Autoencoder-based flow-analogue probabilistic reconstruction of heat waves from pressure fields

Area of Science:

• Meteorology
• Data Science
• Machine Learning

Background:

• Accurate reconstruction of meteorological fields is crucial for weather forecasting and climate studies.
• Traditional methods like the analogue method (AM) can be limited by high-dimensional, noisy predictor data.

Purpose of the Study:

• To introduce and evaluate a novel hybrid approach, AE-AM, for probabilistic reconstruction of meteorological fields.
• To compare the performance of AE-AM against the classical AM in reconstructing temperature fields during heat waves.

Main Methods:

• A deep autoencoder (AE) is trained on predictor fields to create a compressed latent space.
• The analogue method (AM) is applied within this latent space to identify historical analogues for reconstruction.
• The AE-AM approach is tested on reconstructing daily maximum temperature from sea-level pressure during European heat waves (1950-2010).

Main Results:

• The AE-AM approach significantly outperforms the standard AM in reconstructing the magnitude and spatial patterns of heat wave temperature events.
• Improvements in skill score ranged from 7% to 22%, varying by the specific heat wave analyzed.
• The AE-AM method demonstrated enhanced reconstruction accuracy by filtering irrelevant information in the predictor fields.

Conclusions:

• The hybrid AE-AM method offers a significant advancement over classical AM for meteorological field reconstruction.
• Deep autoencoders effectively reduce dimensionality and extract relevant features, improving analogue-based predictions.
• This approach shows considerable potential for improving the probabilistic reconstruction of extreme weather events.