Capacity-Aware Deep Learning for Generalizable Traffic Volume Estimation Across Links and Cities

This paper presents a link-level framework for estimating hourly traffic volumes using territorial data that are broadly available: probe-speed profiles, road and topology descriptors, and weather observations. Sparse fixed-sensor measurements provide supervision. The authors evaluate two spatial generalization settings: unseen links within a training network and an unseen city. Their capacity-aware formulation decomposes traffic volume into a link-specific structural capacity and a regime-aware hourly utilization ratio, incorporating traffic-theoretic constraints into learning. According to the abstract, experiments in both settings consistently outperform a state-of-the-art baseline under spatial distribution shift, although no metric values are provided in the supplied summary.