ExtremeCast: Boosting Extreme Value Prediction for Global Weather Forecast

TL;DR

ExtremeCast tackles the underprediction of extreme weather by combining EVT-informed asymmetric loss with a training-free ensemble booster, integrated into a cascaded diffusion-based global forecast at $0.25^{\circ}$ resolution. The Exloss loss reweights errors to counteract the bias of symmetric losses like MSE, while ExBooster expands forecast uncertainty through multiple random samples and rank-histogram aggregation. Empirical results on ERA5 data show state-of-the-art performance on extreme-value metrics (RQE, SEDI) without sacrificing overall RMSE. The work demonstrates a practical path to more reliable extreme-weather forecasts with implications for disaster risk management and climate resilience.

Abstract

Data-driven weather forecast based on machine learning (ML) has experienced rapid development and demonstrated superior performance in the global medium-range forecast compared to traditional physics-based dynamical models. However, most of these ML models struggle with accurately predicting extreme weather, which is related to training loss and the uncertainty of weather systems. Through mathematical analysis, we prove that the use of symmetric losses, such as the Mean Squared Error (MSE), leads to biased predictions and underestimation of extreme values. To address this issue, we introduce Exloss, a novel loss function that performs asymmetric optimization and highlights extreme values to obtain accurate extreme weather forecast. Beyond the evolution in training loss, we introduce a training-free extreme value enhancement module named ExBooster, which captures the uncertainty in prediction outcomes by employing multiple random samples, thereby increasing the hit rate of low-probability extreme events. Combined with an advanced global weather forecast model, extensive experiments show that our solution can achieve state-of-the-art performance in extreme weather prediction, while maintaining the overall forecast accuracy comparable to the top medium-range forecast models.

Figures (12)

• Figure 1: MSE loss vs. Exloss. Through theoretical analysis, we found that MSE loss will give underestimated extreme value predictions. An asymmetric loss function, Exloss, is designed to address this bias by balancing the data distribution.
• Figure 2: Model Framework. The model consists of three cascaded parts, namely the deterministic forecast model $M_d$, the generation model $M_g$ for enhancing extreme values, and the ExBooster module for modeling uncertainty.
• Figure 3: ExBooster. Multiple random samplings can simulate forecast uncertainty and enhance the accuracy of extreme event predictions. The visual example demonstrates that extreme values are more pronounced in the output.
• Figure 4: RQE ($RQE<0$ means underestimating extreme values, $RQE>0$ means overestimating extreme values). All results are tested on 2018 data and use ERA5 as the target.
• Figure 5: SEDI (the closer to 1 the better). ws10 represents surface wind speed, that is, $ws10=\sqrt{u10^2+v10^2}$. All results are tested on 2018 data and use ERA5 as the target.