Post-disruption lane reversal optimization with surrogate modeling to improve urban traffic resilience

Rapid post-disruption recovery is essential but challenging, given the complex interactions between vehicular flows and the network supply. Simulation-based methods are widely used to assist the planner with realistic user-system interactions in the recovery measure optimization, though the application to large-scale transportation networks remains computationally expensive. This study explores the feasibility of using surrogate models as a time-efficient alternative to resource-intensive simulations. Lane reversal control is employed as a novel recovery measure and an optimization framework prioritizing systematic recovery is developed. A resilience loss indicator based on macroscopic fundamental diagram (MFD) dynamics is used to evaluate the real-time performance of the transportation system. The proposed surrogate model, therefore, also focuses on approximating recovery evaluation indicators, i.e., the resilience loss, other than link flows and density. The surrogate model contains a dynamic analytical network model and a Gaussian process regression (GPR) model. The former provides the analytical resilience loss and considers the temporal correlation of network changes resulting from time-dependent lane reversal decisions. The latter captures the difference between simulated and analytical resilience losses. Experiments are conducted on a large real-world road network in Kyoto City. The proposed approach demonstrates its efficacy by mitigating traffic resilience loss by about 6% under scenarios of 15 and 20 controllable links with a mere five algorithm iterations, requiring only 150 simulation runs. We also illustrate a trade-off between recovery performance and control resources that more controllable links unnecessarily offer better resilience improvement given the short decision-making duration and the very tight computational budget.