Challenges and Opportunities in Large-Scale Deployment of Automated Energy Consumption Scheduling Systems in Smart Grids

Transcription

1 Challenges and Opportuntes n Large-Scale Deployment of Automated Energy Consumpton Schedulng Systems n Smart Grds Hossen Akhavan-Hejaz, Zahra Baharloue, and Hamed Mohsenan-Rad Department of Electrcal Engneerng, Unversty of Calforna at Rversde, Rversde, CA, USA Department of Electrcal and Computer Engneerng, Isfahan Unversty of Technology, Isfahan, Iran Emals: {hossen.akhavan-hejaz, hamed}@ee.ucr.edu, zahra bahar@ec.ut.ac.r, Abstract Recent studes have shown that the lack of knowledge among users on how to respond to tme-varyng prces and the lack of effectve home automaton systems are two major barrers for fully utlzng the advantages of real-tme prcng. Therefore, there has been a growng nterest over the past few years towards developng automated energy consumpton schedulng (ECS) devces to constantly montor the hourly prces and schedule the operaton of users controllable load to mnmze ther energy expendture. Whle the pror results n usng ECS devces are promsng, all pror work are lmted to small-scale deployment of ECS devces. For example, n most cases, the users that are equpped wth the ECS devces are assumed to be part of a mcrogrd or a feeder connected to a sub-staton. In ths paper, we rather nvestgate large-scale deployment of ECS devces n a power grd wth several buses and generators. The prce of electrcty at each bus s set accordng to the locatonal margnal prce (LMP) at that bus. We show that a key challenge n largescale deployment of ECS devces s load synchronzaton. However, we propose to use a movng average smoothng mechansm for LMPs that can fx the load synchronzaton problem and stablze the system. Furthermore, we show that the proposed large-scale ECS system has a close to optmal performance n terms of reducng peak-to-average-rato n load demand, mnmzng the total power generaton cost, and lowerng users electrcty blls. Keywords: Energy consumpton schedulng, large power grd, load synchronzaton, real-tme prcng, locatonal margnal prce. I. INTRODUCTION Real-tme and tme-of-use electrcty prcng models can potentally lead to several economc and envronmental advantages compared to the current commonly used flat rates. In partcular, they can provde power consumers wth the opportunty to reduce ther electrcty expendture by respondng to prcng that vares at dfferent tmes of day and s hgher at peak load hours [1]. Furthermore, they can help utltes and ndependent system operators to reduce the peak-to-average-rato (PAR) n aggregate load demand whch can lead to mnmzng the need for buldng new power generaton capactes [2]. Despte several advantages that real-tme, tme-of-use, and other non-flat prcng models can offer, recent studes have shown that the lack of knowledge among users about how to respond to tme-varyng prces and the lack of effectve home automaton systems are two major barrers for fully utlzng the benefts of non-flat electrcty prcng tarffs [3], [4]. In fact, most of the current resdental load control actvtes are operated manually. Ths makes t dffcult for users to optmally schedule the operaton of ther applances n response to the hourly updated prcng nformaton they may receve from the utltes n a non-flat prcng program. For example, the experence of the real-tme prcng program n Chcago, IL has shown that although the prce values were avalable va telephone and the Internet, only rarely dd households actvely check prces as t was dffcult for the partcpants to constantly montor the hourly prce values to respond properly [5]. To tackle the problems wth manual load control, there has been a growng nterest recently towards usng automated energy consumpton schedulng (ECS) devces [6] [12], smlar to the one shown n Fg. 1. In ths setup, each user s equpped wth an ECS devce, e.g., n ts smart meter, whch s assumed to be connected to a smart power dstrbuton system wth a two-way dgtal communcaton capablty through computer networkng [13]. Based on the updated prcng sgnals that the ECS devce receves from the utlty through the avalable communcatons nfrastructure, and also gven the users personal energy needs, the ECS devce optmally schedules the energy consumpton for the users controllable load such that t can mnmze the users daly or monthly electrcty expenses. The use of ECS devces s recommended not only for resdental consumers [6] but also for ndustral consumers [14]. Furthermore, there have been companes that have already started offerng commercal ECS devces for home automaton products, e.g., see [15]. Whle the pror results n usng automated ECS devces n smart grds have been very promsng, all pror work along ths lne of research have been lmted to small-scale deployment of the ECS devces. For example, n most cases, the users that are equpped wth the ECS devces are assumed to be part of a mcrogrd or part of a small feeder n a dstrbuton lne that s connected to a sngle generator or a sub-staton. Therefore, n ths paper, we rather nvestgate large-scale deployment of ECS devces n power grd such as the one shows n Fg. 2. The prce of electrcty at each bus n ths system s assumed to be set accordng to the locatonal margnal prce (LMP) at that bus. Note that, most exstng deregulated electrcty markets n the Unted States currently use LMPs to settle varous bulk sale and ancllary servce transactons [16]. Although settng retal prces accordng to LMPs s stll not a common practce n most regons, t s recently shown that by reflectng the prces

2 Fg. 1. An automated energy consumpton schedulng devce n a smart meter [6] [12]. The prces are obtaned through a communcatons nfrastructure. n the wholesale market to the consumer sde, users wll be better encouraged to consume electrcty more effcently [17]. We wll show that a key challenge n large-scale deployment of ECS devces s load synchronzaton. Ths problem can be explaned as follows. Every tme the electrcty prces,.e., the LMPs, are set, the ECS devces move ther load from hghprce hours to low-prce hours n an attempt to mnmze ther energy expendture. However, ths wll n turn overload lowprce hours, makng them hgh-prce hours n the next teraton, and underload hgh-prce hours, makng them low-prce hours n the next teraton. Ths causes constant fluctuatons n the electrcty prces and makes the system unstable. To tackle ths problem, we propose to use a movng average smoothng mechansm for LMPs. Our smulaton results show that the proposed approach works well and can assure system stablty. Furthermore, we show that the proposed large-scale deployment of ECS devces has a very close to optmal performance n terms of reducng PAR n the aggregate load demand, mnmzng the total power generaton cost n the system, and reducng each user s ndvdual electrcty bll payments. The rest of ths paper s organzed as follows. The system model s explaned n secton II. The nteractons between the grd operator and the ECS devces s dscussed n Secton III. Smulaton results are presented n Secton IV. The conclusons and future work drectons are dscussed n Secton V. II. SYSTEM MODEL Consder a power grd system, such as the IEEE 24-bus system n Fg. 2(a). Let B, wth cardnalty B, denote the set of buses n the system. For each bus B, let N, wth cardnalty N, denote the set of users connected to bus. Clearly, f bus s not a load bus, then we have N =0. For each load bus, we assume that each user s equpped wth an ECS devce. An example for the case of bus 8 wth N 8 users s shown n Fg. 2(b). The prce of electrcty at each load bus s set accordng to the locatonal margnal prce at that bus. Let LMP h denote the locatonal margnal prce at load bus at hour h. Consder an H>1hours ahead energy consumpton schedulng problem for a user n N connected to bus. Note that for day-ahead plannng, we have H =24. (a) (b) Fg. 2. An example for large-scale deployment of automated ECS devces: (a) An IEEE 24-bus power system wth 16 load buses. (b) The set of N 8 users, equpped wth ECS devces, that are connected to bus 8. The retal prce of electrcty at each bus s set accordng to the LMP at that bus. Gven the followng H 1 prce vector LMP =[LMP 1,LMP 2,...,LMP H ], (1) the ECS devce n user n s smart meter s responsble for schedulng the operaton of all user n s controllable load such that user n s daly energy expendture s mnmzed. For each user n, let A n denote the set of all applances that have controllable / shftable load. Examples for such applances may nclude washer, dryer, dshwasher, and plug-n hybrd electrc vehcles. For each applance a A n, we defne an energy consumpton schedulng vector as x n,a =[x 1 n,a,x 2 n,a,...,x H n,a]. (2)

3 Let E n,a denote the total energy needed to fnsh the operaton of applance a. For example, E n,a =16kWh for a sedan electrc car wth 40 mles daly drvng range [1]. Furthermore, for each applance a, the operaton needs to be scheduled wthn a tme frame [α n,a,β n,a ], where 1 α n,a <β n,a H. These parameters are set by user n based on hs energy consumpton needs for each applance. For example, user n may set α n,a = 1:00 PM and β n,a = 5:00 PM for the operaton of a dshwasher after lunch table and before dner. Of course, the tme duraton β n,a α n,a must be larger than or equal to the tme needed to fnsh the normal operaton of applance a. To assure on tme operaton of applances, t s requred that user n s ECS devce fulflls the followng constrants β n,a x h n,a = E n,a. (3) h=α n,a Furthermore, t s requred that where x h n,a =0, h H\H n,a, (4) H = {1,...,H}, and H n,a = {α n,a,...,β n,a }. (5) Fnally, we note that some applances may have some mnmum standby power γn,a mn and/or some maxmum supported power γn,a max. In that case, t s also requred that γ mn n,a x h n,a γ max n,a, h H n,a. (6) For notatonal smplcty, for each user n, we ntroduce a new vector x n, whch s formed by stackng up energy consumpton schedulng vectors x n,a for all applances a A n. In ths regard, we can defne a feasble energy consumpton schedulng set correspondng to user n as follows: X n = {x n β n,a h=α n,a x h n,a = E n,a, x h n,a =0, h H\H n,a, γ mn n,a x h n,a γ max n,a, h H n,a }. An energy consumpton schedule calculated by the ECS unt n user n s smart meter s vald only f we have x n X n. For each user n N at bus, the total electrcty bll wthn the schedulng horzon of nterest s calculated as ( H LMP h L h n + ) x h n,a, (8) a A h=1 where L h n denotes the total load of user n at hour h due to hs applances that have non-controllable load. Examples for such applances may nclude lghts, refrgerator, televson and other entertanment devces. Note that the operaton of applances wth non-controllable load s not scheduled by ECS devces. To mnmze user n s energy expendture, the ECS devce n user n s smart meter should solve the followng optmzaton problem across applances that have controllable load: ( H mnmze LMP h L h n + ) x h n,a. (9) x n X n a A h=1 (7) (a) (b) Fg. 3. Interactons between the grd operator and the ECS devces. (a) The electrcty prces are set based on the orgnal LMPs. (b) The electrcty prces are set based on a smoothed verson of LMPs n order to enforce stablty. Note that the above optmzaton problem can capture the behavor of each user s ECS devce. Next, we nvestgate the nteractons between the ECS devces and the grd operator when the ECS devces are deployed n a large scale. III. OPERATOR-USER INTERACTIONS If the ECS devces are deployed only n small scales, e.g., n a mcrogrd or n a sngle dstrbuton feeder as n [6] [12], the operaton of ECS devces may not have any mpact on the LMPs. However, f the ECS devces are deployed n a larger scale and at several buses, such as n the power system n Fg. 2, then the operaton of the ECS devces may have a sgnfcant mpact on the LMPs at dfferent buses as we explan next. Let X h denote the total load at bus at hour h. Once all ECS devces set the load by solvng problem (9), we have = ( L h n + ). (10) X h n N a A n x h n,a Usng the standard power system dspatch control model n [18], at each hour h, the grd operator can solve the followng optmzaton problem to calculate the LMPs at each bus: mnmze G h, subject to B ( ) C G h (11a) B B G h X h =0 (11b) B G mn f k, (G h X h ) F max k, k K (11c) G h G max B, (11d) where G h denotes the amount of dspatched power generaton at generator bus at hour h, C ( ) denotes the cost functon for the generator at generator bus, K denotes the set of all transmsson lnes n the system, f k, denotes the [19] njecton shft factor to transmsson lne k from bus, and

4 F max k denotes the transmsson lmt of transmsson lne k. Fnally, G mn and G max denote the mnmum and maxmum generaton range for the generator at bus. Clearly, f bus s not a generaton bus, then we have G mn = G max =0. Assumng that power loss s neglgble on transmsson lnes, the formulaton of LMP at bus can be wrtten as [20], [21]: K LMP h = λ + f k, μ k, (12) k=1 where K denotes the number of transmsson lnes,.e., the cardnalty of set K, λ denotes the Lagrange multpler correspondng to the energy balance constrant n (11b), and μ k denotes the Lagrange multpler correspondng to the lne capacty constrant n (11c) for transmsson lne k K. A. Decentralzed Model The nteractons between the grd operator and ECS devces can be analyzed under the real-tme prcng framework n [22]. Gven the prce values,.e., vector LMP at each bus, the ECS devces schedule the load based on the optmal soluton of problem (9). In turn, f the updated load profles are replaced n optmzaton problem (11), the resulted LMPs can become dfferent from the orgnal values. Ths s shown n Fg. 3(a). Note that, the message exchanges are supported through the two-way dgtal communcatons capablty whch s expected to be avalable n the future smart grd [1]. The key queston s: Do the back and forth teratons between the grd operator and the ECS devces converge to any fxed pont? To answer ths queston, we perform a smulaton based on the power grd topology n Fg. 2. The detaled smulaton setup s explaned n Secton IV. As shown n Fg. 4, the objectve value of the generaton dspatch problem (11),.e., the total cost power generaton n the system, does not converge. The fluctuatons n ths fgure can be explaned as follows. Every tme the prces are set, the ECS devces move ther load from hgh-prce hours to low-prce hours. Ths wll n turn overload low-prce hours, makng them hgh-prce hours n the next teraton, and underload hgh-prce hours, makng them low-prce hours n the next teraton. Ths problem s referred to as load synchronzaton [6]. Whle load synchronzaton does not have a major mpact on electrcty prces when the ECS devces are deployed only n a small scale, large-scale deployment of the ECS devces can cause sgnfcant nstablty n the prce sgnals as well as the aggregate load profles, as t s evdent from the smulaton results n Fg. 4. Next, we propose a movng average smoothng mechansm for LMPs to resolve the load synchronzaton problem. Let LMP [t] denote the locatonal margnal prce vector at bus that s obtaned by solvng optmzaton problem (11) at teraton t 1. We ntroduce a smoothed verson of LMP at teraton t, denoted by LMP [t], to be calculated as follows: LMP [t +1]=(1 η t )LMP [t]+η t LMP [t], (13) where 0 η t 1 s an teraton-dependent step-sze. Choosng a dmnshng step-sze can partcularly assure convergence to a fxed pont. Therefore, we select η t as t 0 η t = t 0 + t 1, (14) Generaton Cost (Thousand Dollars) Average Cost = $282.5K Iteraton Number Fg. 4. The fluctuaton n total power generaton cost n the system when the electrcty prces are set based on the orgnal LMPs as n Fg. 3(a). Generaton Cost (Thousand Dollars) Average Cost = $281.3K Iteraton Number Fg. 5. The total power generaton cost n the system when the electrcty prces are set based on the smoothed verson of LMPs as n Fg. 3(b). where t 0 1 s a fx parameter. As teraton number t, step-sze η t 0. In the new model, the nteractons between the grd operator and the ECS devce becomes as n Fg. 3(b). The smulaton results n ths case are also shown n Fg. 5. Note that, once the prce sgnals sent to the ECS devces converge to a fxed pont, the load profles wll also stop changng and the whole system reaches an equlbrum. B. Centralzed Model Before we conclude ths secton, t s worth emphaszng that the nteracton between the grd operator and the ECS devces shown n Fg. 3 s due to the fact that the utlty / grd operator does not usually have any centralzed control over the operaton of users personal applances. In fact, for each user, the ECS devce n hs smart meter does not follow the utltes commands. Rather t solely responds to the prce sgnals sent by utltes and ams to mnmze the energy expendture specfcally for ts correspondng user. However, f the grd operator does have drect control over the operaton of ECS devces, e.g., as n a drect load control (DLC) framework [23], then the nteractons between the grd operator and the ECS devces would no longer be based on Fg. 3. Instead, the operator would solve the followng global optmzaton problem and t would send the obtaned optmal energy

5 schedules as a command sgnal to each correspondng ECS devce to enforce optmal energy consumpton schedulng: mnmze G h,, x n X n, n subject to B ( ) C G h (15a) B B G h X h =0 (15b) X h B G mn = ( L h n + n N a A n x h n,a ) (15c) f k, (G h X h ) F max k, k K (15d) G h G max B, (15e) where X h acts as an auxlary varable. Recall that, n (11), X h was a known constant. The centralzed desgn n (15) s not the focus of ths paper as t may not be practcal as users could be reluctant to relnqush full control of ther load to utltes. Nevertheless, the soluton of optmzaton problem (15) can provde a benchmark to assess the performance of our proposed dstrbuted desgn n Secton III-A, when t comes to mnmzng the total cost of power generaton n the system. Fg. 6. Load Profle (MW) No ECS Centralzed ECS Dstrbuted ECS Hour of Day The daly load profle for varous ECS deployment scenaros. Generaton Cost (Thousand Dollars) No ECS Centralzed ECS Dstrbuted ECS IV. PERFORMANCE EVALUATION To assess the performance of the dstrbuted ECS system, we consder the IEEE 24-bus relablty test system [24]. It has a maxmum of 2650 MW total load at any hour. To allevate the computaton burden and to better see the mpact of energy consumpton schedulng n the overall system performance, the scale of each user s load s assumed to be relatvely hgh, such as the case for a major ndustral unt. The total load s dstrbuted among 100 users located across all load buses. Each user has both uncontrollable and controllable load. A. Peak Shavng To have a base for comparson, we examne the scenaro where no ECS unt s nstalled and users start ther consumpton rght after the start tme α n,a and contnue untl the operaton of the applance s done. Ths results n the load curve shown n Fg. 6 wth a PAR of On the other hand, when ECS systems are deployed, the PAR decreases as users shft part of ther controllable load from peak hours to off-peak hours. Ths s shown n Fg. 6. For the results n ths fgure, t s assumed that for each user about 50% of the load s controllable. The PAR for the case of dstrbuted ECS s The PAR further decreases to only 1.23 for the case of centralzed ECS. Recall from Secton III-B that whle centralzed ECS deployment may not be practcal t provdes a benchmark to assess the performance of our proposed dstrbuted desgn n Secton III-A. B. Reducng Total Power Generaton Cost The total power generaton cost n the system when the porton of controllable load vares from 0 to 40% s shown Controlable Load (%) Fg. 7. The total power generaton cost n the system versus the porton of controllable load for varous ECS deployment scenaros. n Fg.7. We can see that although the proposed large-scale dstrbuted ECS deployment system cannot acheve the same benchmark performance as n a centralzed energy consumpton schedulng scenaro, ts performance s close to optmal and much better than the case wth no ECS deployment. C. Beneft to Users In addton to shavng the peak load and reducng the total power generaton cost n the system, large-scale deployment of ECS devces can help each user reduce hs electrc bll. Ths s shown n Fg.8, where 50% of the load s controllable. We can see that all users on all buses can reduce ther bll compared to the case wth no ECS deployment. D. Collected Revenue by Utlty Fg.9 shows the collected versus ntended revenue from the users at dfferent controllable load percentages. The collected revenue s what users actually pay based on the smoothed LMPs. The ntended revenues are rather calculated based on what users should have pad f we use the orgnal LMPs. Interestngly, although the smoothed LMPs that are used to stablze the prce do not exactly match the orgnal LMPs, the total collected revenue s very close to (and even sometmes slghtly hgher than) the total ntended revenue n all

6 Average User Bll (Thousand Dollars) No ECS ECS Decentralzed ECS Centralzed Daly Revenue (Thousand Dollars) Intended Revenue Collected Revenue Bus 1 Bus 3 Bus 5 Bus 7 Bus 9 Bus 14 Bus 20 Bus Index % 40% 25% 20% 10% Controlable Load (%) Fg. 8. The average user blls for varous ECS deployment scenaros. Fg. 9. The ntended revenue based on orgnal LMPs versus the collected revenue based on smoothed LMPs n dstrbuted ECS deployment scenaro. scenaros. Therefore, the proposed large-scale dstrbuted ECS deployment system can be of nterest to utltes. V. CONCLUSION AND FUTURE WORK Ths paper represents the frst step towards understandng the challenges and opportuntes n large-scale deployment of automated energy consumpton schedulng devces n smart grds. To gan nsghts, we consdered an IEEE 24-bus relablty test system wth nne generator and 16 load buses. We assumed that all users on each load bus are equpped wth an ECS devce to obtan the updated prce nformaton from the smart grd and accordngly schedule the operaton of the user s controllable load to mnmze the user s electrcty bll. We showed that unlke the case when only a few users are equpped wth ECS devces, the large-scale deployment of ECS devces can drectly mpact the electrcty prces. In partcular, load synchronzaton can cause fluctuatons n locatonal margnal prces at dfferent buses. We proposed to fx ths problem usng a movng average smoothng mechansm for LMPs. We showed that once ths mechansm s appled, the nteractons between the grd operator and the ECS devces can be coordnated such that a very close to optmal performance s acheved n terms of reducng peak-to-average-rato n load demand, mnmzng the total power generaton cost n the system, and lowerng all users electrcty bll payments. The results n ths paper can be extended n several drectons. Frst, n addton to usng a smoothng mechansm, new prcng models can be examned to enforce stablty. Second, larger grd topologes wth renewable power generators can be consdered. Fnally, whle we assume that users are prce taker and gnore the mpact of ther load on LMPs, the scenaro where users are prce antcpator can be consdered. The nteractons n ths case can be studed, e.g., usng game theory. REFERENCES [1] A. Ipakch and F. Albuyeh, Grd of the future, IEEE Power and Energy Magazne, vol. 7, no. 2, pp , Mar [2] J. Medna, N. Muller, and I. Roytelman, Demand response and dstrbuton grd operatons: Opportuntes and challenges, IEEE Trans. on Smart Grd, pp , [3] Quantum Consultng Inc., Demand response program evaluaton - fnal report, LLC Workng Group 2 Measurement and Evaluaton Commttee and Calforna Edson Company, Apr [4] M. Ann-Pette, G. Ghatkar, S. Klccote, D. Watson, E. Koch, and D. Hennage, Desgn and operaton of an open, nteroperable automated demand response nfrastructure for commercal buldngs, Journal of Computng & Informaton Scence n Eng., vol. 9, pp. 1 9, June [5] H. Allcott, Real tme prcng and electrcty markets, Techncal Report, Harvard Unversty, Jan [6] H. Mohsenan-Rad and A. Leon-Garca, Optmal resdental load control wth prce predcton n real-tme electrcty prcng envronments, IEEE Trans. on Smart Grd, vol. 1, no. 2, pp , Sept [7] S. Caron and G. Kesds, Incentve-based energy consumpton schedulng algorthms for the smart grd, n Proc. of IEEE Smart Grd Comm, Gathersburg, MD, Oct [8] C. Ibars, M. Navarro, and L. Guppon, Dstrbuted demand management n smart grd wth a congeston game, n Proc. of IEEE Smart Grd Comm, Gathersburg, MD, Oct [9] D. Ren, H. L, and Y. J, Home energy management system for the resdental load control based on the prce predcton, n Proc. of IEEE Smart Grd Comm, Brussels, Belgum, Oct [10] N. Kumaraguruparan, H. Svaramakrshnan, and S. S. Sapatnekar, Prcng usng the multple knapsack method, n Proc. of IEEE PES Innovatve Smart Grd Technologes, washngton, DC, Jan [11] Z. Baharloue, H. Narman, and H. Mohsenan-Rad, Tacklng Coexstence and Farness Challenges n Autonomous Demand Sde Management, n Proc. of IEEE Globecom 12, Anahem, CA, Dec [12] H. Mohsenan-Rad, V. Wong, J. Jatskevch, R. Schober, and A. Leon- Garca, Autonomous demand sde management based on gametheoretc energy consumpton schedulng for the future smart grd, IEEE Trans. on Smart Grd, vol. 1, no. 3, pp , Dec [13] U.S. Department of Energy, The smart grd: An ntroducton, [14] J. Arnez and S. Bller, Integraton Requrements for Manufacturng- Based Energy Management Systems, n Proc. of IEEE Innovatve Smart Grd Technologes (ISGT 10), Gathersburg, MD, Jan [15] Energy Inc., The energy detectve, [16] M. Shahdehpour, H. Yamn, and Z. L, Market Operatons n Electrc Power Systems. New York, NY: IEEE Press, [17] M. Roozbehan, M. Dahleh, and S. Mtter, Dynamc prcng and stablzaton of supply and demand n modern electrc power grds, n Proc. of IEEE Smart Grd Comm, Gathersburg, MD, Oct [18] F. L and R. Bo, Congeston and prce predcton under load varaton, IEEE Trans. on Power Systems, vol. 24, no. 2, pp , May [19] P. W. Sauer, K. E. Renhard, and T. J. Overbye, Extended factors for lnear contngency analyss, n Proc. of the 34th Hawa Internatonal Conference on System Scence, Mau, HI, Jan [20] T. Orfanogann and G. Gross, A general formulaton for lmp evaluaton, IEEE Trans. on Power Systems, vol. 22, no. 3, Aug [21] Y. Fu and Z. L, Dfferent models and propertes on lmp calculatons, n Proc. of the IEEE Power Engneerng Socety General Meetng, [22] A. Conejo, J. Morales, and L. Barngo, Real-tme demand response model, IEEE Trans. on Smart Grd, vol. 1, no. 3, pp , [23] C. M. Chu, T. L. Jong, and Y. W. Huang, A drect load control of ar-condtonng loads wth thermal comfort control, n Proc. of IEEE PES General Meetng, San Francsco, CA, June [24] Relablty Test System Task Force of the Applcaton of Probablty Methods Subcommttee, IEEE relablty test system, IEEE Trans. on Power Apparatus and Systems, vol. 98, no. 6, pp , Nov