Reconstruction of stochastic wind field based on the probabilistic wind spectra constructed by the PI-WGANDIV model

Due to the inherent uncertainties in boundary-layer winds, the probabilistic information of turbulent wind spectra is a key concern in wind engineering. However, its accuracy is often constrained by insufficient historical strong wind data. Generative Adversarial Networks (GANs) can realize data augmentation by generating high-quality samples. However, the lack of physical constraints in standard GANs might produce invalid samples deviating from the real wind environment. Therefore, a physics-informed Wasserstein GAN with divergence (PI-WGANDIV) is proposed to generate physically consistent spectral parameters under data-scarce conditions for constructing probabilistic wind spectra. It involves an original WGANDIV network and two key physical constraints defined by real wind data. The first is the bounded distributions of spectral parameters, while the other is the correlation coefficients between different spectral parameters. This integration enforces the admissible interpolation space within the observed data envelope, ensuring the generated samples remain physically plausible. The superiority of the proposed method is demonstrated through comparisons with Monte Carlo resampling, WGANDIV, and WGANGP. Subsequently, the probabilistic wind spectra in the high wind velocity domain are constructed through the proposed model for the Jiangyin Bridge site. The stochastic wind field is reconstructed based on the constructed probabilistic spectra. Finally, statistical comparisons within the measured wind speed range indicate that the traditional extrapolation-based model tends to overestimate extreme wind speeds.