As a crucial part of the Intelligent Transportation System,traffic forecasting is of great help for traffic management and guidance.Among the various parameters describing the status of traffic,traffic speed is a straightforward indicator to reflect road users’perception of traffic conditions.With the developing of information technology,urban traffic speed prediction becomes a hotpot in transportation research. However,forecasting short-term traffic speeds on a large-scale road network is challenging due to the complex spatio-temporal dependencies found in traffic data.Previous studies used Euclidean proximity or topological adjacency to explore the spatial correlation of traffic flows,but did not consider the local structures(e.g.Ring,Detour,Two-hop,Converging and Diverging,etc.)exhibited in a road network,which have a significant influence on traffic propagation.Meanwhile,traffic sequences display distinct multiple time-frequency properties,which are presented as Periodicity,Trend characteristic and Random fluctuations,yet few researchers have made full use of this resource. To fill this gap,we propose a novel hybrid approach–Wavelet and Motif-based Spatio-Temporal(WM-ST)forecasting method to accurately predict network-scale traffic speeds.WM-ST first uses discrete wavelet transform(DWT)to decompose raw traffic data into several components with different frequency sub-bands.Then a motif-based graph convolutional recurrent neural network(Motif-GCRNN)is proposed to learn the high-order spatio-temporal dependencies of traffic speeds from low-frequency components,and auto-regressive moving average(ARMA)models are employed to simulate random fluctuations from the high-frequency components. WM-ST was evaluated on two real-world traffic datasets collected in Chengdu and Wuhan,China,respectively.The speed samples were calculated from the raw trajectory data.The experimental results demonstrated that WM-ST performed well compared to eight state-of-art prediction methods in the mean absolute error,the mean absolute percentage error and root mean square error.It reached the prediction level of the latest base line model.Several ablation studies indicated that the motif-based graph convolution and wavelet decomposition was effective for capturing the spatial correlation triggered by the local structures and the multiple temporal dependencies,respectively.
展开▼
