Beyond Traditional Public Transport: A Cost–Benefit Analysis of First and Last-Mile AV Solutions in Periurban Environment

Abstract

With the advent of Autonomous Vehicles (AV) technology, extensive research around the design of on-demand mobility systems powered by such vehicles is performed. An important part of these studies consists in the evaluation of the economic impact of such systems for involved stakeholders. In this work, a cost–benefit analysis (CBA) is applied to the introduction of AV services in Paris-Saclay, an intercommunity, south of Paris, simulated through MATSim, an agent-based model capable of capturing complex travel behaviors and dynamic traffic interactions. AVs would be implemented as a feeder service, first- and last-mile service to public transit, allowing intermodal trips for travelers. The system is designed to target the challenges of public transport accessibility in periurban areas and high private car use, which the AV feeder service is designed to mitigate. To our knowledge, this study is one of the first CBA analyses of an intermodal AV system relying on an agent-based simulation. The introduction of AV in a periurban environment would generate more pressure on the road network (0.8% to 1.7% increase in VKT for all modes, and significant congestion around train stations) but would improve traveler utilities. The utility gains from the new AV users benefiting from a more comfortable mode offsets the longer travel times from private car users. A Stop-Based routing service generates less congestion than a Door-to-Door routing service, but the access/egress time counterbalances this gain. Finally, in a periurban environment where on-demand AV feeder service would be added to reduce the access and egress cost of public transit, the social impact would be nuanced for travelers (over 99% of gains captured by the 10% of most benefiting agents), but externality would increase. This would benefit some travelers but would also involve additional congestion. In that case, a Stop-Based routing on a constrained network (e.g., existing bus network) significantly improves economic viability and reduces infrastructure costs and would be less impacting than a Door-to-Door service.

1. Introduction

The introduction of Autonomous Vehicles (AVs) will have an impact on the different agents of the mobility ecosystem. A few years ago, the introduction of Autonomous Vehicles was widely expected to generate substantial benefits across the board, with all economic agents presumed to gain from the development of this technology [1]. By example, for the users as they would not be in charge of the driving task when traveling by car, we might witness a decrease of their value of time [2,3]. Mobility operators also expect a decrease in operating costs due to the driver’s absence for each vehicle [4]. The cost reduction could lead to a decrease in service price for the customer but also to a change of service attributes such as frequency or routing [5]. As for the externalities, many benefits were expected such as a decrease in greenhouse gases due to a smoother driving but also an increase in the shared use of vehicles [5,6,7] or an increase in road safety [8]. However, this new technology also comes at a cost. As seen in Berlin [9], the introduction of on-demand services will have negative impact due to congestion effect. Indeed, the attractivity generates modal shift from public transit, walk and bike to the AV, leading to more people on the roads. AVs may have heterogenous impacts depending on the type of territories in which they are introduced. In urban areas, the demand is concentrated and the space is scarce. This phenomenon should diminish provided that the demand decreases and the space becomes more available. Such exclusive territory types as defined by the INSEE (the French National Institute of Statistics) may lack finetuning but they provide a good overview of demand density. Periurban territories are structured by several matters such as the predominance of private cars and public transport attractiveness. The AV challenge in such a territory is not to add traffic on the road but rather to massify flows and offers a complementary service to the public transit. In this work, a first- and last-mile use case of AVs in interaction with rail-based public transport is presented. If the periurban territories have been covered in the literature [3,10,11,12], the contribution of this study is the presentation of a feeder service evaluated through the prism of the cost–benefit analysis. The feeder service remains relatively ignored in the simulation literature. In [13] almost a hundred reviewed artocles, only five [3,14,15,16,17] addressed a feeder service use case. Amongst the five [3], only one provided an economic analysis. This work may be pursued further using a cost–benefit analysis for the economic assessment. The base case of this study is based on [18] which introduces AV taxis in the periurban area of the “Communauté d’agglomération Paris Saclay” (inter-township of Saclay, a Paris suburb).

2. Method

The methodology is structured as a two-stage simulation-to-evaluation framework. In the first stage, simulation is conducted using the MATSim agent-based mobility model, which enables to simulate mobility behavior and generate traffic and operational projections associated with the deployment of autonomous vehicles services. These forecasts subsequently inform the second stage—an economic evaluation—carried out using a cost–benefit analysis (CBA) approach, which is specifically adapted to the context of AV service implementation.

2.1. Simulation Model

2.1.1. MATSIM

MATSim is an agent-based mobility simulation framework designed to model large-scale transportation systems [19]. Its agent-based nature provides a robust foundation for simulating on-demand mobility services [13]. MATSim is activity-based, meaning each agent represents an individual with a daily schedule consisting of location- and time-specific activities. Agents travel between these activities, generating travel demand across the network. The model incorporates capacity constraints on both infrastructure and mobility services, enabling the realistic reproduction of congestion dynamics and resource competition.

2.1.2. EQASIM

Eqasim is a Discrete Mode-Choice (DMC) extension of MATSim that substitutes the standard co-evolutionary algorithm with a discrete choice modeling approach. Specifically, it employs a multinomial logit model—originally developed in 1974 [20]—to simulate agents’ travel behavior based on utility functions and travel-related outputs from previous iterations (e.g., cost, travel time, waiting time, and number of transfers). Agents iteratively adjust their mode choices until a stable equilibrium is reached. In contrast to the co-evolutionary algorithm in the standard MATSim framework, the DMC approach offers a theoretically grounded representation of behavior, rooted in decades of research in discrete choice modelling [21,22,23]. Notably, studies such as [24,25] have contributed to integrating discrete choice theory with mode choice modeling within agent-based simulation frameworks like MATSim. Furthermore, the DMC approach enhances computational efficiency by excluding infeasible alternatives prior to simulation, thereby accelerating convergence [26].

The mode choice is based on utility functions and constraints. Based on the utility function, the mode choice can either follow a multinomial model [26] or a maximum utility mode, which chooses the alternative with the highest utility.

For the multinomial model, if the tour (i.e., trip sequence starting at a location and ending at that same location) is considered, the probability to use the modal alternative $i$ for each trip is based on the following formula:

$$P\left(i\right)={\displaystyle \frac{exp\left({\hat{\mathrm{u}}}_i\right)}{{\sum }_{i\prime }exp\left({\hat{\mathrm{u}}}_{i\prime }\right)}}$$

(1)

• $P\left(i\right):$ probability to chose to alternative i
• ${\hat{\mathrm{u}}}_i:$ estimated utility of alternative i
• ${\sum }_{i\prime }\left({\hat{\mathrm{u}}}_{i\prime }\right):$ estimated utilities of all i alternatives

With the goal of assessing the impact of introducing AVs on traveler utilities, by comparing average utility between simulations with and without AVs, the multinomial model poses a few issues. When adding a mode alternative to the set of possibilities, the expected utility can lower depending on how the utilities with the new mode compare to the utilities of other modes. The scenarios were initially simulated using the multinomial mode choice model, which resulted in a counterintuitive loss in overall utility when introducing a new mode of transportation.

As for the maximum utility mode choice, the chosen tour alternative is the one presenting the best utility. If the best alternative is always selected, this can lead to unrealistic behavior (especially if the second-best alternative’s utility is extremely near to the best value). Consequently, we follow the approach proposed in [25] that consists in adding a pseudo-random error term to the utility functions. The extra term is a deterministic function of the traveler ID, the mode that is considered and the trip’s index in the list of the traveler’s trips. This allows consistent error terms not only between iterations of the same simulation but also between different simulations. The utility functions used are described in [26].

2.2. Simulation of On-Demand Mobility Systems

In this work, the simulation of on-demand mobility systems builds on the work performed in [18]. The AV-powered system is used only as a feeder service. No trips could be done exclusively with an AV but only within a public transit trip with the access and/or the egress travel operated through an AV. The feeder service system will be based as a semi-stop-based routing with either starting or destination point based on a rail station service (rail, subway and tram lines).

MATSim enables the simulation and analysis of Mobility-on-Demand (MoD) systems through its Demand-Responsive Transport (DRT) module [27]. This module facilitates the implementation of an on-demand vehicle fleet, allowing users to submit trip requests in real time. An underlying vehicle assignment algorithm is then employed to allocate available vehicles to fulfill the submitted requests. The strategy implemented by default tries to find the best way to insert he request in a vehicle’s plan. Trip insertions within the DRT module are evaluated based on the expected passenger arrival time and are subject to vehicle capacity constraints as well as predefined service level requirements. These include maximum allowable waiting and travel times, as well as a detour factor in the case of shared rides—defined as the difference between the unshared and shared travel distances—which serves as a proxy for the disutility of ridesharing. Additionally, the module incorporates a configurable pricing scheme that can be parameterized to include a cost per kilometer, a cost per unit of time, and a minimum fare per trip. Importantly, the cost per kilometer and per time are calculated based on the estimated unshared distance and unshared travel time, respectively. As a result, the fare remains constant regardless of whether a trip is shared or unshared; however, this also implies that shared rides are not financially incentivized relative to solo trips. Unless otherwise specified, all simulations presented hereafter rely on the default settings of the DRT module—covering fare structure, vehicle speed, fleet size, and dispatching strategies—as provided in the open-source implementation.

2.3. Cost Benefit Analysis

Cost–benefit analysis (CBA) is a systematic and quantitative method to evaluate transport projects. This approach aims to quantify “the change in the well-being of the individuals living in the society, and this involves calculating, in monetary terms, the magnitude of the potential […] gains compared with the opportunity costs of the resources.” The theorical background is well covered [28,29,30,31]. Several criticisms and works have appeared over the last several decades [32,33,34] resulting in the inclusion of dimensions such as equity within the evaluation framework. The method’s replicability and its capacity to assess the impact of public fund allocation have led governments to adopt this appraisal tool since its initial implementation in the United States under the Flood Control Act of 1936 [35]. Since then, various national authorities have developed methodological guidelines [36,37,38] and regularly update reference values to ensure consistency and relevance in evaluations.

In this study, we apply the French CBA guidelines [36], which we adapt by introducing a novel approach to calculating consumer surplus and by incorporating updated reference values specific to autonomous vehicles (AVs). The analysis is structured into three main categories—Consumer Surplus, Operator Profits, and Externalities—each corresponding to a distinct economic agent and comprising several sub-components (Figure 1: Cost–benefit analysis sub-components).

2.3.1. Consumer Surplus

Consumer surplus is defined as the monetized representation of the utility gains or losses experienced by mobility users. This measure is computed using the logsum approach, a welfare measure defined as “the log of the denominator of a logit choice probability, divided by the marginal utility of income, plus arbitrary constants.” [39]. In this study, utility is derived from the MATSim plan scores, in line with the methodology developed [40,41]. For each agent, the highest plan score is selected and converted into monetary units by applying the marginal utility of income. Given that agents may differ in their value of time and marginal utility of income, a time-equivalent approach is employed. Following [41] (pp. 54, 55), the aggregate change in consumer surplus, where individuals are indexed j = 1…J, is

$$\Delta CS={\sum }_{j=1}^J{\Delta m}_j$$

(2)

$\Delta CS$ is the overall Consumer Surplus change

${\Delta m}_j$ represents the monetary compensation required to offset the effects of the policy implementation—effectively capturing the change in consumer welfare between the two states.

The utility functions used to simulate user’s behaviors and the consumer surplus change can be found in Appendix A.2 and

Table A2. Utility functions and utility parameters for car, public transit, bike, walk and Mobility On Demand.

Car	αcar	1.35
	βtravelTime,car	−0.0667	[min−1]
Public Transport	αpt	0.0
	βnumberOfTransfers	−0.17
	βin VehicleTime	−0.017	[min−1]
	βtransferTime	−0.0484	[min−1]
	βaccessEgressTime	−0.0804	[min−1]
Bike	αbike	0.1
	βtravelTime,bike	−0.15	[min−1]
	βage,bike	−0.0496	[a−1]
Walk	αwalk	1.43
	βtravelTime,walk	−0.09	[min−1]
Others	βcost	−0.206	[EUR−1]
	λ	−0.4
	θaverageCrowflyDistance	40	[km]
Calibration	θparkingSearchPenalty	4	[min]
	θaccessEgressWalkTime	4	[min]

Source: [26].

. This method, combined with the use of an agent-based model allows to identify the gain and loss of every agent.

2.3.2. Operator Profits

The operator profit (π) is defined as the difference between total expenditures—which encompass infrastructure investment and maintenance, rolling stock investment, and operating expenses—and the revenues derived from usage fees, sponsorship and subsidies.

Currently, there is no consensus on the precise investment and maintenance costs associated with autonomous vehicle infrastructure. Estimates for upgrading infrastructure range broadly, from approximately USD 3000 per kilometer per year [42] to over two million dollars per kilometer as reported in [43] (p. 10). Similarly, maintenance and replacement costs remain uncertain due to limited empirical data from prior deployments and are thus subject to rough estimation. To address this uncertainty, the present study undertakes a sensitivity analysis whereby the investment cost ($I_{infra}$) varies from 50,000 to 250,000 euros/km and that maintenance $\left(M_{Infra}\right)$ is based on a yearly renovation rate ranging from 10% to 50% of the initial infrastructure investment. A 50% rate implies a full replacement cycle every two years. The total length of the infrastructure is noted as ${Km}_{Infra}$.

The initial fleet size is based on the estimated fleet size required to meet demand (${NB}_{Veh})$, the replacement investment is calculated based on a vehicle lifetime of 300,000 km. The unit cost of an autonomous vehicle is denoted as $P_{veh}$. The baseline vehicle cost is derived from the national average new car price in France, which stands at €26,000 [44], plus a +20% increase for electrification [45,46] and a +7500 euros markup for automation [47,48,49]. The tag price for an autonomous and electric shuttle is fixed at 128,000 euros [45].

Operating costs $\left(OC\right)$ for AVs, private cars and automated shuttles are based on [48]. Since the acquisition costs for AVs and automated shuttles are already accounted for within the investment cost section, the depreciation costs are deducted. However, the model excludes any costs associated with vehicle supervision, to account for this, a surcharge of €0.05 is applied per vehicle kilometer travelled (VKT).

Operator revenues consist of fare revenues (F) generated during the course of service provision.

Finally, the profit can be expressed as

$$\pi =F-\left(\left(I_{infra}\ast {Km}_{Infra}\right)\ast \left(1+M_{Infra}\right)+{NB}_{Veh}\ast P_{veh}+{\displaystyle \frac{VKT}{300000}}\ast P_{veh}+VKT\ast OC\right)$$

(3)

2.3.3. Externalities Impacts

Externalities of AV ($Ext$) are composed of five sub-components, following the segmentation framework proposed by [36]. The ${CO}_2$ emissions component accounts for the monetized valuation of ${CO}_2$-equivalent emissions generated during the vehicle’s use phase ${(Em}_{veh})$. Private car emissions are based on French Guidelines for a reference scenario [50] of 0.049 kg/kWh. Automation could reduce fuel consumption [6]: we consider a decrease of 5% in this paper. Ref. [51] (p. 58)shows that heavier vehicles will emit more ${CO}_2$ eq than cars, a +19% emission ratio increase for shuttles is retained. The 2030 value of ${CO}_2$ is fixed at 250 euros/t [52] ($P_{{CO}_2})$. For the local pollution, a +27% increase local emission (${LP}_{veh}$) between a conventional car and an autonomous shuttle has been retained based on the [4] methodology. The ${LP}_{Impact}$ is derived from [36]. The Lifecycle effects encompass the monetized valuation of ${CO}_2$-equivalent emissions associated with vehicle production and disposal. Emissions from the use phase are excluded, as they are addressed independently. Ref. [36] provides values for upstream and downstream effects for conventional vehicles. Ref. [53] produced a lifecycle analysis for AVs, in which Manufacturing and End-of-Life values are available ($V_{LCE})$. Noise externalities are incorporated as the disutility caused by road traffic, in line with the methodology outlined in [36]. Electrification is assumed to reduce noise impacts by 50%, based on findings from [54]. However, automation is not assumed to contribute to additional noise reduction, consistent with the recommendations in [55]; the resulting noise impact is denoted as ($N_{Impact})$. Road safety is also considered, as AVs are expected to contribute to a reduction in accident rates. A conservative assumption of a 20% decrease in crashes per vehicle-kilometer traveled (VKT) is adopted, informed by the most recent literature review presented in [56]. Value of a statistical life ($V_{SL})$ is set at 3339 million euros [36]. Accident rates are referenced from the French National Interministerial Road Safety Observatory (ONISR, 2019, p. 18) which reports a fatality rate of five deaths per billion VKT (${AR}_D)$.

Consequently, the externalities could be calculated using

$$Ext=VKT({Em}_{veh}\ast P_{{CO}_2}+{LP}_{veh}\ast {LP}_{Impact}+V_{LCE}+N_{Impact}+V_{SL}\ast {AR}_D)$$

(4)

2.3.4. Indicators for Scenario Comparisons

The comparison of scenarios is conducted across three distinct levels of analysis: the macroeconomic level, the economic agent level and the consumer level (Figure 2).

Comparison at the macroeconomic level is made using the social net present value, “equal to the sum of the change in social surplus or the sum of changes in willingness to pay and changes in resources” [36] (p. 28).

$${NPV}_s={\sum }_{t=0}^T{\delta }^t(B_t-C_t)$$

(5)

${NPV}_s$: Social net present value

${\delta }^t$: Discount rate of period t

$B_t$: Aggregated benefits from period t

$C_t$: Aggregated costs from period t

The second basis for comparison is the net present value relative to investment (NPV/Investment), which serves as an indicator of the social profitability of the project.

At the agent level, the evaluation is conducted through the analysis of Agent Surpluses, which capture the distributional effects of the service by quantifying the gains and losses incurred by each economic agent as a result of service provision and associated transactions.

$$aS={\sum }_{t=0}^T{\delta }^t(a{B}_t-{aC}_t)$$

(6)

$aS$: Agent Surplus

${aB}_t$: Agent level aggregated benefits from period t

${aC}_t$: Agent level aggregated costs from period t

Finally, the “Consumers” category represents the most heterogeneous group among the identified economic agents. The agent-based simulation model MATSim enables a detailed evaluation of consumer surplus through the use of agent-specific utility scores. Three key indicators are employed to assess how surplus is distributed across the population:

-

Winners versus Losers: This indicator captures the share of agents who experience an improvement (winners), a deterioration (losers), or no change (indifferent) in their utility compared to the baseline scenario, expressed as a percentage of the total agent population.

-

The 10% Measure: This metric evaluates the concentration of benefits by estimating the proportion of total positive consumer surplus accrued by the top 10% of agents who gain the most from the project, expressed as a percentage of aggregate gains.

-

Gini Index: A Gini coefficient is calculated based on the distribution of individual gains and losses, providing a summary measure of inequality in the welfare impacts among consumers.

3. Use Case

This appraisal methodology has been applied to Berlin [9]. The reasons behind the choice were 1. The use of MATSim, an open-access simulation model (allowing reproducible research) and 2. an application in an already covered territory. The following use case complies to the same criteria by using MATSim combined to an open-source extension that allows DMC models. The territory in itself is also covered in other simulation works [3,10], even if the mode choice model used is different [18]. The list of the data used to build the synthetic population has been added in the Appendix A.1 with their sources, formats and years of collect.

3.1. Territory

The «Communauté d’agglomération Paris-Saclay» (CPS) is an inter-township administrative structure in the Île-de-France region, 20 km south of Paris (Figure 3). It is home to over 450,000 people, making it one of the largest agglomerations in the Paris region.

Demographically, the CPS is home to a diverse population, with a mix of urban and suburban areas. The whole communauté might be defined as a periurban area for its position towards Paris (Figure 3). While some areas, such as Massy and Palaiseau, are already densely inhabited and see a lot of activities, territories south and north of Gif-sur-Yvette are developing themselves. Figure 4 shows that the CPS is a periurban territory in Île-de-France according to the definition of INSEE.

The local economy of the area is strong and diverse, with a range of industries represented, including technology, research and development, and services. CPS is home to a number of major international companies, as well as many smaller businesses, concentrated around the axis Massy–Palaiseau–Orsay–Les Ulis (Figure 5).

This jobs density also follows the RER B design, one of the major local train lines in Île-de-France. Massy also benefits from being on the route of the RER C, another major local train line. This work features the transport services expected in the area by 2030 that were assessed in [10] (see Figure 6).

3.2. Scenarios

The implementation of AVs services in Saclay requires the definition of AV services and also the scenario configuration for each service. As the evaluation step comes after the simulation step, the evaluation scenarios set will be a subset of the simulation scenarios set. The primary objective of this scenario design is to explore the potential of AVs as feeder services in a periurban context. The scenarios have been developed with the same methodology used to developed scenarios for [57], but with adaptations for the territory and use case. For [57], scenarios have been developed to comprehend the economic impact of on-demand autonomous vehicles in an urban territory. The CPS is a periurban area with scarcer public transit supply. If train lines provide a fast mode of transportation, mostly towards Paris center, the bus lines are known to be slow and overcrowded. If some parts of the discontent may be attributed to operational disorganization [58], the 2017 Transport Scheme for the CPS points out the need to reinforce bus frequency, especially during peak hours [59]. On-demand autonomous vehicles could provide a more reliable, comfortable and faster way to access the rail-based modes of transportation (trains, metro, subway). To test this assumption, a feeder service operated with on-demand autonomous vehicles service based on [13] nomenclature. AV taxis with varying capacities (4 seats or 8 seats) and different routings are introduced in the CPS territory. The routings will either be Stop-Based (working from a stations network) or Door-to-Door but both will link at least one of the segment points to a rail station (train, metro or subway) station. The stations of the Stop-Based network are based on the public transit stations of the CPS area. The agent traveling with the on-demand service will need to either have used a public transit mode to access or leave the station. The agents cannot use the on-demand service as a stand-alone service. The 400 vehicle fleet size is based on the first fleet size to reach a sub-five percent rejection rate for AV request with 100 vehicle steps. The parameters used for these simulations were the default ones. The ride must be done in a D2D scheme, with a 10 min maximum waiting time constraint, a fare of 0.3 euros per kilometer travelled and ridesharing enabled. Results from this simulation can be found in the 0.3 D2D scenario. This method remains around the 5% rejection rate threshold for the following simulations when the on-demand service’s attractivity were to be altered. The scenarios set which freed the service from fares allows to assume a fare integration between the public transit fare policy and the AV service. As for the evaluation scenarios, we retained the four scenarios which have the most important impact on the mobility landscape, with the four scenarios with free Avs. In the

Table 1. Simulations and evaluation sets.

Name	Ridesharing	Routing	Capacity	Fare (€/km)	Fleet Size	CBA
BC (Basecase)	N/A	N/A	N/A	N/A	N/A	✓
SAV D2D	Yes	D2D	4	0	400	✓
SAV SB	Yes	SB	4	0	400	✓
0.3 D2D	Yes	D2D	4	0.3	400
0.3 SB	Yes	SB	4	0.3	400
0.6 D2D	Yes	D2D	4	0.6	400
0.6 SB	Yes	SB	4	0.6	400
D2D Shuttles	Yes	D2D	8	0	400	✓
SB Shuttles	Yes	SB	8	0	400	✓

, the “✓” indicates which simulation scenario will be assessed through a Cost–Benefit Analysis.

4. Results

4.1. Simulations Results and Discussion

4.1.1. General Results

The purpose of this study was to simulate the potential impact of autonomous vehicles (AVs) on travel behavior and traffic congestion. Using a maximum utility mode choice model and the Eqasim simulation, we found that AVs are likely to attract passengers mostly from car travel and public transit modes, with a modal share of almost 2% in the most AV favorable scenario which is consistent with [18]. The overall decrease of public transit trips represented in the Figure 7 hides a result consistent with [18,60], where bus trips decrease but trips on train or metro increase thanks to a better attractiveness.

Overall VKT (private cars VKT + AVs VKT, see Figure 8) increased between +0.8% and 1.7%, which is consistent with the literature [13] but at a lower rate. We also found that travel time for car users unexpectedly increased between 7% and 34% depending on the scenario. This increase is difficultly explained by the VKT increase or distance travelled increase (<1% for all scenarios, see Figure 9). This increase of travel time may be due to congestion effects around train stations, where links experienced significant delays and travel times were multiplied by up to 30 times their initial values, which goes against the findings of [61]. This difference may be found in the fleet size deployed and the territory type. If considered as a public transit complementary tool, AV services may require modifications to the train station in order to mitigate this congestion. Despite this increase in travel time, the overall utility observed from the agent’s plan improved after the introduction of the AVs, as we will see in the Consumer Surplus section. The travel time increase is consistent with the literature [13] but the congestion increase might find three explanations. First, the simulation is mainly calibrated based on mode shares. Taking into account traffic flows and congestion in the calibration would require more data which is not currently available. Secondly, the Maximum Utility function makes every agent choose the modal alternative with the sum of the highest deterministic utility plus a random parameter. This parameter is based on a hash function which simulates a random parameter by giving an attribute to each modal alternative for each trip. It simulates a random attribution but with the ability to give the same parameters to each alternative between iterations but also simulations. It also may be at the origin of the additional congestion that we observed with the introduction of the AVs. This random attribution may be disturbed (i.e., the same random parameter may not be attributed to the same alternative) if a new mode if introduced. We tested our reference scenario with different seeds and observed similar impacts on congestion. The congestion is also impacted by random parameters influenced by the hash function.

4.1.2. A Focus on the AVs

AV performances are sensitive to the routing proposed, the fare level and also the capacity of the vehicles. The comparison of D2D and SB routing shows that (1) D2D is more attractive than SB (Figure 10) partly due to a better level of service (Figure 11); however in a higher AV fees scenario (AV2 to AV5), the average total travel time is equal between D2D and SB, but the access/egress team is higher for the SB routing, giving a slight advantage to the D2D in terms of utility (2). D2D is more resource consuming (Figure 12) due to greater demand (Figure 10).

The D2D consistently presents lower Average Vehicle Occupancy rate (AVO) than the SB service which can be caused by the routing which generates higher detours than SB, thereby pushing away agents from AV ridesharing; this result is consistent with [9]. Despite a higher rate of empty VKT that in the Berlin use case (Figure 13), the feeder system where vehicles wait for people to PT station helps to increase the AVO for both SB and D2D. The effect for D2D is straightforward as it corresponds to operate as a semi-SB routing, which increases the AVO. For the SB system, it may help the dispatching algorithm to absorb the demand forcing the vehicles towards station with high level of demand. The empty-VKT are significantly higher (26% to 42%) than for the Berlin use case (8–12%), 7% to 25% for [62], 3% to 13% for [57], ~5% for [63] and 4% for [64]. It might be due to the repositioning of vehicles.

The SB routings, which showed consistent higher AVO by two percentage points in the Berlin use case, presents in the CPS use case a lower AVO, except for the scenarios where shuttles were replacing AV cars in which the trend changes. The AVO seems to be highly dependent on the fare level and vehicle capacity. It shows that shared on-demand vehicles might benefit from economies of scale [57], as the lower the fare, the higher the demand and AVO. The vehicle capacity impact is a novelty in the literature as [3,9,65] showed that bigger than 4-seater vehicles were struggling to use additional seats. Here, we found that the replacement of cars (4-seaters) by shuttles (8-seaters) help to reduce empty-VKT, increase AVO, and to absorb additional demand, even reducing total travel time in some cases (AV1 versus AV9 in Figure 11). The AVO from all scenarios (comprised between 1.21 and 2.26) is significantly higher than most of the literature. Ref. [66] found from ~1.15 to 1.2, Ref. [62] from 1.03 to 1.84, except for [57] which found an AVO of 4 people per vehicle with an fixed attributed demand (a mode choice was not used). This is mainly due to the fact that the AVs are exclusively in interaction with rail-based public transports, increasing the likelihood of receiving many requests in short periods of time (when a train arrives).

Overall, our results suggest that feeder performance is sensitive to routing, fare level, and vehicle capacity, and that D2D and SB services have different strengths and weaknesses that should be carefully considered when designing AV systems. Except for the contribution on the combination of feeder service and vehicle size, all our results are consistent with the existing literature comparing D2D and SB services.

4.1.3. Consumer Surplus (CS)

Before we begin with consumer surplus, here is the impact of the introduction of Avs on utilities. The utilities increase with the number of modal alternatives, the routing and fares. Scenario V0 presents the lowest utility and is the only scenario without AVs. For the routing, as seen previously, the SB routing lowers agent’s utilities (Figure 14: Utilities), due to consistently longer travel times (Figure 15), The AV fares have an impact on AV attractiveness (Figure 7) and also on agent’s utilities (Figure 14: Utilities).

The CS for all four CBA scenarios presents a positive value. The D2D routing appears to increase the gain compared to SB routing, thanks to faster travel time (Figure 16). These results go against the literature which consider that AV introduction would lead to an decrease in Consumer Surplus [9]. The difference can be found in the modal choice model which was made based on different modal alternative constants, number of transfers, length trip or financial expenses and not different value of time. The use of equity value of time during the evaluation marginally altered the consumer surplus value but did not change the trend.

These results contrast with the findings reported in [67,68] where significant consumer gains were observed at the individual vehicle level. The difference can be largely attributed to the differing assumptions regarding the valuation of travel time savings. In [67], the value of time (VoT) was reduced by 75% to 93%, effectively offsetting the modest modal shift from rail to AVs. Similarly, Ref. [68] applied a 50% VoT reduction, which compensated for the associated increases in vehicle miles traveled (VMT), congestion, and travel time. (+10%), congestion and travel time. Conversely, Ref. [43] concluded that improvements in AV traffic flow could generate net positive consumer benefits, even under the assumption of a 2% annual increase in traffic volumes and without applying any VoT reduction for AV users. Likewise, [69] also found an increase in traffic. However, by considering only AV passengers they found a gain for users, which is consistent with the following equity analysis where specific users see their utility increase despite an overall loss.

The equity KPIs showed (Figure 17) that the introduction of AVs has a limited effect on most of the population. The repartition of winners/losers is almost even (19%/18% for AV0 and 18%/19% for AV1, AV8 and AV9) with a stable share of more than 60% of the population which remains indifferent to the new situation. The scope of the feeder service, which aims to complete an already existing offer, does not concern the entire CPS population. As such, more than 99% of the gains are concentrated within the 10% gaining the most (Figure 17). With such a concentration of gain, the analysis has been made for the 5% share and 1% share of people gaining the most for the D2D SAV scenario (the four scenarios have a very similar distribution of gains and losses). The results show that the service introduction really benefits a minority, with a respective 96% and 48% share of the gains. The very low GINI index might surprise with such an important concentration of gains but the gains slop and the share of indifferent agents show that the gains (which are often null) are evenly distributed.

4.1.4. Operator Profits

The repartition of the costs for the operator(s) are homogenous amongst the four scenarios (Figure 18). Over half of the operator(s) costs are composed of infrastructure investments, and a quarter by operational costs, with rolling stock investments representing less than a quarter of the remaining costs. The main difference is the operation of shuttles rather than cars, for which the rolling stock costs doubles when buying shuttles instead of cars, which can be explained by a much higher acquisition price for shuttle despite the fact that they are required to be replaced less frequently due to their lower VKT. Meanwhile the operational costs increase by less than 30% for the shuttle scenarios due to a higher operational cost per kilometer travelled and despite a lower overall VKT.

The implementation of a Stop-Based routing using the public transit station system allows studies to be pursued in which the AVs system would only operate on the bus network. The bus network in the CPS is composed of 286 km compared to the 952 km road network. As the cost of infrastructure is strictly proportional to its size both for the investments and the maintenance, such a diminution of the network size allows to divide the investment and maintenance costs of the infrastructure by more than three (Figure 19). In this situation, if the implementation of the service requires equipping the infrastructure, the Stop-Based routing seems to be a pragmatic solution.

4.1.5. Externalities

The externalities impact is negative for all sub-items (Figure 20). Despite a modal shift from private cars, the introduction of AVs induces additional VKT. The expected improved driving performances of the AVs are not sufficient to offset the additional traffic. It also seems that the use of the shuttle in D2D has a strong impact on all items of the externalities. The highest rate of emissions and accidents of a heavier vehicle plus the difference of VKT (+25% difference between the SB shuttle and the D2D shuttle scenario) explains such a gap.

The lifecycle impact is the more important, although only the vehicle impact has been taken into account. The infrastructure lifecycle and the supervision impacts should add more weight to the sub-item. The negative road safety impact can be surprising as AVs are considered as better drivers than humans. (an assumption that we kept), but the VKT increase offset this effect. In this mixed traffic, a 20% decrease of accident per kilometers travelled for AVs has been retained. A more optimistic assumption could be made but in different scenarios such as the full replacement of the private vehicle fleet. The ${CO}_2$ emissions are the least impacting item. The French energetic mix with a low carbon intensity decreases the variation magnitudes.

Despite the implementation of a feeder service, which aimed to reinforce the public transit attractivity by lowering the cost to access train station, all the externalities studied in this work increased.

These results diverge from the findings of previous studies [43,67,68], which generally report an increase in the positive external effects associated with the introduction of autonomous vehicles. However, Ref. [69], employing a more sophisticated diffusion model to assess the impacts of noise and air pollution, conducted an analysis of AV external cost pricing in Berlin and reached conclusions consistent with ours, despite differing assumptions regarding vehicle propulsion technologies. Similarly, the multi-criteria evaluation presented in [70] concluded that greenhouse gas (GHG) emissions are likely to rise with increased AV usage. In contrast, Ref. [43] projected a decrease in externalities, attributing approximately half of this improvement to enhanced road safety—benefits that are somewhat more pronounced than those considered in our scenarios. Overall, our results are aligned with [71] which states that “AV implementation requires systematic policy frameworks to ensure alignment with sustainability goals”.

4.2. Net Present Value

The net present value is positive for the D2D scenarios and slightly negative for the SB scenarios. The D2D scenarios presents a consumer surplus which offset the loss in financial costs and the impact of additional ${CO}_2$, local pollution, accidents and noise. The NPV/I for SAV, SB SAV, D2D shuttle, SB shuttle are, respectively, 0.27; −0.01; 0.09 and −0.15.

Figure 21, below, shows the impact of AVs introduction with and without the use of a restricted network to equip. The reduction of network length to equip for the Stop-Based scenario has an important impact on the NPV and also on the NPV/I. The four NPV/I are, respectively, 0.27; 1.67; 0.09; 0.97. The use of a constricted network for a Stop-Based service show promising results and should be encouraged.

5. Recommendations for Decision Makers

-

Prioritize Stop-Based routing with constrained networks: Given the significant cost savings in infrastructure investment and maintenance, and the improved net present value, decision-makers should strongly consider implementing Stop-Based AV feeder services that utilize or are restricted to existing public transport networks (e.g., bus networks). This approach offers a pragmatic solution for periurban areas.

-

Evaluate larger vehicle capacities: The findings suggest that 8-seater AV shuttles can offer better operational efficiency compared to 4-seater AV cars, leading to improved average vehicle occupancy and better demand absorption. This could be a key consideration for fleet design.

-

Address potential congestion hotspots: As AV feeder services can increase congestion around train stations, strategies to modify or manage these interchange points should be explored to ensure smooth integration with existing public transit.

-

Account for negative externalities: Be prepared for potential increases in VKT and associated externalities (CO₂, noise, road safety), even with the goal of supporting public transit. Further research into policies that could mitigate these impacts, such as incorporating induced demand into planning, is warranted.

-

Targeted service deployment: Recognize that the benefits of this type of AV feeder service may be concentrated among a minority of users. Policy development should consider how to broaden access to benefits or address potential equity concerns.

-

This study underscores that while on-demand AV feeder services can enhance traveler utility in periurban environments, their successful and sustainable implementation hinges on careful service design, particularly regarding routing strategies and network integration, to manage costs and mitigate negative externalities.

6. Conclusions

In conclusion, this scientific article addresses the issue of the low-level of service bus lines in periurban areas. These territories became auto-centric due to an inadequate mobility offer. This paper investigates the economic relevancy of on-demand autonomous vehicle mobility services working as a feeder for train lines in the periurban area of the Paris region. It brings three empirical contributions: the implementation of a feeder service, the economic evaluation of Autonomous Vehicles in a periurban area, and the consideration of a dedicated Stop-Based infrastructure. Despite the implementation of the feeder system to reinforce public transport attractivity, all scenarios present an increase in externalities, the service was not attractive enough to generate enough modal shift from car users. However, the limitation of the service limited the number of agents which might benefit from the service, but it has shown several advantages such an average vehicle occupancy higher than the literature. The Stop-Based scenario equipping only the bus network infrastructure presents promising results and suggests that such a service might be implemented in periurban territory. Overall, these findings contribute to the ongoing discourse on sustainable transport solutions and highlight the need for further research in this field, as shown by [72] the business model for AV have not been found. For the simulation limits, no induced demand (trips which would not have been made if the alternative did not exist) have been generated despite the fact that the integration of an attractive new service is often followed by new trips. The integration of induced demand would increase the impact on AV on congestion (and the related externalities); see [73] for similar results based on the (non-autonomous) on-demand ridesharing service impact on congestion. The induced demand is traditionally overlooked [74] and/or difficulty handled by transport economists as a few tools have been developed to assess its impacts. It would negatively affect conventional car users, while increasing GHG emissions. The congestion effect measured in this work is higher than might be expected. Electric vehicles were considered in the evaluation whereas no charging behavior has been incorporated into the simulation. Further research will evaluate the economic relevancy of the replacement of bus lines by on-demand autonomous vehicles.