Capacity Estimation for Vehicle-to-Grid Frequency Regulation Services with Smart Charging Mechanism
11 Capacity Estimation for Vehicle-to-Grid FrequencyRegulation Services with Smart ChargingMechanism
Albert Y.S. Lam, Ka-Cheong Leung, and Victor O.K. Li
Abstract —Due to various green initiatives, renewable energywill be massively incorporated into the future smart grid.However, the intermittency of the renewables may result inpower imbalance, thus adversely affecting the stability of a powersystem. Frequency regulation may be used to maintain the powerbalance at all times. As electric vehicles (EVs) become popular,they may be connected to the grid to form a vehicle-to-grid (V2G)system. An aggregation of EVs can be coordinated to providefrequency regulation services. However, V2G is a dynamic systemwhere the participating EVs come and go independently. Thusit is not easy to estimate the regulation capacities for V2G.In a preliminary study, we modeled an aggregation of EVswith a queueing network, whose structure allows us to estimatethe capacities for regulation-up and regulation-down, separately.The estimated capacities from the V2G system can be used forestablishing a regulation contract between an aggregator andthe grid operator, and facilitating a new business model forV2G. In this paper, we extend our previous development bydesigning a smart charging mechanism which can adapt to givencharacteristics of the EVs and make the performance of the actualsystem follow the analytical model.
Index Terms —Capacity, queueing model, regulation services,vehicle-to-grid.
I. I
NTRODUCTION F OR a reliable power system, power balancing needs to bemaintained at all times; power generation and consump-tion must always be equal. Traditional power generations (e.g.,thermal power stations) and the renewables serve in the day-ahead market [2]. One of the most challenging problems ofincorporating the renewables into the power system is theirintermittency, rendering it difficult to predict the amount ofpower generated from the renewables accurately. It is possiblethat the resulting generation from the day-ahead market areexcessive or deficient compared with the predicted amount.The real-time market bridges the residual gap between thepower generation and the actual demand, accomplished bythe ancillary services, including frequency regulation, spinningreserve, supplemental reserve, replacement reserve, and volt-age control [3]. According to the United States (U.S.) FederalEnergy Regulatory Commission, ancillary services are “thoseservices necessary to support the transmission of electricpower from seller to purchaser given the obligations of control
A preliminary version of this paper was presented in [1].A.Y.S. Lam is with the Department of Computer Science, Hong KongBaptist University, Kowloon Tong, Hong Kong (e-mail: [email protected]).K.-C. Leung and V.O.K. Li are with the Department of Electric andElectronic Engineering, The University of Hong Kong, Pokfulam Road, HongKong (e-mail: { kcleung, vli } @eee.hku.hk). areas and transmitting utilities within those control areas tomaintain reliable operations of the interconnected transmissionsystem” [4]. Regarding load balancing, spinning, supplementalreserve, and replacement reserves are for contingency purposeswhile frequency regulation tracks on a minute-to-minute basis.In this paper, we focus on the ancillary service given byfrequency regulation.For a power system to function properly, the operatingfrequency should be maintained close to its nominal value,e.g., 60 Hz in the U.S. excessive power generated (i.e.,generation is larger than consumption) will drive the systemfrequency higher than the nominal setting while a deficiencyof generation results in a smaller system frequency. Frequencyregulation is the measure of adjusting the system frequencyto the nominal value by providing small power (positive ornegative) injections into the grid. Many Regional Transmis-sion Organizations (RTOs) and Independent System Operators(ISOs), e.g., PJM, simply call this service “regulation” [5].The balance of generation and demand between control areasis measured in terms of area control error (ACE). Eachcontrol area generates automatic generation control (AGC)signals based on its ACE values and the regulation resourcesrespond to the AGC signals to perform regulation. This isachieved through a real-time telemetry system and controlledby the grid operator. Purchase and sale of regulation servicesare accomplished in the regulation market managed by thecorresponding ISO/RTO. Consider PJM as an example [5].Resource owners submit offers to the Market Clearing Engine,which optimizes the RTO dispatch profile and determines cor-responding clearing prices. The market is cleared between theregulation resources and service purchasers with the clearingprices. The details can be found in [5], [6], [7].There have been some studies about integrating renewablesinto the grid more reliably and efficiently, such as [8]. Oneproposed solution is the introduction of energy storage to defer the excess for the future deficient. Examples of energy storageinclude batteries, flywheels, and pumped storage. In the nearfuture, one of the most realistic forms is batteries. This can bejustified by the expanding markets of plug-in hybrid electricvehicles or simply electric vehicles (EVs). For example, it isforecast that there will be 2.7 million EVs on the road in theU.S. by 2020 [9]. In California, it is expected that roughly70% of new light-duty vehicles and 60% of the fleets will beEVs [2]. The integration of EVs into the power grid is calledthe vehicle-to-grid (V2G) system depicted in Fig. 1 of [1].Frequency regulation requires power in the order of MW a r X i v : . [ c s . S Y ] M a y while each EV can only supply power around 10-20 kW [10].In order to provide regulation service from the V2G system, anaggregation of EVs is necessary and an aggregator coordinatesa group of EVs. The aggregators thus provide regulationservices to the grid, which are controlled and coordinatedby the operators. In general, an aggregator can be a parkingstructure or a facility coordinating the EV activities of thehouseholds in a residential area.To implement regulation in V2G, the aggregators need tomake contracts with the grid operators. The V2G system cansupport both regulation-up (RU) and regulation-down (RD)services. The former means that the grid does not have enoughpower supply and extra power sources (e.g., V2G) provide theshortfall. The latter refers to the situation in which extra powerloads are needed to absorb the excessive power.According to [11], short-term stored-energy resources, e.g.,batteries, are excellent candidates of regulation resources dueto their very fast response time to AGC and their capabilityof reducing CO emissions. Many ISOs, including New YorkISO, ISO New England, and California ISO, have been inte-grating short-term energy storage resources into their regula-tion markets. A team, composed of experts from the Universityof Delaware and PJM, conducted a practical demonstration toshow that V2G is capable of providing real-time frequencyregulation [12]. Experienced distribution engineers pointed outthat an EV has no difference from a distributed generator oradditional load for regulation in the technological viewpoint[12]. Moreover, [13] revealed that V2G has significant poten-tial for financial return from frequency regulation. Based on[14], private vehicles in the U.S. are driven less than an hour aday on the average and this implies that we may intelligentlyutilize EVs for other purposes when they are parked idly. Fromall these, we can see that V2G, as a regulation resource, willnot bring significant technological challenges to the existingpower system when regulation is taking place. Moreover, it hasa great financial incentive to be implemented in the practicalpower system.In the regulation market, a regulation resource owner needsto submit the capacity of its resource for the locationalmarginal price forecasts, which are in turn used to determinethe market clearing prices. One type of charges for the regula-tion services is capacity payment [15]. It refers to the servicecharges due to the V2G system only guaranteeing powersupport when the grid requires RU or RD. In other words,the V2G system gets paid even without any actual powertransfer. The grid operator pays for the service according tothe expected amount of power to be supplied and absorbedfor RU and RD, respectively. Unlike the traditional generatorswith high controllability, it is not easy to determine theregulation capacities for V2G. The main reason is that V2Gis a dynamic system, in which the governed EVs are allautonomous. The number of EVs managed by an aggregatorvaries from time to time and thus actual regulation capacitiescontributed by the EVs are also varying. Although V2G ispractically feasible [12] and financially favorable [13], we needto estimate the regulation capacity of V2G so as to put itinto operation extensively. In this paper, we focus on capacityestimation of the V2G system for an aggregator which can help estimate the total profit and set up a contract between anaggregator and the grid operator. A smart charging mechanismis designed to enhance the flexibility of the system. Due to thedynamics of EVs and the similarities between the batteriesof V2G (for power) and the buffers of the communicationnetworks (for data packets), we estimate the V2G capacity forregulation with the queueing theoretic approach, which hasbeen widely used for performance analysis in communicationnetworks [16].The rest of this paper is organized as follows. We givesome related work of V2G studies in Section II and a systemoverview in Section III. Section IV presents our analyticalmodel with the RU and RD capacities derived. The smartcharging mechanism is discussed in Section V and a per-formance study of the V2G system for the power regulationservices is presented in Section VI. Finally, Section VIIconcludes our work.II. R ELATED W ORK
The preliminary version of this work can be found in [1].In [1], we defined a queueing network model to estimate theRU and RD capacities. However, we assumed that there existsa smart charging mechanism which makes the service timesat various queues exponentially distributed. This exponentialdistribution property is one of the keys to developing math-ematically tractable closed-form solutions for the capacities.In this paper, we relax this assumption by explaining howsuch smart charging mechanism works. It allows the modelto function even when the attributes of EVs are distributed inunknown distributions. We also perform simulation to verifythe behavior of this mechanism when applied to various queuesin the model.There are many studies on V2G since it is expected tobe a major component in the future smart grid. In [15] and[17], V2G was systematically introduced with studies on thebusiness model for V2G. They gave information of differentkinds of EVs and different power markets, including baseloadpower, peak power, spinning reserves, and regulation. Themerits of V2G are quick response and high-value services withlow capital costs, but V2G has shorter lifespans and higheroperating costs per kWh. They gave some rough idea about thescale of V2G so as to make it comparable with the traditionalregulation from generators. V2G energy trading was studiedas an auction in [18]. Interested readers can refer to [19] fora comprehensive review on the impact of V2G on distributionsystems and utility interfaces of power systems.Queueing theory has been used to study the aggregate be-havior of EVs. In [20], a simple
M/M/c queueing model forEV charging was devised and a similar idea was also adoptedin [21] to determine V2G capacity. Ref. [22] suggested an
M/M/ ∞ queue with random interruptions to model the EVcharging process and analyzed the dynamics with time-scaledecomposition. The “service” process adopted was assumedto be exponential but it may not be practical unless there is aspecial arrangement to conform with the exponential property.In this work, we will design a smart charging mechanism toovercome this problem. Some existing work investigated V2G frequency regulationand the capacities of V2G systems. In [23], an optimalcharging control scheme for maximizing the revenue of anEV from supporting frequency regulation was proposed. In[24], the problem was formulated as a quadratic program andan efficient algorithm considering discharge was devised. Ref.[25] considered the user pattern to develop an approximateprobabilistic model for achievable power capacity, which wasthen utilized to determine the contracted power capacity. How-ever, this contracted power capacity (similar to the capacityaimed to be achieved in this work) depends on the paymentmethodology, which is a business consideration rather than anengineering issue. In [15], the V2G capacity was achieved bytaking the maximum of the current-carrying capacity of theconnecting wires and the power available from the vehiclebattery. The former can be improved with advances in wiringtechnology. The latter simply converts the stored energy intopower with the consideration of efficiency factors. However,the achieved capacity in [15] is for one single vehicle, whichis assumed to be static. For an aggregation of EVs withstochastic arrivals and departures, this estimation may not bevery practical. In [26], the V2G capacity was estimated usingdynamic EV scheduling, where the design was mainly targetedon the building energy management system. Its proposedalgorithm utilized the forecasted building load demand andEV charging profiles to estimate the V2G capacity. This designmay not be applicable to frequency regulation as load demandof the supported region may not be available to an aggregator.In this work, we compute the separate RU and RD capacitiesfor an aggregation of EVs without any assumptions on thestored energy of EVs and their durations of stay. Moreover,our computation does not rely on any payment methodologyand this provides flexibility for other business considerations.To the best of our knowledge, there is no unified study on thecapacity management for both RU and RD in the V2G system.There are many studies about coordinated EV charging. Theissue of time-varying electricity price signals on cost-optimalcharging was considered in [27]. A decentralized algorithmfor coordinating the charging/discharging schedules of EVsin order to meet the regulation demand was devised in [28].Refs. [29] and [30] focused on the network aspects. Ref. [29]investigated the impact of EV charging on distribution gridsand showed that controlled charging can result in significantreduction of overloaded network components. In [30], theauthors proposed a linear-approximation-based framework foronline adaptive EV charging. It could reduce the violations onvarious network limits, e.g., flow limit and voltage magnitudelimit, due to high penetration of EVs. Ref. [31] performedEV charging according to the electricity price. Cao et al. [31]proposed a control method to coordinate charging with thetime-of-use market price. Ref. [32] studied a EV chargingmethod for smart homes and buildings in the presence ofphotovoltaic systems. In [33], Miao et al. proposed a mobility-aware charging strategy for globally optimal energy utilizationby means of appropriately routing mobile EVs to chargingstations with the assistance of vehicular ad-hoc network. Allthese efforts attempt to design EV charging strategies based onthe various system objectives. In this paper, we take a different perspective; we construct a smart charging mechanism forcapacity management for V2G regulation services.III. S
YSTEM O VERVIEW
Each EV is assumed to be autonomous. It can participateand leave the V2G system according to the schedule of theEV owner. Once an EV is connected or plugged to the system,it will be actively charged and/or support regulation until itdeparts. When being actively charged, it pays for the amountof energy consumed. While it is supporting the regulation, itreceives payment for providing the service. Regulation canresult in either EV charging or discharging, depending onwhether RU or RD is requested. Thus, it is possible for an EVto get paid while it is being charged (i.e., in an RD event).Charging events in which an EV requests charging itself andsupports regulation are called active charging and reactivecharging , respectively. A discharging event happens only whenan EV participates in supporting RU. Both the residual energystored in the battery of the EV at the time it arrives and theenergy charged from active or reactive charging can be usedto support RU. However, an EV cannot support RU when itsbattery is fully discharged. Similarly, an EV cannot participatein supporting RD when its battery is fully charged up. Sincethe battery capacity is finite, the amount of energy stored ina battery affects its potential for supporting the RU and RDservices. In this paper, we aim to estimate the capacities of anaggregator for regulation services so that it will be beneficialfor an aggregator to establish a contract with the grid operators.Hence, we only consider the events of active charging. Thecharging and discharging rates due to regulation are smallenough so that the estimated capacities for regulation servicesare not affected by charging and discharging events due toregulation.Now we focus on a particular aggregator. We denote theset of EVs, each of which has registered at the aggregator forproviding the regulation service, by I . The events associatedwith each EV and among EVs are independent with each other.All EVs are assumed to be heterogeneous such that they canbe equipped with batteries of different capacities when fullycharged. The state-of-charge (SOC) of an EV refers to theamount of energy stored in its battery normalized with themaximum capacity. We denote SOC of EV i at time t by x i ( t ) .Without loss of generality, we assume x i ( t ) ∈ [0 , , ∀ i ∈ I .We also define the target SOC of EV i as the amount of energy,normalized with the maximum battery capacity, that the EV’sowner aims to reach when it departs, given in a range [ x i , x i ] ,where x i and x i are the lower and upper limits of the targetSOC of EV i , and ≤ x i ≤ x i ≤ . In other words, if EV i leaves the system at time t (cid:48) , it aims to satisfy x i ≤ x i ( t (cid:48) ) ≤ x i .If the target SOC is merely a value, we have x i = x i .The lower SOC target threshold x i represents the minimumtargeted amount of energy, normalized with the maximumbattery capacity, retained for EV i when it departs from thesystem. Hence, it is designed to meet the mobility pattern ofEV i . For example, an EV which travels a lot in betweentwo successive chargings requires a higher x i . If an EV canbe charged quite frequently, a lower x i may be sufficient to support its operation. On the other hand, x i is defined forregulation. Recall that a fully charged EV cannot provide theRD service. x i < means that EV i reserves room of size (1 − x i ) for later RD opportunities or other purposes. At time t ,if x i ( t ) is smaller than x i , active charging always happensin order to bring SOC to the target range. However, activecharging must stop when x i ( t ) reaches x i , since no future RUevent is guaranteed to happen in order to bring SOC back tothe target range.Recall that regulation can result in charging or dischargingto an EV. We can increase x i ( t ) by both active and reactive(i.e., RD) chargings, while we can only reduce x i ( t ) by RU.Hence, when EV i is still connected to the system, it will bein one of three states according to the value of x i ( t ) , each ofwhich supports different regulation services, as follows: • State 1) x i ( t ) ≤ x i : Only RD (reactive charging) isallowed. • State 2) x i < x i ( t ) < x i : Both RU and RD are allowed. • State 3) x i ( t ) ≥ x i : Only RU (discharging) is allowed.For simplicity in the analysis, we do not consider that theEVs are actually charged or discharged due to regulation inthis paper. Let r i ( t ) ≥ be the active normalized chargingrate of EV i at time t and it is constant over time in eachstate, i.e., r i ( t ) = r i , t ≥ . Consider that EV i is plugged inat time t and its SOC is x i ( t ) . If it is actively charged at rate r i , after a time period ∆ t , we have x i ( t + ∆ t ) = x i ( t ) + r i ∆ t .From the standpoint of an EV owner, the primary concern isto charge its EV such that it has enough battery level to supportits operation. The profit derived from providing the ancillaryservices is of secondary concern. In other words, an ownerconsiders to provide the ancillary services from its EV onlyif the remaining energy (after discharging from providing theancillary services) is enough to support its operation. Hence,we propose the following simple charging policy: When EV i arrives at the system with SOC below x i , it will be activelycharged until x i is reached. Otherwise, no active charging isrequired.In fact, we can always set x i = x i to simplify the system.EV i supports RD when x i ( t ) is below x i , and it supports RUwhen x i ( t ) goes above x i . However, suppose x i ( t ) = x i whenthe system is supporting regulation (i.e., it can be actuallycharged or discharged due to regulation). When there existsa random sequence of RU and RD requests, the EV will beoscillating between States 1 and 3 previously discussed andthis will make the system unstable. The introduction of State2 can help stabilize the system.Note that we aim to perform capacity management byestimating the capacities for regulation to help construct acontract between an aggregator and a grid operator. There aredifferent kinds of regulation contracts in the market: • RD only: An EV always absorbs power from the gridto provide the service. To maximize the profit, we cansimply set x i = x i = 0 so as to reserve the largest roomfor energy absorption. • RU only: An EV always supplies power to the grid whenproviding the service. To maximize the profit, we cansimply set x i = x i = 1 to preserve as much energy inthe battery as possible for future discharging events. • RU and RD: Both RU and RD are allowed. We wouldset < x i < appropriately to balance the demand forRU and RD.We consider the V2G system supporting both RU andRD. We can define two kinds of capacities for the V2Gregulation services, namely, the RD capacity and
RU capacity .The former refers to the total amount of energy that can beabsorbed by the system to support RD. Similarly, the lattercorresponds to the total amount of energy available from thesystem to support RU. Here, we focus on determining the RUand RD capacities of one particular aggregator. The capacityof the whole V2G system can then be seen as the sum of thecapacities of the individual aggregators. In the next section,we propose an analytical model to estimate the two capacitiesof an aggregator for our charging policy.IV. A
NALYTICAL M ODEL
In this section, we model an aggregator with a queueingnetwork. We first define the settings of the model from thesystem discussed in Section III and give some assumptions.Then, we construct a queueing network, which is used toestimate the RU and RD capacities.
A. Settings
The V2G system is modelled as a queueing network withthree queues, namely, the regulation-down queue (RDQ), regulation-up-and-down queue (RUDQ), and regulation-upqueue (RUQ). When an EV is plugged in at time t , the decisionto join which queue depends on its SOC x i ( t ) . If it is inStates 1, 2, and 3 (defined in Section III) at time t , it willjoin the RDQ, RUDQ, and RUQ, respectively. After joining aparticular queue, the following will happen:
1) RDQ:
Each EV i in this queue is actively charged atits own normalized charging rate r i . If its SOC reaches x i attime t (cid:48) , i.e., x i ( t (cid:48) ) = x i , it will leave RDQ and join RUDQ.The duration is determined by: ∆ t = t (cid:48) − t = x i − x i ( t ) r i , x i ( t ) < x i . (1)When an EV is actively charged, it gets served in the queue.It is also possible for it to depart from the queue before itsSOC has reached x i . This represents the situation that it quitsthe system.
2) RUDQ:
When EV i arrives at this queue, it will beactively charged at the normalized charging rate r i until itsSOC reaches x i . If the charging process starts at time t andthe EV is charged to x i at time t (cid:48) , the duration is given by: ∆ t = t (cid:48) − t = x i − x i ( t ) r i , x i < x i ( t ) < x i . (2)If the EV joins from RDQ, we have: ∆ t = t (cid:48) − t = x i − x i r i . (3)After charging up to x i , the EV departs from this queue andgoes to RUQ. Similar to RDQ, a departure of an EV fromthe queue before its SOC reaching x i corresponds to an EVleaving the system. ∞ λ + ∞∞ + λ λ λ λ q − q q RDQ p p p − q λ RUDQRUQ µ µ µ Fig. 1. The queueing model.
3) RUQ:
When an EV joins this queue, no active chargingtakes place. It will stay in this queue until it departs from thesystem.
B. Assumptions
We make the following assumptions to make the analysismathematically tractable:1) The events associated with each EV and among EVsare independent with each other. Each EV arrives atthe system randomly, following a Poisson process atrate λ . Among the EV arrivals, fractions p , p , and p of EVs are in States 1, 2, and 3, respectively, where p , p , p ∈ [0 , and p + p + p = 1 .2) There exists a smart charging mechanism M SC :( x i ( t ) , x i , x i ) (cid:55)→ r i , which assigns the normalized charg-ing rate r i to EV i according to its current SOC x i ( t ) upon its arrival at time t , and its target SOC thresholds x i and x i . With such mechanism, the durations of EVsin States 1 and 2 (refer to (1), and (2) and (3), respec-tively) are exponentially distributed at rates µ and µ ,respectively.3) There exists a fraction q of EVs in State 1 which willdirectly quit the system. This fraction represents thoseEVs whose SOCs do not reach their lower target SOClimits at their departures from the system. Similarly, wehave a fraction q of EVs which depart from the systemin State 2. Note that q and q already capture those EVswhich require fast charging. In other words, they mayonly stay in the system for a short period of time.4) When an EV is in State 3, no charging would happen.It will remain on standby in the system for a periodexponentially distributed with rate µ .The values of λ , p , p , q , and q can be determined bystatistical measurements from the operations of the chargingfacilities. The smart charging mechanism M SC can helpmaintain exponential service times in States 1 and 2 and amodified M SC will be adopted to maintain exponential servicetimes in States 3. We will discuss the design of (modified) M SC in Section V. C. Model
Fig. 1 depicts the queueing model, where RDQ, RUDQ,and RUQ model the behaviors of EVs in States 1, 2, and3, respectively. Assumption 1 states that we randomly splitthe EV arrival process into three subprocesses according to the probability distribution ( p , p , p ) . Since random split-ting results in independent Poisson subprocesses, the externalarrivals at each queue constitute a Poisson process with rate λ for RDQ, λ for RUDQ, and λ for RUQ, where λ = p λ , λ = p λ , and λ = p λ .When an EV enters RDQ, active charging starts immedi-ately. In other words, all EVs in this queue can get servedwithout queueing up. With Assumption 2, an EV resides inthis queue with a duration exponentially distributed with rate µ . Hence, RDQ can be modelled as an M/M/ ∞ queue witharrival rate λ and service rate µ . According to [34], theprobability p ,n of having n EVs in this queue in the steadystate is: p ,n = ( λ µ ) n e − ( λ µ ) n ! . (4)The expected number L of EVs charging in RDQ is: L = ∞ (cid:88) n =1 np ,n = λ µ = p λµ . (5)By Burke’s theorem [35], the departure process of RDQis a Poisson process with rate λ . With Assumption 3, thisPoisson process is split randomly according to the probabilitydistribution ( q , − q ) . (1 − q ) of EVs enter RUDQ withrate λ = (1 − q ) λ , which superposes with the Poissonsubprocess for the external arrivals with rate λ . Since thesuperposition of Poisson processes is still a Poisson process,the combined arrivals to RUDQ constitute a Poisson processwith rate ( λ + λ ) . Similar to RDQ, RUDQ can also bemodelled as an M/M/ ∞ queue with arrival rate ( λ + λ ) and service rate µ . Hence, the probability p ,n of having n EVs in this queue is: p ,n = ( λ + λ µ ) n e − λ λ µ n ! . (6)The expected number L of EVs charging in RUDQ is givenby: L = λ + λ µ = λ ( p + p − p q ) µ . (7)With Assumption 3, the departure process is Poisson withrate ( λ + λ ) , which is split randomly according to theprobability distribution ( q , − q ) . (1 − q ) of EVs enter RUQas a Poisson process with rate λ = ( λ + λ ) · (1 − q ) forRUQ. The combined arrival process of RUQ is also a Poissonprocess with rate ( λ + λ ) . With Assumption 4, RUQ can bemodelled as an M/M/ ∞ queue with arrival rate ( λ + λ ) and service rate µ . Therefore, the probability p ,n of having n EVs in this queue is: p ,n = ( λ + λ µ ) n e − λ λ µ n ! . (8)The expected number L of EVs standing by in RUQ canbe expressed as: L = λ + λ µ = λ (1 − p q − p q − p q + p q q ) µ . (9) The overall system departure process is a Poisson processsuperposed by three individual departure Poisson processesfrom the three queues. The overall departure process has rate: λ = q λ + q ( λ + λ ) + ( λ + λ ) . (10)The duration of each regulation service ∆ t reg is normallyshort, such as a few minutes [3], while EVs are expectedto switch their states in a relatively much lower rate. Thus,the mean service times of the queues, µ , µ , and µ aregenerally much longer than a few minutes. This is justifiableas an EV cannot be charged up nor leave the system within afew minutes on the average. For each EV, the amount of power P EV contributed for a regulation event can be determined withthe amount of energy required ∆ x EV by P EV = ∆ x EV ∆ t reg . As an aggregator normally coordinates hundreds of EVs, P EV contributed by a single EV is small. Hence, ∆ x EV would be even smaller. For a particular regulation contractwith the fixed regulation service duration ∆ t reg , we can fix P EV to be small enough such that the probability of havinga state transition of an EV after an absorption or a removalof energy of ∆ x EV for a regulation service is almost neg-ligible. Therefore, the capacities for the regulation servicescan be estimated based on the numbers of EVs available forregulation. Due to the types of regulation supported by EVsas described in Section III, the steady state RD capacity C RD can be computed as: C RD = P EV ( L + L ) . (11)Similarly, the steady state RU capacity C RU is given by: C RU = P EV ( L + L ) . (12)V. S MART C HARGING M ECHANISM
Recall that in Section IV-B, a smart charging mechanism M SC is adopted to assign charging rates to EVs based on theirbattery statuses so that their durations in States 1 and 2 followexponential distributions. We also apply a modified M SC tothe EVs in State 3 such that their durations in the systemfollow an exponential distribution as well. However, this M SC is posed as an assumption when the model is developed inSection IV. The model will not work and the resultant capacityestimation cannot be validated if M SC does not exist. Tocomplete the model, in this section, we discuss the designof such a smart charging mechanism.When EV i arrives at the system, it specifies its arrival time t ai , its expected departure time t di , its lower and upper SOCtarget thresholds x i and x i , and its initial SOC x i ( t ai ) . Themain purpose of the smart charging mechanism is to assign aservice duration w i to EV i such that:1) the service times assigned to a set of EVs statisticallyfollow the exponential distribution;2) the service time should not be longer than the expectedduration of the EV staying in the system; and3) the required charging rate should fall into the range of [ r, r ] , where r and r are the lower and upper chargingrate limits supported by the system. Interested readers can refer to Section VI for performance results.
Generate a random service time , , , , ( ) a ai i i i i id t t x x tx
Input:
Is empty?
Found? ( ) ai i i x txr ry d ai i y t t
Examine each number y in Q starting from the head satisfyingandNo Y e s No Y e s Remove y from Output y ~ Exp( ) z and d ai i tz t ( ) ai i i x tzxr r ?Put z to No Output z Y e s END
START Fig. 2. Flow chart of the smart charging mechanism for RDQ.
Point 1 ensures that an analytical model for capacity charac-terization can be developed based on Section IV-C. Point 2guarantees that the EVs will not stay longer in the systemthan expected. That is, w i should be smaller than or equal to t di − t ai . For Point 3, when the EVs stay in RDQ and RUDQ,their batteries need to be charged up to certain levels, andthus, the charging rates should be feasible for the system. Fora particular queue, M SC aims to assign the arriving EVs withservice durations, which statistically follow an exponentialdistribution. Upon an arrival, M SC will assign a serviceduration to the EV. In general, not every randomly generatedservice duration fits the condition of the EV. However, wecan reserve any unfit random service duration for another EVwhich comes at a later time. We introduce Lemma 1 andCorollary 1 to formally provide the underlying mathematicalreasoning. Lemma 1.
Consider that the sequence y ( t ) = [ y , y , . . . , y t ] is a realization of the independent and identically distributed(i.i.d.) random process { Y i } . Let σ be any permutation of theindices , , . . . , t . y (cid:48) ( t ) = [ y δ (1) , y δ (2) , . . . , y δ ( t ) ] is also arealization of { Y i } .Proof: Regardless of the order, y ( t ) and y (cid:48) ( t ) contain thesame set of numbers { y , . . . , y t } . As { Y i } is i.i.d., y (cid:48) ( t ) isalso a realization of { Y i } . Corollary 1.
Consider that y ( t ) = [ y , y , . . . , y t ] is arealization of the i.i.d. { Y i } . For ≤ k ≤ t (cid:48) ≤ t , y (cid:48) ( t ) = [ w , . . . , y k − , y k +1 , . . . , y t (cid:48) , y k , y t (cid:48) +1 , . . . , y t ] is alsoa realization of { Y i } . Since RDQ, RUDQ, and RUQ have different expectedservice times and SOC charging ranges (see Section IV), weapply variants of the smart charging mechanism for the threequeues and their design principles are largely similar. We willexplain the detailed design for RDQ and then point out thedifferences for RUDQ and RUQ:
1) RDQ:
The design principle is to set the charging rates r i ’s of the EVs by assigning with corresponding w i ’s, where w i ’s follow the exponential distribution with mean µ as muchas possible. Consider a sequence of exponentially distributednumbers [ y , y , . . . ] with mean µ , and they are potentialservice times of EVs in State 1. The idea is that, when anEV comes, we assign the y j with the smallest index j toits service time such that y j fits its specifications. Once y j has been adopted, we remove it from the sequence. Fig. 2shows the implementation details. We maintain a first-in-first-out queue Ψ to temporarily store the previously unadoptedrandom numbers for y . Suppose that EV i with t ai , t di , x i , x i ,and x i ( t ai ) enters the queue. We first check if there are anyunassigned numbers in Ψ . If so, we examine each number y in Ψ to see if y satisfies y ≤ t di − t ai (13)and r ≤ x i − x i ( t ai ) y ≤ r. (14)If there are multiple qualified y ’s, we select the y which wasthe earliest to be generated. Then we remove y from Ψ and set EV i ’s service time w i = y . Condition (13) meansthat the service time w i will not be longer than the EV’sexpected duration of stay, i.e., t di − t ai . Since EV i in thisqueue will be charged up to x i , the amount of energy willbe charged is x i − x i ( t ai ) , and thus, the required chargingrate is r i = x i − x i ( t ai ) y . Condition (14) ensures that r i canbe supported by the system. We examine from the head ofthe queue to ensure that the earliest qualified number in thequeue can be adopted first. If there is no qualified y in Ψ , wekeep generating random numbers z ’s which are exponentiallydistributed with mean µ until we get one z satisfying (13)and (14). All those unqualified numbers will be queued upin Ψ . Note that a random service time y which violatesthe conditions of a particular EV, i.e., (13) and (14), maybe qualified with another EV. Our goal is to keep the lengthof Ψ at all times as short as possible. In this way, we justpostpone some random numbers used at a later time. Here wejust repeatedly apply Corollary 1 to design M SC .
2) RUDQ:
The M SC design for RUDQ is similar to thatfor RDQ but the random numbers should follow exponentialdistribution with mean µ as much as possible. All EVs willbe charged up to their x i ’s. If EV i comes from RDQ, theamount of energy needed to be charged is x i − x i . Otherwise,the required charged amount is x i − x i ( t ai ) . We maintain aqueue Ψ to store the unqualified random numbers.
3) RUQ:
We modify the previously discussed M SC forRUQ. Instead of setting charging rates, we manipulate theservice times for the EVs only. The design principle is alsosimilar to that for RDQ but the random numbers shouldseemingly follow exponential distribution with mean µ . AllEVs in RUQ do not need to be actively charged, and thus, therate constraint like (14) is not required. We maintain a queue Ψ to store the unqualified random numbers. When an EVis assigned with a service time shorter than its own expectedduration of stay, it can still physically park at the parkinginfrastructure but we just disconnect it from the system.We need a performance metric to determine how well M SC facilitates the analytical model discussed in Section IV-C.We have Theorem 1 to introduce the queue length of Ψ toevaluate the performance of M SC . It says that if the queuelength remains finite, the service durations assigned to theEVs will follow an exponential distribution in the long run.The details can be found as follows: Theorem 1.
Let N ( t ) be the queue length of Ψ at time t , and y ( t ) = [ y , . . . , y t ] and w ( t ) = [ w , . . . , w t ] be the randomnumber sequence generated from an exponential distributionwith mean µ and the set of qualified service times adopted bythe first t participating EVs, respectively, at RDQ. If t tends toinfinity and lim t →∞ N ( t ) < ∞ holds, w ( t ) is exponentiallydistributed with mean µ almost surely.Proof: y k ’s in y ( t ) are split into two sets, either be-coming w k ’s in w ( t ) or being stored in Ψ . The condition lim t →∞ N ( t ) < ∞ implies that only a finite number of y k ’sgo to Ψ and thus w ( t ) must contain an infinite number of w k ’s. When being assigned to w ( t ) , some y k ’s may have beenre-ordered; y k may have been assigned to w k (cid:48) , where k (cid:48) > k .By Corollary 1, permutations of y k ’s in the sequence does notaffect the exponential nature of the sequence. Although y ( t ) and w ( t ) are different, w ( t ) has inherited y ( t ) ’s exponentialdistribution property almost surely.Similar results also hold for RUDQ and RUQ. Note thatwe do not need to specify any requirement of the inputs;the initial SOCs x i ( t ai ) , the duration of stays t di − t ai , andthe lower and upper SOC target thresholds x i and x i canfollow unknown distributions. The smart charging mechanismcan adapt to the characteristics of the inputs and generateexponentially distributed service times with means specifiedfor the queues.VI. P ERFORMANCE E VALUATION
A. System Parameter Settings
We study the performance of the system with a parkingstructure example, where EVs arrive and leave independently.Consider a scenario that there are five EVs entering the parkingstructure per minute following a Poisson distribution on theaverage. 90% of EVs require charging, where their SOCs arebelow their upper target thresholds at their arrivals. One tenthof them do not, since they need parking only and their SOCsare above their respective upper target thresholds. Amongthose requiring a charge, based on [36], we assume theirinitial SOCs x i ( t ai ) ’s follow a truncated Normal distribution in the range of [0, 1] with mean 0.5 and standard deviation0.2. The EVs expect to be charged up to their SOC uppertarget thresholds x i ’s, each of which is set with a truncatedNormal distribution in the range of [ x i ( t ai ) , with mean x i ( t ai )+0 . − x i ( t ai )) and standard deviation . − x i ( t ai )) for EV i . The SOC lower target threshold x i of EV i is setto x i = x i × rand [0 . , . , where rand [0 . , . is a randomnumber uniformly generated in [0 . , . . There are differentcharging standards and the required charging times for typicalEV models vary from 20 minutes to 8 hours [37]. Moreover,the durations of parking also depend on the drivers’ drivingpractice, the purposes of parking, and the initial SOCs. Weassume that the durations of stay (i.e., t di − t ai for EV i )are normally distributed and truncated in the range of [60,780] minutes with mean 420 minutes and standard deviation60 minutes. With these settings, we classify EVs into statesaccording to their SOCs and simulation gives p ≈ . , p ≈ . , and p = 0 . . Based on the current chargingtechnologies, we set the range [ r, r ] of the charging rates as[0, 0.05], where r corresponds to fast charging [38]. We alsohave q = q = 0 . . The above shows how to determinethe system parameters, λ , p , p , p , q , and q , for ourillustrative scenario. Note that we do not require any specificEV arrival rates and distributions for SOC conditions andtotal durations of stays. Different scenarios just give differentparameter settings. B. Results for a Reference Set of µ , µ , and µ The only parameters that we can control are the expectedservice times of the EVs, in terms of µ , µ , and µ , at thethree queues by setting corresponding charging rates and/ordisconnecting EVs from the system. Smaller values of µ , µ , and µ give larger regulation capacities but inappropriatevalues may result in remarkable errors between the analyticaland actual capacities. We first define a reference set of µ = min − , µ = min − , µ = min − . In this way, EVsin State 1 will spend 50 minutes in this state on the average.When they exit State 1, 10% of them leave the system, whilethe other 90% of them transit to State 2 and continue to chargeup their batteries to their upper target thresholds with a meanservice time of 70 minutes. Next 10% of them leave the systemfrom State 2. and the rest of the EVs stay in State 3 (withoutcharging) with the mean residence time equal to 30 minutes.By (5), (7), and (9), the expected numbers of EVs in States 1-3 are L = 127 . , L = 296 . , and L = 129 . in thesteady state.We set P EV = 6 kW and ∆ t reg = 1 min. Thus, each EVabsorbs or delivers . kWh for each regulation service. Whencompared with the EV models already available in the market,such small charging and discharging rates of P EV result in aregulation event where the involved EVs do not switch stateswhen supporting regulation (i.e., no transition to other queuesmerely for regulation). For example, the Tesla Model S has abattery capacity ranging from 40 kWh to 85 kWh [39], andBYD e6 has a battery capacity of 60 kWh [40]. With (11)and (12), the expected RD and RU capacities are C RD =6 kW × (127 . .
55) = 2543 . kW and C RU = 6 kW × (296 .
55 + 129 .
65) = 2557 . kW, respectively. (a) RDQ(b) RUDQ(c) RUQFig. 3. Variations of the number of EVs in each queue. We simulate the instantaneous numbers of EVs gettingserved in the queues for 1440 minutes. The simulation wasimplemented with Matlab. The results are exhibited in Fig. 3.The system is initially empty and it takes about 200 minutesto reach the steady state, where the numbers of EVs in thequeues oscillate around our computed expected values.Fig. 4 shows the queue length variations of Ψ , Ψ , and Ψ for M SC generated from the same simulation run as in Fig.3. Fig. 4(a) illustrates the variations of queue length of Ψ corresponding to the variations of EV population size shownin Fig. 3(a). Similarly, Figs. 4(b) and 4(c) correspond to Figs.3(b) and 3(c). In spite of the existence of some blips, thesizes of Ψ , Ψ , and Ψ do not grow continuously in the timehorizon. By Theorem 1, the actual system supported by thesmart charging mechanism will generally follow the analyticalresults developed in Section IV-C in the long run. Ψ and Ψ can return to empty queues from time to time and this impliesthat all their unqualified random numbers for service times atparticular instants can be adopted in their next few instants.The length of Ψ is stabilized at around 15 and it is likelythat the adoption rate and the generation rate of unqualifiedrandom numbers are similar. (a) Ψ (b) Ψ (c) Ψ Fig. 4. Variations of the queue lengths for M SC . C. Effects of Different Values of µ , µ , and µ Here we examine how the values of µ , µ , and µ affect thesystem performance. We consider the combinations of µ = min − , µ = min − , and µ = min − adopted inSection VI-B as a reference. We vary either µ , µ , or µ from the reference each time to investigate the changes inperformance.Fig. 5 gives the simulated RU and RD capacities and theerrors between analytical and simulated capacities for differentcombinations of values of µ , µ , or µ , where each erroris computed by simulated value - analytical valueanalytical value . In Fig. 5, “C” and“E” stand for “capacity” and “error”, respectively. Thus, forexample, C RD represents the simulated RD capacity. Eachdata point is the average of 100 random cases generated fromthe settings as in Section VI-A. We change the values of µ , µ , and µ , with respect to the reference, in Figs. 5(a),5(b), and 5(c), respectively. In Fig. 5(a), the RD capacityincreases with µ , i.e., decreases with µ . The RU capacitydoes not change as it does not relate to RDQ. In Fig. 5(b),both capacities grow with µ since they both involve RUDQ.In Fig. 5(c), the RU capacity increases with µ while the RDcapacity becomes insensitive. It is because only the former isrelated to RUQ. The trends of the error are similar. On theaverage, the simulated capacities are always smaller than the analytical ones and the discrepancies grow with µ , µ , and µ . A certain part of the error is due to the transient state of thesystem as the analytical results only illustrate the steady stateperformance. The rest comes from the growth of the queuelengths of Ψ , Ψ , and Ψ for M SC because the probabilityof generating large and unqualified service times is higher withsmaller µ , µ , or µ . Therefore, there is a tradeoff betweenthe resultant capacity and error. To construct the system withlarger capacities, we need to tolerate larger errors. In practice,we set the values for µ , µ , and µ based on our requirementsfor the expected capacities and their accuracies. Moreover, fora real parking structure, the characteristics of the arriving EVsmay change with the time of day. Based on some historicaldata, we may divide a day into several periods, where the EVcharacteristics are more or less similar in each period. Thenan appropriate combination of µ , µ , and µ can be set foreach period and the capacities of each period can be deducedaccordingly. VII. C ONCLUSION
With the expected higher penetration of renewable energygeneration in smart grid, the stochastic nature of the re-newables will induce new challenges in matching the actualpower consumption and supply. One measure to enforce powerbalance is through regulation services. Traditional regulationservices are mainly run by power plants and very costly.The increasing social consensus on environmentally friendlytransportation would lead to more reliance on EVs. With theembedded rechargeable batteries in EVs, a fleet of EVs can be-have as a huge energy buffer, absorbing excessive power fromthe smart grid or supplying power to overcome the deficit. Thisimplies that an aggregation of EVs is a practical alternative tosupport the regulation services of smart grid. However, V2Gis a dynamic system. Each EV connects to and disconnectsfrom the system independently. Regulation through V2G canbe realized only if we can capture the aggregate behavior ofthe EVs. However, in general, we cannot directly control theparticipating EVs and make them contribute according to ourrequirements. We can only estimate the collective contributionfrom the EVs while allowing them to behave autonomously. Inthis paper, we model an aggregation of EVs with a queueingnetwork. The structure of the queueing network allows us toestimate the RU and RD capacities separately. The estimatedcapacities can help set up a regulation contract between anaggregator and a grid operator so as to facilitate a newbusiness model for V2G. To make the results analyticallytractable, the EV service durations need to be exponentiallydistributed. We introduce a smart charging mechanism to fulfillthis requirement. This mechanism does not require any specificpatterns for the EVs’ initial SOCs, the duration of stay, andtheir lower and upper SOC thresholds and it can adapt to thecharacteristics of the EVs and make the performance of theactual system follow the analytical model. To summarize, ourcontributions include: 1) proposing a queueing network modelfor V2G regulation services, 2) facilitating various regulationcontracts by separating RU and RD capacities, 3) designinga smart charging mechanism to make the system adaptable to ( k W ) (min) (a) Change in µ (min) ( k W ) (b) Change in µ ( k W ) (min) (c) Change in µ Fig. 5. Simulated capacities and the discrepancies between the analyticaland simulated results. various characteristics of EVs, and 4) performing extensivesimulations to verify the performance of the model.A
CKNOWLEDGMENT
This research is supported in part by the Theme-basedResearch Scheme of the Research Grants Council of HongKong, under Grant No. T23-701/14-N.R
EFERENCES[1] A. Y. S. Lam, K.-C. Leung, and V. O. K. Li, “Capacity management ofvehicle-to-grid system for power regulation services,” in
Proc. of IEEEInt. Conf. on Smart Grid Comm.
Power Systems Analysis , 2nd ed. UpperSaddle River, NJ: Prentice Hall, 2000, ch. 11.[7] H. Bevrani and T. Hiyama,
Intelligent Automatic Generation Control .Boca Raton: CRC Press, 2011.[8] E. Bitar, K. Poolla, P. Khargonekar, R. Rajagopal, P. Varaiya, and F. Wu,“Selling random wind,” in
Proc. 45th Hawaii International Conferenceon Systems Science , Jan. 2012, pp. 1931–1937.[9] T. A. Becker and I. Sidhu, “Electric vehicles in the united states: Anew model with forecasts to 2030,” Center for Entrepreneurship andTechnology, University of California, Berkeley, Tech. Rep. 2009.1.v.2.0,Aug. 2009.[10] C. Silva, M. Ross, and T. Farias, “Evaluation of energy consumption,emissions and cost of plug-in hybrid vehicles,”
Elsevier Energy Conver-sion and Management , vol. 50, no. 7, pp. 1635–1643, Jul. 2009.[11] Y.Chen, M. Keyser, M. H. Tackett, R. Leonard, and J. Gardner,“Incorporating short-term stored energy resource into MISO energyand ancillary service market and development of performance-basedregulation payment,” in
Energy Storage for Smart Grids
J. Power Sources , vol.196, pp. 3972–3980, Apr. 2011.[14] A. Santos, N. McGuckin, H. Y. Nakamoto, D. Gray, and S. Liss,“Summary of travel trends: 2009 national houseehold travel survey,”Federal Highway Administration, U.S. Department of Transportation,Tech. Rep. FHWA-OL-11-022, Jun. 2011.[15] W. Kempton and J. Tomic, “Vehicle-to-grid power fundamentals: Cal-culating capacity and net revenue,”
J. Power Sources , vol. 144, no. 1,pp. 268–279, Jun. 2005.[16] C.-H. Ng and B.-H. Soong,
Queueing Modelling Fundamentals: WithApplications in Communication Networks , 2nd ed. Hoboken, NJ: Wiley,2008.[17] W. Kempton and J. Tomic, “Vehicle-to-grid power implementation: Fromstabilizing the grid to supporting large-scale renewable energy,”
J. PowerSources , vol. 144, no. 1, pp. 280–294, Jun. 2005.[18] A. Y. S. Lam, L. Huang, A. Silva, and W. Saad, “A multi-layer marketfor vehicle-to-grid energy trading in the smart grid,” in
Proc. 1st IEEEINFOCOM Workshop on Green Networking and Smart Grids , Mar.2012.[19] M. Yilmaz and P. T. Krein, “Review of the impact of vehicle-to-gridtechnologies on distribution systems and utility interfaces,”
IEEE Tran.Power Electronics , vol. 28, no. 12, pp. 5673–5689, Dec. 2013.[20] R. Garcia-Valle and J. G. Vlachogiannis, “Letter to the editor: Electricvehicle demand model for load flow studies,”
Electric Power Compo-nents and Systems , vol. 37, no. 5, pp. 577–582, Sep. 2009.[21] U. C. Chukwu and S. M. Mahajan, “V2G parking lot with PV rooftop forcapacity enhancement of a distribution system,”
IEEE Trans. SustainableEnergy , vol. 5, no. 1, pp. 119–127, Jan. 2014.[22] S. J. Baek, D. Kim, S.-J. Oh, and J.-A. Jun, “Modeling of electric vehiclecharging systems in communications enabled smart grids,”
IEICE Trans.on Information and Systems , vol. E94-D, no. 8, pp. 1708–1711, Aug.2011.[23] S. Han, S. Han, and K. Sezaki, “Development of an optimal vehicle-to-grid aggregator for frequency regulation,”
IEEE Trans. Smart Grid ,vol. 1, no. 1, pp. 65–72, Jun. 2010.[24] ——, “Optimal control of the plug-in electric vehicles for V2G fre-quency regulation using quadratic programming,” in
Proc. of IEEE PESInnovative Smart Grid Technologies , Jan. 2011.[25] ——, “Estimation of achievable power capacity from plug-in electricvehicles for V2G frequency regulation: Case studies for market par-ticipation,”
IEEE Trans. Smart Grid , vol. 2, no. 4, pp. 632–641, Dec.2011.[26] K. N. Kumar, B. Sivaneasan, P. H. Cheah, P. L. So, and D. Z. W. Wang,“V2G capacity estimation using dynmaic EV scheduling,”
IEEE Trans.Smart Grid , vol. 5, no. 2, pp. 1051–1060, Mar. 2014.[27] S. Bashash and H. K. Fathy, “Cost-optimal charging of plug-in hybridelectric vehicles under time-varying electricity price signals,”
IEEE Tran.Intelligent Transportation Systems , To appear. [28] J. Lin, K.-C. Leung, and V. O. K. Li, “Online scheduling for vehicle-to-grid regulation service,” in Proc. of IEEE Int. Conf. on Smart GridComm. , Oct. 2013, pp. 43–48.[29] R. A. Verzijlbergh, M. O. W. Grond, Z. Lukszo, J. G. Slootweg,and M. D. Ilic, “Network impacts and cost savings of controlled evcharging,”
IEEE Trans. Smart Grid , vol. 3, no. 3, pp. 1203–1212, Sept.2012.[30] L. Hua, J. Wang, and C. Zhou, “Adaptive electric vehicle chargingcoordination on distribution network,”
IEEE Trans. Smart Grid , vol. 5,no. 6, pp. 2666–2675, Nov. 2014.[31] Y. Cao, S. Tang, C. Li, P. Zhang, Y. Tan, Z. Zhang, and J. Li, “Anoptimized EV charging model considering TOC price and SOC curve,”
IEEE Trans. Smart Grid , vol. 3, no. 1, pp. 388–393, Mar. 2012.[32] Y.-M. Wi, J. uk Lee, and S.-K. Joo, “Electric vehicle charging methodfor smart homes/buildings with a photovoltaic system,”
IEEE Trans.Consum. Electron. , vol. 59, no. 2, pp. 323–328, May 2013.[33] M. Wang, H. Liang, R. Zhang, R. Deng, and X. Shen, “Mobility-awarecoordinated charging for electric vehicles in VANET-enhanced smartgrid,”
IEEE J. Sel. Areas Commun. , vol. 32, no. 7, pp. 1344–1360, Jul.2014.[34] D. Gross, J. F. Shortle, J. M. Thompson, and C. M. Harris,
Fundamentalsof Queueing Theory , 4th ed. Hoboken, NJ: Wiley, 2008.[35] D. P. Bertsekas and R. G. Gallager,
Data Networks , 2nd ed. EnglewoodCliffs, NJ: Prentice Hall, 1992.[36] R.-C. Leou, C.-L. Su, and C. N. Lu, “Stochastic analyses of electricvehicle charging impacts on distribution network,”