Pedestrian-level low-occurrence wind speeds in an urban area predicted by artificial neural networks from fundamental statistics

Promoting urban sustainability requires a comprehensive understanding of the interaction between built and surrounding environments. Low-occurrence wind is a key factor affecting building performance, urban breathability, and human comfort and safety. This study proposes artificial neural network (ANN) models to estimate low-occurrence wind speeds using wind speed statistics from field measurements, and to investigate the relationship between these statistics and pedestrian-level low-occurrence wind speeds. The performance of ANN models was compared with traditional statistical methods, including Gram–Charlier series (GCS) and Weibull models (2W and 3W). It was found that ANN models estimate percentile values more accurately due to their data-driven capabilities. The proportion of negative outliers for 0.1 % percentile wind speeds decreased by approximately 70 % compared to the GCS models and 40 % compared to the 3W model. The accuracy for 99.9 % percentile wind speeds in ANN models and traditional statistical methods with input orders higher than the second order improved by approximately 30 % compared to the Gaussian and 2W models. Additionally, the estimation accuracy of ANN models improved with data clustering. This study highlights the potential of ANN models in advancing urban sustainability by accurately estimating pedestrian-level low-occurrence wind speeds from fundamental statistics, contributing to sustainable urban development.