Roadworks interact with demand, network structure and alternative routes. A planner may need to compare many combinations of closed sections and demand levels before choosing a schedule, yet each high-fidelity simulation is computationally expensive. The SYNCHROMODE work therefore asked whether a fast surrogate could reproduce the important changes in traffic flow closely enough to guide sensitivity and resilience analysis.
Methods
The study used the South Holland Aimsun Next model, whose regional subnetwork contains 137,908 road sections, 57,254 nodes and 3,310 centroids. Scenarios varied total demand from 0.7 to 1.3 times the base level and closed between one and five sections. A Hilbert curve preserved spatial locality and Latin Hypercube Sampling selected a diverse training set of 1,500 scenarios; 250 independently sampled scenarios formed the test set. LightGBM learned non-linear relationships between flows, network characteristics, demand and closures. Spherical kriging then modelled the spatial structure of the prediction errors and produced an uncertainty estimate.
Findings
On the held-out scenarios, kriging reduced mean absolute error from 22.79 to 14.12 flow units and mean absolute percentage error from 18.52% to 12.56%; root mean squared error fell from 48.45 to 44.51. The kriging correction agreed with the direction of the original residual in 81.4% of cases. Feature analysis identified a section’s functional relevance to origin-destination paths as the most influential predictor, followed by section length and the level of demand. Mapping residuals and uncertainty also revealed where the model systematically over- or under-predicted.
