Wavelet Decomposition Self-Supervised Neural Networks for Traffic Flow Forecasting

Accurate traffic flow forecasting is of great significance for improving traffic flow efficiency, optimizing public traffic resources, and improving traffic management capability. However, there are two challenges in current traffic flow forecasting: i) most existing models directly deal with the original sequence where multiple temporal patterns coexist, which cannot effectively handle the interdependence among different temporal patterns or remove the influence of irregular noise in the original sequence; ii) they fail to efficiently and simultaneously capture the spatio-temporal heterogeneity of traffic flow to extract the rich latent spatio-temporal representations.. To address these challenges, we propose an efficient wavelet decomposition self-supervised learning network (waveSSL). Specifically, the discrete wavelet transform is used to convert the original traffic sequence into two components: high frequency and low frequency. These components represent two temporal patterns, short-term changes, and long-term trends, respectively. Each component undergoes a spatio-temporal processing method to fuse the spatio-temporal relationships. Additionally, a self-supervised learner based on a spatio-temporal masking strategy is incorporated to obtain rich latent spatio-temporal representations of traffic flow and enhance the prediction's generalization ability.