Improving Predictions of Convective Storm Wind Gusts through Statistical Post-Processing of Neural Weather Models
Antoine Leclerc, Erwan Koch, Monika Feldmann, Daniele Nerini, Tom Beucler · 2025
Read Improving Predictions of Convective Storm Wind Gusts through Statistical Post-Processing of Neural Weather Models by Antoine Leclerc, Erwan Koch, Monika Feldmann, Daniele Nerini, Tom Beucler (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-00128.txt.
Introduction
Timely warnings for storm gusts save lives and property, but gusts are hard to forecast at a single place. Neural weather models now predict the broad atmosphere cheaply on a global grid. The five authors post-process the Pangu-Weather model's output to predict hourly gust distributions up to three days ahead over five Swiss regions. They test a hierarchy of statistical and deep learning methods and pin the forecasts with generalized extreme value distributions.
A convolutional network that reads the spatial pattern of the predicted environment did best, beating direct forecasts at every lead time and gust speed. The authors conclude that neural weather models add real value for extreme wind forecasting and could power more responsive early warning systems.
Catalog record
| CREATOR | Antoine Leclerc, Erwan Koch, Monika Feldmann, Daniele Nerini, Tom Beucler |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-00128
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Improving Predictions of Convective Storm Wind Gusts through Statistical Post-Processing of Neural Weather Models Antoine Leclerc, Erwan Koch, Monika Feldmann, Daniele Nerini, Tom Beucler Abstract Issuing timely severe weather warnings helps mitigate potentially disastrous consequences. Recent advancements in Neural Weather Models (NWMs) offer a computationally inexpensive and fast approach for forecasting atmospheric environments on a 0.25 ∘ global grid. For thunderstorms, these environments can be empirically post-processed to predict wind gust distributions at specific locations. With the Pangu-Weather NWM, we apply a hierarchy of statistical and deep learning post-processing methods to forecast hourly wind gusts up to three days ahead. To ensure statistical robustness, we constrain our probabilistic forecasts using generalised extreme-value distributions across five regions in Switzerland. Using a convolutional neural network to post-process the predicted atmospheric environment’s spatial patterns yields the best results, outperforming direct forecasting approaches across lead times and wind gust speeds. Our results confirm the added value of NWMs for extreme wind forecasting, especially for designing more responsive early-warning systems. 1 Introduction Figure 1 : Pangu-Weather forecasts for a day with convective conditions are clipped to Switzerland and post-processed using one of our different models. The best results are obtained with the CNN, which takes entire maps of features as input. Five different generalised extreme-value distributions are created, one for each cluster. This example shows predictions for June 2 nd , 2021 at 17:00 with a lead time of 27 h. Accuracy varies with the prediction (e.g., forecast for Cluster 2 is less accurate than for Cluster 4). Severe winds can significantly damage buildings [ 1 , 2 , 3 , 4 , 5 ] and pose risks to human lives [ 6 , 7 , 8 ] . Responding to such extreme events requires providing public authorities with accurate and easily understandable forecasts in time to implement prevention measures [ 9 ] . Effective strategies rely on timely warnings, from long-term to short-term, and accurate medium-range forecasts (2–10 days) have proven particularly efficient when combined with impact forecasts [ 10 ] . Meanwhile, machine learning-based models — especially Neural Weather Models (NWMs) [ 11 ] — are emerging as a viable complement to traditional numerical weather prediction (NWP) models [ 12 ] for short to medium-range weather forecasts. For example, GraphCast [ 13 ] , Pangu-Weather [ 14 ] , FourCastNet [ 15 ] , and AIFS [ 16 ] can issue high-quality, deterministic and global forecasts with a 0.25 ∘ spatial resolution for medium range lead times (up to seven days). Once trained, these models can be run locally on a personal computer on a GPU in seconds. In contrast, physics-based NWP models solve non-linear partial differential equations using dedicated high-performance computing infrastructure. Although their outputs are generally made publicly available by national weather services and thus widely utilised, this complexity and cost prevent the general public from running them, hindering smaller weather prediction centres and hazard-issuing organisations. The use of NWMs for severe weather forecasting is rapidly evolving, from tropical cyclone tracking [ 14 ] to predicting local surface extremes [ 17 ] . However, forecasting thunderstorms — highly localised, short-lived, and inherently uncertain events — presents unique challenges. Although NWMs have typical spatial and temporal resolutions that are reasonable to forecast mesoscale environments conducive to severe weather [ 18 ] , they currently lack the resolution and variables (e.g., precipitation) needed to directly detect individual convective storms, which often last only a few hours and cover areas of just tens of square kilometres. To address this, we propose post-processing the atmospheric environments predicted by NWMs to forecast regional distributions of wind gusts (they can be seen as the distributions of possible wind gusts for a “representative” site within the region) rather than deriving the tracks of individual storms from forecasted fields. Although hail often remains the primary source of thunderstorm damage [ 19 ] , we here focus on wind gusts as they depend on more predictable factors such as mean wind, shear, and topography [ 20 ] , while still posing significant risks to infrastructure and public safety. There are various methods to post-process NWP models to forecast station-wise wind gusts. For example, Friederichs et al. [ 21 ] used Vector Generalised Additive Models (VGAMs) to predict station-wise Generalised Extreme-value (GEV) distributions of wind gusts and Vector Generalised Linear Models (VGLMs), Schulz et al. [ 22 ] reviewed different statistical post-processing techniques (multi-linear regression, machine learning, and neural networks) to post-process wind gust distributions, and Fontana et al. [ 23 ] post-processed deterministic wind gusts forecasts via a Convolutional Neural Network (CNN) to predict distributions. As the considered wind gusts are maxima over 1 hour of the 1-s sustained wind speed (i.e., maxima over 3600 observations), the GEV distribution, which is the asymptotic distribution of block maxima under appropriate conditions [ 24 ] , appears as a natural model for them [ 21 , 22 ] ; their suitability in our context is expounded through diagnostic plots and Kolmogorov–Smirnov tests in the methods section. However, to our knowledge, these methods have not yet been applied to NWMs to predict regional distributions of meteorological variables. Here, we post-process a NWM to issue probabilistic forecasts of regional hourly wind gusts using a hierarchy of empirical models (statistical and machine learning-based). We selected the Pangu-Weather NWM for this study because it is currently the only established NWM that generates medium-range forecasts with an hourly temporal resolution. Our study focuses on Switzerland, allowing us to leverage high-quality wind gust observations over complex topography to assess wind gust hazards under convectively active conditions. However, our framework can easily be adapted to other regions with station-based or gridded wind gust observations, enabling wind gust forecasts in areas where NWP models are not routinely run at the fine resolution needed to resolve extremes. 2 Results Figure 2 : For the specific task of predicting wind-gust GEV distributions over the five pre-defined Swiss meteorological regions, our post-processing models outperform direct forecasts, the ICON NWP reference, and the climatology of wind gusts across lead times, confirming the added value of Pangu-Weather (initialized with ERA5) in predicting atmospheric environments and the ability of statistically constrained, data-driven models to make reliable probabilistic wind gusts forecasts. Panels (a) and (b) show the mean CRPSS and its min/max over the five folds for different models as a function of lead time, on a more restricted ensemble for (a) to use the same timestamps as ICON. Panel (c) shows the mean CRPSS and its min/max over the five folds for different models and different lead times, comparing a direct forecast to a post-processing of a Pangu-Weather forecast. General performance Four post-processing models were trained on five years of Pangu-Weather forecasts (2016–2020) following a k-fold cross-validation scheme with each fold corresponding to one year in the training set. Each model predicts the location, scale, and shape parameters of five wind gust GEV distributions for each time step and each lead time, corresponding respectively to five clusters of stations from the Swiss Meteorological Network (SwissMetNet, [ 25 ] ). The clusters are pre-computed based on measurement correlation between stations (see Section 4 for more details). Performing a forward sequential feature selection with a VGLM led us