Feilian-online

Select Overpass API:

Resolution (m): Filename: Airflow Direction (deg): U10m (m/s):

Show terrain Show wind field

References

Lu, J., Li, W., Hobeichi, S., Azad, S., & Nazarian, N. Machine learning predicts pedestrian wind flow from urban morphology and prevailing wind direction. Environmental Research Letters, 2025. https://doi.org/10.1088/1748-9326/adc148
Snaiki, R., Lu, J., Li, S., & Nazarian, N. A Hierarchical Deep Learning Model for Predicting Pedestrian-Level Urban Winds. Building and Environment, Article 114354. https://doi.org/10.1016/j.buildenv.2026.114354

1. Objective

Given wind direction and a single-point 10 m background wind speed, predict the full-domain pedestrian-level wind speed field.

Input: (θ_wind, U_10m) → Output: U_ped(x, y)

2. Simulation Setup (PALM)

All training cases share identical physics; only wind direction and topography differ.

Parameter	Value
Driving mechanism	Constant pressure gradient ∂P / ∂x = -3.828125 × 10^-4 Pa/m
Friction velocity	u_* = 0.2137 m/s, determined by the pressure gradient
Roughness length	z₀ = 0.01 m
Atmospheric stability	Neutral, top_heatflux = 0
Coriolis	None, Ω = 0

Because the pressure gradient and domain height are identical across all cases, u_* is a fixed constant for the entire training dataset.

3. Undisturbed Reference Wind Speed

For a horizontally homogeneous turbulent boundary layer over flat terrain, the mean wind profile follows the neutral log law:

V_inflow(z) = (u_* / κ) ln((z + z₀) / z₀)

where κ = 0.41 is the von Karman constant. At z_ref = 10 m:

V_inflow(10 m) = (0.2137 / 0.41) ln(10.01 / 0.01) ≈ 3.60 m/s

This is a single scalar reference wind speed shared by all training cases.

4. Scaling Used Online

Feilian predicts the pedestrian wind field under the reference pressure-gradient forcing. The online tool scales that prediction by the user-provided or Open-Meteo-derived 10 m wind speed:

U_scaled(x, y) = U_Feilian(x, y) × U_10m / 3.60

If Open-Meteo returns wind speed in km/h, the tool first converts it to m/s before applying this scaling.

This linear scaling assumes the Reynolds number is sufficiently high that the pedestrian-level wind pattern is approximately Reynolds-number independent. Under this assumption, changing the background forcing mainly rescales velocity magnitude while preserving the normalized spatial wind pattern.

The scaled field is therefore an estimate under a high-Reynolds-number linear-scaling assumption, not a new LES simulation with a different pressure gradient.