Abstract
Short-term traffic flow prediction is of great significance in intelligent transportation. In recent years, with the development of information collection technology and deep learning algorithms, neural network models have become increasingly popular in traffic flow prediction research. However, accurate and fast prediction is a challenge because of the uncertain feature of traffic flow and limitations of the model structure. Motivated by this issue, this paper uses a dual-branch grammar model to extract the deep spatio-temporal features of historical traffic information. Each branch combines the grammar structure with the gated convolution operation to realize the interaction between the implicit features of different traffic parameters. Moreover, scaled exponential linear units (Selu) are used as an activation function for gated convolution operation to enhance the convergence effect of network training. And then, a wide attention module is designed to weigh the extracted deep spatio-temporal features to increase the model's prediction accuracy with a slight increase in computational cost. Finally, actual traffic data from Caltrans Performance Measurement System (PeMS) is used to evaluate the prediction performance with the result that the proposed prediction method outperforms other methods in terms of prediction accuracy. In addition, this paper proves the Selu function's importance by analysing the training error's convergence effect and explains the role of wide attention in the prediction task through visualization and statistical analysis operations.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have