar resource, we obain one forecas value for each 5 minue long ime inerval in he simulaed year using he following equaion: SGF[,+5] r = CIr ( 8) 1 +5 max 5 day (SGr, i) CI r () [0, 1] denoes he clearness facor regarding a solar resource ype r a ime. The higher his facor, he more sunligh peneraes he cloud cover on average a ime. I is obained from our daa se using he following equaion: i= CI r () = SG r / max day (SGr, ) The expression max day (SG r, ) denoes he maximum solar oupu a ime. I is obained by selecing he maximum value found in he corresponding year long solar oupu race a he corresponding minue of he day. To generae he final real-ime solar generaion forecas, we add 5 minue long ramps o he sum of solar resource forecass (cf., Equaion (4)). SGF RT = AR 5min ( r SGF r [,+5] ) (4) The demand for load following a ime, LF D, can be formally described by he following equaion: LF D = LF RT W GF RT LF HA SGF RT + W GF HA + SGF HA (5) We obain all hour-ahead dispachable generaion and wind schedules by adding forecas errors o he hourly averages of he measured quaniies. Similar o he real-ime load forecasing error, we compue he hour-ahead forecas errors based on auo-regressive sochasic processes. Equaion (6) provides he descripion of he sochasic process of he hourahead load forecas error. ɛ HAload = γ HAload s() ɛ 1 + x s() HAload (6) The random variable x HAload s() is drawn from a season-specific runcaed normal disribuion Nx HAload wih zero mean and sandard deviaion σs HAload. The runcaion [ makes sure ha] ɛ HAload lies in he inerval ( 3)σ HAload s ; (+3)σs HAload. The measured season-specific auo-correlaion and sandard deviaion we use o generae he hour-ahead load forecas is repored in Table III. Equaion (7) shows he corresponding formula for he hourahead wind forecasing error. ɛ HAwind = γ HAwind s() ɛ 1 + x s() HAwindW GC Y (7) The random numbers x HAwind s() are drawn from a run- wih zero mean and. The random variable Xs() HAwind caed normal disribuion N HAwind x sandard deviaion σ HAwind s describes he forecas error as he fracion of oal wind generaion capaciy in he year Y, W GC Y (cf., Table II).

5 5 The runcaion makes sure ha ɛ HAwind lies in he inerval [ ] ( 3)σs HAwind W GC Y ; (+3)σs HAwind W GC Y and ha 1 sop(i 1hour ()) i= sar(i 1hour ()) W G i + ɛ HAwind I 1hour () remains in he capaciy inerval [0, W GC Y ]. The measured season-specific auocorrelaion and sandard deviaion parameers for he hourahead load forecas is repored in Table III. We obain he hour-ahead solar forecas using he following procedure described in [13]. For each hour of he simulaed year, we compue a clearness fracion CI r (I 1hour ()) using he following equaion: CI r (I 1hour ()) = 1 1 sop(i 1hour ()) i= sar(i 1hour ()) SGr i sop(i 1hour ()) i= sar(i 1hour ()) max day(sg r, i) Based on he hourly clearness fracions, we can derive he sandard deviaion of he solar forecas error using he correlaion of sky clearness and he variabiliy of solar oupu. Table IV provides he mapping of clearness fracions o sandard deviaions from [13]. Clearness fracion Sandard deviaion CI r (I 1hour ()) (σ HAsolar,r, ) 0.0 CI r (I 1hour ()) CI r (I 1hour ()) CI r (I 1hour ()) CI r (I 1hour ()) TABLE IV STANDARD DEVIATIONS OF SOLAR FORECAST ERRORS BASED ON CLEARNESS FRACTION LEVELS ACCORDING TO [13] To generae he hourly solar forecas error, we use a corresponding runcaed normal disribuion Nx HAsolar,r, (cf., Equaion (8)). ɛ HAsolar,r, = x HAsolar,r, SC r Y (8) The runcaion makes sure ha he forecas error does no exceed 3 imes he sandard deviaion, i.e. each x HAsolar,r, mus lie in he inerval [ ( 3)σ HAsolar,r,, (+3)σ HAsolar,r,] and mus no exceed insalled capaciy, i.e. x HAsolar,r, [0, SCY r ] is fulfilled for all solar resource ypes r a all imes. We compue he final hour-ahead load and wind/solar generaion forecass by adding 20 minue long ramps o he sum of he corresponding hourly averages and forecas errors. The ramp funcion is denoed wih AR 1hour. Equaion (9) describes he hour-ahead load forecas. Equaions (10) and (11) provide he power producion from wind and solar a ime according o he hour-ahead generaion forecas. LF HA = AR 1hour ( 1 sop(i 1hour ()) i= sar(i 1hour ()) (Li + L P i EV ) + ɛ HAload I 1hour () ) (9) W GF HA = AR 1hour ( 1 SGF HA = AR 1hour ( 1 B. PEV Charging Demand sop(i 1hour ()) i= sar(i 1hour ()) sop(i 1hour ()) i= sar(i 1hour ()) W G i + ɛ HAwind I 1hour () ) (10) (SG r i + ɛ HAsolar,r, I 1hour () )) (11) In his secion we describe our approach o obain represenaive driving paerns from he Naional Household Travel Survey (NHTS) daa. The colleced daase is publicly available and conains rip daa repored by more han 150,000 U.S. households [23]. The survey asked household members for all rips hey compleed on a paricular day in The allocaion of reporing days o households was done such ha a roughly equal share of daa was colleced for each weekday/monh ype. In addiion o he rips, households provided informaion abou vehicles in use. Trips are provided in DAYV2PUB.csv, vehicles in VEHV2PUB.csv, all downloadable from he NHTS web sie [23]. To obain he subse of he NHTS daa we are ineresed in, we applied a number of selecion crieria oulined in he following. We seleced all rips compleed by car of equal size as he PEV model considered in his sudy. Daily driving profiles are assembled by collecing he rips of individual vehicles compleed on one day. We do no use any daily driving profiles ha eiher begin a oher locaions han home or end a such locaions. To ensure a sufficien level of daa qualiy, we applied a number of plausibiliy rules. Redundan rips resuling from he repors of several household members who shared he same vehicle were deleed. If rips in a daily driving profile overlapped, he enire profile was removed from he daabase. Some rip repors implied unrealisically high speeds. We herefore deleed enire daily profiles if he average speed of one of is rips exceeded 80 mph. Afer all selecion crieria and qualiy assurance had been applied, we obained a oal of number of 140,707 daily driving profiles from an iniial se of 178,070 profiles. We creaed a sofware ool ha uses randomly sampled daily driving profiles from his daabase o generae one year long driving profiles. To produce a realisic se of such yearly driving profiles wih characerisics corresponding o he daabase of daily driving profiles, we subdivided he laer ino weekday/work (n w/w = 43, 933 profiles), weekday/nonwork (n w/nw = 61, 199 profiles), and weekend profiles (n w = 35, 575 profiles). The NHTS driving profile generaor concaenaes daily driving profiles based on he assumpion ha vehicle usage can be disinguished ino wo ypes: commuing o work and oher purposes from Monday o Friday (work vehicles) and usage for oher purposes only (non-work vehicles). We assume r

6 6 ha no vehicle is used for commuing o work on weekends. The fracion of weekend daily driving profiles ha included a leas one work-relaed rip is less han 10%. For his sudy, we generaed a oal number of 5,000 yearly driving profiles. To accuraely reflec he characerisics of he NHTS daase, we generaed n w/w /(n w/w + n w/nw ) 100% = 41.79% of hese yearly profiles as work vehicle profiles. Before saring he acual generaion of hese yearly driving profiles, each work vehicle is assigned a commue disance drawn from an empirical probabiliy disribuion. We derive his disance disribuion direcly from he se of all weekday/work daily driving profiles in he daabase. We assume ha he disance o work remains he same during he arge year. The weekday pars of he yearly driving profiles of he work vehicles are composed of randomly chosen daily driving profiles of he weekday/work ype wih he same work disance. The weekday pars of he non-work vehicle yearly driving profiles are generaed by concaenaing daily profiles of he weekday/non-work ype. For he weekend pars of boh he work and non-work vehicle driving profiles, he ool appends randomly chosen daily driving profiles of he weekend ype. Many vehicles ha could be replaced by PEVs were no used on he day ha NHTS asked for. We compued he corresponding non-usage fracions on weekdays and weekends by aking he number of deleed profiles in each caegory ino accoun. The non-usage fracions are and , respecively. For he weekday/work profiles we assume ha he non-usage fracion is zero since he corresponding vehicles are used for commuing on every weekday. During he generaion of he yearly driving profiles, he profile generaor randomly insers idle days (insead of weekday/non-work and weekend daily driving profiles from he daabase) based on he corresponding non-usage fracions. Based on he vehicle usage daa, we inferred PEV charging jobs by assuming he relevan characerisics of he firs mass marke PEV equipped wih a range exender: he Chevrole Vol. According o [24], he Vol has an average elecrical range of 35 miles and requires 240 minues o recharge using a level 2 charging device. Each charging job has a sar ime, i.e. when he PEV is plugged in, and a sop ime, i.e. when he PEV leaves for he nex rip. The ime in beween can heoreically be used for charging. In his work, we assume ha all PEVs can be charged whenever and wherever hey are parked. Based on he sae of charge of each vehicle a he sar ime of each parking inerval, he required recharging ime of each vehicle can be compued. As in [6], we assume a greedy charging policy: Each PEV maximizes he baery charge available a he end of each parking inerval. This approach guaranees he non-disrupive operaion as defined in [2]. The difference beween he parking inerval and he inferred charging ime is he charging slack ha can be used o shif charging load. We pariion he parking inerval ino one minue ime inervals ha can eiher be used for charging or no. To our knowledge, here currenly exiss no publicly available daa on acual elecric vehicle charging efficiency over ime using differen schedules. Therefore we assume ha charging efficiency is independen of charge rae, and ha he energy sored in a baery a he end of an inerval depends only on he oal energy delivered o he PEV and, in paricular, is independen of he power rajecory. Finally, we resric he parking inerval usable for load shifing o he firs 24 hours from he beginning of he charging inerval. Thus, if a vehicle is parked for several days, all charging has o be done wihin he firs day. The addiional load resuling from PEV charging is based on he leas expensive level-2 charging device available for purchase oday [18]. Is average power draw is 3.3 kw (240 Vols, 14 Amperes). C. Simulaion Framework Our simulaion framework combines he models for regulaion and load following demand described in Secion III-A wih he PEV charging model described in Secion III-B. Formally, he wo models are conneced by he erm L P EV in Equaion (2) and he consideraion of L P EV in he real-ime and hour-ahead generaion scheduling processes (Equaions (3) and (9)). I allows for simulaing he impac of a populaion of PEVs and he way hey are charged on he regulaion and load following demand in California. The simulaion is based on a number of assumpions: (i) A each ime sep, he sysem operaor can adjus he PEV load only a he curren ime sep. (ii) The sysem operaor can use he enire slack of he charging job, i.e. we assume ha he sysem operaor knows he ime of deparure upon he sar of each parking inerval. However, we resric he look-ahead o 24 hours. (iii) The sysem operaor only considers charging jobs for load shifing if heir sar imes have been reached, i.e. he sysem operaor does no predic fuure charging jobs. (iv) Aside from PEV charging conrol, he sysem operaor s curren operaing procedures remain unchanged. Assumpion (iv), implies ha he real-ime generaion schedule is updaed every 5 minues based on he real-ime load, wind and solar generaion forecass. These forecass are available 8 minues before he sar of he corresponding operaing inerval. The load-following insrucions are sen o he available generaors 5 minues before he sar of each inerval. The unis begin o move o he insruced oupu level 2.5 minues afer he insrucions are sen. Thus, each updae of he real-ime generaion schedule caused by changes o he real-ime load, wind, solar or PEV load forecas shows effec afer 2.5 minues already. We refer o he ime period beween he receip of fresh daa on fuure load, wind and solar and he ime ha he corresponding adjusmen begins o show as he no impac period. Furhermore, we call he ime beween he end of he no impac period and he sar of he par of he schedule ha reflecs all informaion parial impac period. Figure 2(a) provides a schemaic overview of he real-ime generaion scheduling process of he CAISO. The effec of he delayed availabiliy of new forecasing daa on he hour-ahead generaion schedule is similar o he effec on he real-ime generaion schedule. The hour-ahead schedule is updaed hourly based on 2 hours old daa on

7 7 (a) Real-ime (b) Hour-ahead Fig. 2. Time lines for generaion scheduling; Figures are based on [15]. non-pev load, wind and solar as well as he PEV charging schedule known a his ime. The corresponding generaion insrucions are sen o he generaors 75 minues before he sar of he corresponding operaing hour. This resuls in a 65 minues long no impac period and a 20 minues long parial impac period. Changes o he PEV charging schedule ha concern hese periods are no or only parially refleced by he generaion schedule. Figure 2(b) provides a graphical view of he hour-ahead scheduling ime line. D. PEV Charging Algorihm Charging schedules are iniialized when PEVs connec o he grid (cf., Secion I). PEV charging jobs can be iniialized differenly, i.e. he acual charging ime of a job can be disribued in many differen ways over he available ime slos in he parking inerval a firs. To evaluae he impac of differen ways of upfron charging job iniializaion, we consider wo exreme cases: In he firs case, charging is scheduled o sar a he beginning of he parking inerval and coninues unil eiher he baery is fully loaded or he parking inerval ends. In he second case, all required charging is scheduled as lae as possible. We refer o hese wo ypes of charging iniializaion as early and lae, respecively. The wo considered charging job iniializaion mehods provide complemenary opporuniies for load shifing. Early iniializaion gives he algorihm he abiliy o insananeously decrease demand by deferring charging jobs o laer inervals. Lae iniializaion provides capaciy o insananeously increase demand by moving charging jobs originally scheduled for laer inervals o earlier inervals. We assume he sysem operaor mainains a searchable daabase of currenly parked vehicles. For each vehicle, he daabase conains he vehicle s power capaciy and he charging inenions (idle/charging) during each minue of he ime ha he PEV is parked. The PEV charging conrol algorihm we evaluae works as follows: A every ime sep, he sysem operaor deermines he amoun of regulaion demand required in he curren period. For each currenly parked PEV, he algorihm checks wheher i is available, i.e. i can conribue o he required service (regulaion up or regulaion down). A PEV is available if i has no ye reached is maximum sae of charge and will sill be able o mee is argeed sae of charge if charging is deferred for one ime sep. From he pool of available PEVs, he algorihm engages enough PEVs o saisfy he amoun of regulaion demand (unless he available pool is oo small, in which case he algorihm uses as much of he pool as possible). A he vehicle level, regulaion up is provided by deferring charging in he curren ime sep o he nex ime sep in which he vehicle is no currenly scheduled o charge. This reduces he load a he curren ime sep and is hus equivalen o a corresponding increase of generaion. Regulaion down is achieved by moving load from he las ime sep in which he vehicle is scheduled o charge o he curren ime sep. In [25], he auhors explore several ypes of scheduling policies for managing deferrable loads like PEVs. They show ha he well-known Earlies Deadline Firs (EDF) scheduling algorihm performs well in his conex. Using his knowledge as saring poin, we propose a similar selecion crierion ha reflecs he specialies of our seing: For regulaion down, hose PEVs wih he earlies deadlines bu are no ye charging are insruced o begin charging. For regulaion up, hose PEVs ha are currenly charging wih he laes deadlines are insruced o sop charging for one ime sep. In summary, he scheduling algorihm advances or defers demand from he early and lae charge iniializaion baselines. Is goal is o reduce he demand for regulaion from he supply side by as much as possible. The consrains of his algorihm are imposed by he properies of he charging jobs and he requiremen of non-disrupive use. Decisions are made wihou predicing demand for regulaion or changes of he consrains. IV. NUMERICAL STUDY In he following, we presen he resuls of a numerical sudy which we conduced based on he simulaion model described in Secions III-A and III-B.

8 8 A. Seup We simulae one enire year of grid and vehicle flee operaion based on scaled grid daa colleced in 2005 [11] and arificial driving profiles generaed from daa colleced in 2009 [23]. To consider he impac of wind and solar peneraion on regulaion demand, we used he scaling of he corresponding capaciies suggesed in [12] and [21] as provided in Table II. 1 This allows us o analyze he impac of higher renewables peneraion in California on he simulaion resuls and hus provides furher insighs regarding he prospec of inegraing more inermien renewables by leveraging PEVs as disribued energy sorage. Our simulaion framework has hree scaling parameers for PEV populaion size. The firs parameer, a, specifies how many imes each vehicle race is used o insaniae a conrollable group of PEVs. The second parameer, b, denoes he number of yearly driving profiles we generaed using he approach described in Secion III-B. The hird parameer, c, denoes he number of acual PEVs in each of he conrollable groups. Thus, he oal number n of simulaed PEVs is he produc of he number of he hree scaling parameers a, b, and c. The minimum change of PEV charging power draw is equal o 3.3c kw. The accuracy of he maching of regulaion demand and supply from he PEVs herefore negaively correlaes wih c. Increasing a or b o be able o reduce c increases he compuaional cos of he simulaion, since every PEV group is conrolled independenly and memory requiremens increase wih he number of driving profiles used in he simulaion. By using his kind of seup for creaing a conrollable PEV populaion, we make sure ha he informaion conained in our sample of vehicle races is used enirely and he differen races are weighed equally. In our simulaion experimens, we keep b consan a 5,000, c a 100, and vary a in a compuaionally accepable range, namely beween 1 and 6. This allows us o simulae he impac of 0.5 up o 3 million Chevrole Vols in seps of 0.5 million cars. Since he objecive of he load shifing approach is o provide boh up and down regulaion, we iniialize he charging jobs derived from kn driving profiles using early and (1 k)n using lae charging iniializaion. Table V provides an overview of he parameer configuraions we use in he numerical sudy. B. Resuls 1) Effec of Conrolled PEV Charging: In his secion we demonsrae he effec of he PEV load shifing approach described above. Figure 3 shows he impac of load shifing. The figure depics differen merics colleced during wo subsequen days of simulaion (Jan 24-26, 2020). Figures 3(a) and 3(b) give an impression of he regulaion capaciy up and down ha he CAISO could command if one million PEVs were conrolled. I was generaed by summing up he corresponding capaciies a each minue of he 1 Our ool can be used o simulae arbirary peneraions of wind and solar power generaion. However, he characerisics of he forecas errors are likely o change a higher peneraion levels. simulaed day wih load shifing deacivaed. Especially in he morning and early evening hours, when mos PEVs are expeced o recharge afer rips o and from he work place of heir drivers, he PEV regulaion capaciy is relaively high. During he nigh, he available regulaion capaciy is raher low: On he one hand, many charging jobs prescribing early recharging canno be used o provide regulaion up anymore since even a full recharge akes a mos 4 hours. On he oher hand, mos charging jobs ha prescribe lae recharging canno be rescheduled o deliver regulaion down since he ime sep from which onward he corresponding PEVs have o sar recharging heir baeries anyways in order o fulfill he service level has been reached. Figures 3(a) and 3(b) also reveal how changing k from 0.5 o 0.8 increases he regulaion up capaciy, bu a he same ime decreases he regulaion down capaciy of he PEV flee. The change is more dramaic in he case of regulaion down which indicaes ha more PEVs are required o provide regulaion up han regulaion down. Figure 3(c) shows he demand for regulaion if he charging rae of PEVs is no conrolled. The demand srongly flucuaes and can increase from maximum demand for up regulaion o maximum demand for down regulaion in jus a few minues. Figure 3(d) shows he demand for regulaion if he PEVs are conrolled o reduce regulaion demand. Load shifing leads o a clearly observable reducion of he regulaion down demand, especially during imes of high regulaion capaciy in he morning beween 8 and 10 A.M., as well as during he evening hours beween 5 and 10 P.M.. The figures also show ha, a leas a a flee size of 1 million PEVs, here remain imes when he demand for regulaion canno be significanly reduced by load shifing; paricularly in he middle of he he nigh. Figures 3(e) and 3(f) depic he addiional load resuling from PEV charging. One can clearly disinguish he deep valleys in he aggregaed load paern ha are caused by deferring he corresponding load o laer ime periods. The maximum PEV load during he observed ime inerval wihou load shifing is approximaely 500 MW, whereas he PEV load wih load shifing peaks a 1,000 MW. 2) Performance of Conrolled PEV Charging: We presen our resuls regarding he performance of he PEV load shifing approach in his secion. Performance is measured by he reducion of regulaion capaciy ha is necessary o balance supply and demand of elecric power in he grid. The provision of regulaion capaciy mainly depends on he highes demand for regulaion up and he lowes demand for regulaion down ha can be expeced under reasonable condiions. We calculae hese maximal requiremens using he mehod proposed in [12]. For each parameer configuraion, our simulaion runs 10 imes hrough all hours of he year. In each hour, he maximal regulaion up and minimal regulaion down requiremen across all minues of he hour are seleced. These are called he hourly up and down capaciy requiremens. Someimes only one of hem exiss. Hence, we obain a mos 10 hourly up and 10 hourly down capaciy requiremens for each hour of each day in he year. From each of hese ses, we again selec he maximum/minimum value. The overall procedure hus resuls in 365 up and 365 down capaciy requiremens of which we repor he corresponding

9 9 MW MW MW MW MW MW :00:00 06:00:00 12:00:00 18:00:00 00:00:00 06:00:00 12:00:00 18:00:00 00:00: Time (a) Regulaion up capaciy of PEVs k=0.5 k=0.8 00:00:00 06:00:00 12:00:00 18:00:00 00:00:00 06:00:00 12:00:00 18:00:00 00:00: Time (b) Regulaion down capaciy of PEVs k=0.5 k=0.8 00:00:00 06:00:00 12:00:00 18:00:00 00:00:00 06:00:00 12:00:00 18:00:00 00:00: Time (c) Regulaion demand wihou conrolled PEV charging k=0.5 k=0.8 00:00:00 06:00:00 12:00:00 18:00:00 00:00:00 06:00:00 12:00:00 18:00:00 00:00: Time (d) Regulaion demand wih conrolled PEV charging k=0.5 k=0.8 00:00:00 06:00:00 12:00:00 18:00:00 00:00:00 06:00:00 12:00:00 18:00:00 00:00: Time (e) PEV load wihou conrolled charging k=0.5 k=0.8 00:00:00 06:00:00 12:00:00 18:00:00 00:00:00 06:00:00 12:00:00 18:00:00 00:00:00 Time (f) PEV load wih conrolled charging k=0.5 k=0.8 Fig. 3. Effecs of conrolled PEV charging. k is he fracion of vehicles charged according o early charge iniializaion; he remaining fracion follow he lae iniializaion proocol.

10 10 Parameer Values Descripion k 0.5*; 0.6;...; 0.8 Fracion of PEVs wih early charging iniializaion n 0.5 m; 1 m*;...; 3 m Toal number of PEVs S 0 (no PEVs); 1 (PEVs wihou conrol); 2 (PEVs wih conrol) PEV scenario Y 2012; 2020* Simulaed year TABLE V MODEL SENSITIVITY PARAMETERS (* INDICATES DEFAULT VALUE) Per cen change Y=2012, k=0.5 Y=2020, k=0.5 Y=2012, k=0.6 Y=2020, k=0.6 Y=2012, k=0.7 Y=2020, k=0.7 Y=2012, k=0.8 Y=2020, k=0.8 Per cen change Y=2012, k=0.5 Y=2020, k=0.5 Y=2012, k=0.6 Y=2020, k=0.6 Y=2012, k=0.7 Y=2020, k=0.7 Y=2012, k=0.8 Y=2020, k= Number of PEVs in millions Number of PEVs in millions (a) Mean up (b) High up Per cen change Y=2012, k= Y=2020, k=0.5 Y=2012, k= Y=2020, k=0.6 Y=2012, k= Y=2020, k=0.7 Y=2012, k=0.8 Y=2020, k= Number of PEVs in millions Per cen change Y=2012, k= Y=2020, k= Y=2012, k=0.6 Y=2020, k= Y=2012, k=0.7 Y=2020, k= Y=2012, k=0.8 Y=2020, k= Number of PEVs in millions (c) Mean down (d) Low down Fig. 4. Percenage change of required regulaion energy (a & c) and capaciy (b & d). k denoes he fracion of PEVs wih early charging iniializaion. 95h perceniles. This means ha we idenify he highes regulaion up and he lowes regulaion down value afer eliminaing he 2.5% highes regulaion up values and he 2.5% lowes regulaion down values. The mean amoun of regulaion up and down can be inerpreed as a measure of he regulaion energy ha is acually required o balance he grid in differen scenarios. 2 When he acivaion of regulaion capaciy causes addiional coss, in paricular when regulaion up is needed, a higher level of regulaion energy delivered by PEVs is of addiional imporance. Figure 4 shows how he percenage reducion of he mean and high/low regulaion capaciy requiremens evolve as he oal number of PEVs is scaled up from 0.5 o 3 million. The percenage change merics are calculaed according o Equaion (12), where d(s) denoes he corresponding regulaion 2 In fac, his value is no equal bu proporional o he required regulaion energy. However, since we consider only relaive reducions of he regulaion demand, his difference is irrelevan. demand meric in scenario S, e.g. he 95h percenile of he maximal regulaion up capaciy. p = d(2) d(1) d(1) 100% (12) If he percenage change merics approach +/- 100%, he PEV flee is able o supply all regulaion in he corresponding case. The percenage mean regulaion energy demand ha can be saisfied using conrolled PEV recharging is always higher han he corresponding percenage capaciy demand. For insance, wih 1 million conrolled PEVs, beween and 75% of he regulaion up energy can be supplied, whereas only beween 25 and 35% of he capaciy can be covered. Regulaion up appears o be harder o saisfy using PEVs: A a peneraion of 2 million PEVs and k = 0.5, all regulaion down capaciy could be supplied, bu only 35% of regulaion up capaciy. As already indicaed in Figures 3(a) and 3(b), higher levels

11 11 of k increase he regulaion up capaciy a he expense of regulaion down capaciy. Ineresingly, he resuls shown in Figure 4 indicae ha a small increase of he mean or high regulaion up supply has o be raded agains a high reducion of he regulaion down supply. For example, wih 1 million conrollable PEVs, a 10% increase of he regulaion up energy supply by changing k from 0.5 o 0.8 causes a reducion of he regulaion down energy supply by 40%. Figure 4 also reveals he difference in he regulaion supplied in 2010 and 2020: In erms of percenage of required regulaion energy, he same number of PEVs will be able o conribue less in 2020 compared o For insance, he conribuion of 1 million conrollable PEVs o regulaion down energy will decrease by approximaely 5%. This is due o he fac ha boh he capaciy of inermien renewables and he load is expeced o grow which also increases he absolue amoun of regulaion required (cf., he assumpions described in Secion III-A). Surprisingly, he same number of PEVs will be able o saisfy more of he acual boleneck resource in 2020 compared o 2012: regulaion up capaciy. Alhough his difference does no exceed 5%, i is remarkable: I indicaes ha he paerns of peak regulaion up demand evolve in a way ha is advanageous o supply i wih PEVs. 3) Effec of Conrolled PEV Charging on Load Following: Shifing PEV load will no only have an effec on regulaion, bu also on load following. I is imporan o consider his impac since, in order o reduce generaion cos, a reducion of he regulaion demand should no be bough by a significan increase of he load following requiremens. The CAISO updaes he hour-ahead generaion schedule only every hour. Therefore he arrival of load shifing aciviy affec i on much longer ime scales (cf., he no impac period in Figure 2). Using our simulaion model, we are able o quanify he change of load following requiremens caused by load shifing in he same way as he regulaion requiremens (cf., Secion IV-B2). Insead of calculaing he mean and 95h percenile values of RD for S = {1, 2}, we calculae hese values for LF D (cf., (5)). Aferwards, we apply he percenage change formula (12) accordingly. Our resuls show ha load shifing increases he demand for load following up, boh in erms of energy and capaciy, in he range beween 0.5 and 3% depending on n and k. A he same ime, i reduces he demand for load following down in he range beween 0.5 and 1.5%. In he ligh of he subsanial residual load following resources conrolled by he CAISO [12], he negaive impac of PEV load shifing on is average cos of supplying load following up should be limied. One could also argue, ha he negaive impac of increased loadfollowing up capaciy is compensaed by a reducion of load following down requiremens. V. CONCLUSIONS In his aricle, we have proposed a model-based approach o assess he poenial of ICT-conrolled PEVs as suppliers of regulaion in California. Our model considers he acual generaion scheduling processes of he CAISO in use oday and evaluaes our proposed PEV load shifing approach based on represenaive daa. The evaluaed approach for PEV charging conrol greedily saisfies he grid s demand for regulaion up and down using he slack in curren PEV charging jobs wihou reducing he elecric mileage of PEVs (cf., non-disrupive conrol menioned in [2]). I has only limied compuaional and communicaion requiremens and could hus scale o large PEV populaions. The regulaion capaciy depends on he way ha charging ime slos are disribued across he PEVs parking inervals upon arrival ime. We evaluae a simple scenario in which a randomly chosen fracion k of he conrolled PEVs uses early and he remaining fracion 1 k uses lae charging iniializaion. Alhough many more iniializaion mehods are conceivable, his approach should effecively reveal he maximum poenial of PEV load shifing under he given resricions, in paricular no V2G capabiliy, non-disrupive conrol, no forecasing of driving behavior, and no informed selecion of PEVs. Relaxing he resricions could lead o an improved capabiliy of PEVs o saisfy he demand for regulaion. Our resuls indicae ha roughly 3 million PEVs wih he operaional characerisics of a Chevrole Vol would suffice o supply a large par of he regulaion up and down demand in California. Ineresingly, [1] also use California as an example o quanify he number of vehicles needed o provide regulaion capaciy. Based on heir assumpions (V2G capabiliy, half of he flee available anyime, 15 kw power draw and feed-in, 1,200 MW of regulaion capaciy), his number would be (1,200 MW / 15 kw) 2 = 1,000. Our numbers are significanly higher, mainly since we assume a charging rae of 3.3 kw and no V2G capabiliy. Moreover, due o our more advanced approach, we are able o consider he dynamics of regulaion demand as well as he PEVs acual availabiliy for charging (which is lower han he one assumed in [1]). We were also able o show how his resul changes if more inermien renewables are used o generae elecriciy in he near fuure. Moreover, we evaluae he impac of PEV load shifing for supplying regulaion on load following requiremens concluding ha i is limied. Anoher major resul of his work is ha providing regulaion up necessiaes significanly more conrolled PEVs han regulaion down and hus represens a boleneck. The gap in he supply of regulaion up could be efficienly closed by making a subse of he PEV flee V2G capable. A relevan research quesion ha could be answered by exending our model is, for insance, how many V2G-capable PEVs are needed o close he idenified gap in he regulaion up supply and based on which properies he corresponding V2G-capable PEVs can be deermined. Our resuls have significan implicaions for policy making in California. To our knowledge, his is he firs work ha inegraes wo policy dimensions, namely renewable inegraion and elecric mobiliy, a a per minue level of deail. Based on he changes of regulaion and load following capaciy we presen, a policy maker could compue a realisic financial reurn on invesmen of PEV grid inegraion.

12 12 ACKNOWLEDGMENT The auhors would like o hank Yuri Makarov, Clyde Louan, Joe Eo, and Sascha von Meier for heir help wih applying he CAISO models. Furhermore, we hank Anhony Papavasiliou for useful discussions helping o improve his aricle. Finally, we would like o hank Oliver Günher, Moriz Hezer, Ben Fabian and Andreas Paul for providing us wih wih compuaional resources. [22] CAISO, 33% Renewable Inegraion Trajecory Case: Assumpions, Forecas Errors, and Renewable Profiles. hp:// 2b3ed87027c10.xls, Mar Las accessed: 16 Dec [23] U.S. Deparmen of Transporaion, Naional Household Travel Survey (NHTS). hp://nhs.ornl.gov/, Las accessed: 20 Mar [24] General Moors, Chevrole Vol Technology. hp://www. chevrolevolage.com/images/sories/volage U Conen/abou% 20vol%20powerpoin.pdf, Las accessed: 8 Sep [25] A. Subramanian, M. Garcia, A. Dominguez-Garcia, D. Callaway, K. Poolla, and P. Varaiya, Real-ime Scheduling of Deferrable Elecric Loads, REFERENCES [1] W. Kempon and J. Tomic, Vehicle-o-Grid Power Implemenaion: From Sabilizing he Grid o Supporing Large-Scale Renewable Energy, Journal of Power Sources, vol. 144, no. 1, pp , [2] D. Callaway and I. A. Hiskens, Achieving Conrollabiliy of Elecric Loads, Proceedings of he IEEE, vol. 99, no. 1, pp , [3] L. Yao and H.-R. Lu, A Two-Way Direc Conrol of Air-Condiioning Load Via he Inerne, IEEE Transacions on Power Delivery, vol. 24, no. 1, pp , [4] T.-F. Lee, M.-Y. Cho, P.-J. Hsiao, and F.-M. Fang, Opimizaion and Implemenaion of a Load Conrol Scheduler Using Relaxed Dynamic Programming for Large Air Condiioner Loads, IEEE Transacions on Power Sysems, vol. 23, no. 2, pp , [5] D. Callaway, Tapping he Energy Sorage Poenial in Elecric Loads o Deliver Load Following and Regulaion, wih Applicaions o Wind Energy, Energy Conversion and Managemen, vol. 50, no. 5, pp , [6] Z. Ma, D. S. Callaway, and I. A. Hiskens, Decenralized Charging Conrol of Large Populaions of Plug-in Elecric Vehicles, IEEE Transacions on Conrol Sysems Technology, pp. 1 12, [7] A. Brooks, L. E., D. Reicher, C. Spirakis, and B. Weihl, Demand Dispach Using Real-Time Conrol of Demand o Help Balance Generaion and Load, IEEE Power & Energy Magazine, vol. May/June, [8] R. Sioshansi, R. Fagiani, and V. Marano, Cos and Emissions Impacs of Plug-in Hybrid Vehicles on he Ohio Power Sysem, Energy Policy, vol. 38, no. 11, pp , [9] A. Brooks, Vehicle-o-Grid Demonsraion Projec: Grid Regulaion Ancillary Service wih a Baery Elecric Vehicle, ech. rep., AC Propulsion, Inc., [10] M. Galus, S. Koch, and G. Andersson, Provision of Load Frequency Conrol by PHEVs, Conrollable Loads, and a Cogeneraion Uni, IEEE Transacions on Indusrial Elecronics, vol. 58, no. 10, pp , [11] CAISO, Grid Daa. hp:// Las accessed: 23 Mar [12] CAISO, Inegraion of Renewable Resources, Operaional Requiremens and Generaion Flee Capabiliy a 20% RPS, ech. rep., California Independen Sysem Operaor, [13] CAISO, Inegraion of Renewable Resources: Technical Appendices for California ISO Renewable Inegraion Sudies, ech. rep., California Independen Sysem Operaor, [14] CAISO, Operaing Procedure M-410: Day-Ahead Marke. hp://www. caiso.com/hegrid/operaions/opsdoc/markeops/index.hml, Las accessed: 2 May [15] Y. V. Makarov, C. Louan, J. Ma, and P. de Mello, Operaional Impacs of Wind Generaion on California Power Sysems, IEEE Transacions on Power Sysems, vol. 24, no. 2, pp , [16] CAISO, Operaing Procedure M-403: Real-ime Marke. hp://www. caiso.com/hegrid/operaions/opsdoc/markeops/index.hml, Las accessed: 2 May [17] A. Papavasiliou and S. S. Oren, Supplying Renewable Energy o Deferrable Loads: Algorihms and Economic Analysis, in Power and Energy Sociey General Meeing, July 2010, Minneapolis, Minnesoa, USA, [18] C. Goebel, On he Business Value of ICT-Conrolled Plug-In Elecric Vehicle Charging in California, Energy Policy, forhcoming. [19] B. Kirby, Ancillary Services: Technical and Commercial Insighs, ech. rep., Elecric Power Sysem Consuling, [20] CAISO, OASIS Daabase. hp://oasis.caiso.com, Las accessed: 2 Apr [21] CAISO, Summary of Preliminary Resuls of 33% Renewable Inegraion Sudy, ech. rep., California Independen Sysem Operaor, Chrisoph Goebel Chrisoph Goebel received he diploma degree in informaion engineering and managemen from he Universiy of Karlsruhe (now Karlsruhe Insiue of Technology), Germany in 2006 and a docorae degree in informaion sysems from Humbold-Universiy Berlin, Germany in Unil his graduaion, he spen one year sudying compuer science a he Swiss Federal Insiue of Technology, Lausanne and half a year a Carnegie Mellon Universiy, Pisburgh, PA. Afer compleing his disseraion, he worked as a pos-docoral fellow a he Universiy of California, Berkeley for 1.5 years. In Berkeley, he sared research in he area of susainable energy managemen. Currenly, he is a posdocoral fellow a he deparmen of compuer science, Technical Universiy Munich, Germany. His research ineress are in he area of using informaion echnology (i) o faciliae renewable inegraion wih flexible loads and sorage devices, and (ii), o improve energy efficiency in differen conexs. Duncan Callaway Duncan S. Callaway (Member, IEEE) received he B.S. degree in mechanical engineering from he Universiy of Rocheser, Rocheser, NY, in 1995 and he Ph.D. degree in heoreical and applied mechanics from Cornell Universiy, Ihaca NY, in Currenly, he is an Assisan Professor of Energy and Resources and Mechanical Engineering a he Universiy of California, Berkeley and a faculy scienis a Lawrence Berkeley Naional Laboraory. Prior o joining he Universiy of California, he was firs a Naional Science Foundaion (NSF) Posdocoral Fellow a he Deparmen of Environmenal Science and Policy, Universiy of California, Davis, subsequenly worked as a Senior Engineer a Davis Energy Group, Davis, CA, and PowerLigh Corporaion, Berkeley CA, and was mos recenly a Research Scienis a he Universiy of Michigan, Ann Arbor. His curren research ineress are in he areas of (i) modeling and conrol of aggregaed elecriciy loads and sorage devices, (ii) spaially disribued energy resources, (iii) environmenal impac assessmen of energy echnologies, and (iv) using informaion echnology o improve building energy efficiency.