Localized Reduced-Order Modeling via Parameter Space Splitting for the Prediction of Contaminant Dispersion in Urban Areas

Abstract

Numerical simulations of accidental contaminant releases, such as gas leaks at chemical plants, can provide meaningful information about the dispersion of hazardous substances during emergency response actions and in preparatory studies involving many different "what-if" scenarios. Common models couple the incompressible Navier-Stokes equations for the assessment of the wind field with an advection-diffusion problem for evaluating the contaminant transport from an instantaneous source. However, computational times required to evaluate such high-fidelity models grow substantially with increasing model complexity. This limits their feasibility for timecritical applications, and motivates the use of model-order reduction techniques for rapid yet reliable online predictions based on prior offline computations. Therefore, a non-intrusive, purely data-driven reduced-order model (ROM) is developed that accounts for different wind conditions. The ROM is built for a two-dimensional domain that reproduces realistic building footprints, preserving key features necessary for studying the dispersion across developed areas. Due to the complexity of the problem, a global ROM covering the complete parameter space provides insufficient approximations. Thus, a localized approach is employed, setting up and aggregating local reduced-order models each corresponding to a sub-range of a split-up parameter space. Ultimately, the developed model approximates the spatio-temporal concentration field for a certain time frame with a mean relative error below 3 %, while providing predictions around 100 times faster than the original full-order high-fidelity model.