A new data-driven model for rapid urban low-level wind field simulation using conditional generative adversarial networks

Abstract

Rapid and accurate representation of urban low-altitude wind fields is essential for understanding urban wind environments, assessing ventilation performance, and supporting emerging low-altitude aviation applications. Conventional computational fluid dynamics (CFD) provides reliable flow-field predictions but remains computationally expensive for rapid analysis in dense urban areas. In this study, an artificial intelligence-based data-driven wind-field reconstruction model based on a conditional generative adversarial network (cGAN) was developed to enable fast prediction of strong-wind conditions in urban street canyons. High-resolution CFD results, together with urban building morphology and inflow conditions, were used to establish a direct mapping from building-condition inputs to wind-field outputs, with Shenzhen's high-density central district selected as the study area. The cGAN model accurately reproduced major flow structures and local wind features, achieving a 10 m (m) height wind field prediction with a mean absolute error (MAE) of 1.52 m per second (m/s), a root mean square error (RMSE) of 2.71 m/s, and a structural similarity index measure (SSIM) of 0.80; for the typhoon scenario, the corresponding values are an MAE of 1.37 m/s, an RMSE of 2.54 m/s, and an SSIM of 0.80. Moreover, it reduced the computational time from hours to seconds while preserving the key physical characteristics of urban strong-wind flows. These results demonstrate the potential of combining artificial intelligence with physics-based modelling for efficient urban wind field reconstruction. The proposed framework provides a practical tool for ventilation analysis, urban planning, design optimisation, and real-time wind-field inference in low-altitude operational scenarios.