Multistep traffic forecasting by dynamic graph convolution: Interpretations of real-time spatial correlations

• Propose a novel deep learning model for network-level traffic condition forecasting. Implement dynamic modules to learn state-dependent spatial correlations among roads. Decrypt and relate learnt parameters flow theory. Highlight the possibility combine accuracy interpretability in neural networks. Accurate explainable short-term forecasting is pivotal making trustworthy decisions advanced control guidance systems. Recently, approach, as data-driven alternative model-based data assimilation prediction methods, has become popular this domain. Many of these models show promising predictive performance, but inherently suffer from lack interpretability. This difficulty largely originates inconsistency between static input–output mappings encoded networks nature phenomena. Under different conditions, such freely-flowing versus heavily congested traffic, are needed predict propagation congestion resulting speeds over network more accurately. In study, we design variant graph attention mechanism. The major innovation so-called convolution (DGC) module that local area-wide convolutional kernels dynamically generated evolving states capture real-time dependencies. When conditions change, correlation by DGC changes well. Using DGC, propose multistep model, Dynamic Graph Convolutional Network (DGCN). Experiments using real freeway DGCN competitive performance compared other state-of-the-art models. Equally importantly, process trained indeed explainable. It turns out learns mimic upstream–downstream asymmetric information typical road operations. Specifically, there exists speed-dependent optimal receptive field – which governs what assimilate consistent with back-propagation speed stop-and-go waves streams. implies We believe research paves path transparent applied