Citation

Sun T, Yang C, Han K, Ma W, Zhang F. Transp. Res. Rec. 2020; 2674(8): 78-89.

Copyright

(Copyright © 2020, Transportation Research Board, National Research Council, National Academy of Sciences USA, Publisher SAGE Publishing)

DOI

10.1177/0361198120927393

PMID

unavailable

Abstract

Urban traffic congestion has an obvious spatial and temporal relationship and is relevant to real traffic conditions. Traffic speed is a significant parameter for reflecting congestion of road networks, which is feasible to predict. Traditional traffic forecasting methods have poor accuracy for complex urban road networks, and do not take into account weather and other multisource data. This paper proposes a convolutional neural network (CNN)-based bidirectional spatial-temporal network (CNN-BDSTN) using traffic speed and weather data by crawling electric map information. In CNN-BDSTN, the spatial dependence of traffic network is captured by CNN to compose the time-series input dataset. Bidirectional long short-term memory (LSTM) is introduced to train the convolutional time-series dataset. Compared with linear regression, autoregressive integrated moving average, extreme gradient boosting, LSTM, and CNN-LSTM, CNN-BDSTN presents its ability of spatial and temporal extension and achieves more accurately predicted results. In this case study, traffic speed data of 155 roads and weather information in Urumqi, Xinjiang, People's Republic of China, with 1-min interval for 5 months are tested by CNN-BDSTN. The experiment results show that the accuracy of CNN-BDSTN with input of weather information is better than the scenario of no weather information, and the average predicted error is less than 5%.

Language: en