Wind Pressure Prediction for Ridge–Valley Membrane Structures Using a Multi-Mechanism Enhanced Physics-Informed Neural Network
Buildings · 20.07.2026 · 16 (14) · 2887
DOI: 10.3390/buildings16142887 ↗Kurz gefasst
Fachartikel in der Zeitschrift Buildings, online erschienen am 20.07.2026. Verfasst von Fang-Jin Sun und 3 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Guangxi Science and Technology Program, Natural Science Foundation of Guangxi, National Natural Science Foundation of China.
Abstract (Originalsprache)
To address the challenge of reconstructing statistical wind pressure fields of membrane structures under sparse measurement conditions, this study proposes an FF-Res-GP-PINN framework based on wind tunnel test data from a ridge–valley membrane structure. The proposed framework integrates Gaussian Fourier feature mapping, a residual network, spatial gradient penalty, and Laplacian smoothing regularization. The model takes the spatial coordinates and incoming wind direction as inputs and outputs the mean and fluctuating wind pressure coefficients. A total of 675 samples under three wind directions, namely 0°, 45°, and 90°, were used for training and validation. The results show that compared with the conventional PINN, the proposed FF-Res-GP-PINN reduces the validation MSE from 0.226 to 0.155, corresponding to a reduction of 31.4%. The ablation study indicates that Gaussian Fourier feature mapping contributes most significantly to the improvement of pointwise prediction accuracy, whereas the spatial gradient penalty and Laplacian smoothing regularization mainly enhance the spatial continuity of the predicted wind pressure field. Compared with the Fourier-MLP model, the FF-Res-GP-PINN reduces the average gradient norm, Laplacian smoothing residual, and spatial oscillation index by approximately 98.5%, 96.3% and 62.2%, respectively. The sensitivity analysis of regularization weights further demonstrates that increasing λgrad and λlap can enhance spatial smoothness, but may also increase the pointwise prediction error. Therefore, λgrad = 2 × 10−3 and λlap = 1 × 10−4 were adopted in this study as a compromise between prediction accuracy and spatial regularization for the current dataset. The findings indicate that the proposed framework provides a feasible approach for reconstructing statistical wind pressure fields of membrane structures with a specific geometry and tested wind directions. Nevertheless, its predictive capability for unseen wind directions, different structural geometries, and transient wind pressure fields requires further validation.
Förderung
- Guangxi Science and Technology Program
- Natural Science Foundation of Guangxi
- National Natural Science Foundation of China
Zitieren
APA
Sun, F., He, Q., Luo, Q., & Zhang, D. (2026). Wind Pressure Prediction for Ridge–Valley Membrane Structures Using a Multi-Mechanism Enhanced Physics-Informed Neural Network. Buildings, 16(14), 2887. https://doi.org/10.3390/buildings16142887
Vancouver
Sun F, He Q, Luo Q, Zhang D. Wind Pressure Prediction for Ridge–Valley Membrane Structures Using a Multi-Mechanism Enhanced Physics-Informed Neural Network. Buildings. 2026;16(14):2887. doi:10.3390/buildings16142887
BibTeX
@article{Sun2026,
author = {Sun, Fang-Jin and He, Qing-Cheng and Luo, Quan and Zhang, Da-Ming},
title = {Wind Pressure Prediction for Ridge–Valley Membrane Structures Using a Multi-Mechanism Enhanced Physics-Informed Neural Network},
journal = {Buildings},
year = {2026},
volume = {16},
number = {14},
pages = {2887},
doi = {10.3390/buildings16142887}
}
EUVIMED