High-fidelity reconstruction of wind pressure fields on connected super-high-rise structures using machine learning

Abstract

The design of connected super-high-rise structures (CSHSs) is challenged by complex aerodynamic interference, a problem exacerbated by the sparse, discrete data from traditional wind tunnel tests (WTT) which often fail to capture critical pressure peaks. To address this limitation, a machine learning (ML) framework was established for high-fidelity, full-field wind pressure reconstruction. Leveraging comprehensive WTT data from twin- and triple-tower CSHSs, the predictive performance of XGBoost, KNN, and RF algorithms was systematically evaluated for mean, fluctuating, and transient pressures. XGBoost demonstrated superior robustness and generalizability. Critically, a model’s predictive accuracy was found to be intrinsically linked to the local aerodynamic regime. While the performance of KNN degraded in high-turbulence wake regions, where its core spatial proximity assumption is violated, XGBoost successfully captured the non-linear dynamics of vortex shedding. This study therefore establishes a reliable XGBoost-driven reconstruction methodology and provides crucial physical insights into the applicability and limitations of different ML algorithms in wind engineering.