Research map / Time series / Causal and generative time series
HR-Extreme: A High-Resolution Dataset for Extreme Weather Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces HR-Extreme, a 3-km HRRR-based dataset of 17 extreme weather event types over the US, benchmarks NWP, Pangu, FuXi and a new HR-Heim model, and shows extreme-event errors far exceed general errors.
Problem
Deep learning and NWP weather forecasters are rarely evaluated on extreme weather, and existing datasets lack high resolution, event-type diversity, and precise location/time labels.
Approach
Builds a high-resolution (3 km, 320x320 crops, 69 variables) extreme-event dataset from HRRR by combining NOAA Storm Events, Storm Prediction Center reports clustered with DBSCAN, and manually filtered extreme temperatures, then proposes the HR-Heim baseline (causal-conv encoder, Swin transformer, progressive upsampling decoder) to forecast one hour ahead.
Key results
- On HR-Extreme, HR-Heim achieves RMSE 1.60 vs 3.27 for NWP and 6.10 for FuXi and 10.42 for Pangu, with the smallest relative increase (34.3%) over its general RMSE of 1.40.
- Pangu's RMSE rises from 2.77 on the original test set to 10.42 on HR-Extreme (394% mean per-variable increase), showing extreme events strongly degrade deep learning forecasts.
- HR-Heim keeps a lower RMSE than NWP at all lead times 1-4 on HR-Extreme (e.g., 2.86 vs 5.20 at lead time 3).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HR-Heim instead of Pangu | Weather forecasting on HR-Extreme · HR-Extreme | RMSE: 10.42 -> 1.60 | Helped |
| Use HR-Heim instead of FuXi | Weather forecasting on HR-Extreme · HR-Extreme | RMSE: 6.10 -> 1.60 | Helped |
| Use HR-Heim instead of NWP | Weather forecasting on HR-Extreme · HR-Extreme | RMSE: 3.27 -> 1.60 | Helped |
| Evaluate on extreme cases rather than the original test set | Weather forecasting models · Pangu | RMSE: 2.77 -> 10.42 | Hurt |
| Evaluate on extreme cases rather than the original test set | Weather forecasting models · FuXi | RMSE: 2.39 -> 6.10 | Hurt |
| Evaluate on extreme cases rather than the original test set | Weather forecasting models · NWP | RMSE: 2.35 -> 3.27 | Hurt |
| Evaluate on extreme cases rather than the original test set | Weather forecasting models · HR-Heim | RMSE: 1.40 -> 1.60 | Hurt |
| Use HR-Heim instead of NWP | Weather forecasting on HR-Extreme across lead times 1–4 · HR-Extreme | RMSE: 3.27 -> 1.60; 4.12 -> 2.29; 5.20 -> 2.86; 6.04 -> 3.43 | Helped |
Surprising findings
- Pangu's error on extreme events increases roughly fourfold relative to its general error, and NWP is more stable on extremes than the deep learning baselines Pangu and FuXi.
- HR-Heim is not clearly better than NWP at 50 hPa humidity and some wind/temperature variables on the original set, though it wins on nearly all variables overall.
Limitations
- Events are labeled largely from user reports, which introduces uncertainty in range and timing and may omit events in sparse regions such as mountains, oceans and deserts.
- Tropical cyclones and other large-scale phenomena are not covered.
- HR-Heim is only trained and evaluated for single-hour (one-step) prediction.
Experimental setup
Datasets: HRRR NOAA Storm Events Database NOAA Storm Prediction Center reports ERA5
Benchmarks: HR-Extreme HRRR original test set (July-December 2020)
Metrics: RMSE normalized RMSE
Baselines: NWP (WRF-ARW in HRRRv4) Pangu-Weather FuXi
Builds on: FuXi MagViTv2 Pangu-Weather Swin Transformer V2
Closest papers by meaning
- ExtremeWeather: A large-scale climate dataset for semi-supervised detection, localization, and understanding of extreme weather events (NeurIPS 2017 · similarity 0.88)
- RainPro-8: An Efficient Deep Learning Model to Estimate Rainfall Probabilities Over 8 Hours (ICLR 2026 · similarity 0.86)
- OneForecast: A Universal Framework for Global and Regional Weather Forecasting (ICML 2025 · similarity 0.84)
- CasCast: Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling (ICML 2024 · similarity 0.84)
- EWMoE: An Effective Model for Global Weather Forecasting with Mixture-of-Experts (AAAI 2025 · similarity 0.84)
- Quantile-Regression-Ensemble: A Deep Learning Algorithm for Downscaling Extreme Precipitation (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.