cs.LGMar 27, 2026

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

Authors: Yue Li, Shujuan Chen, Akihiro Shimoda, Ying Jin

Organizations: Martin Centre for Architectural and Urban Studies, University of Cambridge, 1–5 Scroope Terrace, Cambridge, CB2 1PX, United Kingdom · Independent Researcher, Cambridge, United Kingdom

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

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, covering 5088 highway segments. Under random cross-validation, DeepDemand achieves an R2 of 0.718 and an MAE of 7406 vehicles, outperforming linear, ridge, random forest, and gravity-style baselines. Performance remains strong under spatial cross-validation (R2 = 0.665), indicating good geographic transferability. Interpretability analysis reveals a stable nonlinear travel-time deterrence pattern, key socioeconomic drivers of demand, and polycentric OD interaction structures aligned with major employment centres and transport hubs. These results highlight the value of integrating transport theory with deep learning for interpretable highway traffic modelling and practical planning applications.

Figures & tables

Figure 1: Framework overview. a , Demographic and socioeconomic features are harmonised to small-area (LSOA) units. b , A simplified road network is constructed with node features and ground truth attached. c , Traffic sensor records are matched and aggregated at edge level as ground truth. d , For each target edge, a local OD region is extracted and candidate OD pairs are screened. e , A deep learning model learns to predict traffic volume on the target edge. f , Model performance is evaluated using random and spatial cross-validation and compared with baseline models. g , Interpretability analyses reveal travel patterns, influential features, and spatial demand distributions.
Figure 2: Data collection and preprocessing workflow.
Figure 3: Construction and preprocessing of the UK national driving network. The raw OSM Great Britain extract is filtered to retain only drivable road segments, converted into a directed graph with key geometric and semantic attributes, simplified to a topologically consistent national-scale network, and exported for downstream spatial joins and modelling.
Feature groupStratification and subcategoriesTotal featuresRelevance to TDMData source
Area and population densityLand area (1) and population density (1)2Trip generation; Trip distributionONS [ 30 ]
PopulationTotal (1); stratified by sex (2) and age group (5: 0–15, 16–24, 25–49, 50–64, ≥\geq 65); stratified by sex–age combinations (10)18Trip generationONS [ 31 ]
EmploymentTotal (1); stratified by work type (2: full-time, part-time) and industrial sector (18); stratified by work type–sector combinations (36)57Trip attractionONS [ 27 ]
Household and car ownershipTotal (1); stratified by car/van availability (2: no cars or vans, one or more cars or vans) and household composition (10); stratified by car–household combinations (20)33Trip generation; Modal splitONS [ 29 ]
Income and employment deprivation2019 domain raw scores for income deprivation and employment deprivation2Trip generation; Trip distributionIMD 2019 [ 20 ]
Land-use areaArea of land-use polygons by class: residential, commercial, industrial, retail4Trip generation; Trip attraction; Trip distributionOSM [ 11 ]
Table 1: Summary of LSOA-level features used for node feature enrichment.
Figure 4: Descriptive statistics and spatial visualisation of ground-truth traffic volumes (AADT). a , Spatial distribution of AADT, with high volumes concentrated on major corridors between large cities, particularly around London. b , c , AADT distributions by highway type. d , Distribution of the coefficient of variation (standard deviation divided by mean daily traffic). e , Relationship between AADT and the standard deviation of daily total traffic, indicating increasing variability with higher traffic volumes.
Figure 5: Local OD region extraction and OD pair screening. a , Illustration of the competitive expansion process from the target edge, shown up to a 20-minute travel-time threshold for simplicity; points indicate potential origin and destination locations (LSOA centroids) identified during the search. b , Example of a valid OD pair whose shortest route necessarily traverses the target edge. c , Example of an invalid OD pair for which a faster bypass route exists, leading to exclusion.
Figure 6: The architecture of the deep learning model.
 TrainTest
ModelMGEHMAER2R^{2}MGEHMAER2R^{2}
Random five-fold CV
Linear regression85.09 (0.36)13326 (67)0.204 (0.006)89.42 (0.74)14255 (126)-0.001 (0.124)
Ridge regression85.11 (0.39)13332 (69)0.203 (0.006)89.21 (0.62)14208 (117)0.012 (0.119)
Random forest26.01 (0.17)3441 (16)0.934 (0.001)56.03 (0.74)8389 (62)0.641 (0.014)
Gravity (log-linear)65.45 (0.22)9374 (51)0.549 (0.005)65.60 (0.31)9399 (99)0.547 (0.016)
Table 2: Baseline comparison under random five-fold and spatial cross-validation. Values are reported as mean (standard deviation).
Figure 7: Model evaluation under random five-fold cross-validation. Reported metrics show the mean and standard deviation across folds. All panels visualise results from the first fold. a , Predicted versus observed AADT for the training set; b , predicted versus observed AADT for the test set; c , distribution of the GEH metric across all target edges; d , MGEH and MAE stratified by road type of the target edge; e , Result of spatial cross-validation.
Figure 8: Learned travel-time deterrence function. The upper panel shows the mean curve from random five-fold cross-validation with pointwise variability across folds. The lower panels show fold-specific curves from spatial cross-validation, each corresponding to a different held-out region.
Figure 9: SHAP value distributions illustrating the influence of origin and destination LSOA characteristics on OD pair scores. SHAP values are computed from random forest models fitted using OD pair scores as the response variable and grouped LSOA-level features as predictors. Panels correspond to different feature groups: a , aggregated first-level feature domains, b , population age structure, c , employment sector composition, d , land-use composition by area, e , POI counts by type, f , deprivation domains, and g , household car ownership. Colour indicates feature magnitude, with red representing higher values and blue lower values.
Figure 10: Spatial distribution of standardised destination potential across LSOAs, shown as quintiles of D-potential density. The national map highlights the overall spatial structure of trip attraction, while zoomed-in panels illustrate detailed patterns for Greater Manchester, Birmingham, and Greater London. High D-potential clusters align with major employment centres, transport hubs, and activity corridors, including key airport locations, illustrating the polycentric organisation of travel attraction recovered by the model.
More in this topic

Traffic Forecasting

Traffic forecasting predicts future flows, speeds, demand, travel times, and congestion from spatial, temporal, and contextual data.

See all 84 papers

Explore similar work

CardsList
May 9, 2026

Data-driven transport modelling without overfit

Macroscopic transport modelling aims to predict traffic flows after proposed public policy interventions, such as a new road or railway section or a temporary road closure. As such, it is a vital step in infrastructure planning and development. Traditionally, building a transport model has relied on complex understanding of socio-economic characteristics of the population requiring expensive data collection via surveys, which are prone to biases. Previous numerical frameworks to optimize transport models to fit observed traffic flows are not easily-interpretable and can lead to overfit. We present here an alternative: a data-driven modelling protocol with objective function based on traffic counts, which can be nowadays cheaply and reliably obtained; explainable model weights; and a controlled path to increase model complexity and accuracy. We demonstrate our approach on several toy and realistic examples, and suggest ways to generalize to multimodal systems including public transport.

Peter Vanya, Katarína Šimková, Rastislav Farkaš

Sev.en Global Investments, V Celnici 4, 11000 Prague, Czechia · Vrije Universiteit Brussels, Brussels, Belgium · Universit´e Libre de Bruxelles, Brussels, Belgium +1

Intelligent Transportation SystemsTraffic Flow EstimationModel Calibration
AskPDF
Feb 27, 2026

Time Series Foundation Models as Strong Baselines in Transportation Forecasting: A Large-Scale Benchmark Analysis

Accurate forecasting of transportation dynamics is essential for urban mobility and infrastructure planning. Although recent work has achieved strong performance with deep learning models, these methods typically require dataset-specific training, architecture design and hyper-parameter tuning. This paper evaluates whether general-purpose time-series foundation models can serve as forecasters for transportation tasks by benchmarking the zero-shot performance of the state-of-the-art model, Chronos-2, across ten real-world datasets covering highway traffic volume and flow, urban traffic speed, bike-sharing demand, and electric vehicle charging station data. Under a consistent evaluation protocol, we find that, even without any task-specific fine-tuning, Chronos-2 delivers state-of-the-art or competitive accuracy across most datasets, frequently outperforming classical statistical baselines and specialized deep learning architectures, particularly at longer horizons. Beyond point forecasting, we evaluate its native probabilistic outputs using prediction-interval coverage and sharpness, demonstrating that Chronos-2 also provides useful uncertainty quantification without dataset-specific training. In general, this study supports the adoption of time-series foundation models as a key baseline for transportation forecasting research.

Javier Yanes-Pulido, Filipe Rodrigues

Technical University of Denmark, Kgs. Lyngby, Denmark

Time-Series Foundation ModelsZero-Shot Time Series ForecastingTraffic ForecastingForecasting BenchmarksProbabilistic Time Series Forecasting
AskPDF
Feb 25, 2026

Learning from Yesterday's Error: An Efficient Online Learning Method for Traffic Demand Prediction

Accurately predicting short-term traffic demand is critical for intelligent transportation systems. While deep learning models achieve strong performance under stationary conditions, their accuracy often degrades significantly when faced with distribution shifts caused by external events or evolving urban dynamics. Frequent model retraining to adapt to such changes incurs prohibitive computational costs, especially for large-scale or foundation models. To address this challenge, we propose FORESEE (Forecasting Online with Residual Smoothing and Ensemble Experts), a lightweight online adaptation framework that is accurate, robust, and computationally efficient. FORESEE operates without any parameter updates to the base model. Instead, it corrects today's forecast in each region using yesterday's prediction error, stabilized through exponential smoothing guided by a mixture-of-experts mechanism that adapts to recent error dynamics. Moreover, an adaptive spatiotemporal smoothing component propagates error signals across neighboring regions and time slots, capturing coherent shifts in demand patterns. Extensive experiments on seven real-world datasets with three backbone models demonstrate that FORESEE consistently improves prediction accuracy, maintains robustness even when distribution shifts are minimal (avoiding performance degradation), and achieves the lowest computational overhead among existing online methods. By enabling real-time adaptation of traffic forecasting models with negligible computational cost, FORESEE paves the way for deploying reliable, up-to-date prediction systems in dynamic urban environments. Code and data are available at https://github.com/xiannanhuang/FORESEE

Xiannan Huang, Quan Yuan, Chao Yang

Key Laboratory of Road and Traffic Engineering, Ministry of Education at Tongji University, 4800 Cao’an Road, Shanghai, 201804, Shanghai, China · Urban Mobility Institute, Tongji University, 1239 Siping Road, Shanghai, 200092, Shanghai, China

Traffic ForecastingTime Series ForecastingNon-Stationary Time Series ForecastingSpatiotemporal Forecasting
AskPDF