Spatio-Temporal Multi-Scale Convolutional Network for Traffic Forecasting

Abstract

Effective traffic forecasting contributes to the public safety and urban management, and the complexity and variability of the traffic system make it difficult to predict traffic over a long-term horizon. In this paper, we focus on analyzing the spatio-temporal features of traffic, and propose a spatio-temporal multi-scale convolutional net-work (ST-MSCN) to solve the problem of traffic flow prediction. First, in order to directly capture the spatial dependence and multi-scale features of urban traffic flow, we propose an MSC unit. In addition, this paper also proposes a multi-level feature fusion strategy to combine low-level surface features and high-level abstract features to effectively avoid feature loss. Finally, we propose an early fusion mechanism and combine it with the MSC unit to ensure the improved prediction results while greatly reducing the complexity of the model. As for the simulation, Beijing taxi trajectory data and New York City bicycle trajectory data are used to carry out the simulation experiments. The experimental results show the advanced nature of our model, and prediction accuracy are 7.58% ∼ 9.23% higher than state-of-the-art.