How to Accurately Predict Traffic Speed Using Simple Input Variables? A Novel Self-Supervised Spatio-Temporal Bilateral Learning Network
Guojian Zou, Ting Wang, Honggang Wang, Jing Fan, Ye Li · 2023
Accurately predicting traffic speed is critical for traffic system scheduling, management, and optimization. Three essential elements should be considered in highway traffic speed prediction: (1) complex traffic spatial diffusion process with time, (2) considerable influence of traffic patterns for predicting, and (3) bi-directed learning mechanism on time series forecasting task occupy a vital place. A self-supervised spatio-temporal bilateral learning network (3S-TBLN) for long-term traffic speed forecasting is proposed to address the above challenges. 3S-TBLN adapts an encoder-decoder, which is bilateral architecture, where both the encoder and the decoder consist of the semantic transformer, multiple spatio-temporal blocks (ST-Blocks), and bridge transformer (BridgeTrans). The semantic transformer is presented convert speeds from source to high-dimension representations; ST-Blocks is designed to model dynamic spatio-temporal correlations in both encoder and decoder; in the encoder, BridgeTrans is applied to learn the forward traffic patterns from the last week's observations, and vice versa; a self-supervised learning method is proposed to reconstruct historical variables as pretext job combined with speed prediction task learn the bi-directed context. Experimental results demonstrate that the proposed 3S-TBLN model significantly outperforms state-of-the-art baselines and can efficiently solve the problem of long-term highway speed prediction.