Online data imputation and traffic prediction based on real-time data streams are essential for the intelligent transportation systems, particularly online navigation applications based on the real-time traffic information. However, the inevitable data missing problem caused by various disturbances undermines the information contained in such real-time data, thereby threatening the reliability of data acquisition as well as the prediction results. Such scenarios raise a strong need for integrating the tasks of network-scale online data imputation and traffic prediction, because the existing two-step approaches that separate the above procedures cannot be implemented in an online manner. In this paper, we propose a customized spatiotemporal deep learning architecture, named the graph convolutional bidirectional recurrent neural network (GCBRNN), to combine network-scale online data imputation and traffic prediction into an integrated task. The imputation mechanism and bidirectional framework are developed to cooperatively estimate missing entries and infer future values. We further design a network-scale graph convolutional gated recurrent unit (NGC-GRU) within the GCBRNN, which applies the graph convolution operation and $1\times 1$ convolution module to capture the spatiotemporal dependencies in the traffic data. Experiments are carried out on two real-world traffic networks, including traffic speed and flow datasets. The comparison results demonstrate that our approach significantly outperforms several classical benchmark models with respect to both the imputation and prediction tasks on two datasets under various missing data rates.

基于实时数据流的在线数据插补和交通预测对于智能交通系统至关重要，尤其是基于实时交通信息的在线导航应用。然而，各种干扰不可避免地导致数据丢失问题，破坏了这些实时数据中包含的信息，从而威胁到数据采集的可靠性和预测结果。此类场景强烈需要集成网络规模的在线数据插补和流量预测任务，因为现有的将上述过程分开的两步方法无法在线实现。在本文中，我们提出了一种定制的时空深度学习架构，称为图卷积双向递归神经网络（GCBRNN），将网络规模的在线数据插补和流量预测结合到一个集成任务中。开发了插补机制和双向框架来协同估计缺失的条目并推断未来值。我们在 GCBRNN 中进一步设计了一个网络规模的图卷积门控循环单元 (NGC-GRU)，它应用了图卷积操作和$1\times 1$卷积模块来捕获交通数据中的时空依赖性。实验在两个真实世界的交通网络上进行，包括交通速度和流量数据集。比较结果表明，在各种缺失数据率下，我们的方法在两个数据集上的插补和预测任务方面明显优于几个经典的基准模型。