Carsharingthroughthedataanalysislens

ChiaraBoldrini*,RaffaeleBruno,andMohamedHaitam LaarabiIIT-CNR,Italy

Carsharing• ShareduseofafleetofcarsbyCSmembers• Obstaclesforcoordinationremovedbytechnology

• CSexplosioninthelast15years• 3maincarsharingmodes

• Two-way:tripsstartandendatthesameCSstation• One-way:tripsmayendatanyCSstation• Freefloating:on-streetparkinganywherewithinthegeofence

1891

Invention of taximeter

1916 1948 1970

Joe Saunder decided to lend out his Ford Model T to local and visiting businessmen

Sefage started its service in Zürich

Witkar(technology-based) in Amsterdam

2000

Zipcar

2008

Car2go(free-floating)

explosion2011

AutolibDriveNow

2013

Enjoy

Motivationbehindthiswork

• OpenproblemsinCSresearch• Vehicleredistribution• Cleaningandmaintenance• Infrastructureplanning

• Carsharingisaweaksignalinthecitylandscape• thefractionofpeoplerelyingoncarsharingfortheirdailytripsisrapidlyincreasing

butitisstillintheorderofsingledigitpercentagepointsinthebestcases.

• Carsharinghasbeenmostlystudiedthroughsurveys anddirectinterviewswithitsmembers.

• Carsharingistypicallynotaccountedforinhouseholdstraveldiariesperiodicallycollectedbycityadministrations.

• Candataminingofferhelpfulinsights?

Thedataset

• AvailabilityovertimeofCSvehicles in10Europeancitiesforoneofthemajorfree-floating carsharingoperators.

• Observationperiod:• May17,2015andJune30,2015(for9cities)• March11,2016toMay12,2016.

• Datacollectedevery1minute usingtheavailablepublicAPI,whichyieldsresponsesintheformofJSONfiles.

• Datacleaning:• technicalproblemsonthebookingwebsite->corrupted

entriesdiscarded• faultyGPSsystems->coordinatesthataremanifestlyinvalid

(e.g.,carsavailableindifferentcountries)havebeendiscarded.

• DatapreprocessingandanalysishasbeencarriedoutinR.

Electricvehiclesonly

Datasetlimitations

• Movementsareinferred fromcarsdisappearingontheCSmap• MovementfromAtoB=acardisappearsfromlocationAtolaterreappearatlocationB

• Noexplicitwayfordistinguishingbetweenregularcustomertripsandmaintenancetrips

• Nodirectinformationaboutthetrajectory followedbythesharedvehicle

• WehavequeriedGoogleMapsaskingfordirectionsandexpectedtraveltimebetweenthesourceanddestinationcoordinatesofeachtrip

• Estimatedthetravelleddistanceusingreal-lifeaverageconsumptionextractedfromhttps://www.spritmonitor.de/en/

<vehicle_id, GPS_coords, engine_type, fuel_level, interior_exterior_state>

Themodeshareinthe10cities

• 3classesofcities:oneinwhichmotorised modesdominate,oneinwhichpublictransport(andhencewalking)aremoreimportant,andoneinwhichpeoplemoveprevalentlybybike

Source:Eurostat’sCityUrbanAuditdatabase

Vehicleutilizationrate

• thenumberofdailytrips pervehicle• indexthatisoftenusedasameasureofcarsharingsuccess,asitcapturesshortandfrequenttrips

Service was in fact shutdown

DoesmodalsplitcorrelatewithsuccessfulCS?

• Bike(Pearsonr=−0.34)

• Publictransport(r=0.22)

• Walkingandmotorcycle(r=0.06andr=0.0051)

• Cars(r=0.09).

Dowereallyneedrelocation?

• Wedividetheoperationalareaincellswithsidelength500m

• Weobservehowthe%ofemptycellsvariesovertime

There are always a lot of empty zones in

the cities!

The CS in City#9 had opened just a few weeks before our

data collection, and its service hadn’t yet stabilized.

Therelocationpotential– part1

Even in the best

case, vehicles remain parked most of the time!

• #emptycells +#idlevehicles =strongconcentrationofvehiclesincertainareas

• Thisisgoodnewsfortheresearchonvehicleredistribution:theoperatorcanindeedexploitalargenumberofvehiclesthatarenotusedmostofthetime…

Therelocationpotential– part2

• …provideditispossibletoaccuratelypredictwherecarswillberequestedinthenearfuture

• WemeasureCELL REGULARITY intermsofthenumberofpickupeventsobservedwithinthecellduringworkingdays.

• Inordertomeasurehowmuchthenumberofpickupsvariesacrosstheobservationperiodweusethetechniquedescribedin[1].

• Wedivideeachdayintobins• ForeachcellN,wecomputethe#oftripsstartingatNforeachbin(1,…,n)ofdayi:

• Wecomputetheaccumulatedvarianceduringthel daysofobservationperiodas:

[1]Zhong,Chen,etal."Variabilityinregularity:MiningtemporalmobilitypatternsinLondon,SingaporeandBeijingusingsmart-carddata." PloS one 11.2(2016):e0149222.[1]Zhong,Chen,etal."Variabilityinregularity:MiningtemporalmobilitypatternsinLondon,SingaporeandBeijingusingsmart-carddata." PloS one 11.2(2016):e0149222.

Squared correlation

the vast majority of cells has an extremely predictable behaviour, with limited variability

the number of outliers is significant, and it should be taken into account when designing supply models for car sharing services

(e.g., unpredictable cells should not be taken into account in the redistribution process).

Howmanydifferentcellusages?• Weusedthefollowingtechniques:

• Wediscretize time into bins with a duration of 10 minutes• We compute the average occupancy (# available vehicles) in each bin across

the observation period• We normalize by the average daily availability in each cell• DynamicTimeWarping:measureshow“close”twotimeseriesare

• Takesintoaccountminorshiftsintimethatcanbeoftenseenintimeseries• PAMClustering:createsk groupsofsimilarstationsbasedontheDTWdistance

• SilhouetteMethod:forselectingthemostinformativek

• The optimal number of clusters in all cities ranges from 2 to 4.

• The fourth cluster, when present, is a very special cluster, composed of just one cell (overlapping with the airportzone)

Bell shaped: above average availability at night and below average availability

during the day

Inverted bell: above average availability at night and below average availability during the day

Flat: no significant difference in usage is detected over the whole day

Business/commercial areas

Residential areas

Mixed usage

Identifyingpotentialserviceareas

• Cleaningandmaintenance isacriticaloperationalaspectinCS

• thecarsharingworkforceisdispatchedtocollectvehiclesthatareinneedofeither

• Movingworkersaroundisexpensive,andmoreefficientsolutionscouldbefoundbasedonthevehicleusageinthecity

• POTENTIAL SERVICE AREA:alocationvehiclepassbywithveryhighprobabilitywithinapredefinedtimewindow

• workshopscouldbedeployedinthisarea,andthiswouldmakecleaningandmaintenanceoperationsmuchmoreefficient

Ourapproach

• WedefineareferencetimewindowW,correspondingtotheacceptedtimebeforetakingoutavehicleformaintenance

• Then,foreachcell,wecountthenumberofdistinctvehiclesseenbythecellsduringW

• Weassumethatathresholdof50%vehicleswouldbeacceptableforthecarsharingoperatortojustifytheopeningofaworkshopinthearea

Only 2 out of 10 when W=15 days

• In both cases, the service area would be located at the airport!

Only 5 cities out of 10 are able to satisfy this requirement when W=30 days

Conclusions

• Carsharinghasn’tbeenmuchdata-drivensofar,butdataanalysiscanofferimportantinsightsforthemanagementofacarsharingsystem

• Inthisworkwehaveprovidedsomeexamples,highlightingtheirimpactonCSoperations:

• Wehaveshowntheimportanceofvehicleutilizationrate andhowitcorrelateswithmodalsharesinthecities

• Wehavehighlightedthehugepotentialforvehiclerelocationduetothehighnumberofemptycellsandalsoofidlevehicles

• Wehaveshownthatthedemandisgenerallyverypredictable,andthisalsocanbeexploitedforrelocation

• Wehavediscussedhowtosmartlydeploycleaningandmaintenancefacilities basedonvehicleflowsintheCSnetwork