Adaptive Fusion-Based Spatio-Temporal Networks For Urban Flow Prediction

Urban flow prediction is of great importance to intelligent traffic system(ITS).Recently,due to the availability of large-scale traffic data sets and the progress of deep learning technology,urban computing and traffic prediction based on various deep learning methods have become the focus of academic circles and industry,with great academic value and wide application prospects.Accurate urban flow prediction is a very challenging spatio-temporal prediction task.In addition to dealing with various complex nonlinear dependencies(such as temporal dependencies,spatial dependencies and external factors),the unique-characteristics of traffic data should be fully considered in the modeling.Previous work has applied various deep neural networks in the task to capture complicated dependencies,but most of them have not been able to efficiently handle various spatio-temporal dependencies and unique characteristics of traffic flow.In order to tackle this problem,this dissertation takes convolutional recurrent network as the basic structure,combines adaptive fusion and attention mechanism,and proposes accurate and efficient traffic prediction networks:(1)Adaptive fusion-based spatio-temporal prediction networks(entitled AdaptST)based on AdaptLSTM(adaptive long short-term memory)for urban flow prediciton.Firstly,according to the needs of traffic prediction,AdaptLSTM is proposed.The internal structure of LSTM(long short-term memory)is redesigned,and original gated units and candidate memory generating block are improved respectively.In this way,it can not only jointly handle multi-scale local and global spatio-temporal features simultaneously,but also fuse them adaptively.Moreover,the parameter quantity is further optimized.Secondly,based on the proposed AdaptLSTM,AdaptST is further realized,which is used for urban flow prediction.Th e experimental results on the real world traffic data set show that,compared with other variants of LSTM and the classical baseline methods,our AdaptST based on AdaptLSTM outperforms them consistently in terms of performance and parameter quantity,which verifies the superiority of the proposed AdaptLSTM.(2)Adaptive fusion-based spatio-temporal prediction networks incorporating periodic patterns for urban flow prediction.Firstly,considering the unique periodic characteristics of traffic flow data,an adaptive fusion-based spatio-temporal network incorporating periodic patterns is proposed by adding an additional periodic pattern branch to the existing prediction network.Secondly,since traffic flow data is not completely strict,an attention mechanism is used to fuse the periodic feature maps of different time steps of the periodic pattern branch,which combines the correlation information between the periodic feature maps and the feature map of short-term branch,so as to further improve the performance of the network.The experimental results show that our adaptive fusionbased spatio-temporal networks incorporating periodic patterns can further improve the prediction accuracy and robustness.