Spatiotemporal dynamics of traffic bottlenecks yields an early signal of heavy congestions

Urban traffic congestion is a pervasive phenomenon where large jams are triggered by uncertain sources (traffic bottlenecks). Prior work has modeled congestion formation with kinematic wave theory, cellular automata, and three-phase traffic theory, as well as queueing, lane-changing, and cell transmission models. More recently, network-based approaches (e.g., cascading failures, epidemic-like models, congestion trees) have been used to capture spatiotemporal propagation of congestion. However, most bottleneck-identification methods focus on single links or simplified corridors and do not capture network-wide propagation, which exhibits strong spatial dependencies. This study aims to quantitatively characterize the full spatiotemporal evolution of congestion components originating from specific bottlenecks—covering their growth and recovery—and to test whether early-stage propagation speed provides an actionable signal to forecast which bottlenecks will evolve into heavy network congestions.

The paper reviews a spectrum of traffic modeling approaches: spontaneous congestion formation via kinematic wave theory, cellular automata, and three-phase traffic theory; bottleneck-caused jams via queue, lane-changing, and cell transmission models. Network-level propagation has been studied with cascading failure and epidemic models, and congestion tree methods that identify origins and branches of jams. Prior prediction and control studies consider propagation and suggest spatiotemporal dynamics are key for forecasting and mitigation. A recent macroscopic SIR framework describes overall congestion spreading, but does not resolve bottleneck-specific dynamics at network scale. Existing bottleneck methods largely assess queues/delays on simplified corridors and may not fit urban networks with multidimensional spatial dependencies. Bayesian inference and propagation-tree approaches identify sources and show that source location shifts can drastically change congestion outcomes, highlighting the need for early identification of critical bottlenecks likely to cause large jams.

Data: Minute-resolution floating-car GPS speed data for all road segments in Beijing (~52,000 links) and Shenzhen (~22,000 links) for 30 days in October 2015 (00:00–24:00). Speeds are aggregated and outliers handled; velocity is generally stable over 5-minute intervals.

Congestion state: For each road e at time window t (5-min aggregation), define congestion weight W(t)=U(t)/U95, where U(t) is the observed speed and U95 is the 95th percentile speed for that link. A link is congested if W(t)<0.5.

Bottleneck identification (jam tree): Using the congestion tree framework, at each time window compute c(t), the duration a link has been continuously congested. The bottleneck is the most downstream link with longest continuous congestion within a congestion component. Upstream neighbors become associated congested links if they turn congested no later than δ=10 minutes after the bottleneck or any already-congested neighbor within the component. Recovery is identified when links (or connecting links) return to uncongested state; recovered branches are no longer associated with the original bottleneck.

Dynamic size model: Track the size S(t) (number of congested links) associated with a bottleneck from its emergence to disappearance. The dynamics follow dS/dt = G(t) − R(t), where G(t) and R(t) are counts of newly congested and newly recovered links in window t; S(t)=∫(G−R)du from emergence. Growth ends at peak time t_p with maximal size S_p; recovery follows until S returns to zero. Growth duration T_G is time from emergence to t_p; recovery duration T_R is t_p to complete dissolution; lifespan T=T_G+T_R.

Statistical analysis: Compute distributions (PDF/CCDF) of T_G and T_R per day and across the month; fit exponential to CCDF of T_G and power-law to CCDF of T_R; assess stability across workdays/holidays; analyze ratio r=T_R/T_G. Sensitivity analyses include absorbing very short roads into junctions, defining S(t) by congestion length, and smoothing in different time intervals.

Spatial–temporal correlation: Define average growth speed V_i=S_pi/T_Gi (increased congested links per 5 minutes during growth). Compute Pearson correlation ρ_{S_p,V} after excluding isolated S_p=1 cases. Also define initial growth speed V_Ti=S_Ti/T_i using first T_i minutes (e.g., 5,10,15,20) and compute ρ_{S_p,V_T} for components with growth duration ≥T_i.

Recurrence: Use Jaccard index to quantify overlap of bottlenecks (J_B) and associated congestion components (J_C) between pairs of workdays as a function of S_p threshold.

Prediction: Train a binary Probit classifier on one workday (e.g., Monday Oct 12, 2015, Beijing) using initial growth speed V_t up to early time T_i to predict the probability a bottleneck becomes major (S_p≥S_L). The model: y_i = α_1 + α_2 V_t + ε_i; P(y_i=1|V_t)=Φ(α_1+α_2 V_t). For each bottleneck, P_i = max_{t≤T_i} P(y_i=1|V_t). Classify as major if P_i ≥ p_thre. Evaluate ROC, AUC, accuracy at fixed FPR (e.g., 5%), and robustness across T_i and S_L values in both cities and across multiple workdays.

• Temporal asymmetry: Growth durations T_G are short and their CCDFs are well approximated by exponentials (e.g., α≈0.24 Beijing, 0.27 Shenzhen on Oct 16), whereas recovery durations T_R are much longer with CCDFs following power laws (β_R≈1.82 Beijing, 1.92 Shenzhen). Jams may take up to ~100 minutes to reach S_p but up to ~1000 minutes to recover.
• Recovery is typically nearly twice as long as growth: Mean ratio ⟨r⟩=⟨T_R/T_G⟩≈2.06 (Beijing) and 1.74 (Shenzhen) on Oct 16; values are stable by day type (workdays > holidays).
• Distributions of r also follow power laws with exponents around 2.1±0.09 (Beijing workdays) and 2.16±0.05 (Shenzhen workdays).
• Spatial–temporal linkage: Strong positive correlation between maximal size S_p and average growth speed V during the growth stage. Pearson ρ_{S_p,V}≈0.75 (Beijing) and 0.79 (Shenzhen) on Oct 16; stable trends across days (higher on holidays and in Shenzhen).
• Early-stage predictability: Initial growth speed within the first 15 minutes is highly predictive of S_p. Correlations ρ_{S_p,V_T} increase with T_i up to ≈15 minutes (example ρ≈0.87 at 15 minutes in Beijing) and plateau thereafter.
• Forecasting performance: Using a Probit model trained on Oct 12 to predict Oct 16 major bottlenecks (S_L=20) yields high accuracy; with FPR set to 5%, detection accuracy ≈88% (Beijing). ROC curves for V_5, V_10, V_15, V_20 show that V_15 approaches ideal classification; AUC values are high (~0.95) and stable across 17 workdays, and robust across different S_L thresholds (10–35) and sensitivity settings.
• Low day-to-day recurrence: Jaccard overlap of heavy bottlenecks across workdays is low and decreases with size; for S_p≥20, J_B≈0.05 and J_C≈0.2, indicating heavy bottlenecks and their exact components seldom repeat at the same locations.

The results demonstrate that bottleneck-driven congestion evolves asymmetrically, with rapid growth but slow, heavy-tailed recovery, consistent with systems near criticality and exhibiting critical slowing down. Stable exponents and ratios across day types suggest self-adaptive network behavior. Crucially, the strong linkage between early growth speed and eventual maximal size provides a practical early-warning signal: observing the first ≈15 minutes of propagation suffices to reliably anticipate whether a bottleneck will become major. This addresses the challenge that heavy bottlenecks are not highly recurrent across days, necessitating real-time early detection rather than reliance on historical location patterns. The predictive framework integrates naturally with traffic control and routing systems to intervene preemptively—mitigating jams before they reach peak extent—while accounting for network-level propagation, which is essential for coordinated control across adjacent intersections and regions.

This work presents a dynamic network framework that tracks the full life cycle of congestion components originating at bottlenecks and quantifies their growth and recovery. Empirically, recovery durations follow power laws and are typically about twice as long as growth, indicating pronounced temporal asymmetry and suggesting proximity to critical points in urban traffic dynamics. A key contribution is identifying initial growth speed (within ≈15 minutes) as a robust predictor of eventual jam size, enabling accurate early detection of major bottlenecks with high and stable performance across cities, days, and thresholds. These insights can inform extensions of mesoscopic traffic models and support integration into real-time traffic control, navigation, and demand management to avert large-scale congestions. Future work could generalize across more cities and seasons, incorporate multimodal data and incident information, explore causal mechanisms behind critical slowing down, and co-design control policies that leverage early warnings for coordinated, system-level mitigation.

• Geographic and temporal scope: Analyses are limited to two Chinese cities (Beijing, Shenzhen) over one month (October 2015), which may affect generalizability to other geographies, seasons, or network designs.
• Data constraints: Raw velocity data are protected and unavailable due to privacy, limiting external validation beyond shared processed datasets and code.
• Model assumptions and thresholds: Congestion state relies on W(t)<0.5 with U95 as a speed baseline; association uses δ=10 minutes and 5-minute aggregation. Although sensitivity tests suggest robustness, different parameterizations or sensing modalities could alter results.
• External factors: The framework does not explicitly model exogenous events (incidents, weather, road works) or multimodal interactions, which may influence propagation and recovery patterns.
• Predictive trade-offs: Early prediction accuracy improves with observation time up to ~15 minutes, implying a practical trade-off between lead time and performance for real-time interventions.