Improving local pedestrian-level wind environment based on probabilistic assessment using Gaussian process regression

Many wind comfort assessment standards have been proposed to evaluate wind comfort and safety in spaces with intended use for pedestrians, such as pathways, building entrance areas, amenity spaces, and outdoor sitting spaces. However, the optimization of the wind environment presents a unique combination of complexities in assessment, simulation, and design. Conventionally, designers rely heavily on wind tunnel experiments or computational fluid dynamics simulations based on which only limited proposals are examined to select the one with the best wind performances. A new design framework for proactively improving the wind conditions at various measuring positions is demonstrated in this paper. Gaussian process regression models are combined with the Latin hypercube sampling method to approximate the relationships between design variables and wind velocities at discrete locations. Gaussian process regression is a non-parametric model well-suited for modeling the non-linear and stochastic wind behaviors. Moreover, it can provide the prediction mean and variance, and the latter is incorporated into a prediction quality function to quantify the modeling uncertainty. In the case study, with the objectives of maximizing the winter and summer wind comfort and minimizing the modeling errors, the robust and near-optimal designs of a target building in an infill development project are explored using the evolutionary search algorithm. We believe that this framework can be migrated to other urban design practice to better predict and simultaneously optimize the microclimate performances at locations of interest, such as solar access, indoor ventilation efficiency, and thermal comfort.

Automated summary

This study introduces a new design framework for improving wind conditions by using Gaussian process regression models and Latin hypercube sampling method to approximate the relationships between design variables and wind velocities. This approach can better predict and optimize microclimate performances at locations of interest.