PDE-constrained traffic assignment optimization for air quality improvement with surrogate models

Abstract

Emissions from road transportation are the major contributors to air pollution, posing the leading risk to public health globally. This issue is particularly severe in mega cities, where overpopulated residents, high-rise buildings, and crowded traffic exacerbate air quality issues. Among the various mitigation strategies, traffic assignment optimization (TAO) is an affordable and practical measure to distribute traffic volumes more judiciously, thereby mitigating the adverse environmental impacts of vehicular emissions.

However, this strategy lacks the capacity to precisely quantify the dispersion process and concentration distribution of emitted air pollutants, especially with versatile meteorological conditions and heterogeneous urban landscapes. To address this deficiency, this thesis proposes an optimization framework that incorporates a surrogate model constructed from computation fluid dynamics (CFD), a high-fidelity module that provides direct and accurate assessments of air quality impacts caused by transportation. The on-road traffic volumes are optimized to minimize carbon monoxide (CO) concentrations at selected sites and systematic accumulative travel time, particularly in areas where vulnerable residents are susceptible to airborne hazards. The Kowloon Peninsula in Hong Kong is used as an example, where the CO concentration is implicitly determined via the surrogate model (Gaussian process regression; GPR). The non-dominated sorting genetic algorithm II (NSGA-II) is then adopted to solve the bi-objective optimization problem.

Although more capable, this proposed strategy faces huge computational costs when dealing with large urban areas. The increased mesh cell number in the aggrandized computational domain leads to deteriorated solution convergence. Hence, more laborious computing is loaded for CFD simulations during surrogate model construction. To address this challenge, the emerging physics informed neural network (PINN) is introduced to solve the advection-diffusion equation (ADE), which