Accurate medium-range global weather forecasting with 3D neural networks.

Area of Science:

• Meteorology
• Artificial Intelligence
• Computational Science

Background:

• Numerical Weather Prediction (NWP) is the current standard for accurate weather forecasting but is computationally intensive.
• Artificial intelligence (AI) methods show promise for accelerating weather forecasts but currently lack the accuracy of NWP.
• Accurate medium-range global weather forecasting remains a significant scientific and societal challenge.

Purpose of the Study:

• To introduce an AI-based method for accurate, medium-range global weather forecasting.
• To demonstrate the effectiveness of deep learning models with Earth-specific priors for weather prediction.
• To reduce accumulated errors in medium-range forecasts using a hierarchical temporal aggregation strategy.

Main Methods:

• Developed Pangu-Weather, a deep learning model utilizing 3D deep networks with Earth-specific priors.
• Implemented a hierarchical temporal aggregation strategy to mitigate error accumulation.
• Trained the model on 39 years of global weather data.

Main Results:

• Pangu-Weather achieved superior deterministic forecast results compared to the European Centre for Medium-Range Weather Forecasts (ECMWF) operational integrated forecasting system.
• The AI model demonstrated strong performance in medium-range forecasting across all tested variables.
• Pangu-Weather also showed effectiveness in extreme weather forecasting, ensemble forecasts, and tropical cyclone tracking.

Conclusions:

• AI-based methods, specifically Pangu-Weather, can achieve highly accurate medium-range global weather forecasts.
• Deep networks with Earth-specific priors and hierarchical temporal aggregation are effective for complex weather pattern analysis.
• This AI approach offers a computationally efficient and accurate alternative to traditional NWP systems.