Experimental Subseasonal Precipitation Accumulation Outlooks
Data and Documentation
These experimental subseasonal outlooks are provided to demonstrate near real-time performance of conventional and deep-learning-based methods to calibrate raw precipitation accumulation forecasts from NOAA’s Global Ensemble Forecast System version 12 (GEFSv12, see Hamill et al. 2022, Guan et al. 2022).
About These Products
To improve guidance of 7-day and 14-day accumulations of precipitation over the contiguous United States, we developed a novel, experimental deep-learning-based based post-processing algorithm to remove inherent systematic biases and ensemble-spread errors of the GEFSv12 ensemble. We also provide probabilistic guidance for a more established and conventional post-processing technique which is based on distributional regression. This webpage shows experimental guidance for week 1 (12-180 h; 0.5-7.5 days) lead times, week 2 (180-348 h; 7.5-14.5 days) lead times, and combined weeks 3 and 4 (348-684 h; 14.5-28.5 days lead times) as well as verification metrics of the real time outlooks and of hindcasts (2000-2019).
Outlooks for each forecast method (described below) show the probability of accumulated precipitation being above or below normal where normal is defined as the median (50th) percentile of observed (PRISM) climatology. Darker green and brown shades indicate higher probabilities of the event. White shaded regions indicate that there was not a >=55% chance of either category.
Method
Raw model: This is the raw model output from the GEFSv12 accumulated over 7 or 14 days.
CSGD (censored-shifted gamma distribution) model: This is a state-of-the-art distributional regression post-processing technique that is used as a baseline for the new deep-learning-based technique (i.e., the RUFCO model). See Scheuerer and Hamill 2015. for more details.
RUFCO (ResUnet, FiLM, and Climatological-Offramp) model: This is a novel neural-network algorithm tailored for subseasonal prediction. The base of the model is made up of three main components: 1) a ResUnet which learns relationships between observed accumulated precipitation and GEFSv12 predictors of weather and geographical variables, 2) a Feature-wise Linear Modulation (FiLM) layer which embeds time-of-year into the network, and 3) a Climatological-Offramp which reverts forecasts towards climatology in situations when the network learns that the predictors are not able to provide enough information to generate a skillful forecast.
For more technical descriptions of the methods, a forthcoming paper will include that information.
Data Source
Total precipitation is used as a predictor to the CSGD model and total precipitation, latitude, longitude, and elevation are used as predictors to the RUFCO model. All realtime predictor variables are from the 00Z initialization of the operational version of GEFSv12 downloaded here from NOAA Operational Model Archive and Distribution System (NOMADS). Parameter-elevation Relationships on Independent Slopes Model (PRISM) gridded precipitation analyses serve as the verification observations and can be downloaded here. All analyses were performed at 0.5 degree rectilinear horizontal grid spacing.Hindcast performance was calculated for past forecasts produced for 20 cross-validated years (2000-2019) of GEFSv12 reforecasts (forecasts retrospectively generated from the same model version of NOAA’s Unified Forecast System) and verified against corresponding PRISM precipitation analyses. GEFSv12 reforecasts can be downloaded here.
Hindcast performance metrics
Spatial maps of average ranked probability skill scores (RPSS) use the 0.33, 0.50, 0.67, and 0.85th percentiles of PRISM climatology. A climatological forecast is used as the reference forecast. The climatological forecast, separately for each grid point and date, is made up of observed (PRISM) precipitation accumulations that fall within +/-24 days centered around the date of interest, during years 2000-2019. This metric provides an overall assessment of the calibration and sharpness of the forecasts. Positive values indicate a more skillful forecast than the reference climatological forecast. Zero indicates the same skill while negative values indicate worse performance compared to a climatological forecast. Data for the skill score maps were pooled over 20 cross-validated years (2000-2019), and for forecasts initialized in each season that were valid for the specified lead time (week 1, week 2, and weeks 3-4).
Reliability diagrams (RD) and corresponding Brier skill scores (BSS) and inset “relative frequency of occurrence histograms” were created for above (>50th percentile of climatology) and below <= 50th percentile of climatology) normal events. Raw, CSGD, and RUFCO forecast events are in reference to the 50th percentile of PRISM climatology while the “bias-corrected” forecast is a simple model calibration of the raw forecast that is in reference to the 50th percentile of GEFSv12 model climatology. Data for the RDs were pooled over all grid points within the contiguous U.S., 20 cross-validated years (2000-2019), and for forecacsts initialized in each season that were valid for the specified lead time (week 1, week 2, and weeks 3-4). Confidence intervals show the 5th and 95th percentiles of a bootstrapped, with replacement, distribution.
Referencing Forecasts
To reference forecast plots, we ask that you acknowledge PSL as in ”image is provided by the NOAA Physical Sciences Laboratory, Boulder, Colorado, USA, from their website at https://psl.noaa.gov/”. You should also reference the publication:
Webpage contributors
- Rochelle Worsnop developed code to run the forecast algorithms, to produce forecast outlooks, and to calculate verification metrics (once available) in real-time.
- Lesley Smith retrieves real-time GEFSv12 forecasts and PRISM data input to algorithms.
- Don Hooper and Cathy Smith created the webpage and run the webpage in real-time.
Note
These forecasts are experimental. NOAA/PSL is not responsible for any loss occasioned by the use of these forecasts.