A Physics-Informed Neural Network Framework for Quasi-Steady Wind Retrieval From Doppler Lidar Observations

Abstract

Accurate and rapid retrieval of quasi-steady wind fields is crucial to many operational applications, such as air traffic management. Classical three-dimensional variational methods (3DVAR) are widely used in such scenarios, but their grid-based optimization framework makes it challenging to balance accuracy and efficiency. To address this problem, this study proposes a mesh-free retrieval framework based on physics-informed neural networks (PINNs) and introduces physical constraints tailored to quasi-steady wind retrieval. Experiments based on simulated datasets verify the superiority of the PINN framework over classical 3DVAR for quasi-steady wind retrieval. Results show that compared with the best performing 3DVAR baseline, the proposed method reduces the retrieval error by 69.8%, with improved end-to-end computational efficiency under the tested implementation setting. At the same time, compared with the existing PINN method constrained by the Navier–Stokes equations, it reduces the computational cost by 74.1% with only a 0.9% increase in retrieval error. Real-data validation further confirms its accuracy–efficiency advantage in the tested field cases.