Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Power Ensemble Aggregation for Improved Extreme Event AI Prediction

Julien Collard, Pierre Gentine, Tian Zheng

TL;DR

This work tackles predicting extreme heat events by defining extremes through the local $q$-th quantile and casting prediction as a binary classification evaluated by AUC. It introduces an adaptive power-mean ensemble aggregation for a generative model that outputs an ensemble of local temperature anomalies, with per-example scores transformed via $ ext{Phi}(ullet)$ and aggregated as $ig( rac{1}{n} ext{sum}( ext{Phi}(hat{x}_i)^p)ig)^{1/p}$, where $p$ is tuned per quantile $q$. The model is built on a light CubeSphere CNN (U-Net) augmented with Perlin noise to produce an ensemble of $n=50$ members and is trained with CRPS loss on ERA5-derived data, achieving higher AUC than the mean ensemble across all $q$ and lead times, with optimal $p$ growing roughly exponentially with $q$. The results show notable improvements for high-extremity prediction and even competitive performance against GraphCast for certain high-quantile, longer-lead scenarios, illustrating the practical potential of adaptive ensemble aggregation for climate extremes.

Abstract

This paper addresses the critical challenge of improving predictions of climate extreme events, specifically heat waves, using machine learning methods. Our work is framed as a classification problem in which we try to predict whether surface air temperature will exceed its q-th local quantile within a specified timeframe. Our key finding is that aggregating ensemble predictions using a power mean significantly enhances the classifier's performance. By making a machine-learning based weather forecasting model generative and applying this non-linear aggregation method, we achieve better accuracy in predicting extreme heat events than with the typical mean prediction from the same model. Our power aggregation method shows promise and adaptability, as its optimal performance varies with the quantile threshold chosen, demonstrating increased effectiveness for higher extremes prediction.

Power Ensemble Aggregation for Improved Extreme Event AI Prediction

TL;DR

This work tackles predicting extreme heat events by defining extremes through the local -th quantile and casting prediction as a binary classification evaluated by AUC. It introduces an adaptive power-mean ensemble aggregation for a generative model that outputs an ensemble of local temperature anomalies, with per-example scores transformed via and aggregated as , where is tuned per quantile . The model is built on a light CubeSphere CNN (U-Net) augmented with Perlin noise to produce an ensemble of members and is trained with CRPS loss on ERA5-derived data, achieving higher AUC than the mean ensemble across all and lead times, with optimal growing roughly exponentially with . The results show notable improvements for high-extremity prediction and even competitive performance against GraphCast for certain high-quantile, longer-lead scenarios, illustrating the practical potential of adaptive ensemble aggregation for climate extremes.

Abstract

This paper addresses the critical challenge of improving predictions of climate extreme events, specifically heat waves, using machine learning methods. Our work is framed as a classification problem in which we try to predict whether surface air temperature will exceed its q-th local quantile within a specified timeframe. Our key finding is that aggregating ensemble predictions using a power mean significantly enhances the classifier's performance. By making a machine-learning based weather forecasting model generative and applying this non-linear aggregation method, we achieve better accuracy in predicting extreme heat events than with the typical mean prediction from the same model. Our power aggregation method shows promise and adaptability, as its optimal performance varies with the quantile threshold chosen, demonstrating increased effectiveness for higher extremes prediction.

Paper Structure

This paper contains 20 sections, 7 equations, 11 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Data
  4. Defining extreme events
  5. Classification problem
  6. Aggregation methods
  7. Model architecture
  8. Results
  9. Optimizing the power exponent
  10. Final performance
  11. Conclusion
  12. Data and grid system
  13. Model architecture
  14. Deterministic baseline
  15. Generative adaptation
  16. ...and 5 more sections

Figures (11)

  • Figure 1: Aggregation method flowchart
  • Figure 2: $log(p_{opt})=f(q)$
  • Figure 3: $\text{RI}_{opt}=f(q)$ on test dataset
  • Figure 4: AUC as a function of the forecast lead time for different quantiles
  • Figure 5: Cube Sphere Grid
  • ...and 6 more figures