Pre-training enhanced physics-informed neural network with refinement mechanism for wind field reconstruction

Accurate and detailed spatiotemporal wind field characterization is of paramount importance for various wind energy applications, including real-time wind turbine control, power prediction, optimal wind farm layout design, etc. However, current wind speed measurement technologies, particularly Light Detection and Ranging (LiDAR) systems, can only provide measurement data at sparse locations. To address this limitation, a novel pre-training enhanced physics-informed neural network with the refinement mechanism is proposed for full-field wind field reconstruction, which effectively integrates sparse measurement data with fundamental fluid dynamics principles. The proposed pre-training paradigm facilitates knowledge transfer from diverse wind field datasets, enabling the model to establish robust initial parameters for target wind field reconstruction, thereby significantly improving both training efficiency and reconstruction accuracy. Furthermore, a diffusion model-based refinement mechanism is introduced to enhance reconstruction quality by restoring fine-grained details and rectifying erroneous regions. Experimental results show that the proposed framework can achieve superior performance in terms of reconstruction accuracy, computational stability, and efficiency. The robustness with respect to variations in spatial and temporal measurement resolution, wind conditions, and measurement noise has also been validated.