DeepTrend 2.0: A light-weighted multi-scale traffic prediction model using detrending. (June 2019)

• Record Type: Journal Article
• Authors: Dai, Xingyuan; Fu, Rui; Zhao, Enmin; Zhang, Zuo; Lin, Yilun; Wang, Fei-Yue; Li, Li
• Is Part Of: Transportation research. Volume 103(2019)
• Page Start: 142
• Page End: 157
• Publication Date: 2019-06
• Subjects: Traffic prediction -- Deep learning -- Detrending -- Multi-scale traffic prediction; Transportation -- Periodicals; Transportation -- Technological innovations -- Periodicals; 388.011
• DOI: 10.1016/j.trc.2019.03.022
• Languages: English
• ISSNs: 0968-090X

Abstract

Highlights: Show that detrending brings the advantage to traffic prediction even when deep learning models are considered. Strike a delicate balance between model complexity and model accuracy in the proposed new prediction model. Propose a light-weighted model that is efficient and easy to transfer and deploy. Abstract: In this paper, we propose a detrending based and deep learning based many-to-many traffic prediction model called DeepTrend 2.0 that accepts information collected from multiple sensors as input and simultaneously generates the prediction for all the sensors as output. First, we demonstrate that detrending brings advantages to traffic prediction, even when deep learning models are considered. Second, the proposed model strikes a delicate balance between model complexity and accuracy. In contrast to the existing models that view a sensor network as a weighted graph and use graph convolutional neural networks (GCNN) to model spatial dependency, we represent a sensor network as an image and propose a convolutional neural network (CNN) as the prediction model. The image is generated by the correlation coefficient between the flow series of sensors, which is different from other CNN based prediction approaches that convert the transportation network into an image by the spatial location of sensors or regions. Compared with the GCNN based model, the CNN based DeepTrend 2.0 can achieve much faster convergence during training, and it guarantees similar prediction quality. Test results indicate that the proposed light-weighted model is efficient and easy to transfer and deploy.