[2603.26440] Interpretable long-term traffic modelling on national road networks using theory-informed deep learning

Computer Science > Machine Learning arXiv:2603.26440 (cs) [Submitted on 27 Mar 2026] Title:Interpretable long-term traffic modelling on national road networks using theory-informed deep learning Authors:Yue Li, Shujuan Chen, Akihiro Shimoda, Ying Jin View a PDF of the paper titled Interpretable long-term traffic modelling on national road networks using theory-informed deep learning, by Yue Li and 3 other authors View PDF HTML (experimental) Abstract:Long-term traffic modelling is fundamental to transport planning, but existing approaches often trade off interpretability, transferability, and predictive accuracy. Classical travel demand models provide behavioural structure but rely on strong assumptions and extensive calibration, whereas generic deep learning models capture complex patterns but often lack theoretical grounding and spatial transferability, limiting their usefulness for long-term planning applications. We propose DeepDemand, a theory-informed deep learning framework that embeds key components of travel demand theory to predict long-term highway traffic volumes using external socioeconomic features and road-network structure. The framework integrates a competitive two-source Dijkstra procedure for local origin-destination (OD) region extraction and OD pair screening with a differentiable architecture modelling OD interactions and travel-time deterrence. The model is evaluated using eight years (2017-2024) of observations on the UK strategic road network, cove...