Machine Learning Predicts Pedestrian Wind Flow from Urban Morphology and Prevailing Wind Direction

The common approach to determining the urban wind field is numerical simulations which are computationally expensive and require expert knowledge. This study develops an alternative convolutional neural network model with U-net architecture, to predict wind speed distributions from urban morphologies and prevailing wind directions. The model is trained and tested with a comprehensive set of simulations, covering 512 realisations of urban configurations with representative morphological parameters seen in global cities. The model is effective in the out-of-sample scenarios, achieving R$^2$ scores higher than 0.94 and 0.65 for idealised and realistic configurations, respectively. The variances in model performance are evaluated from flow patterns and statistics and attributed to geometry complexity, which induces localised areas of wind extreme that are harder to capture. The present model facilitates higher resolution and faster evaluation of the pedestrian wind and the attributions of prediction quality guide future model development and expansion of the airflow dataset.