HARMO21: Sensitivity analysis of microscale pollutant dispersion large-eddy simulations for observation network design

Access

• Extended abstract
• Presentation (Data assimilation for microscale pollutant atmospheric dispersion)

Abstract

High-fidelity, scale-resolving simulations of microscale urban atmospheric flows allow to improve our understanding of the complex multi-physics processes and develop a multi-fidelity approach, which could be operationally used to study accidental exposure to air pollutants.

In this work, we reproduce the MUST (Mock Urban Setting Test) plume dispersion field campaign ¹ thanks to high-fidelity Large-Eddy Simulations (LES) including parametric turbulence injection to have consistent unsteady boundary conditions. Still, important differences remain between LES model predictions and experimental data. These discrepancies are essentially explained by the high uncertainties on the model boundary conditions, the lack of knowledge on the surface heterogeneities and the inherent variability of the atmospheric boundary layer ². A sensitivity study is carried out using an ensemble of LES to explore the model dependencies on boundary condition parameters (wind direction and magnitude, turbulence intensity and surface roughness). Specific treatment is introduced to efficiently implement wind direction uncertainties (through a rotating domain method) and to accurately estimate wind direction statistics (through circular statistics).

The analysis of the LES ensemble variability shows that the model concentration prediction uncertainty has a spatial structure. Sensitivity analysis is applied to identify which boundary condition parameters are most responsible for this model uncertainty in a given area of interest. Results could be used to define optimal observation experiment design, i.e. to define the sensor locations that will provide the most informative data to reduce model prediction uncertainty. The experimental design could eventually be combined with a reduced-order model ³ in a data-driven modelling framework to aggregate fast model predictions with concentration measurements of optimally-located sensors. Such framework, by controlling the most influential parameter uncertainties, should allow to improve microscale pollutant dispersion model predictions.

Figure 1: First-order Sobol indices of the mean concentration field at 𝑧 = 1.6 m with respect to the mean inlet wind direction (left) and to the prescribed friction velocity (right).