Research on Wind Field Correction Method Integrating Position Information and Proxy Divergence.

Area of Science:

• Atmospheric Science and Meteorology
• Artificial Intelligence and Machine Learning

Background:

• Numerical weather prediction accuracy is limited by sparse ground-based wind field observations.
• Current correction models often rely on interpolated reanalysis data (e.g., ERA5), introducing inaccuracies.
• A need exists for high-precision wind field correction using direct observational data.

Purpose of the Study:

• To propose PPWNet, a novel deep learning model for accurate wind field correction.
• To leverage sparse, discrete observation data as the ground truth.
• To integrate positional information and physical consistency into the model.

Main Methods:

• PPWNet encodes observation point positions and uses observation values within the loss function.
• A parallel dual-branch DenseInception network extracts multi-scale grid features.
• An attention mechanism, inspired by PointNet, efficiently processes sparse, irregular observation data.
• A physics-informed approach incorporates a learned physical consistency term into the loss function.
• Hyperparameter tuning is performed using the Bayesian TPE algorithm.

Main Results:

• PPWNet significantly outperforms traditional and existing deep learning methods.
• Mean Absolute Error (MAE) is reduced by 38.65%, and Root Mean Square Error (RMSE) by 28.93%.
• The corrected wind field demonstrates superior agreement with observed wind speed and direction.

Conclusions:

• PPWNet effectively corrects wind fields by directly utilizing sparse observation data.
• Integrating positional information and physical laws enhances deep learning-based wind field correction accuracy.
• The model offers a promising advancement for improving numerical weather model initial conditions.