Shifts in perspective: Operational aspects in (non-) autonomous ride-pooling simulations

14 In this article, we simulate and evaluate the operational challenges of non-autonomous ride-15 pooling systems through driver shifts and breaks and compare their capacity and eﬃciency to 16 automated on-demand services. We introduce shift and break schedules and a new hub return 17 logic to perform the respective tasks at diﬀerent types of vehicle hubs. This way, currently oper-18 ating on-demand services are modelled more realistically and the eﬃciency gains of such services 19 through autonomous vehicles are quantiﬁed. 20 The results suggest that operational challenges substantially limit the ride-pooling capacity in 21 terms of served rides with a given number of vehicles. While results largely depend on the chosen 22 shift plan, the presented operational factors should be considered for the assessment of current 23 operational real-world services. The contribution of this study is threefold - from a technical 24 perspective, it is shown that the explicit simulation of operational constraints of current services 25 is crucial to assess ride-pooling services. From a policy perspective, the study shows the poten-26 tial of future autonomous services in direct comparison with non-autonomous services. Lastly, 27 the paper adds to the literature a realistic ride-pooling simulation use case based on observed 28 real-world demand and shift data. 29

shift plan, the presented operational factors should be considered for the assessment of current 23 operational real-world services. The contribution of this study is threefold -from a technical 24 perspective, it is shown that the explicit simulation of operational constraints of current services 25 is crucial to assess ride-pooling services. From a policy perspective, the study shows the poten-  In order to assess operational challenges, fleet and user behavior or implications on the transport 80 system of a new on-demand mobility system, a common approach is to simulate the proposed ser- 81 vice within a transport model. The minimum requirement for such simulations is a street network, 82 demand and supply and an assignment logic that matches requests and vehicles. In recent years, 83 numerous such simulation studies have been conducted in the field of on-demand mobility, often cruising or heading to a taxi rank to find new customers. Driver shifts were modeled in that a cab 120 becomes inactive as soon as the shift ends and returns either to the cab company or, in the case of an 121 independent cab, to a randomly chosen network node. The model did not include an actual dynamic 122 traffic assignment and assumed fixed travel times. Lokhandwala and Cai (2018) modeled taxi shifts 123 in New York City based on aggregated vehicle availability data. They compared the system with 124 driver shifts with an autonomous service where all vehicles are active during the entire simulation 125 time. They find a lower coverage of low-demand areas in the shift service due to the restricted fleet 126 size since vehicles tend to stay in areas with high demand. However, operational challenges that 127 come with driver shifts such as hub returns for breaks and shift changes were not modeled.
improvements may be achieved through rapid chargers and a battery swapping policy.

158
They optimized the charge scheduling by considering historic electricity price data in Tokyo and 159 also evaluated the vehicle-to-grid potential. By using two model-predictive control optimization 160 algorithms in parallel, one optimizing the transport service and one optimizing charging, charging 161 cost reductions of 10 % are found while service quality reduction is small. In summary, we find that existing simulation studies usually consider autonomous vehicles and do 164 not explicitly account for operational constraints in non-autonomous ride-pooling services. Chal-165 lenges of EVs have been studied more frequently.

167
In order to translate the learnings of the numerous simulation studies to today's non-autonomous 168 ride-pooling systems, we aim to consider the most relevant operational constraints that were learnt 169 from the real-world ride-pooling operator MOIA. For this purpose, we are able to use historical shift 170 and demand data of the service in Hamburg. This way we can investigate how well simulations 171 with autonomous vehicles or pseudo shifts (i.e. vehicles may be active for limited time windows but 172 without driver breaks and shift changeovers at hubs) can be used to describe current driver-based 173 services by comparing against an explicit simulation of driver shifts and breaks. In addition, by 174 direct comparison, this study quantifies the impact that future autonomous vehicles may have on 175 quality and efficiency of ride-pooling services.    It is an agent-based transport simulation framework that utilizes an iterative, co-evolutionary learn-191 ing approach in which each agent tries to maximize their daily score for a given plan of activities.  In our setup we use MATSim as a pure dynamic traffic-assignment model with a fixed demand.

200
In addition, the demand is not represented by full activity schedules but by individual ride-pooling 201 trips as observed by historic real-world MOIA ride requests. As we are only concerned with the 202 ride-pooling service in this study, we ignore other modes such as private cars, public transport or 203 walking and any user adaptation between iterations.  Once selected, the assignment of a customer to a vehicle is binding. If no vehicle is found that can 214 serve an incoming request, the request is rejected.

215
The pre-defined constraints highly impact the DRT system performance (Bischoff et al., 2017; 216 Zwick and Axhausen, 2020b). In order to ensure a good balance between service quality and sys-217 tem performance, we set the maximum wait time to 10 minutes and allow a maximum detour of  The DRT extension comes with a rebalancing algorithm developed by Bischoff and Maciejewski

221
(2020) to ensure that idle vehicles are sent to areas with high expected demand, which has shown 222 to improve the system capacity in terms of acceptance rate (Zwick and Axhausen, 2020a).

229
• A description of hubs and possible in-field break facilities. In-field break facilities can be, for 230 instance, existing parking lots at grocery stores or gas stations with optional charging plugs.

231
While shift starts and ends are fixed, breaks are defined more flexibly inside a given corridor 232 (earliest start time -latest end time) with a fixed duration. In our default setup, the typical break 233 duration is set to 30 minutes. To Each operational facility the type hub or in-field can be assigned.

234
In addition, each facility has a capacity for parked vehicles and, optionally, a number of chargers for 235 electric vehicles.

291
• Vehicles are charged to up to 90 % SoC. We outreach the optimal charging limit of 80 % to 292 avoid capacity shortages during high demand hours. Since the vehicle is already plugged in, 293 charging it up to 90 % is no additional operational effort.

294
• Vehicles can only be picked for a shift if their SoC is above 60 % to ensure that the power 295 lasts for the shift.   The street network is based on OpenStreetMap 5 data and MOIA's more than 10,000 virtual     In order to evaluate the impact of operational duties with non-autonomous ride-pooling services 346 compared to autonomous ride-pooling services, we apply three different service designs as shown in 347   Table 2. 348 In the autonomous service, the entire vehicle fleet is available to pick up and drop off passengers In the pseudo shifts scenario, one autonomous vehicle is generated for each driver shift of the 354 input shifts. These vehicles will have a limited service time that equals the planned shift start/end 355 times. As such, it mimics a service with driver shifts but without driver breaks and shift changeover 356 times including respective hub returns.

Conventional vs. electric fleets 365
After identifying the impacts of explicitly simulating shifts of the ride-pooling service, we add addi-   Table 3.  In addition, we quantify and evaluate the efficiency of the ride-pooling system using several  Table 4 shows the simulation results obtained by the three different scenarios defined in Section 4.2.1.

418
Obviously and as expected, a service running with fully autonomous vehicles is able to serve con-419 siderably more ride requests when compared to services with constrained vehicle availability due to 420 driver shifts and breaks. As such, the rejection rate increases from 1 % for the autonomous service 421 to 13 % and 20 % for the pseudo-shift and explicit-shift simulations, respectively. We can therefore 422 observe that, in terms of served/rejected rides, the pseudo-shift simulation is closer to the explicit 423 simulation of shifts, even though a significant difference persists which would lead to a more opti-424 mistic evaluation of the service.     to the shift services. With the pseudo-shift service we observe a similar occupancy but a lower 465 first evening peak, for which more shifts would be required to serve the entire demand. During 466 the second evening peak, many vehicles are either idle or relocating, which indicates a slight oversupply of shifts. In the explicit-shift service we observe a similar occupancy as with pseudo-shifts.

468
However, vehicles cannot transport passengers throughout their service times but relocations take 469 place to bring drivers to one of the three vehicle hubs for breaks or shift ends. In Section 5.3 we 470 analyze the potential to reduce these hub drives by providing more break and hub facilities in the city.

472
An overview of the sampled shifts including breaks in the explicit-shift scenario can be seen in 473 Figure 7. It becomes obvious that most shifts are active in the late evening/night hours, with a 474 peak of almost 300 simultaneously active shifts. However, it is also clear that with the given shift 475 plan, the high demand of the first peak shortly before 8:00 pm (see autonomous service in Figure 6) 476 cannot be fully served.  in terms of demand compared to the rest of the week. In a next step, we increase the number of hubs in the service area to evaluate the potential to 499 increase service capacity and efficiency through operational facilities. The results of these scenarios 500 are summarized in Table 6. It can be seen that the overall number of rides and rejections as well 1.50 respectively. The effects diminish with an increasing number of hubs as can be seen in Figure 9, 507 which indicates a kind of saturation effect. The overall impact of an increased number of hubs on 508 the ride-pooling service is, therefore, limited.   The application of shifts in the existing ride-pooling extension of MATSim can help to study existing 539 services more realistically and to account for operational challenges. At the same time, we show 540 the potential of current services to operate an even more efficient and resource-saving service with 541 autonomous vehicles. The example scenario with real-world requests and driver shifts applied here 542 shows that operational challenges have major impacts on the number of served rides and efficiency.

543
Due to multiple fictional parameters such as battery size, energy consumption, in-field break fa-  We present updates to current existing ride-pooling simulations to improve realism of results.

562
However, the shown approach still comes with limitations or unsolved questions. One issue is that