Presentation pdf 2.7MB

Surrogate-modelling the effect of building shape on the local wind microclimate through multi-directional, multi-fidelity CFD

Tall buildings affect the local wind microclimate in their vicinity, whereby pedestrian areas feel windier due to downdraughts, funneling and wind accelerating around corners. Studies into the aerodynamic performance of the building’s massing earlier in the design process may allow significant improvements to the wind microclimate before the massing is fixed.

This study aims to characterise the aerodynamic effect of shape modifications to a tall building on the wind microclimate through the construction of a multi-fidelity surrogate model. Surrogate modelling is a technique where a mapped response surface is generated from a limited set of CFD data, allowing the parameter design space to be queried with significantly less expense. With enough data points, a surrogate model will be a good approximation of the true response, however any modelling error associated with the simulations is inherited by the model. An efficient way of reducing the global error is to supplement the simulation data with a small number of high-fidelity simulation data.

In order to test this approach in the context of wind microclimate, a benchmark case was introduced (AIJ, 2005)[a]. Comparisons were made between wind tunnel measurements in the AIJ case and the RANS and time-averaged DES simulations employed in the current study, in order to quantify the model error in the two approaches.

Subsequently, a tall building in the benchmark case was used as the target building whose shape was varied. The surrogate model, the construction of which was automated using a suite of bespoke Python scripts, was populated with low-fidelity RANS and high-fidelity DES results for a number of building shapes from multiple wind directions. The design variable was a single parameter which defined the amount of taper on the top surface of the building. A Latin Hypercube Sampling strategy was employed to uniformly sample the design space. The HELYX Core-3.2.0 fork of OpenFOAM was used for the simulations.

A form of Gaussian process regression called co-kriging was used to build a response surface based on the sampled data points. A unique characteristic of co-kriging is that it can interpolate a particular attribute based on both the low- and high-fidelity data, resulting in a more accurate model when compared to a single low-fidelity approach.

Results from the current study show that predictions of pedestrian wind comfort and safety at ground level are robust using high-fidelity CFD simulations for multiple wind directions. RANS simulations are found to be less reliable, but act as a necessary tool to efficiently populate the surrogate. Model errors are discussed in relation to the amount of taper and the wind direction considered. In general, a more tapered building shape is found to be beneficial for wind comfort at ground level, however a trade-off exists between that and lettable area inside the building.

References [a] – Y. Tominaga et al., Cross Comparisons of CFD Prediction for Wind Environment at Pedestrian Level around Buildings (Part 2), The 6th APAC Conf. on Wind Engineering, Seoul, Korea (2005) 2661-2670