Adaptive lightweight temporal convolutional network with context-aware downsampling strategy for traffic flow prediction

Shuai Zhang, Xiang Yin, Wenyu Zhang, Jiyuan Xu, Xin Jing

Abstract: Accurate traffic flow prediction is crucial for intelligent transportation systems, which remains challenging due to intricate spatiotemporal features. Despite the promising performance of recent advances in spatiotemporal models, expensive computational costs make them difficult to be used with limited hardware resources. Moreover, existing models tend to cope with multiple temporal patterns of traffic flow in a coarse-grained manner, making it difficult to deeply extract temporal features. In this study, a novel deep learning model is proposed to address the aforementioned issues and achieve accurate traffic flow prediction. First, a new context-aware downsampling strategy is proposed for reducing the computational cost of the model, which provides contextual information for the traffic flow at each time step and then performs downsampling to reduce the sequence length, thus making the model more lightweight. Second, a new adaptive lightweight temporal convolutional module is proposed to extract temporal features deeply, which can adaptively update model parameters in a lightweight and fine-grained manner to deal with multiple temporal patterns of traffic flow. Third, the proposed model employs spatiotemporal embedding to efficiently learn the underlying spatiotemporal patterns of traffic flow. Extensive experiments on multiple real-world datasets validate the effectiveness and robustness of the proposed model.