Transferable Traffic State Forecasting Using Meta-Learning with a Multi-Dimensional Spatio-Temporal Graph Attention Model

Abstract

Traffic state forecasting plays a critical role in developing effective traffic control and management strategies. While machine learning (ML) approaches have become popular, owing to automatic spatio-temporal feature extraction, traditional ML approaches fail to generalize and adapt to unseen datasets, limiting their practical applicability. To boost generalization, researchers are increasingly turning to few-shot adaptation techniques, such as meta-learning, which focuses on learning how to learn and enabling rapid adaptation to unseen datasets using limited data. This study applies the Model-Agnostic Meta-Learning framework to a multi-dimensional spatio-temporal graph attention-based traffic prediction model (M-STGAT), producing a new model, called Meta M-STGAT. The goal is to improve forecasting performance through faster adaptation to unseen time periods. This study uses open-access traffic speed and lane closure data from the California Department of Transportation Performance Measurement System and corresponding weather data from the National Oceanic and Atmospheric Administration’s Automated Surface Observing System. Meta M-STGAT is compared against state-of-the-art traffic state forecasting models, including traditional M-STGAT, a multi-dimensional graph attention network, and a multi-dimensional long short-term memory network. Model performance is evaluated for 30-, 45-, and 60-min prediction horizons on one primary and three transfer datasets. Results show that Meta M-STGAT consistently outperforms all alternative state-of-the-art models across all transfer datasets and prediction horizons. The findings underscore the potential of meta-learning in enhancing traffic state forecasting and its practical implications for traffic management systems.