#RRFS #REFS
Ongoing Development of the RRFS
Posted Friday 12/06/24 @ 5:24 PM — RRFS parallel model runs have been halted for planned model evaluation and consideration for release. The model output has been hosted on AWS. I’m wondering if they’re going to make this model operational. It’s had some issues and my guess is that it will be resource intensive for NOAA, especially if the RRFS accuracy issues won’t allow the planned retirement of the NAM and the HIRES-W models that the RRFS was supposed to replace.
Posted Thursday 11/27/24 @ 2:04 PM — The RRFS continues to be under development, but I haven’t seen an updated release date for the model to become operational. I’m increasingly using for my forecasts, despite some shortcomings. The REFS, which uses time-lagged runs of the HRRR and RRFS to create a model group or “ensemble” is also increasingly useful.
Based on a recent journal article, the RRFS release may be further delayed.
Shortcomings of the RRFS revealed in Severe Storms Experiment
Excerpt of article below, published 11/1/24—
In detailed blind (i.e., model labels were hidden) subjective and objective assessments, significant issues were identified in RRFS that degraded forecast quality. While the current rapidly updating operational CAM, the High-Resolution Rapid Refresh (HRRR) model, improved relative to the previous year because of a weather regime associated with more predictable severe weather (i.e., strong vertical wind shear with large instability), the RRFS forecasts were actually worse in 2024, which led to a greater performance gap relative to 2023. It was found that efforts to reduce the high precipitation bias in RRFS by implementing the Grell–Freitas (GF) convective parameterization caused many high-impact severe weather events to be missed by the RRFS. In these cases, the composite reflectivity fields were often characterized by spurious, low-intensity precipitation that seemed to preclude more realistic storm development. Interestingly, the REFS ensemble results did not reveal forecast degradation and subjective ratings showed performance on par with the High-Resolution Ensemble Forecast System version 3 (HREFv3), the current operational CAM ensemble. Upon closer examination of individual REFS members, it was found that performance was closely tied to the specific convective parameterization scheme, with members using the scale-aware Simplified Arakawa–Schubert (saSAS) convective scheme performing much better than those using GF, so that the saSAS members positively impacted ensemble performance. These results have led NOAA’s Environmental Modeling Center to reconsider RRFS/REFS configurations, which will likely delay operational implementation while feasible options to improve performance are explored.
SFE 2024 marked the second consecutive year of tests with MPAS, which were expanded relative to 2023. Similar to SFE 2023, MPAS results for day 1 and 2 lead times were very encouraging. In almost every deterministic comparison, MPAS performance was runner-up to the HRRR and superior to every Finite-Volume Cubed-Sphere (FV3)–based system evaluated. MPAS was also tested within an alternative REFS configuration. While a postprocessing issue prevented meaningful subjective results, objective measures showed that this MPAS REFS ensemble configuration performed similarly to HREFv3 and two other alternative REFS configurations that were considered.
New Update to the RRFS and REFS
Posted Friday 08/16/24 @ 9:53 AM — Yesterday, the RRFS model development group announced a change to the convection parameterization (thunderstorm type precip) of the current experimental RRFS.
They have switched the experimental RRFS and the REFS to a scale-aware SAS deep convection scheme (a somewhat older physics package) from the newer Grell-Freitas (GF) deep convection scheme used by the HRRR. Perhaps the GF convection scheme is not working well with the FV3 core of the RRFS?
As recent verification research indicates, the convection forecasts for the current experimental RRFS are over-done in coverage and intensity. My guess is the switch is motivated by a desire to correct for these excessive forecasts of convective precipitation.
I hope they are also working on the accuracy of the 2m max temperatures which remains about 6º too cool in Philadelphia compared to the HRRR. So far, I don’t see any difference in the current development RRFS-A and their “testbed” RRFS-A temperature forecasts regarding 2m temperature.
Additional concerns about the RRFS and REFS
Posted Friday 08/09/24 @ 3:47 PM — I’ve continued using the RRFS for forecasts. I’ve noticed that the 12z run is perhaps the most accurate of the RRFS runs regarding precipitation. This is something their own model verification pages show.
Today, with the remnants of tropical storm Debby, the 12z RRFS captured the line of showers and storms moving through about 4 PM much better than the 12z HRRR.
Unfortunately, the 12z RRFS first 18 hours forecast isn’t available until after 11:15 AM.
Posted Sunday 07/21/24 @ 9:28 PM — I continue to download the 60 hour REFS and RRFS runs (00z 06z 12z 18z) from AWS daily and I have really been trying to use this new model for my forecasts.
Here’s what I see: model forecast high temperatures for Philadelphia appear to be 6º-10ºF too cool during the past heat wave! That’s especially true for forecasts more than 24 hours out. That departure from the HRRR and NBM forecast temperature highs is unacceptable.
Regarding precipitation, I think the general amounts of precip accumulation are improving and not terribly over-done, but I find that the model seems to get the axis of heavy rainfall off quite a bit and usually the placement is too far southward. I rarely have seen the actual rainfall placement to be correct.
These are hardly statistical valid verified observations, but they’re my impressions based on forecasts for the Philadelphia region.
They’re up to RRFS_A version 0.91. I haven’t seen them update the version for awhile. They have a way to go.
Serious Issues with the RRFS
Posted Monday 03/25/24 @ 5:33 PM — While the RRFS model development (now in version 0.8.5) continues to progress, the lead scientists in model’s development have come to the recent realization that the FV3 core of the model may not allow it to reach the level of accuracy needed for proper prediction of convection (thunderstorms).
Specifically the current experimental RRFS over-predicts areal coverage of convective precipitation and over-predicts amount of precipitation, especially when forecast amounts are over an inch. It does not do as well as the HRRR or in some cases, the NAM-NEST for accurate forecast of thunderstorms.
These issues apparently can’t be solved due to inherent limitations of the FV3 dynamic core in this non-hydrostatic mesoscale model. (The FV3 core is successfully used in the current GFS model, a hydrostatic model.)
An extensive review and discussion of the experiments leading to this realization can be found in this paper published Dec 2023.
Unfortunately, years have already been spent in development using the FV3 and a change away from the FV3 core may be prohibitively difficult. It appears that the NOAA decision to standardize and unify model development using the FV3 core was a poor decision.
An alternate approach, using the MPAS (Model for Prediction Across Scales) is now planned for RRFS version 2.0. Techniques and parameterization of the RRFS version 1.0 will be applied during the development MPAS version RRFS 2.0.
More info here.
The MPAS uses a honeycomb geometry that can be compressed or stretched for greater detail in active weather regions.
Posted Sunday 12/17 @ 10:13 AM — The latest RRFS development page now has the model being operational in early 2025 instead of Fall 2024. Current version is 0.7.7. Updates here.
Posted November 18, 2023
NOAA has been developing a new model, the RRFS (“Rapid Refresh Forecast System“) which will be the beginning of a major advance in high resolution weather forecasting. The model will run hourly forecasts to 18 hours, will have a high (3 km) resolution, and its 00z, 06z, 12z, and 18z forecasts will extend out to 60 hours. It will also have an statistical ensemble component version and a “smoke” forecast version.
Its development was begun in 2016-2017 and early experimental versions have been posted online starting 2021.
The RRFS takes much of its physics from the HRRR and will use the advanced FV3 core used in the recently updated GFS.
The RRFS is going through a series of continued experimentation, testing and fine tuning and the groundwork is being laid for it to be operational sometime around September 2024.
Some of the current ongoing development information is being logged here. When released, it will be RRFS version 1.0, but as of this posting (Nov 2023) , it’s only at version 0.7.3 and is being referred to as “RRFS-A“.
About 6 months after it is released, many of the models I look at regularly ( the HRRR, NAM, NAM-NEST, RAP, SREF, and all three flavors of the HIRESW) will cease operation.
My long-standing mantra when it comes to snowstorms, “Never Ignore the NAM”, will become a relic of the past.
I’ve always felt there are “too many models” and it appears that the RRFS is incorporating and building on all of the best approaches to numerical weather modeling that have been learned from these earlier models.
NOAA/NWS has been making the development version and their testbed variants available on a NOAA- Amazon Web Services site (AWS) for since 2021.
In recent weeks, I’ve been experimenting with downloading these RRFS runs. The experimental version is hosted on Amazon Web Services (AWS) NOAA-Open Data and has certain constraints that the main NOAA NOMADS site doesn’t have, along with different directory structures (buckets). Nonetheless, I hope to be downloading and using the RRFS model in some forecasts soon.
The release of RRFS version 1 will truly be a big deal.
For those of you that wish to explore more, here are some excellent web links related to the RRFS—
https://www.emc.ncep.noaa.gov/users/emc.campara/rrfs/
https://ams.confex.com/ams/WAFNWPMS/meetingapp.cgi/Session/65462
https://ams.confex.com/ams/101ANNUAL/meetingapp.cgi/Paper/381608
https://registry.opendata.aws/noaa-rrfs/
https://doi.org/10.1029/2021MS002483
https://epic.noaa.gov/wp-content/uploads/2022/07/1.-SRW-Carley.pdf