High-resolution Estimation of Average Wind Velocity and Pressure Distributions in a Two-dimensional Urban Street Canyon using Physic-Informed Neural Networks

Abstract

Wind tunnel experiments can be conducted as a precise method for measuring airflow and pressure around buildings. However, the inherent measurement limitations and cost make it challenging to capture detailed distributions. In this study, we implement a physics-informed neural network (PINN) to estimate the average flow field and surface pressure in a twodimensional urban street canyon from sparse velocity information which is assumed to be provided by wind tunnel experiments. High-resolution datasets from a large-eddy simulation provide simulated velocity measurement data for various sensor configurations. As part of the governing equations for the PINN model, a forcing vector method is employed to replace Reynolds stress, thereby reducing computational costs. As a result, the PINN demonstrated high prediction accuracy for the entire detailed flow field. Additionally, the wind pressure on the surfaces was estimated without any pressure measurements. However, regions near building surfaces and shear layers, which exhibit large velocity gradients, still present relatively large errors. To address this issue, supplementary sensors improved performance and increased prediction accuracy. The proposed method shows promising potential in facilitating efficient assessments within wind tunnel experimentation settings.