Traffic Network Speed Prediction via Multi-periodic-component Spatial-temporal Neural Network

To overcome the drawbacks of current traffic network speed prediction methods, such as the lack of accuracy and stability for medium and long period prediction, as well as the low capability of self-adaptive traffic network topological modeling, this paper proposes a traffic network speed prediction approach via a novel multi-periodic-component spatial-temporal neural network which takes multi-scale convolutional operators and gated recurrent units as its building blocks. Firstly, according to the periodic characteristic of traffic network speed, the raw data is transformed into a three-dimensional matrix corresponded to the weekly period, daily period, and recent period before inputting to the period component of the proposed model. Secondly, the multi-scale convolutional kernels are used to capture the spatial correlation between the multi-factor nonlinear correlation and the traffic network nodes with a different spatial field of view in each period component. And then, the gated recurrent units are employed to extract the long-term dependency of traffic data. The residual learning framework is also utilized to improve the training efficiency and prevent gradient dispersion. Finally, the traffic speed prediction results related to each period component via the prediction convolutional unit are adaptively weighted and fused. In order to verify the effectiveness of the proposed model, two public datasets are used for experimental analysis, while the mainstream deep neural network models related to the task are compared. The experimental results show that the average absolute error, the average square error and the average absolute percentage error of the proposed model are 2.55, 3.94 and 10.75%, 1.57, 3.52 and 3.44%, respectively. The proposed model outperforms other baseline models in terms of prediction accuracy and long-term prediction stability. Copyright © 2021 by Science Press.