A Transfer Learning-Augmented Physics-Informed Neural Network With Spatiotemporal Adaptation for Efficient Wind Field Reconstruction

The efficient utilization of wind energy is hindered by the inherent variability and uncertainty of wind, necessitating the acquisition of accurate and detailed wind field data. Wind measurement technologies, such as Light Detection and Ranging (LiDAR), provide precise but sparse wind speed data. To address this problem, physics-informed neural network (PINN) has been employed to reconstruct wind fields by integrating measurement data with the Navier-Stokes (NS) equations. However, PINN faces three critical limitations, including optimization difficulty, challenges in handling long-term temporal dependencies, and diminished reconstruction accuracy in regions far from LiDAR coverage. This paper presents ReWind-PINN, a transfer learning-augmented physics-informed framework with spatiotemporal adaptation for efficient wind field reconstruction. Through the integration of models pre-trained on high-fidelity simulated wind data, our approach enhances PINN optimization efficiency, accelerates convergence, and improves reconstruction accuracy. We further propose a progressive temporal training strategy coupled with an adaptive residual weighting mechanism, designed to effectively address temporal dependencies while reinforcing physical consistency in measurement-sparse regions. Experimental validation confirms that ReWind-PINN delivers enhanced reconstruction accuracy with improved computational efficiency and stability. The framework demonstrates applicability across various PINN variants and different wind conditions, maintaining robust performance in the presence of measurement noise and variations in spatiotemporal resolution.