Insights into carsharing demand dynamics: Outputs of an agent-based model application to Lisbon, Portugal

ABSTRACT

Two important claims for carsharing systems are their increased flexibility and potential contribution to reducing transport externalities such as pollution. Carsharing typically involves a fleet of vehicles in stations around a city that clients may use on an hourly-payment basis. Classical round-trip systems address a niche market of shopping and errand trips. However, a growing market is now arising providing one-way trips to clients. Great uncertainty remains on the economic viability of this type of carsharing given the complex relation between supply and demand, and how this may influence the level of service provided. Realistic modeling tools that include both supply and demand characterization and allow testing several carsharing operational parameters are scarce. In this sense, a detailed agent-based model was developed to simulate one-way carsharing systems. The simulation incorporates a stochastic demand model discretized in time and space and a detailed environment characterization with realistic travel times. The operation includes maintenance operations, relocations and reservations. The model was applied to the case-study city of Lisbon. Our results show that comparing to other modes, carsharing performs worse than private cars both in terms of time and cost. Nevertheless, it clearly outperforms taxis in terms of cost, and outperforms buses, metro and walking in terms of travel time. The competitiveness of carsharing is highly determined by trip length, becoming more competitive than other modes (travel-time wise) as trips become longer. The operational policies as car-fleet relocation and car reservation showed significant effects in enhancing profit while preserving good customers' satisfaction.

KEYWORDS:

• Agent-based modeling
• discrete choice model
• Lisbon
• one-way carsharing
• public transport

1. Introduction

A carsharing service usually provides its members access to a fleet of vehicles, which can be rented for short periods (Barth & Shaheen, 2002). These services increase the mobility options of the users, complementing transit and potentially reducing the need for owning a private car (Shaheen & Cohen, 2007).

There are essentially two types of carsharing programs with respect to the type of trips: round-trip, in which users must return the car to its departing point, and one-way, in which users may drop off the car anywhere within the designated areas. Round-trip systems are generally used for errands or shopping, since the renting scheme makes them unsuitable for commuting or for other long duration activities (Barth & Shaheen, 2002). One-way systems are more flexible for the client, but impose more difficulties to the operator. They are harder to manage, as the freedom given to their users is bound to create imbalances on the distribution of the fleet (Jorge, Correia, & Barnhart, 2014). To compensate this disequilibrium, the operator can relocate the vehicles from areas with excess of supply to those where the demand is greater, but these operations are quite costly even when optimized (Correia & Antunes, 2012).

Round-trip programs are by far more common, but over the last 5 years there has been a significant growth of one-way programs, either station-based (in which cars are grouped in parking lots under the operator's responsibility) or free-floating (in which cars may be parked anywhere on the street, as long as it is legal and within the service area). However, despite this growth, there is still uncertainty on the position of the carsharing systems when compared to the other urban modes. Hence, it is difficult for the public administration to make decisions on supporting or not supporting such systems.

Currently there is a lack of models that can adequately represent the supply and demand of such systems as a function of several operational parameters in order to produce indicators that support decisions on if and how to run these systems (Hampshire & Sinha, 2011; Jorge & Correia, 2013). Several authors propose the use of spatial and temporal disaggregate models to study such systems given the fact that, especially in this mode, supply and demand influence each other in a significant way: as a vehicle is taken from one station to the other depleting the origin and increasing the stock in the destination, demand is implicitly changing the supply which is not a phenomenon observable in classical transit systems (Ayed, Khadraoui, & Aggoune, 2015; Barrios & Doig, 2014; Ciari, Schuessler, & Axhausen, 2013; Jorge & Correia, 2013).

One of such disaggregate techniques is the agent-based modeling (ABM) methodology. It is a natural way of replicating travelers and their movements in an urban space allowing mimicking different behavior rules and interactions with the environment, in this case a road network and vehicle supply in the carsharing stations. This idea is not entirely new and it is possible to find in the literature disaggregate approaches for carsharing such as the one by Ciari et al. (2013), where demand for carsharing is simulated at the level of each traveler. However, in their modeling efforts based on MATSIM there was no particular attention given to the perspective of the operator. Hence in this paper we argue that a more detailed model on the operations side of carsharing, especially one that includes operations which help tackling the effects of the complex supply/demand relation such as relocations (Jorge & Correia, 2013), is needed.

The definition of agent-based simulation model used in this work is the one by Bonabeau (2002): “In agent-based modeling (ABM), a system is modeled as a collection of autonomous decision-making entities called agents. Each agent individually assesses its situation and makes decisions on the basis of a set of rules. Agents may execute various behaviors appropriate for the system they represent—for example, producing, consuming, or selling.” This means that we are not considering other more complex components of the agents as memory and learning, focusing more on the decision-making and spatial interaction.

The main objective of this paper is to propose such new ABM to add in understanding the complex supply—demand relationship. In particular, we analyze the demand perspective by analyzing some performance indicators such as geographical and time distribution of demand, modal share changes, or time and cost gains/losses when compared to competing modes under different operational configurations of carsharing systems. Regarding these operations, we add relocation operations of vehicles along the day and reservations in order to assess the impact of these operational policies in both the client level of service and the company profit. This element allows assessing how different operation strategies may influence the number of clients who cannot find a car.

In the following section, we briefly review the state-of-the-art where the need for such ABM is demonstrated. Subsequently we describe the new ABM that we propose. This model is then applied to the case study city of Lisbon (Portugal) and results are explored in order to understand the role that carsharing may play among the remaining transport options in Lisbon. We end the paper with conclusions and future developments.

2. Literature review

Significant literature can be found on the topic of carsharing systems, especially describing field test studies that were set up mainly in the United States over the last three decades, and also exploring the demand of real round-trip carsharing companies that started operating in the meantime, producing the first use statistics and demand pattern information (Cervero, 2003; de Lorimier & El-Geneidy, 2013; Morency, Habib, Grasset, & Islam, 2012; Morency, Trepanier, & Agard, 2011; Shaheen & Wright, 2001; Zheng et al., 2009). These experiments and businesses were set up mostly for round-trip carsharing systems, while the one-way systems arose during the last decade.

The effect of carsharing systems on the urban mobility patterns has also been studied, observing indicators such as modal share before and after the system has been implemented, car ownership changes, vehicle-kilometers traveled and related vehicle emissions (Martin, Shaheen, & Lidicker, 2010; Shaheen & Rodier, 2005; Sioui, Morency, & Trepanier, 2013; ter Schure, Napolitan, & Hutchinson, 2012). The general trend is a reduction of the vehicle-kilometers traveled in private cars as well as car ownership rates. Modal change comes not only from private car users but also from transit users. There is an estimated reduction of pollutant emissions with the use of carsharing but more detailed studies are needed to draw stronger conclusions on this impact (Martin et al., 2010).

Nowadays there is a profusion of carsharing systems in operation either round-trip or one-way, station-based or floating systems. Along the evolution process, companies faced management problems and research started addressing these by introducing models to analyze and solve these issues (de Lorimier & El-Geneidy, 2013). One of the most relevant, in one-way carsharing systems management, occurs with the imbalance that results from such systems and there is a considerable body of work on studying ways to solve this problem mainly through simulation and optimization techniques (Barth & Todd, 1999; Barth, Todd, & Xue, 2004; Correia & Antunes, 2012; Fan, Machemehl, & Lownes, 2008; Jorge, Molnar, & de Almeida Correia, 2015; Kek, Cheu, & Chor, 2006). They either use travelers as a balancing factor, a staff of drivers (Jorge & Correia, 2013; Kek, Cheu, Meng, & Fung, 2009; Weikl & Bogenberger, 2013), or even changing the network of stations for obtaining a more favorable balanced trip pattern (Correia & Antunes, 2012).

Despite the work that has been developed regarding the understanding of carsharing systems and how to manage them, scant effort has been put in trying to unify this research in a comprehensive approach. Some research focused greatly in describing the demand or estimating the impacts of the systems in the mobility patterns. Others addressed the issue of stocks imbalance, which is more an operational problem. Rarely a bridge was built connecting both topics. Often demand is considered as independent of the supply and vice-versa, which in one-way carsharing systems is a high risk due to the high interdependence between the number of trips and cars available.

In the rare papers where this is considered, the methods may have not been the right ones to translate the highly spatial and temporal complexity of such transportation systems. A case of system dynamics was used by Papanikolaou (2011), where the author acknowledges the importance of feedbacks in the system but is unable to validate his aggregate model for a real system and later recognizes that an aggregate method such as system dynamics is unable to represent the highly spatial and temporal nature of these systems.

Ciari et al. (2013), recognizing these special characteristics of carsharing systems, applied an activity-based microsimulation approach (MATsim) to estimate more accurately the demand for carsharing systems, taking into consideration all transport modes supplied in the city and the characteristics of each traveler. In their first paper on this topic, they applied the model to a round-trip carsharing system for Zurich city center, not considering the supply side. The model assumed that there was always a vehicle available at a station to be used which yields a theoretical potential for carsharing demand.

The same authors later evolved their model to consider station-based and free-float carsharing in both their demand and supply side (Ciari, Bock, & Balmer, 2014). In this research, the simulation model follows the traveler and the vehicle for the case-study city of Berlin. Nevertheless, the improved model does not get into details about the operation side of the system which represents a considerable cost for the private companies that are running these businesses. The distribution of the number of vehicles in every zone, as well as the additional maintenance operations, including vehicle relocations, are examples of operational decisions that the companies must take. Because the model is simulating all the mode choice including private transport departure time and path choice, it is taking 4 days to converge to an equilibrium solution (Balac, Ciari, & Axhausen, 2015). We argue that this computational time makes it difficult for using the model to study in higher detail the operation scenarios that carsharing companies need to choose. However, the authors have also done some interesting work on testing different renting prices and their effect on total demand of the service (Ciari, Balac, & Balmer, 2015). Recently, a new evolution of this model incorporates a heuristic to allocate parking spaces to carsharing stations aiming at minimizing the required parking capacity in order to satisfy the potential demand. The results obtained evidence for the ability of the model to efficiently allocate parking lots to stations, preserving clients' accessibility at the desired level (Balac, Ciari, & Axhausen, 2016).

The linkage of carsharing systems and synergies with autonomous driving systems has also been recently explored. Several studies have analyzed, through simulations, the role of shared mobility services either as additional service in the mobility market (Martinez, Correia, & Viegas, 2015), or by fully deployed systems that would replace all motorized mobility in a city (Fagnant & Kockelman, 2014; International Transport Forum Corporate Partnership Board, 2015; Zachariah, Gao, Kornhauser, & Mufti, 2013).

Worth mentioning as well are the contributions of other teams using other agent-based platforms for modeling carsharing demand, most notably Barrios & Doig (2014) use a well-known agent-based simulation platform called NetLogo to model different relocation policies in one-way carsharing systems. Interaction between clients and vehicles and the demand modeling aspect of the simulated city is very basic which adds to a limited description of the logistics of the process. Heilig, Mallig, Schröder, Kagerbauer, & Vortisch (2015) used MobiTopp, a travel demand model based on the principle of agent-based simulation, to model both round-trip (station-based) and one-way (free-float) carsharing. The authors present a mode choice description of the travelling decisions of each customer using discrete choice modeling and rule-based approach to model the decisions to use one of the two carsharing modes. Moreover, the model observes the existing stocks of vehicles at each location, which is changed as travelers use the cars. Traffic is modeled through an external connection to a VISUM model. Nevertheless, in this model, there is no concern regarding the operation management of the system and the interaction between travelers and vehicles is represented at a zone-based scale.

In our research, we microsimulate one-way carsharing demand, at the same scale, i.e., traveler by traveler, but adding the supply side by representing the stocks of vehicles in the city over a day that result from the vehicles' movements and system management options. This model should allow observing not only the perspective of the level of service offered to the traveler but also the effort that the operator must put into offering such a level of service.

3. Agent-based model of one-way carsharing systems

3.1 Model

The ABM proposed in this paper simulates the daily operation of a hypothetical carsharing system in Lisbon, Portugal. Nearly three million people live in the Lisbon Metropolitan Area (LMA), 550,000 of which are within the city of Lisbon. Every day over 400,000 people travel to Lisbon to work or study (Ine, 2011).

Within the model, the city is characterized by its road network, i.e., a set of nodes and arcs. Travel times were obtained from AIMSUN (calibrated for observed real demand) for each arc of the network, varying with the time of the day. The Dijkstra algorithm generates the optimal path between any pair of nodes of the network. The city is divided in a homogeneous grid of 200 m × 200 m cells.

The model represents a station-based one-way carsharing system: users can pick up a car at any station where there are vehicles available, and are free to drop off the car at any station with available parking spaces.

An optimization model was developed to estimate the location and relative dimension of the stations, based on the mobility patterns of Lisbon. This model was adapted from the traditional p-median formulation, minimizing the distance of travel origins to carsharing stations, selecting the network nodes as potential candidates and imposing the constraint of not having two stations less than 250 m apart. This formulation was adapted from previous case-study applications to Lisbon (Correia & Antunes, 2012; Eiro, Martinez, & Viegasm, 2011).

The capacity of each station is obtained by rounding the ratio between the number of parking spaces, which is defined as a planning parameter at the beginning of the simulation, and the stations dimension obtained from the optimization model. If the capacity of a station equals zero, the station will not be active during that simulation run.

The agents of this model are the users and the carsharing operator who control two other reactive agents—vehicles and staff.

Next, we describe the main features of the model for each agent present in the model.

3.2 Demand

Based on an extensive mobility survey conducted in Lisbon (Câmara Municipal de Lisboa, 2005), we created a synthetic population of trips within the city, aggregated by the aforementioned grids. The used synthetic travel simulation model was developed and calibrated for the LMA in a previous study (Viegas & Martinez, 2010). The model output contains all the trip extremes not only discretized in space (at the census block level) but also in time (presenting different trip departure and arrival times) for a synthetic week day (Viegas & Martinez, 2010).

The model is stochastic because the number of trips that actually occur between two grid-cells at each hour is generated using a Poisson distribution with λ equal to the average number of hourly trips.

Each trip is characterized not only by its time of occurrence, origin and destination, but also by trip purpose, traveler's age and by whether or not the traveler has a transit pass. Additionally, based on Census data and other mobility surveys (Martínez, Viegas, & Silva, 2009; Moura, Duarte, & Viegas, 2007; Santos, Martínez, Viegas, & Alves, 2011), each trip is further characterized by the traveler's gender, income and by whether the traveler has a driver's license, car, motorcycle, parking spaces at home and at work. The purpose of the trip determines the activity duration, which determines parking costs.

The destination grids are characterized by parking cost (in euros) and parking pressure (ratio between dynamic parking demand along the day and parking supply), both depending on time of day, linking the available demand data with the statistics and pricing of Lisbon's parking.

For all modes considered—car, motorcycle, taxi, walking, bus or tram, subway or suburban train, and a combination between light and heavy transport modes—the trip is characterized by real-life data in terms of access time, waiting time, travel time, cost and number of transfers (if applicable) (Santos et al., 2011). Carsharing trips are characterized by access time, travel time and cost, all of these depending solely on the stations' location. It is therefore assumed a perfect level of service, where there are always vehicles and parking spaces available (what we refer to as “potential carsharing trips”).

To determine the modal choice of the passenger, the ABM incorporates the discrete choice model (DCM) described in Eiró & Martínez (2014). The original model aimed at assessing the impact of new shared mobility options in the LMA using stated preferences data. This model was adapted for this paper since we do not take into consideration other alternative modes that the authors considered (minibus, carpooling and shared taxis). The model specification and calibration results are presented in Table 1, where nested logit aggregates transport alternatives into:

• Private transport nest: private car (PC), motorcycle (MT) and taxi (TX);

• Public transport nest: bus (BS), walk (WK), heavy public transport—subway and rail (HV) and bus + heavy public transport (CB) and carsharing (CH).

Table 1. Coefficients of the obtained discrete choice model.

	Transport alternatives
	Private transport (PC)			Public transport (PT)
Attributes	PC	MT	TX	BS	WK	HV	CB	CH
ASC	—	—	−4.230^***	−0.218^*	−1.140^***	−0.270^*	−0.706^***	−1.190^***
Sociodemographic attributes
Age (25–35)	—	—	—	—	—	0.559^***	—	0.559^***
Age (35–65)	—	—	—	—	−0.308^*	—	—	—
Age (+65)	—	—	—	0.195^*	—	—	—	—
Income (thousand €)	—	—	—	0.007^***	—	—	—	—
Land use, car and public transport monthly pass availability
No parking at home	—	—	0.237^***	—	—	—	—	0.237^***
No parking at destination	—	—	0.237^***	—	—	—	—	0.237^***
Own car	—	—	—	—	—	—	—	−0.308^***
Public transport pass	—	—	—	0.562^***	−0.766^***	0.562^***	0.562^***	—
Parking pressure (0–1) ^†	−0.131^***	—	—	—	—	—	—	—
Transport operation attributes
Fuel cost (€)	−0.326^***	−0.326^***	—	—	—	—	—	—
Toll (€)	−0.221^***	−0.221^***	—	—	—	—	—	—
Parking cost (€)	−0.221^***	—	—	—	—	—	—	—
Travel time (min)	−0.026^***	−0.026^***	−0.026^***	−0.024^***	−0.005^***	−0.015^***	−0.015^***	−0.026^***
Access time (min)	—	—	—	−0.051^***	—	−0.053^***	−0.051^***	−0.082^***
Tariff (€)	—	—	−0.115^***	−0.498^***	—	−0.498^***	−0.498^***	−0.229^***
Transfers	—	—	—	−0.199^***	—	−0.150^**	−0.150^**	—
Waiting time (min)	—	—	−0.028^***	−0.028^***	—	−0.045^***	−0.028^***	—
Nested scale (η)	1.000	1.951
0-test significance	—	^***
1-test significance	—	^***

*** Significant at the 99% level.

** Significant at the 95% level.

* Significant at the 90% level.

^† Parking pressure defined as the ratio between estimated demand and supply of parking in a specific area and time period of the day.

The model proved to have an adequate specification: ρ² = 0.37 and the utility function of each transport alternative included socio-demographic variables of the traveler, land use, car and transit availability and instrumental attributes such as travel time and cost, all being statistically significant at a 90% confidence level.

Nonetheless, we should acknowledge that using stated preference data over a new system not experienced by respondents might lead to some biased responses. We consider that the obtained results may present some positive bias toward the carsharing system, as respondents are not aware of the car availability issue. Commonly frustrated mode choice selection may produce a change of mode attractiveness overtime and hence a change of the mode coefficients. As memory is not included in the model and the Bayesian estimate of the coefficients could not be obtained from the available data, they were considered fixed during the whole simulation.

Having the full characterization of the trip as input, the DCM calculates the probability of choosing each mode. A mode is assigned to the traveler by Monte Carlo simulation where modes with higher probability will be chosen more often. Since the option of carsharing can be deactivated, the model can be used to study modal shifts following the new carsharing option entrance.

When the model estimates a carsharing choice, a new user (agent) is generated in the simulation environment, with the attributes: departure node, arrival node, starting time and activity duration.

Currently, one user is equivalent to one trip, i.e., users do not cluster in parties nor have memory. Therefore, previous experience using carsharing services does not have an impact on future choices. It also means that each decision is made individually and is not activity-based, nor the daily routines of the user's family are considered in the choice, which may lead to some overestimation of the potential demand produced by this model.

3.3 Users

All trips that are generated in the demand model, which potentially would be interested in performing a carsharing trip in a perfect level of service system (permanent car availability close to origin and permanent parking space availability at stations near the final destination), are generated as carsharing potential users. A user's first step is to choose his/her pickup and drop-off station. Since it is assumed users have access to information regarding the vehicles' distribution, they select the nearest station that has a vehicle available as departure point and choose the nearest station to the arrival point with an available parking space at the moment of decision (Correia, Jorge, & Antunes, 2013). If there are no such stations within a reasonable walking distance, they leave the system and choose a different mode through the DCM (carsharing excluded from choice set and, as such, the “potential carsharing trip” is discarded).

The reasonable walking distance is determined by the accessibility functions calibrated for Lisbon in a previous study (Martínez & Viegas, 2013). Since in this study the access to carsharing stations was not part of the list of surveyed functions, we considered the willingness to walk to a carsharing station equivalent to the one estimated for a bus stop accessibility function calibrated (generalized logistic function). The used equation is given by:(1) where B is the growth rate, ν affects near which asymptote growth occurs, Q depends on the value f(0), and M is the x value of the maximum growth if Q = ν.

The calibration parameters of the equation for willingness to walk to a bus station in Lisbon are: B = 0.203, ν = 0.002, Q = 0.006 and M = 1.13431.

Once users have chosen the pickup station, they will walk there. Since in our baseline scenario reservations are not allowed, users might not find cars available because these were picked up by other users in the meantime. When that happens, users wait for another car for about 5 min. If no car arrives during this period, users give up and leave the system, choosing a different transport mode. When there is a car available, users pick up the car and start the rental period. When they arrive at the destination station, they drop off the car and trip price is calculated. Afterward they walk to their final destination where their trip is completed.

If the purpose of the trip is shopping or leisure, users will keep the car for the duration of that activity, even while the car is parked. When returning the car, users may find the station full, in which case they search in nearby stations until a free space is found. This can be considered as a round-trip carsharing behavior, as users most probably will return the vehicle to departure station. This type of trips is reported in the literature along with errand trips as the natural market niche of round trip carsharing (Shaheen & Cohen, 2007).

3.4 Carsharing operator: Cars

In the beginning of the simulation, each car is assigned to a station. This distribution is done proportionally to the stations' capacity, which was explained previously.

Cars are available when parked at a station, and not scheduled for maintenance or relocation. After car pick-up, users drive to the drop-off station. When the car is parked at a station, the trip ends and the car becomes available for the next user or for maintenance operations or relocation.

Each time a car enters a station, an assessment is made on its need for maintenance, whether refueling, cleaning or inspection. A car must refuel if it has traveled over 300 km and must be inspected after traveling over 2,500 km since previous inspection. A car is cleaned on average after every 18 rides. When a car needs maintenance, it becomes unavailable to users. A request is sent to the operator and the nearest available staff member is assigned to the car. If no staff members are available, the car will wait until one is free. The time distributions used for each of these operations are:

• uniform distribution between 5 and 10 min for refueling;

• uniform distribution between 10 and 30 min for cleaning;

• uniform distribution between 120 and 240 min for inspection.

If a car is assigned to be relocated (which implies it already has a staff member assigned to that task), it becomes unavailable to customers. After waiting for the assigned staff member to arrive, the car is driven to its new station. If by chance the new station is full, the car will be driven to the nearest station with available parking, until it can finally be parked. When it does so, it becomes available to customers once again.

Car movements are based on a static representation of the traffic environment, where origin-destination flows are allocated to a simple, topologically correct road network representation that accounts for per-link occupancy (and thus for speeds), by time of day. Travel time is attributed to each link, varying with the time of the day, based on which the optimal path between any pair of nodes of the network is identified. This simplification derives from the assumed small impact of carsharing vehicles in traffic (modal share less than 2%, which would be approximately the current weight of taxis' trips in the total traffic).

3.5 Carsharing operator: Staff

Staff members are responsible for maintenance and relocation operations. In the beginning of the simulation, staff members are assigned to one of the company's depots (set up in the simulation model parameters in five depots distributed along the city of Lisbon) and to a working shift. Staff members move around in foldable scooters that fit in the vehicles.

Staff members remain in the depot until they are assigned to a car, considering the time required to initiate the performance of the activity (simulation time of staff's previous task plus travel time to station where the operation takes place). They will then drive their scooter and perform their task. When finished, they become available again, and remain at the same station for a short period, waiting for a new assignment. If this fails to happen, they start to drive back to the nearest depot. However, they can be called while they are moving—it is assumed the company keeps track of the location not only of its cars but also of its staff.

When the shift ends, staff members go off-duty. When their shift starts again, they will resume work at the same depot where they ended their previous shift.

4. Experiment and results

4.1 Set-up

The potential users of the carsharing system were estimated with the DCM, based on a fixed price and distribution of carsharing stations in the city. Taking this number of users, performance of the system would correspond to an ideal carsharing system operating in Lisbon, where users can always find cars and parking available at all stations. This is similar to the research done by Ciari et al. (2013). Although not accurate, it was simulated to serve as a benchmark to compare with the results of the realistic model when the full operation is simulated, i.e., with all supply constraints.

For similar tariffs of Car2Go in 2014 (0.29€/min and 0.19€/min when the car is parked but not available for other users) and the stations' distribution explained above, we tested several different configurations, varying fleet and staff dimension, and the “parking spaces-to-car” ratio. Based on our analysis of operational indicators (e.g., cost, revenue, occupancy rate and unsatisfied trips %), we decided to use as a base scenario a fleet of 1,500 cars, 120 employees and 2 parking spaces per car. Larger fleets increase the number of satisfied trips, but the marginal benefit of each vehicle strongly diminishes as the fleet increases. Increasing the number of parking spaces noticeably inflates the costs while not having a significantly positive effect over users' satisfaction (i.e., percentage of potential clients that find a car available). Some of these trade-offs have been already assessed using this model in a previous paper (Mendes Lopes, Martinez, & Correia, 2014).

To analyze the results we divided Lisbon into three areas: Central Business District (CBD), a consolidated mixed-area around the CBD and the more recent mono-functional residential area in the outskirts.

4.2 Results of the baseline scenario

The model generates around 1,140,000 daily trips within Lisbon, of which around 34,000 are potential carsharing trips (perfect level of service carsharing system) generated by the demand model (3%). Of these potential trips, around 85% were able to be satisfied by the system, 13% chose a different mode after verifying there were no vehicles available at the nearby stations, and 7% were forced to choose a different mode after having walked to the nearest station and finding no cars there.

Each carsharing vehicle was used on average 13 times/day, but they were only occupied 12% of their time. This is due to most trips being very short: on average 4.1 km over 9.4 min, costing around 3€. Walking access time is on average 8.9 min, so users complete their trip in 18.3 min. As an example the average carsharing trip duration in Montreal (Auto-mobile) is 13 min, which indicates the plausibility of the obtained results for a medium city, targeting especially not very long trips (Shaheen, Chan, & Micheaux, 2015). Yet, the distribution on the average number of vehicle trips per day is quite high, mostly due to the demand bias resulting from the stated preferences experiment as discussed above.

The distribution of trips during the day is uneven, peaking significantly around 8 am, around lunchtime, and at 5 pm (see Figure 1).

Figure 1. Trips distribution over a day.

Table 2 presents the modal shares before and after carsharing is added to the model as well as a reference mode share from the census of 2011. The census data in the table serve the purpose of providing a rough validation of the results produced by the model. One should not compare the modal split to the decimal place because these numbers come from different sources and different points in time. The census data refer to the trips that are done inside Lisbon and only with respect to two trip motives: work and study. If one compares the mode shares obtained by the simulation model and the census, we conclude that the model is not far off. The greater percentage of automobile trips and lower percentage of walking trips may be explained by the difference in trip motives that have been included.

Table 2. Modal share before and after the introduction of carsharing (realistic system).

Scenario	Private car (PC)	Taxi	Bus	Metro	Walk	Moto	Heavy + light	Carsharing (CS)
Census 2011 work and study trips (%)	43.7	0.7	21.4	13.8	19.4	1.0		—
Model before CS (%)	38.9	1.4	23.0	7.8	26.3	1.1	1.5	—
Model after CS (%)	38.3	1.3	22.5	7.5	25.5	1.1	1.4	2.4

With regard to the carsharing effect on the mode choice, we see that about 4% of the people who used to walk switch to carsharing. Taxi and Metro are also considerable losers, as 3% of their passengers switch to carsharing. PC, buses and Heavy+Light combination lose around 2% of their users. From the car sharing perspective, 40% of carsharing trips used to walk, 26% used PC, 22% took the bus, 10% took metro, 2% used taxi and the remaining used the Heavy+Light combination or motorbike. Since these two latter modes correspond to a smaller fraction, they will not be analyzed hereon.

Table 3 presents the characteristics of carsharing trips between the three zones considered: CBD—C, Mixed area—M and Residential area—R; for ideal (no operation constraints) and realistic scenarios.

Table 3. Carsharing trips between the three city zones.

	Ideal				Realistic				Difference realistic–ideal (%)
	Number of trips	Time (min)	Cost (€)	Speed (km/h)	Number of trips	Time (min)	Cost (€)	Speed (km/h)	Number of trips	Time	Cost	Speed
Global	34,015	13.4	2.28	11	27,294	18.3	3.01	9	−20	37	32	−18
C–C	4,314	9.8	1.6	6.7	3,395	15.5	2.6	5.2	−21	+58	+63	−22
M–M	8,730	12.7	2	8.7	6,986	16.6	2.7	7.4	−20	+31	+35	−15
R–R	4,550	13.4	1.9	10.1	3,344	18.3	2.5	8.2	−27	+37	+32	−19
C–M	8,709	12.5	2.2	10.4	7,274	17.8	3	7.8	−16	+42	+36	−25
M–R	5,577	16.9	3	14.6	4,598	21.3	3.6	12	−18	+26	+20	−18
C–R	2,135	18.5	3.7	17.9	1,697	24.2	4.5	14.2	−21	+31	+22	−21

Although smaller, CBD concentrates a larger share of economic activity, thus generating a high number of carsharing trips within. These trips are shorter than the rest of the trips in Lisbon. Moreover, these are also the ones whose travel time and cost increases the most when comparing the realistic and ideal scenarios—meaning that supply characterization makes a difference here.

The largest percentage of unsatisfied trips occurs within the residential area (see Figure 2). This area has smaller and more scattered stations, decreasing the odds of having one vehicle available within a reasonable walking distance. Likewise, this area has lower carsharing activity. Trips with origin or destination in this area have higher duration and cost than the city's average, since they are usually longer as well. However, their average speed is higher and their overall cost per km is lower than average, making a priori these trips more advantageous when compared to the remaining transport modes. Still, the residential area is the one where carsharing has a smaller modal share (only 1.9%, while it is 2.5% for the mixed area and 2.9% for the CBD), probably due to the scarcity of available vehicles.

Figure 2. Space and time distribution of unsatisfied carsharing trips: (a) morning peak (7–10 am); (b) lunch (12–2 pm); (c) afternoon peak (4–8 pm); and (d) evening + night (9 pm to 6 am).

Figure 2 shows the number of unsatisfied trips per grid-cell, which happens to change significantly for different periods of the day. The morning peak, being the busiest period, is also the one with more unsatisfied trips. While some of these occur within the CBD or the residential area, the bulk of them are located in the mixed area. The number of unsatisfied trips in the residential area is significantly higher during this period than during the rest of the day, probably due to the existence of many commuter trips, which cannot be satisfied properly if fleet relocation policies are not active.

During the lunch break, there are fewer unsatisfied trips and their spatial distribution is fairly even across the city. The afternoon peak, while being less intensive than the morning one, also presents a large number of unsatisfied trips. Not surprisingly, most of these have their origin in the CBD. Likewise, this area is the origin of the great majority of unsatisfied trips during the evening and night periods, also requiring an active relocation policy.

4.3 How does carsharing compare to other modes?

This section analyzes how car sharing performs when compared to the remaining modes in the Current scenario (current modes available to users). In Figure 3, we show that carsharing performs worse than PC both in terms of time and cost. Carsharing clearly outperforms taxis in terms of cost and outperforms buses, metro and walking in terms of time—these differences being far more pronounced for the ideal scenario (scenario where all potential customers are served by the system and do not lose additional time searching for a car at origin or searching for a parking location at destination). Interestingly, the time losses of a realistic system (results given by the agent-based simulation with the real stocks of vehicles at each station along the day) are enough to turn carsharing travel time from slightly better than taxis to slightly worse.

Figure 3. Carsharing (realistic and ideal) compared to other modes.

We analyzed the system performance for different periods of the day (see Table 4).

Table 4. Carsharing (realistic and ideal) compared to other modes during different periods of the day.

			Car	Taxi	Bus	Metro	Walk
Night (1–6 am)	Time	Realistic	−64%	−38%	—	—	231%
		Ideal	−47%	−10%	—	—	382%
	Cost	Realistic	−67%	233%	—	—	−100%
		Ideal	−57%	341%	—	—	−100%
Morning peak (7–10 am)	Time	Realistic	−39%	−19%	61%	52%	75%
		Ideal	−13%	16%	130%	117%	151%
	Cost	Realistic	−59%	99%	−71%	−71%	−100%
		Ideal	−42%	180%	−59%	−59%	−100%
Lunch (12–2 pm)	Time	Realistic	−40%	−16%	88%	74%	56%
		Ideal	−24%	7%	139%	120%	98%
	Cost	Realistic	−69%	122%	−65%	−65%	−100%
		Ideal	−61%	183%	−55%	−55%	−100%
Afternoon peak (4–8 pm)	Time	Realistic	−29%	−11%	67%	55%	75%
		Ideal	−9%	14%	114%	99%	124%
	Cost	Realistic	−78%	90%	−70%	−70%	−100%
		Ideal	−72%	139%	−63%	−62%	−100%
Evening (9 pm to 12 am)	Time	Realistic	−48%	−23%	81%	66%	131%
		Ideal	−31%	2%	139%	119%	205%
	Cost	Realistic	−71%	189%	−63%	−62%	−100%
		Ideal	−64%	255%	−54%	−54%	−100%

The global patterns stay roughly the same for each period: PC is always more advantageous, taxi is always more expensive, buses and metro are always slower and walking even more so.

When transit is scarce or absent (such as during the evening or night periods), people tend to walk longer distances, therefore increasing the average trip duration. In these circumstances, the advantage of carsharing compared to walking timewise becomes even more pronounced. Conversely, during the lunch break, trips are usually shorter, reducing this advantage. Since short trips penalize transit because of access and waiting times, carsharing is even more advantageous than bus and metro during this period.

Since taxis are subject to a night tariff and carsharing is not, the latter becomes an even better option in terms of cost when compared to taxis during the evening and night periods. This could induce a change in the taxi pricing policy.

Although more carsharing trips are being done in the morning and afternoon peaks, these are the periods where carsharing is relatively less advantageous compared to the remaining modes.

The advantages of carsharing compared to other modes are most decisively conditioned by the trip length. Figure 4 shows the comparison between carsharing and each mode for six categories of trip length.

Figure 4. Carsharing (realistic and ideal) compared to other modes for different trip lengths: (a) private car; (b) taxi; (c) bus; and (d) metro.

Figure 4a shows that PC is always faster than carsharing, because of its short access time. However, this advantage dilutes as the trips' length increases. Up to 3 km, carsharing trips take more time, considering the access and alighting time (67%) than PC, while between 15 and 18 km the difference is much smaller, as the access and alighting time represents a smaller share of the overall time (11%). In the realistic system, users may be forced to search for a station with available parking spaces, increasing the cost of the trip. This extra cost will weigh more significantly when the trip is shorter. For this reason, in the realistic system, longer trips are more cost advantageous than shorter ones, while in the ideal system this remains fairly constant across trip lengths.

As Figure 4b shows, in the ideal system, the advantage of carsharing compared to taxis is greater for short trips since, unlike taxis, carsharing does not have an initial charge. As the trip length increases, the effect of the initial charge is diluted and carsharing becomes relatively less attractive. However, after a certain distance, the relative advantage of carsharing begins to increase again, as its price per km is lower. However, in the realistic system users spend time and money searching for stations with available spaces, increasing the trip cost. Since this extra cost is more significant for shorter trips, these end up being quite less advantageous than the ideal ones.

Figure 4c shows that, when compared to buses, carsharing becomes expensive as the trip's length increases. Buses do not follow necessarily the most direct route between two points, and the longer the trip, the higher the probability of having transfers and therefore increased waiting time. These factors combined lead to carsharing being more advantageous than buses in terms of trip duration for longer trips.

Figure 4d shows that, much like the buses, carsharing becomes more expensive compared to metro as the trips' length increases. However, the time advantages of carsharing do not increase with the trips' length, because eventual transfers and not-so-direct routes are compensated by the higher commercial speed of the metro.

For trips shorter than 3 km, carsharing is only slightly faster than walking, but as the trip distance increases, carsharing becomes significantly more advantageous, being about six times faster than walking for trips longer than 15 km.

4.4 Sensitivity analysis to different operational configurations

In order to test the variability of operational parameters with the introduction of optimized carsharing management features, two alternative scenarios were tested over the baseline scenario (BS): the introduction of a carsharing reservation system and a car-fleet relocation strategy operated by staff members. The tested relocation system only considers short-term booking at departing stations of cars.

The reservation is a simple add-on to the information availability on the internet or smartphone app about cars at stations that allows a client to reserve a car up to 10 min before departure. This policy does not introduce additional fleet management complexity as the operator just needs to block a car that is already at a station, instead of having to plan the presence of a free car as in medium- or long-term booking systems.

Additionally, a relocation strategy policy was defined, adapted from the work of Jorge et al. (2014) to reduce the fleet imbalances during the day. As defined in this paper by the authors, a station at a time period t is classified as a supplier if, on a previous day of operations, the number of customer trips allocated to that station at instant t + x exceeds or equals the number of customer trips that depart that station at the same period, otherwise the station is considered as a demander. In the current test, this definition was implemented with an x of 30 min.

If a car is assigned to be relocated (which implies that it already has a staff member assigned to that task), it becomes unavailable to customers. After waiting for the assigned staff member to arrive, the car is driven to its new station. If by any chance the new station is full, the car will be driven to the nearest one with an available parking space, until it can finally stay idle and become available for new trips.

These three additional operational scenarios were tested over the baseline scenario individually and simultaneously:

• baseline scenario + reservation (BS1);

• baseline scenario + relocation (BS2);

• baseline scenario + reservation + relocation (BS3).

While reservation may have a significant impact on the satisfaction of clients, the operational costs of the system should not vary significantly, which led us to consider the same fleet and station capacity set in the baseline scenario. On the other hand, relocations operations should allow reducing operational costs (fleet size and parking capacity) for the same levels of satisfied demand.

To set the equivalent carsharing fleet for the scenarios with relocation, the fleet was reduced iteratively until the percentage of demand satisfied was approximately the same as in the baseline scenario. The tests led to a reduction of 250 cars (16.7%) of the fleet to ensure the 80% trips satisfaction obtained in the baseline scenario for 5 consecutive simulation days.

The results of the main operational indicators for the several scenarios are presented in Table 5.

Table 5. Summary of the operational indicators for the different tested scenarios.

		Variation to baseline scenario (%)
Indicators (data for an average day of operation)	BS	BS1	BS2	BS3
Carsharing fleet	1,500	+ 0.00	−16.67	−16.67
Carsharing stations capacity	3,000	+ 0.00	−16.67	−16.67
Carsharing fixed costs (worker's salaries + infrastructure costs)	36,708	+ 0.00	−14.58	−14.58
Carsharing variable costs (worker's movements costs + clients movements costs)	14,098.39 €	+ 0.35	+ 3.13	+ 4.68
Carsharing revenues	99,977.27 €	+ 0.48	+ 1.76	+ 2.78
Carsharing profit	85,878.88 €	+ 0.51	+ 1.54	+ 2.47
Number of customers	20,289	+ 0.94	+ 0.29	+ 1.85
Percentage of potential customers that find a car to travel	85.68%	+ 0.94	+ 0.29	+ 1.85
% Customers giving up after trying at least one station	5.72%	−65.44	+ 4.10	−63.77
% Customers giving up with no car nearby at request time	8.88%	+ 28.61	−9.51	+ 20.40
Average number of trials to find a free station at destination	1.26	−0.80	−1.38	−2.72
Vehicle-km of customers	125,059	+ 0.20	+ 2.40	+ 3.85
Vehicle-km of workers	7,962	−1.49	+ 4.48	+ 6.75

Results show that these two operational policies have a significant influence on the system performance. Regarding the reservation scenario, the system produces little impacts on the operational indicators (reduction of vehicle-km of workers and increase in profit). However, the decrease of the number of frustrated clients not finding car at a station is very significant, which might lead to greater customer satisfaction. The side effect of the reservation is that some vehicles are locked without being used for some time before the client arrives, which in some periods of the day leads to clients not being able to find a car although these are present at the station. All in all we believe that it is positive that the frustration of not getting a car happens at the moment of booking the trip and not when arriving at a station like in the base scenario.

The relocation policy has a greater operational impact in the results. This configuration leads to a significant reduction in fixed costs, mainly fleet size (14.58%). Yet, the increase in variable costs derived from the staff movements (an average of 157 relocation movements per day and 1.3 relocations by staff member) hinders this strategy producing a greater profit for the company: 1.54% increase for an average day. Additionally, the relocation policy balances the car stocks over the city, reducing the number of clients who do not find a car for their trip.

The combined scenario of both reservations and relocations produces some synergies. The increase in profit is greater than the one obtained by summing the effect of the two applied separately, showing that they might work well together. The client performance indicators (e.g., % Customers giving up after trying at least one station) tends to be more influenced by the reservation policy, leading to similar results than those obtained when reservations were applied without reservations (scenario BS1).

5. Conclusions

The main objective of this paper was to improve the understanding of the complex relationship between supply and demand in carsharing systems. We proposed a new ABM that simulates a one-way carsharing system as part of the transport supply in a city, accurately describing the demand mode choice and the operation of the system.

The model was applied to the city of Lisbon (Portugal), providing insights on this mode performance through some clarifying indicators such as geographical distribution of demand, modal share changes, or time and cost gains/losses when compared to competing modes.

Our results show that carsharing potential demand can reach 3% of the total number of trips. Notably, 80% of these are able to complete the trip, while the remaining 20% willing to use carsharing were forced to shift to other modes due to car unavailability. Carsharing vehicles are used 18 times/day which corresponds to an average occupancy of 12% of their time only, since most trips are short: average 4.1 km over 9.4 min, costing around 3€. Walking access time is 9 min, thus an average complete trip takes 18 min. Carsharing demand peaks in the morning peak hour, being three times higher than other smaller peaks around lunch and 5 pm. This is potentially a big source of imbalance in the system.

Expectably, Carsharing induces modal shifts. 40% of carsharing users come from walking, 26% used private car, 22% and 10% rode the bus or metro, respectively, while 2% used taxi. Surprisingly, carsharing competes more with PCs and PT, than with taxis (although taxis and metro loose more passengers than other modes).

CBD concentrates a high number of carsharing trips. These trips are shorter and faster than the rest of the trips within the city. Residential area (in the outskirts) shows lower carsharing activity, despite being a priori more competitive than other modes, travel-time wise. This is probably due to the scarcity of available vehicles or to higher costs per trip.

When comparing to other modes, carsharing performs worse than PCs both in terms of time and cost. Differently, it clearly outperforms taxis in terms of cost, and outperforms buses, metro and walking in terms of travel time. The competitiveness of carsharing is highly determined by trip distance, becoming more attractive than other modes (travel-time wise) as trips become longer, although cost-wise it can lose some of its relative advantages. Final share depends greatly on the valuation of time by carsharing users.

One of the advantages of this model is the possibility of describing accurately different operational strategies for the carsharing system. Testing relocations and a very simple reservation system, we conclude that indeed these different ways of managing carsharing play a role on both the level of service and the profit of the carsharing company. Both policies revealed some synergies, showing the users satisfaction (percentage of trips satisfied) can be conciliated with the operator's profit. The results for a large car-fleet (1,500 vehicles) already show improvements with the introduction of these policies. Yet, the impact of these policies may be more significant for reduced fleets.

Future developments will focus on further disaggregating the modeling approach by simulating users' behavior through an activity-based approach. By disaggregating users, further knowledge can be gained on the travelers' specific options regarding carsharing and it will be possible to model particular features of those users such as past experience of services. The model will also evolve to explore free-float carsharing system configurations. This will require including private car agents that compete with carsharing vehicles for street parking locations. The introduction of more sophisticated reservation policies may also be included, which should be coordinated with the relocation policies to ensure the required stocks at a city area or station. Further exploration of pricing policies should also be done in future studies such as imposing a minimum fare to each client, potentially changing the type of satisfied trips, as well as incentivizing users to park at other areas by a dynamic pricing policy.