Predicting Traffic Flow with Real-Time Vehicle Data

by

Author: Denis Avetisyan

A new approach uses connected vehicle information to forecast arterial network traffic, accounting for both typical patterns and unusual events.

The groundwork for predicting traffic flow relies on establishing preliminary conditions, allowing for the anticipation of congestion and the potential optimization of routes.
The groundwork for predicting traffic flow relies on establishing preliminary conditions, allowing for the anticipation of congestion and the potential optimization of routes.

This paper introduces an abnormality-aware spatiotemporal graph convolutional network for accurate traffic state estimation in arterial networks using connected vehicle data.

Existing traffic prediction models struggle to accurately forecast conditions in large-scale arterial networks, particularly when faced with anomalous events. This is addressed in ‘Arterial Network Traffic State Prediction with Connected Vehicle Data: An Abnormality-Aware Spatiotemporal Network’, which introduces a novel framework leveraging connected vehicle data and a dual-expert spatiotemporal graph convolutional network (AASTGCN) to explicitly model both normal and abnormal traffic patterns. Experimental results demonstrate that AASTGCN outperforms existing methods by adaptively balancing historical and real-time information, offering improved prediction accuracy across all traffic conditions. Could this abnormality-aware approach unlock more robust and reliable urban traffic management strategies?

The Inevitable Chaos of Arterial Roads

Conventional traffic prediction models often falter when applied to the realities of arterial roadways, primarily due to their inherent complexity and ever-changing conditions. These networks, unlike freeways with more predictable flow, are characterized by frequent stops, variable demand linked to local activity, and a high sensitivity to even minor disruptions. Consequently, methods designed for consistent traffic patterns struggle with non-recurring congestion – incidents like accidents, construction, or even unusually high retail activity – which introduce unpredictable delays. This creates a significant challenge, as these models often fail to capture the nuanced interplay of factors that define traffic flow on arterial streets, leading to forecasts that quickly diverge from actual conditions and hindering effective traffic management strategies.

Effective traffic management hinges on the ability to anticipate future conditions, demanding prediction models that transcend simple extrapolation of typical travel patterns. While recognizing recurring congestion – the daily commute, for instance – is vital, a truly robust system must also discern and incorporate the influence of anomalous events. These unpredictable incidents – accidents, construction, even large public gatherings – drastically alter traffic flow and can render predictions based solely on historical averages wholly inaccurate. Consequently, advanced methodologies are being developed to not only identify these deviations from the norm, but also to assess their likely magnitude and duration, enabling proactive adjustments to signal timings and rerouting strategies that mitigate congestion and optimize network performance. This dual focus – capturing both the expected and the unexpected – represents a fundamental shift in the pursuit of reliable traffic forecasting.

Conventional traffic prediction systems frequently stumble when confronted with atypical events – accidents, sudden construction, or even large public gatherings – resulting in forecasts that deviate significantly from reality. These systems are often trained on historical data representing normal traffic flow, and lack the adaptability to effectively incorporate the unpredictable nature of anomalies. Consequently, traffic management strategies built upon these inaccurate predictions can prove counterproductive, potentially exacerbating congestion rather than alleviating it. The inability to model abnormal traffic not only limits the effectiveness of real-time control – such as dynamic signal timing – but also hinders proactive planning and resource allocation, ultimately diminishing the overall efficiency of urban transportation networks.

Abnormal traffic is identified through deviations from typical network patterns.
Abnormal traffic is identified through deviations from typical network patterns.

A Pragmatic Approach to Prediction

The proposed Abnormality-Aware Spatiotemporal Graph Convolution Network (ASTGCN) utilizes a dual-expert architecture to address the challenges of traffic prediction under varying conditions. This architecture consists of two distinct graph convolutional networks: one specifically trained to model and predict normal traffic patterns, and another dedicated to abnormal traffic conditions – such as incidents or unexpected congestion. Separating these conditions allows each expert to develop specialized representations and parameters optimized for its respective domain, improving the model’s ability to accurately forecast traffic flow even when faced with atypical events. The outputs of these two experts are then combined using a fusion mechanism to provide a comprehensive traffic prediction.

The model utilizes a Spatiotemporal Graph Convolution Network (STGCN) to represent the arterial road network as a graph, where nodes represent road segments and edges define connectivity. This allows the network to directly model spatial dependencies between adjacent road segments by aggregating feature information from neighboring nodes during convolution. Temporal dynamics are captured through the use of gated recurrent units (GRUs) applied to the graph-convolved features, enabling the network to process sequential traffic data and learn patterns in traffic flow over time. The STGCN effectively combines graph convolution for spatial modeling with recurrent neural networks for temporal modeling, facilitating the prediction of future traffic states based on both the network’s topology and historical traffic patterns.

The Gated Fusion Mechanism functions by dynamically weighting contributions from both historical traffic data and current real-time observations during the prediction process. This is achieved through the implementation of gating units – specifically, sigmoid functions – which generate weights between zero and one. These weights are applied to the historical and real-time data streams, effectively modulating their influence on the final prediction. The gating units are parameterized and learned during training, allowing the model to automatically determine the optimal balance between historical patterns and present conditions. This adaptive weighting scheme enables the model to leverage the stability of historical data while remaining responsive to immediate changes in traffic flow, ultimately improving prediction accuracy compared to methods that treat these data sources equally or rely on fixed weighting schemes.

Explicitly modeling traffic anomalies allows the proposed system to improve performance during unexpected events by differentiating between typical and atypical traffic patterns. This is achieved by maintaining separate analytical pathways for normal and anomalous conditions, enabling the model to avoid being unduly influenced by outliers that would skew predictions based solely on historical averages. By isolating and characterizing anomalies – such as accidents, road closures, or sudden surges in demand – the system can more accurately forecast their impact on traffic flow and proactively adjust predictions, resulting in improved responsiveness and more reliable short-term forecasts under disruptive circumstances.

Varying the threshold for anomaly detection significantly impacts the identification of abnormal traffic patterns.
Varying the threshold for anomaly detection significantly impacts the identification of abnormal traffic patterns.

The Data, As It Is

Traffic state estimation forms the core of this research, and is accomplished via a two-stage process leveraging data from Connected Vehicle (CV) sources. This methodology utilizes the high-resolution trajectory data broadcast by CVs to infer prevailing traffic conditions. The first stage involves data pre-processing and validation to ensure data quality and reliability. The second stage applies a Kalman filter-based algorithm to estimate key traffic state variables, including speed, density, and travel time, from the processed CV data. This two-stage approach provides a robust and accurate estimation of network-wide traffic conditions, enabling real-time traffic monitoring and advanced traffic management strategies.

Vehicle trajectories, obtained from connected vehicle data, serve as the primary input for calculating key traffic performance indicators. Specifically, the system determines Traffic Delay by measuring the difference between actual travel time and ideal free-flow travel time for individual vehicles, then aggregates these values across the network. Queue Length is estimated by analyzing the density and speed of vehicles approaching intersections and bottlenecks, effectively counting the number of vehicles experiencing congestion. The combination of these metrics – delay and queue length – provides a comprehensive and granular assessment of network-wide traffic conditions, enabling detailed monitoring and analysis of congestion patterns.

The evaluation of this model utilized a network comprising 1050 arterial links. This scale represents a substantial increase over previous research in traffic state estimation, which typically focused on smaller, more limited networks. The use of a 1050-link network allows for a more comprehensive and realistic assessment of the model’s performance under varying traffic conditions and provides a stronger basis for generalizing results to larger-scale transportation systems. This larger dataset facilitates the identification of patterns and behaviors that might be obscured in studies utilizing fewer links, enhancing the model’s robustness and practical applicability.

During normal traffic conditions, the implemented model demonstrates a Mean Absolute Error (MAE) of 3.535 for estimated Traffic Delay and 4.819 for Queue Length. These MAE values were calculated across a network of 1050 arterial links using Connected Vehicle Data as input. The reported errors represent the average magnitude of the difference between the model’s predictions and the actual observed values for these two key traffic measures. These results indicate a quantifiable level of accuracy in estimating both delay and queue length under typical operating conditions.

Performance metrics for traffic state estimation during abnormal conditions demonstrate a Mean Absolute Error (MAE) of 4.763 for Traffic Delay and 6.770 for Queue Length. These values were derived from model evaluation on a network consisting of 1050 arterial links and represent the average magnitude of error when predicting these key traffic measures during incidents or other non-typical traffic patterns. The reported MAE values provide a quantitative assessment of the model’s accuracy under stressed network conditions, indicating the level of deviation expected in predicted Delay and Queue Length estimations.

Computation of single-vehicle traffic state measures provides insights into individual vehicle behavior within the traffic stream.
Computation of single-vehicle traffic state measures provides insights into individual vehicle behavior within the traffic stream.

The Inevitable Limits of Prediction, and What We Can Do Anyway

The arterial network, crucial for urban mobility, often experiences unpredictable congestion, making accurate traffic state prediction a significant challenge. This research introduces a model demonstrating a marked improvement in forecasting these conditions, notably during abnormal events like incidents or sudden increases in demand. Unlike conventional approaches, the model effectively captures the complex spatiotemporal dependencies inherent in arterial traffic flow, leading to more reliable short-term predictions. Rigorous testing reveals a substantial performance gain – reductions in prediction error of up to 20.1% compared to existing state-of-the-art methods – suggesting its potential to significantly enhance the responsiveness and efficiency of urban transportation systems, even when faced with unexpected disruptions.

The ability to accurately anticipate traffic congestion unlocks a range of proactive management strategies poised to redefine urban mobility. This model’s predictive capabilities extend beyond simply informing drivers of delays; they enable real-time adjustments to traffic signal timings through adaptive control systems, optimizing flow and minimizing bottlenecks before they fully form. Furthermore, the forecasts facilitate dynamic route guidance, allowing navigation systems to divert vehicles away from congested areas, distributing traffic load across alternative pathways. This shift from reactive responses to preemptive interventions promises not only reduced travel times and improved fuel efficiency, but also a more resilient and efficient transportation infrastructure capable of handling unforeseen events and fluctuating demand with greater ease.

Evaluations demonstrate the AASTGCN model’s significant advancement in traffic state prediction, achieving state-of-the-art performance metrics. Specifically, the model reduced the Mean Absolute Error (MAE) by as much as 20.1% when contrasted with currently established benchmark models. This substantial decrease in prediction error isn’t merely a statistical improvement; it translates directly to a more accurate understanding of traffic flow, allowing for finer-grained and more responsive control strategies. The consistently lower MAE across various testing scenarios highlights the model’s robustness and its ability to generalize beyond the specific datasets used during training, suggesting a reliable tool for real-world implementation and paving the way for more proactive traffic management.

The refinement of traffic state prediction directly translates to tangible benefits for transportation networks and commuters. More accurate forecasts allow for preemptive adjustments to traffic control, minimizing congestion and the associated delays; studies suggest even modest improvements in prediction can yield substantial reductions in overall travel time. Consequently, vehicles spend less time idling in traffic, leading to a demonstrable decrease in fuel consumption and lowered emissions. This enhanced efficiency extends beyond individual trips, contributing to a more sustainable and cost-effective transportation system capable of handling increasing demands with greater resilience and optimized resource allocation.

The development of truly intelligent transportation systems hinges on a capacity for dynamic adaptation, and this research offers a crucial step towards realizing that potential. By providing more accurate and reliable traffic state predictions, particularly under challenging conditions, the proposed model empowers systems to move beyond reactive responses to proactive management. This isn’t simply about smoothing out existing congestion; it’s about building resilience into the network itself, allowing it to anticipate and absorb disruptions – from unexpected incidents to fluctuating demand – with minimal impact on commuters. The resulting framework promises not only reduced travel times and fuel consumption, but also a transportation infrastructure better equipped to handle the complexities of a growing and increasingly mobile population, paving the way for sustainable and efficient urban development.

Model performance varies significantly depending on both the prediction horizon and prevailing traffic conditions.
Model performance varies significantly depending on both the prediction horizon and prevailing traffic conditions.

The pursuit of predictive accuracy, as demonstrated by this AASTGCN model utilizing connected vehicle data, feels perpetually Sisyphean. It’s an attempt to anticipate chaos, to model the unpredictable currents of real-world traffic. As Ada Lovelace observed, “The Analytical Engine has no pretensions whatever to originate anything.” This model, impressive as its spatiotemporal graph convolutional network may be, doesn’t create traffic patterns; it merely extrapolates from them. And the moment it’s deployed, production will inevitably introduce edge cases, anomalies the model hadn’t accounted for, proving once again that even the most sophisticated algorithms are simply delaying the inevitable emergence of tech debt. Better one carefully monitored, thoroughly tested model than a hundred brittle, ‘scalable’ approximations, it seems.

What’s Next?

The pursuit of accurate traffic state prediction, now embellished with ‘abnormality awareness,’ feels…familiar. It’s a constant refinement of models attempting to capture a system fundamentally resistant to complete capture. This AASTGCN, with its spatiotemporal convolutions and connected vehicle data, will undoubtedly improve forecasting – until production decides that a single rogue pothole, or a particularly enthusiastic flash mob, renders the carefully constructed ‘normal’ patterns utterly useless. They’ll call it ‘edge case handling’ and ask for more data. It always happens.

The reliance on connected vehicle data, while promising, introduces a comforting illusion of completeness. The system assumes ubiquitous sensors, perfect data transmission, and the complete absence of malicious actors. One can envision scenarios – a coordinated disruption, a software glitch propagating through the network – where the very data meant to predict chaos becomes a vector for it. The model will dutifully report a ‘highly probable’ congestion event, blissfully unaware it’s predicting a denial-of-service attack.

Ultimately, this work, like so many before it, will become tech debt. A beautifully crafted, meticulously documented system, reduced to a sprawling mess of patches and workarounds, all because reality refuses to conform to the elegance of the graph convolutional network. It began, no doubt, as a simple bash script.

Original article: https://arxiv.org/pdf/2603.04432.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-07 00:48