Traffic Flow Data Prediction Using Residual Deconvolution Based Deep Generative Network

Di Zang, Fang Yang, Zhihua Wei, Keshuang Tang, Jiujun ChengIEEE Access 2019 · DOI

Abstract

Traffic flow prediction is essential for estimating future traffic states and ensuring smooth road network operations. While short-term prediction is well-studied, long-term forecasting—such as predicting traffic for the entire next day—remains a significant challenge due to the weakening correlation of data over time. This paper introduces the Residual Deconvolution Based Deep Generative Network (RDBDGN), a novel framework designed for long-term traffic flow prediction. By treating traffic data as multi-scale spatio-temporal images and employing a Generative Adversarial Network (GAN) architecture, the RDBDGN achieves superior accuracy. Experimental results on real-world traffic data from Shanghai's elevated highways demonstrate that our approach outperforms existing state-of-the-art models.

Introduction

Intelligent Transportation Systems (ITS) rely heavily on accurate traffic flow prediction to optimize routing and manage congestion. Most current research focuses on short-term horizons (5–30 minutes). However, long-term prediction is vital for strategic traffic management and commuter planning. The complexity of traffic systems, characterized by non-linear spatial dependencies and temporal cycles, makes long-term forecasting difficult for traditional linear models like ARIMA. Even standard deep learning models often struggle to maintain the necessary detail for high-resolution future projections.

The RDBDGN Framework

The proposed RDBDGN treats traffic flow as a computer vision problem, converting sensor data into 2D "images" where axes represent physical sensor locations and time intervals. The framework is built on a Generative Adversarial Network (GAN) architecture consisting of two primary components: a generator and a discriminator.

1. The Generator (RDNN)

The generator is composed of multi-channel Residual Deconvolutional Neural Networks (RDNN). To capture patterns at various granularities, the generator processes traffic data at three different time scales (10, 20, and 30-minute intervals).

The RDBDGN architecture, illustrating the adversarial training loop between the multi-scale generator and the discriminator.

Each RDNN subnet utilizes residual units to extract deep features while preventing the vanishing gradient problem. Following feature extraction, deconvolution (transpose convolution) layers are used to reconstruct the high-resolution traffic map for the target day. This "encoder-decoder" style architecture with residual connections ensures that fine-grained spatial and temporal details are preserved throughout the network.

Internal structure of the Residual Deconvolutional Neural Network (RDNN), highlighting the encoder-decoder path with residual connections.

2. The Discriminator

The discriminator is a multi-channel Convolutional Neural Network (CNN) that acts as a critic. During training, it attempts to distinguish between the "fake" traffic patterns produced by the generator and "real" historical traffic data. This adversarial process forces the generator to produce forecasts that are not just statistically accurate but also realistically nuanced, capturing the stochastic fluctuations typical of urban traffic.

3. Multi-Scale Fusion

By fusing the outputs from the three parallel RDNN subnets, the model integrates fine-grained local fluctuations with broader, city-wide congestion trends, leading to a more robust and comprehensive prediction.

Experimental Setup and Datasets

The model was rigorously evaluated using real-world traffic data from 2011 in Shanghai, China, specifically focusing on two major arteries:

  • Yan’an Elevated Highway: Utilizing data from 35 loop detectors.
  • Neihuan Elevated Highway: Utilizing data from 72 loop detectors.

The study focused on the daily window from 7:00 AM to 10:00 PM. To predict the traffic for a target day, the model used input data from the previous day, the same day of the previous week, and the same day two weeks prior, capturing both short-term trends and long-term periodicities.

Results and Discussion

The RDBDGN was compared against several baseline and state-of-the-art models, including traditional Artificial Neural Networks (ANN), standard CNNs, and advanced architectures like ST-ResNet and ConvLSTM.

  • Accuracy Metrics: On the Yan’an highway, the RDBDGN achieved a Mean Relative Error (MRE) of 11.25%, outperforming ConvLSTM (14.49%) and ST-ResNet (15.6%). Similar performance was observed on the Neihuan highway with an MRE of 11.14%.
  • Ablation Study: The authors conducted ablation tests to verify the impact of specific components. Removing the discriminator or the residual units resulted in a significant drop in accuracy, proving that both the adversarial training and the residual deconvolution architecture are critical to the model's success.

The results indicate that treating traffic as a vision task and using GANs helps avoid the "blurry" predictions often produced by standard regression-based deep learning models, which tend to average out peaks and valleys.

Conclusion

The RDBDGN framework represents a significant advancement in long-term traffic flow prediction. By leveraging multi-scale residual deconvolution and adversarial training, the model effectively captures the complex spatio-temporal dynamics of urban highway networks. Future work will focus on integrating external factors, such as weather conditions and special events, to further enhance the model's ability to handle non-periodic traffic anomalies.

Reference: Zang, D., et al. "Traffic Flow Data Prediction Using Residual Deconvolution Based Deep Generative Network," IEEE Access, vol. 7, 2019.

Summary updated

Abstract

Traffic flow prediction is quite crucial for estimating the future traffic states, efficient and accurate prediction models greatly contribute to the smooth traffic of road networks. However, existing methods mainly concentrate on short-term prediction. The challenging task of long-term flow prediction for the next day, as the important reference of traffic management, is still not well solved. In this paper, we present a residual deconvolution based deep generative network (RDBDGN) to handle the problem of long-term traffic flow prediction. The proposed method consists of a generator and a discriminator. The generator is composed of multi-channel residual deconvolutional neural networks, and the discriminator contains a convolutional neural network which aims to optimize the adversarial training process. The experiments are evaluated based on the traffic flow data of elevated highways, presented results demonstrate that our approach outperforms the state-of-the-art works.
BibTeX
@inproceedings{zang2019traffic,
  author = {Di Zang and Fang Yang and Zhihua Wei and Keshuang Tang and Jiujun Cheng},
  title = {{Traffic Flow Data Prediction Using Residual Deconvolution Based Deep Generative Network}},
  booktitle = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2919996},
}
Figure