ECMWF short-term prediction accuracy improvement by deep learning.

Area of Science:

• Meteorology
• Artificial Intelligence
• Data Science

Background:

• Global Numerical Weather Prediction (NWP) models like ECMWF offer wide coverage but can lack localized accuracy.
• Local weather models provide higher precision for specific regions but are often geographically limited and may require payment.
• Improving the accuracy of freely available global NWP forecasts is crucial for broader accessibility and application.

Purpose of the Study:

• To develop and evaluate a neural network-based calibration model for post-processing ECMWF forecasts.
• To enhance the accuracy of short-term (up to 3 days) predictions for near-surface air temperature and 24-hour accumulated precipitation.
• To bridge the accuracy gap between global NWP models and more precise local weather models.

Main Methods:

• A neural network model was designed for post-processing meteorological data.
• The model utilizes raw ECMWF forecast data combined with identified error-correction parameters.
• Training and validation were performed using ground truth data from weather stations in 10 European cities.

Main Results:

• The proposed calibration model demonstrated superior accuracy compared to standard ECMWF predictions.
• The model's performance in cross-validation showed results approaching the precision of local NWP models.
• The enhanced ECMWF forecasts offer improved reliability for key meteorological parameters.

Conclusions:

• The neural network calibration model effectively improves the accuracy of global NWP forecasts.
• This approach offers a viable method to enhance the utility of freely available weather prediction data.
• The findings suggest a pathway for more accurate and accessible short-term weather forecasting.