A Self-Correction Transformer Network for Traffic Flow Prediction Under Dynamic Spatio-Temporal Distributions

Abstract

Precise and timely traffic flow prediction plays a critical role in developing intelligent transportation systems and has attracted considerable attention in recent decades. The traffic flow has a non-stationary character in both time and space, when the drift phenomenon appears, the traffic flow undergoes significant and sudden changes, bringing the challenge to the prediction. This paper proposed a self-supervised learning-based adaptive spatiotemporal self-correction transformer traffic flow prediction network (SCTNet). SCTNet can feel the drift with self-supervised learning, compute distribution features of the test data, obtain the distribution difference signal, feed it into the model as network correction information, and then adjust the spatiotemporal dependence of traffic flow adaptively to enhance prediction accuracy. The self-supervised learning method can adjust the model quickly and smoothly, and be utilized in most existing traffic flow prediction models. The experiments demonstrate that compared to existing models, the proposed self-supervised learning SCTNet has achieved state-of-the-art performance and exhibited strong adaptability to the dynamically changing spatiotemporal distributions of traffic data.