A spatiotemporal transfer learning framework with mixture of experts for traffic flow prediction

Abstract

For traffic management entities, the ability to forecast traffic patterns is crucial to their suite of advanced decision-making solutions. The inherent unpredictability of network traffic makes it challenging to develop a robust predictive model. For this reason, by leveraging a spatiotemporal graph transformer equipped with an array of specialized experts, ensuring more reliable and agile outcomes. In this method, utilizing Louvain algorithm alongside a temporal segmentation approach partition the overarching spatial graph structure of traffic networks into a series of localized spatio-temporal graph subgraphs. Then, multiple expert models are obtained by pre-training each subgraph data using a spatio-temporal synchronous graph transformer. Finally, each expert model is fused in a fine-tuning way to obtain the final predicted value, which ensures the reliability of its forecasts while reducing computational time, demonstrating superior predictive capabilities compared to other state-of-the-art models. Results from simulation experiments on real datasets from PeMS validate its enhanced performance metrics.