A Hybrid Deep Learning Model With Attention-Based Conv-LSTM Networks for Short-Term Traffic Flow Prediction

• Published in: IEEE transactions on intelligent transportation systems Vol. 22; no. 11; pp. 6910 - 6920
• Main Authors: Zheng, Haifeng, Lin, Feng, Feng, Xinxin, Chen, Youjia
• Format: Journal Article
• Language: English
• Published: New York IEEE 01-11-2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; attention mechanism; Bi-LSTM; Conv-LSTM module; Data models; Deep learning; Feature extraction; Forecasting; Intelligent transportation systems; Modules; Neural networks; Predictive models; Traffic flow; Traffic flow prediction; Traffic management; Traffic planning; Transportation; Transportation networks
• ISSN: 1524-9050; 1558-0016
• DOI: 10.1109/TITS.2020.2997352

Accurate short-time traffic flow prediction has gained gradually increasing importance for traffic plan and management with the deployment of intelligent transportation systems (ITSs). However, the existing approaches for short-term traffic flow prediction are unable to efficiently capture the complex nonlinearity of traffic flow, which provide unsatisfactory prediction accuracy. In this paper, we propose a deep learning based model which uses hybrid and multiple-layer architectures to automatically extract inherent features of traffic flow data. Firstly, built on the convolutional neural network (CNN) and the long short-term memory (LSTM) network, we develop an attention-based Conv-LSTM module to extract the spatial and short-term temporal features. The attention mechanism is properly designed to distinguish the importance of flow sequences at different times by automatically assigning different weights. Secondly, to further explore long-term temporal features, we propose a bidirectional LSTM (Bi-LSTM) module to extract daily and weekly periodic features so as to capture variance tendency of the traffic flow from both previous and posterior directions. Finally, extensive experimental results are presented to show that the proposed model combining the attention Conv-LSTM and Bi-LSTM achieves better prediction performance compared with other existing approaches.