Source-linked AI summary
Spatiotemporal Recurrent Convolutional Networks for Traffic Prediction in Transportation Networks
Haiyang Yu, Zhihai Wu, Shuqin Wang, Yunpeng Wang, Xiaolei Ma
TL;DR
Large-scale network-wide traffic prediction requires modeling spatial relationships across transportation links and temporal traffic dynamics. The paper converts traffic speeds into fine-scale grid images and feeds them to SRCNs combining DCNNs with LSTMs. On a Beijing network with 278 links, SRCNs outperformed the compared algorithms in both short- and long-term prediction, while future work targets additional factors, reduced computational burden, and larger networks.
Problem
Large-scale network-wide traffic prediction is important and challenging, while existing methods often cover limited areas or restrict spatial inputs.
Method
The paper maps network-wide traffic speeds to static grid images and uses SRCNs combining DCNNs for spatial dependencies with LSTMs for temporal dynamics.
Results
SRCNs outperformed LSTMs, SAEs, DCNNs, and SVM in prediction accuracy and stability on the Beijing transportation network.
Takeaways & Limitations
The experiments indicate potential for combining DCNNs with LSTMs in large-scale network-wide traffic prediction applications.
Takeaways & Limitations
Future work includes incorporating weather, social events, and traffic control, reducing blank-region computation, and expanding to larger transportation networks.
Abstract
from arXiv · showhide
Predicting large-scale transportation network traffic has become an important and challenging topic in recent decades. Inspired by the domain knowledge of motion prediction, in which the future motion of an object can be predicted based on previous scenes, we propose a network grid representation method that can retain the fine-scale structure of a transportation network. Network-wide traffic speeds are converted into a series of static images and input into a novel deep architecture, namely, spatiotemporal recurrent convolutional networks (SRCNs), for traffic forecasting. The proposed SRCNs inherit the advantages of deep convolutional neural networks (DCNNs) and long short-term memory (LSTM) neural networks. The spatial dependencies of network-wide traffic can be captured by DCNNs, and the temporal dynamics can be learned by LSTMs. An experiment on a Beijing transportation network with 278 links demonstrates that SRCNs outperform other deep learning-based algorithms in both short-term and long-term traffic prediction.
1. Introduction
Large-scale network-wide traffic prediction is important but challenging because existing approaches often focus on limited network areas or restrict spatial inputs. The paper represents traffic as evolving images and combines convolutional and recurrent learning to model network-wide spatiotemporal dependencies.
- Existing traffic prediction approaches are either model-driven or data-driven, with model-driven methods criticized for strong assumptions in real-world applications.
- Many existing methods are designed and validated on expressways or only several intersections rather than large transportation networks.
- Traffic speeds are converted into static images through grid-based segmentation, with each pixel representing the condition of one or more road segments.
- SRCN combines DCNNs and LSTMs to capture near- and far-side spatial dependencies alongside long-term temporal dependencies.
2. Literature review
The literature spans parametric and nonparametric traffic-prediction methods, including statistical models, SVMs, neural networks, and deep architectures. Prior deep methods often struggle to exploit correlations across links and long-term traffic memory, motivating CNN–LSTM integration.
- Traffic forecasting approaches are commonly divided into parametric and nonparametric methods.
- Parametric approaches: Parametric methods include ARIMA, Kalman filters, exponential smoothing, and space-time autoregressive models.
- Nonparametric approaches: SVMs flexibly map nonlinear data into high-dimensional spaces but are sensitive to kernel and parameter choices.
- Nonparametric approaches: Neural networks model complex nonlinear, multidimensional problems and have been applied to traffic speed and congestion prediction.
- Deep neural approaches: Deep neural methods may suit small networks but fail to exploit correlations among links and long-term traffic memory; LSTMs address temporal modeling and outperform MLPs and SVMs in reported studies.
- Proposed direction: The paper proposes combining deep 2D CNNs and deep LSTMs, viewing traffic-network evolution as a video whose frames represent traffic states.
3. Methodology
The methodology converts network traffic into fine-scale grid images, extracts spatial features with convolutional networks, models temporal dependencies with LSTMs, and forecasts future states through an integrated SRCN architecture.
- Network representation: The grid representation divides the network into spatial cells and maps link speeds into scaled image values, assigning zero to blank areas.
- Network representation: Grid segmentation preserves relative link topology and fine-grained road geometry, including sharp U-turns and interchanges.
- Spatial features: DCNNs process traffic-state images with convolution and pooling layers to capture nearby and distant spatial relationships among links.
- Temporal features: LSTMs use memory cells and input, forget, and output gates to learn long-term temporal dependencies while addressing vanishing or exploding gradients.
- Integrated forecasting: DCNN outputs feed the LSTMs, whose outputs enter a fully connected layer that produces predicted speeds through end-to-end training.
- SRCN architecture: SRCNs contain a 2D CNN, two LSTMs, and a fully connected layer for network-wide traffic-state prediction.
- Multistep prediction: SRCN parameters can be set for multistep prediction, such as forecasting time steps 2, 4, and 5 from historical data.
4. Methodology
The study evaluates SRCNs and competing methods for short- and long-term traffic-speed prediction using Beijing network data, with MAPE and RMSE as performance measures. SRCNs achieve the strongest accuracy and most stable error trends across both horizons.
- Experimental setting: Beijing data cover 278 links over 92 days, sampled every two minutes between 6:00 and 22:00.The network exceeds 38.486 km and includes seven arterial roads and hundreds of interchanges.
- Experimental setting: The comparison evaluates SRCNs against LSTMs, SAEs, DCNNs, and SVM under short-term and long-term prediction conditions.Short-term horizons are 2, 4, and 6 minutes; long-term horizons are 20, 40, and 60 minutes, using the previous 30 minutes of data.
- Evaluation: MAPE and RMSE measure forecasting performance across 278 links and 14,896 traffic states.The metrics compare predicted and actual traffic speeds over the tested links and states.
- Short-term prediction: SRCNs produce the most accurate short-term predictions in both MAPE and RMSE, with MAPE approximately 0.1 and RMSE approximately 5.Their prediction error remains lowest with a stable trend, although error increases as the horizon grows.
- Long-term prediction: SRCNs also outperform the alternatives for long-term prediction, achieving approximately 0.2 MAPE and approximately 6 RMSE.Long-term errors increase with prediction horizon and decay more rapidly than short-term performance.
5. Conclusions and Future Studies
The paper presents grid-based network representation and SRCNs for large-scale traffic forecasting, then reports accuracy and stability advantages over four comparison methods. Future work targets richer inputs, more efficient representation, and larger transportation networks.
- Conclusions: The grid-based representation converts network-wide traffic into static images while retaining fine-scale topology such as interchanges, intersections, and ramps.Each image represents traffic conditions across road segments in the transportation network.
- Conclusions: SRCNs combine DCNNs and LSTMs to capture spatial dependencies among links and long-term temporal dependencies for network-wide traffic prediction.The evaluation uses three months of Beijing data from a network with 278 links and compares SRCNs with LSTMs, DCNNs, SAEs, and SVM.
- Conclusions: Numerical experiments show that SRCNs outperform the other algorithms in accuracy and stability.The results support combining DCNNs with LSTMs for large-scale network-wide traffic prediction applications.
- Future studies: Future studies may incorporate weather, social events, and traffic control as additional factors.The authors also propose improving pre-training, redesigning network representations, reducing blank-region computation, and expanding the network scale.
convolutional neural networks., Neural Networks 2017, pp. 105-113.
The cited passages list prior work on traffic prediction, neural networks, convolutional networks, LSTMs, and related applications. They provide bibliographic context rather than substantive findings about this paper.
- Neural architecture literature: The bibliography also includes foundational and applied research on convolutional networks, deep learning, recurrent convolutional networks, and CNN-LSTM models.These references cover computer vision, visual recognition, sentiment analysis, and transportation-network speed prediction.
- Deep learning literature: Several cited studies apply deep neural networks, deep belief networks, and LSTMs to traffic-flow or traffic-speed prediction.The references include work on large-scale road networks, remote microwave sensors, open data, and short-term forecasting.