[Paper Review] Short-term traffic flow forecasting with spatial-temporal correlation in a hybrid deep learning framework

Yuankai Wu, Huachun Tan|arXiv (Cornell University)|Dec 3, 2016

Traffic Prediction and Management Techniques6 references251 citations

TL;DR

The paper proposes CLTFP, a CNN-LSTM hybrid model that fuses spatial features, short-term temporal dynamics, and daily/weekly periodicities to forecast short-term traffic flow, showing superior accuracy on PeMS data compared with several baselines.

ABSTRACT

Deep learning approaches have reached a celebrity status in artificial intelligence field, its success have mostly relied on Convolutional Networks (CNN) and Recurrent Networks. By exploiting fundamental spatial properties of images and videos, the CNN always achieves dominant performance on visual tasks. And the Recurrent Networks (RNN) especially long short-term memory methods (LSTM) can successfully characterize the temporal correlation, thus exhibits superior capability for time series tasks. Traffic flow data have plentiful characteristics on both time and space domain. However, applications of CNN and LSTM approaches on traffic flow are limited. In this paper, we propose a novel deep architecture combined CNN and LSTM to forecast future traffic flow (CLTFP). An 1-dimension CNN is exploited to capture spatial features of traffic flow, and two LSTMs are utilized to mine the short-term variability and periodicities of traffic flow. Given those meaningful features, the feature-level fusion is performed to achieve short-term forecasting. The proposed CLTFP is compared with other popular forecasting methods on an open datasets. Experimental results indicate that the CLTFP has considerable advantages in traffic flow forecasting. in additional, the proposed CLTFP is analyzed from the view of Granger Causality, and several interesting properties of CLTFP are discovered and discussed .

Motivation & Objective

• Motivate accurate short-term traffic flow forecasting for ITS applications (dynamic control, routing, location-based services).
• Leverage spatial locality, short-term variability, and periodic patterns in traffic data.
• Develop a deep learning architecture that combines CNN for spatial features with LSTMs for temporal features and periodicities.
• Enable end-to-end training with feature-level fusion and sparsity regularization to improve generalization.

Proposed method

• Use a 1D CNN to extract spatial features from a p-location traffic data matrix S across time channels.
• Employ two LSTMs to capture short-term temporal features and long-term periodicities (daily and weekly).
• Fuse spatial and temporal features by concatenation and a regression layer with L1 regularization to forecast upcoming traffic flow.
• Train the model end-to-end using Adamax optimizer with early stopping based on a validation set.
• Evaluate against LSTM, SAE, shallow NN, and GBRT on PeMS freeway corridor data.
• Analyze feature contributions via Granger causality-inspired LASSO experiments.

Experimental results

Research questions

• RQ1 Can a CNN-LSTM framework effectively capture spatial, short-term temporal, and periodic patterns in traffic flow for improved short-term forecasting?
• RQ2 How do spatial, short-term temporal, and periodic features contribute to predictive performance, and what is their relative importance?

Key findings

• CLTFP achieves better MAE, MAPE, and ACE than LSTM, SAE, NN, and GBRT on the PeMS dataset.
• Spatial features dominate predictive power, with S+T+P yielding the best performance.
• Including all feature types (S, T, P) via the fusion approach improves both point forecasts and spatial distribution accuracy.
• The model's feature-level sparsity helps reduce redundancy and improve generalization.
• Incremental predictability analysis (Granger causality perspective) shows that combining S, T, and P provides richer predictive information than any single feature set.