HR-Extreme: A High-Resolution Dataset for Extreme Weather Forecasting

TL;DR

This work introduces HR-Extreme, a high-resolution extreme-weather dataset derived from NOAA HRRR data at 3-km resolution to address the underrepresentation of extreme events in forecasting benchmarks. It pairs the dataset with HR-Heim, an improved baseline model that leverages a shared backbone and multi-head decoders to predict multiple physical variables, achieving superior performance on both general and extreme-weather forecasts. The authors show that extreme-weather errors are substantially larger than typical forecast errors, underscoring the need for dedicated extreme-event modeling and evaluation. By providing data, interfaces, and a robust evaluation protocol, HR-Extreme aims to drive development of models that reliably forecast extreme events and improve practical disaster preparedness.

Abstract

The application of large deep learning models in weather forecasting has led to significant advancements in the field, including higher-resolution forecasting and extended prediction periods exemplified by models such as Pangu and Fuxi. Despite these successes, previous research has largely been characterized by the neglect of extreme weather events, and the availability of datasets specifically curated for such events remains limited. Given the critical importance of accurately forecasting extreme weather, this study introduces a comprehensive dataset that incorporates high-resolution extreme weather cases derived from the High-Resolution Rapid Refresh (HRRR) data, a 3-km real-time dataset provided by NOAA. We also evaluate the current state-of-the-art deep learning models and Numerical Weather Prediction (NWP) systems on HR-Extreme, and provide a improved baseline deep learning model called HR-Heim which has superior performance on both general loss and HR-Extreme compared to others. Our results reveal that the errors of extreme weather cases are significantly larger than overall forecast error, highlighting them as an crucial source of loss in weather prediction. These findings underscore the necessity for future research to focus on improving the accuracy of extreme weather forecasts to enhance their practical utility.

Figures (5)

• Figure 1: The performance of clustering by KMeans (right) and DBSCAN (left). It can be seen that the results of DBSCAN are more accurate and reliable, and the approach also identifies noisy points effectively.
• Figure 2: Event area are cropped and covered by a series of squares considering the fixed size fed to the neural network, masks are provided to ensure only event area are calculated.
• Figure 3: The comparison of each variable on original test set and HR-Extreme evaluated on four models, where the blue areas are normalized RMSE on original test set and green areas are normalized RMSE on HR-Extreme.
• Figure 4: The mean error heatmap of all variables on entire U.S for four mdoels.
• Figure 5: Case studies for physical variables in normal and extreme cases in the same area.