How to Accurately Forecast Highway Traffic Speed? A Self-Supervised Spatio-Temporal Bilateral Learning Network

Accurate prediction of traffic speed is critical for traffic system scheduling, management, and optimization. In highway traffic speed prediction, three essential elements should be considered: 1) the complex spatial diffusion of traffic over time; 2) the influence of traffic patterns on prediction accuracy; and 3) the crucial role of bidirectional learning mechanisms in time series forecasting tasks. To address these challenges, a novel self-supervised spatiotemporal bilateral learning network, named 3S-TBLN, is proposed for long-term traffic speed forecasting. 3S-TBLN adopts a bilateral encoder–decoder architecture, in which both the encoder and the decoder consist of a semantic transformer, multiple spatio-temporal blocks (ST-Blocks), and a bridge transformer (BridgeTrans). Specifically, the ST-Blocks are designed to model multi-step dynamic spatio-temporal correlations in both the encoder and the decoder. Additionally, in the encoder, BridgeTrans is applied to learn forward traffic patterns from the previous week’s observations, while the decoder learns reverse patterns in a symmetric manner. Furthermore, a bilateral architecture–based self-supervised learning strategy is employed to reconstruct historical variables as a pretext task and is jointly optimized with the speed prediction task to learn bidirectional temporal context. Experimental results demonstrate that the proposed 3S-TBLN significantly outperforms state-of-the-art baselines on three real-world highway datasets and effectively addresses the challenge of long-term highway traffic speed prediction. The source code is available athttps://github.com/zouguojian/Traffic-speed-prediction/tree/main/3S-TBLN