Multi-precision traffic speed predictions via modified sequence to sequence model and spatial dependency evaluation method

Abstract

Understanding spatial–temporal characteristics of urban traffic will help improve the short-term prediction accuracy in traffic control and management. This paper focuses on mining the spatial dependence of urban traffic networks and establishes an ensemble deep learning model to address multi-precision prediction accuracy requirements. The main contributions of this work are summarized as follows. First, based on the Maximal Information Coefficient (MIC), we design a spatial dependence evaluation algorithm for detecting spatial anomalies, key roads, and the most correlated roads to the key roads. Second, an attention-based Sequence to Sequence (Seq2Seq) model with modified residual units (ARS model) is employed to make network-scale short-term traffic predictions. Third, for the key roads, we combine the MIC-based evaluation algorithm and ARS model to present a multi-task learning-based Key road Attention-based Residual Seq2Seq model (KARS model), which can significantly improve the prediction accuracy of key roads. Consequently, the ensemble model, ARS&KARS, can be implemented with elementary traffic speed data to handle uneven accuracy prediction tasks (without extra topology information of road networks). The experimental results show that the proposed ensemble prediction model, ARS&KARS, outperforms the benchmark model in terms of prediction accuracy and can effectively harness the intrinsic spatial dependencies of urban networks.