Machine Learning Predicts Pedestrian Wind Flowfrom Urban Morphology and Prevailing WindDirection

Pedestrian-level wind plays a critical role in shaping the urban microclimate and is
significantly influenced by urban form and geometry. The most common method for
determining spatial wind speed patterns in cities relies on numerical computational
fluid dynamics (CFD) simulations, which resolve Navier-Stokes equations around
buildings. While effective, these simulations are computationally intensive and require
specialised expertise, limiting their broader applicability. To address these limitations,
this study proposes a more cost-effective alternative while achieving 90% performance
in capturing the mean and maintaining spatial wind patterns captured by CFD. We
developed a machine learning (ML) approach with U-net architecture to predict time
mean wind speed patterns from prevailing wind directions and three-dimensional
urban morphology, which are increasingly available for global cities. The model
is trained and tested using a comprehensive dataset of 512 numerical simulations
of urban neighbourhoods, representing diverse morphological configurations in cities
worldwide. We find that the ML algorithm accurately predicts complex wind patterns,
achieving a normalised mean absolute error of less than 10%, which is comparable
to wind anemometer measurement in a low wind speed environment. In predicting
wind statistics, the ML model also surpasses that of regression models based solely
on statistical representations of urban morphology. The R2 values measuring grid-
level agreement between ML and CFD range from 0.94-0.99 and 0.65-0.95 for the
idealised and whole datasets, respectively. However, we find that grid-based R2 is not
an effective metric for evaluating the 2D model performance due to localised biases
arising from faster wind speed grid regions, which is revealed by the wind probability
density function. These findings demonstrate that complex pedestrian wind patterns
can be effectively predicted using an image-based ML approach, offering the potential
to emulate physics-based LES models, which are computationally expensive, thereby
significantly reducing computing costs.