ebook img

Coordinated Autonomous Vehicle Parking for Vehicle-to-Grid Services: Formulation and Distributed Algorithm PDF

0.74 MB·
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Coordinated Autonomous Vehicle Parking for Vehicle-to-Grid Services: Formulation and Distributed Algorithm

1 Coordinated Autonomous Vehicle Parking for Vehicle-to-Grid Services: Formulation and Distributed Algorithm Albert Y.S. Lam, James J.Q. Yu, Yunhe Hou, and Victor O.K. Li Abstract—Autonomous vehicles (AVs) will revolutionarize m DurationforAVk toreachtheparkingfacility k ground transport and take a substantial role in the future from its initial location. 7 transportationsystem.MostAVsarelikelytobeelectricvehicles m Duration for AV k to reach the return location k 1 (EVs) and they can participate in the vehicle-to-grid (V2G) from the parking facility. 0 systemtosupportvariousV2Gservices.Althoughitisgenerally 2 infeasibleforEVstodictatetheirroutes,wecandesignAVtravel (cid:15)k Amount of energy required for AV k to reach plans to fulfill certain system-wide objectives. In this paper, we the parking facility from its initial location. n focus on the AVs looking for parking and study how they can (cid:15) Amount of energy required for AV k to reach a k be led to appropriate parking facilities to support V2G services. J the return location from the parking facility. 6 Wcaen fboermsoullvaetde bthyeaCsotoarnddianradteidntPegaerrkilningeParropbrleomgra(mCPsPo)l,vewrhbicuht F Set of parking facilities. requires long computational time. To make it more practical, pf Demand profile of parking facility f. ] we develop a distributed algorithm to address CPP based on t A particular time slot. Y dual decomposition. We carry out a series of simulations to ρf NumberofAVsrequiredtosupporttheservices S evaluate the proposed solution methods. Our results show that t at parking facility f in the tth time slot. . the distributed algorithm can produce nearly optimal solutions s D Latest time slot in the time horizon. c with substantially less computational time. A coarser time scale [ canimprovecomputationaltimebutdegradethesolutionquality cf Capacity of parking facility f. resultinginpossibleinfeasiblesolution.Evenwithcommunication βf Estimation function of parking facility f. 1 loss,thedistributedalgorithmcanstillperformwellandconverge mˆf Duration that AV k should stay at parking v with only little degradation in speed. k facility f. 7 2 IndexTerms—Autonomousvehicle,coordinatedparking,smart e(cid:48)k SOC of AV k when arriving at parking facility 5 city, vehicle-to-grid. f. 1 e(cid:48)(cid:48) SOC of AV k when leaving parking facility f. k 0 T Time horizon. 1. NOMENCLATURE xf Binary variable to indicate if AV k is assigned kt 0 to parking facility f at the time slot t. G Thecompletedirectedgraphmodelingtheroad 7 yf Binary variable to indicate if AV k is parked 1 network. k at parking facility f. : N Set of all possible locations. v M A sufficiently large positive number. E Set of paths connecting the locations. Xi dij Expected travel distance from i to j. λfft Lagrangian multiplier corresponding to ρft. r K Set of autonomous vehicles (AVs). λt Lagrangian multiplier corresponding to cf. a k A particular AV. Λ Vector of Lagrangian multiplier λf. t f n Initial location of AV k. Λ Vector of Lagrangian multiplier λ . k t n Return location of AV k. g(Λ,Λ) Dual function. k t Initial time of AV k available for parking. γf(i) Step size for the update rule of λf at Iteration k t t t End time of AV k for parking. i. k e State of charge (SOC) of AV k right before γf(i) Step size for the update rule of λf at Iteration k t t parking. i. e Expected SOC of AV k on return. xf∗(i) Optimal result by solving Problem 2 at Itera- k kt dmax Maximum distance AV k allowed to travel. tion i. k α Estimation function of AV k. δ A small positive number. k f A particular parking facility. γcap(i) Cap of step size at Iteration i. nˆ Location of parking facility f. γinit Initial value of the step size. f (cid:15) A small positive number. This work was supported by the Theme-Based Research Scheme of the Research Grants Council of Hong Kong under Grant T23-701/14-N. A preliminaryversionofthispaperwaspresentedin[1].Theauthorsarewith I. INTRODUCTION the Department of Electric and Electronic Engineering, The University of HongKong,PokfulamRoad,HongKong(e-mail:[email protected]). 2 THANKS to people’s stronger environmental awareness effectivedistributedalgorithmtoaddresstheproblem;and(4) and various governments’ green policies, increasingly conductingextensivesimulationtoevaluatetheperformanceof moreelectricvehicles(EVs)willrunontheroads.EVslargely the distributed algorithm and to compare with the centralized rely on the grid to charge their batteries. Besides, they can and heuristic approaches proposed in [1]. also discharge any excessive energy back to the grid. The EV The rest of the paper is organized as follows. Section batteries become a significant yet flexible energy repository. II provides the related work. We develop models for the Thisvehicle-to-grid(V2G) system whichcancomplementthe road network, AVs, and parking facilities and illustrate the grid with various demand response and auxiliary services. system operation in Section III. Section IV formulates CPP A V2G system may be considered to be associated with a as an optimization problem, and we develop an effective parking facility where a large number of EVs can contribute distributed algorithm in Section V. In Section VI, we evaluate their batteries to support various V2G services [2]. However, theperformanceofthevarioussolutionmethodsandconclude convenience plays a very important role when an EV driver the paper in Section VII. decides where and when to park its vehicle. EV mobility behavior is considered stochastic [3] and it is hard for a II. RELATEDWORK parking facility to predict accurately how many EVs will be available in a particular period, even in the next few hours. There are many related efforts studying the relationship Autonomous vehicles (AVs), also known as driverless cars between V2G and the supported services. [7] investigated and robotic cars, refer to those vehicles which can navigate how demand response helps reduce peak power demand and without human intervention. Based on the recent trend of shape the V2G aggregated demand profile. [8] studied the the automotive industry, e.g., from Tesla, AVs will become impact of EV mobility on demand response for V2G and prevalent on the roads. It has been predicted that AVs will presented a dynamic complex network model of V2G mobile revolutionize the automobile industry in the next two decades energy networks. In [9], an EV scheduling algorithm was [4], [5], [6]. They are equipped with numerous sensors to developed to optimize bidding of V2G for various ancillary facilitatetheirinteractionswiththesurroundingenvironments. services, including frequency regulation and spinning reserve. An AV may be fully or partially driverless; a driver can guide Itmaximizestheaggregator’sprofitwhileprovidingpeakload themovementinthe“normal”modeanditcanalsoimplement shaving to the utility. [10] formulated the optimal combined self-navigationinthe“autonomous”modewithoutthedriver’s bidding of V2G ancillary services and it can enhance the input. AVs enjoy many advantages over conventional cars, profit of the aggregators, utilities, and EV customers. [11] like avoiding collisions due to human errors, lessening traffic designed a V2G aggregator for frequency regulation and a congestion, and reducing physical space for vehicle parking. dynamicprogrammingalgorithmwasdevelopedtocontrolthe AVs are typically electric and they contain batteries to optimalchargingforthevehicles.[2]estimatedthecapacityof store energy for propulsion. Hence, AVs can participate in V2Gforfrequencyregulationwithaqueueingnetworkmodel. V2G. Due to their self-driving ability and advanced vehic- [12] discussed the economical operation of energy resources, ular communication technologies, AVs can be coordinated like batteries, for microgrid. [13] proposed a distributed EV to orchestrate more co-operative exercises. It is possible to coordination management for efficient exploitation of renew- arrange an appropriate number of AVs with parking intention able energy. All these suggest that V2G may potentially be to the right location to support V2G services. Hence, AVs beneficial to the grid and one of the keys to success is to are considered advantageous over ordinary EVs in the sense ensure the availability of EVs to participate in V2G. that the intrinsic uncontrollable EV behaviors, with respect to There are numerous research projects related to AVs. For theirappearanceatV2Ginfrastructure,cannowbeovercome. instance, [14] designed an obstacle avoidance motion control Moreover, different V2G-supporting parking facilities have schemeforAVsoperatinginuncertaindynamicenvironments. diverse V2G objectives and they have different “demands” [15] developed a hierarchical controller for AVs to track of EVs anchoring at the facilities at different times. We can reference paths in uncertain conditions and with external now deploy more effective V2G services by appropriately disturbances. [16] designed a method to detect obstacles and assigning AVs to the parking facilities to meet their EV dangerous areas in the outdoor environments with Kinect demands. Therefore, in this paper, instead of studying how sensors installed on AVs. [17] studied the collective behavior AVs contribute to V2G in parking facilities directly, we of AV flocking under an all-to-all communication scheme. In investigate how to coordinate AV parking to facilitate V2G [18], AVs co-operated in a public transportation system, in services. To the best of our knowledge, we are the first to which AVs were scheduled with centralized control. Admis- study how to manage AVs for supporting V2G services. We sion control of the system was also fully investigated. [19] formulate the Coordinated Parking Problem (CPP) for AVs focused on the pricing issue of the AV public transportation to support V2G. While a centralized and a heuristic solution system and developed a combinatorial auction-based strategy- havebeenproposedinourpreliminaryversion[1],wedevelop proof pricing scheme. The automotive industry is also devel- a distributed algorithm to make the problem solving scalable opingAVtechnologies.[20]reportedthestate-of-the-artdevel- so that this work can become more practical. Compared with opment in the AV industry and AVs will become connected [1], our contributions include: (1) conducting a more compre- vehicles. The success of AVs will rely on connectivity and hensive literature review; (2) providing a neater formulation cooperation of the vehicles. Google launched the self-driving of the problem with fewer constraints; (3) developing an car project and built a fully functioning prototype without a 3 steering wheel and pedals [21]. A Tesla car can enable its d indicating the expected travel distance from i to j. d is ij ij autonomous driving ability with a software update [22]. Thus in general not equal to d and this accounts for the possible ji AVs are not just idle theorizing and they can have practical asymmetryoftraveldistancesindifferentdirections.Notethat use sooner or later. G(N,E) is not a direct representation of the corresponding There is some work studying intelligent parking in general. road network; in G(N,E), a node is always accessible by [23] investigated availability of parking facilities for parking another node in one hop. We can construct G(N,E) from guidanceandinformationsystems.Itdevelopedamultivariate the road system by specifying a route from i to j with the autoregressive model to account for the temporal and spatial corresponding distance, for each (i,j) pair. For instance, we relationship of parking availability. [24] studied the uncoor- may employ Dijkstra’s algorithm [30] to suggest the shortest dinated parking space allocation for inexpensive limited on- route to connect i to j. We assume d ’s are static at the time ij street parking spots and expensive oversized parking lots. ofassignment.d ’scanberevisedtoreflecttheupdatedtraffic ij Some work focuses on AV parking. [25] developed a control conditions in any subsequent assignments. system for AV valet parking with a focus on steering control. [26] designed an intelligent vehicle system to implement the C. Autonomous Vehicles AV valet parking service. However, they mainly targeted AV We denote the set of AVs which need parking by K. Each parking control in a confined parking area. Some investigate k ∈Kisspecifiedbythetuple(cid:104)n ,n ,t ,t ,e ,e ,dmax,α (cid:105). the parking issue for a larger area. [27] proposed intelligent k k k k k k k k The autonomous parking mode of k is turned on at n ∈ N parking assistant architecture to manage parking spots to k at time t with state of charge (SOC) e and it is expected to improve the quality of urban mobility. [28] analyzed the k k return to n ∈ N by time t (t ≥ t ) with SOC e , which impact of charging and discharging of EVs in parking lots on k k k k k represents the minimum allowable SOC of the battery when thepowergridprobabilistically.However,thereisnothorough the driver uses the car again after parking. n is allowed to study on V2G based on AVs. In this work, we aim to bridge k be different from n for the convenience of the driver. As this research gap. k k is expected to park in one of the parking facilities, the driver may desire to confine the total distance that the AV III. SYSTEMMODEL travels during (t ,t ). The maximum distance that AV k is k k The system is composed of three types of components, allowed to travel in the autonomous mode is indicated by namely, a road network, AVs, and parking facilities. In this dmax.1 If the assigned parking facility f is known, the AV k section, we first describe the required infrastructure and then can estimate the amount of time and energy required to reach provide models of these system components. Finally we illus- f from nk and those required to arrive at nk from f based trate how the system operates. on the relevant details (including its locations, driving speed, and energy consumption rate). We define the function α to k accomplish such estimation as A. Infrastructure [m ,m ,(cid:15) ,(cid:15) ]=α (n ,n ,t ,t ,nˆ ), (1) k k k k k k k k k f Consider that we implement the coordinated AV parking wherenˆ ,m ,andm refertothelocationof f,theduration f k k in a smart city [29], in which the road infrastructure is well- for k to reach nˆ from n and the duration for k to return to f k established. There is full communication coverage, backed by n fromnˆ ,respectively.(cid:15) and(cid:15) aretheamountsofenergy k f k k advanced vehicular communication technologies (e.g., IEEE required to support the first and second legs of the parking 802.11p),supportingvariousintelligenttransportationsystems journey, respectively. Thus the reduced energy for mobility applications. A control center, implemented in the cloud, acts needs has been captured. asthecentral“brain”ofthesystemtomanagethefleetofAVs with parking intention and the parking facilities. The Internet D. Parking Facilities ofThingsbackboneprovidesreal-timecommunicationsupport between the AVs (parking facilities) and the control center. We consider a set of parking facilities F, each of which The control center collects the required information from the represents a V2G system connected to the grid as in [2]. AVs and parking facilities, does the computation, and gives Each f ∈ F is described by the tuple (cid:104)nˆf,pf,cf,βf(cid:105). instructions to the AVs for parking arrangement. A similar pf = [ρft]1≤t≤D denotes the demand profile of f, where ρft infrastructure is also adopted in [18] and [19] to implement givesthenumberofAVsrequiredtosupporttheV2Gservices an AV-based public transportation system. atf inthetthtimeslotandDisthelatesttimeslotinthetime horizon (The time slot operation will be explained in Section III-E). There is much work in the literature describing how to B. Road Network utilize EVs to facilitate different kinds of V2G services, e.g., We describe the accessibility of the AVs to and from the frequency regulation [2]. The basic principle is that, for f to parking facilities with a road network. The road network is provide various V2G services, it needs to acquire a certain modeled by a complete directed graph G(N,E), where N is thesetofallpossiblelocationswheretheAVsandtheparking 1dmkax is not the maximum range supported by the energy stored in the battery of AV k. Instead, it is a value set by the owner who tries to cap facilities are located. E represents the set of paths connecting thedistancetraversedforparking.Thisvalueisgenerallysmallandthusthe the locations. Each (i,j) ∈ E is associated with the distance rangelimitduetoenergysufficiencydoesnotmatter. 4 number of vehicles for charging and discharging. Here we 9:00 9:10 9:20 9:30 9:40 9:50 10:00 t t+1 t+2 t+3 t+4 t+5 model the demand on the vehicles for V2G for the given time V2G and horizon by pf. cf denotes the capacity of f dedicated to the Get parked charging Arrive at the current operation. In other words, it represents the number designated Autonomous Leave the of AVs which f can accommodate in the time horizon. We destination parking mode parking assumethatf iscapableofdetermininghowlongAVkshould on facility park at f. In this parking duration, k will be charged up to a level that at least e will be retained when reaching n , Fig.1. Atimeslotconversionexample. k k with the consideration of an appropriate charging rate and the amount of energy charged or discharged to support V2G. thatallnecessaryinformation,fromboththeAVsandparking Consider that f can facilitate the estimation with the function facilities, is available at t=0 and we will assign the AVs of β based on the SOC specifications of AV k as f K to appropriate parking facilities of F for the period T. mˆf =β (e(cid:48),e(cid:48)(cid:48)), (2) Tofitintothetime-slotimplementationofV2G,withoutloss k f k k ofgenerality,weconvertallthetimeparametersintroducedin wheremˆfk isthedurationthatk shouldstayatf.e(cid:48)k =ek−(cid:15)k SectionsIII-CandIII-D,includingtk,tk,mk,mk,andmˆfk,to and e(cid:48)k(cid:48) = ek + (cid:15)k represent the SOCs of k when arriving the time-slotted format. Consider the scenario given in Fig. at f and when leaving from f, respectively, where (cid:15)k and 1 which depicts the schedule of a particular AV k. We set (cid:15)k are computed from (1). In other words, given the SOC the duration of each time slot to 10 minutes for illustrative requirements of AV k in terms of e(cid:48)k and e(cid:48)k(cid:48), f can manage purposes3 and the time slots start at 9:00, 9:10, 9:20, and so theV2Geventsappliedtok (thismaychargeordischargethe on. The AV is ready to park in time slot t and it arrives at a battery of k) and determine an appropriate charging profile parking facility in t+1. It leaves the parking facility in t+4 for k. When k leaves f, f will ensure k’s SOC reached and returns to its designated destination in t+5. k is only e(cid:48)k(cid:48) by keeping k at f for mfk time slots. In the literature, available for V2G and charging in t+2 and t+3. We can a lot of existing work (e.g., [2], [31], [32], [33], [34]) has simplysett =t+1,t =t+5,andmˆf =2.Ittakesoneslot k k k already investigated the energy management of vehicles and for the first leg and one slot for the second leg of its journey, their interactions with the grid for V2G. In this work, we do i.e., m = m = 1. In this way, we have not only reserved k k notplantoreplicatetheseeffortsandsimplyrepresentallthese sufficient time for k to travel, but k can also be made to fit by βf. For a particular V2G application, we can construct the into the V2G slotted operation. corresponding βf based on the relevant published work. In Afteranassignmentforthetimehorizon{t=0,1,...,D}, this way, we can simplify our model and pay our attention another assignment can be performed after time ∆t>0, i.e., to AV parking arrangement, which is the main theme of this for {t = 0+∆t,1+∆t,...,D+∆t} . If ∆t is larger than paper. D, it is like a fresh restart such that the two assignments have no correlation. If ∆t is smaller than D, it is possible that some AVs are still undergoing the schedules settled in the E. Operation first assignment. We can still consider these AVs in the later Supposethatthereisacontrolcenterwhichco-ordinatesthe assignment such that their parameters are revised to reflect parking of AVs. This control center aims to serve a dedicated their updated statuses accordingly. For example, if AV k is groupofAVs,e.g.,theAVPublicTransportationSystem[18], parking at the parking facility f at t = ∆t, we may simply or to provide a kind of parking service to its subscribed AVs. set its starting location to nˆ , i.e., n = nˆ , for the later Similar to many existing V2G implementations (e.g., [2]), f f f assignment. the system is considered to operate in a time-slot basis. The time horizon is described by time slots {t = 0,1,...,D}. IV. PROBLEMFORMULATION As providing auxiliary services is one of the core functions To facilitate the formulation of the problem, we define two of V2G in which the extent of participation needs to be binary variables xf and yf as follows: committed in advance in the corresponding auxiliary service kt k markets, each parking facility f is supposed to be able to 1 if AV k is assigned to Parking Facility f estimate its demand profile pf = [ρft]1≤t≤D by t = 0. xf = in the time slot t, Moreover, with the advancement of vehicular communication kt 0 otherwise, technologies (e.g., vehicular ad-hoc networks [35]), the gov- erned AVs are all connected and they can predict their travel and plans for the near future. Thus it is possible for the system (cid:40) 1 if AV k is parked at f, to determine the set of AVs with parking intention during the yf = period T ={t=1,...,D} by t=0.2 Therefore, we assume k 0 otherwise. Although xf implies yf, the introduction of yf can make the 2AsAVsaremorepredictable,weassumethattheavailabilitiesofallAVs kt k k formulation simpler. areknowninadvance.Thisisvalidwhenitcomestodedicatedtransportation systems,e.g.,theAVPublicTransportationSystem[18].Moreover,wemay adjust D based on the amount of information about the AVs and parking 3Wewillinvestigatethesystemperformancewithdifferenttimescalesin facilities. SectionVI. 5 There are a number of requirements governing the assign- and mentoftheAVstotheparkingfacilities.First,eachAVshould D beallocatedtoaparkingfacilityforproperparking.Ingeneral, (cid:88)xf =0,∀f ∈F,k ∈{k|t −m ≤D}. (9) kt k k an AV k should not impose unnecessary burden to the traffic and should stay stationary in a parking facility most of the t=tk time from t to t . Hence, we consider that an AV will be In fact, since m is known, we can combine (5) and (8) k k k assigned to one and only one parking facility during its off- resulting in duty period. This can be specified by t −1+m (cid:88)ykf =1,∀k ∈K. (3) k (cid:88) kxfkt =0,∀f ∈F,k ∈K. (10) f∈F t=1 If AV k is assigned to Facility f, it will stay at f for a Similarly, combining (6) and (9) can get sufficient number of time slots for charging and supporting D V2G services. Recall that the parked duration mˆfk depends on (cid:88) xfkt =0,∀f ∈F,k ∈{k|tk−mk ≤D}. (11) itsSOCspecifications,thetraveldistancesbetweenitsspecific locations and f, and the expected utilization of k for V2G by t=tk−mk f.Whenthedetailsofkandf aregiven,bycomputing(1)and To meet the demand from the V2G services, we should (2), mˆf is indeed a constant. We can represent such condition secure enough AVs parked at f based on its demand profile k with the following inequality: pf. It is not uncommon to summarize the grid requirements with a total amount of energy required at each aggregator, D mˆfyf ≤(cid:88)xf ≤Myf,∀k ∈K,f ∈F, (4) e.g., in [9]. We can also represent this amount of energy with k k kt k anumberofvehicles,eachofwhichcontributesequalportion, t=1 e.g., in [2]. Moreover, the number of AVs parked at f should where M is a sufficiently large positive number. not exceed its capacity c . These can be ensured with the f IttakestimeforanAVk totravelfromitsoriginalposition following inequality: n toaparkingfacilityf andreturntoadesignatedlocationn k k after parking. The time periods for these two legs of journey ρf ≤ (cid:88)xf ≤c ,∀f ∈F,t∈T. (12) t kt f are specified by m and m , respectively (see Eq. (1)). If k k k k∈K is parked at f at time t, we should reserve at least m time k slots for k to reach f. In other words, if xf = 1, then there AVs should be parked as long as possible. We can do this are at least m time slots with xf =0, whketre s<t. That is by maximizing the occupancy, i.e., assigning the AVs to the f(cid:80)oltsl−o=w1tki(n1g−inxefkqksu)al≥itym: k. This cankbse satisfied by imposing the pisarekqinugivaflaecnitlititeos minaxaims imzinangy(cid:80)timk∈eK,stl∈oTts,f∈asFxpfkots.s4ibWle.eTchailsl the problem the Coordinated Parking Problem (CPP) and its (cid:88)t−1 formulation is given as follows: (1−xf )≥m xf ,∀k ∈K,f ∈F,t∈T. (5) ks k kt Problem 1 (Coordinated Parking Problem). s=t k Similarly, if k is parked at f at time t, we should reserve at maximize (cid:88) xf kt least m time slots for k to get back to n from f by t for k k k k∈K,t∈T,f∈F (13) xf =1. This is equivalent to: kt subject to (3),(4),(7),(10)–(12). (cid:88)tk (1−xf )≥m xf , This problem is equivalent to the one formulated in [1] but ks k kt with much fewer constraints. This allows the problem to be s=t+1 solvedmoreeffectively.CPPisanintegerlinearprogram(ILP) ∀k ∈{k|t −m ≤D},f ∈F,t∈T. (6) k k and it can be solved by a standard ILP solver. An AV k should be assigned to a facility f such that its total travel distance does not exceed dmax. In other words, if yf = 1, then d +d ≤ dmax.kThis can be further V. DISTRIBUTEDALGORITHM k nkf fnk k described by: As will be shown in Section VI, if we solve CPP in a centralized manner, the computational time required grows (d +d )yf ≤dmax,∀f ∈F,k ∈K. (7) nknˆf nˆfnk k k tremendously with the number of AVs. In order to make Since AV k is available for parking from t to t only, it scalable, we are going to develop a distributed algorithm k k it should not be assigned to any parking facility any time to speed up the computational process. We adopt the dual before t and from t onward. This can be specified with decomposition method [36], which have been widely applied k k the following two equalities: toproblemsinpowersystems(e.g.,[37],[38]),todevelopthe distributed algorithm. t −1 (cid:88)k xf =0,∀f ∈F,k ∈K (8) kt 4If economic cost of energy needs to be explicitly considered, we can t=1 simplyreplacetheobjectivefunctionwiththerelatedcostfunction. 6 Based on Problem 1, we first relax Constraint (12) by dual variables Λ and Λ by addressing the dual problem: introducing Lagrangian multipliers λft and λft and construct minimize (cid:88)g (Λ,Λ)+ (cid:88) (λfc −λfρf) (16a) the partial Lagrangian as follows: k t f t t k∈K t∈T,f∈F (cid:32) (cid:33) subject to Λ,Λ≥0, (16b) (cid:88) xf − (cid:88) λf (cid:88)xf −c kt t kt f which is linear. We can solve the dual problem to recover k∈K,t∈T,f∈F t∈T,f∈F k∈K (cid:32) (cid:33) the solution of the original Problem 1. We have the gradients − (cid:88) λft −(cid:88)xfkt+ρft ∂gk∂(λΛft,Λ) =cf−(cid:80)k∈Kxfkt(k)and ∂gk∂(λΛft,Λ) =(cid:80)k∈Kxfkt(k)− t∈T,f∈F k∈K ρf. By projected gradient descent [39], we can generate a t = (cid:88) xf − (cid:88) λfxf sequence of feasible points {Λ(i),Λ(i)} with the following kt t kt update rules: k∈K,t∈T,f∈F k∈K,t∈T,f∈F + (cid:88) λfxf + (cid:88) λfc − (cid:88) λfρf (cid:34) (cid:32) (cid:33)(cid:35)+ t kt t f t t λf(i+1)= λf(i)−γf(i) c −(cid:88)xf∗(i) , k∈K,t∈T,f∈F t∈T,f∈F t∈T,f∈F t t t f kt = (cid:88) (xf −λfxf +λfxf ) k∈K kt t kt t kt ∀t∈T,f ∈F, (17) k∈K,t∈T,f∈F (cid:34) (cid:32) (cid:33)(cid:35)+ + (cid:88) (λftcf −λftρft). λf(i+1)= λf(i)−γf(i) (cid:88)xf∗(i)−ρf , t t t kt t t∈T,f∈F k∈K ∀t∈T,f ∈F, (18) Clearly the rest of the constraints, i.e., (3), (4), (7), (10)–(11), are all separable with respect to k. For each k, we where xf∗(i) is the optimal result by solving Problem 2 at kt representthevariablesandfeasibleregionconfinedby(3),(4), Iteration i while γf(i) > 0 and γf(i) > 0 are the step sizes L(7e)t,(Λ10=)–{(1λ1f)}asσk ={axnfkdt,Λykf=}t∈{λTf,f}∈F andZ.kT,hreusspethcetivdeulayl. at Iteration i. If wfte have (cid:80)k∈Kfxtfkt∗(i) > cf violating (12), t t∈T,f∈F t t∈T,f∈F (17) will make λ (i + 1) > λ (i). Solving Problem 2 at function g(Λ,Λ) of Problem 1 becomes Iteration i+1 tentds to make xft∗(i+1) smaller. Similarly, kt   if we have (cid:80) xf∗(i)<ρf violating (12), (18) will make k∈K kt t g(Λ,Λ)=(cid:88) sup  (cid:88) (xf −λfxf +λfxf ) λft(i + 1) > λft(i). Solving Problem 2 at Iteration i + 1 k∈Kσk∈Zkt∈T,f∈F kt t kt t kt  tends to make xfkt∗(i+1) larger. We can interpret (Λ,Λ) as f + (cid:88) (λftcf −λftρft), (14) aλfseatreofthsehapdroicwe porficreesntifnogr tahepaprakriknigngspraecseouarncdes:thλetpraincde t t∈T,f∈F of selling V2G services at parking facility f in time slot t, respectively.5 On one hand, if the number of required parking which is convex because of the pointwise supremum of affine slots is larger than the capacity (i.e., (cid:80) xf∗ > c ), the functions of (Λ,Λ). We can also see that the first summation k∈K kt f f of (14) clearly decouples with respect to k. Given (Λ,Λ), we parking space selling price (i.e., λt) will increase and this define the subproblem for each k ∈K as follows: maylowerthetotaldemand(cid:80)k∈Kxfkt∗(i).Ontheotherhand, if the number of AVs contributing to V2G is smaller than the Problem 2 (Subproblem for AV k). energy profile (i.e., (cid:80) xf∗ <ρf), the V2G service charge k∈K kt t λf willincreaseandthisencouragesmoreAVstoparkatf in maximize (cid:88) (xf −λfxf +λfxf ) (15a) timt eslott.Asawhole,thedualproblemisusedtocontrolthe kt t kt t kt t∈T,f∈F shadow prices and each AV adjusts its own parking strategy subject to(cid:88)yf =1 (15b) withthesubproblembasedontheparkingfeesΛandtheV2G k service charges Λ. f∈F Suppose there is a control center which manages the whole D mˆfyf ≤(cid:88)xf ≤Myf,∀f ∈F, (15c) system. Fig. 2 depicts how to implement the distributed k k kt k algorithm.Intermsofcomputation,thecontrolcenterupdates t=1 (d +d )yf ≤dmax,∀f ∈F, (15d) (Λ,Λ) with (17) and (18) while each AV solves its own nknˆf nˆfnk k k suproblem, i.e., Problem 2. After updating (Λ,Λ), the control t −1+m k (cid:88) k center distributes (Λ,Λ) to the AVs. Similarly, after solving xf =0,∀f ∈F, (15e) kt thesubproblem,eachAVpassesitsoptimalxf∗ tothecontrol t=1 kt center. The overall distributed algorithm is illustrated by D (cid:88) xf =0,∀f ∈F. (15f) Algorithm 1. kt We first initialize (Λ,Λ) with appropriate non-negative t=tk−mk values at the control center (Step 1). Then the algorithm For those k with t −m > D, (15f) can be ignored. Let k k 5Notethattheshadowpricesservetoprovideanotherwaytointerpret(17) gk(Λ,Λ) be the optimal value of (15) for k. We update the and(18)economicallyonly. 7 and xf = 0,∀t ∈ [t +m ,t −m ), where k∗ is the Compute (18) k∗t k∗ k∗ k∗ k∗ & (19) for each selected AV. The parking capacity constraint is handled in a f and t similarmanner.The(t,f)pairwiththelargestAVoverflowis identified by calculating (cid:80) xf −c . A list of “free” AVs k∈K kt f tf,tf tf,tf tf,tf that can be removed without violating the respective energy profile constraint is developed. Then the AV with shortest possible stay is removed from f. A feasible primal solution xf* xf* xf* kt kt kt is generated when ρf −(cid:80) xf ≤ 0,∀t ∈ T,k ∈ K and ... ... t k∈K kt c −(cid:80) xf ≥0, for all t∈T,k ∈K. f k∈K kt Solve sub- Solve sub- Solve sub- Algorithm 1 requires the minimum amount of information problem (16) problem (16) problem (16) exchange. In each iteration, after receiving the pricing signals fromthecontrolcenter,eachAVaddressesitsownsubproblem Fig.2. Implementationofthedistributedalgorithm. with AV-specific parameters (including mˆf, d , d , k nknˆf nˆfnk and dmax) only. After receving the AVs’ preferences on the Algorithm 1 Distributed Algorithm k parking facility assignments (in terms of xf∗), the control kt 1. Initialize Λ and Λ center updates the shadow prices with the parking facility- 2. while stopping criteria not matched do specificparameters(c andρf)only.Inapracticalsystem,the 3. for each AV k (in parallel) do f t numberofAVsshouldbefarmorethanthenumberofparking 4. Given Λ and Λ, solve (15) 5. Return xf∗,∀t∈T,f ∈F facilities. Asking each vehicle to handle its own subproblem kt 6. end for with their own parameters avoids gathering many scattered 7. for each t∈T,f ∈F (in parallel) do vehicular data, which make the method highly practical. 8. Given xf∗,∀t∈T,f ∈F, update each λf and λf with kt t t (17) and (18), respectively VI. PERFORMANCEEVALUATION 9. Distribute Λ and Λ to the AVs 10. end for We have developed three methods to solve CPP, namely 11. end while (I) centralized, (II) heuristic, and (III) distributed approaches. With Method I, we directly apply a standard ILP solver to Problem 1 and we adopt Gurobi [40] here. Method II is illustrated in [1] while Method III has been introduced in iterates until a stopping criterion has been satisfied (Steps 2– Section V. 11). Each iteration is divided into two parts. The first part We perform four tests to evaluate the performance of the (Steps 3–6) corresponds to solving the subproblems. After solution methods, with emphasis on Method III. In the first receiving (Λ,Λ), each AV solves (15) in parallel and returns the computed xf∗’s to the control center (Step 5). The second test, we assess their performance on different scales of the kt problem with different numbers of AVs and parking facilities. part (Steps 7-10) is for updating (Λ,Λ). After collecting the xf∗’s for particular t and f, the control center can compute The second test aims to investigate the effect of time scaling thketcorresponding λf and λf (Step 8). Hence updating (Λ,Λ) while the third test examines the convergence of Method t t III. In the fourth test, we study the performance of the canbedoneinparallel.Theresultant(Λ,Λ)isthendistributed distributed algorithm in the presence of communication loss. to the AVs (Step 9). Suppose g (i) is the optimal value of the k Wegeneraterandomcasesfortesting.Unlessstatedotherwise, subproblemforAVk atiterationi.Weconsiderthealgorithm we assume that there are 100 time slots (i.e., D = 100) converged if evenly distributed in a horizon of two hours. Consider a |(cid:80) g (i+1)−(cid:80) g (i)| residentialareaof5×5km2,withinwhichwerandomlyplace k k k k <δ,∀k ∈K, (19) |(cid:80) g (i+1)| required numbers of AVs and parking facilities by specifying k k n , n , and nˆ accordingly. Suppose that the AVs travel at where δ is a small positive value, e.g., 10−5. akconkstant spefed of 30 km/h. For AV k, the travel times The primal solution of the original CPP can be retained spent on the two legs for parking, i.e., m and m , are k k by the solutions of the subproblems collectively. We can assigned based on the corresponding distances. We also set recover the primal solution from the dual as follows: We mˆf = rand(1,t −m −(t +m )). These capture α and k k k k k k first determine the (t,f) pair which has the largest AV β . We specify t and t by t = rand(0,D −m −m ) deficit, i.e., ρft − (cid:80)k∈Kxfkt. Then we construct a list of anfd tk = rand(0k,D −mkk −mkk)+tk +mk +mkk, whekre “free” AVs which can be moved to f at t. Each AV k(cid:48) rand(·,·) produces an integer uniformly distributed between in the list can be removed from its original parking facility thetwoinputsinclusively.Ift >D,thenAVk willnotneed k f(cid:48) without violating the respective energy profile constraint to return to n during the time horizon. dmax is randomly ρft(cid:48) −(cid:80)k∈K\k(cid:48)xfkt(cid:48) ≤ 0,∀t ∈ {t|xfk(cid:48)(cid:48)t = 1}, and the AV must set in the rangke of [4,5] km. Finally, the eknergy profile of be able to park in f at t, i.e., tk+mk ≤t<tk−mk. Among Parking Facility f is set as ρft = rand(0,aft/|F|),∀t ∈ T, those AVs in the list, the one with longest possible stay in where af is the number of AVs that are available to park in t f, calculated by tk−mk−(tk+mk), is selected to park in f at t. The parking capacity cf is set to |K|/2 for all f. This f from t + m to t − m − 1. Thus xf(cid:48) = 0,∀t ∈ T allows us to generate feasible instances more easily to inspect k k k k k∗t 8 (I) Value (II) Value (III) Value (I) Value (II) Value (III) Value (I) Time (II) Time (III) Time (I) Time (II) Time (III) Time 1.005 1.005 1.000 (3857)(7758)(11630)(15669)(19446) (27284) (38813) 103 e (s) 1.000(38800)(38737) (38979) (38715) (39035) (38788) 103 e (s) m m 0.995 Ti 0.995 Ti 102 al al Optimality00..998950 101 omputation Optimality00..998950 102 omputation C C 0.980 e 0.980 e g g a a 0.975 100 ver 0.975 ver A A 0.970 0.970 100 200 300 400 500 600 700 800 900 1000 5 10 15 20 25 30 35 40 45 50 Number of Autonomous Vehicles Number of Parking Facilities Fig. 3. Objective function values and computation times with different Fig. 4. Objective function values and computation times with different numbersofAVs. numbersofparkingfacilities. thecomputationalabilitiesofthemethods.Allsimulationsare optimality, they can produce better solutions in those cases performedonacomputerwithIntelCore-i5CPUat2.90GHz with more AVs. The reason is that more AVs provide larger with 8 GB RAM. The simulations are coded with Python on flexibility and it is easier for the algorithms to generate better Linux. solutions.Allmethodsneedmorecomputationaltimewhenthe number of AVs grows. Method I is the most time demanding A. Implementation of the Distributed Algorithm while Method III needs the shortest amount. Therefore, if the true optimal is needed, we will go for Method I, but AsMethodIIIisimplementeddistributedlyineachiteration its computational time grows significantly with problem size. (see Algorithm 1), the subproblem which takes the longest MethodIIIisveryeffectiveinproducinghighqualitysolutions time contributes the time needed for the first part of an and suitable for practical situations. iteration while the update of the λf and λf which needs t t We further consider different numbers of parking facilities the longest time contributes the second part. As only small with a fixed number of AVs. We fix the number of AVs messages containing xf or (Λ,Λ) need to be passed among kt to 1000 and consider cases of 5, 10, 20, 30, 40, and 50 the entities, the communication delay should be small. Based parkingfacilities.Similarly,Fig.4givestherelativecomputed on the average latency of practical cellular systems [41], we objective function values with respect to the optimal and assume each iteration takes 200 ms of communication delay. computational times, where each data point represents the We set δ in (19) to 10−5. For all f and t, we initialize averageof25cases.Theactualoptimalvaluesarealsoshown λf, λf, γf, γf with 0, 0, 0.01, and 0.01, respectively. In a t t t t in brackets for reference. We can see that both the objective subsequent iteration i, we get γf(i + 1) = γf(i) × 1.1 if function value and computational time are not very sensitive t t (cid:80)kgk(i)−(cid:80)kgk(i−1)<0.Otherwise,γft(i+1)=γft(i)× to the number of parking facilities. Thanks to the fact that 0.1.Inaddition,weintroduceγcap(i)tocapthestepsizesuch the occupancy of a vehicle at a parking facility in a time slot that γf(i) ≤ γcap(i). We set γcap(i) = γinit(1−(cid:15))i, where has no difference from any other in the objective function, as t (cid:15) = 10−3 and γinit = 0.01. This satisfies the nonsummable longastheparkingfacilitiesaresufficienttoaccommodatethe diminishing step size rule which guarantees the convergence AVs, more parking facilities available will not help improve of the algorithm [36]. We modify γf similarly. the objective function value. We can understand the trend of t computational time in a similar way. B. Test 1: Different Scales of the Problem C. Test 2: Time Scaling We first examine different numbers of AVs with a fixed number of parking facilities. We consider a setting for a Here we investigate the impact of time scaling. Recall that small neighborhood, where there are five parking facilities. a given time horizon is divided into slots and we can make We generate random cases of 100, 200, 300, 400, 500, 700, the division finer with more time slots for the same period. and 1000 AVs. Fig. 3 depicts the relative objective function We generate 10 random cases for the same horizon of two values(i.e.,occupancies)withrespecttotheoptimalandcom- hours. For each case, we divide the horizon into 10, 20, 30, putational times (in log scale) obtained by the three methods. 40, 50, 80, and 100 time slots with the same settings of 100 The actual optimal values are also shown in brackets for AVs and 5 parking facilities. In other words, we are solving reference. Each point in the figure corresponds to the average the same problem instances with different time scales only. resultsfrom25cases.ItcanbeobservedthatMethodIalways For example, the 51st time slot in the 100-scale corresponds produces the optimal solutions while Method II can generate to the 26th and 5th in the 50- and 10-scale, respectively. Since sub-optimal solutions which are about 97% from the optimal. time scaling is intrinsic to the problem, we demonstrate its MethodIIIisalittleinferiortoMethodIbutmuchbetterthan effects on the optimality and thus we show the results here MethodII.AlthoughbothMethodsIIandIIIcannotguarantee with Method I only. Table I illustrates the percentage (%) 9 TABLEI 80 EFFECTSOFTIMESCALINGON%OPTIMALITYANDCOMPUTATIONAL 0% TIME. 70 10% 60 20% Case 10 20 30 40 50 80 100 e50 30% I N/A 94.27 95.32 97.14 98.83 97.54 100.00 nc 40% e II N/A N/A N/A 96.97 98.67 98.10 100.00 urr40 60% c III N/A 95.54 95.98 96.87 98.73 97.47 100.00 Oc30 80% IV N/A N/A 96.72 98.01 99.79 97.74 100.00 20 V N/A N/A 96.04 97.26 99.00 97.85 100.00 VI N/A N/A 95.29 97.39 99.55 97.62 100.00 10 VII N/A N/A 96.73 96.75 98.66 98.21 100.00 0 1-10 11-20 21-30 31-40 41-50 51-60 VIII N/A 94.72 95.28 96.19 98.93 97.32 100.00 Iterations to converge IX N/A 94.49 95.62 96.65 98.46 97.82 100.00 X N/A 94.94 95.83 96.83 98.65 97.84 100.00 Fig.6. Iterationsneededforconvergencewithdifferentamountofcommu- Avg. nicationloss. N/A 1.41 2.08 2.43 3.17 5.30 9.73 time(s) TABLEII MAXIMUMNUMBEROFITERATIONSREQUIREDTOCONVERGE. 100 AVs Dual 500 AVs Dual %Loss 0 10 20 30 40 60 80 100 AVs Primary 500 AVs Primary 3580 32000 No.ofiterations 50 54 46 56 55 45 49 Vs) Vs) 0 A3570 310000 A 0 0 1 5 e (3560 30000e ( E. Test 4: Communication Loss u u al al n V3550 29000n V Here we evaluate the performance of the distributed al- o o ncti ncti gorithm with the presence of communication loss. Recall e Fu3540 28000e Fu that the algorithm relies on message passing to drive its v v ecti3530 27000ecti convergence. Messages are passed around different entities bj bj O O in a communication network in the form of data packets. 3520 26000 0 5 10 15 20 25 30 However, some packets may be lost during the transmission. Iteration We define p as the probability of having a packet drop. When Fig.5. Primalanddualconvergenceofthedistributedalgorithm. experiencing a packet drop, the involved entity uses the most recently received xf or (Λ,Λ) to do the calculation. We kt consider p equal to 0%, 10%, 20%, 30%, 40%, 60%, and optimality of the different scales with respect to the 100-scale 80%, and for each of which we produce 100 random cases. for the 10 cases. The 100-scale is the finest and gives the best Fig. 6 illustrates the occurrence of the 100 cases for each resultsintermofquality.Whenscalingdown,the%optimality p with respect to the number of iterations required for the drops slightly because the flexibility of assignment decreases. algorithmtoconvergeandTableIIindicatesthecorresponding However, too coarse scaling (e.g., 10-scale) can result in maximum number of iterations among the 100 cases for infeasible solutions. Table I also shows the computational each p. While 40% communication loss results in slightly times averaged over the feasible cases. This suggests that slower convergence, severer communication loss does not scaling-downcanimprovethecomputationaltimesignificantly make significant degradation in performance in general. due to the reduced problem size. Therefore, there exists a VII. CONCLUSION tradeoff between solution quality and computational time. AVswillrepresentasubstantialshareofgroundtransportin the near future. When parked, AVs can participate in V2G as EVs do. The difference is that AVs can be instructed to travel D. Test 3: Convergence of the Distributed Algorithm basedonsomesystem-wideobjectives.Inthispaper,westudy In this test, we study the convergence of the distributed how to coordinate AVs intending to park, to reach parking algorithm.Forillustrativepurposes,weexaminetworepresen- facilities for supporting V2G services. We formulate CPP in tative cases of 100 and 500 AVs with five parking facilities, the form of ILP. Besides solving it in a centralized manner respectively, from Test 1, both of which are accommodated by a standard ILP solver, we propose a distributed algorithm byfiveparkingfacilities.Recallthatthealgorithmmanipulates to overcome efficiency issue of the centralized approach. CPP thedualsolutionsandwecanrecoverthecorrespondingprimal is broken into a number of subproblems, each of which is solutionswiththemethoddiscussedinSectionV.Fig.5shows addressed by an AV, and the convergence of the algorithm the objective function values of the corresponding primal and is controlled by updating the shadow prices at the control dualsolutionsindifferentiterations.Forbothcases,theduality center. Simulations reveal that the distributed algorithm can gaps diminish when the algorithm iterates. It converges faster produce nearly optimal solutions with substantially reduced in the larger case and this is consistent with the results given computational time. A coarser time scale can improve com- in Fig. 3. putational time but degrades the solution quality resulting in 10 possible infeasible solution. Even with communication loss, [21] L. Gannes. (2014, May) Googles new self-driving car ditches the distributed algorithm can still perform well and converge the steering wheel. [Online]. Available: http://recode.net/2014/05/27/ googles-new-self-driving-car-ditches-the-steering-wheel/ with little degradation in speed. [22] R. Bradley. (2014, Oct.) The electric-vehicle maker sent its For future work, we will extend our deterministic formula- cars a software update that suddenly made autonomous driving tion to stochastic or robust optimization. Since some param- a reality. [Online]. Available: https://www.technologyreview.com/s/ 600772/10-breakthrough-technologies-2016-tesla-autopilot/ eters, like travel times, demand profiles, and vehicle return [23] E. Kokolaki, M. Karaliopoulos, and I. Stavrakakis, “Leveraging infor- locations, can become probabilistic in practice, formulating mationinparkingassistancesystems,”IEEETrans.Intell.Transp.Syst., the problem in a probabilistic form may make this line of vol.16,no.5,pp.2913–2924,Oct.2015. [24] T. Rajabioun and P. A. Ioannou, “On-street and off-street parking research more realistic. availabilitypredictionusingmultivariatespatiotemporalmodels,”IEEE Trans.Veh.Technol.,vol.62,no.9,pp.4309–4317,Nov.2013. [25] K. Min and J. Choi, “A control system for autonomous vehicle valet REFERENCES parking,” in Proc. 13th Int. Conf. Control, Automation and Syst., Gwangju,Korea,2013. [1] A. Y. S. Lam, J. J. Q. Yu, Y. Hou, and V. O. K. Li, “Coordinated [26] K.-W.MinandJ.-D.Choi,“Designandimplementationofanintelligent autonomous vehicle parking for vehicle-to-grid services,” in Proc. 7th vehicle system for autonomous valet parking service,” in Proc. 10th IEEEInt.Conf.onSmartGridCommun.,Sydney,Australia,2016. AsianControlConf.,KotaKinabalu,Malaysia,2015. [2] A. Y. S. Lam, K.-C. Leung, and V. O. K. Li, “Capacity estimation [27] R. E. Barone, T. Giuffre, S. M. Siniscalchi, M. A. Morgano, and for vehicle-to-grid frequency regulation services with smart charging G. Tesoriere, “Architecture for parking management in smart cities,” mechanism,”IEEETrans.SmartGrid,vol.7,no.1,pp.156–166,Jan. IETIntell.Transp.Syst.,vol.8,no.5,pp.445–452,2014. 2016. [28] S.Rezaee,E.Farjah,andB.Khorramdel,“Probabilisticanalysisofplug- [3] D.Dallinger,D.Krampe,andM.Wietschel,“Vehicle-to-gridregulation inelectricvehiclesimpactonelectricalgridthroughhomesandparking reserves based on a dynamic simulation of mobility behavior,” IEEE lots,” IEEE Trans. Sustainable Energy, vol. 4, no. 4, pp. 1024–1033, Trans.SmartGrid,vol.2,no.2,pp.302–313,Jun.2011. Oct.2013. [4] D.Mohr,H.-W.Kaas,P.Gao,D.Wee,andT.Mo¨ller,“Automotiverev- [29] S.P.Mohanty,U.Choppali,andE.Kougianos,“Everythingyouwanted olution-persepectivetowards2030:Howtheconvergenceofdisruptive to know about smart cities: The internet of things is the backbone,” technology-driventrendscouldtransformtheautoindustry,”McKinsey IEEEConsum.Electron.Mag.,vol.5,no.3,pp.60–70,Jul.2016. &Company,Tech.Rep.,Jan.2016. [30] T.H.Cormen,C.E.Leiserson,R.L.Rivest,andC.Stein,Introduction [5] T. Litman, “Autonomous vehicle implementation predictions implica- toAlgorithms,2nded. Cambridge,MA:MITPress,2001. tions for transport planning,” Victoria Transport Policy Institute, Tech. [31] E.SortommeandM.A.El-Sharkawi,“Optimalchargingstrategiesfor Rep.,Sept.2016. unidirectional vehicle-to-grid,” IEEE Trans. Smart Grid, vol. 2, no. 1, [6] M. Westervelt, E. Han, D. Gopalakrishna, and J. Klion, “Emerging pp.131–138,Mar.2011. technologytrendsintransportation,”EnoCenterforTransportationand [32] M.Singh,K.Thirugnanam,P.Kumar,andI.Kar,“Real-timecoordina- ICFInternational,Tech.Rep.,Feb.2016. tionofelectricvehiclestosupportthegridatthedistributionsubstation [7] F.Rassaei,W.-S.Soh,andK.-C.Chua,“Demandresponseforresidential level,”IEEESyst.J.,vol.9,no.3,pp.1000–1010,Sept.2015. electricvehicleswithrandomusagepatternsinsmartgrids,”IEEETrans. [33] M. J. E. Alam, K. M. Muttaqi, and D. Sutanto, “Effective utilization SustainableEnergy,vol.6,no.4,pp.1367–1376,Oct.2015. of available PEV battery capacity for mitigation of solar PV impact [8] R.Yu,W.Zhong,S.Xie,C.Yuen,S.Gjessing,andY.Zhang,“Balancing andgridsupportwithintegratedV2Gfunctionality,”IEEETrans.Smart power demand through EV mobility in vehicle-to-grid mobile energy Grid,vol.7,no.3,pp.1562–1571,May2016. networks,” IEEE Trans. Ind. Informat., vol. 12, no. 1, pp. 79–90, Feb. [34] K.Kaur,R.Rana,N.Kumar,M.Singh,andS.Mishra,“Acoloredpetri 2016. net based frequency support scheme using fleet of electric vehicles in [9] E.SortommeandM.A.El-Sharkawi,“Optimalschedulingofvehicle- smart grid environment,” IEEE Trans. Power Syst., vol. 31, no. 6, pp. to-gridenergyandancillaryservices,”IEEETrans.SmartGrid,vol.3, 4638–4649,Nov.2016. no.1,pp.351–359,Mar.2012. [35] D. N. Cottingham, “Vehicular wireless communication,” University of [10] ——,“Optimalcombinedbiddingofvehicle-to-gridancillaryservices,” Cambridge,Tech.Rep.UCAM-CL-TR-741,Jan.2009. IEEETrans.SmartGrid,vol.3,no.1,pp.70–79,Mar.2012. [36] D. P. Bertsekas, Nonlinear Programming. Belmont, MA: Athena [11] S. Han, S. Han, and K. Sezaki, “Development of an optimal vehicle- Scientific,1999. to-grid aggregator for frequency regulation,” IEEE Trans. Smart Grid, [37] A.Y.S.Lam,B.Zhang,andD.Tse,“Distributedalgorithmsforoptimal vol.1,no.1,pp.65–72,Jun.2010. powerflowproblem,”inProc.51stIEEEConf.onDecisionandControl, [12] Z.Yang,R.Wu,J.Yang,K.Long,andP.You,“Economicaloperation Maui,HI,2012. of microgrid with various devices via distributed optimization,” IEEE [38] B.Zhang,A.Y.S.Lam,A.Dominguez-Garcia,andD.Tse,“Anoptimal Trans.SmartGrid,vol.7,no.2,pp.857–867,Mar.2016. and distributed method for voltage regulation in power distribution [13] E. L. Karfopoulos and N. D. Hatziargyriou, “Distributed coordination systems,”IEEETrans.PowerSyst.,vol.30,no.4,pp.1714–1726,Jun. of electric vehicles providing v2g services,” IEEE Trans. Power Syst., 2015. vol.31,no.1,pp.329–338,Jan.2016. [39] S.BoydandL.Vandenberghe,ConvexOptimization. Cambridge,NY: [14] G. Franze and W. Lucia, “A receding horizon control strategy for CambridgeUniversityPress,2004. autonomous vehicles in dynamic environments,” IEEE Trans. Control [40] Gurobioptimization.[Online].Available:http://www.gurobi.com/ Syst.Technol.,vol.24,no.2,pp.695–702,Mar.2016. [41] M. Dano. (2016, Nov.) 3G/4G wireless network latency: [15] C. Chen, Y. Jia, M. Shu, and Y. Wang, “Hierarchical adaptive path- Comparing Verizon, AT&T, Sprint and T-Mobile in February tracking control for autonomous vehicles,” IEEE Trans. Intell. Transp. 2014.[Online].Available:http://www.fiercewireless.com/special-report/ Syst.,vol.16,no.5,pp.2900–2912,Oct.2015. 3g-4g-wireless-network-latency-comparing-verizon-at-t-sprint-and-t-mobile-february [16] J. Hernndez-Aceituno, R. Arnay, J. Toledo, and L. Acosta, “Using kinectonanautonomousvehicleforoutdoorsobstacledetection,”IEEE SensorsJ.,vol.16,no.10,pp.3603–3610,May2016. [17] T.-T. Han and S. S. Ge, “Styled-velocity flocking of autonomous vehicles: A systematic design,” IEEE Trans. Autom. Control, vol. 60, no.8,pp.2015–2030,Aug.2015. [18] A. Y. S. Lam, Y.-W. Leung, and X. Chu, “Autonomous vehicle public transportationsystem:Schedulingandadmissioncontrol,”IEEETrans. Intell.Transp.Syst.,vol.17,no.5,pp.1210–1226,May2016. [19] A. Y. S. Lam, “Combinatorial auction-based pricing for multi-tenant autonomous vehicle public transportation system,” IEEE Trans. Intell. Transp.Syst.,vol.17,no.3,pp.859–869,Mar.2016. [20] E.Uhlemann,“Autonomousvehiclesareconnecting...”IEEEVeh.Tech- nol.Mag.,vol.10,no.2,pp.22–25,Jun.2015.

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.