Authors: Isidora Jankov, Daniel Abdi, Paul Madden, Vanderlei Vargas, Tim Smith, Sergey Frolov, Monte Flora, and Corey Potvin
NOAA’s pioneering High Resolution Rapid Refresh (HRRR) short-term weather forecast model will soon be joined by an experimental, AI-powered sibling.
Last month, NOAA’s Global Systems Laboratory (GSL) released HRRR-Cast, a data-driven model trained on three years of HRRR analysis data, for testing by NOAA’s National Weather Service (NWS). The HRRR has served as the NWS’s flagship operational short-term forecast model for the past decade.
HRRR-Cast output showing simulated reflectivity and wind particles over the contiguous-U.S. domain for a line of strong thunderstorms that occurred May 6, 2024. This image was created using GSL’s DESI tool. Image credit: NOAA/GSL
HRRR-Cast is NOAA’s first regional experimental AI forecast system and a key component of NOAA’s broader Project EAGLE, a long-term project to provide NOAA and the U.S. Weather Enterprise with the ability to rapidly test, develop, and identify the most promising AI models for global to regional ensemble forecasting.
HRRR-Cast emerged from a collaborative effort within NOAA Research led by GSL. GSL created the HRRR model, which is a widely used short-term operational weather model. GSL scientists recognized that the enormous volume of high quality data generated by the HRRR would be an ideal training dataset for an AI-powered high-resolution model, and that they could harness AI capabilities to improve the HRRR system. From there, HRRR-Cast was born.
HRRR-Cast was initially conceived as an emulator, aiming to match the skill of HRRR. It was trained on three years of HRRR analysis data (2021-2024), a period during which the HRRR version remained consistent.
This model is based on the ResNet architecture, specifically ResHRRR, which utilizes convolutional neural networks enhanced with squeeze-and-excitation blocks and Feature-wise Linear Modulation. It supports probabilistic forecasting through the Denoising Diffusion Implicit Model (DDIM).
To improve performance at longer lead times, a single model is trained to predict multiple lead times (1h, 3h, and 6h), and a greedy rollout strategy is employed during inference. For inferences, HRRR analysis is used for initial conditions.
HRRR-Cast advances the pioneering StormCast by:
- Training on the full CONUS domain.
- Training on multiple lead times to enhance long-range performance.
- Utilizing analysis data for training, unlike the +1h post-analysis data inadvertently used in StormCast.
- Incorporating future Global Forecast System (GFS) weather states as inputs.
- Adding a downscaling component that significantly improves longer-lead forecast accuracy.
Early evaluations are promising, GSL scientists say. Regarding one challenging variable, reflectivity forecasts, HRRR-Cast performs at least as well as the operational HRRR up to seven hours, with comparable performance up to 48 hours. Initial results show that HRRR-Cast also performs comparably to the operational HRRR in forecasting humidity, temperature, and wind, and excels in producing realistic depictions of storm structure.
HRRR-Cast is 100 to 1000 times more computationally efficient than the operational HRRR, which allows HRRR-Cast to be run without the need for a supercomputer. More specific information regarding HRRR-Cast can be found in this preprint.
The product of grassroots research in NOAA’s Office of Oceanic and Atmospheric Research (OAR), HRRR-Cast was boosted by support from GSL and the NOAA Office of the Chief Information Officer’s Scientific Engineering and Novel Architecture program. Important contributions have come from scientists in NOAA’s Physical Sciences Laboratory (PSL) and National Severe Storms Laboratory (NSSL), the cooperative institutes Cooperative Institute for Research In Environmental Sciences (CIRES) and Cooperative Institute for Research in the Atmosphere (CIRA), and OAR’s Weather Program Office (WPO). HRRR-Cast’s integration with Project EAGLE will be done in coordination with the Earth Prediction Innovation Center (EPIC).
