Investigation of wind pressures on tall building under interference effects using machine learning techniques

Gang Hu, Lingbo Liu, Dacheng Tao, Jie Song, K. C. S. KwokarXiv 2019 · arXiv

Abstract

This research presents a novel deep learning approach using Generative Adversarial Networks (GANs) to predict wind pressure distributions on tall buildings caused by the interference effects of neighboring structures. By training on only 30% of a standard wind tunnel dataset, the model accurately predicts complex wind loads for unseen scenarios. This methodology demonstrates the potential to reduce the requirement for expensive physical wind tunnel testing by up to 70% while maintaining high predictive accuracy for both mean and fluctuating pressure components.

Introduction

In dense urban environments, skyscrapers do not exist in isolation. The proximity of neighboring structures creates "interference effects" that significantly alter wind flow patterns. These effects can either shield a building or, more critically, amplify wind loads, leading to structural vibrations, cladding damage, and occupant discomfort. Traditionally, engineers rely on extensive wind tunnel testing to map these non-linear interactions. However, the vast number of possible configurations (varying distances, angles, and heights) makes exhaustive physical testing time-consuming and costly. This study investigates machine learning as a more efficient alternative to interpolate and predict these complex aerodynamic loads.

Methodology

The study evaluates four machine learning architectures: Decision Tree Regressor (DTR), Random Forest (RF), XGBoost (XGB), and a novel two-stream Generative Adversarial Network (GAN). Unlike standard tabular models that predict pressure point-by-point, the GAN approach treats pressure data as a 2D image. This allows the model to capture spatial correlations between adjacent pressure taps, preserving the physical structure of the wind flow. The GAN features a "global-local" generator that produces an initial coarse pressure map and subsequently refines local details.

Experimental Setup

The models were trained and validated using a comprehensive aerodynamic database from Tokyo Polytechnic University (TPU). The experimental setup involved two 1:400 scale tall building models: a "principal" building equipped with 252 pressure taps and an "interfering" dummy building. The dataset encompasses 307 configurations across 72 wind attack angles, totaling 2,664 unique test cases.

Experimental setup showing the building configurations (a) and the detailed pressure tap layout (b).

Results and Discussion

The GAN model significantly outperformed traditional regression models, particularly in predicting fluctuating (RMS) pressure coefficients, which are typically harder to model due to turbulence. The model achieved an $R^2$ score of 0.988 for mean pressure coefficients and 0.924 for fluctuating components.

Visualization of mean pressure coefficient (mCp) and fluctuating pressure coefficient (rmsCp) across the building surfaces.

Key findings include:

  • Spatial Intelligence: By treating pressure distributions as images, the GAN effectively captures physical phenomena like flow separation and recirculation bubbles.
  • Data Efficiency: Accuracy improved drastically when increasing the training set from 10% to 30% of the total data, but showed diminishing returns thereafter. This suggests that 30% of a typical experimental matrix provides sufficient physical information for the model to generalize effectively.
  • Visual Fidelity: The generated pressure contour maps were qualitatively indistinguishable from experimental wind tunnel results.

Conclusion and Future Work

The application of GANs to wind engineering offers a transformative shift from brute-force testing to intelligent prediction. By accurately filling the gaps in experimental matrices, this technique can save approximately 70% of the effort required for wind tunnel test cases. Future research will focus on expanding the model to handle buildings of varying geometries and more complex urban clusters involving multiple interfering structures.

Summary updated

Abstract

Interference effects of tall buildings have attracted numerous studies due to the boom of clusters of tall buildings in megacities. To fully understand the interference effects of buildings, it often requires a substantial amount of wind tunnel tests. Limited wind tunnel tests that only cover part of interference scenarios are unable to fully reveal the interference effects. This study used machine learning techniques to resolve the conflicting requirement between limited wind tunnel tests that produce unreliable results and a completed investigation of the interference effects that is costly and time-consuming. Four machine learning models including decision tree, random forest, XGBoost, generative adversarial networks (GANs), were trained based on 30% of a dataset to predict both mean and fluctuating pressure coefficients on the principal building. The GANs model exhibited the best performance in predicting these pressure coefficients. A number of GANs models were then trained based on different portions of the dataset ranging from 10% to 90%. It was found that the GANs model based on 30% of the dataset is capable of predicting both mean and fluctuating pressure coefficients under unseen interference conditions accurately. By using this GANs model, 70% of the wind tunnel test cases can be saved, largely alleviating the cost of this kind of wind tunnel testing study.
BibTeX
@misc{hu2019investigation,
  author = {Gang Hu and Lingbo Liu and Dacheng Tao and Jie Song and K. C. S. Kwok},
  title = {{Investigation of wind pressures on tall building under interference effects using machine learning techniques}},
  year = {2019},
  eprint = {1908.07307},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1908.07307},
}
Figure