Improving seasonal forecasts of air temperature using a genetic algorithm

Area of Science:

• Climate Science
• Atmospheric Science
• Machine Learning Applications

Background:

• Accurate seasonal air temperature forecasts are crucial for societal planning but are challenged by atmospheric internal variability.
• Ensemble forecasting methods are used to capture prediction signals, but managing large prediction spreads requires techniques to mitigate unrealistic members.
• Weighted averaging of ensemble members is a key technique for refining seasonal forecast accuracy.

Purpose of the Study:

• To apply a machine learning technique, specifically a genetic algorithm (GA), to optimize ensemble members for seasonal air temperature forecasts.
• To derive optimal weights for 24 ensemble members of the SINTEX-F2 coupled general circulation model for boreal summer forecasts.
• To evaluate the effectiveness of the GA-derived weights in improving the accuracy and spatial representation of 2m-air temperature anomalies.

Main Methods:

• Utilized a genetic algorithm (GA) to determine optimal weights for ensemble members of the SINTEX-F2 climate model.
• Applied the GA to the Scale Interaction Experiment-Frontier research center for global change version 2 (SINTEX-F2) model's boreal summer forecasts.
• Compared the performance of GA-weighted ensemble means against unweighted ensemble means for predicting 2m-air temperature anomalies.

Main Results:

• The GA technique significantly improved the prediction of 2m-air temperature anomalies over South America, North America, Australia, and Russia.
• The spatial distribution of predicted air temperature anomalies showed better representation using the GA-derived weights.
• Improvements were particularly noted in mid- and high-latitude regions where model skill is typically modest.

Conclusions:

• Machine learning techniques, such as genetic algorithms, can effectively enhance the accuracy of seasonal air temperature forecasts.
• Optimizing ensemble member weights using GA leads to improved regional climate predictions.
• The GA approach offers a promising method for refining climate model outputs, especially in regions with moderate model skill.