Two-stage U-Net predicts pedestrian wind speeds across urban morphologies
AI model reduces computation while improving spatial coherence with iterative refinement.
A two-stage U-Net framework was developed for efficient prediction of time-averaged pedestrian-level wind speed over realistic urban morphologies. Using the UrbanTALES dataset, the first stage (M1) predicts wind fields patch-by-patch, and a second inpainting U-Net (M2) reduces boundary discontinuities via a Gauss-Seidel iterative scheme. The model reproduces mean velocity and spatial variability reasonably well, though maximum velocities remain underestimated. This surrogate enables efficient high-resolution pedestrian-level wind prediction without the computational expense of high-fidelity simulations like LES.
- Two-stage U-Net framework uses patch-by-patch prediction (M1) and inpainting refinement (M2) with Gauss-Seidel iteration.
- Trained on the UrbanTALES dataset with realistic city configurations and multiple wind directions.
- Achieves efficient, high-resolution wind speed prediction while reducing discontinuities; underestimates peak velocities.
Why It Matters
Enables rapid wind-comfort assessment for urban planning, replacing expensive LES simulations with an AI surrogate.