Traffic Speed Prediction Based on Autoencoder and Deep Learning

Traffic prediction is the core of intelligent transportation system, and accurate traffic speed prediction is the key to optimize traffic management. Currently, the traffic speed prediction model based on deep learning has become a research hotspot in the field of transportation. With the rapid development of deep learning and the improvement of computer hardware performance, traffic speed prediction based on deep learning has become a hot spot and mainstream of research. In this paper, a traffic speed prediction model based on autoencoder structure is proposed by combining Causal Convolutional Network (CCN), Graph Convolutional Network (GCN) and Multi-Head Self-Attention (MHSA). The model realizes efficient extraction and fusion of spatiotemporal features through a layered design: GCN handles spatial features, CCN and MHSA handle temporal features. First, in the encoder, multiple 2D causal convolution modules are utilized to capture the core features of traffic flow and remove redundant information. Second, the attention weights are dynamically computed using MHSA to identify important time points and sub-sequences in the traffic sequence, and the spatial features of the traffic flow captured using GCN. Further, when reconstructing potential features in the decoder, jump connections from the encoder are added, so that the decoder multiplexes the shallow features extracted by the encoder in the feature reconstruction stage and retains more detailed information of the original data. Finally, the prediction results are obtained by nonlinear fusion of the autoencoder information by the fully connected network. The experimental results show that compared with many baseline models, the proposed model in this paper is able to capture the spatio-temporal correlation of traffic speed data in traffic flow prediction and has good prediction performance.