[Paper Review] Computing the ensemble spread from deterministic weather predictions using conditional generative adversarial networks

Rüdiger Brecht, Alex Bihlo|arXiv (Cornell University)|May 18, 2022

Meteorological Phenomena and SimulationsEarth and Planetary Sciences32 references34 citations

TL;DR

This paper proposes using conditional generative adversarial networks (cGANs) to predict ensemble spread in weather forecasting directly from a single deterministic 500 hPa geopotential height forecast, bypassing costly ensemble runs. The method adapts the pix2pix architecture to 3D spatial-temporal data and achieves significantly lower RMSE and higher SSIM than climatological and persistence baselines, with model averaging further improving accuracy.

ABSTRACT

Ensemble prediction systems are an invaluable tool for weather forecasting. Practically, ensemble predictions are obtained by running several perturbations of the deterministic control forecast. However, ensemble prediction is associated with a high computational cost and often involves statistical post-processing steps to improve its quality. Here we propose to use deep-learning-based algorithms to learn the statistical properties of an ensemble prediction system, the ensemble spread, given only the deterministic control forecast. Thus, once trained, the costly ensemble prediction system will not be needed anymore to obtain future ensemble forecasts, and the statistical properties of the ensemble can be derived from a single deterministic forecast. We adapt the classical pix2pix architecture to a three-dimensional model and also experiment with a shared latent space encoder-decoder model, and train them against several years of operational (ensemble) weather forecasts for the 500 hPa geopotential height. The results demonstrate that the trained models indeed allow obtaining a highly accurate ensemble spread from the control forecast only.

Motivation & Objective

• To reduce the computational cost of ensemble weather prediction by replacing multiple NWP runs with a single deterministic forecast.
• To learn the statistical spread of ensemble forecasts directly from the deterministic control run using deep learning.
• To improve forecast uncertainty quantification by generating accurate, high-resolution ensemble spread estimates without running full ensemble systems.
• To investigate whether deep learning can effectively capture atmospheric variability and dispersion patterns from limited operational ensemble data.
• To explore the potential of cGANs as a surrogate for traditional ensemble prediction systems in operational meteorology.

Proposed method

• Adapt the 2D pix2pix conditional GAN architecture to 3D spatial-temporal data for paired video-to-video translation of deterministic forecasts to ensemble spread.
• Train the 3D pix2pix3D model on 10 years (2010–2020) of operational ECMWF IFS data at 0.5° resolution, using 500 hPa geopotential height as input and ensemble spread as output.
• Implement a shared latent space encoder–decoder architecture as a secondary model to compare performance and improve generalization.
• Apply Monte-Carlo dropout and model averaging across multiple independently trained networks to enhance robustness and reduce oscillatory behavior.
• Use a conditional GAN framework with a U-Net-based generator and a 3D PatchGAN discriminator to ensure spatial and temporal consistency in predicted spread.
• Optimize the model using adversarial loss, L1 loss, and perceptual similarity (SSIM) to balance fidelity and realism in the generated ensemble spread.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on deterministic forecasts alone accurately reconstruct the statistical spread of an operational ensemble system?
• RQ2 How does the performance of a 3D conditional GAN compare to traditional baselines like climatology and persistence in predicting ensemble spread?
• RQ3 Does model averaging or Monte-Carlo dropout improve the reliability and accuracy of the predicted ensemble spread?
• RQ4 To what extent does the model capture true atmospheric variability, especially given that real ensemble systems are often under-dispersive?
• RQ5 Can a single deterministic forecast generate a realistic, high-resolution ensemble spread that matches operational standards without running multiple simulations?

Key findings

• The 3D pix2pix3D model achieved a test set RMSE of 39.850 and mean SSIM of 0.789, significantly outperforming climatology (RMSE: 51.747, SSIM: 0.722) and persistence (RMSE: 53.184, SSIM: 0.700).
• Model averaging across multiple independently trained networks reduced the RMSE to 35.075 and increased the SSIM to 0.831, demonstrating improved robustness and reduced oscillatory behavior.
• The use of Monte-Carlo dropout further improved performance, yielding an RMSE of 38.773 and SSIM of 0.807, indicating better uncertainty quantification.
• The predicted ensemble spread was slightly larger than the ground truth, suggesting potential for better capture of true atmospheric variability, which is often under-estimated in operational systems.
• Visual comparisons across cities (Oslo, Seattle, Cape Town) confirmed that the pix2pix3D model produced more realistic and dynamically consistent spread patterns than baselines, especially at later forecast steps.
• The results indicate that deep learning-based models can effectively learn complex spatiotemporal statistical dependencies in atmospheric uncertainty from deterministic forecasts alone.