Wind pressure acting on building envelopes often demonstrates extreme peak values and high-frequency fluctuations,posing significant challenges for structural safety assessment.This study proposes a wind pressure time...Wind pressure acting on building envelopes often demonstrates extreme peak values and high-frequency fluctuations,posing significant challenges for structural safety assessment.This study proposes a wind pressure time series prediction framework,termed C-WGAN-timesplit,that combines consistency constraints,Wasserstein generative adversarial networks(WGAN),and a time-series cross-validation strategy.C-WGAN-timesplit predicts the full pressure field using data from sparse pressure taps through learning the mapping from subset sensors to the full pressure field.The performance of C-WGAN-timesplit was evaluated and compared against generative adversarial networks(GAN)and WGAN using boundary layer wind tunnel test pressure data.Results across multiple evaluation metrics confirm that C-WGAN-timesplit achieves superior predictive accuracy under various wind incidence angles.Comprehensive analyses were conducted across input data volume sensitivity and incidence angles through aerodynamic coefficients,pressure statistics,non-Gaussian characteristics,and local pressure time series.The proposed framework achieved accurate mean pressure prediction using only 1 tap(0.25%coverage),reliable fluctuation prediction with 20 taps(5%),and high-fidelity reconstruction of non-Gaussian pressure characteristics with 40 taps(10%).Furthermore,three representative taps exhibiting pronounced non-Gaussian behavior under typical incidence angles were used to assess time-series prediction performance.The model effectively captured fluctuation features under both 20-tap and 40-tap configurations,though some deviations remained in predicting rare extreme values.Overall,the proposed method provides a promising solution for wind load prediction on building surfaces,with future efforts needed to improve the representation of transient extreme events.展开更多
基金the open funding of the Key Laboratory of Transport Industry of Wind Resistant Technology for Bridge Structures(KLWRTBMC24-02)the National Natural Science Foundation of China(52008239,52478534,51978392)Fujian Provincial Department of Science and Technology,China(2023Y0040).
文摘Wind pressure acting on building envelopes often demonstrates extreme peak values and high-frequency fluctuations,posing significant challenges for structural safety assessment.This study proposes a wind pressure time series prediction framework,termed C-WGAN-timesplit,that combines consistency constraints,Wasserstein generative adversarial networks(WGAN),and a time-series cross-validation strategy.C-WGAN-timesplit predicts the full pressure field using data from sparse pressure taps through learning the mapping from subset sensors to the full pressure field.The performance of C-WGAN-timesplit was evaluated and compared against generative adversarial networks(GAN)and WGAN using boundary layer wind tunnel test pressure data.Results across multiple evaluation metrics confirm that C-WGAN-timesplit achieves superior predictive accuracy under various wind incidence angles.Comprehensive analyses were conducted across input data volume sensitivity and incidence angles through aerodynamic coefficients,pressure statistics,non-Gaussian characteristics,and local pressure time series.The proposed framework achieved accurate mean pressure prediction using only 1 tap(0.25%coverage),reliable fluctuation prediction with 20 taps(5%),and high-fidelity reconstruction of non-Gaussian pressure characteristics with 40 taps(10%).Furthermore,three representative taps exhibiting pronounced non-Gaussian behavior under typical incidence angles were used to assess time-series prediction performance.The model effectively captured fluctuation features under both 20-tap and 40-tap configurations,though some deviations remained in predicting rare extreme values.Overall,the proposed method provides a promising solution for wind load prediction on building surfaces,with future efforts needed to improve the representation of transient extreme events.
