Traffic Speed Prediction of Road Cluster with Heterogeneous Sampling Frequency

Abstract

Accurate short-term road traffic prediction is essential for achieving intelligent transportation systems, such as traffic management, travel route planning, and navigation. The existing works typically provide the prediction for an individual road segment each time. Even though some models aim to simultaneously predict the traffic of a cluster of road segments, they usually assume that the road cluster has a regular network topology (e.g., ring network or grid network). These methods cannot be easily extended to road networks of arbitrary graph topology. This paper addresses the problem of traffic speed prediction for a cluster of road segments with arbitrary topology and heterogeneous sampling frequency of traffic states. We propose a novel prediction framework consisting of three modules: network partitioning, feature extraction, and traffic prediction modules. The first module divides the entire traffic network into several disjoint clusters with high intra-clusters similarity and low intercluster similarity, based on our proposed measurement metrics for measuring the similarity of time series with heterogeneous sampling frequency. The second module extract features that capture temporal correlations of speed series and contextual factors (e.g., road network characteristics and extrinsic factors) while considering the heterogeneity in data frequency. The third module relies on the obtained features to simultaneously predict the traffic states of all road segments in a cluster, where the spatial correlations among roadways are captured via an attention mechanism. The performance is evaluated using large-scale real-world traffic data involving 42 bus services.