Deep Learning for Short-term Network-wide Road Traffic Forecasting

Abstract

Traffic forecasting is a critical component of modern intelligent transportation systems for urban traffic management and control. Learning and forecasting network-scale traffic states based on spatial-temporal traffic data is particularly challenging for classical statistical and machine learning models due to the time-varying traffic patterns and the complicated spatial dependencies on road networks. The existence of missing values in traffic data makes this task even harder. With the rise of deep learning, this work attempts to answer: how to design proper deep learning models to deal with complicated network-wide traffic data and extract comprehensive features to enhance prediction performance, and how to evaluate and apply existing deep learning-based traffic prediction models to further facilitate future research? To address those key challenges in short-term road traffic forecasting problems, this work develops deep learning models and applications to: 1) extract comprehensive features from complex spatial-temporal data to enhance prediction performance, 2) address the missing value issue in traffic forecasting tasks, and 3) deal with multi-source data, evaluate existing deep learning-based traffic forecasting models, share model results as benchmarks, and apply those models into practice. This work makes both original methodological and practical contributions to short-term network-wide traffic forecasting research. The traffic feature learning can categorized as learning traffic data as spatial-temporal matrices and learning the traffic network as a graph. Stacked bidirectional recurrent neural network is proposed to capture bidirectional temporal dependencies in traffic data. To learn localized features from the topological structure of the road network, two deep learning frameworks incorporating graph convolution and graph wavelet operations, respectively, are proposed to learn the interactions between roadway segments and predict their traffic states. To deal with missing values in traffic forecasting tasks, an imputation unit is incorporated into the recurrent neural network to increase prediction performance. Further, to fill in missing values in the graph-based traffic network, a graph Markov network is proposed, which can infer missing traffic states step by step along with the prediction process. In summary, the proposed graph-based models not only achieve superior forecasting performance but also increase the interpretability of the interaction between road segments during the forecasting process. From the practical perspective, to further facilitate future research, an open-source data and model sharing platform for evaluating existing traffic forecasting models as benchmarks is established. Additionally, a traffic performance measurement platform is presented which has the capability of taking the proposed network-wide traffic prediction models into practice.