Auto-correlation based spatio-temporal adaptive transformer traffic flow prediction

Abstract

Traffic flow prediction is a crucial technology in intelligent transportation systems. To effectively handle intricate spatio-temporal relationships and dynamic features of traffic flow, an Auto-Correlation Based Spatio-Temporal Adaptive Transformer Prediction Model (Auto-STAT) is established, which considers the periodicity of traffic flow. Auto-STAT encompasses such components as Auto-Correlation, Encoder-Decoder, Dynamic Halting, and Cross-Attention. Auto-Correlation is employed to capture the periodic characteristics of traffic flow. The encoder-decoder architecture incorporates Spatial-Adaptive Transformer (SA-Trans) and Temporal-Adaptive Transformer (TA-Trans) to extract intricate spatio-temporal dynamics. Dynamic Halting is integrated into the encoder to enhance computational efficiency. Cross-Attention module is constructed to mitigate error propagation between the encoder-decoder. Furthermore, two decoders are utilized to simultaneously tackle the Historical Traffic Reconstruction (HTR) task and the Future Traffic Forecasting (FTF) task to recollect historical traffic patterns and predict future traffic patterns. Experimental results demonstrate the proposed Auto-STAT achieves exceptional prediction performance on two datasets.