Theory-guided deep learning for long-term national road traffic modelling

Theory-guided deep learning for long-term national road traffic modelling

This project addresses a central challenge in long-term traffic-demand modelling for national road networks: how to improve predictive performance without sacrificing interpretability. It proposes DeepDemand, a theory-guided deep-learning framework that embeds key mechanisms from classical travel-demand theory into a differentiable neural-network architecture.

The framework is validated on eight years of observations from 5,088 motorway segments in the United Kingdom's strategic road network. Results show that the model can achieve strong predictive accuracy and spatial transferability while revealing how travel-time impedance, population structure, car availability, land use, and employment distribution shape traffic demand.

The study is designed to support accessibility-oriented transport planning, evaluation of public-service coverage, travel-equity analysis, and regional spatial policy by offering a new data-driven tool with stronger theoretical grounding than conventional black-box approaches.

Keywords: long-term traffic modelling; theory-guided deep learning; national road network; explainable AI; accessibility planning