Deep Autoencoder Neural Networks for Short-Term Traffic Congestion Prediction of Transportation Networks

Sen Zhang, Yong Yao, Jie Hu, Yong Zhao, Shaobo Li, Jianjun HuSensors 2019 · DOI

TL;DR

This paper introduces a novel Deep Congestion Prediction Network (DCPN) that leverages deep autoencoders to predict traffic congestion levels in transportation networks. By creating a new large-scale dataset from publicly available traffic map snapshots, the authors demonstrate that their end-to-end model can effectively learn spatial-temporal correlations and outperform existing state-of-the-art methods in prediction accuracy and computational efficiency.

What Problem Does This Paper Address?

Accurate short-term traffic congestion prediction is crucial for intelligent transportation systems, aiming to improve efficiency and capacity. However, research in this area lags behind traffic flow prediction due to a scarcity of large-scale, high-quality congestion data and advanced prediction algorithms. Existing methods often struggle with the complexity of network-wide congestion, the curse of dimensionality, and the need for manual feature engineering. This paper tackles the challenge of predicting congestion for entire transportation networks using readily available online traffic map data, which is often underutilized.

Methods

The authors propose a two-pronged approach:

  1. Data Acquisition and Representation: They developed a general workflow to collect and process traffic congestion map snapshots from online services. This involves creating road network masks, transforming images into grid-based representations, and calculating average congestion levels for each grid. This process results in a time-series of grid-based congestion maps.
  2. Deep Congestion Prediction Network (DCPN): This is a novel deep learning architecture inspired by Deep Autoencoders (DAEs).
  3. Encoder-Decoder Component: This part learns to represent historical congestion patterns and their temporal correlations. It uses symmetrical encoder and decoder layers to extract low-dimensional feature representations.
  4. Dense Layers Component: Two dense layers follow the autoencoder to construct the predicted congestion levels for future time points.
  5. End-to-End Learning: The entire model is trained end-to-end, directly outputting predictions in a format visually similar to the input, eliminating the need for manual feature engineering.

Datasets

The primary dataset used is the Seattle Area Traffic Congestion Status (SATCS) dataset, which was created by the authors. This dataset consists of traffic congestion map snapshots collected from the Washington State Department of Transportation (WSDOT) for the Seattle area, covering morning rush hours on workdays from January 2016 to February 2017. The dataset was preprocessed into grid-based representations.

Key Insights and Ideas

  • Leveraging Publicly Available Maps: The paper demonstrates that readily available online traffic congestion maps, often overlooked for research, can be a valuable source for building large-scale datasets for congestion prediction.
  • End-to-End Prediction: The proposed DCPN model's ability to perform end-to-end prediction, directly outputting congestion maps, simplifies the prediction pipeline and removes the need for manual feature extraction, which is a significant advantage.
  • Grid-Based Representation: Representing the transportation network as a grid of congestion levels allows the model to capture spatial relationships within the network in a structured manner.
  • Autoencoder for Temporal Feature Learning: The use of a DAE-inspired architecture effectively captures the temporal dependencies and correlations inherent in traffic congestion data.

Key Results

  • The DCPN model achieved superior performance compared to two state-of-the-art deep learning models (SRCN and ConvLSTM) on the SATCS dataset.
  • Across different prediction horizons (10, 30, and 60 minutes), DCPN consistently showed lower Mean Absolute Error (MAE) and Weighted Mean Squared Error (wMSE).
  • Specifically, for the 10-minute prediction horizon, DCPN achieved the lowest MAE and wMSE on 32 and 33 out of 42 days, respectively.
  • The wMSE metric, which penalizes incorrect predictions of non-smooth congestion levels more heavily, showed DCPN outperforming the other models on 41 and 40 days for the 30 and 60-minute horizons, respectively.
  • The DCPN model was found to be more computationally efficient, requiring less total training time to converge despite potentially needing more epochs.

Limitations and Future Work

  • Information Loss: The grid-based representation, while necessary due to GPU memory limitations, results in information loss and reduced geographical accuracy. Future work aims to address this by predicting congestion at a finer granularity.
  • Computational Efficiency: While DCPN is more efficient than the compared models, there's still room for improvement. The current approach includes grids with no roads, leading to unnecessary computation. Future work will explore excluding this background information.
  • Data Source Limitation: The study relied on a single data source (WSDOT). Future work plans to incorporate data from multiple service providers.
  • Extended Forecasting: The study focused solely on congestion levels. Future work aims to expand forecasting capabilities to include travel time, volume, speed, and occupancy, potentially integrating weather conditions for more comprehensive predictions.
  • Enhanced Computing Power: The authors suggest that enhanced computing capabilities would allow for more thorough trials and experiments with larger datasets and more complex models.

Summary updated

BibTeX
@article{zhang2019deep,
  author = {Sen Zhang and Yong Yao and Jie Hu and Yong Zhao and Shaobo Li and Jianjun Hu},
  title = {{Deep Autoencoder Neural Networks for Short-Term Traffic Congestion Prediction of Transportation Networks}},
  journal = {Sensors},
  year = {2019},
  doi = {10.3390/s19102229},
}
Figure