 |
|
 |
|
|
Sub-Seasonal Climate Forecast Rodeo Symposium A subseasonal machine learning forecast model for the Forecast Rodeo |
|
|
June 17, 2019 The Sub-Seasonal Climate Forecast Rodeo Symposium was held in NOAA Auditorium, Silver Spring, MD. Dr. Judah Cohen of Team StillLearning, one of the three winning teams, gave a talk about their success in developing a machine learning system to improve subseasonal forecasts of temperature and precipitation over the western U.S. The system is an ensemble of two nonlinear regression models: a local linear regression model with multitask feature selection (MultiLLR) and a weighted local autoregression enhanced with multitask k-nearest neighbor features (AutoKNN). The MultiLLR model introduces candidate regressors from each data source in the Subseasonal Rodeo dataset and then prunes irrelevant predictors using a multitask backward stepwise criterion designed for the forecasting skill objective. The AutoKNN model extracts features only from the target variable (temperature or precipitation), combining lagged measurements with a skill-specific form of nearest-neighbor modeling. The results were encouraging. Each model alone was more accurate than the debiased operational U.S. Climate Forecasting System (CFSv2), and the ensemble skill exceeded that of the top Rodeo competitor. Moreover, over 2011-2018, an ensemble of two regression models and debiased CFSv2 improved debiased CFSv2 skill by 20-38% for temperature and 120-146% for precipitation. In discussions of the next step, Team StillLearning plan to extend the forecasts to the contiguous U.S. For more information, see https://arxiv.org/pdf/1809.07394.pdf. The symposium also honored Salient and Lupoa13 two teams for their outstanding work to improve week 3-4 and 4-6 forecasts, namely “connecting terrestrial rainfall prediction with sea surface temperature and salinity variability” conducted by Salient and “constructing subseasonal forecast using a combination of contingency and analog techniques as well as a crude form of artificial intelligence” pursued by Lupoa13. The next round Rodeo II was announced by Bureau of Reclamationation, Department of the Interior, and will kick off on July 19th, 2019. |
|
|
North American Multi-Model Ensemble Teleconference Verifying the NMME: 2012 – 2018 June 6, 2019 The North American Multi-Model Ensemble (NMME) forecast products have been widely used for decision making since 2011. The quality of the prediction is an important focus for service improvement. Dr. Sarah Strazzo of Climate Prediction Center conducted a study on four types of NMME seasonal forecasts (uncalibrated1, Probability Anomaly Correlation (PAC) calibrated (van den Dool et al. 2017), “Calibration, Bridging, and Merging” (CBaM) post-processed (Strazzo et al. 2018), and “Bridged”2 NMME forecasts) over the period of 2012-2018 and presented her findings in the June teleconference. Her results showed overall more hits and correct negatives than false positives and misses from contingency table of verification. The observation of more above than below normal temperatures was captured by uncalibrated, PAC calibrated, and CBaM post-processed forecasts, among which uncalibrated and PAC calibrated forecasts over-predicted above normal temperatures while CBaM post-processed and “Bridged” over-predicted below normal temperatures. Moreover, PAC calibrated precipitation forecasts achieved impressive Heidke skill scores, and NMME forecasts, particularly the post-processed versions, tended to outperform ENSO-derived forecasts in general. Her results further showed “hits” also increased as predicted probabilities increased, a good indication of “forecast opportunity” potential. In her summary, Dr. Strazzo shared with audience her promising thoughts to leverage other predictability sources, specifically to improve the prediction of winter below normal temperatures. 1) NMME includes CFSv2, CanCM3, CanCM4, GFDL-CM2.1, GFDL-FLOR, NASA-GEOS, and NCAR-RSMAS-CCSM4 participant models. Probabilities are calculated as ensemble frequencies relative to model mean terciles. 2) Uses bridged forecasts (1-month lead) as a stand-in for empirical ENSO-derived forecasts. |
|
|
References Strazzo, S., D. Collins, A. Schepen, Q. J. Wang, E. Becker, and L. Jia, 2018: Seasonal prediction of North American temperature and precipitation using the Calibration, Bridging, and Merging (CBaM) method. NWS Sci. Technol. Infusion Clim. Bull., Climate Prediction S&T Digest, 42nd NOAA Annu. Clim. Diagn. Predict. Workshop, Norman, OK, National Oceanic and Atmospheric Administration, 177-180. van den Dool, H., E. Becker, L.-C. Chen, and Q. Zhang, 2017: The probability anomaly correlation and calibration of probabilistic forecasts. Wea. Forecasting, 199-206. |
* The 43rd NOAA CDPW Digest has been published online at |
