In the urban road network, the traffic may be in the different traffic states at any time, and the influences of neighboring road sections on traffic flow of the target road section will also vary considerably according to the change of traffic state. A desirable way to capture the changes is to separate traffic patterns into a certain number of clusters according to the degree of similarity of spatiotemporal correlation and to perform nonlinear correlation analysis separately in each cluster.
In order to specify the heterogeneity of spatiotemporal correlations in the road network, we proposed a classification-after-prediction method. Our method consists of the offline phase and the online phase.
In the offline phase, we use the correlation coefficient between traffic flows as the clustering indicator and partition the historical traffic patterns into several clusters, where traffic patterns within the same cluster have similar structures of spatiotemporal correlation. The multiple prediction models are built in which each of them corresponds to an individual cluster. Each prediction model is trained separately on the traffic flow data belonging to the corresponding cluster.
In the online phase, current traffic regime is identified by classifying the traffic patterns at the current time into a suitable cluster. Based on the results of the spatiotemporal correlation analysis in a corresponding cluster, a set of the spatiotemporal dimensions most relevant to the future traffic flow at the current traffic state will be determined. The prediction model receives the input vector and outputs the prediction result.
Experimental results on real traffic data show that the proposed method achieves good prediction accuracy by distinguishing the heterogeneity of spatiotemporal correlations among the traffic flows.
The results above were published in the journal "Transportation" under the title of "A clustering based traffic flow prediction method with dynamic spatiotemporal correlation analysis" (https://doi.org/10.1007/s11116-021-10200-9).
