Using Deep Learning Model to Simulate Wind in Urban Area
N. Liu (TU Delft - Architecture and the Built Environment)
Azarakhsh Rafiee – Mentor (Vrije Universiteit Amsterdam)
Frits de Prenter – Mentor (TU Delft - Wind Energy)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Urban wind simulations are essential for assessing pedestrian comfort, pollutant dispersion, and microclimate design, yet high-fidelity CFD remains computationally expensive. This thesis investigates the potential of a Swin-Transformer-based surrogate model to approximate steady-state 2D results for 100 m × 100 m tiles under a single inflow direction. A dataset of 690 urban tiles was extracted from the 3DBAG and simulated in Ansys Fluent at 0.5 m resolution across five inflow speeds (5–15 m/s) and five types of building layouts. Velocity fields were rasterized to 1 m grids. The surrogate architecture preserves the original Swin backbone, evaluated with three loss variants: original RMSE, buffer weighting, and a divergence penalty. Encoding wind as xy-component consistently outperformed magnitude-only training. Errors scaled linearly with inflow speed and were highest in Mixed and Attached urban forms. Architectural resolution was the dominant factor influencing accuracy since it introduced visible artifacts, while the buffer and divergence losses offered only marginal improvements. Limitations remain near building facades, where sharp gradients are smoothed by the patch-based architecture. Nonetheless, the surrogate offers potential for rapid wind flow estimation suitable for early-stage design or preliminary analysis.