ESTIMATION OF INSTANTANEOUS AIRFLOW DISTRIBUTION IN URBAN MODEL USING GENERATIVE ADVERSARIAL NETWORK

Abstract

Urban airflow has strong turbulence characteristics, and its distribution varies with time. Strong winds in the urban area may harm pedestrians. Calculating the urban airflow distribution quickly can protect pedestrians from danger. In this study, a machine learning based method, Generative Adversarial Network (GAN), was used to rapidly estimate the high spatial resolution distribution of instantaneous airflow field in the cubic building group model using velocity measurement obtained from sensors as inputs. We compared the results with large-eddy simulation data as true values and with the results of the Proper Orthogonal Decomposition-Linear Stochastic Estimation (POD-LSE) method which has been validated for rapid prediction of airflow distribution. GAN can estimate instantaneous airflow distribution and streamlines, and the reproduction accuracy of GAN is higher than that of POD-LSE. The probability density function of instantaneous velocity obtained from the GAN at representative points is consistent with that of the true value.