A slow vehicle on a road, such as a bus, a truck, or an automated car driving cautiously, acts like a moving bottleneck. Cars pile up behind it, and the road opens up in front of it. Traffic models built from conservation laws, which treat traffic as a compressible fluid, represent this as a "non-classical" shock: a sharp drop in density across the vehicle that ordinary fluid-model rules would reject, but which a capacity constraint attached to the moving vehicle makes physically correct. Every rigorous treatment of this object lives on a single road. Real slow vehicles cross intersections, and city links can be under 100 meters long. On such a link the queue behind a bus can stretch back past the upstream intersection and block cross-traffic that has a green light. That blocking is called spillback, and network models that store queues as zero-length "point queues" cannot see it.
This note proposes the missing piece: a junction model in which the moving bottleneck's capacity constraint is passed through the intersection, so that the non-classical shock, the junction coupling rule, and finite link storage are resolved together. The deliverable is a conservative numerical solver and a set of benchmark problems. The empirical question is concrete. Does modeling the bottleneck sharply at the junction change predicted spillback onset and blocked vehicle-hours on short urban links enough to matter, compared with the cheap alternative of smearing the bus into a moving capacity drop on a coarse grid?
Core Claim
When a moving bottleneck is within a few cells of a junction, the junction's effective receiving capacity is set by the vehicle's relative passing capacity, not by the downstream link's density. Existing junction solvers get this handover wrong in a predictable way. They either lose mass while the bottleneck sits "inside" the node, or they delay the moment at which the upstream intersection becomes blocked.
In the single-road setting, the bottleneck at position $y(t)$ imposes
$$f(\rho(y,t)) - \dot y\,\rho(y,t) \le F_{\max}(\dot y),$$
where $f$ is the traffic flux, $\rho$ the density and $F_{\max}$ the maximum flow that can pass the vehicle at speed $\dot y$. We propose to insert this inequality as an additional, time-dependent supply bound inside the junction's coupling problem. It is active while the vehicle is within one cell of the node, and it is handed from the incoming link to the chosen outgoing link at the crossing time. The handover is designed so that total vehicle count, including vehicles held in a finite junction buffer of size $M$, is conserved exactly.
The predictive consequence is simple. Let $s$ be the speed of the classical shock at the tail of the bottleneck's queue. If upstream demand exceeds the bottleneck's passing flux, then $s<0$ in the road frame. As soon as the bus clears a node, the queue tail sits upstream of that node and the upstream junction is supply-limited. Coarse models delay this onset by roughly the numerical smearing width divided by $|s|$. On a 60-meter link, that delay can be a large fraction of a signal phase.
Research Program
The method comes from three existing pieces of work, each of which stops short of this setting.
The first supplies the moving-constraint discretization. A conservative scheme for non-classical solutions to a strongly coupled PDE-ODE problem reconstructs two internal states inside the cell that contains the bottleneck, which keeps the non-classical jump sharp and the scheme strictly conservative on one link. We extend that reconstruction to the junction cell. That is where the link-only scheme has no rule.
The second supplies the finite-storage junction. A macroscopic traffic flow model with finite buffers on networks proves well-posedness for junctions with buffers of size $M$ under three architectures: shared space with multiple queues, a single first-in-first-out queue, and independent buffers per exit. It works with cumulative vehicle counts and assumes only classical waves. The file notes that it gives no implementation. We implement its three architectures and add the moving supply bound to them.
The third supplies the junction allocation rule. Priority-based Riemann solver for traffic flow on networks shows that softened priorities stop one blocked exit from freezing a 2×2 junction. On node and route choice models for high-dimensional road networks shows that strict first-in-first-out rules at diverges cause unrealistic spillback. Because these results show the allocation rule itself changes blocking, we treat it as an explicit experimental factor rather than fixing it.
The theory side is motivated by, and constrained by, On the Existence of Solution of Conservation Law with Moving Bottleneck and Discontinuity in FLux. That paper proves existence when a bottleneck crosses a static flux jump on one line. It also shows that this crossing produces resonance, where wave speeds coincide and the usual variation bounds fail. A junction is a harsher version of that interface, so we do not promise a uniqueness proof. Instead we aim for an existence argument for the 1-to-1 and 2-to-1 cases by wave-front tracking, adapting their change-of-variables idea. Uniqueness is left open, as it is even on one road.
The equal-pressure direction enters as a controlled second-order question rather than as the core mechanism. Coupling conditions for gas networks governed by the isothermal Euler equations shows that enforcing continuous pressure, rather than only a subsonic limit, removes spurious pressure jumps at junctions. It also shows that a downstream capacity shortfall launches a backward-moving shock. In their 1-to-1 test with sound speeds 6 and 8, only the equal-pressure condition avoided the jump. That backward "choking" wave is the gas analogue of spillback.
Second-order traffic models such as Aw-Rascle-Zhang have a traffic pressure, and Lagrangian Control of Conservation Laws and Mixed Models identifies extending moving constraints to these models as hard, handling it on open roads with markers carried at the fluid speed. We will implement that model on networks with two closures. The first uses supply and demand alone. The second adds continuity of traffic pressure across the node, modeled on the gas closure and solved implicitly by root-finding, as the gas literature requires once link parameters differ. We then ask whether the closure choice changes spillback predictions once a non-classical shock reaches the node. A null result would be useful too: it would tell modelers that first-order treatment suffices for this phenomenon.
Baselines
The naive reference is a point-queue network. Queues stack at nodes with no physical length, and the bus is modeled as a reduced service rate at the node it is approaching. This mirrors the residual-queue logic of Modeling the residual queue and queue-dependent capacity in a static traffic assignment problem, which reports that conventional static models underestimate queuing delays by about 30% under congestion. That paper explicitly omits spatial spillback. By construction this baseline predicts zero upstream blocking, and it anchors how much the phenomenon matters at all.
The strongest competitor is a standard Godunov cell-transmission network with the bus represented as a moving cell-wise capacity reduction. It is run with the same junction allocation rules and the same finite buffers as our solver. It is what a practitioner would actually build, and on fine enough grids it should converge toward the right answer. The fairness controls are identical demand profiles, signal timings, link lengths, junction rules, and wall-clock budget. The baseline also gets a grid-refinement sweep, so that it can buy accuracy with cells where our method uses reconstruction.
There is no reported result on Lacuna for this baseline on any junction-crossing benchmark. The nearest reported number is the single-link convergence rate of the reconstruction scheme, between $O(\Delta x^{0.8})$ and $O(\Delta x^{1.5})$ (Chalons et al.). Beating this baseline means one of two things. Either our solver reaches a given spillback-onset error with at least five times fewer cells, or at a practical 10–20 m cell size the baseline's onset error exceeds 10% of a signal phase while ours does not.
The key ablation keeps everything in our solver but removes the handover. The constraint stays attached to the incoming link until the bus has fully entered the outgoing cell. This isolates whether the junction-embedded constraint, rather than the sharp link reconstruction alone, produces any improvement.
The microscopic reference is SUMO, run on the 2×3 grid of Scalable Adaptive Traffic Light Control Over a Traffic Network Including Turns, Transit Delays, and Blocking, with buses injected. It serves as an independent check on predicted blocking intervals. For context, that paper reports 0.75%–46% waiting-time reductions over Webster's method, but no spillback-onset figures.
Why not just…?
Why not just run a microscopic simulator like SUMO or CityFlow, which captures bus-induced spillback for free? For prediction alone, this is a genuine competitor, and we use it as ground truth for that reason. What it does not give is a model whose junction behaviour can be written down, analyzed for existence, or embedded in optimization and control. CityFlow is fast, roughly 72 steps per second city-wide. But its output is trajectories, not a coupling rule. If a macroscopic model cannot match SUMO's blocking intervals, it should not be used, and experiment three tests exactly that.
Why not just smear the bus into a moving capacity drop in an ordinary cell-transmission model? This is the most serious threat. The single-link evidence says Godunov-type schemes smear non-classical shocks or fail to enforce the constraint (Chalons et al.). Nobody has measured whether that smearing changes junction-level blocking at realistic cell sizes. If the grid-refinement sweep shows the smeared model matches onset times within tolerance at 10 m cells, we will say so plainly. The contribution then shrinks to the conservation-correct handover and the second-order closure study, and we will not claim a practical accuracy gain.
Why not just add a spillback penalty to signal control, as maximum-pressure-plus-penalty does? Distributed Traffic Signal Control via Coordinated Maximum Pressure-plus-Penalty penalizes full downstream queues and reports 653 s average travel time on a 290-intersection Midtown Manhattan network, against 729 s for capacity-aware backpressure. Its queue counts react to spillback after it happens and carry no information about a bus that is about to enter a short link. Our solver is a prediction model that such a controller could consume. It is not a competing controller, and the stretch experiment below checks whether that prediction helps.
Proposed Experiments
The first experiment is a set of synthetic junction Riemann problems. A single bus crosses a 1-to-1 node with a speed-limit change, a 2-to-1 merge, and a 1-to-2 diverge. Each case is run under the three buffer architectures and both the soft-priority and partial first-in-first-out allocation rules. Reference solutions come from very fine-grid runs and, for 1-to-1, wave-front tracking. The metrics are $L^1$ density error, exact mass-balance error including buffer contents, the error in the time at which the upstream junction becomes supply-limited, and blocked vehicle-hours.
The second experiment repeats these problems on links of 50, 100, and 200 m, sweeping the grid for both our solver and the smeared baseline. It also runs the handover ablation. This is the decisive check. The proposal stands only if the onset-error gap is material on short links and closes on long ones, which is what the smearing-width argument predicts. If the baseline matches at practical resolutions, the main claim is refuted.
The third experiment places buses on the SUMO 2×3 grid with dwell stops near nodes. It compares blocking intervals and upstream queue lengths from our solver, the smeared baseline, and the point-queue model against the microscopic runs, at the demand levels used by Chen and Cassandras.
The fourth experiment is the second-order closure study. On the merge and diverge cases, the Aw-Rascle-Zhang network is run with supply and demand only and with added pressure continuity. Link parameters are made to differ, so that the gas-network finding of spurious junction jumps has a chance to reappear. We will report whether spillback onset differs by more than the first-order discretization error.
As a stretch test, we will feed bus-induced supply forecasts from our solver into the queue-capacity penalty of the maximum-pressure-plus-penalty controller on a small corridor in CityFlow. We will ask whether anticipating the blocking reduces travel time relative to the reactive penalty under the same demand.
References
- A conservative scheme for non-classical solutions to a strongly coupled PDE-ODE problem
- A macroscopic traffic flow model with finite buffers on networks: well-posedness by means of Hamilton–Jacobi equations
- Priority-based Riemann solver for traffic flow on networks
- On node and route choice models for high-dimensional road networks
- On the Existence of Solution of Conservation Law with Moving Bottleneck and Discontinuity in FLux
- Coupling conditions for gas networks governed by the isothermal Euler equations
- Lagrangian Control of Conservation Laws and Mixed Models
- Modeling the residual queue and queue-dependent capacity in a static traffic assignment problem
- Scalable adaptive traffic light control over a traffic network including turns, transit delays, and blocking
- CityFlow: A Multi-Agent Reinforcement Learning Environment for Large Scale City Traffic Scenario
- Distributed Traffic Signal Control via Coordinated Maximum Pressure-Plus-Penalty