Abstract
Traffic prediction on a large-scale road network is of great importance to various applications. However, many factors such as sensor failure and communication errors inevitably resulted in a sparse distribution of effective detection points with missing data, which resulting adversely affects the accuracy of traffic prediction. This study considers the bidirectional connectivity of road networks to construct a two-way network graph topology. Based on the graph representation, the tensor combined temporal similarity revisited graph convolutional gate recurrent unit (T-TRGCGR), ingeniously combining traffic prediction and data completion through the Graph Laplace, is proposed to predict traffic states under partially input data missing circumstances and sparse detector distribution for a large-scale freeway network. Additionally, the proposed model can not only be applicable to traffic data prediction with missing values but also adaptively extract the spatio-temporal characteristics from various traffic periodicities while retaining the topological information of the large-scale network. Experiments on a large intercity network in Jiangsu, China shows that the proposed method outperforms state-of-art baselines on real-world traffic dataset, which can be well adapted to the prediction task of sparse coverage of road network detectors with missing data. Furthermore, through the comprehensive analysis and visualization of model parameters and results, it can be seen that the model adequately identifies the influential road network nodes and automatically learns to determine the importance of past traffic flow. (C) 2021 Elsevier B.V. All rights reserved.
NSTL
Elsevier