Abstract
Traffic congestion event prediction is an important yet challenging task in intelligent transportation systems. Many existing works about traffic prediction integrate various temporal encoders and graph convolution networks (GCNs), called spatio-temporal graph-based neural networks, which focus on predicting dense variables such as flow, speed and demand in time snapshots, but they can hardly forecast the traffic congestion events that are sparsely distributed on the continuous time axis. In recent years, neural point process (NPP) has emerged as an appropriate framework for event prediction in continuous time scenarios. However, most conventional works about NPP cannot model the complex spatio-temporal dependencies and congestion evolution patterns. To address these limitations, we propose a spatio-temporal graph neural point process framework, named STGNPP for traffic congestion event prediction. Specifically, we first design the spatio-temporal graph learning module to fully capture the long-range spatio-temporal dependencies from the historical traffic state data along with the road network. The extracted spatio-temporal hidden representation and congestion event information are then fed into a continuous gated recurrent unit to model the congestion evolution patterns. In particular, to fully exploit the periodic information, we also improve the intensity function calculation of the point process with a periodic gated mechanism. Finally, our model simultaneously predicts the occurrence time and duration of the next congestion. Extensive experiments on two real-world datasets demonstrate that our method achieves superior performance in comparison to existing state-of-the-art approaches.
Introduction
In the realm of Intelligent Transportation Systems (ITS), the ability to accurately forecast traffic conditions is paramount for urban planning and navigation. Traditionally, research has focused on "dense" variables—predicting metrics like speed, volume, or flow at regular intervals (e.g., every 5 or 15 minutes). While effective for general monitoring, these models often fail to capture the "sparse" nature of traffic congestion events, which occur at irregular intervals and vary significantly in duration.
Predicting these discrete events in continuous time presents two primary challenges. First, traditional fixed-window forecasting is ill-suited for events that may occur at any arbitrary point in the future. Second, traffic dynamics are a hybrid of continuous evolution (gradual speed changes) and instantaneous shifts (the sudden onset of a jam). To address these gaps, this paper introduces the Spatio-Temporal Graph Neural Point Process (STGNPP), a framework designed to predict both the timing and duration of future congestion events by merging graph-based spatial learning with neural point processes.
Methodology
Spatio-Temporal Graph Learning
To understand how congestion propagates through a road network, STGNPP utilizes a dual-component learning module. A link-wise Transformer is employed to capture long-range temporal dependencies within historical traffic data, while a Graph Convolutional Network (GCN) models the spatial correlations between neighboring road links. This ensures the model understands that a bottleneck on one artery is likely to impact adjacent streets.
The Spatio-Temporal Inquirer and Continuous Dynamics
A key innovation of STGNPP is the Spatio-Temporal Inquirer, which bridges the gap between the continuous traffic state and the point process. It extracts relevant hidden representations from the graph based on the timing and location of historical events.
To model the evolution of traffic between these events, the framework adopts a "Drift and Jump" philosophy:
- Drift (Continuous Evolution): Using a Continuous GRU (GRU-flow), the model simulates the gradual changes in a road's state between events. By utilizing a neural flow architecture rather than a standard Neural ODE solver, the model achieves high efficiency without the computational overhead of complex differential equation solvers.
- Jump (Instantaneous Change): When a congestion event occurs, a discrete GRU handles the sudden update to the hidden state, capturing the impact of the event itself.
Periodic Gated Intensity Function
Traffic patterns are inherently cyclical, dictated by human schedules such as morning and evening rushes. STGNPP incorporates a Periodic Gated Intensity Function that adjusts the likelihood (intensity) of an event based on the time of day and day of the week, allowing the model to distinguish between a random slowdown and a predictable peak-hour jam.
Experimental Evaluation
Datasets
The model was rigorously tested on two large-scale datasets provided by the Amap mobile app, covering the period from May 12, 2021, to November 12, 2021:
- Beijing: 573 road links and approximately 250,000 congestion events.
- Chengdu: 435 road links and approximately 205,000 congestion events.
Performance Results
STGNPP was compared against ten baseline models, including traditional statistical methods (HA), standard RNNs (GRU), and state-of-the-art spatio-temporal models like DCRNN and STGODE. The results demonstrated a clear advantage for the STGNPP framework:
- Overall Accuracy: STGNPP outperformed all baselines by approximately 10% across key metrics.
- Occurrence Timing: In the Beijing dataset, the model achieved a Mean Absolute Error (MAE) of 16.95 minutes for predicting the start of the next congestion event, outperforming the next best model (FNN-TPP) by nearly two minutes.
- Duration Prediction: The model predicted the length of congestion events with an MAE of 13.15 minutes in Beijing and 32.94 minutes in Chengdu.
Ablation Studies
Ablation tests confirmed that both the spatial (GCN) and long-range temporal (Transformer) components are essential. Removing either led to a significant degradation in performance, proving that congestion is a network-wide phenomenon rather than a localized temporal sequence.
Discussion and Limitations
While STGNPP represents a significant step forward, the authors identified specific constraints. Increasing the depth of the GCN beyond three layers resulted in "over-smoothing," where the representations of different road links became too similar to be useful. Additionally, excessively complex continuous GRU layers led to over-fitting on the training data.
Future research may look toward integrating external factors—such as weather conditions, traffic accidents, or public holidays—directly into the intensity function to further enhance the model's predictive precision in the face of non-periodic shocks.
Conclusion
By treating traffic congestion as a continuous-time point process rather than a series of discrete snapshots, STGNPP provides a more physically grounded approach to traffic forecasting. Its ability to model both the gradual drift of traffic states and the sudden jump of congestion events makes it a robust tool for modern intelligent transportation systems.
Summary updated