Deep learning for bias correction of MJO prediction

Area of Science:

• Atmospheric Science
• Meteorology
• Climate Dynamics

Background:

• Accurate weather prediction beyond two weeks is crucial for socioeconomic applications.
• The Madden-Julian Oscillation (MJO) is a major source of global subseasonal predictability (3-4 weeks).
• Current operational forecast systems show limited MJO prediction skill due to model-induced systematic errors.

Purpose of the Study:

• To enhance Madden-Julian Oscillation (MJO) prediction skill.
• To reduce systematic errors in subseasonal weather forecasts.
• To leverage Deep Learning for bias correction in MJO forecasting.

Main Methods:

• A Deep Learning bias correction method was developed.
• This method blends state-of-the-art dynamical weather forecasts with observational data.
• The approach was applied to multi-model MJO forecasts.

Main Results:

• Deep Learning bias correction significantly reduced MJO forecast errors.
• MJO amplitude errors were reduced by approximately 90%, and phase errors by 77% over four weeks.
• The greatest improvements were observed for MJO events originating in the Indian Ocean and crossing the Maritime Continent.

Conclusions:

• Deep Learning offers a powerful tool for correcting systematic errors in numerical weather prediction models.
• The developed bias correction method substantially improves subseasonal MJO prediction skill.
• Enhanced MJO forecasts have significant potential for improving socioeconomic outcomes reliant on weather prediction.