Research on Regional Traffic Speed Prediction Based on Capsule Network and Attention Mechanism

Abstract

Accurate and efficient prediction of the traffic conditions of complex road networks is critical to the management and control of today’s traffic, especially for the regional traffic speed prediction in the complex road connection, Because of its non-linear time-varying and complex temporal-spatial correlation, it is extremely challenging to predict traffic speed. So, this paper proposes a capsule network model based on attention mechanism(A-CapsNet), which converts traffic speed data into images for learning, and then predicts regional traffic speed. This research extracts the spatial correlation of the complex traffic road network through the capsule network, and embeds the attention mechanism into the capsule network to enhance the importance of the key node information of the convolutional layer in the capsule network. Validated on 12 months of traffic speed data set of the Santander city center road network in Spain, shows that A-CapsNet has superior performance in terms of prediction accuracy and interpretability.