Reconstruction of downburst wind fields using physics-informed neural network

Abstract

Downbursts, as a strong localized wind event, have caused significant damage to engineering structures throughout the world. However, given the spatial and temporal randomness of such strong winds, on-site measurements are often difficult to obtain a sufficient amount of valid wind field information in a short period of time. To refine the resolution of the wind field, this study proposes a physics-informed neural network network-based (PINN-based) approach to reconstruct the downburst from limited observed data. The Navier-Stokes (N-S) equations are embedded into the fully connected neural network as a physical constraint to construct the PINN. The PINN model is then validated by the reconstruction of numerical downburst generated by large eddy simulations. The reconstruction of the sparse downburst wind field by PINN performs well in both interpolation and extrapolation prediction. The optimal construction of the PINN has been evaluated through parameter analysis of the influence of training data and network parameters. Finally, the optimal PINN construction is used to reconstruct the wind field of the experimental data with a relative error of 5% for the horizontal wind velocity.