Fast simulation of high resolution urban wind fields at city scale

Abstract

High resolution urban wind field simulations are limited in small simulation domain, short simulation period, or stable boundary conditions due to the high computational requirements. To address these limitations, non-intrusive reduced order models (NIROMs) are developed in this study and first realized the prediction of high-resolution 3D wind field in a city-scale area (~150 km2) during a long simulation period (4 weeks). Within the NIROM, 3D convolutional autoencoder and eXtreme Gradient Boost regression models (XGBoost) are adopted. The urban airflow can be predicted rapidly by the XGBoost-decoder models with the input of boundary conditions via boundary conditions - latent variable - full-order wind fields. By comparing NIROM simulations to large-eddy simulations, we find that the NIROM with portable model size is capable for capturing the main spatial distribution characteristics of urban airflow. Moreover, the relationship between wind speed and building distribution indicates that the sparse layout with taller buildings can yield a 2-fold increase in the near-ground wind speed compared to the dense layout with lower buildings, which also benefits pedestrian thermal comfort and pollutant dissipation. Furthermore, this study can assist the optimization of urban ventilation and air quality in urban planning.