Predicting wind flow around buildings using deep learning

Abstract

The wind velocity field around buildings provides deep insights into the aerodynamic characteristics of buildings and indicates the pedestrian-level wind environment around buildings. Particle image velocimetry (PIV) is usually employed to measure the wind velocities around building models. Due to laser-light shielding, measuring instantaneous wind velocities at some shielded locations around a building model remains difficult. As a result, analyzing the wind flow pattern with these unmeasured wind velocities is difficult. Using machine learning techniques to impute unmeasured values allows for a comprehensive study of wind flow patterns with laser-light shielding. Unmeasured velocities around building models were imputed in this study using machine learning (ML) models such as the generative adversarial imputation network (GAIN), multiple imputations by chained equations (MICE), and neighbored distanced imputation (NDI). GAIN was the best model with a minimum variance and standard deviation of 1.508 and 1.228, respectively. Compared with experimental wind velocities, GAIN produced the minimum average mean squared error of 2.4%. The correlation between the experimental and predicted wind velocities was 98.2%. Thus, the validated GAIN model is recommended to be integrated into the PIV study to impute the unmeasured wind velocities to obtain a complete wind flow pattern.