Estimating savings in parking demand using shared vehicles for home-work commuting

In this paper, we focus on commuting between home and work, and investigate the possible gains from car-sharing and self-driving on the number of parking spots and vehicles required. Contrary to previous studies, we focus specifically on commuters who contribute a major portion of road traffic and parking demand, yet are not the typical target of car-sharing or even taxi services. A reason for this is that commuting flows are typically imbalanced and traffic demand is highly concentrated in rush hours. These factors make regular commuters a difficult target for current commercial car-sharing solutions, ride-sharing and taxi services; on the other hand, due to the large amount of traffic associated with commuting, even moderate gains in efficiency can have large benefits for cities. Additionally, as there are well established methods to estimate commuting flows from mobile phone usage data, our methodology can be easily applied to provide baseline estimates of possible efficiency gains, in contrast to more detailed case studies which would require accurate data on general purpose trips. We specifically focus on parking, as the decrease in parking needs is expected to be a clear positive outcome; we note that an estimate of decrease in the total number of cars is less meaningful, since each vehicle is expected to travel more, potentially giving rise to similar levels of congestion as private cars today. On the other hand, focusing on parking captures a potential benefit from smaller total fleet sizes.

We use data from mobile phone network logs to estimate home and work locations for a large sample of the population in Singapore and simulate their daily trips assuming private, shared and shared self-driving car usage. In the case of shared cars driven by their users, a main limiting factor for sharing is that the car needs to be parked at a comfortable walking distance from the origin and destination of their users. In the case of self-driving, this limitation is removed as the car can be allowed to travel longer distances to a parking spot or their next customer, at the expense of higher total vehicle miles traveled (VMT); we explore the implications of this trade-off by varying the distance self-driving cars are allowed to travel without a passenger. Furthermore, we repeat simulations with varying presumed adoption rates to estimate which rate is required to gain sizable benefits.

We note that a main limitation in our approach is that, beside taking note of any additional distance traveled, we do not explicitly model any effect on congestion as that would require a detailed microsimulation of traffic and assumptions on the actual performance of autonomous vehicles in real traffic conditions. Furthermore, while we expect that people’s behavior will change in response to availability of shared and self-driving vehicles, we do not aim to model this in our current work yet; we only assume that a certain share of commuting is made with shared vehicles.

Summarizing, the novel contribution of this paper is the development of a methodology that, starting from extensive real-world mobility traces, provides an accurate estimation of parking needs in a variety of sharing scenarios, including the effect of self-driving vehicles.

In accordance with the growing adoption of car sharing and the potential impact of self-driving, there is a significant research interest in assessing potential effects with regards to usage patterns, traffic and emissions. Survey-based methods find that car ownership among car-sharing users decreases significantly, up to 40%, depending on the study and the parameters used to correct for sampling effects [24, 16]. Still, drawing conclusions for the more wide-spread adoption of shared vehicles is not straightforward, since current car-sharing users are probably not a representative sample of the general population. Further studies try to estimate public attitude toward mobility options represented by self-driving vehicles and estimate the potential for adoption based on these [9, 10, 11]. Several studies then try to estimate the fleet size which could serve a certain population given some operational parameters, and the associated costs for travelers. Studies based on randomly generated trips find that about 10% – 15% of cars could serve mobility demands compared to private vehicles, with significantly reduced costs when compared to either privately owned cars or taxi rides [17, 18, 19]. This also prompted some concerns about the possibility of shared self-driving cars inducing significantly more traffic since they offer much cheaper and more convenient means of transportation [15, 10]. A more recent study based on realistic origin-destination flows obtained from travel surveys in Singapore and a theoretical derivation for fleet size finds that a fleet which has a size of about 38% of the number of privately owned vehicles can satisfy mobility demand with a bound of 15 minutes on passenger waiting times [25]. Further work in the central area of Singapore focused on the trade-off between fleet size and utilization using a detailed simulation of people’s mobility [26].

Concentrating on parking, a recent study has shown that by utilizing space much more efficiently, AVs have the potential to significantly reduce to spatial footprint of parking facilities [27]. Regarding shared vehicles, our work is most similar to studies by Zhang et al. [28, 29], who find that parking demand could be reduced by up to 90% for people switching to shared autonomous vehicle usage. The main difference is that the authors in [29] focus on the use of existing parking infrastructure, while in the current study we aim to calculate minimum parking requirements based only on basic assumption about commuter behavior, thus our methodology does not require any previous knowledge of available parking which can be difficult to obtain, especially on large scales [29, 1]. Furthermore, our simulation includes a significantly larger target population and more than 100 times larger fleet size (while the authors in [29] only consider 5% of the population of the city of Atlanta, i.e. about 22 thousand people in total, we consider a sample of over 1 million commuters in Singapore). A further recent study investigating the operational characteristics of a shared autonomous vehicle system in Lisbon, Portugal also considered potential reductions in parking needs with estimating that all on-street parking and a significant amount of off-street parking could be eliminated [30].

For the purpose of this work, we use call record detail records (CDRs) provided by Singtel, the largest mobile network operator in Singapore. The data includes records of several million subscribers for a period of eight weeks. The data includes a record when a user places or receives a call, or sends or receives a text message; data connections or handover information is not included. Each record includes the location of the antenna handling the event; with the high density of antennas in Singapore, spatial accuracy is estimated to be around a few hundred meters. Our dataset does not allow the reconstruction of individual trip data, but can be efficiently used to detect home and work locations of mobile phone users; this is considered standard and well-established practice [31, 32, 33].

Clustering people’s locations and identifying the main nighttime and daytime clusters result in our estimates on home and work locations. To ensure the quality of the results, we use the criteria that the clusters identified as work or home locations should have at least 20 records during working hours or during evenings and at night respectively. Furthermore, for the following work, we only include people whose identified home and work locations are at least 1 km distance apart (using simple geodesic distance) and thus are possible candidates for commuting by car. There are a total of 1,992,950 people in the dataset whose home and work locations could be reliably detected, and 1,066,504 of these fulfill the criteria that the two locations are more than 1 km apart. We show the obtained spatial distribution of home and work locations in Fig. S1 and the distribution of commute distances in Fig. S2 in the Supplementary Material. Furthermore, we display the difference between home and work locations in Fig. 1; as unbalanced flows in the morning and evening present a fundamental challenge to sharing cars and parking spaces, this will pose an inherent limit to the possible gains in efficiency from them. Since the granularity of detected locations is that of antennas in the network (i.e. each location corresponds to an antenna), we add a random noise of the magnitude of $166\,\mathrm{m}$ to users’ locations so that these will be less clustered. We note that the main assumption behind the current work is that the home and work locations obtained from this dataset will be a representative sample of people who would choose commuting by car.

In order to better estimate commute times, we calculate the travel times between people’s home and work locations based on real-world data as well. In the case of Singapore, average travel times between a set of road intersections were provided by the Land Transport Authority, measured at different times of the day and week. There are a total of 11,789 intersections, providing a good coverage of the area. For each user in the dataset, we located the closest intersection to their home and work location and use the travel time between these points as an estimate. We use estimates for times between 7AM and 8AM in the morning for travel from home to work and estimates for times between 4PM and 5PM as for travel from work to home. We display the distribution of these (as compiled for the list of people in the dataset) in Fig. S3. The travel time distributions have a mean of $1199\,\mathrm{s}$ and $1027\,\mathrm{s}$ respectively for the morning and afternoon case, while the medians are $1090\,\mathrm{s}$ and $983\,\mathrm{s}$. Note that these seem relatively low when comparing to typical values people spend by daily commuting. We speculate that this is the effect of Singapore’s highly restrictive policy on private car ownership, but highly car-centric road infrastructure, resulting in cars being a highly efficient means of transport for those who can afford them¹¹ 1 In 2010, there were about 780 thousand private cars in Singapore, a city with a population of about 5 million (3.2 million citizens and 1.8 million permanent residents and visitors), giving a ratio of only 154 cars per 1000 population (241 per 1000 when only counting citizens); this is significantly lower than the value of 500 – 800 found in other developed countries. This is mainly achieved by the government setting quotas on newly registered vehicles and auctioning spots to potential buyers. In October 2017, as the result of the auctioning, the levy to register a new car for a 10-year period was about S$41,000 (US$31,000)..

In this work, we focus on a set of commuters as described in the previous section and estimate the number of parking spaces and vehicles needed to satisfy their mobility demand. In the following, we denote the number of users in our dataset by $N_{U}$, the number of required parking spaces by $N_{P}$, and the number of required cars by $N_{C}$. Furthermore, we measure the total distance traveled by commuters, denoted by $d_{\mathrm{tot}}$. We employ several scenarios for their commuting habits and compare the results and quantify the improvement due to sharing vehicles and self-driving:

We run simulations to determine the demand for parking spaces and the opportunities for sharing in scenarios #2 – #4 and compare results to the constant values in the case of scenario #1. We show the simulation algorithm in the case of private vehicles (#2) as Algorithm 1 and for shared or self-driving vehicles (#3 or #4) as Algorithm 2. In both cases, the input is a set of trips (generated from the home and work locations) and potentially a set of free parking spaces and available shared vehicles (only for Algorithm 2).

In the case of private vehicles (#2) in Algorithm 1, we start the simulation with assuming that everyone has a parking spot at their home location and do not assume any more parking spaces at work locations yet. In accordance with this, we set the total number of parking spots in the city to be $N_{P}=N_{C}=N_{U}$, and the set of available parking ($L_{P}$) is empty. At first, as people leave home in the morning, their home parking spots become available for other to use. We keep track of free parking spots in the list $L_{P}$ (employing a spatial index for efficient searches later). When someone arrives at their work location, they search for free parking spots in $L_{P}$ within a $r_{\textrm{max}}$ radius. If such a parking spot is found (i.e., someone’s home spot that is unused), it can be occupied; in case of more than one parking available within the search radius, we always select the closest one. If there are no free parking spots close to an arriving person’s work location, we add one more which they occupy. Thus, we increase the number $N_{P}$ of parking spots by one. We can assume that this parking spot was there all the time, but no one needed it yet. When moving people back home, we repeat the same procedure: everyone takes their car from where they parked it in the morning (adding that spot to $L_{P}$), drives home and tries to find a free spot. Since leaving from work and arriving at home happens stochastically, it can happen that a person finds their “home” spot occupied. In this case, they again search for the closest available alternative spot, or if none is found within an $r_{\textrm{max}}$ radius, we again add a further parking space to the city, again increasing $N_{P}$. Depending on the timing of commutes, people leaving and arriving at either their home or work locations happens interleaved, meaning that not all parking becomes available. This way, the timing of trips plays a significant role in the result as well. See the next subsection for assigning trip timings.

In the case of car-sharing (#3) and self-driving vehicles (#4), as displayed in Algorithm 2, we not only maintain a list of free parking spots ($L_{P}$), but also of available vehicles, again including the coordinates where they are parked ($L_{C}$). When someone starts a trip, we first search in the list of available cars ($L_{C}$), and if a suitable car $c$ is found within $r_{\textrm{max}}$ distance of the origin of the trip, we select the closest such car $c$, remove it from $L_{C}$ and add its location to $L_{P}$ as a free parking spot. On the other hand, if no such cars are found, we add one more car to the system at the trip origin location, increasing the total number of cars $N_{C}$. We also increase the number of parking spaces $N_{P}$ as we assume the newly added car to have been parked in that location, which again becomes a free parking spot and is added to $L_{P}$. In this case, at the beginning of the simulation, we do not place any parking spaces or cars in the system, i.e., we start with $N_{P}=N_{C}=0$ and the $L_{P}$ and $L_{C}$ lists being empty. This way, during the course of the simulation, only the necessary number of vehicles and parking spaces are added. In this case, we also take into account the extra trip time due to traveling between the origin or destination of a trip and the parking location. This quantity can become significant for self-driving vehicles, especially if we consider a relatively larger $r_{\textrm{max}}$ radius.

In all cases, it is assumed that the agents are able to find the closest available parking and closest available car when using shared cars. While searching for parking is complex problem by itself [34, 35], our assumption basically means that all drivers use an efficient navigation system which also receives real-time updates on parking availability. Implementing such a system is possible already with today’s technology; also, we expect that shared autonomous vehicles will be able to communicate with a “controller” that directs them to the closest available parking.

Since the actual timing of morning and afternoon trips can affect the results – see subsection below –, in case of both algorithms we run the simulation for multiple days in a row with different, randomly generated trip timings each day. It is important to note that we start with empty $L_{P}$ and $L_{C}$ lists only on the first day of the simulations; on subsequent days, we start the simulation with the $L_{P}$ and $L_{C}$ lists and $N_{P}$ and $N_{C}$ values obtained by the end of the previous day. This way, we are testing if the same number of parking, vehicles and actual spatial configuration of parking is sufficient to satisfy the travel demand on the next day, or if further parking and cars need to be added to the system to account for a different sequence of trips. In the experiments reported below, we ran the simulation for $n_{d}=30$ days in each case.

For all scenarios #2–#4, the main parameter that will affect the results is the bound $r_{\textrm{max}}$ on the distance between trip origin and destination and the sought parking spot. In the case of scenarios #2 and #3, this bounds represents the distance people are willing to walk from their parking location and their destination. In case of shared vehicles (scenarios #3 and #4), $r_{\textrm{max}}$ is also the upper bound to the distance to the closest available shared car. For scenario #4, this is the distance that self-driving cars are allowed to travel without a passenger before the start or after the end of a trip to reach their parking location.

The second main parameter in the simulation is the method used to generate commute timings. This is represented by a commute window of length $t_{W}$; all trips are assumed to start inside this window (see below). Furthermore, results are affected by the penetration ratio of shared mobility, i.e., the number of people who use shared or self-driving cars among the group of commuters considered.

As we commented above, a main determinant on the possible efficiency gains is the sequence and timing of individual trips, since it determines if a specific shared vehicle or parking spot is available at the time when a commuter would want to start or finish their journey. Since timings of individual trips on a large scale are hard to obtain, and are still subject to daily variations, we generate these randomly for each person in the simulation. To test for variations in different realizations, we run the simulations for $n_{d}=30$ consequtive days and then repeat the whole process 100 times for better stochastic accuracy. Each day in a single simulation run presents a different realization of random trip start times. Running the simulation for several days helps establish the robustness of spatial configuration of parking and vehicles, while repeating the simulation allows us to test for statistical variations. We find that random variations are very small: standard deviation are less than $1\%$ in all cases, and less than $0.1\%$ in most cases. We report the effect of these variations in the Supplementary Material, in Figs. S5, S7, S8 and S9 and in Table S1²² 2 Available as a separate download https://www.dropbox.com/s/s410s74oxh28h07/parkefficiency_si_table_S1.ods?dl=0.

For the main results of the current work, we generate the start time of each individual trip uniformly at random in a time window of length $t_{W}=1\,\mathrm{hour}$, from 7AM to 8AM for morning commutes and between 4PM and 5PM for afternoon commutes. Beside the main results, we further explore several options for $t_{W}$ and also an option where we generate trip start times based on a dataset of public transportation usage in Singapore.

The main result of the presented estimation methodology is the number of cars and city-wide parking spaces needed to cope with the travel demand. We display the required number of parking spots in the different scenarios as a function of $r_{\textrm{max}}$ in Fig. 2. We see that for reasonably small values of $r_{\textrm{max}}$ (i.e., between $100\,\mathrm{m}$ and $500\,\mathrm{m}$), around 23% of parking spaces can be saved by using private cars and sharing parking spaces, as in scenario #2 (we note that a real city will be between #1 and #2, but we expect that most people still have reserved parking). If we introduce shared cars as well (scenario #3), the reduction in parking demand approaches 40%. Just comparing the case of private and shared cars (#2 and #3), we see that introducing shared cars saves around 20% of parking spaces from an already highly optimized system with shared parking (see the inset in the right panel of Fig. 2).

For private or shared cars driven by their users, the $r_{\textrm{max}}$ distance is essentially the maximum distance people are willing to walk from their parking spot to their final destination. In our main simulations, we considered $r_{\textrm{max}}$ between $100\,\mathrm{m}$ to $1.5\,\mathrm{km}$ for scenarios #2 and #3 and $r_{\textrm{max}}$ up to $10\,\mathrm{km}$ for scenario #4 as shown in Fig. 2. We believe that actually only the smallest values of $r_{\textrm{max}}$ are realistic for walking; while previous studies for transit usage typically consider $500\,\mathrm{m}$ as an acceptable walking distance [36, 37], empirical studies on parking usually reveal distance to the destination as the main factor when deciding where to park (more important than price, time spent searching for parking, etc.) [38, 39, 7]. For this reason, only the leftmost values in Fig. 2 can be considered significant in these cases. We present larger values mainly for comparison between Algorithms 1 and 2.

On the other hand, in the case of self-driving vehicles, much larger $r_{\textrm{max}}$ values are feasible. The results of our analysis show that, already with $r_{\textrm{max}}=1.5\,\mathrm{km}$, parking needs reduction is above 50% compared to scenario #1, and around 33% when compared to scenario #2 with $r_{\textrm{max}}=500\,\mathrm{m}$, a realistic upper bound on walking. When considering larger values of $r_{\mathrm{max}}$ up to $10\,\mathrm{km}$, savings in parking demand increase up to 63% (50% compared to scenario #2). However, these savings would come at the expense of increased traffic, as discussed in the next section. We note that actual walking or extra travel distances can be smaller than $r_{\textrm{max}}$, which only specifies the upper bound. In Fig. S6 in the Supplementary Material, we display the distribution of actual walking distances in a few typical cases of $r_{\textrm{max}}$. For most trips, we find that the actual walking distance is much lower than the $r_{\textrm{max}}$ parameter used for the simulation; on the other hand, a relatively small chance of having to walk excess distances could be still highly discouraging for potential users.

The fleet size resulting from our estimations is reported in Fig. 3; we see that we can achieve about 30% reduction with shared cars and small $r_{\mathrm{max}}$ values suitable for walking, while these gain increase to over 45% for larger $r_{\mathrm{max}}$ values achievable with self-driving.

So far, we have presented results for a limited set of parameters modeling commuting in Singapore. To estimate the robustness of the presented results to changes in the simulation parameters, we repeated the experiments for several different parameter combinations.

First, we considered different penetration rates of shared mobility, repeating the simulations for scenario #3/#4 while varying the number of commuters, out of the total number of commuters considered, who use shared vehicles. The results are reported in Fig. 4. We see that the possible relative gains (in terms of parking spaces) barely change when at least 25% (i.e. about 267,000) of people participate in a shared mobility scheme; a smaller sample of only 10% of people (107,000 people) would instead result in noticeably smaller gains (about 5% difference) when using a radius of $r_{\mathrm{max}}=300\,\mathrm{m}$, which we consider a reasonable value for walking. On the other hand, for radii of at least $500\,\mathrm{m}$, the gains in parking efficiency are only slightly worse even in this case, suggesting that for self-driving cars, a relatively low adoption rate would already bring significant benefits. We note that actual gains might be even better as a smaller fleet could be occupied to a larger degree during the day outside commuting hours, performing taxi-like service as well.

Furthermore, we repeated the simulations using Algorithm 2 for several different commute lengths of the time window $t_{W}$. These results, reported in Fig. 5, indicate that $t_{W}$ is indeed an important parameter as a commute windows value below one hour significantly decreases sharing opportunities. On the other hand, higher values of the commute window will only add moderate reductions in parking needs. Furthermore, using travel timings generated from the transit data does not alter the results significantly. We note that the one hour commute window used to obtain the main results of this paper can be still considered a conservative estimate (e.g. the activity peaks seen in transit data seem significantly longer as we show in Fig. S4 in the Supplementary Material).

We compute a further measure to characterize the inherent inefficiency due to unbalanced commute flows. This bound is obtained by applying the same model under the assumption of instanstaneouos travel (i.e. all trip times are set to zero, but trips are processed in a random order). Since no vehicles are in transit to their destination in this model, the result of this process can be intended as a measure of the inherent inefficiency in parking needs that arises from the mere spatial distribution of trip origins and destinations. This is displayed as the black line in Fig. 5; we see that there is about 20% – 30% difference between the main results (considering $t_{W}=1$ hour) and this theoretical limit.

Self-driving cars would allow parking farther away from the origin or destination of a trip. We have seen that this would further reduce the number of parking spaces required (Figs. 2 and 3). This benefit nevertheless comes at a price of increased traffic, which we quantify here as an increase in the total vehicle miles traveled (VMT). In this section, we present results for estimating this extra VMT to be able to find a good trade-off between less parking (and cars) and more traffic. To obtain this, during the course of the simulation, we recorded the distances between the origin or destination of a trip and the parking spot used; we sum these distances and compare them to the total distance that people have to travel between their home and work locations. We present the relative extra distance traveled as a function of the previously established reduction in parking demand and also in a slightly modified case where the maximum number of parking spots is capped at a number determined from previous simulation runs (see the Supplementary Material for more explanation on this). We see that using self-driving vehicles, achieving about 50% reduction in parking space requirement over scenario #1 will only add about 2% extra VMT, while further gains come at the cost of potentially significantly more vehicle travel. How this extra travel affects congestion will depend on how self-driving cars perform in real-world traffic, i.e., whether they can compensate increased traffic by being more efficient.

We note that allowing longer distances (and more traffic) can correspond to a scenario where instead of on-site parking garages, operators of self-driving fleets have depots placed in strategic locations in the city. Assuming a fleet of interchangeable vehicles (or a few vehicle types), these depots can be highly efficient, have a much smaller total area than traditional parking garages [27]. This would present further reductions in the footprint of parking in cities, introducing both opportunities and challenges in re-using existing parking facilities.