Penn Station Can Handle the Load:
New York is Ready for Through-Running
January 15, 2025
Top Right: Penn Station Track Map (courtesy ETA), Bottom Right: New Penn Station LIRR Concourse (courtesy ETA, Blair Lorenzo)
Executive Summary
The recently funded Gateway Program, which centers around building two new rail tunnels under the Hudson to connect New Jersey to Manhattan's Pennsylvania Station, will enable a massive increase in rail passenger traffic. To handle this growth, ETA has continually advocated for through-running trains from New Jersey to Queens and beyond. Through-running not only makes better use of the region's existing infrastructure, it would also create new services that would transform transportation across the region.
In a recent study, however, Amtrak has claimed that through-running is impossible in the current Penn Station footprint, and that the region instead needs to spend $16.7 billion to demolish a block of Midtown Manhattan and build Penn Expansion. This project adds 7–12 stub-end tracks, maintaining an antiquated operational status quo. ETA's own analysis shows that Amtrak's report is simply incorrect. It builds on faulty assumptions and poor understanding of how through-running rail systems operate across the world. Even the separate $7 billion Penn Reconstruction project to rebuild the track and concourse levels within the station’s footprint, while perhaps desirable for passenger experience, is unnecessary for through-running in our analysis.
ETA's analysis, based on an examination of best practices from peer cities across the globe, shows that today's Penn Station can handle the ridership that Gateway will bring if the region simply changes the way it operates its trains—and it can do so without spending tens of billions of taxpayer dollars.
Background
The Gateway Project’s centerpiece, the new Hudson River Tunnel from New Jersey to New York Penn Station, is now fully funded and under construction. By doubling the amount of tracks under the river, the new tunnel will unlock new services and high traffic, necessitating considerable changes to Penn Station operations.
Through-running, a type of rail service where trains continue in the same direction across the region rather than turning back the way they came, should form the foundation of Greater New York's service pattern. As ETA has elaborated before, through-running would not only improve the network’s efficiency and resiliency, it would introduce travel possibilities that are not practical today. In this document, we model Penn’s capacity, with its current platforms and egress points, to move 48 trains per hour (tph) in each direction across the Hudson, which Gateway will unlock. Our calculations show this is doable without any expansion of the station’s footprint or a major reconstruction of the platform access, with capacity to spare.
Unfortunately, Amtrak, the MTA, and NJ Transit are not modifying their operating practices and embracing through-running. Instead, Amtrak is dead-set on trying to accommodate this new traffic in the most expensive way possible: by demolishing an entire block of Midtown to add more tracks and platforms to Penn Station. This extremely expensive project, named Penn Expansion, is currently projected to cost $16.7 billion, even more than the new tunnel itself, which took decades to plan and fund. Remarkably, that number is only the initial pre-construction cost estimate; the same initial estimate for East Side Access was $4.3 billion, which ultimately ballooned to $11.6 billion two decades later.
The case for this massive expansion rests on dubious assumptions about the existing station’s maximum capacity. Amtrak recently released a report arguing for the station expansion, which relies on claims about how fast trains can enter and exit the station. Amtrak asserts trains would have to remain idle at the station (called the dwell time) for 15–22 minutes, depending on train type. In contrast, according to our analyses of both similar international systems and our own modeling, we find that through trains only need about 3 minutes of dwell time, while terminating trains can turn in 10, both considerably less than Amtrak’s estimate.
With the correct dwell and turnback times, no expansion of Penn Station’s footprint is needed. The design capacity of the Gateway Project can be accommodated on the existing tracks and platforms, rendering new tracks unnecessary. Although there is an ongoing need to improve the passenger experience, as with the $7 billion Penn Reconstruction project to improve the station’s public facilities and improve circulation on the narrow platforms, none of that is necessary for Penn Station to handle capacity. With through-running, Penn Station’s existing platforms are more than capable of meeting future demand.
Regrettably, Amtrak's conclusion that Penn Expansion is necessary is both inaccurate and inconsistent with international best practices. It paints every alternative to Penn Expansion as both extravagant and disruptive based on false assumptions about passengers’ ability to get on and off trains. Amtrak’s review of world regional rail systems contains fundamental errors about both the history and current operations in Paris, London, and Munich and as such draws inaccurate conclusions. Lastly, by not engaging with any planning post-Gateway, the report suggests investments that are, at best, penny-wise and pound-foolish. For instance, it assumes that NJ Transit’s Erie lines and Metro-North’s west-of-Hudson service will remain unelectrified, despite a cost to electrify of about $750 million, using the busy Southern Transcon as a benchmark [FN1], in order to justify spending $16.7 billion to expand Penn Station.
Amtrak’s report is fundamentally flawed and should be disregarded. Instead of a genuine inquiry designed to identify the solution that makes the most of New York’s infrastructure, it instead seeks to preserve the operational status quo, even at a cost of nearly $17 billion. It is of little value for future infrastructure planning at Penn Station or on the Northeast Corridor. Whatever processes led to this report’s conclusions are fundamentally flawed. There’s a possible future at Penn Station of short dwell times, integration between surface and tunnel infrastructure works, and integration between the different railroads leading to through-service with its numerous benefits to passengers. Amtrak, NJ Transit, and the MTA need to work together to make that future a reality.
The Capacity Issue
Today, Penn Station is the only central station for many commuter lines in the New York area. As such, it is already the busiest train station in the country. The Gateway Program will only increase this usage, as new train capacity and more reliable service will attract travellers who currently either ride buses or drive into Manhattan. These conditions create a challenging operating environment that requires careful planning and procurement. However, the claim made in the recently released 2024 report that no through-running is possible using today’s platforms is simply incorrect. Thoughtful management of the existing station infrastructure and modified operations would enable the existing station to accommodate all the additional travel volume that Gateway will bring to the rails.
Egress and Dwell Times
A large part of Amtrak’s argument rests on egress and dwell times. However, the organization’s own findings from a decade ago cast doubt on its own claims. In 2014, Amtrak released a report that concluded that through-running at Penn Station would not sufficiently increase capacity for the Gateway project. Interestingly, the analysis stated that it was safe to assume that dispatchers could cut through-running train slot times to an average of 12 minutes with no platform reconstruction. Moreover, that report predates the Moynihan Station project, which substantially hastened passenger flow from platforms by adding stairwells and escalators. Right away, the claim in the 2024 report that no through-running is possible beyond the few Amtrak trains already doing so seems dubious.
The 2014 report also claims that Amtrak budgets 10 minutes for revenue-to-non-revenue trains that through-run to Sunnyside Yard and make a crew change during the stop. That time includes a train sweep to ensure all passengers have disembarked. While perhaps not the critical path in determining train dwell times, the time taken by that process can be estimated to be, at worst, the same amount of time it would take to board new passengers, meaning revenue-to-revenue through trains should take little to no time more than revenue-to-non-revenue ones.
In fact, it is commonplace for today’s through-running Amtrak trains to dwell for less than 10 minutes, and dwell times as short as 7 minutes have been observed. For example, here is train 146 on Sunday, May 12, 2024, when it would have been quite full:
As we’ll show in the simulation below, passengers on peak commuter trains can get off the train, exit the platform, and reach the concourse within about 5 minutes of the train’s arrival, including stragglers. In fact, there is enough capacity to clear a full train within 2 minutes, but many passengers prefer exits near 7th Avenue (towards the densest parts of Midtown) over exits further west, and the last few percent of passengers leave more slowly. By 2 minutes after arrival, new passengers can begin to board, even as the final passengers are alighting on the platform. By 3 minutes after arrival, the platform is not entirely clear, but its passenger crowding is sufficiently low that the train can safely leave.
Overtakes
Today, Amtrak schedules the Acela, which travels express between Washington and Boston, to overtake the slower, local Northeast Regional at Penn Station. The organization claims that this requires scheduled dwells of 30 minutes for Northeast Regional trains. This is a poor practice that makes inefficient use of busy, expensive infrastructure. Scheduled overtakes should occur at outlying stations such as New Haven or Trenton where there are plenty of tracks. At New Haven, for example, Northeast Regionals often already wait for connections to and from Hartford. Since the Acela and Northeast Regional running time between Penn Station and New Haven should be nearly identical, there is no need for trains to meet in Manhattan, where space and time is precious.
The Consultants’ Gantt Chart
Amtrak and its consultants have overestimated the amount of time required for both passenger circulation and crew operations at Penn Station. This miscalculation goes back further than Amtrak’s report, to a different, unreleased report commissioned by New Jersey Transit for the Penn Station Capacity Improvement project. Since commuter rail trains are the largest capacity constraint for Penn Station, it is worth considering that report’s conclusions.
According to NJ Transit’s consultants, terminating trains have a minimum turn time of 18 minutes: 16 minutes for the turn plus 2 minutes for recovery. This is based on two critical paths: one for alighting and boarding passengers, taking 16 minutes, and another concurrent one for crew change and required brake and signaling tests, taking 13 minutes.
The Gantt chart from NJ Transit’s unreleased report commissioned for the Penn Station Capacity Improvement project, describing the critical paths for passengers and crew.
In truth, both figures are excessive. Regarding the crew path, some American commuter trains turn in 10 minutes when needed, typically at outer terminals such as New Haven or Worcester, versus major central business district ones like Penn or Grand Central. We believe that the minimum time required for the crew change and tests for terminating trains is 8 minutes, with 10 scheduled after padding. The minimum for passenger flow, including on through trains, is 2.5 minutes, with 3 scheduled. The numbers in the Gantt chart’s critical passenger path are excessively conservative.
The Passenger Path
The Gantt chart assumes that passengers take 7 minutes to alight and 5 to board, with 3 minutes between the two processes and a final minute for door closing and signal reaction time. All of these numbers are excessive, on several levels:
As our through-running modeling notes below, passengers on a full rush-hour commuter train will clear the entire platform in 5 minutes. During the last 2 minutes of this process, the platform is nearly empty and the train can leave even without a completely clear platform, as is routine on the subway.
The peaks for boarding and alighting occur at different times of day. During the morning peak, there aren’t many boarding passengers, and they can go onto the platform before passengers have finished going upstairs. Our model excludes a proportion of the egress capacity to account for concurrent boarding passengers, and assumes many are waiting on the platform when the trains arrive, as the crowding level on the platform is high but well within pedestrian crowding limits.
It takes merely a few seconds for doors to close and engineers to acknowledge a departure signal, not an entire minute. For one, the entire dwell time at high platform suburban stations, which includes not just this process but also door opening and passenger boarding or alighting, is 20–45 seconds.
With the correct numbers, as we document more extensively below, this passenger flow path can be reduced to 3 minutes on rush-hour through trains, potentially even less.
The Crew Path
The Gantt chart assumes a succession of required elements for the crew change on terminating trains, plus required safety inspections including a brake test and a signaling test. These sum to 12 minutes, plus the final minute of door closing and reaction time from the passenger path. This path, like the passenger path, is excessive.
On a terminating train, the brake, signaling, and positive train control (PTC) tests are required, as it is changing direction. These, however, only take 5 minutes. The rest of the time is taken by closeup and setup of the train engineer’s position (a total of 4 minutes) and then a 3-minute engineer and conductor job briefing. The job briefing does not need to be done on board a dwelling train, and the train engineer does not need 3 minutes to close up their position. In fact, PATH, which is also regulated by the Federal Railroad Administration (FRA) and thus must conduct the same type of brake test, turns at its Hoboken and 33rd Street terminals in 4 minutes, including time taken for the crew to walk from one end of the train to the other. Clearly, the amount of time it takes to perform closeup is far faster than what the Gantt chart asserts is the minimum.
On through-running SEPTA trains, the brake, signaling, and PTC tests are not required as the train is proceeding in the same direction, and the crew change is done during a 3-minute dwell time. We therefore propose a 3-minute dwell time on through trains as with the passenger path. On terminating trains, the 3-minute dwell time already includes the required elements in the Gantt chart except for the brake, signaling, and PTC tests, which add at most 5 minutes. Thus, trains can turn in a minimum of 8 minutes, close to what they already achieve in practice at some outlying American commuter rail stations, and can be scheduled to turn in 10 minutes with 2 minutes of schedule recovery.
Regional Rail Comparisons
Amtrak’s report misunderstands the largest European commuter rail networks even while attempting to benchmark Penn Expansion against them. The details of the case studies—London’s Crossrail and Thameslink, the Paris RER, and the Munich S-Bahn—are at times just incorrect (for example, on scheduling), and at others not pertinent to the discussion. Based on a flawed understanding of how these commuter rail systems work, the report then proceeds to assert that any alternative to Penn Expansion would itself require disruptive and expensive work, including additional tracks or extensive platform and track reconstruction. In truth, none of these disruptions are necessary. Amtrak’s report is either misinformed about European practices or has painted them in the least-flattering light possible. As a result, its recommendations are decades behind modern commuter rail operating practices and infrastructure.
Scheduling
Incredibly, the report incorrectly asserts that London, Paris, and Munich all run their commuter rail systems on the basis of headway-based scheduling, rather than a fixed timetable. In reality, these branched suburban rail networks are run as scheduled railways. Trains on these systems are meticulously ordered so as to minimize conflict at intersections, create a smooth flow of trains on central trunks, and maximize on-time performance (OTP). While trains sometimes go off-schedule, this is not a part of normal operations. Correctly sequencing trains from different branches to share a central trunk is a vital practice, and is possible on a through-run commuter rail network in New York.
In fact, New York City Transit’s operations planning department has reached the exact same conclusion regarding the subway: a system so intricately branched it must be run as a scheduled railway. The 2000s–10s experiment with Wait Assessment, centering consistent-looking intervals over OTP, was implicated in the slowdowns that became infamous in 2017, and an internal, leaked document [FN2] described the failings of headway-based management. The branched commuter rail networks under discussion in the report have not made this mistake and still run on a timetable as they ought to.
This error highlights just how mistaken Amtrak and its consultants are about through-running commuter rail systems. While their assertions may seem to make sense to a layperson, they are nonsensical to a transit planner.
Track Separation
The report asserts that a through-running commuter rail system must have separate tracks and approaches for inner- and outer-suburban trains, with inner-suburban trains through-running and outer-suburban trains terminating. Based on this belief, it assumes that through-running requires entirely new tracks at Penn Station for the use of through-running trains, while the existing station is used for terminating trains. None of this is true.
A map of Paris’s RER network today overlaid on a map of Paris’s entire rail network today [1].
In fact, some through-running lines shoehorn all branches into a through trunk. For example, in Paris, the RER C has taken over all of the former commuter network into Gare d’Austerlitz. The RER D has taken over most of that of Gare de Lyon, with most trains on the station’s network through-running. The decision of which lines to through-run, and if it is necessary to separate tracks depends on the size of the historic terminal and the capacity of the through-running trunk.
Gare du Nord has a 10-track approach and 28 surface tracks, but the through-running tunnel to the south used by the RER B and D only has 2 tracks, so not all lines can through-run. The separation is not done on a local-express basis. Instead, the western lines connecting to the station, local or express, terminate and are called Transilien H. The eastern lines through-run as RER B or D, including even some express branches (for example, the airport express trains on the RER B).
Zurich is not among the European systems the report considers, but offers valuable lessons as well. Its system has two 2-track through-running tunnels in addition to surface terminating tracks, and the decision of which branch goes in which tunnel (or in no tunnel) depends on timed overtakes and connections, with no neat separation of inner locals and outer expresses.
A map of proposed through-running branches serving Penn Station (Credit: Kara Fischer).
In New York, the most advantageous system for Penn Station involves assigning trains to tracks based on which branch feeds which tunnel under the Hudson, minimizing the need for trains to cross in front of one another. This scheme would send Northeast Corridor and North Jersey Coast Line trains to the existing North River Tunnels and the middle through-running portion of the station. Meanwhile, Raritan Valley and Morris & Essex trains are sent to the new Hudson River Tunnels, with no through-running. Future through-running to the east can be included, but our simulations are designed to neither include nor preclude it.
Misleading Interpretations of History
In another incredible misrepresentation, Amtrak’s report claims that the Paris RER and Munich S-Bahn through-running rail systems took decades to build, heavily implying that a similar system in New York would also take many years to come to fruition. In reality, the cores of both systems, including both new stations and tunnels, were completed quickly. While major extensions did indeed take further decades to occur, the main part of each system was functional far faster than the report implies.
A map of Paris’s RER network in 1970 overlaid on a map of Paris’s entire rail network today [1].
A map of Paris’s RER network today overlaid on a map of Paris’s entire rail network. [1].
Paris built the most heavily used portions of its system in 16 years, from 1961 to 1977, and some routing decisions were only made a few years into this process. The Paris RER has expanded since 1977, but the network that had opened by then comprised 13 miles of new double-track tunnels mostly under the city center, 4 miles of above-ground branch connections, and 6 new city stations. All but one of those stations are located under multiple older Métro lines. Amtrak says that it took 30 years to build, but what took 30 years to build is a multi-line system with four separate through-running tunnels.
Map of the current Munich S- and U-Bahn network [2]
Similarly, the Munich S-Bahn was built in 7 years, from 1965 to 1972. In Munich, there have been additions, too, but the core (comprising the central tunnels, 6 new underground city center stations, and surface branch upgrades) opened by 1972. The central trunk, the Stammstrecke, is 7 miles long, with 2.5 miles in tunnel under the historic center of Munich and the Isar River. We are uncertain why the report claims it took 46 years to build the S-Bahn, but in truth, both the tunnel and station work and the surface upgrades to the stations and electrification were completed by 1972.
New York needs a lot less new construction than what Paris and Munich accomplished. The Gateway tunnel is a shorter project than the tunnels that were built for the initial RER or for Crossrail. No new city center stations are required. The surface improvements, as we detail in our commuter rail report, include electrification of some branches and high platforms, but their combined cost is an order of magnitude less than that of either Gateway or Penn Expansion. It is not a generational project; it merely requires a change in mindset from that of the traditional American railroader to that of a European transit planner. This means learning the right lessons from across the Atlantic.
The Right and Wrong Lessons
Wrong Lesson
Right Lesson
Station infrastructure should assume every branch can run to every part of the station, with plenty of local and express stopping patterns on each branch.
Timetables can and should be simplified—no 13 different stopping patterns as on the New Haven Line. Different branches can be permanently paired with different platforms, to reduce the effective branch-to-trunk ratio. Station track assignments are known months in advance, and are printed on intercity rail tickets.
Penn Station should be compatible with unelectrified branches and dual-mode trains arriving at all times of day, with different electrification standards, and with trains with constrained door locations to serve low-platform stations.
Compared to building new tunnels, it’s both cheap and quick to fully electrify all branches that would use the through tunnel, raise low platforms, and either buy multi-voltage rolling stock (see our Capital Plan Response, Part 2) or install 25 kV 60 Hz catenary on lines with other electrification.
New underground through-running stations are needed below surface terminals, but underground stations can be used without additional tunneling (as in Thameslink). Station expansions are nice to have, but not a necessity.
New underground stations are required underneath any legacy station, and station expansions are always required.
Headway-based management is the only way to make high-frequency through-running work.
Commuter rail lines always run on a fixed timetable, with branches scheduled to merge in consistent patterns.
High-frequency inner commuter lines should limit track sharing with other traffic where possible, but cannot totally eliminate it, and can share some strategic sections with intercity traffic (as in Munich).
The regional metro concept needs to be totally separate from existing suburban trains.
Penn Reconstruction and NFPA 130
Penn Station today is not fully compliant with the NFPA 130 fire code, a series of regulations drawn up by the non-profit National Fire Protection Association (NFPA) for transit infrastructure. NFPA 130 requires train stations to have sufficient egress capacity to evacuate passengers to a place of safety within 4 minutes and to evacuate passengers from the most remote point on the platform to a place of safety within 6 minutes.
A map of the current Penn Station (as of 2017) with labeled tracks and platforms (Credit: ETA) [Full SVG].
Currently, only platforms 9 and 10 meet the 4-minute standard, and platforms 7–11 meet the 6-minute standard, assuming that both tracks facing a platform are occupied by a full train. The one exception to this assumption is platform 9, Penn’s narrowest. It faces tracks 17 and 18, but track 18 also faces the wider platform 10, so trains on this track only open their doors onto platform 10, leaving platform 9 to service only track 17. Station circulation analysis done for the Moynihan Station Phase 2 project has found that, under the Moynihan build alternative (since completed) but without Penn Reconstruction, platform 9 has the second lowest passenger egress capacity at the station after platform 2, which is limited to shorter trains.
Assuming present operations remain as they are, with each train potentially capable of using nearly any platform from any adjoining track at the same time, Penn Reconstruction is required to ensure all platforms meet NFPA 130. If this assumption is dropped, however, and it is possible to guarantee that two full commuter trains will never arrive on opposite sides of the same platform in short succession, then the station becomes NFPA-compliant as it stands.
The model we produce below relies on alternation of tracks, so that under normal operations, any two commuter trains using the same platform are always spaced at least 5 minutes apart. This is possible because although Penn Station’s platforms are narrow, there are 11 of them served by 21 tracks.
A map of Penn Station’s concourses, as of 2022, overlaid over its tracks and platforms on the lower level. Note how the Mezzanine and Exit Concourse span the whole station, but the Central Concourse is only half the length, and the West End Concourse does not reach tracks 1–4, which are too short/eastern (Credit: NJ Transit).
One potential element of Penn Reconstruction that may still be useful is the Central Concourse. The current concourse on the lower level only extends halfway down from the north end of the station, serving platforms 7–11. Extending it further south to connect to platforms 3–6 is a useful element of Penn Reconstruction, which would add access points at the more constrained platforms. This would improve passenger flow, reduce crowding, and make using the station an easier experience. Amtrak should also consider extending both the Central Concourse and West End Concourse even further to serve platforms 1 and 2, which are currently accessible only from the NJ Transit Concourse and the Exit Concourse.
Central Concourse expansion is unnecessary for capacity or fire safety under present operations with or without the Gateway tunnel as currently under construction. However, as a matter of future-proofing, in the event regional agencies choose to extend the Gateway tracks east for through-running, then Central Concourse expansion (including a further extension to platforms 1 and 2) will become necessary for NFPA 130 compliance. If trains stub-end on platforms 1–3, then they can be scheduled to alternate between platforms, avoiding a situation in which two consecutive trains use the same platform. However, if there’s through-running, then eastbound trains will be required to only use platform 1, and westbound trains required to only use platform 2 or 3. Thus, the same scheduling will not be possible.
Modeling Through-Running
Based on mildly conservative assumptions about Penn Station pedestrian flow, ETA is confident that the station as is can handle 48 trans-Hudson trains per hour, can double its current peak throughput, and does not need the $17 billion Penn Expansion project to accompany the Gateway tunnel. We believe the analysis below is sufficient, even with its simplifying assumptions, to warrant saving the taxpayers of the region $17 billion by not building Penn Expansion, and instead coordinating operations to enable through-running. Furthermore, through-running would enable 48 tph as soon as Gateway is finished, while Penn Expansion would first have to be constructed, and if the recently completed East Side Access is anything to go by, this could take decades.
We propose the following straightforward track assignments for initial service when the new Gateway tunnel opens:
Tracks 1–6 for Gateway Tunnel terminating trains and Empire Service terminating trains (6 tracks)
Tracks 7–11 for North River Tunnel to Southern East River Tunnel eastbound revenue-to-revenue through-running (5 tracks)
Tracks 12–16 for Southern East River Tunnel to North River Tunnel westbound revenue-to-revenue through-running (5 tracks)
Tracks 17–21 for Northern East River Tunnel LIRR trains that either run to West Side Yard (WSY) or turn and go back in reverse-peak service (5 tracks, with a wide platform between 18 and 19)
A map of Penn Station’s tracks post-Gateway with platforms for through-running (Credit: ETA) [Full SVG].
Of note, tracks 5–6 currently do connect to the East River Tunnels, but we propose to use them as if they are stub-end tracks, since the terminating trains may require more tracks for extra capacity whereas the through trains do not.
Under this scenario, Sunnyside Yard in Queens (currently used by most NJ Transit trains) could no longer be used to turn most peak trains. The reason is that it connects to Penn Station via the East River Tunnels. Therefore, in the morning peak, a train from Penn Station on tracks 5–6 to the yard would displace a through train using tracks 7–11. In particular, trains using tracks 1–6, which NJ Transit will likely use for the Morris and Essex Lines and Raritan Valley Line, cannot be timetabled into the yard. Simply, ETA sees through-running as a more efficient use of the East River Tunnels than shuffling trains back-and-forth to a yard. Those same NJ Transit trains that currently run through in non-revenue service and turn at Sunnyside Yard service are simply replaced by trains that run through in revenue service to Sunnyside and beyond.
Tracks 7–16, which we propose as revenue-to-revenue through tracks, require a train every 10 minutes per track, or a train every 5 minutes per platform. This, in theory, allows them to handle 30 tph in each direction; in current NJ Transit practice, this only translates to 24 tph. The additional vertical circulation capacity added during the now-complete Moynihan Station project permits such throughput. The station circulation analysis has the worst platform, platform 4, clearing in 4 minutes and 50 seconds per train, which is 9 minutes and 40 seconds for a train on each side of the platform. This already takes into account that passengers do not choose the most efficient egress point but instead prefer to exit near 7th Avenue.
Here, we model egress on platforms 1–3 (serving tracks 1–6) in addition to running the model on the through platforms to see if it produces similar results to those of the Moynihan analysis. We find that the above schema works to provide the necessary capacity:
The through tracks (7–16) can clear 24 tph in practice, mixing intercity and commuter trains. Simultaneous boarding and alighting is possible with enough time for the reverse-peak direction not to add to the dwell time.
The terminating tracks (1–6) can reverse 24 tph in practice, with slots for the hourly peak Empire Service train.
The Penn Reconstruction project could make the station nicer by improving the pedestrian flow level of service (LOS), but it is not necessary. Pedestrian LOS measures crowding relative to the maximum capacity, which is measured as 25 passengers per minute per foot of walkway width on a flat walkway, or 17 per foot of width on a staircase. The maximum capacity occurs at the boundary between LOS E and F; with or without Penn Reconstruction, the worst modeled LOS is E, occurring on the stairways and up escalators in the morning rush hour.
Tracks 17–21 in combination with West Side Yard can handle 24 tph per direction.
Pedestrian Flow Assumptions at Penn Station
ETA uses the Transportation Resource Board (TRB) manual on passenger flow capacity, which developed LOS criteria based on the Highway Capacity Manual and timelapse photography studies. ETA’s model makes several assumptions to simplify and be conservative regarding concerns about crowding:
We assume an exit rate of 1 passenger per second per single-door equivalent and four of those on each side of an 85’ car (12 cars per train) [FN3].
We assume 1,620 passengers onboard each arriving commuter train, [FN4] roughly equating to a crush-loaded 10-car, or seated 12-car NJ Transit MultiLevel EMU [FN5]. A 12-car NJ Transit MultiLevel III seats 1,552. By contrast, a full intercity train only holds 386 (9-car Avelia Liberty) [FN6] to 479 (8-car Airo) passengers, and a full 12-car LIRR M9 seats only 1,302 [FN7].
We start with 200 passengers on the platform waiting for each train and have 200 more passengers per train flowing down from the concourses over the model. This boarding load of 400 passengers would be a major increase over today’s typical reverse-peak boardings bound for either Queens or New Jersey.
No departing passengers board until all arriving passengers are off the train (we assume everyone is courteous), but passengers may board before all arriving passengers are off the platform and on the concourse.
We cut 25% off the platform area to account for columns and stairwells.
Maximum pedestrian flow up and down stairs is 17 passengers/minute/foot of width (LOS E). Maximum pedestrian flow on flat surfaces is 25 passengers/minute /foot of width (also LOS E).
There is no bidirectional flow on stairwells if LOS is worse than C, or 10 passengers/minute/foot of width.
The flow rate on all vertical circulation elements (VCEs) is that of stairways rather than escalators, as this slightly underestimates the flow rate on wide escalators, which are higher capacity per foot of width.
Our model always excludes one VCE per platform and otherwise uses all existing VCEs. In practice, this VCE could be a downward escalator in the morning peak hour, which would be used exclusively by reverse-peak passengers.
Passengers maintain a minimum flow at the LOS C/D boundary but taper off once the number of remaining passengers fits the area of stairways, equal to 20’ times the total VCE width, divided by the mid-LOS E floor area of 5 square feet per passenger.
Interpretations of the model must allow some margin for error. The model does not divide the platform into zones or otherwise account for a likely preference for eastward staircases. Nevertheless, in most simulation runs, the model shows platforms clearing a single train loaded with 1,620 passengers almost completely within 2 minutes, in line with the Moynihan capacity analysis of minimum egress times if the preference for 7th Avenue is assumed away.
Our model almost certainly paints an overly negative picture of the station’s capacity to handle bidirectional flow. For one, excluding one VCE means that there is a free reverse-direction escalator for it. Secondly, reverse-peak passengers may well adjust their trajectory to favor less crowded access points. Outbound passengers in the morning, arriving with a few minutes to spare, are likely to walk through the station to a less crowded 8th Avenue exit. This is reinforced by the facilities at Moynihan Station. A passenger arriving a little early is more likely to buy a coffee at Moynihan than near 7th Avenue and access the platform using the 8th Avenue staircases than to backtrack to 7th Avenue before descending to board their train. Nonetheless, we assume no downward flow on stairwells once upward flow reaches the LOS C/D boundary, even though some would occur even at a LOS D condition.
We modeled 2 arriving trains holding 1,620 passengers each 2 minutes apart at platform 3. We have both tracks servicing this platform handling terminating service from the Hudson River Tunnel or Empire Service trains, so 2 simultaneous arrivals from New Jersey are not possible. Furthermore, the capacity of an Empire Service train is only about 500 passengers.
For this report, we modeled platforms 3, 10, and 11 handling two closely spaced arrivals. Out of all the platforms that could handle through-running trains with the track connections that currently exist, platform 3 is one of the smallest, and has the second lowest vertical circulation capacity, ahead of only platform 9. Were it to handle through-service linking New Jersey and Queens, the minimum possible headway between consecutive trains would be 120 seconds. Our basic service design assigns platform 3 to terminating trains. A more typical headway between trains at the same platform in a terminating station, or portion thereof, is 300 seconds. Prudent sequencing would distribute trains arriving via East River Tunnel Line 4 from Long Island among tracks 19 to 21. Track 19 is served by platform 10, which is wider at 42’.
Under our service vision, platform 6 would be the only platform to handle New Jersey-Queens through traffic in both directions. As a result, platform 6 would often handle two trains arriving simultaneously, which led us to model that scenario. Every other platform would have both of the tracks it serves assigned to either the same through-running approach track from either New Jersey, Long Island, or the West Side Yard or the two-track tunnel from New Jersey for terminating service. For that reason, all other platforms would never see two simultaneous arrivals.
Results
The GitHub repository for ETA’s platform crowding, alighting, and boarding model can be found here.
The modeled up and down rates on the staircases as passengers exit a rush hour train are depicted below. Scenarios 1–3 assume that Penn Reconstruction has been completed with the additional VCE capacity that it includes; scenarios 4 and 5 assume that it has not, meaning a reduced VCE capacity.
Scenario 1: NJ Transit Trains Arriving on Each Platform Every 5 Minutes
Normal 24 tph operations include a train every 2.5 minutes per approach track, with 2–3 platforms per inbound approach track. This means that in scheduled practice, trains arrive on each platform every 5 minutes, using a total of 8 out of 11 platforms. Additional platforms (3, 6, 9) can then be used for emergency overflow or for Amtrak. Here, we see that even without Penn Reconstruction, the LOS is at worst E, and is only worse than C for 119 seconds per cycle. Passengers taper off after 187 seconds, by which point there are practically no passengers still waiting on the platform. More passengers subsequently go from the concourse to the platform, but they’re balanced out by passengers boarding, whereas by assumption, all passengers already alighted from the train to the platform. By then, the platform crowding LOS is B, and the LOS excluding boarding passengers is A.
In this case, there is ample capacity for peak crowding. If the train arrives late, it will still be able to clear the passengers by the time the next one arrives. In theory, the dwell time for through trains is a minimum of 5 minutes, with trains alternating between the two tracks facing the platform. In practice, it is lower. Our model allows reverse-peak flow on stairs as soon as the stairway LOS is C, which happens within 185 seconds of the train’s arrival (Penn Reconstruction improves this to 185 seconds). What’s more, many boarding passengers will already be on the platform by the time their train pulls in, and more can likely squeeze down stairwells as arrivals vacate the platform, especially since our model excludes one VCE. For those reasons, we are confident the dwell time can be about 2.5 minutes. In terms of capacity, there is no difference between a 5-minute dwell and a 2.5-minute dwell, if the tunnels and interlockings are already limited to 24 tph, but it does save through-passengers 2.5 minutes.
Terminating platforms (1–3, 9–11) are then limited to a turnaround time of 10 minutes. This is straightforward from the perspective of passenger flow. The limiting factor is how fast the train can turn based on crew changes and the mandatory brake test, which (as we discuss above) can be done in 8 minutes, with a 10-minute scheduled turnaround.
Graph 1
Graph 2
Graph 3
Graph 4
Scenario 2: Two Trains, 2 Minutes Apart
In normal scheduling, trains should alternate between platforms. However, in an emergency case, two trains may arrive at the same platform in succession. This not only affects NFPA compliance, but is also useful for understanding recovery from service disruption, like if a platform is blocked due to a disabled train.
In this scenario of platform 3, with Penn Reconstruction, the maximum capacity over the total width of all used VCEs is 12.68 passengers/second. Max capacity is reached for 13 seconds during the egress period as the peaks of the two trains overlap, and the passengers taper off 355 seconds after the arrival of the first train (or 235 seconds from the arrival of the second train). Without Penn Reconstruction, the numbers are slightly worse: the maximum capacity is 12.04 passengers/second, reached over a period of 32 seconds, and passengers taper off after 364 seconds from the arrival of the first train.
These rates, with or without Penn Reconstruction, are unsustainable after more than two arrivals. A third train 240 seconds after the first would lead to persistent platform crowding. Even if the trains are spaced 150 rather than 120 seconds apart, they would not reliably clear, and long queues could emerge. Fortunately, this scenario has no reason to happen in regular service under our assumption in which the Gateway tunnel (but no other major tunnel infrastructure) is built, since trains can use other platforms.
One scenario in which trains may enter in such fast succession is if there is through-running from Gateway, with two new tunneled tracks going east toward Grand Central or across the East River to Sunnyside. In that case, platform 1 would be used exclusively for eastbound trains and platform 2 for westbound trains. Through-running trains from Gateway would only be able to use tracks 1–5, and so platform 3 would be orphaned, used only in special circumstances rather than in the regular schedule. However, if the trains go to Grand Central, then only half or even less than half of the train’s passengers unload at Penn Station, the others riding to Grand Central (or beyond). This is in line with how subway systems, as well as the Paris RER and similar systems, achieve low headways even with every train making every stop in succession with only one platform track per running track: there are multiple city center stations, so no station has a reason to clear an entire train in ordinary rush hour circumstances.
Graph 5
Graph 6
Graph 7
Graph 8
Scenario 3: Opposite-Direction Arrivals on Platform 6, Penn Reconstruction Complete
We also modeled two simultaneous full commuter rail arrivals at platform 6. Most other platforms in the station would either only handle terminating or same-direction through arrivals from the same approach track, so that two simultaneously arriving trains at opposite platform faces are not possible. But, under our framework, platform 6 is to handle both eastbound and westbound through-running trains. The possibility of simultaneous peak-direction train arrivals in the morning rush, one coming from the west and one from the east, would lead to a prolonged LOS E period on the platform and egress points.
To resolve this problem, platform 6 should not host the most crowded commuter trains. It should instead be used for less crowded trains, especially intercity trains. While a full commuter train today will have 1,620 passengers (a 12-car MultiLevel III), a full Amtrak train will have 479 (an 8-car Airo).
Graph 9
Graph 10
Graph 11
Graph 12
Scenario 4: LIRR Trains Arriving on Platform Every 5 Minutes
Finally, we model arrivals not just on the terminating NJ Transit platforms (1–3) but also the LIRR ones (9–11). The LIRR platforms are wider and have more VCEs than the NJ Transit platforms, and therefore scheduling trains on them is strictly easier. On platform 10 below, we find that in normal operations, the LOS on the staircases is never worse than C, and the LOS on the platform is at worst at the boundary between LOS B and C. This helps explain our result that the LIRR can turn 24 trains per hour on platforms 9–11 in regular service.
Graph 13
Graph 14
Graph 15
Graph 16
Conclusion
While there is little dispute amongst New Yorkers that Penn Station can be improved significantly, there is no merit in expanding its footprint. Amtrak’s justification for expanding Penn Station’s footprint does not pass technical scrutiny; rather, it is founded on assumptions of operating practices that do not exist in well-run transit systems elsewhere, compounded by a poor and incomplete understanding of how such well-run systems actually operate.
Compared to the situations of Paris, Munich, and other places where through-running involved complex construction, Penn Station only requires modest modifications within its existing footprint to handle the increase in passenger traffic expected from the opening of the Gateway tunnels and modernization of the Northeast Corridor. Shortening dwell times at Penn Station, scheduling overtakes to maximize the Northeast Corridor’s throughput, and overall incorporating practices from successful through-running operations elsewhere can boost Penn Station’s capacity and improve passenger experience without the need to condemn one or two full blocks of Midtown Manhattan.
That federal funding for capital projects is far from certain in the coming years makes it even more paramount to make the best use of already-existing infrastructure. Instead of spending billions to expand Penn Station, Amtrak can pursue far more cost-effective means of addressing Penn Station’s real constraints that also produce benefits for the other passenger rail operators at the station.
Footnotes
[FN1 ↑] $2.046 billion / 1020 km * 370 km = $742 million. This is explained in an FRA UT Austin study, which calculated the cost to wire 635 route-miles of double-track freight mainline, with 60 trains per day each with 3–4 locomotives, to be $2.046 billion all included, or $3.22 million/mile. We judge that the power drawn on medium-traffic commuter branches such as the unelectrified sections of the New York commuter rail network, which total about 230 miles west of the Hudson, is comparable. On the single-track Port Jervis Line, which is 65 of those 230 miles, the traffic is lower and if anything the per-mile cost should be somewhat lower.
[FN2 ↑] This internal document was leaked separately to Dan Rivoli (then at the Daily News) and Alon Levy (then freelance, writing for Vox and the Commercial Observer).
[FN3 ↑] This represents the highest-capacity, smallest door case of a NJT MultiLevel EMU. LIRR EMUs empty faster, as they have more and wider doors, no stairs, and a lower capacity due to being single-level. Intercity trains have fewer doors and slower alighting passengers, often with luggage, but have significantly fewer passengers, as explained in the next bullet point.
[FN4 ↑] The Moynihan Station environmental assessment used a maximum passenger load of 1,620 per train.
[FN5 ↑] Of the terminating tracks, tracks 1–4 are currently 9 cars long and tracks 5–6 are 10 cars long. Of the through tracks, tracks 7–8 and 15–16 are 13 cars long, while tracks 9–14 are 17 cars long. Of the through-to-WSY tracks, tracks 17–21 are 12 cars long. Since the terminating tracks 1–6 are the largest bottleneck, we assume the highest-capacity train, the NJT MultiLevel IIIs, will be limited to 12 cars where the platform length is sufficient to use them.
[FN6 ↑] 9 passenger cars, 2 power cars
[FN7 ↑] 217 seats/married pair * 6 = 1302 seats in a 12-car M9 train comprised of 6 married pairs.