Urban Traffic Predictions Based on Spatio-Temporal Neural Networks


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
Award date4 Sep 2020


Traffic conditions in urban cities have large influences on the cities' developments and citizens' daily lives. For example, many location-based applications provide services for users based on traffic conditions. Therefore, it is very significant to accurately predict traffic data in future. Although the problem of traffic predictions has been studied for decades, it is still very challenging due to the dynamic and complex environments in urban cities, especially for long-term traffic predictions (days to weeks). Inspired by the success of the encoder-decoder framework in machine translation, in traffic predictions, an encoder encodes the historical traffic data and a decoder with the inputs of the encoded data predicts the future traffic values.

To this end, this thesis designs three novel spatio-temporal neural networks based on the encoder-decoder framework to handle the dynamic and complex traffic environments in urban cities and make traffic predictions. The main contents of this thesis are summarized as follows:
(1) Both spatial and temporal dependencies are important for traffic predictions due to the dynamic and complex traffic environments. Existing studies ignore the spatial dependencies by modeling on a single link or study the spatial dependencies with the equal importance for all the links in the road network. Moreover, they exploit the short-term temporal dependencies that may loss the temporal information from long-range historical traffic data. Therefore, this thesis proposes a spatio-temporal attentive neural network (STANN) which is an encoder-decoder framework, to predict the network-wide traffic data. At first, STANN investigates the link-based spatial dependencies with the dynamic roles of links via an attention mechanism in the encoder. Then, STANN enhances the temporal dependencies by considering the temporal attention context with the dynamic correlations between the historical time steps and future time steps via another attention mechanism in the decoder. In short, STANN enhances the exploration of dynamic spatio-temporal dependencies via attention mechanisms to improve the predictive performance. The experiments verify the performance of STANN for traffic predictions compared to other models.

(2) As the traffic conditions in the road network interact with each other, comprehensively exploring the spatio-temporal dependencies is very important for traffic predictions. For example, the traffic of a link is affected by its neighboring regions and the traffic conditions of these regions are also affected by their nearby regions. However, most current studies only learn spatio-temporal dependencies from the view of links or the road network whereas ignore that from the view of regions. Besides, external factors in the road network, e.g., the road type, shape length and speed limit of links, have impacts on the traffic conditions of the road network. To this end, a spatio-temporal neural network (STNN) upgrades STANN for the network-wide traffic predictions by exploring the multi-view spatio-temporal dependencies in perspectives of the road network, regions, and links in the encoder and fusing the external factors into the decoder for making final predictions. The experimental results on three real datasets show the effectiveness of STNN compared to other models.

(3) Most of current studies only focus on short-term traffic predictions whereas perform poorly for long-term traffic predictions, because they take the traffic data at a time step as a unit which does not simultaneously contain the spatial and temporal information and periodic information. Thus, a spatio-temporal convolutional neural network (STCNN) is proposed to achieve better long-term predictive performance. STCNN explores the traffic dynamics including general spatio-temporal traffic dependencies and periodic traffic patterns based on convolutional RNNs over a spatio-temporal traffic matrix series with periodic information (e.g., day-period and week-period patterns). Note that each matrix represents the traffic data during a period (a number of time steps) which contains the information in both spatial and temporal dimensions. The experimental results on real datasets show that STCNN outperforms other models for long-term traffic predictions.