Attention-Enabled Network-level Traffic Speed Prediction

Abstract

Traffic forecasting is critical for the planning and monitoring of modern urban systems. Time-series and junior machine learning methods are either point-based and rely on unrealistic assumptions, or fail to capture the dynamics of the complex traffic network (e.g., non-Euclidean and spatiotemporal). New models need (1) to represent efficiently the spatial dependency of transportation network, and (2) to model nonlinear temporal dynamics simultaneously. They are also expected to forecast for multiple time steps, i.e., long-term. This study investigates a highway sensor network as a graph. Specifically, the level of road network details required for graph deep learning is first discussed. Secondly, this paper proposes a new graph deep learning model enabling attention mechanism to predict speeds in the network. It captures spatial dependencies with adjacency matrices and graph convolutions, and learns temporal information with a recurrent neural network (RNN) structure. Lastly, performance of the proposed model is compared with literature on a real-world dataset. Experiments show that physical roadway linkages are sufficient for the representation, and the proposed attention-enabled model performs better in the prediction task.